trsm_kernel_lt_2x4_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 OLD_M	 4 + STACK + ARGS(%esi)
#define OLD_N	 8 + STACK + ARGS(%esi)
#define OLD_K	12 + STACK + ARGS(%esi)
#define OLD_ALPHA	16 + STACK + ARGS(%esi)
#define OLD_A	24 + STACK + ARGS(%esi)
#define OLD_B	28 + STACK + ARGS(%esi)
#define OLD_C	32 + STACK + ARGS(%esi)
#define OLD_LDC	36 + STACK + ARGS(%esi)
#define OLD_OFFT	40 + STACK + ARGS(%esi)

#define K	16(%esp)
#define N	20(%esp)
#define M	24(%esp)
#define A	28(%esp)
#define C	32(%esp)
#define J	36(%esp)
#define OLD_STACK 40(%esp)
#define OFFSET  44(%esp)
#define KK	48(%esp)
#define KKK	52(%esp)
#define AORIG	56(%esp)
#define BORIG	60(%esp)
#define BUFFER 128(%esp)

#define STACK_ALIGN	4096
#define STACK_OFFSET	1024

#define LOCAL_BUFFER_SIZE  GEMM_Q * GEMM_UNROLL_N * COMPSIZE * 16

#if defined(OPTERON) || defined(BARCELONA)
#define PREFETCH     prefetch
#define PREFETCHSIZE (8 * 10 + 4)
#endif

#define B	%edi
#define AA	%edx
#define BB	%ecx
#define LDC	%ebp
#define CO1	%esi

#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
	addl	$STACK_OFFSET, %esp

	movl	OLD_M, %ebx
	movl	OLD_N, %eax
	movl	OLD_K, %ecx
	movl	OLD_A, %edx

	movl	%ebx, M
	movl	%eax, N
	movl	%ecx, K
	movl	%edx, A
	movl	%esi, OLD_STACK
	movd	OLD_OFFT, %mm4

	movl	OLD_B, B
	movl	OLD_C, %ebx

	movl	%ebx, C
	movl	OLD_LDC, LDC

	movd	%mm4, OFFSET
	movd	%mm4, KK

	leal	(, LDC, SIZE), LDC

#ifdef LN
       movl	M, %eax
       leal	(, %eax, SIZE), %eax
       addl	%eax, C
       imull	K, %eax
       addl	%eax, A
#endif

#ifdef RT
       movl	N, %eax
       leal	(, %eax, SIZE), %eax
       imull	K, %eax
       addl	%eax, B
       movl	N, %eax
       imull	LDC, %eax
       addl	%eax, C
#endif

#ifdef RN
	negl	KK
#endif	

#ifdef RT
       movl	N, %eax
       subl	OFFSET, %eax
       movl	%eax, KK
#endif

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

.L01:
#ifdef LN
	movl	OFFSET, %eax
	addl	M, %eax
	movl	%eax, KK
#endif	

	leal	BUFFER, BB

#ifdef RT
       movl	K, %eax
       sall	$2 + BASE_SHIFT, %eax
       subl	%eax, B
#endif

#if defined(LN) || defined(RT)
	movl	KK, %eax
	movl	B, BORIG
	leal	(, %eax, SIZE), %eax
	leal	(B,  %eax, 4), B
	leal	(BB, %eax, 8), BB
#endif	

#ifdef LT
	movl	OFFSET, %eax
	movl	%eax, KK
#endif

#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	sarl	$1, %eax
	jle	.L05
	ALIGN_4
	
.L02:
#define COPYPREFETCH 40

	prefetchnta	(COPYPREFETCH) * SIZE(B)

	movq	 0 * SIZE(B), %mm0
	movq	 1 * SIZE(B), %mm1
	movq	 2 * SIZE(B), %mm2
	movq	 3 * SIZE(B), %mm3
	movq	 4 * SIZE(B), %mm4
	movq	 5 * SIZE(B), %mm5
	movq	 6 * SIZE(B), %mm6
	movq	 7 * SIZE(B), %mm7

	movq	%mm0,  0 * SIZE(BB)
	movq	%mm0,  1 * SIZE(BB)
	movq	%mm1,  2 * SIZE(BB)
	movq	%mm1,  3 * SIZE(BB)
	movq	%mm2,  4 * SIZE(BB)
	movq	%mm2,  5 * SIZE(BB)
	movq	%mm3,  6 * SIZE(BB)
	movq	%mm3,  7 * SIZE(BB)

	movq	%mm4,  8 * SIZE(BB)
	movq	%mm4,  9 * SIZE(BB)
	movq	%mm5, 10 * SIZE(BB)
	movq	%mm5, 11 * SIZE(BB)
	movq	%mm6, 12 * SIZE(BB)
	movq	%mm6, 13 * SIZE(BB)
	movq	%mm7, 14 * SIZE(BB)
	movq	%mm7, 15 * SIZE(BB)

	addl	$ 8 * SIZE, B
	addl	$16 * SIZE, BB
	decl	%eax
	jne	.L02
	ALIGN_2

.L05:
#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	andl	$1, %eax
	BRANCH
	jle	.L10

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

	movq	%mm0,  0 * SIZE(BB)
	movq	%mm0,  1 * SIZE(BB)
	movq	%mm1,  2 * SIZE(BB)
	movq	%mm1,  3 * SIZE(BB)
	movq	%mm2,  4 * SIZE(BB)
	movq	%mm2,  5 * SIZE(BB)
	movq	%mm3,  6 * SIZE(BB)
	movq	%mm3,  7 * SIZE(BB)

	addl	$4 * SIZE, B
	ALIGN_4
	
.L10:
#if defined(LT) || defined(RN)
	movl	A, AA
#else
	movl	A, %eax
	movl	%eax, AORIG
#endif

	leal	(, LDC, 4), %eax

#ifdef RT
	subl	%eax, C
#endif
	movl	C, CO1
#ifndef RT
	addl	%eax, C
#endif

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

.L11:
#ifdef LN
       movl	K, %eax
       sall	$1 + BASE_SHIFT, %eax
       subl	%eax, AORIG
#endif

#if defined(LN) || defined(RT)
	movl	KK, %eax
	movl	AORIG, AA
	leal	(, %eax, SIZE), %eax
	leal	(AA, %eax, 2), AA
#endif

	leal	BUFFER, BB

#if defined(LN) || defined(RT)
	movl	KK, %eax
	sall	$3 + BASE_SHIFT, %eax
	addl	%eax, BB
#endif	

	pxor	%xmm4, %xmm4
	pxor	%xmm5, %xmm5
	pxor	%xmm6, %xmm6
	pxor	%xmm7, %xmm7

	movapd	 0 * SIZE(AA), %xmm0
	movapd	 8 * SIZE(AA), %xmm1
	movapd	 0 * SIZE(BB), %xmm2
	movapd	 8 * SIZE(BB), %xmm3

	leal	(LDC, LDC, 2), %eax

#ifdef LN
	prefetchw	-2 * SIZE(CO1)
	prefetchw	-2 * SIZE(CO1, LDC)
	prefetchw	-2 * SIZE(CO1, LDC, 2)
	prefetchw	-2 * SIZE(CO1, %eax)
#else
	prefetchw	 1 * SIZE(CO1)
	prefetchw	 1 * SIZE(CO1, LDC)
	prefetchw	 1 * SIZE(CO1, LDC, 2)
	prefetchw	 1 * SIZE(CO1, %eax)
#endif

#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#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
	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)

	addl	$128 * 4  * SIZE, BB
	addl	$128 * 1  * SIZE, AA
	subl	$128 * 8, %eax
	jg	.L1X
	jmp	.L15

.L12:
	leal	(AA, %eax, 1), AA
	leal	(BB, %eax, 4), BB
	ALIGN_4
#else

	sarl	$3, %eax
	je	.L15
	ALIGN_4

.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)

	addl   $64 * SIZE, BB
	addl   $16 * SIZE, AA
	decl   %eax
	jne    .L12
	ALIGN_4
#endif

.L15:
#if defined(LT) || defined(RN)
	movl	KK, %eax
#else
	movl	K, %eax
	subl	KK, %eax
#endif
	andl	$7, %eax		# if (k & 1)
	BRANCH
	je .L18
	ALIGN_3

.L16:
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm4
	movapd	 2 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	addpd	 %xmm2, %xmm5
	movapd	 4 * SIZE(BB), %xmm2
	mulpd	 %xmm0, %xmm2
	mulpd	 6 * SIZE(BB), %xmm0
	addpd	 %xmm2, %xmm6
	movapd	 8 * SIZE(BB), %xmm2
	addpd	 %xmm0, %xmm7
	movapd	 2 * SIZE(AA), %xmm0

	addl	$2 * SIZE, AA
	addl	$8 * SIZE, BB
	decl	%eax
	jg	.L16
	ALIGN_4

.L18:
#if defined(LN) || defined(RT)
	movl	KK, %eax
#ifdef LN
	subl	$2, %eax
#else
	subl	$4, %eax
#endif

	movl	AORIG, AA
	movl	BORIG, B
	leal	BUFFER, BB

	leal	(, %eax, SIZE), %eax
	leal	(AA, %eax, 2), AA
	leal	(B,  %eax, 4), B
	leal	(BB, %eax, 8), BB
#endif

#if defined(LN) || defined(LT)
	movapd	 %xmm4, %xmm0
	unpcklpd %xmm5, %xmm4
	unpckhpd %xmm5, %xmm0

	movapd	 %xmm6, %xmm1
	unpcklpd %xmm7, %xmm6
	unpckhpd %xmm7, %xmm1

	movapd	 0 * SIZE(B), %xmm2
	movapd	 2 * SIZE(B), %xmm5
	movapd	 4 * SIZE(B), %xmm3
	movapd	 6 * SIZE(B), %xmm7

	subpd	%xmm4,  %xmm2
	subpd	%xmm6,  %xmm5
	subpd	%xmm0,  %xmm3
	subpd	%xmm1,  %xmm7
#else
	movapd	 0 * SIZE(AA), %xmm0
	movapd	 2 * SIZE(AA), %xmm1
	movapd	 4 * SIZE(AA), %xmm2
	movapd	 6 * SIZE(AA), %xmm3

	subpd	%xmm4, %xmm0
	subpd	%xmm5, %xmm1
	subpd	%xmm6, %xmm2
	subpd	%xmm7, %xmm3
#endif

#ifdef LN
	movlpd	 3 * SIZE(AA), %xmm4
	movhpd	 3 * SIZE(AA), %xmm4
	mulpd	 %xmm4, %xmm3
	mulpd	 %xmm4, %xmm7

	movlpd	 2 * SIZE(AA), %xmm4
	movhpd	 2 * SIZE(AA), %xmm4
	movapd	 %xmm4, %xmm6