zgemm_ncopy.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 PREFETCHSIZE   64
#define WPREFETCHSIZE  32

#define LD	LDF8
#define ST	STF8_NTA
	
#define TEMP	r2

#define I	r14
#define J	r15
#define PREB	r16
#define PREA	r17

#define A1	r18
#define A2	r19
#define A3	r20
#define A4	r21
#define A5	r22
#define A6	r23
#define A7	r24
#define A8	r25
#define B1	r26

#define COUNT	r28

#define ARLC	r30
#define PR	r31

#define M	r32
#define N	r33
#define A	r34
#define LDA	r35
#define B	r36

	PROLOGUE
	.prologue
	PROFCODE

	.body	
	{ .mii
	shladd	LDA= LDA, ZBASE_SHIFT, r0
	mov	PR = pr
	shr	J = N, 2
	}
	;;
	{ .mii
	mov COUNT=r0
	tbit.nz p10, p0 =M, 1
	tbit.nz p11, p0 =M, 0
	}
	;;
	{ .mmb
	nop	__LINE__
	nop	__LINE__
	nop	__LINE__
	}
	{ .mib
	cmp.eq	p8,p0 = 0, J
	mov	ARLC = ar.lc
	(p8) br.cond.dpnt .L20
	}
	;;
      .align 32

.L11:
	{ .mmi
	mov	A1 = A
	add	A2 = A, LDA
	mov	pr.rot = 0
	}
	{ .mmi
	shladd A3 = LDA, 1, A
	adds   B1 = 4 * SIZE, B
	shr    I  = M, 2
	}
	;;
	{ .mmi
	shladd	A4 = LDA, 1, A2
	cmp.eq	p16,p0 = r0, r0
	mov	ar.ec = 3
	}
	{ .mmi
	cmp.eq	p6,p0 = 0,I
	adds	I =-1, I
	adds	J =-1, J
	}
	;;
	{ .mmi
	shladd	A = LDA, 2, A
	adds	A5 = 4 * SIZE, A1
	adds	A6 = 4 * SIZE, A2
	}
	{ .mmi
	adds	A7 = 4 * SIZE, A3
	adds	A8 = 4 * SIZE, A4
	adds	PREA = PREFETCHSIZE * SIZE,A1
	}
	;;
	{ .mmb
	nop	__LINE__
	nop	__LINE__
	nop	__LINE__
	}
	{ .mib
	adds	PREB = WPREFETCHSIZE * SIZE, B
	mov	ar.lc = I
	(p6) br.cond.dpnt.few .L15
	}
	;;
	.align 32

.L12:
	{ .mmb
	(p16)	lfetch.nt1	[PREA], LDA
	(p16)	lfetch.excl.nt1	[PREB], 16 * SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p18)	ST	[B ] = f34, SIZE
	(p18)	ST	[B1] = f82, SIZE
	nop	__LINE__
	}
	{ .mmb
	(p16)	LD	f32 = [A1], SIZE
	(p16)	LD	f35 = [A5], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p18)	ST	[B ] = f40, SIZE
	(p18)	ST	[B1] = f88, SIZE
	nop	__LINE__
	}
	{ .mmb
	(p16)	LD	f38 = [A1], SIZE
	(p16)	LD	f41 = [A5], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p18)	ST	[B ] = f58,  SIZE
	(p18)	ST	[B1] = f106, SIZE
	nop	__LINE__
	}
	{ .mmb
	(p16)	LD	f44 = [A1], SIZE
	(p16)	LD	f47 = [A5], SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p18)	ST	[B ] = f64,  5 * SIZE
	(p18)	ST	[B1] = f112, 5 * SIZE
	tbit.z	p0,p7 = COUNT,0
	}
	{ .mmb
	(p16)	LD	f50 = [A1], 5 * SIZE
	(p16)	LD	f53 = [A5], 5 * SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p18)	ST	[B ] = f46, SIZE
	(p18)	ST	[B1] = f94, SIZE
	nop	__LINE__
	}
	{ .mmb
	(p16)	LD	f56 = [A2], SIZE
	(p16)	LD	f59 = [A6], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p18)	ST	[B ] = f52, SIZE
	(p18)	ST	[B1] = f100, SIZE
	nop	__LINE__
	}
	{ .mmb
	(p16)	LD	f62 = [A2], SIZE
	(p16)	LD	f65 = [A6], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p18)	ST	[B ] = f70, SIZE
	(p18)	ST	[B1] = f118, SIZE
	nop	__LINE__
	}
	{ .mmb
	(p16)	LD	f68 = [A2], SIZE
	(p16)	LD	f71 = [A6], SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p18)	ST	[B ]  = f76, 5 * SIZE
	(p18)	ST	[B1]  = f124, 5 * SIZE
	shladd	TEMP = LDA, 2, r0
	}
	{ .mmb
	(p16)	LD	f74 = [A2], 5 * SIZE
	(p16)	LD	f77 = [A6], 5 * SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p16)	lfetch.nt1	[PREA], LDA
	(p16)	lfetch.excl.nt1	[PREB], 16 * SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p18)	ST	[B ] = f37, SIZE
	(p18)	ST	[B1] = f85, SIZE
	nop	__LINE__
	}
	{ .mmb
	(p16)	LD	f80 = [A3], SIZE
	(p16)	LD	f83 = [A7], SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p18)	ST	[B ] = f43, SIZE
	(p18)	ST	[B1] = f91, SIZE
	adds	TEMP = -16 * SIZE, TEMP
	}
	{ .mmb
	(p16)	LD	f86 = [A3], SIZE
	(p16)	LD	f89 = [A7], SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p18)	ST	[B ] = f61, SIZE
	(p18)	ST	[B1] = f109, SIZE
	(p7)	sub	PREA = PREA, TEMP
	}
	{ .mmb
	(p16)	LD	f92 = [A3], SIZE
	(p16)	LD	f95 = [A7], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p18)	ST	[B ] = f67, 5 * SIZE
	(p18)	ST	[B1] = f115, 5 * SIZE
	nop	__LINE__
	}
	{ .mmb
	(p16)	LD	f98 = [A3], 5 * SIZE
	(p16)	LD	f101 = [A7], 5 * SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p18)	ST	[B ] = f49, SIZE
	(p18)	ST	[B1] = f97, SIZE
	nop	__LINE__
	}
	{ .mmb
	(p16)	LD	f104 = [A4], SIZE
	(p16)	LD	f107 = [A8], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p18)	ST	[B ] = f55, SIZE
	(p18)	ST	[B1] = f103, SIZE
	nop	__LINE__
	}
	{ .mmb
	(p16)	LD	f110 = [A4], SIZE
	(p16)	LD	f113 = [A8], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p18)	ST	[B ] = f73, SIZE
	(p18)	ST	[B1] = f121, SIZE
	nop	__LINE__
	}
	{ .mmb
	(p16)	LD	f116 = [A4], SIZE
	(p16)	LD	f119 = [A8], SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p18)	ST	[B ] = f79, 5 * SIZE
	(p18)	ST	[B1] = f127, 5 * SIZE
	(p16)	adds	COUNT =  1, COUNT
	}
	{ .mmb
	(p16)	LD	f122 = [A4], 5 * SIZE
	(p16)	LD	f125 = [A8], 5 * SIZE
	br.ctop.sptk.few .L12
	}
	;;
	.align 32

.L15:
	{ .mmb
	(p10)	LD	f32 = [A1], SIZE
	(p10)	LD	f40 = [A2], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	LD	f33 = [A1], SIZE
	(p10)	LD	f41 = [A2], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	LD	f34 = [A1], SIZE
	(p10)	LD	f42 = [A2], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	LD	f35 = [A1], SIZE
	(p10)	LD	f43 = [A2], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	LD	f50 = [A3], SIZE
	(p10)	LD	f60 = [A4], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	LD	f51 = [A3], SIZE
	(p10)	LD	f61 = [A4], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	LD	f52 = [A3], SIZE
	(p10)	LD	f62 = [A4], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	LD	f53 = [A3], SIZE
	(p10)	LD	f63 = [A4], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p11)	LD	f36 = [A1], SIZE
	(p11)	LD	f44 = [A2], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p11)	LD	f37 = [A1]
	(p11)	LD	f45 = [A2]
	nop	__LINE__
	}
	;;
	{ .mmb
	(p11)	LD	f54 = [A3], SIZE
	(p11)	LD	f64 = [A4], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p11)	LD	f55 = [A3]
	(p11)	LD	f65 = [A4]
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	ST	[B ] = f32, SIZE
	(p10)	ST	[B1] = f50, SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	ST	[B ] = f33, SIZE
	(p10)	ST	[B1] = f51, SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	ST	[B ] = f40, SIZE
	(p10)	ST	[B1] = f60, SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	ST	[B ] = f41, 5 * SIZE
	(p10)	ST	[B1] = f61, 5 * SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	ST	[B ] = f34, SIZE
	(p10)	ST	[B1] = f52, SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	ST	[B ] = f35, SIZE
	(p10)	ST	[B1] = f53, SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	ST	[B ] = f42, SIZE
	(p10)	ST	[B1] = f62, SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	ST	[B ] = f43, 5 * SIZE
	(p10)	ST	[B1] = f63, 5 * SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p11)	ST	[B ] = f36, SIZE
	(p11)	ST	[B1] = f54, SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p11)	ST	[B ] = f37, SIZE
	(p11)	ST	[B1] = f55, SIZE
	mov	COUNT = r0
	}
	;;
	{ .mmi
	(p11)	ST	[B ] = f44, SIZE
	(p11)	ST	[B1] = f64, SIZE
	cmp.eq	p0,p6 = 0,J
	}
	;;
	{ .mmb
	(p11)	ST	[B ] = f45, 5 * SIZE
	(p11)	ST	[B1] = f65, 5 * SIZE
	(p6)	br.cond.dptk.few .L11
	}
	;;
	.align 32

.L20:
	{ .mmi
	mov	A1 = A
	add	A2 = A,LDA
	mov	pr.rot = 0
	}
	{ .mmi
	adds	A5 = 4 * SIZE, A
	adds	B1 = 4 * SIZE, B
	tbit.z	p8, p0 = N, 1
	}
	;;
	{ .mmi
	cmp.eq	p16,p0 = r0,r0
	adds	PREA = PREFETCHSIZE * SIZE, A
	mov	ar.ec = 3
	}
	;;
	{ .mib
	adds	PREB = WPREFETCHSIZE * SIZE,B
	shr	I = M, 2
	(p8)	br.cond.dpnt.few .L30
	}
	;;
	{ .mmi
	shladd	A = LDA, 1, A
	cmp.eq	p6, p0 = 0, I
	adds	I = -1, I
	}
	;;
	{ .mib
	adds	A6 = 4 * SIZE, A2
	mov	ar.lc = I
	(p6)	br.cond.dpnt.few .L25
	}
	;;
	.align 32

.L21:
	{ .mmb
	(p16)	lfetch.nt1	[PREA],LDA
	(p16)	lfetch.excl.nt1	[PREB ],16 * SIZE
	nop	__LINE__
	}
	{ .mmb
	nop	__LINE__
	nop	__LINE__
	nop	__LINE__
	}
	;;
	{ .mmb
	(p18)	ST	[B ] = f34, SIZE
	(p18)	ST	[B1] = f46, SIZE
	nop	__LINE__
	}
	{ .mmb
	(p16)	LD	f32 = [A1], SIZE
	(p16)	LD	f35 = [A5], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p18)	ST	[B ] = f40, SIZE
	(p18)	ST	[B1] = f52, SIZE
	nop	__LINE__
	}
	{ .mmb
	(p16)	LD	f38 = [A1], SIZE
	(p16)	LD	f41 = [A5], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p18)	ST	[B ] = f58, SIZE
	(p18)	ST	[B1] = f70, SIZE
	nop	__LINE__
	}
	{ .mmb
	(p16)	LD	f44 = [A1], SIZE
	(p16)	LD	f47 = [A5], SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p18)	ST	[B ] = f64, 5 * SIZE
	(p18)	ST	[B1] = f76, 5 * SIZE
	tbit.z	p0,p7 = COUNT,0
	}
	{ .mmb
	(p16)	LD	f50 = [A1], 5 * SIZE
	(p16)	LD	f53 = [A5], 5 * SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p18)	ST	[B ] = f37, SIZE
	(p18)	ST	[B1] = f49, SIZE
	adds	TEMP = -16 * SIZE,TEMP
	}
	{ .mmb
	(p16)	LD	f56 = [A2], SIZE
	(p16)	LD	f59 = [A6], SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p18)	ST	[B ] = f43, SIZE
	(p18)	ST	[B1] = f55, SIZE
	(p7)	sub	PREA = PREA,TEMP
	}
	{ .mmb
	(p16)	LD	f62 = [A2], SIZE
	(p16)	LD	f65 = [A6], SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p18)	ST	[B ] = f61, SIZE
	(p18)	ST	[B1] = f73, SIZE
	(p16)	adds	COUNT = 1,COUNT
	}
	{ .mmb
	(p16)	LD	f68 = [A2], SIZE
	(p16)	LD	f71 = [A6], SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p18)	ST	[B ] = f67, 5 * SIZE
	(p18)	ST	[B1] = f79, 5 * SIZE
	shladd	TEMP = LDA,2,r0
	}
	{ .mmb
	(p16)	LD	f74 = [A2], 5 * SIZE
	(p16)	LD	f77 = [A6], 5 * SIZE
	br.ctop.sptk.few .L21
	}
	;;
	.align 32

.L25:
	{ .mmb
	(p10)	LD	f32 = [A1], SIZE
	(p10)	LD	f40 = [A2], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	LD	f33 = [A1], SIZE
	(p10)	LD	f41 = [A2], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	LD	f34 = [A1], SIZE
	(p10)	LD	f42 = [A2], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	LD	f35 = [A1], SIZE
	(p10)	LD	f43 = [A2], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p11)	LD	f36 = [A1], SIZE
	(p11)	LD	f44 = [A2], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p11)	LD	f37 = [A1]
	(p11)	LD	f45 = [A2]
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	ST	[B ] = f32, SIZE
	(p10)	ST	[B1] = f34, SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	ST	[B ] = f33, SIZE
	(p10)	ST	[B1] = f35, SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	ST	[B ] = f40, SIZE
	(p10)	ST	[B1] = f42, SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p10)	ST	[B ] = f41, 5 * SIZE
	(p10)	ST	[B1] = f43, 5 * SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p11)	ST	[B ] = f36, SIZE
	;;
	(p11)	ST	[B ] = f37, SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p11)	ST	[B ] = f44, SIZE
	;;
	(p11)	ST	[B ] = f45, SIZE
	nop	__LINE__
	}
	;;
	.align 32

.L30:
	{ .mmi
	mov	A1 = A
	mov	COUNT = r0
	mov	pr.rot = 0
	}
	{ .mmi
	adds	A5 = 4 * SIZE,A
	adds	B1 = 4 * SIZE,B
	tbit.z	p8,p0 = N,0
	}
	;;
	{ .mmi
	cmp.eq	p16,p0 = r0,r0
	nop	__LINE__
	mov	ar.ec = 3
	}
	{ .mib
	nop	__LINE__
	shr	I = M,2
	(p8)	br.cond.dptk.few .L999
	}
	;;
	{ .mmi
	cmp.eq	p6 ,p0 = 0, I
	adds	PREA = PREFETCHSIZE * SIZE, A
	adds	I = -1, I
	}
	;;
	{ .mib
	adds	PREB = WPREFETCHSIZE * SIZE, B
	mov	ar.lc = I
	(p6)	br.cond.dpnt.few .L35
	}
	;;
	.align 32

.L31:
	{ .mmi
	(p16)	lfetch.nt1	[PREA], LDA
	(p16)	lfetch.excl.nt1	[PREB ], 16 * SIZE
	tbit.z	p0, p7 = COUNT, 0
	}
	{ .mmb
	nop	__LINE__
	nop	__LINE__
	nop	__LINE__
	}
	;;
	{ .mmi
	(p18)	ST	[B ] = f34, SIZE
	(p18)	ST	[B1] = f37, SIZE
	shladd	TEMP = LDA,2,r0
	}
	{ .mmb
	(p16)	LD	f32 = [A1], SIZE
	(p16)	LD	f35 = [A5], SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p18)	ST	[B ] = f40, SIZE
	(p18)	ST	[B1] = f43, SIZE
	adds	TEMP = -16 * SIZE,TEMP
	}
	{ .mmb
	(p16)	LD	f38 = [A1], SIZE
	(p16)	LD	f41 = [A5], SIZE
	nop	__LINE__
	}
	;;
	{ .mmb
	(p18)	ST	[B ] = f46, SIZE
	(p18)	ST	[B1] = f49, SIZE
	nop	__LINE__
	}
	{ .mmi
	(p16)	LD	f44 = [A1], SIZE
	(p16)	LD	f47 = [A5], SIZE
	(p7)	sub	PREA = PREA,TEMP
	}
	;;
	{ .mmi
	(p18)	ST	[B ] = f52, 5 * SIZE
	(p18)	ST	[B1] = f55, 5 * SIZE
	(p16)	adds	COUNT = 1,COUNT
	}
	{ .mmb
	(p16)	LD	f50 = [A1], 5 * SIZE
	(p16)	LD	f53 = [A5], 5 * SIZE
	br.ctop.sptk.few .L31
	}
	;;
	.align 32

.L35:
	{ .mmi
	(p10)	LD	f32 = [A1], SIZE
	;;
	(p10)	LD	f33 = [A1], SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p10)	LD	f34 = [A1], SIZE
	;;
	(p10)	LD	f35 = [A1], SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p11)	LD	f36 = [A1], SIZE
	;;
	(p11)	LD	f37 = [A1]
	nop	__LINE__
	}
	;;
	{ .mmi
	(p10)	ST	[B ] = f32, SIZE
	;;
	(p10)	ST	[B ] = f33, SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p10)	ST	[B ] = f34, SIZE
	;;
	(p10)	ST	[B ] = f35, SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	(p11)	ST	[B ] = f36, SIZE
	;;
	(p11)	ST	[B ] = f37, SIZE
	nop	__LINE__
	}
	;;
	.align 32

.L999:
	mov	pr = PR,-1
	mov	ar.lc = ARLC
	br.ret.sptk.many b0
	;;
	EPILOGUE