symv_u_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"

#ifdef CORE2
#define PREFETCH	prefetcht0
#define PREFETCHW	prefetcht0
#define PREFETCHSIZE	(16 * 12)
#endif

#ifdef PENRYN
#define PREFETCH	prefetcht0
#define PREFETCHW	prefetcht0
#define PREFETCHSIZE	(16 * 12)
#endif

#ifdef PENTIUM4
#define PREFETCH	prefetcht0
#define PREFETCHW	prefetcht0
#define PREFETCHSIZE	(16 * 20)
#endif

#ifdef OPTERON
#define PREFETCH	prefetch
#define PREFETCHW	prefetchw
#define PREFETCHSIZE	(16 * 8)
#define movsd		movlps
#endif

#ifdef BARCELONA
#define PREFETCH	prefetch
#define PREFETCHW	prefetchw
#define PREFETCHSIZE	(16 * 16)
#endif

#ifdef GENERIC
#define PREFETCH	prefetcht0
#define PREFETCHW	prefetcht0
#define PREFETCHSIZE	(16 * 20)
#endif

#ifndef WINDOWS_ABI

#define STACKSIZE	80
	
#define OLD_INCY	 8 + STACKSIZE(%rsp)
#define OLD_BUFFER	16 + STACKSIZE(%rsp)

#define M	  ARG1
#define	A	  ARG2
#define LDA	  ARG3	
#define	X	  ARG4
#define INCX	  ARG5	
#define	Y	  ARG6
#define INCY	  %r10
#define BUFFER	  %r11

#else

#define STACKSIZE	256
	
#define OLD_X		 40 + STACKSIZE(%rsp)
#define OLD_INCX	 48 + STACKSIZE(%rsp)
#define OLD_Y		 56 + STACKSIZE(%rsp)
#define OLD_INCY	 64 + STACKSIZE(%rsp)
#define OLD_BUFFER	 72 + STACKSIZE(%rsp)

#define M	  ARG1
#define	A	  ARG3
#define LDA	  ARG4
#define	X	  %rdi
#define INCX	  %rsi	
#define	Y	  %rdx
#define INCY	  %r10
#define BUFFER	  %r11

#endif

#define TEMP	%rax
#define I	%rax
#define IS	%r12
#define A1	%rbx
#define A2	%rbp
#define XX	%r13
#define YY	%r14
#define NEW_X	BUFFER
#define NEW_Y	X

#define ALPHA  %xmm0

#define atemp1 %xmm0
#define atemp2 %xmm1
#define atemp3 %xmm2
#define atemp4 %xmm3

#define xsum1  %xmm4
#define xsum2  %xmm5
#define xsum3  %xmm6
#define xsum4  %xmm7

#define xtemp1 %xmm8
#define xtemp2 %xmm9
#define yy1    %xmm10
#define	xt1    %xmm11

#define a1     %xmm12
#define a2     %xmm13
#define a3     %xmm14
#define a4     %xmm15


	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	OLD_X,     X
	movq	OLD_INCX,  INCX
	movq	OLD_Y,     Y

	movaps	%xmm1, %xmm0
#endif

	movq	OLD_INCY,   INCY
	movq	OLD_BUFFER, BUFFER

	leaq	(,INCX, SIZE), INCX
	leaq	(,INCY, SIZE), INCY
	leaq	(,LDA,  SIZE), LDA

	testq	M, M
	jle	.L999

	shufps	$0, ALPHA, ALPHA

	movq	BUFFER, XX

	movq	M,  %rax
	sarq	$3, %rax
	jle	.L02
	ALIGN_3

.L01:
	movss	0 * SIZE(X), %xmm1
	addq	INCX, X
	movss	0 * SIZE(X), %xmm2
	addq	INCX, X
	movss	0 * SIZE(X), %xmm3
	addq	INCX, X
	movss	0 * SIZE(X), %xmm4
	addq	INCX, X
	movss	0 * SIZE(X), %xmm5
	addq	INCX, X
	movss	0 * SIZE(X), %xmm6
	addq	INCX, X
	movss	0 * SIZE(X), %xmm7
	addq	INCX, X
	movss	0 * SIZE(X), %xmm8
	addq	INCX, X

	mulss	ALPHA, %xmm1
	mulss	ALPHA, %xmm2
	mulss	ALPHA, %xmm3
	mulss	ALPHA, %xmm4
	mulss	ALPHA, %xmm5
	mulss	ALPHA, %xmm6
	mulss	ALPHA, %xmm7
	mulss	ALPHA, %xmm8

	movss	%xmm1, 0 * SIZE(XX)
	movss	%xmm2, 1 * SIZE(XX)
	movss	%xmm3, 2 * SIZE(XX)
	movss	%xmm4, 3 * SIZE(XX)
	movss	%xmm5, 4 * SIZE(XX)
	movss	%xmm6, 5 * SIZE(XX)
	movss	%xmm7, 6 * SIZE(XX)
	movss	%xmm8, 7 * SIZE(XX)

	addq	$8 * SIZE, XX
	decq	%rax
	jg	.L01
	ALIGN_3

.L02:
	movq	M, %rax
	andq	$7, %rax
	jle	.L05
	ALIGN_3

.L03:
	movss	0 * SIZE(X), %xmm1
	addq	INCX, X

	mulss	ALPHA, %xmm1

	movss	%xmm1, 0 * SIZE(XX)

	addq	$1 * SIZE, XX
	decq	%rax
	jg	.L03
	ALIGN_3

.L05:
	/* now we don't need original X */
	movq   Y, NEW_Y

	addq   $512, XX
	andq   $-512, XX

	cmpq   $SIZE, INCY
	je    .L10

	movq   Y,  YY
	movq   XX, NEW_Y

	movq	M,  %rax
	sarq	$3, %rax
	jle	.L07
	ALIGN_3

.L06:
	movss	0 * SIZE(YY), %xmm0
	addq	INCY, YY
	movss	0 * SIZE(YY), %xmm1
	addq	INCY, YY
	movss	0 * SIZE(YY), %xmm2
	addq	INCY, YY
	movss	0 * SIZE(YY), %xmm3
	addq	INCY, YY
	movss	0 * SIZE(YY), %xmm4
	addq	INCY, YY
	movss	0 * SIZE(YY), %xmm5
	addq	INCY, YY
	movss	0 * SIZE(YY), %xmm6
	addq	INCY, YY
	movss	0 * SIZE(YY), %xmm7
	addq	INCY, YY

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

	addq	$8 * SIZE, XX
	decq	%rax
	jg	.L06
	ALIGN_3

.L07:
	movq	M, %rax
	andq	$7, %rax
	jle	.L10
	ALIGN_3

.L08:
	movss	0 * SIZE(YY), %xmm0
	addq	INCY, YY

	movss	%xmm0, 0 * SIZE(XX)

	addq	$1 * SIZE, XX
	decq	%rax
	jg	.L08
	ALIGN_3

.L10:
	xorq	IS, IS		# is = 0

	cmpq	$4, M
	jl	.L20
	ALIGN_3

.L11:
	movq	A,  A1
	leaq	(A, LDA, 2), A2
	leaq	(A, LDA, 4), A

	movaps		0 * SIZE(NEW_X, IS, SIZE), atemp4

	pshufd	$0x00, atemp4, atemp1
	pshufd	$0x55, atemp4, atemp2
	pshufd	$0xaa, atemp4, atemp3
	pshufd	$0xff, atemp4, atemp4

	pxor		xsum1, xsum1
	pxor		xsum2, xsum2
	pxor		xsum3, xsum3
	pxor		xsum4, xsum4

	movaps	 0 * SIZE(NEW_X), xtemp1
	movaps	 4 * SIZE(NEW_X), xtemp2

	movsd	 0 * SIZE(A1), a1
	movhps	 2 * SIZE(A1), a1
	movsd	 0 * SIZE(A1, LDA, 1), a2
	movhps	 2 * SIZE(A1, LDA, 1), a2
	movsd	 0 * SIZE(A2), a3
	movhps	 2 * SIZE(A2), a3
	movsd	 0 * SIZE(A2, LDA, 1), a4
	movhps	 2 * SIZE(A2, LDA, 1), a4

	movsd	 0 * SIZE(NEW_Y), yy1
	movhps	 2 * SIZE(NEW_Y), yy1

	movq		NEW_X, XX
	movq		NEW_Y, YY

	movq	IS,  I
	sarq	$4,  I
	jle	.L14
	ALIGN_3

.L12:
	movaps	 xtemp1, xt1
	mulps	 a1,     xt1
	mulps	 atemp1, a1
	addps	 xt1,    xsum1
	addps	 a1,     yy1
	movsd	 4 * SIZE(A1), a1
	movhps	 6 * SIZE(A1), a1

	PREFETCH	PREFETCHSIZE(A1)

	movaps	 xtemp1, xt1
	mulps	 a2,     xt1
	mulps	 atemp2, a2
	addps	 xt1,    xsum2
	addps	 a2,     yy1
	movsd	 4 * SIZE(A1, LDA, 1), a2
	movhps	 6 * SIZE(A1, LDA, 1), a2

	movaps	 xtemp1, xt1
	mulps	 a3,     xt1
	mulps	 atemp3, a3
	addps	 xt1,    xsum3
	addps	 a3,     yy1
	movsd	 4 * SIZE(A2), a3
	movhps	 6 * SIZE(A2), a3

#if !defined(CORE2) && !defined(PENRYN)
	PREFETCH	PREFETCHSIZE(XX)
#endif

	movaps	 xtemp1, xt1
	movaps	 8 * SIZE(XX), xtemp1
	mulps	 a4,     xt1
	mulps	 atemp4, a4
	addps	 xt1,    xsum4
	addps	 a4,     yy1
	movsd	 4 * SIZE(A2, LDA, 1), a4
	movhps	 6 * SIZE(A2, LDA, 1), a4

	movlps	 yy1, 0 * SIZE(YY)
	movhps	 yy1, 2 * SIZE(YY)
	movsd	 4 * SIZE(YY), yy1
	movhps	 6 * SIZE(YY), yy1

	movaps	 xtemp2, xt1
	mulps	 a1,     xt1
	mulps	 atemp1, a1
	addps	 xt1,    xsum1
	addps	 a1,     yy1
	movsd	 8 * SIZE(A1), a1
	movhps	10 * SIZE(A1), a1

	PREFETCH	PREFETCHSIZE(A1, LDA, 1)

	movaps	 xtemp2, xt1
	mulps	 a2,     xt1
	mulps	 atemp2, a2
	addps	 xt1,    xsum2
	addps	 a2,     yy1
	movsd	 8 * SIZE(A1, LDA, 1), a2
	movhps	10 * SIZE(A1, LDA, 1), a2

	movaps	 xtemp2, xt1
	mulps	 a3,     xt1
	mulps	 atemp3, a3
	addps	 xt1,    xsum3
	addps	 a3,     yy1
	movsd	 8 * SIZE(A2), a3
	movhps	10 * SIZE(A2), a3

	movaps	 xtemp2, xt1
	movaps	12 * SIZE(XX), xtemp2
	mulps	 a4,     xt1
	mulps	 atemp4, a4
	addps	 xt1,    xsum4
	addps	 a4,     yy1
	movsd	 8 * SIZE(A2, LDA, 1), a4
	movhps	10 * SIZE(A2, LDA, 1), a4

	movlps	 yy1, 4 * SIZE(YY)
	movhps	 yy1, 6 * SIZE(YY)
	movsd	 8 * SIZE(YY), yy1
	movhps	10 * SIZE(YY), yy1


	movaps	 xtemp1, xt1
	mulps	 a1,     xt1
	mulps	 atemp1, a1
	addps	 xt1,    xsum1
	addps	 a1,     yy1
	movsd	12 * SIZE(A1), a1
	movhps	14 * SIZE(A1), a1

	PREFETCH	PREFETCHSIZE(A2)

	movaps	 xtemp1, xt1
	mulps	 a2,     xt1
	mulps	 atemp2, a2
	addps	 xt1,    xsum2
	addps	 a2,     yy1
	movsd	12 * SIZE(A1, LDA, 1), a2
	movhps	14 * SIZE(A1, LDA, 1), a2

	movaps	 xtemp1, xt1
	mulps	 a3,     xt1
	mulps	 atemp3, a3
	addps	 xt1,    xsum3
	addps	 a3,     yy1
	movsd	12 * SIZE(A2), a3
	movhps	14 * SIZE(A2), a3

#if !defined(CORE2) && !defined(PENRYN)
	PREFETCHW	PREFETCHSIZE(YY)
#endif

	movaps	 xtemp1, xt1
	movaps	16 * SIZE(XX), xtemp1
	mulps	 a4,     xt1
	mulps	 atemp4, a4
	addps	 xt1,    xsum4
	addps	 a4,     yy1
	movsd	12 * SIZE(A2, LDA, 1), a4
	movhps	14 * SIZE(A2, LDA, 1), a4

	movlps	 yy1,  8 * SIZE(YY)
	movhps	 yy1, 10 * SIZE(YY)
	movsd	12 * SIZE(YY), yy1
	movhps	14 * SIZE(YY), yy1

	movaps	 xtemp2, xt1
	mulps	 a1,     xt1
	mulps	 atemp1, a1
	addps	 xt1,    xsum1
	addps	 a1,     yy1
	movsd	16 * SIZE(A1), a1
	movhps	18 * SIZE(A1), a1

	PREFETCH	PREFETCHSIZE(A2, LDA, 1)

	movaps	 xtemp2, xt1
	mulps	 a2,     xt1
	mulps	 atemp2, a2
	addps	 xt1,    xsum2
	addps	 a2,     yy1
	movsd	16 * SIZE(A1, LDA, 1), a2
	movhps	18 * SIZE(A1, LDA, 1), a2

	movaps	 xtemp2, xt1
	mulps	 a3,     xt1
	mulps	 atemp3, a3
	addps	 xt1,    xsum3