From 11eab4c0199f85ece37453f351ffee6450bf8c7c Mon Sep 17 00:00:00 2001
From: wernsaar <wernsaar@googlemail.com>
Date: Thu, 14 Aug 2014 19:00:30 +0200
Subject: [PATCH] added optimized cgemv_n for haswell

---
 kernel/x86_64/KERNEL.HASWELL             |   2 +
 kernel/x86_64/cgemv_n.c                  | 255 +++++++++++++++++++++++++++++++
 kernel/x86_64/cgemv_n_microk_haswell-2.c | 137 +++++++++++++++++
 3 files changed, 394 insertions(+)
 create mode 100644 kernel/x86_64/cgemv_n.c
 create mode 100644 kernel/x86_64/cgemv_n_microk_haswell-2.c

diff --git a/kernel/x86_64/KERNEL.HASWELL b/kernel/x86_64/KERNEL.HASWELL
index e07448a..d0ac9c7 100644
--- a/kernel/x86_64/KERNEL.HASWELL
+++ b/kernel/x86_64/KERNEL.HASWELL
@@ -1,11 +1,13 @@
 SGEMVNKERNEL = sgemv_n.c
 SGEMVTKERNEL = sgemv_t.c
+
 DGEMVNKERNEL = dgemv_n.c
 DGEMVTKERNEL = dgemv_t.c
 
 ZGEMVNKERNEL = zgemv_n.c
 ZGEMVTKERNEL = zgemv_t.c
 
+CGEMVNKERNEL = cgemv_n.c
 CGEMVTKERNEL = cgemv_t.c
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
diff --git a/kernel/x86_64/cgemv_n.c b/kernel/x86_64/cgemv_n.c
new file mode 100644
index 0000000..47ef0d4
--- /dev/null
+++ b/kernel/x86_64/cgemv_n.c
@@ -0,0 +1,255 @@
+/***************************************************************************
+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 <stdlib.h>
+#include <stdio.h>
+#include "common.h"
+
+#if defined(HASWELL)
+#include "cgemv_n_microk_haswell-2.c"
+#endif
+
+
+#define NBMAX 2048
+
+#ifndef HAVE_KERNEL_16x4
+
+static void cgemv_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];
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		y[i]   += a0[i]*x[0] - a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
+		y[i]   += a1[i]*x[2] - a1[i+1] * x[3];
+		y[i+1] += a1[i]*x[3] + a1[i+1] * x[2];
+		y[i]   += a2[i]*x[4] - a2[i+1] * x[5];
+		y[i+1] += a2[i]*x[5] + a2[i+1] * x[4];
+		y[i]   += a3[i]*x[6] - a3[i+1] * x[7];
+		y[i+1] += a3[i]*x[7] + a3[i+1] * x[6];
+#else 
+		y[i]   += a0[i]*x[0] + a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
+		y[i]   += a1[i]*x[2] + a1[i+1] * x[3];
+		y[i+1] += a1[i]*x[3] - a1[i+1] * x[2];
+		y[i]   += a2[i]*x[4] + a2[i+1] * x[5];
+		y[i+1] += a2[i]*x[5] - a2[i+1] * x[4];
+		y[i]   += a3[i]*x[6] + a3[i+1] * x[7];
+		y[i+1] += a3[i]*x[7] - a3[i+1] * x[6];
+#endif
+	}
+}
+	
+#endif
+
+static void cgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		y[i]   += a0[i]*x[0] - a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
+#else 
+		y[i]   += a0[i]*x[0] + a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
+#endif
+
+	}
+}
+	
+
+static void zero_y(BLASLONG n, FLOAT *dest)
+{
+	BLASLONG i;
+	for ( i=0; i<2*n; i++ )
+	{
+		*dest = 0.0;
+		dest++;
+	}
+}
+
+
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)
+{
+	BLASLONG i;
+	FLOAT temp_r;
+	FLOAT temp_i;
+	for ( i=0; i<n; i++ )
+	{
+#if !defined(XCONJ) 
+		temp_r = alpha_r * src[0] - alpha_i * src[1];
+		temp_i = alpha_r * src[1] + alpha_i * src[0];
+#else
+		temp_r =  alpha_r * src[0] + alpha_i * src[1];
+		temp_i = -alpha_r * src[1] + alpha_i * src[0];
+#endif
+
+		*dest += temp_r;
+		*(dest+1) += temp_i;
+
+		src+=2;
+		dest += inc_dest;
+	}
+}
+
+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[4];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT xbuffer[8],*ybuffer;
+
+
+#if 0
+printf("%s %d %d %.16f %.16f %d %d %d\n","zgemv_n",m,n,alpha_r,alpha_i,lda,inc_x,inc_y);
+#endif
+
+	if ( m < 1 ) return(0);
+	if ( n < 1 ) return(0);
+
+	ybuffer = buffer;
+	
+	inc_x *= 2;
+	inc_y *= 2;
+	lda   *= 2;
+
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+	y_ptr = y;
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		a_ptr = a;
+		x_ptr = x;
+		zero_y(NB,ybuffer);
+		for( i = 0; i < n1 ; i++)
+		{
+
+			xbuffer[0] = x_ptr[0];
+			xbuffer[1] = x_ptr[1];
+			x_ptr += inc_x;	
+			xbuffer[2] = x_ptr[0];
+			xbuffer[3] = x_ptr[1];
+			x_ptr += inc_x;	
+			xbuffer[4] = x_ptr[0];
+			xbuffer[5] = x_ptr[1];
+			x_ptr += inc_x;	
+			xbuffer[6] = x_ptr[0];
+			xbuffer[7] = x_ptr[1];
+			x_ptr += inc_x;	
+
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			cgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			xbuffer[0] = x_ptr[0];
+			xbuffer[1] = x_ptr[1];
+			x_ptr += inc_x;	
+			cgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+
+		}
+		add_y(NB,ybuffer,y_ptr,inc_y,alpha_r,alpha_i);
+		a     += 2 * NB;
+		y_ptr += NB * inc_y;
+	}
+
+	j=0;
+	while ( j < (m % 16))
+	{
+		a_ptr = a;
+		x_ptr = x;
+		FLOAT temp_r = 0.0;
+		FLOAT temp_i = 0.0;
+		for( i = 0; i < n; i++ )
+		{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+			temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+			temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+#else
+			temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+			temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+#endif
+
+			a_ptr += lda;
+			x_ptr += inc_x;
+		}
+
+#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;
+		a+=2;
+		j++;
+	}
+	return(0);
+}
+
+
diff --git a/kernel/x86_64/cgemv_n_microk_haswell-2.c b/kernel/x86_64/cgemv_n_microk_haswell-2.c
new file mode 100644
index 0000000..9b15010
--- /dev/null
+++ b/kernel/x86_64/cgemv_n_microk_haswell-2.c
@@ -0,0 +1,137 @@
+/***************************************************************************
+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 cgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void cgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vbroadcastss	  (%2), %%ymm0                  \n\t"  // real part x0
+	"vbroadcastss	 4(%2), %%ymm1                  \n\t"  // imag part x0
+	"vbroadcastss	 8(%2), %%ymm2                  \n\t"  // real part x1
+	"vbroadcastss	12(%2), %%ymm3                  \n\t"  // imag part x1
+	"vbroadcastss	16(%2), %%ymm4                  \n\t"  // real part x2
+	"vbroadcastss	20(%2), %%ymm5                  \n\t"  // imag part x2
+	"vbroadcastss	24(%2), %%ymm6                  \n\t"  // real part x3
+	"vbroadcastss	28(%2), %%ymm7                  \n\t"  // imag part x3
+
+
+	".align 16				        \n\t"
+	".L01LOOP%=:				        \n\t"
+	"prefetcht0      320(%4,%0,4)			\n\t"
+	"vmovups	(%4,%0,4), %%ymm8	        \n\t" // 4 complex values form a0
+	"vmovups      32(%4,%0,4), %%ymm9	        \n\t" // 4 complex values form a0
+
+	"prefetcht0      320(%5,%0,4)			\n\t"
+	"vmovups	(%5,%0,4), %%ymm10              \n\t" // 4 complex values form a1
+	"vmovups      32(%5,%0,4), %%ymm11              \n\t" // 4 complex values form a1
+
+	"vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vmulps      %%ymm9 , %%ymm0, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vmulps      %%ymm9 , %%ymm1, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      320(%6,%0,4)			\n\t"
+	"vmovups	(%6,%0,4), %%ymm8	        \n\t" // 4 complex values form a2
+	"vmovups      32(%6,%0,4), %%ymm9	        \n\t" // 4 complex values form a2
+
+	"vfmadd231ps      %%ymm10, %%ymm2, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmadd231ps      %%ymm10, %%ymm3, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmadd231ps      %%ymm11, %%ymm2, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmadd231ps      %%ymm11, %%ymm3, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      320(%7,%0,4)			\n\t"
+	"vmovups	(%7,%0,4), %%ymm10              \n\t" // 4 complex values form a3
+	"vmovups      32(%7,%0,4), %%ymm11              \n\t" // 4 complex values form a3
+
+	"vfmadd231ps      %%ymm8 , %%ymm4, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmadd231ps      %%ymm8 , %%ymm5, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmadd231ps      %%ymm9 , %%ymm4, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmadd231ps      %%ymm9 , %%ymm5, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"vfmadd231ps      %%ymm10, %%ymm6, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmadd231ps      %%ymm10, %%ymm7, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmadd231ps      %%ymm11, %%ymm6, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmadd231ps      %%ymm11, %%ymm7, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      320(%3,%0,4)			\n\t"
+	"vmovups	  (%3,%0,4),  %%ymm10           \n\t"
+	"vmovups	32(%3,%0,4),  %%ymm11           \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vpermilps      $0xb1 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+        "vaddsubps      %%ymm15, %%ymm14, %%ymm9              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vpermilps      $0xb1 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vaddsubps      %%ymm14, %%ymm15, %%ymm9              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm9 , %%ymm9                \n\t"
+#endif
+
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+        "vaddps         %%ymm9, %%ymm11, %%ymm13              \n\t"
+
+	"vmovups  %%ymm12,   (%3,%0,4)		        \n\t" // 4 complex values to y	
+	"vmovups  %%ymm13, 32(%3,%0,4)		        \n\t"	
+
+        "addq		$16, %0	  	 	        \n\t"
+	"subq	        $8 , %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"
+	);
+
+} 
+
+
-- 
2.7.4