update knowing framework and application from Tian_Chunyu

it is perfect

APP_Framework/Applications/connection_app/Kconfig

@@ -1,14 +1,3 @@
 menu "connection app"
-    menuconfig APPLICATION_CONNECTION
-        bool "Using connection apps"
-        default n
-
-    menuconfig CONNECTION_COMMUNICATION_ZIGBEE
-        bool "enable zigbee demo"
-        default n
-        select CONFIG_CONNECTION_COMMUNICATION_ZIGBEE
-        if CONNECTION_COMMUNICATION_ZIGBEE
-            source "$KERNEL_DIR/framework/connection/Adapter/zigbee/Kconfig"
-        endif
 
 endmenu

APP_Framework/Applications/knowing_app/Kconfig

@@ -1,7 +1,4 @@
 menu "knowing app"
-
+    source "$APP_DIR/Applications/knowing_app/mnist/Kconfig" 
-    menuconfig APPLICATION_KNOWING
+    source "$APP_DIR/Applications/knowing_app/face_detect/Kconfig" 
-        bool "Using knowing apps"
-        default n
-
 endmenu

APP_Framework/Applications/knowing_app/SConscript

@@ -0,0 +1,14 @@
+import os
+Import('RTT_ROOT')
+from building import *
+
+cwd = GetCurrentDir()
+objs = []
+list = os.listdir(cwd)
+
+for d in list:
+    path = os.path.join(cwd, d)
+    if os.path.isfile(os.path.join(path, 'SConscript')):
+        objs = objs + SConscript(os.path.join(path, 'SConscript'))
+
+Return('objs')

APP_Framework/Applications/knowing_app/face_detect/Kconfig

@@ -0,0 +1,7 @@
+config FACE_DETECT
+    bool "enable apps/face detect"
+    depends on BOARD_K210_EVB
+	depends on DRV_USING_OV2640
+	depends on USING_KPU_POSTPROCESSING
+	depends on USING_YOLOV2
+    default n

APP_Framework/Applications/knowing_app/face_detect/SConscript

@@ -0,0 +1,9 @@
+from building import *
+
+cwd     = GetCurrentDir()
+src     = Glob('*.c') + Glob('*.cpp')
+CPPPATH = [cwd]
+
+group = DefineGroup('Applications', src, depend = ['FACE_DETECT'], LOCAL_CPPPATH = CPPPATH)
+
+Return('group')

APP_Framework/Applications/knowing_app/face_detect/face_detect.c

@@ -0,0 +1,253 @@
+#include <transform.h>
+#include"region_layer.h"
+#define SHOW_RGB_BUF_SIZE (320*240*2)
+#define AI_KPU_RGB_BUF_SIZE (320*240*3)
+#define KMODEL_SIZE (388776)    //face model size
+#define ANCHOR_NUM 5
+#define KPUIMAGEWIDTH (320)
+#define KPUIMAGEHEIGHT (240)
+
+
+static float anchor[ANCHOR_NUM * 2] = {1.889,2.5245,  2.9465,3.94056, 3.99987,5.3658, 5.155437,6.92275, 6.718375,9.01025};
+
+
+#define THREAD_PRIORITY_FACE_D  (11)
+static pthread_t  facetid = 0;
+static void* thread_face_detcet_entry(void *parameter);
+static int g_fd = 0;
+static int kmodel_fd = 0;
+static int  if_exit = 0;
+static unsigned char  * showbuffer = NULL ;
+static unsigned char  * kpurgbbuffer = NULL ;
+
+static _ioctl_shoot_para shoot_para_t = {0};
+unsigned char  * model_data = NULL;     //kpu data  load memory
+unsigned char *model_data_align  = NULL;
+
+kpu_model_context_t face_detect_task;
+static region_layer_t face_detect_rl;
+static obj_info_t face_detect_info;
+volatile uint32_t g_ai_done_flag;
+
+static void ai_done(void *ctx)
+{
+    g_ai_done_flag = 1;
+}
+
+
+
+void face_detect()
+{
+    int ret = 0;
+    int result = 0;
+    int size = 0;
+    g_fd = open("/dev/ov2640",O_RDONLY);
+    if(g_fd < 0)
+    {
+        printf("open ov2640 fail !!");
+        return;
+    }
+    showbuffer = (unsigned char*)malloc(SHOW_RGB_BUF_SIZE);
+    if(NULL ==showbuffer)
+    {
+        close(g_fd);
+        printf("showbuffer apply memory fail !!");
+        return ;
+    }
+    kpurgbbuffer = (unsigned char*)malloc(AI_KPU_RGB_BUF_SIZE);
+    if(NULL ==kpurgbbuffer)
+    {
+        close(g_fd);
+        free(showbuffer);
+        printf("kpurgbbuffer apply memory fail !!");
+        return ;
+    }
+    model_data = (unsigned char *)malloc(KMODEL_SIZE + 255);
+    if(NULL ==model_data)
+    {
+        free(showbuffer);
+        free(kpurgbbuffer);
+        close(g_fd);
+        printf("model_data apply memory fail !!");
+        return ;
+    }
+    memset(model_data,0,KMODEL_SIZE + 255);
+    memset(showbuffer,0,SHOW_RGB_BUF_SIZE);
+    memset(kpurgbbuffer,0,AI_KPU_RGB_BUF_SIZE);
+    shoot_para_t.pdata = (unsigned int *)(showbuffer);
+    shoot_para_t.length = SHOW_RGB_BUF_SIZE;
+    /*
+        load memory 
+    */
+    kmodel_fd = open("/kmodel/detect.kmodel",O_RDONLY);
+    if(kmodel_fd <0)
+    {
+            printf("open kmodel fail");
+            close(g_fd);
+            free(showbuffer);
+            free(kpurgbbuffer);
+            free(model_data);
+            return;
+    }
+    else
+    {
+         size = read(kmodel_fd, model_data, KMODEL_SIZE);
+         if(size != KMODEL_SIZE)
+         {
+            printf("read kmodel error size %d\n",size);
+            close(g_fd);
+            close(kmodel_fd);
+            free(showbuffer);
+            free(kpurgbbuffer);
+            free(model_data);
+            return;
+            
+         }
+         else
+         {
+             printf("read kmodel success \n");
+         }
+
+    }
+    unsigned char *model_data_align = (unsigned char *)(((unsigned int)model_data+255)&(~255));
+    dvp_set_ai_addr((uint32_t)kpurgbbuffer, (uint32_t)(kpurgbbuffer + 320 * 240), (uint32_t)(kpurgbbuffer + 320 * 240 * 2));
+    if (kpu_load_kmodel(&face_detect_task, model_data_align) != 0)
+    {
+        printf("\nmodel init error\n");
+        close(g_fd);
+        close(kmodel_fd);
+        free(showbuffer);
+        free(kpurgbbuffer);
+        free(model_data);
+        return;
+    }
+    face_detect_rl.anchor_number = ANCHOR_NUM;
+    face_detect_rl.anchor = anchor;
+    face_detect_rl.threshold = 0.7;
+    face_detect_rl.nms_value = 0.3;
+    result = region_layer_init(&face_detect_rl, 20, 15, 30, KPUIMAGEWIDTH, KPUIMAGEHEIGHT);
+    printf("region_layer_init result %d \n\r",result);    
+    size_t stack_size = 32*1024;
+    pthread_attr_t attr;      /* 线程属性 */
+    struct sched_param prio;  /* 线程优先级 */
+    prio.sched_priority = 8;  /* 优先级设置为 8 */
+    pthread_attr_init(&attr);  /* 先使用默认值初始化属性 */
+    pthread_attr_setschedparam(&attr,&prio);  /* 修改属性对应的优先级 */
+    pthread_attr_setstacksize(&attr, stack_size);
+
+    /* 创建线程 1, 属性为 attr，入口函数是 thread_entry，入口函数参数是 1 */
+    result = pthread_create(&facetid,&attr,thread_face_detcet_entry,NULL);
+    if (0 == result)
+    {
+        printf("thread_face_detcet_entry successfully!\n");
+    }
+    else
+    {
+        printf("thread_face_detcet_entry failed! error code is %d\n",result);
+        close(g_fd);
+    }     
+}
+#ifdef __RT_THREAD_H__
+MSH_CMD_EXPORT(face_detect,face detect task);
+#endif
+static void* thread_face_detcet_entry(void *parameter)
+{   
+    extern void lcd_draw_picture(uint16_t x1, uint16_t y1, uint16_t width, uint16_t height, uint32_t *ptr);
+    printf("thread_face_detcet_entry start!\n");
+    int ret = 0;
+    //sysctl_enable_irq();
+    while(1)
+    {
+        //memset(showbuffer,0,320*240*2);
+        g_ai_done_flag = 0;
+        ret = ioctl(g_fd,IOCTRL_CAMERA_START_SHOT,&shoot_para_t);
+        if(RT_ERROR == ret)
+        {
+            printf("ov2640 can't wait event flag");
+            rt_free(showbuffer);
+            close(g_fd);
+            pthread_exit(NULL);  
+            return NULL;
+        }
+        kpu_run_kmodel(&face_detect_task, kpurgbbuffer, DMAC_CHANNEL5, ai_done, NULL);
+        while(!g_ai_done_flag);
+        float *output;
+        size_t output_size;
+        kpu_get_output(&face_detect_task, 0, (uint8_t **)&output, &output_size);
+        face_detect_rl.input = output;        
+        region_layer_run(&face_detect_rl, &face_detect_info);
+        /* display result */
+        #ifdef BSP_USING_LCD
+        for (int face_cnt = 0; face_cnt < face_detect_info.obj_number; face_cnt++)
+        {
+            draw_edge((uint32_t *)showbuffer, &face_detect_info, face_cnt, 0xF800);
+        }
+        lcd_draw_picture(0, 0, 320, 240, (unsigned int*)showbuffer); 
+        #endif
+       usleep(1);
+       if(1 == if_exit)
+       {
+          if_exit = 0;
+          printf("thread_face_detcet_entry exit");
+          pthread_exit(NULL);  
+       }
+    }
+
+}
+
+void face_detect_delete()
+{
+    if(showbuffer != NULL)
+    {
+        int ret = 0;
+        close(g_fd);
+        close(kmodel_fd);
+        free(showbuffer);
+        free(kpurgbbuffer);
+        free(model_data);
+        printf("face detect task cancel!!! ret %d ",ret);
+        if_exit = 1;
+    }
+    
+}
+#ifdef __RT_THREAD_H__
+MSH_CMD_EXPORT(face_detect_delete,face detect task delete);
+#endif
+void kmodel_load(unsigned char  * model_data)
+{
+    int kmodel_fd = 0;
+    int size = 0;
+    kmodel_fd = open("/kmodel/detect.kmodel",O_RDONLY);
+    
+    model_data = (unsigned char *)malloc(KMODEL_SIZE + 255);
+    if(NULL ==model_data)
+    {
+        printf("model_data apply memory fail !!");
+        return ;
+    }
+    memset(model_data,0,KMODEL_SIZE + 255);
+    
+    if (kmodel_fd>= 0)
+    {
+         size = read(kmodel_fd, model_data, KMODEL_SIZE);
+         if(size != KMODEL_SIZE)
+         {
+            printf("read kmodel error size %d\n",size);
+            
+         }
+         else
+         {
+              printf("read kmodel success");
+         }
+    }
+    else
+    {
+        free(model_data);
+        printf("open kmodel fail");
+    }
+    
+}
+#ifdef __RT_THREAD_H__
+MSH_CMD_EXPORT(kmodel_load,kmodel load memory);
+#endif
+

APP_Framework/Applications/knowing_app/mnist/Kconfig

@@ -0,0 +1,4 @@
+config APP_MNIST
+    bool "enable apps/mnist"
+    depends on USING_TENSORFLOWLITEMICRO
+    default n

APP_Framework/Applications/knowing_app/mnist/README.md

@@ -1,5 +1,9 @@
 # MNIST 说明
 
+要使用本例程，MCU RAM必须至少500K左右，所以本例程目前在K210上面验证过，stm32f407 目前在rtt上原则上只能采取dlmodule加载的方式。
+
+![K210 mnist ](E:\XIUOS_FRAMEWORK\xiuos\APP_Framework\Applications\knowing_app\mnist\K210 mnist .png)
+
 ## 使用
 
 tools/mnist-train.py 训练生成 mnist 模型。

APP_Framework/Applications/knowing_app/mnist/SConscript

@@ -0,0 +1,9 @@
+from building import *
+
+cwd     = GetCurrentDir()
+src     = Glob('*.c') + Glob('*.cpp')
+CPPPATH = [cwd]
+
+group = DefineGroup('Applications', src, depend = ['APP_MNIST'], LOCAL_CPPPATH = CPPPATH)
+
+Return('group')

APP_Framework/Applications/knowing_app/mnist/digit.h

@@ -1,23 +1,3 @@
-/*
-* Copyright (c) 2020 AIIT XUOS Lab
-* XiOS is licensed under Mulan PSL v2.
-* You can use this software according to the terms and conditions of the Mulan PSL v2.
-* You may obtain a copy of Mulan PSL v2 at:
-*        http://license.coscl.org.cn/MulanPSL2
-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
-* EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
-* MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
-* See the Mulan PSL v2 for more details.
-*/
-
-/**
-* @file:    digit.h
-* @brief:   store digits in this file
-* @version: 1.0
-* @author:  AIIT XUOS Lab
-* @date:    2021/4/30
-*
-*/
 const float mnist_digit[] = {
   0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00,
   0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00, 0.00,

APP_Framework/Applications/knowing_app/mnist/main.cpp

@@ -1,24 +1,5 @@
-/*
+#include <transform.h>
-* Copyright (c) 2020 AIIT XUOS Lab
+#include <stdio.h>
-* XiOS is licensed under Mulan PSL v2.
-* You can use this software according to the terms and conditions of the Mulan PSL v2.
-* You may obtain a copy of Mulan PSL v2 at:
-*        http://license.coscl.org.cn/MulanPSL2
-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
-* EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
-* MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
-* See the Mulan PSL v2 for more details.
-*/
-
-/**
-* @file:    mnistapp.cpp
-* @brief:   mnist function
-* @version: 1.0
-* @author:  AIIT XUOS Lab
-* @date:    2021/4/30
-*
-*/
-#include <xiuos.h>
 
 #include "tensorflow/lite/micro/all_ops_resolver.h"
 #include "tensorflow/lite/micro/micro_error_reporter.h"
@@ -36,8 +17,8 @@ tflite::MicroInterpreter* interpreter = nullptr;
 TfLiteTensor* input = nullptr;
 TfLiteTensor* output = nullptr;
 constexpr int kTensorArenaSize = 110 * 1024;
-//uint8_t *tensor_arena = nullptr;
+uint8_t *tensor_arena = nullptr;
-uint8_t tensor_arena[kTensorArenaSize];
+//uint8_t tensor_arena[kTensorArenaSize];
 }
 
 extern "C" void mnist_app() {
@@ -52,13 +33,12 @@ extern "C" void mnist_app() {
                          model->version(), TFLITE_SCHEMA_VERSION);
     return;
   }
-  /*
+
-  tensor_arena = (uint8_t *)rt_malloc(kTensorArenaSize);
+  tensor_arena = (uint8_t *)malloc(kTensorArenaSize);
   if (tensor_arena == nullptr) {
     TF_LITE_REPORT_ERROR(error_reporter, "malloc for tensor_arena failed");
     return;
   }
-  */
 
   tflite::AllOpsResolver resolver;
   tflite::MicroInterpreter static_interpreter(
@@ -75,15 +55,15 @@ extern "C" void mnist_app() {
   input = interpreter->input(0);
   output = interpreter->output(0);
 
-  KPrintf("\n------- Input Digit -------\n");
+  printf("------- Input Digit -------\n");
   for (int i = 0; i < 28; i++) {
     for (int j = 0; j < 28; j++) {
       if (mnist_digit[i*28+j] > 0.3)
-        KPrintf("#");
+        printf("#");
       else
-        KPrintf(".");
+        printf(".");
     }
-    KPrintf("\n");
+    printf("\n");
   }
 
   for (int i = 0; i < 28*28; i++) {
@@ -105,8 +85,12 @@ extern "C" void mnist_app() {
            index = i;
          }
   }
-
+  printf("------- Output Result -------\n");
-  KPrintf("\n------- Output Result -------\n");
+  printf("result is %d\n", index);
-  KPrintf("result is %d\n\n", index);
+}
-
+
+extern "C" {
+#ifdef __RT_THREAD_H__
+MSH_CMD_EXPORT(mnist_app, run mnist app);
+#endif
 }

APP_Framework/Applications/knowing_app/mnist/model.h

@@ -1,23 +1,3 @@
-/*
-* Copyright (c) 2020 AIIT XUOS Lab
-* XiOS is licensed under Mulan PSL v2.
-* You can use this software according to the terms and conditions of the Mulan PSL v2.
-* You may obtain a copy of Mulan PSL v2 at:
-*        http://license.coscl.org.cn/MulanPSL2
-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
-* EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
-* MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
-* See the Mulan PSL v2 for more details.
-*/
-
-/**
-* @file:    model.h
-* @brief:   store model weights in this file
-* @version: 1.0
-* @author:  AIIT XUOS Lab
-* @date:    2021/4/30
-*
-*/
 unsigned char mnist_model[] = {
   0x1c, 0x00, 0x00, 0x00, 0x54, 0x46, 0x4c, 0x33, 0x14, 0x00, 0x20, 0x00,
   0x04, 0x00, 0x08, 0x00, 0x0c, 0x00, 0x10, 0x00, 0x14, 0x00, 0x00, 0x00,

APP_Framework/Applications/knowing_app/mnist/tools/mnist-c-digit.py

@@ -1,22 +1,4 @@
 #!/usr/bin/env python3
-# ==========================================================================================
-# Copyright (c) 2020 AIIT XUOS Lab
-# XiOS is licensed under Mulan PSL v2.
-# You can use this software according to the terms and conditions of the Mulan PSL v2.
-# You may obtain a copy of Mulan PSL v2 at:
-#        http://license.coscl.org.cn/MulanPSL2
-# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
-# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
-# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
-# See the Mulan PSL v2 for more details.
-
-#  @file:    mnist-c-digit.py
-#  @brief:   print image digit at command line
-#  @version: 1.0
-#  @author:  AIIT XUOS Lab
-#  @date:    2021/4/30
-# ==========================================================================================
-
 
 import tensorflow as tf
 

APP_Framework/Applications/knowing_app/mnist/tools/mnist-c-model.py

@@ -0,0 +1,23 @@
+#!/usr/bin/env python3
+
+#tflite_file_path = 'mnist-default-quan.tflite'
+tflite_file_path = 'mnist.tflite'
+model_file_path = 'model.h'
+
+tflite_file = open(tflite_file_path, 'rb')
+tflite_data = tflite_file.read()
+tflite_file.close()
+tflite_array = [ '0x%02x' % byte for byte in tflite_data ]
+
+model_content = '''unsigned char mnist_model[] = {
+  %s
+};
+unsigned int mnist_model_len = %d;
+'''
+# 12 bytes in a line, the same with xxd
+bytes_of_line = 12
+model_data = (',\n  ').join([ (', ').join(tflite_array[i:i+bytes_of_line]) for i in range(0, len(tflite_array), bytes_of_line) ])
+model_file = open(model_file_path, 'w')
+model_file.write(model_content % (model_data, len(tflite_array)))
+model_file.close()
+

APP_Framework/Applications/knowing_app/mnist/tools/mnist-inference.py

@@ -1,22 +1,4 @@
 #!/usr/bin/env python3
-# ==========================================================================================
-# Copyright (c) 2020 AIIT XUOS Lab
-# XiOS is licensed under Mulan PSL v2.
-# You can use this software according to the terms and conditions of the Mulan PSL v2.
-# You may obtain a copy of Mulan PSL v2 at:
-#        http://license.coscl.org.cn/MulanPSL2
-# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
-# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
-# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
-# See the Mulan PSL v2 for more details.
-
-#  @file:    mnist-inference.py
-#  @brief:   load data amd start model omferemce
-#  @version: 1.0
-#  @author:  AIIT XUOS Lab
-#  @date:    2021/4/30
-# ==========================================================================================
-
 
 import tensorflow as tf
 

APP_Framework/Applications/knowing_app/mnist/tools/mnist-train.py

@@ -1,22 +1,4 @@
 #!/usr/bin/env python3
-# ==========================================================================================
-# Copyright (c) 2020 AIIT XUOS Lab
-# XiOS is licensed under Mulan PSL v2.
-# You can use this software according to the terms and conditions of the Mulan PSL v2.
-# You may obtain a copy of Mulan PSL v2 at:
-#        http://license.coscl.org.cn/MulanPSL2
-# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
-# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
-# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
-# See the Mulan PSL v2 for more details.
-
-#  @file:    mnist-train.py
-#  @brief:   model training
-#  @version: 1.0
-#  @author:  AIIT XUOS Lab
-#  @date:    2021/4/30
-# ==========================================================================================
-
 
 import os
 import tensorflow as tf

APP_Framework/Framework/Kconfig

@@ -20,7 +20,7 @@ menu "Framework"
 
     source "$APP_DIR/Framework/sensor/Kconfig"
     source "$APP_DIR/Framework/connection/Kconfig"
-    source "$APP_DIR/Framework/know/Kconfig"
+    source "$APP_DIR/Framework/knowing/Kconfig"
     source "$APP_DIR/Framework/control/Kconfig"
 
 

APP_Framework/Framework/know/Kconfig

@@ -1,8 +0,0 @@
-menuconfig SUPPORT_KNOWING_FRAMEWORK
-        bool "support knowing framework"
-        default y
-
-if SUPPORT_KNOWING_FRAMEWORK
-        source "$APP_DIR/Framework/know/tflite_sin/Kconfig"
-        source "$APP_DIR/Framework/know/tflite_mnist/Kconfig"
-endif

APP_Framework/Framework/know/tflite_mnist/Kconfig

@@ -1,4 +0,0 @@
-menuconfig USING_TFLITE_MNIST
-    bool "mnist demo app for tflite micro"
-    depends on INTELLIGENT_TFLITE
-    default n

APP_Framework/Framework/know/tflite_mnist/Makefile

@@ -1,8 +0,0 @@
-ifeq ($(CONFIG_USING_TFLITE_MNIST),y)
-	SRC_FILES := \
-		mnistapp.cpp \
-		mnistmain.c
-	CPPPATHS += -I.
-endif
-
-include $(KERNEL_ROOT)/compiler.mk

APP_Framework/Framework/know/tflite_mnist/mnistmain.c

@@ -1,30 +0,0 @@
-/*
-* Copyright (c) 2020 AIIT XUOS Lab
-* XiOS is licensed under Mulan PSL v2.
-* You can use this software according to the terms and conditions of the Mulan PSL v2.
-* You may obtain a copy of Mulan PSL v2 at:
-*        http://license.coscl.org.cn/MulanPSL2
-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
-* EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
-* MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
-* See the Mulan PSL v2 for more details.
-*/
-
-/**
-* @file:    mnistmain.c
-* @brief:   start mnist function
-* @version: 1.0
-* @author:  AIIT XUOS Lab
-* @date:    2021/4/30
-*
-*/
-
-#include <xiuos.h>
-
-void mnist_app(void);
-
-int tfmnist(void) {
-    mnist_app();
-}
-
-SHELL_EXPORT_CMD(SHELL_CMD_PERMISSION(0)|SHELL_CMD_TYPE(SHELL_TYPE_CMD_FUNC)|SHELL_CMD_PARAM_NUM(0), tfmnist, tfmnist, run mnist demo of tflite);

APP_Framework/Framework/know/tflite_mnist/tools/mnist-c-model.py

@@ -1,41 +0,0 @@
-#!/usr/bin/env python3
-# ==========================================================================================
-# Copyright (c) 2020 AIIT XUOS Lab
-# XiOS is licensed under Mulan PSL v2.
-# You can use this software according to the terms and conditions of the Mulan PSL v2.
-# You may obtain a copy of Mulan PSL v2 at:
-#        http://license.coscl.org.cn/MulanPSL2
-# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
-# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
-# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
-# See the Mulan PSL v2 for more details.
-
-#  @file:    mnist-c-model.py
-#  @brief:   open file path and load model
-#  @version: 1.0
-#  @author:  AIIT XUOS Lab
-#  @date:    2021/4/30
-# ==========================================================================================
-
-
-#tflite_file_path = 'mnist-default-quan.tflite'
-tflite_file_path = 'mnist.tflite'
-model_file_path = 'model.h'
-
-tflite_file = open(tflite_file_path, 'rb')
-tflite_data = tflite_file.read()
-tflite_file.close()
-tflite_array = [ '0x%02x' % byte for byte in tflite_data ]
-
-model_content = '''unsigned char mnist_model[] = {
-  %s
-};
-unsigned int mnist_model_len = %d;
-'''
-# 12 bytes in a line, the same with xxd
-bytes_of_line = 12
-model_data = (',\n  ').join([ (', ').join(tflite_array[i:i+bytes_of_line]) for i in range(0, len(tflite_array), bytes_of_line) ])
-model_file = open(model_file_path, 'w')
-model_file.write(model_content % (model_data, len(tflite_array)))
-model_file.close()
-

APP_Framework/Framework/know/tflite_sin/Kconfig

@@ -1,4 +0,0 @@
-menuconfig USING_TFLITE_SIN
-    bool "sin(x) demo app for tflite micro"
-    depends on INTELLIGENT_TFLITE
-    default n

APP_Framework/Framework/know/tflite_sin/Makefile

@@ -1,11 +0,0 @@
-ifeq ($(CONFIG_USING_TFLITE_SIN),y)
-	SRC_FILES := \
-		sinmain.c \
-		main_functions.cc \
-		model.cc \
-		output_handler.cc \
-		constants.cc
-	CPPPATHS += -I.
-endif
-
-include $(KERNEL_ROOT)/compiler.mk

APP_Framework/Framework/knowing/Kconfig

@@ -0,0 +1,8 @@
+menuconfig SUPPORT_KNOWING_FRAMEWORK
+        bool "support knowing framework"
+        default y
+
+if SUPPORT_KNOWING_FRAMEWORK
+        source "$APP_DIR/Framework/knowing/tensorflow-lite/Kconfig"
+        source "$APP_DIR/Framework/knowing/kpu-postprocessing/Kconfig"
+endif

APP_Framework/Framework/knowing/SConscript

@@ -0,0 +1,14 @@
+import os
+Import('RTT_ROOT')
+from building import *
+
+cwd = GetCurrentDir()
+objs = []
+list = os.listdir(cwd)
+
+for d in list:
+    path = os.path.join(cwd, d)
+    if os.path.isfile(os.path.join(path, 'SConscript')):
+        objs = objs + SConscript(os.path.join(path, 'SConscript'))
+
+Return('objs')

APP_Framework/Framework/knowing/kpu-postprocessing/Kconfig

@@ -0,0 +1,5 @@
+menuconfig USING_KPU_POSTPROCESSING
+    bool "kpu model postprocessing"
+    default y
+
+source "$APP_DIR/Framework/knowing/kpu-postprocessing/yolov2/Kconfig"

APP_Framework/Framework/knowing/kpu-postprocessing/SConscript

@@ -0,0 +1,14 @@
+import os
+Import('RTT_ROOT')
+from building import *
+
+cwd = GetCurrentDir()
+objs = []
+list = os.listdir(cwd)
+
+for d in list:
+    path = os.path.join(cwd, d)
+    if os.path.isfile(os.path.join(path, 'SConscript')):
+        objs = objs + SConscript(os.path.join(path, 'SConscript'))
+
+Return('objs')

APP_Framework/Framework/knowing/kpu-postprocessing/yolov2/Kconfig

@@ -0,0 +1,7 @@
+menuconfig USING_YOLOV2
+    bool "yolov2 region layer"
+	depends on USING_KPU_POSTPROCESSING
+    default n
+
+
+

APP_Framework/Framework/knowing/kpu-postprocessing/yolov2/SConscript

@@ -0,0 +1,10 @@
+from building import *
+import os
+
+cwd = GetCurrentDir()
+
+src = Glob('*.c')
+
+group = DefineGroup('yolov2', src, depend = ['USING_YOLOV2'], CPPPATH = [cwd])
+
+Return('group')

APP_Framework/Framework/knowing/kpu-postprocessing/yolov2/region_layer.c

@@ -0,0 +1,437 @@
+#include <stdlib.h>
+#include <math.h>
+#include <stdio.h>
+#include "region_layer.h"
+
+typedef struct
+{
+    float x;
+    float y;
+    float w;
+    float h;
+} box_t;
+
+typedef struct
+{
+    int index;
+    int class;
+    float **probs;
+} sortable_box_t;
+
+
+int region_layer_init(region_layer_t *rl, int width, int height, int channels, int origin_width, int origin_height)
+{
+    int flag = 0;
+
+    rl->coords = 4;
+    rl->image_width = 320;
+    rl->image_height = 240;
+
+    rl->classes = channels / 5 - 5;
+    rl->net_width = origin_width;
+    rl->net_height = origin_height;
+    rl->layer_width = width;
+    rl->layer_height = height;
+    rl->boxes_number = (rl->layer_width * rl->layer_height * rl->anchor_number); 
+    rl->output_number = (rl->boxes_number * (rl->classes + rl->coords + 1));
+
+    rl->output = malloc(rl->output_number * sizeof(float));
+    if (rl->output == NULL)
+    {
+        flag = -1;
+        goto malloc_error;
+    }
+    rl->boxes = malloc(rl->boxes_number * sizeof(box_t));
+    if (rl->boxes == NULL)
+    {
+        flag = -2;
+        goto malloc_error;
+    }
+    rl->probs_buf = malloc(rl->boxes_number * (rl->classes + 1) * sizeof(float));
+    if (rl->probs_buf == NULL)
+    {
+        flag = -3;
+        goto malloc_error;
+    }
+    rl->probs = malloc(rl->boxes_number * sizeof(float *));
+    if (rl->probs == NULL)
+    {
+        flag = -4;
+        goto malloc_error;
+    }
+    for (uint32_t i = 0; i < rl->boxes_number; i++)
+        rl->probs[i] = &(rl->probs_buf[i * (rl->classes + 1)]);
+    return 0;
+malloc_error:
+    free(rl->output);
+    free(rl->boxes);
+    free(rl->probs_buf);
+    free(rl->probs);
+    return flag;
+}
+
+void region_layer_deinit(region_layer_t *rl)
+{
+    free(rl->output);
+    free(rl->boxes);
+    free(rl->probs_buf);
+    free(rl->probs);
+}
+
+static inline float sigmoid(float x)
+{
+    return 1.f / (1.f + expf(-x));
+}
+
+static void activate_array(region_layer_t *rl, int index, int n)
+{
+    float *output = &rl->output[index];
+    float *input = &rl->input[index];
+
+    for (int i = 0; i < n; ++i)
+        output[i] = sigmoid(input[i]);
+}
+
+static int entry_index(region_layer_t *rl, int location, int entry)
+{
+    int wh = rl->layer_width * rl->layer_height;
+    int n   = location / wh;
+    int loc = location % wh;
+
+    return n * wh * (rl->coords + rl->classes + 1) + entry * wh + loc;
+}
+
+static void softmax(region_layer_t *rl, float *input, int n, int stride, float *output)
+{
+    int i;
+    float diff;
+    float e;
+    float sum = 0;
+    float largest_i = input[0];
+
+    for (i = 0; i < n; ++i)
+    {
+        if (input[i * stride] > largest_i)
+            largest_i = input[i * stride];
+    }
+
+    for (i = 0; i < n; ++i) {
+        diff = input[i * stride] - largest_i;
+        e = expf(diff);
+        sum += e;
+        output[i * stride] = e;
+    }
+    for (i = 0; i < n; ++i)
+        output[i * stride] /= sum;
+}
+
+static void softmax_cpu(region_layer_t *rl, float *input, int n, int batch, int batch_offset, int groups, int stride, float *output)
+{
+    int g, b;
+
+    for (b = 0; b < batch; ++b) {
+        for (g = 0; g < groups; ++g)
+            softmax(rl, input + b * batch_offset + g, n, stride, output + b * batch_offset + g);
+    }
+}
+
+static void forward_region_layer(region_layer_t *rl)
+{
+    int index;
+
+    for (index = 0; index < rl->output_number; index++)
+        rl->output[index] = rl->input[index];
+
+    for (int n = 0; n < rl->anchor_number; ++n)
+    {
+        index = entry_index(rl, n * rl->layer_width * rl->layer_height, 0);
+        activate_array(rl, index, 2 * rl->layer_width * rl->layer_height);
+        index = entry_index(rl, n * rl->layer_width * rl->layer_height, 4);
+        activate_array(rl, index, rl->layer_width * rl->layer_height);
+    }
+
+    index = entry_index(rl, 0, rl->coords + 1);
+    softmax_cpu(rl, rl->input + index, rl->classes, rl->anchor_number,
+            rl->output_number / rl->anchor_number, rl->layer_width * rl->layer_height,
+            rl->layer_width * rl->layer_height, rl->output + index);
+}
+
+static void correct_region_boxes(region_layer_t *rl, box_t *boxes)
+{
+    uint32_t net_width = rl->net_width;
+    uint32_t net_height = rl->net_height;
+    uint32_t image_width = rl->image_width;
+    uint32_t image_height = rl->image_height;
+    uint32_t boxes_number = rl->boxes_number;
+    int new_w = 0;
+    int new_h = 0;
+
+    if (((float)net_width / image_width) <
+        ((float)net_height / image_height)) {
+        new_w = net_width;
+        new_h = (image_height * net_width) / image_width;
+    } else {
+        new_h = net_height;
+        new_w = (image_width * net_height) / image_height;
+    }
+    for (int i = 0; i < boxes_number; ++i) {
+        box_t b = boxes[i];
+
+        b.x = (b.x - (net_width - new_w) / 2. / net_width) /
+              ((float)new_w / net_width);
+        b.y = (b.y - (net_height - new_h) / 2. / net_height) /
+              ((float)new_h / net_height);
+        b.w *= (float)net_width / new_w;
+        b.h *= (float)net_height / new_h;
+        boxes[i] = b;
+    }
+}
+
+static box_t get_region_box(float *x, float *biases, int n, int index, int i, int j, int w, int h, int stride)
+{
+    volatile box_t b;
+
+    b.x = (i + x[index + 0 * stride]) / w;
+    b.y = (j + x[index + 1 * stride]) / h;
+    b.w = expf(x[index + 2 * stride]) * biases[2 * n] / w;
+    b.h = expf(x[index + 3 * stride]) * biases[2 * n + 1] / h;
+    return b;
+}
+
+static void get_region_boxes(region_layer_t *rl, float *predictions, float **probs, box_t *boxes)
+{
+    uint32_t layer_width = rl->layer_width;
+    uint32_t layer_height = rl->layer_height;
+    uint32_t anchor_number = rl->anchor_number;
+    uint32_t classes = rl->classes;
+    uint32_t coords = rl->coords;
+    float threshold = rl->threshold;
+
+    for (int i = 0; i < layer_width * layer_height; ++i)
+    {
+        int row = i / layer_width;
+        int col = i % layer_width;
+
+        for (int n = 0; n < anchor_number; ++n)
+        {
+            int index = n * layer_width * layer_height + i;
+
+            for (int j = 0; j < classes; ++j)
+                probs[index][j] = 0;
+            int obj_index = entry_index(rl, n * layer_width * layer_height + i, coords);
+            int box_index = entry_index(rl, n * layer_width * layer_height + i, 0);
+            float scale  = predictions[obj_index];
+
+            boxes[index] = get_region_box(predictions, rl->anchor, n, box_index, col, row,
+                layer_width, layer_height, layer_width * layer_height);
+
+            float max = 0;
+
+            for (int j = 0; j < classes; ++j)
+            {
+                int class_index = entry_index(rl, n * layer_width * layer_height + i, coords + 1 + j);
+                float prob = scale * predictions[class_index];
+
+                probs[index][j] = (prob > threshold) ? prob : 0;
+                if (prob > max)
+                    max = prob;
+            }
+            probs[index][classes] = max;
+        }
+    }
+    correct_region_boxes(rl, boxes);
+}
+
+static int nms_comparator(void *pa, void *pb)
+{
+    sortable_box_t a = *(sortable_box_t *)pa;
+    sortable_box_t b = *(sortable_box_t *)pb;
+    float diff = a.probs[a.index][b.class] - b.probs[b.index][b.class];
+
+    if (diff < 0)
+        return 1;
+    else if (diff > 0)
+        return -1;
+    return 0;
+}
+
+static float overlap(float x1, float w1, float x2, float w2)
+{
+    float l1 = x1 - w1/2;
+    float l2 = x2 - w2/2;
+    float left = l1 > l2 ? l1 : l2;
+    float r1 = x1 + w1/2;
+    float r2 = x2 + w2/2;
+    float right = r1 < r2 ? r1 : r2;
+
+    return right - left;
+}
+
+static float box_intersection(box_t a, box_t b)
+{
+    float w = overlap(a.x, a.w, b.x, b.w);
+    float h = overlap(a.y, a.h, b.y, b.h);
+
+    if (w < 0 || h < 0)
+        return 0;
+    return w * h;
+}
+
+static float box_union(box_t a, box_t b)
+{
+    float i = box_intersection(a, b);
+    float u = a.w * a.h + b.w * b.h - i;
+
+    return u;
+}
+
+static float box_iou(box_t a, box_t b)
+{
+    return box_intersection(a, b) / box_union(a, b);
+}
+
+static void do_nms_sort(region_layer_t *rl, box_t *boxes, float **probs)
+{
+    uint32_t boxes_number = rl->boxes_number;
+    uint32_t classes = rl->classes;
+    float nms_value = rl->nms_value;
+    int i, j, k;
+    sortable_box_t s[boxes_number];
+
+    for (i = 0; i < boxes_number; ++i)
+    {
+        s[i].index = i;
+        s[i].class = 0;
+        s[i].probs = probs;
+    }
+
+    for (k = 0; k < classes; ++k)
+    {
+        for (i = 0; i < boxes_number; ++i)
+            s[i].class = k;
+        qsort(s, boxes_number, sizeof(sortable_box_t), nms_comparator);
+        for (i = 0; i < boxes_number; ++i)
+        {
+            if (probs[s[i].index][k] == 0)
+                continue;
+            box_t a = boxes[s[i].index];
+
+            for (j = i + 1; j < boxes_number; ++j)
+            {
+                box_t b = boxes[s[j].index];
+
+                if (box_iou(a, b) > nms_value)
+                    probs[s[j].index][k] = 0;
+            }
+        }
+    }
+}
+
+static int max_index(float *a, int n)
+{
+    int i, max_i = 0;
+    float max = a[0];
+
+    for (i = 1; i < n; ++i)
+    {
+        if (a[i] > max)
+        {
+            max   = a[i];
+            max_i = i;
+        }
+    }
+    return max_i;
+}
+
+static void region_layer_output(region_layer_t *rl, obj_info_t *obj_info)
+{
+    uint32_t obj_number = 0;
+    uint32_t image_width = rl->image_width;
+    uint32_t image_height = rl->image_height;
+    uint32_t boxes_number = rl->boxes_number;
+    float threshold = rl->threshold;
+    box_t *boxes = (box_t *)rl->boxes;
+    
+    for (int i = 0; i < rl->boxes_number; ++i)
+    {
+        int class  = max_index(rl->probs[i], rl->classes);
+        float prob = rl->probs[i][class];
+
+        if (prob > threshold)
+        {
+            box_t *b = boxes + i;
+            obj_info->obj[obj_number].x1 = b->x * image_width - (b->w * image_width / 2);
+            obj_info->obj[obj_number].y1 = b->y * image_height - (b->h * image_height / 2);
+            obj_info->obj[obj_number].x2 = b->x * image_width + (b->w * image_width / 2);
+            obj_info->obj[obj_number].y2 = b->y * image_height + (b->h * image_height / 2);
+            obj_info->obj[obj_number].class_id = class;
+            obj_info->obj[obj_number].prob = prob;
+            obj_number++;
+        }
+    }
+    obj_info->obj_number = obj_number;
+}
+
+void region_layer_run(region_layer_t *rl, obj_info_t *obj_info)
+{
+    forward_region_layer(rl);
+    get_region_boxes(rl, rl->output, rl->probs, rl->boxes);
+    do_nms_sort(rl, rl->boxes, rl->probs);
+    region_layer_output(rl, obj_info);
+}
+
+void draw_edge(uint32_t *gram, obj_info_t *obj_info, uint32_t index, uint16_t color)
+{
+    uint32_t data = ((uint32_t)color << 16) | (uint32_t)color;
+    uint32_t *addr1, *addr2, *addr3, *addr4, x1, y1, x2, y2;
+
+    x1 = obj_info->obj[index].x1;
+    y1 = obj_info->obj[index].y1;
+    x2 = obj_info->obj[index].x2;
+    y2 = obj_info->obj[index].y2;
+
+    if (x1 <= 0)
+        x1 = 1;
+    if (x2 >= 319)
+        x2 = 318;
+    if (y1 <= 0)
+        y1 = 1;
+    if (y2 >= 239)
+        y2 = 238;
+
+    addr1 = gram + (320 * y1 + x1) / 2;
+    addr2 = gram + (320 * y1 + x2 - 8) / 2;
+    addr3 = gram + (320 * (y2 - 1) + x1) / 2;
+    addr4 = gram + (320 * (y2 - 1) + x2 - 8) / 2;
+    for (uint32_t i = 0; i < 4; i++)
+    {
+        *addr1 = data;
+        *(addr1 + 160) = data;
+        *addr2 = data;
+        *(addr2 + 160) = data;
+        *addr3 = data;
+        *(addr3 + 160) = data;
+        *addr4 = data;
+        *(addr4 + 160) = data;
+        addr1++;
+        addr2++;
+        addr3++;
+        addr4++;
+    }
+    addr1 = gram + (320 * y1 + x1) / 2;
+    addr2 = gram + (320 * y1 + x2 - 2) / 2;
+    addr3 = gram + (320 * (y2 - 8) + x1) / 2;
+    addr4 = gram + (320 * (y2 - 8) + x2 - 2) / 2;
+    for (uint32_t i = 0; i < 8; i++)
+    {
+        *addr1 = data;
+        *addr2 = data;
+        *addr3 = data;
+        *addr4 = data;
+        addr1 += 160;
+        addr2 += 160;
+        addr3 += 160;
+        addr4 += 160;
+    }
+}
+

APP_Framework/Framework/knowing/kpu-postprocessing/yolov2/region_layer.h

@@ -0,0 +1,49 @@
+#ifndef _REGION_LAYER
+#define _REGION_LAYER
+
+#include <stdint.h>
+#include "kpu.h"
+
+typedef struct
+{
+    uint32_t obj_number;
+    struct
+    {
+        uint32_t x1;
+        uint32_t y1;
+        uint32_t x2;
+        uint32_t y2;
+        uint32_t class_id;
+        float prob;
+    } obj[10];
+} obj_info_t;
+
+typedef struct
+{
+    float threshold;
+    float nms_value;
+    uint32_t coords;
+    uint32_t anchor_number;
+    float *anchor;
+    uint32_t image_width;
+    uint32_t image_height;
+    uint32_t classes;
+    uint32_t net_width;
+    uint32_t net_height;
+    uint32_t layer_width;
+    uint32_t layer_height;
+    uint32_t boxes_number;
+    uint32_t output_number;
+    void *boxes;
+    float *input;
+    float *output;
+    float *probs_buf;
+    float **probs;
+} region_layer_t;
+
+int region_layer_init(region_layer_t *rl, int width, int height, int channels, int origin_width, int origin_height);
+void region_layer_deinit(region_layer_t *rl);
+void region_layer_run(region_layer_t *rl, obj_info_t *obj_info);
+void draw_edge(uint32_t *gram, obj_info_t *obj_info, uint32_t index, uint16_t color);
+
+#endif // _REGION_LAYER

APP_Framework/Framework/knowing/tensorflow-lite/Kconfig

@@ -0,0 +1,24 @@
+menuconfig USING_TENSORFLOWLITEMICRO
+    bool "Tensorflow Lite for Micro"
+    select RT_USING_CPLUSPLUS
+    default n
+
+if USING_TENSORFLOWLITEMICRO
+
+    choice
+        prompt "Select Tensorflow Lite Operators Type"
+        default USING_TENSORFLOWLITEMICRO_NORMAL
+
+        config USING_TENSORFLOWLITEMICRO_NORMAL
+            bool "Using Tensorflow Lite normal operations"
+
+        config USING_TENSORFLOWLITEMICRO_CMSISNN
+            bool "Using Tensorflow Lite CMSIS NN operations"
+    endchoice
+
+    config USING_TENSORFLOWLITEMICRO_DEMOAPP
+        bool "Using tensorflow lite for micro demo app"
+        default n
+
+endif
+

APP_Framework/Framework/knowing/tensorflow-lite/SConscript

@@ -0,0 +1,207 @@
+from building import *
+import os
+
+cwd = GetCurrentDir()
+
+common = Split('''
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/all_ops_resolver.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/debug_log.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/memory_helpers.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/micro_allocator.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/micro_error_reporter.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/micro_interpreter.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/micro_profiler.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/micro_string.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/micro_time.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/micro_utils.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/recording_micro_allocator.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/recording_simple_memory_allocator.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/simple_memory_allocator.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/test_helpers.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/benchmarks/keyword_scrambled_model_data.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/memory_planner/greedy_memory_planner.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/memory_planner/linear_memory_planner.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/testing/test_conv_model.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/c/common.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/core/api/error_reporter.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/core/api/flatbuffer_conversions.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/core/api/op_resolver.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/core/api/tensor_utils.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/quantization_util.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/kernel_util.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/schema/schema_utils.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/activations.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/arg_min_max.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/ceil.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/circular_buffer.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/comparisons.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/concatenation.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/conv_test_common.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/dequantize.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/detection_postprocess.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/elementwise.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/ethosu.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/flexbuffers_generated_data.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/floor.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/hard_swish.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/kernel_runner.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/kernel_util.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/l2norm.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/logical.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/logistic.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/maximum_minimum.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/neg.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/pack.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/pad.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/prelu.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/quantize.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/quantize_common.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/reduce.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/reshape.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/resize_nearest_neighbor.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/round.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/shape.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/split.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/split_v.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/strided_slice.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/sub.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/svdf_common.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/tanh.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/transpose_conv.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/unpack.cc
+''')
+
+app = Split('''
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/examples/hello_world/main.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/examples/hello_world/main_functions.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/examples/hello_world/model.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/examples/hello_world/output_handler.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/examples/hello_world/constants.cc
+''')
+
+normal_ops = Split('''
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/normal/add.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/normal/conv.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/normal/depthwise_conv.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/normal/fully_connected.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/normal/mul.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/normal/pooling.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/normal/softmax.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/normal/svdf.cc
+''')
+
+cmsis_ops = Split('''
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/cmsis-nn/add.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/cmsis-nn/conv.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/cmsis-nn/depthwise_conv.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/cmsis-nn/fully_connected.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/cmsis-nn/mul.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/cmsis-nn/pooling.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/cmsis-nn/softmax.cc
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/kernels/cmsis-nn/svdf.cc
+''')
+
+cmsis = Split('''
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_nn_activations_q15.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_nn_activations_q7.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu6_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu_q15.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu_q7.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/BasicMathFunctions/arm_elementwise_add_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/BasicMathFunctions/arm_elementwise_mul_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_w.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_x.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_y.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_z.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1_x_n_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1x1_HWC_q7_fast_nonsquare.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1x1_s8_fast.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_basic.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast_nonsquare.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_RGB.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic_nonsquare.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast_nonsquare.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_wrapper_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_3x3_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_s8_opt.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_u8_basic_ver1.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_wrapper_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7_nonsquare.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_depthwise_conv_s8_core.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15_reordered.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_s8_s16.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_s8_s16_reordered.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15_opt.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15_opt.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7_opt.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_accumulate_q7_to_q15.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_add_q7.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_depthwise_conv_nt_t_padded_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_depthwise_conv_nt_t_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mul_core_1x_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mul_core_4x_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mult_nt_t_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mult_q15.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mult_q7.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_vec_mat_mult_t_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nntables.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_no_shift.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_reordered_no_shift.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_reordered_with_offset.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_with_offset.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_avgpool_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_max_pool_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_pool_q7_HWC.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ReshapeFunctions/arm_reshape_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SVDFunctions/arm_svdf_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q15.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q7.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_u8.c
+tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_with_batch_q7.c
+''')
+
+CPPPATH = [
+    os.path.join(cwd, 'tensorflow-lite-for-mcu/source'),
+    os.path.join(cwd, 'tensorflow-lite-for-mcu/patch'),
+    os.path.join(cwd, 'tensorflow-lite-for-mcu/source/third_party/gemmlowp'),
+    os.path.join(cwd, 'tensorflow-lite-for-mcu/source/third_party/flatbuffers/include'),
+    os.path.join(cwd, 'tensorflow-lite-for-mcu/source/third_party/ruy'),
+    os.path.join(cwd, 'tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis'),
+    os.path.join(cwd, 'tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include'),
+    os.path.join(cwd, 'tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include'),
+    os.path.join(cwd, 'tensorflow-lite-for-mcu/source/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/Core/Include'),
+    ]
+
+# embedded C++ std don't have some math functions, use global math functions instead
+CPPDEFINES = ['TF_LITE_USE_GLOBAL_CMATH_FUNCTIONS', 'TF_LITE_USE_GLOBAL_MAX', 'TF_LITE_USE_GLOBAL_MIN']
+src = []
+
+if GetDepend(['USING_TENSORFLOWLITEMICRO_NORMAL']):
+    src += common + normal_ops
+elif GetDepend(['USING_TENSORFLOWLITEMICRO_CMSISNN']):
+    CPPDEFINES += ['CMSIS_NN']
+    src += common + cmsis_ops + cmsis
+
+if GetDepend(['USING_TENSORFLOWLITEMICRO_DEMOAPP']):
+    src += app
+
+# building scripts don't support building LIBRARY with LOCAL FLAGS(like LOCAL_CPPDEFINES) currently
+# so LIBRARY use GLOBAL FLAGS(like CPPDEFINES)
+#group = DefineGroup('tensorflow-lite-for-mcu', src, depend = ['USING_TENSORFLOWLITEMICRO'], CPPPATH = CPPPATH, CPPDEFINES = CPPDEFINES, LIBRARY = True)
+# static library link order matters, apps/mnist/main.o should be placed before libtensorflow.a, we should fix this later
+group = DefineGroup('tensorflow-lite-for-mcu', src, depend = ['USING_TENSORFLOWLITEMICRO'], CPPPATH = CPPPATH, CPPDEFINES = CPPDEFINES)
+
+Return('group')

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/.gitignore

@@ -0,0 +1 @@
+*.o

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/README.md

@@ -0,0 +1,203 @@
+# Tensorflow Lite for MCU
+
+## 仓库说明
+
+该仓库是由 tensorflow 仓库生成：
+
+```
+# 生成源码项目
+# 不指定 'hello_world' 则生成 hello_world/image_recognition/micro_speech/... 等多个项目
+# 不指定 'make' 则生成 arduino、esp-idf、keil、make、mbed 等多个工程
+# 生成结果在 tensorflow/tensorflow/lite/micro/tools/make/gen/osx_x86_64/prj/hello_world/make
+# 以下指令生成 hello_world 的源码项目，其中算子使用常规算子
+make -f tensorflow/lite/micro/tools/make/Makefile generate_hello_world_make_project
+
+# 以下指令生成 针对 arm cortex-m 平台使用 cmsis 算子优化的源码项目，详见 tensorflow 仓库的 tensorflow/tensorflow/lite/micro/kernels/cmsis-nn/README.md
+make -f tensorflow/lite/micro/tools/make/Makefile TAGS=cmsis-nn generate_hello_world_make_project
+```
+
+当前使用的 tensorflow 版本为：`e30e1d1aedb` （2021-01-07）。
+
+我们在其他平台使用常规算子，在 cortex-m 平台使用 cmsis 优化算子，需要将上述生成的两个源码项目整合起来，具体代码如何融合，见本文最后文件列表。
+
+## 问题说明
+
+### 问题一：缺少头文件路径
+
+tensorflow生成的源码项目中的 makefile 的头文件路径是不全的，需要添加以下头文件路径：
+
+```
+/tensorflow/lite/micro/tools/make/downloads/cmsis
+/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/DSP/Include
+/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Include
+/tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/Core/Include
+```
+
+### 问题二：cmsis 缺少头文件
+
+tensorflow生成的项目缺少 cmsis_gcc.h 头文件，我们将该文件放在了 patch/cmsis_gcc.h。
+
+## 文件列表
+
+```
+# 共有文件
+tensorflow/lite/micro/all_ops_resolver.cc
+tensorflow/lite/micro/debug_log.cc
+tensorflow/lite/micro/memory_helpers.cc
+tensorflow/lite/micro/micro_allocator.cc
+tensorflow/lite/micro/micro_error_reporter.cc
+tensorflow/lite/micro/micro_interpreter.cc
+tensorflow/lite/micro/micro_profiler.cc
+tensorflow/lite/micro/micro_string.cc
+tensorflow/lite/micro/micro_time.cc
+tensorflow/lite/micro/micro_utils.cc
+tensorflow/lite/micro/recording_micro_allocator.cc
+tensorflow/lite/micro/recording_simple_memory_allocator.cc
+tensorflow/lite/micro/simple_memory_allocator.cc
+tensorflow/lite/micro/test_helpers.cc
+tensorflow/lite/micro/benchmarks/keyword_scrambled_model_data.cc
+tensorflow/lite/micro/memory_planner/greedy_memory_planner.cc
+tensorflow/lite/micro/memory_planner/linear_memory_planner.cc
+tensorflow/lite/micro/testing/test_conv_model.cc
+tensorflow/lite/c/common.c
+tensorflow/lite/core/api/error_reporter.cc
+tensorflow/lite/core/api/flatbuffer_conversions.cc
+tensorflow/lite/core/api/op_resolver.cc
+tensorflow/lite/core/api/tensor_utils.cc
+tensorflow/lite/kernels/internal/quantization_util.cc
+tensorflow/lite/kernels/kernel_util.cc
+tensorflow/lite/schema/schema_utils.cc
+tensorflow/lite/micro/kernels/activations.cc
+tensorflow/lite/micro/kernels/arg_min_max.cc
+tensorflow/lite/micro/kernels/ceil.cc
+tensorflow/lite/micro/kernels/circular_buffer.cc
+tensorflow/lite/micro/kernels/comparisons.cc
+tensorflow/lite/micro/kernels/concatenation.cc
+tensorflow/lite/micro/kernels/conv_test_common.cc
+tensorflow/lite/micro/kernels/dequantize.cc
+tensorflow/lite/micro/kernels/detection_postprocess.cc
+tensorflow/lite/micro/kernels/elementwise.cc
+tensorflow/lite/micro/kernels/ethosu.cc
+tensorflow/lite/micro/kernels/flexbuffers_generated_data.cc
+tensorflow/lite/micro/kernels/floor.cc
+tensorflow/lite/micro/kernels/hard_swish.cc
+tensorflow/lite/micro/kernels/kernel_runner.cc
+tensorflow/lite/micro/kernels/kernel_util.cc
+tensorflow/lite/micro/kernels/l2norm.cc
+tensorflow/lite/micro/kernels/logical.cc
+tensorflow/lite/micro/kernels/logistic.cc
+tensorflow/lite/micro/kernels/maximum_minimum.cc
+tensorflow/lite/micro/kernels/neg.cc
+tensorflow/lite/micro/kernels/pack.cc
+tensorflow/lite/micro/kernels/pad.cc
+tensorflow/lite/micro/kernels/prelu.cc
+tensorflow/lite/micro/kernels/quantize.cc
+tensorflow/lite/micro/kernels/quantize_common.cc
+tensorflow/lite/micro/kernels/reduce.cc
+tensorflow/lite/micro/kernels/reshape.cc
+tensorflow/lite/micro/kernels/resize_nearest_neighbor.cc
+tensorflow/lite/micro/kernels/round.cc
+tensorflow/lite/micro/kernels/shape.cc
+tensorflow/lite/micro/kernels/split.cc
+tensorflow/lite/micro/kernels/split_v.cc
+tensorflow/lite/micro/kernels/strided_slice.cc
+tensorflow/lite/micro/kernels/sub.cc
+tensorflow/lite/micro/kernels/svdf_common.cc
+tensorflow/lite/micro/kernels/tanh.cc
+tensorflow/lite/micro/kernels/transpose_conv.cc
+tensorflow/lite/micro/kernels/unpack.cc
+tensorflow/lite/micro/examples/hello_world/main.cc
+tensorflow/lite/micro/examples/hello_world/main_functions.cc
+tensorflow/lite/micro/examples/hello_world/model.cc
+tensorflow/lite/micro/examples/hello_world/output_handler.cc
+tensorflow/lite/micro/examples/hello_world/constants.cc
+
+# 常规算子，将这些文件移入 tensorflow/lite/micro/kernels/normal 中
+tensorflow/lite/micro/kernels/add.cc
+tensorflow/lite/micro/kernels/conv.cc
+tensorflow/lite/micro/kernels/depthwise_conv.cc
+tensorflow/lite/micro/kernels/fully_connected.cc
+tensorflow/lite/micro/kernels/mul.cc
+tensorflow/lite/micro/kernels/pooling.cc
+tensorflow/lite/micro/kernels/softmax.cc
+tensorflow/lite/micro/kernels/svdf.cc
+
+# cmsis 优化算子
+tensorflow/lite/micro/kernels/cmsis-nn/add.cc
+tensorflow/lite/micro/kernels/cmsis-nn/conv.cc
+tensorflow/lite/micro/kernels/cmsis-nn/depthwise_conv.cc
+tensorflow/lite/micro/kernels/cmsis-nn/fully_connected.cc
+tensorflow/lite/micro/kernels/cmsis-nn/mul.cc
+tensorflow/lite/micro/kernels/cmsis-nn/pooling.cc
+tensorflow/lite/micro/kernels/cmsis-nn/softmax.cc
+tensorflow/lite/micro/kernels/cmsis-nn/svdf.cc
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_nn_activations_q15.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_nn_activations_q7.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu6_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu_q15.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ActivationFunctions/arm_relu_q7.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/BasicMathFunctions/arm_elementwise_add_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/BasicMathFunctions/arm_elementwise_mul_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_w.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_x.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_y.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConcatenationFunctions/arm_concatenation_s8_z.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1_x_n_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1x1_HWC_q7_fast_nonsquare.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_1x1_s8_fast.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_basic.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q15_fast_nonsquare.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_RGB.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_basic_nonsquare.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_HWC_q7_fast_nonsquare.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_convolve_wrapper_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_3x3_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_s8_opt.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_u8_basic_ver1.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_conv_wrapper_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_depthwise_separable_conv_HWC_q7_nonsquare.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_depthwise_conv_s8_core.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_q7_q15_reordered.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_s8_s16.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_kernel_s8_s16_reordered.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ConvolutionFunctions/arm_nn_mat_mult_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_mat_q7_vec_q15_opt.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q15_opt.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_q7_opt.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/FullyConnectedFunctions/arm_fully_connected_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_accumulate_q7_to_q15.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_add_q7.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_depthwise_conv_nt_t_padded_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_depthwise_conv_nt_t_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mul_core_1x_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mul_core_4x_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mult_nt_t_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mult_q15.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mult_q7.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_vec_mat_mult_t_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nntables.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_no_shift.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_reordered_no_shift.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_reordered_with_offset.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_q7_to_q15_with_offset.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_avgpool_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_max_pool_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/PoolingFunctions/arm_pool_q7_HWC.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/ReshapeFunctions/arm_reshape_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SVDFunctions/arm_svdf_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q15.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_q7.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_s8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_u8.c
+tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/SoftmaxFunctions/arm_softmax_with_batch_q7.c
+```

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/LICENSE

@@ -0,0 +1,203 @@
+Copyright 2019 The TensorFlow Authors.  All rights reserved.
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/README_MAKE.md

@@ -0,0 +1,29 @@
+# TensorFlow Lite Micro Make Project
+
+This folder has been autogenerated by TensorFlow, and contains source, header,
+and project files needed to build a single TensorFlow Lite Micro target using
+the make tool.
+
+## Usage
+
+To build this, run:
+
+```
+make
+```
+
+This should attempt to build the target locally on your platform, using the
+standard Makefile variables like CFLAGS, CC, CXX, and so on.
+
+## Project Generation
+
+See
+[tensorflow/lite/micro](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/micro)
+for details on how projects like this can be generated from the main source
+tree.
+
+## License
+
+TensorFlow's code is covered by the Apache2 License included in the repository,
+and third party dependencies are covered by their respective licenses, in the
+third_party folder of this package.

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/core/public/version.h

@@ -0,0 +1,139 @@
+/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_CORE_PUBLIC_VERSION_H_
+#define TENSORFLOW_CORE_PUBLIC_VERSION_H_
+
+// TensorFlow uses semantic versioning, see http://semver.org/.
+
+// Also update tensorflow/tensorflow.bzl and
+// tensorflow/tools/pip_package/setup.py
+#define TF_MAJOR_VERSION 2
+#define TF_MINOR_VERSION 5
+#define TF_PATCH_VERSION 0
+
+// TF_VERSION_SUFFIX is non-empty for pre-releases (e.g. "-alpha", "-alpha.1",
+// "-beta", "-rc", "-rc.1")
+#define TF_VERSION_SUFFIX ""
+
+#define TF_STR_HELPER(x) #x
+#define TF_STR(x) TF_STR_HELPER(x)
+
+// e.g. "0.5.0" or "0.6.0-alpha".
+#define TF_VERSION_STRING                                            \
+  (TF_STR(TF_MAJOR_VERSION) "." TF_STR(TF_MINOR_VERSION) "." TF_STR( \
+      TF_PATCH_VERSION) TF_VERSION_SUFFIX)
+
+// GraphDef compatibility versions (the versions field in graph.proto).
+//
+// Each graph has producer and min_consumer versions, and each
+// consumer has its own version and a min_producer.  In addition, graphs can
+// mark specific consumer versions as bad (to prevent bugs from executing).
+// A consumer will execute a graph if the consumer's version is at least the
+// graph's min_consumer, the graph's producer version is at least the consumer's
+// min_producer, and the consumer version isn't specifically disallowed by the
+// graph.
+//
+// By default, newly created graphs have producer version TF_GRAPH_DEF_VERSION
+// min_consumer TF_GRAPH_DEF_MIN_CONSUMER, and no other bad consumer versions.
+//
+// Version history:
+//
+// 0. Graphs created before GraphDef versioning
+// 1. First real version (2dec2015)
+// 2. adjust_contrast only takes float, doesn't perform clamping (11dec2015)
+// 3. Remove TileGrad, since it was equivalent to reduce_sum (30dec2015)
+// 4. When support for this version is removed, we can safely make AttrValue
+//    parsing more strict with respect to empty list values (see
+//    111635679, 7jan2016).
+// 5. Graphs are wholly-validated during Session::Create() (7jan2016).
+// 6. TensorFlow is scalar strict within Google (27jan2016).
+// 7. Remove TopK in favor of TopKV2 (5feb2016).
+// 8. Replace RandomCrop from C++ with pure Python (5feb2016).
+// 9. Deprecate batch_norm_with_global_normalization (16feb2016).
+// 10. Deprecate conv3d_backprop_{filter,input} (10jun2016).
+// 11. Deprecate {batch}_self_adjoint_eig (3aug2016).
+// 12. Graph consumers understand the node_def field of FunctionDef (22aug2016).
+// 13. Deprecate multiple batch linear algebra ops (9sep2016).
+// 14. Deprecate batch_matrix_* ops. (10sep2016).
+// 15. Deprecate batch_fft_* ops. (14sep2016).
+// 16. Deprecate tensor_array (v1) ops in favor of v2 (10nov2016).
+// 17. Deprecate inv (11nov2016).
+// 17. Expose reverse_v2 (10nov2016)
+// 18. Add VariableV2 (30nov2016)
+// 19. Deprecated ops created by models moved out of core SkipGram, NegTrain.
+//     (08dec2016)
+// 20. Catch all version 1.0 changes to Python API generation. SplitV is now
+//     used for tf.split, ReverseV2 is now used by tf.reverse, ConcatV2 is
+//     now used by tf.concat. Graphs use flooring
+//     division and mod semantics. TensorArrayV3. (12dec2016)
+//     Also considered the version for when it is required for reduction
+//     ops' indices to be scalar or vector, and not higher rank.
+//     Some earlier graph def versions allowed this.
+// 21. Dropped FunctionDef.Node support, switched to node_def introduced
+//     in version 12. (11jan2017)
+// 22. Placeholder now can specify and enforce scalar and partial
+//     shapes, particularly when restoring a graph from GraphDef
+//     produced at version 22 or later.  (04/10/2016)
+// 23. Remove NonMaxSuppression in favor of NonMaxSuppressionV2.
+// 24. Deprecate lookup ops (v1) ops in favor of v2 (30may2017)
+// 25. Deprecate stack (v1) ops in favor of v2 (2017/6/15).
+// 25. Deprecate RandomPoisson (v1) ops in favor of v2 (2017/10/25).
+// 26. Add a bool 'stripped_default_attrs' to MetaInfoDef indicating
+//     whether default-valued attrs have been stripped from the nodes in the
+//     GraphDef. (7dec2017)
+// 27. Deprecate TensorArray ops v2 in favor of v3 and deprecated io_ops
+//     deprecated in favor of V2 ops. (2018/01/23)
+// 28. Deprecate MatrixExponential op in favor of Python implementation.
+//     (2018/08/21).
+// (2019/02/15). Added `control_ret` field to FunctionDef proto, and
+//     `control_output` field to OpDef proto.
+// 29. Deprecate StatefulStandardNormal op in favor of StatefulStandardNormalV2.
+//     (2019/03/25).
+// (2019/04/17). Added `arg_attr` field to FunctionDefProto.
+// 30. (2019/05/09) First date based GraphDef version. GraphDef
+//     versions advance by 1 each day after this point.
+
+#define TF_GRAPH_DEF_VERSION_MIN_PRODUCER 0
+#define TF_GRAPH_DEF_VERSION_MIN_CONSUMER 0
+#define TF_GRAPH_DEF_VERSION 639  // Updated: 2021/1/7
+
+// Checkpoint compatibility versions (the versions field in SavedSliceMeta).
+//
+// The checkpoint versions have the same semantics as GraphDef versions, but the
+// numbering scheme is separate.  We have no plans to ever deprecate checkpoint
+// versions, but it's good to have this in place in case we ever need to.
+//
+// Version history:
+//
+// 0. Checkpoints saved before checkpoint versioning.
+// 1. First real version (10feb2015).
+#define TF_CHECKPOINT_VERSION_MIN_PRODUCER 0
+#define TF_CHECKPOINT_VERSION_MIN_CONSUMER 0
+#define TF_CHECKPOINT_VERSION 1
+
+/// Version query functions (defined in generated version_info.cc)
+
+// Host compiler version (declared elsewhere to be __VERSION__)
+extern const char* tf_compiler_version();
+// The git commit designator when tensorflow was built
+// If no git repository, this will be "internal".
+extern const char* tf_git_version();
+// Value of the _GLIBCXX_USE_CXX11_ABI flag, or 0 if it's not set.
+extern int tf_cxx11_abi_flag();
+// Returns 1 if build is monolithic, or 0 otherwise.
+extern int tf_monolithic_build();
+
+#endif  // TENSORFLOW_CORE_PUBLIC_VERSION_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/c/builtin_op_data.h

@@ -0,0 +1,484 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_C_BUILTIN_OP_DATA_H_
+#define TENSORFLOW_LITE_C_BUILTIN_OP_DATA_H_
+
+#include <stdint.h>
+
+#include "tensorflow/lite/c/common.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif  // __cplusplus
+
+// TfLiteReshapeParams can't have dynamic data so we fix the maximum possible
+// number of dimensions.
+#define TFLITE_RESHAPE_PARAMS_MAX_DIMENSION_COUNT 8
+
+// TODO(aselle): Consider using "if this then that" for testing.
+
+// Useful placeholder to put in otherwise empty structs to avoid size warnings.
+typedef struct {
+  char dummy;
+} EmptyStructPlaceholder;
+
+// IMPORTANT: All new members of structs must be added at the end to ensure
+// backwards compatibility.
+
+// Possible padding types (for convolutions)
+typedef enum {
+  kTfLitePaddingUnknown = 0,
+  kTfLitePaddingSame,
+  kTfLitePaddingValid,
+} TfLitePadding;
+
+typedef enum {
+  kTfLiteMirrorPaddingUnknown = 0,
+  kTfLiteMirrorPaddingReflect,
+  kTfLiteMirrorPaddingSymmetric,
+} TfLiteMirrorPaddingMode;
+
+// TODO(b/130259536): We should move this out of builtin_op_data.
+typedef struct {
+  int width;
+  int height;
+  int width_offset;
+  int height_offset;
+} TfLitePaddingValues;
+
+typedef struct {
+  TfLiteMirrorPaddingMode mode;
+} TfLiteMirrorPaddingParams;
+
+// Possible fused activation functions.
+// TODO(aselle): rename to TfLiteActivation
+typedef enum {
+  kTfLiteActNone = 0,
+  kTfLiteActRelu,
+  kTfLiteActReluN1To1,                    // min(max(-1, x), 1)
+  kTfLiteActRelu6,                        // min(max(0, x), 6)
+  kTfLiteActTanh,
+  kTfLiteActSignBit,
+  kTfLiteActSigmoid,
+} TfLiteFusedActivation;
+
+typedef struct {
+  // Parameters for CONV_2D version 1.
+  TfLitePadding padding;
+  int stride_width;
+  int stride_height;
+  TfLiteFusedActivation activation;
+
+  // Parameters for CONV_2D version 2.
+  // Note: Version 2 supports dilation values not equal to 1.
+  int dilation_width_factor;
+  int dilation_height_factor;
+} TfLiteConvParams;
+
+typedef struct {
+  TfLitePadding padding;
+  int stride_width;
+  int stride_height;
+  int filter_width;
+  int filter_height;
+  TfLiteFusedActivation activation;
+  struct {
+    TfLitePaddingValues padding;
+  } computed;
+} TfLitePoolParams;
+
+typedef struct {
+  // Parameters for DepthwiseConv version 1 or above.
+  TfLitePadding padding;
+  int stride_width;
+  int stride_height;
+  // `depth_multiplier` is redundant. It's used by CPU kernels in
+  // TensorFlow 2.0 or below, but ignored in versions above.
+  //
+  // The information can be deduced from the shape of input and the shape of
+  // weights. Since the TFLiteConverter toolchain doesn't support partially
+  // specified shapes, relying on `depth_multiplier` stops us from supporting
+  // graphs with dynamic shape tensors.
+  //
+  // Note: Some of the delegates (e.g. NNAPI, GPU) are still relying on this
+  // field.
+  int depth_multiplier;
+  TfLiteFusedActivation activation;
+  // Parameters for DepthwiseConv version 2 or above.
+  int dilation_width_factor;
+  int dilation_height_factor;
+} TfLiteDepthwiseConvParams;
+
+typedef struct {
+  int rank;
+  TfLiteFusedActivation activation;
+
+  // Parameter for SVDF version 4.
+  bool asymmetric_quantize_inputs;
+} TfLiteSVDFParams;
+
+typedef struct {
+  TfLiteFusedActivation activation;
+
+  // Parameter for RNN version 3.
+  bool asymmetric_quantize_inputs;
+} TfLiteRNNParams;
+
+typedef struct {
+  bool time_major;
+  TfLiteFusedActivation activation;
+
+  // Parameter for Sequence RNN version 3.
+  bool asymmetric_quantize_inputs;
+} TfLiteSequenceRNNParams;
+
+typedef struct {
+  bool time_major;
+  TfLiteFusedActivation activation;
+  bool merge_outputs;
+
+  // Parameter for Bidirectional RNN verison 3.
+  bool asymmetric_quantize_inputs;
+} TfLiteBidirectionalSequenceRNNParams;
+
+typedef enum {
+  kTfLiteFullyConnectedWeightsFormatDefault = 0,
+  kTfLiteFullyConnectedWeightsFormatShuffled4x16Int8 = 1,
+} TfLiteFullyConnectedWeightsFormat;
+
+typedef struct {
+  // Parameters for FullyConnected version 1 or above.
+  TfLiteFusedActivation activation;
+
+  // Parameters for FullyConnected version 2 or above.
+  TfLiteFullyConnectedWeightsFormat weights_format;
+
+  // Parameters for FullyConnected version 5 or above.
+  // If set to true, then the number of dimensions in the input and the output
+  // tensors are the same. Furthermore, all but the last dimension of the input
+  // and output shapes will be equal.
+  bool keep_num_dims;
+
+  // Parameters for FullyConnected version 7 or above.
+  // If set to true and the weights are quantized, then non constant inputs
+  // are quantized at evaluation time with asymmetric quantization.
+  bool asymmetric_quantize_inputs;
+} TfLiteFullyConnectedParams;
+
+typedef enum {
+  kTfLiteLshProjectionUnknown = 0,
+  kTfLiteLshProjectionSparse = 1,
+  kTfLiteLshProjectionDense = 2,
+} TfLiteLSHProjectionType;
+
+typedef struct {
+  TfLiteLSHProjectionType type;
+} TfLiteLSHProjectionParams;
+
+typedef struct {
+  float beta;
+} TfLiteSoftmaxParams;
+
+typedef struct {
+  int axis;
+  TfLiteFusedActivation activation;
+} TfLiteConcatenationParams;
+
+typedef struct {
+  TfLiteFusedActivation activation;
+  // Parameter added for the version 4.
+  bool pot_scale_int16;
+} TfLiteAddParams;
+
+typedef struct {
+  EmptyStructPlaceholder placeholder;
+} TfLiteSpaceToBatchNDParams;
+
+typedef struct {
+  EmptyStructPlaceholder placeholder;
+} TfLiteBatchToSpaceNDParams;
+
+typedef struct {
+  bool adj_x;
+  bool adj_y;
+  // Parameters for BatchMatMul version 4 or above.
+  // If set to true and the weights are quantized, then non constant inputs
+  // are quantized at evaluation time with asymmetric quantization.
+  bool asymmetric_quantize_inputs;
+} TfLiteBatchMatMulParams;
+
+typedef struct {
+  TfLiteFusedActivation activation;
+} TfLiteMulParams;
+
+typedef struct {
+  TfLiteFusedActivation activation;
+  // Parameter added for the version 5.
+  bool pot_scale_int16;
+} TfLiteSubParams;
+
+typedef struct {
+  TfLiteFusedActivation activation;
+} TfLiteDivParams;
+
+typedef struct {
+  TfLiteFusedActivation activation;
+} TfLiteL2NormParams;
+
+typedef struct {
+  int radius;
+  float bias;
+  float alpha;
+  float beta;
+} TfLiteLocalResponseNormParams;
+
+typedef enum {
+  kTfLiteLSTMFullKernel = 0,
+  kTfLiteLSTMBasicKernel
+} TfLiteLSTMKernelType;
+
+typedef struct {
+  // Parameters for LSTM version 1.
+  TfLiteFusedActivation activation;
+  float cell_clip;
+  float proj_clip;
+
+  // Parameters for LSTM version 2.
+  // kTfLiteLSTMBasicKernel is only supported in version 2 or above.
+  TfLiteLSTMKernelType kernel_type;
+
+  // Parameters for LSTM version 4.
+  bool asymmetric_quantize_inputs;
+} TfLiteLSTMParams;
+
+typedef struct {
+  // Parameters needed for the underlying LSTM.
+  TfLiteFusedActivation activation;
+  float cell_clip;
+  float proj_clip;
+
+  // If set to true then the first dimension is time, otherwise batch.
+  bool time_major;
+
+  // Parameter for unidirectional sequence RNN version 3.
+  bool asymmetric_quantize_inputs;
+} TfLiteUnidirectionalSequenceLSTMParams;
+
+typedef struct {
+  // Parameters supported by version 1:
+  // Parameters inherited for the LSTM kernel.
+  TfLiteFusedActivation activation;
+  float cell_clip;
+  float proj_clip;
+
+  // If true, store the outputs of both directions in the first output.
+  bool merge_outputs;
+
+  // Parameters supported by version 2:
+  // If set to true then the first dimension is time, otherwise batch.
+  bool time_major;
+
+  // Parameters supported by version 4:
+  // If set to true, then hybrid ops use asymmetric quantization for inputs.
+  bool asymmetric_quantize_inputs;
+} TfLiteBidirectionalSequenceLSTMParams;
+
+typedef struct {
+  bool align_corners;
+  // half_pixel_centers assumes pixels are of half the actual dimensions, and
+  // yields more accurate resizes. Corresponds to the same argument for the
+  // original TensorFlow op in TF2.0.
+  bool half_pixel_centers;
+} TfLiteResizeBilinearParams;
+
+typedef struct {
+  bool align_corners;
+  bool half_pixel_centers;
+} TfLiteResizeNearestNeighborParams;
+
+typedef struct {
+  EmptyStructPlaceholder placeholder;
+} TfLitePadParams;
+
+typedef struct {
+  EmptyStructPlaceholder placeholder;
+} TfLitePadV2Params;
+
+typedef struct {
+  // TODO(ahentz): We can't have dynamic data in this struct, at least not yet.
+  // For now we will fix the maximum possible number of dimensions.
+  int shape[TFLITE_RESHAPE_PARAMS_MAX_DIMENSION_COUNT];
+  int num_dimensions;
+} TfLiteReshapeParams;
+
+typedef struct {
+  int ngram_size;
+  int max_skip_size;
+  bool include_all_ngrams;
+} TfLiteSkipGramParams;
+
+typedef struct {
+  int block_size;
+} TfLiteSpaceToDepthParams;
+
+typedef struct {
+  int block_size;
+} TfLiteDepthToSpaceParams;
+
+typedef struct {
+  TfLiteType in_data_type;
+  TfLiteType out_data_type;
+} TfLiteCastParams;
+
+typedef enum {
+  kTfLiteCombinerTypeSum = 0,
+  kTfLiteCombinerTypeMean = 1,
+  kTfLiteCombinerTypeSqrtn = 2,
+} TfLiteCombinerType;
+
+typedef struct {
+  TfLiteCombinerType combiner;
+} TfLiteEmbeddingLookupSparseParams;
+
+typedef struct {
+  int axis;
+} TfLiteGatherParams;
+
+typedef struct {
+  EmptyStructPlaceholder placeholder;
+} TfLiteTransposeParams;
+
+typedef struct {
+  bool keep_dims;
+} TfLiteReducerParams;
+
+typedef struct {
+  int num_splits;
+} TfLiteSplitParams;
+
+typedef struct {
+  int num_splits;
+} TfLiteSplitVParams;
+
+typedef struct {
+  // TODO(ahentz): We can't have dynamic data in this struct, at least not yet.
+  // For now we will fix the maximum possible number of dimensions.
+  int squeeze_dims[8];
+  int num_squeeze_dims;
+} TfLiteSqueezeParams;
+
+typedef struct {
+  int begin_mask;
+  int end_mask;
+  int ellipsis_mask;
+  int new_axis_mask;
+  int shrink_axis_mask;
+} TfLiteStridedSliceParams;
+
+typedef struct {
+  TfLiteType output_type;
+} TfLiteArgMaxParams;
+
+typedef struct {
+  TfLiteType output_type;
+} TfLiteArgMinParams;
+
+typedef struct {
+  TfLitePadding padding;
+  int stride_width;
+  int stride_height;
+} TfLiteTransposeConvParams;
+
+typedef struct {
+  bool validate_indices;
+} TfLiteSparseToDenseParams;
+
+typedef struct {
+  TfLiteType out_type;
+} TfLiteShapeParams;
+
+typedef struct {
+  EmptyStructPlaceholder placeholder;
+} TfLiteRankParams;
+
+typedef struct {
+  // Parameters supported by version 1:
+  float min;
+  float max;
+  int num_bits;
+
+  // Parameters supported by version 2:
+  bool narrow_range;
+} TfLiteFakeQuantParams;
+
+typedef struct {
+  int values_count;
+  int axis;
+} TfLitePackParams;
+
+typedef struct {
+  int axis;
+} TfLiteOneHotParams;
+
+typedef struct {
+  int num;
+  int axis;
+} TfLiteUnpackParams;
+
+typedef struct {
+  float alpha;
+} TfLiteLeakyReluParams;
+
+typedef struct {
+  TfLiteType index_out_type;
+} TfLiteUniqueParams;
+
+typedef struct {
+  int seq_dim;
+  int batch_dim;
+} TfLiteReverseSequenceParams;
+
+typedef struct {
+  EmptyStructPlaceholder placeholder;
+} TfLiteMatrixDiagParams;
+
+typedef struct {
+  EmptyStructPlaceholder placeholder;
+} TfLiteMatrixSetDiagParams;
+
+typedef struct {
+  int then_subgraph_index;
+  int else_subgraph_index;
+} TfLiteIfParams;
+
+typedef struct {
+  int cond_subgraph_index;
+  int body_subgraph_index;
+} TfLiteWhileParams;
+
+typedef struct {
+  bool exclusive;
+  bool reverse;
+} TfLiteCumsumParams;
+
+typedef struct {
+  int init_subgraph_index;
+} TfLiteCallOnceParams;
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif  // __cplusplus
+
+#endif  // TENSORFLOW_LITE_C_BUILTIN_OP_DATA_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/c/c_api_types.h

@@ -0,0 +1,92 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// This file declares types used by the pure C inference API defined in c_api.h,
+// some of which are also used in the C++ and C kernel and interpreter APIs.
+
+#ifndef TENSORFLOW_LITE_C_C_API_TYPES_H_
+#define TENSORFLOW_LITE_C_C_API_TYPES_H_
+
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Define TFL_CAPI_EXPORT macro to export a function properly with a shared
+// library.
+#ifdef SWIG
+#define TFL_CAPI_EXPORT
+#else
+#if defined(_WIN32)
+#ifdef TFL_COMPILE_LIBRARY
+#define TFL_CAPI_EXPORT __declspec(dllexport)
+#else
+#define TFL_CAPI_EXPORT __declspec(dllimport)
+#endif  // TFL_COMPILE_LIBRARY
+#else
+#define TFL_CAPI_EXPORT __attribute__((visibility("default")))
+#endif  // _WIN32
+#endif  // SWIG
+
+typedef enum TfLiteStatus {
+  kTfLiteOk = 0,
+
+  // Generally referring to an error in the runtime (i.e. interpreter)
+  kTfLiteError = 1,
+
+  // Generally referring to an error from a TfLiteDelegate itself.
+  kTfLiteDelegateError = 2,
+
+  // Generally referring to an error in applying a delegate due to
+  // incompatibility between runtime and delegate, e.g., this error is returned
+  // when trying to apply a TfLite delegate onto a model graph that's already
+  // immutable.
+  kTfLiteApplicationError = 3
+} TfLiteStatus;
+
+// Types supported by tensor
+typedef enum {
+  kTfLiteNoType = 0,
+  kTfLiteFloat32 = 1,
+  kTfLiteInt32 = 2,
+  kTfLiteUInt8 = 3,
+  kTfLiteInt64 = 4,
+  kTfLiteString = 5,
+  kTfLiteBool = 6,
+  kTfLiteInt16 = 7,
+  kTfLiteComplex64 = 8,
+  kTfLiteInt8 = 9,
+  kTfLiteFloat16 = 10,
+  kTfLiteFloat64 = 11,
+  kTfLiteComplex128 = 12,
+  kTfLiteUInt64 = 13,
+} TfLiteType;
+
+// Legacy. Will be deprecated in favor of TfLiteAffineQuantization.
+// If per-layer quantization is specified this field will still be populated in
+// addition to TfLiteAffineQuantization.
+// Parameters for asymmetric quantization. Quantized values can be converted
+// back to float using:
+//     real_value = scale * (quantized_value - zero_point)
+typedef struct TfLiteQuantizationParams {
+  float scale;
+  int32_t zero_point;
+} TfLiteQuantizationParams;
+
+#ifdef __cplusplus
+}  // extern C
+#endif
+#endif  // TENSORFLOW_LITE_C_C_API_TYPES_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/c/common.c

@@ -0,0 +1,236 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/c/c_api_types.h"
+
+#ifndef TF_LITE_STATIC_MEMORY
+#include <stdlib.h>
+#include <string.h>
+#endif  // TF_LITE_STATIC_MEMORY
+
+int TfLiteIntArrayGetSizeInBytes(int size) {
+  static TfLiteIntArray dummy;
+  return sizeof(dummy) + sizeof(dummy.data[0]) * size;
+}
+
+int TfLiteIntArrayEqual(const TfLiteIntArray* a, const TfLiteIntArray* b) {
+  if (a == b) return 1;
+  if (a == NULL || b == NULL) return 0;
+  return TfLiteIntArrayEqualsArray(a, b->size, b->data);
+}
+
+int TfLiteIntArrayEqualsArray(const TfLiteIntArray* a, int b_size,
+                              const int b_data[]) {
+  if (a == NULL) return (b_size == 0);
+  if (a->size != b_size) return 0;
+  int i = 0;
+  for (; i < a->size; i++)
+    if (a->data[i] != b_data[i]) return 0;
+  return 1;
+}
+
+#ifndef TF_LITE_STATIC_MEMORY
+
+TfLiteIntArray* TfLiteIntArrayCreate(int size) {
+  TfLiteIntArray* ret =
+      (TfLiteIntArray*)malloc(TfLiteIntArrayGetSizeInBytes(size));
+  ret->size = size;
+  return ret;
+}
+
+TfLiteIntArray* TfLiteIntArrayCopy(const TfLiteIntArray* src) {
+  if (!src) return NULL;
+  TfLiteIntArray* ret = TfLiteIntArrayCreate(src->size);
+  if (ret) {
+    memcpy(ret->data, src->data, src->size * sizeof(int));
+  }
+  return ret;
+}
+
+void TfLiteIntArrayFree(TfLiteIntArray* a) { free(a); }
+
+#endif  // TF_LITE_STATIC_MEMORY
+
+int TfLiteFloatArrayGetSizeInBytes(int size) {
+  static TfLiteFloatArray dummy;
+  return sizeof(dummy) + sizeof(dummy.data[0]) * size;
+}
+
+#ifndef TF_LITE_STATIC_MEMORY
+
+TfLiteFloatArray* TfLiteFloatArrayCreate(int size) {
+  TfLiteFloatArray* ret =
+      (TfLiteFloatArray*)malloc(TfLiteFloatArrayGetSizeInBytes(size));
+  ret->size = size;
+  return ret;
+}
+
+void TfLiteFloatArrayFree(TfLiteFloatArray* a) { free(a); }
+
+void TfLiteTensorDataFree(TfLiteTensor* t) {
+  if (t->allocation_type == kTfLiteDynamic ||
+      t->allocation_type == kTfLitePersistentRo) {
+    free(t->data.raw);
+  }
+  t->data.raw = NULL;
+}
+
+void TfLiteQuantizationFree(TfLiteQuantization* quantization) {
+  if (quantization->type == kTfLiteAffineQuantization) {
+    TfLiteAffineQuantization* q_params =
+        (TfLiteAffineQuantization*)(quantization->params);
+    if (q_params->scale) {
+      TfLiteFloatArrayFree(q_params->scale);
+      q_params->scale = NULL;
+    }
+    if (q_params->zero_point) {
+      TfLiteIntArrayFree(q_params->zero_point);
+      q_params->zero_point = NULL;
+    }
+    free(q_params);
+  }
+  quantization->params = NULL;
+  quantization->type = kTfLiteNoQuantization;
+}
+
+void TfLiteSparsityFree(TfLiteSparsity* sparsity) {
+  if (sparsity == NULL) {
+    return;
+  }
+
+  if (sparsity->traversal_order) {
+    TfLiteIntArrayFree(sparsity->traversal_order);
+    sparsity->traversal_order = NULL;
+  }
+
+  if (sparsity->block_map) {
+    TfLiteIntArrayFree(sparsity->block_map);
+    sparsity->block_map = NULL;
+  }
+
+  if (sparsity->dim_metadata) {
+    int i = 0;
+    for (; i < sparsity->dim_metadata_size; i++) {
+      TfLiteDimensionMetadata metadata = sparsity->dim_metadata[i];
+      if (metadata.format == kTfLiteDimSparseCSR) {
+        TfLiteIntArrayFree(metadata.array_segments);
+        metadata.array_segments = NULL;
+        TfLiteIntArrayFree(metadata.array_indices);
+        metadata.array_indices = NULL;
+      }
+    }
+    free(sparsity->dim_metadata);
+    sparsity->dim_metadata = NULL;
+  }
+
+  free(sparsity);
+}
+
+void TfLiteTensorFree(TfLiteTensor* t) {
+  TfLiteTensorDataFree(t);
+  if (t->dims) TfLiteIntArrayFree(t->dims);
+  t->dims = NULL;
+
+  if (t->dims_signature) {
+    TfLiteIntArrayFree((TfLiteIntArray *) t->dims_signature);
+  }
+  t->dims_signature = NULL;
+
+  TfLiteQuantizationFree(&t->quantization);
+  TfLiteSparsityFree(t->sparsity);
+  t->sparsity = NULL;
+}
+
+void TfLiteTensorReset(TfLiteType type, const char* name, TfLiteIntArray* dims,
+                       TfLiteQuantizationParams quantization, char* buffer,
+                       size_t size, TfLiteAllocationType allocation_type,
+                       const void* allocation, bool is_variable,
+                       TfLiteTensor* tensor) {
+  TfLiteTensorFree(tensor);
+  tensor->type = type;
+  tensor->name = name;
+  tensor->dims = dims;
+  tensor->params = quantization;
+  tensor->data.raw = buffer;
+  tensor->bytes = size;
+  tensor->allocation_type = allocation_type;
+  tensor->allocation = allocation;
+  tensor->is_variable = is_variable;
+
+  tensor->quantization.type = kTfLiteNoQuantization;
+  tensor->quantization.params = NULL;
+}
+
+void TfLiteTensorRealloc(size_t num_bytes, TfLiteTensor* tensor) {
+  if (tensor->allocation_type != kTfLiteDynamic &&
+      tensor->allocation_type != kTfLitePersistentRo) {
+    return;
+  }
+  // TODO(b/145340303): Tensor data should be aligned.
+  if (!tensor->data.raw) {
+    tensor->data.raw = malloc(num_bytes);
+  } else if (num_bytes > tensor->bytes) {
+    tensor->data.raw = realloc(tensor->data.raw, num_bytes);
+  }
+  tensor->bytes = num_bytes;
+}
+#endif  // TF_LITE_STATIC_MEMORY
+
+const char* TfLiteTypeGetName(TfLiteType type) {
+  switch (type) {
+    case kTfLiteNoType:
+      return "NOTYPE";
+    case kTfLiteFloat32:
+      return "FLOAT32";
+    case kTfLiteInt16:
+      return "INT16";
+    case kTfLiteInt32:
+      return "INT32";
+    case kTfLiteUInt8:
+      return "UINT8";
+    case kTfLiteInt8:
+      return "INT8";
+    case kTfLiteInt64:
+      return "INT64";
+    case kTfLiteUInt64:
+      return "UINT64";
+    case kTfLiteBool:
+      return "BOOL";
+    case kTfLiteComplex64:
+      return "COMPLEX64";
+    case kTfLiteComplex128:
+      return "COMPLEX128";
+    case kTfLiteString:
+      return "STRING";
+    case kTfLiteFloat16:
+      return "FLOAT16";
+    case kTfLiteFloat64:
+      return "FLOAT64";
+  }
+  return "Unknown type";
+}
+
+TfLiteDelegate TfLiteDelegateCreate() {
+  TfLiteDelegate d = {
+      .data_ = NULL,
+      .Prepare = NULL,
+      .CopyFromBufferHandle = NULL,
+      .CopyToBufferHandle = NULL,
+      .FreeBufferHandle = NULL,
+      .flags = kTfLiteDelegateFlagsNone,
+  };
+  return d;
+}

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/c/common.h

@@ -0,0 +1,913 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+// This file defines common C types and APIs for implementing operations,
+// delegates and other constructs in TensorFlow Lite. The actual operations and
+// delegates can be defined using C++, but the interface between the interpreter
+// and the operations are C.
+//
+// Summary of abstractions
+// TF_LITE_ENSURE - Self-sufficient error checking
+// TfLiteStatus - Status reporting
+// TfLiteIntArray - stores tensor shapes (dims),
+// TfLiteContext - allows an op to access the tensors
+// TfLiteTensor - tensor (a multidimensional array)
+// TfLiteNode - a single node or operation
+// TfLiteRegistration - the implementation of a conceptual operation.
+// TfLiteDelegate - allows delegation of nodes to alternative backends.
+//
+// Some abstractions in this file are created and managed by Interpreter.
+//
+// NOTE: The order of values in these structs are "semi-ABI stable". New values
+// should be added only to the end of structs and never reordered.
+
+#ifndef TENSORFLOW_LITE_C_COMMON_H_
+#define TENSORFLOW_LITE_C_COMMON_H_
+
+#include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#include "tensorflow/lite/c/c_api_types.h"  // IWYU pragma: export
+
+#ifdef __cplusplus
+extern "C" {
+#endif  // __cplusplus
+
+// The list of external context types known to TF Lite. This list exists solely
+// to avoid conflicts and to ensure ops can share the external contexts they
+// need. Access to the external contexts is controlled by one of the
+// corresponding support files.
+typedef enum TfLiteExternalContextType {
+  kTfLiteEigenContext = 0,       // include eigen_support.h to use.
+  kTfLiteGemmLowpContext = 1,    // include gemm_support.h to use.
+  kTfLiteEdgeTpuContext = 2,     // Placeholder for Edge TPU support.
+  kTfLiteCpuBackendContext = 3,  // include cpu_backend_context.h to use.
+  kTfLiteMaxExternalContexts = 4
+} TfLiteExternalContextType;
+
+// Forward declare so dependent structs and methods can reference these types
+// prior to the struct definitions.
+struct TfLiteContext;
+struct TfLiteDelegate;
+struct TfLiteRegistration;
+
+// An external context is a collection of information unrelated to the TF Lite
+// framework, but useful to a subset of the ops. TF Lite knows very little
+// about the actual contexts, but it keeps a list of them, and is able to
+// refresh them if configurations like the number of recommended threads
+// change.
+typedef struct TfLiteExternalContext {
+  TfLiteExternalContextType type;
+  TfLiteStatus (*Refresh)(struct TfLiteContext* context);
+} TfLiteExternalContext;
+
+#define kTfLiteOptionalTensor (-1)
+
+// Fixed size list of integers. Used for dimensions and inputs/outputs tensor
+// indices
+typedef struct TfLiteIntArray {
+  int size;
+// gcc 6.1+ have a bug where flexible members aren't properly handled
+// https://github.com/google/re2/commit/b94b7cd42e9f02673cd748c1ac1d16db4052514c
+#if (!defined(__clang__) && defined(__GNUC__) && __GNUC__ == 6 && \
+     __GNUC_MINOR__ >= 1) ||                                      \
+    defined(HEXAGON) || (__clang_major__ == 7 && __clang_minor__ == 1)
+  int data[0];
+#else
+  int data[];
+#endif
+} TfLiteIntArray;
+
+// Given the size (number of elements) in a TfLiteIntArray, calculate its size
+// in bytes.
+int TfLiteIntArrayGetSizeInBytes(int size);
+
+#ifndef TF_LITE_STATIC_MEMORY
+// Create a array of a given `size` (uninitialized entries).
+// This returns a pointer, that you must free using TfLiteIntArrayFree().
+TfLiteIntArray* TfLiteIntArrayCreate(int size);
+#endif
+
+// Check if two intarrays are equal. Returns 1 if they are equal, 0 otherwise.
+int TfLiteIntArrayEqual(const TfLiteIntArray* a, const TfLiteIntArray* b);
+
+// Check if an intarray equals an array. Returns 1 if equals, 0 otherwise.
+int TfLiteIntArrayEqualsArray(const TfLiteIntArray* a, int b_size,
+                              const int b_data[]);
+
+#ifndef TF_LITE_STATIC_MEMORY
+// Create a copy of an array passed as `src`.
+// You are expected to free memory with TfLiteIntArrayFree
+TfLiteIntArray* TfLiteIntArrayCopy(const TfLiteIntArray* src);
+
+// Free memory of array `a`.
+void TfLiteIntArrayFree(TfLiteIntArray* a);
+#endif  // TF_LITE_STATIC_MEMORY
+
+// Fixed size list of floats. Used for per-channel quantization.
+typedef struct TfLiteFloatArray {
+  int size;
+// gcc 6.1+ have a bug where flexible members aren't properly handled
+// https://github.com/google/re2/commit/b94b7cd42e9f02673cd748c1ac1d16db4052514c
+// This also applies to the toolchain used for Qualcomm Hexagon DSPs.
+#if !defined(__clang__) && defined(__GNUC__) && __GNUC__ == 6 && \
+    __GNUC_MINOR__ >= 1
+  float data[0];
+#else
+  float data[];
+#endif
+} TfLiteFloatArray;
+
+// Given the size (number of elements) in a TfLiteFloatArray, calculate its size
+// in bytes.
+int TfLiteFloatArrayGetSizeInBytes(int size);
+
+#ifndef TF_LITE_STATIC_MEMORY
+// Create a array of a given `size` (uninitialized entries).
+// This returns a pointer, that you must free using TfLiteFloatArrayFree().
+TfLiteFloatArray* TfLiteFloatArrayCreate(int size);
+
+// Free memory of array `a`.
+void TfLiteFloatArrayFree(TfLiteFloatArray* a);
+#endif  // TF_LITE_STATIC_MEMORY
+
+// Since we must not depend on any libraries, define a minimal subset of
+// error macros while avoiding names that have pre-conceived meanings like
+// assert and check.
+
+// Try to make all reporting calls through TF_LITE_KERNEL_LOG rather than
+// calling the context->ReportError function directly, so that message strings
+// can be stripped out if the binary size needs to be severely optimized.
+#ifndef TF_LITE_STRIP_ERROR_STRINGS
+#define TF_LITE_KERNEL_LOG(context, ...)            \
+  do {                                              \
+    (context)->ReportError((context), __VA_ARGS__); \
+  } while (false)
+
+#define TF_LITE_MAYBE_KERNEL_LOG(context, ...)        \
+  do {                                                \
+    if ((context) != nullptr) {                       \
+      (context)->ReportError((context), __VA_ARGS__); \
+    }                                                 \
+  } while (false)
+#else  // TF_LITE_STRIP_ERROR_STRINGS
+#define TF_LITE_KERNEL_LOG(context, ...)
+#define TF_LITE_MAYBE_KERNEL_LOG(context, ...)
+#endif  // TF_LITE_STRIP_ERROR_STRINGS
+
+// Check whether value is true, and if not return kTfLiteError from
+// the current function (and report the error string msg).
+#define TF_LITE_ENSURE_MSG(context, value, msg)        \
+  do {                                                 \
+    if (!(value)) {                                    \
+      TF_LITE_KERNEL_LOG((context), __FILE__ " " msg); \
+      return kTfLiteError;                             \
+    }                                                  \
+  } while (0)
+
+// Check whether the value `a` is true, and if not return kTfLiteError from
+// the current function, while also reporting the location of the error.
+#define TF_LITE_ENSURE(context, a)                                      \
+  do {                                                                  \
+    if (!(a)) {                                                         \
+      TF_LITE_KERNEL_LOG((context), "%s:%d %s was not true.", __FILE__, \
+                         __LINE__, #a);                                 \
+      return kTfLiteError;                                              \
+    }                                                                   \
+  } while (0)
+
+#define TF_LITE_ENSURE_STATUS(a) \
+  do {                           \
+    const TfLiteStatus s = (a);  \
+    if (s != kTfLiteOk) {        \
+      return s;                  \
+    }                            \
+  } while (0)
+
+// Check whether the value `a == b` is true, and if not return kTfLiteError from
+// the current function, while also reporting the location of the error.
+// `a` and `b` may be evaluated more than once, so no side effects or
+// extremely expensive computations should be done.
+// NOTE: Use TF_LITE_ENSURE_TYPES_EQ if comparing TfLiteTypes.
+#define TF_LITE_ENSURE_EQ(context, a, b)                                   \
+  do {                                                                     \
+    if ((a) != (b)) {                                                      \
+      TF_LITE_KERNEL_LOG((context), "%s:%d %s != %s (%d != %d)", __FILE__, \
+                         __LINE__, #a, #b, (a), (b));                      \
+      return kTfLiteError;                                                 \
+    }                                                                      \
+  } while (0)
+
+#define TF_LITE_ENSURE_TYPES_EQ(context, a, b)                             \
+  do {                                                                     \
+    if ((a) != (b)) {                                                      \
+      TF_LITE_KERNEL_LOG((context), "%s:%d %s != %s (%s != %s)", __FILE__, \
+                         __LINE__, #a, #b, TfLiteTypeGetName(a),           \
+                         TfLiteTypeGetName(b));                            \
+      return kTfLiteError;                                                 \
+    }                                                                      \
+  } while (0)
+
+#define TF_LITE_ENSURE_NEAR(context, a, b, epsilon)                          \
+  do {                                                                       \
+    auto delta = ((a) > (b)) ? ((a) - (b)) : ((b) - (a));                    \
+    if (delta > epsilon) {                                                   \
+      TF_LITE_KERNEL_LOG((context), "%s:%d %s not near %s (%f != %f)",       \
+                         __FILE__, __LINE__, #a, #b, static_cast<double>(a), \
+                         static_cast<double>(b));                            \
+      return kTfLiteError;                                                   \
+    }                                                                        \
+  } while (0)
+
+#define TF_LITE_ENSURE_OK(context, status) \
+  do {                                     \
+    const TfLiteStatus s = (status);       \
+    if ((s) != kTfLiteOk) {                \
+      return s;                            \
+    }                                      \
+  } while (0)
+
+// Single-precision complex data type compatible with the C99 definition.
+typedef struct TfLiteComplex64 {
+  float re, im;  // real and imaginary parts, respectively.
+} TfLiteComplex64;
+
+// Double-precision complex data type compatible with the C99 definition.
+typedef struct TfLiteComplex128 {
+  double re, im;  // real and imaginary parts, respectively.
+} TfLiteComplex128;
+
+// Half precision data type compatible with the C99 definition.
+typedef struct TfLiteFloat16 {
+  uint16_t data;
+} TfLiteFloat16;
+
+// Return the name of a given type, for error reporting purposes.
+const char* TfLiteTypeGetName(TfLiteType type);
+
+// SupportedQuantizationTypes.
+typedef enum TfLiteQuantizationType {
+  // No quantization.
+  kTfLiteNoQuantization = 0,
+  // Affine quantization (with support for per-channel quantization).
+  // Corresponds to TfLiteAffineQuantization.
+  kTfLiteAffineQuantization = 1,
+} TfLiteQuantizationType;
+
+// Structure specifying the quantization used by the tensor, if-any.
+typedef struct TfLiteQuantization {
+  // The type of quantization held by params.
+  TfLiteQuantizationType type;
+  // Holds an optional reference to a quantization param structure. The actual
+  // type depends on the value of the `type` field (see the comment there for
+  // the values and corresponding types).
+  void* params;
+} TfLiteQuantization;
+
+// Parameters for asymmetric quantization across a dimension (i.e per output
+// channel quantization).
+// quantized_dimension specifies which dimension the scales and zero_points
+// correspond to.
+// For a particular value in quantized_dimension, quantized values can be
+// converted back to float using:
+//     real_value = scale * (quantized_value - zero_point)
+typedef struct TfLiteAffineQuantization {
+  TfLiteFloatArray* scale;
+  TfLiteIntArray* zero_point;
+  int32_t quantized_dimension;
+} TfLiteAffineQuantization;
+
+/* A union of pointers that points to memory for a given tensor. */
+typedef union TfLitePtrUnion {
+  /* Do not access these members directly, if possible, use
+   * GetTensorData<TYPE>(tensor) instead, otherwise only access .data, as other
+   * members are deprecated. */
+  int32_t* i32;
+  int64_t* i64;
+  uint64_t* u64;
+  float* f;
+  TfLiteFloat16* f16;
+  double* f64;
+  char* raw;
+  const char* raw_const;
+  uint8_t* uint8;
+  bool* b;
+  int16_t* i16;
+  TfLiteComplex64* c64;
+  TfLiteComplex128* c128;
+  int8_t* int8;
+  /* Only use this member. */
+  void* data;
+} TfLitePtrUnion;
+
+// Memory allocation strategies.
+//  * kTfLiteMmapRo: Read-only memory-mapped data, or data externally allocated.
+//  * kTfLiteArenaRw: Arena allocated with no guarantees about persistence,
+//        and available during eval.
+//  * kTfLiteArenaRwPersistent: Arena allocated but persistent across eval, and
+//        only available during eval.
+//  * kTfLiteDynamic: Allocated during eval, or for string tensors.
+//  * kTfLitePersistentRo: Allocated and populated during prepare. This is
+//        useful for tensors that can be computed during prepare and treated
+//        as constant inputs for downstream ops (also in prepare).
+//  * kTfLiteCustom: Custom memory allocation provided by the user. See
+//        TfLiteCustomAllocation below.
+typedef enum TfLiteAllocationType {
+  kTfLiteMemNone = 0,
+  kTfLiteMmapRo,
+  kTfLiteArenaRw,
+  kTfLiteArenaRwPersistent,
+  kTfLiteDynamic,
+  kTfLitePersistentRo,
+  kTfLiteCustom,
+} TfLiteAllocationType;
+
+// The delegates should use zero or positive integers to represent handles.
+// -1 is reserved from unallocated status.
+typedef int TfLiteBufferHandle;
+enum {
+  kTfLiteNullBufferHandle = -1,
+};
+
+// Storage format of each dimension in a sparse tensor.
+typedef enum TfLiteDimensionType {
+  kTfLiteDimDense = 0,
+  kTfLiteDimSparseCSR,
+} TfLiteDimensionType;
+
+// Metadata to encode each dimension in a sparse tensor.
+typedef struct TfLiteDimensionMetadata {
+  TfLiteDimensionType format;
+  int dense_size;
+  TfLiteIntArray* array_segments;
+  TfLiteIntArray* array_indices;
+} TfLiteDimensionMetadata;
+
+// Parameters used to encode a sparse tensor. For detailed explanation of each
+// field please refer to lite/schema/schema.fbs.
+typedef struct TfLiteSparsity {
+  TfLiteIntArray* traversal_order;
+  TfLiteIntArray* block_map;
+  TfLiteDimensionMetadata* dim_metadata;
+  int dim_metadata_size;
+} TfLiteSparsity;
+
+// Defines a custom memory allocation not owned by the runtime.
+// `data` should be aligned to kDefaultTensorAlignment defined in
+// lite/util.h. (Currently 64 bytes)
+// NOTE: See Interpreter.SetCustomAllocationForTensor for details on usage.
+typedef struct TfLiteCustomAllocation {
+  void* data;
+  size_t bytes;
+} TfLiteCustomAllocation;
+
+// A tensor in the interpreter system which is a wrapper around a buffer of
+// data including a dimensionality (or NULL if not currently defined).
+#ifndef TF_LITE_STATIC_MEMORY
+typedef struct TfLiteTensor {
+  // The data type specification for data stored in `data`. This affects
+  // what member of `data` union should be used.
+  TfLiteType type;
+  // A union of data pointers. The appropriate type should be used for a typed
+  // tensor based on `type`.
+  TfLitePtrUnion data;
+  // A pointer to a structure representing the dimensionality interpretation
+  // that the buffer should have. NOTE: the product of elements of `dims`
+  // and the element datatype size should be equal to `bytes` below.
+  TfLiteIntArray* dims;
+  // Quantization information.
+  TfLiteQuantizationParams params;
+  // How memory is mapped
+  //  kTfLiteMmapRo: Memory mapped read only.
+  //  i.e. weights
+  //  kTfLiteArenaRw: Arena allocated read write memory
+  //  (i.e. temporaries, outputs).
+  TfLiteAllocationType allocation_type;
+  // The number of bytes required to store the data of this Tensor. I.e.
+  // (bytes of each element) * dims[0] * ... * dims[n-1].  For example, if
+  // type is kTfLiteFloat32 and dims = {3, 2} then
+  // bytes = sizeof(float) * 3 * 2 = 4 * 3 * 2 = 24.
+  size_t bytes;
+
+  // An opaque pointer to a tflite::MMapAllocation
+  const void* allocation;
+
+  // Null-terminated name of this tensor.
+  const char* name;
+
+  // The delegate which knows how to handle `buffer_handle`.
+  // WARNING: This is an experimental interface that is subject to change.
+  struct TfLiteDelegate* delegate;
+
+  // An integer buffer handle that can be handled by `delegate`.
+  // The value is valid only when delegate is not null.
+  // WARNING: This is an experimental interface that is subject to change.
+  TfLiteBufferHandle buffer_handle;
+
+  // If the delegate uses its own buffer (e.g. GPU memory), the delegate is
+  // responsible to set data_is_stale to true.
+  // `delegate->CopyFromBufferHandle` can be called to copy the data from
+  // delegate buffer.
+  // WARNING: This is an // experimental interface that is subject to change.
+  bool data_is_stale;
+
+  // True if the tensor is a variable.
+  bool is_variable;
+
+  // Quantization information. Replaces params field above.
+  TfLiteQuantization quantization;
+
+  // Parameters used to encode a sparse tensor.
+  // This is optional. The field is NULL if a tensor is dense.
+  // WARNING: This is an experimental interface that is subject to change.
+  TfLiteSparsity* sparsity;
+
+  // Optional. Encodes shapes with unknown dimensions with -1. This field is
+  // only populated when unknown dimensions exist in a read-write tensor (i.e.
+  // an input or output tensor). (e.g.  `dims` contains [1, 1, 1, 3] and
+  // `dims_signature` contains [1, -1, -1, 3]).
+  const TfLiteIntArray* dims_signature;
+} TfLiteTensor;
+
+// A structure representing an instance of a node.
+// This structure only exhibits the inputs, outputs and user defined data, not
+// other features like the type.
+typedef struct TfLiteNode {
+  // Inputs to this node expressed as indices into the simulator's tensors.
+  TfLiteIntArray* inputs;
+
+  // Outputs to this node expressed as indices into the simulator's tensors.
+  TfLiteIntArray* outputs;
+
+  // intermediate tensors to this node expressed as indices into the simulator's
+  // tensors.
+  TfLiteIntArray* intermediates;
+
+  // Temporary tensors uses during the computations. This usually contains no
+  // tensors, but ops are allowed to change that if they need scratch space of
+  // any sort.
+  TfLiteIntArray* temporaries;
+
+  // Opaque data provided by the node implementer through `Registration.init`.
+  void* user_data;
+
+  // Opaque data provided to the node if the node is a builtin. This is usually
+  // a structure defined in builtin_op_data.h
+  void* builtin_data;
+
+  // Custom initial data. This is the opaque data provided in the flatbuffer.
+  // WARNING: This is an experimental interface that is subject to change.
+  const void* custom_initial_data;
+  int custom_initial_data_size;
+
+  // The pointer to the delegate. This is non-null only when the node is
+  // created by calling `interpreter.ModifyGraphWithDelegate`.
+  // WARNING: This is an experimental interface that is subject to change.
+  struct TfLiteDelegate* delegate;
+} TfLiteNode;
+#else   // defined(TF_LITE_STATIC_MEMORY)?
+// NOTE: This flag is opt-in only at compile time.
+//
+// Specific reduced TfLiteTensor struct for TF Micro runtime. This struct
+// contains only the minimum fields required to initialize and prepare a micro
+// inference graph. The fields in this struct have been ordered from
+// largest-to-smallest for optimal struct sizeof.
+//
+// This struct does not use:
+// - allocation
+// - buffer_handle
+// - data_is_stale
+// - delegate
+// - dims_signature
+// - name
+// - sparsity
+typedef struct TfLiteTensor {
+  // TODO(b/155784997): Consider consolidating these quantization fields:
+  // Quantization information. Replaces params field above.
+  TfLiteQuantization quantization;
+
+  // Quantization information.
+  TfLiteQuantizationParams params;
+
+  // A union of data pointers. The appropriate type should be used for a typed
+  // tensor based on `type`.
+  TfLitePtrUnion data;
+
+  // A pointer to a structure representing the dimensionality interpretation
+  // that the buffer should have. NOTE: the product of elements of `dims`
+  // and the element datatype size should be equal to `bytes` below.
+  TfLiteIntArray* dims;
+
+  // The number of bytes required to store the data of this Tensor. I.e.
+  // (bytes of each element) * dims[0] * ... * dims[n-1].  For example, if
+  // type is kTfLiteFloat32 and dims = {3, 2} then
+  // bytes = sizeof(float) * 3 * 2 = 4 * 3 * 2 = 24.
+  size_t bytes;
+
+  // The data type specification for data stored in `data`. This affects
+  // what member of `data` union should be used.
+  TfLiteType type;
+
+  // How memory is mapped
+  //  kTfLiteMmapRo: Memory mapped read only.
+  //  i.e. weights
+  //  kTfLiteArenaRw: Arena allocated read write memory
+  //  (i.e. temporaries, outputs).
+  TfLiteAllocationType allocation_type;
+
+  // True if the tensor is a variable.
+  bool is_variable;
+} TfLiteTensor;
+
+// Specific reduced TfLiteNode struct for TF Micro runtime. This struct contains
+// only the minimum fields required to represent a node.
+//
+// This struct does not use:
+// - delegate
+// - intermediates
+// - temporaries
+typedef struct TfLiteNode {
+  // Inputs to this node expressed as indices into the simulator's tensors.
+  TfLiteIntArray* inputs;
+
+  // Outputs to this node expressed as indices into the simulator's tensors.
+  TfLiteIntArray* outputs;
+
+  // Opaque data provided by the node implementer through `Registration.init`.
+  void* user_data;
+
+  // Opaque data provided to the node if the node is a builtin. This is usually
+  // a structure defined in builtin_op_data.h
+  void* builtin_data;
+
+  // Custom initial data. This is the opaque data provided in the flatbuffer.
+  // WARNING: This is an experimental interface that is subject to change.
+  const void* custom_initial_data;
+  int custom_initial_data_size;
+} TfLiteNode;
+#endif  // TF_LITE_STATIC_MEMORY
+
+// Light-weight tensor struct for TF Micro runtime. Provides the minimal amount
+// of information required for a kernel to run during TfLiteRegistration::Eval.
+// TODO(b/160955687): Move this field into TF_LITE_STATIC_MEMORY when TFLM
+// builds with this flag by default internally.
+typedef struct TfLiteEvalTensor {
+  // A union of data pointers. The appropriate type should be used for a typed
+  // tensor based on `type`.
+  TfLitePtrUnion data;
+
+  // A pointer to a structure representing the dimensionality interpretation
+  // that the buffer should have.
+  TfLiteIntArray* dims;
+
+  // The data type specification for data stored in `data`. This affects
+  // what member of `data` union should be used.
+  TfLiteType type;
+} TfLiteEvalTensor;
+
+#ifndef TF_LITE_STATIC_MEMORY
+// Free data memory of tensor `t`.
+void TfLiteTensorDataFree(TfLiteTensor* t);
+
+// Free quantization data.
+void TfLiteQuantizationFree(TfLiteQuantization* quantization);
+
+// Free sparsity parameters.
+void TfLiteSparsityFree(TfLiteSparsity* sparsity);
+
+// Free memory of tensor `t`.
+void TfLiteTensorFree(TfLiteTensor* t);
+
+// Set all of a tensor's fields (and free any previously allocated data).
+void TfLiteTensorReset(TfLiteType type, const char* name, TfLiteIntArray* dims,
+                       TfLiteQuantizationParams quantization, char* buffer,
+                       size_t size, TfLiteAllocationType allocation_type,
+                       const void* allocation, bool is_variable,
+                       TfLiteTensor* tensor);
+
+// Resize the allocated data of a (dynamic) tensor. Tensors with allocation
+// types other than kTfLiteDynamic will be ignored.
+void TfLiteTensorRealloc(size_t num_bytes, TfLiteTensor* tensor);
+#endif  // TF_LITE_STATIC_MEMORY
+
+// WARNING: This is an experimental interface that is subject to change.
+//
+// Currently, TfLiteDelegateParams has to be allocated in a way that it's
+// trivially destructable. It will be stored as `builtin_data` field in
+// `TfLiteNode` of the delegate node.
+//
+// See also the `CreateDelegateParams` function in `interpreter.cc` details.
+typedef struct TfLiteDelegateParams {
+  struct TfLiteDelegate* delegate;
+  TfLiteIntArray* nodes_to_replace;
+  TfLiteIntArray* input_tensors;
+  TfLiteIntArray* output_tensors;
+} TfLiteDelegateParams;
+
+typedef struct TfLiteContext {
+  // Number of tensors in the context.
+  size_t tensors_size;
+
+  // The execution plan contains a list of the node indices in execution
+  // order. execution_plan->size is the current number of nodes. And,
+  // execution_plan->data[0] is the first node that needs to be run.
+  // TfLiteDelegates can traverse the current execution plan by iterating
+  // through each member of this array and using GetNodeAndRegistration() to
+  // access details about a node. i.e.
+  // TfLiteIntArray* execution_plan;
+  // TF_LITE_ENSURE_STATUS(context->GetExecutionPlan(context, &execution_plan));
+  // for (int exec_index = 0; exec_index < execution_plan->size; exec_index++) {
+  //    int node_index = execution_plan->data[exec_index];
+  //    TfLiteNode* node;
+  //    TfLiteRegistration* reg;
+  //    context->GetNodeAndRegistration(context, node_index, &node, &reg);
+  // }
+  // WARNING: This is an experimental interface that is subject to change.
+  TfLiteStatus (*GetExecutionPlan)(struct TfLiteContext* context,
+                                   TfLiteIntArray** execution_plan);
+
+  // An array of tensors in the interpreter context (of length `tensors_size`)
+  TfLiteTensor* tensors;
+
+  // opaque full context ptr (an opaque c++ data structure)
+  void* impl_;
+
+  // Request memory pointer be resized. Updates dimensions on the tensor.
+  // NOTE: ResizeTensor takes ownership of newSize.
+  TfLiteStatus (*ResizeTensor)(struct TfLiteContext*, TfLiteTensor* tensor,
+                               TfLiteIntArray* new_size);
+  // Request that an error be reported with format string msg.
+  void (*ReportError)(struct TfLiteContext*, const char* msg, ...);
+
+  // Add `tensors_to_add` tensors, preserving pre-existing Tensor entries.  If
+  // non-null, the value pointed to by `first_new_tensor_index` will be set to
+  // the index of the first new tensor.
+  TfLiteStatus (*AddTensors)(struct TfLiteContext*, int tensors_to_add,
+                             int* first_new_tensor_index);
+
+  // Get a Tensor node by node_index.
+  // WARNING: This is an experimental interface that is subject to change.
+  TfLiteStatus (*GetNodeAndRegistration)(
+      struct TfLiteContext*, int node_index, TfLiteNode** node,
+      struct TfLiteRegistration** registration);
+
+  // Replace ops with one or more stub delegate operations. This function
+  // does not take ownership of `nodes_to_replace`.
+  TfLiteStatus (*ReplaceNodeSubsetsWithDelegateKernels)(
+      struct TfLiteContext*, struct TfLiteRegistration registration,
+      const TfLiteIntArray* nodes_to_replace, struct TfLiteDelegate* delegate);
+
+  // Number of threads that are recommended to subsystems like gemmlowp and
+  // eigen.
+  int recommended_num_threads;
+
+  // Access external contexts by type.
+  // WARNING: This is an experimental interface that is subject to change.
+  TfLiteExternalContext* (*GetExternalContext)(struct TfLiteContext*,
+                                               TfLiteExternalContextType);
+  // Set the value of a external context. Does not take ownership of the
+  // pointer.
+  // WARNING: This is an experimental interface that is subject to change.
+  void (*SetExternalContext)(struct TfLiteContext*, TfLiteExternalContextType,
+                             TfLiteExternalContext*);
+
+  // Flag for allowing float16 precision for FP32 calculation.
+  // default: false.
+  // WARNING: This is an experimental API and subject to change.
+  bool allow_fp32_relax_to_fp16;
+
+  // Pointer to the op-level profiler, if set; nullptr otherwise.
+  void* profiler;
+
+  // Allocate persistent buffer which has the same life time as the interpreter.
+  // Returns nullptr on failure.
+  // The memory is allocated from heap for TFL, and from tail in TFLM.
+  // This method is only available in Init or Prepare stage.
+  // WARNING: This is an experimental interface that is subject to change.
+  void* (*AllocatePersistentBuffer)(struct TfLiteContext* ctx, size_t bytes);
+
+  // Allocate a buffer which will be deallocated right after invoke phase.
+  // The memory is allocated from heap in TFL, and from volatile arena in TFLM.
+  // This method is only available in invoke stage.
+  // NOTE: If possible use RequestScratchBufferInArena method to avoid memory
+  // allocation during inference time.
+  // WARNING: This is an experimental interface that is subject to change.
+  TfLiteStatus (*AllocateBufferForEval)(struct TfLiteContext* ctx, size_t bytes,
+                                        void** ptr);
+
+  // Request a scratch buffer in the arena through static memory planning.
+  // This method is only available in Prepare stage and the buffer is allocated
+  // by the interpreter between Prepare and Eval stage. In Eval stage,
+  // GetScratchBuffer API can be used to fetch the address.
+  // WARNING: This is an experimental interface that is subject to change.
+  TfLiteStatus (*RequestScratchBufferInArena)(struct TfLiteContext* ctx,
+                                              size_t bytes, int* buffer_idx);
+
+  // Get the scratch buffer pointer.
+  // This method is only available in Eval stage.
+  // WARNING: This is an experimental interface that is subject to change.
+  void* (*GetScratchBuffer)(struct TfLiteContext* ctx, int buffer_idx);
+
+  // Resize the memory pointer of the `tensor`. This method behaves the same as
+  // `ResizeTensor`, except that it makes a copy of the shape array internally
+  // so the shape array could be deallocated right afterwards.
+  // WARNING: This is an experimental interface that is subject to change.
+  TfLiteStatus (*ResizeTensorExplicit)(struct TfLiteContext* ctx,
+                                       TfLiteTensor* tensor, int dims,
+                                       const int* shape);
+
+  // This method provides a preview of post-delegation partitioning. Each
+  // TfLiteDelegateParams in the referenced array corresponds to one instance of
+  // the delegate kernel.
+  // Example usage:
+  //
+  // TfLiteIntArray* nodes_to_replace = ...;
+  // TfLiteDelegateParams* params_array;
+  // int num_partitions = 0;
+  // TF_LITE_ENSURE_STATUS(context->PreviewDelegatePartitioning(
+  //    context, delegate, nodes_to_replace, &params_array, &num_partitions));
+  // for (int idx = 0; idx < num_partitions; idx++) {
+  //    const auto& partition_params = params_array[idx];
+  //    ...
+  // }
+  //
+  // NOTE: The context owns the memory referenced by partition_params_array. It
+  // will be cleared with another call to PreviewDelegateParitioning, or after
+  // TfLiteDelegateParams::Prepare returns.
+  //
+  // WARNING: This is an experimental interface that is subject to change.
+  TfLiteStatus (*PreviewDelegatePartitioning)(
+      struct TfLiteContext* context, const TfLiteIntArray* nodes_to_replace,
+      TfLiteDelegateParams** partition_params_array, int* num_partitions);
+
+  // Returns a TfLiteTensor struct for a given index.
+  // WARNING: This is an experimental interface that is subject to change.
+  // WARNING: This method may not be available on all platforms.
+  TfLiteTensor* (*GetTensor)(const struct TfLiteContext* context,
+                             int tensor_idx);
+
+  // Returns a TfLiteEvalTensor struct for a given index.
+  // WARNING: This is an experimental interface that is subject to change.
+  // WARNING: This method may not be available on all platforms.
+  TfLiteEvalTensor* (*GetEvalTensor)(const struct TfLiteContext* context,
+                                     int tensor_idx);
+} TfLiteContext;
+
+typedef struct TfLiteRegistration {
+  // Initializes the op from serialized data.
+  // If a built-in op:
+  //   `buffer` is the op's params data (TfLiteLSTMParams*).
+  //   `length` is zero.
+  // If custom op:
+  //   `buffer` is the op's `custom_options`.
+  //   `length` is the size of the buffer.
+  //
+  // Returns a type-punned (i.e. void*) opaque data (e.g. a primitive pointer
+  // or an instance of a struct).
+  //
+  // The returned pointer will be stored with the node in the `user_data` field,
+  // accessible within prepare and invoke functions below.
+  // NOTE: if the data is already in the desired format, simply implement this
+  // function to return `nullptr` and implement the free function to be a no-op.
+  void* (*init)(TfLiteContext* context, const char* buffer, size_t length);
+
+  // The pointer `buffer` is the data previously returned by an init invocation.
+  void (*free)(TfLiteContext* context, void* buffer);
+
+  // prepare is called when the inputs this node depends on have been resized.
+  // context->ResizeTensor() can be called to request output tensors to be
+  // resized.
+  //
+  // Returns kTfLiteOk on success.
+  TfLiteStatus (*prepare)(TfLiteContext* context, TfLiteNode* node);
+
+  // Execute the node (should read node->inputs and output to node->outputs).
+  // Returns kTfLiteOk on success.
+  TfLiteStatus (*invoke)(TfLiteContext* context, TfLiteNode* node);
+
+  // profiling_string is called during summarization of profiling information
+  // in order to group executions together. Providing a value here will cause a
+  // given op to appear multiple times is the profiling report. This is
+  // particularly useful for custom ops that can perform significantly
+  // different calculations depending on their `user-data`.
+  const char* (*profiling_string)(const TfLiteContext* context,
+                                  const TfLiteNode* node);
+
+  // Builtin codes. If this kernel refers to a builtin this is the code
+  // of the builtin. This is so we can do marshaling to other frameworks like
+  // NN API.
+  // Note: It is the responsibility of the registration binder to set this
+  // properly.
+  int32_t builtin_code;
+
+  // Custom op name. If the op is a builtin, this will be null.
+  // Note: It is the responsibility of the registration binder to set this
+  // properly.
+  // WARNING: This is an experimental interface that is subject to change.
+  const char* custom_name;
+
+  // The version of the op.
+  // Note: It is the responsibility of the registration binder to set this
+  // properly.
+  int version;
+} TfLiteRegistration;
+
+// The flags used in `TfLiteDelegate`. Note that this is a bitmask, so the
+// values should be 1, 2, 4, 8, ...etc.
+typedef enum TfLiteDelegateFlags {
+  kTfLiteDelegateFlagsNone = 0,
+  // The flag is set if the delegate can handle dynamic sized tensors.
+  // For example, the output shape of a `Resize` op with non-constant shape
+  // can only be inferred when the op is invoked.
+  // In this case, the Delegate is responsible for calling
+  // `SetTensorToDynamic` to mark the tensor as a dynamic tensor, and calling
+  // `ResizeTensor` when invoking the op.
+  //
+  // If the delegate isn't capable to handle dynamic tensors, this flag need
+  // to be set to false.
+  kTfLiteDelegateFlagsAllowDynamicTensors = 1,
+
+  // This flag can be used by delegates (that allow dynamic tensors) to ensure
+  // applicable tensor shapes are automatically propagated in the case of tensor
+  // resizing.
+  // This means that non-dynamic (allocation_type != kTfLiteDynamic) I/O tensors
+  // of a delegate kernel will have correct shapes before its Prepare() method
+  // is called. The runtime leverages TFLite builtin ops in the original
+  // execution plan to propagate shapes.
+  //
+  // A few points to note:
+  // 1. This requires kTfLiteDelegateFlagsAllowDynamicTensors. If that flag is
+  // false, this one is redundant since the delegate kernels are re-initialized
+  // every time tensors are resized.
+  // 2. Enabling this flag adds some overhead to AllocateTensors(), since extra
+  // work is required to prepare the original execution plan.
+  // 3. This flag requires that the original execution plan only have ops with
+  // valid registrations (and not 'dummy' custom ops like with Flex).
+  // WARNING: This feature is experimental and subject to change.
+  kTfLiteDelegateFlagsRequirePropagatedShapes = 2
+} TfLiteDelegateFlags;
+
+// WARNING: This is an experimental interface that is subject to change.
+typedef struct TfLiteDelegate {
+  // Data that delegate needs to identify itself. This data is owned by the
+  // delegate. The delegate is owned in the user code, so the delegate is
+  // responsible for doing this when it is destroyed.
+  void* data_;
+
+  // Invoked by ModifyGraphWithDelegate. This prepare is called, giving the
+  // delegate a view of the current graph through TfLiteContext*. It typically
+  // will look at the nodes and call ReplaceNodeSubsetsWithDelegateKernels()
+  // to ask the TensorFlow lite runtime to create macro-nodes to represent
+  // delegated subgraphs of the original graph.
+  TfLiteStatus (*Prepare)(TfLiteContext* context,
+                          struct TfLiteDelegate* delegate);
+
+  // Copy the data from delegate buffer handle into raw memory of the given
+  // 'tensor'. Note that the delegate is allowed to allocate the raw bytes as
+  // long as it follows the rules for kTfLiteDynamic tensors, in which case this
+  // cannot be null.
+  TfLiteStatus (*CopyFromBufferHandle)(TfLiteContext* context,
+                                       struct TfLiteDelegate* delegate,
+                                       TfLiteBufferHandle buffer_handle,
+                                       TfLiteTensor* tensor);
+
+  // Copy the data from raw memory of the given 'tensor' to delegate buffer
+  // handle. This can be null if the delegate doesn't use its own buffer.
+  TfLiteStatus (*CopyToBufferHandle)(TfLiteContext* context,
+                                     struct TfLiteDelegate* delegate,
+                                     TfLiteBufferHandle buffer_handle,
+                                     TfLiteTensor* tensor);
+
+  // Free the Delegate Buffer Handle. Note: This only frees the handle, but
+  // this doesn't release the underlying resource (e.g. textures). The
+  // resources are either owned by application layer or the delegate.
+  // This can be null if the delegate doesn't use its own buffer.
+  void (*FreeBufferHandle)(TfLiteContext* context,
+                           struct TfLiteDelegate* delegate,
+                           TfLiteBufferHandle* handle);
+
+  // Bitmask flags. See the comments in `TfLiteDelegateFlags`.
+  int64_t flags;
+} TfLiteDelegate;
+
+// Build a 'null' delegate, with all the fields properly set to their default
+// values.
+TfLiteDelegate TfLiteDelegateCreate();
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif  // __cplusplus
+#endif  // TENSORFLOW_LITE_C_COMMON_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/core/api/error_reporter.cc

@@ -0,0 +1,38 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#include "tensorflow/lite/core/api/error_reporter.h"
+#include <cstdarg>
+
+namespace tflite {
+
+int ErrorReporter::Report(const char* format, ...) {
+  va_list args;
+  va_start(args, format);
+  int code = Report(format, args);
+  va_end(args);
+  return code;
+}
+
+// TODO(aselle): Make the name of ReportError on context the same, so
+// we can use the ensure functions w/o a context and w/ a reporter.
+int ErrorReporter::ReportError(void*, const char* format, ...) {
+  va_list args;
+  va_start(args, format);
+  int code = Report(format, args);
+  va_end(args);
+  return code;
+}
+
+}  // namespace tflite

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/core/api/error_reporter.h

@@ -0,0 +1,59 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_CORE_API_ERROR_REPORTER_H_
+#define TENSORFLOW_LITE_CORE_API_ERROR_REPORTER_H_
+
+#include <cstdarg>
+
+namespace tflite {
+
+/// A functor that reports error to supporting system. Invoked similar to
+/// printf.
+///
+/// Usage:
+///  ErrorReporter foo;
+///  foo.Report("test %d", 5);
+/// or
+///  va_list args;
+///  foo.Report("test %d", args); // where args is va_list
+///
+/// Subclass ErrorReporter to provide another reporting destination.
+/// For example, if you have a GUI program, you might redirect to a buffer
+/// that drives a GUI error log box.
+class ErrorReporter {
+ public:
+  virtual ~ErrorReporter() {}
+  virtual int Report(const char* format, va_list args) = 0;
+  int Report(const char* format, ...);
+  int ReportError(void*, const char* format, ...);
+};
+
+}  // namespace tflite
+
+// You should not make bare calls to the error reporter, instead use the
+// TF_LITE_REPORT_ERROR macro, since this allows message strings to be
+// stripped when the binary size has to be optimized. If you are looking to
+// reduce binary size, define TF_LITE_STRIP_ERROR_STRINGS when compiling and
+// every call will be stubbed out, taking no memory.
+#ifndef TF_LITE_STRIP_ERROR_STRINGS
+#define TF_LITE_REPORT_ERROR(reporter, ...)                             \
+  do {                                                                  \
+    static_cast<tflite::ErrorReporter*>(reporter)->Report(__VA_ARGS__); \
+  } while (false)
+#else  // TF_LITE_STRIP_ERROR_STRINGS
+#define TF_LITE_REPORT_ERROR(reporter, ...)
+#endif  // TF_LITE_STRIP_ERROR_STRINGS
+
+#endif  // TENSORFLOW_LITE_CORE_API_ERROR_REPORTER_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/core/api/flatbuffer_conversions.h

@@ -0,0 +1,301 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_CORE_API_FLATBUFFER_CONVERSIONS_H_
+#define TENSORFLOW_LITE_CORE_API_FLATBUFFER_CONVERSIONS_H_
+
+// These functions transform codes and data structures that are defined in the
+// flatbuffer serialization format into in-memory values that are used by the
+// runtime API and interpreter.
+
+#include <cstddef>
+#include <new>
+#include <type_traits>
+
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/core/api/error_reporter.h"
+#include "tensorflow/lite/schema/schema_generated.h"
+
+namespace tflite {
+
+// Interface class for builtin data allocations.
+class BuiltinDataAllocator {
+ public:
+  virtual void* Allocate(size_t size, size_t alignment_hint) = 0;
+  virtual void Deallocate(void* data) = 0;
+
+  // Allocate a structure, but make sure it is a POD structure that doesn't
+  // require constructors to run. The reason we do this, is that Interpreter's C
+  // extension part will take ownership so destructors  will not be run during
+  // deallocation.
+  template <typename T>
+  T* AllocatePOD() {
+    // TODO(b/154346074): Change this to is_trivially_destructible when all
+    // platform targets support that properly.
+    static_assert(std::is_pod<T>::value, "Builtin data structure must be POD.");
+    void* allocated_memory = this->Allocate(sizeof(T), alignof(T));
+    return new (allocated_memory) T();
+  }
+
+  virtual ~BuiltinDataAllocator() {}
+};
+
+// Parse the appropriate data out of the op.
+//
+// This handles builtin data explicitly as there are flatbuffer schemas.
+// If it returns kTfLiteOk, it passes the data out with `builtin_data`. The
+// calling function has to pass in an allocator object, and this allocator
+// will be called to reserve space for the output data. If the calling
+// function's allocator reserves memory on the heap, then it's the calling
+// function's responsibility to free it.
+// If it returns kTfLiteError, `builtin_data` will be `nullptr`.
+TfLiteStatus ParseOpData(const Operator* op, BuiltinOperator op_type,
+                         ErrorReporter* error_reporter,
+                         BuiltinDataAllocator* allocator, void** builtin_data);
+
+// Converts the tensor data type used in the flat buffer to the representation
+// used by the runtime.
+TfLiteStatus ConvertTensorType(TensorType tensor_type, TfLiteType* type,
+                               ErrorReporter* error_reporter);
+
+TfLiteStatus ParseAbs(const Operator* op, ErrorReporter* error_reporter,
+                      BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseAdd(const Operator* op, ErrorReporter* error_reporter,
+                      BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseArgMax(const Operator* op, ErrorReporter* error_reporter,
+                         BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseArgMin(const Operator* op, ErrorReporter* error_reporter,
+                         BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseCeil(const Operator* op, ErrorReporter* error_reporter,
+                       BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseCast(const Operator* op, ErrorReporter* error_reporter,
+                       BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseConcatenation(const Operator* op,
+                                ErrorReporter* error_reporter,
+                                BuiltinDataAllocator* allocator,
+                                void** builtin_data);
+
+TfLiteStatus ParseConv2D(const Operator* op, ErrorReporter* error_reporter,
+                         BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseCos(const Operator* op, ErrorReporter* error_reporter,
+                      BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseDepthwiseConv2D(const Operator* op,
+                                  ErrorReporter* error_reporter,
+                                  BuiltinDataAllocator* allocator,
+                                  void** builtin_data);
+
+TfLiteStatus ParseDequantize(const Operator* op, ErrorReporter* error_reporter,
+                             BuiltinDataAllocator* allocator,
+                             void** builtin_data);
+
+TfLiteStatus ParseDiv(const Operator* op, ErrorReporter* error_reporter,
+                      BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseEqual(const Operator* op, ErrorReporter* error_reporter,
+                        BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseExp(const Operator* op, ErrorReporter* error_reporter,
+                      BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseFill(const Operator* op, ErrorReporter* error_reporter,
+                       BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseFloor(const Operator* op, ErrorReporter* error_reporter,
+                        BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseFloorDiv(const Operator* op, ErrorReporter* error_reporter,
+                           BuiltinDataAllocator* allocator,
+                           void** builtin_data);
+
+TfLiteStatus ParseFloorMod(const Operator* op, ErrorReporter* error_reporter,
+                           BuiltinDataAllocator* allocator,
+                           void** builtin_data);
+
+TfLiteStatus ParseFullyConnected(const Operator* op,
+                                 ErrorReporter* error_reporter,
+                                 BuiltinDataAllocator* allocator,
+                                 void** builtin_data);
+
+TfLiteStatus ParseGreater(const Operator* op, ErrorReporter* error_reporter,
+                          BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseGreaterEqual(const Operator* op,
+                               ErrorReporter* error_reporter,
+                               BuiltinDataAllocator* allocator,
+                               void** builtin_data);
+
+TfLiteStatus ParseHardSwish(const Operator* op, ErrorReporter* error_reporter,
+                            BuiltinDataAllocator* allocator,
+                            void** builtin_data);
+
+TfLiteStatus ParseL2Normalization(const Operator* op,
+                                  ErrorReporter* error_reporter,
+                                  BuiltinDataAllocator* allocator,
+                                  void** builtin_data);
+
+TfLiteStatus ParseLess(const Operator* op, ErrorReporter* error_reporter,
+                       BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseLessEqual(const Operator* op, ErrorReporter* error_reporter,
+                            BuiltinDataAllocator* allocator,
+                            void** builtin_data);
+
+TfLiteStatus ParseLog(const Operator* op, ErrorReporter* error_reporter,
+                      BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseLogicalAnd(const Operator* op, ErrorReporter* error_reporter,
+                             BuiltinDataAllocator* allocator,
+                             void** builtin_data);
+
+TfLiteStatus ParseLogicalNot(const Operator* op, ErrorReporter* error_reporter,
+                             BuiltinDataAllocator* allocator,
+                             void** builtin_data);
+
+TfLiteStatus ParseLogicalOr(const Operator* op, ErrorReporter* error_reporter,
+                            BuiltinDataAllocator* allocator,
+                            void** builtin_data);
+
+TfLiteStatus ParseLogistic(const Operator* op, ErrorReporter* error_reporter,
+                           BuiltinDataAllocator* allocator,
+                           void** builtin_data);
+
+TfLiteStatus ParseMaximum(const Operator* op, ErrorReporter* error_reporter,
+                          BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseMinimum(const Operator* op, ErrorReporter* error_reporter,
+                          BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseMul(const Operator* op, ErrorReporter* error_reporter,
+                      BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseNeg(const Operator* op, ErrorReporter* error_reporter,
+                      BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseNotEqual(const Operator* op, ErrorReporter* error_reporter,
+                           BuiltinDataAllocator* allocator,
+                           void** builtin_data);
+
+TfLiteStatus ParsePack(const Operator* op, ErrorReporter* error_reporter,
+                       BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParsePad(const Operator* op, ErrorReporter* error_reporter,
+                      BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParsePadV2(const Operator* op, ErrorReporter* error_reporter,
+                        BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParsePool(const Operator* op, ErrorReporter* error_reporter,
+                       BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParsePow(const Operator* op, ErrorReporter* error_reporter,
+                      BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParsePrelu(const Operator* op, ErrorReporter* error_reporter,
+                        BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseQuantize(const Operator* op, ErrorReporter* error_reporter,
+                           BuiltinDataAllocator* allocator,
+                           void** builtin_data);
+
+TfLiteStatus ParseReducer(const Operator* op, ErrorReporter* error_reporter,
+                          BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseRelu(const Operator* op, ErrorReporter* error_reporter,
+                       BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseRelu6(const Operator* op, ErrorReporter* error_reporter,
+                        BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseReshape(const Operator* op, ErrorReporter* error_reporter,
+                          BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseResizeBilinear(const Operator* op,
+                                 ErrorReporter* error_reporter,
+                                 BuiltinDataAllocator* allocator,
+                                 void** builtin_data);
+
+TfLiteStatus ParseResizeNearestNeighbor(const Operator* op,
+                                        ErrorReporter* error_reporter,
+                                        BuiltinDataAllocator* allocator,
+                                        void** builtin_data);
+
+TfLiteStatus ParseRound(const Operator* op, ErrorReporter* error_reporter,
+                        BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseRsqrt(const Operator* op, ErrorReporter* error_reporter,
+                        BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseShape(const Operator* op, ErrorReporter* error_reporter,
+                        BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseSin(const Operator* op, ErrorReporter* error_reporter,
+                      BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseSoftmax(const Operator* op, ErrorReporter* error_reporter,
+                          BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseSpaceToDepth(const Operator* op,
+                               ErrorReporter* error_reporter,
+                               BuiltinDataAllocator* allocator,
+                               void** builtin_data);
+
+TfLiteStatus ParseSplit(const Operator* op, ErrorReporter* error_reporter,
+                        BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseSplitV(const Operator* op, ErrorReporter* error_reporter,
+                         BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseSqrt(const Operator* op, ErrorReporter* error_reporter,
+                       BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseSquare(const Operator* op, ErrorReporter* error_reporter,
+                         BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseStridedSlice(const Operator* op,
+                               ErrorReporter* error_reporter,
+                               BuiltinDataAllocator* allocator,
+                               void** builtin_data);
+
+TfLiteStatus ParseSub(const Operator* op, ErrorReporter* error_reporter,
+                      BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseSvdf(const Operator* op, ErrorReporter* error_reporter,
+                       BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseTanh(const Operator* op, ErrorReporter* error_reporter,
+                       BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseTransposeConv(const Operator* op,
+                                ErrorReporter* error_reporter,
+                                BuiltinDataAllocator* allocator,
+                                void** builtin_data);
+
+TfLiteStatus ParseUnpack(const Operator* op, ErrorReporter* error_reporter,
+                         BuiltinDataAllocator* allocator, void** builtin_data);
+
+TfLiteStatus ParseZerosLike(const Operator* op, ErrorReporter* error_reporter,
+                            BuiltinDataAllocator* allocator,
+                            void** builtin_data);
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_CORE_API_FLATBUFFER_CONVERSIONS_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/core/api/op_resolver.cc

@@ -0,0 +1,67 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/lite/core/api/op_resolver.h"
+
+#include "flatbuffers/flatbuffers.h"  // from @flatbuffers
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/core/api/error_reporter.h"
+#include "tensorflow/lite/schema/schema_utils.h"
+
+namespace tflite {
+
+TfLiteStatus GetRegistrationFromOpCode(
+    const OperatorCode* opcode, const OpResolver& op_resolver,
+    ErrorReporter* error_reporter, const TfLiteRegistration** registration) {
+  TfLiteStatus status = kTfLiteOk;
+  *registration = nullptr;
+  auto builtin_code = GetBuiltinCode(opcode);
+  int version = opcode->version();
+
+  if (builtin_code > BuiltinOperator_MAX ||
+      builtin_code < BuiltinOperator_MIN) {
+    TF_LITE_REPORT_ERROR(
+        error_reporter,
+        "Op builtin_code out of range: %d. Are you using old TFLite binary "
+        "with newer model?",
+        builtin_code);
+    status = kTfLiteError;
+  } else if (builtin_code != BuiltinOperator_CUSTOM) {
+    *registration = op_resolver.FindOp(builtin_code, version);
+    if (*registration == nullptr) {
+      TF_LITE_REPORT_ERROR(
+          error_reporter,
+          "Didn't find op for builtin opcode '%s' version '%d'\n",
+          EnumNameBuiltinOperator(builtin_code), version);
+      status = kTfLiteError;
+    }
+  } else if (!opcode->custom_code()) {
+    TF_LITE_REPORT_ERROR(
+        error_reporter,
+        "Operator with CUSTOM builtin_code has no custom_code.\n");
+    status = kTfLiteError;
+  } else {
+    const char* name = opcode->custom_code()->c_str();
+    *registration = op_resolver.FindOp(name, version);
+    if (*registration == nullptr) {
+      // Do not report error for unresolved custom op, we do the final check
+      // while preparing ops.
+      status = kTfLiteError;
+    }
+  }
+  return status;
+}
+
+}  // namespace tflite

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/core/api/op_resolver.h

@@ -0,0 +1,60 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_CORE_API_OP_RESOLVER_H_
+#define TENSORFLOW_LITE_CORE_API_OP_RESOLVER_H_
+
+#include <vector>
+
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/core/api/error_reporter.h"
+#include "tensorflow/lite/schema/schema_generated.h"
+
+namespace tflite {
+
+/// Abstract interface that returns TfLiteRegistrations given op codes or custom
+/// op names. This is the mechanism that ops being referenced in the flatbuffer
+/// model are mapped to executable function pointers (TfLiteRegistrations).
+class OpResolver {
+ public:
+  /// Finds the op registration for a builtin operator by enum code.
+  virtual const TfLiteRegistration* FindOp(tflite::BuiltinOperator op,
+                                           int version) const = 0;
+  /// Finds the op registration of a custom operator by op name.
+  virtual const TfLiteRegistration* FindOp(const char* op,
+                                           int version) const = 0;
+
+  // Returns optional delegates for resolving and handling ops in the flatbuffer
+  // model. This may be used in addition to the standard TfLiteRegistration
+  // lookup for graph resolution.
+  using TfLiteDelegatePtrVector =
+      std::vector<std::unique_ptr<TfLiteDelegate, void (*)(TfLiteDelegate*)>>;
+  virtual TfLiteDelegatePtrVector GetDelegates(int num_threads) const {
+    return TfLiteDelegatePtrVector();
+  }
+
+  virtual ~OpResolver() {}
+};
+
+// Handles the logic for converting between an OperatorCode structure extracted
+// from a flatbuffer and information about a registered operator
+// implementation.
+TfLiteStatus GetRegistrationFromOpCode(const OperatorCode* opcode,
+                                       const OpResolver& op_resolver,
+                                       ErrorReporter* error_reporter,
+                                       const TfLiteRegistration** registration);
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_CORE_API_OP_RESOLVER_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/core/api/profiler.h

@@ -0,0 +1,194 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_CORE_API_PROFILER_H_
+#define TENSORFLOW_LITE_CORE_API_PROFILER_H_
+
+#include <cstdint>
+
+namespace tflite {
+
+// A simple utility for enabling profiled event tracing in TensorFlow Lite.
+class Profiler {
+ public:
+  // As certain Profiler instance might be only interested in certain event
+  // types, we define each event type value to allow a Profiler to use
+  // bitmasking bitwise operations to determine whether an event should be
+  // recorded or not.
+  enum class EventType {
+    // Default event type, the metadata field has no special significance.
+    DEFAULT = 1,
+
+    // The event is an operator invocation and the event_metadata field is the
+    // index of operator node.
+    OPERATOR_INVOKE_EVENT = 2,
+
+    // The event is an invocation for an internal operator of a TFLite delegate.
+    // The event_metadata field is the index of operator node that's specific to
+    // the delegate.
+    DELEGATE_OPERATOR_INVOKE_EVENT = 4,
+
+    // The event is a recording of runtime instrumentation such as the overall
+    // TFLite runtime status, the TFLite delegate status (if a delegate
+    // is applied), and the overall model inference latency etc.
+    // Note, the delegate status and overall status are stored as separate
+    // event_metadata fields. In particular, the delegate status is encoded
+    // as DelegateStatus::full_status().
+    GENERAL_RUNTIME_INSTRUMENTATION_EVENT = 8,
+  };
+
+  virtual ~Profiler() {}
+
+  // Signals the beginning of an event and returns a handle to the profile
+  // event. The `event_metadata1` and `event_metadata2` have different
+  // interpretations based on the actual Profiler instance and the `event_type`.
+  // For example, as for the 'SubgraphAwareProfiler' defined in
+  // lite/core/subgraph.h, when the event_type is OPERATOR_INVOKE_EVENT,
+  // `event_metadata1` represents the index of a TFLite node, and
+  // `event_metadata2` represents the index of the subgraph that this event
+  // comes from.
+  virtual uint32_t BeginEvent(const char* tag, EventType event_type,
+                              int64_t event_metadata1,
+                              int64_t event_metadata2) = 0;
+  // Similar w/ the above, but `event_metadata2` defaults to 0.
+  uint32_t BeginEvent(const char* tag, EventType event_type,
+                      int64_t event_metadata) {
+    return BeginEvent(tag, event_type, event_metadata, /*event_metadata2*/ 0);
+  }
+
+  // Signals an end to the specified profile event with 'event_metadata's, This
+  // is useful when 'event_metadata's are not available when the event begins
+  // or when one wants to overwrite the 'event_metadata's set at the beginning.
+  virtual void EndEvent(uint32_t event_handle, int64_t event_metadata1,
+                        int64_t event_metadata2) {}
+  // Signals an end to the specified profile event.
+  virtual void EndEvent(uint32_t event_handle) = 0;
+
+  // Appends an event of type 'event_type' with 'tag' and 'event_metadata'
+  // which started at 'start' and ended at 'end'
+  // Note:
+  // In cases were ProfileSimmarizer and tensorflow::StatsCalculator are used
+  // they assume the value is in "usec", if in any case subclasses
+  // didn't put usec, then the values are not meaningful.
+  // TODO karimnosseir: Revisit and make the function more clear.
+  void AddEvent(const char* tag, EventType event_type, uint64_t start,
+                uint64_t end, int64_t event_metadata) {
+    AddEvent(tag, event_type, start, end, event_metadata,
+             /*event_metadata2*/ 0);
+  }
+
+  virtual void AddEvent(const char* tag, EventType event_type, uint64_t start,
+                        uint64_t end, int64_t event_metadata1,
+                        int64_t event_metadata2) {}
+
+ protected:
+  friend class ScopedProfile;
+};
+
+// Adds a profile event to `profiler` that begins with the construction
+// of the object and ends when the object goes out of scope.
+// The lifetime of tag should be at least the lifetime of `profiler`.
+// `profiler` may be null, in which case nothing is profiled.
+class ScopedProfile {
+ public:
+  ScopedProfile(Profiler* profiler, const char* tag,
+                Profiler::EventType event_type = Profiler::EventType::DEFAULT,
+                int64_t event_metadata = 0)
+      : profiler_(profiler), event_handle_(0) {
+    if (profiler) {
+      event_handle_ = profiler_->BeginEvent(tag, event_type, event_metadata);
+    }
+  }
+
+  ~ScopedProfile() {
+    if (profiler_) {
+      profiler_->EndEvent(event_handle_);
+    }
+  }
+
+ protected:
+  Profiler* profiler_;
+  uint32_t event_handle_;
+};
+
+class ScopedOperatorProfile : public ScopedProfile {
+ public:
+  ScopedOperatorProfile(Profiler* profiler, const char* tag, int node_index)
+      : ScopedProfile(profiler, tag, Profiler::EventType::OPERATOR_INVOKE_EVENT,
+                      static_cast<uint32_t>(node_index)) {}
+};
+
+class ScopedDelegateOperatorProfile : public ScopedProfile {
+ public:
+  ScopedDelegateOperatorProfile(Profiler* profiler, const char* tag,
+                                int node_index)
+      : ScopedProfile(profiler, tag,
+                      Profiler::EventType::DELEGATE_OPERATOR_INVOKE_EVENT,
+                      static_cast<uint32_t>(node_index)) {}
+};
+
+class ScopedRuntimeInstrumentationProfile : public ScopedProfile {
+ public:
+  ScopedRuntimeInstrumentationProfile(Profiler* profiler, const char* tag)
+      : ScopedProfile(
+            profiler, tag,
+            Profiler::EventType::GENERAL_RUNTIME_INSTRUMENTATION_EVENT, -1) {}
+
+  void set_runtime_status(int64_t delegate_status, int64_t interpreter_status) {
+    if (profiler_) {
+      delegate_status_ = delegate_status;
+      interpreter_status_ = interpreter_status;
+    }
+  }
+
+  ~ScopedRuntimeInstrumentationProfile() {
+    if (profiler_) {
+      profiler_->EndEvent(event_handle_, delegate_status_, interpreter_status_);
+    }
+  }
+
+ private:
+  int64_t delegate_status_;
+  int64_t interpreter_status_;
+};
+
+}  // namespace tflite
+
+#define TFLITE_VARNAME_UNIQ_IMPL(name, ctr) name##ctr
+#define TFLITE_VARNAME_UNIQ(name, ctr) TFLITE_VARNAME_UNIQ_IMPL(name, ctr)
+
+#define TFLITE_SCOPED_TAGGED_DEFAULT_PROFILE(profiler, tag)          \
+  tflite::ScopedProfile TFLITE_VARNAME_UNIQ(_profile_, __COUNTER__)( \
+      (profiler), (tag))
+
+#define TFLITE_SCOPED_TAGGED_OPERATOR_PROFILE(profiler, tag, node_index)     \
+  tflite::ScopedOperatorProfile TFLITE_VARNAME_UNIQ(_profile_, __COUNTER__)( \
+      (profiler), (tag), (node_index))
+
+#define TFLITE_SCOPED_DELEGATE_OPERATOR_PROFILE(profiler, tag, node_index) \
+  tflite::ScopedDelegateOperatorProfile TFLITE_VARNAME_UNIQ(               \
+      _profile_, __COUNTER__)((profiler), (tag), (node_index))
+
+#define TFLITE_ADD_RUNTIME_INSTRUMENTATION_EVENT(                          \
+    profiler, tag, event_metadata1, event_metadata2)                       \
+  do {                                                                     \
+    if (profiler) {                                                        \
+      const auto handle = profiler->BeginEvent(                            \
+          tag, Profiler::EventType::GENERAL_RUNTIME_INSTRUMENTATION_EVENT, \
+          event_metadata1, event_metadata2);                               \
+      profiler->EndEvent(handle);                                          \
+    }                                                                      \
+  } while (false);
+
+#endif  // TENSORFLOW_LITE_CORE_API_PROFILER_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/core/api/tensor_utils.cc

@@ -0,0 +1,50 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/lite/core/api/tensor_utils.h"
+
+#include <string.h>
+
+#include "tensorflow/lite/c/common.h"
+
+namespace tflite {
+
+TfLiteStatus ResetVariableTensor(TfLiteTensor* tensor) {
+  if (!tensor->is_variable) {
+    return kTfLiteOk;
+  }
+  // TODO(b/115961645): Implement - If a variable tensor has a buffer, reset it
+  // to the value of the buffer.
+  int value = 0;
+  if (tensor->type == kTfLiteInt8) {
+    value = tensor->params.zero_point;
+  }
+  // TODO(b/139446230): Provide a platform header to better handle these
+  // specific scenarios.
+#if __ANDROID__ || defined(__x86_64__) || defined(__i386__) || \
+    defined(__i386) || defined(__x86__) || defined(__X86__) || \
+    defined(_X86_) || defined(_M_IX86) || defined(_M_X64)
+  memset(tensor->data.raw, value, tensor->bytes);
+#else
+  char* raw_ptr = tensor->data.raw;
+  for (size_t i = 0; i < tensor->bytes; ++i) {
+    *raw_ptr = value;
+    raw_ptr++;
+  }
+#endif
+  return kTfLiteOk;
+}
+
+}  // namespace tflite

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/core/api/tensor_utils.h

@@ -0,0 +1,28 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_LITE_CORE_API_TENSOR_UTILS_H_
+#define TENSORFLOW_LITE_CORE_API_TENSOR_UTILS_H_
+
+#include "tensorflow/lite/c/common.h"
+
+namespace tflite {
+
+// Resets a variable tensor to the default value.
+TfLiteStatus ResetVariableTensor(TfLiteTensor* tensor);
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_CORE_API_TENSOR_UTILS_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/compatibility.h

@@ -0,0 +1,112 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_COMPATIBILITY_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_COMPATIBILITY_H_
+
+#include <cstdint>
+
+#include "tensorflow/lite/kernels/op_macros.h"
+
+#ifndef TFLITE_DCHECK
+#define TFLITE_DCHECK(condition) (condition) ? (void)0 : TFLITE_ASSERT_FALSE
+#endif
+
+#ifndef TFLITE_DCHECK_EQ
+#define TFLITE_DCHECK_EQ(x, y) ((x) == (y)) ? (void)0 : TFLITE_ASSERT_FALSE
+#endif
+
+#ifndef TFLITE_DCHECK_NE
+#define TFLITE_DCHECK_NE(x, y) ((x) != (y)) ? (void)0 : TFLITE_ASSERT_FALSE
+#endif
+
+#ifndef TFLITE_DCHECK_GE
+#define TFLITE_DCHECK_GE(x, y) ((x) >= (y)) ? (void)0 : TFLITE_ASSERT_FALSE
+#endif
+
+#ifndef TFLITE_DCHECK_GT
+#define TFLITE_DCHECK_GT(x, y) ((x) > (y)) ? (void)0 : TFLITE_ASSERT_FALSE
+#endif
+
+#ifndef TFLITE_DCHECK_LE
+#define TFLITE_DCHECK_LE(x, y) ((x) <= (y)) ? (void)0 : TFLITE_ASSERT_FALSE
+#endif
+
+#ifndef TFLITE_DCHECK_LT
+#define TFLITE_DCHECK_LT(x, y) ((x) < (y)) ? (void)0 : TFLITE_ASSERT_FALSE
+#endif
+
+// TODO(ahentz): Clean up: We should stick to the DCHECK versions.
+#ifndef TFLITE_CHECK
+#define TFLITE_CHECK(condition) (condition) ? (void)0 : TFLITE_ABORT
+#endif
+
+#ifndef TFLITE_CHECK_EQ
+#define TFLITE_CHECK_EQ(x, y) ((x) == (y)) ? (void)0 : TFLITE_ABORT
+#endif
+
+#ifndef TFLITE_CHECK_NE
+#define TFLITE_CHECK_NE(x, y) ((x) != (y)) ? (void)0 : TFLITE_ABORT
+#endif
+
+#ifndef TFLITE_CHECK_GE
+#define TFLITE_CHECK_GE(x, y) ((x) >= (y)) ? (void)0 : TFLITE_ABORT
+#endif
+
+#ifndef TFLITE_CHECK_GT
+#define TFLITE_CHECK_GT(x, y) ((x) > (y)) ? (void)0 : TFLITE_ABORT
+#endif
+
+#ifndef TFLITE_CHECK_LE
+#define TFLITE_CHECK_LE(x, y) ((x) <= (y)) ? (void)0 : TFLITE_ABORT
+#endif
+
+#ifndef TFLITE_CHECK_LT
+#define TFLITE_CHECK_LT(x, y) ((x) < (y)) ? (void)0 : TFLITE_ABORT
+#endif
+
+#ifndef TF_LITE_STATIC_MEMORY
+// TODO(b/162019032): Consider removing these type-aliases.
+using int8 = std::int8_t;
+using uint8 = std::uint8_t;
+using int16 = std::int16_t;
+using uint16 = std::uint16_t;
+using int32 = std::int32_t;
+using uint32 = std::uint32_t;
+#endif  // !defined(TF_LITE_STATIC_MEMORY)
+
+// TFLITE_DEPRECATED()
+//
+// Duplicated from absl/base/macros.h to avoid pulling in that library.
+// Marks a deprecated class, struct, enum, function, method and variable
+// declarations. The macro argument is used as a custom diagnostic message (e.g.
+// suggestion of a better alternative).
+//
+// Example:
+//
+//   class TFLITE_DEPRECATED("Use Bar instead") Foo {...};
+//   TFLITE_DEPRECATED("Use Baz instead") void Bar() {...}
+//
+// Every usage of a deprecated entity will trigger a warning when compiled with
+// clang's `-Wdeprecated-declarations` option. This option is turned off by
+// default, but the warnings will be reported by clang-tidy.
+#if defined(__clang__) && __cplusplus >= 201103L
+#define TFLITE_DEPRECATED(message) __attribute__((deprecated(message)))
+#endif
+
+#ifndef TFLITE_DEPRECATED
+#define TFLITE_DEPRECATED(message)
+#endif
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_COMPATIBILITY_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/cppmath.h

@@ -0,0 +1,40 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_CPPMATH_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_CPPMATH_H_
+
+#include <cmath>
+
+namespace tflite {
+
+#if defined(TF_LITE_USE_GLOBAL_CMATH_FUNCTIONS) ||                           \
+    (defined(__ANDROID__) && !defined(__NDK_MAJOR__)) || defined(ARDUINO) || \
+    defined(__ZEPHYR__)
+#define TF_LITE_GLOBAL_STD_PREFIX
+#else
+#define TF_LITE_GLOBAL_STD_PREFIX std
+#endif
+
+#define DECLARE_STD_GLOBAL_SWITCH1(tf_name, std_name) \
+  template <class T>                                  \
+  inline T tf_name(const T x) {                       \
+    return TF_LITE_GLOBAL_STD_PREFIX::std_name(x);    \
+  }
+
+DECLARE_STD_GLOBAL_SWITCH1(TfLiteRound, round);
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_CPPMATH_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/max.h

@@ -0,0 +1,35 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_MAX_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_MAX_H_
+
+#include <cmath>
+
+namespace tflite {
+
+#if defined(TF_LITE_USE_GLOBAL_MAX) || defined(__ZEPHYR__)
+inline float TfLiteMax(const float& x, const float& y) {
+  return std::max(x, y);
+}
+#else
+template <class T>
+inline T TfLiteMax(const T& x, const T& y) {
+  return std::fmax(x, y);
+}
+#endif
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_MAX_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/min.h

@@ -0,0 +1,35 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_MIN_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_MIN_H_
+
+#include <cmath>
+
+namespace tflite {
+
+#if defined(TF_LITE_USE_GLOBAL_MIN) || defined(__ZEPHYR__)
+inline float TfLiteMin(const float& x, const float& y) {
+  return std::min(x, y);
+}
+#else
+template <class T>
+inline T TfLiteMin(const T& x, const T& y) {
+  return std::fmin(x, y);
+}
+#endif
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_MIN_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/optimized/neon_check.h

@@ -0,0 +1,40 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_OPTIMIZED_NEON_CHECK_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_OPTIMIZED_NEON_CHECK_H_
+
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+#define USE_NEON
+#include <arm_neon.h>
+#endif
+
+#if defined __GNUC__ && defined __SSE4_1__ && !defined TF_LITE_DISABLE_X86_NEON
+#define USE_NEON
+#include "NEON_2_SSE.h"
+#endif
+
+// NEON_OR_PORTABLE(SomeFunc, args) calls NeonSomeFunc(args) if USE_NEON is
+// defined, PortableSomeFunc(args) otherwise.
+#ifdef USE_NEON
+// Always use Neon code
+#define NEON_OR_PORTABLE(funcname, ...) Neon##funcname(__VA_ARGS__)
+
+#else
+// No NEON available: Use Portable code
+#define NEON_OR_PORTABLE(funcname, ...) Portable##funcname(__VA_ARGS__)
+
+#endif  // defined(USE_NEON)
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_OPTIMIZED_NEON_CHECK_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/portable_tensor.h

@@ -0,0 +1,122 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_PORTABLE_TENSOR_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_PORTABLE_TENSOR_H_
+
+#include <vector>
+
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/kernels/internal/tensor_ctypes.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+inline RuntimeShape GetTensorShape(std::vector<int32_t> data) {
+  return RuntimeShape(data.size(), data.data());
+}
+
+// A list of tensors in a format that can be used by kernels like split and
+// concatenation.
+template <typename T>
+class VectorOfTensors {
+ public:
+  // Build with the tensors in 'tensor_list'.
+  VectorOfTensors(const TfLiteContext& context,
+                  const TfLiteIntArray& tensor_list) {
+    int num_tensors = tensor_list.size;
+
+    all_data_.reserve(num_tensors);
+    all_shape_.reserve(num_tensors);
+    all_shape_ptr_.reserve(num_tensors);
+
+    for (int i = 0; i < num_tensors; ++i) {
+      TfLiteTensor* t = &context.tensors[tensor_list.data[i]];
+      all_data_.push_back(GetTensorData<T>(t));
+      all_shape_.push_back(GetTensorShape(t));
+    }
+
+    // Taking the pointer from inside a std::vector is only OK if the vector is
+    // never modified, so we populate all_shape in the previous loop and then we
+    // are free to grab iterators here.
+    for (int i = 0; i < num_tensors; ++i) {
+      all_shape_ptr_.push_back(&all_shape_[i]);
+    }
+  }
+  // Return a pointer to the data pointers of all tensors in the list. For
+  // example:
+  //   float* const* f = v.data();
+  //   f[0][1] is the second element of the first tensor.
+  T* const* data() const { return all_data_.data(); }
+
+  // Return a pointer the shape pointers of all tensors in the list. For
+  // example:
+  //   const RuntimeShape* const* d = v.dims();
+  //   dims[1] are the dimensions of the second tensor in the list.
+  const RuntimeShape* const* shapes() const { return all_shape_ptr_.data(); }
+
+ private:
+  std::vector<T*> all_data_;
+  std::vector<RuntimeShape> all_shape_;
+  std::vector<RuntimeShape*> all_shape_ptr_;
+};
+
+// A list of quantized tensors in a format that can be used by kernels like
+// split and concatenation.
+class VectorOfQuantizedTensors : public VectorOfTensors<uint8_t> {
+ public:
+  // Build with the tensors in 'tensor_list'.
+  VectorOfQuantizedTensors(const TfLiteContext& context,
+                           const TfLiteIntArray& tensor_list)
+      : VectorOfTensors<uint8_t>(context, tensor_list) {
+    for (int i = 0; i < tensor_list.size; ++i) {
+      TfLiteTensor* t = &context.tensors[tensor_list.data[i]];
+      zero_point_.push_back(t->params.zero_point);
+      scale_.push_back(t->params.scale);
+    }
+  }
+
+  const float* scale() const { return scale_.data(); }
+  const int32_t* zero_point() const { return zero_point_.data(); }
+
+ private:
+  std::vector<int32_t> zero_point_;
+  std::vector<float> scale_;
+};
+
+// Writes randomly accessed values from `input` sequentially into `output`.
+template <typename T>
+class SequentialTensorWriter {
+ public:
+  SequentialTensorWriter(const TfLiteTensor* input, TfLiteTensor* output) {
+    input_data_ = GetTensorData<T>(input);
+    output_ptr_ = GetTensorData<T>(output);
+  }
+  SequentialTensorWriter(const T* input_data, T* output_data)
+      : input_data_(input_data), output_ptr_(output_data) {}
+
+  void Write(int position) { *output_ptr_++ = input_data_[position]; }
+  void WriteN(int position, int len) {
+    memcpy(output_ptr_, &input_data_[position], sizeof(T) * len);
+    output_ptr_ += len;
+  }
+
+ private:
+  const T* input_data_;
+  T* output_ptr_;
+};
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_PORTABLE_TENSOR_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/quantization_util.cc

@@ -0,0 +1,395 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/lite/kernels/internal/quantization_util.h"
+
+#include <algorithm>
+#include <cmath>
+#include <limits>
+
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+
+namespace tflite {
+
+namespace {
+// These constants are used to manipulate the binary representation of doubles.
+// Double-precision binary64 floating point format is:
+// Bit |  63  |  62-52   |   51-0   |
+//     | Sign | Exponent | Fraction |
+// To avoid 64-bit integers as much as possible, I break this into high and
+// low 32-bit chunks. High is:
+// Bit |  31  |  30-20   |      19-0     |
+//     | Sign | Exponent | High Fraction |
+// Low is:
+// Bit |     31-0     |
+//     | Low Fraction |
+// We then access the components through logical bit-wise operations to
+// extract the parts needed, with the positions and masks derived from the
+// layout shown above.
+constexpr uint64_t kSignMask = 0x8000000000000000LL;
+constexpr uint64_t kExponentMask = 0x7ff0000000000000LL;
+constexpr int32_t kExponentShift = 52;
+constexpr int32_t kExponentBias = 1023;
+constexpr uint32_t kExponentIsBadNum = 0x7ff;
+constexpr uint64_t kFractionMask = 0x000fffffffc00000LL;
+constexpr uint32_t kFractionShift = 22;
+constexpr uint32_t kFractionRoundingMask = 0x003fffff;
+constexpr uint32_t kFractionRoundingThreshold = 0x00200000;
+}  // namespace
+
+void QuantizeMultiplier(double double_multiplier, int32_t* quantized_multiplier,
+                        int* shift) {
+  if (double_multiplier == 0.) {
+    *quantized_multiplier = 0;
+    *shift = 0;
+    return;
+  }
+#ifdef TFLITE_EMULATE_FLOAT
+  // If we're trying to avoid the use of floating-point instructions (for
+  // example on microcontrollers) then use an alternative implementation
+  // that only requires integer and bitwise operations. To enable this, you
+  // need to set the define during the build process for your platform.
+  int64_t q_fixed = IntegerFrExp(double_multiplier, shift);
+#else   // TFLITE_EMULATE_FLOAT
+  const double q = std::frexp(double_multiplier, shift);
+  auto q_fixed = static_cast<int64_t>(TfLiteRound(q * (1ll << 31)));
+#endif  // TFLITE_EMULATE_FLOAT
+  TFLITE_CHECK(q_fixed <= (1ll << 31));
+  if (q_fixed == (1ll << 31)) {
+    q_fixed /= 2;
+    ++*shift;
+  }
+  TFLITE_CHECK_LE(q_fixed, std::numeric_limits<int32_t>::max());
+  // A shift amount smaller than -31 would cause all bits to be shifted out
+  // and thus all results would be zero. We implement that instead with
+  // q_fixed==0, so as to avoid hitting issues with right-shift
+  // operations with shift amounts greater than 31. Note that this happens
+  // roughly when abs(double_multiplier) < 2^-31 and the present handling means
+  // that we're effectively flushing tiny double_multiplier's to zero.
+  // We could conceivably handle values in the range (roughly) [32, 63]
+  // as 'denormals' i.e. (shift==0, q_fixed < 2^30). In that point of view
+  // the present handling is just doing 'flush denormals to zero'. We could
+  // reconsider and actually generate nonzero denormals if a need arises.
+  if (*shift < -31) {
+    *shift = 0;
+    q_fixed = 0;
+  }
+  *quantized_multiplier = static_cast<int32_t>(q_fixed);
+}
+
+void QuantizeMultiplierGreaterThanOne(double double_multiplier,
+                                      int32_t* quantized_multiplier,
+                                      int* left_shift) {
+  TFLITE_CHECK_GT(double_multiplier, 1.);
+  QuantizeMultiplier(double_multiplier, quantized_multiplier, left_shift);
+  TFLITE_CHECK_GE(*left_shift, 0);
+}
+
+void QuantizeMultiplierSmallerThanOneExp(double double_multiplier,
+                                         int32_t* quantized_multiplier,
+                                         int* left_shift) {
+  TFLITE_CHECK_LT(double_multiplier, 1.);
+  TFLITE_CHECK_GT(double_multiplier, 0.);
+  int shift;
+  QuantizeMultiplier(double_multiplier, quantized_multiplier, &shift);
+  TFLITE_CHECK_LE(shift, 0);
+  *left_shift = shift;
+}
+
+int64_t IntegerFrExp(double input, int* shift) {
+  // Make sure our assumptions about the double layout hold.
+  TFLITE_CHECK_EQ(8, sizeof(double));
+
+  // We want to access the bits of the input double value directly, which is
+  // tricky to do safely, so use a union to handle the casting.
+  union {
+    double double_value;
+    uint64_t double_as_uint;
+  } cast_union;
+  cast_union.double_value = input;
+  const uint64_t u = cast_union.double_as_uint;
+
+  // If the bitfield is all zeros apart from the sign bit, this is a normalized
+  // zero value, so return standard values for this special case.
+  if ((u & ~kSignMask) == 0) {
+    *shift = 0;
+    return 0;
+  }
+
+  // Deal with NaNs and Infs, which are always indicated with a fixed pattern in
+  // the exponent, and distinguished by whether the fractions are zero or
+  // non-zero.
+  const uint32_t exponent_part = ((u & kExponentMask) >> kExponentShift);
+  if (exponent_part == kExponentIsBadNum) {
+    *shift = std::numeric_limits<int>::max();
+    if (u & kFractionMask) {
+      // NaN, so just return zero (with the exponent set to INT_MAX).
+      return 0;
+    } else {
+      // Infinity, so return +/- INT_MAX.
+      if (u & kSignMask) {
+        return std::numeric_limits<int64_t>::min();
+      } else {
+        return std::numeric_limits<int64_t>::max();
+      }
+    }
+  }
+
+  // The shift is fairly easy to extract from the high bits of the double value,
+  // just by masking it out and applying a bias. The std::frexp() implementation
+  // always returns values between 0.5 and 1.0 though, whereas the exponent
+  // assumes 1.0 to 2.0 is the standard range, so I add on one to match that
+  // interface.
+  *shift = (exponent_part - kExponentBias) + 1;
+
+  // There's an implicit high bit in the double format definition, so make sure
+  // we include that at the top, and then reconstruct the rest of the fractional
+  // value from the remaining fragments.
+  int64_t fraction = 0x40000000 + ((u & kFractionMask) >> kFractionShift);
+
+  // We're cutting off some bits at the bottom, so to exactly match the standard
+  // frexp implementation here we'll apply rounding by adding one to the least
+  // significant bit of the result if the discarded portion is over half of the
+  // maximum.
+  if ((u & kFractionRoundingMask) > kFractionRoundingThreshold) {
+    fraction += 1;
+  }
+  // Negate the fraction if the sign bit was set.
+  if (u & kSignMask) {
+    fraction *= -1;
+  }
+
+  return fraction;
+}
+
+double DoubleFromFractionAndShift(int64_t fraction, int shift) {
+  union {
+    double double_value;
+    uint64_t double_as_uint;
+  } result;
+
+  // Detect NaNs and infinities.
+  if (shift == std::numeric_limits<int>::max()) {
+    if (fraction == 0) {
+      return std::numeric_limits<double>::quiet_NaN();
+    } else if (fraction > 0) {
+      return std::numeric_limits<double>::infinity();
+    } else {
+      return -std::numeric_limits<double>::infinity();
+    }
+  }
+
+  // Return a normalized zero for a zero fraction.
+  if (fraction == 0) {
+    result.double_as_uint = 0;
+    return result.double_value;
+  }
+
+  bool is_negative = (fraction < 0);
+  int64_t encoded_fraction = is_negative ? -fraction : fraction;
+  int64_t encoded_shift = (shift - 1);
+  while (encoded_fraction < 0x40000000) {
+    encoded_fraction *= 2;
+    encoded_shift -= 1;
+  }
+  while (encoded_fraction > 0x80000000) {
+    encoded_fraction /= 2;
+    encoded_shift += 1;
+  }
+  encoded_fraction -= 0x40000000;
+  if (encoded_shift < -1022) {
+    encoded_shift = -1023;
+  } else if (encoded_shift > 1022) {
+    encoded_shift = 1023;
+  }
+  encoded_shift += kExponentBias;
+  uint64_t encoded_sign = is_negative ? kSignMask : 0;
+  result.double_as_uint = encoded_sign | (encoded_shift << kExponentShift) |
+                          (encoded_fraction << kFractionShift);
+  return result.double_value;
+}
+
+double IntegerDoubleMultiply(double a, double b) {
+  int a_shift;
+  const int64_t a_fraction = IntegerFrExp(a, &a_shift);
+  int b_shift;
+  const int64_t b_fraction = IntegerFrExp(b, &b_shift);
+  // Detect NaNs and infinities.
+  if (a_shift == std::numeric_limits<int>::max() ||
+      (b_shift == std::numeric_limits<int>::max())) {
+    return std::numeric_limits<double>::quiet_NaN();
+  }
+  const int result_shift = a_shift + b_shift + 1;
+  const int64_t result_fraction = (a_fraction * b_fraction) >> 32;
+  return DoubleFromFractionAndShift(result_fraction, result_shift);
+}
+
+int IntegerDoubleCompare(double a, double b) {
+  int a_shift;
+  const int64_t a_fraction = IntegerFrExp(a, &a_shift);
+  int b_shift;
+  const int64_t b_fraction = IntegerFrExp(b, &b_shift);
+
+  // Detect NaNs and infinities.
+  if (a_shift == std::numeric_limits<int>::max() ||
+      (b_shift == std::numeric_limits<int>::max())) {
+    return 1;
+  }
+
+  if ((a_fraction == 0) && (b_fraction < 0)) {
+    return 1;
+  } else if ((a_fraction < 0) && (b_fraction == 0)) {
+    return -1;
+  } else if (a_shift < b_shift) {
+    return -1;
+  } else if (a_shift > b_shift) {
+    return 1;
+  } else if (a_fraction < b_fraction) {
+    return -1;
+  } else if (a_fraction > b_fraction) {
+    return 1;
+  } else {
+    return 0;
+  }
+}
+
+void PreprocessSoftmaxScaling(double beta, double input_scale,
+                              int input_integer_bits,
+                              int32_t* quantized_multiplier, int* left_shift) {
+  // If the overall multiplier (input and beta) is large, then exp() of an
+  // input difference of 1 scaled by this will be large.  In other words, we
+  // can cap the multiplier and know that, when it is used, the output will be
+  // (round to) zero wherever the input is not at the maximum value.
+
+  // If the overall scale is less than one, and input_integer_bits=0, then the
+  // result is double equivalent of Q0.31 (actually with more precision). Thus
+  // this generates a Q(input_integer_bits).(31-input_integer_bits)
+  // representation.
+#ifdef TFLITE_EMULATE_FLOAT
+  const double input_beta = IntegerDoubleMultiply(beta, input_scale);
+  int shift;
+  int64_t fraction = IntegerFrExp(input_beta, &shift);
+  shift += (31 - input_integer_bits);
+  double input_beta_real_multiplier =
+      DoubleFromFractionAndShift(fraction, shift);
+  if (IntegerDoubleCompare(input_beta_real_multiplier, (1ll << 31) - 1.0) > 0) {
+    input_beta_real_multiplier = (1ll << 31) - 1.0;
+  }
+#else   // TFLITE_EMULATE_FLOAT
+  const double input_beta_real_multiplier = std::min<double>(
+      beta * input_scale * (1 << (31 - input_integer_bits)), (1ll << 31) - 1.0);
+#endif  // TFLITE_EMULATE_FLOAT
+
+  QuantizeMultiplierGreaterThanOne(input_beta_real_multiplier,
+                                   quantized_multiplier, left_shift);
+}
+
+void PreprocessLogSoftmaxScalingExp(double beta, double input_scale,
+                                    int input_integer_bits,
+                                    int32_t* quantized_multiplier,
+                                    int* left_shift,
+                                    int32_t* reverse_scaling_divisor,
+                                    int* reverse_scaling_left_shift) {
+  PreprocessSoftmaxScaling(beta, input_scale, input_integer_bits,
+                           quantized_multiplier, left_shift);
+
+  // Also calculate what amounts to the inverse scaling factor for the input.
+  const double real_reverse_scaling_divisor =
+      (1 << (31 - *left_shift)) / static_cast<double>(*quantized_multiplier);
+  tflite::QuantizeMultiplierSmallerThanOneExp(real_reverse_scaling_divisor,
+                                              reverse_scaling_divisor,
+                                              reverse_scaling_left_shift);
+}
+
+int CalculateInputRadius(int input_integer_bits, int input_left_shift,
+                         int total_signed_bits) {
+#ifdef TFLITE_EMULATE_FLOAT
+  int64_t result = (1 << input_integer_bits) - 1;
+  result <<= (total_signed_bits - input_integer_bits);
+  result >>= input_left_shift;
+  return result;
+#else   // TFLITE_EMULATE_FLOAT
+  const double max_input_rescaled =
+      1.0 * ((1 << input_integer_bits) - 1) *
+      (1ll << (total_signed_bits - input_integer_bits)) /
+      (1ll << input_left_shift);
+  // Tighten bound using floor.  Suppose that we could use the exact value.
+  // After scaling the difference, the result would be at the maximum.  Thus we
+  // must ensure that our value has lower magnitude.
+  return static_cast<int>(std::floor(max_input_rescaled));
+#endif  // TFLITE_EMULATE_FLOAT
+}
+
+void NudgeQuantizationRange(const float min, const float max,
+                            const int quant_min, const int quant_max,
+                            float* nudged_min, float* nudged_max,
+                            float* nudged_scale) {
+  // This code originates from tensorflow/core/kernels/fake_quant_ops_functor.h.
+  const float quant_min_float = static_cast<float>(quant_min);
+  const float quant_max_float = static_cast<float>(quant_max);
+  *nudged_scale = (max - min) / (quant_max_float - quant_min_float);
+  const float zero_point_from_min = quant_min_float - min / *nudged_scale;
+  uint16_t nudged_zero_point;
+  if (zero_point_from_min < quant_min_float) {
+    nudged_zero_point = static_cast<uint16_t>(quant_min);
+  } else if (zero_point_from_min > quant_max_float) {
+    nudged_zero_point = static_cast<uint16_t>(quant_max);
+  } else {
+    nudged_zero_point = static_cast<uint16_t>(TfLiteRound(zero_point_from_min));
+  }
+  *nudged_min = (quant_min_float - nudged_zero_point) * (*nudged_scale);
+  *nudged_max = (quant_max_float - nudged_zero_point) * (*nudged_scale);
+}
+
+void FakeQuantizeArray(const float nudged_scale, const float nudged_min,
+                       const float nudged_max, const float* input_data,
+                       float* output_data, const float size) {
+  // This code originates from tensorflow/core/kernels/fake_quant_ops_functor.h.
+  const float inv_nudged_scale = 1.0f / nudged_scale;
+
+  for (int i = 0; i < size; i++) {
+    const float src_val = input_data[i];
+    const float clamped = std::min(nudged_max, std::max(nudged_min, src_val));
+    const float clamped_shifted = clamped - nudged_min;
+    const float dst_val =
+        TfLiteRound(clamped_shifted * inv_nudged_scale) * nudged_scale +
+        nudged_min;
+    output_data[i] = dst_val;
+  }
+}
+
+bool CheckedLog2(const float x, int* log2_result) {
+  // Using TfLiteRound instead of std::round and std::log instead of
+  // std::log2 to work around these functions being missing in a toolchain
+  // used in some TensorFlow tests as of May 2018.
+  const float x_log2 = std::log(x) * (1.0f / std::log(2.0f));
+  const float x_log2_rounded = TfLiteRound(x_log2);
+  const float x_log2_fracpart = x_log2 - x_log2_rounded;
+
+  *log2_result = static_cast<int>(x_log2_rounded);
+  return std::abs(x_log2_fracpart) < 1e-3f;
+}
+
+void QuantizeMultiplierArray(const double* effective_scales, size_t size,
+                             int32_t* effective_scale_significand,
+                             int* effective_shift) {
+  for (size_t i = 0; i < size; ++i) {
+    QuantizeMultiplier(effective_scales[i], &effective_scale_significand[i],
+                       &effective_shift[i]);
+  }
+}
+
+}  // namespace tflite

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/quantization_util.h

@@ -0,0 +1,292 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_QUANTIZATION_UTIL_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_QUANTIZATION_UTIL_H_
+
+#include <cmath>
+#include <cstdint>
+#include <limits>
+
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+// Given the min and max values of a float array, return
+// reasonable quantization parameters to use for this array.
+template <typename T>
+QuantizationParams ChooseQuantizationParams(double rmin, double rmax,
+                                            bool narrow_range) {
+  const T qmin = std::numeric_limits<T>::min() + (narrow_range ? 1 : 0);
+  const T qmax = std::numeric_limits<T>::max();
+  const double qmin_double = qmin;
+  const double qmax_double = qmax;
+  // 0 should always be a representable value. Let's assume that the initial
+  // min,max range contains 0.
+  TFLITE_CHECK_LE(rmin, 0.);
+  TFLITE_CHECK_GE(rmax, 0.);
+  if (rmin == rmax) {
+    // Special case where the min,max range is a point. Should be {0}.
+    TFLITE_CHECK_EQ(rmin, 0.);
+    TFLITE_CHECK_EQ(rmax, 0.);
+    QuantizationParams quantization_params;
+    quantization_params.zero_point = 0;
+    quantization_params.scale = 0.;
+    return quantization_params;
+  }
+
+  // General case.
+  //
+  // First determine the scale.
+  const double scale = (rmax - rmin) / (qmax_double - qmin_double);
+
+  // Zero-point computation.
+  // First the initial floating-point computation. The zero-point can be
+  // determined from solving an affine equation for any known pair
+  // (real value, corresponding quantized value).
+  // We know two such pairs: (rmin, qmin) and (rmax, qmax).
+  // The arithmetic error on the zero point computed from either pair
+  // will be roughly machine_epsilon * (sum of absolute values of terms)
+  // so we want to use the variant that adds the smaller terms.
+  const double zero_point_from_min = qmin_double - rmin / scale;
+  const double zero_point_from_max = qmax_double - rmax / scale;
+  const double zero_point_from_min_error =
+      std::abs(qmin_double) + std::abs(rmin / scale);
+  const double zero_point_from_max_error =
+      std::abs(qmax_double) + std::abs(rmax / scale);
+
+  const double zero_point_double =
+      zero_point_from_min_error < zero_point_from_max_error
+          ? zero_point_from_min
+          : zero_point_from_max;
+
+  // Now we need to nudge the zero point to be an integer
+  // (our zero points are integer, and this is motivated by the requirement
+  // to be able to represent the real value "0" exactly as a quantized value,
+  // which is required in multiple places, for example in Im2col with SAME
+  // padding).
+  T nudged_zero_point = 0;
+  if (zero_point_double < qmin_double) {
+    nudged_zero_point = qmin;
+  } else if (zero_point_double > qmax_double) {
+    nudged_zero_point = qmax;
+  } else {
+    nudged_zero_point = static_cast<T>(round(zero_point_double));
+  }
+  // The zero point should always be in the range of quantized value,
+  // [qmin, qmax].
+  TFLITE_CHECK_GE(nudged_zero_point, qmin);
+  TFLITE_CHECK_LE(nudged_zero_point, qmax);
+
+  // Finally, store the result nudged quantization params.
+  QuantizationParams quantization_params;
+  quantization_params.zero_point = nudged_zero_point;
+  quantization_params.scale = scale;
+  return quantization_params;
+}
+
+template <typename T>
+QuantizationParams ChooseQuantizationParams(double rmin, double rmax) {
+  return ChooseQuantizationParams<T>(rmin, rmax, false);
+}
+
+// Converts a floating-point number to an integer. For all inputs x where
+// static_cast<IntOut>(x) is legal according to the C++ standard, the result
+// is identical to that cast (i.e. the result is x with its fractional part
+// truncated whenever that is representable as IntOut).
+//
+// static_cast would cause undefined behavior for the following cases, which
+// have well-defined behavior for this function:
+//
+//  1. If x is NaN, the result is zero.
+//
+//  2. If the truncated form of x is above the representable range of IntOut,
+//     the result is std::numeric_limits<IntOut>::max().
+//
+//  3. If the truncated form of x is below the representable range of IntOut,
+//     the result is std::numeric_limits<IntOut>::min().
+//
+// Note that cases #2 and #3 cover infinities as well as finite numbers.
+//
+// The range of FloatIn must include the range of IntOut, otherwise
+// the results are undefined.
+// TODO(sfeuz): Replace by absl::SafeCast once available.
+template <class IntOut, class FloatIn>
+IntOut SafeCast(FloatIn x) {
+  static_assert(!std::numeric_limits<FloatIn>::is_integer,
+                "FloatIn is integer");
+  static_assert(std::numeric_limits<IntOut>::is_integer,
+                "IntOut is not integer");
+  static_assert(std::numeric_limits<IntOut>::radix == 2, "IntOut is base 2");
+
+  // Special case NaN, for which the logic below doesn't work.
+  if (std::isnan(x)) {
+    return 0;
+  }
+
+  // Negative values all clip to zero for unsigned results.
+  if (!std::numeric_limits<IntOut>::is_signed && x < 0) {
+    return 0;
+  }
+
+  // Handle infinities.
+  if (std::isinf(x)) {
+    return x < 0 ? std::numeric_limits<IntOut>::min()
+                 : std::numeric_limits<IntOut>::max();
+  }
+
+  // Set exp such that x == f * 2^exp for some f with |f| in [0.5, 1.0),
+  // unless x is zero in which case exp == 0. Note that this implies that the
+  // magnitude of x is strictly less than 2^exp.
+  int exp = 0;
+  std::frexp(x, &exp);
+
+  // Let N be the number of non-sign bits in the representation of IntOut. If
+  // the magnitude of x is strictly less than 2^N, the truncated version of x
+  // is representable as IntOut. The only representable integer for which this
+  // is not the case is kMin for signed types (i.e. -2^N), but that is covered
+  // by the fall-through below.
+  if (exp <= std::numeric_limits<IntOut>::digits) {
+    return x;
+  }
+
+  // Handle numbers with magnitude >= 2^N.
+  return x < 0 ? std::numeric_limits<IntOut>::min()
+               : std::numeric_limits<IntOut>::max();
+}
+
+// Decompose a double multiplier into a Q0.31 int32 representation of its
+// significand, and shift representation of NEGATIVE its exponent ---
+// this is intended as a RIGHT-shift.
+//
+// Restricted to the case where the multiplier < 1 (and non-negative).
+void QuantizeMultiplierSmallerThanOneExp(double double_multiplier,
+                                         int32_t* quantized_multiplier,
+                                         int* left_shift);
+
+// Decompose a double multiplier into a Q0.31 int32 representation of its
+// significand, and shift representation of its exponent.
+//
+// Restricted to the case where the multiplier > 1.
+void QuantizeMultiplierGreaterThanOne(double double_multiplier,
+                                      int32_t* quantized_multiplier,
+                                      int* left_shift);
+
+// Decompose a double multiplier into a Q0.31 int32 representation of its
+// significand, and shift representation of its exponent.
+//
+// Handles an arbitrary positive multiplier. The 'shift' output-value is
+// basically the 'floating-point exponent' of the multiplier:
+// Negative for a right-shift (when the multiplier is <1), positive for a
+// left-shift (when the multiplier is >1)
+void QuantizeMultiplier(double double_multiplier, int32_t* quantized_multiplier,
+                        int* shift);
+
+// Splits a double input value into a returned fraction, and a shift value from
+// the exponent, using only bitwise and integer operations to support
+// microcontrollers and other environments without floating-point support.
+//
+// This is designed to be a replacement for how std::frexp() is used within the
+// QuantizeMultiplier() function, and so has a different signature than the
+// standard version, returning a 64-bit integer rather than a double. This
+// result has a maximum value of 1<<31, with the fraction expressed as a
+// proportion of that maximum.
+//
+// std::frexp() returns NaNs and infinities unmodified, but since we're
+// returning integers that can't represent those values, instead we return
+// a shift of std::numeric_limits<int>::max() for all bad numbers, with an int64
+// result of 0 for NaNs, std:numeric_limits<int64_t>::max() for +INFINITY, and
+// std::numeric_limits<int64_t>::min() for -INFINITY. Denormalized inputs will
+// result in return values that end up truncating some bits at the end,
+// reflecting the loss of precision inherent in denormalization.
+int64_t IntegerFrExp(double input, int* shift);
+
+// Converts an integer fraction in the format produced by IntegerFrExp (where
+// 0x40000000 is 1.0) and an exponent shift (between -1022 and +1022) into an
+// IEEE binary64 double format result. The implementation uses only integer and
+// bitwise operators, so no floating point hardware support or emulation is
+// needed. This is here so quantized operations can run non-time-critical
+// preparation calculations on microcontrollers and other platforms without
+// float support.
+double DoubleFromFractionAndShift(int64_t fraction, int shift);
+
+// Performs a multiplication of two numbers in double format, using only integer
+// and bitwise instructions. This is aimed at supporting housekeeping functions
+// for quantized operations on microcontrollers without floating-point hardware.
+double IntegerDoubleMultiply(double a, double b);
+
+// Returns -1 if a is less than b, 0 if a and b are equal, and +1 if a is
+// greater than b. It is implemented using only integer and logical instructions
+// so that it can be easily run on microcontrollers for quantized operations.
+int IntegerDoubleCompare(double a, double b);
+
+// This first creates a multiplier in a double equivalent of
+// Q(input_integer_bits).(31-input_integer_bits) representation, with extra
+// precision in the double's fractional bits.  It then splits the result into
+// significand and exponent.
+void PreprocessSoftmaxScaling(double beta, double input_scale,
+                              int input_integer_bits,
+                              int32_t* quantized_multiplier, int* left_shift);
+// Like PreprocessSoftmaxScaling, but inverse scaling factors also calculated.
+void PreprocessLogSoftmaxScalingExp(double beta, double input_scale,
+                                    int input_integer_bits,
+                                    int32_t* quantized_multiplier,
+                                    int* left_shift,
+                                    int32_t* reverse_scaling_divisor,
+                                    int* reverse_scaling_left_shift);
+// Calculate the largest input that will result in a within-bounds intermediate
+// result within MultiplyByQuantizedMultiplierGreaterThanOne.  In other words,
+// it must not overflow before we reduce the value by multiplication by the
+// input multiplier.  The negative radius is used as the minimum difference in
+// Softmax.
+int CalculateInputRadius(int input_integer_bits, int input_left_shift,
+                         int total_signed_bits = 31);
+
+// Nudges a min/max quantization range to ensure zero is zero.
+// Gymnastics with nudged zero point is to ensure that real zero maps to
+// an integer, which is required for e.g. zero-padding in convolutional layers.
+// Outputs nudged_min, nudged_max, nudged_scale.
+void NudgeQuantizationRange(const float min, const float max,
+                            const int quant_min, const int quant_max,
+                            float* nudged_min, float* nudged_max,
+                            float* nudged_scale);
+
+// Fake quantizes (quantizes and dequantizes) input_data using the scale,
+// nudged_min, and nudged_max from NudgeQuantizationRange. This matches the code
+// in TensorFlow's FakeQuantizeWithMinMaxVarsFunctor.
+void FakeQuantizeArray(const float nudged_scale, const float nudged_min,
+                       const float nudged_max, const float* input_data,
+                       float* output_data, const float size);
+
+// If x is approximately a power of two (with any positive or negative
+// exponent), stores that exponent (i.e. log2(x)) in *log2_result, otherwise
+// returns false.
+bool CheckedLog2(const float x, int* log2_result);
+
+// Decomposes an array of double multipliers into a Q0.31 int32 representation
+// of its significand, and shift representation of its exponent.
+//
+// Handles an arbitrary multiplier. The 'shift' output-value is
+// basically the 'floating-point exponent' of the multiplier:
+// Negative for a right-shift (when the multiplier is <1), positive for a
+// left-shift (when the multiplier is >1)
+void QuantizeMultiplierArray(const double* effective_scales, size_t size,
+                             int32_t* effective_scale_significand,
+                             int* effective_shift);
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_QUANTIZATION_UTIL_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/add.h

@@ -0,0 +1,446 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ADD_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ADD_H_
+
+#include "fixedpoint/fixedpoint.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename T>
+inline void Add(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const T* input1_data,
+                const RuntimeShape& input2_shape, const T* input2_data,
+                const RuntimeShape& output_shape, T* output_data) {
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    output_data[i] = ActivationFunctionWithMinMax(
+        input1_data[i] + input2_data[i], params.quantized_activation_min,
+        params.quantized_activation_max);
+  }
+}
+
+inline void Add(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const float* input1_data,
+                const RuntimeShape& input2_shape, const float* input2_data,
+                const RuntimeShape& output_shape, float* output_data) {
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+  for (int i = 0; i < flat_size; i++) {
+    auto x = input1_data[i] + input2_data[i];
+    output_data[i] = ActivationFunctionWithMinMax(
+        x, params.float_activation_min, params.float_activation_max);
+  }
+}
+
+// Element-wise add that can often be used for inner loop of broadcast add as
+// well as the non-broadcast add.
+
+// This function is used for 8-bit as well as for 16-bit, but the accumulator
+// is 32-bit for both cases. The overflow does not happen due to the
+// choice of the shift (20 or 15, accordingly - see add.cc for more comments).
+template <typename T>
+inline void AddElementwise(int size, const ArithmeticParams& params,
+                           const T* input1_data, const T* input2_data,
+                           T* output_data) {
+  TFLITE_DCHECK_GT(params.input1_offset, -std::numeric_limits<T>::max());
+  TFLITE_DCHECK_GT(params.input2_offset, -std::numeric_limits<T>::max());
+  TFLITE_DCHECK_LT(params.input1_offset, std::numeric_limits<T>::max());
+  TFLITE_DCHECK_LT(params.input2_offset, std::numeric_limits<T>::max());
+
+  for (int i = 0; i < size; ++i) {
+    const int32_t input1_val = params.input1_offset + input1_data[i];
+    const int32_t input2_val = params.input2_offset + input2_data[i];
+    const int32_t shifted_input1_val = input1_val * (1 << params.left_shift);
+    const int32_t shifted_input2_val = input2_val * (1 << params.left_shift);
+    const int32_t scaled_input1_val =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            shifted_input1_val, params.input1_multiplier, params.input1_shift);
+    const int32_t scaled_input2_val =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            shifted_input2_val, params.input2_multiplier, params.input2_shift);
+    const int32_t raw_sum = scaled_input1_val + scaled_input2_val;
+    const int32_t raw_output =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            raw_sum, params.output_multiplier, params.output_shift) +
+        params.output_offset;
+    const int32_t clamped_output =
+        std::min(params.quantized_activation_max,
+                 std::max(params.quantized_activation_min, raw_output));
+    output_data[i] = static_cast<T>(clamped_output);
+  }
+}
+
+// Scalar-broadcast add that can be used for inner loop of more general
+// broadcast add, so that, for example, scalar-broadcast with batch will still
+// be fast.
+inline void AddScalarBroadcast(int size, const ArithmeticParams& params,
+                               uint8_t input1_data, const uint8_t* input2_data,
+                               uint8_t* output_data) {
+  TFLITE_DCHECK_GT(params.input1_offset, -256);
+  TFLITE_DCHECK_GT(params.input2_offset, -256);
+  TFLITE_DCHECK_LT(params.input1_offset, 256);
+  TFLITE_DCHECK_LT(params.input2_offset, 256);
+
+  const int32_t input1_val = params.input1_offset + input1_data;
+  const int32_t shifted_input1_val = input1_val * (1 << params.left_shift);
+  const int32_t scaled_input1_val =
+      MultiplyByQuantizedMultiplierSmallerThanOneExp(
+          shifted_input1_val, params.input1_multiplier, params.input1_shift);
+  for (int i = 0; i < size; ++i) {
+    const int32_t input2_val = params.input2_offset + input2_data[i];
+    const int32_t shifted_input2_val = input2_val * (1 << params.left_shift);
+    const int32_t scaled_input2_val =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            shifted_input2_val, params.input2_multiplier, params.input2_shift);
+    const int32_t raw_sum = scaled_input1_val + scaled_input2_val;
+    const int32_t raw_output =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            raw_sum, params.output_multiplier, params.output_shift) +
+        params.output_offset;
+    const int32_t clamped_output =
+        std::min(params.quantized_activation_max,
+                 std::max(params.quantized_activation_min, raw_output));
+    output_data[i] = static_cast<uint8_t>(clamped_output);
+  }
+}
+
+inline void Add(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const uint8_t* input1_data,
+                const RuntimeShape& input2_shape, const uint8_t* input2_data,
+                const RuntimeShape& output_shape, uint8_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+
+  TFLITE_DCHECK_GT(params.input1_offset, -256);
+  TFLITE_DCHECK_GT(params.input2_offset, -256);
+  TFLITE_DCHECK_LT(params.input1_offset, 256);
+  TFLITE_DCHECK_LT(params.input2_offset, 256);
+  AddElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+inline void AddGeneralParamScale(const ArithmeticParams& params,
+                                 const RuntimeShape& input1_shape,
+                                 const int16_t* input1_data,
+                                 const RuntimeShape& input2_shape,
+                                 const int16_t* input2_data,
+                                 const RuntimeShape& output_shape,
+                                 int16_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+
+  int max_value = std::numeric_limits<int16_t>::max();
+
+  TFLITE_DCHECK_GT(params.input1_offset, -max_value);
+  TFLITE_DCHECK_GT(params.input2_offset, -max_value);
+  TFLITE_DCHECK_LT(params.input1_offset, max_value);
+  TFLITE_DCHECK_LT(params.input2_offset, max_value);
+  AddElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+inline void Add(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const int16_t* input1_data,
+                const RuntimeShape& input2_shape, const int16_t* input2_data,
+                const RuntimeShape& output_shape, int16_t* output_data,
+                bool pot_scale = true) {
+  if (!pot_scale) {
+    AddGeneralParamScale(params, input1_shape, input1_data, input2_shape,
+                         input2_data, output_shape, output_data);
+    return;
+  }
+
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+
+  const int input1_shift = params.input1_shift;
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+  const int16_t output_activation_min = params.quantized_activation_min;
+  const int16_t output_activation_max = params.quantized_activation_max;
+
+  TFLITE_DCHECK(input1_shift == 0 || params.input2_shift == 0);
+  TFLITE_DCHECK_LE(input1_shift, 0);
+  TFLITE_DCHECK_LE(params.input2_shift, 0);
+  const int16_t* not_shift_input =
+      input1_shift == 0 ? input1_data : input2_data;
+  const int16_t* shift_input = input1_shift == 0 ? input2_data : input1_data;
+  const int input_right_shift =
+      input1_shift == 0 ? -params.input2_shift : -input1_shift;
+
+  for (int i = 0; i < flat_size; i++) {
+    // F0 uses 0 integer bits, range [-1, 1].
+    using F0 = gemmlowp::FixedPoint<std::int16_t, 0>;
+
+    F0 input_ready_scaled = F0::FromRaw(not_shift_input[i]);
+    F0 scaled_input = F0::FromRaw(
+        gemmlowp::RoundingDivideByPOT(shift_input[i], input_right_shift));
+    F0 result = gemmlowp::SaturatingAdd(scaled_input, input_ready_scaled);
+    const int16_t raw_output = result.raw();
+    const int16_t clamped_output = std::min(
+        output_activation_max, std::max(output_activation_min, raw_output));
+    output_data[i] = clamped_output;
+  }
+}
+
+inline void BroadcastAdd4DSlow(const ArithmeticParams& params,
+                               const RuntimeShape& input1_shape,
+                               const float* input1_data,
+                               const RuntimeShape& input2_shape,
+                               const float* input2_data,
+                               const RuntimeShape& output_shape,
+                               float* output_data) {
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(4, output_shape);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
+    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
+      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
+        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
+          output_data[Offset(extended_output_shape, b, y, x, c)] =
+              ActivationFunctionWithMinMax(
+                  input1_data[SubscriptToIndex(desc1, b, y, x, c)] +
+                      input2_data[SubscriptToIndex(desc2, b, y, x, c)],
+                  params.float_activation_min, params.float_activation_max);
+        }
+      }
+    }
+  }
+}
+
+inline void BroadcastAdd4DSlow(const ArithmeticParams& params,
+                               const RuntimeShape& input1_shape,
+                               const int32_t* input1_data,
+                               const RuntimeShape& input2_shape,
+                               const int32_t* input2_data,
+                               const RuntimeShape& output_shape,
+                               int32_t* output_data) {
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(4, output_shape);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
+    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
+      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
+        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
+          output_data[Offset(extended_output_shape, b, y, x, c)] =
+              ActivationFunctionWithMinMax(
+                  input1_data[SubscriptToIndex(desc1, b, y, x, c)] +
+                      input2_data[SubscriptToIndex(desc2, b, y, x, c)],
+                  params.quantized_activation_min,
+                  params.quantized_activation_max);
+        }
+      }
+    }
+  }
+}
+
+// This function is used for 8-bit as well as for 16-bit, but the accumulator
+// is 32-bit for both cases. The overflow does not happen due to the
+// choice of the shift (20 or 15, accordingly - see add.cc for more comments).
+template <typename T>
+inline void BroadcastAdd4DSlow(
+    const ArithmeticParams& params, const RuntimeShape& input1_shape,
+    const T* input1_data, const RuntimeShape& input2_shape,
+    const T* input2_data, const RuntimeShape& output_shape, T* output_data) {
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(4, output_shape);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
+    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
+      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
+        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
+          const int32_t input1_val =
+              params.input1_offset +
+              input1_data[SubscriptToIndex(desc1, b, y, x, c)];
+          const int32_t input2_val =
+              params.input2_offset +
+              input2_data[SubscriptToIndex(desc2, b, y, x, c)];
+          const int32_t shifted_input1_val =
+              input1_val * (1 << params.left_shift);
+          const int32_t shifted_input2_val =
+              input2_val * (1 << params.left_shift);
+          const int32_t scaled_input1_val =
+              MultiplyByQuantizedMultiplierSmallerThanOneExp(
+                  shifted_input1_val, params.input1_multiplier,
+                  params.input1_shift);
+          const int32_t scaled_input2_val =
+              MultiplyByQuantizedMultiplierSmallerThanOneExp(
+                  shifted_input2_val, params.input2_multiplier,
+                  params.input2_shift);
+          const int32_t raw_sum = scaled_input1_val + scaled_input2_val;
+          const int32_t raw_output =
+              MultiplyByQuantizedMultiplierSmallerThanOneExp(
+                  raw_sum, params.output_multiplier, params.output_shift) +
+              params.output_offset;
+          const int32_t clamped_output =
+              std::min(params.quantized_activation_max,
+                       std::max(params.quantized_activation_min, raw_output));
+          output_data[Offset(extended_output_shape, b, y, x, c)] =
+              static_cast<T>(clamped_output);
+        }
+      }
+    }
+  }
+}
+
+inline void BroadcastAddFivefold(const ArithmeticParams& unswitched_params,
+                                 const RuntimeShape& unswitched_input1_shape,
+                                 const uint8_t* unswitched_input1_data,
+                                 const RuntimeShape& unswitched_input2_shape,
+                                 const uint8_t* unswitched_input2_data,
+                                 const RuntimeShape& output_shape,
+                                 uint8_t* output_data) {
+  ArithmeticParams switched_params = unswitched_params;
+  switched_params.input1_offset = unswitched_params.input2_offset;
+  switched_params.input1_multiplier = unswitched_params.input2_multiplier;
+  switched_params.input1_shift = unswitched_params.input2_shift;
+  switched_params.input2_offset = unswitched_params.input1_offset;
+  switched_params.input2_multiplier = unswitched_params.input1_multiplier;
+  switched_params.input2_shift = unswitched_params.input1_shift;
+
+  const bool use_unswitched =
+      unswitched_params.broadcast_category ==
+      tflite::BroadcastableOpCategory::kFirstInputBroadcastsFast;
+
+  const ArithmeticParams& params =
+      use_unswitched ? unswitched_params : switched_params;
+  const uint8_t* input1_data =
+      use_unswitched ? unswitched_input1_data : unswitched_input2_data;
+  const uint8_t* input2_data =
+      use_unswitched ? unswitched_input2_data : unswitched_input1_data;
+
+  // Fivefold nested loops. The second input resets its position for each
+  // iteration of the second loop. The first input resets its position at the
+  // beginning of the fourth loop. The innermost loop is an elementwise add of
+  // sections of the arrays.
+  uint8_t* output_data_ptr = output_data;
+  const uint8_t* input1_data_ptr = input1_data;
+  const uint8_t* input2_data_reset = input2_data;
+  // In the fivefold pattern, y0, y2 and y4 are not broadcast, and so shared
+  // between input shapes. y3 for input 1 is always broadcast, and so the
+  // dimension there is 1, whereas optionally y1 might be broadcast for input 2.
+  // Put another way,
+  // input1.shape.FlatSize = y0 * y1 * y2 * y4,
+  // input2.shape.FlatSize = y0 * y2 * y3 * y4.
+  int y0 = params.broadcast_shape[0];
+  int y1 = params.broadcast_shape[1];
+  int y2 = params.broadcast_shape[2];
+  int y3 = params.broadcast_shape[3];
+  int y4 = params.broadcast_shape[4];
+  if (y4 > 1) {
+    // General fivefold pattern, with y4 > 1 so there is a non-broadcast inner
+    // dimension.
+    for (int i0 = 0; i0 < y0; ++i0) {
+      const uint8_t* input2_data_ptr;
+      for (int i1 = 0; i1 < y1; ++i1) {
+        input2_data_ptr = input2_data_reset;
+        for (int i2 = 0; i2 < y2; ++i2) {
+          for (int i3 = 0; i3 < y3; ++i3) {
+            AddElementwise(y4, params, input1_data_ptr, input2_data_ptr,
+                           output_data_ptr);
+            input2_data_ptr += y4;
+            output_data_ptr += y4;
+          }
+          // We have broadcast y4 of input1 data y3 times, and now move on.
+          input1_data_ptr += y4;
+        }
+      }
+      // We have broadcast y2*y3*y4 of input2 data y1 times, and now move on.
+      input2_data_reset = input2_data_ptr;
+    }
+  } else {
+    // Special case of y4 == 1, in which the innermost loop is a single element
+    // and can be combined with the next (y3) as an inner broadcast.
+    //
+    // Note that this handles the case of pure scalar broadcast when
+    // y0 == y1 == y2 == 1. With low overhead it handles cases such as scalar
+    // broadcast with batch (as y2 > 1).
+    //
+    // NOTE The process is the same as the above general case except simplified
+    // for y4 == 1 and the loop over y3 is contained within the
+    // AddScalarBroadcast function.
+    for (int i0 = 0; i0 < y0; ++i0) {
+      const uint8_t* input2_data_ptr;
+      for (int i1 = 0; i1 < y1; ++i1) {
+        input2_data_ptr = input2_data_reset;
+        for (int i2 = 0; i2 < y2; ++i2) {
+          AddScalarBroadcast(y3, params, *input1_data_ptr, input2_data_ptr,
+                             output_data_ptr);
+          input2_data_ptr += y3;
+          output_data_ptr += y3;
+          input1_data_ptr += 1;
+        }
+      }
+      input2_data_reset = input2_data_ptr;
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ADD_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/arg_min_max.h

@@ -0,0 +1,68 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ARG_MIN_MAX_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ARG_MIN_MAX_H_
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename T1, typename T2, typename T3, typename Cmp>
+void ArgMinMax(const RuntimeShape& input1_shape, const T1* input1_data,
+               const T3* input2_data, const RuntimeShape& output_shape,
+               T2* output_data, const Cmp& cmp) {
+  TFLITE_DCHECK_GT(input1_shape.DimensionsCount(), 0);
+  TFLITE_DCHECK_EQ(input1_shape.DimensionsCount() - 1,
+                   output_shape.DimensionsCount());
+  int axis = input2_data[0];
+  if (axis < 0) {
+    axis += input1_shape.DimensionsCount();
+  }
+  const int axis_size = input1_shape.Dims(axis);
+
+  int outer_size = 1;
+  for (int i = 0; i < axis; ++i) {
+    TFLITE_DCHECK_EQ(input1_shape.Dims(i), output_shape.Dims(i));
+    outer_size *= input1_shape.Dims(i);
+  }
+
+  int inner_size = 1;
+  const int dims_count = input1_shape.DimensionsCount();
+  for (int i = axis + 1; i < dims_count; ++i) {
+    TFLITE_DCHECK_EQ(input1_shape.Dims(i), output_shape.Dims(i - 1));
+    inner_size *= input1_shape.Dims(i);
+  }
+  for (int outer = 0; outer < outer_size; ++outer) {
+    for (int inner = 0; inner < inner_size; ++inner) {
+      auto min_max_value = input1_data[outer * axis_size * inner_size + inner];
+      T2 min_max_index = 0;
+      for (int i = 1; i < axis_size; ++i) {
+        const auto& curr_value =
+            input1_data[(outer * axis_size + i) * inner_size + inner];
+        if (cmp(curr_value, min_max_value)) {
+          min_max_value = curr_value;
+          min_max_index = static_cast<T2>(i);
+        }
+      }
+      output_data[outer * inner_size + inner] = min_max_index;
+    }
+  }
+}
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ARG_MIN_MAX_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/binary_function.h

@@ -0,0 +1,80 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BINARY_FUNCTION_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BINARY_FUNCTION_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+// Also appears to duplicate MinimumMaximum.
+//
+// R: Result type. T1: Input 1 type. T2: Input 2 type.
+template <typename R, typename T1, typename T2>
+inline void BroadcastBinaryFunction4DSlow(
+    const RuntimeShape& unextended_input1_shape, const T1* input1_data,
+    const RuntimeShape& unextended_input2_shape, const T2* input2_data,
+    const RuntimeShape& unextended_output_shape, R* output_data,
+    R (*func)(T1, T2)) {
+  TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_input2_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
+  const RuntimeShape output_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_shape);
+
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(unextended_input1_shape,
+                                      unextended_input2_shape, &desc1, &desc2);
+
+  for (int b = 0; b < output_shape.Dims(0); ++b) {
+    for (int y = 0; y < output_shape.Dims(1); ++y) {
+      for (int x = 0; x < output_shape.Dims(2); ++x) {
+        for (int c = 0; c < output_shape.Dims(3); ++c) {
+          auto out_idx = Offset(output_shape, b, y, x, c);
+          auto in1_idx = SubscriptToIndex(desc1, b, y, x, c);
+          auto in2_idx = SubscriptToIndex(desc2, b, y, x, c);
+          auto in1_val = input1_data[in1_idx];
+          auto in2_val = input2_data[in2_idx];
+          output_data[out_idx] = func(in1_val, in2_val);
+        }
+      }
+    }
+  }
+}
+
+// R: Result type. T1: Input 1 type. T2: Input 2 type.
+template <typename R, typename T1, typename T2>
+inline void BinaryFunction(const RuntimeShape& input1_shape,
+                           const T1* input1_data,
+                           const RuntimeShape& input2_shape,
+                           const T2* input2_data,
+                           const RuntimeShape& output_shape, R* output_data,
+                           R (*func)(T1, T2)) {
+  const int flat_size =
+      MatchingFlatSize(input1_shape, input2_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    output_data[i] = func(input1_data[i], input2_data[i]);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_BINARY_FUNCTION_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/ceil.h

@@ -0,0 +1,37 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CEIL_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CEIL_H_
+
+#include <cmath>
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+inline void Ceil(const RuntimeShape& input_shape, const float* input_data,
+                 const RuntimeShape& output_shape, float* output_data) {
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; ++i) {
+    output_data[i] = std::ceil(input_data[i]);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CEIL_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/comparisons.h

@@ -0,0 +1,280 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_COMPARISONS_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_COMPARISONS_H_
+
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename T>
+inline bool EqualFn(T lhs, T rhs) {
+  return lhs == rhs;
+}
+
+template <typename T>
+inline bool NotEqualFn(T lhs, T rhs) {
+  return lhs != rhs;
+}
+
+template <typename T>
+inline bool GreaterFn(T lhs, T rhs) {
+  return lhs > rhs;
+}
+template <typename T>
+inline bool GreaterEqualFn(T lhs, T rhs) {
+  return lhs >= rhs;
+}
+template <typename T>
+inline bool LessFn(T lhs, T rhs) {
+  return lhs < rhs;
+}
+template <typename T>
+inline bool LessEqualFn(T lhs, T rhs) {
+  return lhs <= rhs;
+}
+
+template <typename T>
+using ComparisonFn = bool (*)(T, T);
+
+template <typename T, ComparisonFn<T> F>
+inline void ComparisonImpl(
+    const ComparisonParams& op_params, const RuntimeShape& input1_shape,
+    const T* input1_data, const RuntimeShape& input2_shape,
+    const T* input2_data, const RuntimeShape& output_shape, bool* output_data) {
+  const int64_t flatsize =
+      MatchingFlatSize(input1_shape, input2_shape, output_shape);
+  for (int64_t i = 0; i < flatsize; ++i) {
+    output_data[i] = F(input1_data[i], input2_data[i]);
+  }
+}
+
+template <ComparisonFn<float> F>
+inline void Comparison(const ComparisonParams& op_params,
+                       const RuntimeShape& input1_shape,
+                       const float* input1_data,
+                       const RuntimeShape& input2_shape,
+                       const float* input2_data,
+                       const RuntimeShape& output_shape, bool* output_data) {
+  ComparisonImpl<float, F>(op_params, input1_shape, input1_data, input2_shape,
+                           input2_data, output_shape, output_data);
+}
+
+template <typename T, ComparisonFn<int32_t> F>
+inline void ComparisonWithScaling(
+    const ComparisonParams& op_params, const RuntimeShape& input1_shape,
+    const T* input1_data, const RuntimeShape& input2_shape,
+    const T* input2_data, const RuntimeShape& output_shape, bool* output_data) {
+  int left_shift = op_params.left_shift;
+  int32_t input1_offset = op_params.input1_offset;
+  int32_t input1_multiplier = op_params.input1_multiplier;
+  int input1_shift = op_params.input1_shift;
+  int32_t input2_offset = op_params.input2_offset;
+  int32_t input2_multiplier = op_params.input2_multiplier;
+  int input2_shift = op_params.input2_shift;
+
+  const int64_t flatsize =
+      MatchingFlatSize(input1_shape, input2_shape, output_shape);
+  for (int64_t i = 0; i < flatsize; ++i) {
+    const int32_t input1_val = input1_offset + input1_data[i];
+    const int32_t input2_val = input2_offset + input2_data[i];
+    const int32_t shifted_input1_val = input1_val * (1 << left_shift);
+    const int32_t shifted_input2_val = input2_val * (1 << left_shift);
+    const int32_t scaled_input1_val =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            shifted_input1_val, input1_multiplier, input1_shift);
+    const int32_t scaled_input2_val =
+        MultiplyByQuantizedMultiplierSmallerThanOneExp(
+            shifted_input2_val, input2_multiplier, input2_shift);
+    output_data[i] = F(scaled_input1_val, scaled_input2_val);
+  }
+}
+
+struct BroadcastComparison4DSlowCommon {
+  const RuntimeShape output_shape;
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+};
+
+inline BroadcastComparison4DSlowCommon BroadcastComparison4DSlowPreprocess(
+    const RuntimeShape& unextended_input1_shape,
+    const RuntimeShape& unextended_input2_shape,
+    const RuntimeShape& unextended_output_shape) {
+  TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_input2_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(unextended_input1_shape,
+                                      unextended_input2_shape, &desc1, &desc2);
+  return {RuntimeShape::ExtendedShape(4, unextended_output_shape), desc1,
+          desc2};
+}
+
+template <typename T, ComparisonFn<T> F>
+inline void BroadcastComparison4DSlowImpl(
+    const ComparisonParams& op_params,
+    const RuntimeShape& unextended_input1_shape, const T* input1_data,
+    const RuntimeShape& unextended_input2_shape, const T* input2_data,
+    const RuntimeShape& unextended_output_shape, bool* output_data) {
+  const BroadcastComparison4DSlowCommon dims =
+      BroadcastComparison4DSlowPreprocess(unextended_input1_shape,
+                                          unextended_input2_shape,
+                                          unextended_output_shape);
+
+  for (int b = 0; b < dims.output_shape.Dims(0); ++b) {
+    for (int y = 0; y < dims.output_shape.Dims(1); ++y) {
+      for (int x = 0; x < dims.output_shape.Dims(2); ++x) {
+        for (int c = 0; c < dims.output_shape.Dims(3); ++c) {
+          output_data[Offset(dims.output_shape, b, y, x, c)] =
+              F(input1_data[SubscriptToIndex(dims.desc1, b, y, x, c)],
+                input2_data[SubscriptToIndex(dims.desc2, b, y, x, c)]);
+        }
+      }
+    }
+  }
+}
+
+template <ComparisonFn<float> F>
+inline void BroadcastComparison4DSlow(const ComparisonParams& op_params,
+                                      const RuntimeShape& input1_shape,
+                                      const float* input1_data,
+                                      const RuntimeShape& input2_shape,
+                                      const float* input2_data,
+                                      const RuntimeShape& output_shape,
+                                      bool* output_data) {
+  BroadcastComparison4DSlowImpl<float, F>(op_params, input1_shape, input1_data,
+                                          input2_shape, input2_data,
+                                          output_shape, output_data);
+}
+
+template <typename T, ComparisonFn<int32_t> F>
+inline void BroadcastComparison4DSlowWithScaling(
+    const ComparisonParams& op_params,
+    const RuntimeShape& unextended_input1_shape, const T* input1_data,
+    const RuntimeShape& unextended_input2_shape, const T* input2_data,
+    const RuntimeShape& unextended_output_shape, bool* output_data) {
+  const BroadcastComparison4DSlowCommon dims =
+      BroadcastComparison4DSlowPreprocess(unextended_input1_shape,
+                                          unextended_input2_shape,
+                                          unextended_output_shape);
+
+  int left_shift = op_params.left_shift;
+  int32_t input1_offset = op_params.input1_offset;
+  int32_t input1_multiplier = op_params.input1_multiplier;
+  int input1_shift = op_params.input1_shift;
+  int32_t input2_offset = op_params.input2_offset;
+  int32_t input2_multiplier = op_params.input2_multiplier;
+  int input2_shift = op_params.input2_shift;
+
+  for (int b = 0; b < dims.output_shape.Dims(0); ++b) {
+    for (int y = 0; y < dims.output_shape.Dims(1); ++y) {
+      for (int x = 0; x < dims.output_shape.Dims(2); ++x) {
+        for (int c = 0; c < dims.output_shape.Dims(3); ++c) {
+          const int32_t input1_val =
+              input1_offset +
+              input1_data[SubscriptToIndex(dims.desc1, b, y, x, c)];
+          const int32_t input2_val =
+              input2_offset +
+              input2_data[SubscriptToIndex(dims.desc2, b, y, x, c)];
+          const int32_t shifted_input1_val = input1_val * (1 << left_shift);
+          const int32_t shifted_input2_val = input2_val * (1 << left_shift);
+          const int32_t scaled_input1_val =
+              MultiplyByQuantizedMultiplierSmallerThanOneExp(
+                  shifted_input1_val, input1_multiplier, input1_shift);
+          const int32_t scaled_input2_val =
+              MultiplyByQuantizedMultiplierSmallerThanOneExp(
+                  shifted_input2_val, input2_multiplier, input2_shift);
+          output_data[Offset(dims.output_shape, b, y, x, c)] =
+              F(scaled_input1_val, scaled_input2_val);
+        }
+      }
+    }
+  }
+}
+
+#define TFLITE_COMPARISON_OP(name)                                             \
+  inline void name(const ComparisonParams& op_params,                          \
+                   const RuntimeShape& input1_shape, const float* input1_data, \
+                   const RuntimeShape& input2_shape, const float* input2_data, \
+                   const RuntimeShape& output_shape, bool* output_data) {      \
+    Comparison<name##Fn>(op_params, input1_shape, input1_data, input2_shape,   \
+                         input2_data, output_shape, output_data);              \
+  }                                                                            \
+  template <typename T>                                                        \
+  inline void name##NoScaling(                                                 \
+      const ComparisonParams& op_params, const RuntimeShape& input1_shape,     \
+      const T* input1_data, const RuntimeShape& input2_shape,                  \
+      const T* input2_data, const RuntimeShape& output_shape,                  \
+      bool* output_data) {                                                     \
+    ComparisonImpl<T, name##Fn>(op_params, input1_shape, input1_data,          \
+                                input2_shape, input2_data, output_shape,       \
+                                output_data);                                  \
+  }                                                                            \
+  template <typename T>                                                        \
+  inline void name##WithScaling(                                               \
+      const ComparisonParams& op_params, const RuntimeShape& input1_shape,     \
+      const T* input1_data, const RuntimeShape& input2_shape,                  \
+      const T* input2_data, const RuntimeShape& output_shape,                  \
+      bool* output_data) {                                                     \
+    ComparisonWithScaling<T, name##Fn>(op_params, input1_shape, input1_data,   \
+                                       input2_shape, input2_data,              \
+                                       output_shape, output_data);             \
+  }                                                                            \
+  template <typename T>                                                        \
+  inline void Broadcast4DSlow##name##NoScaling(                                \
+      const ComparisonParams& op_params, const RuntimeShape& input1_shape,     \
+      const T* input1_data, const RuntimeShape& input2_shape,                  \
+      const T* input2_data, const RuntimeShape& output_shape,                  \
+      bool* output_data) {                                                     \
+    BroadcastComparison4DSlowImpl<T, name##Fn>(                                \
+        op_params, input1_shape, input1_data, input2_shape, input2_data,       \
+        output_shape, output_data);                                            \
+  }                                                                            \
+  inline void Broadcast4DSlow##name(                                           \
+      const ComparisonParams& op_params, const RuntimeShape& input1_shape,     \
+      const float* input1_data, const RuntimeShape& input2_shape,              \
+      const float* input2_data, const RuntimeShape& output_shape,              \
+      bool* output_data) {                                                     \
+    BroadcastComparison4DSlow<name##Fn>(op_params, input1_shape, input1_data,  \
+                                        input2_shape, input2_data,             \
+                                        output_shape, output_data);            \
+  }                                                                            \
+  template <typename T>                                                        \
+  inline void Broadcast4DSlow##name##WithScaling(                              \
+      const ComparisonParams& op_params, const RuntimeShape& input1_shape,     \
+      const T* input1_data, const RuntimeShape& input2_shape,                  \
+      const T* input2_data, const RuntimeShape& output_shape,                  \
+      bool* output_data) {                                                     \
+    BroadcastComparison4DSlowWithScaling<T, name##Fn>(                         \
+        op_params, input1_shape, input1_data, input2_shape, input2_data,       \
+        output_shape, output_data);                                            \
+  }
+TFLITE_COMPARISON_OP(Equal);
+TFLITE_COMPARISON_OP(NotEqual);
+TFLITE_COMPARISON_OP(Greater);
+TFLITE_COMPARISON_OP(GreaterEqual);
+TFLITE_COMPARISON_OP(Less);
+TFLITE_COMPARISON_OP(LessEqual);
+#undef TFLITE_COMPARISON_OP
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_COMPARISONS_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/concatenation.h

@@ -0,0 +1,139 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONCATENATION_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONCATENATION_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename Scalar>
+inline void Concatenation(const ConcatenationParams& params,
+                          const RuntimeShape* const* input_shapes,
+                          const Scalar* const* input_data,
+                          const RuntimeShape& output_shape,
+                          Scalar* output_data) {
+  int axis = params.axis;
+  int inputs_count = params.inputs_count;
+  const int concat_dimensions = output_shape.DimensionsCount();
+  TFLITE_DCHECK_LT(axis, concat_dimensions);
+
+  int64_t concat_size = 0;
+  for (int i = 0; i < inputs_count; i++) {
+    TFLITE_DCHECK_EQ(input_shapes[i]->DimensionsCount(), concat_dimensions);
+    for (int j = 0; j < concat_dimensions; j++) {
+      if (j != axis) {
+        MatchingDim(*input_shapes[i], j, output_shape, j);
+      }
+    }
+    concat_size += input_shapes[i]->Dims(axis);
+  }
+  TFLITE_DCHECK_EQ(concat_size, output_shape.Dims(axis));
+  int64_t outer_size = 1;
+  for (int i = 0; i < axis; ++i) {
+    outer_size *= output_shape.Dims(i);
+  }
+  // For all input arrays,
+  // FlatSize() = outer_size * Dims(axis) * base_inner_size;
+  int64_t base_inner_size = 1;
+  for (int i = axis + 1; i < concat_dimensions; ++i) {
+    base_inner_size *= output_shape.Dims(i);
+  }
+
+  Scalar* output_ptr = output_data;
+  for (int k = 0; k < outer_size; k++) {
+    for (int i = 0; i < inputs_count; ++i) {
+      const int copy_size = input_shapes[i]->Dims(axis) * base_inner_size;
+      const Scalar* input_ptr = input_data[i] + k * copy_size;
+      memcpy(output_ptr, input_ptr, copy_size * sizeof(Scalar));
+      output_ptr += copy_size;
+    }
+  }
+}
+
+// TODO(b/174275780): The quantized implementation of concatentation isn't fully
+// quantized as it takes scale as a floating point value. This should be fixed
+// when optimizng this routine further.
+inline void ConcatenationWithScaling(const ConcatenationParams& params,
+                                     const RuntimeShape* const* input_shapes,
+                                     const uint8_t* const* input_data,
+                                     const RuntimeShape& output_shape,
+                                     uint8_t* output_data) {
+  int axis = params.axis;
+  const int32_t* input_zeropoint = params.input_zeropoint;
+  const float* input_scale = params.input_scale;
+  int inputs_count = params.inputs_count;
+  const int32_t output_zeropoint = params.output_zeropoint;
+  const float output_scale = params.output_scale;
+
+  const int concat_dimensions = output_shape.DimensionsCount();
+  TFLITE_DCHECK_LT(axis, concat_dimensions);
+
+  int64_t concat_size = 0;
+  for (int i = 0; i < inputs_count; i++) {
+    TFLITE_DCHECK_EQ(input_shapes[i]->DimensionsCount(), concat_dimensions);
+    for (int j = 0; j < concat_dimensions; j++) {
+      if (j != axis) {
+        MatchingDim(*input_shapes[i], j, output_shape, j);
+      }
+    }
+    concat_size += input_shapes[i]->Dims(axis);
+  }
+  TFLITE_DCHECK_EQ(concat_size, output_shape.Dims(axis));
+  int64_t outer_size = 1;
+  for (int i = 0; i < axis; ++i) {
+    outer_size *= output_shape.Dims(i);
+  }
+  // For all input arrays,
+  // FlatSize() = outer_size * Dims(axis) * base_inner_size;
+  int64_t base_inner_size = 1;
+  for (int i = axis + 1; i < concat_dimensions; ++i) {
+    base_inner_size *= output_shape.Dims(i);
+  }
+
+  const float inverse_output_scale = 1.f / output_scale;
+  uint8_t* output_ptr = output_data;
+  for (int k = 0; k < outer_size; k++) {
+    for (int i = 0; i < inputs_count; ++i) {
+      const int copy_size = input_shapes[i]->Dims(axis) * base_inner_size;
+      const uint8_t* input_ptr = input_data[i] + k * copy_size;
+      if (input_zeropoint[i] == output_zeropoint &&
+          input_scale[i] == output_scale) {
+        memcpy(output_ptr, input_ptr, copy_size);
+      } else {
+        const float scale = input_scale[i] * inverse_output_scale;
+        const float bias = -input_zeropoint[i] * scale;
+        for (int j = 0; j < copy_size; ++j) {
+          const int32_t value = static_cast<int32_t>(tflite::TfLiteRound(
+                                    input_ptr[j] * scale + bias)) +
+                                output_zeropoint;
+          output_ptr[j] = static_cast<uint8_t>(
+              std::max<int32_t>(std::min<int32_t>(255, value), 0));
+        }
+      }
+      output_ptr += copy_size;
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONCATENATION_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/conv.h

@@ -0,0 +1,264 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+inline void Conv(const ConvParams& params, const RuntimeShape& input_shape,
+                 const float* input_data, const RuntimeShape& filter_shape,
+                 const float* filter_data, const RuntimeShape& bias_shape,
+                 const float* bias_data, const RuntimeShape& output_shape,
+                 float* output_data, const RuntimeShape& im2col_shape,
+                 float* im2col_data) {
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const float output_activation_min = params.float_activation_min;
+  const float output_activation_max = params.float_activation_max;
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+
+  (void)im2col_data;   // only used in optimized code.
+  (void)im2col_shape;  // only used in optimized code.
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      const int in_y_origin = (out_y * stride_height) - pad_height;
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        const int in_x_origin = (out_x * stride_width) - pad_width;
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          float total = 0.f;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            const int in_y = in_y_origin + dilation_height_factor * filter_y;
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              const int in_x = in_x_origin + dilation_width_factor * filter_x;
+
+              // Zero padding by omitting the areas outside the image.
+              const bool is_point_inside_image =
+                  (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                  (in_y < input_height);
+
+              if (!is_point_inside_image) {
+                continue;
+              }
+
+              for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+                float input_value = input_data[Offset(input_shape, batch, in_y,
+                                                      in_x, in_channel)];
+                float filter_value = filter_data[Offset(
+                    filter_shape, out_channel, filter_y, filter_x, in_channel)];
+                total += (input_value * filter_value);
+              }
+            }
+          }
+          float bias_value = 0.0f;
+          if (bias_data) {
+            bias_value = bias_data[out_channel];
+          }
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              ActivationFunctionWithMinMax(total + bias_value,
+                                           output_activation_min,
+                                           output_activation_max);
+        }
+      }
+    }
+  }
+}
+
+inline void Conv(const ConvParams& params, const RuntimeShape& input_shape,
+                 const uint8_t* input_data, const RuntimeShape& filter_shape,
+                 const uint8_t* filter_data, const RuntimeShape& bias_shape,
+                 const int32_t* bias_data, const RuntimeShape& output_shape,
+                 uint8_t* output_data, const RuntimeShape& im2col_shape,
+                 uint8_t* im2col_data, void* cpu_backend_context) {
+  (void)cpu_backend_context;  // only used in optimized code.
+  (void)im2col_data;          // only used in optimized code.
+  (void)im2col_shape;         // only used in optimized code.
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const int32_t input_offset = params.input_offset;
+  const int32_t filter_offset = params.weights_offset;
+  const int32_t output_offset = params.output_offset;
+  const int32_t output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      const int in_y_origin = (out_y * stride_height) - pad_height;
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        const int in_x_origin = (out_x * stride_width) - pad_width;
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          int32_t acc = 0;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            const int in_y = in_y_origin + dilation_height_factor * filter_y;
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              const int in_x = in_x_origin + dilation_width_factor * filter_x;
+
+              // Zero padding by omitting the areas outside the image.
+              const bool is_point_inside_image =
+                  (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                  (in_y < input_height);
+
+              if (!is_point_inside_image) {
+                continue;
+              }
+
+              for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+                int32_t input_val = input_data[Offset(input_shape, batch, in_y,
+                                                      in_x, in_channel)];
+                int32_t filter_val = filter_data[Offset(
+                    filter_shape, out_channel, filter_y, filter_x, in_channel)];
+                acc +=
+                    (filter_val + filter_offset) * (input_val + input_offset);
+              }
+            }
+          }
+          if (bias_data) {
+            acc += bias_data[out_channel];
+          }
+          acc = MultiplyByQuantizedMultiplier(acc, output_multiplier,
+                                              output_shift);
+          acc += output_offset;
+          acc = std::max(acc, output_activation_min);
+          acc = std::min(acc, output_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              static_cast<uint8_t>(acc);
+        }
+      }
+    }
+  }
+}
+
+inline void HybridConvPerChannel(
+    const ConvParams& params, float* scaling_factors_ptr,
+    const RuntimeShape& input_shape, const int8_t* input_data,
+    const RuntimeShape& filter_shape, const int8_t* filter_data,
+    const RuntimeShape& bias_shape, const float* bias_data,
+    const RuntimeShape& output_shape, float* output_data,
+    const RuntimeShape& im2col_shape, int8_t* im2col_data,
+    const float* per_channel_scale, int32_t* input_offset) {
+  (void)im2col_data;   // only used in optimized code.
+  (void)im2col_shape;  // only used in optimized code.
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const float output_activation_min = params.float_activation_min;
+  const float output_activation_max = params.float_activation_max;
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          const int in_x_origin = (out_x * stride_width) - pad_width;
+          const int in_y_origin = (out_y * stride_height) - pad_height;
+          int32_t acc = 0;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+                const int in_x = in_x_origin + dilation_width_factor * filter_x;
+                const int in_y =
+                    in_y_origin + dilation_height_factor * filter_y;
+                // If the location is outside the bounds of the input image,
+                // use zero as a default value.
+                if ((in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                    (in_y < input_height)) {
+                  int32_t input_val = input_data[Offset(
+                      input_shape, batch, in_y, in_x, in_channel)];
+                  int32_t filter_val =
+                      filter_data[Offset(filter_shape, out_channel, filter_y,
+                                         filter_x, in_channel)];
+                  acc += filter_val * (input_val - input_offset[batch]);
+                }
+              }
+            }
+          }
+          float acc_float =
+              acc * per_channel_scale[out_channel] * scaling_factors_ptr[batch];
+          if (bias_data) {
+            acc_float += bias_data[out_channel];
+          }
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              ActivationFunctionWithMinMax(acc_float, output_activation_min,
+                                           output_activation_max);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/depthwiseconv_float.h

@@ -0,0 +1,100 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_FLOAT_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_FLOAT_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline void DepthwiseConv(
+    const DepthwiseParams& params, const RuntimeShape& input_shape,
+    const float* input_data, const RuntimeShape& filter_shape,
+    const float* filter_data, const RuntimeShape& bias_shape,
+    const float* bias_data, const RuntimeShape& output_shape,
+    float* output_data) {
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const int depth_multiplier = params.depth_multiplier;
+  const float output_activation_min = params.float_activation_min;
+  const float output_activation_max = params.float_activation_max;
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int input_depth = input_shape.Dims(3);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier);
+  TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+
+  for (int b = 0; b < batches; ++b) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int ic = 0; ic < input_depth; ++ic) {
+          for (int m = 0; m < depth_multiplier; m++) {
+            const int oc = m + ic * depth_multiplier;
+            const int in_x_origin = (out_x * stride_width) - pad_width;
+            const int in_y_origin = (out_y * stride_height) - pad_height;
+            float total = 0.f;
+            for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+              for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+                const int in_x = in_x_origin + dilation_width_factor * filter_x;
+                const int in_y =
+                    in_y_origin + dilation_height_factor * filter_y;
+                // If the location is outside the bounds of the input image,
+                // use zero as a default value.
+                if ((in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                    (in_y < input_height)) {
+                  float input_value =
+                      input_data[Offset(input_shape, b, in_y, in_x, ic)];
+                  float filter_value = filter_data[Offset(
+                      filter_shape, 0, filter_y, filter_x, oc)];
+                  total += (input_value * filter_value);
+                }
+              }
+            }
+            float bias_value = 0.0f;
+            if (bias_data) {
+              bias_value = bias_data[oc];
+            }
+            output_data[Offset(output_shape, b, out_y, out_x, oc)] =
+                ActivationFunctionWithMinMax(total + bias_value,
+                                             output_activation_min,
+                                             output_activation_max);
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // end namespace reference_ops
+}  // end namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_FLOAT_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h

@@ -0,0 +1,297 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_UINT8_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_UINT8_H_
+
+#include <algorithm>
+
+#include "fixedpoint/fixedpoint.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+// Used in tests and template parameters to control which version of depthwise
+// convolution is called. Primarily for reference code, and specializations
+// forced in tests.
+enum class DepthwiseConvImplementation {
+  // Run all tests against kUseStandardEntry even if also testing another
+  // kernel, since we need to be sure that the main DepthwiseConv() function in
+  // optimized_ops.h dispatches to a correctly-executing kernel.
+  kNone = 0,                 // The "default" option: use the normal
+                             // DepthwiseConv kernel (entry) function.
+  kUseGenericKernel,         // Forced use of generic kernel.
+  kUseNeon3x3,               // 3x3 kernel that uses NEON when available.
+  kUseNeon3x3DotProduct,     // 3x3 kernel that uses dot-product enabled NEON
+                             // when available.
+  kUseCModel3x3DotProduct,   // 3x3 kernel, reference C model that is intended
+                             // to match overall design NEON code.
+  kUseUnwound3x3DotProduct,  // 3x3 kernel, reference C model with unwound loops
+                             // and some arrays.
+  kUseIntrinsics3x3DotProduct,  // 3x3 kernel using NEON intrinsics.
+};
+
+// Category of depthwise convolution output rounding.
+enum class DepthwiseConvOutputRounding {
+  kNone = 0,      // Invalid: specific method must be specified.
+  kAwayFromZero,  // Original method: exact halves rounded away from zero.
+  kUpward,        // Halves towards +infinity: adds 0.5 before truncate.
+  // This is where a future kNearestEven would be placed.
+};
+
+// Category of depthwise convolution depth multiplication.
+enum class DepthwiseConvDepthMultiplication {
+  kNoMultiplication = 0,  // Depth multiplier = 1.
+  kUnitInputDepth,        // Input depth = 1, output depth = depth multiplier.
+};
+
+namespace reference_ops {
+namespace depthwise_conv {
+
+template <DepthwiseConvOutputRounding output_rounding>
+inline int32_t DepthwiseConvRound(int32_t x, int32_t quantized_multiplier,
+                                  int shift) {
+  TFLITE_DCHECK_NE(output_rounding, DepthwiseConvOutputRounding::kNone);
+  return MultiplyByQuantizedMultiplier(x, quantized_multiplier, shift);
+}
+
+template <>
+inline int32_t DepthwiseConvRound<DepthwiseConvOutputRounding::kAwayFromZero>(
+    int32_t x, int32_t quantized_multiplier, int shift) {
+  return MultiplyByQuantizedMultiplier(x, quantized_multiplier, shift);
+}
+
+template <>
+inline int32_t DepthwiseConvRound<DepthwiseConvOutputRounding::kUpward>(
+    int32_t x, int32_t quantized_multiplier, int shift) {
+  using gemmlowp::SaturatingRoundingDoublingHighMul;
+  const int left_shift = shift > 0 ? shift : 0;
+  const int right_shift = shift > 0 ? 0 : -shift;
+  const int rounding_offset = right_shift > 0 ? 1 << (right_shift - 1) : 0;
+  return (SaturatingRoundingDoublingHighMul(x * (1 << left_shift),
+                                            quantized_multiplier) +
+          rounding_offset) >>
+         right_shift;
+}
+
+template <DepthwiseConvOutputRounding output_rounding>
+struct DepthwiseConvBasicKernel {
+  static inline void Run(
+      const DepthwiseParams& params, const RuntimeShape& input_shape,
+      const uint8_t* input_data, const RuntimeShape& filter_shape,
+      const uint8_t* filter_data, const RuntimeShape& bias_shape,
+      const int32_t* bias_data, const RuntimeShape& output_shape,
+      uint8_t* output_data) {
+    const int stride_width = params.stride_width;
+    const int stride_height = params.stride_height;
+    const int dilation_width_factor = params.dilation_width_factor;
+    const int dilation_height_factor = params.dilation_height_factor;
+    const int pad_width = params.padding_values.width;
+    const int pad_height = params.padding_values.height;
+    const int depth_multiplier = params.depth_multiplier;
+    const int32_t output_activation_min = params.quantized_activation_min;
+    const int32_t output_activation_max = params.quantized_activation_max;
+    const int32_t input_offset = params.input_offset;
+    const int32_t filter_offset = params.weights_offset;
+    const int32_t output_offset = params.output_offset;
+    const int32_t output_multiplier = params.output_multiplier;
+    const int output_shift = params.output_shift;
+    TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+    TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+    TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+
+    TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+    const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+    const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3);
+    const int input_height = input_shape.Dims(1);
+    const int input_width = input_shape.Dims(2);
+    const int input_depth = input_shape.Dims(3);
+    const int filter_height = filter_shape.Dims(1);
+    const int filter_width = filter_shape.Dims(2);
+    const int output_height = output_shape.Dims(1);
+    const int output_width = output_shape.Dims(2);
+    TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier);
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+
+    for (int b = 0; b < batches; ++b) {
+      for (int out_y = 0; out_y < output_height; ++out_y) {
+        for (int out_x = 0; out_x < output_width; ++out_x) {
+          for (int ic = 0; ic < input_depth; ++ic) {
+            for (int m = 0; m < depth_multiplier; m++) {
+              const int oc = m + ic * depth_multiplier;
+              const int in_x_origin = (out_x * stride_width) - pad_width;
+              const int in_y_origin = (out_y * stride_height) - pad_height;
+              int32_t acc = 0;
+              for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+                for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+                  const int in_x =
+                      in_x_origin + dilation_width_factor * filter_x;
+                  const int in_y =
+                      in_y_origin + dilation_height_factor * filter_y;
+                  // If the location is outside the bounds of the input image,
+                  // use zero as a default value.
+                  if ((in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                      (in_y < input_height)) {
+                    int32_t input_val =
+                        input_data[Offset(input_shape, b, in_y, in_x, ic)];
+                    int32_t filter_val = filter_data[Offset(
+                        filter_shape, 0, filter_y, filter_x, oc)];
+                    acc += (filter_val + filter_offset) *
+                           (input_val + input_offset);
+                  }
+                }
+              }
+              if (bias_data) {
+                acc += bias_data[oc];
+              }
+              acc = DepthwiseConvRound<output_rounding>(acc, output_multiplier,
+                                                        output_shift);
+              acc += output_offset;
+              acc = std::max(acc, output_activation_min);
+              acc = std::min(acc, output_activation_max);
+              output_data[Offset(output_shape, b, out_y, out_x, oc)] =
+                  static_cast<uint8_t>(acc);
+            }
+          }
+        }
+      }
+    }
+  }
+
+  // TODO(b/148596273): Reconcile reference versions, perhaps with common
+  // MultiplyByQuantizedMultiplier or DepthwiseConvRound function.
+  static inline void RunPerChannel(
+      const DepthwiseParams& params, const RuntimeShape& input_shape,
+      const int8_t* input_data, const RuntimeShape& filter_shape,
+      const int8_t* filter_data, const RuntimeShape& bias_shape,
+      const int32_t* bias_data, const RuntimeShape& output_shape,
+      int8_t* output_data) {
+    // Get parameters.
+    // TODO(b/141565753): Re-introduce ScopedProfilingLabel on Micro.
+    const int stride_width = params.stride_width;
+    const int stride_height = params.stride_height;
+    const int dilation_width_factor = params.dilation_width_factor;
+    const int dilation_height_factor = params.dilation_height_factor;
+    const int pad_width = params.padding_values.width;
+    const int pad_height = params.padding_values.height;
+    const int depth_multiplier = params.depth_multiplier;
+    const int32_t input_offset = params.input_offset;
+    const int32_t output_offset = params.output_offset;
+    const int32_t output_activation_min = params.quantized_activation_min;
+    const int32_t output_activation_max = params.quantized_activation_max;
+    const int32_t* output_multiplier = params.output_multiplier_per_channel;
+    const int32_t* output_shift = params.output_shift_per_channel;
+
+    // Check dimensions of the tensors.
+    TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+    TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+    TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+
+    TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+    const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+    const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3);
+    const int input_height = input_shape.Dims(1);
+    const int input_width = input_shape.Dims(2);
+    const int input_depth = input_shape.Dims(3);
+    const int filter_height = filter_shape.Dims(1);
+    const int filter_width = filter_shape.Dims(2);
+    const int output_height = output_shape.Dims(1);
+    const int output_width = output_shape.Dims(2);
+    TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier);
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+
+    for (int batch = 0; batch < batches; ++batch) {
+      for (int out_y = 0; out_y < output_height; ++out_y) {
+        for (int out_x = 0; out_x < output_width; ++out_x) {
+          for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+            for (int m = 0; m < depth_multiplier; ++m) {
+              const int output_channel = m + in_channel * depth_multiplier;
+              const int in_x_origin = (out_x * stride_width) - pad_width;
+              const int in_y_origin = (out_y * stride_height) - pad_height;
+              int32_t acc = 0;
+              for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+                for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+                  const int in_x =
+                      in_x_origin + dilation_width_factor * filter_x;
+                  const int in_y =
+                      in_y_origin + dilation_height_factor * filter_y;
+                  // Zero padding by omitting the areas outside the image.
+                  const bool is_point_inside_image =
+                      (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                      (in_y < input_height);
+                  if (is_point_inside_image) {
+                    int32_t input_val = input_data[Offset(
+                        input_shape, batch, in_y, in_x, in_channel)];
+                    int32_t filter_val = filter_data[Offset(
+                        filter_shape, 0, filter_y, filter_x, output_channel)];
+                    // Accumulate with 32 bits accumulator.
+                    // In the nudging process during model quantization, we
+                    // force real value of 0.0 be represented by a quantized
+                    // value. This guarantees that the input_offset is a int8_t,
+                    // even though it is represented using int32_t. int32_t +=
+                    // int8_t
+                    // * (int8_t - int8_t) so the highest value we can get from
+                    // each accumulation is [-127, 127] * ([-128, 127] -
+                    // [-128, 127]), which is [-32512, 32512]. log2(32512)
+                    // = 14.98, which means we can accumulate at least 2^16
+                    // multiplications without overflow. The accumulator is
+                    // applied to a filter so the accumulation logic will hold
+                    // as long as the filter size (filter_y * filter_x *
+                    // in_channel) does not exceed 2^16, which is the case in
+                    // all the models we have seen so far.
+                    acc += filter_val * (input_val + input_offset);
+                  }
+                }
+              }
+              if (bias_data) {
+                acc += bias_data[output_channel];
+              }
+              acc = DepthwiseConvRound<output_rounding>(
+                  acc, output_multiplier[output_channel],
+                  output_shift[output_channel]);
+              acc += output_offset;
+              acc = std::max(acc, output_activation_min);
+              acc = std::min(acc, output_activation_max);
+              output_data[Offset(output_shape, batch, out_y, out_x,
+                                 output_channel)] = static_cast<int8_t>(acc);
+            }
+          }
+        }
+      }
+    }
+  }
+};
+
+}  // namespace depthwise_conv
+
+inline void DepthwiseConv(
+    const DepthwiseParams& params, const RuntimeShape& input_shape,
+    const uint8_t* input_data, const RuntimeShape& filter_shape,
+    const uint8_t* filter_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    uint8_t* output_data) {
+  return depthwise_conv::DepthwiseConvBasicKernel<
+      DepthwiseConvOutputRounding::kAwayFromZero>::Run(params, input_shape,
+                                                       input_data, filter_shape,
+                                                       filter_data, bias_shape,
+                                                       bias_data, output_shape,
+                                                       output_data);
+}
+
+}  // namespace reference_ops
+}  // end namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEPTHWISECONV_UINT8_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/dequantize.h

@@ -0,0 +1,78 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEQUANTIZE_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEQUANTIZE_H_
+
+#include <limits.h>
+
+#include <vector>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+// Dequantizes into a float without rounding.
+template <typename InputT, typename OutputT>
+inline void Dequantize(const tflite::DequantizationParams& op_params,
+                       const RuntimeShape& input_shape,
+                       const InputT* input_data,
+                       const RuntimeShape& output_shape, OutputT* output_data) {
+  int32_t zero_point = op_params.zero_point;
+  const double scale = op_params.scale;
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; i++) {
+    const int32_t val = input_data[i];
+    const OutputT result = static_cast<OutputT>(scale * (val - zero_point));
+    output_data[i] = result;
+  }
+}
+
+// Dequantizes per-channel quantized tensor to float.
+template <typename T>
+inline void PerChannelDequantize(
+    const tflite::PerChannelDequantizationParams& op_params,
+    const RuntimeShape& input_shape, const T* input_data,
+    const RuntimeShape& output_shape, float* output_data) {
+  // Ensure flat size is same.
+  MatchingFlatSize(input_shape, output_shape);
+
+  const int32_t* zero_point = op_params.zero_point;
+  const float* scale = op_params.scale;
+  const int32_t quantized_dimension = op_params.quantized_dimension;
+  const int32_t num_dims = input_shape.DimensionsCount();
+  const int32_t* dims_data = input_shape.DimsData();
+  std::vector<int> current_dim(num_dims, 0);
+
+  do {
+    size_t offset =
+        ReducedOutputOffset(num_dims, reinterpret_cast<const int*>(dims_data),
+                            current_dim.data(), 0, nullptr);
+    const int channel = current_dim[quantized_dimension];
+    const int32_t val = input_data[offset];
+    const float result =
+        static_cast<float>(scale[channel] * (val - zero_point[channel]));
+    output_data[offset] = result;
+  } while (NextIndex(num_dims, reinterpret_cast<const int*>(dims_data),
+                     current_dim.data()));
+}
+
+}  // namespace reference_ops
+
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_DEQUANTIZE_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/floor.h

@@ -0,0 +1,39 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FLOOR_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FLOOR_H_
+
+#include <cmath>
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+inline void Floor(const RuntimeShape& input_shape, const float* input_data,
+                  const RuntimeShape& output_shape, float* output_data) {
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; i++) {
+    int offset = i;
+    output_data[offset] = std::floor(input_data[offset]);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FLOOR_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/fully_connected.h

@@ -0,0 +1,320 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FULLY_CONNECTED_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FULLY_CONNECTED_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/quantization_util.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline void FullyConnected(
+    const FullyConnectedParams& params, const RuntimeShape& input_shape,
+    const float* input_data, const RuntimeShape& weights_shape,
+    const float* weights_data, const RuntimeShape& bias_shape,
+    const float* bias_data, const RuntimeShape& output_shape,
+    float* output_data) {
+  const float output_activation_min = params.float_activation_min;
+  const float output_activation_max = params.float_activation_max;
+  // TODO(b/62193649): This really should be:
+  //     const int batches = ArraySize(output_dims, 1);
+  // but the current --variable_batch hack consists in overwriting the 3rd
+  // dimension with the runtime batch size, as we don't keep track for each
+  // array of which dimension is the batch dimension in it.
+  const int output_dims_count = output_shape.DimensionsCount();
+  const int weights_dims_count = weights_shape.DimensionsCount();
+  const int batches = FlatSizeSkipDim(output_shape, output_dims_count - 1);
+  const int output_depth = MatchingDim(weights_shape, weights_dims_count - 2,
+                                       output_shape, output_dims_count - 1);
+  const int accum_depth = weights_shape.Dims(weights_dims_count - 1);
+  for (int b = 0; b < batches; ++b) {
+    for (int out_c = 0; out_c < output_depth; ++out_c) {
+      float total = 0.f;
+      for (int d = 0; d < accum_depth; ++d) {
+        total += input_data[b * accum_depth + d] *
+                 weights_data[out_c * accum_depth + d];
+      }
+      float bias_value = 0.0f;
+      if (bias_data) {
+        bias_value = bias_data[out_c];
+      }
+      output_data[out_c + output_depth * b] = ActivationFunctionWithMinMax(
+          total + bias_value, output_activation_min, output_activation_max);
+    }
+  }
+}
+
+inline void FullyConnected(
+    const FullyConnectedParams& params, const RuntimeShape& input_shape,
+    const uint8_t* input_data, const RuntimeShape& filter_shape,
+    const uint8_t* filter_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    uint8_t* output_data) {
+  const int32_t input_offset = params.input_offset;
+  const int32_t filter_offset = params.weights_offset;
+  const int32_t output_offset = params.output_offset;
+  const int32_t output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_GE(filter_shape.DimensionsCount(), 2);
+  TFLITE_DCHECK_GE(output_shape.DimensionsCount(), 1);
+
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  // TODO(b/62193649): This really should be:
+  //     const int batches = ArraySize(output_dims, 1);
+  // but the current --variable_batch hack consists in overwriting the 3rd
+  // dimension with the runtime batch size, as we don't keep track for each
+  // array of which dimension is the batch dimension in it.
+  const int output_dim_count = output_shape.DimensionsCount();
+  const int filter_dim_count = filter_shape.DimensionsCount();
+  const int batches = FlatSizeSkipDim(output_shape, output_dim_count - 1);
+  const int output_depth = MatchingDim(filter_shape, filter_dim_count - 2,
+                                       output_shape, output_dim_count - 1);
+  const int accum_depth = filter_shape.Dims(filter_dim_count - 1);
+  for (int b = 0; b < batches; ++b) {
+    for (int out_c = 0; out_c < output_depth; ++out_c) {
+      int32_t acc = 0;
+      for (int d = 0; d < accum_depth; ++d) {
+        int32_t input_val = input_data[b * accum_depth + d];
+        int32_t filter_val = filter_data[out_c * accum_depth + d];
+        acc += (filter_val + filter_offset) * (input_val + input_offset);
+      }
+      if (bias_data) {
+        acc += bias_data[out_c];
+      }
+      acc = MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift);
+      acc += output_offset;
+      acc = std::max(acc, output_activation_min);
+      acc = std::min(acc, output_activation_max);
+      output_data[out_c + output_depth * b] = static_cast<uint8_t>(acc);
+    }
+  }
+}
+
+inline void FullyConnected(
+    const FullyConnectedParams& params, const RuntimeShape& input_shape,
+    const uint8_t* input_data, const RuntimeShape& filter_shape,
+    const uint8_t* filter_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    int16_t* output_data) {
+  const int32_t input_offset = params.input_offset;
+  const int32_t filter_offset = params.weights_offset;
+  const int32_t output_offset = params.output_offset;
+  const int32_t output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  TFLITE_DCHECK_EQ(output_offset, 0);
+  // TODO(b/62193649): This really should be:
+  //     const int batches = ArraySize(output_dims, 1);
+  // but the current --variable_batch hack consists in overwriting the 3rd
+  // dimension with the runtime batch size, as we don't keep track for each
+  // array of which dimension is the batch dimension in it.
+  const int output_dim_count = output_shape.DimensionsCount();
+  const int filter_dim_count = filter_shape.DimensionsCount();
+  const int batches = FlatSizeSkipDim(output_shape, output_dim_count - 1);
+  const int output_depth = MatchingDim(filter_shape, filter_dim_count - 2,
+                                       output_shape, output_dim_count - 1);
+  const int accum_depth = filter_shape.Dims(filter_dim_count - 1);
+  for (int b = 0; b < batches; ++b) {
+    for (int out_c = 0; out_c < output_depth; ++out_c) {
+      // Internal accumulation.
+      // Initialize accumulator with the bias-value.
+      int32_t accum = bias_data[out_c];
+      // Accumulation loop.
+      for (int d = 0; d < accum_depth; ++d) {
+        int16_t input_val = input_data[b * accum_depth + d] + input_offset;
+        int16_t filter_val =
+            filter_data[out_c * accum_depth + d] + filter_offset;
+        accum += filter_val * input_val;
+      }
+      // Down-scale the final int32_t accumulator to the scale used by our
+      // (16-bit, typically 3 integer bits) fixed-point format. The quantized
+      // multiplier and shift here have been pre-computed offline
+      // (e.g. by toco).
+      accum =
+          MultiplyByQuantizedMultiplier(accum, output_multiplier, output_shift);
+      // Saturate, cast to int16_t, and store to output array.
+      accum = std::max(accum, output_activation_min - output_offset);
+      accum = std::min(accum, output_activation_max - output_offset);
+      accum += output_offset;
+      output_data[out_c + output_depth * b] = accum;
+    }
+  }
+}
+
+inline void ShuffledFullyConnected(
+    const FullyConnectedParams& params, const RuntimeShape& input_shape,
+    const uint8_t* input_data, const RuntimeShape& weights_shape,
+    const uint8_t* shuffled_weights_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    int16_t* output_data, uint8_t* shuffled_input_workspace_data) {
+  const int32_t output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+
+  TFLITE_DCHECK_GE(input_shape.DimensionsCount(), 1);
+  TFLITE_DCHECK_GE(weights_shape.DimensionsCount(), 2);
+  TFLITE_DCHECK_GE(output_shape.DimensionsCount(), 1);
+  // TODO(b/62193649): This really should be:
+  //     const int batches = ArraySize(output_dims, 1);
+  // but the current --variable_batch hack consists in overwriting the 3rd
+  // dimension with the runtime batch size, as we don't keep track for each
+  // array of which dimension is the batch dimension in it.
+  const int output_dim_count = output_shape.DimensionsCount();
+  const int weights_dim_count = weights_shape.DimensionsCount();
+  const int batches = FlatSizeSkipDim(output_shape, output_dim_count - 1);
+  const int output_depth = MatchingDim(weights_shape, weights_dim_count - 2,
+                                       output_shape, output_dim_count - 1);
+  const int accum_depth = weights_shape.Dims(weights_dim_count - 1);
+  TFLITE_DCHECK((accum_depth % 16) == 0);
+  TFLITE_DCHECK((output_depth % 4) == 0);
+
+  // Shuffling and xoring of input activations into the workspace buffer
+  uint8_t* shuffled_input_workspace_ptr = shuffled_input_workspace_data;
+  if (batches == 1) {
+    for (int i = 0; i < accum_depth; i++) {
+      shuffled_input_workspace_data[i] = input_data[i] ^ 0x80;
+    }
+  } else if (batches == 4) {
+    for (int c = 0; c < accum_depth; c += 16) {
+      for (int b = 0; b < 4; b++) {
+        const uint8_t* src_data_ptr = input_data + b * accum_depth + c;
+        for (int j = 0; j < 16; j++) {
+          uint8_t src_val = *src_data_ptr++;
+          // Flip the sign bit, so that the kernel will only need to
+          // reinterpret these uint8_t values as int8_t, getting for free the
+          // subtraction of the zero_point value 128.
+          uint8_t dst_val = src_val ^ 0x80;
+          *shuffled_input_workspace_ptr++ = dst_val;
+        }
+      }
+    }
+  } else {
+    TFLITE_DCHECK(false);
+    return;
+  }
+
+  // Actual computation
+  if (batches == 1) {
+    int16_t* output_ptr = output_data;
+    // Shuffled weights have had their sign bit (0x80) pre-flipped (xor'd)
+    // so that just reinterpreting them as int8_t values is equivalent to
+    // subtracting 128 from them, thus implementing for free the subtraction of
+    // the zero_point value 128.
+    const int8_t* shuffled_weights_ptr =
+        reinterpret_cast<const int8_t*>(shuffled_weights_data);
+    // Likewise, we preshuffled and pre-xored the input data above.
+    const int8_t* shuffled_input_data =
+        reinterpret_cast<const int8_t*>(shuffled_input_workspace_data);
+    for (int c = 0; c < output_depth; c += 4) {
+      // Internal accumulation.
+      // Initialize accumulator with the bias-value.
+      int32_t accum[4] = {0};
+      // Accumulation loop.
+      for (int d = 0; d < accum_depth; d += 16) {
+        for (int i = 0; i < 4; i++) {
+          for (int j = 0; j < 16; j++) {
+            int8_t input_val = shuffled_input_data[d + j];
+            int8_t weights_val = *shuffled_weights_ptr++;
+            accum[i] += weights_val * input_val;
+          }
+        }
+      }
+      for (int i = 0; i < 4; i++) {
+        // Add bias value
+        int32_t acc = accum[i] + bias_data[c + i];
+        // Down-scale the final int32_t accumulator to the scale used by our
+        // (16-bit, typically 3 integer bits) fixed-point format. The quantized
+        // multiplier and shift here have been pre-computed offline
+        // (e.g. by toco).
+        acc =
+            MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift);
+        // Saturate, cast to int16_t, and store to output array.
+        acc = std::max(acc, output_activation_min);
+        acc = std::min(acc, output_activation_max);
+        output_ptr[c + i] = acc;
+      }
+    }
+  } else if (batches == 4) {
+    int16_t* output_ptr = output_data;
+    // Shuffled weights have had their sign bit (0x80) pre-flipped (xor'd)
+    // so that just reinterpreting them as int8_t values is equivalent to
+    // subtracting 128 from them, thus implementing for free the subtraction of
+    // the zero_point value 128.
+    const int8_t* shuffled_weights_ptr =
+        reinterpret_cast<const int8_t*>(shuffled_weights_data);
+    // Likewise, we preshuffled and pre-xored the input data above.
+    const int8_t* shuffled_input_data =
+        reinterpret_cast<const int8_t*>(shuffled_input_workspace_data);
+    for (int c = 0; c < output_depth; c += 4) {
+      const int8_t* shuffled_input_ptr = shuffled_input_data;
+      // Accumulation loop.
+      // Internal accumulation.
+      // Initialize accumulator with the bias-value.
+      int32_t accum[4][4];
+      for (int i = 0; i < 4; i++) {
+        for (int b = 0; b < 4; b++) {
+          accum[i][b] = 0;
+        }
+      }
+      for (int d = 0; d < accum_depth; d += 16) {
+        for (int i = 0; i < 4; i++) {
+          for (int b = 0; b < 4; b++) {
+            for (int j = 0; j < 16; j++) {
+              int8_t input_val = shuffled_input_ptr[16 * b + j];
+              int8_t weights_val = shuffled_weights_ptr[16 * i + j];
+              accum[i][b] += weights_val * input_val;
+            }
+          }
+        }
+        shuffled_input_ptr += 64;
+        shuffled_weights_ptr += 64;
+      }
+      for (int i = 0; i < 4; i++) {
+        for (int b = 0; b < 4; b++) {
+          // Add bias value
+          int32_t acc = accum[i][b] + bias_data[c + i];
+          // Down-scale the final int32_t accumulator to the scale used by our
+          // (16-bit, typically 3 integer bits) fixed-point format. The
+          // quantized multiplier and shift here have been pre-computed offline
+          // (e.g. by toco).
+          acc = MultiplyByQuantizedMultiplier(acc, output_multiplier,
+                                              output_shift);
+          // Saturate, cast to int16_t, and store to output array.
+          acc = std::max(acc, output_activation_min);
+          acc = std::min(acc, output_activation_max);
+          output_ptr[b * output_depth + c + i] = acc;
+        }
+      }
+    }
+  } else {
+    TFLITE_DCHECK(false);
+    return;
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FULLY_CONNECTED_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/hard_swish.h

@@ -0,0 +1,166 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ACTIVATIONS_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_ACTIVATIONS_H_
+
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline int16_t SaturatingLeftShift(int16_t value, int amount) {
+  int32_t result = static_cast<int32_t>(value) * (1 << amount);
+  result = std::min<int32_t>(result, std::numeric_limits<int16_t>::max());
+  result = std::max<int32_t>(result, std::numeric_limits<int16_t>::min());
+  return result;
+}
+
+// Similar to ARM instruction SQDMULH.
+// Similar to gemmlowp::SaturatingRoundingDoublingHighMul except
+// rounding to zero instead of to nearest (SQRDMULH).
+inline std::int16_t SaturatingDoublingHighMul(std::int16_t a, std::int16_t b) {
+  bool overflow = a == b && a == std::numeric_limits<std::int16_t>::min();
+  std::int32_t a_32(a);
+  std::int32_t b_32(b);
+  std::int32_t ab_32 = a_32 * b_32;
+  std::int16_t ab_x2_high16 = static_cast<std::int16_t>((ab_32) / (1 << 15));
+  return overflow ? std::numeric_limits<std::int16_t>::max() : ab_x2_high16;
+}
+
+template <typename T>
+inline void HardSwish(const RuntimeShape& input_shape, const T* input_data,
+                      const RuntimeShape& output_shape, T* output_data) {
+  ruy::profiler::ScopeLabel label("ReferenceHardSwish/Float");
+  auto matching_size = MatchingFlatSize(input_shape, output_shape);
+  const T* in_end = input_data + matching_size;
+  for (; input_data < in_end; input_data++, output_data++) {
+    const float in = *input_data;
+    *output_data =
+        in * std::min(static_cast<T>(6), std::max(static_cast<T>(0), in + 3)) /
+        6;
+  }
+}
+
+template <typename T>
+inline void HardSwish(const HardSwishParams& params,
+                      const RuntimeShape& input_shape, const T* input_data,
+                      const RuntimeShape& output_shape, T* output_data) {
+  ruy::profiler::ScopeLabel label("ReferenceHardSwish/Quantized");
+
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; i++) {
+    const int16_t input_value = input_data[i] - params.input_zero_point;
+    // Left-shift as much as we can without overflow/saturation to put
+    // significant bits in the high bits of our 16-bit fixedpoint values, so
+    // that fixed-point approximate computations below are as accurate as
+    // possible.
+    const int16_t input_value_on_hires_input_scale = input_value * (1 << 7);
+    // Compute the input value on essentially the output scale, just not
+    // right-shifted yet. This is the value that we'll use in the (x >= +3)
+    // case, and that in the general case we'll multiply against the "relu-ish"
+    // fixed-point multiplier in [0, 1].
+    const int16_t input_value_on_preshift_output_scale =
+        gemmlowp::SaturatingRoundingDoublingHighMul(
+            input_value_on_hires_input_scale,
+            params.output_multiplier_fixedpoint_int16);
+    // Now compute the "relu-ish multiplier". In the (-3 <= x <= +3) case, that
+    // is just an affine rescaling of x from [-3, 3] to [0, 1]. In the general
+    // case, it is just that plus saturation at the boundaries of [-3, 3].
+    // First, we rescale from [-3, 3] to [-1, 1], saturating.
+    // That is done by rescaling the input value with a fixed-point multiplier
+    // (reluish_multiplier_fixedpoint) and bit-shift such that we represent
+    // that input value on the scale where the real value 3.0f is represented
+    // by the quantized value 32768.  (+32768 is actually not representable as
+    // int16_t, so this saturates at +32767, and that is seen empirically to be
+    // a negligible contribution to numerical error/bias).
+    //
+    // This code is careful to correctly implement any magnitude of multiplier,
+    // involving either a right shift or a left shift, with correct saturation
+    // behavior in the left-shift case. This forces this code to be more
+    // complicated, but is necessary for real applications: a partially
+    // trained quantized MobileNet v3-small model that motivated this code
+    // exhibits some large [min, max] range boundaries, of the order of
+    // magnitude of 10 or 100 depending on layers.
+    //
+    // The next few lines are basically just an ordinary
+    // MultiplyByQuantizedMultiplier, except that we are more careful here
+    // about the fine details of saturation when left-shifting, because here
+    // overflow in left-shift is a common case, not an anomaly as
+    // MultiplyByQuantizedMultiplier assumes.
+    int16_t reluish_value = input_value_on_hires_input_scale;
+    // Shift left, saturating, as much as we can while ensuring that this
+    // saturation will not contribute to the result. That is, left shift amount
+    // reduced by 1.
+    if (params.reluish_multiplier_exponent > 0) {
+      reluish_value = SaturatingLeftShift(
+          reluish_value, params.reluish_multiplier_exponent - 1);
+    }
+    // Apply the fixed-point multiplier, dividing the value by a divisor
+    // ranging in [1, 2].
+    reluish_value = gemmlowp::SaturatingRoundingDoublingHighMul(
+        reluish_value, params.reluish_multiplier_fixedpoint_int16);
+    // Apply the last bit of left-shift. Thus, in the left-shifting case, if
+    // any saturation affects the result, it is happening here --- any
+    // saturation having occurred above is overwritten here, not affecting the
+    // result.
+    if (params.reluish_multiplier_exponent > 0) {
+      reluish_value = SaturatingLeftShift(reluish_value, 1);
+    }
+    // Shift right, in the right-shifting case.
+    if (params.reluish_multiplier_exponent < 0) {
+      reluish_value = gemmlowp::RoundingDivideByPOT(
+          reluish_value, -params.reluish_multiplier_exponent);
+    }
+    // At this point we have rescaled the value into a 16bit fixedpoint
+    // reluish_value in [-1, 1].
+    // We now convert that to a 16bit fixedpoint value in [0, 1].
+    reluish_value = (reluish_value + (1 << 15)) >> 1;
+    // Use of SaturatingDoublingHighMul here is important to cancel the biases
+    // from the above SaturatingRoundingDoublingHighMul.
+    //
+    // On a partially trained MobileNet-v3-small,
+    //
+    //                                       | bias on    |  ImageNet
+    //                                       | quantized  |  Top-1
+    // Operation used here                   | values     |  accuracy (50k)
+    // --------------------------------------+------------+-----------
+    // SaturatingDoublingHighMul             | -0.0024    |  58.920
+    // SaturatingRoundingDoublingHighMul     | -0.0067    |  58.064
+    //
+    // In activations_test, this is covered by this testcase:
+    //     QuantizedActivationsOpTest.HardSwishBias
+    //
+    const int16_t preshift_output_value = SaturatingDoublingHighMul(
+        reluish_value, input_value_on_preshift_output_scale);
+    // We were so far operating on the pre-shift output scale. Now we finally
+    // apply that output shift, arriving at the final output scale.
+    int16_t output_value = gemmlowp::RoundingDivideByPOT(
+        preshift_output_value, -params.output_multiplier_exponent);
+    output_value += params.output_zero_point;
+    output_value =
+        std::min<int16_t>(output_value, std::numeric_limits<T>::max());
+    output_value =
+        std::max<int16_t>(output_value, std::numeric_limits<T>::min());
+    output_data[i] = output_value;
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_CONV_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/integer_ops/add.h

@@ -0,0 +1,144 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_ADD_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_ADD_H_
+
+#include <limits>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+inline void CheckArithmeticParams(const ArithmeticParams& params) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  // Input offset is negative input zero point. Activation tensors are
+  // asymmetric quantized so they span the full int8 range.
+  TFLITE_DCHECK_GE(-params.input1_offset, std::numeric_limits<int8_t>::min());
+  TFLITE_DCHECK_GE(-params.input2_offset, std::numeric_limits<int8_t>::min());
+  TFLITE_DCHECK_LE(-params.input1_offset, std::numeric_limits<int8_t>::max());
+  TFLITE_DCHECK_LE(-params.input2_offset, std::numeric_limits<int8_t>::max());
+}
+
+inline void ElementWise(
+    int size, const ArithmeticParams& params, const int8_t* input1_data,
+    const int8_t* input2_data, int8_t* output_data,
+    void (*check_arithmetic_params)(const ArithmeticParams&),
+    int8_t (*binary_func)(int8_t, int8_t, const ArithmeticParams&)) {
+  CheckArithmeticParams(params);
+  for (int i = 0; i < size; ++i) {
+    output_data[i] = binary_func(input1_data[i], input2_data[i], params);
+  }
+}
+
+inline void BroadcastBinaryFunction4DSlow(
+    const ArithmeticParams& params, const RuntimeShape& input1_shape,
+    const int8_t* input1_data, const RuntimeShape& input2_shape,
+    const int8_t* input2_data, const RuntimeShape& output_shape,
+    int8_t* output_data,
+    void (*check_arithmetic_params)(const ArithmeticParams&),
+    int8_t (*binary_func)(int8_t, int8_t, const ArithmeticParams&)) {
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(4, output_shape);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
+    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
+      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
+        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
+          output_data[Offset(extended_output_shape, b, y, x, c)] = binary_func(
+              input1_data[SubscriptToIndex(desc1, b, y, x, c)],
+              input2_data[SubscriptToIndex(desc2, b, y, x, c)], params);
+        }
+      }
+    }
+  }
+}
+
+inline int8_t AddFunc(int8_t x, int8_t y, const ArithmeticParams& params) {
+  const int32_t input1_val = params.input1_offset + x;
+  const int32_t input2_val = params.input2_offset + y;
+  const int32_t shifted_input1_val = input1_val * (1 << params.left_shift);
+  const int32_t shifted_input2_val = input2_val * (1 << params.left_shift);
+  const int32_t scaled_input1_val =
+      MultiplyByQuantizedMultiplierSmallerThanOneExp(
+          shifted_input1_val, params.input1_multiplier, params.input1_shift);
+  const int32_t scaled_input2_val =
+      MultiplyByQuantizedMultiplierSmallerThanOneExp(
+          shifted_input2_val, params.input2_multiplier, params.input2_shift);
+  const int32_t raw_sum = scaled_input1_val + scaled_input2_val;
+  const int32_t raw_output =
+      MultiplyByQuantizedMultiplierSmallerThanOneExp(
+          raw_sum, params.output_multiplier, params.output_shift) +
+      params.output_offset;
+  const int32_t clamped_output =
+      std::min(params.quantized_activation_max,
+               std::max(params.quantized_activation_min, raw_output));
+  return static_cast<int8_t>(clamped_output);
+}
+
+// Element-wise add that can often be used for inner loop of broadcast add as
+// well as the non-broadcast add.
+inline void AddElementwise(int size, const ArithmeticParams& params,
+                           const int8_t* input1_data, const int8_t* input2_data,
+                           int8_t* output_data) {
+  ElementWise(size, params, input1_data, input2_data, output_data,
+              CheckArithmeticParams, AddFunc);
+}
+
+inline void Add(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const int8_t* input1_data,
+                const RuntimeShape& input2_shape, const int8_t* input2_data,
+                const RuntimeShape& output_shape, int8_t* output_data) {
+  CheckArithmeticParams(params);
+
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+
+  AddElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+inline void BroadcastAdd4DSlow(const ArithmeticParams& params,
+                               const RuntimeShape& input1_shape,
+                               const int8_t* input1_data,
+                               const RuntimeShape& input2_shape,
+                               const int8_t* input2_data,
+                               const RuntimeShape& output_shape,
+                               int8_t* output_data) {
+  BroadcastBinaryFunction4DSlow(params, input1_shape, input1_data, input2_shape,
+                                input2_data, output_shape, output_data,
+                                CheckArithmeticParams, AddFunc);
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_ADD_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/integer_ops/conv.h

@@ -0,0 +1,221 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_CONV_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_CONV_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+// Fixed-point per-channel-quantization convolution reference kernel.
+inline void ConvPerChannel(
+    const ConvParams& params, const int32_t* output_multiplier,
+    const int32_t* output_shift, const RuntimeShape& input_shape,
+    const int8_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    int8_t* output_data) {
+  // Get parameters.
+  const int32_t input_offset = params.input_offset;  // r = s(q - Z)
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const int32_t output_offset = params.output_offset;
+
+  // Set min and max value of the output.
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+
+  // Consistency check.
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+
+  // Check dimensions of the tensors.
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      const int in_y_origin = (out_y * stride_height) - pad_height;
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        const int in_x_origin = (out_x * stride_width) - pad_width;
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          int32_t acc = 0;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            const int in_y = in_y_origin + dilation_height_factor * filter_y;
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              const int in_x = in_x_origin + dilation_width_factor * filter_x;
+
+              // Zero padding by omitting the areas outside the image.
+              const bool is_point_inside_image =
+                  (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                  (in_y < input_height);
+
+              if (!is_point_inside_image) {
+                continue;
+              }
+
+              for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+                int32_t input_val = input_data[Offset(input_shape, batch, in_y,
+                                                      in_x, in_channel)];
+                int32_t filter_val = filter_data[Offset(
+                    filter_shape, out_channel, filter_y, filter_x, in_channel)];
+                // Accumulate with 32 bits accumulator.
+                // In the nudging process during model quantization, we force
+                // real value of 0.0 be represented by a quantized value. This
+                // guarantees that the input_offset is a int8_t, even though
+                // it is represented using int32_t. int32_t += int8_t *
+                // (int8_t - int8_t) so the highest value we can get from each
+                // accumulation is [-127, 127] * ([-128, 127] -
+                // [-128, 127]), which is [-32512, 32512]. log2(32512)
+                // = 14.98, which means we can accumulate at least 2^16
+                // multiplications without overflow. The accumulator is
+                // applied to a filter so the accumulation logic will hold as
+                // long as the filter size (filter_y * filter_x * in_channel)
+                // does not exceed 2^16, which is the case in all the models
+                // we have seen so far.
+                // TODO(b/174275578): Add a check to make sure the
+                // accumulator depth is smaller than 2^16.
+                acc += filter_val * (input_val + input_offset);
+              }
+            }
+          }
+
+          if (bias_data) {
+            acc += bias_data[out_channel];
+          }
+          acc = MultiplyByQuantizedMultiplier(
+              acc, output_multiplier[out_channel], output_shift[out_channel]);
+          acc += output_offset;
+          acc = std::max(acc, output_activation_min);
+          acc = std::min(acc, output_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              static_cast<int8_t>(acc);
+        }
+      }
+    }
+  }
+}
+
+// Fixed-point per-channel-quantization convolution reference kernel.
+// 16-bit data and 8-bit filter
+inline void ConvPerChannel(
+    const ConvParams& params, const int32_t* output_multiplier,
+    const int32_t* output_shift, const RuntimeShape& input_shape,
+    const int16_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const std::int64_t* bias_data, const RuntimeShape& output_shape,
+    int16_t* output_data) {
+  // Get parameters.
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+
+  // Set min and max value of the output.
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+
+  // Consistency check.
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+
+  // Check dimensions of the tensors.
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      const int in_y_origin = (out_y * stride_height) - pad_height;
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        const int in_x_origin = (out_x * stride_width) - pad_width;
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          std::int64_t acc = 0;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            const int in_y = in_y_origin + dilation_height_factor * filter_y;
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              const int in_x = in_x_origin + dilation_width_factor * filter_x;
+
+              // Zero padding by omitting the areas outside the image.
+              const bool is_point_inside_image =
+                  (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                  (in_y < input_height);
+
+              if (!is_point_inside_image) {
+                continue;
+              }
+
+              for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+                int32_t input_val = input_data[Offset(input_shape, batch, in_y,
+                                                      in_x, in_channel)];
+                int32_t filter_val = filter_data[Offset(
+                    filter_shape, out_channel, filter_y, filter_x, in_channel)];
+                // Accumulate with 64 bits accumulator.
+                // int64_t += int8_t * int16_t so the highest value we can
+                // get from each accumulation is [-127, 127] * ([-32768,
+                // 32767] -
+                // [-32768, 32767]), which is [-8322945, 8322945].
+                // log2(8322945) = 22.99.
+                acc += filter_val * input_val;
+              }
+            }
+          }
+          if (bias_data) {
+            acc += bias_data[out_channel];
+          }
+          int32_t scaled_acc = MultiplyByQuantizedMultiplier(
+              acc, output_multiplier[out_channel], output_shift[out_channel]);
+          scaled_acc = std::max(scaled_acc, output_activation_min);
+          scaled_acc = std::min(scaled_acc, output_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              static_cast<int16_t>(scaled_acc);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_CONV_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/integer_ops/depthwise_conv.h

@@ -0,0 +1,289 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_DEPTHWISE_CONV_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_DEPTHWISE_CONV_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+inline void DepthwiseConvPerChannel(
+    const DepthwiseParams& params, const int32_t* output_multiplier,
+    const int32_t* output_shift, const RuntimeShape& input_shape,
+    const int8_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    int8_t* output_data) {
+  // Get parameters.
+  // TODO(b/141565753): Re-introduce ScopedProfilingLabel on Micro.
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const int depth_multiplier = params.depth_multiplier;
+  const int32_t input_offset = params.input_offset;
+  const int32_t output_offset = params.output_offset;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+
+  // Check dimensions of the tensors.
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int input_depth = input_shape.Dims(3);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier);
+  TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+          for (int m = 0; m < depth_multiplier; ++m) {
+            const int output_channel = m + in_channel * depth_multiplier;
+            const int in_x_origin = (out_x * stride_width) - pad_width;
+            const int in_y_origin = (out_y * stride_height) - pad_height;
+            int32_t acc = 0;
+            for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+              for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+                const int in_x = in_x_origin + dilation_width_factor * filter_x;
+                const int in_y =
+                    in_y_origin + dilation_height_factor * filter_y;
+                // Zero padding by omitting the areas outside the image.
+                const bool is_point_inside_image =
+                    (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                    (in_y < input_height);
+                if (is_point_inside_image) {
+                  int32_t input_val = input_data[Offset(
+                      input_shape, batch, in_y, in_x, in_channel)];
+                  int32_t filter_val = filter_data[Offset(
+                      filter_shape, 0, filter_y, filter_x, output_channel)];
+                  // Accumulate with 32 bits accumulator.
+                  // In the nudging process during model quantization, we force
+                  // real value of 0.0 be represented by a quantized value. This
+                  // guarantees that the input_offset is a int8_t, even though
+                  // it is represented using int32_t. int32_t += int8_t *
+                  // (int8_t - int8_t) so the highest value we can get from each
+                  // accumulation is [-127, 127] * ([-128, 127] -
+                  // [-128, 127]), which is [-32512, 32512]. log2(32512)
+                  // = 14.98, which means we can accumulate at least 2^16
+                  // multiplications without overflow. The accumulator is
+                  // applied to a filter so the accumulation logic will hold as
+                  // long as the filter size (filter_y * filter_x * in_channel)
+                  // does not exceed 2^16, which is the case in all the models
+                  // we have seen so far.
+                  // TODO(b/174275578): Add a check to make sure the
+                  // accumulator depth is smaller than 2^16.
+                  acc += filter_val * (input_val + input_offset);
+                }
+              }
+            }
+            if (bias_data) {
+              acc += bias_data[output_channel];
+            }
+            acc = MultiplyByQuantizedMultiplier(
+                acc, output_multiplier[output_channel],
+                output_shift[output_channel]);
+            acc += output_offset;
+            acc = std::max(acc, output_activation_min);
+            acc = std::min(acc, output_activation_max);
+            output_data[Offset(output_shape, batch, out_y, out_x,
+                               output_channel)] = static_cast<int8_t>(acc);
+          }
+        }
+      }
+    }
+  }
+}
+
+inline void DepthwiseConvPerChannel(
+    const DepthwiseParams& params, const int32_t* output_multiplier,
+    const int32_t* output_shift, const RuntimeShape& input_shape,
+    const int16_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const std::int64_t* bias_data, const RuntimeShape& output_shape,
+    int16_t* output_data) {
+  // Get parameters.
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const int depth_multiplier = params.depth_multiplier;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+
+  // Check dimensions of the tensors.
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int input_depth = input_shape.Dims(3);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier);
+  TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+          for (int m = 0; m < depth_multiplier; ++m) {
+            const int output_channel = m + in_channel * depth_multiplier;
+            const int in_x_origin = (out_x * stride_width) - pad_width;
+            const int in_y_origin = (out_y * stride_height) - pad_height;
+            std::int64_t acc = 0;
+            for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+              for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+                const int in_x = in_x_origin + dilation_width_factor * filter_x;
+                const int in_y =
+                    in_y_origin + dilation_height_factor * filter_y;
+                // Zero padding by omitting the areas outside the image.
+                const bool is_point_inside_image =
+                    (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                    (in_y < input_height);
+                if (is_point_inside_image) {
+                  int32_t input_val = input_data[Offset(
+                      input_shape, batch, in_y, in_x, in_channel)];
+                  int32_t filter_val = filter_data[Offset(
+                      filter_shape, 0, filter_y, filter_x, output_channel)];
+                  // Accumulate with 64 bits accumulator.
+                  // We assume maximum of 2^16 accumulations as with the 8-bit
+                  // case so actually the value in the accumulator should not
+                  // exceed 40 bits
+                  acc += static_cast<int64_t>(filter_val) *
+                         static_cast<int64_t>(input_val);
+                }
+              }
+            }
+            if (bias_data) {
+              acc += bias_data[output_channel];
+            }
+            int32_t scaled_acc = MultiplyByQuantizedMultiplier(
+                acc, output_multiplier[output_channel],
+                output_shift[output_channel]);
+            scaled_acc = std::max(scaled_acc, output_activation_min);
+            scaled_acc = std::min(scaled_acc, output_activation_max);
+            output_data[Offset(output_shape, batch, out_y, out_x,
+                               output_channel)] =
+                static_cast<int16_t>(scaled_acc);
+          }
+        }
+      }
+    }
+  }
+}
+
+inline void DepthwiseConvHybridPerChannel(
+    const DepthwiseParams& params, float* scaling_factors_ptr,
+    const RuntimeShape& input_shape, const int8_t* input_data,
+    const RuntimeShape& filter_shape, const int8_t* filter_data,
+    const RuntimeShape& bias_shape, const float* bias_data,
+    const RuntimeShape& output_shape, float* output_data,
+    const float* per_channel_scale, int32_t* input_offset) {
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int dilation_width_factor = params.dilation_width_factor;
+  const int dilation_height_factor = params.dilation_height_factor;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  const int depth_multiplier = params.depth_multiplier;
+  const float output_activation_min = params.float_activation_min;
+  const float output_activation_max = params.float_activation_max;
+  // Check dimensions of the tensors.
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int output_depth = MatchingDim(filter_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int input_depth = input_shape.Dims(3);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int bias_depth = bias_shape.FlatSize();
+  TFLITE_DCHECK_EQ(output_depth, input_depth * depth_multiplier);
+  TFLITE_DCHECK_EQ(bias_depth, output_depth);
+
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+          for (int m = 0; m < depth_multiplier; ++m) {
+            const int output_channel = m + in_channel * depth_multiplier;
+            const int in_x_origin = (out_x * stride_width) - pad_width;
+            const int in_y_origin = (out_y * stride_height) - pad_height;
+            int32_t acc = 0;
+            for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+              for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+                const int in_x = in_x_origin + dilation_width_factor * filter_x;
+                const int in_y =
+                    in_y_origin + dilation_height_factor * filter_y;
+                // Zero padding by omitting the areas outside the image.
+                const bool is_point_inside_image =
+                    (in_x >= 0) && (in_x < input_width) && (in_y >= 0) &&
+                    (in_y < input_height);
+                if (is_point_inside_image) {
+                  int32_t input_val = input_data[Offset(
+                      input_shape, batch, in_y, in_x, in_channel)];
+                  int32_t filter_val = filter_data[Offset(
+                      filter_shape, 0, filter_y, filter_x, output_channel)];
+                  acc += filter_val * (input_val - input_offset[batch]);
+                }
+              }
+            }
+            float acc_float = static_cast<float>(acc);
+            acc_float *=
+                per_channel_scale[output_channel] * scaling_factors_ptr[batch];
+            if (bias_data && output_channel < bias_depth) {
+              acc_float += bias_data[output_channel];
+            }
+            output_data[Offset(output_shape, batch, out_y, out_x,
+                               output_channel)] =
+                ActivationFunctionWithMinMax(acc_float, output_activation_min,
+                                             output_activation_max);
+          }
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_DEPTHWISE_CONV_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h

@@ -0,0 +1,108 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_FULLY_CONNECTED_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_FULLY_CONNECTED_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+inline void FullyConnected(
+    const FullyConnectedParams& params, const RuntimeShape& input_shape,
+    const int8_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    int8_t* output_data) {
+  const int32_t input_offset = params.input_offset;
+  const int32_t filter_offset = params.weights_offset;
+  const int32_t output_offset = params.output_offset;
+  const int32_t output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_GE(filter_shape.DimensionsCount(), 2);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 2);
+
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  const int filter_dim_count = filter_shape.DimensionsCount();
+  const int batches = output_shape.Dims(0);
+  const int output_depth = output_shape.Dims(1);
+  TFLITE_DCHECK_LE(output_depth, filter_shape.Dims(filter_dim_count - 2));
+  const int accum_depth = filter_shape.Dims(filter_dim_count - 1);
+  for (int b = 0; b < batches; ++b) {
+    for (int out_c = 0; out_c < output_depth; ++out_c) {
+      int32_t acc = 0;
+      for (int d = 0; d < accum_depth; ++d) {
+        int32_t input_val = input_data[b * accum_depth + d];
+        int32_t filter_val = filter_data[out_c * accum_depth + d];
+        acc += (filter_val + filter_offset) * (input_val + input_offset);
+      }
+      if (bias_data) {
+        acc += bias_data[out_c];
+      }
+      acc = MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift);
+      acc += output_offset;
+      acc = std::max(acc, output_activation_min);
+      acc = std::min(acc, output_activation_max);
+      output_data[out_c + output_depth * b] = static_cast<int8_t>(acc);
+    }
+  }
+}
+
+inline void FullyConnected(
+    const FullyConnectedParams& params, const RuntimeShape& input_shape,
+    const int16_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const int64_t* bias_data, const RuntimeShape& output_shape,
+    int16_t* output_data) {
+  const int32_t filter_offset = params.weights_offset;
+  const int32_t output_multiplier = params.output_multiplier;
+  const int output_shift = params.output_shift;
+  const int32_t output_activation_min = params.quantized_activation_min;
+  const int32_t output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_GE(filter_shape.DimensionsCount(), 2);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 2);
+
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+  const int filter_dim_count = filter_shape.DimensionsCount();
+  const int batches = output_shape.Dims(0);
+  const int output_depth = output_shape.Dims(1);
+  TFLITE_DCHECK_LE(output_depth, filter_shape.Dims(filter_dim_count - 2));
+  const int accum_depth = filter_shape.Dims(filter_dim_count - 1);
+  for (int b = 0; b < batches; ++b) {
+    for (int out_c = 0; out_c < output_depth; ++out_c) {
+      int64_t acc = 0;
+      for (int d = 0; d < accum_depth; ++d) {
+        int32_t input_val = input_data[b * accum_depth + d];
+        int32_t filter_val = filter_data[out_c * accum_depth + d];
+        acc += (filter_val + filter_offset) * input_val;
+      }
+      if (bias_data) {
+        acc += bias_data[out_c];
+      }
+      int32_t acc_scaled =
+          MultiplyByQuantizedMultiplier(acc, output_multiplier, output_shift);
+      acc_scaled = std::max(acc_scaled, output_activation_min);
+      acc_scaled = std::min(acc_scaled, output_activation_max);
+      output_data[out_c + output_depth * b] = static_cast<int16_t>(acc_scaled);
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_FULLY_CONNECTED_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/integer_ops/l2normalization.h

@@ -0,0 +1,65 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_L2NORMALIZATION_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_L2NORMALIZATION_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+inline void L2Normalization(int32_t input_zero_point, int32_t outer_size,
+                            int32_t depth, const int8_t* input_data,
+                            int8_t* output_data) {
+  static constexpr int8_t kMinInt8 = std::numeric_limits<int8_t>::min();
+  static constexpr int8_t kMaxInt8 = std::numeric_limits<int8_t>::max();
+  // The output scale must be in sync with Prepare().
+  // Output is in 1/128 scale so the actual output range is nudged from [-1, 1]
+  // to [-1, 127/128].
+  static constexpr int32_t kOutputScale = 7;
+  for (int outer_index = 0; outer_index < outer_size; ++outer_index) {
+    // int32_t = (int8_t - int8_t) ^ 2.
+    // ([-128, 127] - [-128, 127]) ^ 2 = [0, (2^8 - 1)^2] so the accumulator is
+    // safe from overflowing in at least 2^16 steps.
+    int32_t acc = 0;
+    for (int inner_index = 0; inner_index < depth; ++inner_index) {
+      int32_t input =
+          input_data[depth * outer_index + inner_index] - input_zero_point;
+      acc += input * input;
+    }
+    int32_t inv_l2norm_multiplier;
+    int inv_l2norm_shift;
+    GetInvSqrtQuantizedMultiplierExp(acc, kReverseShift, &inv_l2norm_multiplier,
+                                     &inv_l2norm_shift);
+
+    for (int inner_index = 0; inner_index < depth; ++inner_index) {
+      int32_t input =
+          input_data[depth * outer_index + inner_index] - input_zero_point;
+
+      // Rescale and downcast. Rescale is folded into the division.
+      int32_t output_in_q24 = MultiplyByQuantizedMultiplier(
+          input, inv_l2norm_multiplier, inv_l2norm_shift + kOutputScale);
+      output_in_q24 =
+          std::min(static_cast<int32_t>(kMaxInt8),
+                   std::max(static_cast<int32_t>(kMinInt8), output_in_q24));
+      output_data[depth * outer_index + inner_index] =
+          static_cast<int8_t>(output_in_q24);
+    }
+  }
+}
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_L2NORMALIZATION_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/integer_ops/logistic.h

@@ -0,0 +1,106 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_LOGISTIC_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_LOGISTIC_H_
+
+#include <limits>
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+inline void Logistic(int32_t input_zero_point, int32_t input_range_radius,
+                     int32_t input_multiplier, int32_t input_left_shift,
+                     int32_t input_size, const int8_t* input_data,
+                     int8_t* output_data) {
+  // Integer bits must be in sync with Prepare() function.
+  static constexpr int32_t kInputIntegerBits = 4;
+  static constexpr int32_t kOutputIntegerBits = 8;
+  static constexpr int8_t kMinInt8 = std::numeric_limits<int8_t>::min();
+  static constexpr int8_t kMaxInt8 = std::numeric_limits<int8_t>::max();
+  static constexpr int32_t kOutputZeroPoint = -128;
+
+  for (int i = 0; i < input_size; ++i) {
+    const int32_t input =
+        static_cast<int32_t>(input_data[i]) - input_zero_point;
+    if (input <= -input_range_radius) {
+      output_data[i] = kMinInt8;
+    } else if (input >= input_range_radius) {
+      output_data[i] = kMaxInt8;
+    } else {
+      const int32_t input_in_q4 = MultiplyByQuantizedMultiplier(
+          input, input_multiplier, input_left_shift);
+      using FixedPoint4 = gemmlowp::FixedPoint<int32_t, kInputIntegerBits>;
+      const int32_t output_in_q0 =
+          gemmlowp::logistic(FixedPoint4::FromRaw(input_in_q4)).raw();
+
+      // Rescale and downcast.
+      using gemmlowp::RoundingDivideByPOT;
+      int32_t output_in_q23 =
+          RoundingDivideByPOT(output_in_q0, 31 - kOutputIntegerBits);
+      output_in_q23 = std::min(std::max(output_in_q23 + kOutputZeroPoint,
+                                        static_cast<int32_t>(kMinInt8)),
+                               static_cast<int32_t>(kMaxInt8));
+      output_data[i] = static_cast<int8_t>(output_in_q23);
+    }
+  }
+}
+
+inline void Logistic(int32_t input_multiplier, int32_t input_size,
+                     const int16_t* ptr_input_data, int16_t* ptr_output_data) {
+  // We use the LUT for sigmoid and take into account, that
+  // tanh(x) = 2*sigmoid(2*x) - 1
+
+  int32_t input_data_mul = (input_multiplier > 0) ? input_multiplier : 1;
+
+  for (int i = 0; i < input_size; ++i, ptr_input_data++, ptr_output_data++) {
+    int32_t input_data = (*ptr_input_data) * input_data_mul;
+
+    // Scale by 3/4 to expand range [-8,8]->[-10.7,10.7] and
+    // we do interpolation on unsigned values.
+    uint32_t abs_input_data = 3 * abs(input_data);
+
+    // We divide by 2 power of 9, because
+    // we need to divide by 2 in power of 7 for
+    // the input conversion + 1/4 from the scale above.
+    // Define uh as uint32_t type not to make this function overflow.
+    uint32_t uh = abs_input_data >> 9;
+    uint32_t result;
+
+    if (uh >= 255) {
+      // Saturate to maximum.
+      result = 0x7FFF << 10;
+    } else {
+      uint32_t ua = sigmoid_table_uint16[uh];
+      uint32_t ub = sigmoid_table_uint16[uh + 1];
+      uint32_t ut = abs_input_data & 0x1ff;
+      // Interpolation is done using the fractional bit.
+      result = (ua << 9) + ut * (ub - ua);
+    }
+
+    result = (input_data >= 0) ? (result + (1 << 9))
+                               : ((1 << (16 + 9)) - result + (1 << 9) - 1);
+
+    // Back to 16-bit.
+    result >>= 10;
+
+    *ptr_output_data = result;
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_LOGISTIC_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/integer_ops/mean.h

@@ -0,0 +1,77 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MEAN_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MEAN_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+template <typename integer_type>
+inline void Mean(const tflite::MeanParams& op_params, int32_t multiplier,
+                 int32_t shift, const RuntimeShape& unextended_input_shape,
+                 const integer_type* input_data, int32_t input_zero_point,
+                 const RuntimeShape& unextended_output_shape,
+                 integer_type* output_data, int32_t output_zero_point) {
+  // Current implementation only supports dimension equals 4 and simultaneous
+  // reduction over width and height.
+  TFLITE_CHECK_EQ(unextended_input_shape.DimensionsCount(), 4);
+  TFLITE_CHECK_LE(unextended_output_shape.DimensionsCount(), 4);
+  const RuntimeShape input_shape =
+      RuntimeShape::ExtendedShape(4, unextended_input_shape);
+  const RuntimeShape output_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_shape);
+  const int output_batch = output_shape.Dims(0);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int output_depth = output_shape.Dims(3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int num_elements_in_axis = input_width * input_height;
+
+  TFLITE_CHECK_EQ(op_params.axis_count, 2);
+  TFLITE_CHECK((op_params.axis[0] == 1 && op_params.axis[1] == 2) ||
+               (op_params.axis[0] == 2 && op_params.axis[1] == 1));
+  TFLITE_CHECK_EQ(output_height, 1);
+  TFLITE_CHECK_EQ(output_width, 1);
+
+  static constexpr int32_t kMinInt = std::numeric_limits<integer_type>::min();
+  static constexpr int32_t kMaxInt = std::numeric_limits<integer_type>::max();
+
+  for (int out_b = 0; out_b < output_batch; ++out_b) {
+    for (int out_d = 0; out_d < output_depth; ++out_d) {
+      int32_t acc = 0;
+      for (int in_h = 0; in_h < input_height; ++in_h) {
+        for (int in_w = 0; in_w < input_width; ++in_w) {
+          acc += input_data[Offset(input_shape, out_b, in_h, in_w, out_d)] -
+                 input_zero_point;
+        }
+      }
+      acc = MultiplyByQuantizedMultiplier(acc, multiplier, shift);
+      acc = acc > 0 ? (acc + num_elements_in_axis / 2) / num_elements_in_axis
+                    : (acc - num_elements_in_axis / 2) / num_elements_in_axis;
+      acc += output_zero_point;
+      acc = std::min(std::max(acc, kMinInt), kMaxInt);
+      output_data[Offset(output_shape, out_b, 0, 0, out_d)] =
+          static_cast<integer_type>(acc);
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MEAN_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/integer_ops/mul.h

@@ -0,0 +1,131 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MUL_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MUL_H_
+
+#include "fixedpoint/fixedpoint.h"
+#include "ruy/profiler/instrumentation.h"  // from @ruy
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+template <typename T>
+inline void MulElementwise(int size, const ArithmeticParams& params,
+                           const T* input1_data, const T* input2_data,
+                           T* output_data) {
+  for (int i = 0; i < size; ++i) {
+    const int32_t input1_val = params.input1_offset + input1_data[i];
+    const int32_t input2_val = params.input2_offset + input2_data[i];
+    const int32_t unclamped_result =
+        params.output_offset +
+        MultiplyByQuantizedMultiplier(input1_val * input2_val,
+                                      params.output_multiplier,
+                                      params.output_shift);
+    const int32_t clamped_output =
+        std::min(params.quantized_activation_max,
+                 std::max(params.quantized_activation_min, unclamped_result));
+    output_data[i] = static_cast<T>(clamped_output);
+  }
+}
+
+template <typename T>
+inline void Mul(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const T* input1_data,
+                const RuntimeShape& input2_shape, const T* input2_data,
+                const RuntimeShape& output_shape, T* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  ruy::profiler::ScopeLabel label("Mul/8bit");
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+
+  MulElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+// Mul with 16 bit inputs and int8_t outputs.
+inline void Mul(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const int16_t* input1_data,
+                const RuntimeShape& input2_shape, const int16_t* input2_data,
+                const RuntimeShape& output_shape, int8_t* output_data) {
+  ruy::profiler::ScopeLabel label("Mul/Int16Int8");
+  int32_t output_offset = params.output_offset;
+  int32_t output_activation_min = params.quantized_activation_min;
+  int32_t output_activation_max = params.quantized_activation_max;
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+
+  const int flat_size =
+      MatchingElementsSize(input1_shape, input2_shape, output_shape);
+
+  for (int i = 0; i < flat_size; i++) {
+    // F0 uses 0 integer bits, range [-1, 1].
+    using F0 = gemmlowp::FixedPoint<std::int16_t, 0>;
+
+    F0 unclamped_result =
+        F0::FromRaw(input1_data[i]) * F0::FromRaw(input2_data[i]);
+    int16_t rescaled_result =
+        gemmlowp::RoundingDivideByPOT(unclamped_result.raw(), 8);
+    int16_t clamped_result = std::min<int16_t>(
+        output_activation_max - output_offset, rescaled_result);
+    clamped_result = std::max<int16_t>(output_activation_min - output_offset,
+                                       clamped_result);
+    output_data[i] = output_offset + clamped_result;
+  }
+}
+
+template <typename T>
+inline void BroadcastMul4DSlow(
+    const ArithmeticParams& params, const RuntimeShape& input1_shape,
+    const T* input1_data, const RuntimeShape& input2_shape,
+    const T* input2_data, const RuntimeShape& output_shape, T* output_data) {
+  ruy::profiler::ScopeLabel label("BroadcastMul4DSlow");
+
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  // The input shapes are extended as part of NdArrayDesc initialization.
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(4, output_shape);
+
+  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
+    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
+      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
+        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
+          const int32_t input1_val =
+              params.input1_offset +
+              input1_data[SubscriptToIndex(desc1, b, y, x, c)];
+          const int32_t input2_val =
+              params.input2_offset +
+              input2_data[SubscriptToIndex(desc2, b, y, x, c)];
+          const int32_t unclamped_result =
+              params.output_offset +
+              MultiplyByQuantizedMultiplier(input1_val * input2_val,
+                                            params.output_multiplier,
+                                            params.output_shift);
+          const int32_t clamped_output = std::min(
+              params.quantized_activation_max,
+              std::max(params.quantized_activation_min, unclamped_result));
+          output_data[Offset(extended_output_shape, b, y, x, c)] =
+              static_cast<T>(clamped_output);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_MUL_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/integer_ops/pooling.h

@@ -0,0 +1,258 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_POOLING_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_POOLING_H_
+
+#include <limits>
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+inline void AveragePool(const PoolParams& params,
+                        const RuntimeShape& input_shape,
+                        const int8_t* input_data,
+                        const RuntimeShape& output_shape, int8_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          int32_t acc = 0;
+          int filter_count = 0;
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              acc +=
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)];
+              filter_count++;
+            }
+          }
+          // Round to the closest integer value.
+          acc = acc > 0 ? (acc + filter_count / 2) / filter_count
+                        : (acc - filter_count / 2) / filter_count;
+          acc = std::max(acc, params.quantized_activation_min);
+          acc = std::min(acc, params.quantized_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              static_cast<int8_t>(acc);
+        }
+      }
+    }
+  }
+}
+
+inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape,
+                    const int8_t* input_data, const RuntimeShape& output_shape,
+                    int8_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  TFLITE_DCHECK_GE(params.quantized_activation_min,
+                   std::numeric_limits<int8_t>::min());
+  TFLITE_DCHECK_LE(params.quantized_activation_max,
+                   std::numeric_limits<int8_t>::max());
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          int8_t max = std::numeric_limits<int8_t>::lowest();
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              max = std::max(
+                  max,
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)]);
+            }
+          }
+          max = std::max<int8_t>(max, params.quantized_activation_min);
+          max = std::min<int8_t>(max, params.quantized_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              static_cast<int8_t>(max);
+        }
+      }
+    }
+  }
+}
+
+inline void AveragePool(const PoolParams& params,
+                        const RuntimeShape& input_shape,
+                        const int16_t* input_data,
+                        const RuntimeShape& output_shape,
+                        int16_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          int32_t acc = 0;
+          int filter_count = 0;
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              acc +=
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)];
+              filter_count++;
+            }
+          }
+          // Round to the closest integer value.
+          acc = acc > 0 ? (acc + filter_count / 2) / filter_count
+                        : (acc - filter_count / 2) / filter_count;
+          acc = std::max(acc, params.quantized_activation_min);
+          acc = std::min(acc, params.quantized_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              static_cast<int16_t>(acc);
+        }
+      }
+    }
+  }
+}
+
+inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape,
+                    const int16_t* input_data, const RuntimeShape& output_shape,
+                    int16_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  TFLITE_DCHECK_GE(params.quantized_activation_min,
+                   std::numeric_limits<int16_t>::min());
+  TFLITE_DCHECK_LE(params.quantized_activation_max,
+                   std::numeric_limits<int16_t>::max());
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          int16_t max = std::numeric_limits<int16_t>::lowest();
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              max = std::max(
+                  max,
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)]);
+            }
+          }
+          max = std::max<int16_t>(max, params.quantized_activation_min);
+          max = std::min<int16_t>(max, params.quantized_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              static_cast<int16_t>(max);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_POOLING_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/integer_ops/tanh.h

@@ -0,0 +1,110 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TANH_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TANH_H_
+
+#include <limits>
+
+#include "fixedpoint/fixedpoint.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+inline void Tanh(int32_t input_zero_point, int32_t input_range_radius,
+                 int32_t input_multiplier, int32_t input_shift,
+                 const RuntimeShape& input_shape, const int8_t* input_data,
+                 const RuntimeShape& output_shape, int8_t* output_data) {
+  // Integer bits must be in sync with Prepare() function.
+  static constexpr int32_t kInputIntegerBits = 4;
+  static constexpr int32_t kOutputScale = 7;
+  static constexpr int32_t kMinInt8 = std::numeric_limits<int8_t>::min();
+  static constexpr int32_t kMaxInt8 = std::numeric_limits<int8_t>::max();
+  using F4 = gemmlowp::FixedPoint<int32_t, kInputIntegerBits>;
+
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; ++i) {
+    const int32_t input =
+        static_cast<int32_t>(input_data[i]) - input_zero_point;
+    if (input <= -input_range_radius) {
+      output_data[i] = kMinInt8;
+    } else if (input >= input_range_radius) {
+      output_data[i] = kMaxInt8;
+    } else {
+      const int32_t input_in_q4 =
+          MultiplyByQuantizedMultiplier(input, input_multiplier, input_shift);
+      const int32_t output_in_q0 =
+          gemmlowp::tanh(F4::FromRaw(input_in_q4)).raw();
+
+      // Rescale and downcast.
+      using gemmlowp::RoundingDivideByPOT;
+      int32_t output_in_q24 =
+          RoundingDivideByPOT(output_in_q0, 31 - kOutputScale);
+      output_in_q24 = std::min(std::max(output_in_q24, kMinInt8), kMaxInt8);
+      output_data[i] = static_cast<int8_t>(output_in_q24);
+    }
+  }
+}
+
+inline void Tanh(int32_t input_multiplier, int32_t input_left_shift,
+                 const RuntimeShape& input_shape, const int16_t* ptr_input_data,
+                 const RuntimeShape& output_shape, int16_t* ptr_output_data) {
+  // We use the LUT for sigmoid and take into account, that
+  // tanh(x) = 2*sigmoid(2*x) - 1
+
+  int32_t input_data_mul = (input_multiplier > 0) ? input_multiplier : 1;
+
+  int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; ++i, ptr_input_data++, ptr_output_data++) {
+    int32_t input_data = (*ptr_input_data) * input_data_mul;
+
+    if (input_left_shift == 1) {
+      input_data <<= 1;
+    }
+
+    // Scale by 3/4 to expand range [-8,8]->[-10.7,10.7].
+    uint32_t abs_input_data = 3 * abs(input_data);
+    uint32_t uh = abs_input_data >> 8;
+    int32_t result;
+
+    if (uh >= 255) {
+      // Saturate to maximum.
+      result = 0xFFFF << 8;
+    } else {
+      uint32_t ua = sigmoid_table_uint16[uh];
+      uint32_t ub = sigmoid_table_uint16[uh + 1];
+
+      uint8_t ut = abs_input_data & 0xFF;
+
+      result = (ua << 8) + ut * (ub - ua);
+    }
+
+    result = (input_data >= 0)
+                 ? (result - (1 << (14 + 9)) + (1 << (9 - 2)))
+                 : (-result + (1 << (14 + 9)) + (1 << (9 - 2)) - 1);
+
+    // Convert back to 16-bit.
+    result >>= (9 - 1);
+
+    *ptr_output_data = result;
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TANH_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/integer_ops/transpose_conv.h

@@ -0,0 +1,221 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TRANSPOSE_CONV_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TRANSPOSE_CONV_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+namespace reference_integer_ops {
+
+// Fixed-point per-channel-quantization transpose convolution reference kernel.
+inline void TransposeConv(
+    const ConvParams& params, const int32_t* output_multiplier,
+    const int32_t* output_shift, const RuntimeShape& input_shape,
+    const int8_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const int32_t* bias_data, const RuntimeShape& output_shape,
+    int8_t* output_data, const RuntimeShape& im2col_shape, int8_t* im2col_data,
+    int32_t* scratch_buffer) {
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  (void)im2col_data;   // only used in optimized code.
+  (void)im2col_shape;  // only used in optimized code.
+
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int32_t input_offset = params.input_offset;
+  const int32_t output_offset = params.output_offset;
+  const int32_t output_activation_min = std::numeric_limits<int8_t>::min();
+  const int32_t output_activation_max = std::numeric_limits<int8_t>::max();
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+
+  const int num_elements = output_shape.FlatSize();
+  // We need to initialize scratch_buffer to all 0s, as we apply the same
+  // 'scatter' based trick as in float version.
+  memset(scratch_buffer, 0, num_elements * sizeof(int32_t));
+
+  // Loop through input elements one at a time.
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int in_y = 0; in_y < input_height; ++in_y) {
+      for (int in_x = 0; in_x < input_width; ++in_x) {
+        for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+          // Loop through the output elements it will influence.
+          const int out_x_origin = (in_x * stride_width) - pad_width;
+          const int out_y_origin = (in_y * stride_height) - pad_height;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              for (int out_channel = 0; out_channel < output_depth;
+                   ++out_channel) {
+                // Compute output element location.
+                const int out_x = out_x_origin + filter_x;
+                const int out_y = out_y_origin + filter_y;
+                // We cannot accumulate out of bounds.
+                if ((out_x >= 0) && (out_x < output_width) && (out_y >= 0) &&
+                    (out_y < output_height)) {
+                  const int8_t input_value = input_data[Offset(
+                      input_shape, batch, in_y, in_x, in_channel)];
+                  const int8_t filter_value =
+                      filter_data[Offset(filter_shape, out_channel, filter_y,
+                                         filter_x, in_channel)];
+                  scratch_buffer[Offset(output_shape, batch, out_y, out_x,
+                                        out_channel)] +=
+                      (input_value + input_offset) * filter_value;
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          int32_t acc = scratch_buffer[Offset(output_shape, batch, out_y, out_x,
+                                              out_channel)];
+          if (bias_data) {
+            acc += bias_data[out_channel];
+          }
+          acc = MultiplyByQuantizedMultiplier(
+              acc, output_multiplier[out_channel], output_shift[out_channel]);
+          acc += output_offset;
+          acc = std::max(acc, output_activation_min);
+          acc = std::min(acc, output_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              static_cast<int8_t>(acc);
+        }
+      }
+    }
+  }
+}
+
+// int16_t input (zero_point=0), int8_t filter, int64 accumulator
+inline void TransposeConv(
+    const ConvParams& params, const int32_t* output_multiplier,
+    const int32_t* output_shift, const RuntimeShape& input_shape,
+    const int16_t* input_data, const RuntimeShape& filter_shape,
+    const int8_t* filter_data, const RuntimeShape& bias_shape,
+    const std::int64_t* bias_data, const RuntimeShape& output_shape,
+    int16_t* output_data, const RuntimeShape& im2col_shape, int8_t* im2col_data,
+    std::int64_t* scratch_buffer) {
+  const int stride_width = params.stride_width;
+  const int stride_height = params.stride_height;
+  const int pad_width = params.padding_values.width;
+  const int pad_height = params.padding_values.height;
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(filter_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  (void)im2col_data;   // only used in optimized code.
+  (void)im2col_shape;  // only used in optimized code.
+
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int input_depth = MatchingDim(input_shape, 3, filter_shape, 3);
+  const int output_depth = MatchingDim(filter_shape, 0, output_shape, 3);
+  if (bias_data) {
+    TFLITE_DCHECK_EQ(bias_shape.FlatSize(), output_depth);
+  }
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int filter_height = filter_shape.Dims(1);
+  const int filter_width = filter_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int32_t output_activation_min = std::numeric_limits<int16_t>::min();
+  const int32_t output_activation_max = std::numeric_limits<int16_t>::max();
+  TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
+
+  const int num_elements = output_shape.FlatSize();
+  // We need to initialize scratch_buffer to all 0s, as we apply the same
+  // 'scatter' based trick as in float version.
+  memset(scratch_buffer, 0, num_elements * sizeof(std::int64_t));
+
+  // Loop through input elements one at a time.
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int in_y = 0; in_y < input_height; ++in_y) {
+      for (int in_x = 0; in_x < input_width; ++in_x) {
+        for (int in_channel = 0; in_channel < input_depth; ++in_channel) {
+          // Loop through the output elements it will influence.
+          const int out_x_origin = (in_x * stride_width) - pad_width;
+          const int out_y_origin = (in_y * stride_height) - pad_height;
+          for (int filter_y = 0; filter_y < filter_height; ++filter_y) {
+            for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
+              for (int out_channel = 0; out_channel < output_depth;
+                   ++out_channel) {
+                // Compute output element location.
+                const int out_x = out_x_origin + filter_x;
+                const int out_y = out_y_origin + filter_y;
+                // We cannot accumulate out of bounds.
+                if ((out_x >= 0) && (out_x < output_width) && (out_y >= 0) &&
+                    (out_y < output_height)) {
+                  const int32_t input_value = input_data[Offset(
+                      input_shape, batch, in_y, in_x, in_channel)];
+                  const int32_t filter_value =
+                      filter_data[Offset(filter_shape, out_channel, filter_y,
+                                         filter_x, in_channel)];
+                  scratch_buffer[Offset(output_shape, batch, out_y, out_x,
+                                        out_channel)] +=
+                      input_value * filter_value;
+                }
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int out_channel = 0; out_channel < output_depth; ++out_channel) {
+          std::int64_t acc = scratch_buffer[Offset(output_shape, batch, out_y,
+                                                   out_x, out_channel)];
+          if (bias_data) {
+            acc += bias_data[out_channel];
+          }
+          int32_t scaled_acc = MultiplyByQuantizedMultiplier(
+              acc, output_multiplier[out_channel], output_shift[out_channel]);
+          scaled_acc = std::max(scaled_acc, output_activation_min);
+          scaled_acc = std::min(scaled_acc, output_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, out_channel)] =
+              static_cast<int16_t>(scaled_acc);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_integer_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_INTEGER_OPS_TRANSPOSE_CONV_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/l2normalization.h

@@ -0,0 +1,90 @@
+/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_L2NORMALIZATION_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_L2NORMALIZATION_H_
+
+#include <algorithm>
+#include <cmath>
+
+#include "tensorflow/lite/c/common.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+inline void L2Normalization(const tflite::L2NormalizationParams& op_params,
+                            const RuntimeShape& input_shape,
+                            const float* input_data,
+                            const RuntimeShape& output_shape,
+                            float* output_data, float epsilon = 1e-6) {
+  const int trailing_dim = input_shape.DimensionsCount() - 1;
+  const int outer_size =
+      MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape);
+  const int depth =
+      MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim);
+  for (int i = 0; i < outer_size; ++i) {
+    float squared_l2_norm = 0;
+    for (int c = 0; c < depth; ++c) {
+      const float val = input_data[depth * i + c];
+      squared_l2_norm += val * val;
+    }
+    float l2_norm = std::sqrt(squared_l2_norm);
+    l2_norm = std::max(l2_norm, epsilon);
+    for (int c = 0; c < depth; ++c) {
+      output_data[depth * i + c] = input_data[depth * i + c] / l2_norm;
+    }
+  }
+}
+
+inline void L2Normalization(const tflite::L2NormalizationParams& op_params,
+                            const RuntimeShape& input_shape,
+                            const uint8_t* input_data,
+                            const RuntimeShape& output_shape,
+                            uint8_t* output_data) {
+  const int trailing_dim = input_shape.DimensionsCount() - 1;
+  const int depth =
+      MatchingDim(input_shape, trailing_dim, output_shape, trailing_dim);
+  const int outer_size =
+      MatchingFlatSizeSkipDim(input_shape, trailing_dim, output_shape);
+  const int32_t input_zero_point = op_params.input_zero_point;
+
+  for (int i = 0; i < outer_size; ++i) {
+    int32_t square_l2_norm = 0;
+    for (int c = 0; c < depth; c++) {
+      int32_t diff = input_data[depth * i + c] - input_zero_point;
+      square_l2_norm += diff * diff;
+    }
+    int32_t inv_l2norm_multiplier;
+    int inv_l2norm_shift;
+    GetInvSqrtQuantizedMultiplierExp(square_l2_norm, kReverseShift,
+                                     &inv_l2norm_multiplier, &inv_l2norm_shift);
+    for (int c = 0; c < depth; c++) {
+      int32_t diff = input_data[depth * i + c] - input_zero_point;
+      int32_t rescaled_diff = MultiplyByQuantizedMultiplierSmallerThanOneExp(
+          128 * diff, inv_l2norm_multiplier, inv_l2norm_shift);
+      int32_t unclamped_output_val = 128 + rescaled_diff;
+      int32_t output_val =
+          std::min(static_cast<int32_t>(255),
+                   std::max(static_cast<int32_t>(0), unclamped_output_val));
+      output_data[depth * i + c] = static_cast<uint8_t>(output_val);
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_L2NORMALIZATION_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/logistic.h

@@ -0,0 +1,132 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LOGISTIC_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LOGISTIC_H_
+
+#include <cmath>
+
+#include "fixedpoint/fixedpoint.h"
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/quantization_util.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+#include "tensorflow/lite/kernels/op_macros.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline void Logistic(const RuntimeShape& input_shape, const float* input_data,
+                     const RuntimeShape& output_shape, float* output_data) {
+  const float cutoff_upper = 16.619047164916992188f;
+  const float cutoff_lower = -9.f;
+
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  // Rational for using approximation in reference kernel.
+  // 0. This approximation gives enough precision for float.
+  // 1. This works around an issue on an embedded chipset where exp() does not
+  // return correctly as expected - exp(x) should return inf when overflown
+  // not 1.701417   IEEE 754 defines representation for inf.
+  // 2. This will speed up calculation and is matching the behavior in the
+  // optimized kernels. (check the definition of scalar_logistic_op<float>)
+
+  for (int i = 0; i < flat_size; i++) {
+    float val = input_data[i];
+    float result;
+    if (val > cutoff_upper) {
+      result = 1.0f;
+    } else if (val < cutoff_lower) {
+      result = std::exp(val);
+    } else {
+      result = 1.f / (1.f + std::exp(-val));
+    }
+    output_data[i] = result;
+  }
+}
+
+// Convenience version that allows, for example, generated-code calls to be
+// uniform between data types.
+inline void Logistic(const LogisticParams&, const RuntimeShape& input_shape,
+                     const float* input_data, const RuntimeShape& output_shape,
+                     float* output_data) {
+  // Drop params: not needed.
+  Logistic(input_shape, input_data, output_shape, output_data);
+}
+
+inline void Logistic(const LogisticParams& params,
+                     const RuntimeShape& input_shape, const int16_t* input_data,
+                     const RuntimeShape& output_shape, int16_t* output_data) {
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; i++) {
+    // F0 uses 0 integer bits, range [-1, 1].
+    // This is the return type of math functions such as tanh, logistic,
+    // whose range is in [-1, 1].
+    using F0 = gemmlowp::FixedPoint<std::int16_t, 0>;
+    // F3 uses 3 integer bits, range [-8, 8], the input range expected here.
+    using F3 = gemmlowp::FixedPoint<std::int16_t, 3>;
+
+    const F3 input = F3::FromRaw(input_data[i]);
+    F0 output = gemmlowp::logistic(input);
+    output_data[i] = output.raw();
+  }
+}
+
+// Quantized int8_t logistic activation.  Cheats by dequantizing and
+// requantizing around the floating point logistic method.  This implementation
+// is slow on platforms without a floating point unit.
+
+// TODO(b/141211002): Delete this int8_t implementation once we can reuse the
+// approach used in TFLite for int8_t Logistic.
+inline void Logistic(const RuntimeShape& input_shape, const int8_t* input_data,
+                     float input_scale, int input_zero_point,
+                     const RuntimeShape& output_shape, int8_t* output_data,
+                     float output_scale, int output_zero_point) {
+  const float cutoff_upper = 16.619047164916992188f;
+  const float cutoff_lower = -9.f;
+
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  // Rational for using approximation in reference kernel.
+  // 0. This approximation gives enough precision for float.
+  // 1. This works around an issue on an embedded chipset where exp() does not
+  // return correctly as expected - exp(x) should return inf when overflown
+  // not 1.701417   IEEE 754 defines representation for inf.
+  // 2. This will speed up calculation and is matching the behavior in the
+  // optimized kernels. (check the definition of scalar_logistic_op<float>)
+
+  for (int i = 0; i < flat_size; i++) {
+    // Dequantize.
+    float val =
+        static_cast<float>((input_data[i] - input_zero_point) * input_scale);
+    float result;
+    if (val > cutoff_upper) {
+      result = 1.0f;
+    } else if (val < cutoff_lower) {
+      result = std::exp(val);
+    } else {
+      result = 1.f / (1.f + std::exp(-val));
+    }
+    // Requantize
+    int8_t output =
+        static_cast<int8_t>(result / output_scale + output_zero_point);
+    output_data[i] = output;
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_LOGISTIC_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/maximum_minimum.h

@@ -0,0 +1,64 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MAXIMUM_MINIMUM_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MAXIMUM_MINIMUM_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+template <typename T, typename Op, int N = 5>
+void MaximumMinimumBroadcastSlow(const RuntimeShape& unextended_input1_shape,
+                                 const T* input1_data,
+                                 const RuntimeShape& unextended_input2_shape,
+                                 const T* input2_data,
+                                 const RuntimeShape& unextended_output_shape,
+                                 T* output_data, Op op) {
+  // Uses element-wise calculation if broadcast is not required.
+  if (unextended_input1_shape == unextended_input2_shape) {
+    const int flat_size =
+        MatchingElementsSize(unextended_input1_shape, unextended_input2_shape,
+                             unextended_output_shape);
+    for (int i = 0; i < flat_size; ++i) {
+      output_data[i] = op(input1_data[i], input2_data[i]);
+    }
+  } else {
+    TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), N);
+    TFLITE_DCHECK_LE(unextended_input2_shape.DimensionsCount(), N);
+    TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), N);
+
+    NdArrayDesc<N> desc1;
+    NdArrayDesc<N> desc2;
+    NdArrayDesc<N> output_desc;
+    NdArrayDescsForElementwiseBroadcast(
+        unextended_input1_shape, unextended_input2_shape, &desc1, &desc2);
+    CopyDimsToDesc(RuntimeShape::ExtendedShape(N, unextended_output_shape),
+                   &output_desc);
+
+    auto maxmin_func = [&](int indexes[N]) {
+      output_data[SubscriptToIndex(output_desc, indexes)] =
+          op(input1_data[SubscriptToIndex(desc1, indexes)],
+             input2_data[SubscriptToIndex(desc2, indexes)]);
+    };
+    NDOpsHelper<N>(output_desc, maxmin_func);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MAXIMUM_MINIMUM_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/mul.h

@@ -0,0 +1,166 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MUL_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MUL_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+// Element-wise mul that can often be used for inner loop of broadcast Mul as
+// well as the non-broadcast Mul.
+inline void MulElementwise(int size, const ArithmeticParams& params,
+                           const uint8_t* input1_data,
+                           const uint8_t* input2_data, uint8_t* output_data) {
+  for (int i = 0; i < size; ++i) {
+    const int32_t input1_val = params.input1_offset + input1_data[i];
+    const int32_t input2_val = params.input2_offset + input2_data[i];
+    const int32_t unclamped_result =
+        params.output_offset +
+        MultiplyByQuantizedMultiplier(input1_val * input2_val,
+                                      params.output_multiplier,
+                                      params.output_shift);
+    const int32_t clamped_output =
+        std::min(params.quantized_activation_max,
+                 std::max(params.quantized_activation_min, unclamped_result));
+    output_data[i] = static_cast<uint8_t>(clamped_output);
+  }
+}
+
+template <typename T>
+inline void Mul(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const T* input1_data,
+                const RuntimeShape& input2_shape, const T* input2_data,
+                const RuntimeShape& output_shape, T* output_data) {
+  T output_activation_min;
+  T output_activation_max;
+  GetActivationParams(params, &output_activation_min, &output_activation_max);
+
+  const int flat_size =
+      MatchingFlatSize(input1_shape, input2_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    output_data[i] = ActivationFunctionWithMinMax(
+        input1_data[i] * input2_data[i], output_activation_min,
+        output_activation_max);
+  }
+}
+
+inline void Mul(const ArithmeticParams& params,
+                const RuntimeShape& input1_shape, const uint8_t* input1_data,
+                const RuntimeShape& input2_shape, const uint8_t* input2_data,
+                const RuntimeShape& output_shape, uint8_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  const int flat_size =
+      MatchingFlatSize(input1_shape, input2_shape, output_shape);
+
+  MulElementwise(flat_size, params, input1_data, input2_data, output_data);
+}
+
+inline void BroadcastMul4DSlow(const ArithmeticParams& params,
+                               const RuntimeShape& input1_shape,
+                               const uint8_t* input1_data,
+                               const RuntimeShape& input2_shape,
+                               const uint8_t* input2_data,
+                               const RuntimeShape& output_shape,
+                               uint8_t* output_data) {
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input1_shape, input2_shape, &desc1,
+                                      &desc2);
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(4, output_shape);
+
+  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
+    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
+      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
+        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
+          const int32_t input1_val =
+              params.input1_offset +
+              input1_data[SubscriptToIndex(desc1, b, y, x, c)];
+          const int32_t input2_val =
+              params.input2_offset +
+              input2_data[SubscriptToIndex(desc2, b, y, x, c)];
+          const int32_t unclamped_result =
+              params.output_offset +
+              MultiplyByQuantizedMultiplier(input1_val * input2_val,
+                                            params.output_multiplier,
+                                            params.output_shift);
+          const int32_t clamped_output = std::min(
+              params.quantized_activation_max,
+              std::max(params.quantized_activation_min, unclamped_result));
+          output_data[Offset(extended_output_shape, b, y, x, c)] =
+              static_cast<uint8_t>(clamped_output);
+        }
+      }
+    }
+  }
+}
+
+template <typename T>
+void BroadcastMul4DSlow(const ArithmeticParams& params,
+                        const RuntimeShape& unextended_input1_shape,
+                        const T* input1_data,
+                        const RuntimeShape& unextended_input2_shape,
+                        const T* input2_data,
+                        const RuntimeShape& unextended_output_shape,
+                        T* output_data) {
+  T output_activation_min;
+  T output_activation_max;
+  GetActivationParams(params, &output_activation_min, &output_activation_max);
+
+  TFLITE_DCHECK_LE(unextended_input1_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_input2_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(unextended_output_shape.DimensionsCount(), 4);
+  const RuntimeShape output_shape =
+      RuntimeShape::ExtendedShape(4, unextended_output_shape);
+
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(unextended_input1_shape,
+                                      unextended_input2_shape, &desc1, &desc2);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  for (int b = 0; b < output_shape.Dims(0); ++b) {
+    for (int y = 0; y < output_shape.Dims(1); ++y) {
+      for (int x = 0; x < output_shape.Dims(2); ++x) {
+        for (int c = 0; c < output_shape.Dims(3); ++c) {
+          output_data[Offset(output_shape, b, y, x, c)] =
+              ActivationFunctionWithMinMax(
+                  input1_data[SubscriptToIndex(desc1, b, y, x, c)] *
+                      input2_data[SubscriptToIndex(desc2, b, y, x, c)],
+                  output_activation_min, output_activation_max);
+        }
+      }
+    }
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_MUL_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/neg.h

@@ -0,0 +1,37 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_NEG_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_NEG_H_
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename T>
+inline void Negate(const RuntimeShape& input_shape, const T* input_data,
+                   const RuntimeShape& output_shape, T* output_data) {
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+
+  for (int i = 0; i < flat_size; ++i) {
+    output_data[i] = -input_data[i];
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_NEG_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/pad.h

@@ -0,0 +1,162 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PAD_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PAD_H_
+
+#include <vector>
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+// TFLite Pad supports activation tensors with up to 4 dimensions.
+constexpr int PadKernelMaxDimensionCount() { return 4; }
+
+// There are two versions of pad: Pad and PadV2.  In PadV2 there is a second
+// scalar input that provides the padding value.  Therefore pad_value_ptr can be
+// equivalent to a simple input1_data.  For Pad, it should point to a zero
+// value.
+//
+// Note that two typenames are required, so that T=P=int32_t is considered a
+// specialization distinct from P=int32_t.
+template <typename T, typename P>
+inline void PadImpl(const tflite::PadParams& op_params,
+                    const RuntimeShape& input_shape, const T* input_data,
+                    const P* pad_value_ptr, const RuntimeShape& output_shape,
+                    T* output_data) {
+  const RuntimeShape ext_input_shape =
+      RuntimeShape::ExtendedShape(PadKernelMaxDimensionCount(), input_shape);
+  const RuntimeShape ext_output_shape =
+      RuntimeShape::ExtendedShape(PadKernelMaxDimensionCount(), output_shape);
+  TFLITE_DCHECK_LE(op_params.left_padding_count, PadKernelMaxDimensionCount());
+  TFLITE_DCHECK_LE(op_params.right_padding_count, PadKernelMaxDimensionCount());
+
+  // Runtime calls are currently fixed at 4 dimensions. Copy inputs so we can
+  // pad them to 4 dims (yes, we are "padding the padding").
+  int left_padding_copy[PadKernelMaxDimensionCount()];
+  for (int i = 0; i < PadKernelMaxDimensionCount(); i++) {
+    left_padding_copy[i] = 0;
+  }
+  for (int i = 0; i < op_params.left_padding_count; ++i) {
+    left_padding_copy[i + PadKernelMaxDimensionCount() -
+                      op_params.left_padding_count] = op_params.left_padding[i];
+  }
+  int right_padding_copy[PadKernelMaxDimensionCount()];
+  for (int i = 0; i < PadKernelMaxDimensionCount(); i++) {
+    right_padding_copy[i] = 0;
+  }
+  for (int i = 0; i < op_params.right_padding_count; ++i) {
+    right_padding_copy[i + PadKernelMaxDimensionCount() -
+                       op_params.right_padding_count] =
+        op_params.right_padding[i];
+  }
+
+  const int output_batch = ext_output_shape.Dims(0);
+  const int output_height = ext_output_shape.Dims(1);
+  const int output_width = ext_output_shape.Dims(2);
+  const int output_depth = ext_output_shape.Dims(3);
+
+  const int left_b_padding = left_padding_copy[0];
+  const int left_h_padding = left_padding_copy[1];
+  const int left_w_padding = left_padding_copy[2];
+  const int left_d_padding = left_padding_copy[3];
+
+  const int right_b_padding = right_padding_copy[0];
+  const int right_h_padding = right_padding_copy[1];
+  const int right_w_padding = right_padding_copy[2];
+  const int right_d_padding = right_padding_copy[3];
+
+  const T pad_value = *pad_value_ptr;
+
+  const T* in_ptr = input_data;
+  T* out_ptr = output_data;
+  for (int out_b = 0; out_b < output_batch; ++out_b) {
+    for (int out_h = 0; out_h < output_height; ++out_h) {
+      for (int out_w = 0; out_w < output_width; ++out_w) {
+        for (int out_d = 0; out_d < output_depth; ++out_d) {
+          if (out_b < left_b_padding ||
+              out_b >= output_batch - right_b_padding ||
+              out_h < left_h_padding ||
+              out_h >= output_height - right_h_padding ||
+              out_w < left_w_padding ||
+              out_w >= output_width - right_w_padding ||
+              out_d < left_d_padding ||
+              out_d >= output_depth - right_d_padding) {
+            *out_ptr++ = pad_value;
+          } else {
+            *out_ptr++ = *in_ptr++;
+          }
+        }
+      }
+    }
+  }
+}
+
+template <typename T, typename P>
+inline void Pad(const tflite::PadParams& op_params,
+                const RuntimeShape& input_shape, const T* input_data,
+                const P* pad_value_ptr, const RuntimeShape& output_shape,
+                T* output_data) {
+  PadImpl(op_params, input_shape, input_data, pad_value_ptr, output_shape,
+          output_data);
+}
+
+// The second (pad-value) input can be int32_t when, say, the first is uint8_t.
+template <typename T>
+inline void Pad(const tflite::PadParams& op_params,
+                const RuntimeShape& input_shape, const T* input_data,
+                const int32_t* pad_value_ptr, const RuntimeShape& output_shape,
+                T* output_data) {
+  const T converted_pad_value = static_cast<T>(*pad_value_ptr);
+  PadImpl(op_params, input_shape, input_data, &converted_pad_value,
+          output_shape, output_data);
+}
+
+// This version avoids conflicting template matching.
+template <>
+inline void Pad(const tflite::PadParams& op_params,
+                const RuntimeShape& input_shape, const int32_t* input_data,
+                const int32_t* pad_value_ptr, const RuntimeShape& output_shape,
+                int32_t* output_data) {
+  PadImpl(op_params, input_shape, input_data, pad_value_ptr, output_shape,
+          output_data);
+}
+
+template <typename T, typename P>
+inline void PadImageStyle(const tflite::PadParams& op_params,
+                          const RuntimeShape& input_shape, const T* input_data,
+                          const P* pad_value_ptr,
+                          const RuntimeShape& output_shape, T* output_data) {
+  Pad(op_params, input_shape, input_data, pad_value_ptr, output_shape,
+      output_data);
+}
+
+template <typename P>
+inline void PadImageStyle(const tflite::PadParams& op_params,
+                          const RuntimeShape& input_shape,
+                          const float* input_data, const P* pad_value_ptr,
+                          const RuntimeShape& output_shape,
+                          float* output_data) {
+  Pad(op_params, input_shape, input_data, pad_value_ptr, output_shape,
+      output_data);
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PAD_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/pooling.h

@@ -0,0 +1,297 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_POOLING_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_POOLING_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/quantization_util.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+namespace reference_ops {
+
+inline void AveragePool(const PoolParams& params,
+                        const RuntimeShape& input_shape,
+                        const float* input_data,
+                        const RuntimeShape& output_shape, float* output_data) {
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          float total = 0.f;
+          float filter_count = 0;
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              total +=
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)];
+              filter_count++;
+            }
+          }
+          const float average = total / filter_count;
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              ActivationFunctionWithMinMax(average, params.float_activation_min,
+                                           params.float_activation_max);
+        }
+      }
+    }
+  }
+}
+
+inline void AveragePool(const PoolParams& params,
+                        const RuntimeShape& input_shape,
+                        const uint8_t* input_data,
+                        const RuntimeShape& output_shape,
+                        uint8_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          int32_t acc = 0;
+          int filter_count = 0;
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              acc +=
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)];
+              filter_count++;
+            }
+          }
+          acc = (acc + filter_count / 2) / filter_count;
+          acc = std::max(acc, params.quantized_activation_min);
+          acc = std::min(acc, params.quantized_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              static_cast<uint8_t>(acc);
+        }
+      }
+    }
+  }
+}
+
+inline void L2Pool(const PoolParams& params, const RuntimeShape& input_shape,
+                   const float* input_data, const RuntimeShape& output_shape,
+                   float* output_data) {
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          float sum_squares = 0.f;
+          int filter_count = 0;
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              const float val =
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)];
+              sum_squares += val * val;
+              filter_count++;
+            }
+          }
+          const float l2pool_result = std::sqrt(sum_squares / filter_count);
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              ActivationFunctionWithMinMax(l2pool_result,
+                                           params.float_activation_min,
+                                           params.float_activation_max);
+        }
+      }
+    }
+  }
+}
+
+inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape,
+                    const float* input_data, const RuntimeShape& output_shape,
+                    float* output_data) {
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          float max = std::numeric_limits<float>::lowest();
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              max = std::max(
+                  max,
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)]);
+            }
+          }
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              ActivationFunctionWithMinMax(max, params.float_activation_min,
+                                           params.float_activation_max);
+        }
+      }
+    }
+  }
+}
+
+inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape,
+                    const uint8_t* input_data, const RuntimeShape& output_shape,
+                    uint8_t* output_data) {
+  TFLITE_DCHECK_LE(params.quantized_activation_min,
+                   params.quantized_activation_max);
+  TFLITE_DCHECK_GE(params.quantized_activation_min, 0);
+  TFLITE_DCHECK_LE(params.quantized_activation_max, 255);
+  TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4);
+  const int batches = MatchingDim(input_shape, 0, output_shape, 0);
+  const int depth = MatchingDim(input_shape, 3, output_shape, 3);
+  const int input_height = input_shape.Dims(1);
+  const int input_width = input_shape.Dims(2);
+  const int output_height = output_shape.Dims(1);
+  const int output_width = output_shape.Dims(2);
+  const int stride_height = params.stride_height;
+  const int stride_width = params.stride_width;
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        for (int channel = 0; channel < depth; ++channel) {
+          const int in_x_origin =
+              (out_x * stride_width) - params.padding_values.width;
+          const int in_y_origin =
+              (out_y * stride_height) - params.padding_values.height;
+          // Compute the boundaries of the filter region clamped so as to
+          // ensure that the filter window fits in the input array.
+          const int filter_x_start = std::max(0, -in_x_origin);
+          const int filter_x_end =
+              std::min(params.filter_width, input_width - in_x_origin);
+          const int filter_y_start = std::max(0, -in_y_origin);
+          const int filter_y_end =
+              std::min(params.filter_height, input_height - in_y_origin);
+          uint8_t max = 0;
+          for (int filter_y = filter_y_start; filter_y < filter_y_end;
+               ++filter_y) {
+            for (int filter_x = filter_x_start; filter_x < filter_x_end;
+                 ++filter_x) {
+              const int in_x = in_x_origin + filter_x;
+              const int in_y = in_y_origin + filter_y;
+              max = std::max(
+                  max,
+                  input_data[Offset(input_shape, batch, in_y, in_x, channel)]);
+            }
+          }
+          max = std::max<uint8_t>(max, params.quantized_activation_min);
+          max = std::min<uint8_t>(max, params.quantized_activation_max);
+          output_data[Offset(output_shape, batch, out_y, out_x, channel)] =
+              static_cast<uint8_t>(max);
+        }
+      }
+    }
+  }
+}
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_POOLING_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/prelu.h

@@ -0,0 +1,109 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PRELU_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PRELU_H_
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+// Broadcast prelu to output_shape for quantized uint8_t/int8_t data.
+template <typename T>
+inline void BroadcastPrelu4DSlow(
+    const PreluParams& params, const RuntimeShape& input_shape,
+    const T* input_data, const RuntimeShape& alpha_shape, const T* alpha_data,
+    const RuntimeShape& output_shape, T* output_data) {
+  TFLITE_DCHECK_LE(input_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(alpha_shape.DimensionsCount(), 4);
+  TFLITE_DCHECK_LE(output_shape.DimensionsCount(), 4);
+  const RuntimeShape extended_output_shape =
+      RuntimeShape::ExtendedShape(4, output_shape);
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input_shape, alpha_shape, &desc1, &desc2);
+
+  for (int b = 0; b < extended_output_shape.Dims(0); ++b) {
+    for (int y = 0; y < extended_output_shape.Dims(1); ++y) {
+      for (int x = 0; x < extended_output_shape.Dims(2); ++x) {
+        for (int c = 0; c < extended_output_shape.Dims(3); ++c) {
+          int output_index = Offset(extended_output_shape, b, y, x, c);
+          int input_index = SubscriptToIndex(desc1, b, y, x, c);
+          const int32_t input_value =
+              params.input_offset + input_data[input_index];
+          int32_t output_value;
+          if (input_value >= 0) {
+            output_value = MultiplyByQuantizedMultiplier(
+                input_value, params.output_multiplier_1, params.output_shift_1);
+          } else {
+            auto alpha_index = SubscriptToIndex(desc2, b, y, x, c);
+            const int32_t alpha_value =
+                params.alpha_offset + alpha_data[alpha_index];
+
+            output_value = MultiplyByQuantizedMultiplier(
+                input_value * alpha_value, params.output_multiplier_2,
+                params.output_shift_2);
+          }
+          output_value += params.output_offset;
+
+          const int32_t quantized_min = std::numeric_limits<T>::min();
+          const int32_t quantized_max = std::numeric_limits<T>::max();
+          const int32_t clamped_output =
+              std::min(quantized_max, std::max(quantized_min, output_value));
+          output_data[output_index] = static_cast<T>(clamped_output);
+        }
+      }
+    }
+  }
+}
+
+template <typename T>
+inline void Prelu(const PreluParams& params, const RuntimeShape& input_shape,
+                  const T* input_data, const RuntimeShape& alpha_shape,
+                  const T* alpha_data, const RuntimeShape& output_shape,
+                  T* output_data) {
+  const int32_t quantized_min = std::numeric_limits<T>::min();
+  const int32_t quantized_max = std::numeric_limits<T>::max();
+
+  const int flat_size =
+      MatchingElementsSize(input_shape, alpha_shape, output_shape);
+  for (int i = 0; i < flat_size; ++i) {
+    const int32_t input_value = params.input_offset + input_data[i];
+    int32_t output_value;
+    if (input_value >= 0) {
+      output_value = MultiplyByQuantizedMultiplier(
+          input_value, params.output_multiplier_1, params.output_shift_1);
+    } else {
+      const int32_t alpha_value = params.alpha_offset + alpha_data[i];
+
+      output_value = MultiplyByQuantizedMultiplier(input_value * alpha_value,
+                                                   params.output_multiplier_2,
+                                                   params.output_shift_2);
+    }
+    output_value += params.output_offset;
+
+    const int32_t clamped_output =
+        std::min(quantized_max, std::max(quantized_min, output_value));
+    output_data[i] = static_cast<T>(clamped_output);
+  }
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PRELU_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/process_broadcast_shapes.h

@@ -0,0 +1,138 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PROCESS_BROADCAST_SHAPES_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PROCESS_BROADCAST_SHAPES_H_
+
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+// Consolidates dimensions in broadcast inputs, checks for five-fold pattern.
+//
+// For example, if sequence of dimensions of one input is
+// ..., 1, 3, 1, 7, 9, 5,... and the other is ..., 2, 3, 1, 7, 1, 1, ...
+// we can consolidate these as
+// ..., 1, 3*7, 9*5, ... and 2, 3*7, 1.
+//
+// The category is updated in the less-frequent case of shapes that are
+// not suited to a fivefold-loop broadcast.
+//
+// Falls back to generic pattern when it does not know how to process properly.
+//
+// Returns true iff there is some sort of broadcast, which includes five-fold
+// patterns and falling back to generic broadcast.
+inline bool ProcessBroadcastShapes(const RuntimeShape& shape0,
+                                   const RuntimeShape& shape1,
+                                   tflite::ArithmeticParams* params) {
+  const int dims_count =
+      std::max(shape0.DimensionsCount(), shape1.DimensionsCount());
+
+  params->broadcast_category = BroadcastableOpCategory::kGenericBroadcast;
+  RuntimeShape scalar_shape(dims_count, 1);
+
+  auto extended_shape0 = RuntimeShape::ExtendedShape(dims_count, shape0);
+  auto extended_shape1 = RuntimeShape::ExtendedShape(dims_count, shape1);
+
+  // Check for "exact" match, implicitly accepting any scalar shapes.
+  if (extended_shape0 == extended_shape1) {
+    params->broadcast_category = BroadcastableOpCategory::kNonBroadcast;
+    return false;
+  }
+
+  for (int i = dims_count - 1; i >= 0; --i) {
+    if (extended_shape0.Dims(i) == extended_shape1.Dims(i)) {
+      continue;
+    } else if (extended_shape0.Dims(i) == 1) {
+      params->broadcast_category =
+          BroadcastableOpCategory::kFirstInputBroadcastsFast;
+      break;
+    } else if (extended_shape1.Dims(i) == 1) {
+      params->broadcast_category =
+          BroadcastableOpCategory::kSecondInputBroadcastsFast;
+      break;
+    } else {
+      // This case is erroneous: there is a dimension that does not match and
+      // is not a broadcast from one shape to the other.
+      params->broadcast_category = BroadcastableOpCategory::kGenericBroadcast;
+      return true;
+    }
+  }
+
+  if (params->broadcast_category !=
+          BroadcastableOpCategory::kFirstInputBroadcastsFast &&
+      params->broadcast_category !=
+          BroadcastableOpCategory::kSecondInputBroadcastsFast) {
+    // This is unreachable because at least one else clause in the above loop
+    // must be reached.
+    TFLITE_DCHECK(false);
+    params->broadcast_category = BroadcastableOpCategory::kNonBroadcast;
+    return false;
+  }
+
+  // From this point it is assumed contractually that corresponding dimensions
+  // in shape0 and shape1 are either (a) equal or (b) one or other equals 1.
+  const bool swap_inputs = params->broadcast_category ==
+                           BroadcastableOpCategory::kSecondInputBroadcastsFast;
+  const RuntimeShape* shape_a =
+      swap_inputs ? &extended_shape1 : &extended_shape0;
+  const RuntimeShape* shape_b =
+      swap_inputs ? &extended_shape0 : &extended_shape1;
+
+  int i = dims_count - 1;
+  params->broadcast_shape[0] = 1;
+  params->broadcast_shape[1] = 1;
+  params->broadcast_shape[2] = 1;
+  params->broadcast_shape[3] = 1;
+  params->broadcast_shape[4] = 1;
+  // y_0 is greedy: include dims if both or neither equal 1: in other words,
+  // test for equality rather than (shape_a->Dims(i) != 1).
+  while (i >= 0 && shape_a->Dims(i) == shape_b->Dims(i)) {
+    params->broadcast_shape[4] *= shape_b->Dims(i);
+    --i;
+  }
+  // Here either input_a or input_b has dim of 1 (if i >= 0).  If it is input_b
+  // that has the unit dimension, the next two loops are not entered.
+  while (i >= 0 && shape_a->Dims(i) == 1) {
+    params->broadcast_shape[3] *= shape_b->Dims(i);
+    --i;
+  }
+  while (i >= 0 && shape_a->Dims(i) == shape_b->Dims(i)) {
+    params->broadcast_shape[2] *= shape_a->Dims(i);
+    --i;
+  }
+  // Here either input_a or input_b has dim of 1 (if i >= 0).
+  while (i >= 0 && shape_b->Dims(i) == 1) {
+    params->broadcast_shape[1] *= shape_a->Dims(i);
+    --i;
+  }
+  while (i >= 0 && shape_a->Dims(i) == shape_b->Dims(i)) {
+    params->broadcast_shape[0] *= shape_b->Dims(i);
+    --i;
+  }
+
+  // Rarer case is when the broadcast dimensions cannot be handled by a fivefold
+  // loop.
+  if (i >= 0) {
+    params->broadcast_category = BroadcastableOpCategory::kGenericBroadcast;
+  }
+  return true;
+}
+
+}  // namespace reference_ops
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_PROCESS_BROADCAST_SHAPES_H_

APP_Framework/Framework/knowing/tensorflow-lite/tensorflow-lite-for-mcu/source/tensorflow/lite/kernels/internal/reference/quantize.h

@@ -0,0 +1,55 @@
+/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+#ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_QUANTIZE_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_QUANTIZE_H_
+
+#include <algorithm>
+#include <limits>
+
+#include "tensorflow/lite/kernels/internal/common.h"
+#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/cppmath.h"
+#include "tensorflow/lite/kernels/internal/types.h"
+
+namespace tflite {
+
+namespace reference_ops {
+
+template <typename InputT, typename OutputT>
+inline void AffineQuantize(const tflite::QuantizationParams& op_params,
+                           const RuntimeShape& input_shape,
+                           const InputT* input_data,
+                           const RuntimeShape& output_shape,
+                           OutputT* output_data) {
+  const int32_t zero_point = op_params.zero_point;
+  const double scale = op_params.scale;
+  const int flat_size = MatchingFlatSize(input_shape, output_shape);
+  static constexpr int32_t min_val = std::numeric_limits<OutputT>::min();
+  static constexpr int32_t max_val = std::numeric_limits<OutputT>::max();
+
+  for (int i = 0; i < flat_size; i++) {
+    const InputT val = input_data[i];
+    int32_t unclamped =
+        static_cast<int32_t>(TfLiteRound(val / static_cast<float>(scale))) +
+        zero_point;
+    int32_t clamped = std::min(std::max(unclamped, min_val), max_val);
+    output_data[i] = clamped;
+  }
+}
+
+}  // namespace reference_ops
+
+}  // namespace tflite
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_QUANTIZE_H_