zgemm_kernel_1x2_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 STACK	16
#define ARGS	 0
	
#define STACK_M	 4 + STACK + ARGS(%esi)
#define STACK_N	 8 + STACK + ARGS(%esi)
#define STACK_K	12 + STACK + ARGS(%esi)
#define STACK_ALPHA_R	16 + STACK + ARGS(%esi)
#define STACK_ALPHA_I	24 + STACK + ARGS(%esi)
#define STACK_A	32 + STACK + ARGS(%esi)
#define STACK_B	36 + STACK + ARGS(%esi)
#define STACK_C	40 + STACK + ARGS(%esi)
#define STACK_LDC	44 + STACK + ARGS(%esi)
#define STACK_OFFT	48 + STACK + ARGS(%esi)

#define POSINV	 0(%esp)
#define ALPHA_R	16(%esp)
#define ALPHA_I	32(%esp)
#define K	48(%esp)
#define N	52(%esp)
#define M	56(%esp)
#define A	60(%esp)
#define C	64(%esp)
#define J	68(%esp)
#define OLD_STACK 72(%esp)
#define OFFSET  76(%esp)
#define KK	80(%esp)
#define KKK	84(%esp)
#define BUFFER 128(%esp)

#define STACK_ALIGN	4096
#define STACK_OFFSET	1024

#define LOCAL_BUFFER_SIZE  GEMM_Q * GEMM_UNROLL_N * COMPSIZE * 16

#define PREFETCHSIZE (8 * 10 + 4)

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


#define KERNEL1(address) \
	mulpd	%xmm0, %xmm2; \
	addpd	%xmm2, %xmm4; \
	PREFETCH (PREFETCHSIZE +  0) * SIZE + (address) * 1 * SIZE(AA); \
	movapd	 2 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulpd	%xmm0, %xmm2; \
	addpd	%xmm2, %xmm5; \
	movapd	 4 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulpd	%xmm0, %xmm2; \
	mulpd	 6 * SIZE + (address) * 4 * SIZE(BB), %xmm0; \
	addpd	%xmm2, %xmm6; \
	movapd	16 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	addpd	%xmm0, %xmm7; \
	movapd	 2 * SIZE + (address) * 1 * SIZE(AA), %xmm0

#define KERNEL2(address) \
	mulpd	%xmm0, %xmm3; \
	addpd	%xmm3, %xmm4; \
	movapd	10 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulpd	%xmm0, %xmm3; \
	addpd	%xmm3, %xmm5; \
	movapd	12 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulpd	%xmm0, %xmm3; \
	mulpd	14 * SIZE + (address) * 4 * SIZE(BB), %xmm0; \
	addpd	%xmm3, %xmm6; \
	movapd	24 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	addpd	%xmm0, %xmm7; \
	movapd	 4 * SIZE + (address) * 1 * SIZE(AA), %xmm0

#define KERNEL3(address) \
	mulpd	%xmm0, %xmm2; \
	addpd	%xmm2, %xmm4; \
	movapd	18 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulpd	%xmm0, %xmm2; \
	addpd	%xmm2, %xmm5; \
	movapd	20 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulpd	%xmm0, %xmm2; \
	mulpd	22 * SIZE + (address) * 4 * SIZE(BB), %xmm0; \
	addpd	%xmm2, %xmm6; \
	movapd	32 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	addpd	%xmm0, %xmm7; \
	movapd	 6 * SIZE + (address) * 1 * SIZE(AA), %xmm0

#define KERNEL4(address) \
	mulpd	%xmm0, %xmm3; \
	addpd	%xmm3, %xmm4; \
	movapd	26 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulpd	%xmm0, %xmm3; \
	addpd	%xmm3, %xmm5; \
	movapd	28 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulpd	%xmm0, %xmm3; \
	mulpd	30 * SIZE + (address) * 4 * SIZE(BB), %xmm0; \
	addpd	%xmm3, %xmm6; \
	movapd	40 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	addpd	%xmm0, %xmm7; \
	movapd	16 * SIZE + (address) * 1 * SIZE(AA), %xmm0

#define KERNEL5(address) \
	PREFETCH (PREFETCHSIZE + 8) * SIZE + (address) * 1 * SIZE(AA); \
	mulpd	%xmm1, %xmm2; \
	addpd	%xmm2, %xmm4; \
	movapd	34 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulpd	%xmm1, %xmm2; \
	addpd	%xmm2, %xmm5; \
	movapd	36 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulpd	%xmm1, %xmm2; \
	mulpd	38 * SIZE + (address) * 4 * SIZE(BB), %xmm1; \
	addpd	%xmm2, %xmm6; \
	movapd	48 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	addpd	%xmm1, %xmm7; \
	movapd	10 * SIZE + (address) * 1 * SIZE(AA), %xmm1

#define KERNEL6(address) \
	mulpd	%xmm1, %xmm3; \
	addpd	%xmm3, %xmm4; \
	movapd	42 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulpd	%xmm1, %xmm3; \
	addpd	%xmm3, %xmm5; \
	movapd	44 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulpd	%xmm1, %xmm3; \
	mulpd	46 * SIZE + (address) * 4 * SIZE(BB), %xmm1; \
	addpd	%xmm3, %xmm6; \
	movapd	56 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	addpd	%xmm1, %xmm7; \
	movapd	12 * SIZE + (address) * 1 * SIZE(AA), %xmm1

#define KERNEL7(address) \
	mulpd	%xmm1, %xmm2; \
	addpd	%xmm2, %xmm4; \
	movapd	50 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulpd	%xmm1, %xmm2; \
	addpd	%xmm2, %xmm5; \
	movapd	52 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulpd	%xmm1, %xmm2; \
	mulpd	54 * SIZE + (address) * 4 * SIZE(BB), %xmm1; \
	addpd	%xmm2, %xmm6; \
	movapd	64 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	addpd	%xmm1, %xmm7; \
	movapd	14 * SIZE + (address) * 1 * SIZE(AA), %xmm1

#define KERNEL8(address) \
	mulpd	%xmm1, %xmm3; \
	addpd	%xmm3, %xmm4; \
	movapd	58 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulpd	%xmm1, %xmm3; \
	addpd	%xmm3, %xmm5; \
	movapd	60 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulpd	%xmm1, %xmm3; \
	mulpd	62 * SIZE + (address) * 4 * SIZE(BB), %xmm1; \
	addpd	%xmm3, %xmm6; \
	movapd	72 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	addpd	%xmm1, %xmm7; \
	movapd	24 * SIZE + (address) * 1 * SIZE(AA), %xmm1

	PROLOGUE

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

	PROFCODE

	EMMS

	movl	%esp, %esi	# save old stack

	subl	$128 + LOCAL_BUFFER_SIZE + STACK_OFFSET, %esp
	andl	$-STACK_ALIGN, %esp	# align stack
	addl	$STACK_OFFSET, %esp

#ifdef WINDOWS_ABI
#if LOCAL_BUFFER_SIZE > 12288
	movl	$0,  4096 * 3(%esp)
#endif
#if LOCAL_BUFFER_SIZE >  8192
	movl	$0,  4096 * 2(%esp)
#endif
#if LOCAL_BUFFER_SIZE >  4096
	movl	$0,  4096 * 1(%esp)
#endif
	movl	$0,  4096 * 0(%esp)
#endif

	movl	STACK_M, %ebx
	movl	STACK_N, %eax
	movl	STACK_K, %ecx
	movl	STACK_A, %edx

	movl	%ebx, M
	movl	%eax, N
	movl	%ecx, K
	movl	%edx, A
	movl	%esi, OLD_STACK

	movl	STACK_B, B
	movl	STACK_C, %ebx
#ifdef TRMMKERNEL
	movss	STACK_OFFT, %xmm4
#endif

	movlpd	STACK_ALPHA_R, %xmm0
	movlpd	STACK_ALPHA_I, %xmm1

	pcmpeqb	%xmm7, %xmm7
	psllq	$63, %xmm7	# Generate mask
	pxor	%xmm2, %xmm2

	movlpd	 %xmm0, 0 + ALPHA_R
	movlpd	 %xmm0, 8 + ALPHA_R
	
	movlpd	 %xmm1, 8 + ALPHA_I
	xorpd	 %xmm7, %xmm1
	movlpd	 %xmm1, 0 + ALPHA_I

	movlpd	  %xmm2,  0 + POSINV
	movlpd	  %xmm7,  8 + POSINV

	movl	%ebx, C
	movl	STACK_LDC, LDC

#ifdef TRMMKERNEL
	movss	%xmm4, OFFSET
	movss	%xmm4, KK
#ifndef LEFT
	negl	KK
#endif	
#endif

	sall	$ZBASE_SHIFT, LDC

	sarl	$1, %eax
	movl	%eax, J			# j = n
	jle	.L100
	ALIGN_4

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

	leal	BUFFER, %ecx

	movapd	POSINV, %xmm7

	movl	K, %eax
	sarl	$1, %eax
	jle	.L03
	ALIGN_4

.L02:
	prefetchnta	 56 * SIZE(B)

	movlpd	 0 * SIZE(B), %xmm0
	movlpd	 1 * SIZE(B), %xmm1
	movlpd	 2 * SIZE(B), %xmm2
	movlpd	 3 * SIZE(B), %xmm3
	movlpd	 4 * SIZE(B), %xmm4
	movlpd	 5 * SIZE(B), %xmm5
	movlpd	 6 * SIZE(B), %xmm6
	movlpd	 7 * SIZE(B), %xmm7

	movlpd	%xmm0,  0 * SIZE(BB)
	movlpd	%xmm0,  1 * SIZE(BB)
	movlpd	%xmm1,  2 * SIZE(BB)
	movlpd	%xmm1,  3 * SIZE(BB)
	movlpd	%xmm2,  4 * SIZE(BB)
	movlpd	%xmm2,  5 * SIZE(BB)
	movlpd	%xmm3,  6 * SIZE(BB)
	movlpd	%xmm3,  7 * SIZE(BB)
	movlpd	%xmm4,  8 * SIZE(BB)
	movlpd	%xmm4,  9 * SIZE(BB)
	movlpd	%xmm5, 10 * SIZE(BB)
	movlpd	%xmm5, 11 * SIZE(BB)
	movlpd	%xmm6, 12 * SIZE(BB)
	movlpd	%xmm6, 13 * SIZE(BB)
	movlpd	%xmm7, 14 * SIZE(BB)
	movlpd	%xmm7, 15 * SIZE(BB)

	addl	$  8 * SIZE, B
	subl	$-16 * SIZE, BB

	decl	%eax
	jne	.L02
	ALIGN_4

.L03:
	movl	K, %eax
	andl	$1, %eax
	BRANCH
	jle	.L05

	movlpd	 0 * SIZE(B), %xmm0
	movlpd	 1 * SIZE(B), %xmm1
	movlpd	 2 * SIZE(B), %xmm2
	movlpd	 3 * SIZE(B), %xmm3

	movlpd	%xmm0,  0 * SIZE(BB)
	movlpd	%xmm0,  1 * SIZE(BB)
	movlpd	%xmm1,  2 * SIZE(BB)
	movlpd	%xmm1,  3 * SIZE(BB)
	movlpd	%xmm2,  4 * SIZE(BB)
	movlpd	%xmm2,  5 * SIZE(BB)
	movlpd	%xmm3,  6 * SIZE(BB)
	movlpd	%xmm3,  7 * SIZE(BB)

	addl	$4 * SIZE, B
	ALIGN_4

.L05:
	movl	C, %esi		# coffset = c
	movl	A, AA		# aoffset = a
	movl	M,  %ebx
	testl	%ebx, %ebx
	jle	.L100
	ALIGN_4

.L10:
#if !defined(TRMMKERNEL) || \
	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	leal	BUFFER, BB	# boffset1 = boffset
#else
	leal	BUFFER, BB	# boffset1 = boffset
	movl	KK, %eax
	leal	(, %eax, SIZE), %eax
	leal	(AA, %eax, 2), AA
	leal	(BB, %eax, 8), BB
#endif	

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

	prefetchw 2 * SIZE(%esi)
	prefetchw 2 * SIZE(%esi, LDC)

#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	$1, %eax
#else
	addl	$2, %eax
#endif
	movl	%eax, KKK
#endif

#if 1
	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