zgemm_kernel_2x2_sse.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	20 + STACK + ARGS(%esi)
#define STACK_A	24 + STACK + ARGS(%esi)
#define STACK_B	28 + STACK + ARGS(%esi)
#define STACK_C	32 + STACK + ARGS(%esi)
#define STACK_LDC	36 + STACK + ARGS(%esi)
#define STACK_OFFT	40 + 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 B	%edi
#define	LDC	%ebp
#define AA	%edx
#define BB	%ecx

#define LOCAL_BUFFER_SIZE  GEMM_Q * GEMM_UNROLL_N * COMPSIZE * 16

#define STACK_ALIGN	4096
#define STACK_OFFSET	1024

#ifdef ATHLON
#define PREFETCHSIZE  64
#define WPREFETCHSIZE 80
#endif

#if defined(OPTERON) || defined(BARCELONA)
#define PREFETCHSIZE (16 * 10 + 8)
#define WPREFETCHSIZE 112
#define movsd  movlps
#endif

#ifdef PENTIUM4
#define PREFETCH	prefetcht1
#define PREFETCHSIZE 168
#endif

#ifdef HAVE_SSE2
#define XORPS	pxor
#else
#define XORPS	xorps
#define movsd	movlps
#endif

#if defined(ATHLON) || defined(OPTERON) || defined(BARCELONA)
#define KERNEL1(address) \
	mulps	%xmm0, %xmm2; \
	addps	%xmm2, %xmm4; \
	PREFETCH (PREFETCHSIZE +  0) * SIZE + (address) * 1 * SIZE(AA); \
	movaps	 4 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulps	%xmm0, %xmm2; \
	addps	%xmm2, %xmm5; \
	movaps	 8 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulps	%xmm0, %xmm2; \
	mulps	12 * SIZE + (address) * 4 * SIZE(BB), %xmm0; \
	addps	%xmm2, %xmm6; \
	movaps	32 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	addps	%xmm0, %xmm7; \
	movaps	 4 * SIZE + (address) * 1 * SIZE(AA), %xmm0

#define KERNEL2(address) \
	mulps	%xmm0, %xmm3; \
	addps	%xmm3, %xmm4; \
	movaps	20 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulps	%xmm0, %xmm3; \
	addps	%xmm3, %xmm5; \
	movaps	24 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulps	%xmm0, %xmm3; \
	mulps	28 * SIZE + (address) * 4 * SIZE(BB), %xmm0; \
	addps	%xmm3, %xmm6; \
	movaps	48 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	addps	%xmm0, %xmm7; \
	movaps	 8 * SIZE + (address) * 1 * SIZE(AA), %xmm0

#define KERNEL3(address) \
	mulps	%xmm0, %xmm2; \
	addps	%xmm2, %xmm4; \
	movaps	36 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulps	%xmm0, %xmm2; \
	addps	%xmm2, %xmm5; \
	movaps	40 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulps	%xmm0, %xmm2; \
	mulps	44 * SIZE + (address) * 4 * SIZE(BB), %xmm0; \
	addps	%xmm2, %xmm6; \
	movaps	64 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	addps	%xmm0, %xmm7; \
	movaps	12 * SIZE + (address) * 1 * SIZE(AA), %xmm0

#define KERNEL4(address) \
	mulps	%xmm0, %xmm3; \
	addps	%xmm3, %xmm4; \
	movaps	52 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulps	%xmm0, %xmm3; \
	addps	%xmm3, %xmm5; \
	movaps	56 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulps	%xmm0, %xmm3; \
	mulps	60 * SIZE + (address) * 4 * SIZE(BB), %xmm0; \
	addps	%xmm3, %xmm6; \
	movaps	80 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	addps	%xmm0, %xmm7; \
	movaps	 32 * SIZE + (address) * 1 * SIZE(AA), %xmm0

#define KERNEL5(address) \
	PREFETCH (PREFETCHSIZE + 16) * SIZE + (address) * 1 * SIZE(AA); \
	mulps	%xmm1, %xmm2; \
	addps	%xmm2, %xmm4; \
	movaps	68 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulps	%xmm1, %xmm2; \
	addps	%xmm2, %xmm5; \
	movaps	72 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulps	%xmm1, %xmm2; \
	mulps	76 * SIZE + (address) * 4 * SIZE(BB), %xmm1; \
	addps	%xmm2, %xmm6; \
	movaps	 96 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	addps	%xmm1, %xmm7; \
	movaps	20 * SIZE + (address) * 1 * SIZE(AA), %xmm1

#define KERNEL6(address) \
	mulps	%xmm1, %xmm3; \
	addps	%xmm3, %xmm4; \
	movaps	84 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulps	%xmm1, %xmm3; \
	addps	%xmm3, %xmm5; \
	movaps	88 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulps	%xmm1, %xmm3; \
	mulps	92 * SIZE + (address) * 4 * SIZE(BB), %xmm1; \
	addps	%xmm3, %xmm6; \
	movaps	112 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	addps	%xmm1, %xmm7; \
	movaps	 24 * SIZE + (address) * 1 * SIZE(AA), %xmm1

#define KERNEL7(address) \
	mulps	%xmm1, %xmm2; \
	addps	%xmm2, %xmm4; \
	movaps	100 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulps	%xmm1, %xmm2; \
	addps	%xmm2, %xmm5; \
	movaps	104 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulps	%xmm1, %xmm2; \
	mulps	108 * SIZE + (address) * 4 * SIZE(BB), %xmm1; \
	addps	%xmm2, %xmm6; \
	movaps	128 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	addps	%xmm1, %xmm7; \
	movaps	28 * SIZE + (address) * 1 * SIZE(AA), %xmm1

#define KERNEL8(address) \
	mulps	%xmm1, %xmm3; \
	addps	%xmm3, %xmm4; \
	movaps	116 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulps	%xmm1, %xmm3; \
	addps	%xmm3, %xmm5; \
	movaps	120 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulps	%xmm1, %xmm3; \
	mulps	124 * SIZE + (address) * 4 * SIZE(BB), %xmm1; \
	addps	%xmm3, %xmm6; \
	movaps	144 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	addps	%xmm1, %xmm7; \
	movaps	 48 * SIZE + (address) * 1 * SIZE(AA), %xmm1;
#endif
	
#ifdef PENTIUM4
#define KERNEL1(address) \
	mulps	%xmm0, %xmm2; \
	addps	%xmm2, %xmm4; \
	movaps	 4 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulps	%xmm0, %xmm2; \
	PREFETCH (PREFETCHSIZE +  0) * SIZE + (address) * 1 * SIZE(AA); \
	addps	%xmm2, %xmm5; \
	movaps	 8 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulps	%xmm0, %xmm2; \
	mulps	12 * SIZE + (address) * 4 * SIZE(BB), %xmm0; \
	addps	%xmm2, %xmm6; \
	movaps	32 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	addps	%xmm0, %xmm7; \
	movaps	 4 * SIZE + (address) * 1 * SIZE(AA), %xmm0

#define KERNEL2(address) \
	mulps	%xmm0, %xmm3; \
	addps	%xmm3, %xmm4; \
	movaps	20 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulps	%xmm0, %xmm3; \
	addps	%xmm3, %xmm5; \
	movaps	24 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulps	%xmm0, %xmm3; \
	mulps	28 * SIZE + (address) * 4 * SIZE(BB), %xmm0; \
	addps	%xmm3, %xmm6; \
	movaps	48 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	addps	%xmm0, %xmm7; \
	movaps	 8 * SIZE + (address) * 1 * SIZE(AA), %xmm0

#define KERNEL3(address) \
	mulps	%xmm0, %xmm2; \
	addps	%xmm2, %xmm4; \
	movaps	36 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulps	%xmm0, %xmm2; \
	addps	%xmm2, %xmm5; \
	movaps	40 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulps	%xmm0, %xmm2; \
	mulps	44 * SIZE + (address) * 4 * SIZE(BB), %xmm0; \
	addps	%xmm2, %xmm6; \
	movaps	64 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	addps	%xmm0, %xmm7; \
	movaps	12 * SIZE + (address) * 1 * SIZE(AA), %xmm0

#define KERNEL4(address) \
	mulps	%xmm0, %xmm3; \
	addps	%xmm3, %xmm4; \
	movaps	52 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulps	%xmm0, %xmm3; \
	addps	%xmm3, %xmm5; \
	movaps	56 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulps	%xmm0, %xmm3; \
	mulps	60 * SIZE + (address) * 4 * SIZE(BB), %xmm0; \
	addps	%xmm3, %xmm6; \
	movaps	80 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	addps	%xmm0, %xmm7; \
	movaps	 32 * SIZE + (address) * 1 * SIZE(AA), %xmm0

#define KERNEL5(address) \
	mulps	%xmm1, %xmm2; \
	addps	%xmm2, %xmm4; \
	movaps	68 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulps	%xmm1, %xmm2; \
	addps	%xmm2, %xmm5; \
	movaps	72 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulps	%xmm1, %xmm2; \
	mulps	76 * SIZE + (address) * 4 * SIZE(BB), %xmm1; \
	addps	%xmm2, %xmm6; \
	movaps	 96 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	addps	%xmm1, %xmm7; \
	movaps	20 * SIZE + (address) * 1 * SIZE(AA), %xmm1

#define KERNEL6(address) \
	mulps	%xmm1, %xmm3; \
	addps	%xmm3, %xmm4; \
	movaps	84 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulps	%xmm1, %xmm3; \
	addps	%xmm3, %xmm5; \
	movaps	88 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulps	%xmm1, %xmm3; \
	mulps	92 * SIZE + (address) * 4 * SIZE(BB), %xmm1; \
	addps	%xmm3, %xmm6; \
	movaps	112 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	addps	%xmm1, %xmm7; \
	movaps	 24 * SIZE + (address) * 1 * SIZE(AA), %xmm1

#define KERNEL7(address) \
	mulps	%xmm1, %xmm2; \
	addps	%xmm2, %xmm4; \
	movaps	100 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulps	%xmm1, %xmm2; \
	addps	%xmm2, %xmm5; \
	movaps	104 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	mulps	%xmm1, %xmm2; \
	mulps	108 * SIZE + (address) * 4 * SIZE(BB), %xmm1; \
	addps	%xmm2, %xmm6; \
	movaps	128 * SIZE + (address) * 4 * SIZE(BB), %xmm2; \
	addps	%xmm1, %xmm7; \
	movaps	28 * SIZE + (address) * 1 * SIZE(AA), %xmm1

#define KERNEL8(address) \
	mulps	%xmm1, %xmm3; \
	addps	%xmm3, %xmm4; \
	movaps	116 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulps	%xmm1, %xmm3; \
	addps	%xmm3, %xmm5; \
	movaps	120 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	mulps	%xmm1, %xmm3; \
	mulps	124 * SIZE + (address) * 4 * SIZE(BB), %xmm1; \
	addps	%xmm3, %xmm6; \
	movaps	144 * SIZE + (address) * 4 * SIZE(BB), %xmm3; \
	addps	%xmm1, %xmm7; \
	movaps	 48 * SIZE + (address) * 1 * SIZE(AA), %xmm1
#endif

	PROLOGUE

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

	PROFCODE

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

	movss	STACK_ALPHA_R, %xmm0
	movss	STACK_ALPHA_I, %xmm1

	xorps	%xmm7, %xmm7
	cmpeqps	%xmm7, %xmm7
	pslld	$31, %xmm7	# Generate mask
	xorps	%xmm2, %xmm2

	shufps	$0, %xmm0, %xmm0

	movaps	 %xmm0,  0 + ALPHA_R
	movss	 %xmm1,  4 + ALPHA_I
	movss	 %xmm1, 12 + ALPHA_I
	xorps	 %xmm7, %xmm1
	movss	 %xmm1,  0 + ALPHA_I
	movss	 %xmm1,  8 + ALPHA_I

#if   defined(NN) || defined(NT) || defined(NR) || defined(NC) || \
      defined(TN) || defined(TT) || defined(TR) || defined(TC)
	movss	  %xmm7,   0 + POSINV
	movss	  %xmm2,   4 + POSINV
	movss	  %xmm7,   8 + POSINV
	movss	  %xmm2,  12 + POSINV
#else
	movss	  %xmm2,   0 + POSINV
	movss	  %xmm7,   4 + POSINV
	movss	  %xmm2,   8 + POSINV
	movss	  %xmm7,  12 + POSINV
#endif

	EMMS

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

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

/* Copying to Sub Buffer */
	leal	BUFFER, %ecx
	movaps	POSINV, %xmm7

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

.L02:
	movss	0 * SIZE(B), %xmm0
	movss	1 * SIZE(B), %xmm1
	movss	2 * SIZE(B), %xmm2
	movss	3 * SIZE(B), %xmm3

	shufps	$0, %xmm0, %xmm0
	shufps	$0, %xmm1, %xmm1
	shufps	$0, %xmm2, %xmm2
	shufps	$0, %xmm3, %xmm3

#if defined(NN) || defined(NT) || defined(NR) || defined(NC)  || \
    defined(TN) || defined(TT) || defined(TR) || defined(TC) 
	xorps	 %xmm7, %xmm1
	xorps	 %xmm7, %xmm3
#else
	xorps	 %xmm7, %xmm0
	xorps	 %xmm7, %xmm2
#endif

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

	movss	4 * SIZE(B), %xmm0
	movss	5 * SIZE(B), %xmm1
	movss	6 * SIZE(B), %xmm2
	movss	7 * SIZE(B), %xmm3

	shufps	$0, %xmm0, %xmm0
	shufps	$0, %xmm1, %xmm1
	shufps	$0, %xmm2, %xmm2
	shufps	$0, %xmm3, %xmm3

#if defined(NN) || defined(NT) || defined(NR) || defined(NC)  || \
    defined(TN) || defined(TT) || defined(TR) || defined(TC) 
	xorps	 %xmm7, %xmm1
	xorps	 %xmm7, %xmm3
#else
	xorps	 %xmm7, %xmm0
	xorps	 %xmm7, %xmm2
#endif

	movaps	%xmm0, 16 * SIZE(BB)
	movaps	%xmm1, 20 * SIZE(BB)
	movaps	%xmm2, 24 * SIZE(BB)
	movaps	%xmm3, 28 * SIZE(BB)

#ifdef PENTIUM4
	prefetcht1	104 * SIZE(BB)
#endif

	addl	$ 8 * SIZE, %edi
	addl	$32 * SIZE, %ecx

	decl	%eax
	jne	.L02
	ALIGN_4

.L03:
	movl	K, %eax
	andl	$1, %eax
	BRANCH
	jle	.L05
	ALIGN_4

.L04:
	movss	0 * SIZE(B), %xmm0
	movss	1 * SIZE(B), %xmm1
	movss	2 * SIZE(B), %xmm2
	movss	3 * SIZE(B), %xmm3

	shufps	$0, %xmm0, %xmm0
	shufps	$0, %xmm1, %xmm1
	shufps	$0, %xmm2, %xmm2
	shufps	$0, %xmm3, %xmm3

#if defined(NN) || defined(NT) || defined(NR) || defined(NC)  || \
    defined(TN) || defined(TT) || defined(TR) || defined(TC) 
	xorps	 %xmm7, %xmm1
	xorps	 %xmm7, %xmm3
#else
	xorps	 %xmm7, %xmm0
	xorps	 %xmm7, %xmm2
#endif

	movaps	%xmm0,  0 * SIZE(BB)