-/*********************************************************************************\r
-Copyright (c) 2013, The OpenBLAS Project\r
-All rights reserved.\r
-Redistribution and use in source and binary forms, with or without\r
-modification, are permitted provided that the following conditions are\r
-met:\r
-1. Redistributions of source code must retain the above copyright\r
-notice, this list of conditions and the following disclaimer.\r
-2. Redistributions in binary form must reproduce the above copyright\r
-notice, this list of conditions and the following disclaimer in\r
-the documentation and/or other materials provided with the\r
-distribution.\r
-3. Neither the name of the OpenBLAS project nor the names of\r
-its contributors may be used to endorse or promote products\r
-derived from this software without specific prior written permission.\r
-THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"\r
-AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE\r
-IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE\r
-ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE\r
-LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL\r
-DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR\r
-SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER\r
-CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,\r
-OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE\r
-USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.\r
-**********************************************************************************/\r
-\r
-/*********************************************************************\r
-* 2014/06/28 Saar\r
-* BLASTEST : OK\r
-* CTEST : OK\r
-* TEST : OK\r
-*\r
-* 2013/10/28 Saar\r
-* Parameter:\r
-* CGEMM_DEFAULT_UNROLL_N 2\r
-* CGEMM_DEFAULT_UNROLL_M 8\r
-* CGEMM_DEFAULT_P 384\r
-* CGEMM_DEFAULT_Q 192\r
-* A_PR1 512\r
-* B_PR1 512\r
-*\r
-* Performance at 6912x6912x6912:\r
-* 1 thread: 84 GFLOPS (SANDYBRIDGE: 60) (MKL: 86)\r
-* 2 threads: 153 GFLOPS (SANDYBRIDGE: 114) (MKL: 155)\r
-* 3 threads: 224 GFLOPS (SANDYBRIDGE: 162) (MKL: 222)\r
-* 4 threads: 278 GFLOPS (SANDYBRIDGE: 223) (MKL: 279)\r
-*\r
-*\r
-*********************************************************************/\r
-\r
-\r
-\r
-#define ASSEMBLER\r
-#include "common.h"\r
- \r
-#define OLD_M %rdi\r
-#define OLD_N %rsi\r
-#define M %r13\r
-#define J %r14\r
-#define OLD_K %rdx\r
-\r
-#define A %rcx\r
-#define B %r8\r
-#define C %r9\r
-#define LDC %r10\r
- \r
-#define I %r11\r
-#define AO %rdi\r
-#define BO %rsi\r
-#define CO1 %r15\r
-#define K %r12\r
-#define BI %rbp\r
-#define SP %rbx\r
-\r
-#define BO1 %rdi\r
-#define BO2 %r15\r
-\r
-#ifndef WINDOWS_ABI\r
-\r
-#define STACKSIZE 96\r
-\r
-#else\r
-\r
-#define STACKSIZE 320\r
-\r
-#define OLD_ALPHA_I 40 + STACKSIZE(%rsp)\r
-#define OLD_A 48 + STACKSIZE(%rsp)\r
-#define OLD_B 56 + STACKSIZE(%rsp)\r
-#define OLD_C 64 + STACKSIZE(%rsp)\r
-#define OLD_LDC 72 + STACKSIZE(%rsp)\r
-#define OLD_OFFSET 80 + STACKSIZE(%rsp)\r
-\r
-#endif\r
-\r
-#define L_BUFFER_SIZE 8192\r
-\r
-#define Ndiv6 24(%rsp)\r
-#define Nmod6 32(%rsp)\r
-#define N 40(%rsp)\r
-#define ALPHA_R 48(%rsp)\r
-#define ALPHA_I 56(%rsp)\r
-#define OFFSET 64(%rsp)\r
-#define KK 72(%rsp)\r
-#define KKK 80(%rsp)\r
-#define BUFFER1 128(%rsp)\r
-\r
-#if defined(OS_WINDOWS)\r
-#if L_BUFFER_SIZE > 16384\r
-#define STACK_TOUCH \\r
- movl $ 0, 4096 * 4(%rsp);\\r
- movl $ 0, 4096 * 3(%rsp);\\r
- movl $ 0, 4096 * 2(%rsp);\\r
- movl $ 0, 4096 * 1(%rsp);\r
-#elif L_BUFFER_SIZE > 12288\r
-#define STACK_TOUCH \\r
- movl $ 0, 4096 * 3(%rsp);\\r
- movl $ 0, 4096 * 2(%rsp);\\r
- movl $ 0, 4096 * 1(%rsp);\r
-#elif L_BUFFER_SIZE > 8192\r
-#define STACK_TOUCH \\r
- movl $ 0, 4096 * 2(%rsp);\\r
- movl $ 0, 4096 * 1(%rsp);\r
-#elif L_BUFFER_SIZE > 4096\r
-#define STACK_TOUCH \\r
- movl $ 0, 4096 * 1(%rsp);\r
-#else\r
-#define STACK_TOUCH\r
-#endif\r
-#else\r
-#define STACK_TOUCH\r
-#endif\r
-\r
-\r
-#if defined(BULLDOZER)\r
-\r
-#if defined(NN) || defined(NT) || defined(TN) || defined(TT)\r
-\r
-#define VFMADDPS_R( y0,y1,y2 ) vfmaddps y0,y1,y2,y0\r
-\r
-#define VFMADDPS_I( y0,y1,y2 ) vfmaddps y0,y1,y2,y0\r
-\r
-#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)\r
-\r
-#define VFMADDPS_R( y0,y1,y2 ) vfnmaddps y0,y1,y2,y0\r
-\r
-#define VFMADDPS_I( y0,y1,y2 ) vfmaddps y0,y1,y2,y0\r
-\r
-#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)\r
-\r
-#define VFMADDPS_R( y0,y1,y2 ) vfmaddps y0,y1,y2,y0\r
-\r
-#define VFMADDPS_I( y0,y1,y2 ) vfnmaddps y0,y1,y2,y0\r
-\r
-#else\r
-\r
-#define VFMADDPS_R( y0,y1,y2 ) vfnmaddps y0,y1,y2,y0\r
-\r
-#define VFMADDPS_I( y0,y1,y2 ) vfnmaddps y0,y1,y2,y0\r
-\r
-#endif\r
-\r
-#else\r
-\r
-#if defined(NN) || defined(NT) || defined(TN) || defined(TT)\r
-\r
-#define VFMADDPS_R( y0,y1,y2 ) vfmadd231ps y1,y2,y0\r
-\r
-#define VFMADDPS_I( y0,y1,y2 ) vfmadd231ps y1,y2,y0\r
-\r
-#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)\r
-\r
-#define VFMADDPS_R( y0,y1,y2 ) vfnmadd231ps y1,y2,y0\r
-\r
-#define VFMADDPS_I( y0,y1,y2 ) vfmadd231ps y1,y2,y0\r
-\r
-#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)\r
-\r
-#define VFMADDPS_R( y0,y1,y2 ) vfmadd231ps y1,y2,y0\r
-\r
-#define VFMADDPS_I( y0,y1,y2 ) vfnmadd231ps y1,y2,y0\r
-\r
-#else\r
-\r
-#define VFMADDPS_R( y0,y1,y2 ) vfnmadd231ps y1,y2,y0\r
-\r
-#define VFMADDPS_I( y0,y1,y2 ) vfnmadd231ps y1,y2,y0\r
-\r
-#endif\r
-\r
-#endif\r
-\r
-\r
-#define A_PR1 512\r
-#define B_PR1 512\r
-\r
-/***************************************************************************************************************************/\r
-\r
-.macro KERNEL8x2_SUB\r
-\r
- vmovups -16 * SIZE(AO, %rax, SIZE), %ymm0\r
- vbroadcastss -8 * SIZE(BO, BI, SIZE), %ymm4\r
- VFMADDPS_R( %ymm8,%ymm4,%ymm0 )\r
- vmovups -8 * SIZE(AO, %rax, SIZE), %ymm1\r
- VFMADDPS_R( %ymm12,%ymm4,%ymm1 )\r
- vbroadcastss -7 * SIZE(BO, BI, SIZE), %ymm5\r
- VFMADDPS_I( %ymm9,%ymm5,%ymm0 )\r
- VFMADDPS_I( %ymm13,%ymm5,%ymm1 )\r
- vbroadcastss -6 * SIZE(BO, BI, SIZE), %ymm6\r
- VFMADDPS_R( %ymm10,%ymm6,%ymm0 )\r
- VFMADDPS_R( %ymm14,%ymm6,%ymm1 )\r
- vbroadcastss -5 * SIZE(BO, BI, SIZE), %ymm7\r
- VFMADDPS_I( %ymm11,%ymm7,%ymm0 )\r
- VFMADDPS_I( %ymm15,%ymm7,%ymm1 )\r
- addq $ 4 , BI \r
- addq $ 16, %rax \r
-.endm\r
-\r
-.macro SAVE8x2\r
-\r
- vbroadcastss ALPHA_R, %ymm0\r
- vbroadcastss ALPHA_I, %ymm1\r
-\r
- // swap high and low 64 bytes\r
- vshufps $ 0xb1, %ymm9 , %ymm9, %ymm9\r
- vshufps $ 0xb1, %ymm11, %ymm11, %ymm11\r
- vshufps $ 0xb1, %ymm13, %ymm13, %ymm13\r
- vshufps $ 0xb1, %ymm15, %ymm15, %ymm15\r
-\r
-#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \\r
- defined(NR) || defined(NC) || defined(TR) || defined(TC)\r
-\r
- vaddsubps %ymm9, %ymm8 , %ymm8\r
- vaddsubps %ymm11,%ymm10, %ymm10\r
- vaddsubps %ymm13,%ymm12, %ymm12\r
- vaddsubps %ymm15,%ymm14, %ymm14\r
-\r
- vshufps $ 0xb1, %ymm8 , %ymm8, %ymm9\r
- vshufps $ 0xb1, %ymm10, %ymm10, %ymm11\r
- vshufps $ 0xb1, %ymm12, %ymm12, %ymm13\r
- vshufps $ 0xb1, %ymm14, %ymm14, %ymm15\r
-\r
-#else\r
- vaddsubps %ymm8, %ymm9 ,%ymm9\r
- vaddsubps %ymm10, %ymm11,%ymm11\r
- vaddsubps %ymm12, %ymm13,%ymm13\r
- vaddsubps %ymm14, %ymm15,%ymm15\r
-\r
- vmovaps %ymm9, %ymm8\r
- vmovaps %ymm11, %ymm10\r
- vmovaps %ymm13, %ymm12\r
- vmovaps %ymm15, %ymm14\r
-\r
- // swap high and low 64 bytes\r
- vshufps $ 0xb1, %ymm9 , %ymm9, %ymm9\r
- vshufps $ 0xb1, %ymm11, %ymm11, %ymm11\r
- vshufps $ 0xb1, %ymm13, %ymm13, %ymm13\r
- vshufps $ 0xb1, %ymm15, %ymm15, %ymm15\r
-\r
-#endif\r
-\r
- // multiply with ALPHA_R\r
- vmulps %ymm8 , %ymm0, %ymm8\r
- vmulps %ymm10, %ymm0, %ymm10\r
- vmulps %ymm12, %ymm0, %ymm12\r
- vmulps %ymm14, %ymm0, %ymm14\r
-\r
- // multiply with ALPHA_I\r
- vmulps %ymm9 , %ymm1, %ymm9\r
- vmulps %ymm11, %ymm1, %ymm11\r
- vmulps %ymm13, %ymm1, %ymm13\r
- vmulps %ymm15, %ymm1, %ymm15\r
-\r
- vaddsubps %ymm9, %ymm8 , %ymm8\r
- vaddsubps %ymm11,%ymm10, %ymm10\r
- vaddsubps %ymm13,%ymm12, %ymm12\r
- vaddsubps %ymm15,%ymm14, %ymm14\r
-\r
-\r
-\r
-#ifndef TRMMKERNEL\r
-\r
- vaddps (CO1), %ymm8 , %ymm8\r
- vaddps 8 * SIZE(CO1), %ymm12, %ymm12\r
-\r
- vaddps (CO1, LDC), %ymm10, %ymm10\r
- vaddps 8 * SIZE(CO1, LDC), %ymm14, %ymm14\r
-\r
-#endif\r
-\r
- vmovups %ymm8 , (CO1)\r
- vmovups %ymm12 , 8 * SIZE(CO1)\r
-\r
- vmovups %ymm10 , (CO1, LDC)\r
- vmovups %ymm14 , 8 * SIZE(CO1, LDC)\r
-\r
- prefetcht0 64(CO1)\r
- prefetcht0 64(CO1, LDC)\r
-\r
-.endm\r
-\r
-/***************************************************************************************************************************/\r
-\r
-.macro KERNEL4x2_SUB\r
- vmovups -16 * SIZE(AO, %rax, SIZE), %xmm0\r
- vbroadcastss -8 * SIZE(BO, BI, SIZE), %xmm4\r
- VFMADDPS_R( %xmm8,%xmm4,%xmm0 )\r
- vmovups -12 * SIZE(AO, %rax, SIZE), %xmm1\r
- VFMADDPS_R( %xmm12,%xmm4,%xmm1 )\r
- vbroadcastss -7 * SIZE(BO, BI, SIZE), %xmm5\r
- VFMADDPS_I( %xmm9,%xmm5,%xmm0 )\r
- VFMADDPS_I( %xmm13,%xmm5,%xmm1 )\r
- vbroadcastss -6 * SIZE(BO, BI, SIZE), %xmm6\r
- VFMADDPS_R( %xmm10,%xmm6,%xmm0 )\r
- VFMADDPS_R( %xmm14,%xmm6,%xmm1 )\r
- vbroadcastss -5 * SIZE(BO, BI, SIZE), %xmm7\r
- VFMADDPS_I( %xmm11,%xmm7,%xmm0 )\r
- VFMADDPS_I( %xmm15,%xmm7,%xmm1 )\r
- addq $ 4, BI \r
- addq $ 8, %rax \r
-.endm\r
-\r
-.macro SAVE4x2\r
-\r
- vbroadcastss ALPHA_R, %xmm0\r
- vbroadcastss ALPHA_I, %xmm1\r
-\r
- // swap high and low 64 bytes\r
- vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9\r
- vshufps $ 0xb1, %xmm11, %xmm11, %xmm11\r
- vshufps $ 0xb1, %xmm13, %xmm13, %xmm13\r
- vshufps $ 0xb1, %xmm15, %xmm15, %xmm15\r
-\r
-#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \\r
- defined(NR) || defined(NC) || defined(TR) || defined(TC)\r
-\r
- vaddsubps %xmm9, %xmm8 , %xmm8\r
- vaddsubps %xmm11,%xmm10, %xmm10\r
- vaddsubps %xmm13,%xmm12, %xmm12\r
- vaddsubps %xmm15,%xmm14, %xmm14\r
-\r
- vshufps $ 0xb1, %xmm8 , %xmm8, %xmm9\r
- vshufps $ 0xb1, %xmm10, %xmm10, %xmm11\r
- vshufps $ 0xb1, %xmm12, %xmm12, %xmm13\r
- vshufps $ 0xb1, %xmm14, %xmm14, %xmm15\r
-\r
-#else\r
- vaddsubps %xmm8, %xmm9 ,%xmm9\r
- vaddsubps %xmm10, %xmm11,%xmm11\r
- vaddsubps %xmm12, %xmm13,%xmm13\r
- vaddsubps %xmm14, %xmm15,%xmm15\r
-\r
- vmovaps %xmm9, %xmm8\r
- vmovaps %xmm11, %xmm10\r
- vmovaps %xmm13, %xmm12\r
- vmovaps %xmm15, %xmm14\r
-\r
- // swap high and low 64 bytes\r
- vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9\r
- vshufps $ 0xb1, %xmm11, %xmm11, %xmm11\r
- vshufps $ 0xb1, %xmm13, %xmm13, %xmm13\r
- vshufps $ 0xb1, %xmm15, %xmm15, %xmm15\r
-\r
-#endif\r
-\r
- // multiply with ALPHA_R\r
- vmulps %xmm8 , %xmm0, %xmm8\r
- vmulps %xmm10, %xmm0, %xmm10\r
- vmulps %xmm12, %xmm0, %xmm12\r
- vmulps %xmm14, %xmm0, %xmm14\r
-\r
- // multiply with ALPHA_I\r
- vmulps %xmm9 , %xmm1, %xmm9\r
- vmulps %xmm11, %xmm1, %xmm11\r
- vmulps %xmm13, %xmm1, %xmm13\r
- vmulps %xmm15, %xmm1, %xmm15\r
-\r
- vaddsubps %xmm9, %xmm8 , %xmm8\r
- vaddsubps %xmm11,%xmm10, %xmm10\r
- vaddsubps %xmm13,%xmm12, %xmm12\r
- vaddsubps %xmm15,%xmm14, %xmm14\r
-\r
-#ifndef TRMMKERNEL\r
-\r
- vaddps (CO1), %xmm8 , %xmm8\r
- vaddps 4 * SIZE(CO1), %xmm12, %xmm12\r
-\r
- vaddps (CO1, LDC), %xmm10, %xmm10\r
- vaddps 4 * SIZE(CO1, LDC), %xmm14, %xmm14\r
-\r
-#endif\r
-\r
- vmovups %xmm8 , (CO1)\r
- vmovups %xmm12 , 4 * SIZE(CO1)\r
-\r
- vmovups %xmm10 , (CO1, LDC)\r
- vmovups %xmm14 , 4 * SIZE(CO1, LDC)\r
-\r
-.endm\r
-\r
-/************************************************************************************************/\r
-\r
-.macro KERNEL2x2_SUB\r
- vmovups -16 * SIZE(AO, %rax, SIZE), %xmm0\r
- vbroadcastss -8 * SIZE(BO, BI, SIZE), %xmm4\r
- VFMADDPS_R( %xmm8,%xmm4,%xmm0 )\r
- vbroadcastss -7 * SIZE(BO, BI, SIZE), %xmm5\r
- VFMADDPS_I( %xmm9,%xmm5,%xmm0 )\r
- vbroadcastss -6 * SIZE(BO, BI, SIZE), %xmm6\r
- VFMADDPS_R( %xmm10,%xmm6,%xmm0 )\r
- vbroadcastss -5 * SIZE(BO, BI, SIZE), %xmm7\r
- VFMADDPS_I( %xmm11,%xmm7,%xmm0 )\r
- addq $ 4, BI \r
- addq $ 4, %rax \r
-.endm\r
-\r
-.macro SAVE2x2\r
-\r
- vbroadcastss ALPHA_R, %xmm0\r
- vbroadcastss ALPHA_I, %xmm1\r
-\r
- // swap high and low 4 bytes\r
- vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9\r
- vshufps $ 0xb1, %xmm11, %xmm11, %xmm11\r
-\r
-#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \\r
- defined(NR) || defined(NC) || defined(TR) || defined(TC)\r
-\r
- vaddsubps %xmm9, %xmm8 , %xmm8\r
- vaddsubps %xmm11,%xmm10, %xmm10\r
-\r
- vshufps $ 0xb1, %xmm8 , %xmm8, %xmm9\r
- vshufps $ 0xb1, %xmm10, %xmm10, %xmm11\r
-\r
-#else\r
- vaddsubps %xmm8, %xmm9 ,%xmm9\r
- vaddsubps %xmm10, %xmm11,%xmm11\r
-\r
- vmovaps %xmm9, %xmm8\r
- vmovaps %xmm11, %xmm10\r
-\r
- // swap high and low 4 bytes\r
- vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9\r
- vshufps $ 0xb1, %xmm11, %xmm11, %xmm11\r
-\r
-#endif\r
-\r
- // multiply with ALPHA_R\r
- vmulps %xmm8 , %xmm0, %xmm8\r
- vmulps %xmm10, %xmm0, %xmm10\r
-\r
- // multiply with ALPHA_I\r
- vmulps %xmm9 , %xmm1, %xmm9\r
- vmulps %xmm11, %xmm1, %xmm11\r
-\r
- vaddsubps %xmm9, %xmm8 , %xmm8\r
- vaddsubps %xmm11,%xmm10, %xmm10\r
-\r
-#ifndef TRMMKERNEL\r
-\r
- vaddps (CO1), %xmm8 , %xmm8\r
-\r
- vaddps (CO1, LDC), %xmm10, %xmm10\r
-\r
-#endif\r
-\r
- vmovups %xmm8 , (CO1)\r
-\r
- vmovups %xmm10 , (CO1, LDC)\r
-\r
-.endm\r
-\r
-/************************************************************************************************/\r
-\r
-.macro KERNEL1x2_SUB\r
- vmovsd -16 * SIZE(AO, %rax, SIZE), %xmm0\r
- vbroadcastss -8 * SIZE(BO, BI, SIZE), %xmm4\r
- VFMADDPS_R( %xmm8,%xmm4,%xmm0 )\r
- vbroadcastss -7 * SIZE(BO, BI, SIZE), %xmm5\r
- VFMADDPS_I( %xmm9,%xmm5,%xmm0 )\r
- vbroadcastss -6 * SIZE(BO, BI, SIZE), %xmm6\r
- VFMADDPS_R( %xmm10,%xmm6,%xmm0 )\r
- vbroadcastss -5 * SIZE(BO, BI, SIZE), %xmm7\r
- VFMADDPS_I( %xmm11,%xmm7,%xmm0 )\r
- addq $ 4, BI \r
- addq $ 2, %rax \r
-.endm\r
-\r
-.macro SAVE1x2\r
-\r
- vbroadcastss ALPHA_R, %xmm0\r
- vbroadcastss ALPHA_I, %xmm1\r
-\r
- // swap high and low 64 bytes\r
- vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9\r
- vshufps $ 0xb1, %xmm11, %xmm11, %xmm11\r
-\r
-#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \\r
- defined(NR) || defined(NC) || defined(TR) || defined(TC)\r
-\r
- vaddsubps %xmm9, %xmm8 , %xmm8\r
- vaddsubps %xmm11,%xmm10, %xmm10\r
-\r
- vshufps $ 0xb1, %xmm8 , %xmm8, %xmm9\r
- vshufps $ 0xb1, %xmm10, %xmm10, %xmm11\r
-\r
-#else\r
- vaddsubps %xmm8, %xmm9 ,%xmm9\r
- vaddsubps %xmm10, %xmm11,%xmm11\r
-\r
- vmovaps %xmm9, %xmm8\r
- vmovaps %xmm11, %xmm10\r
-\r
- // swap high and low 64 bytes\r
- vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9\r
- vshufps $ 0xb1, %xmm11, %xmm11, %xmm11\r
-\r
-#endif\r
-\r
- // multiply with ALPHA_R\r
- vmulps %xmm8 , %xmm0, %xmm8\r
- vmulps %xmm10, %xmm0, %xmm10\r
-\r
- // multiply with ALPHA_I\r
- vmulps %xmm9 , %xmm1, %xmm9\r
- vmulps %xmm11, %xmm1, %xmm11\r
-\r
- vaddsubps %xmm9, %xmm8 , %xmm8\r
- vaddsubps %xmm11,%xmm10, %xmm10\r
-\r
-#ifndef TRMMKERNEL\r
-\r
- vmovsd (CO1), %xmm14\r
- vaddps %xmm14, %xmm8 , %xmm8\r
-\r
- vmovsd (CO1, LDC), %xmm15\r
- vaddps %xmm15, %xmm10, %xmm10\r
-\r
-#endif\r
-\r
- vmovsd %xmm8 , (CO1)\r
- vmovsd %xmm10 , (CO1, LDC)\r
-\r
-.endm\r
-\r
-/************************************************************************************************/\r
-\r
-.macro KERNEL8x1_SUB\r
- vmovups -16 * SIZE(AO, %rax, SIZE), %ymm0\r
- vmovups -8 * SIZE(AO, %rax, SIZE), %ymm1\r
- vbroadcastss -4 * SIZE(BO, BI, SIZE), %ymm4\r
- VFMADDPS_R( %ymm8,%ymm4,%ymm0 )\r
- VFMADDPS_R( %ymm12,%ymm4,%ymm1 )\r
- vbroadcastss -3 * SIZE(BO, BI, SIZE), %ymm5\r
- VFMADDPS_I( %ymm9,%ymm5,%ymm0 )\r
- VFMADDPS_I( %ymm13,%ymm5,%ymm1 )\r
- addq $ 2 , BI \r
- addq $ 16, %rax \r
-.endm\r
-\r
-.macro SAVE8x1\r
-\r
- vbroadcastss ALPHA_R, %ymm0\r
- vbroadcastss ALPHA_I, %ymm1\r
-\r
- // swap high and low 64 bytes\r
- vshufps $ 0xb1, %ymm9 , %ymm9, %ymm9\r
- vshufps $ 0xb1, %ymm13, %ymm13, %ymm13\r
-\r
-#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \\r
- defined(NR) || defined(NC) || defined(TR) || defined(TC)\r
-\r
- vaddsubps %ymm9, %ymm8 , %ymm8\r
- vaddsubps %ymm13,%ymm12, %ymm12\r
-\r
- vshufps $ 0xb1, %ymm8 , %ymm8, %ymm9\r
- vshufps $ 0xb1, %ymm12, %ymm12, %ymm13\r
-\r
-#else\r
- vaddsubps %ymm8, %ymm9 ,%ymm9\r
- vaddsubps %ymm12, %ymm13,%ymm13\r
-\r
- vmovaps %ymm9, %ymm8\r
- vmovaps %ymm13, %ymm12\r
-\r
- // swap high and low 64 bytes\r
- vshufps $ 0xb1, %ymm9 , %ymm9, %ymm9\r
- vshufps $ 0xb1, %ymm13, %ymm13, %ymm13\r
-\r
-#endif\r
-\r
- // multiply with ALPHA_R\r
- vmulps %ymm8 , %ymm0, %ymm8\r
- vmulps %ymm12, %ymm0, %ymm12\r
-\r
- // multiply with ALPHA_I\r
- vmulps %ymm9 , %ymm1, %ymm9\r
- vmulps %ymm13, %ymm1, %ymm13\r
-\r
- vaddsubps %ymm9, %ymm8 , %ymm8\r
- vaddsubps %ymm13,%ymm12, %ymm12\r
-\r
-\r
-\r
-#ifndef TRMMKERNEL\r
-\r
- vaddps (CO1), %ymm8 , %ymm8\r
- vaddps 8 * SIZE(CO1), %ymm12, %ymm12\r
-\r
-#endif\r
-\r
- vmovups %ymm8 , (CO1)\r
- vmovups %ymm12 , 8 * SIZE(CO1)\r
-\r
-.endm\r
-\r
-\r
-/************************************************************************************************/\r
-\r
-.macro KERNEL4x1_SUB\r
- vmovups -16 * SIZE(AO, %rax, SIZE), %xmm0\r
- vbroadcastss -4 * SIZE(BO, BI, SIZE), %xmm4\r
- VFMADDPS_R( %xmm8,%xmm4,%xmm0 )\r
- vmovups -12 * SIZE(AO, %rax, SIZE), %xmm1\r
- VFMADDPS_R( %xmm12,%xmm4,%xmm1 )\r
- vbroadcastss -3 * SIZE(BO, BI, SIZE), %xmm5\r
- VFMADDPS_I( %xmm9,%xmm5,%xmm0 )\r
- VFMADDPS_I( %xmm13,%xmm5,%xmm1 )\r
- addq $ 2, BI \r
- addq $ 8, %rax \r
-.endm\r
-\r
-.macro SAVE4x1\r
-\r
- vbroadcastss ALPHA_R, %xmm0\r
- vbroadcastss ALPHA_I, %xmm1\r
-\r
- // swap high and low 4 bytes\r
- vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9\r
- vshufps $ 0xb1, %xmm13, %xmm13, %xmm13\r
-\r
-#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \\r
- defined(NR) || defined(NC) || defined(TR) || defined(TC)\r
-\r
- vaddsubps %xmm9, %xmm8 , %xmm8\r
- vaddsubps %xmm13,%xmm12, %xmm12\r
-\r
- vshufps $ 0xb1, %xmm8 , %xmm8, %xmm9\r
- vshufps $ 0xb1, %xmm12, %xmm12, %xmm13\r
-\r
-#else\r
- vaddsubps %xmm8, %xmm9 ,%xmm9\r
- vaddsubps %xmm12, %xmm13,%xmm13\r
-\r
- vmovaps %xmm9, %xmm8\r
- vmovaps %xmm13, %xmm12\r
-\r
- // swap high and low 4 bytes\r
- vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9\r
- vshufps $ 0xb1, %xmm13, %xmm13, %xmm13\r
-\r
-#endif\r
-\r
- // multiply with ALPHA_R\r
- vmulps %xmm8 , %xmm0, %xmm8\r
- vmulps %xmm12, %xmm0, %xmm12\r
-\r
- // multiply with ALPHA_I\r
- vmulps %xmm9 , %xmm1, %xmm9\r
- vmulps %xmm13, %xmm1, %xmm13\r
-\r
- vaddsubps %xmm9, %xmm8 , %xmm8\r
- vaddsubps %xmm13,%xmm12, %xmm12\r
-\r
-#ifndef TRMMKERNEL\r
-\r
- vaddps (CO1), %xmm8 , %xmm8\r
- vaddps 4 * SIZE(CO1), %xmm12, %xmm12\r
-\r
-#endif\r
-\r
- vmovups %xmm8 , (CO1)\r
- vmovups %xmm12 , 4 * SIZE(CO1)\r
-\r
-.endm\r
-\r
-/************************************************************************************************/\r
-\r
-.macro KERNEL2x1_SUB\r
- vmovups -16 * SIZE(AO, %rax, SIZE), %xmm0\r
- vbroadcastss -4 * SIZE(BO, BI, SIZE), %xmm4\r
- VFMADDPS_R( %xmm8,%xmm4,%xmm0 )\r
- vbroadcastss -3 * SIZE(BO, BI, SIZE), %xmm5\r
- VFMADDPS_I( %xmm9,%xmm5,%xmm0 )\r
- addq $ 2, BI \r
- addq $ 4, %rax \r
-.endm\r
-\r
-.macro SAVE2x1\r
-\r
- vbroadcastss ALPHA_R, %xmm0\r
- vbroadcastss ALPHA_I, %xmm1\r
-\r
- // swap high and low 64 bytes\r
- vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9\r
-\r
-#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \\r
- defined(NR) || defined(NC) || defined(TR) || defined(TC)\r
-\r
- vaddsubps %xmm9, %xmm8 , %xmm8\r
-\r
- vshufps $ 0xb1, %xmm8 , %xmm8, %xmm9\r
-\r
-#else\r
- vaddsubps %xmm8, %xmm9 ,%xmm9\r
-\r
- vmovaps %xmm9, %xmm8\r
-\r
- // swap high and low 64 bytes\r
- vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9\r
-\r
-#endif\r
-\r
- // multiply with ALPHA_R\r
- vmulps %xmm8 , %xmm0, %xmm8\r
-\r
- // multiply with ALPHA_I\r
- vmulps %xmm9 , %xmm1, %xmm9\r
-\r
- vaddsubps %xmm9, %xmm8 , %xmm8\r
-\r
-#ifndef TRMMKERNEL\r
-\r
- vaddps (CO1), %xmm8 , %xmm8\r
-\r
-#endif\r
-\r
- vmovups %xmm8 , (CO1)\r
-\r
-.endm\r
-\r
-/************************************************************************************************/\r
-\r
-.macro KERNEL1x1_SUB\r
- vmovsd -16 * SIZE(AO, %rax, SIZE), %xmm0\r
- vbroadcastss -4 * SIZE(BO, BI, SIZE), %xmm4\r
- VFMADDPS_R( %xmm8,%xmm4,%xmm0 )\r
- vbroadcastss -3 * SIZE(BO, BI, SIZE), %xmm5\r
- VFMADDPS_I( %xmm9,%xmm5,%xmm0 )\r
- addq $ 2, BI \r
- addq $ 2, %rax \r
-.endm\r
-\r
-.macro SAVE1x1\r
-\r
- vbroadcastss ALPHA_R, %xmm0\r
- vbroadcastss ALPHA_I, %xmm1\r
-\r
- // swap high and low 64 bytes\r
- vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9\r
-\r
-#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \\r
- defined(NR) || defined(NC) || defined(TR) || defined(TC)\r
-\r
- vaddsubps %xmm9, %xmm8 , %xmm8\r
-\r
- vshufps $ 0xb1, %xmm8 , %xmm8, %xmm9\r
-\r
-#else\r
- vaddsubps %xmm8, %xmm9 ,%xmm9\r
-\r
- vmovaps %xmm9, %xmm8\r
-\r
- // swap high and low 64 bytes\r
- vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9\r
-\r
-#endif\r
-\r
- // multiply with ALPHA_R\r
- vmulps %xmm8 , %xmm0, %xmm8\r
-\r
- // multiply with ALPHA_I\r
- vmulps %xmm9 , %xmm1, %xmm9\r
-\r
- vaddsubps %xmm9, %xmm8 , %xmm8\r
-\r
-#ifndef TRMMKERNEL\r
-\r
- vmovsd (CO1), %xmm14\r
- vaddps %xmm14, %xmm8 , %xmm8\r
-\r
-#endif\r
-\r
- vmovsd %xmm8 , (CO1)\r
-\r
-.endm\r
-\r
-/************************************************************************************************/\r
-\r
-\r
-\r
-\r
- PROLOGUE\r
- PROFCODE\r
- \r
- subq $ STACKSIZE, %rsp\r
- movq %rbx, (%rsp)\r
- movq %rbp, 8(%rsp)\r
- movq %r12, 16(%rsp)\r
- movq %r13, 24(%rsp)\r
- movq %r14, 32(%rsp)\r
- movq %r15, 40(%rsp)\r
-\r
- vzeroupper\r
-\r
-#ifdef WINDOWS_ABI\r
- movq %rdi, 48(%rsp)\r
- movq %rsi, 56(%rsp)\r
- vmovups %xmm6, 64(%rsp)\r
- vmovups %xmm7, 80(%rsp)\r
- vmovups %xmm8, 96(%rsp)\r
- vmovups %xmm9, 112(%rsp)\r
- vmovups %xmm10, 128(%rsp)\r
- vmovups %xmm11, 144(%rsp)\r
- vmovups %xmm12, 160(%rsp)\r
- vmovups %xmm13, 176(%rsp)\r
- vmovups %xmm14, 192(%rsp)\r
- vmovups %xmm15, 208(%rsp)\r
-\r
- movq ARG1, OLD_M\r
- movq ARG2, OLD_N\r
- movq ARG3, OLD_K\r
- movq OLD_A, A\r
- movq OLD_B, B\r
- movq OLD_C, C\r
- movq OLD_LDC, LDC\r
-#ifdef TRMMKERNEL\r
- movsd OLD_OFFSET, %xmm12\r
-#endif\r
- vmovaps %xmm3, %xmm0\r
- vmovsd OLD_ALPHA_I, %xmm1\r
-\r
-#else\r
- movq STACKSIZE + 8(%rsp), LDC\r
-#ifdef TRMMKERNEL\r
- movsd STACKSIZE + 16(%rsp), %xmm12\r
-#endif\r
-\r
-#endif\r
-\r
- movq %rsp, SP # save old stack\r
- subq $ 128 + L_BUFFER_SIZE, %rsp\r
- andq $ -4096, %rsp # align stack\r
-\r
- STACK_TOUCH\r
-\r
- cmpq $ 0, OLD_M\r
- je .L999\r
-\r
- cmpq $ 0, OLD_N\r
- je .L999\r
-\r
- cmpq $ 0, OLD_K\r
- je .L999\r
-\r
- movq OLD_M, M\r
- movq OLD_N, N\r
- movq OLD_K, K\r
-\r
- vmovss %xmm0, ALPHA_R\r
- vmovss %xmm1, ALPHA_I\r
-\r
- salq $ ZBASE_SHIFT, LDC\r
-\r
- movq N, %rax\r
- xorq %rdx, %rdx\r
- movq $ 2, %rdi\r
- divq %rdi // N / 2\r
- movq %rax, Ndiv6 // N / 2\r
- movq %rdx, Nmod6 // N % 2\r
-\r
- \r
-\r
-#ifdef TRMMKERNEL\r
- vmovsd %xmm12, OFFSET\r
- vmovsd %xmm12, KK\r
-#ifndef LEFT\r
- negq KK\r
-#endif \r
-#endif\r
-\r
-.L2_0:\r
-\r
- movq Ndiv6, J\r
- cmpq $ 0, J\r
- je .L1_0\r
- ALIGN_4\r
-\r
-\r
-\r
-.L2_01:\r
- // copy to sub buffer\r
- movq B, BO1\r
- leaq BUFFER1, BO // first buffer to BO\r
- movq K, %rax\r
- ALIGN_4\r
-\r
-.L2_02b:\r
-\r
- vmovups (BO1), %xmm0\r
- vmovups %xmm0, (BO)\r
- addq $ 4*SIZE,BO1\r
- addq $ 4*SIZE,BO\r
- decq %rax\r
- jnz .L2_02b\r
-\r
-.L2_02c:\r
-\r
- movq BO1, B // next offset of B\r
-\r
-.L2_10:\r
- movq C, CO1\r
- leaq (C, LDC, 2), C // c += 2 * ldc\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- movq OFFSET, %rax\r
- movq %rax, KK\r
-#endif\r
- \r
- movq A, AO // aoffset = a\r
- addq $ 16 * SIZE, AO\r
-\r
- movq M, I\r
- sarq $ 3, I // i = (m >> 3)\r
- je .L2_4_10\r
-\r
- ALIGN_4\r
-/**********************************************************************************************************/\r
-\r
-.L2_8_11:\r
-\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $ 8 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $ 8 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (,BI,4), BI // BI = BI * 4 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $ 4, %rax // rax = rax *16 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
- vzeroall\r
-\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
- movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $ 8, %rax // number of values in AO\r
-#else\r
- addq $ 2, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
-\r
-\r
- andq $ -8, %rax // K = K - ( K % 8 )\r
- je .L2_8_16\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,4), BI // BI = BI * 4 ; number of values\r
-\r
- salq $ 4, %rax // rax = rax *16 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
- leaq (BO, BI, SIZE), BO\r
- negq BI\r
- negq %rax\r
- ALIGN_4\r
-\r
-.L2_8_12:\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL8x2_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x2_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x2_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x2_SUB\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL8x2_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x2_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x2_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x2_SUB\r
-\r
- je .L2_8_16\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL8x2_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x2_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x2_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x2_SUB\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL8x2_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x2_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x2_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x2_SUB\r
-\r
- je .L2_8_16\r
-\r
- jmp .L2_8_12\r
- ALIGN_4\r
-\r
-.L2_8_16:\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
-\r
- andq $ 7, %rax # if (k & 1)\r
- je .L2_8_19\r
-\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,4), BI // BI = BI * 4 ; number of values\r
-\r
- salq $ 4, %rax // rax = rax *16 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
- leaq (BO, BI, SIZE), BO\r
- negq BI\r
- negq %rax\r
- ALIGN_4\r
-\r
-.L2_8_17:\r
-\r
- KERNEL8x2_SUB\r
-\r
- jl .L2_8_17\r
- ALIGN_4\r
-\r
-\r
-.L2_8_19:\r
-\r
- SAVE8x2\r
-\r
-\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax\r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,4), BI // BI = BI * 4 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $ 4, %rax // rax = rax *16 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $ 8, KK\r
-#endif\r
-\r
- addq $ 16 * SIZE, CO1 # coffset += 16\r
- decq I # i --\r
- jg .L2_8_11\r
- ALIGN_4 \r
-\r
-\r
-/**********************************************************************************************************/\r
-\r
-\r
-\r
-\r
-.L2_4_10:\r
- testq $ 7, M \r
- jz .L2_4_60 // to next 2 lines of N\r
-\r
- testq $ 4, M \r
- jz .L2_4_20\r
- ALIGN_4\r
-\r
-\r
-.L2_4_11:\r
-\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $ 8 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $ 8 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (,BI,4), BI // BI = BI * 4 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $ 3, %rax // rax = rax * 8 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
- vzeroall\r
-\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
- movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $ 4, %rax // number of values in AO\r
-#else\r
- addq $ 2, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
-\r
-\r
- andq $ -8, %rax // K = K - ( K % 8 )\r
- je .L2_4_16\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,4), BI // BI = BI * 4 ; number of values\r
-\r
- salq $ 3, %rax // rax = rax * 8 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
- leaq (BO, BI, SIZE), BO\r
- negq BI\r
- negq %rax\r
- ALIGN_4\r
-\r
-.L2_4_12:\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL4x2_SUB\r
- KERNEL4x2_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL4x2_SUB\r
- KERNEL4x2_SUB\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL4x2_SUB\r
- KERNEL4x2_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL4x2_SUB\r
- KERNEL4x2_SUB\r
-\r
- je .L2_4_16\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL4x2_SUB\r
- KERNEL4x2_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL4x2_SUB\r
- KERNEL4x2_SUB\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL4x2_SUB\r
- KERNEL4x2_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL4x2_SUB\r
- KERNEL4x2_SUB\r
-\r
- je .L2_4_16\r
-\r
- jmp .L2_4_12\r
- ALIGN_4\r
-\r
-.L2_4_16:\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
-\r
- andq $ 7, %rax # if (k & 1)\r
- je .L2_4_19\r
-\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,4), BI // BI = BI * 4 ; number of values\r
-\r
- salq $ 3, %rax // rax = rax * 8 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
- leaq (BO, BI, SIZE), BO\r
- negq BI\r
- negq %rax\r
- ALIGN_4\r
-\r
-.L2_4_17:\r
-\r
- KERNEL4x2_SUB\r
-\r
- jl .L2_4_17\r
- ALIGN_4\r
-\r
-\r
-.L2_4_19:\r
-\r
- SAVE4x2\r
-\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax\r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,4), BI // BI = BI * 4 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $ 3, %rax // rax = rax * 8 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $ 4, KK\r
-#endif\r
-\r
- addq $ 8 * SIZE, CO1 # coffset += 8\r
- ALIGN_4 \r
-\r
-\r
-\r
-/**************************************************************************\r
-* Rest of M \r
-***************************************************************************/\r
-\r
-.L2_4_20:\r
-\r
- testq $ 2, M \r
- jz .L2_4_40\r
- ALIGN_4\r
-\r
-.L2_4_21:\r
-\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $ 8 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $ 8 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (,BI,4), BI // BI = BI * 4 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $ 2, %rax // rax = rax * 4 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
- vzeroall\r
-\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
- movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $ 2, %rax // number of values in AO\r
-#else\r
- addq $ 2, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
-\r
-\r
- andq $ -8, %rax // K = K - ( K % 8 )\r
- je .L2_4_26\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,4), BI // BI = BI * 4 ; number of values\r
-\r
- salq $ 2, %rax // rax = rax * 4 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
- leaq (BO, BI, SIZE), BO\r
- negq BI\r
- negq %rax\r
- ALIGN_4\r
-\r
-.L2_4_22:\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL2x2_SUB\r
- KERNEL2x2_SUB\r
- KERNEL2x2_SUB\r
- KERNEL2x2_SUB\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL2x2_SUB\r
- KERNEL2x2_SUB\r
- KERNEL2x2_SUB\r
- KERNEL2x2_SUB\r
-\r
- je .L2_4_26\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL2x2_SUB\r
- KERNEL2x2_SUB\r
- KERNEL2x2_SUB\r
- KERNEL2x2_SUB\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL2x2_SUB\r
- KERNEL2x2_SUB\r
- KERNEL2x2_SUB\r
- KERNEL2x2_SUB\r
-\r
- je .L2_4_26\r
-\r
- jmp .L2_4_22\r
- ALIGN_4\r
-\r
-.L2_4_26:\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
-\r
- andq $ 7, %rax # if (k & 1)\r
- je .L2_4_29\r
-\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,4), BI // BI = BI * 4 ; number of values\r
-\r
- salq $ 2, %rax // rax = rax * 4 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
- leaq (BO, BI, SIZE), BO\r
- negq BI\r
- negq %rax\r
- ALIGN_4\r
-\r
-.L2_4_27:\r
-\r
- KERNEL2x2_SUB\r
-\r
- jl .L2_4_27\r
- ALIGN_4\r
-\r
-\r
-.L2_4_29:\r
-\r
- vbroadcastss ALPHA_R, %xmm0\r
- vbroadcastss ALPHA_I, %xmm1\r
-\r
- // swap high and low 64 bytes\r
- vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9\r
- vshufps $ 0xb1, %xmm11, %xmm11, %xmm11\r
-\r
-#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \\r
- defined(NR) || defined(NC) || defined(TR) || defined(TC)\r
-\r
- vaddsubps %xmm9, %xmm8 , %xmm8\r
- vaddsubps %xmm11,%xmm10, %xmm10\r
-\r
- vshufps $ 0xb1, %xmm8 , %xmm8, %xmm9\r
- vshufps $ 0xb1, %xmm10, %xmm10, %xmm11\r
-\r
-#else\r
- vaddsubps %xmm8, %xmm9 ,%xmm9\r
- vaddsubps %xmm10, %xmm11,%xmm11\r
-\r
- vmovaps %xmm9, %xmm8\r
- vmovaps %xmm11, %xmm10\r
-\r
- // swap high and low 64 bytes\r
- vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9\r
- vshufps $ 0xb1, %xmm11, %xmm11, %xmm11\r
-\r
-#endif\r
-\r
- // multiply with ALPHA_R\r
- vmulps %xmm8 , %xmm0, %xmm8\r
- vmulps %xmm10, %xmm0, %xmm10\r
-\r
- // multiply with ALPHA_I\r
- vmulps %xmm9 , %xmm1, %xmm9\r
- vmulps %xmm11, %xmm1, %xmm11\r
-\r
- vaddsubps %xmm9, %xmm8 , %xmm8\r
- vaddsubps %xmm11,%xmm10, %xmm10\r
-\r
-\r
-\r
-#ifndef TRMMKERNEL\r
-\r
- vaddps (CO1), %xmm8 , %xmm8\r
-\r
- vaddps (CO1, LDC), %xmm10, %xmm10\r
-\r
-#endif\r
-\r
- vmovups %xmm8 , (CO1)\r
-\r
- vmovups %xmm10 , (CO1, LDC)\r
-\r
-\r
-\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax\r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,4), BI // BI = BI * 4 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $ 2, %rax // rax = rax * 4 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $ 2, KK\r
-#endif\r
-\r
- addq $ 4 * SIZE, CO1 # coffset += 4\r
- decq I # i --\r
- jg .L2_4_21\r
- ALIGN_4 \r
-\r
-\r
-\r
-/**************************************************************************/\r
-.L2_4_40:\r
- testq $ 1, M \r
- jz .L2_4_60 // to next 2 lines of N\r
-\r
- ALIGN_4\r
-\r
-.L2_4_41:\r
-\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $ 8 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $ 8 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (,BI,4), BI // BI = BI * 4 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $ 1, %rax // rax = rax * 2 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
- vzeroall\r
-\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
- movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $ 1, %rax // number of values in AO\r
-#else\r
- addq $ 2, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
-\r
-\r
- andq $ -8, %rax // K = K - ( K % 8 )\r
- je .L2_4_46\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,4), BI // BI = BI * 4 ; number of values\r
-\r
- salq $ 1, %rax // rax = rax * 2 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
- leaq (BO, BI, SIZE), BO\r
- negq BI\r
- negq %rax\r
- ALIGN_4\r
-\r
-.L2_4_42:\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL1x2_SUB\r
- KERNEL1x2_SUB\r
- KERNEL1x2_SUB\r
- KERNEL1x2_SUB\r
-\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL1x2_SUB\r
- KERNEL1x2_SUB\r
- KERNEL1x2_SUB\r
- KERNEL1x2_SUB\r
-\r
- je .L2_4_46\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL1x2_SUB\r
- KERNEL1x2_SUB\r
- KERNEL1x2_SUB\r
- KERNEL1x2_SUB\r
-\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL1x2_SUB\r
- KERNEL1x2_SUB\r
- KERNEL1x2_SUB\r
- KERNEL1x2_SUB\r
-\r
- je .L2_4_46\r
-\r
- jmp .L2_4_42\r
- ALIGN_4\r
-\r
-.L2_4_46:\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
-\r
- andq $ 7, %rax # if (k & 1)\r
- je .L2_4_49\r
-\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,4), BI // BI = BI * 4 ; number of values\r
-\r
- salq $ 1, %rax // rax = rax * 2 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
- leaq (BO, BI, SIZE), BO\r
- negq BI\r
- negq %rax\r
- ALIGN_4\r
-\r
-.L2_4_47:\r
-\r
- KERNEL1x2_SUB\r
-\r
- jl .L2_4_47\r
- ALIGN_4\r
-\r
-\r
-.L2_4_49:\r
-\r
- SAVE1x2\r
-\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax\r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,4), BI // BI = BI * 4 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $ 1, %rax // rax = rax * 2 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $ 1, KK\r
-#endif\r
-\r
- addq $ 2 * SIZE, CO1 # coffset += 2\r
- decq I # i --\r
- jg .L2_4_41\r
- ALIGN_4 \r
-\r
-\r
-\r
- \r
-.L2_4_60:\r
-#if defined(TRMMKERNEL) && !defined(LEFT)\r
- addq $ 2, KK\r
-#endif\r
-\r
- decq J // j --\r
- jg .L2_01 // next 2 lines of N\r
-\r
-\r
-\r
-.L1_0:\r
-\r
-/************************************************************************************************\r
-* Loop for Nmod6 % 2 > 0\r
-*************************************************************************************************/\r
-\r
- movq Nmod6, J \r
- andq $ 1, J // j % 2\r
- je .L999\r
- ALIGN_4\r
-\r
-.L1_01:\r
- // copy to sub buffer\r
- movq B, BO1\r
- leaq BUFFER1, BO // first buffer to BO\r
- movq K, %rax\r
- ALIGN_4\r
-\r
-.L1_02b:\r
-\r
- vmovsd (BO1), %xmm0\r
- vmovsd %xmm0, (BO)\r
- addq $ 2*SIZE,BO1\r
- addq $ 2*SIZE,BO\r
- decq %rax\r
- jnz .L1_02b\r
-\r
-.L1_02c:\r
-\r
- movq BO1, B // next offset of B\r
-\r
-.L1_10:\r
- movq C, CO1\r
- leaq (C, LDC, 1), C // c += 1 * ldc\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- movq OFFSET, %rax\r
- movq %rax, KK\r
-#endif\r
- \r
- movq A, AO // aoffset = a\r
- addq $ 16 * SIZE, AO\r
-\r
- movq M, I\r
- sarq $ 3, I // i = (m >> 3)\r
- je .L1_4_10\r
-\r
- ALIGN_4\r
-\r
-/**************************************************************************************************/\r
-\r
-.L1_8_11:\r
-\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $ 4 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $ 4 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (,BI,2), BI // BI = BI * 2 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $ 4, %rax // rax = rax *16 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
- vzeroall\r
-\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
- movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $ 8, %rax // number of values in AO\r
-#else\r
- addq $ 1, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
-\r
-\r
- andq $ -8, %rax // K = K - ( K % 8 )\r
- je .L1_8_16\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,2), BI // BI = BI * 2 ; number of values\r
-\r
- salq $ 4, %rax // rax = rax *16 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
- leaq (BO, BI, SIZE), BO\r
- negq BI\r
- negq %rax\r
- ALIGN_4\r
-\r
-.L1_8_12:\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL8x1_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x1_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x1_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x1_SUB\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x1_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x1_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x1_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x1_SUB\r
-\r
- je .L1_8_16\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL8x1_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x1_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x1_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x1_SUB\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x1_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x1_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x1_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL8x1_SUB\r
-\r
- je .L1_8_16\r
-\r
- jmp .L1_8_12\r
- ALIGN_4\r
-\r
-.L1_8_16:\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
-\r
- andq $ 7, %rax # if (k & 1)\r
- je .L1_8_19\r
-\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,2), BI // BI = BI * 4 ; number of values\r
-\r
- salq $ 4, %rax // rax = rax *16 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
- leaq (BO, BI, SIZE), BO\r
- negq BI\r
- negq %rax\r
- ALIGN_4\r
-\r
-.L1_8_17:\r
-\r
- KERNEL8x1_SUB\r
-\r
- jl .L1_8_17\r
- ALIGN_4\r
-\r
-\r
-.L1_8_19:\r
-\r
- SAVE8x1\r
-\r
-\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax\r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,2), BI // BI = BI * 2 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $ 4, %rax // rax = rax *16 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $ 8, KK\r
-#endif\r
-\r
- addq $ 16 * SIZE, CO1 # coffset += 16\r
- decq I # i --\r
- jg .L1_8_11\r
- ALIGN_4 \r
-\r
-\r
-\r
-/**************************************************************************************************/\r
-.L1_4_10:\r
-\r
- testq $ 7, M \r
- jz .L999\r
-\r
- testq $ 4, M \r
- jz .L1_4_20\r
-\r
-\r
-.L1_4_11:\r
-\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $ 4 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $ 4 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (,BI,2), BI // BI = BI * 2 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $ 3, %rax // rax = rax * 8 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
- vzeroall\r
-\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
- movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $ 4, %rax // number of values in AO\r
-#else\r
- addq $ 1, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
-\r
-\r
- andq $ -8, %rax // K = K - ( K % 8 )\r
- je .L1_4_16\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,2), BI // BI = BI * 2 ; number of values\r
-\r
- salq $ 3, %rax // rax = rax * 8 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
- leaq (BO, BI, SIZE), BO\r
- negq BI\r
- negq %rax\r
- ALIGN_4\r
-\r
-.L1_4_12:\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL4x1_SUB\r
- KERNEL4x1_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL4x1_SUB\r
- KERNEL4x1_SUB\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL4x1_SUB\r
- KERNEL4x1_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL4x1_SUB\r
- KERNEL4x1_SUB\r
-\r
- je .L1_4_16\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL4x1_SUB\r
- KERNEL4x1_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL4x1_SUB\r
- KERNEL4x1_SUB\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL4x1_SUB\r
- KERNEL4x1_SUB\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL4x1_SUB\r
- KERNEL4x1_SUB\r
-\r
- je .L1_4_16\r
-\r
- jmp .L1_4_12\r
- ALIGN_4\r
-\r
-.L1_4_16:\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
-\r
- andq $ 7, %rax # if (k & 1)\r
- je .L1_4_19\r
-\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,2), BI // BI = BI * 4 ; number of values\r
-\r
- salq $ 3, %rax // rax = rax * 8 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
- leaq (BO, BI, SIZE), BO\r
- negq BI\r
- negq %rax\r
- ALIGN_4\r
-\r
-.L1_4_17:\r
-\r
- KERNEL4x1_SUB\r
-\r
- jl .L1_4_17\r
- ALIGN_4\r
-\r
-\r
-.L1_4_19:\r
-\r
- SAVE4x1\r
-\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax\r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,2), BI // BI = BI * 2 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $ 3, %rax // rax = rax * 8 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $ 4, KK\r
-#endif\r
-\r
- addq $ 8 * SIZE, CO1 # coffset += 8\r
- ALIGN_4 \r
-\r
-\r
-\r
-/**************************************************************************\r
-* Rest of M \r
-***************************************************************************/\r
-\r
-.L1_4_20:\r
-\r
- testq $ 2, M \r
- jz .L1_4_40\r
- ALIGN_4\r
-\r
-.L1_4_21:\r
-\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $ 4 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $ 4 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (,BI,2), BI // BI = BI * 2 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $ 2, %rax // rax = rax * 4 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
- vzeroall\r
-\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
- movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $ 2, %rax // number of values in AO\r
-#else\r
- addq $ 1, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
-\r
-\r
- andq $ -8, %rax // K = K - ( K % 8 )\r
- je .L1_4_26\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,2), BI // BI = BI * 2 ; number of values\r
-\r
- salq $ 2, %rax // rax = rax * 4 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
- leaq (BO, BI, SIZE), BO\r
- negq BI\r
- negq %rax\r
- ALIGN_4\r
-\r
-.L1_4_22:\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL2x1_SUB\r
- KERNEL2x1_SUB\r
- KERNEL2x1_SUB\r
- KERNEL2x1_SUB\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL2x1_SUB\r
- KERNEL2x1_SUB\r
- KERNEL2x1_SUB\r
- KERNEL2x1_SUB\r
-\r
- je .L1_4_26\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL2x1_SUB\r
- KERNEL2x1_SUB\r
- KERNEL2x1_SUB\r
- KERNEL2x1_SUB\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- KERNEL2x1_SUB\r
- KERNEL2x1_SUB\r
- KERNEL2x1_SUB\r
- KERNEL2x1_SUB\r
-\r
- je .L1_4_26\r
-\r
- jmp .L1_4_22\r
- ALIGN_4\r
-\r
-.L1_4_26:\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
-\r
- andq $ 7, %rax # if (k & 1)\r
- je .L1_4_29\r
-\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,2), BI // BI = BI * 2; number of values\r
-\r
- salq $ 2, %rax // rax = rax * 4 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
- leaq (BO, BI, SIZE), BO\r
- negq BI\r
- negq %rax\r
- ALIGN_4\r
-\r
-.L1_4_27:\r
-\r
- KERNEL2x1_SUB\r
-\r
- jl .L1_4_27\r
- ALIGN_4\r
-\r
-\r
-.L1_4_29:\r
-\r
- SAVE2x1\r
-\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax\r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,2), BI // BI = BI * 2 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $ 2, %rax // rax = rax * 4 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $ 2, KK\r
-#endif\r
-\r
- addq $ 4 * SIZE, CO1 # coffset += 4\r
- ALIGN_4 \r
-\r
-\r
-\r
-/**************************************************************************/\r
-.L1_4_40:\r
- testq $ 1, M \r
- jz .L999 // to next 2 lines of N\r
-\r
- ALIGN_4\r
-\r
-.L1_4_41:\r
-\r
-#if !defined(TRMMKERNEL) || \\r
- (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $ 4 * SIZE, BO\r
-#else\r
- movq KK, %rax\r
- leaq BUFFER1, BO // first buffer to BO\r
- addq $ 4 * SIZE, BO\r
- movq %rax, BI // Index for BO\r
- leaq (,BI,2), BI // BI = BI * 2 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $ 1, %rax // rax = rax * 2 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
- vzeroall\r
-\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))\r
- movq K, %rax\r
- subq KK, %rax\r
- movq %rax, KKK\r
-#else\r
- movq KK, %rax\r
-#ifdef LEFT\r
- addq $ 1, %rax // number of values in AO\r
-#else\r
- addq $ 1, %rax // number of values in BO\r
-#endif\r
- movq %rax, KKK\r
-#endif\r
-\r
-\r
- andq $ -8, %rax // K = K - ( K % 8 )\r
- je .L1_4_46\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,2), BI // BI = BI * 2 ; number of values\r
-\r
- salq $ 1, %rax // rax = rax * 2 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
- leaq (BO, BI, SIZE), BO\r
- negq BI\r
- negq %rax\r
- ALIGN_4\r
-\r
-.L1_4_42:\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL1x1_SUB\r
- KERNEL1x1_SUB\r
- KERNEL1x1_SUB\r
- KERNEL1x1_SUB\r
-\r
- KERNEL1x1_SUB\r
- KERNEL1x1_SUB\r
- KERNEL1x1_SUB\r
- KERNEL1x1_SUB\r
-\r
- je .L1_4_46\r
-\r
- prefetcht0 A_PR1(AO,%rax,SIZE)\r
- prefetcht0 B_PR1(BO,BI,SIZE)\r
- KERNEL1x1_SUB\r
- KERNEL1x1_SUB\r
- KERNEL1x1_SUB\r
- KERNEL1x1_SUB\r
-\r
- KERNEL1x1_SUB\r
- KERNEL1x1_SUB\r
- KERNEL1x1_SUB\r
- KERNEL1x1_SUB\r
-\r
- je .L1_4_46\r
-\r
- jmp .L1_4_42\r
- ALIGN_4\r
-\r
-.L1_4_46:\r
-#ifndef TRMMKERNEL\r
- movq K, %rax\r
-#else\r
- movq KKK, %rax\r
-#endif\r
-\r
- andq $ 7, %rax # if (k & 1)\r
- je .L1_4_49\r
-\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,2), BI // BI = BI * 2 ; number of values\r
-\r
- salq $ 1, %rax // rax = rax * 2 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
- leaq (BO, BI, SIZE), BO\r
- negq BI\r
- negq %rax\r
- ALIGN_4\r
-\r
-.L1_4_47:\r
-\r
- KERNEL1x1_SUB\r
-\r
- jl .L1_4_47\r
- ALIGN_4\r
-\r
-\r
-.L1_4_49:\r
-\r
- SAVE1x1\r
-\r
-\r
-\r
-#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \\r
- (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))\r
- movq K, %rax\r
- subq KKK, %rax\r
- movq %rax, BI // Index for BO\r
- leaq ( ,BI,2), BI // BI = BI * 2 ; number of values\r
- leaq (BO, BI, SIZE), BO\r
- salq $ 1, %rax // rax = rax * 2 ; number of values\r
- leaq (AO, %rax, SIZE), AO\r
-#endif\r
-\r
-\r
-#if defined(TRMMKERNEL) && defined(LEFT)\r
- addq $ 1, KK\r
-#endif\r
-\r
- addq $ 2 * SIZE, CO1 # coffset += 2\r
- ALIGN_4 \r
-\r
-\r
-.L999:\r
- vzeroupper\r
-\r
- movq SP, %rsp\r
- movq (%rsp), %rbx\r
- movq 8(%rsp), %rbp\r
- movq 16(%rsp), %r12\r
- movq 24(%rsp), %r13\r
- movq 32(%rsp), %r14\r
- movq 40(%rsp), %r15\r
-\r
-#ifdef WINDOWS_ABI\r
- movq 48(%rsp), %rdi\r
- movq 56(%rsp), %rsi\r
- vmovups 64(%rsp), %xmm6\r
- vmovups 80(%rsp), %xmm7\r
- vmovups 96(%rsp), %xmm8\r
- vmovups 112(%rsp), %xmm9\r
- vmovups 128(%rsp), %xmm10\r
- vmovups 144(%rsp), %xmm11\r
- vmovups 160(%rsp), %xmm12\r
- vmovups 176(%rsp), %xmm13\r
- vmovups 192(%rsp), %xmm14\r
- vmovups 208(%rsp), %xmm15\r
-#endif\r
-\r
- addq $ STACKSIZE, %rsp\r
- ret\r
-\r
- EPILOGUE\r
+/*********************************************************************************
+Copyright (c) 2013, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+**********************************************************************************/
+
+/*********************************************************************
+* 2014/07/29 Saar
+* BLASTEST : OK
+* CTEST : OK
+* TEST : OK
+*
+* 2013/10/28 Saar
+* Parameter:
+* CGEMM_DEFAULT_UNROLL_N 2
+* CGEMM_DEFAULT_UNROLL_M 8
+* CGEMM_DEFAULT_P 384
+* CGEMM_DEFAULT_Q 192
+* A_PR1 512
+* B_PR1 512
+*
+* 2014/07/29 Saar
+* Performance at 6912x6912x6912:
+* 1 thread: 107 GFLOPS (SANDYBRIDGE: 60) (MKL: 86)
+* 2 threads: 208 GFLOPS (SANDYBRIDGE: 114) (MKL: 155)
+* 3 threads: 289 GFLOPS (SANDYBRIDGE: 162) (MKL: 222)
+* 4 threads: 377 GFLOPS (SANDYBRIDGE: 223) (MKL: 279)
+*
+*
+*********************************************************************/
+
+
+
+#define ASSEMBLER
+#include "common.h"
+
+#define OLD_M %rdi
+#define OLD_N %rsi
+#define M %r13
+#define J %r14
+#define OLD_K %rdx
+
+#define A %rcx
+#define B %r8
+#define C %r9
+#define LDC %r10
+
+#define I %r11
+#define AO %rdi
+#define BO %rsi
+#define CO1 %r15
+#define K %r12
+#define BI %rbp
+#define SP %rbx
+
+#define BO1 %rdi
+#define BO2 %rbp
+
+#ifndef WINDOWS_ABI
+
+#define STACKSIZE 96
+
+#else
+
+#define STACKSIZE 320
+
+#define OLD_ALPHA_I 40 + STACKSIZE(%rsp)
+#define OLD_A 48 + STACKSIZE(%rsp)
+#define OLD_B 56 + STACKSIZE(%rsp)
+#define OLD_C 64 + STACKSIZE(%rsp)
+#define OLD_LDC 72 + STACKSIZE(%rsp)
+#define OLD_OFFSET 80 + STACKSIZE(%rsp)
+
+#endif
+
+#define L_BUFFER_SIZE 8192
+
+#define Ndiv6 24(%rsp)
+#define Nmod6 32(%rsp)
+#define N 40(%rsp)
+#define ALPHA_R 48(%rsp)
+#define ALPHA_I 56(%rsp)
+#define OFFSET 64(%rsp)
+#define KK 72(%rsp)
+#define KKK 80(%rsp)
+#define BUFFER1 128(%rsp)
+
+#if defined(OS_WINDOWS)
+#if L_BUFFER_SIZE > 16384
+#define STACK_TOUCH \
+ movl $ 0, 4096 * 4(%rsp);\
+ movl $ 0, 4096 * 3(%rsp);\
+ movl $ 0, 4096 * 2(%rsp);\
+ movl $ 0, 4096 * 1(%rsp);
+#elif L_BUFFER_SIZE > 12288
+#define STACK_TOUCH \
+ movl $ 0, 4096 * 3(%rsp);\
+ movl $ 0, 4096 * 2(%rsp);\
+ movl $ 0, 4096 * 1(%rsp);
+#elif L_BUFFER_SIZE > 8192
+#define STACK_TOUCH \
+ movl $ 0, 4096 * 2(%rsp);\
+ movl $ 0, 4096 * 1(%rsp);
+#elif L_BUFFER_SIZE > 4096
+#define STACK_TOUCH \
+ movl $ 0, 4096 * 1(%rsp);
+#else
+#define STACK_TOUCH
+#endif
+#else
+#define STACK_TOUCH
+#endif
+
+
+#if defined(BULLDOZER)
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT)
+
+#define VFMADDPS_R( y0,y1,y2 ) vfmaddps y0,y1,y2,y0
+
+#define VFMADDPS_I( y0,y1,y2 ) vfmaddps y0,y1,y2,y0
+
+#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)
+
+#define VFMADDPS_R( y0,y1,y2 ) vfnmaddps y0,y1,y2,y0
+
+#define VFMADDPS_I( y0,y1,y2 ) vfmaddps y0,y1,y2,y0
+
+#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+#define VFMADDPS_R( y0,y1,y2 ) vfmaddps y0,y1,y2,y0
+
+#define VFMADDPS_I( y0,y1,y2 ) vfnmaddps y0,y1,y2,y0
+
+#else
+
+#define VFMADDPS_R( y0,y1,y2 ) vfnmaddps y0,y1,y2,y0
+
+#define VFMADDPS_I( y0,y1,y2 ) vfnmaddps y0,y1,y2,y0
+
+#endif
+
+#else
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT)
+
+#define VFMADDPS_R( y0,y1,y2 ) vfmadd231ps y1,y2,y0
+
+#define VFMADDPS_I( y0,y1,y2 ) vfmadd231ps y1,y2,y0
+
+#elif defined(RN) || defined(RT) || defined(CN) || defined(CT)
+
+#define VFMADDPS_R( y0,y1,y2 ) vfnmadd231ps y1,y2,y0
+
+#define VFMADDPS_I( y0,y1,y2 ) vfmadd231ps y1,y2,y0
+
+#elif defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+#define VFMADDPS_R( y0,y1,y2 ) vfmadd231ps y1,y2,y0
+
+#define VFMADDPS_I( y0,y1,y2 ) vfnmadd231ps y1,y2,y0
+
+#else
+
+#define VFMADDPS_R( y0,y1,y2 ) vfnmadd231ps y1,y2,y0
+
+#define VFMADDPS_I( y0,y1,y2 ) vfnmadd231ps y1,y2,y0
+
+#endif
+
+#endif
+
+
+#define A_PR1 512
+#define B_PR1 512
+
+
+
+/***************************************************************************************************************************/
+
+.macro KERNEL8x3_SUB
+
+ vmovups -16 * SIZE(AO), %ymm0
+ vmovups -8 * SIZE(AO), %ymm1
+ vbroadcastss -8 * SIZE(BO), %ymm2
+ vbroadcastss -7 * SIZE(BO), %ymm3
+ prefetcht0 A_PR1(AO)
+
+ VFMADDPS_R( %ymm8 ,%ymm2,%ymm0 )
+ VFMADDPS_R( %ymm12,%ymm2,%ymm1 )
+ VFMADDPS_I( %ymm9 ,%ymm3,%ymm0 )
+ VFMADDPS_I( %ymm13,%ymm3,%ymm1 )
+
+ vbroadcastss -6 * SIZE(BO), %ymm2
+ vbroadcastss -5 * SIZE(BO), %ymm3
+ VFMADDPS_R( %ymm10,%ymm2,%ymm0 )
+ VFMADDPS_R( %ymm14,%ymm2,%ymm1 )
+ VFMADDPS_I( %ymm11,%ymm3,%ymm0 )
+ VFMADDPS_I( %ymm15,%ymm3,%ymm1 )
+
+ vbroadcastss -4 * SIZE(BO), %ymm2
+ vbroadcastss -3 * SIZE(BO), %ymm3
+ VFMADDPS_R( %ymm4 ,%ymm2,%ymm0 )
+ VFMADDPS_R( %ymm6 ,%ymm2,%ymm1 )
+ VFMADDPS_I( %ymm5 ,%ymm3,%ymm0 )
+ VFMADDPS_I( %ymm7 ,%ymm3,%ymm1 )
+
+
+ addq $6*SIZE, BO
+ addq $16*SIZE, AO
+ decq %rax
+.endm
+
+.macro SAVE8x3
+
+ vbroadcastss ALPHA_R, %ymm0
+ vbroadcastss ALPHA_I, %ymm1
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %ymm9 , %ymm9 , %ymm9
+ vshufps $ 0xb1, %ymm11, %ymm11, %ymm11
+ vshufps $ 0xb1, %ymm13, %ymm13, %ymm13
+ vshufps $ 0xb1, %ymm15, %ymm15, %ymm15
+ vshufps $ 0xb1, %ymm5 , %ymm5 , %ymm5
+ vshufps $ 0xb1, %ymm7 , %ymm7 , %ymm7
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
+ defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+ vaddsubps %ymm9, %ymm8 , %ymm8
+ vaddsubps %ymm11,%ymm10, %ymm10
+ vaddsubps %ymm13,%ymm12, %ymm12
+ vaddsubps %ymm15,%ymm14, %ymm14
+ vaddsubps %ymm5, %ymm4 , %ymm4
+ vaddsubps %ymm7, %ymm6 , %ymm6
+
+ vshufps $ 0xb1, %ymm8 , %ymm8 , %ymm9
+ vshufps $ 0xb1, %ymm10, %ymm10, %ymm11
+ vshufps $ 0xb1, %ymm12, %ymm12, %ymm13
+ vshufps $ 0xb1, %ymm14, %ymm14, %ymm15
+ vshufps $ 0xb1, %ymm4 , %ymm4 , %ymm5
+ vshufps $ 0xb1, %ymm6 , %ymm6 , %ymm7
+
+#else
+ vaddsubps %ymm8, %ymm9 ,%ymm9
+ vaddsubps %ymm10, %ymm11,%ymm11
+ vaddsubps %ymm12, %ymm13,%ymm13
+ vaddsubps %ymm14, %ymm15,%ymm15
+ vaddsubps %ymm4, %ymm5 ,%ymm5
+ vaddsubps %ymm6, %ymm7 ,%ymm7
+
+ vmovaps %ymm9, %ymm8
+ vmovaps %ymm11, %ymm10
+ vmovaps %ymm13, %ymm12
+ vmovaps %ymm15, %ymm14
+ vmovaps %ymm5, %ymm4
+ vmovaps %ymm7, %ymm6
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %ymm9 , %ymm9 , %ymm9
+ vshufps $ 0xb1, %ymm11, %ymm11, %ymm11
+ vshufps $ 0xb1, %ymm13, %ymm13, %ymm13
+ vshufps $ 0xb1, %ymm15, %ymm15, %ymm15
+ vshufps $ 0xb1, %ymm5 , %ymm5 , %ymm5
+ vshufps $ 0xb1, %ymm7 , %ymm7 , %ymm7
+
+#endif
+
+ // multiply with ALPHA_R
+ vmulps %ymm8 , %ymm0, %ymm8
+ vmulps %ymm10, %ymm0, %ymm10
+ vmulps %ymm12, %ymm0, %ymm12
+ vmulps %ymm14, %ymm0, %ymm14
+ vmulps %ymm4 , %ymm0, %ymm4
+ vmulps %ymm6 , %ymm0, %ymm6
+
+ // multiply with ALPHA_I
+ vmulps %ymm9 , %ymm1, %ymm9
+ vmulps %ymm11, %ymm1, %ymm11
+ vmulps %ymm13, %ymm1, %ymm13
+ vmulps %ymm15, %ymm1, %ymm15
+ vmulps %ymm5 , %ymm1, %ymm5
+ vmulps %ymm7 , %ymm1, %ymm7
+
+ vaddsubps %ymm9, %ymm8 , %ymm8
+ vaddsubps %ymm11,%ymm10, %ymm10
+ vaddsubps %ymm13,%ymm12, %ymm12
+ vaddsubps %ymm15,%ymm14, %ymm14
+ vaddsubps %ymm5, %ymm4 , %ymm4
+ vaddsubps %ymm7, %ymm6 , %ymm6
+
+#if !defined(TRMMKERNEL)
+
+ vaddps (CO1), %ymm8 , %ymm8
+ vaddps 8 * SIZE(CO1), %ymm12, %ymm12
+
+ vaddps (CO1, LDC), %ymm10, %ymm10
+ vaddps 8 * SIZE(CO1, LDC), %ymm14, %ymm14
+
+ vaddps (CO1, LDC,2), %ymm4, %ymm4
+ vaddps 8 * SIZE(CO1, LDC,2), %ymm6, %ymm6
+
+#endif
+
+ vmovups %ymm8 , (CO1)
+ vmovups %ymm12 , 8 * SIZE(CO1)
+
+ vmovups %ymm10 , (CO1, LDC)
+ vmovups %ymm14 , 8 * SIZE(CO1, LDC)
+
+ vmovups %ymm4 , (CO1, LDC,2)
+ vmovups %ymm6 , 8 * SIZE(CO1, LDC,2)
+
+.endm
+
+
+/***************************************************************************************************************************/
+
+.macro KERNEL4x3_SUB
+
+ vmovups -16 * SIZE(AO), %ymm0
+ vbroadcastss -8 * SIZE(BO), %ymm2
+ vbroadcastss -7 * SIZE(BO), %ymm3
+
+ VFMADDPS_R( %ymm8 ,%ymm2,%ymm0 )
+ VFMADDPS_I( %ymm9 ,%ymm3,%ymm0 )
+
+ vbroadcastss -6 * SIZE(BO), %ymm2
+ vbroadcastss -5 * SIZE(BO), %ymm3
+ VFMADDPS_R( %ymm12,%ymm2,%ymm0 )
+ VFMADDPS_I( %ymm13,%ymm3,%ymm0 )
+
+ vbroadcastss -4 * SIZE(BO), %ymm2
+ vbroadcastss -3 * SIZE(BO), %ymm3
+ VFMADDPS_R( %ymm4 ,%ymm2,%ymm0 )
+ VFMADDPS_I( %ymm5 ,%ymm3,%ymm0 )
+
+ addq $6*SIZE, BO
+ addq $8*SIZE, AO
+ decq %rax
+.endm
+
+.macro SAVE4x3
+
+ vbroadcastss ALPHA_R, %ymm0
+ vbroadcastss ALPHA_I, %ymm1
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %ymm9 , %ymm9 , %ymm9
+ vshufps $ 0xb1, %ymm13, %ymm13, %ymm13
+ vshufps $ 0xb1, %ymm5 , %ymm5 , %ymm5
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
+ defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+ vaddsubps %ymm9, %ymm8 , %ymm8
+ vaddsubps %ymm13,%ymm12, %ymm12
+ vaddsubps %ymm5, %ymm4 , %ymm4
+
+ vshufps $ 0xb1, %ymm8 , %ymm8 , %ymm9
+ vshufps $ 0xb1, %ymm12, %ymm12, %ymm13
+ vshufps $ 0xb1, %ymm4 , %ymm4 , %ymm5
+
+#else
+ vaddsubps %ymm8, %ymm9 ,%ymm9
+ vaddsubps %ymm12, %ymm13,%ymm13
+ vaddsubps %ymm4, %ymm5 ,%ymm5
+
+ vmovaps %ymm9, %ymm8
+ vmovaps %ymm13, %ymm12
+ vmovaps %ymm5, %ymm4
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %ymm9 , %ymm9 , %ymm9
+ vshufps $ 0xb1, %ymm13, %ymm13, %ymm13
+ vshufps $ 0xb1, %ymm5 , %ymm5 , %ymm5
+
+#endif
+
+ // multiply with ALPHA_R
+ vmulps %ymm8 , %ymm0, %ymm8
+ vmulps %ymm12, %ymm0, %ymm12
+ vmulps %ymm4 , %ymm0, %ymm4
+
+ // multiply with ALPHA_I
+ vmulps %ymm9 , %ymm1, %ymm9
+ vmulps %ymm13, %ymm1, %ymm13
+ vmulps %ymm5 , %ymm1, %ymm5
+
+ vaddsubps %ymm9, %ymm8 , %ymm8
+ vaddsubps %ymm13,%ymm12, %ymm12
+ vaddsubps %ymm5, %ymm4 , %ymm4
+
+#if !defined(TRMMKERNEL)
+
+ vaddps (CO1), %ymm8 , %ymm8
+ vaddps (CO1, LDC), %ymm12, %ymm12
+ vaddps (CO1, LDC,2), %ymm4, %ymm4
+
+#endif
+
+ vmovups %ymm8 , (CO1)
+ vmovups %ymm12 , (CO1, LDC)
+ vmovups %ymm4 , (CO1, LDC,2)
+
+.endm
+
+/***************************************************************************************************************************/
+
+.macro KERNEL2x3_SUB
+
+ vmovups -16 * SIZE(AO), %xmm0
+ vbroadcastss -8 * SIZE(BO), %xmm2
+ vbroadcastss -7 * SIZE(BO), %xmm3
+
+ VFMADDPS_R( %xmm8 ,%xmm2,%xmm0 )
+ VFMADDPS_I( %xmm9 ,%xmm3,%xmm0 )
+
+ vbroadcastss -6 * SIZE(BO), %xmm2
+ vbroadcastss -5 * SIZE(BO), %xmm3
+ VFMADDPS_R( %xmm12,%xmm2,%xmm0 )
+ VFMADDPS_I( %xmm13,%xmm3,%xmm0 )
+
+ vbroadcastss -4 * SIZE(BO), %xmm2
+ vbroadcastss -3 * SIZE(BO), %xmm3
+ VFMADDPS_R( %xmm4 ,%xmm2,%xmm0 )
+ VFMADDPS_I( %xmm5 ,%xmm3,%xmm0 )
+
+ addq $6*SIZE, BO
+ addq $4*SIZE, AO
+ decq %rax
+
+.endm
+
+.macro SAVE2x3
+
+ vbroadcastss ALPHA_R, %xmm0
+ vbroadcastss ALPHA_I, %xmm1
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9 , %xmm9
+ vshufps $ 0xb1, %xmm13, %xmm13, %xmm13
+ vshufps $ 0xb1, %xmm5 , %xmm5 , %xmm5
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
+ defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+ vaddsubps %xmm13,%xmm12, %xmm12
+ vaddsubps %xmm5, %xmm4 , %xmm4
+
+ vshufps $ 0xb1, %xmm8 , %xmm8 , %xmm9
+ vshufps $ 0xb1, %xmm12, %xmm12, %xmm13
+ vshufps $ 0xb1, %xmm4 , %xmm4 , %xmm5
+
+#else
+ vaddsubps %xmm8, %xmm9 ,%xmm9
+ vaddsubps %xmm12, %xmm13,%xmm13
+ vaddsubps %xmm4, %xmm5 ,%xmm5
+
+ vmovaps %xmm9, %xmm8
+ vmovaps %xmm13, %xmm12
+ vmovaps %xmm5, %xmm4
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9 , %xmm9
+ vshufps $ 0xb1, %xmm13, %xmm13, %xmm13
+ vshufps $ 0xb1, %xmm5 , %xmm5 , %xmm5
+
+#endif
+
+ // multiply with ALPHA_R
+ vmulps %xmm8 , %xmm0, %xmm8
+ vmulps %xmm12, %xmm0, %xmm12
+ vmulps %xmm4 , %xmm0, %xmm4
+
+ // multiply with ALPHA_I
+ vmulps %xmm9 , %xmm1, %xmm9
+ vmulps %xmm13, %xmm1, %xmm13
+ vmulps %xmm5 , %xmm1, %xmm5
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+ vaddsubps %xmm13,%xmm12, %xmm12
+ vaddsubps %xmm5, %xmm4 , %xmm4
+
+#if !defined(TRMMKERNEL)
+
+ vaddps (CO1), %xmm8 , %xmm8
+ vaddps (CO1, LDC), %xmm12, %xmm12
+ vaddps (CO1, LDC,2), %xmm4, %xmm4
+
+#endif
+
+ vmovups %xmm8 , (CO1)
+ vmovups %xmm12 , (CO1, LDC)
+ vmovups %xmm4 , (CO1, LDC,2)
+
+.endm
+
+
+/***************************************************************************************************************************/
+
+.macro KERNEL1x3_SUB
+
+ vmovsd -16 * SIZE(AO), %xmm0
+ vbroadcastss -8 * SIZE(BO), %xmm2
+ vbroadcastss -7 * SIZE(BO), %xmm3
+
+ VFMADDPS_R( %xmm8 ,%xmm2,%xmm0 )
+ VFMADDPS_I( %xmm9 ,%xmm3,%xmm0 )
+
+ vbroadcastss -6 * SIZE(BO), %xmm2
+ vbroadcastss -5 * SIZE(BO), %xmm3
+ VFMADDPS_R( %xmm12,%xmm2,%xmm0 )
+ VFMADDPS_I( %xmm13,%xmm3,%xmm0 )
+
+ vbroadcastss -4 * SIZE(BO), %xmm2
+ vbroadcastss -3 * SIZE(BO), %xmm3
+ VFMADDPS_R( %xmm4 ,%xmm2,%xmm0 )
+ VFMADDPS_I( %xmm5 ,%xmm3,%xmm0 )
+
+ addq $6*SIZE, BO
+ addq $2*SIZE, AO
+ decq %rax
+
+.endm
+
+.macro SAVE1x3
+
+ vbroadcastss ALPHA_R, %xmm0
+ vbroadcastss ALPHA_I, %xmm1
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9 , %xmm9
+ vshufps $ 0xb1, %xmm13, %xmm13, %xmm13
+ vshufps $ 0xb1, %xmm5 , %xmm5 , %xmm5
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
+ defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+ vaddsubps %xmm13,%xmm12, %xmm12
+ vaddsubps %xmm5, %xmm4 , %xmm4
+
+ vshufps $ 0xb1, %xmm8 , %xmm8 , %xmm9
+ vshufps $ 0xb1, %xmm12, %xmm12, %xmm13
+ vshufps $ 0xb1, %xmm4 , %xmm4 , %xmm5
+
+#else
+ vaddsubps %xmm8, %xmm9 ,%xmm9
+ vaddsubps %xmm12, %xmm13,%xmm13
+ vaddsubps %xmm4, %xmm5 ,%xmm5
+
+ vmovaps %xmm9, %xmm8
+ vmovaps %xmm13, %xmm12
+ vmovaps %xmm5, %xmm4
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9 , %xmm9
+ vshufps $ 0xb1, %xmm13, %xmm13, %xmm13
+ vshufps $ 0xb1, %xmm5 , %xmm5 , %xmm5
+
+#endif
+
+ // multiply with ALPHA_R
+ vmulps %xmm8 , %xmm0, %xmm8
+ vmulps %xmm12, %xmm0, %xmm12
+ vmulps %xmm4 , %xmm0, %xmm4
+
+ // multiply with ALPHA_I
+ vmulps %xmm9 , %xmm1, %xmm9
+ vmulps %xmm13, %xmm1, %xmm13
+ vmulps %xmm5 , %xmm1, %xmm5
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+ vaddsubps %xmm13,%xmm12, %xmm12
+ vaddsubps %xmm5, %xmm4 , %xmm4
+
+#if !defined(TRMMKERNEL)
+
+ vmovsd (CO1) , %xmm9
+ vmovsd (CO1,LDC) , %xmm13
+ vmovsd (CO1,LDC,2), %xmm5
+ vaddps %xmm9 , %xmm8 , %xmm8
+ vaddps %xmm13, %xmm12, %xmm12
+ vaddps %xmm5 , %xmm4, %xmm4
+
+#endif
+
+ vmovsd %xmm8 , (CO1)
+ vmovsd %xmm12 , (CO1, LDC)
+ vmovsd %xmm4 , (CO1, LDC,2)
+
+.endm
+
+
+/***************************************************************************************************************************/
+
+.macro KERNEL8x2_SUB
+
+ vmovups -16 * SIZE(AO, %rax, SIZE), %ymm0
+ vbroadcastss -8 * SIZE(BO, BI, SIZE), %ymm4
+ VFMADDPS_R( %ymm8,%ymm4,%ymm0 )
+ vmovups -8 * SIZE(AO, %rax, SIZE), %ymm1
+ VFMADDPS_R( %ymm12,%ymm4,%ymm1 )
+ vbroadcastss -7 * SIZE(BO, BI, SIZE), %ymm5
+ VFMADDPS_I( %ymm9,%ymm5,%ymm0 )
+ VFMADDPS_I( %ymm13,%ymm5,%ymm1 )
+ vbroadcastss -6 * SIZE(BO, BI, SIZE), %ymm6
+ VFMADDPS_R( %ymm10,%ymm6,%ymm0 )
+ VFMADDPS_R( %ymm14,%ymm6,%ymm1 )
+ vbroadcastss -5 * SIZE(BO, BI, SIZE), %ymm7
+ VFMADDPS_I( %ymm11,%ymm7,%ymm0 )
+ VFMADDPS_I( %ymm15,%ymm7,%ymm1 )
+ addq $ 4 , BI
+ addq $ 16, %rax
+.endm
+
+.macro SAVE8x2
+
+ vbroadcastss ALPHA_R, %ymm0
+ vbroadcastss ALPHA_I, %ymm1
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %ymm9 , %ymm9, %ymm9
+ vshufps $ 0xb1, %ymm11, %ymm11, %ymm11
+ vshufps $ 0xb1, %ymm13, %ymm13, %ymm13
+ vshufps $ 0xb1, %ymm15, %ymm15, %ymm15
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
+ defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+ vaddsubps %ymm9, %ymm8 , %ymm8
+ vaddsubps %ymm11,%ymm10, %ymm10
+ vaddsubps %ymm13,%ymm12, %ymm12
+ vaddsubps %ymm15,%ymm14, %ymm14
+
+ vshufps $ 0xb1, %ymm8 , %ymm8, %ymm9
+ vshufps $ 0xb1, %ymm10, %ymm10, %ymm11
+ vshufps $ 0xb1, %ymm12, %ymm12, %ymm13
+ vshufps $ 0xb1, %ymm14, %ymm14, %ymm15
+
+#else
+ vaddsubps %ymm8, %ymm9 ,%ymm9
+ vaddsubps %ymm10, %ymm11,%ymm11
+ vaddsubps %ymm12, %ymm13,%ymm13
+ vaddsubps %ymm14, %ymm15,%ymm15
+
+ vmovaps %ymm9, %ymm8
+ vmovaps %ymm11, %ymm10
+ vmovaps %ymm13, %ymm12
+ vmovaps %ymm15, %ymm14
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %ymm9 , %ymm9, %ymm9
+ vshufps $ 0xb1, %ymm11, %ymm11, %ymm11
+ vshufps $ 0xb1, %ymm13, %ymm13, %ymm13
+ vshufps $ 0xb1, %ymm15, %ymm15, %ymm15
+
+#endif
+
+ // multiply with ALPHA_R
+ vmulps %ymm8 , %ymm0, %ymm8
+ vmulps %ymm10, %ymm0, %ymm10
+ vmulps %ymm12, %ymm0, %ymm12
+ vmulps %ymm14, %ymm0, %ymm14
+
+ // multiply with ALPHA_I
+ vmulps %ymm9 , %ymm1, %ymm9
+ vmulps %ymm11, %ymm1, %ymm11
+ vmulps %ymm13, %ymm1, %ymm13
+ vmulps %ymm15, %ymm1, %ymm15
+
+ vaddsubps %ymm9, %ymm8 , %ymm8
+ vaddsubps %ymm11,%ymm10, %ymm10
+ vaddsubps %ymm13,%ymm12, %ymm12
+ vaddsubps %ymm15,%ymm14, %ymm14
+
+
+
+#if !defined(TRMMKERNEL)
+
+ vaddps (CO1), %ymm8 , %ymm8
+ vaddps 8 * SIZE(CO1), %ymm12, %ymm12
+
+ vaddps (CO1, LDC), %ymm10, %ymm10
+ vaddps 8 * SIZE(CO1, LDC), %ymm14, %ymm14
+
+#endif
+
+ vmovups %ymm8 , (CO1)
+ vmovups %ymm12 , 8 * SIZE(CO1)
+
+ vmovups %ymm10 , (CO1, LDC)
+ vmovups %ymm14 , 8 * SIZE(CO1, LDC)
+
+ prefetcht0 64(CO1)
+ prefetcht0 64(CO1, LDC)
+
+.endm
+
+/***************************************************************************************************************************/
+
+.macro KERNEL4x2_SUB
+ vmovups -16 * SIZE(AO, %rax, SIZE), %xmm0
+ vbroadcastss -8 * SIZE(BO, BI, SIZE), %xmm4
+ VFMADDPS_R( %xmm8,%xmm4,%xmm0 )
+ vmovups -12 * SIZE(AO, %rax, SIZE), %xmm1
+ VFMADDPS_R( %xmm12,%xmm4,%xmm1 )
+ vbroadcastss -7 * SIZE(BO, BI, SIZE), %xmm5
+ VFMADDPS_I( %xmm9,%xmm5,%xmm0 )
+ VFMADDPS_I( %xmm13,%xmm5,%xmm1 )
+ vbroadcastss -6 * SIZE(BO, BI, SIZE), %xmm6
+ VFMADDPS_R( %xmm10,%xmm6,%xmm0 )
+ VFMADDPS_R( %xmm14,%xmm6,%xmm1 )
+ vbroadcastss -5 * SIZE(BO, BI, SIZE), %xmm7
+ VFMADDPS_I( %xmm11,%xmm7,%xmm0 )
+ VFMADDPS_I( %xmm15,%xmm7,%xmm1 )
+ addq $ 4, BI
+ addq $ 8, %rax
+.endm
+
+.macro SAVE4x2
+
+ vbroadcastss ALPHA_R, %xmm0
+ vbroadcastss ALPHA_I, %xmm1
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9
+ vshufps $ 0xb1, %xmm11, %xmm11, %xmm11
+ vshufps $ 0xb1, %xmm13, %xmm13, %xmm13
+ vshufps $ 0xb1, %xmm15, %xmm15, %xmm15
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
+ defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+ vaddsubps %xmm11,%xmm10, %xmm10
+ vaddsubps %xmm13,%xmm12, %xmm12
+ vaddsubps %xmm15,%xmm14, %xmm14
+
+ vshufps $ 0xb1, %xmm8 , %xmm8, %xmm9
+ vshufps $ 0xb1, %xmm10, %xmm10, %xmm11
+ vshufps $ 0xb1, %xmm12, %xmm12, %xmm13
+ vshufps $ 0xb1, %xmm14, %xmm14, %xmm15
+
+#else
+ vaddsubps %xmm8, %xmm9 ,%xmm9
+ vaddsubps %xmm10, %xmm11,%xmm11
+ vaddsubps %xmm12, %xmm13,%xmm13
+ vaddsubps %xmm14, %xmm15,%xmm15
+
+ vmovaps %xmm9, %xmm8
+ vmovaps %xmm11, %xmm10
+ vmovaps %xmm13, %xmm12
+ vmovaps %xmm15, %xmm14
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9
+ vshufps $ 0xb1, %xmm11, %xmm11, %xmm11
+ vshufps $ 0xb1, %xmm13, %xmm13, %xmm13
+ vshufps $ 0xb1, %xmm15, %xmm15, %xmm15
+
+#endif
+
+ // multiply with ALPHA_R
+ vmulps %xmm8 , %xmm0, %xmm8
+ vmulps %xmm10, %xmm0, %xmm10
+ vmulps %xmm12, %xmm0, %xmm12
+ vmulps %xmm14, %xmm0, %xmm14
+
+ // multiply with ALPHA_I
+ vmulps %xmm9 , %xmm1, %xmm9
+ vmulps %xmm11, %xmm1, %xmm11
+ vmulps %xmm13, %xmm1, %xmm13
+ vmulps %xmm15, %xmm1, %xmm15
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+ vaddsubps %xmm11,%xmm10, %xmm10
+ vaddsubps %xmm13,%xmm12, %xmm12
+ vaddsubps %xmm15,%xmm14, %xmm14
+
+#if !defined(TRMMKERNEL)
+
+ vaddps (CO1), %xmm8 , %xmm8
+ vaddps 4 * SIZE(CO1), %xmm12, %xmm12
+
+ vaddps (CO1, LDC), %xmm10, %xmm10
+ vaddps 4 * SIZE(CO1, LDC), %xmm14, %xmm14
+
+#endif
+
+ vmovups %xmm8 , (CO1)
+ vmovups %xmm12 , 4 * SIZE(CO1)
+
+ vmovups %xmm10 , (CO1, LDC)
+ vmovups %xmm14 , 4 * SIZE(CO1, LDC)
+
+.endm
+
+/************************************************************************************************/
+
+.macro KERNEL2x2_SUB
+ vmovups -16 * SIZE(AO, %rax, SIZE), %xmm0
+ vbroadcastss -8 * SIZE(BO, BI, SIZE), %xmm4
+ VFMADDPS_R( %xmm8,%xmm4,%xmm0 )
+ vbroadcastss -7 * SIZE(BO, BI, SIZE), %xmm5
+ VFMADDPS_I( %xmm9,%xmm5,%xmm0 )
+ vbroadcastss -6 * SIZE(BO, BI, SIZE), %xmm6
+ VFMADDPS_R( %xmm10,%xmm6,%xmm0 )
+ vbroadcastss -5 * SIZE(BO, BI, SIZE), %xmm7
+ VFMADDPS_I( %xmm11,%xmm7,%xmm0 )
+ addq $ 4, BI
+ addq $ 4, %rax
+.endm
+
+.macro SAVE2x2
+
+ vbroadcastss ALPHA_R, %xmm0
+ vbroadcastss ALPHA_I, %xmm1
+
+ // swap high and low 4 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9
+ vshufps $ 0xb1, %xmm11, %xmm11, %xmm11
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
+ defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+ vaddsubps %xmm11,%xmm10, %xmm10
+
+ vshufps $ 0xb1, %xmm8 , %xmm8, %xmm9
+ vshufps $ 0xb1, %xmm10, %xmm10, %xmm11
+
+#else
+ vaddsubps %xmm8, %xmm9 ,%xmm9
+ vaddsubps %xmm10, %xmm11,%xmm11
+
+ vmovaps %xmm9, %xmm8
+ vmovaps %xmm11, %xmm10
+
+ // swap high and low 4 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9
+ vshufps $ 0xb1, %xmm11, %xmm11, %xmm11
+
+#endif
+
+ // multiply with ALPHA_R
+ vmulps %xmm8 , %xmm0, %xmm8
+ vmulps %xmm10, %xmm0, %xmm10
+
+ // multiply with ALPHA_I
+ vmulps %xmm9 , %xmm1, %xmm9
+ vmulps %xmm11, %xmm1, %xmm11
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+ vaddsubps %xmm11,%xmm10, %xmm10
+
+#if !defined(TRMMKERNEL)
+
+ vaddps (CO1), %xmm8 , %xmm8
+
+ vaddps (CO1, LDC), %xmm10, %xmm10
+
+#endif
+
+ vmovups %xmm8 , (CO1)
+
+ vmovups %xmm10 , (CO1, LDC)
+
+.endm
+
+/************************************************************************************************/
+
+.macro KERNEL1x2_SUB
+ vmovsd -16 * SIZE(AO, %rax, SIZE), %xmm0
+ vbroadcastss -8 * SIZE(BO, BI, SIZE), %xmm4
+ VFMADDPS_R( %xmm8,%xmm4,%xmm0 )
+ vbroadcastss -7 * SIZE(BO, BI, SIZE), %xmm5
+ VFMADDPS_I( %xmm9,%xmm5,%xmm0 )
+ vbroadcastss -6 * SIZE(BO, BI, SIZE), %xmm6
+ VFMADDPS_R( %xmm10,%xmm6,%xmm0 )
+ vbroadcastss -5 * SIZE(BO, BI, SIZE), %xmm7
+ VFMADDPS_I( %xmm11,%xmm7,%xmm0 )
+ addq $ 4, BI
+ addq $ 2, %rax
+.endm
+
+.macro SAVE1x2
+
+ vbroadcastss ALPHA_R, %xmm0
+ vbroadcastss ALPHA_I, %xmm1
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9
+ vshufps $ 0xb1, %xmm11, %xmm11, %xmm11
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
+ defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+ vaddsubps %xmm11,%xmm10, %xmm10
+
+ vshufps $ 0xb1, %xmm8 , %xmm8, %xmm9
+ vshufps $ 0xb1, %xmm10, %xmm10, %xmm11
+
+#else
+ vaddsubps %xmm8, %xmm9 ,%xmm9
+ vaddsubps %xmm10, %xmm11,%xmm11
+
+ vmovaps %xmm9, %xmm8
+ vmovaps %xmm11, %xmm10
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9
+ vshufps $ 0xb1, %xmm11, %xmm11, %xmm11
+
+#endif
+
+ // multiply with ALPHA_R
+ vmulps %xmm8 , %xmm0, %xmm8
+ vmulps %xmm10, %xmm0, %xmm10
+
+ // multiply with ALPHA_I
+ vmulps %xmm9 , %xmm1, %xmm9
+ vmulps %xmm11, %xmm1, %xmm11
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+ vaddsubps %xmm11,%xmm10, %xmm10
+
+#if !defined(TRMMKERNEL)
+
+ vmovsd (CO1), %xmm14
+ vaddps %xmm14, %xmm8 , %xmm8
+
+ vmovsd (CO1, LDC), %xmm15
+ vaddps %xmm15, %xmm10, %xmm10
+
+#endif
+
+ vmovsd %xmm8 , (CO1)
+ vmovsd %xmm10 , (CO1, LDC)
+
+.endm
+
+/************************************************************************************************/
+
+.macro KERNEL8x1_SUB
+ vmovups -16 * SIZE(AO, %rax, SIZE), %ymm0
+ vmovups -8 * SIZE(AO, %rax, SIZE), %ymm1
+ vbroadcastss -4 * SIZE(BO, BI, SIZE), %ymm4
+ VFMADDPS_R( %ymm8,%ymm4,%ymm0 )
+ VFMADDPS_R( %ymm12,%ymm4,%ymm1 )
+ vbroadcastss -3 * SIZE(BO, BI, SIZE), %ymm5
+ VFMADDPS_I( %ymm9,%ymm5,%ymm0 )
+ VFMADDPS_I( %ymm13,%ymm5,%ymm1 )
+ addq $ 2 , BI
+ addq $ 16, %rax
+.endm
+
+.macro SAVE8x1
+
+ vbroadcastss ALPHA_R, %ymm0
+ vbroadcastss ALPHA_I, %ymm1
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %ymm9 , %ymm9, %ymm9
+ vshufps $ 0xb1, %ymm13, %ymm13, %ymm13
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
+ defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+ vaddsubps %ymm9, %ymm8 , %ymm8
+ vaddsubps %ymm13,%ymm12, %ymm12
+
+ vshufps $ 0xb1, %ymm8 , %ymm8, %ymm9
+ vshufps $ 0xb1, %ymm12, %ymm12, %ymm13
+
+#else
+ vaddsubps %ymm8, %ymm9 ,%ymm9
+ vaddsubps %ymm12, %ymm13,%ymm13
+
+ vmovaps %ymm9, %ymm8
+ vmovaps %ymm13, %ymm12
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %ymm9 , %ymm9, %ymm9
+ vshufps $ 0xb1, %ymm13, %ymm13, %ymm13
+
+#endif
+
+ // multiply with ALPHA_R
+ vmulps %ymm8 , %ymm0, %ymm8
+ vmulps %ymm12, %ymm0, %ymm12
+
+ // multiply with ALPHA_I
+ vmulps %ymm9 , %ymm1, %ymm9
+ vmulps %ymm13, %ymm1, %ymm13
+
+ vaddsubps %ymm9, %ymm8 , %ymm8
+ vaddsubps %ymm13,%ymm12, %ymm12
+
+
+
+#if !defined(TRMMKERNEL)
+
+ vaddps (CO1), %ymm8 , %ymm8
+ vaddps 8 * SIZE(CO1), %ymm12, %ymm12
+
+#endif
+
+ vmovups %ymm8 , (CO1)
+ vmovups %ymm12 , 8 * SIZE(CO1)
+
+.endm
+
+
+/************************************************************************************************/
+
+.macro KERNEL4x1_SUB
+ vmovups -16 * SIZE(AO, %rax, SIZE), %xmm0
+ vbroadcastss -4 * SIZE(BO, BI, SIZE), %xmm4
+ VFMADDPS_R( %xmm8,%xmm4,%xmm0 )
+ vmovups -12 * SIZE(AO, %rax, SIZE), %xmm1
+ VFMADDPS_R( %xmm12,%xmm4,%xmm1 )
+ vbroadcastss -3 * SIZE(BO, BI, SIZE), %xmm5
+ VFMADDPS_I( %xmm9,%xmm5,%xmm0 )
+ VFMADDPS_I( %xmm13,%xmm5,%xmm1 )
+ addq $ 2, BI
+ addq $ 8, %rax
+.endm
+
+.macro SAVE4x1
+
+ vbroadcastss ALPHA_R, %xmm0
+ vbroadcastss ALPHA_I, %xmm1
+
+ // swap high and low 4 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9
+ vshufps $ 0xb1, %xmm13, %xmm13, %xmm13
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
+ defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+ vaddsubps %xmm13,%xmm12, %xmm12
+
+ vshufps $ 0xb1, %xmm8 , %xmm8, %xmm9
+ vshufps $ 0xb1, %xmm12, %xmm12, %xmm13
+
+#else
+ vaddsubps %xmm8, %xmm9 ,%xmm9
+ vaddsubps %xmm12, %xmm13,%xmm13
+
+ vmovaps %xmm9, %xmm8
+ vmovaps %xmm13, %xmm12
+
+ // swap high and low 4 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9
+ vshufps $ 0xb1, %xmm13, %xmm13, %xmm13
+
+#endif
+
+ // multiply with ALPHA_R
+ vmulps %xmm8 , %xmm0, %xmm8
+ vmulps %xmm12, %xmm0, %xmm12
+
+ // multiply with ALPHA_I
+ vmulps %xmm9 , %xmm1, %xmm9
+ vmulps %xmm13, %xmm1, %xmm13
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+ vaddsubps %xmm13,%xmm12, %xmm12
+
+#ifndef TRMMKERNEL
+
+ vaddps (CO1), %xmm8 , %xmm8
+ vaddps 4 * SIZE(CO1), %xmm12, %xmm12
+
+#endif
+
+ vmovups %xmm8 , (CO1)
+ vmovups %xmm12 , 4 * SIZE(CO1)
+
+.endm
+
+/************************************************************************************************/
+
+.macro KERNEL2x1_SUB
+ vmovups -16 * SIZE(AO, %rax, SIZE), %xmm0
+ vbroadcastss -4 * SIZE(BO, BI, SIZE), %xmm4
+ VFMADDPS_R( %xmm8,%xmm4,%xmm0 )
+ vbroadcastss -3 * SIZE(BO, BI, SIZE), %xmm5
+ VFMADDPS_I( %xmm9,%xmm5,%xmm0 )
+ addq $ 2, BI
+ addq $ 4, %rax
+.endm
+
+.macro SAVE2x1
+
+ vbroadcastss ALPHA_R, %xmm0
+ vbroadcastss ALPHA_I, %xmm1
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
+ defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+
+ vshufps $ 0xb1, %xmm8 , %xmm8, %xmm9
+
+#else
+ vaddsubps %xmm8, %xmm9 ,%xmm9
+
+ vmovaps %xmm9, %xmm8
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9
+
+#endif
+
+ // multiply with ALPHA_R
+ vmulps %xmm8 , %xmm0, %xmm8
+
+ // multiply with ALPHA_I
+ vmulps %xmm9 , %xmm1, %xmm9
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+
+#if !defined(TRMMKERNEL)
+
+ vaddps (CO1), %xmm8 , %xmm8
+
+#endif
+
+ vmovups %xmm8 , (CO1)
+
+.endm
+
+/************************************************************************************************/
+
+.macro KERNEL1x1_SUB
+ vmovsd -16 * SIZE(AO, %rax, SIZE), %xmm0
+ vbroadcastss -4 * SIZE(BO, BI, SIZE), %xmm4
+ VFMADDPS_R( %xmm8,%xmm4,%xmm0 )
+ vbroadcastss -3 * SIZE(BO, BI, SIZE), %xmm5
+ VFMADDPS_I( %xmm9,%xmm5,%xmm0 )
+ addq $ 2, BI
+ addq $ 2, %rax
+.endm
+
+.macro SAVE1x1
+
+ vbroadcastss ALPHA_R, %xmm0
+ vbroadcastss ALPHA_I, %xmm1
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
+ defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+
+ vshufps $ 0xb1, %xmm8 , %xmm8, %xmm9
+
+#else
+ vaddsubps %xmm8, %xmm9 ,%xmm9
+
+ vmovaps %xmm9, %xmm8
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9
+
+#endif
+
+ // multiply with ALPHA_R
+ vmulps %xmm8 , %xmm0, %xmm8
+
+ // multiply with ALPHA_I
+ vmulps %xmm9 , %xmm1, %xmm9
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+
+#if !defined(TRMMKERNEL)
+
+ vmovsd (CO1), %xmm14
+ vaddps %xmm14, %xmm8 , %xmm8
+
+#endif
+
+ vmovsd %xmm8 , (CO1)
+
+.endm
+
+
+#if !defined(TRMMKERNEL)
+
+ PROLOGUE
+ PROFCODE
+
+ subq $STACKSIZE, %rsp
+ movq %rbx, (%rsp)
+ movq %rbp, 8(%rsp)
+ movq %r12, 16(%rsp)
+ movq %r13, 24(%rsp)
+ movq %r14, 32(%rsp)
+ movq %r15, 40(%rsp)
+
+ vzeroupper
+
+#ifdef WINDOWS_ABI
+ movq %rdi, 48(%rsp)
+ movq %rsi, 56(%rsp)
+ vmovups %xmm6, 64(%rsp)
+ vmovups %xmm7, 80(%rsp)
+ vmovups %xmm8, 96(%rsp)
+ vmovups %xmm9, 112(%rsp)
+ vmovups %xmm10, 128(%rsp)
+ vmovups %xmm11, 144(%rsp)
+ vmovups %xmm12, 160(%rsp)
+ vmovups %xmm13, 176(%rsp)
+ vmovups %xmm14, 192(%rsp)
+ vmovups %xmm15, 208(%rsp)
+
+ movq ARG1, OLD_M
+ movq ARG2, OLD_N
+ movq ARG3, OLD_K
+ movq OLD_A, A
+ movq OLD_B, B
+ movq OLD_C, C
+ movq OLD_LDC, LDC
+ vmovaps %xmm3, %xmm0
+ vmovsd OLD_ALPHA_I, %xmm1
+
+#else
+ movq STACKSIZE + 8(%rsp), LDC
+
+#endif
+
+ movq %rsp, SP # save old stack
+ subq $ 128 + L_BUFFER_SIZE, %rsp
+ andq $ -4096, %rsp # align stack
+
+ STACK_TOUCH
+
+ cmpq $ 0, OLD_M
+ je .L999
+
+ cmpq $ 0, OLD_N
+ je .L999
+
+ cmpq $ 0, OLD_K
+ je .L999
+
+ movq OLD_M, M
+ movq OLD_N, N
+ movq OLD_K, K
+
+ vmovss %xmm0, ALPHA_R
+ vmovss %xmm1, ALPHA_I
+
+ salq $ ZBASE_SHIFT, LDC
+
+ movq N, %rax
+ xorq %rdx, %rdx
+ movq $ 6, %rdi
+ divq %rdi // N / 6
+ movq %rax, Ndiv6 // N / 6
+ movq %rdx, Nmod6 // N % 6
+
+/************************************************************************************************/
+
+.L6_0:
+
+ movq Ndiv6, J
+ cmpq $ 0, J
+ je .L2_00
+ ALIGN_4
+
+
+
+.L6_01:
+ // copy to sub buffer
+ movq B, BO1
+ leaq BUFFER1, BO // first buffer to BO
+ movq K, %rax
+ salq $2, %rax // 2 * COMPSIZE
+ leaq (B, %rax,4), BO2
+ movq BO2, B // next offset of B
+ movq K, %rax
+
+ ALIGN_4
+
+.L6_02b:
+
+ vmovups (BO1), %xmm0
+ vmovsd (BO2), %xmm1
+ vmovups %xmm0, (BO)
+ vmovsd %xmm1, 4*SIZE(BO)
+ addq $ 4*SIZE,BO1
+ addq $ 4*SIZE,BO2
+ addq $ 6*SIZE,BO
+ decq %rax
+ jnz .L6_02b
+
+
+.L6_10:
+ movq C, CO1
+ leaq (C, LDC, 2), C // c += 2 * ldc
+ leaq (C, LDC, 1), C // c += 1 * ldc
+
+ movq A, AO // aoffset = a
+ addq $ 16 * SIZE, AO
+
+ movq M, I
+ sarq $ 3, I // i = (m >> 3)
+ je .L6_4_10
+
+ ALIGN_4
+/**********************************************************************************************************/
+
+.L6_8_11:
+
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L6_8_16
+
+ ALIGN_4
+
+.L6_8_12:
+
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+
+ je .L6_8_16
+
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+
+ je .L6_8_16
+
+ jmp .L6_8_12
+ ALIGN_4
+
+.L6_8_16:
+ movq K, %rax
+
+ andq $ 7, %rax # if (k & 1)
+ je .L6_8_19
+
+ ALIGN_4
+
+.L6_8_17:
+
+ KERNEL8x3_SUB
+
+ jnz .L6_8_17
+ ALIGN_4
+
+
+.L6_8_19:
+
+ SAVE8x3
+
+ addq $ 16 * SIZE, CO1 # coffset += 16
+ decq I # i --
+ jg .L6_8_11
+ ALIGN_4
+
+
+/**********************************************************************************************************/
+
+
+.L6_4_10:
+ testq $ 7, M
+ jz .L6_4_60 // to next 2 lines of N
+
+ testq $ 4, M
+ jz .L6_4_20
+ ALIGN_4
+
+
+.L6_4_11:
+
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L6_4_16
+
+ ALIGN_4
+
+.L6_4_12:
+
+ prefetcht0 A_PR1(AO)
+ KERNEL4x3_SUB
+ KERNEL4x3_SUB
+ prefetcht0 A_PR1(AO)
+ KERNEL4x3_SUB
+ KERNEL4x3_SUB
+
+ prefetcht0 A_PR1(AO)
+ KERNEL4x3_SUB
+ KERNEL4x3_SUB
+ prefetcht0 A_PR1(AO)
+ KERNEL4x3_SUB
+ KERNEL4x3_SUB
+
+ je .L6_4_16
+
+ prefetcht0 A_PR1(AO)
+ KERNEL4x3_SUB
+ KERNEL4x3_SUB
+ prefetcht0 A_PR1(AO)
+ KERNEL4x3_SUB
+ KERNEL4x3_SUB
+
+ prefetcht0 A_PR1(AO)
+ KERNEL4x3_SUB
+ KERNEL4x3_SUB
+ prefetcht0 A_PR1(AO)
+ KERNEL4x3_SUB
+ KERNEL4x3_SUB
+
+ je .L6_4_16
+
+ jmp .L6_4_12
+ ALIGN_4
+
+.L6_4_16:
+ movq K, %rax
+
+ andq $ 7, %rax # if (k & 1)
+ je .L6_4_19
+
+ ALIGN_4
+
+.L6_4_17:
+
+ KERNEL4x3_SUB
+
+ jnz .L6_4_17
+ ALIGN_4
+
+
+.L6_4_19:
+
+ SAVE4x3
+
+ addq $ 8 * SIZE, CO1 # coffset += 8
+ ALIGN_4
+
+
+
+/**************************************************************************
+* Rest of M
+***************************************************************************/
+
+.L6_4_20:
+
+ testq $ 2, M
+ jz .L6_4_40
+ ALIGN_4
+
+.L6_4_21:
+
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L6_4_26
+
+ ALIGN_4
+
+.L6_4_22:
+
+ prefetcht0 A_PR1(AO)
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+
+ prefetcht0 A_PR1(AO)
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+
+ je .L6_4_26
+
+ prefetcht0 A_PR1(AO)
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+
+ prefetcht0 A_PR1(AO)
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+
+ je .L6_4_26
+
+ jmp .L6_4_22
+ ALIGN_4
+
+.L6_4_26:
+ movq K, %rax
+
+ andq $ 7, %rax # if (k & 1)
+ je .L6_4_29
+
+ ALIGN_4
+
+.L6_4_27:
+
+ KERNEL2x3_SUB
+
+ jnz .L6_4_27
+ ALIGN_4
+
+
+.L6_4_29:
+
+ SAVE2x3
+
+ addq $ 4 * SIZE, CO1 # coffset += 4
+ decq I # i --
+ jg .L6_4_21
+ ALIGN_4
+
+
+
+/**************************************************************************/
+.L6_4_40:
+ testq $ 1, M
+ jz .L6_4_60 // to next 2 lines of N
+
+ ALIGN_4
+
+.L6_4_41:
+
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L6_4_46
+
+ ALIGN_4
+
+.L6_4_42:
+
+ prefetcht0 A_PR1(AO)
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+
+ je .L6_4_46
+
+ prefetcht0 A_PR1(AO)
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+
+ je .L6_4_46
+
+ jmp .L6_4_42
+ ALIGN_4
+
+.L6_4_46:
+ movq K, %rax
+
+ andq $ 7, %rax # if (k & 1)
+ je .L6_4_49
+ ALIGN_4
+
+.L6_4_47:
+
+ KERNEL1x3_SUB
+
+ jnz .L6_4_47
+ ALIGN_4
+
+
+.L6_4_49:
+
+ SAVE1x3
+
+ addq $ 2 * SIZE, CO1 # coffset += 2
+ decq I # i --
+ jg .L6_4_41
+ ALIGN_4
+
+
+
+
+.L6_4_60:
+
+
+/*******************************************************************************************/
+
+.L7_01:
+ // copy to sub buffer
+ movq B, BO1
+ leaq BUFFER1, BO // first buffer to BO
+ movq K, %rax
+ salq $2, %rax // 2 * COMPSIZE
+ leaq (B, %rax,4), BO2
+ movq K, %rax
+
+ ALIGN_4
+
+.L7_02b:
+
+ vmovsd 2*SIZE(BO1), %xmm0
+ vmovups (BO2), %xmm1
+ vmovsd %xmm0, (BO)
+ vmovups %xmm1, 2*SIZE(BO)
+ addq $ 4*SIZE,BO1
+ addq $ 4*SIZE,BO2
+ addq $ 6*SIZE,BO
+ decq %rax
+ jnz .L7_02b
+
+ movq BO2, B // next offset of B
+
+.L7_10:
+ movq C, CO1
+ leaq (C, LDC, 2), C // c += 2 * ldc
+ leaq (C, LDC, 1), C // c += 1 * ldc
+
+ movq A, AO // aoffset = a
+ addq $ 16 * SIZE, AO
+
+ movq M, I
+ sarq $ 3, I // i = (m >> 3)
+ je .L7_4_10
+
+ ALIGN_4
+/**********************************************************************************************************/
+
+.L7_8_11:
+
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L7_8_16
+
+ ALIGN_4
+
+.L7_8_12:
+
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+
+ je .L7_8_16
+
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+ KERNEL8x3_SUB
+
+ je .L7_8_16
+
+ jmp .L7_8_12
+ ALIGN_4
+
+.L7_8_16:
+ movq K, %rax
+
+ andq $ 7, %rax # if (k & 1)
+ je .L7_8_19
+
+ ALIGN_4
+
+.L7_8_17:
+
+ KERNEL8x3_SUB
+
+ jnz .L7_8_17
+ ALIGN_4
+
+
+.L7_8_19:
+
+ SAVE8x3
+
+ addq $ 16 * SIZE, CO1 # coffset += 16
+ decq I # i --
+ jg .L7_8_11
+ ALIGN_4
+
+
+/**********************************************************************************************************/
+
+
+.L7_4_10:
+ testq $ 7, M
+ jz .L7_4_60 // to next 2 lines of N
+
+ testq $ 4, M
+ jz .L7_4_20
+ ALIGN_4
+
+
+.L7_4_11:
+
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L7_4_16
+
+ ALIGN_4
+
+.L7_4_12:
+
+ prefetcht0 A_PR1(AO)
+ KERNEL4x3_SUB
+ KERNEL4x3_SUB
+ prefetcht0 A_PR1(AO)
+ KERNEL4x3_SUB
+ KERNEL4x3_SUB
+
+ prefetcht0 A_PR1(AO)
+ KERNEL4x3_SUB
+ KERNEL4x3_SUB
+ prefetcht0 A_PR1(AO)
+ KERNEL4x3_SUB
+ KERNEL4x3_SUB
+
+ je .L7_4_16
+
+ prefetcht0 A_PR1(AO)
+ KERNEL4x3_SUB
+ KERNEL4x3_SUB
+ prefetcht0 A_PR1(AO)
+ KERNEL4x3_SUB
+ KERNEL4x3_SUB
+
+ prefetcht0 A_PR1(AO)
+ KERNEL4x3_SUB
+ KERNEL4x3_SUB
+ prefetcht0 A_PR1(AO)
+ KERNEL4x3_SUB
+ KERNEL4x3_SUB
+
+ je .L7_4_16
+
+ jmp .L7_4_12
+ ALIGN_4
+
+.L7_4_16:
+ movq K, %rax
+
+ andq $ 7, %rax # if (k & 1)
+ je .L7_4_19
+
+ ALIGN_4
+
+.L7_4_17:
+
+ KERNEL4x3_SUB
+
+ jnz .L7_4_17
+ ALIGN_4
+
+
+.L7_4_19:
+
+ SAVE4x3
+
+ addq $ 8 * SIZE, CO1 # coffset += 8
+ ALIGN_4
+
+
+
+/**************************************************************************
+* Rest of M
+***************************************************************************/
+
+.L7_4_20:
+
+ testq $ 2, M
+ jz .L7_4_40
+ ALIGN_4
+
+.L7_4_21:
+
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L7_4_26
+
+ ALIGN_4
+
+.L7_4_22:
+
+ prefetcht0 A_PR1(AO)
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+
+ prefetcht0 A_PR1(AO)
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+
+ je .L7_4_26
+
+ prefetcht0 A_PR1(AO)
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+
+ prefetcht0 A_PR1(AO)
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+ KERNEL2x3_SUB
+
+ je .L7_4_26
+
+ jmp .L7_4_22
+ ALIGN_4
+
+.L7_4_26:
+ movq K, %rax
+
+ andq $ 7, %rax # if (k & 1)
+ je .L7_4_29
+
+ ALIGN_4
+
+.L7_4_27:
+
+ KERNEL2x3_SUB
+
+ jnz .L7_4_27
+ ALIGN_4
+
+
+.L7_4_29:
+
+ SAVE2x3
+
+ addq $ 4 * SIZE, CO1 # coffset += 4
+ decq I # i --
+ jg .L7_4_21
+ ALIGN_4
+
+
+
+/**************************************************************************/
+.L7_4_40:
+ testq $ 1, M
+ jz .L7_4_60 // to next 2 lines of N
+
+ ALIGN_4
+
+.L7_4_41:
+
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L7_4_46
+
+ ALIGN_4
+
+.L7_4_42:
+
+ prefetcht0 A_PR1(AO)
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+
+ je .L7_4_46
+
+ prefetcht0 A_PR1(AO)
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+ KERNEL1x3_SUB
+
+ je .L7_4_46
+
+ jmp .L7_4_42
+ ALIGN_4
+
+.L7_4_46:
+ movq K, %rax
+
+ andq $ 7, %rax # if (k & 1)
+ je .L7_4_49
+ ALIGN_4
+
+.L7_4_47:
+
+ KERNEL1x3_SUB
+
+ jnz .L7_4_47
+ ALIGN_4
+
+
+.L7_4_49:
+
+ SAVE1x3
+
+ addq $ 2 * SIZE, CO1 # coffset += 2
+ decq I # i --
+ jg .L7_4_41
+ ALIGN_4
+
+
+
+
+.L7_4_60:
+
+ decq J // j --
+ jg .L6_01 // next 6 lines of N
+
+
+
+/************************************************************************************************/
+
+.L2_00:
+
+ movq Nmod6, J
+ sarq $1, J // j = j / 2
+ cmpq $ 0, J
+ je .L1_0
+ ALIGN_4
+
+
+
+.L2_01:
+ // copy to sub buffer
+ movq B, BO1
+ leaq BUFFER1, BO // first buffer to BO
+ movq K, %rax
+ ALIGN_4
+
+.L2_02b:
+
+ vmovups (BO1), %xmm0
+ vmovups %xmm0, (BO)
+ addq $ 4*SIZE,BO1
+ addq $ 4*SIZE,BO
+ decq %rax
+ jnz .L2_02b
+
+.L2_02c:
+
+ movq BO1, B // next offset of B
+
+.L2_10:
+ movq C, CO1
+ leaq (C, LDC, 2), C // c += 2 * ldc
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ movq OFFSET, %rax
+ movq %rax, KK
+#endif
+
+ movq A, AO // aoffset = a
+ addq $ 16 * SIZE, AO
+
+ movq M, I
+ sarq $ 3, I // i = (m >> 3)
+ je .L2_4_10
+
+ ALIGN_4
+/**********************************************************************************************************/
+
+.L2_8_11:
+
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (,BI,4), BI // BI = BI * 4 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 4, %rax // rax = rax *16 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+ vzeroall
+
+#ifndef TRMMKERNEL
+ movq K, %rax
+#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ movq K, %rax
+ subq KK, %rax
+ movq %rax, KKK
+#else
+ movq KK, %rax
+#ifdef LEFT
+ addq $ 8, %rax // number of values in AO
+#else
+ addq $ 2, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L2_8_16
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+
+ salq $ 4, %rax // rax = rax *16 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_8_12:
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+
+ je .L2_8_16
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+
+ je .L2_8_16
+
+ jmp .L2_8_12
+ ALIGN_4
+
+.L2_8_16:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $ 7, %rax # if (k & 1)
+ je .L2_8_19
+
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+
+ salq $ 4, %rax // rax = rax *16 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_8_17:
+
+ KERNEL8x2_SUB
+
+ jl .L2_8_17
+ ALIGN_4
+
+
+.L2_8_19:
+
+ SAVE8x2
+
+
+#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ movq K, %rax
+ subq KKK, %rax
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 4, %rax // rax = rax *16 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $ 8, KK
+#endif
+
+ addq $ 16 * SIZE, CO1 # coffset += 16
+ decq I # i --
+ jg .L2_8_11
+ ALIGN_4
+
+
+/**********************************************************************************************************/
+
+
+
+
+.L2_4_10:
+ testq $ 7, M
+ jz .L2_4_60 // to next 2 lines of N
+
+ testq $ 4, M
+ jz .L2_4_20
+ ALIGN_4
+
+
+.L2_4_11:
+
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (,BI,4), BI // BI = BI * 4 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 3, %rax // rax = rax * 8 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+ vzeroall
+
+#ifndef TRMMKERNEL
+ movq K, %rax
+#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ movq K, %rax
+ subq KK, %rax
+ movq %rax, KKK
+#else
+ movq KK, %rax
+#ifdef LEFT
+ addq $ 4, %rax // number of values in AO
+#else
+ addq $ 2, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L2_4_16
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+
+ salq $ 3, %rax // rax = rax * 8 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_4_12:
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL4x2_SUB
+ KERNEL4x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x2_SUB
+ KERNEL4x2_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL4x2_SUB
+ KERNEL4x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x2_SUB
+ KERNEL4x2_SUB
+
+ je .L2_4_16
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL4x2_SUB
+ KERNEL4x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x2_SUB
+ KERNEL4x2_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL4x2_SUB
+ KERNEL4x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x2_SUB
+ KERNEL4x2_SUB
+
+ je .L2_4_16
+
+ jmp .L2_4_12
+ ALIGN_4
+
+.L2_4_16:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $ 7, %rax # if (k & 1)
+ je .L2_4_19
+
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+
+ salq $ 3, %rax // rax = rax * 8 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_4_17:
+
+ KERNEL4x2_SUB
+
+ jl .L2_4_17
+ ALIGN_4
+
+
+.L2_4_19:
+
+ SAVE4x2
+
+#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ movq K, %rax
+ subq KKK, %rax
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 3, %rax // rax = rax * 8 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $ 4, KK
+#endif
+
+ addq $ 8 * SIZE, CO1 # coffset += 8
+ ALIGN_4
+
+
+
+/**************************************************************************
+* Rest of M
+***************************************************************************/
+
+.L2_4_20:
+
+ testq $ 2, M
+ jz .L2_4_40
+ ALIGN_4
+
+.L2_4_21:
+
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (,BI,4), BI // BI = BI * 4 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 2, %rax // rax = rax * 4 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+ vzeroall
+
+#ifndef TRMMKERNEL
+ movq K, %rax
+#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ movq K, %rax
+ subq KK, %rax
+ movq %rax, KKK
+#else
+ movq KK, %rax
+#ifdef LEFT
+ addq $ 2, %rax // number of values in AO
+#else
+ addq $ 2, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L2_4_26
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+
+ salq $ 2, %rax // rax = rax * 4 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_4_22:
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+
+ je .L2_4_26
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+
+ je .L2_4_26
+
+ jmp .L2_4_22
+ ALIGN_4
+
+.L2_4_26:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $ 7, %rax # if (k & 1)
+ je .L2_4_29
+
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+
+ salq $ 2, %rax // rax = rax * 4 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_4_27:
+
+ KERNEL2x2_SUB
+
+ jl .L2_4_27
+ ALIGN_4
+
+
+.L2_4_29:
+
+ vbroadcastss ALPHA_R, %xmm0
+ vbroadcastss ALPHA_I, %xmm1
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9
+ vshufps $ 0xb1, %xmm11, %xmm11, %xmm11
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
+ defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+ vaddsubps %xmm11,%xmm10, %xmm10
+
+ vshufps $ 0xb1, %xmm8 , %xmm8, %xmm9
+ vshufps $ 0xb1, %xmm10, %xmm10, %xmm11
+
+#else
+ vaddsubps %xmm8, %xmm9 ,%xmm9
+ vaddsubps %xmm10, %xmm11,%xmm11
+
+ vmovaps %xmm9, %xmm8
+ vmovaps %xmm11, %xmm10
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9
+ vshufps $ 0xb1, %xmm11, %xmm11, %xmm11
+
+#endif
+
+ // multiply with ALPHA_R
+ vmulps %xmm8 , %xmm0, %xmm8
+ vmulps %xmm10, %xmm0, %xmm10
+
+ // multiply with ALPHA_I
+ vmulps %xmm9 , %xmm1, %xmm9
+ vmulps %xmm11, %xmm1, %xmm11
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+ vaddsubps %xmm11,%xmm10, %xmm10
+
+
+
+#ifndef TRMMKERNEL
+
+ vaddps (CO1), %xmm8 , %xmm8
+
+ vaddps (CO1, LDC), %xmm10, %xmm10
+
+#endif
+
+ vmovups %xmm8 , (CO1)
+
+ vmovups %xmm10 , (CO1, LDC)
+
+
+
+#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ movq K, %rax
+ subq KKK, %rax
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 2, %rax // rax = rax * 4 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $ 2, KK
+#endif
+
+ addq $ 4 * SIZE, CO1 # coffset += 4
+ decq I # i --
+ jg .L2_4_21
+ ALIGN_4
+
+
+
+/**************************************************************************/
+.L2_4_40:
+ testq $ 1, M
+ jz .L2_4_60 // to next 2 lines of N
+
+ ALIGN_4
+
+.L2_4_41:
+
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (,BI,4), BI // BI = BI * 4 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 1, %rax // rax = rax * 2 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+ vzeroall
+
+#ifndef TRMMKERNEL
+ movq K, %rax
+#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ movq K, %rax
+ subq KK, %rax
+ movq %rax, KKK
+#else
+ movq KK, %rax
+#ifdef LEFT
+ addq $ 1, %rax // number of values in AO
+#else
+ addq $ 2, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L2_4_46
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+
+ salq $ 1, %rax // rax = rax * 2 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_4_42:
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+
+ je .L2_4_46
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+
+ je .L2_4_46
+
+ jmp .L2_4_42
+ ALIGN_4
+
+.L2_4_46:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $ 7, %rax # if (k & 1)
+ je .L2_4_49
+
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+
+ salq $ 1, %rax // rax = rax * 2 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_4_47:
+
+ KERNEL1x2_SUB
+
+ jl .L2_4_47
+ ALIGN_4
+
+
+.L2_4_49:
+
+ SAVE1x2
+
+#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ movq K, %rax
+ subq KKK, %rax
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 1, %rax // rax = rax * 2 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $ 1, KK
+#endif
+
+ addq $ 2 * SIZE, CO1 # coffset += 2
+ decq I # i --
+ jg .L2_4_41
+ ALIGN_4
+
+
+
+
+.L2_4_60:
+#if defined(TRMMKERNEL) && !defined(LEFT)
+ addq $ 2, KK
+#endif
+
+ decq J // j --
+ jg .L2_01 // next 2 lines of N
+
+
+
+.L1_0:
+
+/************************************************************************************************
+* Loop for Nmod6 % 2 > 0
+*************************************************************************************************/
+
+ movq Nmod6, J
+ andq $ 1, J // j % 2
+ je .L999
+ ALIGN_4
+
+.L1_01:
+ // copy to sub buffer
+ movq B, BO1
+ leaq BUFFER1, BO // first buffer to BO
+ movq K, %rax
+ ALIGN_4
+
+.L1_02b:
+
+ vmovsd (BO1), %xmm0
+ vmovsd %xmm0, (BO)
+ addq $ 2*SIZE,BO1
+ addq $ 2*SIZE,BO
+ decq %rax
+ jnz .L1_02b
+
+.L1_02c:
+
+ movq BO1, B // next offset of B
+
+.L1_10:
+ movq C, CO1
+ leaq (C, LDC, 1), C // c += 1 * ldc
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ movq OFFSET, %rax
+ movq %rax, KK
+#endif
+
+ movq A, AO // aoffset = a
+ addq $ 16 * SIZE, AO
+
+ movq M, I
+ sarq $ 3, I // i = (m >> 3)
+ je .L1_4_10
+
+ ALIGN_4
+
+/**************************************************************************************************/
+
+.L1_8_11:
+
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 4 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 4 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (,BI,2), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 4, %rax // rax = rax *16 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+ vzeroall
+
+#ifndef TRMMKERNEL
+ movq K, %rax
+#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ movq K, %rax
+ subq KK, %rax
+ movq %rax, KKK
+#else
+ movq KK, %rax
+#ifdef LEFT
+ addq $ 8, %rax // number of values in AO
+#else
+ addq $ 1, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L1_8_16
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+
+ salq $ 4, %rax // rax = rax *16 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_8_12:
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+
+ je .L1_8_16
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+
+ je .L1_8_16
+
+ jmp .L1_8_12
+ ALIGN_4
+
+.L1_8_16:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $ 7, %rax # if (k & 1)
+ je .L1_8_19
+
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 4 ; number of values
+
+ salq $ 4, %rax // rax = rax *16 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_8_17:
+
+ KERNEL8x1_SUB
+
+ jl .L1_8_17
+ ALIGN_4
+
+
+.L1_8_19:
+
+ SAVE8x1
+
+
+#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ movq K, %rax
+ subq KKK, %rax
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 4, %rax // rax = rax *16 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $ 8, KK
+#endif
+
+ addq $ 16 * SIZE, CO1 # coffset += 16
+ decq I # i --
+ jg .L1_8_11
+ ALIGN_4
+
+
+
+/**************************************************************************************************/
+.L1_4_10:
+
+ testq $ 7, M
+ jz .L999
+
+ testq $ 4, M
+ jz .L1_4_20
+
+
+.L1_4_11:
+
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 4 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 4 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (,BI,2), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 3, %rax // rax = rax * 8 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+ vzeroall
+
+#ifndef TRMMKERNEL
+ movq K, %rax
+#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ movq K, %rax
+ subq KK, %rax
+ movq %rax, KKK
+#else
+ movq KK, %rax
+#ifdef LEFT
+ addq $ 4, %rax // number of values in AO
+#else
+ addq $ 1, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L1_4_16
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+
+ salq $ 3, %rax // rax = rax * 8 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_4_12:
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL4x1_SUB
+ KERNEL4x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x1_SUB
+ KERNEL4x1_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x1_SUB
+ KERNEL4x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x1_SUB
+ KERNEL4x1_SUB
+
+ je .L1_4_16
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL4x1_SUB
+ KERNEL4x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x1_SUB
+ KERNEL4x1_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x1_SUB
+ KERNEL4x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x1_SUB
+ KERNEL4x1_SUB
+
+ je .L1_4_16
+
+ jmp .L1_4_12
+ ALIGN_4
+
+.L1_4_16:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $ 7, %rax # if (k & 1)
+ je .L1_4_19
+
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 4 ; number of values
+
+ salq $ 3, %rax // rax = rax * 8 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_4_17:
+
+ KERNEL4x1_SUB
+
+ jl .L1_4_17
+ ALIGN_4
+
+
+.L1_4_19:
+
+ SAVE4x1
+
+#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ movq K, %rax
+ subq KKK, %rax
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 3, %rax // rax = rax * 8 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $ 4, KK
+#endif
+
+ addq $ 8 * SIZE, CO1 # coffset += 8
+ ALIGN_4
+
+
+
+/**************************************************************************
+* Rest of M
+***************************************************************************/
+
+.L1_4_20:
+
+ testq $ 2, M
+ jz .L1_4_40
+ ALIGN_4
+
+.L1_4_21:
+
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 4 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 4 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (,BI,2), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 2, %rax // rax = rax * 4 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+ vzeroall
+
+#ifndef TRMMKERNEL
+ movq K, %rax
+#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ movq K, %rax
+ subq KK, %rax
+ movq %rax, KKK
+#else
+ movq KK, %rax
+#ifdef LEFT
+ addq $ 2, %rax // number of values in AO
+#else
+ addq $ 1, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L1_4_26
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+
+ salq $ 2, %rax // rax = rax * 4 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_4_22:
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+
+ je .L1_4_26
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+
+ je .L1_4_26
+
+ jmp .L1_4_22
+ ALIGN_4
+
+.L1_4_26:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $ 7, %rax # if (k & 1)
+ je .L1_4_29
+
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2; number of values
+
+ salq $ 2, %rax // rax = rax * 4 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_4_27:
+
+ KERNEL2x1_SUB
+
+ jl .L1_4_27
+ ALIGN_4
+
+
+.L1_4_29:
+
+ SAVE2x1
+
+#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ movq K, %rax
+ subq KKK, %rax
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 2, %rax // rax = rax * 4 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $ 2, KK
+#endif
+
+ addq $ 4 * SIZE, CO1 # coffset += 4
+ ALIGN_4
+
+
+
+/**************************************************************************/
+.L1_4_40:
+ testq $ 1, M
+ jz .L999 // to next 2 lines of N
+
+ ALIGN_4
+
+.L1_4_41:
+
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 4 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 4 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (,BI,2), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 1, %rax // rax = rax * 2 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+ vzeroall
+
+#ifndef TRMMKERNEL
+ movq K, %rax
+#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ movq K, %rax
+ subq KK, %rax
+ movq %rax, KKK
+#else
+ movq KK, %rax
+#ifdef LEFT
+ addq $ 1, %rax // number of values in AO
+#else
+ addq $ 1, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L1_4_46
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+
+ salq $ 1, %rax // rax = rax * 2 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_4_42:
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+
+ je .L1_4_46
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+
+ je .L1_4_46
+
+ jmp .L1_4_42
+ ALIGN_4
+
+.L1_4_46:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $ 7, %rax # if (k & 1)
+ je .L1_4_49
+
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+
+ salq $ 1, %rax // rax = rax * 2 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_4_47:
+
+ KERNEL1x1_SUB
+
+ jl .L1_4_47
+ ALIGN_4
+
+
+.L1_4_49:
+
+ SAVE1x1
+
+
+
+#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ movq K, %rax
+ subq KKK, %rax
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 1, %rax // rax = rax * 2 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $ 1, KK
+#endif
+
+ addq $ 2 * SIZE, CO1 # coffset += 2
+ ALIGN_4
+
+
+.L999:
+ vzeroupper
+
+ movq SP, %rsp
+ movq (%rsp), %rbx
+ movq 8(%rsp), %rbp
+ movq 16(%rsp), %r12
+ movq 24(%rsp), %r13
+ movq 32(%rsp), %r14
+ movq 40(%rsp), %r15
+
+#ifdef WINDOWS_ABI
+ movq 48(%rsp), %rdi
+ movq 56(%rsp), %rsi
+ vmovups 64(%rsp), %xmm6
+ vmovups 80(%rsp), %xmm7
+ vmovups 96(%rsp), %xmm8
+ vmovups 112(%rsp), %xmm9
+ vmovups 128(%rsp), %xmm10
+ vmovups 144(%rsp), %xmm11
+ vmovups 160(%rsp), %xmm12
+ vmovups 176(%rsp), %xmm13
+ vmovups 192(%rsp), %xmm14
+ vmovups 208(%rsp), %xmm15
+#endif
+
+ addq $ STACKSIZE, %rsp
+ ret
+
+ EPILOGUE
+
+#else
+
+/************************************************************************************************/
+
+
+ PROLOGUE
+ PROFCODE
+
+ subq $ STACKSIZE, %rsp
+ movq %rbx, (%rsp)
+ movq %rbp, 8(%rsp)
+ movq %r12, 16(%rsp)
+ movq %r13, 24(%rsp)
+ movq %r14, 32(%rsp)
+ movq %r15, 40(%rsp)
+
+ vzeroupper
+
+#ifdef WINDOWS_ABI
+ movq %rdi, 48(%rsp)
+ movq %rsi, 56(%rsp)
+ vmovups %xmm6, 64(%rsp)
+ vmovups %xmm7, 80(%rsp)
+ vmovups %xmm8, 96(%rsp)
+ vmovups %xmm9, 112(%rsp)
+ vmovups %xmm10, 128(%rsp)
+ vmovups %xmm11, 144(%rsp)
+ vmovups %xmm12, 160(%rsp)
+ vmovups %xmm13, 176(%rsp)
+ vmovups %xmm14, 192(%rsp)
+ vmovups %xmm15, 208(%rsp)
+
+ movq ARG1, OLD_M
+ movq ARG2, OLD_N
+ movq ARG3, OLD_K
+ movq OLD_A, A
+ movq OLD_B, B
+ movq OLD_C, C
+ movq OLD_LDC, LDC
+#ifdef TRMMKERNEL
+ movsd OLD_OFFSET, %xmm12
+#endif
+ vmovaps %xmm3, %xmm0
+ vmovsd OLD_ALPHA_I, %xmm1
+
+#else
+ movq STACKSIZE + 8(%rsp), LDC
+#ifdef TRMMKERNEL
+ movsd STACKSIZE + 16(%rsp), %xmm12
+#endif
+
+#endif
+
+ movq %rsp, SP # save old stack
+ subq $ 128 + L_BUFFER_SIZE, %rsp
+ andq $ -4096, %rsp # align stack
+
+ STACK_TOUCH
+
+ cmpq $ 0, OLD_M
+ je .L999
+
+ cmpq $ 0, OLD_N
+ je .L999
+
+ cmpq $ 0, OLD_K
+ je .L999
+
+ movq OLD_M, M
+ movq OLD_N, N
+ movq OLD_K, K
+
+ vmovss %xmm0, ALPHA_R
+ vmovss %xmm1, ALPHA_I
+
+ salq $ ZBASE_SHIFT, LDC
+
+ movq N, %rax
+ xorq %rdx, %rdx
+ movq $ 2, %rdi
+ divq %rdi // N / 2
+ movq %rax, Ndiv6 // N / 2
+ movq %rdx, Nmod6 // N % 2
+
+
+
+#ifdef TRMMKERNEL
+ vmovsd %xmm12, OFFSET
+ vmovsd %xmm12, KK
+#ifndef LEFT
+ negq KK
+#endif
+#endif
+
+.L2_0:
+
+ movq Ndiv6, J
+ cmpq $ 0, J
+ je .L1_0
+ ALIGN_4
+
+
+
+.L2_01:
+ // copy to sub buffer
+ movq B, BO1
+ leaq BUFFER1, BO // first buffer to BO
+ movq K, %rax
+ ALIGN_4
+
+.L2_02b:
+
+ vmovups (BO1), %xmm0
+ vmovups %xmm0, (BO)
+ addq $ 4*SIZE,BO1
+ addq $ 4*SIZE,BO
+ decq %rax
+ jnz .L2_02b
+
+.L2_02c:
+
+ movq BO1, B // next offset of B
+
+.L2_10:
+ movq C, CO1
+ leaq (C, LDC, 2), C // c += 2 * ldc
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ movq OFFSET, %rax
+ movq %rax, KK
+#endif
+
+ movq A, AO // aoffset = a
+ addq $ 16 * SIZE, AO
+
+ movq M, I
+ sarq $ 3, I // i = (m >> 3)
+ je .L2_4_10
+
+ ALIGN_4
+/**********************************************************************************************************/
+
+.L2_8_11:
+
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (,BI,4), BI // BI = BI * 4 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 4, %rax // rax = rax *16 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+ vzeroall
+
+#ifndef TRMMKERNEL
+ movq K, %rax
+#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ movq K, %rax
+ subq KK, %rax
+ movq %rax, KKK
+#else
+ movq KK, %rax
+#ifdef LEFT
+ addq $ 8, %rax // number of values in AO
+#else
+ addq $ 2, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L2_8_16
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+
+ salq $ 4, %rax // rax = rax *16 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_8_12:
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+
+ je .L2_8_16
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x2_SUB
+
+ je .L2_8_16
+
+ jmp .L2_8_12
+ ALIGN_4
+
+.L2_8_16:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $ 7, %rax # if (k & 1)
+ je .L2_8_19
+
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+
+ salq $ 4, %rax // rax = rax *16 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_8_17:
+
+ KERNEL8x2_SUB
+
+ jl .L2_8_17
+ ALIGN_4
+
+
+.L2_8_19:
+
+ SAVE8x2
+
+
+#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ movq K, %rax
+ subq KKK, %rax
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 4, %rax // rax = rax *16 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $ 8, KK
+#endif
+
+ addq $ 16 * SIZE, CO1 # coffset += 16
+ decq I # i --
+ jg .L2_8_11
+ ALIGN_4
+
+
+/**********************************************************************************************************/
+
+
+
+
+.L2_4_10:
+ testq $ 7, M
+ jz .L2_4_60 // to next 2 lines of N
+
+ testq $ 4, M
+ jz .L2_4_20
+ ALIGN_4
+
+
+.L2_4_11:
+
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (,BI,4), BI // BI = BI * 4 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 3, %rax // rax = rax * 8 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+ vzeroall
+
+#ifndef TRMMKERNEL
+ movq K, %rax
+#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ movq K, %rax
+ subq KK, %rax
+ movq %rax, KKK
+#else
+ movq KK, %rax
+#ifdef LEFT
+ addq $ 4, %rax // number of values in AO
+#else
+ addq $ 2, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L2_4_16
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+
+ salq $ 3, %rax // rax = rax * 8 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_4_12:
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL4x2_SUB
+ KERNEL4x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x2_SUB
+ KERNEL4x2_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL4x2_SUB
+ KERNEL4x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x2_SUB
+ KERNEL4x2_SUB
+
+ je .L2_4_16
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL4x2_SUB
+ KERNEL4x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x2_SUB
+ KERNEL4x2_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL4x2_SUB
+ KERNEL4x2_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x2_SUB
+ KERNEL4x2_SUB
+
+ je .L2_4_16
+
+ jmp .L2_4_12
+ ALIGN_4
+
+.L2_4_16:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $ 7, %rax # if (k & 1)
+ je .L2_4_19
+
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+
+ salq $ 3, %rax // rax = rax * 8 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_4_17:
+
+ KERNEL4x2_SUB
+
+ jl .L2_4_17
+ ALIGN_4
+
+
+.L2_4_19:
+
+ SAVE4x2
+
+#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ movq K, %rax
+ subq KKK, %rax
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 3, %rax // rax = rax * 8 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $ 4, KK
+#endif
+
+ addq $ 8 * SIZE, CO1 # coffset += 8
+ ALIGN_4
+
+
+
+/**************************************************************************
+* Rest of M
+***************************************************************************/
+
+.L2_4_20:
+
+ testq $ 2, M
+ jz .L2_4_40
+ ALIGN_4
+
+.L2_4_21:
+
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (,BI,4), BI // BI = BI * 4 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 2, %rax // rax = rax * 4 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+ vzeroall
+
+#ifndef TRMMKERNEL
+ movq K, %rax
+#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ movq K, %rax
+ subq KK, %rax
+ movq %rax, KKK
+#else
+ movq KK, %rax
+#ifdef LEFT
+ addq $ 2, %rax // number of values in AO
+#else
+ addq $ 2, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L2_4_26
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+
+ salq $ 2, %rax // rax = rax * 4 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_4_22:
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+
+ je .L2_4_26
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+ KERNEL2x2_SUB
+
+ je .L2_4_26
+
+ jmp .L2_4_22
+ ALIGN_4
+
+.L2_4_26:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $ 7, %rax # if (k & 1)
+ je .L2_4_29
+
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+
+ salq $ 2, %rax // rax = rax * 4 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_4_27:
+
+ KERNEL2x2_SUB
+
+ jl .L2_4_27
+ ALIGN_4
+
+
+.L2_4_29:
+
+ vbroadcastss ALPHA_R, %xmm0
+ vbroadcastss ALPHA_I, %xmm1
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9
+ vshufps $ 0xb1, %xmm11, %xmm11, %xmm11
+
+#if defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
+ defined(NR) || defined(NC) || defined(TR) || defined(TC)
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+ vaddsubps %xmm11,%xmm10, %xmm10
+
+ vshufps $ 0xb1, %xmm8 , %xmm8, %xmm9
+ vshufps $ 0xb1, %xmm10, %xmm10, %xmm11
+
+#else
+ vaddsubps %xmm8, %xmm9 ,%xmm9
+ vaddsubps %xmm10, %xmm11,%xmm11
+
+ vmovaps %xmm9, %xmm8
+ vmovaps %xmm11, %xmm10
+
+ // swap high and low 64 bytes
+ vshufps $ 0xb1, %xmm9 , %xmm9, %xmm9
+ vshufps $ 0xb1, %xmm11, %xmm11, %xmm11
+
+#endif
+
+ // multiply with ALPHA_R
+ vmulps %xmm8 , %xmm0, %xmm8
+ vmulps %xmm10, %xmm0, %xmm10
+
+ // multiply with ALPHA_I
+ vmulps %xmm9 , %xmm1, %xmm9
+ vmulps %xmm11, %xmm1, %xmm11
+
+ vaddsubps %xmm9, %xmm8 , %xmm8
+ vaddsubps %xmm11,%xmm10, %xmm10
+
+
+
+#ifndef TRMMKERNEL
+
+ vaddps (CO1), %xmm8 , %xmm8
+
+ vaddps (CO1, LDC), %xmm10, %xmm10
+
+#endif
+
+ vmovups %xmm8 , (CO1)
+
+ vmovups %xmm10 , (CO1, LDC)
+
+
+
+#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ movq K, %rax
+ subq KKK, %rax
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 2, %rax // rax = rax * 4 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $ 2, KK
+#endif
+
+ addq $ 4 * SIZE, CO1 # coffset += 4
+ decq I # i --
+ jg .L2_4_21
+ ALIGN_4
+
+
+
+/**************************************************************************/
+.L2_4_40:
+ testq $ 1, M
+ jz .L2_4_60 // to next 2 lines of N
+
+ ALIGN_4
+
+.L2_4_41:
+
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 8 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (,BI,4), BI // BI = BI * 4 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 1, %rax // rax = rax * 2 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+ vzeroall
+
+#ifndef TRMMKERNEL
+ movq K, %rax
+#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ movq K, %rax
+ subq KK, %rax
+ movq %rax, KKK
+#else
+ movq KK, %rax
+#ifdef LEFT
+ addq $ 1, %rax // number of values in AO
+#else
+ addq $ 2, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L2_4_46
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+
+ salq $ 1, %rax // rax = rax * 2 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_4_42:
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+
+ je .L2_4_46
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+ KERNEL1x2_SUB
+
+ je .L2_4_46
+
+ jmp .L2_4_42
+ ALIGN_4
+
+.L2_4_46:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $ 7, %rax # if (k & 1)
+ je .L2_4_49
+
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+
+ salq $ 1, %rax // rax = rax * 2 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_4_47:
+
+ KERNEL1x2_SUB
+
+ jl .L2_4_47
+ ALIGN_4
+
+
+.L2_4_49:
+
+ SAVE1x2
+
+#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ movq K, %rax
+ subq KKK, %rax
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,4), BI // BI = BI * 4 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 1, %rax // rax = rax * 2 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $ 1, KK
+#endif
+
+ addq $ 2 * SIZE, CO1 # coffset += 2
+ decq I # i --
+ jg .L2_4_41
+ ALIGN_4
+
+
+
+
+.L2_4_60:
+#if defined(TRMMKERNEL) && !defined(LEFT)
+ addq $ 2, KK
+#endif
+
+ decq J // j --
+ jg .L2_01 // next 2 lines of N
+
+
+
+.L1_0:
+
+/************************************************************************************************
+* Loop for Nmod6 % 2 > 0
+*************************************************************************************************/
+
+ movq Nmod6, J
+ andq $ 1, J // j % 2
+ je .L999
+ ALIGN_4
+
+.L1_01:
+ // copy to sub buffer
+ movq B, BO1
+ leaq BUFFER1, BO // first buffer to BO
+ movq K, %rax
+ ALIGN_4
+
+.L1_02b:
+
+ vmovsd (BO1), %xmm0
+ vmovsd %xmm0, (BO)
+ addq $ 2*SIZE,BO1
+ addq $ 2*SIZE,BO
+ decq %rax
+ jnz .L1_02b
+
+.L1_02c:
+
+ movq BO1, B // next offset of B
+
+.L1_10:
+ movq C, CO1
+ leaq (C, LDC, 1), C // c += 1 * ldc
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ movq OFFSET, %rax
+ movq %rax, KK
+#endif
+
+ movq A, AO // aoffset = a
+ addq $ 16 * SIZE, AO
+
+ movq M, I
+ sarq $ 3, I // i = (m >> 3)
+ je .L1_4_10
+
+ ALIGN_4
+
+/**************************************************************************************************/
+
+.L1_8_11:
+
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 4 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 4 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (,BI,2), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 4, %rax // rax = rax *16 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+ vzeroall
+
+#ifndef TRMMKERNEL
+ movq K, %rax
+#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ movq K, %rax
+ subq KK, %rax
+ movq %rax, KKK
+#else
+ movq KK, %rax
+#ifdef LEFT
+ addq $ 8, %rax // number of values in AO
+#else
+ addq $ 1, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L1_8_16
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+
+ salq $ 4, %rax // rax = rax *16 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_8_12:
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+
+ je .L1_8_16
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL8x1_SUB
+
+ je .L1_8_16
+
+ jmp .L1_8_12
+ ALIGN_4
+
+.L1_8_16:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $ 7, %rax # if (k & 1)
+ je .L1_8_19
+
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 4 ; number of values
+
+ salq $ 4, %rax // rax = rax *16 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_8_17:
+
+ KERNEL8x1_SUB
+
+ jl .L1_8_17
+ ALIGN_4
+
+
+.L1_8_19:
+
+ SAVE8x1
+
+
+#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ movq K, %rax
+ subq KKK, %rax
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 4, %rax // rax = rax *16 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $ 8, KK
+#endif
+
+ addq $ 16 * SIZE, CO1 # coffset += 16
+ decq I # i --
+ jg .L1_8_11
+ ALIGN_4
+
+
+
+/**************************************************************************************************/
+.L1_4_10:
+
+ testq $ 7, M
+ jz .L999
+
+ testq $ 4, M
+ jz .L1_4_20
+
+
+.L1_4_11:
+
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 4 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 4 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (,BI,2), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 3, %rax // rax = rax * 8 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+ vzeroall
+
+#ifndef TRMMKERNEL
+ movq K, %rax
+#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ movq K, %rax
+ subq KK, %rax
+ movq %rax, KKK
+#else
+ movq KK, %rax
+#ifdef LEFT
+ addq $ 4, %rax // number of values in AO
+#else
+ addq $ 1, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L1_4_16
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+
+ salq $ 3, %rax // rax = rax * 8 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_4_12:
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL4x1_SUB
+ KERNEL4x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x1_SUB
+ KERNEL4x1_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x1_SUB
+ KERNEL4x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x1_SUB
+ KERNEL4x1_SUB
+
+ je .L1_4_16
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL4x1_SUB
+ KERNEL4x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x1_SUB
+ KERNEL4x1_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x1_SUB
+ KERNEL4x1_SUB
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL4x1_SUB
+ KERNEL4x1_SUB
+
+ je .L1_4_16
+
+ jmp .L1_4_12
+ ALIGN_4
+
+.L1_4_16:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $ 7, %rax # if (k & 1)
+ je .L1_4_19
+
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 4 ; number of values
+
+ salq $ 3, %rax // rax = rax * 8 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_4_17:
+
+ KERNEL4x1_SUB
+
+ jl .L1_4_17
+ ALIGN_4
+
+
+.L1_4_19:
+
+ SAVE4x1
+
+#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ movq K, %rax
+ subq KKK, %rax
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 3, %rax // rax = rax * 8 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $ 4, KK
+#endif
+
+ addq $ 8 * SIZE, CO1 # coffset += 8
+ ALIGN_4
+
+
+
+/**************************************************************************
+* Rest of M
+***************************************************************************/
+
+.L1_4_20:
+
+ testq $ 2, M
+ jz .L1_4_40
+ ALIGN_4
+
+.L1_4_21:
+
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 4 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 4 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (,BI,2), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 2, %rax // rax = rax * 4 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+ vzeroall
+
+#ifndef TRMMKERNEL
+ movq K, %rax
+#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ movq K, %rax
+ subq KK, %rax
+ movq %rax, KKK
+#else
+ movq KK, %rax
+#ifdef LEFT
+ addq $ 2, %rax // number of values in AO
+#else
+ addq $ 1, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L1_4_26
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+
+ salq $ 2, %rax // rax = rax * 4 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_4_22:
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+
+ je .L1_4_26
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+ KERNEL2x1_SUB
+
+ je .L1_4_26
+
+ jmp .L1_4_22
+ ALIGN_4
+
+.L1_4_26:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $ 7, %rax # if (k & 1)
+ je .L1_4_29
+
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2; number of values
+
+ salq $ 2, %rax // rax = rax * 4 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_4_27:
+
+ KERNEL2x1_SUB
+
+ jl .L1_4_27
+ ALIGN_4
+
+
+.L1_4_29:
+
+ SAVE2x1
+
+#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ movq K, %rax
+ subq KKK, %rax
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 2, %rax // rax = rax * 4 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $ 2, KK
+#endif
+
+ addq $ 4 * SIZE, CO1 # coffset += 4
+ ALIGN_4
+
+
+
+/**************************************************************************/
+.L1_4_40:
+ testq $ 1, M
+ jz .L999 // to next 2 lines of N
+
+ ALIGN_4
+
+.L1_4_41:
+
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 4 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $ 4 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (,BI,2), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 1, %rax // rax = rax * 2 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+ vzeroall
+
+#ifndef TRMMKERNEL
+ movq K, %rax
+#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
+ movq K, %rax
+ subq KK, %rax
+ movq %rax, KKK
+#else
+ movq KK, %rax
+#ifdef LEFT
+ addq $ 1, %rax // number of values in AO
+#else
+ addq $ 1, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $ -8, %rax // K = K - ( K % 8 )
+ je .L1_4_46
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+
+ salq $ 1, %rax // rax = rax * 2 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_4_42:
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+
+ je .L1_4_46
+
+ prefetcht0 A_PR1(AO,%rax,SIZE)
+ prefetcht0 B_PR1(BO,BI,SIZE)
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+ KERNEL1x1_SUB
+
+ je .L1_4_46
+
+ jmp .L1_4_42
+ ALIGN_4
+
+.L1_4_46:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $ 7, %rax # if (k & 1)
+ je .L1_4_49
+
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+
+ salq $ 1, %rax // rax = rax * 2 ; number of values
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_4_47:
+
+ KERNEL1x1_SUB
+
+ jl .L1_4_47
+ ALIGN_4
+
+
+.L1_4_49:
+
+ SAVE1x1
+
+
+
+#if (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ movq K, %rax
+ subq KKK, %rax
+ movq %rax, BI // Index for BO
+ leaq ( ,BI,2), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ salq $ 1, %rax // rax = rax * 2 ; number of values
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $ 1, KK
+#endif
+
+ addq $ 2 * SIZE, CO1 # coffset += 2
+ ALIGN_4
+
+
+.L999:
+ vzeroupper
+
+ movq SP, %rsp
+ movq (%rsp), %rbx
+ movq 8(%rsp), %rbp
+ movq 16(%rsp), %r12
+ movq 24(%rsp), %r13
+ movq 32(%rsp), %r14
+ movq 40(%rsp), %r15
+
+#ifdef WINDOWS_ABI
+ movq 48(%rsp), %rdi
+ movq 56(%rsp), %rsi
+ vmovups 64(%rsp), %xmm6
+ vmovups 80(%rsp), %xmm7
+ vmovups 96(%rsp), %xmm8
+ vmovups 112(%rsp), %xmm9
+ vmovups 128(%rsp), %xmm10
+ vmovups 144(%rsp), %xmm11
+ vmovups 160(%rsp), %xmm12
+ vmovups 176(%rsp), %xmm13
+ vmovups 192(%rsp), %xmm14
+ vmovups 208(%rsp), %xmm15
+#endif
+
+ addq $ STACKSIZE, %rsp
+ ret
+
+ EPILOGUE
+
+
+#endif
+
projects / platform / upstream / openblas.git / commitdiff