Merge pull request #732 from wernsaar/develop

added optimized trsm_kernels

kernel/x86_64/KERNEL.BULLDOZER

@@ -69,24 +69,24 @@ ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
 CGEMM3MKERNEL    =  zgemm3m_kernel_8x4_barcelona.S
 ZGEMM3MKERNEL    =  zgemm3m_kernel_4x4_barcelona.S
 
-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
+STRSMKERNEL_LN  = strsm_kernel_LN_bulldozer.c 
+STRSMKERNEL_LT  = strsm_kernel_LT_bulldozer.c 
+STRSMKERNEL_RN  = strsm_kernel_RN_bulldozer.c 
+STRSMKERNEL_RT  = strsm_kernel_RT_bulldozer.c
 
-DTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
+DTRSMKERNEL_LN  = dtrsm_kernel_LN_bulldozer.c
 DTRSMKERNEL_LT  = dtrsm_kernel_LT_8x2_bulldozer.S
 DTRSMKERNEL_RN  = dtrsm_kernel_RN_8x2_bulldozer.S
-DTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
+DTRSMKERNEL_RT  = dtrsm_kernel_RT_bulldozer.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
+CTRSMKERNEL_LN  = ctrsm_kernel_LN_bulldozer.c
+CTRSMKERNEL_LT  = ctrsm_kernel_LT_bulldozer.c
+CTRSMKERNEL_RN  = ctrsm_kernel_RN_bulldozer.c
+CTRSMKERNEL_RT  = ctrsm_kernel_RT_bulldozer.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
+ZTRSMKERNEL_LN  = ztrsm_kernel_LN_bulldozer.c
+ZTRSMKERNEL_LT  = ztrsm_kernel_LT_bulldozer.c
+ZTRSMKERNEL_RN  = ztrsm_kernel_RN_bulldozer.c
+ZTRSMKERNEL_RT  = ztrsm_kernel_RT_bulldozer.c
 
 

kernel/x86_64/KERNEL.HASWELL

@@ -80,7 +80,7 @@ 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_RN	= dtrsm_kernel_RN_haswell.c
 DTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
 
 CTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c

kernel/x86_64/KERNEL.PILEDRIVER

@@ -66,25 +66,23 @@ ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
 CGEMM3MKERNEL    =  zgemm3m_kernel_8x4_barcelona.S
 ZGEMM3MKERNEL    =  zgemm3m_kernel_4x4_barcelona.S
 
-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
+STRSMKERNEL_LN  = strsm_kernel_LN_bulldozer.c 
+STRSMKERNEL_LT  = strsm_kernel_LT_bulldozer.c 
+STRSMKERNEL_RN  = strsm_kernel_RN_bulldozer.c 
+STRSMKERNEL_RT  = strsm_kernel_RT_bulldozer.c
 
-
-DTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
+DTRSMKERNEL_LN  = dtrsm_kernel_LN_bulldozer.c
 DTRSMKERNEL_LT  = dtrsm_kernel_LT_8x2_bulldozer.S
 DTRSMKERNEL_RN  = dtrsm_kernel_RN_8x2_bulldozer.S
-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
+DTRSMKERNEL_RT  = dtrsm_kernel_RT_bulldozer.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
+CTRSMKERNEL_LN  = ctrsm_kernel_LN_bulldozer.c
+CTRSMKERNEL_LT  = ctrsm_kernel_LT_bulldozer.c
+CTRSMKERNEL_RN  = ctrsm_kernel_RN_bulldozer.c
+CTRSMKERNEL_RT  = ctrsm_kernel_RT_bulldozer.c
 
+ZTRSMKERNEL_LN  = ztrsm_kernel_LN_bulldozer.c
+ZTRSMKERNEL_LT  = ztrsm_kernel_LT_bulldozer.c
+ZTRSMKERNEL_RN  = ztrsm_kernel_RN_bulldozer.c
+ZTRSMKERNEL_RT  = ztrsm_kernel_RT_bulldozer.c
 

kernel/x86_64/KERNEL.STEAMROLLER

@@ -72,25 +72,23 @@ ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
 CGEMM3MKERNEL    =  zgemm3m_kernel_8x4_barcelona.S
 ZGEMM3MKERNEL    =  zgemm3m_kernel_4x4_barcelona.S
 
-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
+STRSMKERNEL_LN  = strsm_kernel_LN_bulldozer.c 
+STRSMKERNEL_LT  = strsm_kernel_LT_bulldozer.c 
+STRSMKERNEL_RN  = strsm_kernel_RN_bulldozer.c 
+STRSMKERNEL_RT  = strsm_kernel_RT_bulldozer.c
 
-
-DTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
+DTRSMKERNEL_LN  = dtrsm_kernel_LN_bulldozer.c
 DTRSMKERNEL_LT  = dtrsm_kernel_LT_8x2_bulldozer.S
 DTRSMKERNEL_RN  = dtrsm_kernel_RN_8x2_bulldozer.S
-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
+DTRSMKERNEL_RT  = dtrsm_kernel_RT_bulldozer.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
+CTRSMKERNEL_LN  = ctrsm_kernel_LN_bulldozer.c
+CTRSMKERNEL_LT  = ctrsm_kernel_LT_bulldozer.c
+CTRSMKERNEL_RN  = ctrsm_kernel_RN_bulldozer.c
+CTRSMKERNEL_RT  = ctrsm_kernel_RT_bulldozer.c
 
+ZTRSMKERNEL_LN  = ztrsm_kernel_LN_bulldozer.c
+ZTRSMKERNEL_LT  = ztrsm_kernel_LT_bulldozer.c
+ZTRSMKERNEL_RN  = ztrsm_kernel_RN_bulldozer.c
+ZTRSMKERNEL_RT  = ztrsm_kernel_RT_bulldozer.c
 

kernel/x86_64/ctrsm_kernel_LN_bulldozer.c

@@ -0,0 +1,472 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* 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.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``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 UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN 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.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_L
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+
+#ifndef CONJ
+
+static void ctrsm_LN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void ctrsm_LN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc*2 ;
+	BLASLONG n1 = n * 4;
+	BLASLONG i=0;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorps	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorps	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorps	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorps	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorps	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorps	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorps	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorps	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       3f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	vbroadcastss        (%3,%1,4), %%xmm0				\n\t"	// b0 real, b0 real
+	"	vbroadcastss       4(%3,%1,4), %%xmm1				\n\t"	// b0 imag, b0 imag
+	"	vbroadcastss       8(%3,%1,4), %%xmm2				\n\t"	// b1 real, b1 real
+	"	vbroadcastss      12(%3,%1,4), %%xmm3				\n\t"	// b1 imag, b1 imag
+
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"	// a0 real , a0 imag
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"	// a1 real , a1 imag
+
+	"	vfnmaddps	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm9 , %%xmm1 , %%xmm4 , %%xmm9		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm10, %%xmm0 , %%xmm5 , %%xmm10		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm11, %%xmm1 , %%xmm5 , %%xmm11		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm12, %%xmm2 , %%xmm4 , %%xmm12		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm13, %%xmm3 , %%xmm4 , %%xmm13		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm14, %%xmm2 , %%xmm5 , %%xmm14		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm15, %%xmm3 , %%xmm5 , %%xmm15		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	addq		$4, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+
+	"2:									\n\t"
+
+	"	vshufps		$0xb1 , %%xmm9  , %%xmm9, %%xmm9		\n\t"
+	"	vshufps		$0xb1 , %%xmm11 , %%xmm11 , %%xmm11		\n\t"
+	"	vshufps		$0xb1 , %%xmm13 , %%xmm13 , %%xmm13		\n\t"
+	"	vshufps		$0xb1 , %%xmm15 , %%xmm15 , %%xmm15		\n\t"
+
+	"	vaddsubps		%%xmm8 , %%xmm9 , %%xmm8		\n\t"
+	"	vaddsubps		%%xmm10, %%xmm11, %%xmm10		\n\t"
+	"	vaddsubps		%%xmm12, %%xmm13, %%xmm12		\n\t"
+	"	vaddsubps		%%xmm14, %%xmm15, %%xmm14		\n\t"
+
+	"	vxorps		%%xmm7 , %%xmm7 , %%xmm7			\n\t"
+
+	"	vaddsubps		%%xmm8 , %%xmm7 , %%xmm8		\n\t"
+	"	vaddsubps		%%xmm10, %%xmm7 , %%xmm10		\n\t"
+	"	vaddsubps		%%xmm12, %%xmm7 , %%xmm12		\n\t"
+	"	vaddsubps		%%xmm14, %%xmm7 , %%xmm14		\n\t"
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+
+	"	vaddps		%%xmm0 , %%xmm8 , %%xmm8			\n\t"
+	"	vaddps		%%xmm1 , %%xmm10, %%xmm10			\n\t"
+	"	vaddps		%%xmm4 , %%xmm12, %%xmm12			\n\t"
+	"	vaddps		%%xmm5 , %%xmm14, %%xmm14			\n\t"
+
+	"	vmovups		%%xmm8  ,  (%4)					\n\t"
+	"	vmovups		%%xmm10 ,16(%4)					\n\t"
+
+	"	vmovups		%%xmm12 ,  (%5)					\n\t"
+	"	vmovups		%%xmm14 ,16(%5)					\n\t"
+
+	"3:									\n\t"	
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+#endif
+
+
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa,  bb;
+
+  int i, j, k;
+
+  a += (m - 1) * m;
+  b += (m - 1) * n;
+
+  for (i = m - 1; i >= 0; i--) {
+
+    aa = *(a + i);
+
+    for (j = 0; j < n; j ++) {
+      bb = *(c + i + j * ldc);
+      bb *= aa;
+      *b             = bb;
+      *(c + i + j * ldc) = bb;
+      b ++;
+
+      for (k = 0; k < i; k ++){
+	*(c + k + j * ldc) -= bb * *(a + k);
+      }
+
+    }
+    a -= m;
+    b -= 2 * n;
+  }
+
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+  a += (m - 1) * m * 2;
+  b += (m - 1) * n * 2;
+
+  for (i = m - 1; i >= 0; i--) {
+
+    aa1 = *(a + i * 2 + 0);
+    aa2 = *(a + i * 2 + 1);
+
+    for (j = 0; j < n; j ++) {
+      bb1 = *(c + i * 2 + 0 + j * ldc);
+      bb2 = *(c + i * 2 + 1 + j * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 = aa1 * bb1 + aa2 * bb2;
+      cc2 = aa1 * bb2 - aa2 * bb1;
+#endif
+
+
+      *(b + 0) = cc1;
+      *(b + 1) = cc2;
+      *(c + i * 2 + 0 + j * ldc) = cc1;
+      *(c + i * 2 + 1 + j * ldc) = cc2;
+      b += 2;
+
+      for (k = 0; k < i; k ++){
+#ifndef CONJ
+	*(c + k * 2 + 0 + j * ldc) -= cc1 * *(a + k * 2 + 0) - cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#else
+	*(c + k * 2 + 0 + j * ldc) -=   cc1 * *(a + k * 2 + 0) + cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= - cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#endif
+      }
+
+    }
+    a -= m * 2;
+    b -= 4 * n;
+  }
+
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k,  FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  BLASLONG i, j;
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+
+#if 0
+  fprintf(stderr, "TRSM KERNEL LN : m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  while (j > 0) {
+
+    kk = m + offset;
+
+    if (m & (GEMM_UNROLL_M - 1)) {
+      for (i = 1; i < GEMM_UNROLL_M; i *= 2){
+	if (m & i) {
+	  aa = a + ((m & ~(i - 1)) - i) * k * COMPSIZE;
+	  cc = c + ((m & ~(i - 1)) - i)     * COMPSIZE;
+
+	  if (k - kk > 0) {
+	    GEMM_KERNEL(i, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa + i             * kk * COMPSIZE,
+			b  + GEMM_UNROLL_N * kk * COMPSIZE,
+			cc,
+			ldc);
+	  }
+
+	  solve(i, GEMM_UNROLL_N,
+		aa + (kk - i) * i             * COMPSIZE,
+		b  + (kk - i) * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+	  kk -= i;
+	}
+      }
+    }
+
+    i = (m >> GEMM_UNROLL_M_SHIFT);
+    if (i > 0) {
+      aa = a + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M) * k * COMPSIZE;
+      cc = c + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M)     * COMPSIZE;
+
+      do {
+
+#ifdef CONJ
+	if (k - kk > 0) {
+	  GEMM_KERNEL(GEMM_UNROLL_M, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+		      ZERO,
+#endif
+		      aa + GEMM_UNROLL_M * kk * COMPSIZE,
+		      b +  GEMM_UNROLL_N * kk * COMPSIZE,
+		      cc,
+		      ldc);
+	}
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_M * COMPSIZE,
+	      b  + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#else
+
+	ctrsm_LN_solve_opt(k-kk, aa + GEMM_UNROLL_M * kk * COMPSIZE, b +  GEMM_UNROLL_N * kk * COMPSIZE, cc, ldc,
+			   aa + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_M * COMPSIZE, b  + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_N * COMPSIZE);
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_M * COMPSIZE,
+	      b  + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#endif
+
+	aa -= GEMM_UNROLL_M * k * COMPSIZE;
+	cc -= GEMM_UNROLL_M     * COMPSIZE;
+	kk -= GEMM_UNROLL_M;
+	i --;
+      } while (i > 0);
+    }
+
+    b += GEMM_UNROLL_N * k * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+	kk = m + offset;
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  for (i = 1; i < GEMM_UNROLL_M; i *= 2){
+	    if (m & i) {
+	      aa = a + ((m & ~(i - 1)) - i) * k * COMPSIZE;
+	      cc = c + ((m & ~(i - 1)) - i)     * COMPSIZE;
+
+	      if (k - kk > 0) {
+		GEMM_KERNEL(i, j, k - kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa + i * kk * COMPSIZE,
+			    b  + j * kk * COMPSIZE,
+			    cc, ldc);
+	      }
+
+	      solve(i, j,
+		    aa + (kk - i) * i * COMPSIZE,
+		    b  + (kk - i) * j * COMPSIZE,
+		    cc, ldc);
+
+	      kk -= i;
+	    }
+	  }
+	}
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+	if (i > 0) {
+	  aa = a + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M) * k * COMPSIZE;
+	  cc = c + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M)     * COMPSIZE;
+
+	  do {
+	    if (k - kk > 0) {
+	      GEMM_KERNEL(GEMM_UNROLL_M, j, k - kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa + GEMM_UNROLL_M * kk * COMPSIZE,
+			  b +  j             * kk * COMPSIZE,
+			  cc,
+			  ldc);
+	    }
+
+	    solve(GEMM_UNROLL_M, j,
+		  aa + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_M * COMPSIZE,
+		  b  + (kk - GEMM_UNROLL_M) * j             * COMPSIZE,
+		  cc, ldc);
+
+	    aa -= GEMM_UNROLL_M * k * COMPSIZE;
+	    cc -= GEMM_UNROLL_M     * COMPSIZE;
+	    kk -= GEMM_UNROLL_M;
+	    i --;
+	  } while (i > 0);
+	}
+
+	b += j * k   * COMPSIZE;
+	c += j * ldc * COMPSIZE;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/x86_64/ctrsm_kernel_LT_bulldozer.c

@@ -0,0 +1,455 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* 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.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``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 UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN 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.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_L
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+
+#ifndef CONJ
+
+static void ctrsm_LT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void ctrsm_LT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc*2 ;
+	BLASLONG n1 = n * 4;
+	BLASLONG i=0;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorps	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorps	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorps	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorps	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorps	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorps	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorps	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorps	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       3f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	vbroadcastss        (%3,%1,4), %%xmm0				\n\t"	// b0 real, b0 real
+	"	vbroadcastss       4(%3,%1,4), %%xmm1				\n\t"	// b0 imag, b0 imag
+	"	vbroadcastss       8(%3,%1,4), %%xmm2				\n\t"	// b1 real, b1 real
+	"	vbroadcastss      12(%3,%1,4), %%xmm3				\n\t"	// b1 imag, b1 imag
+
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"	// a0 real , a0 imag
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"	// a1 real , a1 imag
+
+	"	vfnmaddps	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm9 , %%xmm1 , %%xmm4 , %%xmm9		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm10, %%xmm0 , %%xmm5 , %%xmm10		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm11, %%xmm1 , %%xmm5 , %%xmm11		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm12, %%xmm2 , %%xmm4 , %%xmm12		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm13, %%xmm3 , %%xmm4 , %%xmm13		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm14, %%xmm2 , %%xmm5 , %%xmm14		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm15, %%xmm3 , %%xmm5 , %%xmm15		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	addq		$4, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+
+	"2:									\n\t"
+
+	"	vshufps		$0xb1 , %%xmm9  , %%xmm9, %%xmm9		\n\t"
+	"	vshufps		$0xb1 , %%xmm11 , %%xmm11 , %%xmm11		\n\t"
+	"	vshufps		$0xb1 , %%xmm13 , %%xmm13 , %%xmm13		\n\t"
+	"	vshufps		$0xb1 , %%xmm15 , %%xmm15 , %%xmm15		\n\t"
+
+	"	vaddsubps		%%xmm8 , %%xmm9 , %%xmm8		\n\t"
+	"	vaddsubps		%%xmm10, %%xmm11, %%xmm10		\n\t"
+	"	vaddsubps		%%xmm12, %%xmm13, %%xmm12		\n\t"
+	"	vaddsubps		%%xmm14, %%xmm15, %%xmm14		\n\t"
+
+	"	vxorps		%%xmm7 , %%xmm7 , %%xmm7			\n\t"
+
+	"	vaddsubps		%%xmm8 , %%xmm7 , %%xmm8		\n\t"
+	"	vaddsubps		%%xmm10, %%xmm7 , %%xmm10		\n\t"
+	"	vaddsubps		%%xmm12, %%xmm7 , %%xmm12		\n\t"
+	"	vaddsubps		%%xmm14, %%xmm7 , %%xmm14		\n\t"
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+
+	"	vaddps		%%xmm0 , %%xmm8 , %%xmm8			\n\t"
+	"	vaddps		%%xmm1 , %%xmm10, %%xmm10			\n\t"
+	"	vaddps		%%xmm4 , %%xmm12, %%xmm12			\n\t"
+	"	vaddps		%%xmm5 , %%xmm14, %%xmm14			\n\t"
+
+	"	vmovups		%%xmm8  ,  (%4)					\n\t"
+	"	vmovups		%%xmm10 ,16(%4)					\n\t"
+
+	"	vmovups		%%xmm12 ,  (%5)					\n\t"
+	"	vmovups		%%xmm14 ,16(%5)					\n\t"
+
+	"3:									\n\t"	
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+#endif
+
+
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa, bb;
+
+  int i, j, k;
+
+  for (i = 0; i < m; i++) {
+
+    aa = *(a + i);
+
+    for (j = 0; j < n; j ++) {
+      bb = *(c + i + j * ldc);
+      bb *= aa;
+      *b             = bb;
+      *(c + i + j * ldc) = bb;
+      b ++;
+
+      for (k = i + 1; k < m; k ++){
+	*(c + k + j * ldc) -= bb * *(a + k);
+      }
+
+    }
+    a += m;
+  }
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  for (i = 0; i < m; i++) {
+
+    aa1 = *(a + i * 2 + 0);
+    aa2 = *(a + i * 2 + 1);
+
+    for (j = 0; j < n; j ++) {
+      bb1 = *(c + i * 2 + 0 + j * ldc);
+      bb2 = *(c + i * 2 + 1 + j * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 = aa1 * bb1 + aa2 * bb2;
+      cc2 = aa1 * bb2 - aa2 * bb1;
+#endif
+
+      *(b + 0) = cc1;
+      *(b + 1) = cc2;
+      *(c + i * 2 + 0 + j * ldc) = cc1;
+      *(c + i * 2 + 1 + j * ldc) = cc2;
+      b += 2;
+
+      for (k = i + 1; k < m; k ++){
+#ifndef CONJ
+	*(c + k * 2 + 0 + j * ldc) -= cc1 * *(a + k * 2 + 0) - cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#else
+	*(c + k * 2 + 0 + j * ldc) -= cc1 * *(a + k * 2 + 0) + cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= -cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#endif
+      }
+
+    }
+    a += m * 2;
+  }
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+  BLASLONG i, j, jj;
+
+#if 0
+  fprintf(stderr, "TRSM KERNEL LT : m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  jj = 0;
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  while (j > 0) {
+
+    kk = offset;
+    aa = a;
+    cc = c;
+
+    i = (m >> GEMM_UNROLL_M_SHIFT);
+
+    while (i > 0) {
+
+#ifdef CONJ
+
+	if (kk > 0) {
+	  GEMM_KERNEL(GEMM_UNROLL_M, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+		      ZERO,
+#endif
+		      aa, b, cc, ldc);
+	}
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + kk * GEMM_UNROLL_M * COMPSIZE,
+	      b  + kk * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#else
+
+	ctrsm_LT_solve_opt(kk, aa, b, cc, ldc, aa + kk * GEMM_UNROLL_M * COMPSIZE, b  + kk * GEMM_UNROLL_N * COMPSIZE);
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + kk * GEMM_UNROLL_M * COMPSIZE,
+	      b  + kk * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#endif
+
+      aa += GEMM_UNROLL_M * k * COMPSIZE;
+      cc += GEMM_UNROLL_M     * COMPSIZE;
+      kk += GEMM_UNROLL_M;
+      i --;
+    }
+
+    if (m & (GEMM_UNROLL_M - 1)) {
+      i = (GEMM_UNROLL_M >> 1);
+      while (i > 0) {
+	if (m & i) {
+	    if (kk > 0) {
+	      GEMM_KERNEL(i, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa, b, cc, ldc);
+	    }
+	  solve(i, GEMM_UNROLL_N,
+		aa + kk * i             * COMPSIZE,
+		b  + kk * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+	  aa += i * k * COMPSIZE;
+	  cc += i     * COMPSIZE;
+	  kk += i;
+	}
+	i >>= 1;
+      }
+    }
+
+    b += GEMM_UNROLL_N * k   * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+    jj += GEMM_UNROLL_M;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+	kk = offset;
+	aa = a;
+	cc = c;
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+
+	while (i > 0) {
+	  if (kk > 0) {
+	    GEMM_KERNEL(GEMM_UNROLL_M, j, kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa,
+			b,
+			cc,
+			ldc);
+	  }
+
+	  solve(GEMM_UNROLL_M, j,
+		aa + kk * GEMM_UNROLL_M * COMPSIZE,
+		b  + kk * j             * COMPSIZE, cc, ldc);
+
+	  aa += GEMM_UNROLL_M * k * COMPSIZE;
+	  cc += GEMM_UNROLL_M     * COMPSIZE;
+	  kk += GEMM_UNROLL_M;
+	  i --;
+	}
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  i = (GEMM_UNROLL_M >> 1);
+	  while (i > 0) {
+	    if (m & i) {
+	      if (kk > 0) {
+		GEMM_KERNEL(i, j, kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa,
+			    b,
+			    cc,
+			    ldc);
+	      }
+
+	      solve(i, j,
+		    aa + kk * i * COMPSIZE,
+		    b  + kk * j * COMPSIZE, cc, ldc);
+
+	      aa += i * k * COMPSIZE;
+	      cc += i     * COMPSIZE;
+	      kk += i;
+	      }
+	    i >>= 1;
+	  }
+	}
+
+	b += j * k   * COMPSIZE;
+	c += j * ldc * COMPSIZE;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/x86_64/ctrsm_kernel_RN_bulldozer.c

@@ -0,0 +1,454 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* 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.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``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 UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN 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.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_R
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+
+#ifndef CONJ
+
+static void ctrsm_RN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void ctrsm_RN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc*2 ;
+	BLASLONG n1 = n * 4;
+	BLASLONG i=0;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorps	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorps	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorps	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorps	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorps	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorps	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorps	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorps	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       3f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	vbroadcastss        (%3,%1,4), %%xmm0				\n\t"	// b0 real, b0 real
+	"	vbroadcastss       4(%3,%1,4), %%xmm1				\n\t"	// b0 imag, b0 imag
+	"	vbroadcastss       8(%3,%1,4), %%xmm2				\n\t"	// b1 real, b1 real
+	"	vbroadcastss      12(%3,%1,4), %%xmm3				\n\t"	// b1 imag, b1 imag
+
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"	// a0 real , a0 imag
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"	// a1 real , a1 imag
+
+	"	vfnmaddps	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm9 , %%xmm1 , %%xmm4 , %%xmm9		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm10, %%xmm0 , %%xmm5 , %%xmm10		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm11, %%xmm1 , %%xmm5 , %%xmm11		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm12, %%xmm2 , %%xmm4 , %%xmm12		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm13, %%xmm3 , %%xmm4 , %%xmm13		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm14, %%xmm2 , %%xmm5 , %%xmm14		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm15, %%xmm3 , %%xmm5 , %%xmm15		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	addq		$4, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+
+	"2:									\n\t"
+
+	"	vshufps		$0xb1 , %%xmm9  , %%xmm9, %%xmm9		\n\t"
+	"	vshufps		$0xb1 , %%xmm11 , %%xmm11 , %%xmm11		\n\t"
+	"	vshufps		$0xb1 , %%xmm13 , %%xmm13 , %%xmm13		\n\t"
+	"	vshufps		$0xb1 , %%xmm15 , %%xmm15 , %%xmm15		\n\t"
+
+	"	vaddsubps		%%xmm8 , %%xmm9 , %%xmm8		\n\t"
+	"	vaddsubps		%%xmm10, %%xmm11, %%xmm10		\n\t"
+	"	vaddsubps		%%xmm12, %%xmm13, %%xmm12		\n\t"
+	"	vaddsubps		%%xmm14, %%xmm15, %%xmm14		\n\t"
+
+	"	vxorps		%%xmm7 , %%xmm7 , %%xmm7			\n\t"
+
+	"	vaddsubps		%%xmm8 , %%xmm7 , %%xmm8		\n\t"
+	"	vaddsubps		%%xmm10, %%xmm7 , %%xmm10		\n\t"
+	"	vaddsubps		%%xmm12, %%xmm7 , %%xmm12		\n\t"
+	"	vaddsubps		%%xmm14, %%xmm7 , %%xmm14		\n\t"
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+
+	"	vaddps		%%xmm0 , %%xmm8 , %%xmm8			\n\t"
+	"	vaddps		%%xmm1 , %%xmm10, %%xmm10			\n\t"
+	"	vaddps		%%xmm4 , %%xmm12, %%xmm12			\n\t"
+	"	vaddps		%%xmm5 , %%xmm14, %%xmm14			\n\t"
+
+	"	vmovups		%%xmm8  ,  (%4)					\n\t"
+	"	vmovups		%%xmm10 ,16(%4)					\n\t"
+
+	"	vmovups		%%xmm12 ,  (%5)					\n\t"
+	"	vmovups		%%xmm14 ,16(%5)					\n\t"
+
+	"3:									\n\t"	
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+#endif
+
+
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa, bb;
+
+  int i, j, k;
+
+  for (i = 0; i < n; i++) {
+
+    bb = *(b + i);
+
+    for (j = 0; j < m; j ++) {
+      aa = *(c + j + i * ldc);
+      aa *= bb;
+      *a  = aa;
+      *(c + j + i * ldc) = aa;
+      a ++;
+
+      for (k = i + 1; k < n; k ++){
+	*(c + j + k * ldc) -= aa * *(b + k);
+      }
+
+    }
+    b += n;
+  }
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  for (i = 0; i < n; i++) {
+
+    bb1 = *(b + i * 2 + 0);
+    bb2 = *(b + i * 2 + 1);
+
+    for (j = 0; j < m; j ++) {
+      aa1 = *(c + j * 2 + 0 + i * ldc);
+      aa2 = *(c + j * 2 + 1 + i * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 =  aa1 * bb1 + aa2 * bb2;
+      cc2 = -aa1 * bb2 + aa2 * bb1;
+#endif
+
+      *(a + 0) = cc1;
+      *(a + 1) = cc2;
+      *(c + j * 2 + 0 + i * ldc) = cc1;
+      *(c + j * 2 + 1 + i * ldc) = cc2;
+      a += 2;
+
+      for (k = i + 1; k < n; k ++){
+#ifndef CONJ
+	*(c + j * 2 + 0 + k * ldc) -= cc1 * *(b + k * 2 + 0) - cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#else
+	*(c + j * 2 + 0 + k * ldc) -=   cc1 * *(b + k * 2 + 0) + cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= - cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#endif
+      }
+
+    }
+    b += n * 2;
+  }
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+  BLASLONG i, j, jj;
+
+#if 0
+  fprintf(stderr, "TRSM RN KERNEL m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  jj = 0;
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+  kk = -offset;
+
+  while (j > 0) {
+
+    aa = a;
+    cc = c;
+
+    i = (m >> GEMM_UNROLL_M_SHIFT);
+
+    if (i > 0) {
+      do {
+
+#ifndef CONJ
+
+        ctrsm_RN_solve_opt(kk, aa, b, cc, ldc, aa + kk * GEMM_UNROLL_M * COMPSIZE, b  + kk * GEMM_UNROLL_N * COMPSIZE);
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + kk * GEMM_UNROLL_M * COMPSIZE,
+	      b  + kk * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#else
+
+        if (kk > 0) {
+          GEMM_KERNEL(GEMM_UNROLL_M, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+                      ZERO,
+#endif
+                      aa, b, cc, ldc);
+        }
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + kk * GEMM_UNROLL_M * COMPSIZE,
+	      b  + kk * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#endif
+
+	aa += GEMM_UNROLL_M * k * COMPSIZE;
+	cc += GEMM_UNROLL_M     * COMPSIZE;
+	i --;
+      } while (i > 0);
+    }
+
+
+    if (m & (GEMM_UNROLL_M - 1)) {
+      i = (GEMM_UNROLL_M >> 1);
+      while (i > 0) {
+	if (m & i) {
+	    if (kk > 0) {
+	      GEMM_KERNEL(i, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa, b, cc, ldc);
+	    }
+	  solve(i, GEMM_UNROLL_N,
+		aa + kk * i             * COMPSIZE,
+		b  + kk * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+	  aa += i * k * COMPSIZE;
+	  cc += i     * COMPSIZE;
+	}
+	i >>= 1;
+      }
+    }
+
+    kk += GEMM_UNROLL_N;
+    b += GEMM_UNROLL_N * k   * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+    jj += GEMM_UNROLL_M;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+	aa = a;
+	cc = c;
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+
+	while (i > 0) {
+	  if (kk > 0) {
+	    GEMM_KERNEL(GEMM_UNROLL_M, j, kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa,
+			b,
+			cc,
+			ldc);
+	  }
+
+	  solve(GEMM_UNROLL_M, j,
+		aa + kk * GEMM_UNROLL_M * COMPSIZE,
+		b  + kk * j             * COMPSIZE, cc, ldc);
+
+	  aa += GEMM_UNROLL_M * k * COMPSIZE;
+	  cc += GEMM_UNROLL_M     * COMPSIZE;
+	  i --;
+	}
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  i = (GEMM_UNROLL_M >> 1);
+	  while (i > 0) {
+	    if (m & i) {
+	      if (kk > 0) {
+		GEMM_KERNEL(i, j, kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa,
+			    b,
+			    cc,
+			    ldc);
+	      }
+
+	      solve(i, j,
+		    aa + kk * i * COMPSIZE,
+		    b  + kk * j * COMPSIZE, cc, ldc);
+
+	      aa += i * k * COMPSIZE;
+	      cc += i     * COMPSIZE;
+	      }
+	    i >>= 1;
+	  }
+	}
+
+	b += j * k   * COMPSIZE;
+	c += j * ldc * COMPSIZE;
+	kk += j;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/x86_64/ctrsm_kernel_RT_bulldozer.c

@@ -0,0 +1,482 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* 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.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``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 UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN 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.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_R
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+
+#ifndef CONJ
+
+static void ctrsm_RT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void ctrsm_RT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc*2 ;
+	BLASLONG n1 = n * 4;
+	BLASLONG i=0;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorps	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorps	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorps	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorps	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorps	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorps	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorps	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorps	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       3f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	vbroadcastss        (%3,%1,4), %%xmm0				\n\t"	// b0 real, b0 real
+	"	vbroadcastss       4(%3,%1,4), %%xmm1				\n\t"	// b0 imag, b0 imag
+	"	vbroadcastss       8(%3,%1,4), %%xmm2				\n\t"	// b1 real, b1 real
+	"	vbroadcastss      12(%3,%1,4), %%xmm3				\n\t"	// b1 imag, b1 imag
+
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"	// a0 real , a0 imag
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"	// a1 real , a1 imag
+
+	"	vfnmaddps	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm9 , %%xmm1 , %%xmm4 , %%xmm9		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm10, %%xmm0 , %%xmm5 , %%xmm10		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm11, %%xmm1 , %%xmm5 , %%xmm11		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm12, %%xmm2 , %%xmm4 , %%xmm12		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm13, %%xmm3 , %%xmm4 , %%xmm13		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm14, %%xmm2 , %%xmm5 , %%xmm14		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm15, %%xmm3 , %%xmm5 , %%xmm15		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	addq		$4, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+
+	"2:									\n\t"
+
+	"	vshufps		$0xb1 , %%xmm9  , %%xmm9, %%xmm9		\n\t"
+	"	vshufps		$0xb1 , %%xmm11 , %%xmm11 , %%xmm11		\n\t"
+	"	vshufps		$0xb1 , %%xmm13 , %%xmm13 , %%xmm13		\n\t"
+	"	vshufps		$0xb1 , %%xmm15 , %%xmm15 , %%xmm15		\n\t"
+
+	"	vaddsubps		%%xmm8 , %%xmm9 , %%xmm8		\n\t"
+	"	vaddsubps		%%xmm10, %%xmm11, %%xmm10		\n\t"
+	"	vaddsubps		%%xmm12, %%xmm13, %%xmm12		\n\t"
+	"	vaddsubps		%%xmm14, %%xmm15, %%xmm14		\n\t"
+
+	"	vxorps		%%xmm7 , %%xmm7 , %%xmm7			\n\t"
+
+	"	vaddsubps		%%xmm8 , %%xmm7 , %%xmm8		\n\t"
+	"	vaddsubps		%%xmm10, %%xmm7 , %%xmm10		\n\t"
+	"	vaddsubps		%%xmm12, %%xmm7 , %%xmm12		\n\t"
+	"	vaddsubps		%%xmm14, %%xmm7 , %%xmm14		\n\t"
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+
+	"	vaddps		%%xmm0 , %%xmm8 , %%xmm8			\n\t"
+	"	vaddps		%%xmm1 , %%xmm10, %%xmm10			\n\t"
+	"	vaddps		%%xmm4 , %%xmm12, %%xmm12			\n\t"
+	"	vaddps		%%xmm5 , %%xmm14, %%xmm14			\n\t"
+
+	"	vmovups		%%xmm8  ,  (%4)					\n\t"
+	"	vmovups		%%xmm10 ,16(%4)					\n\t"
+
+	"	vmovups		%%xmm12 ,  (%5)					\n\t"
+	"	vmovups		%%xmm14 ,16(%5)					\n\t"
+
+	"3:									\n\t"	
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+#endif
+
+
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa,  bb;
+
+  int i, j, k;
+
+  a += (n - 1) * m;
+  b += (n - 1) * n;
+
+  for (i = n - 1; i >= 0; i--) {
+
+    bb = *(b + i);
+
+    for (j = 0; j < m; j ++) {
+      aa = *(c + j + i * ldc);
+      aa *= bb;
+      *a   = aa;
+      *(c + j + i * ldc) = aa;
+      a ++;
+
+      for (k = 0; k < i; k ++){
+	*(c + j + k * ldc) -= aa * *(b + k);
+      }
+
+    }
+    b -= n;
+    a -= 2 * m;
+  }
+
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  a += (n - 1) * m * 2;
+  b += (n - 1) * n * 2;
+
+  for (i = n - 1; i >= 0; i--) {
+
+    bb1 = *(b + i * 2 + 0);
+    bb2 = *(b + i * 2 + 1);
+
+    for (j = 0; j < m; j ++) {
+
+      aa1 = *(c + j * 2 + 0 + i * ldc);
+      aa2 = *(c + j * 2 + 1 + i * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 =  aa1 * bb1  + aa2 * bb2;
+      cc2 = - aa1 * bb2 + aa2 * bb1;
+#endif
+
+      *(a + 0) = cc1;
+      *(a + 1) = cc2;
+
+      *(c + j * 2 + 0 + i * ldc) = cc1;
+      *(c + j * 2 + 1 + i * ldc) = cc2;
+      a += 2;
+
+      for (k = 0; k < i; k ++){
+#ifndef CONJ
+	*(c + j * 2 + 0 + k * ldc) -= cc1 * *(b + k * 2 + 0) - cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#else
+	*(c + j * 2 + 0 + k * ldc) -=   cc1 * *(b + k * 2 + 0) + cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -=  -cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#endif
+      }
+
+    }
+    b -= n * 2;
+    a -= 4 * m;
+  }
+
+}
+
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k,  FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  BLASLONG i, j;
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+
+#if 0
+  fprintf(stderr, "TRSM RT KERNEL m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  kk = n - offset;
+  c += n * ldc * COMPSIZE;
+  b += n * k   * COMPSIZE;
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = 1;
+    while (j < GEMM_UNROLL_N) {
+      if (n & j) {
+
+	aa  = a;
+	b -= j * k  * COMPSIZE;
+	c -= j * ldc* COMPSIZE;
+	cc  = c;
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+	if (i > 0) {
+
+	  do {
+	    if (k - kk > 0) {
+	      GEMM_KERNEL(GEMM_UNROLL_M, j, k - kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa + GEMM_UNROLL_M * kk * COMPSIZE,
+			  b  +  j            * kk * COMPSIZE,
+			  cc,
+			  ldc);
+	    }
+
+	    solve(GEMM_UNROLL_M, j,
+		  aa + (kk - j) * GEMM_UNROLL_M * COMPSIZE,
+		  b  + (kk - j) * j             * COMPSIZE,
+		  cc, ldc);
+
+	    aa += GEMM_UNROLL_M * k * COMPSIZE;
+	    cc += GEMM_UNROLL_M     * COMPSIZE;
+	    i --;
+	  } while (i > 0);
+	}
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  i = (GEMM_UNROLL_M >> 1);
+	  do {
+	    if (m & i) {
+
+	      if (k - kk > 0) {
+		GEMM_KERNEL(i, j, k - kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa + i * kk * COMPSIZE,
+			    b  + j * kk * COMPSIZE,
+			    cc, ldc);
+	      }
+
+	      solve(i, j,
+		    aa + (kk - j) * i * COMPSIZE,
+		    b  + (kk - j) * j * COMPSIZE,
+		    cc, ldc);
+
+	      aa += i * k * COMPSIZE;
+	      cc += i     * COMPSIZE;
+
+	    }
+	    i >>= 1;
+	  } while (i > 0);
+	}
+	kk -= j;
+      }
+      j <<= 1;
+    }
+  }
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  if (j > 0) {
+
+    do {
+      aa  = a;
+      b -= GEMM_UNROLL_N * k   * COMPSIZE;
+      c -= GEMM_UNROLL_N * ldc * COMPSIZE;
+      cc  = c;
+
+      i = (m >> GEMM_UNROLL_M_SHIFT);
+      if (i > 0) {
+	do {
+
+#ifndef CONJ
+
+	  ctrsm_RT_solve_opt(k-kk, aa + GEMM_UNROLL_M * kk * COMPSIZE, b  + GEMM_UNROLL_N * kk * COMPSIZE, cc, ldc,
+                             aa + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_M * COMPSIZE, b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE);
+ 
+	  solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+		aa + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_M * COMPSIZE,
+		b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+#else
+	  if (k - kk > 0) {
+	    GEMM_KERNEL(GEMM_UNROLL_M, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa + GEMM_UNROLL_M * kk * COMPSIZE,
+			b  + GEMM_UNROLL_N * kk * COMPSIZE,
+			cc,
+			ldc);
+	  }
+
+	  solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+		aa + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_M * COMPSIZE,
+		b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+#endif
+
+	  aa += GEMM_UNROLL_M * k * COMPSIZE;
+	  cc += GEMM_UNROLL_M     * COMPSIZE;
+	  i --;
+	} while (i > 0);
+      }
+
+      if (m & (GEMM_UNROLL_M - 1)) {
+	i = (GEMM_UNROLL_M >> 1);
+	do {
+	  if (m & i) {
+	    if (k - kk > 0) {
+	      GEMM_KERNEL(i, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa + i             * kk * COMPSIZE,
+			  b  + GEMM_UNROLL_N * kk * COMPSIZE,
+			  cc,
+			  ldc);
+	    }
+
+	    solve(i, GEMM_UNROLL_N,
+		  aa + (kk - GEMM_UNROLL_N) * i             * COMPSIZE,
+		  b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE,
+		  cc, ldc);
+
+	    aa += i * k * COMPSIZE;
+	    cc += i     * COMPSIZE;
+	  }
+	  i >>= 1;
+	} while (i > 0);
+      }
+
+      kk -= GEMM_UNROLL_N;
+      j --;
+    } while (j > 0);
+  }
+
+  return 0;
+}
+
+

kernel/x86_64/dtrsm_kernel_LN_bulldozer.c

@@ -0,0 +1,697 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* 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.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``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 UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN 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.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_L
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+static void dtrsm_LN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void dtrsm_LN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc ;
+	BLASLONG n1 = n * 8;
+	BLASLONG i=0;
+
+  	bs += 14;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorpd	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorpd	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorpd	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorpd	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorpd	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorpd	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorpd	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorpd	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       2f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	prefetcht0	384(%2,%1,8)					\n\t"
+	"	prefetcht0	384(%3,%1,8)					\n\t"
+	"	vmovddup	(%3,%1,2), %%xmm0				\n\t"	// read b
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"
+	"	vmovddup       8(%3,%1,2), %%xmm1				\n\t"	
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"
+	"	vmovups       32(%2,%1,8), %%xmm6				\n\t"
+	"	vmovups       48(%2,%1,8), %%xmm7				\n\t"
+
+	"	vfmaddpd	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"
+	"	vfmaddpd	%%xmm12, %%xmm1 , %%xmm4 , %%xmm12		\n\t"
+	"	vfmaddpd	%%xmm9 , %%xmm0 , %%xmm5 , %%xmm9		\n\t"
+	"	vfmaddpd	%%xmm13, %%xmm1 , %%xmm5 , %%xmm13		\n\t"
+	"	vfmaddpd	%%xmm10, %%xmm0 , %%xmm6 , %%xmm10		\n\t"
+	"	vfmaddpd	%%xmm14, %%xmm1 , %%xmm6 , %%xmm14		\n\t"
+	"	addq		$8, %1						\n\t"
+	"	vfmaddpd	%%xmm11, %%xmm0 , %%xmm7 , %%xmm11		\n\t"
+	"	vfmaddpd	%%xmm15, %%xmm1 , %%xmm7 , %%xmm15		\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jz		2f						\n\t"
+
+	"	prefetcht0	384(%2,%1,8)					\n\t"
+	"	vmovddup	(%3,%1,2), %%xmm0				\n\t"	// read b
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"
+	"	vmovddup       8(%3,%1,2), %%xmm1				\n\t"	
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"
+	"	vmovups       32(%2,%1,8), %%xmm6				\n\t"
+	"	vmovups       48(%2,%1,8), %%xmm7				\n\t"
+
+	"	vfmaddpd	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"
+	"	vfmaddpd	%%xmm12, %%xmm1 , %%xmm4 , %%xmm12		\n\t"
+	"	vfmaddpd	%%xmm9 , %%xmm0 , %%xmm5 , %%xmm9		\n\t"
+	"	vfmaddpd	%%xmm13, %%xmm1 , %%xmm5 , %%xmm13		\n\t"
+	"	vfmaddpd	%%xmm10, %%xmm0 , %%xmm6 , %%xmm10		\n\t"
+	"	vfmaddpd	%%xmm14, %%xmm1 , %%xmm6 , %%xmm14		\n\t"
+	"	addq		$8, %1						\n\t"
+	"	vfmaddpd	%%xmm11, %%xmm0 , %%xmm7 , %%xmm11		\n\t"
+	"	vfmaddpd	%%xmm15, %%xmm1 , %%xmm7 , %%xmm15		\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jz		2f						\n\t"
+
+	"	prefetcht0	384(%2,%1,8)					\n\t"
+	"	vmovddup	(%3,%1,2), %%xmm0				\n\t"	// read b
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"
+	"	vmovddup       8(%3,%1,2), %%xmm1				\n\t"	
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"
+	"	vmovups       32(%2,%1,8), %%xmm6				\n\t"
+	"	vmovups       48(%2,%1,8), %%xmm7				\n\t"
+
+	"	vfmaddpd	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"
+	"	vfmaddpd	%%xmm12, %%xmm1 , %%xmm4 , %%xmm12		\n\t"
+	"	vfmaddpd	%%xmm9 , %%xmm0 , %%xmm5 , %%xmm9		\n\t"
+	"	vfmaddpd	%%xmm13, %%xmm1 , %%xmm5 , %%xmm13		\n\t"
+	"	vfmaddpd	%%xmm10, %%xmm0 , %%xmm6 , %%xmm10		\n\t"
+	"	vfmaddpd	%%xmm14, %%xmm1 , %%xmm6 , %%xmm14		\n\t"
+	"	addq		$8, %1						\n\t"
+	"	vfmaddpd	%%xmm11, %%xmm0 , %%xmm7 , %%xmm11		\n\t"
+	"	vfmaddpd	%%xmm15, %%xmm1 , %%xmm7 , %%xmm15		\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jz		2f						\n\t"
+
+	"	prefetcht0	384(%2,%1,8)					\n\t"
+	"	vmovddup	(%3,%1,2), %%xmm0				\n\t"	// read b
+	"	vmovddup       8(%3,%1,2), %%xmm1				\n\t"	
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"
+	"	vmovups       32(%2,%1,8), %%xmm6				\n\t"
+	"	vmovups       48(%2,%1,8), %%xmm7				\n\t"
+
+	"	vfmaddpd	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"
+	"	vfmaddpd	%%xmm12, %%xmm1 , %%xmm4 , %%xmm12		\n\t"
+	"	vfmaddpd	%%xmm9 , %%xmm0 , %%xmm5 , %%xmm9		\n\t"
+	"	vfmaddpd	%%xmm13, %%xmm1 , %%xmm5 , %%xmm13		\n\t"
+	"	vfmaddpd	%%xmm10, %%xmm0 , %%xmm6 , %%xmm10		\n\t"
+	"	vfmaddpd	%%xmm14, %%xmm1 , %%xmm6 , %%xmm14		\n\t"
+	"	addq		$8, %1						\n\t"
+	"	vfmaddpd	%%xmm11, %%xmm0 , %%xmm7 , %%xmm11		\n\t"
+	"	vfmaddpd	%%xmm15, %%xmm1 , %%xmm7 , %%xmm15		\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+	"2:									\n\t"
+
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+	"	vmovups		32(%4) , %%xmm2					\n\t"
+	"	vmovups		48(%4) , %%xmm3					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+	"	vmovups		32(%5) , %%xmm6					\n\t"
+	"	vmovups		48(%5) , %%xmm7					\n\t"
+
+	"	vsubpd		%%xmm8 , %%xmm0 , %%xmm8			\n\t"
+	"	vsubpd		%%xmm9 , %%xmm1 , %%xmm9			\n\t"
+	"	vsubpd		%%xmm10, %%xmm2 , %%xmm10			\n\t"
+	"	vsubpd		%%xmm11, %%xmm3 , %%xmm11			\n\t"
+
+	"	vsubpd		%%xmm12, %%xmm4 , %%xmm12			\n\t"
+	"	vsubpd		%%xmm13, %%xmm5 , %%xmm13			\n\t"
+	"	vsubpd		%%xmm14, %%xmm6 , %%xmm14			\n\t"
+	"	vsubpd		%%xmm15, %%xmm7 , %%xmm15			\n\t"
+
+	"3:									\n\t"
+
+	"	movq		$56, %1						\n\t"	// i = 7
+	"	xorq		%0, %0						\n\t"	// pointer for bs
+
+	"	vmovddup	56(%6, %1, 8)  , %%xmm0 			\n\t"	// read aa[i]
+	"	vunpckhpd	%%xmm11 , %%xmm11, %%xmm1			\n\t"	// extract bb0
+	"	vunpckhpd	%%xmm15 , %%xmm15, %%xmm2			\n\t"	// extract bb1
+	"	vmulpd		%%xmm0  , %%xmm1 , %%xmm1			\n\t"	// bb0 * aa
+	"	vmulpd		%%xmm0  , %%xmm2 , %%xmm2			\n\t"	// bb0 * aa
+	"	vmovsd		%%xmm1  , 56(%4)				\n\t"   // c[i] = bb0 * aa
+	"	vmovsd		%%xmm2  , 56(%5)				\n\t"   // c[i] = bb1 * aa
+	"	vmovsd		%%xmm1  ,   (%7 , %0, 8)			\n\t"   // b[0] = bb0 * aa
+	"	vmovsd		%%xmm2  ,  8(%7 , %0, 8)			\n\t"   // b[1] = bb1 * aa
+
+	"	vmovups	 	0(%6, %1, 8)  , %%xmm4				\n\t"   // read a[k]
+	"	vmovups	       16(%6, %1, 8)  , %%xmm5				\n\t"   // read a[k]
+	"	vmovups	       32(%6, %1, 8)  , %%xmm6				\n\t"   // read a[k]
+	"	vmovsd	       48(%6, %1, 8)  , %%xmm7				\n\t"   // read a[k]
+	"	vfnmaddpd	%%xmm8  , %%xmm1 , %%xmm4 , %%xmm8		\n\t"
+	"	vfnmaddpd	%%xmm12 , %%xmm2 , %%xmm4 , %%xmm12		\n\t"
+	"	vfnmaddpd	%%xmm9  , %%xmm1 , %%xmm5 , %%xmm9		\n\t"
+	"	vfnmaddpd	%%xmm13 , %%xmm2 , %%xmm5 , %%xmm13		\n\t"
+	"	vfnmaddpd	%%xmm10 , %%xmm1 , %%xmm6 , %%xmm10		\n\t"
+	"	vfnmaddpd	%%xmm14 , %%xmm2 , %%xmm6 , %%xmm14		\n\t"
+	"	vfnmaddsd	%%xmm11 , %%xmm1 , %%xmm7 , %%xmm11		\n\t"
+	"	vfnmaddsd	%%xmm15 , %%xmm2 , %%xmm7 , %%xmm15		\n\t"
+
+	"	subq		$8, %1						\n\t" // i = 6
+	"	subq		$2, %0						\n\t" // b-= n
+
+	"	vmovddup	48(%6, %1, 8)  , %%xmm0 			\n\t"	// read aa[i]
+	"	vunpcklpd	%%xmm11 , %%xmm11, %%xmm1			\n\t"	// extract bb0
+	"	vunpcklpd	%%xmm15 , %%xmm15, %%xmm2			\n\t"	// extract bb1
+	"	vmulpd		%%xmm0  , %%xmm1 , %%xmm1			\n\t"	// bb0 * aa
+	"	vmulpd		%%xmm0  , %%xmm2 , %%xmm2			\n\t"	// bb1 * aa
+	"	vmovsd		%%xmm1  , 48(%4)				\n\t"   // c[i] = bb0 * aa
+	"	vmovsd		%%xmm2  , 48(%5)				\n\t"   // c[i] = bb1 * aa
+	"	vmovsd		%%xmm1  ,   (%7 , %0, 8)			\n\t"   // b[0] = bb0 * aa
+	"	vmovsd		%%xmm2  ,  8(%7 , %0, 8)			\n\t"   // b[1] = bb1 * aa
+
+	"	vmovups	 	0(%6, %1, 8)  , %%xmm4				\n\t"   // read a[k]
+	"	vmovups	       16(%6, %1, 8)  , %%xmm5				\n\t"   // read a[k]
+	"	vmovups	       32(%6, %1, 8)  , %%xmm6				\n\t"   // read a[k]
+	"	vfnmaddpd	%%xmm8  , %%xmm1 , %%xmm4 , %%xmm8		\n\t"
+	"	vfnmaddpd	%%xmm12 , %%xmm2 , %%xmm4 , %%xmm12		\n\t"
+	"	vfnmaddpd	%%xmm9  , %%xmm1 , %%xmm5 , %%xmm9		\n\t"
+	"	vfnmaddpd	%%xmm13 , %%xmm2 , %%xmm5 , %%xmm13		\n\t"
+	"	vfnmaddpd	%%xmm10 , %%xmm1 , %%xmm6 , %%xmm10		\n\t"
+	"	vfnmaddpd	%%xmm14 , %%xmm2 , %%xmm6 , %%xmm14		\n\t"
+
+	"	subq		$8, %1						\n\t" // i = 5
+	"	subq		$2, %0						\n\t" // b-= n
+
+	"	vmovddup	40(%6, %1, 8)  , %%xmm0 			\n\t"	// read aa[i]
+	"	vunpckhpd	%%xmm10 , %%xmm10, %%xmm1			\n\t"	// extract bb0
+	"	vunpckhpd	%%xmm14 , %%xmm14, %%xmm2			\n\t"	// extract bb1
+	"	vmulpd		%%xmm0  , %%xmm1 , %%xmm1			\n\t"	// bb0 * aa
+	"	vmulpd		%%xmm0  , %%xmm2 , %%xmm2			\n\t"	// bb0 * aa
+	"	vmovsd		%%xmm1  , 40(%4)				\n\t"   // c[i] = bb0 * aa
+	"	vmovsd		%%xmm2  , 40(%5)				\n\t"   // c[i] = bb1 * aa
+	"	vmovsd		%%xmm1  ,   (%7 , %0, 8)			\n\t"   // b[0] = bb0 * aa
+	"	vmovsd		%%xmm2  ,  8(%7 , %0, 8)			\n\t"   // b[1] = bb1 * aa
+
+	"	vmovups	 	0(%6, %1, 8)  , %%xmm4				\n\t"   // read a[k]
+	"	vmovups	       16(%6, %1, 8)  , %%xmm5				\n\t"   // read a[k]
+	"	vmovsd 	       32(%6, %1, 8)  , %%xmm6				\n\t"   // read a[k]
+	"	vfnmaddpd	%%xmm8  , %%xmm1 , %%xmm4 , %%xmm8		\n\t"
+	"	vfnmaddpd	%%xmm12 , %%xmm2 , %%xmm4 , %%xmm12		\n\t"
+	"	vfnmaddpd	%%xmm9  , %%xmm1 , %%xmm5 , %%xmm9		\n\t"
+	"	vfnmaddpd	%%xmm13 , %%xmm2 , %%xmm5 , %%xmm13		\n\t"
+	"	vfnmaddsd	%%xmm10 , %%xmm1 , %%xmm6 , %%xmm10		\n\t"
+	"	vfnmaddsd	%%xmm14 , %%xmm2 , %%xmm6 , %%xmm14		\n\t"
+
+	"	subq		$8, %1						\n\t" // i = 4
+	"	subq		$2, %0						\n\t" // b-= n
+
+	"	vmovddup	32(%6, %1, 8)  , %%xmm0 			\n\t"	// read aa[i]
+	"	vunpcklpd	%%xmm10 , %%xmm10, %%xmm1			\n\t"	// extract bb0
+	"	vunpcklpd	%%xmm14 , %%xmm14, %%xmm2			\n\t"	// extract bb1
+	"	vmulpd		%%xmm0  , %%xmm1 , %%xmm1			\n\t"	// bb0 * aa
+	"	vmulpd		%%xmm0  , %%xmm2 , %%xmm2			\n\t"	// bb0 * aa
+	"	vmovsd		%%xmm1  , 32(%4)				\n\t"   // c[i] = bb0 * aa
+	"	vmovsd		%%xmm2  , 32(%5)				\n\t"   // c[i] = bb1 * aa
+	"	vmovsd		%%xmm1  ,   (%7 , %0, 8)			\n\t"   // b[0] = bb0 * aa
+	"	vmovsd		%%xmm2  ,  8(%7 , %0, 8)			\n\t"   // b[1] = bb1 * aa
+
+
+	"	vmovups	 	0(%6, %1, 8)  , %%xmm4				\n\t"   // read a[k]
+	"	vmovups	       16(%6, %1, 8)  , %%xmm5				\n\t"   // read a[k]
+	"	vfnmaddpd	%%xmm8  , %%xmm1 , %%xmm4 , %%xmm8		\n\t"
+	"	vfnmaddpd	%%xmm12 , %%xmm2 , %%xmm4 , %%xmm12		\n\t"
+	"	vfnmaddpd	%%xmm9  , %%xmm1 , %%xmm5 , %%xmm9		\n\t"
+	"	vfnmaddpd	%%xmm13 , %%xmm2 , %%xmm5 , %%xmm13		\n\t"
+
+	"	subq		$8, %1						\n\t" // i = 3
+	"	subq		$2, %0						\n\t" // b-= n
+
+	"	vmovddup	24(%6, %1, 8)  , %%xmm0 			\n\t"	// read aa[i]
+	"	vunpckhpd	%%xmm9  , %%xmm9 , %%xmm1			\n\t"	// extract bb0
+	"	vunpckhpd	%%xmm13 , %%xmm13, %%xmm2			\n\t"	// extract bb1
+	"	vmulpd		%%xmm0  , %%xmm1 , %%xmm1			\n\t"	// bb0 * aa
+	"	vmulpd		%%xmm0  , %%xmm2 , %%xmm2			\n\t"	// bb0 * aa
+	"	vmovsd		%%xmm1  , 24(%4)				\n\t"   // c[i] = bb0 * aa
+	"	vmovsd		%%xmm2  , 24(%5)				\n\t"   // c[i] = bb1 * aa
+	"	vmovsd		%%xmm1  ,   (%7 , %0, 8)			\n\t"   // b[0] = bb0 * aa
+	"	vmovsd		%%xmm2  ,  8(%7 , %0, 8)			\n\t"   // b[1] = bb1 * aa
+
+	"	vmovups	 	0(%6, %1, 8)  , %%xmm4				\n\t"   // read a[k]
+	"	vmovsd 	       16(%6, %1, 8)  , %%xmm5				\n\t"   // read a[k]
+	"	vfnmaddpd	%%xmm8  , %%xmm1 , %%xmm4 , %%xmm8		\n\t"
+	"	vfnmaddpd	%%xmm12 , %%xmm2 , %%xmm4 , %%xmm12		\n\t"
+	"	vfnmaddsd	%%xmm9  , %%xmm1 , %%xmm5 , %%xmm9		\n\t"
+	"	vfnmaddsd	%%xmm13 , %%xmm2 , %%xmm5 , %%xmm13		\n\t"
+
+	"	subq		$8, %1						\n\t" // i = 2
+	"	subq		$2, %0						\n\t" // b-= n
+
+	"	vmovddup	16(%6, %1, 8)  , %%xmm0 			\n\t"	// read aa[i]
+	"	vunpcklpd	%%xmm9  , %%xmm9 , %%xmm1			\n\t"	// extract bb0
+	"	vunpcklpd	%%xmm13 , %%xmm13, %%xmm2			\n\t"	// extract bb1
+	"	vmulpd		%%xmm0  , %%xmm1 , %%xmm1			\n\t"	// bb0 * aa
+	"	vmulpd		%%xmm0  , %%xmm2 , %%xmm2			\n\t"	// bb0 * aa
+	"	vmovsd		%%xmm1  , 16(%4)				\n\t"   // c[i] = bb0 * aa
+	"	vmovsd		%%xmm2  , 16(%5)				\n\t"   // c[i] = bb1 * aa
+	"	vmovsd		%%xmm1  ,   (%7 , %0, 8)			\n\t"   // b[0] = bb0 * aa
+	"	vmovsd		%%xmm2  ,  8(%7 , %0, 8)			\n\t"   // b[1] = bb1 * aa
+
+	"	vmovups	 	0(%6, %1, 8)  , %%xmm4				\n\t"   // read a[k]
+	"	vfnmaddpd	%%xmm8  , %%xmm1 , %%xmm4 , %%xmm8		\n\t"
+	"	vfnmaddpd	%%xmm12 , %%xmm2 , %%xmm4 , %%xmm12		\n\t"
+
+	"	subq		$8, %1						\n\t" // i = 1
+	"	subq		$2, %0						\n\t" // b-= n
+
+	"	vmovddup	 8(%6, %1, 8)  , %%xmm0 			\n\t"	// read aa[i]
+	"	vunpckhpd	%%xmm8  , %%xmm8 , %%xmm1			\n\t"	// extract bb0
+	"	vunpckhpd	%%xmm12 , %%xmm12, %%xmm2			\n\t"	// extract bb1
+	"	vmulpd		%%xmm0  , %%xmm1 , %%xmm1			\n\t"	// bb0 * aa
+	"	vmulpd		%%xmm0  , %%xmm2 , %%xmm2			\n\t"	// bb0 * aa
+	"	vmovsd		%%xmm1  ,  8(%4)				\n\t"   // c[i] = bb0 * aa
+	"	vmovsd		%%xmm2  ,  8(%5)				\n\t"   // c[i] = bb1 * aa
+	"	vmovsd		%%xmm1  ,   (%7 , %0, 8)			\n\t"   // b[0] = bb0 * aa
+	"	vmovsd		%%xmm2  ,  8(%7 , %0, 8)			\n\t"   // b[1] = bb1 * aa
+
+	"	vmovsd 	 	0(%6, %1, 8)  , %%xmm4				\n\t"   // read a[k]
+	"	vfnmaddsd	%%xmm8  , %%xmm1 , %%xmm4 , %%xmm8		\n\t"
+	"	vfnmaddsd	%%xmm12 , %%xmm2 , %%xmm4 , %%xmm12		\n\t"
+
+	"	subq		$8, %1						\n\t" // i = 0
+	"	subq		$2, %0						\n\t" // b-= n
+
+	"	vmovddup	 0(%6, %1, 8)  , %%xmm0 			\n\t"	// read aa[i]
+	"	vunpcklpd	%%xmm8  , %%xmm8 , %%xmm1			\n\t"	// extract bb0
+	"	vunpcklpd	%%xmm12 , %%xmm12, %%xmm2			\n\t"	// extract bb1
+	"	vmulpd		%%xmm0  , %%xmm1 , %%xmm1			\n\t"	// bb0 * aa
+	"	vmulpd		%%xmm0  , %%xmm2 , %%xmm2			\n\t"	// bb0 * aa
+	"	vmovsd		%%xmm1  ,  0(%4)				\n\t"   // c[i] = bb0 * aa
+	"	vmovsd		%%xmm2  ,  0(%5)				\n\t"   // c[i] = bb1 * aa
+	"	vmovsd		%%xmm1  ,   (%7 , %0, 8)			\n\t"   // b[0] = bb0 * aa
+	"	vmovsd		%%xmm2  ,  8(%7 , %0, 8)			\n\t"   // b[1] = bb1 * aa
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+
+#ifndef COMPLEX
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa,  bb;
+  FLOAT *cj;
+
+
+  BLASLONG i, j, k;
+
+  a += (m - 1) * m;
+  b += (m - 1) * n;
+
+  for (i = m - 1; i >= 0; i--) {
+
+    aa = *(a + i);
+
+    for (j = 0; j < n; j ++) {
+      cj = c + j * ldc;
+      bb = *(cj + i);
+      bb *= aa;
+      *b             = bb;
+      *(cj + i) = bb;
+      b ++;
+
+      BLASLONG i1 = i & -4 ;
+      FLOAT t0,t1,t2,t3;
+
+      k=0;
+
+
+      if ( i & 4 )
+      {
+	t0 = cj[k];
+	t1 = cj[k+1];
+	t2 = cj[k+2];
+	t3 = cj[k+3];
+
+	t0 -= bb * a[k+0];
+	t1 -= bb * a[k+1];
+	t2 -= bb * a[k+2];
+	t3 -= bb * a[k+3];
+
+	cj[k+0] = t0;
+	cj[k+1] = t1;
+	cj[k+2] = t2;
+	cj[k+3] = t3;
+
+	k+=4;
+      }
+
+      if ( i & 2 )
+      {
+	t0 = a[k];
+	t1 = a[k+1];
+
+	t0 *= bb;
+	t1 *= bb;
+
+	cj[k+0] -= t0;
+	cj[k+1] -= t1;
+
+	k+=2;
+      }
+
+      if ( i & 1 ) 
+      {
+	t0 = bb * a[k];
+	cj[k+0] -= t0;
+
+      }
+
+
+    }
+    a -= m;
+    b -= 2 * n;
+  }
+
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+  a += (m - 1) * m * 2;
+  b += (m - 1) * n * 2;
+
+  for (i = m - 1; i >= 0; i--) {
+
+    aa1 = *(a + i * 2 + 0);
+    aa2 = *(a + i * 2 + 1);
+
+    for (j = 0; j < n; j ++) {
+      bb1 = *(c + i * 2 + 0 + j * ldc);
+      bb2 = *(c + i * 2 + 1 + j * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 = aa1 * bb1 + aa2 * bb2;
+      cc2 = aa1 * bb2 - aa2 * bb1;
+#endif
+
+
+      *(b + 0) = cc1;
+      *(b + 1) = cc2;
+      *(c + i * 2 + 0 + j * ldc) = cc1;
+      *(c + i * 2 + 1 + j * ldc) = cc2;
+      b += 2;
+
+      for (k = 0; k < i; k ++){
+#ifndef CONJ
+	*(c + k * 2 + 0 + j * ldc) -= cc1 * *(a + k * 2 + 0) - cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#else
+	*(c + k * 2 + 0 + j * ldc) -=   cc1 * *(a + k * 2 + 0) + cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= - cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#endif
+      }
+
+    }
+    a -= m * 2;
+    b -= 4 * n;
+  }
+
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k,  FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  BLASLONG i, j;
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+
+#if 0
+  fprintf(stderr, "TRSM KERNEL LN : m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  while (j > 0) {
+
+    kk = m + offset;
+
+    if (m & (GEMM_UNROLL_M - 1)) {
+      for (i = 1; i < GEMM_UNROLL_M; i *= 2){
+	if (m & i) {
+	  aa = a + ((m & ~(i - 1)) - i) * k * COMPSIZE;
+	  cc = c + ((m & ~(i - 1)) - i)     * COMPSIZE;
+
+	  if (k - kk > 0) {
+
+	    GEMM_KERNEL(i, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa + i             * kk * COMPSIZE,
+			b  + GEMM_UNROLL_N * kk * COMPSIZE,
+			cc,
+			ldc);
+
+	  }
+
+	  solve(i, GEMM_UNROLL_N,
+		aa + (kk - i) * i             * COMPSIZE,
+		b  + (kk - i) * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+	  kk -= i;
+	}
+      }
+    }
+
+    i = (m >> GEMM_UNROLL_M_SHIFT);
+    if (i > 0) {
+      aa = a + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M) * k * COMPSIZE;
+      cc = c + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M)     * COMPSIZE;
+
+      do {
+		dtrsm_LN_solve_opt(k-kk, aa + GEMM_UNROLL_M * kk * COMPSIZE, b +  GEMM_UNROLL_N * kk * COMPSIZE, cc, ldc
+                             , aa + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_M * COMPSIZE,b  + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_N * COMPSIZE);
+	
+
+		aa -= GEMM_UNROLL_M * k * COMPSIZE;
+		cc -= GEMM_UNROLL_M     * COMPSIZE;
+		kk -= GEMM_UNROLL_M;
+		i --;
+         } while (i > 0);
+    }
+
+    b += GEMM_UNROLL_N * k * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+	kk = m + offset;
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  for (i = 1; i < GEMM_UNROLL_M; i *= 2){
+	    if (m & i) {
+	      aa = a + ((m & ~(i - 1)) - i) * k * COMPSIZE;
+	      cc = c + ((m & ~(i - 1)) - i)     * COMPSIZE;
+
+	      if (k - kk > 0) {
+		GEMM_KERNEL(i, j, k - kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa + i * kk * COMPSIZE,
+			    b  + j * kk * COMPSIZE,
+			    cc, ldc);
+	      }
+
+	      solve(i, j,
+		    aa + (kk - i) * i * COMPSIZE,
+		    b  + (kk - i) * j * COMPSIZE,
+		    cc, ldc);
+
+	      kk -= i;
+	    }
+	  }
+	}
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+	if (i > 0) {
+	  aa = a + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M) * k * COMPSIZE;
+	  cc = c + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M)     * COMPSIZE;
+
+	  do {
+	    if (k - kk > 0) {
+	      GEMM_KERNEL(GEMM_UNROLL_M, j, k - kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa + GEMM_UNROLL_M * kk * COMPSIZE,
+			  b +  j             * kk * COMPSIZE,
+			  cc,
+			  ldc);
+	    }
+
+	    solve(GEMM_UNROLL_M, j,
+		  aa + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_M * COMPSIZE,
+		  b  + (kk - GEMM_UNROLL_M) * j             * COMPSIZE,
+		  cc, ldc);
+
+	    aa -= GEMM_UNROLL_M * k * COMPSIZE;
+	    cc -= GEMM_UNROLL_M     * COMPSIZE;
+	    kk -= GEMM_UNROLL_M;
+	    i --;
+	  } while (i > 0);
+	}
+
+	b += j * k   * COMPSIZE;
+	c += j * ldc * COMPSIZE;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/x86_64/dtrsm_kernel_RN_haswell.c

@@ -0,0 +1,677 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* 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.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``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 UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN 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.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_R
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+
+static void dtrsm_RN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void dtrsm_RN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c3 = c + ldc + ldc*2 ;
+	FLOAT *c6 = c + ldc*4 + ldc*2 ;
+	ldc = ldc *8;
+	BLASLONG n1 = n * 8;
+	BLASLONG i=0;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+
+	"	vxorpd	%%ymm8 , %%ymm8 , %%ymm8				\n\t"
+	"	vxorpd	%%ymm9 , %%ymm9 , %%ymm9				\n\t"
+	"	vxorpd	%%ymm10, %%ymm10, %%ymm10				\n\t"
+	"	vxorpd	%%ymm11, %%ymm11, %%ymm11				\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"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       4f						\n\t"
+
+	"	vmovups         (%2,%1,4), %%ymm0				\n\t"	// read a
+	"	vmovups         (%3,%1,8), %%ymm1				\n\t"	// read b0
+	"	vmovups       32(%3,%1,8), %%ymm2				\n\t"	// read b1
+
+
+	"	addq		$8, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+	"	je	       21f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	vmovups         (%2,%1,4), %%ymm4				\n\t"	// read a
+        "       vpermpd         $0xb1  , %%ymm0 , %%ymm3                	\n\t"
+
+	"	vfmadd231pd	%%ymm0 , %%ymm1 , %%ymm8			\n\t"
+	"	vfmadd231pd	%%ymm0 , %%ymm2 , %%ymm12			\n\t"
+
+	"	vmovups         (%3,%1,8), %%ymm5				\n\t"	// read b0
+	"	vfmadd231pd	%%ymm3 , %%ymm1 , %%ymm9			\n\t"
+	"	vfmadd231pd	%%ymm3 , %%ymm2 , %%ymm13			\n\t"
+
+        "       vpermpd         $0x1b  , %%ymm3 , %%ymm0                	\n\t"
+	"	vmovups       32(%3,%1,8), %%ymm6				\n\t"	// read b1
+        "       vpermpd         $0xb1  , %%ymm0 , %%ymm3                	\n\t"
+	"	vfmadd231pd	%%ymm0 , %%ymm1 , %%ymm10			\n\t"
+	"	vfmadd231pd	%%ymm0 , %%ymm2 , %%ymm14			\n\t"
+
+	"	addq		$8, %1						\n\t"
+	"	vfmadd231pd	%%ymm3 , %%ymm1 , %%ymm11			\n\t"
+	"	vfmadd231pd	%%ymm3 , %%ymm2 , %%ymm15			\n\t"
+
+	"	cmpq		%1, %0						\n\t"
+
+	"	jz		22f						\n\t"
+
+	"	vmovups         (%2,%1,4), %%ymm0				\n\t"	// read a
+
+	"	vfmadd231pd	%%ymm4 , %%ymm5 , %%ymm8			\n\t"
+	"	vfmadd231pd	%%ymm4 , %%ymm6 , %%ymm12			\n\t"
+
+        "       vpermpd         $0xb1  , %%ymm4 , %%ymm4                	\n\t"
+	"	vmovups         (%3,%1,8), %%ymm1				\n\t"	// read b0
+	"	vfmadd231pd	%%ymm4 , %%ymm5 , %%ymm9			\n\t"
+	"	vfmadd231pd	%%ymm4 , %%ymm6 , %%ymm13			\n\t"
+
+        "       vpermpd         $0x1b  , %%ymm4 , %%ymm4                	\n\t"
+	"	vmovups       32(%3,%1,8), %%ymm2				\n\t"	// read b1
+	"	vfmadd231pd	%%ymm4 , %%ymm5 , %%ymm10			\n\t"
+	"	vfmadd231pd	%%ymm4 , %%ymm6 , %%ymm14			\n\t"
+
+        "       vpermpd         $0xb1  , %%ymm4 , %%ymm4                	\n\t"
+	"	addq		$8, %1						\n\t"
+	"	vfmadd231pd	%%ymm4 , %%ymm5 , %%ymm11			\n\t"
+	"	vfmadd231pd	%%ymm4 , %%ymm6 , %%ymm15			\n\t"
+
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+
+	"21:									\n\t"
+
+	"	vfmadd231pd	%%ymm0 , %%ymm1 , %%ymm8			\n\t"
+	"	vfmadd231pd	%%ymm0 , %%ymm2 , %%ymm12			\n\t"
+
+        "       vpermpd         $0xb1  , %%ymm0 , %%ymm0                	\n\t"
+	"	vfmadd231pd	%%ymm0 , %%ymm1 , %%ymm9			\n\t"
+	"	vfmadd231pd	%%ymm0 , %%ymm2 , %%ymm13			\n\t"
+
+        "       vpermpd         $0x1b  , %%ymm0 , %%ymm0                	\n\t"
+	"	vfmadd231pd	%%ymm0 , %%ymm1 , %%ymm10			\n\t"
+	"	vfmadd231pd	%%ymm0 , %%ymm2 , %%ymm14			\n\t"
+
+        "       vpermpd         $0xb1  , %%ymm0 , %%ymm0                	\n\t"
+	"	vfmadd231pd	%%ymm0 , %%ymm1 , %%ymm11			\n\t"
+	"	vfmadd231pd	%%ymm0 , %%ymm2 , %%ymm15			\n\t"
+
+	"	jmp	3f							\n\t"
+
+	"22:									\n\t"
+	
+	"	vfmadd231pd	%%ymm4 , %%ymm5 , %%ymm8			\n\t"
+	"	vfmadd231pd	%%ymm4 , %%ymm6 , %%ymm12			\n\t"
+
+        "       vpermpd         $0xb1  , %%ymm4 , %%ymm4                	\n\t"
+	"	vfmadd231pd	%%ymm4 , %%ymm5 , %%ymm9			\n\t"
+	"	vfmadd231pd	%%ymm4 , %%ymm6 , %%ymm13			\n\t"
+
+        "       vpermpd         $0x1b  , %%ymm4 , %%ymm4                	\n\t"
+	"	vfmadd231pd	%%ymm4 , %%ymm5 , %%ymm10			\n\t"
+	"	vfmadd231pd	%%ymm4 , %%ymm6 , %%ymm14			\n\t"
+
+        "       vpermpd         $0xb1  , %%ymm4 , %%ymm4                	\n\t"
+	"	vfmadd231pd	%%ymm4 , %%ymm5 , %%ymm11			\n\t"
+	"	vfmadd231pd	%%ymm4 , %%ymm6 , %%ymm15			\n\t"
+
+	"3:								\n\t"	
+
+        "       vpermpd         $0xb1  , %%ymm9 , %%ymm9                \n\t"
+        "       vpermpd         $0xb1  , %%ymm11, %%ymm11               \n\t"
+
+        "       vblendpd        $0x0a  , %%ymm9 , %%ymm8 , %%ymm0       \n\t"
+        "       vblendpd        $0x05  , %%ymm9 , %%ymm8 , %%ymm1       \n\t"
+        "       vblendpd        $0x0a  , %%ymm11, %%ymm10, %%ymm2       \n\t"
+        "       vblendpd        $0x05  , %%ymm11, %%ymm10, %%ymm3       \n\t"
+
+        "       vpermpd         $0x1b  , %%ymm2 , %%ymm2                \n\t"
+        "       vpermpd         $0x1b  , %%ymm3 , %%ymm3                \n\t"
+        "       vpermpd         $0xb1  , %%ymm2 , %%ymm2                \n\t"
+        "       vpermpd         $0xb1  , %%ymm3 , %%ymm3                \n\t"
+
+        "       vblendpd        $0x03  , %%ymm0 , %%ymm2 , %%ymm8       \n\t"
+        "       vblendpd        $0x03  , %%ymm1 , %%ymm3 , %%ymm9       \n\t"
+        "       vblendpd        $0x03  , %%ymm2 , %%ymm0 , %%ymm10      \n\t"
+        "       vblendpd        $0x03  , %%ymm3 , %%ymm1 , %%ymm11      \n\t"
+
+        "       vpermpd         $0xb1  , %%ymm13, %%ymm13               \n\t"
+        "       vpermpd         $0xb1  , %%ymm15, %%ymm15               \n\t"
+
+        "       vblendpd        $0x0a  , %%ymm13, %%ymm12, %%ymm0       \n\t"
+        "       vblendpd        $0x05  , %%ymm13, %%ymm12, %%ymm1       \n\t"
+        "       vblendpd        $0x0a  , %%ymm15, %%ymm14, %%ymm2       \n\t"
+        "       vblendpd        $0x05  , %%ymm15, %%ymm14, %%ymm3       \n\t"
+
+        "       vpermpd         $0x1b  , %%ymm2 , %%ymm2                \n\t"
+        "       vpermpd         $0x1b  , %%ymm3 , %%ymm3                \n\t"
+        "       vpermpd         $0xb1  , %%ymm2 , %%ymm2                \n\t"
+        "       vpermpd         $0xb1  , %%ymm3 , %%ymm3                \n\t"
+
+        "       vblendpd        $0x03  , %%ymm0 , %%ymm2 , %%ymm12      \n\t"
+        "       vblendpd        $0x03  , %%ymm1 , %%ymm3 , %%ymm13      \n\t"
+        "       vblendpd        $0x03  , %%ymm2 , %%ymm0 , %%ymm14      \n\t"
+        "       vblendpd        $0x03  , %%ymm3 , %%ymm1 , %%ymm15      \n\t"
+
+
+	"4:								\n\t"	
+
+	"	vmovups		  (%4)      , %%ymm0			\n\t"	// read c0
+	"	vmovups		  (%4,%7,1) , %%ymm1			\n\t"	// read c1
+	"	vmovups		  (%4,%7,2) , %%ymm2			\n\t"	// read c2
+	"	vmovups		  (%5)      , %%ymm3			\n\t"	// read c3
+
+	"	vmovups		  (%5,%7,1) , %%ymm4			\n\t"	// read c4
+	"	vmovups		  (%5,%7,2) , %%ymm5			\n\t"	// read c5
+	"	vmovups		  (%6)      , %%ymm6			\n\t"	// read c6
+	"	vmovups		  (%6,%7,1) , %%ymm7			\n\t"	// read c7
+
+	"	vsubpd		%%ymm8 , %%ymm0 , %%ymm8		\n\t"
+	"	vmovups           (%9),  %%ymm0				\n\t"
+	"	vsubpd		%%ymm9 , %%ymm1 , %%ymm9		\n\t"
+	"	vpermpd		$0x55 ,  %%ymm0 , %%ymm1		\n\t"
+	"	vsubpd		%%ymm10, %%ymm2 , %%ymm10		\n\t"
+	"	vpermpd		$0xaa ,  %%ymm0 , %%ymm2		\n\t"
+	"	vsubpd		%%ymm11, %%ymm3 , %%ymm11		\n\t"
+	"	vpermpd		$0xff ,  %%ymm0 , %%ymm3		\n\t"
+	"	vpermpd		$0x00 ,  %%ymm0 , %%ymm0		\n\t"
+
+	"	vsubpd		%%ymm12, %%ymm4 , %%ymm12		\n\t"
+	"	vmovups         32(%9),  %%ymm4				\n\t"
+	"	vsubpd		%%ymm13, %%ymm5 , %%ymm13		\n\t"
+	"	vpermpd		$0x55 ,  %%ymm4 , %%ymm5		\n\t"
+	"	vsubpd		%%ymm14, %%ymm6 , %%ymm14		\n\t"
+	"	vpermpd		$0xaa ,  %%ymm4 , %%ymm6		\n\t"
+	"	vsubpd		%%ymm15, %%ymm7 , %%ymm15		\n\t"
+	"	vpermpd		$0xff ,  %%ymm4 , %%ymm7		\n\t"
+	"	vpermpd		$0x00 ,  %%ymm4 , %%ymm4		\n\t"
+
+
+	"5:								\n\t"	// i = 0
+
+	"	addq	$64, %9						\n\t"	// b=b+8
+
+	"	vmulpd		%%ymm8 , %%ymm0, %%ymm8			\n\t"	// a *bb
+	"	vmovups           (%9),  %%ymm0				\n\t"
+	"	vmovups		%%ymm8 , (%8)				\n\t"	// write a
+	"	vmovups		%%ymm8 , (%4)				\n\t"	// write c
+
+	"	vfnmadd231pd	%%ymm8 , %%ymm1 , %%ymm9		\n\t"
+	"	vmovups         32(%9),  %%ymm1				\n\t"
+	"	vfnmadd231pd	%%ymm8 , %%ymm2 , %%ymm10		\n\t"
+	"	vpermpd		$0xaa ,  %%ymm0 , %%ymm2		\n\t"
+	"	vfnmadd231pd	%%ymm8 , %%ymm3 , %%ymm11		\n\t"
+	"	vpermpd		$0xff ,  %%ymm0 , %%ymm3		\n\t"
+	"	vfnmadd231pd	%%ymm8 , %%ymm4 , %%ymm12		\n\t"
+	"	vpermpd		$0x55 ,  %%ymm0 , %%ymm0		\n\t"
+	"	vfnmadd231pd	%%ymm8 , %%ymm5 , %%ymm13		\n\t"
+	"	vpermpd		$0x55 ,  %%ymm1 , %%ymm5		\n\t"
+	"	vfnmadd231pd	%%ymm8 , %%ymm6 , %%ymm14		\n\t"
+	"	vpermpd		$0xaa ,  %%ymm1 , %%ymm6		\n\t"
+	"	vfnmadd231pd	%%ymm8 , %%ymm7 , %%ymm15		\n\t"
+	"	vpermpd		$0xff ,  %%ymm1 , %%ymm7		\n\t"
+	"	vpermpd		$0x00 ,  %%ymm1 , %%ymm4		\n\t"
+
+	"	addq	$64, %9						\n\t"	// b=b+8
+	"	addq	$32, %8						\n\t"	// a=a+8
+
+
+
+	"	vmulpd		%%ymm9 , %%ymm0, %%ymm9			\n\t"	// a *bb
+	"	vmovups           (%9),  %%ymm0				\n\t"
+	"	vmovups         32(%9),  %%ymm1				\n\t"
+	"	vmovups		%%ymm9 , (%8)				\n\t"	// write a
+	"	vmovups		%%ymm9 , (%4,%7,1)			\n\t"	// write c
+
+	"	vfnmadd231pd	%%ymm9 , %%ymm2 , %%ymm10		\n\t"
+	"	vfnmadd231pd	%%ymm9 , %%ymm3 , %%ymm11		\n\t"
+	"	vpermpd		$0xff ,  %%ymm0 , %%ymm3		\n\t"
+	"	vfnmadd231pd	%%ymm9 , %%ymm4 , %%ymm12		\n\t"
+	"	vpermpd		$0xaa ,  %%ymm0 , %%ymm0		\n\t"
+	"	vfnmadd231pd	%%ymm9 , %%ymm5 , %%ymm13		\n\t"
+	"	vpermpd		$0x55 ,  %%ymm1 , %%ymm5		\n\t"
+	"	vfnmadd231pd	%%ymm9 , %%ymm6 , %%ymm14		\n\t"
+	"	vpermpd		$0xaa ,  %%ymm1 , %%ymm6		\n\t"
+	"	vfnmadd231pd	%%ymm9 , %%ymm7 , %%ymm15		\n\t"
+	"	vpermpd		$0xff ,  %%ymm1 , %%ymm7		\n\t"
+	"	vpermpd		$0x00 ,  %%ymm1 , %%ymm4		\n\t"
+
+	"	addq	$64, %9						\n\t"	// b=b+8
+	"	addq	$32, %8						\n\t"	// a=a+8
+
+	"	vmulpd		%%ymm10, %%ymm0, %%ymm10		\n\t"	// a *bb
+	"	vmovups           (%9),  %%ymm0				\n\t"
+	"	vmovups         32(%9),  %%ymm1				\n\t"
+	"	vmovups		%%ymm10, (%8)				\n\t"	// write a
+	"	vmovups		%%ymm10, (%4,%7,2)			\n\t"	// write c
+
+	"	vfnmadd231pd	%%ymm10, %%ymm3 , %%ymm11		\n\t"
+	"	vpermpd		$0xff ,  %%ymm0 , %%ymm0		\n\t"
+	"	vfnmadd231pd	%%ymm10, %%ymm4 , %%ymm12		\n\t"
+	"	vfnmadd231pd	%%ymm10, %%ymm5 , %%ymm13		\n\t"
+	"	vpermpd		$0x55 ,  %%ymm1 , %%ymm5		\n\t"
+	"	vfnmadd231pd	%%ymm10, %%ymm6 , %%ymm14		\n\t"
+	"	vpermpd		$0xaa ,  %%ymm1 , %%ymm6		\n\t"
+	"	vfnmadd231pd	%%ymm10, %%ymm7 , %%ymm15		\n\t"
+	"	vpermpd		$0xff ,  %%ymm1 , %%ymm7		\n\t"
+	"	vpermpd		$0x00 ,  %%ymm1 , %%ymm4		\n\t"
+
+
+	"	addq	$64, %9						\n\t"	// b=b+8
+	"	addq	$32, %8						\n\t"	// a=a+8
+
+
+
+	"	vmulpd		%%ymm11, %%ymm0, %%ymm11		\n\t"	// a *bb
+	"	vmovups         32(%9),  %%ymm1				\n\t"
+	"	vmovups		%%ymm11, (%8)				\n\t"	// write a
+	"	vmovups		%%ymm11, (%5)     			\n\t"	// write c
+
+	"	vfnmadd231pd	%%ymm11, %%ymm4 , %%ymm12		\n\t"
+	"	vfnmadd231pd	%%ymm11, %%ymm5 , %%ymm13		\n\t"
+	"	vpermpd		$0x55 ,  %%ymm1 , %%ymm5		\n\t"
+	"	vfnmadd231pd	%%ymm11, %%ymm6 , %%ymm14		\n\t"
+	"	vpermpd		$0xaa ,  %%ymm1 , %%ymm6		\n\t"
+	"	vfnmadd231pd	%%ymm11, %%ymm7 , %%ymm15		\n\t"
+	"	vpermpd		$0xff ,  %%ymm1 , %%ymm7		\n\t"
+	"	vpermpd		$0x00 ,  %%ymm1 , %%ymm0		\n\t"
+
+
+	"	addq	$64, %9						\n\t"	// b=b+8
+	"	addq	$32, %8						\n\t"	// a=a+8
+
+
+	"	vmulpd		%%ymm12, %%ymm0, %%ymm12		\n\t"	// a *bb
+	"	vmovups         32(%9),  %%ymm1				\n\t"
+	"	vmovups		%%ymm12, (%8)				\n\t"	// write a
+	"	vmovups		%%ymm12, (%5,%7,1)			\n\t"	// write c
+
+	"	vfnmadd231pd	%%ymm12, %%ymm5 , %%ymm13		\n\t"
+	"	vfnmadd231pd	%%ymm12, %%ymm6 , %%ymm14		\n\t"
+	"	vpermpd		$0xaa ,  %%ymm1 , %%ymm6		\n\t"
+	"	vfnmadd231pd	%%ymm12, %%ymm7 , %%ymm15		\n\t"
+	"	vpermpd		$0xff ,  %%ymm1 , %%ymm7		\n\t"
+	"	vpermpd		$0x55 ,  %%ymm1 , %%ymm0		\n\t"
+
+	"	addq	$64, %9						\n\t"	// b=b+8
+	"	addq	$32, %8						\n\t"	// a=a+8
+
+	"	vmulpd		%%ymm13, %%ymm0, %%ymm13		\n\t"	// a *bb
+	"	vmovups         32(%9),  %%ymm1				\n\t"
+	"	vmovups		%%ymm13, (%8)				\n\t"	// write a
+	"	vmovups		%%ymm13, (%5,%7,2)			\n\t"	// write c
+
+	"	vfnmadd231pd	%%ymm13, %%ymm6 , %%ymm14		\n\t"
+	"	vfnmadd231pd	%%ymm13, %%ymm7 , %%ymm15		\n\t"
+	"	vpermpd		$0xff ,  %%ymm1 , %%ymm7		\n\t"
+	"	vpermpd		$0xaa ,  %%ymm1 , %%ymm0		\n\t"
+
+
+	"	addq	$64, %9						\n\t"	// b=b+8
+	"	addq	$32, %8						\n\t"	// a=a+8
+
+
+	"	vmulpd		%%ymm14, %%ymm0, %%ymm14		\n\t"	// a *bb
+	"	vmovups         32(%9),  %%ymm1				\n\t"
+	"	vmovups		%%ymm14, (%8)				\n\t"	// write a
+	"	vmovups		%%ymm14, (%6)     			\n\t"	// write c
+
+	"	vfnmadd231pd	%%ymm14, %%ymm7 , %%ymm15		\n\t"
+
+	"	vpermpd		$0xff ,  %%ymm1 , %%ymm0		\n\t"
+
+	"	addq	$32, %8						\n\t"	// a=a+8
+
+	"	vmulpd		%%ymm15, %%ymm0, %%ymm15		\n\t"	// a *bb
+	"	vmovups		%%ymm15, (%8)				\n\t"	// write a
+	"	vmovups		%%ymm15, (%6,%7,1)			\n\t"	// write c
+
+	"	vzeroupper						\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c3),     // 5
+          "r" (c6),     // 6
+          "r" (ldc),    // 7
+          "r" (as),     // 8
+          "r" (bs)      // 9
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa, bb;
+
+  int i, j, k;
+
+  for (i = 0; i < n; i++) {
+
+    bb = *(b + i);
+
+    for (j = 0; j < m; j ++) {
+      aa = *(c + j + i * ldc);
+      aa *= bb;
+      *a  = aa;
+      *(c + j + i * ldc) = aa;
+      a ++;
+
+      for (k = i + 1; k < n; k ++){
+	*(c + j + k * ldc) -= aa * *(b + k);
+      }
+
+    }
+    b += n;
+  }
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  for (i = 0; i < n; i++) {
+
+    bb1 = *(b + i * 2 + 0);
+    bb2 = *(b + i * 2 + 1);
+
+    for (j = 0; j < m; j ++) {
+      aa1 = *(c + j * 2 + 0 + i * ldc);
+      aa2 = *(c + j * 2 + 1 + i * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 =  aa1 * bb1 + aa2 * bb2;
+      cc2 = -aa1 * bb2 + aa2 * bb1;
+#endif
+
+      *(a + 0) = cc1;
+      *(a + 1) = cc2;
+      *(c + j * 2 + 0 + i * ldc) = cc1;
+      *(c + j * 2 + 1 + i * ldc) = cc2;
+      a += 2;
+
+      for (k = i + 1; k < n; k ++){
+#ifndef CONJ
+	*(c + j * 2 + 0 + k * ldc) -= cc1 * *(b + k * 2 + 0) - cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#else
+	*(c + j * 2 + 0 + k * ldc) -=   cc1 * *(b + k * 2 + 0) + cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= - cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#endif
+      }
+
+    }
+    b += n * 2;
+  }
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+  BLASLONG i, j, jj;
+
+#if 0
+  fprintf(stderr, "TRSM RN KERNEL m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  jj = 0;
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+  kk = -offset;
+
+  while (j > 0) {
+
+    aa = a;
+    cc = c;
+
+    i = (m >> GEMM_UNROLL_M_SHIFT);
+
+    if (i > 0) {
+      do {
+
+	dtrsm_RN_solve_opt(kk, aa, b, cc, ldc, aa + kk * GEMM_UNROLL_M * COMPSIZE, b + kk * GEMM_UNROLL_N * COMPSIZE);
+/* 
+        solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+              aa + kk * GEMM_UNROLL_M * COMPSIZE,
+              b  + kk * GEMM_UNROLL_N * COMPSIZE,
+              cc, ldc);
+*/
+	aa += GEMM_UNROLL_M * k * COMPSIZE;
+	cc += GEMM_UNROLL_M     * COMPSIZE;
+	i --;
+      } while (i > 0);
+    }
+
+
+    if (m & (GEMM_UNROLL_M - 1)) {
+      i = (GEMM_UNROLL_M >> 1);
+      while (i > 0) {
+	if (m & i) {
+	    if (kk > 0) {
+	      GEMM_KERNEL(i, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa, b, cc, ldc);
+	    }
+	  solve(i, GEMM_UNROLL_N,
+		aa + kk * i             * COMPSIZE,
+		b  + kk * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+	  aa += i * k * COMPSIZE;
+	  cc += i     * COMPSIZE;
+	}
+	i >>= 1;
+      }
+    }
+
+    kk += GEMM_UNROLL_N;
+    b += GEMM_UNROLL_N * k   * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+    jj += GEMM_UNROLL_M;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+	aa = a;
+	cc = c;
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+
+	while (i > 0) {
+	  if (kk > 0) {
+	    GEMM_KERNEL(GEMM_UNROLL_M, j, kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa,
+			b,
+			cc,
+			ldc);
+	  }
+
+	  solve(GEMM_UNROLL_M, j,
+		aa + kk * GEMM_UNROLL_M * COMPSIZE,
+		b  + kk * j             * COMPSIZE, cc, ldc);
+
+	  aa += GEMM_UNROLL_M * k * COMPSIZE;
+	  cc += GEMM_UNROLL_M     * COMPSIZE;
+	  i --;
+	}
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  i = (GEMM_UNROLL_M >> 1);
+	  while (i > 0) {
+	    if (m & i) {
+	      if (kk > 0) {
+		GEMM_KERNEL(i, j, kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa,
+			    b,
+			    cc,
+			    ldc);
+	      }
+
+	      solve(i, j,
+		    aa + kk * i * COMPSIZE,
+		    b  + kk * j * COMPSIZE, cc, ldc);
+
+	      aa += i * k * COMPSIZE;
+	      cc += i     * COMPSIZE;
+	      }
+	    i >>= 1;
+	  }
+	}
+
+	b += j * k   * COMPSIZE;
+	c += j * ldc * COMPSIZE;
+	kk += j;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/x86_64/dtrsm_kernel_RT_bulldozer.c

@@ -0,0 +1,546 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* 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.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``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 UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN 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.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_R
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+
+static void dtrsm_RT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void dtrsm_RT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc ;
+	BLASLONG n1 = n * 8;
+	BLASLONG i=0;
+
+  	as += (2 - 1) * 8;
+  	bs += (2 - 1) * 2;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorpd	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorpd	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorpd	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorpd	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorpd	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorpd	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorpd	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorpd	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       2f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	prefetcht0	384(%2,%1,8)					\n\t"
+	"	prefetcht0	384(%3,%1,8)					\n\t"
+	"	vmovddup	(%3,%1,2), %%xmm0				\n\t"	// read b
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"
+	"	vmovddup       8(%3,%1,2), %%xmm1				\n\t"	
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"
+	"	vmovups       32(%2,%1,8), %%xmm6				\n\t"
+	"	vmovups       48(%2,%1,8), %%xmm7				\n\t"
+
+	"	vfmaddpd	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"
+	"	vfmaddpd	%%xmm12, %%xmm1 , %%xmm4 , %%xmm12		\n\t"
+	"	vfmaddpd	%%xmm9 , %%xmm0 , %%xmm5 , %%xmm9		\n\t"
+	"	vfmaddpd	%%xmm13, %%xmm1 , %%xmm5 , %%xmm13		\n\t"
+	"	vfmaddpd	%%xmm10, %%xmm0 , %%xmm6 , %%xmm10		\n\t"
+	"	vfmaddpd	%%xmm14, %%xmm1 , %%xmm6 , %%xmm14		\n\t"
+	"	addq		$8, %1						\n\t"
+	"	vfmaddpd	%%xmm11, %%xmm0 , %%xmm7 , %%xmm11		\n\t"
+	"	vfmaddpd	%%xmm15, %%xmm1 , %%xmm7 , %%xmm15		\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jz		2f						\n\t"
+
+	"	prefetcht0	384(%2,%1,8)					\n\t"
+	"	vmovddup	(%3,%1,2), %%xmm0				\n\t"	// read b
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"
+	"	vmovddup       8(%3,%1,2), %%xmm1				\n\t"	
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"
+	"	vmovups       32(%2,%1,8), %%xmm6				\n\t"
+	"	vmovups       48(%2,%1,8), %%xmm7				\n\t"
+
+	"	vfmaddpd	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"
+	"	vfmaddpd	%%xmm12, %%xmm1 , %%xmm4 , %%xmm12		\n\t"
+	"	vfmaddpd	%%xmm9 , %%xmm0 , %%xmm5 , %%xmm9		\n\t"
+	"	vfmaddpd	%%xmm13, %%xmm1 , %%xmm5 , %%xmm13		\n\t"
+	"	vfmaddpd	%%xmm10, %%xmm0 , %%xmm6 , %%xmm10		\n\t"
+	"	vfmaddpd	%%xmm14, %%xmm1 , %%xmm6 , %%xmm14		\n\t"
+	"	addq		$8, %1						\n\t"
+	"	vfmaddpd	%%xmm11, %%xmm0 , %%xmm7 , %%xmm11		\n\t"
+	"	vfmaddpd	%%xmm15, %%xmm1 , %%xmm7 , %%xmm15		\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jz		2f						\n\t"
+
+	"	prefetcht0	384(%2,%1,8)					\n\t"
+	"	vmovddup	(%3,%1,2), %%xmm0				\n\t"	// read b
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"
+	"	vmovddup       8(%3,%1,2), %%xmm1				\n\t"	
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"
+	"	vmovups       32(%2,%1,8), %%xmm6				\n\t"
+	"	vmovups       48(%2,%1,8), %%xmm7				\n\t"
+
+	"	vfmaddpd	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"
+	"	vfmaddpd	%%xmm12, %%xmm1 , %%xmm4 , %%xmm12		\n\t"
+	"	vfmaddpd	%%xmm9 , %%xmm0 , %%xmm5 , %%xmm9		\n\t"
+	"	vfmaddpd	%%xmm13, %%xmm1 , %%xmm5 , %%xmm13		\n\t"
+	"	vfmaddpd	%%xmm10, %%xmm0 , %%xmm6 , %%xmm10		\n\t"
+	"	vfmaddpd	%%xmm14, %%xmm1 , %%xmm6 , %%xmm14		\n\t"
+	"	addq		$8, %1						\n\t"
+	"	vfmaddpd	%%xmm11, %%xmm0 , %%xmm7 , %%xmm11		\n\t"
+	"	vfmaddpd	%%xmm15, %%xmm1 , %%xmm7 , %%xmm15		\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jz		2f						\n\t"
+
+	"	prefetcht0	384(%2,%1,8)					\n\t"
+	"	vmovddup	(%3,%1,2), %%xmm0				\n\t"	// read b
+	"	vmovddup       8(%3,%1,2), %%xmm1				\n\t"	
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"
+	"	vmovups       32(%2,%1,8), %%xmm6				\n\t"
+	"	vmovups       48(%2,%1,8), %%xmm7				\n\t"
+
+	"	vfmaddpd	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"
+	"	vfmaddpd	%%xmm12, %%xmm1 , %%xmm4 , %%xmm12		\n\t"
+	"	vfmaddpd	%%xmm9 , %%xmm0 , %%xmm5 , %%xmm9		\n\t"
+	"	vfmaddpd	%%xmm13, %%xmm1 , %%xmm5 , %%xmm13		\n\t"
+	"	vfmaddpd	%%xmm10, %%xmm0 , %%xmm6 , %%xmm10		\n\t"
+	"	vfmaddpd	%%xmm14, %%xmm1 , %%xmm6 , %%xmm14		\n\t"
+	"	addq		$8, %1						\n\t"
+	"	vfmaddpd	%%xmm11, %%xmm0 , %%xmm7 , %%xmm11		\n\t"
+	"	vfmaddpd	%%xmm15, %%xmm1 , %%xmm7 , %%xmm15		\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+	"2:									\n\t"
+
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+	"	vmovups		32(%4) , %%xmm2					\n\t"
+	"	vmovups		48(%4) , %%xmm3					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+	"	vmovups		32(%5) , %%xmm6					\n\t"
+	"	vmovups		48(%5) , %%xmm7					\n\t"
+
+	"	vsubpd		%%xmm8 , %%xmm0 , %%xmm8			\n\t"
+	"	vsubpd		%%xmm9 , %%xmm1 , %%xmm9			\n\t"
+	"	vsubpd		%%xmm10, %%xmm2 , %%xmm10			\n\t"
+	"	vsubpd		%%xmm11, %%xmm3 , %%xmm11			\n\t"
+
+	"	vsubpd		%%xmm12, %%xmm4 , %%xmm12			\n\t"
+	"	vsubpd		%%xmm13, %%xmm5 , %%xmm13			\n\t"
+	"	vsubpd		%%xmm14, %%xmm6 , %%xmm14			\n\t"
+	"	vsubpd		%%xmm15, %%xmm7 , %%xmm15			\n\t"
+
+	"3:									\n\t"	// i = 1
+
+	"	vmovddup	(%7), %%xmm1					\n\t"	// read b
+	"	vmovddup       8(%7), %%xmm0					\n\t"	// read bb
+
+	"	vmulpd		%%xmm12 ,  %%xmm0 ,  %%xmm12			\n\t"	// aa * bb 
+	"	vmulpd		%%xmm13 ,  %%xmm0 ,  %%xmm13			\n\t"	// aa * bb 
+	"	vmulpd		%%xmm14 ,  %%xmm0 ,  %%xmm14			\n\t"	// aa * bb 
+	"	vmulpd		%%xmm15 ,  %%xmm0 ,  %%xmm15			\n\t"	// aa * bb 
+
+	"	vmovups		%%xmm12 ,    (%6)				\n\t"	// write a
+	"	vmovups		%%xmm13 ,  16(%6)				\n\t"	// write a
+	"	vmovups		%%xmm14 ,  32(%6)				\n\t"	// write a
+	"	vmovups		%%xmm15 ,  48(%6)				\n\t"	// write a
+
+	"	vmovups		%%xmm12 ,    (%5)				\n\t"	// write c1
+	"	vmovups		%%xmm13 ,  16(%5)				\n\t"	
+	"	vmovups		%%xmm14 ,  32(%5)				\n\t"	
+	"	vmovups		%%xmm15 ,  48(%5)				\n\t"	
+
+	"	vfnmaddpd	%%xmm8  ,  %%xmm12 , %%xmm1 , %%xmm8		\n\t"  // c = c - aa * b 
+	"	vfnmaddpd	%%xmm9  ,  %%xmm13 , %%xmm1 , %%xmm9		\n\t"   
+	"	vfnmaddpd	%%xmm10 ,  %%xmm14 , %%xmm1 , %%xmm10		\n\t"   
+	"	vfnmaddpd	%%xmm11 ,  %%xmm15 , %%xmm1 , %%xmm11		\n\t"   
+
+	"									\n\t" // i = 0
+	"	subq		$16 , %7					\n\t" // b = b - 2
+	"	subq		$64 , %6					\n\t" // a = a - 8
+
+	"	vmovddup        (%7), %%xmm0					\n\t"	// read bb
+
+	"	vmulpd		%%xmm8  ,  %%xmm0 ,  %%xmm8 			\n\t"	// aa * bb 
+	"	vmulpd		%%xmm9  ,  %%xmm0 ,  %%xmm9 			\n\t"
+	"	vmulpd		%%xmm10 ,  %%xmm0 ,  %%xmm10			\n\t"
+	"	vmulpd		%%xmm11 ,  %%xmm0 ,  %%xmm11			\n\t"
+
+	"	vmovups		%%xmm8  ,    (%6)				\n\t"	// write a
+	"	vmovups		%%xmm9  ,  16(%6)				\n\t"
+	"	vmovups		%%xmm10 ,  32(%6)				\n\t"
+	"	vmovups		%%xmm11 ,  48(%6)				\n\t"
+
+	"	vmovups		%%xmm8  ,    (%4)				\n\t"	// write c0
+	"	vmovups		%%xmm9  ,  16(%4)				\n\t"
+	"	vmovups		%%xmm10 ,  32(%4)				\n\t"
+	"	vmovups		%%xmm11 ,  48(%4)				\n\t"
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+
+
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa,  bb;
+
+  int i, j, k;
+
+  a += (n - 1) * m;
+  b += (n - 1) * n;
+
+  for (i = n - 1; i >= 0; i--) {
+
+    bb = *(b + i);
+
+    for (j = 0; j < m; j ++) {
+      aa = *(c + j + i * ldc);
+      aa *= bb;
+      *a   = aa;
+      *(c + j + i * ldc) = aa;
+      a ++;
+
+      for (k = 0; k < i; k ++){
+	*(c + j + k * ldc) -= aa * *(b + k);
+      }
+
+    }
+    b -= n;
+    a -= 2 * m;
+  }
+
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  a += (n - 1) * m * 2;
+  b += (n - 1) * n * 2;
+
+  for (i = n - 1; i >= 0; i--) {
+
+    bb1 = *(b + i * 2 + 0);
+    bb2 = *(b + i * 2 + 1);
+
+    for (j = 0; j < m; j ++) {
+
+      aa1 = *(c + j * 2 + 0 + i * ldc);
+      aa2 = *(c + j * 2 + 1 + i * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 =  aa1 * bb1  + aa2 * bb2;
+      cc2 = - aa1 * bb2 + aa2 * bb1;
+#endif
+
+      *(a + 0) = cc1;
+      *(a + 1) = cc2;
+
+      *(c + j * 2 + 0 + i * ldc) = cc1;
+      *(c + j * 2 + 1 + i * ldc) = cc2;
+      a += 2;
+
+      for (k = 0; k < i; k ++){
+#ifndef CONJ
+	*(c + j * 2 + 0 + k * ldc) -= cc1 * *(b + k * 2 + 0) - cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#else
+	*(c + j * 2 + 0 + k * ldc) -=   cc1 * *(b + k * 2 + 0) + cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -=  -cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#endif
+      }
+
+    }
+    b -= n * 2;
+    a -= 4 * m;
+  }
+
+}
+
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k,  FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  BLASLONG i, j;
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+
+#if 0
+  fprintf(stderr, "TRSM RT KERNEL m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  kk = n - offset;
+  c += n * ldc * COMPSIZE;
+  b += n * k   * COMPSIZE;
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = 1;
+    while (j < GEMM_UNROLL_N) {
+      if (n & j) {
+
+	aa  = a;
+	b -= j * k  * COMPSIZE;
+	c -= j * ldc* COMPSIZE;
+	cc  = c;
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+	if (i > 0) {
+
+	  do {
+	    if (k - kk > 0) {
+	      GEMM_KERNEL(GEMM_UNROLL_M, j, k - kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa + GEMM_UNROLL_M * kk * COMPSIZE,
+			  b  +  j            * kk * COMPSIZE,
+			  cc,
+			  ldc);
+	    }
+
+	    solve(GEMM_UNROLL_M, j,
+		  aa + (kk - j) * GEMM_UNROLL_M * COMPSIZE,
+		  b  + (kk - j) * j             * COMPSIZE,
+		  cc, ldc);
+
+	    aa += GEMM_UNROLL_M * k * COMPSIZE;
+	    cc += GEMM_UNROLL_M     * COMPSIZE;
+	    i --;
+	  } while (i > 0);
+	}
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  i = (GEMM_UNROLL_M >> 1);
+	  do {
+	    if (m & i) {
+
+	      if (k - kk > 0) {
+		GEMM_KERNEL(i, j, k - kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa + i * kk * COMPSIZE,
+			    b  + j * kk * COMPSIZE,
+			    cc, ldc);
+	      }
+
+	      solve(i, j,
+		    aa + (kk - j) * i * COMPSIZE,
+		    b  + (kk - j) * j * COMPSIZE,
+		    cc, ldc);
+
+	      aa += i * k * COMPSIZE;
+	      cc += i     * COMPSIZE;
+
+	    }
+	    i >>= 1;
+	  } while (i > 0);
+	}
+	kk -= j;
+      }
+      j <<= 1;
+    }
+  }
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  if (j > 0) {
+
+    do {
+      aa  = a;
+      b -= GEMM_UNROLL_N * k   * COMPSIZE;
+      c -= GEMM_UNROLL_N * ldc * COMPSIZE;
+      cc  = c;
+
+      i = (m >> GEMM_UNROLL_M_SHIFT);
+      if (i > 0) {
+	do {
+
+	  dtrsm_RT_solve_opt(k - kk, aa + GEMM_UNROLL_M * kk * COMPSIZE, b  + GEMM_UNROLL_N * kk * COMPSIZE, cc, ldc,
+                            aa + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_M * COMPSIZE , b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE );
+
+	  aa += GEMM_UNROLL_M * k * COMPSIZE;
+	  cc += GEMM_UNROLL_M     * COMPSIZE;
+	  i --;
+	} while (i > 0);
+      }
+
+      if (m & (GEMM_UNROLL_M - 1)) {
+	i = (GEMM_UNROLL_M >> 1);
+	do {
+	  if (m & i) {
+	    if (k - kk > 0) {
+	      GEMM_KERNEL(i, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa + i             * kk * COMPSIZE,
+			  b  + GEMM_UNROLL_N * kk * COMPSIZE,
+			  cc,
+			  ldc);
+	    }
+
+	    solve(i, GEMM_UNROLL_N,
+		  aa + (kk - GEMM_UNROLL_N) * i             * COMPSIZE,
+		  b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE,
+		  cc, ldc);
+
+	    aa += i * k * COMPSIZE;
+	    cc += i     * COMPSIZE;
+	  }
+	  i >>= 1;
+	} while (i > 0);
+      }
+
+      kk -= GEMM_UNROLL_N;
+      j --;
+    } while (j > 0);
+  }
+
+  return 0;
+}
+
+

kernel/x86_64/strsm_kernel_LN_bulldozer.c

@@ -0,0 +1,756 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* 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.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``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 UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN 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.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_L
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+#ifndef COMPLEX
+
+
+
+static void strsm_LN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void strsm_LN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc ;
+	BLASLONG n1 = n * 8;
+	BLASLONG i=0;
+
+  	as += (16 - 1) * 16;
+  	bs += (16 - 1) * 2;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorps	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorps	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorps	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorps	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorps	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorps	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorps	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorps	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       2f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	vbroadcastss	(%3,%1,1), %%xmm0				\n\t"	// read b
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"
+	"	vbroadcastss   4(%3,%1,1), %%xmm1				\n\t"	
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"
+	"	vmovups       32(%2,%1,8), %%xmm6				\n\t"
+	"	vmovups       48(%2,%1,8), %%xmm7				\n\t"
+
+	"	vfmaddps	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"
+	"	vfmaddps	%%xmm12, %%xmm1 , %%xmm4 , %%xmm12		\n\t"
+	"	vfmaddps	%%xmm9 , %%xmm0 , %%xmm5 , %%xmm9		\n\t"
+	"	vfmaddps	%%xmm13, %%xmm1 , %%xmm5 , %%xmm13		\n\t"
+	"	vfmaddps	%%xmm10, %%xmm0 , %%xmm6 , %%xmm10		\n\t"
+	"	vfmaddps	%%xmm14, %%xmm1 , %%xmm6 , %%xmm14		\n\t"
+	"	addq		$8, %1						\n\t"
+	"	vfmaddps	%%xmm11, %%xmm0 , %%xmm7 , %%xmm11		\n\t"
+	"	vfmaddps	%%xmm15, %%xmm1 , %%xmm7 , %%xmm15		\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+	"2:									\n\t"
+
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+	"	vmovups		32(%4) , %%xmm2					\n\t"
+	"	vmovups		48(%4) , %%xmm3					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+	"	vmovups		32(%5) , %%xmm6					\n\t"
+	"	vmovups		48(%5) , %%xmm7					\n\t"
+
+	"	vsubps		%%xmm8 , %%xmm0 , %%xmm8			\n\t"
+	"	vsubps		%%xmm9 , %%xmm1 , %%xmm9			\n\t"
+	"	vsubps		%%xmm10, %%xmm2 , %%xmm10			\n\t"
+	"	vsubps		%%xmm11, %%xmm3 , %%xmm11			\n\t"
+
+	"	vsubps		%%xmm12, %%xmm4 , %%xmm12			\n\t"
+	"	vsubps		%%xmm13, %%xmm5 , %%xmm13			\n\t"
+	"	vsubps		%%xmm14, %%xmm6 , %%xmm14			\n\t"
+	"	vsubps		%%xmm15, %%xmm7 , %%xmm15			\n\t"
+
+	"3:									\n\t"	
+
+	"	vbroadcastss    60(%6) , %%xmm0					\n\t" // i=15, read aa[i]		
+	"	vshufps		$0xff  , %%xmm11 , %%xmm11 , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0xff  , %%xmm15 , %%xmm15 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 60(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 60(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	subq		$64 , %6					\n\t"   // a -= m
+	"	subq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    56(%6) , %%xmm0					\n\t" // i=14, read aa[i]		
+	"	vshufps		$0xaa  , %%xmm11 , %%xmm11 , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0xaa  , %%xmm15 , %%xmm15 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 56(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 56(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	subq		$64 , %6					\n\t"   // a -= m
+	"	subq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    52(%6) , %%xmm0					\n\t" // i=13, read aa[i]		
+	"	vshufps		$0x55  , %%xmm11 , %%xmm11 , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0x55  , %%xmm15 , %%xmm15 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 52(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 52(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	subq		$64 , %6					\n\t"   // a -= m
+	"	subq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    48(%6) , %%xmm0					\n\t" // i=12, read aa[i]		
+	"	vshufps		$0x00  , %%xmm11 , %%xmm11 , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0x00  , %%xmm15 , %%xmm15 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 48(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 48(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+
+	"	subq		$64 , %6					\n\t"   // a -= m
+	"	subq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    44(%6) , %%xmm0					\n\t" // i=11, read aa[i]		
+	"	vshufps		$0xff  , %%xmm10 , %%xmm10 , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0xff  , %%xmm14 , %%xmm14 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 44(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 44(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+
+	"	subq		$64 , %6					\n\t"   // a -= m
+	"	subq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    40(%6) , %%xmm0					\n\t" // i=10, read aa[i]		
+	"	vshufps		$0xaa  , %%xmm10 , %%xmm10 , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0xaa  , %%xmm14 , %%xmm14 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 40(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 40(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+
+	"	subq		$64 , %6					\n\t"   // a -= m
+	"	subq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    36(%6) , %%xmm0					\n\t" // i=9 , read aa[i]		
+	"	vshufps		$0x55  , %%xmm10 , %%xmm10 , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0x55  , %%xmm14 , %%xmm14 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 36(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 36(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+
+	"	subq		$64 , %6					\n\t"   // a -= m
+	"	subq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    32(%6) , %%xmm0					\n\t" // i=8 , read aa[i]		
+	"	vshufps		$0x00  , %%xmm10 , %%xmm10 , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0x00  , %%xmm14 , %%xmm14 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 32(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 32(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+
+	"	subq		$64 , %6					\n\t"   // a -= m
+	"	subq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    28(%6) , %%xmm0					\n\t" // i=7 , read aa[i]		
+	"	vshufps		$0xff  , %%xmm9  , %%xmm9  , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0xff  , %%xmm13 , %%xmm13 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 28(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 28(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+
+	"	subq		$64 , %6					\n\t"   // a -= m
+	"	subq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    24(%6) , %%xmm0					\n\t" // i=6 , read aa[i]		
+	"	vshufps		$0xaa  , %%xmm9  , %%xmm9  , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0xaa  , %%xmm13 , %%xmm13 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 24(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 24(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+
+	"	subq		$64 , %6					\n\t"   // a -= m
+	"	subq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    20(%6) , %%xmm0					\n\t" // i=5 , read aa[i]		
+	"	vshufps		$0x55  , %%xmm9  , %%xmm9  , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0x55  , %%xmm13 , %%xmm13 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 20(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 20(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+
+	"	subq		$64 , %6					\n\t"   // a -= m
+	"	subq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    16(%6) , %%xmm0					\n\t" // i=4 , read aa[i]		
+	"	vshufps		$0x00  , %%xmm9  , %%xmm9  , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0x00  , %%xmm13 , %%xmm13 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 16(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 16(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+
+	"	subq		$64 , %6					\n\t"   // a -= m
+	"	subq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    12(%6) , %%xmm0					\n\t" // i=3 , read aa[i]		
+	"	vshufps		$0xff  , %%xmm8  , %%xmm8  , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0xff  , %%xmm12 , %%xmm12 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 12(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 12(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+
+	"	subq		$64 , %6					\n\t"   // a -= m
+	"	subq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss     8(%6) , %%xmm0					\n\t" // i=2 , read aa[i]		
+	"	vshufps		$0xaa  , %%xmm8  , %%xmm8  , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0xaa  , %%xmm12 , %%xmm12 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  ,  8(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  ,  8(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+
+	"	subq		$64 , %6					\n\t"   // a -= m
+	"	subq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss     4(%6) , %%xmm0					\n\t" // i=1 , read aa[i]		
+	"	vshufps		$0x55  , %%xmm8  , %%xmm8  , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0x55  , %%xmm12 , %%xmm12 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  ,  4(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+
+	"	subq		$64 , %6					\n\t"   // a -= m
+	"	subq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss     0(%6) , %%xmm0					\n\t" // i=0 , read aa[i]		
+	"	vshufps		$0x00  , %%xmm8  , %%xmm8  , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0x00  , %%xmm12 , %%xmm12 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  ,  0(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  ,  0(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+
+
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa,  bb;
+
+  int i, j, k;
+
+  a += (m - 1) * m;
+  b += (m - 1) * n;
+
+  for (i = m - 1; i >= 0; i--) {
+
+    aa = *(a + i);
+
+    for (j = 0; j < n; j ++) {
+      bb = *(c + i + j * ldc);
+      bb *= aa;
+      *b             = bb;
+      *(c + i + j * ldc) = bb;
+      b ++;
+
+      for (k = 0; k < i; k ++){
+	*(c + k + j * ldc) -= bb * *(a + k);
+      }
+
+    }
+    a -= m;
+    b -= 2 * n;
+  }
+
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+  a += (m - 1) * m * 2;
+  b += (m - 1) * n * 2;
+
+  for (i = m - 1; i >= 0; i--) {
+
+    aa1 = *(a + i * 2 + 0);
+    aa2 = *(a + i * 2 + 1);
+
+    for (j = 0; j < n; j ++) {
+      bb1 = *(c + i * 2 + 0 + j * ldc);
+      bb2 = *(c + i * 2 + 1 + j * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 = aa1 * bb1 + aa2 * bb2;
+      cc2 = aa1 * bb2 - aa2 * bb1;
+#endif
+
+
+      *(b + 0) = cc1;
+      *(b + 1) = cc2;
+      *(c + i * 2 + 0 + j * ldc) = cc1;
+      *(c + i * 2 + 1 + j * ldc) = cc2;
+      b += 2;
+
+      for (k = 0; k < i; k ++){
+#ifndef CONJ
+	*(c + k * 2 + 0 + j * ldc) -= cc1 * *(a + k * 2 + 0) - cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#else
+	*(c + k * 2 + 0 + j * ldc) -=   cc1 * *(a + k * 2 + 0) + cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= - cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#endif
+      }
+
+    }
+    a -= m * 2;
+    b -= 4 * n;
+  }
+
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k,  FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  BLASLONG i, j;
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+
+#if 0
+  fprintf(stderr, "TRSM KERNEL LN : m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  while (j > 0) {
+
+    kk = m + offset;
+
+    if (m & (GEMM_UNROLL_M - 1)) {
+      for (i = 1; i < GEMM_UNROLL_M; i *= 2){
+	if (m & i) {
+	  aa = a + ((m & ~(i - 1)) - i) * k * COMPSIZE;
+	  cc = c + ((m & ~(i - 1)) - i)     * COMPSIZE;
+
+	  if (k - kk > 0) {
+	    GEMM_KERNEL(i, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa + i             * kk * COMPSIZE,
+			b  + GEMM_UNROLL_N * kk * COMPSIZE,
+			cc,
+			ldc);
+	  }
+
+	  solve(i, GEMM_UNROLL_N,
+		aa + (kk - i) * i             * COMPSIZE,
+		b  + (kk - i) * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+	  kk -= i;
+	}
+      }
+    }
+
+    i = (m >> GEMM_UNROLL_M_SHIFT);
+    if (i > 0) {
+      aa = a + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M) * k * COMPSIZE;
+      cc = c + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M)     * COMPSIZE;
+
+      do {
+	strsm_LN_solve_opt(k-kk, aa + GEMM_UNROLL_M * kk * COMPSIZE, b +  GEMM_UNROLL_N * kk * COMPSIZE, cc, ldc,
+                           aa + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_M * COMPSIZE,b  + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_N * COMPSIZE);
+
+	aa -= GEMM_UNROLL_M * k * COMPSIZE;
+	cc -= GEMM_UNROLL_M     * COMPSIZE;
+	kk -= GEMM_UNROLL_M;
+	i --;
+      } while (i > 0);
+    }
+
+    b += GEMM_UNROLL_N * k * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+	kk = m + offset;
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  for (i = 1; i < GEMM_UNROLL_M; i *= 2){
+	    if (m & i) {
+	      aa = a + ((m & ~(i - 1)) - i) * k * COMPSIZE;
+	      cc = c + ((m & ~(i - 1)) - i)     * COMPSIZE;
+
+	      if (k - kk > 0) {
+		GEMM_KERNEL(i, j, k - kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa + i * kk * COMPSIZE,
+			    b  + j * kk * COMPSIZE,
+			    cc, ldc);
+	      }
+
+	      solve(i, j,
+		    aa + (kk - i) * i * COMPSIZE,
+		    b  + (kk - i) * j * COMPSIZE,
+		    cc, ldc);
+
+	      kk -= i;
+	    }
+	  }
+	}
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+	if (i > 0) {
+	  aa = a + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M) * k * COMPSIZE;
+	  cc = c + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M)     * COMPSIZE;
+
+	  do {
+	    if (k - kk > 0) {
+	      GEMM_KERNEL(GEMM_UNROLL_M, j, k - kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa + GEMM_UNROLL_M * kk * COMPSIZE,
+			  b +  j             * kk * COMPSIZE,
+			  cc,
+			  ldc);
+	    }
+
+	    solve(GEMM_UNROLL_M, j,
+		  aa + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_M * COMPSIZE,
+		  b  + (kk - GEMM_UNROLL_M) * j             * COMPSIZE,
+		  cc, ldc);
+
+	    aa -= GEMM_UNROLL_M * k * COMPSIZE;
+	    cc -= GEMM_UNROLL_M     * COMPSIZE;
+	    kk -= GEMM_UNROLL_M;
+	    i --;
+	  } while (i > 0);
+	}
+
+	b += j * k   * COMPSIZE;
+	c += j * ldc * COMPSIZE;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/x86_64/strsm_kernel_LT_bulldozer.c

@@ -0,0 +1,739 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* 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.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``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 UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN 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.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_L
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+
+static void strsm_LT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void strsm_LT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc ;
+	BLASLONG n1 = n * 8;
+	BLASLONG i=0;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorps	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorps	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorps	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorps	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorps	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorps	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorps	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorps	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       2f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	vbroadcastss	(%3,%1,1), %%xmm0				\n\t"	// read b
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"
+	"	vbroadcastss   4(%3,%1,1), %%xmm1				\n\t"	
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"
+	"	vmovups       32(%2,%1,8), %%xmm6				\n\t"
+	"	vmovups       48(%2,%1,8), %%xmm7				\n\t"
+
+	"	vfmaddps	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"
+	"	vfmaddps	%%xmm12, %%xmm1 , %%xmm4 , %%xmm12		\n\t"
+	"	vfmaddps	%%xmm9 , %%xmm0 , %%xmm5 , %%xmm9		\n\t"
+	"	vfmaddps	%%xmm13, %%xmm1 , %%xmm5 , %%xmm13		\n\t"
+	"	vfmaddps	%%xmm10, %%xmm0 , %%xmm6 , %%xmm10		\n\t"
+	"	vfmaddps	%%xmm14, %%xmm1 , %%xmm6 , %%xmm14		\n\t"
+	"	addq		$8, %1						\n\t"
+	"	vfmaddps	%%xmm11, %%xmm0 , %%xmm7 , %%xmm11		\n\t"
+	"	vfmaddps	%%xmm15, %%xmm1 , %%xmm7 , %%xmm15		\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+	"2:									\n\t"
+
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+	"	vmovups		32(%4) , %%xmm2					\n\t"
+	"	vmovups		48(%4) , %%xmm3					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+	"	vmovups		32(%5) , %%xmm6					\n\t"
+	"	vmovups		48(%5) , %%xmm7					\n\t"
+
+	"	vsubps		%%xmm8 , %%xmm0 , %%xmm8			\n\t"
+	"	vsubps		%%xmm9 , %%xmm1 , %%xmm9			\n\t"
+	"	vsubps		%%xmm10, %%xmm2 , %%xmm10			\n\t"
+	"	vsubps		%%xmm11, %%xmm3 , %%xmm11			\n\t"
+
+	"	vsubps		%%xmm12, %%xmm4 , %%xmm12			\n\t"
+	"	vsubps		%%xmm13, %%xmm5 , %%xmm13			\n\t"
+	"	vsubps		%%xmm14, %%xmm6 , %%xmm14			\n\t"
+	"	vsubps		%%xmm15, %%xmm7 , %%xmm15			\n\t"
+
+	"3:									\n\t"	
+
+	"	vbroadcastss     0(%6) , %%xmm0					\n\t" // i=0, read aa[i]		
+	"	vshufps		$0x00  , %%xmm8  , %%xmm8  , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0x00  , %%xmm12 , %%xmm12 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  ,  0(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  ,  0(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	addq		$64 , %6					\n\t"   // a -= m
+	"	addq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss     4(%6) , %%xmm0					\n\t" // i=1, read aa[i]		
+	"	vshufps		$0x55  , %%xmm8  , %%xmm8  , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0x55  , %%xmm12 , %%xmm12 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  ,  4(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	addq		$64 , %6					\n\t"   // a -= m
+	"	addq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss     8(%6) , %%xmm0					\n\t" // i=2, read aa[i]		
+	"	vshufps		$0xaa  , %%xmm8  , %%xmm8  , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0xaa  , %%xmm12 , %%xmm12 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  ,  8(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  ,  8(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups         0(%6)  , %%xmm4                                 \n\t"   // read a[k]
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+	"       vfnmaddps       %%xmm8  , %%xmm1 , %%xmm4 , %%xmm8              \n\t"
+        "       vfnmaddps       %%xmm12 , %%xmm2 , %%xmm4 , %%xmm12             \n\t"
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	addq		$64 , %6					\n\t"   // a -= m
+	"	addq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    12(%6) , %%xmm0					\n\t" // i=3, read aa[i]		
+	"	vshufps		$0xff  , %%xmm8  , %%xmm8  , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0xff  , %%xmm12 , %%xmm12 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 12(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 12(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	addq		$64 , %6					\n\t"   // a -= m
+	"	addq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    16(%6) , %%xmm0					\n\t" // i=4, read aa[i]		
+	"	vshufps		$0x00  , %%xmm9  , %%xmm9  , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0x00  , %%xmm13 , %%xmm13 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 16(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 16(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	addq		$64 , %6					\n\t"   // a -= m
+	"	addq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    20(%6) , %%xmm0					\n\t" // i=5, read aa[i]		
+	"	vshufps		$0x55  , %%xmm9  , %%xmm9  , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0x55  , %%xmm13 , %%xmm13 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 20(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 20(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	addq		$64 , %6					\n\t"   // a -= m
+	"	addq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    24(%6) , %%xmm0					\n\t" // i=6, read aa[i]		
+	"	vshufps		$0xaa  , %%xmm9  , %%xmm9  , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0xaa  , %%xmm13 , %%xmm13 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 24(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 24(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups        16(%6)  , %%xmm5                                 \n\t"   // read a[k]
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+        "       vfnmaddps       %%xmm9  , %%xmm1 , %%xmm5 , %%xmm9              \n\t"
+        "       vfnmaddps       %%xmm13 , %%xmm2 , %%xmm5 , %%xmm13             \n\t"
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	addq		$64 , %6					\n\t"   // a -= m
+	"	addq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    28(%6) , %%xmm0					\n\t" // i=7, read aa[i]		
+	"	vshufps		$0xff  , %%xmm9  , %%xmm9  , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0xff  , %%xmm13 , %%xmm13 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 28(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 28(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	addq		$64 , %6					\n\t"   // a -= m
+	"	addq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    32(%6) , %%xmm0					\n\t" // i=8, read aa[i]		
+	"	vshufps		$0x00  , %%xmm10 , %%xmm10 , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0x00  , %%xmm14 , %%xmm14 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 32(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 32(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	addq		$64 , %6					\n\t"   // a -= m
+	"	addq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    36(%6) , %%xmm0					\n\t" // i=9, read aa[i]		
+	"	vshufps		$0x55  , %%xmm10 , %%xmm10 , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0x55  , %%xmm14 , %%xmm14 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 36(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 36(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	addq		$64 , %6					\n\t"   // a -= m
+	"	addq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    40(%6) , %%xmm0					\n\t" // i=10, read aa[i]		
+	"	vshufps		$0xaa  , %%xmm10 , %%xmm10 , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0xaa  , %%xmm14 , %%xmm14 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 40(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 40(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups        32(%6)  , %%xmm6                                 \n\t"   // read a[k]
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+        "       vfnmaddps       %%xmm10 , %%xmm1 , %%xmm6 , %%xmm10             \n\t"
+        "       vfnmaddps       %%xmm14 , %%xmm2 , %%xmm6 , %%xmm14             \n\t"
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	addq		$64 , %6					\n\t"   // a -= m
+	"	addq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    44(%6) , %%xmm0					\n\t" // i=11, read aa[i]		
+	"	vshufps		$0xff  , %%xmm10 , %%xmm10 , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0xff  , %%xmm14 , %%xmm14 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 44(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 44(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	addq		$64 , %6					\n\t"   // a -= m
+	"	addq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    48(%6) , %%xmm0					\n\t" // i=12, read aa[i]		
+	"	vshufps		$0x00  , %%xmm11 , %%xmm11 , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0x00  , %%xmm15 , %%xmm15 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 48(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 48(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	addq		$64 , %6					\n\t"   // a -= m
+	"	addq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    52(%6) , %%xmm0					\n\t" // i=13, read aa[i]		
+	"	vshufps		$0x55  , %%xmm11 , %%xmm11 , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0x55  , %%xmm15 , %%xmm15 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 52(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 52(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	addq		$64 , %6					\n\t"   // a -= m
+	"	addq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    56(%6) , %%xmm0					\n\t" // i=14, read aa[i]		
+	"	vshufps		$0xaa  , %%xmm11 , %%xmm11 , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0xaa  , %%xmm15 , %%xmm15 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 56(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 56(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"       vmovups        48(%6)  , %%xmm7                                 \n\t"   // read a[k]
+        "       vfnmaddps       %%xmm11 , %%xmm1 , %%xmm7 , %%xmm11             \n\t"
+        "       vfnmaddps       %%xmm15 , %%xmm2 , %%xmm7 , %%xmm15             \n\t"
+
+	"	addq		$64 , %6					\n\t"   // a -= m
+	"	addq		$8  , %7					\n\t"   // b -= n
+
+	"	vbroadcastss    60(%6) , %%xmm0					\n\t" // i=15, read aa[i]		
+	"	vshufps		$0xff  , %%xmm11 , %%xmm11 , %%xmm1		\n\t" // extract bb0
+	"	vshufps		$0xff  , %%xmm15 , %%xmm15 , %%xmm2		\n\t" // extract bb1
+	"       vmulps          %%xmm0  , %%xmm1 , %%xmm1                       \n\t"   // bb0 * aa
+	"       vmulps          %%xmm0  , %%xmm2 , %%xmm2                       \n\t"   // bb1 * aa
+        "       vmovss          %%xmm1  , 60(%4)                                \n\t"   // c[i] = bb0 * aa
+        "       vmovss          %%xmm2  , 60(%5)                                \n\t"   // c[i] = bb1 * aa
+        "       vmovss          %%xmm1  ,   (%7)                        	\n\t"   // b[0] = bb0 * aa
+        "       vmovss          %%xmm2  ,  4(%7)                        	\n\t"   // b[1] = bb1 * aa
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa, bb;
+
+  int i, j, k;
+
+  for (i = 0; i < m; i++) {
+
+    aa = *(a + i);
+
+    for (j = 0; j < n; j ++) {
+      bb = *(c + i + j * ldc);
+      bb *= aa;
+      *b             = bb;
+      *(c + i + j * ldc) = bb;
+      b ++;
+
+      for (k = i + 1; k < m; k ++){
+	*(c + k + j * ldc) -= bb * *(a + k);
+      }
+
+    }
+    a += m;
+  }
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  for (i = 0; i < m; i++) {
+
+    aa1 = *(a + i * 2 + 0);
+    aa2 = *(a + i * 2 + 1);
+
+    for (j = 0; j < n; j ++) {
+      bb1 = *(c + i * 2 + 0 + j * ldc);
+      bb2 = *(c + i * 2 + 1 + j * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 = aa1 * bb1 + aa2 * bb2;
+      cc2 = aa1 * bb2 - aa2 * bb1;
+#endif
+
+      *(b + 0) = cc1;
+      *(b + 1) = cc2;
+      *(c + i * 2 + 0 + j * ldc) = cc1;
+      *(c + i * 2 + 1 + j * ldc) = cc2;
+      b += 2;
+
+      for (k = i + 1; k < m; k ++){
+#ifndef CONJ
+	*(c + k * 2 + 0 + j * ldc) -= cc1 * *(a + k * 2 + 0) - cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#else
+	*(c + k * 2 + 0 + j * ldc) -= cc1 * *(a + k * 2 + 0) + cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= -cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#endif
+      }
+
+    }
+    a += m * 2;
+  }
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+  BLASLONG i, j, jj;
+
+#if 0
+  fprintf(stderr, "TRSM KERNEL LT : m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  jj = 0;
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  while (j > 0) {
+
+    kk = offset;
+    aa = a;
+    cc = c;
+
+    i = (m >> GEMM_UNROLL_M_SHIFT);
+
+    while (i > 0) {
+
+	strsm_LT_solve_opt(kk , aa , b , cc, ldc, aa + kk * GEMM_UNROLL_M * COMPSIZE, b  + kk * GEMM_UNROLL_N * COMPSIZE);
+
+
+      aa += GEMM_UNROLL_M * k * COMPSIZE;
+      cc += GEMM_UNROLL_M     * COMPSIZE;
+      kk += GEMM_UNROLL_M;
+      i --;
+    }
+
+    if (m & (GEMM_UNROLL_M - 1)) {
+      i = (GEMM_UNROLL_M >> 1);
+      while (i > 0) {
+	if (m & i) {
+	    if (kk > 0) {
+	      GEMM_KERNEL(i, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa, b, cc, ldc);
+	    }
+	  solve(i, GEMM_UNROLL_N,
+		aa + kk * i             * COMPSIZE,
+		b  + kk * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+	  aa += i * k * COMPSIZE;
+	  cc += i     * COMPSIZE;
+	  kk += i;
+	}
+	i >>= 1;
+      }
+    }
+
+    b += GEMM_UNROLL_N * k   * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+    jj += GEMM_UNROLL_M;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+	kk = offset;
+	aa = a;
+	cc = c;
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+
+	while (i > 0) {
+	  if (kk > 0) {
+	    GEMM_KERNEL(GEMM_UNROLL_M, j, kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa,
+			b,
+			cc,
+			ldc);
+	  }
+
+	  solve(GEMM_UNROLL_M, j,
+		aa + kk * GEMM_UNROLL_M * COMPSIZE,
+		b  + kk * j             * COMPSIZE, cc, ldc);
+
+	  aa += GEMM_UNROLL_M * k * COMPSIZE;
+	  cc += GEMM_UNROLL_M     * COMPSIZE;
+	  kk += GEMM_UNROLL_M;
+	  i --;
+	}
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  i = (GEMM_UNROLL_M >> 1);
+	  while (i > 0) {
+	    if (m & i) {
+	      if (kk > 0) {
+		GEMM_KERNEL(i, j, kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa,
+			    b,
+			    cc,
+			    ldc);
+	      }
+
+	      solve(i, j,
+		    aa + kk * i * COMPSIZE,
+		    b  + kk * j * COMPSIZE, cc, ldc);
+
+	      aa += i * k * COMPSIZE;
+	      cc += i     * COMPSIZE;
+	      kk += i;
+	      }
+	    i >>= 1;
+	  }
+	}
+
+	b += j * k   * COMPSIZE;
+	c += j * ldc * COMPSIZE;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/x86_64/strsm_kernel_RN_bulldozer.c

@@ -0,0 +1,454 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* 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.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``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 UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN 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.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_R
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+
+static void strsm_RN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void strsm_RN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc ;
+	BLASLONG n1 = n * 8;
+	BLASLONG i=0;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorps	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorps	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorps	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorps	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorps	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorps	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorps	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorps	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       2f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	vbroadcastss	(%3,%1,1), %%xmm0				\n\t"	// read b
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"
+	"	vbroadcastss   4(%3,%1,1), %%xmm1				\n\t"	
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"
+	"	vmovups       32(%2,%1,8), %%xmm6				\n\t"
+	"	vmovups       48(%2,%1,8), %%xmm7				\n\t"
+
+	"	vfmaddps	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"
+	"	vfmaddps	%%xmm12, %%xmm1 , %%xmm4 , %%xmm12		\n\t"
+	"	vfmaddps	%%xmm9 , %%xmm0 , %%xmm5 , %%xmm9		\n\t"
+	"	vfmaddps	%%xmm13, %%xmm1 , %%xmm5 , %%xmm13		\n\t"
+	"	vfmaddps	%%xmm10, %%xmm0 , %%xmm6 , %%xmm10		\n\t"
+	"	vfmaddps	%%xmm14, %%xmm1 , %%xmm6 , %%xmm14		\n\t"
+	"	addq		$8, %1						\n\t"
+	"	vfmaddps	%%xmm11, %%xmm0 , %%xmm7 , %%xmm11		\n\t"
+	"	vfmaddps	%%xmm15, %%xmm1 , %%xmm7 , %%xmm15		\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+	"2:									\n\t"
+
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+	"	vmovups		32(%4) , %%xmm2					\n\t"
+	"	vmovups		48(%4) , %%xmm3					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+	"	vmovups		32(%5) , %%xmm6					\n\t"
+	"	vmovups		48(%5) , %%xmm7					\n\t"
+
+	"	vsubps		%%xmm8 , %%xmm0 , %%xmm8			\n\t"
+	"	vsubps		%%xmm9 , %%xmm1 , %%xmm9			\n\t"
+	"	vsubps		%%xmm10, %%xmm2 , %%xmm10			\n\t"
+	"	vsubps		%%xmm11, %%xmm3 , %%xmm11			\n\t"
+
+	"	vsubps		%%xmm12, %%xmm4 , %%xmm12			\n\t"
+	"	vsubps		%%xmm13, %%xmm5 , %%xmm13			\n\t"
+	"	vsubps		%%xmm14, %%xmm6 , %%xmm14			\n\t"
+	"	vsubps		%%xmm15, %%xmm7 , %%xmm15			\n\t"
+
+	"3:									\n\t"	// i = 0
+
+	"	vbroadcastss	(%7), %%xmm0					\n\t"	// read bb
+	"	vbroadcastss   4(%7), %%xmm1					\n\t"	// read b
+
+	"	vmulps		%%xmm8  ,  %%xmm0 ,  %%xmm8 			\n\t"	// aa * bb 
+	"	vmulps		%%xmm9  ,  %%xmm0 ,  %%xmm9 			\n\t"
+	"	vmulps		%%xmm10 ,  %%xmm0 ,  %%xmm10			\n\t"
+	"	vmulps		%%xmm11 ,  %%xmm0 ,  %%xmm11			\n\t"
+
+	"	vmovups		%%xmm8  ,    (%6)				\n\t"	// write a
+	"	vmovups		%%xmm9  ,  16(%6)				\n\t"
+	"	vmovups		%%xmm10 ,  32(%6)				\n\t"
+	"	vmovups		%%xmm11 ,  48(%6)				\n\t"
+
+	"	vmovups		%%xmm8  ,    (%4)				\n\t"	// write c0
+	"	vmovups		%%xmm9  ,  16(%4)				\n\t"
+	"	vmovups		%%xmm10 ,  32(%4)				\n\t"
+	"	vmovups		%%xmm11 ,  48(%4)				\n\t"
+
+	"	vfnmaddps	%%xmm12 ,  %%xmm8  , %%xmm1 , %%xmm12		\n\t"  // c = c - aa * b 
+	"	vfnmaddps	%%xmm13 ,  %%xmm9  , %%xmm1 , %%xmm13		\n\t"   
+	"	vfnmaddps	%%xmm14 ,  %%xmm10 , %%xmm1 , %%xmm14		\n\t"   
+	"	vfnmaddps	%%xmm15 ,  %%xmm11 , %%xmm1 , %%xmm15		\n\t"   
+
+	"									\n\t" // i = 1
+	"	addq		$8  , %7					\n\t" // b = b + 2
+	"	addq	       $64  , %6					\n\t" // a = a + 16
+
+	"	vbroadcastss      4(%7), %%xmm0					\n\t"	// read bb
+
+	"	vmulps		%%xmm12 ,  %%xmm0 ,  %%xmm12			\n\t"	// aa * bb 
+	"	vmulps		%%xmm13 ,  %%xmm0 ,  %%xmm13			\n\t"	// aa * bb 
+	"	vmulps		%%xmm14 ,  %%xmm0 ,  %%xmm14			\n\t"	// aa * bb 
+	"	vmulps		%%xmm15 ,  %%xmm0 ,  %%xmm15			\n\t"	// aa * bb 
+
+	"	vmovups		%%xmm12 ,    (%6)				\n\t"	// write a
+	"	vmovups		%%xmm13 ,  16(%6)				\n\t"	// write a
+	"	vmovups		%%xmm14 ,  32(%6)				\n\t"	// write a
+	"	vmovups		%%xmm15 ,  48(%6)				\n\t"	// write a
+
+	"	vmovups		%%xmm12 ,    (%5)				\n\t"	// write c1
+	"	vmovups		%%xmm13 ,  16(%5)				\n\t"	
+	"	vmovups		%%xmm14 ,  32(%5)				\n\t"	
+	"	vmovups		%%xmm15 ,  48(%5)				\n\t"	
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa, bb;
+
+  int i, j, k;
+
+  for (i = 0; i < n; i++) {
+
+    bb = *(b + i);
+
+    for (j = 0; j < m; j ++) {
+      aa = *(c + j + i * ldc);
+      aa *= bb;
+      *a  = aa;
+      *(c + j + i * ldc) = aa;
+      a ++;
+
+      for (k = i + 1; k < n; k ++){
+	*(c + j + k * ldc) -= aa * *(b + k);
+      }
+
+    }
+    b += n;
+  }
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  for (i = 0; i < n; i++) {
+
+    bb1 = *(b + i * 2 + 0);
+    bb2 = *(b + i * 2 + 1);
+
+    for (j = 0; j < m; j ++) {
+      aa1 = *(c + j * 2 + 0 + i * ldc);
+      aa2 = *(c + j * 2 + 1 + i * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 =  aa1 * bb1 + aa2 * bb2;
+      cc2 = -aa1 * bb2 + aa2 * bb1;
+#endif
+
+      *(a + 0) = cc1;
+      *(a + 1) = cc2;
+      *(c + j * 2 + 0 + i * ldc) = cc1;
+      *(c + j * 2 + 1 + i * ldc) = cc2;
+      a += 2;
+
+      for (k = i + 1; k < n; k ++){
+#ifndef CONJ
+	*(c + j * 2 + 0 + k * ldc) -= cc1 * *(b + k * 2 + 0) - cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#else
+	*(c + j * 2 + 0 + k * ldc) -=   cc1 * *(b + k * 2 + 0) + cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= - cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#endif
+      }
+
+    }
+    b += n * 2;
+  }
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+  BLASLONG i, j, jj;
+
+#if 0
+  fprintf(stderr, "TRSM RN KERNEL m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  jj = 0;
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+  kk = -offset;
+
+  while (j > 0) {
+
+    aa = a;
+    cc = c;
+
+    i = (m >> GEMM_UNROLL_M_SHIFT);
+
+    if (i > 0) {
+      do {
+
+	strsm_RN_solve_opt(kk, aa, b, cc, ldc, aa + kk * GEMM_UNROLL_M * COMPSIZE, b + kk * GEMM_UNROLL_N * COMPSIZE);
+
+
+	aa += GEMM_UNROLL_M * k * COMPSIZE;
+	cc += GEMM_UNROLL_M     * COMPSIZE;
+	i --;
+      } while (i > 0);
+    }
+
+
+    if (m & (GEMM_UNROLL_M - 1)) {
+      i = (GEMM_UNROLL_M >> 1);
+      while (i > 0) {
+	if (m & i) {
+	    if (kk > 0) {
+	      GEMM_KERNEL(i, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa, b, cc, ldc);
+	    }
+	  solve(i, GEMM_UNROLL_N,
+		aa + kk * i             * COMPSIZE,
+		b  + kk * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+	  aa += i * k * COMPSIZE;
+	  cc += i     * COMPSIZE;
+	}
+	i >>= 1;
+      }
+    }
+
+    kk += GEMM_UNROLL_N;
+    b += GEMM_UNROLL_N * k   * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+    jj += GEMM_UNROLL_M;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+	aa = a;
+	cc = c;
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+
+	while (i > 0) {
+	  if (kk > 0) {
+	    GEMM_KERNEL(GEMM_UNROLL_M, j, kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa,
+			b,
+			cc,
+			ldc);
+	  }
+
+	  solve(GEMM_UNROLL_M, j,
+		aa + kk * GEMM_UNROLL_M * COMPSIZE,
+		b  + kk * j             * COMPSIZE, cc, ldc);
+
+	  aa += GEMM_UNROLL_M * k * COMPSIZE;
+	  cc += GEMM_UNROLL_M     * COMPSIZE;
+	  i --;
+	}
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  i = (GEMM_UNROLL_M >> 1);
+	  while (i > 0) {
+	    if (m & i) {
+	      if (kk > 0) {
+		GEMM_KERNEL(i, j, kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa,
+			    b,
+			    cc,
+			    ldc);
+	      }
+
+	      solve(i, j,
+		    aa + kk * i * COMPSIZE,
+		    b  + kk * j * COMPSIZE, cc, ldc);
+
+	      aa += i * k * COMPSIZE;
+	      cc += i     * COMPSIZE;
+	      }
+	    i >>= 1;
+	  }
+	}
+
+	b += j * k   * COMPSIZE;
+	c += j * ldc * COMPSIZE;
+	kk += j;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/x86_64/strsm_kernel_RT_bulldozer.c

@@ -0,0 +1,481 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* 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.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``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 UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN 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.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_R
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+
+static void strsm_RT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void strsm_RT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc ;
+	BLASLONG n1 = n * 8;
+	BLASLONG i=0;
+
+  	as += (2 - 1) * 16;
+  	bs += (2 - 1) * 2;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorps	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorps	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorps	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorps	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorps	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorps	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorps	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorps	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       2f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	vbroadcastss	(%3,%1,1), %%xmm0				\n\t"	// read b
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"
+	"	vbroadcastss   4(%3,%1,1), %%xmm1				\n\t"	
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"
+	"	vmovups       32(%2,%1,8), %%xmm6				\n\t"
+	"	vmovups       48(%2,%1,8), %%xmm7				\n\t"
+
+	"	vfmaddps	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"
+	"	vfmaddps	%%xmm12, %%xmm1 , %%xmm4 , %%xmm12		\n\t"
+	"	vfmaddps	%%xmm9 , %%xmm0 , %%xmm5 , %%xmm9		\n\t"
+	"	vfmaddps	%%xmm13, %%xmm1 , %%xmm5 , %%xmm13		\n\t"
+	"	vfmaddps	%%xmm10, %%xmm0 , %%xmm6 , %%xmm10		\n\t"
+	"	vfmaddps	%%xmm14, %%xmm1 , %%xmm6 , %%xmm14		\n\t"
+	"	addq		$8, %1						\n\t"
+	"	vfmaddps	%%xmm11, %%xmm0 , %%xmm7 , %%xmm11		\n\t"
+	"	vfmaddps	%%xmm15, %%xmm1 , %%xmm7 , %%xmm15		\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+	"2:									\n\t"
+
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+	"	vmovups		32(%4) , %%xmm2					\n\t"
+	"	vmovups		48(%4) , %%xmm3					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+	"	vmovups		32(%5) , %%xmm6					\n\t"
+	"	vmovups		48(%5) , %%xmm7					\n\t"
+
+	"	vsubps		%%xmm8 , %%xmm0 , %%xmm8			\n\t"
+	"	vsubps		%%xmm9 , %%xmm1 , %%xmm9			\n\t"
+	"	vsubps		%%xmm10, %%xmm2 , %%xmm10			\n\t"
+	"	vsubps		%%xmm11, %%xmm3 , %%xmm11			\n\t"
+
+	"	vsubps		%%xmm12, %%xmm4 , %%xmm12			\n\t"
+	"	vsubps		%%xmm13, %%xmm5 , %%xmm13			\n\t"
+	"	vsubps		%%xmm14, %%xmm6 , %%xmm14			\n\t"
+	"	vsubps		%%xmm15, %%xmm7 , %%xmm15			\n\t"
+
+	"3:									\n\t"	// i = 1
+
+	"	vbroadcastss	(%7), %%xmm1					\n\t"	// read b
+	"	vbroadcastss   4(%7), %%xmm0					\n\t"	// read bb
+
+	"	vmulps		%%xmm12 ,  %%xmm0 ,  %%xmm12			\n\t"	// aa * bb 
+	"	vmulps		%%xmm13 ,  %%xmm0 ,  %%xmm13			\n\t"	// aa * bb 
+	"	vmulps		%%xmm14 ,  %%xmm0 ,  %%xmm14			\n\t"	// aa * bb 
+	"	vmulps		%%xmm15 ,  %%xmm0 ,  %%xmm15			\n\t"	// aa * bb 
+
+	"	vmovups		%%xmm12 ,    (%6)				\n\t"	// write a
+	"	vmovups		%%xmm13 ,  16(%6)				\n\t"	// write a
+	"	vmovups		%%xmm14 ,  32(%6)				\n\t"	// write a
+	"	vmovups		%%xmm15 ,  48(%6)				\n\t"	// write a
+
+	"	vmovups		%%xmm12 ,    (%5)				\n\t"	// write c1
+	"	vmovups		%%xmm13 ,  16(%5)				\n\t"	
+	"	vmovups		%%xmm14 ,  32(%5)				\n\t"	
+	"	vmovups		%%xmm15 ,  48(%5)				\n\t"	
+
+	"	vfnmaddps	%%xmm8  ,  %%xmm12 , %%xmm1 , %%xmm8		\n\t"  // c = c - aa * b 
+	"	vfnmaddps	%%xmm9  ,  %%xmm13 , %%xmm1 , %%xmm9		\n\t"   
+	"	vfnmaddps	%%xmm10 ,  %%xmm14 , %%xmm1 , %%xmm10		\n\t"   
+	"	vfnmaddps	%%xmm11 ,  %%xmm15 , %%xmm1 , %%xmm11		\n\t"   
+
+	"									\n\t" // i = 0
+	"	subq		$8  , %7					\n\t" // b = b - 2
+	"	subq	       $64  , %6					\n\t" // a = a - 16
+
+	"	vbroadcastss       (%7), %%xmm0					\n\t"	// read bb
+
+	"	vmulps		%%xmm8  ,  %%xmm0 ,  %%xmm8 			\n\t"	// aa * bb 
+	"	vmulps		%%xmm9  ,  %%xmm0 ,  %%xmm9 			\n\t"
+	"	vmulps		%%xmm10 ,  %%xmm0 ,  %%xmm10			\n\t"
+	"	vmulps		%%xmm11 ,  %%xmm0 ,  %%xmm11			\n\t"
+
+	"	vmovups		%%xmm8  ,    (%6)				\n\t"	// write a
+	"	vmovups		%%xmm9  ,  16(%6)				\n\t"
+	"	vmovups		%%xmm10 ,  32(%6)				\n\t"
+	"	vmovups		%%xmm11 ,  48(%6)				\n\t"
+
+	"	vmovups		%%xmm8  ,    (%4)				\n\t"	// write c0
+	"	vmovups		%%xmm9  ,  16(%4)				\n\t"
+	"	vmovups		%%xmm10 ,  32(%4)				\n\t"
+	"	vmovups		%%xmm11 ,  48(%4)				\n\t"
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+
+
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa,  bb;
+
+  int i, j, k;
+
+  a += (n - 1) * m;
+  b += (n - 1) * n;
+
+  for (i = n - 1; i >= 0; i--) {
+
+    bb = *(b + i);
+
+    for (j = 0; j < m; j ++) {
+      aa = *(c + j + i * ldc);
+      aa *= bb;
+      *a   = aa;
+      *(c + j + i * ldc) = aa;
+      a ++;
+
+      for (k = 0; k < i; k ++){
+	*(c + j + k * ldc) -= aa * *(b + k);
+      }
+
+    }
+    b -= n;
+    a -= 2 * m;
+  }
+
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  a += (n - 1) * m * 2;
+  b += (n - 1) * n * 2;
+
+  for (i = n - 1; i >= 0; i--) {
+
+    bb1 = *(b + i * 2 + 0);
+    bb2 = *(b + i * 2 + 1);
+
+    for (j = 0; j < m; j ++) {
+
+      aa1 = *(c + j * 2 + 0 + i * ldc);
+      aa2 = *(c + j * 2 + 1 + i * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 =  aa1 * bb1  + aa2 * bb2;
+      cc2 = - aa1 * bb2 + aa2 * bb1;
+#endif
+
+      *(a + 0) = cc1;
+      *(a + 1) = cc2;
+
+      *(c + j * 2 + 0 + i * ldc) = cc1;
+      *(c + j * 2 + 1 + i * ldc) = cc2;
+      a += 2;
+
+      for (k = 0; k < i; k ++){
+#ifndef CONJ
+	*(c + j * 2 + 0 + k * ldc) -= cc1 * *(b + k * 2 + 0) - cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#else
+	*(c + j * 2 + 0 + k * ldc) -=   cc1 * *(b + k * 2 + 0) + cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -=  -cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#endif
+      }
+
+    }
+    b -= n * 2;
+    a -= 4 * m;
+  }
+
+}
+
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k,  FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  BLASLONG i, j;
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+
+#if 0
+  fprintf(stderr, "TRSM RT KERNEL m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  kk = n - offset;
+  c += n * ldc * COMPSIZE;
+  b += n * k   * COMPSIZE;
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = 1;
+    while (j < GEMM_UNROLL_N) {
+      if (n & j) {
+
+	aa  = a;
+	b -= j * k  * COMPSIZE;
+	c -= j * ldc* COMPSIZE;
+	cc  = c;
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+	if (i > 0) {
+
+	  do {
+	    if (k - kk > 0) {
+	      GEMM_KERNEL(GEMM_UNROLL_M, j, k - kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa + GEMM_UNROLL_M * kk * COMPSIZE,
+			  b  +  j            * kk * COMPSIZE,
+			  cc,
+			  ldc);
+	    }
+
+	    solve(GEMM_UNROLL_M, j,
+		  aa + (kk - j) * GEMM_UNROLL_M * COMPSIZE,
+		  b  + (kk - j) * j             * COMPSIZE,
+		  cc, ldc);
+
+	    aa += GEMM_UNROLL_M * k * COMPSIZE;
+	    cc += GEMM_UNROLL_M     * COMPSIZE;
+	    i --;
+	  } while (i > 0);
+	}
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  i = (GEMM_UNROLL_M >> 1);
+	  do {
+	    if (m & i) {
+
+	      if (k - kk > 0) {
+		GEMM_KERNEL(i, j, k - kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa + i * kk * COMPSIZE,
+			    b  + j * kk * COMPSIZE,
+			    cc, ldc);
+	      }
+
+	      solve(i, j,
+		    aa + (kk - j) * i * COMPSIZE,
+		    b  + (kk - j) * j * COMPSIZE,
+		    cc, ldc);
+
+	      aa += i * k * COMPSIZE;
+	      cc += i     * COMPSIZE;
+
+	    }
+	    i >>= 1;
+	  } while (i > 0);
+	}
+	kk -= j;
+      }
+      j <<= 1;
+    }
+  }
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  if (j > 0) {
+
+    do {
+      aa  = a;
+      b -= GEMM_UNROLL_N * k   * COMPSIZE;
+      c -= GEMM_UNROLL_N * ldc * COMPSIZE;
+      cc  = c;
+
+      i = (m >> GEMM_UNROLL_M_SHIFT);
+      if (i > 0) {
+	do {
+
+	  strsm_RT_solve_opt(k - kk, aa + GEMM_UNROLL_M * kk * COMPSIZE, b  + GEMM_UNROLL_N * kk * COMPSIZE, cc, ldc,
+                            aa + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_M * COMPSIZE , b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE );
+
+	  aa += GEMM_UNROLL_M * k * COMPSIZE;
+	  cc += GEMM_UNROLL_M     * COMPSIZE;
+	  i --;
+	} while (i > 0);
+      }
+
+      if (m & (GEMM_UNROLL_M - 1)) {
+	i = (GEMM_UNROLL_M >> 1);
+	do {
+	  if (m & i) {
+	    if (k - kk > 0) {
+	      GEMM_KERNEL(i, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa + i             * kk * COMPSIZE,
+			  b  + GEMM_UNROLL_N * kk * COMPSIZE,
+			  cc,
+			  ldc);
+	    }
+
+	    solve(i, GEMM_UNROLL_N,
+		  aa + (kk - GEMM_UNROLL_N) * i             * COMPSIZE,
+		  b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE,
+		  cc, ldc);
+
+	    aa += i * k * COMPSIZE;
+	    cc += i     * COMPSIZE;
+	  }
+	  i >>= 1;
+	} while (i > 0);
+      }
+
+      kk -= GEMM_UNROLL_N;
+      j --;
+    } while (j > 0);
+  }
+
+  return 0;
+}
+
+

kernel/x86_64/ztrsm_kernel_LN_bulldozer.c

@@ -0,0 +1,497 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* 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.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``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 UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN 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.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_L
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+
+#ifndef CONJ
+
+static void ztrsm_LN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void ztrsm_LN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc*2 ;
+	BLASLONG n1 = n * 4;
+	BLASLONG i=0;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorpd	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorpd	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorpd	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorpd	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorpd	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorpd	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorpd	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorpd	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       3f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	prefetcht0	256(%3,%1,8)					\n\t"
+	"	prefetcht0	256(%2,%1,8)					\n\t"
+	"	vmovddup	(%3,%1,8), %%xmm0				\n\t"	// b0 real, b0 real
+	"	vmovddup       8(%3,%1,8), %%xmm1				\n\t"	// b0 imag, b0 imag
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"	// a0 real , a0 imag
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"	// a1 real , a1 imag
+	"	vmovddup      16(%3,%1,8), %%xmm2				\n\t"	// b1 real, b1 real
+	"	vmovddup      24(%3,%1,8), %%xmm3				\n\t"	// b1 imag, b1 imag
+
+	"	vfnmaddpd	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm9 , %%xmm1 , %%xmm4 , %%xmm9		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm10, %%xmm0 , %%xmm5 , %%xmm10		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm11, %%xmm1 , %%xmm5 , %%xmm11		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm12, %%xmm2 , %%xmm4 , %%xmm12		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm13, %%xmm3 , %%xmm4 , %%xmm13		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm14, %%xmm2 , %%xmm5 , %%xmm14		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm15, %%xmm3 , %%xmm5 , %%xmm15		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	addq		$4, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jz		2f						\n\t"
+
+	"	vmovddup	(%3,%1,8), %%xmm0				\n\t"	// b0 real, b0 real
+	"	vmovddup       8(%3,%1,8), %%xmm1				\n\t"	// b0 imag, b0 imag
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"	// a0 real , a0 imag
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"	// a1 real , a1 imag
+	"	vmovddup      16(%3,%1,8), %%xmm2				\n\t"	// b1 real, b1 real
+	"	vmovddup      24(%3,%1,8), %%xmm3				\n\t"	// b1 imag, b1 imag
+
+	"	vfnmaddpd	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm9 , %%xmm1 , %%xmm4 , %%xmm9		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm10, %%xmm0 , %%xmm5 , %%xmm10		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm11, %%xmm1 , %%xmm5 , %%xmm11		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm12, %%xmm2 , %%xmm4 , %%xmm12		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm13, %%xmm3 , %%xmm4 , %%xmm13		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm14, %%xmm2 , %%xmm5 , %%xmm14		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm15, %%xmm3 , %%xmm5 , %%xmm15		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	addq		$4, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+
+	"2:									\n\t"
+
+	"	vshufpd		$0x01 , %%xmm9  , %%xmm9, %%xmm9		\n\t"
+	"	vshufpd		$0x01 , %%xmm11 , %%xmm11 , %%xmm11		\n\t"
+	"	vshufpd		$0x01 , %%xmm13 , %%xmm13 , %%xmm13		\n\t"
+	"	vshufpd		$0x01 , %%xmm15 , %%xmm15 , %%xmm15		\n\t"
+
+	"	vaddsubpd		%%xmm8 , %%xmm9 , %%xmm8		\n\t"
+	"	vaddsubpd		%%xmm10, %%xmm11, %%xmm10		\n\t"
+	"	vaddsubpd		%%xmm12, %%xmm13, %%xmm12		\n\t"
+	"	vaddsubpd		%%xmm14, %%xmm15, %%xmm14		\n\t"
+
+	"	vxorpd		%%xmm7 , %%xmm7 , %%xmm7			\n\t"
+
+	"	vaddsubpd		%%xmm8 , %%xmm7 , %%xmm8		\n\t"
+	"	vaddsubpd		%%xmm10, %%xmm7 , %%xmm10		\n\t"
+	"	vaddsubpd		%%xmm12, %%xmm7 , %%xmm12		\n\t"
+	"	vaddsubpd		%%xmm14, %%xmm7 , %%xmm14		\n\t"
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+
+	"	vaddpd		%%xmm0 , %%xmm8 , %%xmm8			\n\t"
+	"	vaddpd		%%xmm1 , %%xmm10, %%xmm10			\n\t"
+	"	vaddpd		%%xmm4 , %%xmm12, %%xmm12			\n\t"
+	"	vaddpd		%%xmm5 , %%xmm14, %%xmm14			\n\t"
+
+	"	vmovups		%%xmm8  ,  (%4)					\n\t"
+	"	vmovups		%%xmm10 ,16(%4)					\n\t"
+
+	"	vmovups		%%xmm12 ,  (%5)					\n\t"
+	"	vmovups		%%xmm14 ,16(%5)					\n\t"
+
+	"3:									\n\t"	
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+#endif
+
+
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa,  bb;
+
+  int i, j, k;
+
+  a += (m - 1) * m;
+  b += (m - 1) * n;
+
+  for (i = m - 1; i >= 0; i--) {
+
+    aa = *(a + i);
+
+    for (j = 0; j < n; j ++) {
+      bb = *(c + i + j * ldc);
+      bb *= aa;
+      *b             = bb;
+      *(c + i + j * ldc) = bb;
+      b ++;
+
+      for (k = 0; k < i; k ++){
+	*(c + k + j * ldc) -= bb * *(a + k);
+      }
+
+    }
+    a -= m;
+    b -= 2 * n;
+  }
+
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+  a += (m - 1) * m * 2;
+  b += (m - 1) * n * 2;
+
+  for (i = m - 1; i >= 0; i--) {
+
+    aa1 = *(a + i * 2 + 0);
+    aa2 = *(a + i * 2 + 1);
+
+    for (j = 0; j < n; j ++) {
+      bb1 = *(c + i * 2 + 0 + j * ldc);
+      bb2 = *(c + i * 2 + 1 + j * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 = aa1 * bb1 + aa2 * bb2;
+      cc2 = aa1 * bb2 - aa2 * bb1;
+#endif
+
+
+      *(b + 0) = cc1;
+      *(b + 1) = cc2;
+      *(c + i * 2 + 0 + j * ldc) = cc1;
+      *(c + i * 2 + 1 + j * ldc) = cc2;
+      b += 2;
+
+      for (k = 0; k < i; k ++){
+#ifndef CONJ
+	*(c + k * 2 + 0 + j * ldc) -= cc1 * *(a + k * 2 + 0) - cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#else
+	*(c + k * 2 + 0 + j * ldc) -=   cc1 * *(a + k * 2 + 0) + cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= - cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#endif
+      }
+
+    }
+    a -= m * 2;
+    b -= 4 * n;
+  }
+
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k,  FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  BLASLONG i, j;
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+
+#if 0
+  fprintf(stderr, "TRSM KERNEL LN : m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  while (j > 0) {
+
+    kk = m + offset;
+
+    if (m & (GEMM_UNROLL_M - 1)) {
+      for (i = 1; i < GEMM_UNROLL_M; i *= 2){
+	if (m & i) {
+	  aa = a + ((m & ~(i - 1)) - i) * k * COMPSIZE;
+	  cc = c + ((m & ~(i - 1)) - i)     * COMPSIZE;
+
+	  if (k - kk > 0) {
+	    GEMM_KERNEL(i, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa + i             * kk * COMPSIZE,
+			b  + GEMM_UNROLL_N * kk * COMPSIZE,
+			cc,
+			ldc);
+	  }
+
+	  solve(i, GEMM_UNROLL_N,
+		aa + (kk - i) * i             * COMPSIZE,
+		b  + (kk - i) * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+	  kk -= i;
+	}
+      }
+    }
+
+    i = (m >> GEMM_UNROLL_M_SHIFT);
+    if (i > 0) {
+      aa = a + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M) * k * COMPSIZE;
+      cc = c + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M)     * COMPSIZE;
+
+      do {
+
+#ifdef CONJ
+	if (k - kk > 0) {
+	  GEMM_KERNEL(GEMM_UNROLL_M, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+		      ZERO,
+#endif
+		      aa + GEMM_UNROLL_M * kk * COMPSIZE,
+		      b +  GEMM_UNROLL_N * kk * COMPSIZE,
+		      cc,
+		      ldc);
+	}
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_M * COMPSIZE,
+	      b  + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#else
+
+	ztrsm_LN_solve_opt(k-kk, aa + GEMM_UNROLL_M * kk * COMPSIZE, b +  GEMM_UNROLL_N * kk * COMPSIZE, cc, ldc,
+			   aa + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_M * COMPSIZE, b  + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_N * COMPSIZE);
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_M * COMPSIZE,
+	      b  + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#endif
+
+	aa -= GEMM_UNROLL_M * k * COMPSIZE;
+	cc -= GEMM_UNROLL_M     * COMPSIZE;
+	kk -= GEMM_UNROLL_M;
+	i --;
+      } while (i > 0);
+    }
+
+    b += GEMM_UNROLL_N * k * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+	kk = m + offset;
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  for (i = 1; i < GEMM_UNROLL_M; i *= 2){
+	    if (m & i) {
+	      aa = a + ((m & ~(i - 1)) - i) * k * COMPSIZE;
+	      cc = c + ((m & ~(i - 1)) - i)     * COMPSIZE;
+
+	      if (k - kk > 0) {
+		GEMM_KERNEL(i, j, k - kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa + i * kk * COMPSIZE,
+			    b  + j * kk * COMPSIZE,
+			    cc, ldc);
+	      }
+
+	      solve(i, j,
+		    aa + (kk - i) * i * COMPSIZE,
+		    b  + (kk - i) * j * COMPSIZE,
+		    cc, ldc);
+
+	      kk -= i;
+	    }
+	  }
+	}
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+	if (i > 0) {
+	  aa = a + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M) * k * COMPSIZE;
+	  cc = c + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M)     * COMPSIZE;
+
+	  do {
+	    if (k - kk > 0) {
+	      GEMM_KERNEL(GEMM_UNROLL_M, j, k - kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa + GEMM_UNROLL_M * kk * COMPSIZE,
+			  b +  j             * kk * COMPSIZE,
+			  cc,
+			  ldc);
+	    }
+
+	    solve(GEMM_UNROLL_M, j,
+		  aa + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_M * COMPSIZE,
+		  b  + (kk - GEMM_UNROLL_M) * j             * COMPSIZE,
+		  cc, ldc);
+
+	    aa -= GEMM_UNROLL_M * k * COMPSIZE;
+	    cc -= GEMM_UNROLL_M     * COMPSIZE;
+	    kk -= GEMM_UNROLL_M;
+	    i --;
+	  } while (i > 0);
+	}
+
+	b += j * k   * COMPSIZE;
+	c += j * ldc * COMPSIZE;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/x86_64/ztrsm_kernel_LT_bulldozer.c

@@ -0,0 +1,480 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* 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.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``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 UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN 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.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_L
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+
+#ifndef CONJ
+
+static void ztrsm_LT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void ztrsm_LT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc*2 ;
+	BLASLONG n1 = n * 4;
+	BLASLONG i=0;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorpd	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorpd	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorpd	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorpd	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorpd	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorpd	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorpd	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorpd	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       3f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	prefetcht0	256(%3,%1,8)					\n\t"
+	"	prefetcht0	256(%2,%1,8)					\n\t"
+	"	vmovddup	(%3,%1,8), %%xmm0				\n\t"	// b0 real, b0 real
+	"	vmovddup       8(%3,%1,8), %%xmm1				\n\t"	// b0 imag, b0 imag
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"	// a0 real , a0 imag
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"	// a1 real , a1 imag
+	"	vmovddup      16(%3,%1,8), %%xmm2				\n\t"	// b1 real, b1 real
+	"	vmovddup      24(%3,%1,8), %%xmm3				\n\t"	// b1 imag, b1 imag
+
+	"	vfnmaddpd	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm9 , %%xmm1 , %%xmm4 , %%xmm9		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm10, %%xmm0 , %%xmm5 , %%xmm10		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm11, %%xmm1 , %%xmm5 , %%xmm11		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm12, %%xmm2 , %%xmm4 , %%xmm12		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm13, %%xmm3 , %%xmm4 , %%xmm13		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm14, %%xmm2 , %%xmm5 , %%xmm14		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm15, %%xmm3 , %%xmm5 , %%xmm15		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	addq		$4, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jz		2f						\n\t"
+
+	"	vmovddup	(%3,%1,8), %%xmm0				\n\t"	// b0 real, b0 real
+	"	vmovddup       8(%3,%1,8), %%xmm1				\n\t"	// b0 imag, b0 imag
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"	// a0 real , a0 imag
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"	// a1 real , a1 imag
+	"	vmovddup      16(%3,%1,8), %%xmm2				\n\t"	// b1 real, b1 real
+	"	vmovddup      24(%3,%1,8), %%xmm3				\n\t"	// b1 imag, b1 imag
+
+	"	vfnmaddpd	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm9 , %%xmm1 , %%xmm4 , %%xmm9		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm10, %%xmm0 , %%xmm5 , %%xmm10		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm11, %%xmm1 , %%xmm5 , %%xmm11		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm12, %%xmm2 , %%xmm4 , %%xmm12		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm13, %%xmm3 , %%xmm4 , %%xmm13		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm14, %%xmm2 , %%xmm5 , %%xmm14		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm15, %%xmm3 , %%xmm5 , %%xmm15		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	addq		$4, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+
+	"2:									\n\t"
+
+	"	vshufpd		$0x01 , %%xmm9  , %%xmm9, %%xmm9		\n\t"
+	"	vshufpd		$0x01 , %%xmm11 , %%xmm11 , %%xmm11		\n\t"
+	"	vshufpd		$0x01 , %%xmm13 , %%xmm13 , %%xmm13		\n\t"
+	"	vshufpd		$0x01 , %%xmm15 , %%xmm15 , %%xmm15		\n\t"
+
+	"	vaddsubpd		%%xmm8 , %%xmm9 , %%xmm8		\n\t"
+	"	vaddsubpd		%%xmm10, %%xmm11, %%xmm10		\n\t"
+	"	vaddsubpd		%%xmm12, %%xmm13, %%xmm12		\n\t"
+	"	vaddsubpd		%%xmm14, %%xmm15, %%xmm14		\n\t"
+
+	"	vxorpd		%%xmm7 , %%xmm7 , %%xmm7			\n\t"
+
+	"	vaddsubpd		%%xmm8 , %%xmm7 , %%xmm8		\n\t"
+	"	vaddsubpd		%%xmm10, %%xmm7 , %%xmm10		\n\t"
+	"	vaddsubpd		%%xmm12, %%xmm7 , %%xmm12		\n\t"
+	"	vaddsubpd		%%xmm14, %%xmm7 , %%xmm14		\n\t"
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+
+	"	vaddpd		%%xmm0 , %%xmm8 , %%xmm8			\n\t"
+	"	vaddpd		%%xmm1 , %%xmm10, %%xmm10			\n\t"
+	"	vaddpd		%%xmm4 , %%xmm12, %%xmm12			\n\t"
+	"	vaddpd		%%xmm5 , %%xmm14, %%xmm14			\n\t"
+
+	"	vmovups		%%xmm8  ,  (%4)					\n\t"
+	"	vmovups		%%xmm10 ,16(%4)					\n\t"
+
+	"	vmovups		%%xmm12 ,  (%5)					\n\t"
+	"	vmovups		%%xmm14 ,16(%5)					\n\t"
+
+	"3:									\n\t"	
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+#endif
+
+
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa, bb;
+
+  int i, j, k;
+
+  for (i = 0; i < m; i++) {
+
+    aa = *(a + i);
+
+    for (j = 0; j < n; j ++) {
+      bb = *(c + i + j * ldc);
+      bb *= aa;
+      *b             = bb;
+      *(c + i + j * ldc) = bb;
+      b ++;
+
+      for (k = i + 1; k < m; k ++){
+	*(c + k + j * ldc) -= bb * *(a + k);
+      }
+
+    }
+    a += m;
+  }
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  for (i = 0; i < m; i++) {
+
+    aa1 = *(a + i * 2 + 0);
+    aa2 = *(a + i * 2 + 1);
+
+    for (j = 0; j < n; j ++) {
+      bb1 = *(c + i * 2 + 0 + j * ldc);
+      bb2 = *(c + i * 2 + 1 + j * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 = aa1 * bb1 + aa2 * bb2;
+      cc2 = aa1 * bb2 - aa2 * bb1;
+#endif
+
+      *(b + 0) = cc1;
+      *(b + 1) = cc2;
+      *(c + i * 2 + 0 + j * ldc) = cc1;
+      *(c + i * 2 + 1 + j * ldc) = cc2;
+      b += 2;
+
+      for (k = i + 1; k < m; k ++){
+#ifndef CONJ
+	*(c + k * 2 + 0 + j * ldc) -= cc1 * *(a + k * 2 + 0) - cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#else
+	*(c + k * 2 + 0 + j * ldc) -= cc1 * *(a + k * 2 + 0) + cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= -cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#endif
+      }
+
+    }
+    a += m * 2;
+  }
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+  BLASLONG i, j, jj;
+
+#if 0
+  fprintf(stderr, "TRSM KERNEL LT : m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  jj = 0;
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  while (j > 0) {
+
+    kk = offset;
+    aa = a;
+    cc = c;
+
+    i = (m >> GEMM_UNROLL_M_SHIFT);
+
+    while (i > 0) {
+
+#ifdef CONJ
+
+	if (kk > 0) {
+	  GEMM_KERNEL(GEMM_UNROLL_M, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+		      ZERO,
+#endif
+		      aa, b, cc, ldc);
+	}
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + kk * GEMM_UNROLL_M * COMPSIZE,
+	      b  + kk * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#else
+
+	ztrsm_LT_solve_opt(kk, aa, b, cc, ldc, aa + kk * GEMM_UNROLL_M * COMPSIZE, b  + kk * GEMM_UNROLL_N * COMPSIZE);
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + kk * GEMM_UNROLL_M * COMPSIZE,
+	      b  + kk * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#endif
+
+      aa += GEMM_UNROLL_M * k * COMPSIZE;
+      cc += GEMM_UNROLL_M     * COMPSIZE;
+      kk += GEMM_UNROLL_M;
+      i --;
+    }
+
+    if (m & (GEMM_UNROLL_M - 1)) {
+      i = (GEMM_UNROLL_M >> 1);
+      while (i > 0) {
+	if (m & i) {
+	    if (kk > 0) {
+	      GEMM_KERNEL(i, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa, b, cc, ldc);
+	    }
+	  solve(i, GEMM_UNROLL_N,
+		aa + kk * i             * COMPSIZE,
+		b  + kk * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+	  aa += i * k * COMPSIZE;
+	  cc += i     * COMPSIZE;
+	  kk += i;
+	}
+	i >>= 1;
+      }
+    }
+
+    b += GEMM_UNROLL_N * k   * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+    jj += GEMM_UNROLL_M;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+	kk = offset;
+	aa = a;
+	cc = c;
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+
+	while (i > 0) {
+	  if (kk > 0) {
+	    GEMM_KERNEL(GEMM_UNROLL_M, j, kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa,
+			b,
+			cc,
+			ldc);
+	  }
+
+	  solve(GEMM_UNROLL_M, j,
+		aa + kk * GEMM_UNROLL_M * COMPSIZE,
+		b  + kk * j             * COMPSIZE, cc, ldc);
+
+	  aa += GEMM_UNROLL_M * k * COMPSIZE;
+	  cc += GEMM_UNROLL_M     * COMPSIZE;
+	  kk += GEMM_UNROLL_M;
+	  i --;
+	}
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  i = (GEMM_UNROLL_M >> 1);
+	  while (i > 0) {
+	    if (m & i) {
+	      if (kk > 0) {
+		GEMM_KERNEL(i, j, kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa,
+			    b,
+			    cc,
+			    ldc);
+	      }
+
+	      solve(i, j,
+		    aa + kk * i * COMPSIZE,
+		    b  + kk * j * COMPSIZE, cc, ldc);
+
+	      aa += i * k * COMPSIZE;
+	      cc += i     * COMPSIZE;
+	      kk += i;
+	      }
+	    i >>= 1;
+	  }
+	}
+
+	b += j * k   * COMPSIZE;
+	c += j * ldc * COMPSIZE;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/x86_64/ztrsm_kernel_RN_bulldozer.c

@@ -0,0 +1,476 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* 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.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``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 UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN 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.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_R
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+#ifndef CONJ
+
+static void ztrsm_RN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void ztrsm_RN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc*2 ;
+	BLASLONG n1 = n * 4;
+	BLASLONG i=0;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorpd	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorpd	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorpd	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorpd	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorpd	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorpd	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorpd	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorpd	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       3f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	prefetcht0	256(%3,%1,8)					\n\t"
+	"	prefetcht0	256(%2,%1,8)					\n\t"
+	"	vmovddup	(%3,%1,8), %%xmm0				\n\t"	// b0 real, b0 real
+	"	vmovddup       8(%3,%1,8), %%xmm1				\n\t"	// b0 imag, b0 imag
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"	// a0 real , a0 imag
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"	// a1 real , a1 imag
+	"	vmovddup      16(%3,%1,8), %%xmm2				\n\t"	// b1 real, b1 real
+	"	vmovddup      24(%3,%1,8), %%xmm3				\n\t"	// b1 imag, b1 imag
+
+	"	vfnmaddpd	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm9 , %%xmm1 , %%xmm4 , %%xmm9		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm10, %%xmm0 , %%xmm5 , %%xmm10		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm11, %%xmm1 , %%xmm5 , %%xmm11		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm12, %%xmm2 , %%xmm4 , %%xmm12		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm13, %%xmm3 , %%xmm4 , %%xmm13		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm14, %%xmm2 , %%xmm5 , %%xmm14		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm15, %%xmm3 , %%xmm5 , %%xmm15		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	addq		$4, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jz		2f						\n\t"
+
+	"	vmovddup	(%3,%1,8), %%xmm0				\n\t"	// b0 real, b0 real
+	"	vmovddup       8(%3,%1,8), %%xmm1				\n\t"	// b0 imag, b0 imag
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"	// a0 real , a0 imag
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"	// a1 real , a1 imag
+	"	vmovddup      16(%3,%1,8), %%xmm2				\n\t"	// b1 real, b1 real
+	"	vmovddup      24(%3,%1,8), %%xmm3				\n\t"	// b1 imag, b1 imag
+
+	"	vfnmaddpd	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm9 , %%xmm1 , %%xmm4 , %%xmm9		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm10, %%xmm0 , %%xmm5 , %%xmm10		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm11, %%xmm1 , %%xmm5 , %%xmm11		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm12, %%xmm2 , %%xmm4 , %%xmm12		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm13, %%xmm3 , %%xmm4 , %%xmm13		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm14, %%xmm2 , %%xmm5 , %%xmm14		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm15, %%xmm3 , %%xmm5 , %%xmm15		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	addq		$4, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+
+	"2:									\n\t"
+
+	"	vshufpd		$0x01 , %%xmm9  , %%xmm9, %%xmm9		\n\t"
+	"	vshufpd		$0x01 , %%xmm11 , %%xmm11 , %%xmm11		\n\t"
+	"	vshufpd		$0x01 , %%xmm13 , %%xmm13 , %%xmm13		\n\t"
+	"	vshufpd		$0x01 , %%xmm15 , %%xmm15 , %%xmm15		\n\t"
+
+	"	vaddsubpd		%%xmm8 , %%xmm9 , %%xmm8		\n\t"
+	"	vaddsubpd		%%xmm10, %%xmm11, %%xmm10		\n\t"
+	"	vaddsubpd		%%xmm12, %%xmm13, %%xmm12		\n\t"
+	"	vaddsubpd		%%xmm14, %%xmm15, %%xmm14		\n\t"
+
+	"	vxorpd		%%xmm7 , %%xmm7 , %%xmm7			\n\t"
+
+	"	vaddsubpd		%%xmm8 , %%xmm7 , %%xmm8		\n\t"
+	"	vaddsubpd		%%xmm10, %%xmm7 , %%xmm10		\n\t"
+	"	vaddsubpd		%%xmm12, %%xmm7 , %%xmm12		\n\t"
+	"	vaddsubpd		%%xmm14, %%xmm7 , %%xmm14		\n\t"
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+
+	"	vaddpd		%%xmm0 , %%xmm8 , %%xmm8			\n\t"
+	"	vaddpd		%%xmm1 , %%xmm10, %%xmm10			\n\t"
+	"	vaddpd		%%xmm4 , %%xmm12, %%xmm12			\n\t"
+	"	vaddpd		%%xmm5 , %%xmm14, %%xmm14			\n\t"
+
+	"	vmovups		%%xmm8  ,  (%4)					\n\t"
+	"	vmovups		%%xmm10 ,16(%4)					\n\t"
+
+	"	vmovups		%%xmm12 ,  (%5)					\n\t"
+	"	vmovups		%%xmm14 ,16(%5)					\n\t"
+
+	"3:									\n\t"	
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+#endif
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa, bb;
+
+  int i, j, k;
+
+  for (i = 0; i < n; i++) {
+
+    bb = *(b + i);
+
+    for (j = 0; j < m; j ++) {
+      aa = *(c + j + i * ldc);
+      aa *= bb;
+      *a  = aa;
+      *(c + j + i * ldc) = aa;
+      a ++;
+
+      for (k = i + 1; k < n; k ++){
+	*(c + j + k * ldc) -= aa * *(b + k);
+      }
+
+    }
+    b += n;
+  }
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  for (i = 0; i < n; i++) {
+
+    bb1 = *(b + i * 2 + 0);
+    bb2 = *(b + i * 2 + 1);
+
+    for (j = 0; j < m; j ++) {
+      aa1 = *(c + j * 2 + 0 + i * ldc);
+      aa2 = *(c + j * 2 + 1 + i * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 =  aa1 * bb1 + aa2 * bb2;
+      cc2 = -aa1 * bb2 + aa2 * bb1;
+#endif
+
+      *(a + 0) = cc1;
+      *(a + 1) = cc2;
+      *(c + j * 2 + 0 + i * ldc) = cc1;
+      *(c + j * 2 + 1 + i * ldc) = cc2;
+      a += 2;
+
+      for (k = i + 1; k < n; k ++){
+#ifndef CONJ
+	*(c + j * 2 + 0 + k * ldc) -= cc1 * *(b + k * 2 + 0) - cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#else
+	*(c + j * 2 + 0 + k * ldc) -=   cc1 * *(b + k * 2 + 0) + cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= - cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#endif
+      }
+
+    }
+    b += n * 2;
+  }
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+  BLASLONG i, j, jj;
+
+#if 0
+  fprintf(stderr, "TRSM RN KERNEL m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  jj = 0;
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+  kk = -offset;
+
+  while (j > 0) {
+
+    aa = a;
+    cc = c;
+
+    i = (m >> GEMM_UNROLL_M_SHIFT);
+
+    if (i > 0) {
+      do {
+
+#ifndef CONJ
+
+        ztrsm_RN_solve_opt(kk, aa, b, cc, ldc, aa + kk * GEMM_UNROLL_M * COMPSIZE, b  + kk * GEMM_UNROLL_N * COMPSIZE);
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + kk * GEMM_UNROLL_M * COMPSIZE,
+	      b  + kk * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#else
+
+        if (kk > 0) {
+          GEMM_KERNEL(GEMM_UNROLL_M, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+                      ZERO,
+#endif
+                      aa, b, cc, ldc);
+        }
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + kk * GEMM_UNROLL_M * COMPSIZE,
+	      b  + kk * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#endif
+
+	aa += GEMM_UNROLL_M * k * COMPSIZE;
+	cc += GEMM_UNROLL_M     * COMPSIZE;
+	i --;
+      } while (i > 0);
+    }
+
+
+    if (m & (GEMM_UNROLL_M - 1)) {
+      i = (GEMM_UNROLL_M >> 1);
+      while (i > 0) {
+	if (m & i) {
+	    if (kk > 0) {
+	      GEMM_KERNEL(i, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa, b, cc, ldc);
+	    }
+	  solve(i, GEMM_UNROLL_N,
+		aa + kk * i             * COMPSIZE,
+		b  + kk * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+	  aa += i * k * COMPSIZE;
+	  cc += i     * COMPSIZE;
+	}
+	i >>= 1;
+      }
+    }
+
+    kk += GEMM_UNROLL_N;
+    b += GEMM_UNROLL_N * k   * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+    jj += GEMM_UNROLL_M;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+	aa = a;
+	cc = c;
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+
+	while (i > 0) {
+	  if (kk > 0) {
+	    GEMM_KERNEL(GEMM_UNROLL_M, j, kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa,
+			b,
+			cc,
+			ldc);
+	  }
+
+	  solve(GEMM_UNROLL_M, j,
+		aa + kk * GEMM_UNROLL_M * COMPSIZE,
+		b  + kk * j             * COMPSIZE, cc, ldc);
+
+	  aa += GEMM_UNROLL_M * k * COMPSIZE;
+	  cc += GEMM_UNROLL_M     * COMPSIZE;
+	  i --;
+	}
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  i = (GEMM_UNROLL_M >> 1);
+	  while (i > 0) {
+	    if (m & i) {
+	      if (kk > 0) {
+		GEMM_KERNEL(i, j, kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa,
+			    b,
+			    cc,
+			    ldc);
+	      }
+
+	      solve(i, j,
+		    aa + kk * i * COMPSIZE,
+		    b  + kk * j * COMPSIZE, cc, ldc);
+
+	      aa += i * k * COMPSIZE;
+	      cc += i     * COMPSIZE;
+	      }
+	    i >>= 1;
+	  }
+	}
+
+	b += j * k   * COMPSIZE;
+	c += j * ldc * COMPSIZE;
+	kk += j;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/x86_64/ztrsm_kernel_RT_bulldozer.c

@@ -0,0 +1,506 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* 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.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``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 UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN 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.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_R
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+
+#ifndef CONJ
+
+static void ztrsm_RT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void ztrsm_RT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc*2 ;
+	BLASLONG n1 = n * 4;
+	BLASLONG i=0;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorpd	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorpd	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorpd	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorpd	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorpd	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorpd	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorpd	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorpd	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       3f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	prefetcht0	256(%3,%1,8)					\n\t"
+	"	prefetcht0	256(%2,%1,8)					\n\t"
+	"	vmovddup	(%3,%1,8), %%xmm0				\n\t"	// b0 real, b0 real
+	"	vmovddup       8(%3,%1,8), %%xmm1				\n\t"	// b0 imag, b0 imag
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"	// a0 real , a0 imag
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"	// a1 real , a1 imag
+	"	vmovddup      16(%3,%1,8), %%xmm2				\n\t"	// b1 real, b1 real
+	"	vmovddup      24(%3,%1,8), %%xmm3				\n\t"	// b1 imag, b1 imag
+
+	"	vfnmaddpd	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm9 , %%xmm1 , %%xmm4 , %%xmm9		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm10, %%xmm0 , %%xmm5 , %%xmm10		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm11, %%xmm1 , %%xmm5 , %%xmm11		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm12, %%xmm2 , %%xmm4 , %%xmm12		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm13, %%xmm3 , %%xmm4 , %%xmm13		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm14, %%xmm2 , %%xmm5 , %%xmm14		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm15, %%xmm3 , %%xmm5 , %%xmm15		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	addq		$4, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jz		2f						\n\t"
+
+	"	vmovddup	(%3,%1,8), %%xmm0				\n\t"	// b0 real, b0 real
+	"	vmovddup       8(%3,%1,8), %%xmm1				\n\t"	// b0 imag, b0 imag
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"	// a0 real , a0 imag
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"	// a1 real , a1 imag
+	"	vmovddup      16(%3,%1,8), %%xmm2				\n\t"	// b1 real, b1 real
+	"	vmovddup      24(%3,%1,8), %%xmm3				\n\t"	// b1 imag, b1 imag
+
+	"	vfnmaddpd	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm9 , %%xmm1 , %%xmm4 , %%xmm9		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm10, %%xmm0 , %%xmm5 , %%xmm10		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm11, %%xmm1 , %%xmm5 , %%xmm11		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm12, %%xmm2 , %%xmm4 , %%xmm12		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm13, %%xmm3 , %%xmm4 , %%xmm13		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddpd	%%xmm14, %%xmm2 , %%xmm5 , %%xmm14		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddpd	%%xmm15, %%xmm3 , %%xmm5 , %%xmm15		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	addq		$4, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+
+	"2:									\n\t"
+
+	"	vshufpd		$0x01 , %%xmm9  , %%xmm9, %%xmm9		\n\t"
+	"	vshufpd		$0x01 , %%xmm11 , %%xmm11 , %%xmm11		\n\t"
+	"	vshufpd		$0x01 , %%xmm13 , %%xmm13 , %%xmm13		\n\t"
+	"	vshufpd		$0x01 , %%xmm15 , %%xmm15 , %%xmm15		\n\t"
+
+	"	vaddsubpd		%%xmm8 , %%xmm9 , %%xmm8		\n\t"
+	"	vaddsubpd		%%xmm10, %%xmm11, %%xmm10		\n\t"
+	"	vaddsubpd		%%xmm12, %%xmm13, %%xmm12		\n\t"
+	"	vaddsubpd		%%xmm14, %%xmm15, %%xmm14		\n\t"
+
+	"	vxorpd		%%xmm7 , %%xmm7 , %%xmm7			\n\t"
+
+	"	vaddsubpd		%%xmm8 , %%xmm7 , %%xmm8		\n\t"
+	"	vaddsubpd		%%xmm10, %%xmm7 , %%xmm10		\n\t"
+	"	vaddsubpd		%%xmm12, %%xmm7 , %%xmm12		\n\t"
+	"	vaddsubpd		%%xmm14, %%xmm7 , %%xmm14		\n\t"
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+
+	"	vaddpd		%%xmm0 , %%xmm8 , %%xmm8			\n\t"
+	"	vaddpd		%%xmm1 , %%xmm10, %%xmm10			\n\t"
+	"	vaddpd		%%xmm4 , %%xmm12, %%xmm12			\n\t"
+	"	vaddpd		%%xmm5 , %%xmm14, %%xmm14			\n\t"
+
+	"	vmovups		%%xmm8  ,  (%4)					\n\t"
+	"	vmovups		%%xmm10 ,16(%4)					\n\t"
+
+	"	vmovups		%%xmm12 ,  (%5)					\n\t"
+	"	vmovups		%%xmm14 ,16(%5)					\n\t"
+
+	"3:									\n\t"	
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+#endif
+
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa,  bb;
+
+  int i, j, k;
+
+  a += (n - 1) * m;
+  b += (n - 1) * n;
+
+  for (i = n - 1; i >= 0; i--) {
+
+    bb = *(b + i);
+
+    for (j = 0; j < m; j ++) {
+      aa = *(c + j + i * ldc);
+      aa *= bb;
+      *a   = aa;
+      *(c + j + i * ldc) = aa;
+      a ++;
+
+      for (k = 0; k < i; k ++){
+	*(c + j + k * ldc) -= aa * *(b + k);
+      }
+
+    }
+    b -= n;
+    a -= 2 * m;
+  }
+
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  a += (n - 1) * m * 2;
+  b += (n - 1) * n * 2;
+
+  for (i = n - 1; i >= 0; i--) {
+
+    bb1 = *(b + i * 2 + 0);
+    bb2 = *(b + i * 2 + 1);
+
+    for (j = 0; j < m; j ++) {
+
+      aa1 = *(c + j * 2 + 0 + i * ldc);
+      aa2 = *(c + j * 2 + 1 + i * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 =  aa1 * bb1  + aa2 * bb2;
+      cc2 = - aa1 * bb2 + aa2 * bb1;
+#endif
+
+      *(a + 0) = cc1;
+      *(a + 1) = cc2;
+
+      *(c + j * 2 + 0 + i * ldc) = cc1;
+      *(c + j * 2 + 1 + i * ldc) = cc2;
+      a += 2;
+
+      for (k = 0; k < i; k ++){
+#ifndef CONJ
+	*(c + j * 2 + 0 + k * ldc) -= cc1 * *(b + k * 2 + 0) - cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#else
+	*(c + j * 2 + 0 + k * ldc) -=   cc1 * *(b + k * 2 + 0) + cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -=  -cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#endif
+      }
+
+    }
+    b -= n * 2;
+    a -= 4 * m;
+  }
+
+}
+
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k,  FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  BLASLONG i, j;
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+
+#if 0
+  fprintf(stderr, "TRSM RT KERNEL m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  kk = n - offset;
+  c += n * ldc * COMPSIZE;
+  b += n * k   * COMPSIZE;
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = 1;
+    while (j < GEMM_UNROLL_N) {
+      if (n & j) {
+
+	aa  = a;
+	b -= j * k  * COMPSIZE;
+	c -= j * ldc* COMPSIZE;
+	cc  = c;
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+	if (i > 0) {
+
+	  do {
+	    if (k - kk > 0) {
+	      GEMM_KERNEL(GEMM_UNROLL_M, j, k - kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa + GEMM_UNROLL_M * kk * COMPSIZE,
+			  b  +  j            * kk * COMPSIZE,
+			  cc,
+			  ldc);
+	    }
+
+	    solve(GEMM_UNROLL_M, j,
+		  aa + (kk - j) * GEMM_UNROLL_M * COMPSIZE,
+		  b  + (kk - j) * j             * COMPSIZE,
+		  cc, ldc);
+
+	    aa += GEMM_UNROLL_M * k * COMPSIZE;
+	    cc += GEMM_UNROLL_M     * COMPSIZE;
+	    i --;
+	  } while (i > 0);
+	}
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  i = (GEMM_UNROLL_M >> 1);
+	  do {
+	    if (m & i) {
+
+	      if (k - kk > 0) {
+		GEMM_KERNEL(i, j, k - kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa + i * kk * COMPSIZE,
+			    b  + j * kk * COMPSIZE,
+			    cc, ldc);
+	      }
+
+	      solve(i, j,
+		    aa + (kk - j) * i * COMPSIZE,
+		    b  + (kk - j) * j * COMPSIZE,
+		    cc, ldc);
+
+	      aa += i * k * COMPSIZE;
+	      cc += i     * COMPSIZE;
+
+	    }
+	    i >>= 1;
+	  } while (i > 0);
+	}
+	kk -= j;
+      }
+      j <<= 1;
+    }
+  }
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  if (j > 0) {
+
+    do {
+      aa  = a;
+      b -= GEMM_UNROLL_N * k   * COMPSIZE;
+      c -= GEMM_UNROLL_N * ldc * COMPSIZE;
+      cc  = c;
+
+      i = (m >> GEMM_UNROLL_M_SHIFT);
+      if (i > 0) {
+	do {
+
+#ifndef CONJ
+
+	  ztrsm_RT_solve_opt(k-kk, aa + GEMM_UNROLL_M * kk * COMPSIZE, b  + GEMM_UNROLL_N * kk * COMPSIZE, cc, ldc,
+                             aa + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_M * COMPSIZE, b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE);
+ 
+	  solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+		aa + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_M * COMPSIZE,
+		b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+#else
+	  if (k - kk > 0) {
+	    GEMM_KERNEL(GEMM_UNROLL_M, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa + GEMM_UNROLL_M * kk * COMPSIZE,
+			b  + GEMM_UNROLL_N * kk * COMPSIZE,
+			cc,
+			ldc);
+	  }
+
+	  solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+		aa + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_M * COMPSIZE,
+		b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+#endif
+
+	  aa += GEMM_UNROLL_M * k * COMPSIZE;
+	  cc += GEMM_UNROLL_M     * COMPSIZE;
+	  i --;
+	} while (i > 0);
+      }
+
+      if (m & (GEMM_UNROLL_M - 1)) {
+	i = (GEMM_UNROLL_M >> 1);
+	do {
+	  if (m & i) {
+	    if (k - kk > 0) {
+	      GEMM_KERNEL(i, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa + i             * kk * COMPSIZE,
+			  b  + GEMM_UNROLL_N * kk * COMPSIZE,
+			  cc,
+			  ldc);
+	    }
+
+	    solve(i, GEMM_UNROLL_N,
+		  aa + (kk - GEMM_UNROLL_N) * i             * COMPSIZE,
+		  b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE,
+		  cc, ldc);
+
+	    aa += i * k * COMPSIZE;
+	    cc += i     * COMPSIZE;
+	  }
+	  i >>= 1;
+	} while (i > 0);
+      }
+
+      kk -= GEMM_UNROLL_N;
+      j --;
+    } while (j > 0);
+  }
+
+  return 0;
+}
+
+