--- /dev/null
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin. */
+/* All rights reserved. */
+/* */
+/* Redistribution and use in source and binary forms, with or */
+/* without modification, are permitted provided that the following */
+/* conditions are met: */
+/* */
+/* 1. Redistributions of source code must retain the above */
+/* copyright notice, this list of conditions and the following */
+/* disclaimer. */
+/* */
+/* 2. Redistributions in binary form must reproduce the above */
+/* copyright notice, this list of conditions and the following */
+/* disclaimer in the documentation and/or other materials */
+/* provided with the distribution. */
+/* */
+/* THIS SOFTWARE IS PROVIDED BY THE UNIVERSITY OF TEXAS AT */
+/* AUSTIN ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, */
+/* INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF */
+/* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE */
+/* DISCLAIMED. IN NO EVENT SHALL THE UNIVERSITY OF TEXAS AT */
+/* AUSTIN OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, */
+/* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES */
+/* (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE */
+/* GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR */
+/* BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF */
+/* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT */
+/* (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT */
+/* OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE */
+/* POSSIBILITY OF SUCH DAMAGE. */
+/* */
+/* The views and conclusions contained in the software and */
+/* documentation are those of the authors and should not be */
+/* interpreted as representing official policies, either expressed */
+/* or implied, of The University of Texas at Austin. */
+/*********************************************************************/
+
+/*********************************************************************
+* 2013/10/20 Saar
+* BLASTEST : OK
+* CTEST : OK
+* TEST : OK
+
+*
+*
+* 2013/10/20 Saar
+* Parameter:
+* DGEMM_DEFAULT_UNROLL_N 2
+* DGEMM_DEFAULT_UNROLL_M 16
+* DGEMM_DEFAULT_P 192
+* DGEMM_DEFAULT_Q 128
+* A_PR1 512
+*
+*
+* Performance without prefetch of B:
+* 1 thread: 45.8 GFLOPS (MKL: 45)
+* 2 threads: 80.0 GFLOPS (MKL: 91)
+* 4 threads: 135.0 GFLOPS (MKL: 135)
+*********************************************************************/
+
+
+#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 %r15
+
+#ifndef WINDOWS_ABI
+
+#define STACKSIZE 96
+
+#else
+
+#define STACKSIZE 256
+
+#define OLD_A 40 + STACKSIZE(%rsp)
+#define OLD_B 48 + STACKSIZE(%rsp)
+#define OLD_C 56 + STACKSIZE(%rsp)
+#define OLD_LDC 64 + STACKSIZE(%rsp)
+#define OLD_OFFSET 72 + STACKSIZE(%rsp)
+
+#endif
+
+#define L_BUFFER_SIZE 512*8*4
+#define LB2_OFFSET 512*8*2
+
+#define Ndiv6 24(%rsp)
+#define Nmod6 32(%rsp)
+#define N 40(%rsp)
+#define ALPHA 48(%rsp)
+#define OFFSET 56(%rsp)
+#define KK 64(%rsp)
+#define KKK 72(%rsp)
+#define BUFFER1 128(%rsp)
+#define BUFFER2 LB2_OFFSET+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)
+
+.macro VFMADD231PD_ y0,y1,y2
+ vfmaddpd \y0,\y1,\y2,\y0
+.endm
+
+.macro VFMADD231SD_ x0,x1,x2
+ vfmaddsd \x0,\x1,\x2,\x0
+.endm
+
+#else
+
+.macro VFMADD231PD_ y0,y1,y2
+ vfmadd231pd \y2,\y1,\y0
+.endm
+
+.macro VFMADD231SD_ x0,x1,x2
+ vfmadd231sd \x2,\x1,\x0
+.endm
+
+#endif
+
+
+#define A_PR1 1024
+#define B_PR1 256
+
+/*******************************************************************************************
+* 3 lines of N
+*******************************************************************************************/
+
+.macro KERNEL16x3_SUBN
+ vbroadcastsd -12 * SIZE(BO), %zmm1
+ vbroadcastsd -11 * SIZE(BO), %zmm2
+ vbroadcastsd -10 * SIZE(BO), %zmm3
+
+ vmovaps -16 * SIZE(AO), %zmm0
+ VFMADD231PD_ %zmm4,%zmm1,%zmm0
+ VFMADD231PD_ %zmm5,%zmm2,%zmm0
+ VFMADD231PD_ %zmm6,%zmm3,%zmm0
+
+ vmovaps -8 * SIZE(AO), %zmm9
+ VFMADD231PD_ %zmm10,%zmm1,%zmm9
+ VFMADD231PD_ %zmm11,%zmm2,%zmm9
+ VFMADD231PD_ %zmm12,%zmm3,%zmm9
+ addq $ 3*SIZE , BO
+ addq $ 16*SIZE, AO
+.endm
+
+
+.macro KERNEL8x3_SUBN
+ vbroadcastsd -12 * SIZE(BO), %ymm1
+ vmovaps -16 * SIZE(AO), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -11 * SIZE(BO), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vbroadcastsd -10 * SIZE(BO), %ymm3
+ VFMADD231PD_ %ymm6,%ymm3,%ymm0
+ vmovaps -12 * SIZE(AO), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ VFMADD231PD_ %ymm8,%ymm2,%ymm0
+ VFMADD231PD_ %ymm9,%ymm3,%ymm0
+ prefetcht0 B_PR1(BO)
+ addq $ 3*SIZE , BO
+ addq $ 8*SIZE, AO
+.endm
+
+.macro KERNEL4x3_SUBN
+ vbroadcastsd -12 * SIZE(BO), %ymm1
+ vmovaps -16 * SIZE(AO), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -11 * SIZE(BO), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vbroadcastsd -10 * SIZE(BO), %ymm3
+ VFMADD231PD_ %ymm6,%ymm3,%ymm0
+ addq $ 3*SIZE , BO
+ addq $ 4*SIZE, AO
+.endm
+
+.macro KERNEL2x3_SUBN
+ vmovsd -12 * SIZE(BO), %xmm1
+ vmovsd -16 * SIZE(AO), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -11 * SIZE(BO), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd -10 * SIZE(BO), %xmm3
+ VFMADD231SD_ %xmm6,%xmm3,%xmm0
+ vmovsd -15 * SIZE(AO), %xmm0
+ VFMADD231SD_ %xmm8,%xmm1,%xmm0
+ VFMADD231SD_ %xmm10,%xmm2,%xmm0
+ VFMADD231SD_ %xmm12,%xmm3,%xmm0
+ addq $ 3*SIZE , BO
+ addq $ 2*SIZE, AO
+.endm
+
+.macro KERNEL1x3_SUBN
+ vmovsd -12 * SIZE(BO), %xmm1
+ vmovsd -16 * SIZE(AO), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -11 * SIZE(BO), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd -10 * SIZE(BO), %xmm3
+ VFMADD231SD_ %xmm6,%xmm3,%xmm0
+ addq $ 3*SIZE , BO
+ addq $ 1*SIZE, AO
+.endm
+
+
+
+
+
+
+/******************************************************************************************/
+
+.macro KERNEL16x3_1
+ vbroadcastsd -6 * SIZE(BO, BI, SIZE), %zmm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %zmm0
+ VFMADD231PD_ %zmm4,%zmm1,%zmm0
+ vbroadcastsd -5 * SIZE(BO, BI, SIZE), %zmm2
+ VFMADD231PD_ %zmm5,%zmm2,%zmm0
+ vbroadcastsd -4 * SIZE(BO, BI, SIZE), %zmm3
+ VFMADD231PD_ %zmm6,%zmm3,%zmm0
+ vmovups -24 * SIZE(AO, %rax, SIZE), %zmm0
+ VFMADD231PD_ %zmm10,%zmm1,%zmm0
+ VFMADD231PD_ %zmm11,%zmm2,%zmm0
+ VFMADD231PD_ %zmm12,%zmm3,%zmm0
+ vbroadcastsd -3 * SIZE(BO, BI, SIZE), %zmm1
+ vbroadcastsd -2 * SIZE(BO, BI, SIZE), %zmm2
+.endm
+
+
+
+
+.macro KERNEL16x3_2
+ vmovups -16 * SIZE(AO, %rax, SIZE), %zmm0
+ VFMADD231PD_ %zmm4,%zmm1,%zmm0
+ VFMADD231PD_ %zmm5,%zmm2,%zmm0
+ vbroadcastsd -1 * SIZE(BO, BI, SIZE), %zmm3
+ VFMADD231PD_ %zmm6,%zmm3,%zmm0
+ vmovups -8 * SIZE(AO, %rax, SIZE), %zmm0
+ VFMADD231PD_ %zmm10,%zmm1,%zmm0
+ VFMADD231PD_ %zmm11,%zmm2,%zmm0
+ VFMADD231PD_ %zmm12,%zmm3,%zmm0
+ vbroadcastsd 0 * SIZE(BO, BI, SIZE), %zmm1
+ vbroadcastsd 1 * SIZE(BO, BI, SIZE), %zmm2
+.endm
+
+.macro KERNEL16x3_3
+ vmovups 0 * SIZE(AO, %rax, SIZE), %zmm0
+ VFMADD231PD_ %zmm4,%zmm1,%zmm0
+ VFMADD231PD_ %zmm5,%zmm2,%zmm0
+ vbroadcastsd 2 * SIZE(BO, BI, SIZE), %zmm3
+ VFMADD231PD_ %zmm6,%zmm3,%zmm0
+ vmovups 8 * SIZE(AO, %rax, SIZE), %zmm0
+ VFMADD231PD_ %zmm10,%zmm1,%zmm0
+ VFMADD231PD_ %zmm11,%zmm2,%zmm0
+ VFMADD231PD_ %zmm12,%zmm3,%zmm0
+ vbroadcastsd 3 * SIZE(BO, BI, SIZE), %zmm1
+ vbroadcastsd 4 * SIZE(BO, BI, SIZE), %zmm2
+.endm
+
+.macro KERNEL16x3_4
+ vmovups 16 * SIZE(AO, %rax, SIZE), %zmm0
+ VFMADD231PD_ %zmm4,%zmm1,%zmm0
+ VFMADD231PD_ %zmm5,%zmm2,%zmm0
+ vbroadcastsd 5 * SIZE(BO, BI, SIZE), %zmm3
+ VFMADD231PD_ %zmm6,%zmm3,%zmm0
+ vmovups 24 * SIZE(AO, %rax, SIZE), %zmm0
+ VFMADD231PD_ %zmm10,%zmm1,%zmm0
+ VFMADD231PD_ %zmm11,%zmm2,%zmm0
+ addq $12, BI
+ VFMADD231PD_ %zmm12,%zmm3,%zmm0
+ addq $64, %rax
+.endm
+
+.macro KERNEL16x3_SUB
+ vbroadcastsd -6 * SIZE(BO, BI, SIZE), %zmm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %zmm0
+ VFMADD231PD_ %zmm4,%zmm1,%zmm0
+ vbroadcastsd -5 * SIZE(BO, BI, SIZE), %zmm2
+ VFMADD231PD_ %zmm5,%zmm2,%zmm0
+ vbroadcastsd -4 * SIZE(BO, BI, SIZE), %zmm3
+ VFMADD231PD_ %zmm6,%zmm3,%zmm0
+ vmovups -24 * SIZE(AO, %rax, SIZE), %zmm0
+ VFMADD231PD_ %zmm10,%zmm1,%zmm0
+ VFMADD231PD_ %zmm11,%zmm2,%zmm0
+ VFMADD231PD_ %zmm12,%zmm3,%zmm0
+ addq $3 , BI
+ addq $16, %rax
+.endm
+
+.macro SAVE16x3
+
+ vbroadcastsd ALPHA, %zmm0
+
+ vmulpd %zmm0 , %zmm4 , %zmm4
+ vmulpd %zmm0 , %zmm10, %zmm10
+
+ vmulpd %zmm0 , %zmm5 , %zmm5
+ vmulpd %zmm0 , %zmm11, %zmm11
+
+ vmulpd %zmm0 , %zmm6 , %zmm6
+ vmulpd %zmm0 , %zmm12, %zmm12
+
+#if !defined(TRMMKERNEL)
+
+ vaddpd (CO1), %zmm4,%zmm4
+ vaddpd 8 * SIZE(CO1), %zmm10,%zmm10
+
+ vaddpd (CO1, LDC), %zmm5,%zmm5
+ vaddpd 8 * SIZE(CO1, LDC), %zmm11,%zmm11
+
+ vaddpd (CO1, LDC, 2), %zmm6,%zmm6
+ vaddpd 8 * SIZE(CO1, LDC, 2), %zmm12,%zmm12
+
+#endif
+
+ vmovups %zmm4 , (CO1)
+ vmovups %zmm10, 8 * SIZE(CO1)
+
+ vmovups %zmm5 , (CO1, LDC)
+ vmovups %zmm11, 8 * SIZE(CO1, LDC)
+
+ vmovups %zmm6 , (CO1, LDC, 2)
+ vmovups %zmm12, 8 * SIZE(CO1, LDC, 2)
+
+.endm
+
+
+
+/*******************************************************************************************/
+
+.macro KERNEL8x3_1
+ prefetcht0 A_PR1(AO, %rax, SIZE)
+ vbroadcastsd -6 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -5 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vbroadcastsd -4 * SIZE(BO, BI, SIZE), %ymm3
+ VFMADD231PD_ %ymm6,%ymm3,%ymm0
+ vmovups -28 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ VFMADD231PD_ %ymm8,%ymm2,%ymm0
+ VFMADD231PD_ %ymm9,%ymm3,%ymm0
+.endm
+
+.macro KERNEL8x3_2
+ prefetcht0 64+A_PR1(AO, %rax, SIZE)
+ vbroadcastsd -3 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -24 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -2 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vbroadcastsd -1 * SIZE(BO, BI, SIZE), %ymm3
+ VFMADD231PD_ %ymm6,%ymm3,%ymm0
+ vmovups -20 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ VFMADD231PD_ %ymm8,%ymm2,%ymm0
+ VFMADD231PD_ %ymm9,%ymm3,%ymm0
+.endm
+
+.macro KERNEL8x3_3
+ prefetcht0 128+A_PR1(AO, %rax, SIZE)
+ vbroadcastsd 0 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -16 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd 1 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vbroadcastsd 2 * SIZE(BO, BI, SIZE), %ymm3
+ VFMADD231PD_ %ymm6,%ymm3,%ymm0
+ vmovups -12 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ VFMADD231PD_ %ymm8,%ymm2,%ymm0
+ VFMADD231PD_ %ymm9,%ymm3,%ymm0
+.endm
+
+.macro KERNEL8x3_4
+ prefetcht0 192+A_PR1(AO, %rax, SIZE)
+ vbroadcastsd 3 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -8 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd 4 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vbroadcastsd 5 * SIZE(BO, BI, SIZE), %ymm3
+ VFMADD231PD_ %ymm6,%ymm3,%ymm0
+ vmovups -4 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ VFMADD231PD_ %ymm8,%ymm2,%ymm0
+ VFMADD231PD_ %ymm9,%ymm3,%ymm0
+ addq $12, BI
+ addq $32, %rax
+.endm
+
+.macro KERNEL8x3_SUB
+ vbroadcastsd -6 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -5 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vbroadcastsd -4 * SIZE(BO, BI, SIZE), %ymm3
+ VFMADD231PD_ %ymm6,%ymm3,%ymm0
+ vmovups -28 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ VFMADD231PD_ %ymm8,%ymm2,%ymm0
+ VFMADD231PD_ %ymm9,%ymm3,%ymm0
+ addq $3 , BI
+ addq $8 , %rax
+.endm
+
+.macro SAVE8x3
+
+ vbroadcastsd ALPHA, %ymm0
+
+ vmulpd %ymm0 , %ymm4 , %ymm4
+ vmulpd %ymm0 , %ymm7 , %ymm7
+
+ vmulpd %ymm0 , %ymm5 , %ymm5
+ vmulpd %ymm0 , %ymm8 , %ymm8
+
+ vmulpd %ymm0 , %ymm6 , %ymm6
+ vmulpd %ymm0 , %ymm9 , %ymm9
+
+#if !defined(TRMMKERNEL)
+
+ vaddpd (CO1), %ymm4,%ymm4
+ vaddpd 4 * SIZE(CO1), %ymm7,%ymm7
+
+ vaddpd (CO1, LDC), %ymm5,%ymm5
+ vaddpd 4 * SIZE(CO1, LDC), %ymm8,%ymm8
+
+ vaddpd (CO1, LDC, 2), %ymm6,%ymm6
+ vaddpd 4 * SIZE(CO1, LDC, 2), %ymm9,%ymm9
+
+#endif
+
+ vmovups %ymm4 , (CO1)
+ vmovups %ymm7 , 4 * SIZE(CO1)
+
+ vmovups %ymm5 , (CO1, LDC)
+ vmovups %ymm8 , 4 * SIZE(CO1, LDC)
+
+ vmovups %ymm6 , (CO1, LDC, 2)
+ vmovups %ymm9 , 4 * SIZE(CO1, LDC, 2)
+
+.endm
+
+
+
+/*******************************************************************************************/
+
+.macro KERNEL4x3_1
+ prefetcht0 A_PR1(AO, %rax, SIZE)
+ vbroadcastsd -6 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -5 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vbroadcastsd -4 * SIZE(BO, BI, SIZE), %ymm3
+ VFMADD231PD_ %ymm6,%ymm3,%ymm0
+.endm
+
+.macro KERNEL4x3_2
+ vbroadcastsd -3 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -28 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -2 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vbroadcastsd -1 * SIZE(BO, BI, SIZE), %ymm3
+ VFMADD231PD_ %ymm6,%ymm3,%ymm0
+.endm
+
+.macro KERNEL4x3_3
+ prefetcht0 A_PR1(AO, %rax, SIZE)
+ vbroadcastsd 0 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -24 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd 1 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vbroadcastsd 2 * SIZE(BO, BI, SIZE), %ymm3
+ VFMADD231PD_ %ymm6,%ymm3,%ymm0
+.endm
+
+.macro KERNEL4x3_4
+ vbroadcastsd 3 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -20 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd 4 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vbroadcastsd 5 * SIZE(BO, BI, SIZE), %ymm3
+ VFMADD231PD_ %ymm6,%ymm3,%ymm0
+ addq $12, BI
+ addq $16, %rax
+.endm
+
+.macro KERNEL4x3_SUB
+ vbroadcastsd -6 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -5 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vbroadcastsd -4 * SIZE(BO, BI, SIZE), %ymm3
+ VFMADD231PD_ %ymm6,%ymm3,%ymm0
+ addq $3 , BI
+ addq $4 , %rax
+.endm
+
+.macro SAVE4x3
+
+ vbroadcastsd ALPHA, %ymm0
+
+ vmulpd %ymm0 , %ymm4 , %ymm4
+ vmulpd %ymm0 , %ymm5 , %ymm5
+ vmulpd %ymm0 , %ymm6 , %ymm6
+
+#if !defined(TRMMKERNEL)
+
+ vaddpd (CO1), %ymm4,%ymm4
+ vaddpd (CO1, LDC), %ymm5,%ymm5
+ vaddpd (CO1, LDC, 2), %ymm6,%ymm6
+
+#endif
+
+ vmovups %ymm4 , (CO1)
+ vmovups %ymm5 , (CO1, LDC)
+ vmovups %ymm6 , (CO1, LDC, 2)
+
+.endm
+
+
+/*******************************************************************************************/
+
+.macro KERNEL2x3_1
+ prefetcht0 A_PR1(AO, %rax, SIZE)
+ vmovsd -6 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -32 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -5 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd -4 * SIZE(BO, BI, SIZE), %xmm3
+ VFMADD231SD_ %xmm6,%xmm3,%xmm0
+ vmovsd -31 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm8,%xmm1,%xmm0
+ VFMADD231SD_ %xmm10,%xmm2,%xmm0
+ VFMADD231SD_ %xmm12,%xmm3,%xmm0
+.endm
+
+.macro KERNEL2x3_2
+ vmovsd -3 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -30 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -2 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd -1 * SIZE(BO, BI, SIZE), %xmm3
+ VFMADD231SD_ %xmm6,%xmm3,%xmm0
+ vmovsd -29 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm8,%xmm1,%xmm0
+ VFMADD231SD_ %xmm10,%xmm2,%xmm0
+ VFMADD231SD_ %xmm12,%xmm3,%xmm0
+.endm
+
+.macro KERNEL2x3_3
+ vmovsd 0 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -28 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd 1 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd 2 * SIZE(BO, BI, SIZE), %xmm3
+ VFMADD231SD_ %xmm6,%xmm3,%xmm0
+ vmovsd -27 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm8,%xmm1,%xmm0
+ VFMADD231SD_ %xmm10,%xmm2,%xmm0
+ VFMADD231SD_ %xmm12,%xmm3,%xmm0
+.endm
+
+.macro KERNEL2x3_4
+ vmovsd 3 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -26 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd 4 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd 5 * SIZE(BO, BI, SIZE), %xmm3
+ VFMADD231SD_ %xmm6,%xmm3,%xmm0
+ vmovsd -25 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm8,%xmm1,%xmm0
+ VFMADD231SD_ %xmm10,%xmm2,%xmm0
+ VFMADD231SD_ %xmm12,%xmm3,%xmm0
+ addq $12, BI
+ addq $8, %rax
+.endm
+
+.macro KERNEL2x3_SUB
+ vmovsd -6 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -32 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -5 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd -4 * SIZE(BO, BI, SIZE), %xmm3
+ VFMADD231SD_ %xmm6,%xmm3,%xmm0
+ vmovsd -31 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm8,%xmm1,%xmm0
+ VFMADD231SD_ %xmm10,%xmm2,%xmm0
+ VFMADD231SD_ %xmm12,%xmm3,%xmm0
+ addq $3 , BI
+ addq $2 , %rax
+.endm
+
+.macro SAVE2x3
+
+ vmovsd ALPHA, %xmm0
+
+ vmulsd %xmm0 , %xmm4 , %xmm4
+ vmulsd %xmm0 , %xmm8 , %xmm8
+ vmulsd %xmm0 , %xmm5 , %xmm5
+ vmulsd %xmm0 , %xmm10, %xmm10
+ vmulsd %xmm0 , %xmm6 , %xmm6
+ vmulsd %xmm0 , %xmm12, %xmm12
+
+#if !defined(TRMMKERNEL)
+
+ vaddsd (CO1), %xmm4,%xmm4
+ vaddsd 1 * SIZE(CO1), %xmm8,%xmm8
+ vaddsd (CO1, LDC), %xmm5,%xmm5
+ vaddsd 1 * SIZE(CO1, LDC), %xmm10,%xmm10
+ vaddsd (CO1, LDC, 2), %xmm6,%xmm6
+ vaddsd 1 * SIZE(CO1, LDC, 2), %xmm12,%xmm12
+
+#endif
+
+ vmovsd %xmm4 , (CO1)
+ vmovsd %xmm8 , 1 * SIZE(CO1)
+ vmovsd %xmm5 , (CO1, LDC)
+ vmovsd %xmm10, 1 * SIZE(CO1, LDC)
+ vmovsd %xmm6 , (CO1, LDC, 2)
+ vmovsd %xmm12, 1 * SIZE(CO1, LDC, 2)
+
+.endm
+
+/*******************************************************************************************/
+
+.macro KERNEL1x3_1
+ vmovsd -6 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -32 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -5 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd -4 * SIZE(BO, BI, SIZE), %xmm3
+ VFMADD231SD_ %xmm6,%xmm3,%xmm0
+.endm
+
+.macro KERNEL1x3_2
+ vmovsd -3 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -31 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -2 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd -1 * SIZE(BO, BI, SIZE), %xmm3
+ VFMADD231SD_ %xmm6,%xmm3,%xmm0
+.endm
+
+.macro KERNEL1x3_3
+ vmovsd 0 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -30 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd 1 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd 2 * SIZE(BO, BI, SIZE), %xmm3
+ VFMADD231SD_ %xmm6,%xmm3,%xmm0
+.endm
+
+.macro KERNEL1x3_4
+ vmovsd 3 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -29 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd 4 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd 5 * SIZE(BO, BI, SIZE), %xmm3
+ VFMADD231SD_ %xmm6,%xmm3,%xmm0
+ addq $12, BI
+ addq $4, %rax
+.endm
+
+.macro KERNEL1x3_SUB
+ vmovsd -6 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -32 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -5 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd -4 * SIZE(BO, BI, SIZE), %xmm3
+ VFMADD231SD_ %xmm6,%xmm3,%xmm0
+ addq $3 , BI
+ addq $1 , %rax
+.endm
+
+.macro SAVE1x3
+
+ vmovsd ALPHA, %xmm0
+
+ vmulsd %xmm0 , %xmm4 , %xmm4
+ vmulsd %xmm0 , %xmm5 , %xmm5
+ vmulsd %xmm0 , %xmm6 , %xmm6
+
+#if !defined(TRMMKERNEL)
+
+ vaddsd (CO1), %xmm4,%xmm4
+ vaddsd (CO1, LDC), %xmm5,%xmm5
+ vaddsd (CO1, LDC, 2), %xmm6,%xmm6
+
+#endif
+
+ vmovsd %xmm4 , (CO1)
+ vmovsd %xmm5 , (CO1, LDC)
+ vmovsd %xmm6 , (CO1, LDC, 2)
+
+.endm
+
+
+/*******************************************************************************************/
+
+/*******************************************************************************************
+* 2 lines of N
+*******************************************************************************************/
+
+.macro KERNEL16x2_1
+ prefetcht0 A_PR1(AO, %rax, SIZE)
+ vbroadcastsd -4 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -3 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vmovups -28 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ prefetcht0 64+A_PR1(AO, %rax, SIZE)
+ VFMADD231PD_ %ymm8,%ymm2,%ymm0
+ vmovups -24 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm10,%ymm1,%ymm0
+ VFMADD231PD_ %ymm11,%ymm2,%ymm0
+ vmovups -20 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm13,%ymm1,%ymm0
+ VFMADD231PD_ %ymm14,%ymm2,%ymm0
+.endm
+
+.macro KERNEL16x2_2
+ prefetcht0 128+A_PR1(AO, %rax, SIZE)
+ vbroadcastsd -2 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -16 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -1 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vmovups -12 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ prefetcht0 192+A_PR1(AO, %rax, SIZE)
+ VFMADD231PD_ %ymm8,%ymm2,%ymm0
+ vmovups -8 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm10,%ymm1,%ymm0
+ VFMADD231PD_ %ymm11,%ymm2,%ymm0
+ vmovups -4 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm13,%ymm1,%ymm0
+ VFMADD231PD_ %ymm14,%ymm2,%ymm0
+.endm
+
+.macro KERNEL16x2_3
+ prefetcht0 256+A_PR1(AO, %rax, SIZE)
+ vbroadcastsd 0 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups 0 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd 1 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vmovups 4 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ prefetcht0 320+A_PR1(AO, %rax, SIZE)
+ VFMADD231PD_ %ymm8,%ymm2,%ymm0
+ vmovups 8 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm10,%ymm1,%ymm0
+ VFMADD231PD_ %ymm11,%ymm2,%ymm0
+ vmovups 12 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm13,%ymm1,%ymm0
+ VFMADD231PD_ %ymm14,%ymm2,%ymm0
+.endm
+
+.macro KERNEL16x2_4
+ prefetcht0 384+A_PR1(AO, %rax, SIZE)
+ vbroadcastsd 2 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups 16 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd 3 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vmovups 20 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ prefetcht0 448+A_PR1(AO, %rax, SIZE)
+ VFMADD231PD_ %ymm8,%ymm2,%ymm0
+ vmovups 24 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm10,%ymm1,%ymm0
+ VFMADD231PD_ %ymm11,%ymm2,%ymm0
+ vmovups 28 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm13,%ymm1,%ymm0
+ VFMADD231PD_ %ymm14,%ymm2,%ymm0
+ addq $8, BI
+ addq $64, %rax
+.endm
+
+.macro KERNEL16x2_SUB
+ vbroadcastsd -4 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -3 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vmovups -28 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ VFMADD231PD_ %ymm8,%ymm2,%ymm0
+ vmovups -24 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm10,%ymm1,%ymm0
+ VFMADD231PD_ %ymm11,%ymm2,%ymm0
+ vmovups -20 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm13,%ymm1,%ymm0
+ VFMADD231PD_ %ymm14,%ymm2,%ymm0
+ addq $2, BI
+ addq $16, %rax
+.endm
+
+.macro SAVE16x2
+
+ vbroadcastsd ALPHA, %ymm0
+
+ vmulpd %ymm0 , %ymm4 , %ymm4
+ vmulpd %ymm0 , %ymm7 , %ymm7
+ vmulpd %ymm0 , %ymm10, %ymm10
+ vmulpd %ymm0 , %ymm13, %ymm13
+
+ vmulpd %ymm0 , %ymm5 , %ymm5
+ vmulpd %ymm0 , %ymm8 , %ymm8
+ vmulpd %ymm0 , %ymm11, %ymm11
+ vmulpd %ymm0 , %ymm14, %ymm14
+
+#if !defined(TRMMKERNEL)
+
+ vaddpd (CO1), %ymm4,%ymm4
+ vaddpd 4 * SIZE(CO1), %ymm7,%ymm7
+ vaddpd 8 * SIZE(CO1), %ymm10,%ymm10
+ vaddpd 12 * SIZE(CO1), %ymm13,%ymm13
+
+ vaddpd (CO1, LDC), %ymm5,%ymm5
+ vaddpd 4 * SIZE(CO1, LDC), %ymm8,%ymm8
+ vaddpd 8 * SIZE(CO1, LDC), %ymm11,%ymm11
+ vaddpd 12 * SIZE(CO1, LDC), %ymm14,%ymm14
+
+#endif
+
+ vmovups %ymm4 , (CO1)
+ vmovups %ymm7 , 4 * SIZE(CO1)
+ vmovups %ymm10, 8 * SIZE(CO1)
+ vmovups %ymm13,12 * SIZE(CO1)
+
+ vmovups %ymm5 , (CO1, LDC)
+ vmovups %ymm8 , 4 * SIZE(CO1, LDC)
+ vmovups %ymm11, 8 * SIZE(CO1, LDC)
+ vmovups %ymm14,12 * SIZE(CO1, LDC)
+
+.endm
+
+
+
+/*******************************************************************************************/
+
+.macro KERNEL8x2_1
+ prefetcht0 A_PR1(AO, %rax, SIZE)
+ vbroadcastsd -4 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -3 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vmovups -28 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ VFMADD231PD_ %ymm8,%ymm2,%ymm0
+.endm
+
+.macro KERNEL8x2_2
+ prefetcht0 64+A_PR1(AO, %rax, SIZE)
+ vbroadcastsd -2 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -24 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -1 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vmovups -20 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ VFMADD231PD_ %ymm8,%ymm2,%ymm0
+.endm
+
+.macro KERNEL8x2_3
+ prefetcht0 128+A_PR1(AO, %rax, SIZE)
+ vbroadcastsd 0 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -16 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd 1 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vmovups -12 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ VFMADD231PD_ %ymm8,%ymm2,%ymm0
+.endm
+
+.macro KERNEL8x2_4
+ prefetcht0 192+A_PR1(AO, %rax, SIZE)
+ vbroadcastsd 2 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -8 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd 3 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vmovups -4 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ VFMADD231PD_ %ymm8,%ymm2,%ymm0
+ addq $8, BI
+ addq $32, %rax
+.endm
+
+.macro KERNEL8x2_SUB
+ vbroadcastsd -4 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -3 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ vmovups -28 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ VFMADD231PD_ %ymm8,%ymm2,%ymm0
+ addq $2, BI
+ addq $8 , %rax
+.endm
+
+.macro SAVE8x2
+
+ vbroadcastsd ALPHA, %ymm0
+
+ vmulpd %ymm0 , %ymm4 , %ymm4
+ vmulpd %ymm0 , %ymm7 , %ymm7
+
+ vmulpd %ymm0 , %ymm5 , %ymm5
+ vmulpd %ymm0 , %ymm8 , %ymm8
+
+#if !defined(TRMMKERNEL)
+
+ vaddpd (CO1), %ymm4,%ymm4
+ vaddpd 4 * SIZE(CO1), %ymm7,%ymm7
+
+ vaddpd (CO1, LDC), %ymm5,%ymm5
+ vaddpd 4 * SIZE(CO1, LDC), %ymm8,%ymm8
+
+#endif
+
+ vmovups %ymm4 , (CO1)
+ vmovups %ymm7 , 4 * SIZE(CO1)
+
+ vmovups %ymm5 , (CO1, LDC)
+ vmovups %ymm8 , 4 * SIZE(CO1, LDC)
+
+.endm
+
+
+
+/*******************************************************************************************/
+
+.macro KERNEL4x2_1
+ prefetcht0 A_PR1(AO, %rax, SIZE)
+ vbroadcastsd -4 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -3 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+.endm
+
+.macro KERNEL4x2_2
+ vbroadcastsd -2 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -28 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -1 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+.endm
+
+.macro KERNEL4x2_3
+ prefetcht0 64+A_PR1(AO, %rax, SIZE)
+ vbroadcastsd 0 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -24 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd 1 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+.endm
+
+.macro KERNEL4x2_4
+ vbroadcastsd 2 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -20 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd 3 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ addq $8, BI
+ addq $16, %rax
+.endm
+
+.macro KERNEL4x2_SUB
+ vbroadcastsd -4 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vbroadcastsd -3 * SIZE(BO, BI, SIZE), %ymm2
+ VFMADD231PD_ %ymm5,%ymm2,%ymm0
+ addq $2, BI
+ addq $4 , %rax
+.endm
+
+.macro SAVE4x2
+
+ vbroadcastsd ALPHA, %ymm0
+
+ vmulpd %ymm0 , %ymm4 , %ymm4
+ vmulpd %ymm0 , %ymm5 , %ymm5
+
+#if !defined(TRMMKERNEL)
+
+ vaddpd (CO1), %ymm4,%ymm4
+ vaddpd (CO1, LDC), %ymm5,%ymm5
+
+#endif
+
+ vmovups %ymm4 , (CO1)
+ vmovups %ymm5 , (CO1, LDC)
+
+.endm
+
+
+/*******************************************************************************************/
+
+.macro KERNEL2x2_1
+ prefetcht0 A_PR1(AO, %rax, SIZE)
+ vmovsd -4 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -32 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -3 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd -31 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm8,%xmm1,%xmm0
+ VFMADD231SD_ %xmm10,%xmm2,%xmm0
+.endm
+
+.macro KERNEL2x2_2
+ vmovsd -2 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -30 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -1 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd -29 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm8,%xmm1,%xmm0
+ VFMADD231SD_ %xmm10,%xmm2,%xmm0
+.endm
+
+.macro KERNEL2x2_3
+ vmovsd 0 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -28 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd 1 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd -27 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm8,%xmm1,%xmm0
+ VFMADD231SD_ %xmm10,%xmm2,%xmm0
+.endm
+
+.macro KERNEL2x2_4
+ vmovsd 2 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -26 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd 3 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd -25 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm8,%xmm1,%xmm0
+ VFMADD231SD_ %xmm10,%xmm2,%xmm0
+ addq $8, BI
+ addq $8, %rax
+.endm
+
+.macro KERNEL2x2_SUB
+ vmovsd -4 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -32 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -3 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ vmovsd -31 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm8,%xmm1,%xmm0
+ VFMADD231SD_ %xmm10,%xmm2,%xmm0
+ addq $2, BI
+ addq $2, %rax
+.endm
+
+.macro SAVE2x2
+
+ vmovsd ALPHA, %xmm0
+
+ vmulsd %xmm0 , %xmm4 , %xmm4
+ vmulsd %xmm0 , %xmm8 , %xmm8
+ vmulsd %xmm0 , %xmm5 , %xmm5
+ vmulsd %xmm0 , %xmm10, %xmm10
+
+#if !defined(TRMMKERNEL)
+
+ vaddsd (CO1), %xmm4,%xmm4
+ vaddsd 1 * SIZE(CO1), %xmm8,%xmm8
+ vaddsd (CO1, LDC), %xmm5,%xmm5
+ vaddsd 1 * SIZE(CO1, LDC), %xmm10,%xmm10
+
+#endif
+
+ vmovsd %xmm4 , (CO1)
+ vmovsd %xmm8 , 1 * SIZE(CO1)
+ vmovsd %xmm5 , (CO1, LDC)
+ vmovsd %xmm10, 1 * SIZE(CO1, LDC)
+
+.endm
+
+
+/*******************************************************************************************/
+
+.macro KERNEL1x2_1
+ vmovsd -4 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -32 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -3 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+.endm
+
+.macro KERNEL1x2_2
+ vmovsd -2 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -31 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -1 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+.endm
+
+.macro KERNEL1x2_3
+ vmovsd 0 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -30 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd 1 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+.endm
+
+.macro KERNEL1x2_4
+ vmovsd 2 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -29 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd 3 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ addq $8, BI
+ addq $4, %rax
+.endm
+
+.macro KERNEL1x2_SUB
+ vmovsd -4 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -32 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -3 * SIZE(BO, BI, SIZE), %xmm2
+ VFMADD231SD_ %xmm5,%xmm2,%xmm0
+ addq $2, BI
+ addq $1, %rax
+.endm
+
+.macro SAVE1x2
+
+ vmovsd ALPHA, %xmm0
+
+ vmulsd %xmm0 , %xmm4 , %xmm4
+ vmulsd %xmm0 , %xmm5 , %xmm5
+
+#if !defined(TRMMKERNEL)
+
+ vaddsd (CO1), %xmm4,%xmm4
+ vaddsd (CO1, LDC), %xmm5,%xmm5
+
+#endif
+
+ vmovsd %xmm4 , (CO1)
+ vmovsd %xmm5 , (CO1, LDC)
+
+.endm
+
+
+/*******************************************************************************************/
+
+/*******************************************************************************************
+* 1 line of N
+*******************************************************************************************/
+
+.macro KERNEL16x1_1
+ vbroadcastsd -2 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vmovups -28 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ vmovups -24 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm10,%ymm1,%ymm0
+ vmovups -20 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm13,%ymm1,%ymm0
+.endm
+
+.macro KERNEL16x1_2
+ vbroadcastsd -1 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -16 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vmovups -12 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ vmovups -8 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm10,%ymm1,%ymm0
+ vmovups -4 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm13,%ymm1,%ymm0
+.endm
+
+.macro KERNEL16x1_3
+ vbroadcastsd 0 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups 0 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vmovups 4 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ vmovups 8 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm10,%ymm1,%ymm0
+ vmovups 12 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm13,%ymm1,%ymm0
+.endm
+
+.macro KERNEL16x1_4
+ vbroadcastsd 1 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups 16 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vmovups 20 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ vmovups 24 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm10,%ymm1,%ymm0
+ vmovups 28 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm13,%ymm1,%ymm0
+ addq $4, BI
+ addq $64, %rax
+.endm
+
+.macro KERNEL16x1_SUB
+ vbroadcastsd -2 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vmovups -28 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ vmovups -24 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm10,%ymm1,%ymm0
+ vmovups -20 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm13,%ymm1,%ymm0
+ addq $1, BI
+ addq $16, %rax
+.endm
+
+.macro SAVE16x1
+
+ vbroadcastsd ALPHA, %ymm0
+
+ vmulpd %ymm0 , %ymm4 , %ymm4
+ vmulpd %ymm0 , %ymm7 , %ymm7
+ vmulpd %ymm0 , %ymm10, %ymm10
+ vmulpd %ymm0 , %ymm13, %ymm13
+
+#if !defined(TRMMKERNEL)
+
+ vaddpd (CO1), %ymm4,%ymm4
+ vaddpd 4 * SIZE(CO1), %ymm7,%ymm7
+ vaddpd 8 * SIZE(CO1), %ymm10,%ymm10
+ vaddpd 12 * SIZE(CO1), %ymm13,%ymm13
+
+#endif
+
+ vmovups %ymm4 , (CO1)
+ vmovups %ymm7 , 4 * SIZE(CO1)
+ vmovups %ymm10, 8 * SIZE(CO1)
+ vmovups %ymm13,12 * SIZE(CO1)
+
+.endm
+
+
+
+/*******************************************************************************************/
+
+.macro KERNEL8x1_1
+ vbroadcastsd -2 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vmovups -28 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+.endm
+
+.macro KERNEL8x1_2
+ vbroadcastsd -1 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -24 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vmovups -20 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+.endm
+
+.macro KERNEL8x1_3
+ vbroadcastsd 0 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -16 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vmovups -12 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+.endm
+
+.macro KERNEL8x1_4
+ vbroadcastsd 1 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -8 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vmovups -4 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ addq $4, BI
+ addq $32, %rax
+.endm
+
+.macro KERNEL8x1_SUB
+ vbroadcastsd -2 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ vmovups -28 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm7,%ymm1,%ymm0
+ addq $1, BI
+ addq $8 , %rax
+.endm
+
+.macro SAVE8x1
+
+ vbroadcastsd ALPHA, %ymm0
+
+ vmulpd %ymm0 , %ymm4 , %ymm4
+ vmulpd %ymm0 , %ymm7 , %ymm7
+
+#if !defined(TRMMKERNEL)
+
+ vaddpd (CO1), %ymm4,%ymm4
+ vaddpd 4 * SIZE(CO1), %ymm7,%ymm7
+
+#endif
+
+ vmovups %ymm4 , (CO1)
+ vmovups %ymm7 , 4 * SIZE(CO1)
+
+.endm
+
+
+
+/*******************************************************************************************/
+
+.macro KERNEL4x1_1
+ vbroadcastsd -2 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+.endm
+
+.macro KERNEL4x1_2
+ vbroadcastsd -1 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -28 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+.endm
+
+.macro KERNEL4x1_3
+ vbroadcastsd 0 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -24 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+.endm
+
+.macro KERNEL4x1_4
+ vbroadcastsd 1 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -20 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ addq $4, BI
+ addq $16, %rax
+.endm
+
+.macro KERNEL4x1_SUB
+ vbroadcastsd -2 * SIZE(BO, BI, SIZE), %ymm1
+ vmovups -32 * SIZE(AO, %rax, SIZE), %ymm0
+ VFMADD231PD_ %ymm4,%ymm1,%ymm0
+ addq $1, BI
+ addq $4 , %rax
+.endm
+
+.macro SAVE4x1
+
+ vbroadcastsd ALPHA, %ymm0
+
+ vmulpd %ymm0 , %ymm4 , %ymm4
+
+#if !defined(TRMMKERNEL)
+
+ vaddpd (CO1), %ymm4,%ymm4
+
+#endif
+
+ vmovups %ymm4 , (CO1)
+
+.endm
+
+
+/*******************************************************************************************/
+
+.macro KERNEL2x1_1
+ vmovsd -2 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -32 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -31 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm8,%xmm1,%xmm0
+.endm
+
+.macro KERNEL2x1_2
+ vmovsd -1 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -30 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -29 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm8,%xmm1,%xmm0
+.endm
+
+.macro KERNEL2x1_3
+ vmovsd 0 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -28 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -27 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm8,%xmm1,%xmm0
+.endm
+
+.macro KERNEL2x1_4
+ vmovsd 1 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -26 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -25 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm8,%xmm1,%xmm0
+ addq $4, BI
+ addq $8, %rax
+.endm
+
+.macro KERNEL2x1_SUB
+ vmovsd -2 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -32 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ vmovsd -31 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm8,%xmm1,%xmm0
+ addq $1, BI
+ addq $2 , %rax
+.endm
+
+.macro SAVE2x1
+
+ vmovsd ALPHA, %xmm0
+
+ vmulsd %xmm0 , %xmm4 , %xmm4
+ vmulsd %xmm0 , %xmm8 , %xmm8
+
+#if !defined(TRMMKERNEL)
+
+ vaddsd (CO1), %xmm4,%xmm4
+ vaddsd 1 * SIZE(CO1), %xmm8,%xmm8
+
+#endif
+
+ vmovsd %xmm4 , (CO1)
+ vmovsd %xmm8 , 1 * SIZE(CO1)
+
+.endm
+
+
+/*******************************************************************************************/
+
+.macro KERNEL1x1_1
+ vmovsd -2 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -32 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+.endm
+
+.macro KERNEL1x1_2
+ vmovsd -1 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -31 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+.endm
+
+.macro KERNEL1x1_3
+ vmovsd 0 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -30 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+.endm
+
+.macro KERNEL1x1_4
+ vmovsd 1 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -29 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ addq $ 4, BI
+ addq $ 4, %rax
+.endm
+
+.macro KERNEL1x1_SUB
+ vmovsd -2 * SIZE(BO, BI, SIZE), %xmm1
+ vmovsd -32 * SIZE(AO, %rax, SIZE), %xmm0
+ VFMADD231SD_ %xmm4,%xmm1,%xmm0
+ addq $ 1, BI
+ addq $ 1 , %rax
+.endm
+
+.macro SAVE1x1
+
+ vmovsd ALPHA, %xmm0
+
+ vmulsd %xmm0 , %xmm4 , %xmm4
+
+#if !defined(TRMMKERNEL)
+
+ vaddsd (CO1), %xmm4,%xmm4
+
+#endif
+
+ vmovsd %xmm4 , (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)
+ movups %xmm6, 64(%rsp)
+ movups %xmm7, 80(%rsp)
+ movups %xmm8, 96(%rsp)
+ movups %xmm9, 112(%rsp)
+ movups %xmm10, 128(%rsp)
+ movups %xmm11, 144(%rsp)
+ movups %xmm12, 160(%rsp)
+ movups %xmm13, 176(%rsp)
+ movups %xmm14, 192(%rsp)
+ movups %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
+
+#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
+
+ vmovsd %xmm0, ALPHA
+
+ salq $BASE_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
+
+
+ movq Ndiv6, J
+ cmpq $0, J
+ je .L2_0
+ ALIGN_4
+
+.L6_01:
+ // copy to sub buffer
+ movq K, %rax
+ salq $1,%rax // K * 2 ; read 2 values
+ movq B, BO1
+ leaq (B,%rax, SIZE), BO2 // next offset to BO2
+ leaq BUFFER1, BO // first buffer to BO
+ movq K, %rax
+ sarq $3 , %rax // K / 8
+ jz .L6_01a_2
+ ALIGN_4
+
+.L6_01a_1:
+
+ prefetcht0 512(BO1)
+ prefetcht0 512(BO2)
+ prefetchw 512(BO)
+
+
+ vmovups 0 * SIZE(BO1), %xmm0
+ vmovups 2 * SIZE(BO1), %xmm2
+ vmovups 4 * SIZE(BO1), %xmm4
+ vmovups 6 * SIZE(BO1), %xmm6
+ vmovsd 0 * SIZE(BO2), %xmm1
+ vmovsd 2 * SIZE(BO2), %xmm3
+ vmovsd 4 * SIZE(BO2), %xmm5
+ vmovsd 6 * SIZE(BO2), %xmm7
+ vmovups %xmm0, 0*SIZE(BO)
+ vmovsd %xmm1, 2*SIZE(BO)
+ vmovups %xmm2, 3*SIZE(BO)
+ vmovsd %xmm3, 5*SIZE(BO)
+ vmovups %xmm4, 6*SIZE(BO)
+ vmovsd %xmm5, 8*SIZE(BO)
+ vmovups %xmm6, 9*SIZE(BO)
+ vmovsd %xmm7,11*SIZE(BO)
+ addq $ 8*SIZE,BO1
+ addq $ 8*SIZE,BO2
+ addq $ 12*SIZE,BO
+
+ vmovups 0 * SIZE(BO1), %xmm0
+ vmovups 2 * SIZE(BO1), %xmm2
+ vmovups 4 * SIZE(BO1), %xmm4
+ vmovups 6 * SIZE(BO1), %xmm6
+ vmovsd 0 * SIZE(BO2), %xmm1
+ vmovsd 2 * SIZE(BO2), %xmm3
+ vmovsd 4 * SIZE(BO2), %xmm5
+ vmovsd 6 * SIZE(BO2), %xmm7
+ vmovups %xmm0, 0*SIZE(BO)
+ vmovsd %xmm1, 2*SIZE(BO)
+ vmovups %xmm2, 3*SIZE(BO)
+ vmovsd %xmm3, 5*SIZE(BO)
+ vmovups %xmm4, 6*SIZE(BO)
+ vmovsd %xmm5, 8*SIZE(BO)
+ vmovups %xmm6, 9*SIZE(BO)
+ vmovsd %xmm7,11*SIZE(BO)
+ addq $ 8*SIZE,BO1
+ addq $ 8*SIZE,BO2
+ addq $ 12*SIZE,BO
+
+ decq %rax
+ jnz .L6_01a_1
+
+
+
+.L6_01a_2:
+
+ movq K, %rax
+ andq $7, %rax // K % 8
+ jz .L6_02c
+ ALIGN_4
+
+
+.L6_02b:
+
+ vmovups 0 * SIZE(BO1), %xmm0
+ vmovsd 0 * SIZE(BO2), %xmm2
+ vmovups %xmm0, 0*SIZE(BO)
+ vmovsd %xmm2, 2*SIZE(BO)
+ addq $ 2*SIZE,BO1
+ addq $ 2*SIZE,BO2
+ addq $ 3*SIZE,BO
+ decq %rax
+ jnz .L6_02b
+
+.L6_02c:
+
+ movq K, %rax
+ salq $1,%rax // K * 2
+ leaq (B,%rax, SIZE), BO1 // next offset to BO1
+ leaq (BO1,%rax, SIZE), BO2 // next offset to BO2
+ leaq BUFFER2, BO // second buffer to BO
+ movq K, %rax
+ sarq $3 , %rax // K / 8
+ jz .L6_02c_2
+ ALIGN_4
+
+.L6_02c_1:
+
+ prefetcht0 512(BO2)
+ prefetchw 512(BO)
+
+ vmovups 0 * SIZE(BO2), %xmm0
+ vmovups 2 * SIZE(BO2), %xmm2
+ vmovups 4 * SIZE(BO2), %xmm4
+ vmovups 6 * SIZE(BO2), %xmm6
+ vmovsd 1 * SIZE(BO1), %xmm1
+ vmovsd 3 * SIZE(BO1), %xmm3
+ vmovsd 5 * SIZE(BO1), %xmm5
+ vmovsd 7 * SIZE(BO1), %xmm7
+ vmovsd %xmm1, 0*SIZE(BO)
+ vmovups %xmm0, 1*SIZE(BO)
+ vmovsd %xmm3, 3*SIZE(BO)
+ vmovups %xmm2, 4*SIZE(BO)
+ vmovsd %xmm5, 6*SIZE(BO)
+ vmovups %xmm4, 7*SIZE(BO)
+ vmovsd %xmm7, 9*SIZE(BO)
+ vmovups %xmm6,10*SIZE(BO)
+ addq $8*SIZE,BO1
+ addq $8*SIZE,BO2
+ addq $12*SIZE,BO
+
+
+ vmovups 0 * SIZE(BO2), %xmm0
+ vmovups 2 * SIZE(BO2), %xmm2
+ vmovups 4 * SIZE(BO2), %xmm4
+ vmovups 6 * SIZE(BO2), %xmm6
+ vmovsd 1 * SIZE(BO1), %xmm1
+ vmovsd 3 * SIZE(BO1), %xmm3
+ vmovsd 5 * SIZE(BO1), %xmm5
+ vmovsd 7 * SIZE(BO1), %xmm7
+ vmovsd %xmm1, 0*SIZE(BO)
+ vmovups %xmm0, 1*SIZE(BO)
+ vmovsd %xmm3, 3*SIZE(BO)
+ vmovups %xmm2, 4*SIZE(BO)
+ vmovsd %xmm5, 6*SIZE(BO)
+ vmovups %xmm4, 7*SIZE(BO)
+ vmovsd %xmm7, 9*SIZE(BO)
+ vmovups %xmm6,10*SIZE(BO)
+ addq $8*SIZE,BO1
+ addq $8*SIZE,BO2
+ addq $12*SIZE,BO
+
+ decq %rax
+ jnz .L6_02c_1
+
+
+.L6_02c_2:
+
+ movq K, %rax
+ andq $7, %rax // K % 8
+ jz .L6_03c
+ ALIGN_4
+
+.L6_03b:
+
+ vmovsd 1*SIZE(BO1), %xmm0
+ vmovups 0*SIZE(BO2), %xmm1
+ vmovsd %xmm0, 0*SIZE(BO)
+ vmovups %xmm1, 1*SIZE(BO)
+ addq $2*SIZE,BO1
+ addq $2*SIZE,BO2
+ addq $3*SIZE,BO
+ decq %rax
+ jnz .L6_03b
+
+
+.L6_03c:
+
+ movq BO2, B // next offset of B
+
+.L6_10:
+ movq C, CO1
+ leaq (C, LDC, 2), C
+ leaq (C, LDC, 1), C // c += 3 * ldc
+
+
+ movq A, AO // aoffset = a
+ addq $16 * SIZE, AO
+
+ movq M, I
+ sarq $4, I // i = (m >> 4)
+ je .L6_20
+
+ ALIGN_4
+
+.L6_11:
+ leaq BUFFER1, BO // first buffer to BO
+ addq $12 * SIZE, BO
+
+ prefetcht0 (CO1)
+ prefetcht0 (CO1,LDC,1)
+ prefetcht0 (CO1,LDC,2)
+ prefetcht0 64(CO1)
+ prefetcht0 64(CO1,LDC,1)
+ prefetcht0 64(CO1,LDC,2)
+
+ vzeroall
+
+ movq K, %rax
+
+ sarq $1, %rax // K / 8
+ je .L6_16
+
+ ALIGN_5
+
+.L6_12:
+/*
+ prefetcht0 B_PR1(BO)
+ prefetcht0 B_PR1+64(BO)
+ prefetcht0 B_PR1+128(BO)
+*/
+ KERNEL16x3_SUBN
+ KERNEL16x3_SUBN
+/*
+ KERNEL16x3_SUBN
+ KERNEL16x3_SUBN
+
+ KERNEL16x3_SUBN
+ KERNEL16x3_SUBN
+ KERNEL16x3_SUBN
+ KERNEL16x3_SUBN
+*/
+ dec %rax
+ jne .L6_12
+
+.L6_16:
+ movq K, %rax
+
+ andq $1, %rax # if (k & 1)
+ je .L6_19
+
+ ALIGN_4
+
+.L6_17:
+
+ KERNEL16x3_SUBN
+
+ dec %rax
+ jne .L6_17
+ ALIGN_4
+
+
+.L6_19:
+
+ SAVE16x3
+
+ addq $16 * SIZE, CO1 # coffset += 16
+ decq I # i --
+ jg .L6_11
+ ALIGN_4
+
+/**************************************************************************
+* Rest of M
+***************************************************************************/
+.L6_20:
+ // Test rest of M
+
+ testq $15, M
+ jz .L7_10 // to next 3 lines of N
+
+ testq $8, M
+ jz .L6_21pre
+ ALIGN_4
+
+/**************************************************************************/
+
+.L6_20_1:
+ leaq BUFFER1, BO // first buffer to BO
+ addq $12 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ sarq $3, %rax
+ je .L6_20_6
+
+ ALIGN_4
+
+.L6_20_2:
+
+ KERNEL8x3_SUBN
+ KERNEL8x3_SUBN
+ KERNEL8x3_SUBN
+ KERNEL8x3_SUBN
+
+ KERNEL8x3_SUBN
+ KERNEL8x3_SUBN
+ KERNEL8x3_SUBN
+ KERNEL8x3_SUBN
+ dec %rax
+ jne .L6_20_2
+ ALIGN_4
+
+.L6_20_6:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L6_20_9
+
+
+ ALIGN_4
+
+.L6_20_7:
+
+ KERNEL8x3_SUBN
+
+ dec %rax
+ jne .L6_20_7
+ ALIGN_4
+
+
+.L6_20_9:
+
+ SAVE8x3
+
+ addq $8 * SIZE, CO1 # coffset += 8
+ ALIGN_4
+
+
+
+/**************************************************************************/
+
+.L6_21pre:
+
+ testq $4, M
+ jz .L6_30
+ ALIGN_4
+
+.L6_21:
+ leaq BUFFER1, BO // first buffer to BO
+ addq $12 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ sarq $3, %rax
+ je .L6_26
+
+ ALIGN_4
+
+.L6_22:
+
+ KERNEL4x3_SUBN
+ KERNEL4x3_SUBN
+ KERNEL4x3_SUBN
+ KERNEL4x3_SUBN
+
+ KERNEL4x3_SUBN
+ KERNEL4x3_SUBN
+ KERNEL4x3_SUBN
+ KERNEL4x3_SUBN
+ dec %rax
+ jne .L6_22
+ ALIGN_4
+
+.L6_26:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L6_29
+
+ ALIGN_4
+
+.L6_27:
+
+ KERNEL4x3_SUBN
+
+ dec %rax
+ jne .L6_27
+ ALIGN_4
+
+
+.L6_29:
+
+ SAVE4x3
+
+ addq $4 * SIZE, CO1 # coffset += 4
+ ALIGN_4
+
+
+.L6_30:
+ testq $2, M
+ jz .L6_40
+
+ ALIGN_4
+
+.L6_31:
+ leaq BUFFER1, BO // first buffer to BO
+ addq $12 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ sarq $3, %rax
+ je .L6_36
+ ALIGN_4
+
+.L6_32:
+
+ KERNEL2x3_SUBN
+ KERNEL2x3_SUBN
+ KERNEL2x3_SUBN
+ KERNEL2x3_SUBN
+
+ KERNEL2x3_SUBN
+ KERNEL2x3_SUBN
+ KERNEL2x3_SUBN
+ KERNEL2x3_SUBN
+ dec %rax
+ jne .L6_32
+ ALIGN_4
+
+.L6_36:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L6_39
+
+ ALIGN_4
+
+.L6_37:
+
+ KERNEL2x3_SUBN
+
+ dec %rax
+ jne .L6_37
+ ALIGN_4
+
+
+.L6_39:
+
+ SAVE2x3
+
+ addq $2 * SIZE, CO1 # coffset += 2
+ ALIGN_4
+
+.L6_40:
+ testq $1, M
+ jz .L7_10 // to next 3 lines of N
+
+ ALIGN_4
+
+.L6_41:
+ leaq BUFFER1, BO // first buffer to BO
+ addq $12 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ sarq $3,%rax
+ je .L6_46
+
+ ALIGN_4
+
+.L6_42:
+
+ KERNEL1x3_SUBN
+ KERNEL1x3_SUBN
+ KERNEL1x3_SUBN
+ KERNEL1x3_SUBN
+
+ KERNEL1x3_SUBN
+ KERNEL1x3_SUBN
+ KERNEL1x3_SUBN
+ KERNEL1x3_SUBN
+
+ dec %rax
+ jne .L6_42
+ ALIGN_4
+
+.L6_46:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L6_49
+
+ ALIGN_4
+
+.L6_47:
+
+ KERNEL1x3_SUBN
+
+ dec %rax
+ jne .L6_47
+ ALIGN_4
+
+
+.L6_49:
+
+ SAVE1x3
+
+ addq $1 * SIZE, CO1 # coffset += 1
+ ALIGN_4
+
+
+
+
+/***************************************************************************************************************/
+
+.L7_10:
+ movq C, CO1
+ leaq (C, LDC, 2), C
+ leaq (C, LDC, 1), C // c += 3 * ldc
+
+
+ movq A, AO // aoffset = a
+ addq $16 * SIZE, AO
+
+ movq M, I
+ sarq $4, I // i = (m >> 4)
+ je .L7_20
+
+ ALIGN_4
+
+.L7_11:
+ leaq BUFFER2, BO // second buffer to BO
+ addq $12 * SIZE, BO
+
+ prefetcht0 (CO1)
+ prefetcht0 (CO1,LDC,1)
+ prefetcht0 (CO1,LDC,2)
+ prefetcht0 64(CO1)
+ prefetcht0 64(CO1,LDC,1)
+ prefetcht0 64(CO1,LDC,2)
+
+ vzeroall
+
+ movq K, %rax
+
+ sarq $3, %rax // K / 8
+ je .L7_16
+ ALIGN_5
+
+.L7_12:
+/*
+ prefetcht0 B_PR1(BO)
+ prefetcht0 B_PR1+64(BO)
+ prefetcht0 B_PR1+128(BO)
+*/
+ KERNEL16x3_SUBN
+ KERNEL16x3_SUBN
+ KERNEL16x3_SUBN
+ KERNEL16x3_SUBN
+
+ KERNEL16x3_SUBN
+ KERNEL16x3_SUBN
+ KERNEL16x3_SUBN
+ KERNEL16x3_SUBN
+ dec %rax
+ jne .L7_12
+ ALIGN_4
+
+.L7_16:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L7_19
+
+ ALIGN_5
+
+.L7_17:
+
+ KERNEL16x3_SUBN
+
+ dec %rax
+ jne .L7_17
+
+
+.L7_19:
+
+ SAVE16x3
+
+ addq $16 * SIZE, CO1 # coffset += 16
+ decq I # i --
+ jg .L7_11
+ ALIGN_4
+
+/**************************************************************************
+* Rest of M
+***************************************************************************/
+.L7_20:
+ // Test rest of M
+
+ testq $15, M
+ jz .L7_60 // to next 3 lines of N
+
+ testq $8, M
+ jz .L7_21pre
+ ALIGN_4
+
+/**************************************************************************/
+
+.L7_20_1:
+ leaq BUFFER2, BO // first buffer to BO
+ addq $12 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ sarq $3, %rax
+ je .L7_20_6
+
+ ALIGN_4
+
+.L7_20_2:
+
+ KERNEL8x3_SUBN
+ KERNEL8x3_SUBN
+ KERNEL8x3_SUBN
+ KERNEL8x3_SUBN
+
+ KERNEL8x3_SUBN
+ KERNEL8x3_SUBN
+ KERNEL8x3_SUBN
+ KERNEL8x3_SUBN
+
+ dec %rax
+ jne .L7_20_2
+ ALIGN_4
+
+.L7_20_6:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L7_20_9
+
+ ALIGN_4
+
+.L7_20_7:
+
+ KERNEL8x3_SUBN
+
+ dec %rax
+ jne .L7_20_7
+ ALIGN_4
+
+.L7_20_9:
+
+ SAVE8x3
+
+ addq $8 * SIZE, CO1 # coffset += 8
+ ALIGN_4
+
+
+
+/**************************************************************************/
+
+.L7_21pre:
+
+ testq $4, M
+ jz .L7_30
+ ALIGN_4
+
+.L7_21:
+ leaq BUFFER2, BO // second buffer to BO
+ addq $12 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ sarq $3, %rax
+ je .L7_26
+
+ ALIGN_4
+
+.L7_22:
+
+ KERNEL4x3_SUBN
+ KERNEL4x3_SUBN
+ KERNEL4x3_SUBN
+ KERNEL4x3_SUBN
+
+ KERNEL4x3_SUBN
+ KERNEL4x3_SUBN
+ KERNEL4x3_SUBN
+ KERNEL4x3_SUBN
+
+ dec %rax
+ jne .L7_22
+ ALIGN_4
+
+.L7_26:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L7_29
+
+ ALIGN_4
+
+.L7_27:
+
+ KERNEL4x3_SUBN
+
+ dec %rax
+ jne .L7_27
+ ALIGN_4
+
+
+.L7_29:
+
+ SAVE4x3
+
+ addq $4 * SIZE, CO1 # coffset += 4
+ ALIGN_4
+
+
+.L7_30:
+ testq $2, M
+ jz .L7_40
+
+ ALIGN_4
+
+.L7_31:
+ leaq BUFFER2, BO // second buffer to BO
+ addq $12 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ sarq $3, %rax
+ je .L7_36
+
+ ALIGN_4
+
+.L7_32:
+
+ KERNEL2x3_SUBN
+ KERNEL2x3_SUBN
+ KERNEL2x3_SUBN
+ KERNEL2x3_SUBN
+
+ KERNEL2x3_SUBN
+ KERNEL2x3_SUBN
+ KERNEL2x3_SUBN
+ KERNEL2x3_SUBN
+
+ dec %rax
+ jne .L7_32
+ ALIGN_4
+
+.L7_36:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L7_39
+
+ ALIGN_4
+
+.L7_37:
+
+ KERNEL2x3_SUBN
+
+ dec %rax
+ jne .L7_37
+ ALIGN_4
+
+
+.L7_39:
+
+ SAVE2x3
+
+ addq $2 * SIZE, CO1 # coffset += 2
+ ALIGN_4
+
+.L7_40:
+ testq $1, M
+ jz .L7_60 // to next 3 lines of N
+
+ ALIGN_4
+
+.L7_41:
+ leaq BUFFER2, BO // second buffer to BO
+ addq $12 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ sarq $3, %rax
+ je .L7_46
+
+ ALIGN_4
+
+.L7_42:
+ KERNEL1x3_SUBN
+ KERNEL1x3_SUBN
+ KERNEL1x3_SUBN
+ KERNEL1x3_SUBN
+
+ KERNEL1x3_SUBN
+ KERNEL1x3_SUBN
+ KERNEL1x3_SUBN
+ KERNEL1x3_SUBN
+
+ dec %rax
+ jne .L7_42
+ ALIGN_4
+
+.L7_46:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L7_49
+
+ ALIGN_4
+
+.L7_47:
+
+ KERNEL1x3_SUBN
+
+ dec %rax
+ jne .L7_47
+ ALIGN_4
+
+
+.L7_49:
+
+ SAVE1x3
+
+ addq $1 * SIZE, CO1 # coffset += 1
+ ALIGN_4
+
+
+
+.L7_60:
+
+ decq J // j --
+ jg .L6_01
+
+
+.L2_0:
+ cmpq $0, Nmod6 // N % 6 == 0
+ je .L999
+
+/************************************************************************************************
+* Loop for Nmod6 / 2 > 0
+*************************************************************************************************/
+
+ movq Nmod6, J
+ sarq $1, J // j = j / 2
+ je .L1_0
+ ALIGN_4
+
+.L2_01:
+ // copy to sub buffer
+ movq B, BO1
+ leaq BUFFER1, BO // first buffer to BO
+ movq K, %rax
+ sarq $2, %rax // K / 4
+ jz .L2_01b
+ ALIGN_4
+
+.L2_01a:
+ prefetcht0 512(BO1)
+ prefetchw 512(BO)
+
+ vmovups (BO1), %xmm0
+ vmovups 2*SIZE(BO1), %xmm1
+ vmovups 4*SIZE(BO1), %xmm2
+ vmovups 6*SIZE(BO1), %xmm3
+
+ vmovups %xmm0, (BO)
+ vmovups %xmm1, 2*SIZE(BO)
+ vmovups %xmm2, 4*SIZE(BO)
+ vmovups %xmm3, 6*SIZE(BO)
+
+ addq $8*SIZE,BO1
+ addq $8*SIZE,BO
+ decq %rax
+ jnz .L2_01a
+
+
+.L2_01b:
+
+ movq K, %rax
+ andq $3, %rax // K % 4
+ jz .L2_02d
+ ALIGN_4
+
+.L2_02c:
+
+ vmovups (BO1), %xmm0
+ vmovups %xmm0, (BO)
+ addq $2*SIZE,BO1
+ addq $2*SIZE,BO
+ decq %rax
+ jnz .L2_02c
+
+.L2_02d:
+
+ movq BO1, B // next offset of B
+
+.L2_10:
+ movq C, CO1
+ leaq (C, LDC, 2), C // c += 2 * ldc
+
+
+ movq A, AO // aoffset = a
+ addq $32 * SIZE, AO
+
+ movq M, I
+ sarq $4, I // i = (m >> 4)
+ je .L2_20
+
+ ALIGN_4
+
+.L2_11:
+ leaq BUFFER1, BO // first buffer to BO
+ addq $4 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $-8, %rax // K = K - ( K % 8 )
+ je .L2_16
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), 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
+
+.L2_12:
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL16x2_1
+ KERNEL16x2_2
+ KERNEL16x2_3
+ KERNEL16x2_4
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL16x2_1
+ KERNEL16x2_2
+ KERNEL16x2_3
+ KERNEL16x2_4
+
+ je .L2_16
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL16x2_1
+ KERNEL16x2_2
+ KERNEL16x2_3
+ KERNEL16x2_4
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL16x2_1
+ KERNEL16x2_2
+ KERNEL16x2_3
+ KERNEL16x2_4
+
+ je .L2_16
+
+ jmp .L2_12
+ ALIGN_4
+
+.L2_16:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L2_19
+
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), 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
+
+.L2_17:
+
+ KERNEL16x2_SUB
+
+ jl .L2_17
+ ALIGN_4
+
+
+.L2_19:
+
+ SAVE16x2
+
+ addq $16 * SIZE, CO1 # coffset += 16
+ decq I # i --
+ jg .L2_11
+ ALIGN_4
+
+/**************************************************************************
+* Rest of M
+***************************************************************************/
+.L2_20:
+ // Test rest of M
+
+ testq $15, M
+ jz .L2_60 // to next 3 lines of N
+
+ testq $8, M
+ jz .L2_21pre
+ ALIGN_4
+
+/**************************************************************************/
+
+.L2_20_1:
+ leaq BUFFER1, BO // first buffer to BO
+ addq $4 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $-8, %rax
+ je .L2_20_6
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), 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
+
+.L2_20_2:
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL8x2_1
+ KERNEL8x2_2
+ KERNEL8x2_3
+ KERNEL8x2_4
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL8x2_1
+ KERNEL8x2_2
+ KERNEL8x2_3
+ KERNEL8x2_4
+
+ je .L2_20_6
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL8x2_1
+ KERNEL8x2_2
+ KERNEL8x2_3
+ KERNEL8x2_4
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL8x2_1
+ KERNEL8x2_2
+ KERNEL8x2_3
+ KERNEL8x2_4
+
+ je .L2_20_6
+
+ jmp .L2_20_2
+ ALIGN_4
+
+.L2_20_6:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L2_20_9
+
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), 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
+
+.L2_20_7:
+
+ KERNEL8x2_SUB
+
+ jl .L2_20_7
+ ALIGN_4
+
+
+.L2_20_9:
+
+ SAVE8x2
+
+ addq $8 * SIZE, CO1 # coffset += 8
+ ALIGN_4
+
+
+
+/**************************************************************************/
+
+.L2_21pre:
+
+ testq $4, M
+ jz .L2_30
+ ALIGN_4
+
+.L2_21:
+ leaq BUFFER1, BO // first buffer to BO
+ addq $4 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $-8, %rax
+ je .L2_26
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), BI // BI = BI * 1 ; 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_22:
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL4x2_1
+ KERNEL4x2_2
+ KERNEL4x2_3
+ KERNEL4x2_4
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL4x2_1
+ KERNEL4x2_2
+ KERNEL4x2_3
+ KERNEL4x2_4
+
+ je .L2_26
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL4x2_1
+ KERNEL4x2_2
+ KERNEL4x2_3
+ KERNEL4x2_4
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL4x2_1
+ KERNEL4x2_2
+ KERNEL4x2_3
+ KERNEL4x2_4
+
+ je .L2_26
+
+ jmp .L2_22
+ ALIGN_4
+
+.L2_26:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L2_29
+
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), 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
+
+.L2_27:
+
+ KERNEL4x2_SUB
+
+ jl .L2_27
+ ALIGN_4
+
+
+.L2_29:
+
+ SAVE4x2
+
+ addq $4 * SIZE, CO1 # coffset += 4
+ ALIGN_4
+
+
+.L2_30:
+ testq $2, M
+ jz .L2_40
+
+ ALIGN_4
+
+.L2_31:
+ leaq BUFFER1, BO // first buffer to BO
+ addq $4 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $-8, %rax
+ je .L2_36
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), 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
+
+.L2_32:
+
+ KERNEL2x2_1
+ KERNEL2x2_2
+ KERNEL2x2_3
+ KERNEL2x2_4
+
+ KERNEL2x2_1
+ KERNEL2x2_2
+ KERNEL2x2_3
+ KERNEL2x2_4
+
+ je .L2_36
+
+ KERNEL2x2_1
+ KERNEL2x2_2
+ KERNEL2x2_3
+ KERNEL2x2_4
+
+ KERNEL2x2_1
+ KERNEL2x2_2
+ KERNEL2x2_3
+ KERNEL2x2_4
+
+ je .L2_36
+
+ jmp .L2_32
+ ALIGN_4
+
+.L2_36:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L2_39
+
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), 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
+
+.L2_37:
+
+ KERNEL2x2_SUB
+
+ jl .L2_37
+ ALIGN_4
+
+
+.L2_39:
+
+ SAVE2x2
+
+ addq $2 * SIZE, CO1 # coffset += 2
+ ALIGN_4
+
+.L2_40:
+ testq $1, M
+ jz .L2_60 // to next 2 lines of N
+
+ ALIGN_4
+
+.L2_41:
+ leaq BUFFER1, BO // first buffer to BO
+ addq $4 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $-8, %rax
+ je .L2_46
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), BI // BI = BI * 2 ; number of values
+
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_42:
+
+ KERNEL1x2_1
+ KERNEL1x2_2
+ KERNEL1x2_3
+ KERNEL1x2_4
+
+ KERNEL1x2_1
+ KERNEL1x2_2
+ KERNEL1x2_3
+ KERNEL1x2_4
+
+ je .L2_46
+
+ KERNEL1x2_1
+ KERNEL1x2_2
+ KERNEL1x2_3
+ KERNEL1x2_4
+
+ KERNEL1x2_1
+ KERNEL1x2_2
+ KERNEL1x2_3
+ KERNEL1x2_4
+
+ je .L2_46
+
+ jmp .L2_42
+ ALIGN_4
+
+.L2_46:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L2_49
+
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), BI // BI = BI * 2 ; number of values
+
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_47:
+
+ KERNEL1x2_SUB
+
+ jl .L2_47
+ ALIGN_4
+
+
+.L2_49:
+
+ SAVE1x2
+
+ addq $1 * SIZE, CO1 # coffset += 1
+ ALIGN_4
+
+.L2_60:
+
+ 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 $1*SIZE,BO1
+ addq $1*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
+
+
+ movq A, AO // aoffset = a
+ addq $32 * SIZE, AO
+
+ movq M, I
+ sarq $4, I // i = (m >> 4)
+ je .L1_20
+
+ ALIGN_4
+
+.L1_11:
+ leaq BUFFER1, BO // first buffer to BO
+ addq $2 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $-8, %rax // K = K - ( K % 8 )
+ je .L1_16
+ movq %rax, BI // Index for BO
+
+ 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_12:
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL16x1_1
+ KERNEL16x1_2
+ KERNEL16x1_3
+ KERNEL16x1_4
+
+ KERNEL16x1_1
+ KERNEL16x1_2
+ KERNEL16x1_3
+ KERNEL16x1_4
+
+ je .L1_16
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL16x1_1
+ KERNEL16x1_2
+ KERNEL16x1_3
+ KERNEL16x1_4
+
+ KERNEL16x1_1
+ KERNEL16x1_2
+ KERNEL16x1_3
+ KERNEL16x1_4
+
+ je .L1_16
+
+ jmp .L1_12
+ ALIGN_4
+
+.L1_16:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L1_19
+
+ movq %rax, BI // Index for BO
+
+ 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_17:
+
+ KERNEL16x1_SUB
+
+ jl .L1_17
+ ALIGN_4
+
+
+.L1_19:
+
+ SAVE16x1
+
+ addq $16 * SIZE, CO1 # coffset += 16
+ decq I # i --
+ jg .L1_11
+ ALIGN_4
+
+/**************************************************************************
+* Rest of M
+***************************************************************************/
+.L1_20:
+ // Test rest of M
+
+ testq $15, M
+ jz .L999
+
+ testq $8, M
+ jz .L1_21pre
+ ALIGN_4
+
+/**************************************************************************/
+
+.L1_20_1:
+ leaq BUFFER1, BO // first buffer to BO
+ addq $2 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $-8, %rax
+ je .L1_20_6
+ movq %rax, BI // Index for BO
+
+ 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_20_2:
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL8x1_1
+ KERNEL8x1_2
+ KERNEL8x1_3
+ KERNEL8x1_4
+
+ KERNEL8x1_1
+ KERNEL8x1_2
+ KERNEL8x1_3
+ KERNEL8x1_4
+
+ je .L1_20_6
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL8x1_1
+ KERNEL8x1_2
+ KERNEL8x1_3
+ KERNEL8x1_4
+
+ KERNEL8x1_1
+ KERNEL8x1_2
+ KERNEL8x1_3
+ KERNEL8x1_4
+
+ je .L1_20_6
+
+ jmp .L1_20_2
+ ALIGN_4
+
+.L1_20_6:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L1_20_9
+
+ movq %rax, BI // Index for BO
+
+ 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_20_7:
+
+ KERNEL8x1_SUB
+
+ jl .L1_20_7
+ ALIGN_4
+
+
+.L1_20_9:
+
+ SAVE8x1
+
+ addq $8 * SIZE, CO1 # coffset += 8
+ ALIGN_4
+
+
+
+/**************************************************************************/
+
+.L1_21pre:
+
+ testq $4, M
+ jz .L1_30
+ ALIGN_4
+
+.L1_21:
+ leaq BUFFER1, BO // first buffer to BO
+ addq $2 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $-8, %rax
+ je .L1_26
+ movq %rax, BI // Index for BO
+
+ 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_22:
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL4x1_1
+ KERNEL4x1_2
+ KERNEL4x1_3
+ KERNEL4x1_4
+
+ KERNEL4x1_1
+ KERNEL4x1_2
+ KERNEL4x1_3
+ KERNEL4x1_4
+
+ je .L1_26
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL4x1_1
+ KERNEL4x1_2
+ KERNEL4x1_3
+ KERNEL4x1_4
+
+ KERNEL4x1_1
+ KERNEL4x1_2
+ KERNEL4x1_3
+ KERNEL4x1_4
+
+ je .L1_26
+
+ jmp .L1_22
+ ALIGN_4
+
+.L1_26:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L1_29
+
+ movq %rax, BI // Index for BO
+
+ 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_27:
+
+ KERNEL4x1_SUB
+
+ jl .L1_27
+ ALIGN_4
+
+
+.L1_29:
+
+ SAVE4x1
+
+ addq $4 * SIZE, CO1 # coffset += 4
+ ALIGN_4
+
+
+.L1_30:
+ testq $2, M
+ jz .L1_40
+
+ ALIGN_4
+
+.L1_31:
+ leaq BUFFER1, BO // first buffer to BO
+ addq $2 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $-8, %rax
+ je .L1_36
+ movq %rax, BI // Index for BO
+
+ 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_32:
+
+ KERNEL2x1_1
+ KERNEL2x1_2
+ KERNEL2x1_3
+ KERNEL2x1_4
+
+ KERNEL2x1_1
+ KERNEL2x1_2
+ KERNEL2x1_3
+ KERNEL2x1_4
+
+ je .L1_36
+
+ KERNEL2x1_1
+ KERNEL2x1_2
+ KERNEL2x1_3
+ KERNEL2x1_4
+
+ KERNEL2x1_1
+ KERNEL2x1_2
+ KERNEL2x1_3
+ KERNEL2x1_4
+
+ je .L1_36
+
+ jmp .L1_32
+ ALIGN_4
+
+.L1_36:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L1_39
+
+ movq %rax, BI // Index for BO
+
+ 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_37:
+
+ KERNEL2x1_SUB
+
+ jl .L1_37
+ ALIGN_4
+
+
+.L1_39:
+
+ SAVE2x1
+
+ addq $2 * SIZE, CO1 # coffset += 2
+ ALIGN_4
+
+.L1_40:
+ testq $1, M
+ jz .L999
+
+ ALIGN_4
+
+.L1_41:
+ leaq BUFFER1, BO // first buffer to BO
+ addq $2 * SIZE, BO
+
+ vzeroall
+
+ movq K, %rax
+
+ andq $-8, %rax
+ je .L1_46
+ movq %rax, BI // Index for BO
+
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_42:
+
+ KERNEL1x1_1
+ KERNEL1x1_2
+ KERNEL1x1_3
+ KERNEL1x1_4
+
+ KERNEL1x1_1
+ KERNEL1x1_2
+ KERNEL1x1_3
+ KERNEL1x1_4
+
+ je .L1_46
+
+ KERNEL1x1_1
+ KERNEL1x1_2
+ KERNEL1x1_3
+ KERNEL1x1_4
+
+ KERNEL1x1_1
+ KERNEL1x1_2
+ KERNEL1x1_3
+ KERNEL1x1_4
+
+ je .L1_46
+
+ jmp .L1_42
+ ALIGN_4
+
+.L1_46:
+ movq K, %rax
+
+ andq $7, %rax # if (k & 1)
+ je .L1_49
+
+ movq %rax, BI // Index for BO
+
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_47:
+
+ KERNEL1x1_SUB
+
+ jl .L1_47
+ ALIGN_4
+
+
+.L1_49:
+
+ SAVE1x1
+
+ addq $1 * SIZE, CO1 # coffset += 1
+ ALIGN_4
+
+
+.L999:
+ 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
+ movups 64(%rsp), %xmm6
+ movups 80(%rsp), %xmm7
+ movups 96(%rsp), %xmm8
+ movups 112(%rsp), %xmm9
+ movups 128(%rsp), %xmm10
+ movups 144(%rsp), %xmm11
+ movups 160(%rsp), %xmm12
+ movups 176(%rsp), %xmm13
+ movups 192(%rsp), %xmm14
+ movups 208(%rsp), %xmm15
+#endif
+
+ addq $STACKSIZE, %rsp
+ ret
+
+ EPILOGUE
+
+
+#else
+/*************************************************************************************
+* TRMM Kernel
+*************************************************************************************/
+
+
+ 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)
+ movups %xmm6, 64(%rsp)
+ movups %xmm7, 80(%rsp)
+ movups %xmm8, 96(%rsp)
+ movups %xmm9, 112(%rsp)
+ movups %xmm10, 128(%rsp)
+ movups %xmm11, 144(%rsp)
+ movups %xmm12, 160(%rsp)
+ movups %xmm13, 176(%rsp)
+ movups %xmm14, 192(%rsp)
+ movups %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
+
+#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
+
+ vmovsd %xmm0, ALPHA
+
+ salq $BASE_SHIFT, LDC
+
+ movq N, %rax
+ xorq %rdx, %rdx
+ movq $2, %rdi
+ divq %rdi // N / 6
+ movq %rax, Ndiv6 // N / 6
+ movq %rdx, Nmod6 // N % 6
+
+
+
+#ifdef TRMMKERNEL
+ vmovsd %xmm12, OFFSET
+ vmovsd %xmm12, KK
+#ifndef LEFT
+ negq KK
+#endif
+#endif
+
+ 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
+ sarq $2, %rax // K / 4
+ jz .L2_01b
+ ALIGN_4
+
+.L2_01a:
+ prefetcht0 512(BO1)
+ prefetchw 512(BO)
+
+ vmovups (BO1), %xmm0
+ vmovups 2*SIZE(BO1), %xmm1
+ vmovups 4*SIZE(BO1), %xmm2
+ vmovups 6*SIZE(BO1), %xmm3
+
+ vmovups %xmm0, (BO)
+ vmovups %xmm1, 2*SIZE(BO)
+ vmovups %xmm2, 4*SIZE(BO)
+ vmovups %xmm3, 6*SIZE(BO)
+
+ addq $8*SIZE,BO1
+ addq $8*SIZE,BO
+ decq %rax
+ jnz .L2_01a
+
+
+.L2_01b:
+
+ movq K, %rax
+ andq $3, %rax // K % 4
+ jz .L2_02d
+ ALIGN_4
+
+.L2_02c:
+
+ vmovups (BO1), %xmm0
+ vmovups %xmm0, (BO)
+ addq $2*SIZE,BO1
+ addq $2*SIZE,BO
+ decq %rax
+ jnz .L2_02c
+
+.L2_02d:
+
+ 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 $32 * SIZE, AO
+
+ movq M, I
+ sarq $4, I // i = (m >> 4)
+ je .L2_20
+
+ ALIGN_4
+
+.L2_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,BI,1), 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 $16, %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_16
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), 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
+
+.L2_12:
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL16x2_1
+ KERNEL16x2_2
+ KERNEL16x2_3
+ KERNEL16x2_4
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL16x2_1
+ KERNEL16x2_2
+ KERNEL16x2_3
+ KERNEL16x2_4
+
+ je .L2_16
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL16x2_1
+ KERNEL16x2_2
+ KERNEL16x2_3
+ KERNEL16x2_4
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL16x2_1
+ KERNEL16x2_2
+ KERNEL16x2_3
+ KERNEL16x2_4
+
+ je .L2_16
+
+ jmp .L2_12
+ ALIGN_4
+
+.L2_16:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $7, %rax # if (k & 1)
+ je .L2_19
+
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), 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
+
+.L2_17:
+
+ KERNEL16x2_SUB
+
+ jl .L2_17
+ ALIGN_4
+
+
+.L2_19:
+
+ SAVE16x2
+
+#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,BI,1), 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 $16, KK
+#endif
+
+ addq $16 * SIZE, CO1 # coffset += 16
+ decq I # i --
+ jg .L2_11
+ ALIGN_4
+
+/**************************************************************************
+* Rest of M
+***************************************************************************/
+.L2_20:
+ // Test rest of M
+
+ testq $15, M
+ jz .L2_60 // to next 3 lines of N
+
+ testq $8, M
+ jz .L2_21pre
+ ALIGN_4
+
+/**************************************************************************/
+
+.L2_20_1:
+#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,BI,1), 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 $8, %rax // number of values in A
+#else
+ addq $2, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $-8, %rax
+ je .L2_20_6
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), 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
+
+.L2_20_2:
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL8x2_1
+ KERNEL8x2_2
+ KERNEL8x2_3
+ KERNEL8x2_4
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL8x2_1
+ KERNEL8x2_2
+ KERNEL8x2_3
+ KERNEL8x2_4
+
+ je .L2_20_6
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL8x2_1
+ KERNEL8x2_2
+ KERNEL8x2_3
+ KERNEL8x2_4
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL8x2_1
+ KERNEL8x2_2
+ KERNEL8x2_3
+ KERNEL8x2_4
+
+ je .L2_20_6
+
+ jmp .L2_20_2
+ ALIGN_4
+
+.L2_20_6:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $7, %rax # if (k & 1)
+ je .L2_20_9
+
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), 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
+
+.L2_20_7:
+
+ KERNEL8x2_SUB
+
+ jl .L2_20_7
+ ALIGN_4
+
+
+.L2_20_9:
+
+ 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,BI,1), 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 $8, KK
+#endif
+
+ addq $8 * SIZE, CO1 # coffset += 8
+ ALIGN_4
+
+
+
+/**************************************************************************/
+
+.L2_21pre:
+
+ testq $4, M
+ jz .L2_30
+ ALIGN_4
+
+.L2_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,BI,1), 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 $4, %rax // number of values in A
+#else
+ addq $2, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $-8, %rax
+ je .L2_26
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), BI // BI = BI * 1 ; 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_22:
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL4x2_1
+ KERNEL4x2_2
+ KERNEL4x2_3
+ KERNEL4x2_4
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL4x2_1
+ KERNEL4x2_2
+ KERNEL4x2_3
+ KERNEL4x2_4
+
+ je .L2_26
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL4x2_1
+ KERNEL4x2_2
+ KERNEL4x2_3
+ KERNEL4x2_4
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL4x2_1
+ KERNEL4x2_2
+ KERNEL4x2_3
+ KERNEL4x2_4
+
+ je .L2_26
+
+ jmp .L2_22
+ ALIGN_4
+
+.L2_26:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $7, %rax # if (k & 1)
+ je .L2_29
+
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), 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
+
+.L2_27:
+
+ KERNEL4x2_SUB
+
+ jl .L2_27
+ ALIGN_4
+
+
+.L2_29:
+
+ 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,BI,1), 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 $4, KK
+#endif
+
+ addq $4 * SIZE, CO1 # coffset += 4
+ ALIGN_4
+
+
+.L2_30:
+ testq $2, M
+ jz .L2_40
+
+ ALIGN_4
+
+.L2_31:
+#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,BI,1), 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 $2, %rax // number of values in AO
+#else
+ addq $2, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $-8, %rax
+ je .L2_36
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), 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
+
+.L2_32:
+
+ KERNEL2x2_1
+ KERNEL2x2_2
+ KERNEL2x2_3
+ KERNEL2x2_4
+
+ KERNEL2x2_1
+ KERNEL2x2_2
+ KERNEL2x2_3
+ KERNEL2x2_4
+
+ je .L2_36
+
+ KERNEL2x2_1
+ KERNEL2x2_2
+ KERNEL2x2_3
+ KERNEL2x2_4
+
+ KERNEL2x2_1
+ KERNEL2x2_2
+ KERNEL2x2_3
+ KERNEL2x2_4
+
+ je .L2_36
+
+ jmp .L2_32
+ ALIGN_4
+
+.L2_36:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $7, %rax # if (k & 1)
+ je .L2_39
+
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), 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
+
+.L2_37:
+
+ KERNEL2x2_SUB
+
+ jl .L2_37
+ ALIGN_4
+
+
+.L2_39:
+
+ SAVE2x2
+
+#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,BI,1), 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 $2, KK
+#endif
+
+ addq $2 * SIZE, CO1 # coffset += 2
+ ALIGN_4
+
+.L2_40:
+ testq $1, M
+ jz .L2_60 // to next 2 lines of N
+
+ ALIGN_4
+
+.L2_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,BI,1), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ 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
+ je .L2_46
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), BI // BI = BI * 2 ; number of values
+
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_42:
+
+ KERNEL1x2_1
+ KERNEL1x2_2
+ KERNEL1x2_3
+ KERNEL1x2_4
+
+ KERNEL1x2_1
+ KERNEL1x2_2
+ KERNEL1x2_3
+ KERNEL1x2_4
+
+ je .L2_46
+
+ KERNEL1x2_1
+ KERNEL1x2_2
+ KERNEL1x2_3
+ KERNEL1x2_4
+
+ KERNEL1x2_1
+ KERNEL1x2_2
+ KERNEL1x2_3
+ KERNEL1x2_4
+
+ je .L2_46
+
+ jmp .L2_42
+ ALIGN_4
+
+.L2_46:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $7, %rax # if (k & 1)
+ je .L2_49
+
+ movq %rax, BI // Index for BO
+ leaq (BI,BI,1), BI // BI = BI * 2 ; number of values
+
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L2_47:
+
+ KERNEL1x2_SUB
+
+ jl .L2_47
+ ALIGN_4
+
+
+.L2_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,BI,1), BI // BI = BI * 2 ; number of values
+ leaq (BO, BI, SIZE), BO
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $1, KK
+#endif
+
+ addq $1 * SIZE, CO1 # coffset += 1
+ ALIGN_4
+
+
+
+
+
+.L2_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 $1*SIZE,BO1
+ addq $1*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 $32 * SIZE, AO
+
+ movq M, I
+ sarq $4, I // i = (m >> 4)
+ je .L1_20
+
+ ALIGN_4
+
+.L1_11:
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $2 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $2 * SIZE, BO
+ movq %rax, BI // Index for BO
+ 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 $16, %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_16
+ movq %rax, BI // Index for BO
+
+ 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_12:
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL16x1_1
+ KERNEL16x1_2
+ KERNEL16x1_3
+ KERNEL16x1_4
+
+ KERNEL16x1_1
+ KERNEL16x1_2
+ KERNEL16x1_3
+ KERNEL16x1_4
+
+ je .L1_16
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL16x1_1
+ KERNEL16x1_2
+ KERNEL16x1_3
+ KERNEL16x1_4
+
+ KERNEL16x1_1
+ KERNEL16x1_2
+ KERNEL16x1_3
+ KERNEL16x1_4
+
+ je .L1_16
+
+ jmp .L1_12
+ ALIGN_4
+
+.L1_16:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $7, %rax # if (k & 1)
+ je .L1_19
+
+ movq %rax, BI // Index for BO
+
+ 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_17:
+
+ KERNEL16x1_SUB
+
+ jl .L1_17
+ ALIGN_4
+
+
+.L1_19:
+
+ SAVE16x1
+
+#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 (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 $16, KK
+#endif
+
+ addq $16 * SIZE, CO1 # coffset += 16
+ decq I # i --
+ jg .L1_11
+ ALIGN_4
+
+/**************************************************************************
+* Rest of M
+***************************************************************************/
+.L1_20:
+ // Test rest of M
+
+ testq $15, M
+ jz .L999
+
+ testq $8, M
+ jz .L1_21pre
+ ALIGN_4
+
+/**************************************************************************/
+
+.L1_20_1:
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $2 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $2 * SIZE, BO
+ movq %rax, BI // Index for BO
+ 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 $8, %rax // number of values in A
+#else
+ addq $1, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $-8, %rax
+ je .L1_20_6
+ movq %rax, BI // Index for BO
+
+ 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_20_2:
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL8x1_1
+ KERNEL8x1_2
+ KERNEL8x1_3
+ KERNEL8x1_4
+
+ KERNEL8x1_1
+ KERNEL8x1_2
+ KERNEL8x1_3
+ KERNEL8x1_4
+
+ je .L1_20_6
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL8x1_1
+ KERNEL8x1_2
+ KERNEL8x1_3
+ KERNEL8x1_4
+
+ KERNEL8x1_1
+ KERNEL8x1_2
+ KERNEL8x1_3
+ KERNEL8x1_4
+
+ je .L1_20_6
+
+ jmp .L1_20_2
+ ALIGN_4
+
+.L1_20_6:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $7, %rax # if (k & 1)
+ je .L1_20_9
+
+ movq %rax, BI // Index for BO
+
+ 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_20_7:
+
+ KERNEL8x1_SUB
+
+ jl .L1_20_7
+ ALIGN_4
+
+
+.L1_20_9:
+
+ 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 (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 $8, KK
+#endif
+
+ addq $8 * SIZE, CO1 # coffset += 8
+ ALIGN_4
+
+
+
+/**************************************************************************/
+
+.L1_21pre:
+
+ testq $4, M
+ jz .L1_30
+ ALIGN_4
+
+.L1_21:
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $2 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $2 * SIZE, BO
+ movq %rax, BI // Index for BO
+ 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 $4, %rax // number of values in A
+#else
+ addq $1, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $-8, %rax
+ je .L1_26
+ movq %rax, BI // Index for BO
+
+ 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_22:
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL4x1_1
+ KERNEL4x1_2
+ KERNEL4x1_3
+ KERNEL4x1_4
+
+ KERNEL4x1_1
+ KERNEL4x1_2
+ KERNEL4x1_3
+ KERNEL4x1_4
+
+ je .L1_26
+
+ prefetcht0 B_PR1(BO,BI,8)
+ KERNEL4x1_1
+ KERNEL4x1_2
+ KERNEL4x1_3
+ KERNEL4x1_4
+
+ KERNEL4x1_1
+ KERNEL4x1_2
+ KERNEL4x1_3
+ KERNEL4x1_4
+
+ je .L1_26
+
+ jmp .L1_22
+ ALIGN_4
+
+.L1_26:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $7, %rax # if (k & 1)
+ je .L1_29
+
+ movq %rax, BI // Index for BO
+
+ 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_27:
+
+ KERNEL4x1_SUB
+
+ jl .L1_27
+ ALIGN_4
+
+
+.L1_29:
+
+ 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 (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 $4, KK
+#endif
+
+ addq $4 * SIZE, CO1 # coffset += 4
+ ALIGN_4
+
+
+.L1_30:
+ testq $2, M
+ jz .L1_40
+
+ ALIGN_4
+
+.L1_31:
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $2 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $2 * SIZE, BO
+ movq %rax, BI // Index for BO
+ 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 $2, %rax // number of values in AO
+#else
+ addq $1, %rax // number of values in BO
+#endif
+ movq %rax, KKK
+#endif
+
+
+ andq $-8, %rax
+ je .L1_36
+ movq %rax, BI // Index for BO
+
+ 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_32:
+
+ KERNEL2x1_1
+ KERNEL2x1_2
+ KERNEL2x1_3
+ KERNEL2x1_4
+
+ KERNEL2x1_1
+ KERNEL2x1_2
+ KERNEL2x1_3
+ KERNEL2x1_4
+
+ je .L1_36
+
+ KERNEL2x1_1
+ KERNEL2x1_2
+ KERNEL2x1_3
+ KERNEL2x1_4
+
+ KERNEL2x1_1
+ KERNEL2x1_2
+ KERNEL2x1_3
+ KERNEL2x1_4
+
+ je .L1_36
+
+ jmp .L1_32
+ ALIGN_4
+
+.L1_36:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $7, %rax # if (k & 1)
+ je .L1_39
+
+ movq %rax, BI // Index for BO
+
+ 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_37:
+
+ KERNEL2x1_SUB
+
+ jl .L1_37
+ ALIGN_4
+
+
+.L1_39:
+
+ 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 (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 $2, KK
+#endif
+
+ addq $2 * SIZE, CO1 # coffset += 2
+ ALIGN_4
+
+.L1_40:
+ testq $1, M
+ jz .L999
+
+ ALIGN_4
+
+.L1_41:
+#if !defined(TRMMKERNEL) || \
+ (defined(TRMMKERNEL) && defined(LEFT) && defined(TRANSA)) || \
+ (defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))
+ leaq BUFFER1, BO // first buffer to BO
+ addq $2 * SIZE, BO
+#else
+ movq KK, %rax
+ leaq BUFFER1, BO // first buffer to BO
+ addq $2 * SIZE, BO
+ movq %rax, BI // Index for BO
+ leaq (BO, BI, SIZE), BO
+ 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
+ je .L1_46
+ movq %rax, BI // Index for BO
+
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_42:
+
+ KERNEL1x1_1
+ KERNEL1x1_2
+ KERNEL1x1_3
+ KERNEL1x1_4
+
+ KERNEL1x1_1
+ KERNEL1x1_2
+ KERNEL1x1_3
+ KERNEL1x1_4
+
+ je .L1_46
+
+ KERNEL1x1_1
+ KERNEL1x1_2
+ KERNEL1x1_3
+ KERNEL1x1_4
+
+ KERNEL1x1_1
+ KERNEL1x1_2
+ KERNEL1x1_3
+ KERNEL1x1_4
+
+ je .L1_46
+
+ jmp .L1_42
+ ALIGN_4
+
+.L1_46:
+#ifndef TRMMKERNEL
+ movq K, %rax
+#else
+ movq KKK, %rax
+#endif
+
+ andq $7, %rax # if (k & 1)
+ je .L1_49
+
+ movq %rax, BI // Index for BO
+
+ leaq (AO, %rax, SIZE), AO
+ leaq (BO, BI, SIZE), BO
+ negq BI
+ negq %rax
+ ALIGN_4
+
+.L1_47:
+
+ KERNEL1x1_SUB
+
+ jl .L1_47
+ ALIGN_4
+
+
+.L1_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 (BO, BI, SIZE), BO
+ leaq (AO, %rax, SIZE), AO
+#endif
+
+
+#if defined(TRMMKERNEL) && defined(LEFT)
+ addq $1, KK
+#endif
+
+ addq $1 * SIZE, CO1 # coffset += 1
+ ALIGN_4
+
+
+.L999:
+ 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
+ movups 64(%rsp), %xmm6
+ movups 80(%rsp), %xmm7
+ movups 96(%rsp), %xmm8
+ movups 112(%rsp), %xmm9
+ movups 128(%rsp), %xmm10
+ movups 144(%rsp), %xmm11
+ movups 160(%rsp), %xmm12
+ movups 176(%rsp), %xmm13
+ movups 192(%rsp), %xmm14
+ movups 208(%rsp), %xmm15
+#endif
+
+ addq $STACKSIZE, %rsp
+ ret
+
+ EPILOGUE
+
+
+
+
+
+#endif
projects / platform / upstream / openblas.git / commitdiff