gemm_kernel_8x4_barcelona.s

Optimized GotoBLAS libraries

/*********************************************************************/
/*                                                                   */
/*             Optimized BLAS libraries                              */
/*                     By Kazushige Goto <[e-mail]>     */
/*                                                                   */
/* Copyright (c) The University of Texas, 2005. All rights reserved. */
/* UNIVERSITY EXPRESSLY DISCLAIMS ANY AND ALL WARRANTIES CONCERNING  */
/* THIS SOFTWARE AND DOCUMENTATION, INCLUDING ANY WARRANTIES OF      */
/* MERCHANTABILITY, FITNESS FOR ANY PARTICULAR PURPOSE,              */
/* NON-INFRINGEMENT AND WARRANTIES OF PERFORMANCE, AND ANY WARRANTY  */
/* THAT MIGHT OTHERWISE ARISE FROM COURSE OF DEALING OR USAGE OF     */
/* TRADE. NO WARRANTY IS EITHER EXPRESS OR IMPLIED WITH RESPECT TO   */
/* THE USE OF THE SOFTWARE OR DOCUMENTATION.                         */
/* Under no circumstances shall University be liable for incidental, */
/* special, indirect, direct or consequential damages or loss of     */
/* profits, interruption of business, or related expenses which may  */
/* arise from use of Software or Documentation, including but not    */
/* limited to those resulting from defects in Software and/or        */
/* Documentation, or loss or inaccuracy of data of any kind.         */
/*********************************************************************/

#define ASSEMBLER
#include "common.h"
 
#define OLD_M	%rdi
#define OLD_N	%rsi
#define M	%r13
#define N	%r14
#define 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 CO2	%r12
#define BB	%rbp
	
#ifndef WINDOWS_ABI

#define STACKSIZE 64

#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 ALPHA	  0(%rsp)
#define J	 16(%rsp)
#define OFFSET	 24(%rsp)
#define KK	 32(%rsp)
#define KKK	 40(%rsp)
#define BUFFER	128(%rsp)

#ifdef OPTERON
#define movsd movlps
#endif

#define PREFETCH     prefetch
#define PREFETCHSIZE  (16 *  17 + 0)

#define RPREFETCHSIZE (16 *  9 + 0)
#define WPREFETCHSIZE (16 *  9 + 0)

#define KERNEL1(xx) \
	mulps	%xmm1, %xmm0 ;\
	mulps	-28 * SIZE(AO, %rax, 4), %xmm1 ;\
	addps	%xmm0, %xmm8 ;\
	PREFETCH (PREFETCHSIZE +  0) * SIZE(AO, %rax, 4) ;\
	movaps	%xmm2, %xmm0 ;\
	addps	%xmm1, %xmm12 ;\
	movaps	-24 * SIZE(BO, %rax, 8), %xmm1 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	-28 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm9 ;\
	movaps	%xmm0, %xmm2 ;\
	addps	%xmm3, %xmm13 ;\
	movaps	-20 * SIZE(BO, %rax, 8), %xmm3 ;\
	mulps	%xmm1, %xmm0 ;\
	mulps	-28 * SIZE(AO, %rax, 4), %xmm1 ;\
	addps	%xmm0, %xmm10 ;\
	movaps	-24 * SIZE(AO, %rax, 4), %xmm0 ;\
	addps	%xmm1, %xmm14 ;\
	movaps	-16 * SIZE(BO, %rax, 8), %xmm1 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	-28 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm11 ;\
	addps	%xmm3, %xmm15 ;\
 	movaps	-12 * SIZE(BO, %rax, 8), %xmm3 ;\
	movaps	%xmm0, %xmm2

#define KERNEL2(xx) \
	mulps	%xmm1, %xmm0 ;\
	mulps	-20 * SIZE(AO, %rax, 4), %xmm1 ;\
	addps	%xmm0, %xmm8 ;\
	movaps	%xmm2, %xmm0 ;\
	addps	%xmm1, %xmm12 ;\
	movaps	 -8 * SIZE(BO, %rax, 8), %xmm1 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	-20 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm9 ;\
	movaps	%xmm0, %xmm2 ;\
	addps	%xmm3, %xmm13 ;\
	movaps	 -4 * SIZE(BO, %rax, 8), %xmm3 ;\
	mulps	%xmm1, %xmm0 ;\
	mulps	-20 * SIZE(AO, %rax, 4), %xmm1 ;\
	addps	%xmm0, %xmm10 ;\
	addps	%xmm1, %xmm14 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	-20 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm11 ;\
	addps	%xmm3, %xmm15 ;\
 	movaps	  4 * SIZE(BO, %rax, 8), %xmm3 ;\
	movaps	%xmm4, %xmm2

#define KERNEL3(xx) \
	mulps	%xmm5, %xmm4 ;\
	mulps	-12 * SIZE(AO, %rax, 4), %xmm5 ;\
	addps	%xmm4, %xmm8 ;\
	movaps	 32 * SIZE(BO, %rax, 8), %xmm1 ;\
	movaps	%xmm2, %xmm4 ;\
	addps	%xmm5, %xmm12 ;\
	movaps	  8 * SIZE(BO, %rax, 8), %xmm5 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	-12 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm9 ;\
	movaps	%xmm4, %xmm2 ;\
	addps	%xmm3, %xmm13 ;\
	movaps	 12 * SIZE(BO, %rax, 8), %xmm3 ;\
	mulps	%xmm5, %xmm4 ;\
	mulps	-12 * SIZE(AO, %rax, 4), %xmm5 ;\
	addps	%xmm4, %xmm10 ;\
	movaps	 -8 * SIZE(AO, %rax, 4), %xmm4 ;\
	addps	%xmm5, %xmm14 ;\
	movaps	 16 * SIZE(BO, %rax, 8), %xmm5 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	-12 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm11 ;\
	addps	%xmm3, %xmm15 ;\
 	movaps	 20 * SIZE(BO, %rax, 8), %xmm3 ;\
	movaps	%xmm4, %xmm2

#define KERNEL4(xx) \
	mulps	%xmm5, %xmm4 ;\
	mulps	 -4 * SIZE(AO, %rax, 4), %xmm5 ;\
	movaps	  (AO, %rax, 4), %xmm6 ;\
	addps	%xmm4, %xmm8 ;\
	movaps	%xmm2, %xmm4 ;\
	addps	%xmm5, %xmm12 ;\
	movaps	 24 * SIZE(BO, %rax, 8), %xmm5 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	 -4 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm9 ;\
	movaps	%xmm4, %xmm2 ;\
	addps	%xmm3, %xmm13 ;\
	movaps	 28 * SIZE(BO, %rax, 8), %xmm3 ;\
	mulps	%xmm5, %xmm4 ;\
	mulps	 -4 * SIZE(AO, %rax, 4), %xmm5 ;\
	addps	%xmm4, %xmm10 ;\
	addps	%xmm5, %xmm14 ;\
	movaps	 64 * SIZE(BO, %rax, 8), %xmm5 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	 -4 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm11 ;\
	addps	%xmm3, %xmm15 ;\
 	movaps	 36 * SIZE(BO, %rax, 8), %xmm3 ;\
	movaps	%xmm6, %xmm2

#define KERNEL5(xx) \
	mulps	%xmm1, %xmm6 ;\
	mulps	  4 * SIZE(AO, %rax, 4), %xmm1 ;\
	addps	%xmm6, %xmm8 ;\
	movaps	%xmm2, %xmm6 ;\
	addps	%xmm1, %xmm12 ;\
	movaps	 40 * SIZE(BO, %rax, 8), %xmm1 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	  4 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm9 ;\
	movaps	 16 * SIZE(AO, %rax, 4), %xmm7 ;\
	movaps	%xmm6, %xmm2 ;\
	addps	%xmm3, %xmm13 ;\
	movaps	 44 * SIZE(BO, %rax, 8), %xmm3 ;\
	mulps	%xmm1, %xmm6 ;\
	mulps	  4 * SIZE(AO, %rax, 4), %xmm1 ;\
	addps	%xmm6, %xmm10 ;\
	movaps	  8 * SIZE(AO, %rax, 4), %xmm6 ;\
	addps	%xmm1, %xmm14 ;\
	movaps	 48 * SIZE(BO, %rax, 8), %xmm1 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	  4 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm11 ;\
	addps	%xmm3, %xmm15 ;\
 	movaps	 52 * SIZE(BO, %rax, 8), %xmm3 ;\
	movaps	%xmm6, %xmm2

#define KERNEL6(xx) \
	mulps	%xmm1, %xmm6 ;\
	mulps	 12 * SIZE(AO, %rax, 4), %xmm1 ;\
	addps	%xmm6, %xmm8 ;\
	movaps	%xmm2, %xmm6 ;\
	addps	%xmm1, %xmm12 ;\
	movaps	 56 * SIZE(BO, %rax, 8), %xmm1 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	 12 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm9 ;\
	movaps	%xmm6, %xmm2 ;\
	addps	%xmm3, %xmm13 ;\
	movaps	 60 * SIZE(BO, %rax, 8), %xmm3 ;\
	mulps	%xmm1, %xmm6 ;\
	mulps	 12 * SIZE(AO, %rax, 4), %xmm1 ;\
	addps	%xmm6, %xmm10 ;\
	movaps	 32 * SIZE(AO, %rax, 4), %xmm0 ;\
	addps	%xmm1, %xmm14 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	 12 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm11 ;\
	addps	%xmm3, %xmm15 ;\
 	movaps	 68 * SIZE(BO, %rax, 8), %xmm3 ;\
	movaps	%xmm7, %xmm2

#define KERNEL7(xx) \
	mulps	%xmm5, %xmm7 ;\
	mulps	 20 * SIZE(AO, %rax, 4), %xmm5 ;\
	addps	%xmm7, %xmm8 ;\
	movaps	 96 * SIZE(BO, %rax, 8), %xmm1 ;\
	movaps	%xmm2, %xmm7 ;\
	addps	%xmm5, %xmm12 ;\
	movaps	 72 * SIZE(BO, %rax, 8), %xmm5 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	 20 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm9 ;\
	movaps	%xmm7, %xmm2 ;\
	addps	%xmm3, %xmm13 ;\
	movaps	 76 * SIZE(BO, %rax, 8), %xmm3 ;\
	mulps	%xmm5, %xmm7 ;\
	mulps	 20 * SIZE(AO, %rax, 4), %xmm5 ;\
	addps	%xmm7, %xmm10 ;\
	movaps	 24 * SIZE(AO, %rax, 4), %xmm7 ;\
	addps	%xmm5, %xmm14 ;\
	movaps	 80 * SIZE(BO, %rax, 8), %xmm5 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	 20 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm11 ;\
	addps	%xmm3, %xmm15 ;\
 	movaps	 84 * SIZE(BO, %rax, 8), %xmm3 ;\
	movaps	%xmm7, %xmm2

#define KERNEL8(xx) \
	mulps	%xmm5, %xmm7 ;\
	mulps	 28 * SIZE(AO, %rax, 4), %xmm5 ;\
	addps	%xmm7, %xmm8 ;\
	movaps	%xmm2, %xmm7 ;\
	addps	%xmm5, %xmm12 ;\
	movaps	 88 * SIZE(BO, %rax, 8), %xmm5 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	 28 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm9 ;\
	movaps	%xmm7, %xmm2 ;\
	addps	%xmm3, %xmm13 ;\
	movaps	 92 * SIZE(BO, %rax, 8), %xmm3 ;\
	mulps	%xmm5, %xmm7 ;\
	mulps	 28 * SIZE(AO, %rax, 4), %xmm5 ;\
	addps	%xmm7, %xmm10 ;\
	movaps	 48 * SIZE(AO, %rax, 4), %xmm4 ;\
	addps	%xmm5, %xmm14 ;\
	movaps	128 * SIZE(BO, %rax, 8), %xmm5 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	 28 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm11 ;\
	addps	%xmm3, %xmm15 ;\
 	movaps	100 * SIZE(BO, %rax, 8), %xmm3 ;\
	movaps	%xmm0, %xmm2 ;\
	addq	$16 * SIZE, %rax
	
#define KERNEL_SUB1(xx) \
	mulps	%xmm1, %xmm0 ;\
	mulps	-28 * SIZE(AO, %rax, 4), %xmm1 ;\
	addps	%xmm0, %xmm8 ;\
	movaps	%xmm2, %xmm0 ;\
	addps	%xmm1, %xmm12 ;\
	movaps	-24 * SIZE(BO, %rax, 8), %xmm1 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	-28 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm9 ;\
	movaps	%xmm0, %xmm2 ;\
	addps	%xmm3, %xmm13 ;\
	movaps	-20 * SIZE(BO, %rax, 8), %xmm3 ;\
	mulps	%xmm1, %xmm0 ;\
	mulps	-28 * SIZE(AO, %rax, 4), %xmm1 ;\
	addps	%xmm0, %xmm10 ;\
	movaps	-24 * SIZE(AO, %rax, 4), %xmm0 ;\
	addps	%xmm1, %xmm14 ;\
	movaps	-16 * SIZE(BO, %rax, 8), %xmm1 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	-28 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm11 ;\
	addps	%xmm3, %xmm15 ;\
 	movaps	-12 * SIZE(BO, %rax, 8), %xmm3 ;\
	movaps	%xmm0, %xmm2

#define KERNEL_SUB2(xx) \
	mulps	%xmm1, %xmm0 ;\
	mulps	-20 * SIZE(AO, %rax, 4), %xmm1 ;\
	addps	%xmm0, %xmm8 ;\
	movaps	%xmm2, %xmm0 ;\
	addps	%xmm1, %xmm12 ;\
	movaps	 -8 * SIZE(BO, %rax, 8), %xmm1 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	-20 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm9 ;\
	movaps	%xmm0, %xmm2 ;\
	addps	%xmm3, %xmm13 ;\
	movaps	 -4 * SIZE(BO, %rax, 8), %xmm3 ;\
	mulps	%xmm1, %xmm0 ;\
	mulps	-20 * SIZE(AO, %rax, 4), %xmm1 ;\
	addps	%xmm0, %xmm10 ;\
	movaps	  (AO, %rax, 4), %xmm0 ;\
	addps	%xmm1, %xmm14 ;\
	movaps	 32 * SIZE(BO, %rax, 8), %xmm1 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	-20 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm11 ;\
	addps	%xmm3, %xmm15 ;\
 	movaps	  4 * SIZE(BO, %rax, 8), %xmm3 ;\
	movaps	%xmm4, %xmm2

#define KERNEL_SUB3(xx) \
	mulps	%xmm5, %xmm4 ;\
	mulps	-12 * SIZE(AO, %rax, 4), %xmm5 ;\
	addps	%xmm4, %xmm8 ;\
	movaps	%xmm2, %xmm4 ;\
	addps	%xmm5, %xmm12 ;\
	movaps	  8 * SIZE(BO, %rax, 8), %xmm5 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	-12 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm9 ;\
	movaps	%xmm4, %xmm2 ;\
	addps	%xmm3, %xmm13 ;\
	movaps	 12 * SIZE(BO, %rax, 8), %xmm3 ;\
	mulps	%xmm5, %xmm4 ;\
	mulps	-12 * SIZE(AO, %rax, 4), %xmm5 ;\
	addps	%xmm4, %xmm10 ;\
	movaps	 -8 * SIZE(AO, %rax, 4), %xmm4 ;\
	addps	%xmm5, %xmm14 ;\
	movaps	 16 * SIZE(BO, %rax, 8), %xmm5 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	-12 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm11 ;\
	addps	%xmm3, %xmm15 ;\
 	movaps	 20 * SIZE(BO, %rax, 8), %xmm3 ;\
	movaps	%xmm4, %xmm2

#define KERNEL_SUB4(xx) \
	mulps	%xmm5, %xmm4 ;\
	mulps	 -4 * SIZE(AO, %rax, 4), %xmm5 ;\
	addps	%xmm4, %xmm8 ;\
	movaps	%xmm2, %xmm4 ;\
	addps	%xmm5, %xmm12 ;\
	movaps	 24 * SIZE(BO, %rax, 8), %xmm5 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	 -4 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm9 ;\
	movaps	%xmm4, %xmm2 ;\
	addps	%xmm3, %xmm13 ;\
	movaps	 28 * SIZE(BO, %rax, 8), %xmm3 ;\
	mulps	%xmm5, %xmm4 ;\
	mulps	 -4 * SIZE(AO, %rax, 4), %xmm5 ;\
	addps	%xmm4, %xmm10 ;\
	addps	%xmm5, %xmm14 ;\
	mulps	%xmm3, %xmm2 ;\
	mulps	 -4 * SIZE(AO, %rax, 4), %xmm3 ;\
	addps	%xmm2, %xmm11 ;\
	addps	%xmm3, %xmm15 ;\
 	movaps	 36 * SIZE(BO, %rax, 8), %xmm3 ;\
	movaps	%xmm0, %xmm2

	PROLOGUE
	PROFCODE
	
	subq	$STACKSIZE, %rsp

	movq	%rbx,  0(%rsp)
	movq	%rbp,  8(%rsp)
	movq	%r12, 16(%rsp)
	movq	%r13, 24(%rsp)
	movq	%r14, 32(%rsp)
	movq	%r15, 40(%rsp)

#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,      K
	movq	OLD_A,     A
	movq	OLD_B,     B
	movq	OLD_C,     C
	movq	OLD_LDC,   LDC
#ifdef TRMMKERNEL
	movsd	OLD_OFFSET, %xmm12
#endif
	movaps	%xmm3, %xmm0

#else
	movq	72(%rsp), LDC
#ifdef TRMMKERNEL
	movsd	80(%rsp), %xmm12
#endif

#endif

	movq	%rsp, %rbx	# save old stack
	subq	$128 + GEMM_Q * 64, %rsp
	andq	$-4096, %rsp	# align stack

#ifdef WINDOWS_ABI
#if GEMM_Q > 192
	movq	$0,  4096 * 3(%rsp)
#endif
#if GEMM_Q > 128
	movq	$0,  4096 * 2(%rsp)
#endif
#if GEMM_Q >  64
	movq	$0,  4096 * 1(%rsp)
#endif
	movq	$0,  4096 * 0(%rsp)
#endif

	EMMS
	
	movq	OLD_M, M
	movq	OLD_N, N

	shufps	$0, %xmm0, %xmm0
	movaps	%xmm0, ALPHA
	
#ifdef TRMMKERNEL
	movsd	%xmm4, OFFSET
	movsd	%xmm4, KK
#ifndef LEFT
	negq	KK
#endif	
#endif

	subq	$-32 * SIZE, A

	leaq	(, LDC, SIZE), LDC

	movq	N,  J
	sarq	$2, J		# j = (n >> 2)
	jle	.L50

.L01:
#if defined(TRMMKERNEL) && defined(LEFT)
	movq	OFFSET, %rax
	movq	%rax, KK
#endif	

/* Copying to Sub Buffer */
	leaq	BUFFER, BO
	
	movq	K,  %rax
	sarq	$2, %rax
	jle	.L03
	ALIGN_4
	
.L02:
	prefetch	 (RPREFETCHSIZE +  0)  * SIZE(B)

	movaps	 0 * SIZE(B), %xmm3
	movaps	 4 * SIZE(B), %xmm7
	movaps	 8 * SIZE(B), %xmm11
	movaps	12 * SIZE(B), %xmm15

	prefetchw	 (WPREFETCHSIZE +  0) * SIZE(BO)

	pshufd	$0x00, %xmm3, %xmm0
	pshufd	$0x55, %xmm3, %xmm1
	pshufd	$0xaa, %xmm3, %xmm2
	pshufd	$0xff, %xmm3, %xmm3

	prefetchw	 (WPREFETCHSIZE + 16) * SIZE(BO)

	pshufd	$0x00, %xmm7, %xmm4
	pshufd	$0x55, %xmm7, %xmm5
	pshufd	$0xaa, %xmm7, %xmm6
	pshufd	$0xff, %xmm7, %xmm7

	movaps	%xmm0,   0 * SIZE(BO)
	movaps	%xmm1,   4 * SIZE(BO)
	movaps	%xmm2,   8 * SIZE(BO)
	movaps	%xmm3,  12 * SIZE(BO)
	movaps	%xmm4,  16 * SIZE(BO)
	movaps	%xmm5,  20 * SIZE(BO)
	movaps	%xmm6,  24 * SIZE(BO)
	movaps	%xmm7,  28 * SIZE(BO)

	prefetchw	 (WPREFETCHSIZE + 32) * SIZE(BO)

	pshufd	$0x00, %xmm11, %xmm0
	pshufd	$0x55, %xmm11, %xmm1
	pshufd	$0xaa, %xmm11, %xmm2
	pshufd	$0xff, %xmm11, %xmm3

	prefetchw	 (WPREFETCHSIZE + 48) * SIZE(BO)

	pshufd	$0x00, %xmm15, %xmm4
	pshufd	$0x55, %xmm15, %xmm5
	pshufd	$0xaa, %xmm15, %xmm6
	pshufd	$0xff, %xmm15, %xmm7

	movaps	%xmm0,  32 * SIZE(BO)
	movaps	%xmm1,  36 * SIZE(BO)
	movaps	%xmm2,  40 * SIZE(BO)
	movaps	%xmm3,  44 * SIZE(BO)
	movaps	%xmm4,  48 * SIZE(BO)
	movaps	%xmm5,  52 * SIZE(BO)
	movaps	%xmm6,  56 * SIZE(BO)
	movaps	%xmm7,  60 * SIZE(BO)

	addq	$16 * SIZE, B
	addq	$64 * SIZE, BO

	decq	%rax
	jne	.L02
	ALIGN_4

.L03:
	movq	K,  %rax
	andq	$3, %rax
	BRANCH
	jle	.L10
	ALIGN_4

.L04:
	movaps	 0 * SIZE(B), %xmm3

	pshufd	 $0x00, %xmm3, %xmm0
	pshufd	 $0x55, %xmm3, %xmm1
	pshufd	 $0xaa, %xmm3, %xmm2
	pshufd	 $0xff, %xmm3, %xmm3

	movaps	%xmm0,  0 * SIZE(BO)
	movaps	%xmm1,  4 * SIZE(BO)
	movaps	%xmm2,  8 * SIZE(BO)
	movaps	%xmm3, 12 * SIZE(BO)

	addq	$ 4 * SIZE, B
	addq	$16 * SIZE, BO
	decq	%rax
	jne	.L04
	ALIGN_4
	
.L10:
	movq	C, CO1
	leaq	(C, LDC, 1), CO2
	movq	A, AO

	leaq	 (RPREFETCHSIZE +  0)  * SIZE(B), BB

	movq	M,  I
	sarq	$3, I	# i = (m >> 3)
	jle	.L20
	ALIGN_4

.L11:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	leaq	32 * SIZE + BUFFER, BO
#else
	leaq	32 * SIZE + BUFFER, BO
	movq	KK, %rax
	leaq	(, %rax,   8), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 8), BO
#endif	

	prefetch	  0  * SIZE(BB)
	prefetch	 16  * SIZE(BB)
	subq		 $-32 * SIZE, BB

	movaps	-32 * SIZE(AO), %xmm0
	movaps	-32 * SIZE(BO), %xmm1
	pxor	%xmm8, %xmm8
 	movaps	-28 * SIZE(BO), %xmm3
	pxor	%xmm9, %xmm9
	movaps	-16 * SIZE(AO), %xmm4
	pxor	%xmm10, %xmm10
	movaps	  0 * SIZE(BO), %xmm5
	pxor	%xmm11, %xmm11

	prefetchw      7 * SIZE(CO1)
	pxor	%xmm12, %xmm12
	prefetchw      7 * SIZE(CO2)
	pxor	%xmm13, %xmm13
	prefetchw      7 * SIZE(CO1, LDC, 2)
	pxor	%xmm14, %xmm14
	prefetchw      7 * SIZE(CO2, LDC, 2)
	pxor	%xmm15, %xmm15

	movaps	%xmm0, %xmm2

#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
#else