From b06550519eb6a02d87f73a48b73fcef2fdedb9c9 Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Tue, 12 Aug 2014 12:15:41 +0200
Subject: [PATCH] added optimized cgemv_t c-kernel

---
 kernel/x86_64/cgemv_t.c                | 269 +++++++++++++++++++++++++
 kernel/x86_64/zgemv_n_microk_sandy-2.c | 161 +++++++++++++++
 2 files changed, 430 insertions(+)
 create mode 100644 kernel/x86_64/cgemv_t.c
 create mode 100644 kernel/x86_64/zgemv_n_microk_sandy-2.c

diff --git a/kernel/x86_64/cgemv_t.c b/kernel/x86_64/cgemv_t.c
new file mode 100644
index 000000000..ccdf13a57
--- /dev/null
+++ b/kernel/x86_64/cgemv_t.c
@@ -0,0 +1,269 @@
+/***************************************************************************
+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 "common.h"
+
+/*
+#if defined(BULLDOZER)
+#include "zgemv_t_microk_bulldozer-2.c"
+#elif defined(HASWELL)
+#include "zgemv_t_microk_haswell-2.c"
+#endif
+*/
+
+#define NBMAX 2048
+
+#ifndef HAVE_KERNEL_16x4
+
+static void zgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+	FLOAT temp_r0 = 0.0;
+	FLOAT temp_r1 = 0.0;
+	FLOAT temp_r2 = 0.0;
+	FLOAT temp_r3 = 0.0;
+	FLOAT temp_i0 = 0.0;
+	FLOAT temp_i1 = 0.0;
+	FLOAT temp_i2 = 0.0;
+	FLOAT temp_i3 = 0.0;
+
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		temp_r0 += a0[i]*x[i]   - a0[i+1]*x[i+1];		
+		temp_i0 += a0[i]*x[i+1] + a0[i+1]*x[i];		
+		temp_r1 += a1[i]*x[i]   - a1[i+1]*x[i+1];		
+		temp_i1 += a1[i]*x[i+1] + a1[i+1]*x[i];		
+		temp_r2 += a2[i]*x[i]   - a2[i+1]*x[i+1];		
+		temp_i2 += a2[i]*x[i+1] + a2[i+1]*x[i];		
+		temp_r3 += a3[i]*x[i]   - a3[i+1]*x[i+1];		
+		temp_i3 += a3[i]*x[i+1] + a3[i+1]*x[i];		
+#else
+		temp_r0 += a0[i]*x[i]   + a0[i+1]*x[i+1];		
+		temp_i0 += a0[i]*x[i+1] - a0[i+1]*x[i];		
+		temp_r1 += a1[i]*x[i]   + a1[i+1]*x[i+1];		
+		temp_i1 += a1[i]*x[i+1] - a1[i+1]*x[i];		
+		temp_r2 += a2[i]*x[i]   + a2[i+1]*x[i+1];		
+		temp_i2 += a2[i]*x[i+1] - a2[i+1]*x[i];		
+		temp_r3 += a3[i]*x[i]   + a3[i+1]*x[i+1];		
+		temp_i3 += a3[i]*x[i+1] - a3[i+1]*x[i];		
+#endif
+	}
+	y[0] = temp_r0;
+	y[1] = temp_i0;
+	y[2] = temp_r1;
+	y[3] = temp_i1;
+	y[4] = temp_r2;
+	y[5] = temp_i2;
+	y[6] = temp_r3;
+	y[7] = temp_i3;
+}
+	
+#endif
+
+static void zgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+	FLOAT temp_r = 0.0;
+	FLOAT temp_i = 0.0;
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		temp_r += a0[i]*x[i]   - a0[i+1]*x[i+1];		
+		temp_i += a0[i]*x[i+1] + a0[i+1]*x[i];		
+#else
+		temp_r += a0[i]*x[i]   + a0[i+1]*x[i+1];		
+		temp_i += a0[i]*x[i+1] - a0[i+1]*x[i];		
+#endif
+	}
+	*y      = temp_r;
+	*(y+1)  = temp_i;
+}
+	
+static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
+{
+        BLASLONG i;
+        for ( i=0; i<n; i++ )
+        {
+                *dest     = *src;
+                *(dest+1) = *(src+1);
+                dest+=2;
+                src += inc_src;
+        }
+}
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[8];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT ybuffer[8],*xbuffer;
+
+        inc_x *= 2;
+        inc_y *= 2;
+        lda   *= 2;
+
+	xbuffer = buffer;
+	
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		y_ptr = y;
+		a_ptr = a;
+		x_ptr = x;
+		copy_x(NB,x_ptr,xbuffer,inc_x);
+		for( i = 0; i < n1 ; i++)
+		{
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			zgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+
+#if !defined(XCONJ)
+			y_ptr[0] += alpha_r * ybuffer[0] - alpha_i * ybuffer[1];
+			y_ptr[1] += alpha_r * ybuffer[1] + alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[2] - alpha_i * ybuffer[3];
+			y_ptr[1] += alpha_r * ybuffer[3] + alpha_i * ybuffer[2];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[4] - alpha_i * ybuffer[5];
+			y_ptr[1] += alpha_r * ybuffer[5] + alpha_i * ybuffer[4];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[6] - alpha_i * ybuffer[7];
+			y_ptr[1] += alpha_r * ybuffer[7] + alpha_i * ybuffer[6];
+			y_ptr  += inc_y;
+#else
+			y_ptr[0] += alpha_r * ybuffer[0] + alpha_i * ybuffer[1];
+			y_ptr[1] -= alpha_r * ybuffer[1] - alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[2] + alpha_i * ybuffer[3];
+			y_ptr[1] -= alpha_r * ybuffer[3] - alpha_i * ybuffer[2];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[4] + alpha_i * ybuffer[5];
+			y_ptr[1] -= alpha_r * ybuffer[5] - alpha_i * ybuffer[4];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[6] + alpha_i * ybuffer[7];
+			y_ptr[1] -= alpha_r * ybuffer[7] - alpha_i * ybuffer[6];
+			y_ptr  += inc_y;
+#endif
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			zgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+
+#if !defined(XCONJ)
+			y_ptr[0] += alpha_r * ybuffer[0] - alpha_i * ybuffer[1];
+			y_ptr[1] += alpha_r * ybuffer[1] + alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+#else
+			y_ptr[0] += alpha_r * ybuffer[0] + alpha_i * ybuffer[1];
+			y_ptr[1] -= alpha_r * ybuffer[1] - alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+#endif
+
+		}
+		a += 2* NB;
+		x += NB * inc_x;	
+	}
+
+	BLASLONG m3 = m % 16;
+	if ( m3 == 0 ) return(0);
+
+	x_ptr = x;
+	copy_x(m3,x_ptr,xbuffer,inc_x);
+	j=0;
+	a_ptr = a;
+	y_ptr = y;
+	while ( j < n)
+	{
+		FLOAT temp_r = 0.0;
+		FLOAT temp_i = 0.0;
+		for( i = 0; i < m3*2; i+=2 )
+		{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+			temp_r += a_ptr[i] * xbuffer[i]   - a_ptr[i+1] * xbuffer[i+1];
+			temp_i += a_ptr[i] * xbuffer[i+1] + a_ptr[i+1] * xbuffer[i];
+#else
+			temp_r += a_ptr[i] * xbuffer[i]   + a_ptr[i+1] * xbuffer[i+1];
+			temp_i += a_ptr[i] * xbuffer[i+1] - a_ptr[i+1] * xbuffer[i];
+#endif
+		}
+		a_ptr += lda;
+
+#if !defined(XCONJ) 
+                y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+                y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+                y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+                y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+
+		y_ptr += inc_y;
+		j++;
+	}
+	return(0);
+}
+
+
diff --git a/kernel/x86_64/zgemv_n_microk_sandy-2.c b/kernel/x86_64/zgemv_n_microk_sandy-2.c
new file mode 100644
index 000000000..8061ed4fa
--- /dev/null
+++ b/kernel/x86_64/zgemv_n_microk_sandy-2.c
@@ -0,0 +1,161 @@
+/***************************************************************************
+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.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vbroadcastsd	  (%2), %%ymm0                  \n\t"  // real part x0
+	"vbroadcastsd	 8(%2), %%ymm1                  \n\t"  // imag part x0
+	"vbroadcastsd	16(%2), %%ymm2                  \n\t"  // real part x1
+	"vbroadcastsd	24(%2), %%ymm3                  \n\t"  // imag part x1
+	"vbroadcastsd	32(%2), %%ymm4                  \n\t"  // real part x2
+	"vbroadcastsd	40(%2), %%ymm5                  \n\t"  // imag part x2
+	"vbroadcastsd	48(%2), %%ymm6                  \n\t"  // real part x3
+	"vbroadcastsd	56(%2), %%ymm7                  \n\t"  // imag part x3
+
+
+	".align 16				        \n\t"
+	".L01LOOP%=:				        \n\t"
+
+	"vxorpd		%%ymm12, %%ymm12, %%ymm12	\n\t"
+	"vxorpd		%%ymm13, %%ymm13, %%ymm13	\n\t"
+	"vxorpd		%%ymm14, %%ymm14, %%ymm14	\n\t"
+	"vxorpd		%%ymm15, %%ymm15, %%ymm15	\n\t"
+
+	"vmovups	(%4,%0,8), %%ymm8	        \n\t" // 2 complex values form a0
+	"vmovups      32(%4,%0,8), %%ymm9	        \n\t" // 2 complex values form a0
+
+	"vmulpd		  %%ymm8 , %%ymm0 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm12, %%ymm10, %%ymm12	\n\t"
+	"vmulpd		  %%ymm8 , %%ymm1 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm13, %%ymm11, %%ymm13	\n\t"
+	"vmulpd		  %%ymm9 , %%ymm0 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm14, %%ymm10, %%ymm14	\n\t"
+	"vmulpd		  %%ymm9 , %%ymm1 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm15, %%ymm11, %%ymm15	\n\t"
+
+	"vmovups	(%5,%0,8), %%ymm8	        \n\t" // 2 complex values form a0
+	"vmovups      32(%5,%0,8), %%ymm9	        \n\t" // 2 complex values form a0
+
+	"vmulpd		  %%ymm8 , %%ymm2 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm12, %%ymm10, %%ymm12	\n\t"
+	"vmulpd		  %%ymm8 , %%ymm3 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm13, %%ymm11, %%ymm13	\n\t"
+	"vmulpd		  %%ymm9 , %%ymm2 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm14, %%ymm10, %%ymm14	\n\t"
+	"vmulpd		  %%ymm9 , %%ymm3 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm15, %%ymm11, %%ymm15	\n\t"
+
+	"vmovups	(%6,%0,8), %%ymm8	        \n\t" // 2 complex values form a0
+	"vmovups      32(%6,%0,8), %%ymm9	        \n\t" // 2 complex values form a0
+
+	"vmulpd		  %%ymm8 , %%ymm4 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm12, %%ymm10, %%ymm12	\n\t"
+	"vmulpd		  %%ymm8 , %%ymm5 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm13, %%ymm11, %%ymm13	\n\t"
+	"vmulpd		  %%ymm9 , %%ymm4 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm14, %%ymm10, %%ymm14	\n\t"
+	"vmulpd		  %%ymm9 , %%ymm5 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm15, %%ymm11, %%ymm15	\n\t"
+
+	"vmovups	(%7,%0,8), %%ymm8	        \n\t" // 2 complex values form a0
+	"vmovups      32(%7,%0,8), %%ymm9	        \n\t" // 2 complex values form a0
+
+	"vmulpd		  %%ymm8 , %%ymm6 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm12, %%ymm10, %%ymm12	\n\t"
+	"vmulpd		  %%ymm8 , %%ymm7 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm13, %%ymm11, %%ymm13	\n\t"
+	"vmulpd		  %%ymm9 , %%ymm6 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm14, %%ymm10, %%ymm14	\n\t"
+	"vmulpd		  %%ymm9 , %%ymm7 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm15, %%ymm11, %%ymm15	\n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilpd      $0x5 , %%ymm13, %%ymm13               \n\t"
+        "vpermilpd      $0x5 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubpd      %%ymm13, %%ymm12, %%ymm8              \n\t"
+        "vaddsubpd      %%ymm15, %%ymm14, %%ymm9              \n\t"
+#else
+        "vpermilpd      $0x5 , %%ymm12, %%ymm12               \n\t"
+        "vpermilpd      $0x5 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubpd      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vaddsubpd      %%ymm14, %%ymm15, %%ymm9              \n\t"
+        "vpermilpd      $0x5 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilpd      $0x5 , %%ymm9 , %%ymm9                \n\t"
+#endif
+
+	"prefetcht0      192(%3,%0,8)			\n\t"
+	"vmovups	  (%3,%0,8),  %%ymm12           \n\t"
+	"vmovups	32(%3,%0,8),  %%ymm13           \n\t"
+
+#if !defined(XCONJ)
+        "vaddpd         %%ymm8, %%ymm12, %%ymm12              \n\t"
+        "vaddpd         %%ymm9, %%ymm13, %%ymm13              \n\t"
+#else
+        "vaddsubpd              %%ymm12, %%ymm8, %%ymm12              \n\t"
+        "vaddsubpd              %%ymm13, %%ymm9, %%ymm13              \n\t"
+#endif
+
+
+	"vmovups  %%ymm12,   (%3,%0,8)		        \n\t" // 2 complex values to y	
+	"vmovups  %%ymm13, 32(%3,%0,8)		        \n\t"	
+
+        "addq		$8 , %0	  	 	        \n\t"
+	"subq	        $4 , %1			        \n\t"		
+	"jnz		.L01LOOP%=		        \n\t"
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+