gemm_kernel_2x4_sse3.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 STACK	16
#define ARGS	16
	
#define M	 4 + STACK + ARGS(%esp)
#define N	 8 + STACK + ARGS(%esp)
#define K	12 + STACK + ARGS(%esp)
#define ALPHA	16 + STACK + ARGS(%esp)
#define A	24 + STACK + ARGS(%esp)
#define ARG_B	28 + STACK + ARGS(%esp)
#define C	32 + STACK + ARGS(%esp)
#define ARG_LDC	36 + STACK + ARGS(%esp)
#define OFFSET	40 + STACK + ARGS(%esp)

#define J	 0 + STACK(%esp)
#define BX	 4 + STACK(%esp)
#define KK	 8 + STACK(%esp)
#define KKK	12 + STACK(%esp)

#ifdef PENTIUM4
#define PREFETCH_R   (8 * 4)
#define PREFETCH     prefetcht1
#define PREFETCHSIZE 84
#endif

#ifdef PENTIUMM
#define PREFETCH_R   (8 * 4)
#define PREFETCH     prefetcht1
#define PREFETCHSIZE 84
#endif

#define AA	%edx
#define BB	%ecx
#define LDC	%ebp
#define B	%edi

#define KERNEL1(address) \
	mulpd	 %xmm0, %xmm2; \
	PREFETCH  (PREFETCHSIZE + 0) * SIZE + (address) * 1 * SIZE(AA); \
	addpd	 %xmm2, %xmm4; \
	movddup	 1 * SIZE + (address) * 2 * SIZE(BB), %xmm2; \
	mulpd	 %xmm0, %xmm2; \
	addpd	 %xmm2, %xmm5; \
	movddup	 2 * SIZE + (address) * 2 * SIZE(BB), %xmm2; \
	mulpd	 %xmm0, %xmm2; \
	addpd	 %xmm2, %xmm6; \
	movddup	 3 * SIZE + (address) * 2 * SIZE(BB), %xmm2; \
	mulpd	 %xmm0, %xmm2; \
	movapd	 2 * SIZE + (address) * 1 * SIZE(AA), %xmm0; \
	addpd	 %xmm2, %xmm7; \
	movddup	 4 * SIZE + (address) * 2 * SIZE(BB), %xmm2

#define KERNEL2(address) \
	mulpd	 %xmm0, %xmm2; \
	addpd	 %xmm2, %xmm4; \
	movddup	 5 * SIZE + (address) * 2 * SIZE(BB), %xmm2; \
	mulpd	 %xmm0, %xmm2; \
	addpd	 %xmm2, %xmm5; \
	movddup	 6 * SIZE + (address) * 2 * SIZE(BB), %xmm2; \
	mulpd	 %xmm0, %xmm2; \
	addpd	 %xmm2, %xmm6; \
	movddup	 7 * SIZE + (address) * 2 * SIZE(BB), %xmm2; \
	mulpd	 %xmm0, %xmm2; \
	movapd	 4 * SIZE + (address) * 1 * SIZE(AA), %xmm0; \
	addpd	 %xmm2, %xmm7; \
	movddup	16 * SIZE + (address) * 2 * SIZE(BB), %xmm2

#define KERNEL3(address) \
	mulpd	 %xmm0, %xmm3; \
	addpd	 %xmm3, %xmm4; \
	movddup	 9 * SIZE + (address) * 2 * SIZE(BB), %xmm3; \
	mulpd	 %xmm0, %xmm3; \
	addpd	 %xmm3, %xmm5; \
	movddup	10 * SIZE + (address) * 2 * SIZE(BB), %xmm3; \
	mulpd	 %xmm0, %xmm3; \
	addpd	 %xmm3, %xmm6; \
	movddup	11 * SIZE + (address) * 2 * SIZE(BB), %xmm3; \
	mulpd	 %xmm0, %xmm3; \
	movapd	 6 * SIZE + (address) * 1 * SIZE(AA), %xmm0; \
	addpd	 %xmm3, %xmm7; \
	movddup	12 * SIZE + (address) * 2 * SIZE(BB), %xmm3

#define KERNEL4(address) \
	mulpd	 %xmm0, %xmm3; \
	addpd	 %xmm3, %xmm4; \
	movddup	13 * SIZE + (address) * 2 * SIZE(BB), %xmm3; \
	mulpd	 %xmm0, %xmm3; \
	addpd	 %xmm3, %xmm5; \
	movddup	14 * SIZE + (address) * 2 * SIZE(BB), %xmm3; \
	mulpd	 %xmm0, %xmm3; \
	addpd	 %xmm3, %xmm6; \
	movddup	15 * SIZE + (address) * 2 * SIZE(BB), %xmm3; \
	mulpd	 %xmm0, %xmm3; \
	movapd	16 * SIZE + (address) * 1 * SIZE(AA), %xmm0; \
	addpd	 %xmm3, %xmm7; \
	movddup	24 * SIZE + (address) * 2 * SIZE(BB), %xmm3

#define KERNEL5(address) \
	mulpd	 %xmm1, %xmm2; \
	addpd	 %xmm2, %xmm4; \
	movddup	17 * SIZE + (address) * 2 * SIZE(BB), %xmm2; \
	mulpd	 %xmm1, %xmm2; \
	addpd	 %xmm2, %xmm5; \
	movddup	18 * SIZE + (address) * 2 * SIZE(BB), %xmm2; \
	mulpd	 %xmm1, %xmm2; \
	addpd	 %xmm2, %xmm6; \
	movddup	19 * SIZE + (address) * 2 * SIZE(BB), %xmm2; \
	mulpd	 %xmm1, %xmm2; \
	movapd	10 * SIZE + (address) * 1 * SIZE(AA), %xmm1; \
	addpd	 %xmm2, %xmm7

#define KERNEL6(address) \
	movddup	20 * SIZE + (address) * 2 * SIZE(BB), %xmm2; \
	mulpd	 %xmm1, %xmm2; \
	addpd	 %xmm2, %xmm4; \
	movddup	21 * SIZE + (address) * 2 * SIZE(BB), %xmm2; \
	mulpd	 %xmm1, %xmm2; \
	addpd	 %xmm2, %xmm5; \
	movddup	22 * SIZE + (address) * 2 * SIZE(BB), %xmm2; \
	mulpd	 %xmm1, %xmm2; \
	addpd	 %xmm2, %xmm6; \
	movddup	23 * SIZE + (address) * 2 * SIZE(BB), %xmm2; \
	mulpd	 %xmm1, %xmm2; \
	movapd	12 * SIZE + (address) * 1 * SIZE(AA), %xmm1; \
	addpd	 %xmm2, %xmm7; \
	movddup	32 * SIZE + (address) * 2 * SIZE(BB), %xmm2

#define KERNEL7(address) \
	mulpd	 %xmm1, %xmm3; \
	addpd	 %xmm3, %xmm4; \
	movddup	25 * SIZE + (address) * 2 * SIZE(BB), %xmm3; \
	mulpd	 %xmm1, %xmm3; \
	addpd	 %xmm3, %xmm5; \
	movddup	26 * SIZE + (address) * 2 * SIZE(BB), %xmm3; \
	mulpd	 %xmm1, %xmm3; \
	addpd	 %xmm3, %xmm6; \
	movddup	27 * SIZE + (address) * 2 * SIZE(BB), %xmm3; \
	mulpd	 %xmm1, %xmm3; \
	movapd	14 * SIZE + (address) * 1 * SIZE(AA), %xmm1; \
	addpd	 %xmm3, %xmm7; \
	movddup	28 * SIZE + (address) * 2 * SIZE(BB), %xmm3

#define KERNEL8(address) \
	mulpd	 %xmm1, %xmm3; \
	addpd	 %xmm3, %xmm4; \
	movddup	29 * SIZE + (address) * 2 * SIZE(BB), %xmm3; \
	mulpd	 %xmm1, %xmm3; \
	addpd	 %xmm3, %xmm5; \
	movddup	30 * SIZE + (address) * 2 * SIZE(BB), %xmm3; \
	mulpd	 %xmm1, %xmm3; \
	addpd	 %xmm3, %xmm6; \
	movddup	31 * SIZE + (address) * 2 * SIZE(BB), %xmm3; \
	mulpd	 %xmm1, %xmm3; \
	movapd	24 * SIZE + (address) * 1 * SIZE(AA), %xmm1; \
	addpd	 %xmm3, %xmm7; \
	movddup	40 * SIZE + (address) * 2 * SIZE(BB), %xmm3

	PROLOGUE

	subl	$ARGS, %esp

	pushl	%ebp
	pushl	%edi
	pushl	%esi
	pushl	%ebx

	PROFCODE

	movl	ARG_B,   B
	movl	ARG_LDC, LDC

#ifdef TRMMKERNEL
	movl	OFFSET, %eax
#ifndef LEFT
	negl	%eax
#endif	
	movl	%eax, KK
#endif

	leal	(, LDC, SIZE), LDC

	movl	N,  %eax
	sarl	$2, %eax
	movl	%eax, J
	jle	.L30
	ALIGN_2

.L10:
#if defined(TRMMKERNEL) && defined(LEFT)
	movl	OFFSET, %eax
	movl	%eax, KK
#endif	

	movl	K, %eax
	sall	$BASE_SHIFT + 2, %eax
	leal	(B, %eax), %eax
	movl	%eax, BX

	movl	C, %esi		# coffset = c
	movl	A, AA		# aoffset = a

	movl	M,  %ebx
	sarl	$1, %ebx	# i = (m >> 2)
	jle	.L20
	ALIGN_4

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

	movl	B, BB
#else
	movl	KK, %eax
	leal	(, %eax, SIZE), %eax
	leal	(AA, %eax, 2), AA
	leal	(B,  %eax, 4), BB
#endif	

	movl	BX, %eax

	prefetcht2  0 * SIZE(%eax)

	subl	$-8 * SIZE, BX

	movapd	 0 * SIZE(AA), %xmm0
	pxor	%xmm4, %xmm4
	movapd	 8 * SIZE(AA), %xmm1
	pxor	%xmm5, %xmm5
	movddup	 0 * SIZE(BB), %xmm2
	pxor	%xmm6, %xmm6
	movddup	 8 * SIZE(BB), %xmm3
	pxor	%xmm7, %xmm7

	leal	(LDC, LDC, 2), %eax

#ifdef PENTIUM4
	prefetchnta	3 * SIZE(%esi)
	prefetchnta	3 * SIZE(%esi, LDC, 1)
	prefetchnta	3 * SIZE(%esi, LDC, 2)
	prefetchnta	3 * SIZE(%esi, %eax, 1)
#endif

#ifndef TRMMKERNEL
	movl	K, %eax
#elif (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	movl	K, %eax
	subl	KK, %eax
	movl	%eax, KKK	
#else
	movl	KK, %eax
#ifdef LEFT
	addl	$2, %eax
#else
	addl	$4, %eax
#endif
	movl	%eax, KKK
#endif

#ifdef CORE_PRESCOTT
	andl	$-8, %eax
	sall	$4, %eax
	je	.L15
	
.L1X:
	KERNEL1(16  *  0)
	KERNEL2(16  *  0)
	KERNEL3(16  *  0)
	KERNEL4(16  *  0)
	KERNEL5(16  *  0)
	KERNEL6(16  *  0)
	KERNEL7(16  *  0)
	KERNEL8(16  *  0)
	cmpl	$128 *  1, %eax
	jle	.L12
	KERNEL1(16  *  1)
	KERNEL2(16  *  1)
	KERNEL3(16  *  1)
	KERNEL4(16  *  1)
	KERNEL5(16  *  1)
	KERNEL6(16  *  1)
	KERNEL7(16  *  1)
	KERNEL8(16  *  1)
	cmpl	$128 *  2, %eax
	jle	.L12
	KERNEL1(16  *  2)
	KERNEL2(16  *  2)
	KERNEL3(16  *  2)
	KERNEL4(16  *  2)
	KERNEL5(16  *  2)
	KERNEL6(16  *  2)
	KERNEL7(16  *  2)
	KERNEL8(16  *  2)
	cmpl	$128 *  3, %eax
	jle	.L12
	KERNEL1(16  *  3)
	KERNEL2(16  *  3)
	KERNEL3(16  *  3)
	KERNEL4(16  *  3)
	KERNEL5(16  *  3)
	KERNEL6(16  *  3)
	KERNEL7(16  *  3)
	KERNEL8(16  *  3)
	cmpl	$128 *  4, %eax
	jle	.L12
	KERNEL1(16  *  4)
	KERNEL2(16  *  4)
	KERNEL3(16  *  4)
	KERNEL4(16  *  4)
	KERNEL5(16  *  4)
	KERNEL6(16  *  4)
	KERNEL7(16  *  4)
	KERNEL8(16  *  4)
	cmpl	$128 *  5, %eax
	jle	.L12
	KERNEL1(16  *  5)
	KERNEL2(16  *  5)
	KERNEL3(16  *  5)
	KERNEL4(16  *  5)
	KERNEL5(16  *  5)
	KERNEL6(16  *  5)
	KERNEL7(16  *  5)
	KERNEL8(16  *  5)
	cmpl	$128 *  6, %eax
	jle	.L12
	KERNEL1(16  *  6)
	KERNEL2(16  *  6)
	KERNEL3(16  *  6)
	KERNEL4(16  *  6)
	KERNEL5(16  *  6)
	KERNEL6(16  *  6)
	KERNEL7(16  *  6)
	KERNEL8(16  *  6)
	cmpl	$128 *  7, %eax
	jle	.L12
	KERNEL1(16  *  7)
	KERNEL2(16  *  7)
	KERNEL3(16  *  7)
	KERNEL4(16  *  7)
	KERNEL5(16  *  7)
	KERNEL6(16  *  7)
	KERNEL7(16  *  7)
	KERNEL8(16  *  7)
#if 1
	cmpl	$128 *  8, %eax
	jle	.L12
	KERNEL1(16  *  8)
	KERNEL2(16  *  8)
	KERNEL3(16  *  8)
	KERNEL4(16  *  8)
	KERNEL5(16  *  8)
	KERNEL6(16  *  8)
	KERNEL7(16  *  8)
	KERNEL8(16  *  8)
	cmpl	$128 *  9, %eax
	jle	.L12
	KERNEL1(16  *  9)
	KERNEL2(16  *  9)
	KERNEL3(16  *  9)
	KERNEL4(16  *  9)
	KERNEL5(16  *  9)
	KERNEL6(16  *  9)
	KERNEL7(16  *  9)
	KERNEL8(16  *  9)
	cmpl	$128 * 10, %eax
	jle	.L12
	KERNEL1(16  * 10)
	KERNEL2(16  * 10)
	KERNEL3(16  * 10)
	KERNEL4(16  * 10)
	KERNEL5(16  * 10)
	KERNEL6(16  * 10)
	KERNEL7(16  * 10)
	KERNEL8(16  * 10)
	cmpl	$128 * 11, %eax
	jle	.L12
	KERNEL1(16  * 11)
	KERNEL2(16  * 11)
	KERNEL3(16  * 11)
	KERNEL4(16  * 11)
	KERNEL5(16  * 11)
	KERNEL6(16  * 11)
	KERNEL7(16  * 11)
	KERNEL8(16  * 11)
	cmpl	$128 * 12, %eax
	jle	.L12
	KERNEL1(16  * 12)
	KERNEL2(16  * 12)
	KERNEL3(16  * 12)
	KERNEL4(16  * 12)
	KERNEL5(16  * 12)
	KERNEL6(16  * 12)
	KERNEL7(16  * 12)
	KERNEL8(16  * 12)
	cmpl	$128 * 13, %eax
	jle	.L12
	KERNEL1(16  * 13)
	KERNEL2(16  * 13)
	KERNEL3(16  * 13)
	KERNEL4(16  * 13)
	KERNEL5(16  * 13)
	KERNEL6(16  * 13)
	KERNEL7(16  * 13)
	KERNEL8(16  * 13)
	cmpl	$128 * 14, %eax
	jle	.L12
	KERNEL1(16  * 14)
	KERNEL2(16  * 14)
	KERNEL3(16  * 14)
	KERNEL4(16  * 14)
	KERNEL5(16  * 14)
	KERNEL6(16  * 14)
	KERNEL7(16  * 14)
	KERNEL8(16  * 14)
	cmpl	$128 * 15, %eax
	jle	.L12
	KERNEL1(16  * 15)
	KERNEL2(16  * 15)
	KERNEL3(16  * 15)
	KERNEL4(16  * 15)
	KERNEL5(16  * 15)
	KERNEL6(16  * 15)
	KERNEL7(16  * 15)
	KERNEL8(16  * 15)
#else
	addl	$32 * 4  * SIZE, AA
	addl	$32 * 8  * SIZE, BB
	subl	$128 * 8, %eax
	jg	.L1X
#endif

.L12:
	leal	(AA, %eax, 1), AA	# * 16
	leal	(BB, %eax, 2), BB	# * 64

#else

	sarl	$3, %eax
	je	.L15
	ALIGN_4

.L12:
	mulpd	 %xmm0, %xmm2
	PREFETCH  (PREFETCHSIZE + 0) * SIZE(AA)
	addpd	 %xmm2, %xmm4
	movddup	 1 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm5
	movddup	 2 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm6
	movddup	 3 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	movapd	 2 * SIZE(AA), %xmm0
	addpd	 %xmm2, %xmm7
	movddup	 4 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm4
	movddup	 5 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm5
	movddup	 6 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm6
	movddup	 7 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	movapd	 4 * SIZE(AA), %xmm0
	addpd	 %xmm2, %xmm7
	movddup	16 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm3
	addpd	 %xmm3, %xmm4
	movddup	 9 * SIZE(BB), %xmm3
	mulpd	 %xmm0, %xmm3
	addpd	 %xmm3, %xmm5
	movddup	10 * SIZE(BB), %xmm3
	mulpd	 %xmm0, %xmm3
	addpd	 %xmm3, %xmm6
	movddup	11 * SIZE(BB), %xmm3
	mulpd	 %xmm0, %xmm3
	movapd	 6 * SIZE(AA), %xmm0
	addpd	 %xmm3, %xmm7
	movddup	12 * SIZE(BB), %xmm3
	mulpd	 %xmm0, %xmm3
	addpd	 %xmm3, %xmm4
	movddup	13 * SIZE(BB), %xmm3
	mulpd	 %xmm0, %xmm3
	addpd	 %xmm3, %xmm5
	movddup	14 * SIZE(BB), %xmm3
	mulpd	 %xmm0, %xmm3
	addpd	 %xmm3, %xmm6
	movddup	15 * SIZE(BB), %xmm3
	mulpd	 %xmm0, %xmm3
	movapd	16 * SIZE(AA), %xmm0
	addpd	 %xmm3, %xmm7
	movddup	24 * SIZE(BB), %xmm3
	mulpd	 %xmm1, %xmm2
	addpd	 %xmm2, %xmm4
	movddup	17 * SIZE(BB), %xmm2
	mulpd	 %xmm1, %xmm2
	addpd	 %xmm2, %xmm5
	movddup	18 * SIZE(BB), %xmm2
	mulpd	 %xmm1, %xmm2
	addpd	 %xmm2, %xmm6
	movddup	19 * SIZE(BB), %xmm2