Merge pull request #35 from xianyi/develop

rebase
2020-03-01 23:44:10 +01:00 · 2020-03-01 23:44:10 +01:00 · 02d60c1563
parent 531c6b96d6 2e6963259b
commit 02d60c1563
22 changed files with 1423 additions and 153 deletions
--- a/Makefile.arm64
+++ b/Makefile.arm64
@ -24,6 +24,23 @@ CCOMMON_OPT += -march=armv8-a -mtune=cortex-a73
 FCOMMON_OPT += -march=armv8-a -mtune=cortex-a73
 endif

+# Use a72 tunings because Neoverse-N1 is only available
+# in GCC>=9
+ifeq ($(CORE), NEOVERSEN1)
+ifeq ($(GCCVERSIONGTEQ7), 1)
+ifeq ($(GCCVERSIONGTEQ9), 1)
+CCOMMON_OPT += -march=armv8.2-a -mtune=neoverse-n1
+FCOMMON_OPT += -march=armv8.2-a -mtune=neoverse-n1
+else
+CCOMMON_OPT += -march=armv8.2-a -mtune=cortex-a72
+FCOMMON_OPT += -march=armv8.2-a -mtune=cortex-a72
+endif
+else
+CCOMMON_OPT += -march=armv8-a -mtune=cortex-a72
+FCOMMON_OPT += -march=armv8-a -mtune=cortex-a72
+endif
+endif
+
 ifeq ($(CORE), THUNDERX)
 CCOMMON_OPT += -march=armv8-a -mtune=thunderx
 FCOMMON_OPT += -march=armv8-a -mtune=thunderx
--- a/Makefile.system
+++ b/Makefile.system
@ -328,6 +328,7 @@ ifeq ($(C_COMPILER), GCC)
 GCCVERSIONGTEQ4 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 4)
 GCCVERSIONGT4 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \> 4)
 GCCVERSIONGT5 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \> 5)
+GCCVERSIONGTEQ7 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 7)
 GCCVERSIONGTEQ9 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 9)
 GCCMINORVERSIONGTEQ7 := $(shell expr `$(CC) -dumpversion | cut -f2 -d.` \>= 7)
 ifeq ($(GCCVERSIONGT4), 1)
@ -554,6 +555,7 @@ DYNAMIC_CORE += CORTEXA53
 DYNAMIC_CORE += CORTEXA57
 DYNAMIC_CORE += CORTEXA72
 DYNAMIC_CORE += CORTEXA73
+DYNAMIC_CORE += NEOVERSEN1
 DYNAMIC_CORE += FALKOR
 DYNAMIC_CORE += THUNDERX
 DYNAMIC_CORE += THUNDERX2T99
--- a/Makefile.x86_64
+++ b/Makefile.x86_64
@ -15,10 +15,12 @@ CCOMMON_OPT += -march=skylake-avx512
 FCOMMON_OPT += -march=skylake-avx512
 ifeq ($(OSNAME), CYGWIN_NT)
 CCOMMON_OPT += -fno-asynchronous-unwind-tables
+FCOMMON_OPT += -fno-asynchronous-unwind-tables
 endif
 ifeq ($(OSNAME), WINNT)
 ifeq ($(C_COMPILER), GCC)
 CCOMMON_OPT += -fno-asynchronous-unwind-tables
+FCOMMON_OPT += -fno-asynchronous-unwind-tables
 endif
 endif
 endif
--- a/TargetList.txt
+++ b/TargetList.txt
@ -88,6 +88,7 @@ CORTEXA53
 CORTEXA57
 CORTEXA72
 CORTEXA73
+NEOVERSEN1
 FALKOR
 THUNDERX
 THUNDERX2T99
--- a/benchmark/Makefile
+++ b/benchmark/Makefile
@ -12,9 +12,9 @@ include $(TOPDIR)/Makefile.system
 # ACML 6.1 custom
 ACML=/home/saar/acml6.1/gfortran64_mp/lib
 LIBACML = -fopenmp $(ACML)/libacml_mp.so -lgfortran -lm
- 

-# Atlas Ubuntu 
+
+# Atlas Ubuntu
 #ATLAS=/usr/lib/atlas-base
 #LIBATLAS = -fopenmp $(ATLAS)/liblapack_atlas.a  $(ATLAS)/libptcblas.a  $(ATLAS)/libptf77blas.a  $(ATLAS)/libatlas.a -lgfortran -lm

@ -56,6 +56,7 @@ goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \
       sgemm.goto dgemm.goto cgemm.goto zgemm.goto \
       strmm.goto dtrmm.goto ctrmm.goto ztrmm.goto \
       strsm.goto dtrsm.goto ctrsm.goto ztrsm.goto \
+       ssyr.goto dsyr.goto \
       ssyrk.goto dsyrk.goto csyrk.goto zsyrk.goto \
       ssyr2k.goto dsyr2k.goto csyr2k.goto zsyr2k.goto \
       sger.goto dger.goto cger.goto zger.goto \
@ -83,6 +84,7 @@ acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
       sgemm.acml dgemm.acml cgemm.acml zgemm.acml \
       strmm.acml dtrmm.acml ctrmm.acml ztrmm.acml \
       strsm.acml dtrsm.acml ctrsm.acml ztrsm.acml \
+       ssyr.acml dsyr.acml \
       ssyrk.acml dsyrk.acml csyrk.acml zsyrk.acml \
       ssyr2k.acml dsyr2k.acml csyr2k.acml zsyr2k.acml \
       sger.acml dger.acml cger.acml zger.acml \
@ -109,6 +111,7 @@ atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \
       sgemm.atlas dgemm.atlas cgemm.atlas zgemm.atlas \
       strmm.atlas dtrmm.atlas ctrmm.atlas ztrmm.atlas \
       strsm.atlas dtrsm.atlas ctrsm.atlas ztrsm.atlas \
+       ssyr.goto dsyr.atlas \
       ssyrk.atlas dsyrk.atlas csyrk.atlas zsyrk.atlas \
       ssyr2k.atlas dsyr2k.atlas csyr2k.atlas zsyr2k.atlas \
       sger.atlas dger.atlas cger.atlas zger.atlas\
@ -136,6 +139,7 @@ mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \
       sgemm.mkl dgemm.mkl cgemm.mkl zgemm.mkl \
       strmm.mkl dtrmm.mkl ctrmm.mkl ztrmm.mkl \
       strsm.mkl dtrsm.mkl ctrsm.mkl ztrsm.mkl \
+       ssyr.mkl dsyr.mkl \
       ssyrk.mkl dsyrk.mkl csyrk.mkl zsyrk.mkl \
       ssyr2k.mkl dsyr2k.mkl csyr2k.mkl zsyr2k.mkl \
       sger.mkl dger.mkl cger.mkl zger.mkl \
@ -162,6 +166,7 @@ else
 goto :: sgemm.goto dgemm.goto cgemm.goto zgemm.goto \
       strmm.goto dtrmm.goto ctrmm.goto ztrmm.goto \
       strsm.goto dtrsm.goto ctrsm.goto ztrsm.goto \
+       ssyr.goto dsyr.goto \
       ssyrk.goto dsyrk.goto csyrk.goto zsyrk.goto \
       ssyr2k.goto dsyr2k.goto csyr2k.goto zsyr2k.goto \
       sger.goto dger.goto cger.goto zger.goto \
@ -188,6 +193,7 @@ acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
       sgemm.acml dgemm.acml cgemm.acml zgemm.acml \
       strmm.acml dtrmm.acml ctrmm.acml ztrmm.acml \
       strsm.acml dtrsm.acml ctrsm.acml ztrsm.acml \
+       ssyr.acml dsyr.acml \
       ssyrk.acml dsyrk.acml csyrk.acml zsyrk.acml \
       ssyr2k.acml dsyr2k.acml csyr2k.acml zsyr2k.acml \
       sger.acml dger.acml cger.acml zger.acml \
@ -214,6 +220,7 @@ atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \
       sgemm.atlas dgemm.atlas cgemm.atlas zgemm.atlas \
       strmm.atlas dtrmm.atlas ctrmm.atlas ztrmm.atlas \
       strsm.atlas dtrsm.atlas ctrsm.atlas ztrsm.atlas \
+       ssyr.atlas dsyr.atlas \
       ssyrk.atlas dsyrk.atlas csyrk.atlas zsyrk.atlas \
       ssyr2k.atlas dsyr2k.atlas csyr2k.atlas zsyr2k.atlas \
       sger.atlas dger.atlas cger.atlas zger.atlas\
@ -243,6 +250,7 @@ mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \
       sgemm.mkl dgemm.mkl cgemm.mkl zgemm.mkl \
       strmm.mkl dtrmm.mkl ctrmm.mkl ztrmm.mkl \
       strsm.mkl dtrsm.mkl ctrsm.mkl ztrsm.mkl \
+       ssyr.mkl dsyr.mkl \
       ssyrk.mkl dsyrk.mkl csyrk.mkl zsyrk.mkl \
       ssyr2k.mkl dsyr2k.mkl csyr2k.mkl zsyr2k.mkl \
       sger.mkl dger.mkl cger.mkl zger.mkl \
@ -280,6 +288,7 @@ veclib :: slinpack.veclib dlinpack.veclib clinpack.veclib zlinpack.veclib \
       sgemm.veclib dgemm.veclib cgemm.veclib zgemm.veclib \
       strmm.veclib dtrmm.veclib ctrmm.veclib ztrmm.veclib \
       strsm.veclib dtrsm.veclib ctrsm.veclib ztrsm.veclib \
+       ssyr.veclib dsyr.veclib \
       ssyrk.veclib dsyrk.veclib csyrk.veclib zsyrk.veclib \
       ssyr2k.veclib dsyr2k.veclib csyr2k.veclib zsyr2k.veclib \
       sger.veclib dger.veclib cger.veclib zger.veclib \
@ -768,6 +777,36 @@ ztrsm.veclib : ztrsm.$(SUFFIX)

 ztrsm.essl : ztrsm.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBESSL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+##################################### Ssyr ####################################################
+ssyr.goto : ssyr.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
+
+ssyr.acml : ssyr.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+ssyr.atlas : ssyr.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+ssyr.mkl : ssyr.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+ssyr.veclib : ssyr.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+##################################### Dsyr ####################################################
+dsyr.goto : dsyr.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
+
+dsyr.acml : dsyr.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dsyr.atlas : dsyr.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dsyr.mkl : dsyr.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dsyr.veclib : dsyr.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBVECLIB) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)

 ##################################### Ssyrk ####################################################
 ssyrk.goto : ssyrk.$(SUFFIX) ../$(LIBNAME)
@ -2078,6 +2117,12 @@ ctrsm.$(SUFFIX) : trsm.c
 ztrsm.$(SUFFIX) : trsm.c
 	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^

+ssyr.$(SUFFIX) : syr.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^
+
+dsyr.$(SUFFIX) : syr.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^
+
 ssyrk.$(SUFFIX) : syrk.c
 	$(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^

--- a/benchmark/syr.c
+++ b/benchmark/syr.c
@ -0,0 +1,187 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include <stdio.h>
+#include <stdlib.h>
+#ifdef __CYGWIN32__
+#include <sys/time.h>
+#endif
+#include "common.h"
+
+
+#undef SYR
+
+#ifdef DOUBLE
+#define SYR   BLASFUNC(dsyr)
+#else
+#define SYR   BLASFUNC(ssyr)
+#endif
+
+
+
+#if defined(__WIN32__) || defined(__WIN64__)
+
+#ifndef DELTA_EPOCH_IN_MICROSECS
+#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL
+#endif
+
+int gettimeofday(struct timeval *tv, void *tz){
+
+  FILETIME ft;
+  unsigned __int64 tmpres = 0;
+  static int tzflag;
+
+  if (NULL != tv)
+    {
+      GetSystemTimeAsFileTime(&ft);
+
+      tmpres |= ft.dwHighDateTime;
+      tmpres <<= 32;
+      tmpres |= ft.dwLowDateTime;
+
+      /*converting file time to unix epoch*/
+      tmpres /= 10;  /*convert into microseconds*/
+      tmpres -= DELTA_EPOCH_IN_MICROSECS;
+      tv->tv_sec = (long)(tmpres / 1000000UL);
+      tv->tv_usec = (long)(tmpres % 1000000UL);
+    }
+
+  return 0;
+}
+
+#endif
+
+#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0
+
+static void *huge_malloc(BLASLONG size){
+  int shmid;
+  void *address;
+
+#ifndef SHM_HUGETLB
+#define SHM_HUGETLB 04000
+#endif
+
+  if ((shmid =shmget(IPC_PRIVATE,
+		     (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1),
+		     SHM_HUGETLB | IPC_CREAT |0600)) < 0) {
+    printf( "Memory allocation failed(shmget).\n");
+    exit(1);
+  }
+
+  address = shmat(shmid, NULL, SHM_RND);
+
+  if ((BLASLONG)address == -1){
+    printf( "Memory allocation failed(shmat).\n");
+    exit(1);
+  }
+
+  shmctl(shmid, IPC_RMID, 0);
+
+  return address;
+}
+
+#define malloc huge_malloc
+
+#endif
+
+int main(int argc, char *argv[]){
+
+  FLOAT *x,*a;
+  FLOAT alpha[] = {1.0, 1.0};
+  char *p;
+
+  char uplo='U';
+
+  if ((p = getenv("OPENBLAS_UPLO"))) uplo=*p;
+
+  blasint m, i, j;
+  blasint inc_x= 1;
+  int from =   1;
+  int to   = 200;
+  int step =   1;
+  
+  struct timeval start, stop;
+  double time1;
+
+  argc--;argv++;
+
+  if (argc > 0) { from     = atol(*argv);		argc--; argv++;}
+  if (argc > 0) { to       = MAX(atol(*argv), from);	argc--; argv++;}
+  if (argc > 0) { step     = atol(*argv);		argc--; argv++;}
+
+  fprintf(stderr, "From : %3d  To : %3d Step = %3d Uplo = %c Inc_x = %d\n", from, to, step,uplo,inc_x);
+
+
+  if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to *  COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+   if (( x = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+#ifdef linux
+  srandom(getpid());
+#endif
+
+  fprintf(stderr, "   SIZE       Flops\n");
+
+  for(m = from; m <= to; m += step)
+  {
+
+    fprintf(stderr, " %6d : ", (int)m);
+	
+	for(i = 0; i < m * COMPSIZE * abs(inc_x); i++){
+	x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+   	}
+	
+    for(j = 0; j < m; j++){
+      for(i = 0; i < m * COMPSIZE; i++){
+	a[(long)i + (long)j * (long)m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+      }
+    }
+
+    gettimeofday( &start, (struct timezone *)0);
+
+    SYR (&uplo, &m, alpha, x, &inc_x, a, &m );
+
+    gettimeofday( &stop, (struct timezone *)0);
+
+    time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
+
+    gettimeofday( &start, (struct timezone *)0);
+
+    fprintf(stderr,
+	    " %10.2f MFlops\n",
+	    COMPSIZE * COMPSIZE * 1. * (double)m * (double)m * (double)m / time1 * 1.e-6);
+
+  }
+
+  return 0;
+}
+
+// void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
--- a/cmake/arch.cmake
+++ b/cmake/arch.cmake
@ -45,7 +45,7 @@ endif ()

 if (DYNAMIC_ARCH)
  if (ARM64)
-    set(DYNAMIC_CORE ARMV8 CORTEXA53 CORTEXA57 CORTEXA72 CORTEXA73 FALKOR THUNDERX THUNDERX2T99 TSV110 EMAG8180)
+    set(DYNAMIC_CORE ARMV8 CORTEXA53 CORTEXA57 CORTEXA72 CORTEXA73 FALKOR THUNDERX THUNDERX2T99 TSV110 EMAG8180 NEOVERSEN1)
  endif ()
  
  if (POWER)
--- a/cmake/prebuild.cmake
+++ b/cmake/prebuild.cmake
@ -229,6 +229,33 @@ if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSS
    set(ZGEMM_UNROLL_M 4)
    set(ZGEMM_UNROLL_N 4)
    set(SYMV_P 16)
+  elseif ("${TCORE}" STREQUAL "NEOVERSEN1")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L1_CODE_SIZE\t65536\n"
+      "#define L1_CODE_LINESIZE\t64\n"
+      "#define L1_CODE_ASSOCIATIVE\t4\n"
+      "#define L1_DATA_SIZE\t65536\n"
+      "#define L1_DATA_LINESIZE\t64\n"
+      "#define L1_DATA_ASSOCIATIVE\t2\n"
+      "#define L2_SIZE\t1048576\n\n"
+      "#define L2_LINESIZE\t64\n"
+      "#define L2_ASSOCIATIVE\t16\n"
+      "#define DTB_DEFAULT_ENTRIES\t64\n"
+      "#define DTB_SIZE\t4096\n"
+      "#define HAVE_VFPV4\n"
+      "#define HAVE_VFPV3\n"
+      "#define HAVE_VFP\n"
+      "#define HAVE_NEON\n"
+      "#define ARMV8\n")
+    set(SGEMM_UNROLL_M 16)
+    set(SGEMM_UNROLL_N 4)
+    set(DGEMM_UNROLL_M 8)
+    set(DGEMM_UNROLL_N 4)
+    set(CGEMM_UNROLL_M 8)
+    set(CGEMM_UNROLL_N 4)
+    set(ZGEMM_UNROLL_M 4)
+    set(ZGEMM_UNROLL_N 4)
+    set(SYMV_P 16)
  elseif ("${TCORE}" STREQUAL "FALKOR")
    file(APPEND ${TARGET_CONF_TEMP}
      "#define L1_CODE_SIZE\t65536\n"
--- a/common_arm64.h
+++ b/common_arm64.h
@ -53,16 +53,16 @@ static void __inline blas_lock(volatile BLASULONG *address){
  BLASULONG ret;

  do {
-    while (*address) {YIELDING;};
-
    __asm__ __volatile__(
 			 "mov	x4, #1							\n\t"
+			 "sevl								\n\t"
                         "1:                                                            \n\t"
+                         "wfe                                                           \n\t"
+			 "2:								\n\t"
                         "ldaxr x2, [%1]                                                \n\t"
                         "cbnz  x2, 1b                                                  \n\t"
-			 "2:								\n\t"
                         "stxr  w3, x4, [%1]                                            \n\t"
-                         "cbnz  w3, 1b                                                  \n\t"
+                         "cbnz  w3, 2b                                                  \n\t"
                         "mov   %0, #0                                                  \n\t"
                         : "=r"(ret), "=r"(address)
                         : "1"(address)
@ -81,10 +81,12 @@ static void __inline blas_lock(volatile BLASULONG *address){
 #if !defined(OS_DARWIN) && !defined (OS_ANDROID)
 static __inline BLASULONG rpcc(void){
  BLASULONG ret = 0;
+  blasint shift;
 
  __asm__ __volatile__ ("isb; mrs %0,cntvct_el0":"=r"(ret));
+  __asm__ __volatile__ ("mrs %0,cntfrq_el0; clz %w0, %w0":"=&r"(shift));

-  return ret;
+  return ret << shift;
 }

 #define RPCC_DEFINED
--- a/cpuid_arm64.c
+++ b/cpuid_arm64.c
@ -34,6 +34,7 @@
 #define CPU_CORTEXA57     3
 #define CPU_CORTEXA72     4
 #define CPU_CORTEXA73     5
+#define CPU_NEOVERSEN1    11
 // Qualcomm
 #define CPU_FALKOR        6
 // Cavium
@ -55,7 +56,8 @@ static char *cpuname[] = {
  "THUNDERX",
  "THUNDERX2T99",
  "TSV110",
-  "EMAG8180"
+  "EMAG8180",
+  "NEOVERSEN1"
 };

 static char *cpuname_lower[] = {
@ -69,7 +71,8 @@ static char *cpuname_lower[] = {
  "thunderx",
  "thunderx2t99",
  "tsv110",
-  "emag8180"
+  "emag8180",
+  "neoversen1"
 };

 int get_feature(char *search)
@ -144,6 +147,8 @@ int detect(void)
        return CPU_CORTEXA72;
      else if (strstr(cpu_part, "0xd09"))
        return CPU_CORTEXA73;
+      else if (strstr(cpu_part, "0xd0c"))
+        return CPU_NEOVERSEN1;
    }
    // Qualcomm
    else if (strstr(cpu_implementer, "0x51") && strstr(cpu_part, "0xc00"))
@ -285,6 +290,20 @@ void get_cpuconfig(void)
 			printf("#define DTB_DEFAULT_ENTRIES 64\n");
 			printf("#define DTB_SIZE 4096\n");
 			break;
+		case CPU_NEOVERSEN1:
+			printf("#define %s\n", cpuname[d]);
+			printf("#define L1_CODE_SIZE 65536\n");
+			printf("#define L1_CODE_LINESIZE 64\n");
+			printf("#define L1_CODE_ASSOCIATIVE 4\n");
+			printf("#define L1_DATA_SIZE 65536\n");
+			printf("#define L1_DATA_LINESIZE 64\n");
+			printf("#define L1_DATA_ASSOCIATIVE 4\n");
+			printf("#define L2_SIZE 1048576\n");
+			printf("#define L2_LINESIZE 64\n");
+			printf("#define L2_ASSOCIATIVE 16\n");
+			printf("#define DTB_DEFAULT_ENTRIES 64\n");
+			printf("#define DTB_SIZE 4096\n");
+			break;

    case CPU_FALKOR:
      printf("#define FALKOR\n");
--- a/driver/level3/level3_thread.c
+++ b/driver/level3/level3_thread.c
@ -351,8 +351,9 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
      /* Make sure if no one is using workspace */
      START_RPCC();
      for (i = 0; i < args -> nthreads; i++)
-	while (job[mypos].working[i][CACHE_LINE_SIZE * bufferside]) {YIELDING;MB;};
+	while (job[mypos].working[i][CACHE_LINE_SIZE * bufferside]) {YIELDING;};
      STOP_RPCC(waiting1);
+      MB;

 #if defined(FUSED_GEMM) && !defined(TIMING)

@ -395,10 +396,10 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
      }
 #endif

+      WMB;
      /* Set flag so other threads can access local region of B */
      for (i = mypos_n * nthreads_m; i < (mypos_n + 1) * nthreads_m; i++)
        job[mypos].working[i][CACHE_LINE_SIZE * bufferside] = (BLASLONG)buffer[bufferside];
-      WMB;
    }

    /* Get regions of B from other threads and apply kernel */
@ -417,8 +418,9 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,

 	  /* Wait until other region of B is initialized */
 	  START_RPCC();
-	  while(job[current].working[mypos][CACHE_LINE_SIZE * bufferside] == 0) {YIELDING;MB;};
+	  while(job[current].working[mypos][CACHE_LINE_SIZE * bufferside] == 0) {YIELDING;};
 	  STOP_RPCC(waiting2);
+	  MB;

          /* Apply kernel with local region of A and part of other region of B */
 	  START_RPCC();
@ -434,8 +436,8 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,

        /* Clear synchronization flag if this thread is done with other region of B */
 	if (m_to - m_from == min_i) {
-	  job[current].working[mypos][CACHE_LINE_SIZE * bufferside] &= 0;
 	  WMB;
+	  job[current].working[mypos][CACHE_LINE_SIZE * bufferside] &= 0;
 	}
      }
    } while (current != mypos);
@ -477,8 +479,8 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
          
          /* Clear synchronization flag if this thread is done with region of B */
          if (is + min_i >= m_to) {
-            job[current].working[mypos][CACHE_LINE_SIZE * bufferside] &= 0;
            WMB;
+            job[current].working[mypos][CACHE_LINE_SIZE * bufferside] &= 0;
          }
 	}

@ -497,10 +499,11 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
  START_RPCC();
  for (i = 0; i < args -> nthreads; i++) {
    for (js = 0; js < DIVIDE_RATE; js++) {
-      while (job[mypos].working[i][CACHE_LINE_SIZE * js] ) {YIELDING;MB;};
+      while (job[mypos].working[i][CACHE_LINE_SIZE * js] ) {YIELDING;};
    }
  }
  STOP_RPCC(waiting3);
+  MB;

 #ifdef TIMING
  BLASLONG waiting = waiting1 + waiting2 + waiting3;
@ -705,7 +708,7 @@ EnterCriticalSection((PCRITICAL_SECTION)&level3_lock);
 	}
      }
    }
-
+    WMB;
    /* Execute parallel computation */
    exec_blas(nthreads, queue);
  }
--- a/driver/others/dynamic_arm64.c
+++ b/driver/others/dynamic_arm64.c
@ -52,10 +52,11 @@ extern gotoblas_t  gotoblas_THUNDERX;
 extern gotoblas_t  gotoblas_THUNDERX2T99;
 extern gotoblas_t  gotoblas_TSV110;
 extern gotoblas_t  gotoblas_EMAG8180;
+extern gotoblas_t  gotoblas_NEOVERSEN1;

 extern void openblas_warning(int verbose, const char * msg);

-#define NUM_CORETYPES   10
+#define NUM_CORETYPES   11

 /*
 * In case asm/hwcap.h is outdated on the build system, make sure
@ -80,6 +81,7 @@ static char *corename[] = {
  "thunderx2t99",
  "tsv110",
  "emag8180",
+  "neoversen1",
  "unknown"
 };

@ -94,6 +96,7 @@ char *gotoblas_corename(void) {
  if (gotoblas == &gotoblas_THUNDERX2T99) return corename[ 7];
  if (gotoblas == &gotoblas_TSV110)       return corename[ 8];
  if (gotoblas == &gotoblas_EMAG8180)     return corename[ 9];
+  if (gotoblas == &gotoblas_NEOVERSEN1)   return corename[10];
  return corename[NUM_CORETYPES];
 }

@ -123,6 +126,7 @@ static gotoblas_t *force_coretype(char *coretype) {
    case  7: return (&gotoblas_THUNDERX2T99);
    case  8: return (&gotoblas_TSV110);
    case  9: return (&gotoblas_EMAG8180);
+    case 10: return (&gotoblas_NEOVERSEN1);
  }
  snprintf(message, 128, "Core not found: %s\n", coretype);
  openblas_warning(1, message);
@ -168,6 +172,8 @@ static gotoblas_t *get_coretype(void) {
          return &gotoblas_CORTEXA72;
        case 0xd09: // Cortex A73
          return &gotoblas_CORTEXA73;
+        case 0xd0c: // Neoverse N1
+          return &gotoblas_NEOVERSEN1;
      }
      break;
    case 0x42: // Broadcom
--- a/getarch.c
+++ b/getarch.c
@ -1028,6 +1028,24 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #else
 #endif

+#ifdef FORCE_NEOVERSEN1
+#define FORCE
+#define ARCHITECTURE    "ARM64"
+#define SUBARCHITECTURE "NEOVERSEN1"
+#define SUBDIRNAME      "arm64"
+#define ARCHCONFIG   "-DNEOVERSEN1 " \
+       "-DL1_CODE_SIZE=65536 -DL1_CODE_LINESIZE=64 -DL1_CODE_ASSOCIATIVE=4 " \
+       "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 -DL1_DATA_ASSOCIATIVE=4 " \
+       "-DL2_SIZE=1048576 -DL2_LINESIZE=64 -DL2_ASSOCIATIVE=16 " \
+       "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
+       "-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DARMV8" \
+       "-march=armv8.2-a -mtune=cortex-a72"
+#define LIBNAME   "neoversen1"
+#define CORENAME  "NEOVERSEN1"
+#else
+#endif
+
+
 #ifdef FORCE_FALKOR
 #define FORCE
 #define ARCHITECTURE    "ARM64"
--- a/kernel/arm64/KERNEL.NEOVERSEN1
+++ b/kernel/arm64/KERNEL.NEOVERSEN1
@ -0,0 +1,189 @@
+SAMINKERNEL  = ../arm/amin.c
+DAMINKERNEL  = ../arm/amin.c
+CAMINKERNEL  = ../arm/zamin.c
+ZAMINKERNEL  = ../arm/zamin.c
+
+SMAXKERNEL   = ../arm/max.c
+DMAXKERNEL   = ../arm/max.c
+
+SMINKERNEL   = ../arm/min.c
+DMINKERNEL   = ../arm/min.c
+
+ISAMINKERNEL = ../arm/iamin.c
+IDAMINKERNEL = ../arm/iamin.c
+ICAMINKERNEL = ../arm/izamin.c
+IZAMINKERNEL = ../arm/izamin.c
+
+ISMAXKERNEL  = ../arm/imax.c
+IDMAXKERNEL  = ../arm/imax.c
+
+ISMINKERNEL  = ../arm/imin.c
+IDMINKERNEL  = ../arm/imin.c
+
+STRSMKERNEL_LN	=  ../generic/trsm_kernel_LN.c
+STRSMKERNEL_LT	=  ../generic/trsm_kernel_LT.c
+STRSMKERNEL_RN	=  ../generic/trsm_kernel_RN.c
+STRSMKERNEL_RT	=  ../generic/trsm_kernel_RT.c
+
+DTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+DTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+DTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+DTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+CTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+CTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+CTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+CTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+ZTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+ZTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+ZTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+ZTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+SAMAXKERNEL  = amax.S
+DAMAXKERNEL  = amax.S
+CAMAXKERNEL  = zamax.S
+ZAMAXKERNEL  = zamax.S
+
+SAXPYKERNEL  = axpy.S
+DAXPYKERNEL  = daxpy_thunderx2t99.S
+CAXPYKERNEL  = zaxpy.S
+ZAXPYKERNEL  = zaxpy.S
+
+SROTKERNEL   = rot.S
+DROTKERNEL   = rot.S
+CROTKERNEL   = zrot.S
+ZROTKERNEL   = zrot.S
+
+SSCALKERNEL  = scal.S
+DSCALKERNEL  = scal.S
+CSCALKERNEL  = zscal.S
+ZSCALKERNEL  = zscal.S
+
+SGEMVNKERNEL = gemv_n.S
+DGEMVNKERNEL = gemv_n.S
+CGEMVNKERNEL = zgemv_n.S
+ZGEMVNKERNEL = zgemv_n.S
+
+SGEMVTKERNEL = gemv_t.S
+DGEMVTKERNEL = gemv_t.S
+CGEMVTKERNEL = zgemv_t.S
+ZGEMVTKERNEL = zgemv_t.S
+
+
+SASUMKERNEL    = sasum_thunderx2t99.c
+DASUMKERNEL    = dasum_thunderx2t99.c
+CASUMKERNEL    = casum_thunderx2t99.c
+ZASUMKERNEL    = zasum_thunderx2t99.c
+
+SCOPYKERNEL    = copy_thunderx2t99.c
+DCOPYKERNEL    = copy_thunderx2t99.c
+CCOPYKERNEL    = copy_thunderx2t99.c
+ZCOPYKERNEL    = copy_thunderx2t99.c
+
+SSWAPKERNEL    = swap_thunderx2t99.S
+DSWAPKERNEL    = swap_thunderx2t99.S
+CSWAPKERNEL    = swap_thunderx2t99.S
+ZSWAPKERNEL    = swap_thunderx2t99.S
+
+ISAMAXKERNEL   = iamax_thunderx2t99.c
+IDAMAXKERNEL   = iamax_thunderx2t99.c
+ICAMAXKERNEL   = izamax_thunderx2t99.c
+IZAMAXKERNEL   = izamax_thunderx2t99.c
+
+SNRM2KERNEL    = scnrm2_thunderx2t99.c
+DNRM2KERNEL    = dznrm2_thunderx2t99.c
+CNRM2KERNEL    = scnrm2_thunderx2t99.c
+ZNRM2KERNEL    = dznrm2_thunderx2t99.c
+
+DDOTKERNEL     = dot_thunderx2t99.c
+SDOTKERNEL     = dot_thunderx2t99.c
+CDOTKERNEL     = zdot_thunderx2t99.c
+ZDOTKERNEL     = zdot_thunderx2t99.c
+DSDOTKERNEL    = dot.S
+
+DGEMM_BETA     = dgemm_beta.S
+SGEMM_BETA     = sgemm_beta.S
+
+SGEMMKERNEL    =  sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
+STRMMKERNEL    =  strmm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
+ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
+ifeq ($(SGEMM_UNROLL_M), 16)
+SGEMMITCOPY    =  sgemm_tcopy_$(SGEMM_UNROLL_M).S
+else
+SGEMMITCOPY    =  ../generic/gemm_tcopy_$(SGEMM_UNROLL_M).c
+endif
+ifeq ($(SGEMM_UNROLL_M), 4)
+SGEMMINCOPY    =  sgemm_ncopy_$(SGEMM_UNROLL_M).S
+else
+SGEMMINCOPY    =  ../generic/gemm_ncopy_$(SGEMM_UNROLL_M).c
+endif
+SGEMMINCOPYOBJ =  sgemm_incopy$(TSUFFIX).$(SUFFIX)
+SGEMMITCOPYOBJ =  sgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+ifeq ($(SGEMM_UNROLL_N), 16)
+SGEMMOTCOPY    =  sgemm_tcopy_$(SGEMM_UNROLL_N).S
+else
+SGEMMOTCOPY    =  ../generic/gemm_tcopy_$(SGEMM_UNROLL_N).c
+endif
+ifeq ($(SGEMM_UNROLL_N), 4)
+SGEMMONCOPY    =  sgemm_ncopy_$(SGEMM_UNROLL_N).S
+else
+SGEMMONCOPY    =  ../generic/gemm_ncopy_$(SGEMM_UNROLL_N).c
+endif
+SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+DGEMMKERNEL    =  dgemm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N).S
+DTRMMKERNEL    =  dtrmm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N).S
+
+ifneq ($(DGEMM_UNROLL_M), $(DGEMM_UNROLL_N))
+
+ifeq ($(DGEMM_UNROLL_M), 8)
+DGEMMINCOPY    =  dgemm_ncopy_$(DGEMM_UNROLL_M).S
+DGEMMITCOPY    =  dgemm_tcopy_$(DGEMM_UNROLL_M).S
+else
+DGEMMINCOPY    =  ../generic/gemm_ncopy_$(DGEMM_UNROLL_M).c
+DGEMMITCOPY    =  ../generic/gemm_tcopy_$(DGEMM_UNROLL_M).c
+endif
+
+DGEMMINCOPYOBJ =  dgemm_incopy$(TSUFFIX).$(SUFFIX)
+DGEMMITCOPYOBJ =  dgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+
+ifeq ($(DGEMM_UNROLL_N), 4)
+DGEMMONCOPY    =  dgemm_ncopy_$(DGEMM_UNROLL_N).S
+DGEMMOTCOPY    =  dgemm_tcopy_$(DGEMM_UNROLL_N).S
+else
+DGEMMONCOPY    =  ../generic/gemm_ncopy_$(DGEMM_UNROLL_N).c
+DGEMMOTCOPY    =  ../generic/gemm_tcopy_$(DGEMM_UNROLL_N).c
+endif
+
+DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
+DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+CGEMMKERNEL    =  cgemm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S
+CTRMMKERNEL    =  ctrmm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S
+ifneq ($(CGEMM_UNROLL_M), $(CGEMM_UNROLL_N))
+CGEMMINCOPY    =  ../generic/zgemm_ncopy_$(CGEMM_UNROLL_M).c
+CGEMMITCOPY    =  ../generic/zgemm_tcopy_$(CGEMM_UNROLL_M).c
+CGEMMINCOPYOBJ =  cgemm_incopy$(TSUFFIX).$(SUFFIX)
+CGEMMITCOPYOBJ =  cgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+CGEMMONCOPY    =  ../generic/zgemm_ncopy_$(CGEMM_UNROLL_N).c
+CGEMMOTCOPY    =  ../generic/zgemm_tcopy_$(CGEMM_UNROLL_N).c
+CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
+CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+ZGEMMKERNEL    =  zgemm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N).S
+ZTRMMKERNEL    =  ztrmm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N).S
+ifneq ($(ZGEMM_UNROLL_M), $(ZGEMM_UNROLL_N))
+ZGEMMINCOPY    =  ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_M).c
+ZGEMMITCOPY    =  ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_M).c
+ZGEMMINCOPYOBJ =  zgemm_incopy$(TSUFFIX).$(SUFFIX)
+ZGEMMITCOPYOBJ =  zgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+ZGEMMONCOPY    =  ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_N).c
+ZGEMMOTCOPY    =  ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_N).c
+ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
+ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@ -32,7 +32,7 @@ CAXPYKERNEL = caxpy.c
 ZAXPYKERNEL = zaxpy.c

 STRMMKERNEL    =  sgemm_kernel_8x4_haswell.c
-SGEMMKERNEL    =  sgemm_kernel_8x4_haswell.c
+SGEMMKERNEL    =  sgemm_kernel_8x4_haswell_2.c
 SGEMM_BETA     =  sgemm_beta_skylakex.c
 SGEMMINCOPY    =  ../generic/gemm_ncopy_8.c
 SGEMMITCOPY    =  ../generic/gemm_tcopy_8.c
--- a/kernel/x86_64/KERNEL.SKYLAKEX
+++ b/kernel/x86_64/KERNEL.SKYLAKEX
@ -1,7 +1,7 @@
 include $(KERNELDIR)/KERNEL.HASWELL

 SGEMMKERNEL    =  sgemm_kernel_16x4_skylakex_2.c
-STRMMKERNEL    =  sgemm_kernel_16x4_haswell.S
+STRMMKERNEL    =  sgemm_kernel_16x4_skylakex_2.c
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c
 SGEMMITCOPY    =  sgemm_tcopy_16_skylakex.c
 SGEMMONCOPY    =  sgemm_ncopy_4_skylakex.c
--- a/kernel/x86_64/KERNEL.ZEN
+++ b/kernel/x86_64/KERNEL.ZEN
@ -31,7 +31,7 @@ CAXPYKERNEL = caxpy.c
 ZAXPYKERNEL = zaxpy.c

 STRMMKERNEL    =  sgemm_kernel_8x4_haswell.c
-SGEMMKERNEL    =  sgemm_kernel_8x4_haswell.c
+SGEMMKERNEL    =  sgemm_kernel_8x4_haswell_2.c
 SGEMMINCOPY    =  ../generic/gemm_ncopy_8.c
 SGEMMITCOPY    =  ../generic/gemm_tcopy_8.c
 SGEMMONCOPY    =  ../generic/gemm_ncopy_4.c
--- a/kernel/x86_64/cgemm_kernel_8x2_haswell.c
+++ b/kernel/x86_64/cgemm_kernel_8x2_haswell.c
@ -50,7 +50,7 @@
  "vmovups (%0),%%ymm0; vmovups 32(%0),%%ymm1; prefetcht0 512(%0); addq $64,%0;"\
  acc_m8n2_con(0,1,4,5,6,7,0,8,%1) acc_m8n2_con(0,1,8,9,10,11,0,8,%1,%%r12,1)
 #define KERNEL_2_k1m8n4 \
-  "vpermilps $177,%%ymm0,%%ymm0; vpermilps $177,%%ymm1,%%ymm1;"\
+  "vpermilps $177,-64(%0),%%ymm0; vpermilps $177,-32(%0),%%ymm1;"\
  acc_m8n2_con(0,1,4,5,6,7,4,12,%1) acc_m8n2_con(0,1,8,9,10,11,4,12,%1,%%r12,1)
 #define KERNEL_1_k1m8n6 KERNEL_1_k1m8n4 acc_m8n2_con(0,1,12,13,14,15,0,8,%1,%%r12,2)
 #define KERNEL_2_k1m8n6 KERNEL_2_k1m8n4 acc_m8n2_con(0,1,12,13,14,15,4,12,%1,%%r12,2)
@ -93,7 +93,6 @@
  "movq $10,%5; movq $84,%%r15;"\
  #ndim"8881:\n\t"\
  "prefetcht1 (%3); subq $63,%3; addq %%r15,%3;"\
-  "prefetcht0 64(%1); prefetcht0 64(%1,%%r12,1); prefetcht0 64(%1,%%r12,2);"\
  KERNEL_k1m8n##ndim KERNEL_k1m8n##ndim\
  "testq $12,%5; movq $84,%%r15; cmovz %4,%%r15; prefetcht1 (%8); addq $16,%8;"\
  KERNEL_k1m8n##ndim KERNEL_k1m8n##ndim\
--- a/kernel/x86_64/sgemm_kernel_16x4_skylakex_2.c
+++ b/kernel/x86_64/sgemm_kernel_16x4_skylakex_2.c
@ -1,8 +1,152 @@
-/* %0 = "+r"(a_pointer), %1 = "+r"(b_pointer), %2 = "+r"(c_pointer), %3 = "+r"(ldc_in_bytes), %4 for k_count, %5 for c_store */
-/* r10 to assist prefetch, r12 = k << 4(const), r13 = k(const), r14 = b_head_pos(const), r15 = %1 + 3r12 */
+/* %0 = "+r"(a_pointer), %1 = "+r"(b_pointer), %2 = "+r"(c_pointer), %3 = "+r"(ldc_in_bytes), %4 for k_count, %5 for c_store, %6 = b_pref */
+/* r10 to assist prefetch, r11 = m_counter, r12 = k << 4(const), r13 = k_todo, r14 = b_head_pos(const), r15 = %1 + 3r12 */

 #include "common.h"
 #include <stdint.h>
+#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+  #define BACKWARDS 1
+#else
+  #define BACKWARDS 0
+#endif
+#define REC_POINTER_1(ptr) "salq $2,%%r13; subq %%r13,"#ptr"; sarq $2,%%r13;"
+#define REC_POINTER_2(ptr) "salq $3,%%r13; subq %%r13,"#ptr"; sarq $3,%%r13;"
+#define REC_POINTER_4(ptr) "salq $4,%%r13; subq %%r13,"#ptr"; sarq $4,%%r13;"
+#define REC_POINTER_8(ptr) "salq $5,%%r13; subq %%r13,"#ptr"; sarq $5,%%r13;"
+#define REC_POINTER_16(ptr) "salq $6,%%r13; subq %%r13,"#ptr"; sarq $6,%%r13;"
+#define INC_POINTER_1(ptr) "sarq $2,%%r12; addq %%r12,"#ptr"; salq $2,%%r12;"
+#define INC_POINTER_2(ptr) "sarq $1,%%r12; addq %%r12,"#ptr"; salq $1,%%r12;"
+#define INC_POINTER_4(ptr) "addq %%r12,"#ptr";"
+#define INC_POINTER_8(ptr) "leaq ("#ptr",%%r12,2),"#ptr";"
+#define INC_POINTER_16(ptr) "leaq ("#ptr",%%r12,4),"#ptr";"
+#define SET_POINTER(ptr,dim) REC_POINTER_##dim(ptr) INC_POINTER_##dim(ptr)
+#define SET_PB_1 SET_POINTER(%1,1)
+#define SET_PB_2 SET_POINTER(%1,2)
+#define SET_PB_4 SET_POINTER(%1,4)
+#define SET_PB_8 SET_POINTER(%1,4)
+#define SET_PB_12 SET_POINTER(%1,4)
+#define SET_PB_16 SET_POINTER(%1,4)
+#define SET_PB_20 SET_POINTER(%1,4)
+#define SET_PB_24 SET_POINTER(%1,4)
+#ifdef TRMMKERNEL
+  #if BACKWARDS == 1
+    #define START_SET_PAPB(mdim,ndim) SET_POINTER(%0,mdim) "movq %%r14,%1;" SET_PB_##ndim "leaq (%1,%%r12,2),%%r15; addq %%r12,%%r15;"
+    #define END_SET_PA(mdim) ""
+  #else
+    #define START_SET_PAPB(mdim,ndim) "movq %%r14,%1; leaq (%1,%%r12,2),%%r15; addq %%r12,%%r15;"
+    #define END_SET_PA(mdim) SET_POINTER(%0,mdim)
+  #endif
+#else
+  #define START_SET_PAPB(mdim,ndim) "movq %%r14,%1; leaq (%1,%%r12,2),%%r15; addq %%r12,%%r15;"
+  #define END_SET_PA(mdim) ""
+#endif
+#define RECOVER_PA(mdim) REC_POINTER_##mdim(%0)
+
+#if defined(TRMMKERNEL) && !defined(LEFT)
+  #if BACKWARDS == 1
+    #define KERNEL_HEAD_C_n8(mdim) \
+      KERNEL_k1m##mdim##n4 KERNEL_k1m##mdim##n4 KERNEL_k1m##mdim##n4 KERNEL_k1m##mdim##n4 "subq $4,%4; addq $64,%%r15;"
+    #define KERNEL_HEAD_C_n12(mdim) KERNEL_HEAD_C_n8(mdim)\
+      KERNEL_k1m##mdim##n8 KERNEL_k1m##mdim##n8 KERNEL_k1m##mdim##n8 KERNEL_k1m##mdim##n8 "subq $4,%4; addq $64,%%r15;"
+    #define KERNEL_HEAD_C_n16(mdim) KERNEL_HEAD_C_n12(mdim)\
+      KERNEL_k1m##mdim##n12 KERNEL_k1m##mdim##n12 KERNEL_k1m##mdim##n12 KERNEL_k1m##mdim##n12 "subq $4,%4; addq $64,%%r15;"
+    #define KERNEL_HEAD_C_n20(mdim) KERNEL_HEAD_C_n16(mdim)\
+      KERNEL_k1m##mdim##n16 KERNEL_k1m##mdim##n16 KERNEL_k1m##mdim##n16 KERNEL_k1m##mdim##n16 "subq $4,%4;"
+    #define KERNEL_HEAD_C_n24(mdim) KERNEL_HEAD_C_n20(mdim)\
+      KERNEL_k1m##mdim##n20 KERNEL_k1m##mdim##n20 KERNEL_k1m##mdim##n20 KERNEL_k1m##mdim##n20 "subq $4,%4;"
+    #define KERNEL_HEAD_R_n4(mdim) "subq $12,%4; addq $64,%%r15; addq $"#mdim"*48,%0;"
+    #define KERNEL_HEAD_R_n8(mdim) KERNEL_HEAD_R_n4(mdim)\
+      kernel_k1m##mdim##n4(%%r15) kernel_k1m##mdim##n4(%%r15) kernel_k1m##mdim##n4(%%r15) kernel_k1m##mdim##n4(%%r15) "subq $4,%4;"
+    #define KERNEL_HEAD_R_n12(mdim) KERNEL_HEAD_R_n8(mdim)\
+      kernel_k1m##mdim##n8(%%r15) kernel_k1m##mdim##n8(%%r15) kernel_k1m##mdim##n8(%%r15) kernel_k1m##mdim##n8(%%r15) "subq $4,%4;"
+    #define KERNEL_TAIL_C_n8(mdim) ""
+    #define KERNEL_TAIL_C_n12(mdim) ""
+    #define KERNEL_TAIL_C_n16(mdim) ""
+    #define KERNEL_TAIL_C_n20(mdim) ""
+    #define KERNEL_TAIL_C_n24(mdim) ""
+    #define KERNEL_TAIL_R_n4(mdim) ""
+    #define KERNEL_TAIL_R_n8(mdim) ""
+    #define KERNEL_TAIL_R_n12(mdim) ""
+  #else
+    #define KERNEL_HEAD_C_n8(mdim) ""
+    #define KERNEL_HEAD_C_n12(mdim) ""
+    #define KERNEL_HEAD_C_n16(mdim) ""
+    #define KERNEL_HEAD_C_n20(mdim) ""
+    #define KERNEL_HEAD_C_n24(mdim) ""
+    #define KERNEL_HEAD_R_n4(mdim) ""
+    #define KERNEL_HEAD_R_n8(mdim) ""
+    #define KERNEL_HEAD_R_n12(mdim) ""
+    #define end_kernel_k4_ncx1(k_0,k_1,k_2,k_3,n1,mdim) \
+      end_load_a_k1m##mdim(k_0) end_acc_nc##n1##_k1m##mdim(k_0)\
+      end_load_a_k1m##mdim(k_1) end_acc_nc##n1##_k1m##mdim(k_1)\
+      end_load_a_k1m##mdim(k_2) end_acc_nc##n1##_k1m##mdim(k_2)\
+      end_load_a_k1m##mdim(k_3) end_acc_nc##n1##_k1m##mdim(k_3)
+    #define end_kernel_k4_ncx2(k_0,k_1,k_2,k_3,n1,n2,mdim) \
+      end_load_a_k1m##mdim(k_0) end_acc_nc##n1##_k1m##mdim(k_0) end_acc_nc##n2##_k1m##mdim(k_0)\
+      end_load_a_k1m##mdim(k_1) end_acc_nc##n1##_k1m##mdim(k_1) end_acc_nc##n2##_k1m##mdim(k_1)\
+      end_load_a_k1m##mdim(k_2) end_acc_nc##n1##_k1m##mdim(k_2) end_acc_nc##n2##_k1m##mdim(k_2)\
+      end_load_a_k1m##mdim(k_3) end_acc_nc##n1##_k1m##mdim(k_3) end_acc_nc##n2##_k1m##mdim(k_3)
+    #define end_kernel_k4_ncx3(k_0,k_1,k_2,k_3,n1,n2,n3,mdim) \
+      end_load_a_k1m##mdim(k_0) end_acc_nc##n1##_k1m##mdim(k_0) end_acc_nc##n2##_k1m##mdim(k_0) end_acc_nc##n3##_k1m##mdim(k_0)\
+      end_load_a_k1m##mdim(k_1) end_acc_nc##n1##_k1m##mdim(k_1) end_acc_nc##n2##_k1m##mdim(k_1) end_acc_nc##n3##_k1m##mdim(k_1)\
+      end_load_a_k1m##mdim(k_2) end_acc_nc##n1##_k1m##mdim(k_2) end_acc_nc##n2##_k1m##mdim(k_2) end_acc_nc##n3##_k1m##mdim(k_2)\
+      end_load_a_k1m##mdim(k_3) end_acc_nc##n1##_k1m##mdim(k_3) end_acc_nc##n2##_k1m##mdim(k_3) end_acc_nc##n3##_k1m##mdim(k_3)
+    #define end_kernel_k4_ncx4(k_0,k_1,k_2,k_3,n1,n2,n3,n4,mdim) \
+      end_load_a_k1m##mdim(k_0) end_acc_nc##n1##_k1m##mdim(k_0) end_acc_nc##n2##_k1m##mdim(k_0) end_acc_nc##n3##_k1m##mdim(k_0) end_acc_nc##n4##_k1m##mdim(k_0)\
+      end_load_a_k1m##mdim(k_1) end_acc_nc##n1##_k1m##mdim(k_1) end_acc_nc##n2##_k1m##mdim(k_1) end_acc_nc##n3##_k1m##mdim(k_1) end_acc_nc##n4##_k1m##mdim(k_1)\
+      end_load_a_k1m##mdim(k_2) end_acc_nc##n1##_k1m##mdim(k_2) end_acc_nc##n2##_k1m##mdim(k_2) end_acc_nc##n3##_k1m##mdim(k_2) end_acc_nc##n4##_k1m##mdim(k_2)\
+      end_load_a_k1m##mdim(k_3) end_acc_nc##n1##_k1m##mdim(k_3) end_acc_nc##n2##_k1m##mdim(k_3) end_acc_nc##n3##_k1m##mdim(k_3) end_acc_nc##n4##_k1m##mdim(k_3)
+    #define end_kernel_k4_ncx5(k_0,k_1,k_2,k_3,n1,n2,n3,n4,n5,mdim) \
+      end_load_a_k1m##mdim(k_0)     end_acc_nc##n1##_k1m##mdim(k_0) end_acc_nc##n2##_k1m##mdim(k_0)\
+      end_acc_nc##n3##_k1m##mdim(k_0) end_acc_nc##n4##_k1m##mdim(k_0) end_acc_nc##n5##_k1m##mdim(k_0)\
+      end_load_a_k1m##mdim(k_1)     end_acc_nc##n1##_k1m##mdim(k_1) end_acc_nc##n2##_k1m##mdim(k_1)\
+      end_acc_nc##n3##_k1m##mdim(k_1) end_acc_nc##n4##_k1m##mdim(k_1) end_acc_nc##n5##_k1m##mdim(k_1)\
+      end_load_a_k1m##mdim(k_2)     end_acc_nc##n1##_k1m##mdim(k_2) end_acc_nc##n2##_k1m##mdim(k_2)\
+      end_acc_nc##n3##_k1m##mdim(k_2) end_acc_nc##n4##_k1m##mdim(k_2) end_acc_nc##n5##_k1m##mdim(k_2)\
+      end_load_a_k1m##mdim(k_3)     end_acc_nc##n1##_k1m##mdim(k_3) end_acc_nc##n2##_k1m##mdim(k_3)\
+      end_acc_nc##n3##_k1m##mdim(k_3) end_acc_nc##n4##_k1m##mdim(k_3) end_acc_nc##n5##_k1m##mdim(k_3)
+    #define KERNEL_TAIL_C_n8(mdim) end_kernel_k4_ncx1(0,1,2,3,2,mdim)
+    #define KERNEL_TAIL_C_n12(mdim) \
+      end_kernel_k4_ncx2(0,1,2,3,2,3,mdim) end_kernel_k4_ncx1(4,5,6,7,3,mdim)
+    #define KERNEL_TAIL_C_n16(mdim) \
+      end_kernel_k4_ncx3(0,1,2,3,2,3,4,mdim) end_kernel_k4_ncx2(4,5,6,7,3,4,mdim) end_kernel_k4_ncx1(8,9,10,11,4,mdim)
+    #define KERNEL_TAIL_C_n20(mdim) \
+      end_kernel_k4_ncx4(0,1,2,3,2,3,4,5,mdim) end_kernel_k4_ncx3(4,5,6,7,3,4,5,mdim)\
+      end_kernel_k4_ncx2(8,9,10,11,4,5,mdim) end_kernel_k4_ncx1(12,13,14,15,5,mdim)
+    #define KERNEL_TAIL_C_n24(mdim) \
+      end_kernel_k4_ncx5(0,1,2,3,2,3,4,5,6,mdim) end_kernel_k4_ncx4(4,5,6,7,3,4,5,6,mdim) end_kernel_k4_ncx3(8,9,10,11,4,5,6,mdim)\
+      end_kernel_k4_ncx2(12,13,14,15,5,6,mdim) end_kernel_k4_ncx1(16,17,18,19,6,mdim)
+    #define KERNEL_TAIL_R_n4(mdim) \
+      end_kernel_k4_ncx1(0,1,2,3,4,mdim) end_kernel_k4_ncx1(4,5,6,7,4,mdim) end_kernel_k4_ncx1(8,9,10,11,4,mdim)
+    #define KERNEL_TAIL_R_n8(mdim) \
+      end_kernel_k4_ncx2(0,1,2,3,4,5,mdim) end_kernel_k4_ncx2(4,5,6,7,4,5,mdim) end_kernel_k4_ncx2(8,9,10,11,4,5,mdim) end_kernel_k4_ncx1(12,13,14,15,5,mdim)
+    #define KERNEL_TAIL_R_n12(mdim) \
+      end_kernel_k4_ncx3(0,1,2,3,4,5,6,mdim) end_kernel_k4_ncx3(4,5,6,7,4,5,6,mdim) end_kernel_k4_ncx3(8,9,10,11,4,5,6,mdim)\
+      end_kernel_k4_ncx2(12,13,14,15,5,6,mdim) end_kernel_k4_ncx1(16,17,18,19,6,mdim)
+  #endif
+#else
+  #define KERNEL_HEAD_C_n8(mdim) ""
+  #define KERNEL_HEAD_C_n12(mdim) ""
+  #define KERNEL_HEAD_C_n16(mdim) ""
+  #define KERNEL_HEAD_C_n20(mdim) ""
+  #define KERNEL_HEAD_C_n24(mdim) ""
+  #define KERNEL_HEAD_R_n4(mdim) ""
+  #define KERNEL_HEAD_R_n8(mdim) ""
+  #define KERNEL_HEAD_R_n12(mdim) ""
+  #define KERNEL_TAIL_C_n8(mdim) ""
+  #define KERNEL_TAIL_C_n12(mdim) ""
+  #define KERNEL_TAIL_C_n16(mdim) ""
+  #define KERNEL_TAIL_C_n20(mdim) ""
+  #define KERNEL_TAIL_C_n24(mdim) ""
+  #define KERNEL_TAIL_R_n4(mdim) ""
+  #define KERNEL_TAIL_R_n8(mdim) ""
+  #define KERNEL_TAIL_R_n12(mdim) ""
+#endif
+#define KERNEL_HEAD_C_n1(mdim) ""
+#define KERNEL_HEAD_C_n2(mdim) ""
+#define KERNEL_HEAD_C_n4(mdim) ""
+#define KERNEL_TAIL_C_n1(mdim) ""
+#define KERNEL_TAIL_C_n2(mdim) ""
+#define KERNEL_TAIL_C_n4(mdim) ""

 /* m = 16 */ /* zmm8-zmm31 for accumulators, zmm1-zmm7 for temporary use, zmm0 for alpha */
 #define KERNEL_k1m16n1 \
@ -15,9 +159,10 @@
 #define KERNEL_k1m16n2 KERNEL_h_k1m16n2 "addq $8,%1;"
 #define KERNEL_h_k1m16n4 KERNEL_h_k1m16n2 "vbroadcastsd 8(%1),%%zmm7; vfmadd231ps %%zmm4,%%zmm7,%%zmm10; vfmadd231ps %%zmm5,%%zmm7,%%zmm11;"
 #define KERNEL_k1m16n4 KERNEL_h_k1m16n4 "addq $16,%1;"
-#define unit_kernel_k1m16n4(c1,c2,c3,c4, ...) \
-    "vbroadcastsd  ("#__VA_ARGS__"),%%zmm6; vfmadd231ps %%zmm4,%%zmm6,"#c1"; vfmadd231ps %%zmm5,%%zmm6,"#c2";"\
-    "vbroadcastsd 8("#__VA_ARGS__"),%%zmm7; vfmadd231ps %%zmm4,%%zmm7,"#c3"; vfmadd231ps %%zmm5,%%zmm7,"#c4";"
+#define unit_gen_kernel_k1m16n4(c1,c2,c3,c4,k_no,...) \
+    "vbroadcastsd "#k_no"*16  ("#__VA_ARGS__"),%%zmm6; vfmadd231ps %%zmm4,%%zmm6,"#c1"; vfmadd231ps %%zmm5,%%zmm6,"#c2";"\
+    "vbroadcastsd "#k_no"*16+8("#__VA_ARGS__"),%%zmm7; vfmadd231ps %%zmm4,%%zmm7,"#c3"; vfmadd231ps %%zmm5,%%zmm7,"#c4";"
+#define unit_kernel_k1m16n4(c1,c2,c3,c4, ...) unit_gen_kernel_k1m16n4(c1,c2,c3,c4,0,__VA_ARGS__)
 #define KERNEL_h_k1m16n8 KERNEL_h_k1m16n4 unit_kernel_k1m16n4(%%zmm12,%%zmm13,%%zmm14,%%zmm15,%1,%%r12,1)
 #define KERNEL_k1m16n8 KERNEL_h_k1m16n8 "addq $16,%1;"
 #define KERNEL_h_k1m16n12 KERNEL_h_k1m16n8 unit_kernel_k1m16n4(%%zmm16,%%zmm17,%%zmm18,%%zmm19,%1,%%r12,2)
@ -28,6 +173,12 @@
 #define KERNEL_k1m16n20 KERNEL_h_k1m16n20 "addq $16,%%r15;"
 #define KERNEL_h_k1m16n24 KERNEL_h_k1m16n20 unit_kernel_k1m16n4(%%zmm28,%%zmm29,%%zmm30,%%zmm31,%%r15,%%r12,2)
 #define KERNEL_k1m16n24 KERNEL_h_k1m16n24 "addq $16,%%r15;"
+#define end_load_a_k1m16(k_no) "vmovsldup "#k_no"*64(%0),%%zmm4; vmovshdup "#k_no"*64(%0),%%zmm5;"
+#define end_acc_nc2_k1m16(k_no) unit_gen_kernel_k1m16n4(%%zmm12,%%zmm13,%%zmm14,%%zmm15,k_no,%1,%%r12,1)
+#define end_acc_nc3_k1m16(k_no) unit_gen_kernel_k1m16n4(%%zmm16,%%zmm17,%%zmm18,%%zmm19,k_no,%1,%%r12,2)
+#define end_acc_nc4_k1m16(k_no) unit_gen_kernel_k1m16n4(%%zmm20,%%zmm21,%%zmm22,%%zmm23,k_no,%%r15)
+#define end_acc_nc5_k1m16(k_no) unit_gen_kernel_k1m16n4(%%zmm24,%%zmm25,%%zmm26,%%zmm27,k_no,%%r15,%%r12,1)
+#define end_acc_nc6_k1m16(k_no) unit_gen_kernel_k1m16n4(%%zmm28,%%zmm29,%%zmm30,%%zmm31,k_no,%%r15,%%r12,2)
 #define INIT_m16n1 "vpxorq %%zmm8,%%zmm8,%%zmm8;"
 #define INIT_m16n2 INIT_m16n1 "vpxorq %%zmm9,%%zmm9,%%zmm9;"
 #define INIT_m16n4 INIT_m16n2 "vpxorq %%zmm10,%%zmm10,%%zmm10;vpxorq %%zmm11,%%zmm11,%%zmm11;"
@ -38,11 +189,19 @@
 #define INIT_m16n16 INIT_m16n12 unit_init_m16n4(%%zmm20,%%zmm21,%%zmm22,%%zmm23)
 #define INIT_m16n20 INIT_m16n16 unit_init_m16n4(%%zmm24,%%zmm25,%%zmm26,%%zmm27)
 #define INIT_m16n24 INIT_m16n20 unit_init_m16n4(%%zmm28,%%zmm29,%%zmm30,%%zmm31)
-#define SAVE_h_m16n1 "vfmadd213ps (%2),%%zmm0,%%zmm8; vmovups %%zmm8,(%2);"
-#define unit_save_m16n2(c1,c2) \
+#ifdef TRMMKERNEL
+  #define SAVE_h_m16n1 "vmulps %%zmm8,%%zmm0,%%zmm8; vmovups %%zmm8,(%2);"
+  #define unit_save_m16n2(c1,c2) \
+    "vunpcklps "#c2","#c1",%%zmm6; vunpckhps "#c2","#c1",%%zmm7; vunpcklpd %%zmm7,%%zmm6,%%zmm4; vunpckhpd %%zmm7,%%zmm6,%%zmm5;"\
+    "vmulps %%zmm4,%%zmm0,%%zmm4; vmulps %%zmm5,%%zmm0,%%zmm5;"\
+    "vmovups %%zmm4,(%5); vmovups %%zmm5,(%5,%3,1); leaq (%5,%3,2),%5;"
+#else
+  #define SAVE_h_m16n1 "vfmadd213ps (%2),%%zmm0,%%zmm8; vmovups %%zmm8,(%2);"
+  #define unit_save_m16n2(c1,c2) \
    "vunpcklps "#c2","#c1",%%zmm6; vunpckhps "#c2","#c1",%%zmm7; vunpcklpd %%zmm7,%%zmm6,%%zmm4; vunpckhpd %%zmm7,%%zmm6,%%zmm5;"\
    "vfmadd213ps (%5),%%zmm0,%%zmm4; vfmadd213ps (%5,%3,1),%%zmm0,%%zmm5;"\
    "vmovups %%zmm4,(%5); vmovups %%zmm5,(%5,%3,1); leaq (%5,%3,2),%5;"
+#endif
 #define SAVE_h_m16n2 "movq %2,%5;" unit_save_m16n2(%%zmm8,%%zmm9)
 #define SAVE_h_m16n4  SAVE_h_m16n2  unit_save_m16n2(%%zmm10,%%zmm11)
 #define SAVE_h_m16n8  SAVE_h_m16n4  unit_save_m16n2(%%zmm12,%%zmm13) unit_save_m16n2(%%zmm14,%%zmm15)
@ -52,8 +211,9 @@
 #define SAVE_h_m16n24 SAVE_h_m16n20 unit_save_m16n2(%%zmm28,%%zmm29) unit_save_m16n2(%%zmm30,%%zmm31)
 #define SAVE_m16(ndim) SAVE_h_m16n##ndim "addq $64,%2;"
 #define COMPUTE_m16(ndim) \
-    INIT_m16n##ndim\
-    "movq %%r13,%4; movq %%r14,%1; leaq (%1,%%r12,2),%%r15; addq %%r12,%%r15; movq %2,%5; xorq %%r10,%%r10;"\
+    INIT_m16n##ndim START_SET_PAPB(16,ndim)\
+    "movq %%r13,%4; movq %2,%5; xorq %%r10,%%r10;"\
+    KERNEL_HEAD_C_n##ndim(16)\
    "cmpq $16,%4; jb "#ndim"016162f;"\
    #ndim"016161:\n\t"\
    "cmpq $126,%%r10; movq $126,%%r10; cmoveq %3,%%r10;"\
@ -72,28 +232,41 @@
    KERNEL_k1m16n##ndim\
    "leaq (%5,%3,2),%5; decq %4; jnz "#ndim"016163b;"\
    #ndim"016164:\n\t"\
+    KERNEL_TAIL_C_n##ndim(16)\
    "prefetcht0 (%%r14); prefetcht0 64(%%r14);"\
-    SAVE_m16(ndim)
+    SAVE_m16(ndim) END_SET_PA(16)

 /* m = 8 *//* ymm0 for alpha, ymm1-ymm3 for temporary use, ymm4-ymm15 for accumulators */
-#define KERNEL_k1m8n1(b_addr) \
+#define kernel_k1m8n1(b_addr) \
    "vmovups (%0),%%ymm1; addq $32,%0;"\
    "vbroadcastss ("#b_addr"),%%ymm2; vfmadd231ps %%ymm1,%%ymm2,%%ymm4;"\
    "addq $4,"#b_addr";"
-#define KERNEL_h_k1m8n2(b_addr) \
+#define kernel_h_k1m8n2(b_addr) \
    "vmovsldup (%0),%%ymm1; vmovshdup (%0),%%ymm2; addq $32,%0;"\
    "vbroadcastsd ("#b_addr"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4; vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"
-#define KERNEL_k1m8n2(b_addr) KERNEL_h_k1m8n2(b_addr) "addq $8,"#b_addr";"
-#define KERNEL_h_k1m8n4(b_addr) \
-    KERNEL_h_k1m8n2(b_addr) "vbroadcastsd 8("#b_addr"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6; vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"
-#define KERNEL_k1m8n4(b_addr) KERNEL_h_k1m8n4(b_addr) "addq $16,"#b_addr";"
-#define unit_kernel_k1m8n4(c1,c2,c3,c4,...) \
-    "vbroadcastsd  ("#__VA_ARGS__"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,"#c1"; vfmadd231ps %%ymm2,%%ymm3,"#c2";"\
-    "vbroadcastsd 8("#__VA_ARGS__"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,"#c3"; vfmadd231ps %%ymm2,%%ymm3,"#c4";"
-#define KERNEL_h_k1m8n8(b_addr) KERNEL_h_k1m8n4(b_addr) unit_kernel_k1m8n4(%%ymm8,%%ymm9,%%ymm10,%%ymm11,b_addr,%%r12,1)
-#define KERNEL_k1m8n8(b_addr) KERNEL_h_k1m8n8(b_addr) "addq $16,"#b_addr";"
-#define KERNEL_h_k1m8n12(b_addr) KERNEL_h_k1m8n8(b_addr) unit_kernel_k1m8n4(%%ymm12,%%ymm13,%%ymm14,%%ymm15,b_addr,%%r12,2)
-#define KERNEL_k1m8n12(b_addr) KERNEL_h_k1m8n12(b_addr) "addq $16,"#b_addr";"
+#define kernel_k1m8n2(b_addr) kernel_h_k1m8n2(b_addr) "addq $8,"#b_addr";"
+#define kernel_h_k1m8n4(b_addr) \
+    kernel_h_k1m8n2(b_addr) "vbroadcastsd 8("#b_addr"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6; vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"
+#define kernel_k1m8n4(b_addr) kernel_h_k1m8n4(b_addr) "addq $16,"#b_addr";"
+#define unit_gen_kernel_k1m8n4(c1,c2,c3,c4,k_no,...) \
+    "vbroadcastsd "#k_no"*16  ("#__VA_ARGS__"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,"#c1"; vfmadd231ps %%ymm2,%%ymm3,"#c2";"\
+    "vbroadcastsd "#k_no"*16+8("#__VA_ARGS__"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,"#c3"; vfmadd231ps %%ymm2,%%ymm3,"#c4";"
+#define unit_kernel_k1m8n4(c1,c2,c3,c4,...) unit_gen_kernel_k1m8n4(c1,c2,c3,c4,0,__VA_ARGS__)
+#define kernel_h_k1m8n8(b_addr) kernel_h_k1m8n4(b_addr) unit_kernel_k1m8n4(%%ymm8,%%ymm9,%%ymm10,%%ymm11,b_addr,%%r12,1)
+#define kernel_k1m8n8(b_addr) kernel_h_k1m8n8(b_addr) "addq $16,"#b_addr";"
+#define kernel_h_k1m8n12(b_addr) kernel_h_k1m8n8(b_addr) unit_kernel_k1m8n4(%%ymm12,%%ymm13,%%ymm14,%%ymm15,b_addr,%%r12,2)
+#define kernel_k1m8n12(b_addr) kernel_h_k1m8n12(b_addr) "addq $16,"#b_addr";"
+#define KERNEL_k1m8n1 kernel_k1m8n1(%1)
+#define KERNEL_k1m8n2 kernel_k1m8n2(%1)
+#define KERNEL_k1m8n4 kernel_k1m8n4(%1)
+#define KERNEL_k1m8n8 kernel_k1m8n8(%1)
+#define KERNEL_k1m8n12 kernel_k1m8n12(%1)
+#define end_load_a_k1m8(k_no) "vmovsldup "#k_no"*32(%0),%%ymm1; vmovshdup "#k_no"*32(%0),%%ymm2;"
+#define end_acc_nc2_k1m8(k_no) unit_gen_kernel_k1m8n4(%%ymm8,%%ymm9,%%ymm10,%%ymm11,k_no,%1,%%r12,1)
+#define end_acc_nc3_k1m8(k_no) unit_gen_kernel_k1m8n4(%%ymm12,%%ymm13,%%ymm14,%%ymm15,k_no,%1,%%r12,2)
+#define end_acc_nc4_k1m8(k_no) unit_gen_kernel_k1m8n4(%%ymm4,%%ymm5,%%ymm6,%%ymm7,k_no,%%r15)
+#define end_acc_nc5_k1m8(k_no) unit_gen_kernel_k1m8n4(%%ymm8,%%ymm9,%%ymm10,%%ymm11,k_no,%%r15,%%r12,1)
+#define end_acc_nc6_k1m8(k_no) unit_gen_kernel_k1m8n4(%%ymm12,%%ymm13,%%ymm14,%%ymm15,k_no,%%r15,%%r12,2)
 #define INIT_m8n1 "vpxor %%ymm4,%%ymm4,%%ymm4;"
 #define INIT_m8n2 INIT_m8n1 "vpxor %%ymm5,%%ymm5,%%ymm5;"
 #define INIT_m8n4 INIT_m8n2 "vpxor %%ymm6,%%ymm6,%%ymm6;vpxor %%ymm7,%%ymm7,%%ymm7;"
@ -101,12 +274,21 @@
    "vpxor "#c1","#c1","#c1";vpxor "#c2","#c2","#c2";vpxor "#c3","#c3","#c3";vpxor "#c4","#c4","#c4";"
 #define INIT_m8n8  INIT_m8n4 unit_init_m8n4(%%ymm8,%%ymm9,%%ymm10,%%ymm11)
 #define INIT_m8n12 INIT_m8n8 unit_init_m8n4(%%ymm12,%%ymm13,%%ymm14,%%ymm15)
-#define SAVE_L_m8n1 "vfmadd213ps (%2),%%ymm0,%%ymm4; vmovups %%ymm4,(%2);"
-#define unit_save_m8n2(c1,c2) \
+#ifdef TRMMKERNEL
+  #define SAVE_L_m8n1 "vmulps %%ymm4,%%ymm0,%%ymm4; vmovups %%ymm4,(%2);"
+  #define unit_save_m8n2(c1,c2) \
    "vunpcklps "#c2","#c1",%%ymm2; vunpckhps "#c2","#c1",%%ymm3;"\
-    "vunpcklpd %%ymm3,%%ymm2,%%ymm1;vfmadd213ps (%5),     %%ymm0,%%ymm1;vmovups %%ymm1,(%5);"\
-    "vunpckhpd %%ymm3,%%ymm2,%%ymm1;vfmadd213ps (%5,%3,1),%%ymm0,%%ymm1;vmovups %%ymm1,(%5,%3,1);"\
+    "vunpcklpd %%ymm3,%%ymm2,%%ymm1; vmulps %%ymm1,%%ymm0,%%ymm1; vmovups %%ymm1,(%5);"\
+    "vunpckhpd %%ymm3,%%ymm2,%%ymm1; vmulps %%ymm1,%%ymm0,%%ymm1; vmovups %%ymm1,(%5,%3,1);"\
    "leaq (%5,%3,2),%5;"
+#else
+  #define SAVE_L_m8n1 "vfmadd213ps (%2),%%ymm0,%%ymm4; vmovups %%ymm4,(%2);"
+  #define unit_save_m8n2(c1,c2) \
+    "vunpcklps "#c2","#c1",%%ymm2; vunpckhps "#c2","#c1",%%ymm3;"\
+    "vunpcklpd %%ymm3,%%ymm2,%%ymm1; vfmadd213ps (%5),     %%ymm0,%%ymm1; vmovups %%ymm1,(%5);"\
+    "vunpckhpd %%ymm3,%%ymm2,%%ymm1; vfmadd213ps (%5,%3,1),%%ymm0,%%ymm1; vmovups %%ymm1,(%5,%3,1);"\
+    "leaq (%5,%3,2),%5;"
+#endif
 #define SAVE_L_m8n2 "movq %2,%5;" unit_save_m8n2(%%ymm4,%%ymm5)
 #define SAVE_L_m8n4  SAVE_L_m8n2  unit_save_m8n2(%%ymm6,%%ymm7)
 #define SAVE_L_m8n8  SAVE_L_m8n4  unit_save_m8n2(%%ymm8,%%ymm9)   unit_save_m8n2(%%ymm10,%%ymm11)
@ -115,53 +297,68 @@
 #define SAVE_R_m8n8  SAVE_R_m8n4  unit_save_m8n2(%%ymm8,%%ymm9)   unit_save_m8n2(%%ymm10,%%ymm11)
 #define SAVE_R_m8n12 SAVE_R_m8n8  unit_save_m8n2(%%ymm12,%%ymm13) unit_save_m8n2(%%ymm14,%%ymm15)
 #define COMPUTE_L_m8(ndim,sim) \
-    INIT_m8n##ndim\
-    "movq %%r13,%4; movq %%r14,%1;"\
-    #ndim""#sim"882:\n\t"\
+    INIT_m8n##ndim START_SET_PAPB(8,ndim)\
+    "movq %%r13,%4;"\
+    KERNEL_HEAD_C_n##ndim(8)\
    "testq %4,%4; jz "#ndim""#sim"883f;"\
-    KERNEL_k1m8n##ndim(%1)\
-    "decq %4; jmp "#ndim""#sim"882b;"\
+    #ndim""#sim"882:\n\t"\
+    kernel_k1m8n##ndim(%1)\
+    "decq %4; jnz "#ndim""#sim"882b;"\
    #ndim""#sim"883:\n\t"\
+    KERNEL_TAIL_C_n##ndim(8)\
    SAVE_L_m8n##ndim "addq $32,%2;"
 #define COMPUTE_R_m8(ndim,sim) \
-    "subq %%r12,%0; subq %%r12,%0;"\
-    INIT_m8n##ndim\
-    "movq %%r13,%4; leaq (%%r14,%%r12,2),%%r15; addq %%r12,%%r15;"\
-    #ndim""#sim"882:\n\t"\
+    INIT_m8n##ndim RECOVER_PA(8)\
+    "movq %%r13,%4;"\
+    KERNEL_HEAD_R_n##ndim(8)\
    "testq %4,%4; jz "#ndim""#sim"883f;"\
-    KERNEL_k1m8n##ndim(%%r15)\
-    "decq %4; jmp "#ndim""#sim"882b;"\
+    #ndim""#sim"882:\n\t"\
+    kernel_k1m8n##ndim(%%r15)\
+    "decq %4; jnz "#ndim""#sim"882b;"\
    #ndim""#sim"883:\n\t"\
-    SAVE_R_m8n##ndim
-#define COMPUTE_m8_n1  COMPUTE_L_m8(1,33833)
-#define COMPUTE_m8_n2  COMPUTE_L_m8(2,33833)
-#define COMPUTE_m8_n4  COMPUTE_L_m8(4,33833)
-#define COMPUTE_m8_n8  COMPUTE_L_m8(8,33833)
-#define COMPUTE_m8_n12 COMPUTE_L_m8(12,33833)
+    KERNEL_TAIL_R_n##ndim(8)\
+    SAVE_R_m8n##ndim END_SET_PA(8)
+#define COMPUTE_m8_n1  COMPUTE_L_m8(1,33833) END_SET_PA(8)
+#define COMPUTE_m8_n2  COMPUTE_L_m8(2,33833) END_SET_PA(8)
+#define COMPUTE_m8_n4  COMPUTE_L_m8(4,33833) END_SET_PA(8)
+#define COMPUTE_m8_n8  COMPUTE_L_m8(8,33833) END_SET_PA(8)
+#define COMPUTE_m8_n12 COMPUTE_L_m8(12,33833) END_SET_PA(8)
 #define COMPUTE_m8_n16 COMPUTE_L_m8(12,33733) COMPUTE_R_m8(4,33933)
 #define COMPUTE_m8_n20 COMPUTE_L_m8(12,33633) COMPUTE_R_m8(8,33933)
 #define COMPUTE_m8_n24 COMPUTE_L_m8(12,33533) COMPUTE_R_m8(12,33933)
 #define COMPUTE_m8(ndim) COMPUTE_m8_n##ndim

 /* m = 4 *//* xmm0 for alpha, xmm1-xmm3 for temporary use, xmm4-xmm15 for accumulators */
-#define KERNEL_k1m4n1(b_addr) \
+#define kernel_k1m4n1(b_addr) \
    "vmovups (%0),%%xmm1; addq $16,%0;"\
    "vbroadcastss ("#b_addr"),%%xmm2; vfmadd231ps %%xmm1,%%xmm2,%%xmm4;"\
    "addq $4,"#b_addr";"
-#define KERNEL_h_k1m4n2(b_addr) \
+#define kernel_h_k1m4n2(b_addr) \
    "vmovsldup (%0),%%xmm1; vmovshdup (%0),%%xmm2; addq $16,%0;"\
    "vmovddup ("#b_addr"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm4; vfmadd231ps %%xmm2,%%xmm3,%%xmm5;"
-#define KERNEL_k1m4n2(b_addr) KERNEL_h_k1m4n2(b_addr) "addq $8,"#b_addr";"
-#define KERNEL_h_k1m4n4(b_addr) \
-    KERNEL_h_k1m4n2(b_addr) "vmovddup 8("#b_addr"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm6; vfmadd231ps %%xmm2,%%xmm3,%%xmm7;"
-#define KERNEL_k1m4n4(b_addr) KERNEL_h_k1m4n4(b_addr) "addq $16,"#b_addr";"
-#define unit_kernel_k1m4n4(c1,c2,c3,c4,...) \
-    "vmovddup  ("#__VA_ARGS__"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,"#c1"; vfmadd231ps %%xmm2,%%xmm3,"#c2";"\
-    "vmovddup 8("#__VA_ARGS__"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,"#c3"; vfmadd231ps %%xmm2,%%xmm3,"#c4";"
-#define KERNEL_h_k1m4n8(b_addr) KERNEL_h_k1m4n4(b_addr) unit_kernel_k1m4n4(%%xmm8,%%xmm9,%%xmm10,%%xmm11,b_addr,%%r12,1)
-#define KERNEL_k1m4n8(b_addr) KERNEL_h_k1m4n8(b_addr) "addq $16,"#b_addr";"
-#define KERNEL_h_k1m4n12(b_addr) KERNEL_h_k1m4n8(b_addr) unit_kernel_k1m4n4(%%xmm12,%%xmm13,%%xmm14,%%xmm15,b_addr,%%r12,2)
-#define KERNEL_k1m4n12(b_addr) KERNEL_h_k1m4n12(b_addr) "addq $16,"#b_addr";"
+#define kernel_k1m4n2(b_addr) kernel_h_k1m4n2(b_addr) "addq $8,"#b_addr";"
+#define kernel_h_k1m4n4(b_addr) \
+    kernel_h_k1m4n2(b_addr) "vmovddup 8("#b_addr"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm6; vfmadd231ps %%xmm2,%%xmm3,%%xmm7;"
+#define kernel_k1m4n4(b_addr) kernel_h_k1m4n4(b_addr) "addq $16,"#b_addr";"
+#define unit_gen_kernel_k1m4n4(c1,c2,c3,c4,k_no,...) \
+    "vmovddup "#k_no"*16  ("#__VA_ARGS__"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,"#c1"; vfmadd231ps %%xmm2,%%xmm3,"#c2";"\
+    "vmovddup "#k_no"*16+8("#__VA_ARGS__"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,"#c3"; vfmadd231ps %%xmm2,%%xmm3,"#c4";"
+#define unit_kernel_k1m4n4(c1,c2,c3,c4,...) unit_gen_kernel_k1m4n4(c1,c2,c3,c4,0,__VA_ARGS__)
+#define kernel_h_k1m4n8(b_addr) kernel_h_k1m4n4(b_addr) unit_kernel_k1m4n4(%%xmm8,%%xmm9,%%xmm10,%%xmm11,b_addr,%%r12,1)
+#define kernel_k1m4n8(b_addr) kernel_h_k1m4n8(b_addr) "addq $16,"#b_addr";"
+#define kernel_h_k1m4n12(b_addr) kernel_h_k1m4n8(b_addr) unit_kernel_k1m4n4(%%xmm12,%%xmm13,%%xmm14,%%xmm15,b_addr,%%r12,2)
+#define kernel_k1m4n12(b_addr) kernel_h_k1m4n12(b_addr) "addq $16,"#b_addr";"
+#define KERNEL_k1m4n1 kernel_k1m4n1(%1)
+#define KERNEL_k1m4n2 kernel_k1m4n2(%1)
+#define KERNEL_k1m4n4 kernel_k1m4n4(%1)
+#define KERNEL_k1m4n8 kernel_k1m4n8(%1)
+#define KERNEL_k1m4n12 kernel_k1m4n12(%1)
+#define end_load_a_k1m4(k_no) "vmovsldup "#k_no"*16(%0),%%xmm1; vmovshdup "#k_no"*16(%0),%%xmm2;"
+#define end_acc_nc2_k1m4(k_no) unit_gen_kernel_k1m4n4(%%xmm8,%%xmm9,%%xmm10,%%xmm11,k_no,%1,%%r12,1)
+#define end_acc_nc3_k1m4(k_no) unit_gen_kernel_k1m4n4(%%xmm12,%%xmm13,%%xmm14,%%xmm15,k_no,%1,%%r12,2)
+#define end_acc_nc4_k1m4(k_no) unit_gen_kernel_k1m4n4(%%xmm4,%%xmm5,%%xmm6,%%xmm7,k_no,%%r15)
+#define end_acc_nc5_k1m4(k_no) unit_gen_kernel_k1m4n4(%%xmm8,%%xmm9,%%xmm10,%%xmm11,k_no,%%r15,%%r12,1)
+#define end_acc_nc6_k1m4(k_no) unit_gen_kernel_k1m4n4(%%xmm12,%%xmm13,%%xmm14,%%xmm15,k_no,%%r15,%%r12,2)
 #define INIT_m4n1 "vpxor %%xmm4,%%xmm4,%%xmm4;"
 #define INIT_m4n2 INIT_m4n1 "vpxor %%xmm5,%%xmm5,%%xmm5;"
 #define INIT_m4n4 INIT_m4n2 "vpxor %%xmm6,%%xmm6,%%xmm6;vpxor %%xmm7,%%xmm7,%%xmm7;"
@ -169,12 +366,21 @@
    "vpxor "#c1","#c1","#c1";vpxor "#c2","#c2","#c2";vpxor "#c3","#c3","#c3";vpxor "#c4","#c4","#c4";"
 #define INIT_m4n8  INIT_m4n4 unit_init_m4n4(%%xmm8,%%xmm9,%%xmm10,%%xmm11)
 #define INIT_m4n12 INIT_m4n8 unit_init_m4n4(%%xmm12,%%xmm13,%%xmm14,%%xmm15)
-#define SAVE_L_m4n1 "vfmadd213ps (%2),%%xmm0,%%xmm4; vmovups %%xmm4,(%2);"
-#define unit_save_m4n2(c1,c2) \
+#ifdef TRMMKERNEL
+  #define SAVE_L_m4n1 "vmulps %%xmm4,%%xmm0,%%xmm4; vmovups %%xmm4,(%2);"
+  #define unit_save_m4n2(c1,c2) \
+    "vunpcklps "#c2","#c1",%%xmm2; vunpckhps "#c2","#c1",%%xmm3;"\
+    "vunpcklpd %%xmm3,%%xmm2,%%xmm1;vmulps %%xmm1,%%xmm0,%%xmm1;vmovups %%xmm1,(%5);"\
+    "vunpckhpd %%xmm3,%%xmm2,%%xmm1;vmulps %%xmm1,%%xmm0,%%xmm1;vmovups %%xmm1,(%5,%3,1);"\
+    "leaq (%5,%3,2),%5;"
+#else
+  #define SAVE_L_m4n1 "vfmadd213ps (%2),%%xmm0,%%xmm4; vmovups %%xmm4,(%2);"
+  #define unit_save_m4n2(c1,c2) \
    "vunpcklps "#c2","#c1",%%xmm2; vunpckhps "#c2","#c1",%%xmm3;"\
    "vunpcklpd %%xmm3,%%xmm2,%%xmm1;vfmadd213ps (%5),     %%xmm0,%%xmm1;vmovups %%xmm1,(%5);"\
    "vunpckhpd %%xmm3,%%xmm2,%%xmm1;vfmadd213ps (%5,%3,1),%%xmm0,%%xmm1;vmovups %%xmm1,(%5,%3,1);"\
    "leaq (%5,%3,2),%5;"
+#endif
 #define SAVE_L_m4n2 "movq %2,%5;" unit_save_m4n2(%%xmm4,%%xmm5)
 #define SAVE_L_m4n4  SAVE_L_m4n2  unit_save_m4n2(%%xmm6,%%xmm7)
 #define SAVE_L_m4n8  SAVE_L_m4n4  unit_save_m4n2(%%xmm8,%%xmm9)   unit_save_m4n2(%%xmm10,%%xmm11)
@ -183,29 +389,32 @@
 #define SAVE_R_m4n8  SAVE_R_m4n4  unit_save_m4n2(%%xmm8,%%xmm9)   unit_save_m4n2(%%xmm10,%%xmm11)
 #define SAVE_R_m4n12 SAVE_R_m4n8  unit_save_m4n2(%%xmm12,%%xmm13) unit_save_m4n2(%%xmm14,%%xmm15)
 #define COMPUTE_L_m4(ndim,sim) \
-    INIT_m4n##ndim\
-    "movq %%r13,%4; movq %%r14,%1;"\
-    #ndim""#sim"442:\n\t"\
+    INIT_m4n##ndim START_SET_PAPB(4,ndim)\
+    "movq %%r13,%4;"\
+    KERNEL_HEAD_C_n##ndim(4)\
    "testq %4,%4; jz "#ndim""#sim"443f;"\
-    KERNEL_k1m4n##ndim(%1)\
-    "decq %4; jmp "#ndim""#sim"442b;"\
+    #ndim""#sim"442:\n\t"\
+    kernel_k1m4n##ndim(%1)\
+    "decq %4; jnz "#ndim""#sim"442b;"\
    #ndim""#sim"443:\n\t"\
+    KERNEL_TAIL_C_n##ndim(4)\
    SAVE_L_m4n##ndim "addq $16,%2;"
 #define COMPUTE_R_m4(ndim,sim) \
-    "subq %%r12,%0;"\
-    INIT_m4n##ndim\
-    "movq %%r13,%4; leaq (%%r14,%%r12,2),%%r15; addq %%r12,%%r15;"\
-    #ndim""#sim"442:\n\t"\
+    INIT_m4n##ndim RECOVER_PA(4)\
+    "movq %%r13,%4;"\
+    KERNEL_HEAD_R_n##ndim(4)\
    "testq %4,%4; jz "#ndim""#sim"443f;"\
-    KERNEL_k1m4n##ndim(%%r15)\
-    "decq %4; jmp "#ndim""#sim"442b;"\
+    #ndim""#sim"442:\n\t"\
+    kernel_k1m4n##ndim(%%r15)\
+    "decq %4; jnz "#ndim""#sim"442b;"\
    #ndim""#sim"443:\n\t"\
-    SAVE_R_m4n##ndim
-#define COMPUTE_m4_n1  COMPUTE_L_m4(1,55855)
-#define COMPUTE_m4_n2  COMPUTE_L_m4(2,55855)
-#define COMPUTE_m4_n4  COMPUTE_L_m4(4,55855)
-#define COMPUTE_m4_n8  COMPUTE_L_m4(8,55855)
-#define COMPUTE_m4_n12 COMPUTE_L_m4(12,55855)
+    KERNEL_TAIL_R_n##ndim(4)\
+    SAVE_R_m4n##ndim END_SET_PA(4)
+#define COMPUTE_m4_n1  COMPUTE_L_m4(1,55855) END_SET_PA(4)
+#define COMPUTE_m4_n2  COMPUTE_L_m4(2,55855) END_SET_PA(4)
+#define COMPUTE_m4_n4  COMPUTE_L_m4(4,55855) END_SET_PA(4)
+#define COMPUTE_m4_n8  COMPUTE_L_m4(8,55855) END_SET_PA(4)
+#define COMPUTE_m4_n12 COMPUTE_L_m4(12,55855) END_SET_PA(4)
 #define COMPUTE_m4_n16 COMPUTE_L_m4(12,55755) COMPUTE_R_m4(4,55955)
 #define COMPUTE_m4_n20 COMPUTE_L_m4(12,55655) COMPUTE_R_m4(8,55955)
 #define COMPUTE_m4_n24 COMPUTE_L_m4(12,55555) COMPUTE_R_m4(12,55955)
@ -217,40 +426,60 @@
    "vmovsd (%0),%%xmm1; addq $8,%0;"\
    "vbroadcastss (%1),%%xmm2; vfmadd231ps %%xmm1,%%xmm2,%%xmm4;"\
    "addq $4,%1;"
-#define SAVE_h_m2n1 "vmovsd (%2),%%xmm1; vfmadd213ps %%xmm1,%%xmm0,%%xmm4; vmovsd %%xmm4,(%2);"
 #define INIT_m2n2 INIT_m2n1 "vpxor %%xmm5,%%xmm5,%%xmm5;"
 #define KERNEL_k1m2n2 \
    "vmovsd (%0),%%xmm1; addq $8,%0;"\
    "vbroadcastss  (%1),%%xmm2; vfmadd231ps %%xmm1,%%xmm2,%%xmm4;"\
    "vbroadcastss 4(%1),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm5;"\
    "addq $8,%1;"
-#define SAVE_h_m2n2 SAVE_h_m2n1 "vmovsd (%2,%3,1),%%xmm1; vfmadd213ps %%xmm1,%%xmm0,%%xmm5; vmovsd %%xmm5,(%2,%3,1);"
+#ifdef TRMMKERNEL
+  #define SAVE_h_m2n1 "vmulps %%xmm4,%%xmm0,%%xmm4; vmovsd %%xmm4,(%2);"
+  #define SAVE_h_m2n2 SAVE_h_m2n1 "vmulps %%xmm5,%%xmm0,%%xmm5; vmovsd %%xmm5,(%2,%3,1);"
+#else
+  #define SAVE_h_m2n1 "vmovsd (%2),%%xmm1; vfmadd213ps %%xmm1,%%xmm0,%%xmm4; vmovsd %%xmm4,(%2);"
+  #define SAVE_h_m2n2 SAVE_h_m2n1 "vmovsd (%2,%3,1),%%xmm1; vfmadd213ps %%xmm1,%%xmm0,%%xmm5; vmovsd %%xmm5,(%2,%3,1);"
+#endif
 #define INIT_m2n4  INIT_m2n2
 #define INIT_m2n8  INIT_m2n4 "vpxor %%xmm6,%%xmm6,%%xmm6; vpxor %%xmm7,%%xmm7,%%xmm7;"
 #define INIT_m2n12 INIT_m2n8 "vpxor %%xmm8,%%xmm8,%%xmm8; vpxor %%xmm9,%%xmm9,%%xmm9;"
 #define INIT_m2n16 INIT_m2n12 "vpxor %%xmm10,%%xmm10,%%xmm10; vpxor %%xmm11,%%xmm11,%%xmm11;"
 #define INIT_m2n20 INIT_m2n16 "vpxor %%xmm12,%%xmm12,%%xmm12; vpxor %%xmm13,%%xmm13,%%xmm13;"
 #define INIT_m2n24 INIT_m2n20 "vpxor %%xmm14,%%xmm14,%%xmm14; vpxor %%xmm15,%%xmm15,%%xmm15;"
+#define unit_gen_kernel_k1m2n4(c1,c2,k_no,...) \
+    "vmovups "#k_no"*16("#__VA_ARGS__"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,"#c1"; vfmadd231ps %%xmm2,%%xmm3,"#c2";"
 #define KERNEL_h_k1m2n4 \
-    "vbroadcastss (%0),%%xmm1; vbroadcastss 4(%0),%%xmm2; addq $8,%0;"\
-    "vmovups (%1),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm4; vfmadd231ps %%xmm2,%%xmm3,%%xmm5;"
+    "vbroadcastss (%0),%%xmm1; vbroadcastss 4(%0),%%xmm2; addq $8,%0;" unit_gen_kernel_k1m2n4(%%xmm4,%%xmm5,0,%1)
 #define KERNEL_k1m2n4 KERNEL_h_k1m2n4 "addq $16,%1;"
-#define KERNEL_h_k1m2n8 KERNEL_h_k1m2n4 "vmovups (%1,%%r12,1),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm6; vfmadd231ps %%xmm2,%%xmm3,%%xmm7;"
+#define KERNEL_h_k1m2n8 KERNEL_h_k1m2n4 unit_gen_kernel_k1m2n4(%%xmm6,%%xmm7,0,%1,%%r12,1)
 #define KERNEL_k1m2n8 KERNEL_h_k1m2n8 "addq $16,%1;"
-#define KERNEL_k1m2n12 KERNEL_h_k1m2n8 \
-    "vmovups (%1,%%r12,2),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm8; vfmadd231ps %%xmm2,%%xmm3,%%xmm9; addq $16,%1;"
-#define KERNEL_h_k1m2n16 KERNEL_k1m2n12 "vmovups (%%r15),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm10; vfmadd231ps %%xmm2,%%xmm3,%%xmm11;"
+#define KERNEL_k1m2n12 KERNEL_h_k1m2n8 unit_gen_kernel_k1m2n4(%%xmm8,%%xmm9,0,%1,%%r12,2) "addq $16,%1;"
+#define KERNEL_h_k1m2n16 KERNEL_k1m2n12 unit_gen_kernel_k1m2n4(%%xmm10,%%xmm11,0,%%r15)
 #define KERNEL_k1m2n16 KERNEL_h_k1m2n16 "addq $16,%%r15;"
-#define KERNEL_h_k1m2n20 KERNEL_h_k1m2n16 "vmovups (%%r15,%%r12,1),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm12; vfmadd231ps %%xmm2,%%xmm3,%%xmm13;"
+#define KERNEL_h_k1m2n20 KERNEL_h_k1m2n16 unit_gen_kernel_k1m2n4(%%xmm12,%%xmm13,0,%%r15,%%r12,1)
 #define KERNEL_k1m2n20 KERNEL_h_k1m2n20 "addq $16,%%r15;"
-#define KERNEL_h_k1m2n24 KERNEL_h_k1m2n20 "vmovups (%%r15,%%r12,2),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm14; vfmadd231ps %%xmm2,%%xmm3,%%xmm15;"
+#define KERNEL_h_k1m2n24 KERNEL_h_k1m2n20 unit_gen_kernel_k1m2n4(%%xmm14,%%xmm15,0,%%r15,%%r12,2)
 #define KERNEL_k1m2n24 KERNEL_h_k1m2n24 "addq $16,%%r15;"
-#define unit_save_m2n4(c1,c2) \
+#define end_load_a_k1m2(k_no) "vbroadcastss "#k_no"*8(%0),%%xmm1; vbroadcastss "#k_no"*8+4(%0),%%xmm2;"
+#define end_acc_nc2_k1m2(k_no) unit_gen_kernel_k1m2n4(%%xmm6,%%xmm7,k_no,%1,%%r12,1)
+#define end_acc_nc3_k1m2(k_no) unit_gen_kernel_k1m2n4(%%xmm8,%%xmm9,k_no,%1,%%r12,2)
+#define end_acc_nc4_k1m2(k_no) unit_gen_kernel_k1m2n4(%%xmm10,%%xmm11,k_no,%%r15)
+#define end_acc_nc5_k1m2(k_no) unit_gen_kernel_k1m2n4(%%xmm12,%%xmm13,k_no,%%r15,%%r12,1)
+#define end_acc_nc6_k1m2(k_no) unit_gen_kernel_k1m2n4(%%xmm14,%%xmm15,k_no,%%r15,%%r12,2)
+#ifdef TRMMKERNEL
+  #define unit_save_m2n4(c1,c2) \
+    "vunpcklps "#c2","#c1",%%xmm1; vunpckhps "#c2","#c1",%%xmm2;"\
+    "vmulps %%xmm1,%%xmm0,%%xmm1; vmovsd %%xmm1,(%5); vmovhpd %%xmm1,(%5,%3,1);"\
+    "leaq (%5,%3,2),%5;"\
+    "vmulps %%xmm2,%%xmm0,%%xmm2; vmovsd %%xmm2,(%5); vmovhpd %%xmm2,(%5,%3,1);"\
+    "leaq (%5,%3,2),%5;"
+#else
+  #define unit_save_m2n4(c1,c2) \
    "vunpcklps "#c2","#c1",%%xmm1; vunpckhps "#c2","#c1",%%xmm2;"\
    "vmovsd (%5),%%xmm3; vmovhpd (%5,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm1; vmovsd %%xmm1,(%5); vmovhpd %%xmm1,(%5,%3,1);"\
    "leaq (%5,%3,2),%5;"\
    "vmovsd (%5),%%xmm3; vmovhpd (%5,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm2; vmovsd %%xmm2,(%5); vmovhpd %%xmm2,(%5,%3,1);"\
    "leaq (%5,%3,2),%5;"
+#endif
 #define SAVE_h_m2n4  "movq %2,%5;" unit_save_m2n4(%%xmm4,%%xmm5)
 #define SAVE_h_m2n8  SAVE_h_m2n4   unit_save_m2n4(%%xmm6,%%xmm7)
 #define SAVE_h_m2n12 SAVE_h_m2n8   unit_save_m2n4(%%xmm8,%%xmm9)
@ -259,13 +488,15 @@
 #define SAVE_h_m2n24 SAVE_h_m2n20  unit_save_m2n4(%%xmm14,%%xmm15)
 #define SAVE_m2(ndim) SAVE_h_m2n##ndim "addq $8,%2;"
 #define COMPUTE_m2(ndim) \
-    INIT_m2n##ndim\
-    "movq %%r13,%4; movq %%r14,%1; leaq (%1,%%r12,2),%%r15; addq %%r12,%%r15;"\
+    INIT_m2n##ndim START_SET_PAPB(2,ndim)\
+    "movq %%r13,%4;"\
+    KERNEL_HEAD_C_n##ndim(2)\
    "testq %4,%4; jz "#ndim"002022f;"\
    #ndim"002021:\n\t"\
    KERNEL_k1m2n##ndim "decq %4; jnz "#ndim"002021b;"\
    #ndim"002022:\n\t"\
-    SAVE_m2(ndim)
+    KERNEL_TAIL_C_n##ndim(2)\
+    SAVE_m2(ndim) END_SET_PA(2)

 /* m = 1 *//* xmm0 for alpha, xmm1-xmm3 and xmm10 for temporary use, xmm4-xmm9 for accumulators */
 #define INIT_m1n1 "vpxor %%xmm4,%%xmm4,%%xmm4;"
@ -273,15 +504,25 @@
    "vmovss (%1),%%xmm3; addq $4,%1;"\
    "vmovss (%0),%%xmm1; vfmadd231ss %%xmm3,%%xmm1,%%xmm4;"\
    "addq $4,%0;"
-#define SAVE_h_m1n1 "vfmadd213ss (%2),%%xmm0,%%xmm4; vmovss %%xmm4,(%2);"
+#ifdef TRMMKERNEL
+  #define SAVE_h_m1n1 "vmulss %%xmm4,%%xmm0,%%xmm4; vmovss %%xmm4,(%2);"
+#else
+  #define SAVE_h_m1n1 "vfmadd213ss (%2),%%xmm0,%%xmm4; vmovss %%xmm4,(%2);"
+#endif
 #define INIT_m1n2 INIT_m1n1
 #define KERNEL_k1m1n2 \
    "vmovsd (%1),%%xmm3; addq $8,%1;"\
    "vbroadcastss  (%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm4;"\
    "addq $4,%0;"
-#define SAVE_h_m1n2 \
+#ifdef TRMMKERNEL
+  #define SAVE_h_m1n2 \
+    "vmulps %%xmm4,%%xmm0,%%xmm4;"\
+    "vmovss %%xmm4,(%2); vextractps $1,%%xmm4,(%2,%3,1);"
+#else
+  #define SAVE_h_m1n2 \
    "vmovss (%2),%%xmm3; vinsertps $16,(%2,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm4;"\
    "vmovss %%xmm4,(%2); vextractps $1,%%xmm4,(%2,%3,1);"
+#endif
 #define INIT_m1n4  INIT_m1n2
 #define INIT_m1n8  INIT_m1n4 "vpxor %%xmm5,%%xmm5,%%xmm5;"
 #define INIT_m1n12 INIT_m1n8 "vpxor %%xmm6,%%xmm6,%%xmm6;"
@ -300,12 +541,25 @@
 #define KERNEL_k1m1n20 KERNEL_h_k1m1n20 "addq $16,%%r15;"
 #define KERNEL_h_k1m1n24 KERNEL_h_k1m1n20 "vfmadd231ps (%%r15,%%r12,2),%%xmm1,%%xmm9;"
 #define KERNEL_k1m1n24 KERNEL_h_k1m1n24 "addq $16,%%r15;"
-#define unit_save_m1n4(c1) \
+#define end_load_a_k1m1(k_no) "vbroadcastss "#k_no"*4(%0),%%xmm1;"
+#define end_acc_nc2_k1m1(k_no) "vfmadd231ps "#k_no"*16(%1,%%r12,1),%%xmm1,%%xmm5;"
+#define end_acc_nc3_k1m1(k_no) "vfmadd231ps "#k_no"*16(%1,%%r12,2),%%xmm1,%%xmm6;"
+#define end_acc_nc4_k1m1(k_no) "vfmadd231ps "#k_no"*16(%%r15),%%xmm1,%%xmm7;"
+#define end_acc_nc5_k1m1(k_no) "vfmadd231ps "#k_no"*16(%%r15,%%r12,1),%%xmm1,%%xmm8;"
+#define end_acc_nc6_k1m1(k_no) "vfmadd231ps "#k_no"*16(%%r15,%%r12,2),%%xmm1,%%xmm9;"
+#ifdef TRMMKERNEL
+  #define unit_save_m1n4(c1) \
+    "vmulps "#c1",%%xmm0,"#c1"; vpxor %%xmm10,%%xmm10,%%xmm10; vmovsd "#c1",%%xmm10,%%xmm2; vmovhlps "#c1",%%xmm10,%%xmm1;"\
+    "vmovss %%xmm2,(%5); vextractps $1,%%xmm2,(%5,%3,1); leaq (%5,%3,2),%5;"\
+    "vmovss %%xmm1,(%5); vextractps $1,%%xmm1,(%5,%3,1); leaq (%5,%3,2),%5;"
+#else
+  #define unit_save_m1n4(c1) \
    "vpxor %%xmm10,%%xmm10,%%xmm10; vmovsd "#c1",%%xmm10,%%xmm2; vmovhlps "#c1",%%xmm10,%%xmm1;"\
    "vmovss (%5),%%xmm3; vinsertps $16,(%5,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm2;"\
    "vmovss %%xmm2,(%5); vextractps $1,%%xmm2,(%5,%3,1); leaq (%5,%3,2),%5;"\
    "vmovss (%5),%%xmm3; vinsertps $16,(%5,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm1;"\
    "vmovss %%xmm1,(%5); vextractps $1,%%xmm1,(%5,%3,1); leaq (%5,%3,2),%5;"
+#endif
 #define SAVE_h_m1n4 "movq %2,%5;" unit_save_m1n4(%%xmm4)
 #define SAVE_h_m1n8  SAVE_h_m1n4  unit_save_m1n4(%%xmm5)
 #define SAVE_h_m1n12 SAVE_h_m1n8  unit_save_m1n4(%%xmm6)
@ -314,58 +568,102 @@
 #define SAVE_h_m1n24 SAVE_h_m1n20 unit_save_m1n4(%%xmm9)
 #define SAVE_m1(ndim) SAVE_h_m1n##ndim "addq $4,%2;"
 #define COMPUTE_m1(ndim) \
-    INIT_m1n##ndim\
-    "movq %%r13,%4; movq %%r14,%1; leaq (%1,%%r12,2),%%r15; addq %%r12,%%r15;"\
+    INIT_m1n##ndim START_SET_PAPB(1,ndim)\
+    "movq %%r13,%4;"\
+    KERNEL_HEAD_C_n##ndim(1)\
    "testq %4,%4; jz "#ndim"001012f;"\
    #ndim"001011:\n\t"\
    KERNEL_k1m1n##ndim "decq %4; jnz "#ndim"001011b;"\
    #ndim"001012:\n\t"\
-    SAVE_m1(ndim)
+    KERNEL_TAIL_C_n##ndim(1)\
+    SAVE_m1(ndim) END_SET_PA(1)

-/* %0 = "+r"(a_pointer), %1 = "+r"(b_pointer), %2 = "+r"(c_pointer), %3 = "+r"(ldc_in_bytes), %4 = "+r"(K), %5 = "+r"(ctemp) */
-/* %6 = "+r"(next_b), %7 = "m"(ALPHA), %8 = "m"(M) */
-/* r11 = m_counter, r12 = k << 4(const), r13 = k(const), r14 = b_head_pos(const), r15 = %1 + 3r12 */
+/* %7 = "m"(ALPHA), %8 = "m"(M), %9 = "m"(K), %10 = "m"(off) */

+#ifdef TRMMKERNEL
+  #if BACKWARDS == 1
+    #define OFFSET_TO_K "movq %9,%%r13; subq %10,%%r13;"
+  #else
+    #define OFFSET_TO_K "movq %10,%%r13;"
+  #endif
+#else
+  #define OFFSET_TO_K "movq %9,%%r13;"
+#endif
+#if defined(TRMMKERNEL) && !defined(LEFT)
+  #if BACKWARDS == 1
+    #define START_UPDATE_OFFSET(ndim) {}
+    #define END_UPDATE_OFFSET(ndim) {off += (ndim);}
+  #else
+    #define START_UPDATE_OFFSET(ndim) {off += (ndim)>4 ? 4:(ndim);}
+    #define END_UPDATE_OFFSET(ndim) {off += (ndim)>4 ? ((ndim)-4):0;}
+  #endif
+#else
+  #define START_UPDATE_OFFSET(ndim) {}
+  #define END_UPDATE_OFFSET(ndim) {}
+#endif
+#if defined(TRMMKERNEL) && defined(LEFT)
+  #if BACKWARDS == 1
+    #define START_UPDATE_K(mdim) ""
+    #define END_UPDATE_K(mdim) "subq $"#mdim",%%r13;"
+  #else
+    #define START_UPDATE_K(mdim) "addq $"#mdim",%%r13;"
+    #define END_UPDATE_K(mdim) ""
+  #endif
+#else
+  #define START_UPDATE_K(mdim) ""
+  #define END_UPDATE_K(mdim) ""
+#endif
 #define COMPUTE(ndim) {\
-    next_b = b_pointer + ndim * K;\
-    __asm__ __volatile__(\
+  next_b = b_pointer + ndim * K; START_UPDATE_OFFSET(ndim)\
+  __asm__ __volatile__(\
    "vbroadcastss %7,%%zmm0;"\
-    "movq %4,%%r13; movq %4,%%r12; salq $4,%%r12; movq %1,%%r14; movq %8,%%r11;"\
+    OFFSET_TO_K "movq %9,%%r12; salq $4,%%r12; movq %1,%%r14; movq %8,%%r11;"\
    "cmpq $16,%%r11;jb 33101"#ndim"f;"\
    "33109"#ndim":\n\t"\
-    COMPUTE_m16(ndim)\
+    START_UPDATE_K(16) COMPUTE_m16(ndim) END_UPDATE_K(16)\
    "subq $16,%%r11;cmpq $16,%%r11;jnb 33109"#ndim"b;"\
    "33101"#ndim":\n\t"\
    "cmpq $8,%%r11;jb 33102"#ndim"f;"\
-    COMPUTE_m8(ndim)\
+    START_UPDATE_K(8) COMPUTE_m8(ndim) END_UPDATE_K(8)\
    "subq $8,%%r11;"\
    "33102"#ndim":\n\t"\
    "cmpq $4,%%r11;jb 33103"#ndim"f;"\
-    COMPUTE_m4(ndim)\
+    START_UPDATE_K(4) COMPUTE_m4(ndim) END_UPDATE_K(4)\
    "subq $4,%%r11;"\
    "33103"#ndim":\n\t"\
    "cmpq $2,%%r11;jb 33104"#ndim"f;"\
-    COMPUTE_m2(ndim)\
+    START_UPDATE_K(2) COMPUTE_m2(ndim) END_UPDATE_K(2)\
    "subq $2,%%r11;"\
    "33104"#ndim":\n\t"\
    "testq %%r11,%%r11;jz 33105"#ndim"f;"\
-    COMPUTE_m1(ndim)\
+    START_UPDATE_K(1) COMPUTE_m1(ndim) END_UPDATE_K(1)\
    "33105"#ndim":\n\t"\
-    "movq %%r13,%4; movq %%r14,%1; vzeroupper;"\
-    :"+r"(a_pointer),"+r"(b_pointer),"+r"(c_pointer),"+r"(ldc_in_bytes),"+r"(K),"+r"(ctemp),"+r"(next_b):"m"(ALPHA),"m"(M)\
-    :"r10","r11","r12","r13","r14","r15","zmm0","zmm1","zmm2","zmm3","zmm4","zmm5","zmm6","zmm7","zmm8","zmm9","zmm10","zmm11","zmm12","zmm13","zmm14",\
-    "zmm15","zmm16","zmm17","zmm18","zmm19","zmm20","zmm21","zmm22","zmm23","zmm24","zmm25","zmm26","zmm27","zmm28","zmm29","zmm30","zmm31",\
-    "cc","memory");\
-    a_pointer -= M * K; b_pointer += ndim * K; c_pointer += LDC * ndim - M;\
+    "movq %%r14,%1; vzeroupper;"\
+    :"+r"(a_pointer),"+r"(b_pointer),"+r"(c_pointer),"+r"(ldc_in_bytes),"+r"(k_counter),"+r"(ctemp),"+r"(next_b)\
+    :"m"(ALPHA),"m"(M),"m"(K),"m"(off):"r10","r11","r12","r13","r14","r15","cc","memory",\
+    "zmm0","zmm1","zmm2","zmm3","zmm4","zmm5","zmm6","zmm7","zmm8","zmm9","zmm10","zmm11","zmm12","zmm13","zmm14","zmm15",\
+    "zmm16","zmm17","zmm18","zmm19","zmm20","zmm21","zmm22","zmm23","zmm24","zmm25","zmm26","zmm27","zmm28","zmm29","zmm30","zmm31");\
+  a_pointer -= M * K; b_pointer += ndim * K; c_pointer += LDC * ndim - M; END_UPDATE_OFFSET(ndim)\
 }
 int __attribute__ ((noinline))
-CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict__ A, float * __restrict__ B, float * __restrict__ C, BLASLONG LDC)
+CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict__ A, float * __restrict__ B, float * __restrict__ C, BLASLONG LDC
+#ifdef TRMMKERNEL
+,BLASLONG offset
+#endif
+)
 {
-    if(m==0||n==0||k==0||alpha==(float)0.0) return 0;
+    if(m==0||n==0) return 0;
    int64_t ldc_in_bytes = (int64_t)LDC * sizeof(float);float ALPHA = alpha;
-    int64_t M = (int64_t)m, K = (int64_t)k;
+    int64_t M = (int64_t)m, K = (int64_t)k, k_counter = K, off = 0;
    BLASLONG n_count = n;
    float *a_pointer = A,*b_pointer = B,*c_pointer = C,*ctemp = C,*next_b = B;
+#ifdef TRMMKERNEL
+  #ifdef LEFT
+    off = offset;
+  #else
+    off = -offset;
+  #endif
+#endif
    for(;n_count>23;n_count-=24) COMPUTE(24)
    for(;n_count>19;n_count-=20) COMPUTE(20)
    for(;n_count>15;n_count-=16) COMPUTE(16)
@ -376,5 +674,7 @@ CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict__ A, f
    if(n_count>0) COMPUTE(1)
    return 0;
 }
-#include <immintrin.h>
-#include "sgemm_direct_skylakex.c"
+#ifndef TRMMKERNEL
+  #include <immintrin.h>
+  #include "sgemm_direct_skylakex.c"
+#endif
--- a/kernel/x86_64/sgemm_kernel_8x4_haswell_2.c
+++ b/kernel/x86_64/sgemm_kernel_8x4_haswell_2.c
@ -0,0 +1,424 @@
+/* %0 = "+r"(a_pointer), %1 = "+r"(b_pointer), %2 = "+r"(c_pointer), %3 = "+r"(ldc_in_bytes), %4 for k_count, %5 for c_store, %6 = b_pref */
+/* r10 = tmp, r11 = m_counter, r12 = k << 2(const), r13 = tmp, r14 = b_head_pos(const), r15 = tmp */
+
+/* m = 8 *//* ymm0 for alpha, ymm1-ymm3 for temporary use, ymm4-ymm15 for accumulators */
+#define KERNEL_k1m8n1 \
+    "vmovups (%0),%%ymm1; addq $32,%0;"\
+    "vbroadcastss (%1),%%ymm2; vfmadd231ps %%ymm1,%%ymm2,%%ymm4;"\
+    "addq $4,%1;"
+#define KERNEL_h_k1m8n2 \
+    "vmovsldup (%0),%%ymm1; vmovshdup (%0),%%ymm2; addq $32,%0;"\
+    "vbroadcastsd (%1),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4; vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"
+#define KERNEL_k1m8n2 KERNEL_h_k1m8n2 "addq $8,%1;"
+#define KERNEL_h_k1m8n4 \
+    KERNEL_h_k1m8n2 "vbroadcastsd 8(%1),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6; vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"
+#define KERNEL_k1m8n4 KERNEL_h_k1m8n4 "addq $16,%1;"
+#define unit_kernel_k1m8n4(c1,c2,c3,c4,boff,...) \
+    "vbroadcastsd "#boff"("#__VA_ARGS__"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,"#c1"; vfmadd231ps %%ymm2,%%ymm3,"#c2";"\
+    "vbroadcastsd "#boff"+8("#__VA_ARGS__"),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,"#c3"; vfmadd231ps %%ymm2,%%ymm3,"#c4";"
+#define KERNEL_h_k1m8n8 KERNEL_h_k1m8n4 unit_kernel_k1m8n4(%%ymm8,%%ymm9,%%ymm10,%%ymm11,0,%1,%%r12,4)
+#define KERNEL_k1m8n8 KERNEL_h_k1m8n8 "addq $16,%1;"
+#define KERNEL_h_k1m8n12 KERNEL_h_k1m8n8 unit_kernel_k1m8n4(%%ymm12,%%ymm13,%%ymm14,%%ymm15,0,%1,%%r12,8)
+#define KERNEL_k1m8n12 KERNEL_h_k1m8n12 "addq $16,%1;"
+#define KERNEL_k2m8n4 \
+    "vmovsldup (%0),%%ymm1; vmovshdup (%0),%%ymm2; prefetcht0 512(%0);"\
+    unit_kernel_k1m8n4(%%ymm4,%%ymm5,%%ymm6,%%ymm7,0,%1)\
+    "vmovsldup 32(%0),%%ymm1; vmovshdup 32(%0),%%ymm2; addq $64,%0;"\
+    unit_kernel_k1m8n4(%%ymm8,%%ymm9,%%ymm10,%%ymm11,16,%1)\
+    "addq $32,%1;"
+#define KERNEL_L_k1m8n6 \
+    "vmovsldup (%0),%%ymm1; vmovshdup (%0),%%ymm2; prefetcht0 512(%0); addq $32,%0;"\
+    "vbroadcastsd    (%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4;  vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\
+    "vbroadcastsd   8(%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6;  vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\
+    "vbroadcastsd    (%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8;  vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\
+    "addq $16,%1;"
+#define KERNEL_L_k2m8n6 \
+    "vmovsldup (%0),%%ymm1; vmovshdup (%0),%%ymm2; prefetcht0 512(%0);"\
+    "vbroadcastsd    (%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4;  vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\
+    "vbroadcastsd   8(%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6;  vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\
+    "vbroadcastsd    (%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8;  vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\
+    "vmovsldup 32(%0),%%ymm1; vmovshdup 32(%0),%%ymm2; addq $64,%0;"\
+    "vbroadcastsd  16(%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm10; vfmadd231ps %%ymm2,%%ymm3,%%ymm11;"\
+    "vbroadcastsd  24(%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm12; vfmadd231ps %%ymm2,%%ymm3,%%ymm13;"\
+    "vbroadcastsd  16(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm14; vfmadd231ps %%ymm2,%%ymm3,%%ymm15;"\
+    "addq $32,%1;"
+#define KERNEL_L_k1m16n6 \
+    "vmovups (%0),%%ymm1; vmovups (%0,%%r12,8),%%ymm2; prefetcht0 512(%0,%%r12,8); addq $32,%0;"\
+    "vbroadcastss    (%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4;  vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\
+    "vbroadcastss   4(%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6;  vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\
+    "vbroadcastss   8(%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8;  vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\
+    "vbroadcastss  12(%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm10; vfmadd231ps %%ymm2,%%ymm3,%%ymm11;"\
+    "vbroadcastss    (%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm12; vfmadd231ps %%ymm2,%%ymm3,%%ymm13;"\
+    "vbroadcastss   4(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm14; vfmadd231ps %%ymm2,%%ymm3,%%ymm15;"\
+    "addq $16,%1;"
+#define KERNEL_L_k2m16n6 \
+    "vmovups (%0),%%ymm1; vmovups (%0,%%r12,8),%%ymm2; prefetcht0 512(%0,%%r12,8);"\
+    "vbroadcastss    (%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4;  vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\
+    "vbroadcastss   4(%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6;  vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\
+    "vbroadcastss   8(%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8;  vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\
+    "vbroadcastss  12(%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm10; vfmadd231ps %%ymm2,%%ymm3,%%ymm11;"\
+    "vbroadcastss    (%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm12; vfmadd231ps %%ymm2,%%ymm3,%%ymm13;"\
+    "vbroadcastss   4(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm14; vfmadd231ps %%ymm2,%%ymm3,%%ymm15;"\
+    "vmovups 32(%0),%%ymm1; vmovups 32(%0,%%r12,8),%%ymm2; addq $64,%0;"\
+    "vbroadcastss  16(%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4;  vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\
+    "vbroadcastss  20(%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6;  vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\
+    "vbroadcastss  24(%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8;  vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\
+    "vbroadcastss  28(%1)        ,%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm10; vfmadd231ps %%ymm2,%%ymm3,%%ymm11;"\
+    "vbroadcastss  16(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm12; vfmadd231ps %%ymm2,%%ymm3,%%ymm13;"\
+    "vbroadcastss  20(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm14; vfmadd231ps %%ymm2,%%ymm3,%%ymm15;"\
+    "addq $32,%1;"
+#define KERNEL_R_k1m16n6 \
+    "vmovups (%0),%%ymm1; vmovups (%0,%%r12,8),%%ymm2; prefetcht0 512(%0,%%r12,8); addq $32,%0;"\
+    "vbroadcastss   8(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4;  vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\
+    "vbroadcastss  12(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6;  vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\
+    "vbroadcastss    (%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8;  vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\
+    "vbroadcastss   4(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm10; vfmadd231ps %%ymm2,%%ymm3,%%ymm11;"\
+    "vbroadcastss   8(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm12; vfmadd231ps %%ymm2,%%ymm3,%%ymm13;"\
+    "vbroadcastss  12(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm14; vfmadd231ps %%ymm2,%%ymm3,%%ymm15;"\
+    "addq $16,%1;"
+#define KERNEL_R_k2m16n6 \
+    "vmovups (%0),%%ymm1; vmovups (%0,%%r12,8),%%ymm2; prefetcht0 512(%0,%%r12,8);"\
+    "vbroadcastss   8(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4;  vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\
+    "vbroadcastss  12(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6;  vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\
+    "vbroadcastss    (%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8;  vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\
+    "vbroadcastss   4(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm10; vfmadd231ps %%ymm2,%%ymm3,%%ymm11;"\
+    "vbroadcastss   8(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm12; vfmadd231ps %%ymm2,%%ymm3,%%ymm13;"\
+    "vbroadcastss  12(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm14; vfmadd231ps %%ymm2,%%ymm3,%%ymm15;"\
+    "vmovups 32(%0),%%ymm1; vmovups 32(%0,%%r12,8),%%ymm2; addq $64,%0;"\
+    "vbroadcastss  24(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4;  vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\
+    "vbroadcastss  28(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6;  vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\
+    "vbroadcastss  16(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8;  vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\
+    "vbroadcastss  20(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm10; vfmadd231ps %%ymm2,%%ymm3,%%ymm11;"\
+    "vbroadcastss  24(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm12; vfmadd231ps %%ymm2,%%ymm3,%%ymm13;"\
+    "vbroadcastss  28(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm14; vfmadd231ps %%ymm2,%%ymm3,%%ymm15;"\
+    "addq $32,%1;"
+#define KERNEL_R_k1m8n6 \
+    "vmovsldup (%0),%%ymm1; vmovshdup (%0),%%ymm2; prefetcht0 512(%0); addq $32,%0;"\
+    "vbroadcastsd   8(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4;  vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\
+    "vbroadcastsd    (%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6;  vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\
+    "vbroadcastsd   8(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8;  vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\
+    "addq $16,%1;"
+#define KERNEL_R_k2m8n6 \
+    "vmovsldup (%0),%%ymm1; vmovshdup (%0),%%ymm2; prefetcht0 512(%0);"\
+    "vbroadcastsd   8(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm4;  vfmadd231ps %%ymm2,%%ymm3,%%ymm5;"\
+    "vbroadcastsd    (%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm6;  vfmadd231ps %%ymm2,%%ymm3,%%ymm7;"\
+    "vbroadcastsd   8(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm8;  vfmadd231ps %%ymm2,%%ymm3,%%ymm9;"\
+    "vmovsldup 32(%0),%%ymm1; vmovshdup 32(%0),%%ymm2; addq $64,%0;"\
+    "vbroadcastsd  24(%1,%%r12,4),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm10; vfmadd231ps %%ymm2,%%ymm3,%%ymm11;"\
+    "vbroadcastsd  16(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm12; vfmadd231ps %%ymm2,%%ymm3,%%ymm13;"\
+    "vbroadcastsd  24(%1,%%r12,8),%%ymm3; vfmadd231ps %%ymm1,%%ymm3,%%ymm14; vfmadd231ps %%ymm2,%%ymm3,%%ymm15;"\
+    "addq $32,%1;"
+#define INIT_m8n1 "vpxor %%ymm4,%%ymm4,%%ymm4;"
+#define INIT_m8n2 INIT_m8n1 "vpxor %%ymm5,%%ymm5,%%ymm5;"
+#define unit_init_m8n4(c1,c2,c3,c4) \
+    "vpxor "#c1","#c1","#c1";vpxor "#c2","#c2","#c2";vpxor "#c3","#c3","#c3";vpxor "#c4","#c4","#c4";"
+#define INIT_m8n8  unit_init_m8n4(%%ymm4,%%ymm5,%%ymm6,%%ymm7) unit_init_m8n4(%%ymm8,%%ymm9,%%ymm10,%%ymm11)
+#define INIT_m8n4  INIT_m8n8
+#define INIT_m8n12 INIT_m8n8 unit_init_m8n4(%%ymm12,%%ymm13,%%ymm14,%%ymm15)
+#define INIT_m8n6  INIT_m8n12
+#define INIT_m16n6 INIT_m8n12
+#define SAVE_m8n1 "vfmadd213ps (%2),%%ymm0,%%ymm4; vmovups %%ymm4,(%2);"
+#define unit_save_m8n2(c1,c2) \
+    "vunpcklps "#c2","#c1",%%ymm2; vunpckhps "#c2","#c1",%%ymm3; vunpcklpd %%ymm3,%%ymm2,"#c1"; vunpckhpd %%ymm3,%%ymm2,"#c2";"\
+    "vfmadd213ps (%5),%%ymm0,"#c1"; vfmadd213ps (%5,%3,1),%%ymm0,"#c2"; vmovups "#c1",(%5); vmovups "#c2",(%5,%3,1); leaq (%5,%3,2),%5;"
+#define SAVE_m8n2 "movq %2,%5;" unit_save_m8n2(%%ymm4,%%ymm5)
+#define SAVE_m8n4 "movq %2,%5;"\
+    "vaddps %%ymm4,%%ymm8,%%ymm4; vaddps %%ymm5,%%ymm9,%%ymm5; vaddps %%ymm6,%%ymm10,%%ymm6; vaddps %%ymm7,%%ymm11,%%ymm7;"\
+    unit_save_m8n2(%%ymm4,%%ymm5)  unit_save_m8n2(%%ymm6,%%ymm7)
+#define SAVE_m8n8 "movq %2,%5;"\
+    unit_save_m8n2(%%ymm4,%%ymm5)  unit_save_m8n2(%%ymm6,%%ymm7)  unit_save_m8n2(%%ymm8,%%ymm9)   unit_save_m8n2(%%ymm10,%%ymm11)
+#define SAVE_m8n12 SAVE_m8n8  unit_save_m8n2(%%ymm12,%%ymm13) unit_save_m8n2(%%ymm14,%%ymm15)
+#define unit_save_m16n2(c1,c2,c3,c4) \
+    "vfmadd213ps (%5),%%ymm0,"#c1"; vfmadd213ps 32(%5),%%ymm0,"#c2"; vmovups "#c1",(%5); vmovups "#c2",32(%5);"\
+    "vfmadd213ps (%5,%3,1),%%ymm0,"#c3"; vfmadd213ps 32(%5,%3,1),%%ymm0,"#c4"; vmovups "#c3",(%5,%3,1); vmovups "#c4",32(%5,%3,1); leaq (%5,%3,2),%5;"
+#define SAVE_L_m16n6 "movq %2,%5;"\
+    unit_save_m16n2(%%ymm4,%%ymm5,%%ymm6,%%ymm7) unit_save_m16n2(%%ymm8,%%ymm9,%%ymm10,%%ymm11) unit_save_m16n2(%%ymm12,%%ymm13,%%ymm14,%%ymm15)
+#define SAVE_R_m16n6 "leaq (%2,%3,4),%5; leaq (%5,%3,2),%5;"\
+    unit_save_m16n2(%%ymm4,%%ymm5,%%ymm6,%%ymm7) unit_save_m16n2(%%ymm8,%%ymm9,%%ymm10,%%ymm11) unit_save_m16n2(%%ymm12,%%ymm13,%%ymm14,%%ymm15)
+#define SAVE_L_m8n6 "movq %2,%5;"\
+    "vaddps %%ymm4,%%ymm10,%%ymm4; vaddps %%ymm5,%%ymm11,%%ymm5; vaddps %%ymm6,%%ymm12,%%ymm6;"\
+    "vaddps %%ymm7,%%ymm13,%%ymm7; vaddps %%ymm8,%%ymm14,%%ymm8; vaddps %%ymm9,%%ymm15,%%ymm9;"\
+    unit_save_m8n2(%%ymm4,%%ymm5) unit_save_m8n2(%%ymm6,%%ymm7) unit_save_m8n2(%%ymm8,%%ymm9)
+#define SAVE_R_m8n6 "leaq (%2,%3,4),%5; leaq (%5,%3,2),%5;"\
+    "vaddps %%ymm4,%%ymm10,%%ymm4; vaddps %%ymm5,%%ymm11,%%ymm5; vaddps %%ymm6,%%ymm12,%%ymm6;"\
+    "vaddps %%ymm7,%%ymm13,%%ymm7; vaddps %%ymm8,%%ymm14,%%ymm8; vaddps %%ymm9,%%ymm15,%%ymm9;"\
+    unit_save_m8n2(%%ymm4,%%ymm5) unit_save_m8n2(%%ymm6,%%ymm7) unit_save_m8n2(%%ymm8,%%ymm9)
+
+/* m = 4 *//* xmm0 for alpha, xmm1-xmm3 for temporary use, xmm4-xmm15 for accumulators */
+#define KERNEL_k1m4n1 \
+    "vmovups (%0),%%xmm1; addq $16,%0;"\
+    "vbroadcastss (%1),%%xmm2; vfmadd231ps %%xmm1,%%xmm2,%%xmm4;"\
+    "addq $4,%1;"
+#define KERNEL_h_k1m4n2 \
+    "vmovsldup (%0),%%xmm1; vmovshdup (%0),%%xmm2; addq $16,%0;"\
+    "vmovddup (%1),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm4; vfmadd231ps %%xmm2,%%xmm3,%%xmm5;"
+#define KERNEL_k1m4n2 KERNEL_h_k1m4n2 "addq $8,%1;"
+#define KERNEL_h_k1m4n4 \
+    KERNEL_h_k1m4n2 "vmovddup 8(%1),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm6; vfmadd231ps %%xmm2,%%xmm3,%%xmm7;"
+#define KERNEL_k1m4n4 KERNEL_h_k1m4n4 "addq $16,%1;"
+#define unit_kernel_k1m4n4(c1,c2,c3,c4,...) \
+    "vmovddup ("#__VA_ARGS__"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,"#c1"; vfmadd231ps %%xmm2,%%xmm3,"#c2";"\
+    "vmovddup 8("#__VA_ARGS__"),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,"#c3"; vfmadd231ps %%xmm2,%%xmm3,"#c4";"
+#define KERNEL_h_k1m4n8 KERNEL_h_k1m4n4 unit_kernel_k1m4n4(%%xmm8,%%xmm9,%%xmm10,%%xmm11,%1,%%r12,4)
+#define KERNEL_k1m4n8 KERNEL_h_k1m4n8 "addq $16,%1;"
+#define KERNEL_h_k1m4n12 KERNEL_h_k1m4n8 unit_kernel_k1m4n4(%%xmm12,%%xmm13,%%xmm14,%%xmm15,%1,%%r12,8)
+#define KERNEL_k1m4n12 KERNEL_h_k1m4n12 "addq $16,%1;"
+#define INIT_m4n1 "vpxor %%xmm4,%%xmm4,%%xmm4;"
+#define INIT_m4n2 INIT_m4n1 "vpxor %%xmm5,%%xmm5,%%xmm5;"
+#define INIT_m4n4 INIT_m4n2 "vpxor %%xmm6,%%xmm6,%%xmm6;vpxor %%xmm7,%%xmm7,%%xmm7;"
+#define unit_init_m4n4(c1,c2,c3,c4) \
+    "vpxor "#c1","#c1","#c1";vpxor "#c2","#c2","#c2";vpxor "#c3","#c3","#c3";vpxor "#c4","#c4","#c4";"
+#define INIT_m4n8  INIT_m4n4 unit_init_m4n4(%%xmm8,%%xmm9,%%xmm10,%%xmm11)
+#define INIT_m4n12 INIT_m4n8 unit_init_m4n4(%%xmm12,%%xmm13,%%xmm14,%%xmm15)
+#define SAVE_m4n1 "vfmadd213ps (%2),%%xmm0,%%xmm4; vmovups %%xmm4,(%2);"
+#define unit_save_m4n2(c1,c2) \
+    "vunpcklps "#c2","#c1",%%xmm2; vunpckhps "#c2","#c1",%%xmm3; vunpcklpd %%xmm3,%%xmm2,"#c1"; vunpckhpd %%xmm3,%%xmm2,"#c2";"\
+    "vfmadd213ps (%5),%%xmm0,"#c1"; vmovups "#c1",(%5);"\
+    "vfmadd213ps (%5,%3,1),%%xmm0,"#c2"; vmovups "#c2",(%5,%3,1);"\
+    "leaq (%5,%3,2),%5;"
+#define SAVE_m4n2 "movq %2,%5;" unit_save_m4n2(%%xmm4,%%xmm5)
+#define SAVE_m4n4  SAVE_m4n2  unit_save_m4n2(%%xmm6,%%xmm7)
+#define SAVE_m4n8  SAVE_m4n4  unit_save_m4n2(%%xmm8,%%xmm9)   unit_save_m4n2(%%xmm10,%%xmm11)
+#define SAVE_m4n12 SAVE_m4n8  unit_save_m4n2(%%xmm12,%%xmm13) unit_save_m4n2(%%xmm14,%%xmm15)
+
+/* m = 2 *//* xmm0 for alpha, xmm1-xmm3 and xmm10 for temporary use, xmm4-xmm9 for accumulators */
+#define INIT_m2n1 "vpxor %%xmm4,%%xmm4,%%xmm4;"
+#define KERNEL_k1m2n1 \
+    "vmovsd (%0),%%xmm1; addq $8,%0;"\
+    "vbroadcastss (%1),%%xmm2; vfmadd231ps %%xmm1,%%xmm2,%%xmm4;"\
+    "addq $4,%1;"
+#define SAVE_m2n1 "vmovsd (%2),%%xmm1; vfmadd213ps %%xmm1,%%xmm0,%%xmm4; vmovsd %%xmm4,(%2);"
+#define INIT_m2n2 INIT_m2n1 "vpxor %%xmm5,%%xmm5,%%xmm5;"
+#define KERNEL_k1m2n2 \
+    "vmovsd (%0),%%xmm1; addq $8,%0;"\
+    "vbroadcastss  (%1),%%xmm2; vfmadd231ps %%xmm1,%%xmm2,%%xmm4;"\
+    "vbroadcastss 4(%1),%%xmm3; vfmadd231ps %%xmm1,%%xmm3,%%xmm5;"\
+    "addq $8,%1;"
+#define SAVE_m2n2 SAVE_m2n1 "vmovsd (%2,%3,1),%%xmm1; vfmadd213ps %%xmm1,%%xmm0,%%xmm5; vmovsd %%xmm5,(%2,%3,1);"
+#define INIT_m2n4  INIT_m2n2
+#define INIT_m2n8  INIT_m2n4 "vpxor %%xmm6,%%xmm6,%%xmm6; vpxor %%xmm7,%%xmm7,%%xmm7;"
+#define INIT_m2n12 INIT_m2n8 "vpxor %%xmm8,%%xmm8,%%xmm8; vpxor %%xmm9,%%xmm9,%%xmm9;"
+#define KERNEL_k1m2n4 \
+    "vmovups (%1),%%xmm3; addq $16,%1;"\
+    "vbroadcastss  (%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm4;"\
+    "vbroadcastss 4(%0),%%xmm2; vfmadd231ps %%xmm3,%%xmm2,%%xmm5;"\
+    "addq $8,%0;"
+#define KERNEL_k1m2n8 \
+    "vmovups (%1),%%xmm3; vmovups (%1,%%r12,4),%%xmm2; addq $16,%1;"\
+    "vbroadcastss  (%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm4; vfmadd231ps %%xmm2,%%xmm1,%%xmm6;"\
+    "vbroadcastss 4(%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm5; vfmadd231ps %%xmm2,%%xmm1,%%xmm7;"\
+    "addq $8,%0;"
+#define KERNEL_k1m2n12 \
+    "vmovups (%1),%%xmm3; vmovups (%1,%%r12,4),%%xmm2; vmovups (%1,%%r12,8),%%xmm1; addq $16,%1;"\
+    "vbroadcastss  (%0),%%xmm10; vfmadd231ps %%xmm3,%%xmm10,%%xmm4; vfmadd231ps %%xmm2,%%xmm10,%%xmm6; vfmadd231ps %%xmm1,%%xmm10,%%xmm8;"\
+    "vbroadcastss 4(%0),%%xmm10; vfmadd231ps %%xmm3,%%xmm10,%%xmm5; vfmadd231ps %%xmm2,%%xmm10,%%xmm7; vfmadd231ps %%xmm1,%%xmm10,%%xmm9;"\
+    "addq $8,%0;"
+#define unit_save_m2n4(c1,c2) \
+    "vunpcklps "#c2","#c1",%%xmm1; vunpckhps "#c2","#c1",%%xmm2;"\
+    "vmovsd (%5),%%xmm3; vmovhpd (%5,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm1;"\
+    "vmovsd %%xmm1,(%5); vmovhpd %%xmm1,(%5,%3,1); leaq (%5,%3,2),%5;"\
+    "vmovsd (%5),%%xmm3; vmovhpd (%5,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm2;"\
+    "vmovsd %%xmm2,(%5); vmovhpd %%xmm2,(%5,%3,1); leaq (%5,%3,2),%5;"
+#define SAVE_m2n4 "movq %2,%5;" unit_save_m2n4(%%xmm4,%%xmm5)
+#define SAVE_m2n8   SAVE_m2n4   unit_save_m2n4(%%xmm6,%%xmm7)
+#define SAVE_m2n12  SAVE_m2n8   unit_save_m2n4(%%xmm8,%%xmm9)
+
+/* m = 1 *//* xmm0 for alpha, xmm1-xmm3 and xmm10 for temporary use, xmm4-xmm6 for accumulators */
+#define INIT_m1n1 "vpxor %%xmm4,%%xmm4,%%xmm4;"
+#define KERNEL_k1m1n1 \
+    "vmovss (%1),%%xmm3; addq $4,%1;"\
+    "vmovss (%0),%%xmm1; vfmadd231ss %%xmm3,%%xmm1,%%xmm4;"\
+    "addq $4,%0;"
+#define SAVE_m1n1 "vfmadd213ss (%2),%%xmm0,%%xmm4; vmovss %%xmm4,(%2);"
+#define INIT_m1n2 INIT_m1n1
+#define KERNEL_k1m1n2 \
+    "vmovsd (%1),%%xmm3; addq $8,%1;"\
+    "vbroadcastss  (%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm4;"\
+    "addq $4,%0;"
+#define SAVE_m1n2 \
+    "vmovss (%2),%%xmm3; vinsertps $16,(%2,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm4;"\
+    "vmovss %%xmm4,(%2); vextractps $1,%%xmm4,(%2,%3,1);"
+#define INIT_m1n4  INIT_m1n2
+#define INIT_m1n8  INIT_m1n4 "vpxor %%xmm5,%%xmm5,%%xmm5;"
+#define INIT_m1n12 INIT_m1n8 "vpxor %%xmm6,%%xmm6,%%xmm6;"
+#define KERNEL_k1m1n4 \
+    "vmovups (%1),%%xmm3; addq $16,%1;"\
+    "vbroadcastss  (%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm4;"\
+    "addq $4,%0;"
+#define KERNEL_k1m1n8 \
+    "vmovups (%1),%%xmm3; vmovups (%1,%%r12,4),%%xmm2; addq $16,%1;"\
+    "vbroadcastss  (%0),%%xmm1; vfmadd231ps %%xmm3,%%xmm1,%%xmm4; vfmadd231ps %%xmm2,%%xmm1,%%xmm5;"\
+    "addq $4,%0;"
+#define KERNEL_k1m1n12 \
+    "vmovups (%1),%%xmm3; vmovups (%1,%%r12,4),%%xmm2; vmovups (%1,%%r12,8),%%xmm1; addq $16,%1;"\
+    "vbroadcastss  (%0),%%xmm10; vfmadd231ps %%xmm3,%%xmm10,%%xmm4; vfmadd231ps %%xmm2,%%xmm10,%%xmm5; vfmadd231ps %%xmm1,%%xmm10,%%xmm6;"\
+    "addq $4,%0;"
+#define unit_save_m1n4(c1) \
+    "vpxor %%xmm10,%%xmm10,%%xmm10; vmovsd "#c1",%%xmm10,%%xmm2; vmovhlps "#c1",%%xmm10,%%xmm1;"\
+    "vmovss (%5),%%xmm3; vinsertps $16,(%5,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm2;"\
+    "vmovss %%xmm2,(%5); vextractps $1,%%xmm2,(%5,%3,1); leaq (%5,%3,2),%5;"\
+    "vmovss (%5),%%xmm3; vinsertps $16,(%5,%3,1),%%xmm3,%%xmm3; vfmadd213ps %%xmm3,%%xmm0,%%xmm1;"\
+    "vmovss %%xmm1,(%5); vextractps $1,%%xmm1,(%5,%3,1); leaq (%5,%3,2),%5;"
+#define SAVE_m1n4 "movq %2,%5;" unit_save_m1n4(%%xmm4)
+#define SAVE_m1n8  SAVE_m1n4    unit_save_m1n4(%%xmm5)
+#define SAVE_m1n12 SAVE_m1n8    unit_save_m1n4(%%xmm6)
+
+/* %0 = "+r"(a_pointer), %1 = "+r"(b_pointer), %2 = "+r"(c_pointer), %3 = "+r"(ldc_in_bytes), %4 for k_count, %5 for c_store, %6 = b_pref */
+/* r10 = tmp, r11 = m_counter, r12 = k << 2(const), r13 = tmp, r14 = b_head_pos(const), r15 = tmp */
+
+#define COMPUTE_SIMPLE(mdim,ndim) \
+    "movq %%r12,%4; sarq $2,%4; movq %%r14,%1;" INIT_m##mdim##n##ndim\
+    "testq %4,%4; jz 7"#mdim"7"#ndim"2f;"\
+    "7"#mdim"7"#ndim"1:\n\t"\
+    KERNEL_k1m##mdim##n##ndim "decq %4; jnz 7"#mdim"7"#ndim"1b;"\
+    "7"#mdim"7"#ndim"2:\n\t"\
+    SAVE_m##mdim##n##ndim "addq $"#mdim"*4,%2;"
+#define COMPUTE_m8n1 COMPUTE_SIMPLE(8,1)
+#define COMPUTE_m8n2 COMPUTE_SIMPLE(8,2)
+#define COMPUTE_m8n8 COMPUTE_SIMPLE(8,8)
+#define COMPUTE_m8n12 COMPUTE_SIMPLE(8,12)
+#define COMPUTE_m8n4 \
+    "movq %%r12,%4; sarq $2,%4; movq %%r14,%1;" INIT_m8n4\
+    "cmpq $8,%4; jb 78740f;"\
+    "78749:\n\t"\
+    KERNEL_k2m8n4 KERNEL_k2m8n4 KERNEL_k2m8n4 KERNEL_k2m8n4\
+    "subq $8,%4; cmpq $8,%4; jnb 78749b;"\
+    "78740:\n\t"\
+    "testq %4,%4; jz 78742f;"\
+    "78741:\n\t"\
+    KERNEL_k1m8n4 "decq %4; jnz 78741b;"\
+    "78742:\n\t"\
+    SAVE_m8n4 "addq $32,%2;"
+#define COMPUTE_L_m16n6 \
+    "movq %%r12,%%r13; sarq $2,%%r13; movq %%r14,%1;" INIT_m16n6\
+    "movq %%r13,%4; movq %2,%5; cmpq $16,%%r13; jb 7116762f; movq $14,%4;"\
+    "7116761:\n\t"\
+    KERNEL_L_k2m16n6 "prefetcht0 128(%1); testq $24,%4; movq $84,%%r15; cmovz %3,%%r15;"\
+    KERNEL_L_k2m16n6 "prefetcht1 (%5); subq $63,%5; addq %%r15,%5;"\
+    KERNEL_L_k2m16n6 "prefetcht0 128(%1); prefetcht1 (%6); cmpq $198,%4; cmoveq %2,%5;"\
+    KERNEL_L_k2m16n6 "addq $16,%6; addq $8,%4; cmpq %4,%%r13; jnb 7116761b;"\
+    "movq %2,%5; negq %4; leaq 14(%%r13,%4,1),%4;"\
+    "7116762:\n\t"\
+    "xorq %%r15,%%r15; testq %4,%4; jz 7116764f;"\
+    "7116763:\n\t"\
+    "prefetcht0 (%5); prefetcht0 63(%5); addq %3,%5; incq %%r15;"\
+    KERNEL_L_k1m16n6 "cmpq $6,%%r15; cmoveq %2,%5; decq %4; jnz 7116763b;"\
+    "7116764:\n\t"\
+    SAVE_L_m16n6 "addq $32,%2;"
+#define COMPUTE_R_m16n6 \
+    "movq %%r12,%%r13; sarq $2,%%r13; movq %%r14,%1;" INIT_m16n6\
+    "movq %%r13,%4; leaq (%2,%3,4),%5; leaq (%5,%3,2),%5; movq %5,%%r10; cmpq $16,%%r13; jb 7216762f; movq $14,%4;"\
+    "7216761:\n\t"\
+    KERNEL_R_k2m16n6 "prefetcht0 128(%1,%%r12,8); testq $24,%4; movq $84,%%r15; cmovz %3,%%r15;"\
+    KERNEL_R_k2m16n6 "prefetcht1 (%5); subq $63,%5; addq %%r15,%5;"\
+    KERNEL_R_k2m16n6 "prefetcht0 128(%1,%%r12,8); prefetcht1 (%6); cmpq $198,%4; cmoveq %%r10,%5;"\
+    KERNEL_R_k2m16n6 "addq $16,%6; addq $8,%4; cmpq %4,%%r13; jnb 7216761b;"\
+    "movq %%r10,%5; negq %4; leaq 14(%%r13,%4,1),%4;"\
+    "7216762:\n\t"\
+    "xorq %%r15,%%r15; testq %4,%4; jz 7216764f;"\
+    "7216763:\n\t"\
+    "prefetcht0 (%5); prefetcht0 63(%5); addq %3,%5; incq %%r15;"\
+    KERNEL_R_k1m16n6 "cmpq $6,%%r15; cmoveq %%r10,%5; decq %4; jnz 7216763b;"\
+    "7216764:\n\t"\
+    "prefetcht0 (%%r14); prefetcht0 64(%%r14);" SAVE_R_m16n6 "addq $32,%2;"
+#define COMPUTE_H_m8n6 \
+    "movq %%r12,%4; sarq $2,%4; movq %%r14,%1;" INIT_m8n6\
+    "cmpq $8,%4; jb 718760f; movq %2,%5; xorq %%r15,%%r15;"\
+    "718769:\n\t"\
+    KERNEL_L_k2m8n6 KERNEL_L_k2m8n6 "cmpq $62,%%r15; movq $62,%%r15; cmoveq %3,%%r15;"\
+    KERNEL_L_k2m8n6 KERNEL_L_k2m8n6 "prefetcht2 (%5); leaq -31(%5,%%r15,1),%5;"\
+    "subq $8,%4; cmpq $8,%4; jnb 718769b;"\
+    "718760:\n\t"\
+    "testq %4,%4; jz 718762f;"\
+    "718761:\n\t"\
+    KERNEL_L_k1m8n6 "decq %4; jnz 718761b;"\
+    "718762:\n\t"\
+    SAVE_L_m8n6 "negq %%r12; leaq (%0,%%r12,8),%0; negq %%r12;"
+#define COMPUTE_T_m8n6(side,sim) \
+    "movq %%r12,%4; sarq $2,%4; movq %%r14,%1;" INIT_m8n6\
+    "cmpq $8,%4; jb 72"#sim"8760f;"\
+    "72"#sim"8769:\n\t"\
+    KERNEL_##side##_k2m8n6 KERNEL_##side##_k2m8n6 KERNEL_##side##_k2m8n6 KERNEL_##side##_k2m8n6\
+    "subq $8,%4; cmpq $8,%4; jnb 72"#sim"8769b;"\
+    "72"#sim"8760:\n\t"\
+    "testq %4,%4; jz 72"#sim"8762f;"\
+    "72"#sim"8761:\n\t"\
+    KERNEL_##side##_k1m8n6 "decq %4; jnz 72"#sim"8761b;"\
+    "72"#sim"8762:\n\t"\
+    SAVE_##side##_m8n6 "addq $32,%2;"
+#define COMPUTE_NORMAL(ndim) {\
+    next_b = b_pointer + ndim * K;\
+    __asm__ __volatile__(\
+    "vbroadcastss %9,%%ymm0;"\
+    "movq %8,%%r12; salq $2,%%r12; movq %1,%%r14; movq %7,%%r11;"\
+    "cmpq $8,%%r11;jb 33101"#ndim"f;"\
+    "33109"#ndim":\n\t"\
+    COMPUTE_m8n##ndim\
+    "subq $8,%%r11;cmpq $8,%%r11;jnb 33109"#ndim"b;"\
+    "33101"#ndim":\n\t"\
+    "cmpq $4,%%r11;jb 33103"#ndim"f;"\
+    COMPUTE_SIMPLE(4,ndim) "subq $4,%%r11;"\
+    "33103"#ndim":\n\t"\
+    "cmpq $2,%%r11;jb 33104"#ndim"f;"\
+    COMPUTE_SIMPLE(2,ndim) "subq $2,%%r11;"\
+    "33104"#ndim":\n\t"\
+    "testq %%r11,%%r11;jz 33105"#ndim"f;"\
+    COMPUTE_SIMPLE(1,ndim)\
+    "33105"#ndim":\n\t"\
+    "movq %%r14,%1; vzeroupper;"\
+    :"+r"(a_pointer),"+r"(b_pointer),"+r"(c_pointer),"+r"(ldc_in_bytes),"+r"(k_count),"+r"(ctemp),"+r"(next_b)\
+    :"m"(M),"m"(K),"m"(ALPHA):"r10","r11","r12","r13","r14","r15",\
+    "xmm0","xmm1","xmm2","xmm3","xmm4","xmm5","xmm6","xmm7","xmm8","xmm9","xmm10","xmm11","xmm12","xmm13","xmm14","xmm15","cc","memory");\
+    a_pointer -= M * K; b_pointer += ndim * K; c_pointer += (LDC * ndim - M);\
+}
+#define COMPUTE_n12 {\
+    next_b = b_pointer + 12 * K;\
+    __asm__ __volatile__(\
+    "vbroadcastss %9,%%ymm0;"\
+    "movq %8,%%r12; salq $2,%%r12; movq %1,%%r14; movq %7,%%r11;"\
+    "cmpq $16,%%r11;jb 3310112f;"\
+    COMPUTE_H_m8n6\
+    "3310612:\n\t"\
+    COMPUTE_R_m16n6 "subq $8,%%r11; cmpq $16,%%r11;jb 3310712f;"\
+    COMPUTE_L_m16n6 "subq $8,%%r11; cmpq $16,%%r11;jnb 3310612b;"\
+    COMPUTE_T_m8n6(R,5) "subq $8,%%r11; jmp 3310212f;"\
+    "3310712:\n\t"\
+    COMPUTE_T_m8n6(L,7) "subq $8,%%r11; jmp 3310212f;"\
+    "3310112:\n\t"\
+    "cmpq $8,%%r11;jb 3310212f;"\
+    COMPUTE_SIMPLE(8,12) "subq $8,%%r11;"\
+    "3310212:\n\t"\
+    "cmpq $4,%%r11;jb 3310312f;"\
+    COMPUTE_SIMPLE(4,12) "subq $4,%%r11;"\
+    "3310312:\n\t"\
+    "cmpq $2,%%r11;jb 3310412f;"\
+    COMPUTE_SIMPLE(2,12) "subq $2,%%r11;"\
+    "3310412:\n\t"\
+    "testq %%r11,%%r11;jz 3310512f;"\
+    COMPUTE_SIMPLE(1,12)\
+    "3310512:\n\t"\
+    "movq %%r14,%1; vzeroupper;"\
+    :"+r"(a_pointer),"+r"(b_pointer),"+r"(c_pointer),"+r"(ldc_in_bytes),"+r"(k_count),"+r"(ctemp),"+r"(next_b)\
+    :"m"(M),"m"(K),"m"(ALPHA):"r10","r11","r12","r13","r14","r15",\
+    "xmm0","xmm1","xmm2","xmm3","xmm4","xmm5","xmm6","xmm7","xmm8","xmm9","xmm10","xmm11","xmm12","xmm13","xmm14","xmm15","cc","memory");\
+    a_pointer -= M * K; b_pointer += 12 * K; c_pointer += (LDC * 12 - M);\
+}
+
+#include "common.h"
+#include <stdint.h>
+int __attribute__ ((noinline))
+CNAME(BLASLONG m, BLASLONG n, BLASLONG k, float alpha, float * __restrict__ A, float * __restrict__ B, float * __restrict__ C, BLASLONG LDC){
+    if(m==0||n==0||k==0||alpha==(float)0.0) return 0;
+    int64_t ldc_in_bytes = (int64_t)LDC * sizeof(float);
+    float ALPHA = alpha;
+    int64_t M = (int64_t)m, K = (int64_t)k, k_count = 0;
+    BLASLONG n_count = n;
+    float *a_pointer = A,*b_pointer = B,*c_pointer = C,*ctemp = C,*next_b = B;
+    for(;n_count>11;n_count-=12) COMPUTE_n12
+    for(;n_count>7;n_count-=8) COMPUTE_NORMAL(8)
+    for(;n_count>3;n_count-=4) COMPUTE_NORMAL(4)
+    for(;n_count>1;n_count-=2) COMPUTE_NORMAL(2)
+    if(n_count>0) COMPUTE_NORMAL(1)
+    return 0;
+}
+
--- a/kernel/x86_64/zgemm_kernel_4x2_haswell.c
+++ b/kernel/x86_64/zgemm_kernel_4x2_haswell.c
@ -50,7 +50,7 @@
  "vmovupd (%0),%%ymm0; vmovupd 32(%0),%%ymm1; prefetcht0 512(%0); addq $64,%0;"\
  acc_m4n2_con(0,1,4,5,6,7,0,16,%1) acc_m4n2_con(0,1,8,9,10,11,0,16,%1,%%r12,1)
 #define KERNEL_2_k1m4n4 \
-  "vpermilpd $5,%%ymm0,%%ymm0; vpermilpd $5,%%ymm1,%%ymm1;"\
+  "vpermilpd $5,-64(%0),%%ymm0; vpermilpd $5,-32(%0),%%ymm1;"\
  acc_m4n2_con(0,1,4,5,6,7,8,24,%1) acc_m4n2_con(0,1,8,9,10,11,8,24,%1,%%r12,1)
 #define KERNEL_1_k1m4n6 KERNEL_1_k1m4n4 acc_m4n2_con(0,1,12,13,14,15,0,16,%1,%%r12,2)
 #define KERNEL_2_k1m4n6 KERNEL_2_k1m4n4 acc_m4n2_con(0,1,12,13,14,15,8,24,%1,%%r12,2)
@ -93,9 +93,9 @@
  "movq $10,%5; movq $84,%%r15;"\
  #ndim"4441:\n\t"\
  "prefetcht1 (%3); subq $63,%3; addq %%r15,%3;"\
-  "prefetcht0 96(%1); prefetcht0 96(%1,%%r12,1); prefetcht0 96(%1,%%r12,2);" KERNEL_k1m4n##ndim KERNEL_k1m4n##ndim\
+  KERNEL_k1m4n##ndim KERNEL_k1m4n##ndim\
  "testq $12,%5; movq $84,%%r15; cmovz %4,%%r15; prefetcht1 (%8); addq $16,%8;"\
-  "prefetcht0 96(%1); prefetcht0 96(%1,%%r12,1); prefetcht0 96(%1,%%r12,2);" KERNEL_k1m4n##ndim KERNEL_k1m4n##ndim\
+  KERNEL_k1m4n##ndim KERNEL_k1m4n##ndim\
  "addq $4,%5; cmpq %5,%%r13; jnb "#ndim"4441b;"\
  "movq %2,%3; negq %5; leaq 10(%%r13,%5,1),%5; prefetcht0 (%6); prefetcht0 15(%6);"\
  #ndim"4442:\n\t"\
--- a/param.h
+++ b/param.h
@ -1722,16 +1722,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define XGEMM_DEFAULT_R xgemm_r
 #define XGEMM_DEFAULT_Q 128

-#define CGEMM3M_DEFAULT_UNROLL_N 8
-#define CGEMM3M_DEFAULT_UNROLL_M 4
-#define ZGEMM3M_DEFAULT_UNROLL_N 8
-#define ZGEMM3M_DEFAULT_UNROLL_M 2
+#define CGEMM3M_DEFAULT_UNROLL_N 4
+#define CGEMM3M_DEFAULT_UNROLL_M 8
+#define ZGEMM3M_DEFAULT_UNROLL_N 4
+#define ZGEMM3M_DEFAULT_UNROLL_M 4

-#define CGEMM3M_DEFAULT_P 448
-#define ZGEMM3M_DEFAULT_P 224
+#define CGEMM3M_DEFAULT_P 320
+#define ZGEMM3M_DEFAULT_P 256
 #define XGEMM3M_DEFAULT_P 112
-#define CGEMM3M_DEFAULT_Q 224
-#define ZGEMM3M_DEFAULT_Q 224
+#define CGEMM3M_DEFAULT_Q 320
+#define ZGEMM3M_DEFAULT_Q 256
 #define XGEMM3M_DEFAULT_Q 224
 #define CGEMM3M_DEFAULT_R 12288
 #define ZGEMM3M_DEFAULT_R 12288
@ -2620,7 +2620,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 /*FIXME: this should be using the cache size, but there is currently no easy way to
 query that on ARM. So if getarch counted more than 8 cores we simply assume the host
 is a big desktop or server with abundant cache rather than a phone or embedded device */ 
-#if NUM_CORES > 8
+#if NUM_CORES > 8 || defined(TSV110) || defined(EMAG8180)
  #define SGEMM_DEFAULT_P 512
  #define DGEMM_DEFAULT_P 256
  #define CGEMM_DEFAULT_P 256
@ -2705,6 +2705,35 @@ is a big desktop or server with abundant cache rather than a phone or embedded d
 #define CGEMM_DEFAULT_R 4096
 #define ZGEMM_DEFAULT_R 4096

+#elif defined(NEOVERSEN1)
+
+#define SGEMM_DEFAULT_UNROLL_M  16
+#define SGEMM_DEFAULT_UNROLL_N  4
+
+#define DGEMM_DEFAULT_UNROLL_M  8
+#define DGEMM_DEFAULT_UNROLL_N  4
+
+#define CGEMM_DEFAULT_UNROLL_M  8
+#define CGEMM_DEFAULT_UNROLL_N  4
+
+#define ZGEMM_DEFAULT_UNROLL_M  4
+#define ZGEMM_DEFAULT_UNROLL_N  4
+
+#define SGEMM_DEFAULT_P	128
+#define DGEMM_DEFAULT_P	160
+#define CGEMM_DEFAULT_P 128
+#define ZGEMM_DEFAULT_P 128
+
+#define SGEMM_DEFAULT_Q 352
+#define DGEMM_DEFAULT_Q 128
+#define CGEMM_DEFAULT_Q 224
+#define ZGEMM_DEFAULT_Q 112
+
+#define SGEMM_DEFAULT_R 4096
+#define DGEMM_DEFAULT_R 4096
+#define CGEMM_DEFAULT_R 4096
+#define ZGEMM_DEFAULT_R 4096
+
 #else // Other/undetected ARMv8 cores

 #define SGEMM_DEFAULT_UNROLL_M  16