zgemm3m_kernel_4x4_sse2.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	%rbp
#define BB	%r12

#ifndef WINDOWS_ABI

#define STACKSIZE 64

#define OLD_LDC		 8 + STACKSIZE(%rsp)
#define OLD_OFFSET	16 + STACKSIZE(%rsp)

#else

#define STACKSIZE 256

#define OLD_ALPHA_I	40 + STACKSIZE(%rsp)
#define OLD_A		48 + STACKSIZE(%rsp)
#define OLD_B		56 + STACKSIZE(%rsp)
#define OLD_C		64 + STACKSIZE(%rsp)
#define OLD_LDC		72 + STACKSIZE(%rsp)
#define OLD_OFFSET	80 + STACKSIZE(%rsp)

#endif

#define 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 PREFETCH     prefetch
#define PREFETCHW    prefetchw
#define PREFETCHSIZE (8 * 5 + 4)
#define movsd	movlps
#define movapd	movaps
#endif

#ifdef GENERIC
#define PREFETCH     prefetcht0
#define PREFETCHW    prefetcht0
#define PREFETCHSIZE (8 * 5 + 4)
#define movapd	movaps
#endif

#define LOCAL_BUFFER_SIZE  DGEMM_Q * DGEMM_UNROLL_N * 16

#ifndef GENERIC
#define KERNEL1(xx) \
	mulpd	%xmm0, %xmm1 ;\
	addpd	%xmm1, %xmm8 ;\
	movapd	-16 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm1 ;\
	mulpd	%xmm0, %xmm3 ;\
	addpd	%xmm3, %xmm9 ;\
	movapd	-14 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm3 ;\
	mulpd	%xmm0, %xmm5 ;\
	PREFETCH	(PREFETCHSIZE     +  0) * SIZE + 1 * (xx) * SIZE(AO, %rax, 4) ;\
	mulpd	-10 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm0 ;\
	addpd	%xmm5, %xmm10 ;\
	movapd	-12 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm5 ;\
	addpd	%xmm0, %xmm11 ;\
	movapd	-8 * SIZE + 1 * (xx) * SIZE(AO, %rax, 4), %xmm0

#define KERNEL2(xx) \
	mulpd	%xmm2, %xmm1 ;\
	addpd	%xmm1, %xmm12 ;\
	movapd	 0 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm1 ;\
	mulpd	%xmm2, %xmm3 ;\
	addpd	%xmm3, %xmm13 ;\
	movapd	-6 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm3 ;\
	mulpd	%xmm2, %xmm5 ;\
	mulpd	-10 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm2 ;\
	addpd	%xmm5, %xmm14 ;\
	movapd	 -4 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm5 ;\
	addpd	%xmm2, %xmm15 ;\
	movapd	-6 * SIZE + 1 * (xx) * SIZE(AO, %rax, 4), %xmm2

#define KERNEL3(xx) \
	mulpd	%xmm4, %xmm7 ;\
	addpd	%xmm7, %xmm8 ;\
	movapd	-8 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm7 ;\
	mulpd	%xmm4, %xmm3 ;\
	addpd	%xmm3, %xmm9 ;\
	movapd	-6 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm3 ;\
	mulpd	%xmm4, %xmm5 ;\
	mulpd	-2 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm4 ;\
	addpd	%xmm5, %xmm10 ;\
	movapd	-4 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm5 ;\
	addpd	%xmm4, %xmm11 ;\
	movapd	-4 * SIZE + 1 * (xx) * SIZE(AO, %rax, 4), %xmm4

#define KERNEL4(xx) \
	mulpd	%xmm6, %xmm7 ;\
	addpd	%xmm7, %xmm12 ;\
	movapd	 8 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm7 ;\
	mulpd	%xmm6, %xmm3 ;\
	addpd	%xmm3, %xmm13 ;\
	movapd	 2 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm3 ;\
	mulpd	%xmm6, %xmm5 ;\
	mulpd	-2 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm6 ;\
	addpd	%xmm5, %xmm14 ;\
	movapd	 4 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm5 ;\
 	PREFETCH	(PREFETCHSIZE     +  8) * SIZE + 1 * (xx) * SIZE(AO, %rax, 4) ;\
	addpd	%xmm6, %xmm15 ;\
	movapd	-2 * SIZE + 1 * (xx) * SIZE(AO, %rax, 4), %xmm6

#define KERNEL5(xx) \
	mulpd	%xmm0, %xmm1 ;\
	addpd	%xmm1, %xmm8 ;\
	movapd	 0 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm1 ;\
	mulpd	%xmm0, %xmm3 ;\
	addpd	%xmm3, %xmm9 ;\
	movapd	 2 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm3 ;\
	mulpd	%xmm0, %xmm5 ;\
	mulpd	 6 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm0 ;\
	addpd	%xmm5, %xmm10 ;\
	movapd	 4 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm5 ;\
	addpd	%xmm0, %xmm11 ;\
	movapd	 0 * SIZE + 1 * (xx) * SIZE(AO, %rax, 4), %xmm0

#define KERNEL6(xx) \
	mulpd	%xmm2, %xmm1 ;\
	addpd	%xmm1, %xmm12 ;\
	movapd	16 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm1 ;\
	mulpd	%xmm2, %xmm3 ;\
	addpd	%xmm3, %xmm13 ;\
	movapd	10 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm3 ;\
	mulpd	%xmm2, %xmm5 ;\
	mulpd	 6 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm2 ;\
	addpd	%xmm5, %xmm14 ;\
	movapd	12 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm5 ;\
	addpd	%xmm2, %xmm15 ;\
	movapd	 2 * SIZE + 1 * (xx) * SIZE(AO, %rax, 4), %xmm2

#define KERNEL7(xx) \
	mulpd	%xmm4, %xmm7 ;\
	addpd	%xmm7, %xmm8 ;\
	movapd	 8 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm7 ;\
	mulpd	%xmm4, %xmm3 ;\
	addpd	%xmm3, %xmm9 ;\
	movapd	10 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm3 ;\
	mulpd	%xmm4, %xmm5 ;\
	mulpd	14 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm4 ;\
	addpd	%xmm5, %xmm10 ;\
	movapd	12 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm5 ;\
	addpd	%xmm4, %xmm11 ;\
	movapd	 4 * SIZE + 1 * (xx) * SIZE(AO, %rax, 4), %xmm4

#define KERNEL8(xx) \
	mulpd	%xmm6, %xmm7 ;\
	addpd	%xmm7, %xmm12 ;\
	movapd	24 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm7 ;\
	mulpd	%xmm6, %xmm3 ;\
	addpd	%xmm3, %xmm13 ;\
	movapd	18 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm3 ;\
	mulpd	%xmm6, %xmm5 ;\
	mulpd	14 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm6 ;\
	addpd	%xmm5, %xmm14 ;\
	movapd	20 * SIZE + 2 * (xx) * SIZE(BO, %rax, 8), %xmm5 ;\
	addpd	%xmm6, %xmm15 ;\
	movapd	 6 * SIZE + 1 * (xx) * SIZE(AO, %rax, 4), %xmm6

#else

#define KERNEL1(xx) \
	mulpd	%xmm0, %xmm1 ;\
	addpd	%xmm1, %xmm8 ;\
	movapd	-16 * SIZE + 2 * (xx) * SIZE(BO), %xmm1 ;\
	mulpd	%xmm0, %xmm3 ;\
	addpd	%xmm3, %xmm9 ;\
	movapd	-14 * SIZE + 2 * (xx) * SIZE(BO), %xmm3 ;\
	mulpd	%xmm0, %xmm5 ;\
	PREFETCH	(PREFETCHSIZE     +  0) * SIZE + 1 * (xx) * SIZE(AO) ;\
	mulpd	-10 * SIZE + 2 * (xx) * SIZE(BO), %xmm0 ;\
	addpd	%xmm5, %xmm10 ;\
	movapd	-12 * SIZE + 2 * (xx) * SIZE(BO), %xmm5 ;\
	addpd	%xmm0, %xmm11 ;\
	movapd	-8 * SIZE + 1 * (xx) * SIZE(AO), %xmm0

#define KERNEL2(xx) \
	mulpd	%xmm2, %xmm1 ;\
	addpd	%xmm1, %xmm12 ;\
	movapd	 0 * SIZE + 2 * (xx) * SIZE(BO), %xmm1 ;\
	mulpd	%xmm2, %xmm3 ;\
	addpd	%xmm3, %xmm13 ;\
	movapd	-6 * SIZE + 2 * (xx) * SIZE(BO), %xmm3 ;\
	mulpd	%xmm2, %xmm5 ;\
	mulpd	-10 * SIZE + 2 * (xx) * SIZE(BO), %xmm2 ;\
	addpd	%xmm5, %xmm14 ;\
	movapd	 -4 * SIZE + 2 * (xx) * SIZE(BO), %xmm5 ;\
	addpd	%xmm2, %xmm15 ;\
	movapd	-6 * SIZE + 1 * (xx) * SIZE(AO), %xmm2

#define KERNEL3(xx) \
	mulpd	%xmm4, %xmm7 ;\
	addpd	%xmm7, %xmm8 ;\
	movapd	-8 * SIZE + 2 * (xx) * SIZE(BO), %xmm7 ;\
	mulpd	%xmm4, %xmm3 ;\
	addpd	%xmm3, %xmm9 ;\
	movapd	-6 * SIZE + 2 * (xx) * SIZE(BO), %xmm3 ;\
	mulpd	%xmm4, %xmm5 ;\
	mulpd	-2 * SIZE + 2 * (xx) * SIZE(BO), %xmm4 ;\
	addpd	%xmm5, %xmm10 ;\
	movapd	-4 * SIZE + 2 * (xx) * SIZE(BO), %xmm5 ;\
	addpd	%xmm4, %xmm11 ;\
	movapd	-4 * SIZE + 1 * (xx) * SIZE(AO), %xmm4

#define KERNEL4(xx) \
	mulpd	%xmm6, %xmm7 ;\
	addpd	%xmm7, %xmm12 ;\
	movapd	 8 * SIZE + 2 * (xx) * SIZE(BO), %xmm7 ;\
	mulpd	%xmm6, %xmm3 ;\
	addpd	%xmm3, %xmm13 ;\
	movapd	 2 * SIZE + 2 * (xx) * SIZE(BO), %xmm3 ;\
	mulpd	%xmm6, %xmm5 ;\
	mulpd	-2 * SIZE + 2 * (xx) * SIZE(BO), %xmm6 ;\
	addpd	%xmm5, %xmm14 ;\
	movapd	 4 * SIZE + 2 * (xx) * SIZE(BO), %xmm5 ;\
 	PREFETCH	(PREFETCHSIZE     +  8) * SIZE + 1 * (xx) * SIZE(AO) ;\
	addpd	%xmm6, %xmm15 ;\
	movapd	-2 * SIZE + 1 * (xx) * SIZE(AO), %xmm6

#define KERNEL5(xx) \
	mulpd	%xmm0, %xmm1 ;\
	addpd	%xmm1, %xmm8 ;\
	movapd	 0 * SIZE + 2 * (xx) * SIZE(BO), %xmm1 ;\
	mulpd	%xmm0, %xmm3 ;\
	addpd	%xmm3, %xmm9 ;\
	movapd	 2 * SIZE + 2 * (xx) * SIZE(BO), %xmm3 ;\
	mulpd	%xmm0, %xmm5 ;\
	mulpd	 6 * SIZE + 2 * (xx) * SIZE(BO), %xmm0 ;\
	addpd	%xmm5, %xmm10 ;\
	movapd	 4 * SIZE + 2 * (xx) * SIZE(BO), %xmm5 ;\
	addpd	%xmm0, %xmm11 ;\
	movapd	 0 * SIZE + 1 * (xx) * SIZE(AO), %xmm0

#define KERNEL6(xx) \
	mulpd	%xmm2, %xmm1 ;\
	addpd	%xmm1, %xmm12 ;\
	movapd	16 * SIZE + 2 * (xx) * SIZE(BO), %xmm1 ;\
	mulpd	%xmm2, %xmm3 ;\
	addpd	%xmm3, %xmm13 ;\
	movapd	10 * SIZE + 2 * (xx) * SIZE(BO), %xmm3 ;\
	mulpd	%xmm2, %xmm5 ;\
	mulpd	 6 * SIZE + 2 * (xx) * SIZE(BO), %xmm2 ;\
	addpd	%xmm5, %xmm14 ;\
	movapd	12 * SIZE + 2 * (xx) * SIZE(BO), %xmm5 ;\
	addpd	%xmm2, %xmm15 ;\
	movapd	 2 * SIZE + 1 * (xx) * SIZE(AO), %xmm2

#define KERNEL7(xx) \
	mulpd	%xmm4, %xmm7 ;\
	addpd	%xmm7, %xmm8 ;\
	movapd	 8 * SIZE + 2 * (xx) * SIZE(BO), %xmm7 ;\
	mulpd	%xmm4, %xmm3 ;\
	addpd	%xmm3, %xmm9 ;\
	movapd	10 * SIZE + 2 * (xx) * SIZE(BO), %xmm3 ;\
	mulpd	%xmm4, %xmm5 ;\
	mulpd	14 * SIZE + 2 * (xx) * SIZE(BO), %xmm4 ;\
	addpd	%xmm5, %xmm10 ;\
	movapd	12 * SIZE + 2 * (xx) * SIZE(BO), %xmm5 ;\
	addpd	%xmm4, %xmm11 ;\
	movapd	 4 * SIZE + 1 * (xx) * SIZE(AO), %xmm4

#define KERNEL8(xx) \
	mulpd	%xmm6, %xmm7 ;\
	addpd	%xmm7, %xmm12 ;\
	movapd	24 * SIZE + 2 * (xx) * SIZE(BO), %xmm7 ;\
	mulpd	%xmm6, %xmm3 ;\
	addpd	%xmm3, %xmm13 ;\
	movapd	18 * SIZE + 2 * (xx) * SIZE(BO), %xmm3 ;\
	mulpd	%xmm6, %xmm5 ;\
	mulpd	14 * SIZE + 2 * (xx) * SIZE(BO), %xmm6 ;\
	addpd	%xmm5, %xmm14 ;\
	movapd	20 * SIZE + 2 * (xx) * SIZE(BO), %xmm5 ;\
	addpd	%xmm6, %xmm15 ;\
	movapd	 6 * SIZE + 1 * (xx) * SIZE(AO), %xmm6
#endif	
	
	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
	movsd	OLD_ALPHA_I, %xmm1
#else
	movq	OLD_LDC,   LDC
#ifdef TRMMKERNEL
	movsd	OLD_OFFSET, %xmm12
#endif

#endif

	EMMS

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

	movq	OLD_M, M
	movq	OLD_N, N

	subq	$-16 * SIZE, A

	movsd	 %xmm0, 0 + ALPHA
	movsd	 %xmm1, 8 + ALPHA
	
	salq	$ZBASE_SHIFT, LDC

#ifdef TRMMKERNEL
	movsd	%xmm12, OFFSET
	movsd	%xmm12, KK
#ifndef LEFT
	negq	KK
#endif	
#endif
	movq	N,  J
	sarq	$2, J		# j = (n >> 2)
	jle	.L40
	ALIGN_3

.L01:
/* Copying to Sub Buffer */
	leaq	16 * SIZE + BUFFER, BO
	movq	C, CO1			# coffset1 = c
	leaq	(C, LDC, 1), CO2	# coffset2 = c + ldc
	
#if defined(TRMMKERNEL) && defined(LEFT)
	movq	OFFSET, %rax
	movq	%rax, KK
#endif	

	movq	K, %rax
	sarq	$2, %rax
	jle	.L03
	ALIGN_3
	

#define RPREFETCHSIZE (8 *  7 + 4)
#define WPREFETCHSIZE (8 *  8 + 4)

.L02:
	PREFETCH	 (RPREFETCHSIZE +  0)  * SIZE(B)

	movq	 0 * SIZE(B), %mm0
	movq	 1 * SIZE(B), %mm1
	movq	%mm0,  -16 * SIZE(BO)
	movq	%mm0,  -15 * SIZE(BO)
	movq	%mm1,  -14 * SIZE(BO)
	movq	%mm1,  -13 * SIZE(BO)

	movq	 2 * SIZE(B), %mm2
	movq	 3 * SIZE(B), %mm3
	movq	%mm2,  -12 * SIZE(BO)
	movq	%mm2,  -11 * SIZE(BO)
	movq	%mm3,  -10 * SIZE(BO)
	movq	%mm3,   -9 * SIZE(BO)

	PREFETCHW	 (WPREFETCHSIZE +  0)  * SIZE(BO)

	movq	 4 * SIZE(B), %mm4
	movq	 5 * SIZE(B), %mm5
	movq	%mm4,   -8 * SIZE(BO)
	movq	%mm4,   -7 * SIZE(BO)
	movq	%mm5,   -6 * SIZE(BO)
	movq	%mm5,   -5 * SIZE(BO)

	PREFETCHW	 (WPREFETCHSIZE +  8)  * SIZE(BO)

	movq	 6 * SIZE(B), %mm6
	movq	 7 * SIZE(B), %mm7
	movq	%mm6,   -4 * SIZE(BO)
	movq	%mm6,   -3 * SIZE(BO)
	movq	%mm7,   -2 * SIZE(BO)
	movq	%mm7,   -1 * SIZE(BO)

	PREFETCH	 (RPREFETCHSIZE +  8)  * SIZE(B)

	movq	 8 * SIZE(B), %mm0
	movq	 9 * SIZE(B), %mm1

	movq	%mm0,   0 * SIZE(BO)
	movq	%mm0,   1 * SIZE(BO)
	movq	%mm1,   2 * SIZE(BO)
	movq	%mm1,   3 * SIZE(BO)

	movq	10 * SIZE(B), %mm2
	movq	11 * SIZE(B), %mm3
	movq	%mm2,   4 * SIZE(BO)
	movq	%mm2,   5 * SIZE(BO)
	movq	%mm3,   6 * SIZE(BO)
	movq	%mm3,   7 * SIZE(BO)

	PREFETCHW	 (WPREFETCHSIZE + 16)  * SIZE(BO)

	movq	12 * SIZE(B), %mm4
	movq	13 * SIZE(B), %mm5
	movq	%mm4,   8 * SIZE(BO)
	movq	%mm4,   9 * SIZE(BO)
	movq	%mm5,  10 * SIZE(BO)
	movq	%mm5,  11 * SIZE(BO)

	PREFETCHW	 (WPREFETCHSIZE + 24)  * SIZE(BO)

	movq	14 * SIZE(B), %mm6
	movq	15 * SIZE(B), %mm7
	movq	%mm6,  12 * SIZE(BO)
	movq	%mm6,  13 * SIZE(BO)
	movq	%mm7,  14 * SIZE(BO)
	movq	%mm7,  15 * SIZE(BO)

	addq	$ 32 * SIZE, BO
	subq	$-16 * SIZE, B

	subq	$1, %rax
	jne	.L02
	ALIGN_3

.L03:
	movq	K, %rax
	andq	$3, %rax
	BRANCH
	jle	.L10
	ALIGN_3

.L04:
	movq	 0 * SIZE(B), %mm0
	movq	 1 * SIZE(B), %mm1
	movq	 2 * SIZE(B), %mm2
	movq	 3 * SIZE(B), %mm3

	movq	%mm0, -16 * SIZE(BO)
	movq	%mm0, -15 * SIZE(BO)
	movq	%mm1, -14 * SIZE(BO)
	movq	%mm1, -13 * SIZE(BO)
	movq	%mm2, -12 * SIZE(BO)
	movq	%mm2, -11 * SIZE(BO)
	movq	%mm3, -10 * SIZE(BO)
	movq	%mm3,  -9 * SIZE(BO)

	addq	$4 * SIZE, B
	addq	$8 * SIZE, BO
	subq	$1, %rax
	jne	.L04
	ALIGN_3
	
.L10:
	movq	A, AO		# aoffset = a
	leaq	 (RPREFETCHSIZE +  0)  * SIZE(B), BB

	movq	M,  I
	sarq	$2, I	# i = (m >> 2)
	jle	.L20
	ALIGN_3

.L11:
	PREFETCH	  0  * SIZE(BB)
	PREFETCH	  8  * SIZE(BB)
	subq		 $-16 * SIZE, BB

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

	leaq	16 * SIZE + BUFFER, BO
#else
	leaq	16 * SIZE + BUFFER, BO
	movq	KK, %rax
	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 8), BO
#endif	

	movapd	-16 * SIZE(AO), %xmm0
	movapd	-16 * SIZE(BO), %xmm1
	pxor	%xmm8, %xmm8
	movapd	-14 * SIZE(AO), %xmm2
	movapd	-14 * SIZE(BO), %xmm3
	pxor	%xmm9, %xmm9
	movapd	-12 * SIZE(AO), %xmm4
	movapd	-12 * SIZE(BO), %xmm5
	pxor	%xmm10, %xmm10
	movapd	-10 * SIZE(AO), %xmm6
	movapd	 -8 * SIZE(BO), %xmm7
	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

#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
#else
	addq	$4, %rax
#endif
	movq	%rax, KKK
#endif
#ifndef GENERIC
	andq	$-8, %rax

	leaq	(, %rax, SIZE), %rax
	leaq	(AO, %rax, 4), AO
	leaq	(BO, %rax, 8), BO
	negq	%rax
	je,pn	.L15
	ALIGN_3

.L12:
	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)

	KERNEL1(16 *  1)
	KERNEL2(16 *  1)
	KERNEL3(16 *  1)
	KERNEL4(16 *  1)
	KERNEL5(16 *  1)
	KERNEL6(16 *  1)
	KERNEL7(16 *  1)
	KERNEL8(16 *  1)

	addq	$8 * SIZE, %rax
	je,pn	.L15

	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)

	KERNEL1(16 *  1)
	KERNEL2(16 *  1)
	KERNEL3(16 *  1)
	KERNEL4(16 *  1)
	KERNEL5(16 *  1)
	KERNEL6(16 *  1)
	KERNEL7(16 *  1)
	KERNEL8(16 *  1)

	addq	$8 * SIZE, %rax
	je,pn	.L15

	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)

	KERNEL1(16 *  1)
	KERNEL2(16 *  1)
	KERNEL3(16 *  1)
	KERNEL4(16 *  1)
	KERNEL5(16 *  1)
	KERNEL6(16 *  1)
	KERNEL7(16 *  1)
	KERNEL8(16 *  1)

	addq	$8 * SIZE, %rax
	je,pn	.L15

	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)

	KERNEL1(16 *  1)
	KERNEL2(16 *  1)
	KERNEL3(16 *  1)
	KERNEL4(16 *  1)
	KERNEL5(16 *  1)
	KERNEL6(16 *  1)
	KERNEL7(16 *  1)
	KERNEL8(16 *  1)

	addq	$8 * SIZE, %rax
	je,pn	.L15

	KERNEL1(16 *  0)
	KERNEL2(16 *  0)
	KERNEL3(16 *  0)
	KERNEL4(16 *  0)
	KERNEL5(16 *  0)
	KERNEL6(16 *  0)
	KERNEL7(16 *  0)
	KERNEL8(16 *  0)

	KERNEL1(16 *  1)
	KERNEL2(16 *  1)
	KERNEL3(16 *  1)
	KERNEL4(16 *  1)
	KERNEL5(16 *  1)
	KERNEL6(16 *  1)
	KERNEL7(16 *  1)
	KERNEL8(16 *  1)

	addq	$8 * SIZE, %rax
	je,pn	.L15

	KERNEL1(16 *  0)