qgemv_t.s

来自「Optimized GotoBLAS libraries」· S 代码 · 共 1,271 行 · 第 1/2 页

	STFD [YY1] = f35
	add	YY1 = YY1, INCY5
	;;
	STFD [YY2] = f36
	add	YY2 = YY2, INCY
	;;
	STFD [YY2] = f37
	add	YY2 = YY2, INCY
	;;
	STFD [YY2] = f38
	add	YY2 = YY2, INCY
	;;
	STFD [YY2] = f39
	add	YY2 = YY2, INCY5
	;;
	adds	J = -1, J
	;;
	cmp.lt p6, p0 = 0, J
	(p6) br.cond.dptk .L11
	;;
	.align 16

.L20:
	tbit.z	p6, p0  = N, 2
	;;
	(p6) br.cond.dpnt .L30
	;;
	mov	AO1 = A
	adds	AO2 = 1 * SIZE, A
	adds	AO3 = 2 * SIZE, A
	adds	AO4 = 3 * SIZE, A
	adds	AO5 = 4 * SIZE, A
	adds	AO6 = 5 * SIZE, A
	adds	AO7 = 6 * SIZE, A
	adds	AO8 = 7 * SIZE, A
	shladd	A   = LDA, 2, A
	;;
	shladd	LDA7M8 = LDA, 2, r0
	;;
	sub	LDA7M8 = LDA, LDA7M8	
	;;
	adds	LDA7M8 = 8 * SIZE, LDA7M8
	;;	
	mov	f8  = f0
	mov	f9  = f0
	mov	f10 = f0
	mov	f11 = f0
	mov	f12 = f0
	mov	f13 = f0
	mov	f14 = f0
	mov	f15 = f0

	mov	pr.rot= 0
	shr	I = M, 3
	mov	ar.ec = 2
	;;
	mov	X1  = BUFFER
	adds	X2  = 2 * SIZE, BUFFER
	;;
	cmp.eq	p16, p0 = r0, r0
	;;
	adds	I = -1, I
	;;
	mov	ar.lc = I
	cmp.eq	p6, p0 = -1, I
	(p6) br.cond.dpnt .L25
	;;
	.align 16

.L22:
	(p16) LDFD	f32  = [AO1], LDA
	(p16) LDFD	f34  = [AO3], LDA
	(p16) LDFD	f36  = [AO5], LDA
	(p16) LDFD	f38  = [AO7], LDA
	;;
	(p16) LDFD	f33  = [AO2], LDA
	(p16) LDFD	f35  = [AO4], LDA
	(p16) LDFD	f37  = [AO6], LDA
	(p16) LDFD	f39  = [AO8], LDA
	;;
	(p16) LDFD	f40  = [AO1], LDA
	(p16) LDFD	f42  = [AO3], LDA
	(p16) LDFD	f44  = [AO5], LDA
	(p16) LDFD	f46  = [AO7], LDA
	;;
	(p16) LDFD	f41  = [AO2], LDA
	(p16) LDFD	f43  = [AO4], LDA
	(p16) LDFD	f45  = [AO6], LDA
	(p16) LDFD	f47  = [AO8], LDA
	;;
	(p16) LDFD	f48  = [AO1], LDA
	(p16) LDFD	f50  = [AO3], LDA
	(p16) LDFD	f52  = [AO5], LDA
	(p16) LDFD	f54  = [AO7], LDA
	;;
	(p16) LDFD	f49  = [AO2], LDA
	(p16) LDFD	f51  = [AO4], LDA
	(p16) LDFD	f53  = [AO6], LDA
	(p16) LDFD	f55  = [AO8], LDA
	;;
	(p16) LDFD	f56  = [AO1], LDA7M8
	(p16) LDFD	f58  = [AO3], LDA7M8
	(p16) LDFD	f60  = [AO5], LDA7M8
	(p16) LDFD	f62  = [AO7], LDA7M8
	;;
	(p16) LDFD	f57  = [AO2], LDA7M8
	(p16) LDFD	f59  = [AO4], LDA7M8
	(p16) LDFD	f61  = [AO6], LDA7M8
	(p16) LDFD	f63  = [AO8], LDA7M8
	;;
	(p16) LDFD	f96  = [X1], 1 * SIZE
	(p16) LDFD	f98  = [X2], 1 * SIZE
	;;
	(p16) LDFD	f97  = [X1], 3 * SIZE
	(p16) LDFD	f99  = [X2], 3 * SIZE
	;;
	(p16) LDFD	f100 = [X1], 1 * SIZE
	(p16) LDFD	f102 = [X2], 1 * SIZE
	;;
	(p16) LDFD	f101 = [X1], 3 * SIZE
	(p16) LDFD	f103 = [X2], 3 * SIZE
	;;
	(p16) FMA	f8  = f96,  f32, f8
	(p16) FMA	f9  = f96,  f40, f9
	(p16) FMA	f10 = f96,  f48, f10
	(p16) FMA	f11 = f96,  f56, f11
	;;
	(p16) FMA	f8  = f97,  f33, f8 
	(p16) FMA	f9  = f97,  f41, f9 
	(p16) FMA	f10 = f97,  f49, f10
	(p16) FMA	f11 = f97,  f57, f11
	;;
	(p16) FMA	f8  = f98,  f34, f8
	(p16) FMA	f9  = f98,  f42, f9
	(p16) FMA	f10 = f98,  f50, f10
	(p16) FMA	f11 = f98,  f58, f11
	;;
	(p16) FMA	f8  = f99,  f35, f8 
	(p16) FMA	f9  = f99,  f43, f9 
	(p16) FMA	f10 = f99,  f51, f10
	(p16) FMA	f11 = f99,  f59, f11
	;;
	(p16) FMA	f8  = f100, f36, f8
	(p16) FMA	f9  = f100, f44, f9
	(p16) FMA	f10 = f100, f52, f10
	(p16) FMA	f11 = f100, f60, f11

	;;
	(p16) FMA	f8  = f101, f37, f8 
	(p16) FMA	f9  = f101, f45, f9 
	(p16) FMA	f10 = f101, f53, f10
	(p16) FMA	f11 = f101, f61, f11
	;;
	(p16) FMA	f8  = f102, f38, f8
	(p16) FMA	f9  = f102, f46, f9
	(p16) FMA	f10 = f102, f54, f10
	(p16) FMA	f11 = f102, f62, f11
	;;
	(p16) FMA	f8  = f103, f39, f8 
	(p16) FMA	f9  = f103, f47, f9 
	(p16) FMA	f10 = f103, f55, f10
	(p16) FMA	f11 = f103, f63, f11
	br.ctop.sptk.few .L22
	;;
	.align 16

.L25:
	tbit.nz	p13, p11 = M, 2
	tbit.nz	p14, p12 = M, 1
	;;
	{ .mmi
	(p11) adds	AO5  = - 4 * SIZE, AO5
	}
	{ .mbb
	(p11) adds	AO7  = - 4 * SIZE, AO7
	}
	;;
	{ .mmi
	(p13) LDFD	f32  = [AO1], LDA
	(p13) LDFD	f34  = [AO3], LDA
	tbit.nz	p15, p0  = M, 0
	}
	{ .mmi
	(p14) LDFD	f36  = [AO5], LDA
	(p11) adds	AO6  = - 4 * SIZE, AO6
	(p12) adds	AO7  = - 2 * SIZE, AO7
	}
	;;
	(p13) LDFD	f33  = [AO2], LDA
	(p13) LDFD	f35  = [AO4], LDA
	(p14) LDFD	f37  = [AO6], LDA
	(p15) LDFD	f38  = [AO7], LDA
	;;
	(p13) LDFD	f40  = [AO1], LDA
	(p13) LDFD	f42  = [AO3], LDA
	(p14) LDFD	f44  = [AO5], LDA
	(p15) LDFD	f46  = [AO7], LDA
	;;
	(p13) LDFD	f41  = [AO2], LDA
	(p13) LDFD	f43  = [AO4], LDA
	(p14) LDFD	f45  = [AO6], LDA
	;;
	(p13) LDFD	f48  = [AO1], LDA
	(p13) LDFD	f50  = [AO3], LDA
	(p14) LDFD	f52  = [AO5], LDA
	(p15) LDFD	f54  = [AO7], LDA
	;;
	(p13) LDFD	f49  = [AO2], LDA
	(p13) LDFD	f51  = [AO4], LDA
	(p14) LDFD	f53  = [AO6], LDA
	;;
	(p13) LDFD	f56  = [AO1]
	(p13) LDFD	f58  = [AO3]
	(p14) LDFD	f60  = [AO5]
	(p15) LDFD	f62  = [AO7]
	;;
	(p13) LDFD	f57  = [AO2]
	(p13) LDFD	f59  = [AO4]
	(p14) LDFD	f61  = [AO6]
	;;
	(p13) LDFD	f96  = [X1], 1 * SIZE
	(p13) LDFD	f98  = [X2], 1 * SIZE
	;;
	(p13) LDFD	f97  = [X1], 3 * SIZE
	(p13) LDFD	f99  = [X2], 3 * SIZE
	;;
	(p14) LDFD	f100 = [X1], 1 * SIZE
	;;
	(p14) LDFD	f101 = [X1], 1 * SIZE
	;;
	(p15) LDFD	f102 = [X1], 1 * SIZE
	;;
	(p13) FMA	f8  = f96,  f32, f8
	(p13) FMA	f9  = f96,  f40, f9
	(p13) FMA	f10 = f96,  f48, f10
	(p13) FMA	f11 = f96,  f56, f11
	;;
	(p13) FMA	f8  = f97,  f33, f8 
	(p13) FMA	f9  = f97,  f41, f9 
	(p13) FMA	f10 = f97,  f49, f10
	(p13) FMA	f11 = f97,  f57, f11
	;;
	(p13) FMA	f8  = f98,  f34, f8
	(p13) FMA	f9  = f98,  f42, f9
	(p13) FMA	f10 = f98,  f50, f10
	(p13) FMA	f11 = f98,  f58, f11
	;;
	(p13) FMA	f8  = f99,  f35, f8 
	(p13) FMA	f9  = f99,  f43, f9 
	(p13) FMA	f10 = f99,  f51, f10
	(p13) FMA	f11 = f99,  f59, f11
	;;
	(p14) FMA	f8  = f100, f36, f8
	(p14) FMA	f9  = f100, f44, f9
	(p14) FMA	f10 = f100, f52, f10
	(p14) FMA	f11 = f100, f60, f11
	;;
	(p14) FMA	f8  = f101, f37, f8 
	(p14) FMA	f9  = f101, f45, f9 
	(p14) FMA	f10 = f101, f53, f10
	(p14) FMA	f11 = f101, f61, f11
	;;
	(p15) FMA	f8  = f102, f38, f8
	(p15) FMA	f9  = f102, f46, f9
	(p15) FMA	f10 = f102, f54, f10
	(p15) FMA	f11 = f102, f62, f11
	;;
	LDFD	f32 = [Y1], INCY
	;;
	LDFD	f33 = [Y1], INCY
	;;
	LDFD	f34 = [Y1], INCY
	;;
	LDFD	f35 = [Y1], INCY
	;;
	FMA	f32 = ALPHA, f8,  f32
	FMA	f33 = ALPHA, f9,  f33
	FMA	f34 = ALPHA, f10, f34
	FMA	f35 = ALPHA, f11, f35
	;;
	STFD [YY1] = f32
	add	YY1 = YY1, INCY
	;;
	STFD [YY1] = f33
	add	YY1 = YY1, INCY
	;;
	STFD [YY1] = f34
	add	YY1 = YY1, INCY
	;;
	STFD [YY1] = f35
	add	YY1 = YY1, INCY
	;;
	.align 16

.L30:
	tbit.z	p6, p0  = N, 1
	;;
	(p6) br.cond.dpnt .L40
	;;
	mov	AO1 = A
	adds	AO2 = 1 * SIZE, A
	adds	AO3 = 2 * SIZE, A
	adds	AO4 = 3 * SIZE, A
	adds	AO5 = 4 * SIZE, A
	adds	AO6 = 5 * SIZE, A
	adds	AO7 = 6 * SIZE, A
	adds	AO8 = 7 * SIZE, A
	shladd	A   = LDA, 1, A
	;;
	shladd	LDA7M8 = LDA, 1, r0
	;;
	sub	LDA7M8 = LDA, LDA7M8	
	;;
	adds	LDA7M8 = 8 * SIZE, LDA7M8
	;;	
	mov	f8  = f0
	mov	f9  = f0
	mov	f10 = f0
	mov	f11 = f0
	mov	f12 = f0
	mov	f13 = f0
	mov	f14 = f0
	mov	f15 = f0

	mov	pr.rot= 0
	shr	I = M, 3
	mov	ar.ec = 2
	;;
	mov	X1  = BUFFER
	adds	X2  = 2 * SIZE, BUFFER
	;;
	cmp.eq	p16, p0 = r0, r0
	;;
	adds	I = -1, I
	;;
	mov	ar.lc = I
	cmp.eq	p6, p0 = -1, I
	(p6) br.cond.dpnt .L35
	;;
	.align 16

.L32:
	(p16) LDFD	f32  = [AO1], LDA
	(p16) LDFD	f34  = [AO3], LDA
	(p16) LDFD	f36  = [AO5], LDA
	(p16) LDFD	f38  = [AO7], LDA
	;;
	(p16) LDFD	f33  = [AO2], LDA
	(p16) LDFD	f35  = [AO4], LDA
	(p16) LDFD	f37  = [AO6], LDA
	(p16) LDFD	f39  = [AO8], LDA
	;;
	(p16) LDFD	f40  = [AO1], LDA7M8
	(p16) LDFD	f42  = [AO3], LDA7M8
	(p16) LDFD	f44  = [AO5], LDA7M8
	(p16) LDFD	f46  = [AO7], LDA7M8
	;;
	(p16) LDFD	f41  = [AO2], LDA7M8
	(p16) LDFD	f43  = [AO4], LDA7M8
	(p16) LDFD	f45  = [AO6], LDA7M8
	(p16) LDFD	f47  = [AO8], LDA7M8
	;;
	(p16) LDFD	f96  = [X1], 1 * SIZE
	(p16) LDFD	f98  = [X2], 1 * SIZE
	;;
	(p16) LDFD	f97  = [X1], 3 * SIZE
	(p16) LDFD	f99  = [X2], 3 * SIZE
	;;
	(p16) LDFD	f100 = [X1], 1 * SIZE
	(p16) LDFD	f102 = [X2], 1 * SIZE
	;;
	(p16) LDFD	f101 = [X1], 3 * SIZE
	(p16) LDFD	f103 = [X2], 3 * SIZE
	;;
	(p16) FMA	f8  = f96,  f32, f8
	(p16) FMA	f9  = f96,  f40, f9
	;;
	(p16) FMA	f8  = f97,  f33, f8 
	(p16) FMA	f9  = f97,  f41, f9 
	;;
	(p16) FMA	f8  = f98,  f34, f8
	(p16) FMA	f9  = f98,  f42, f9
	;;
	(p16) FMA	f8  = f99,  f35, f8 
	(p16) FMA	f9  = f99,  f43, f9 
	;;
	(p16) FMA	f8  = f100, f36, f8
	(p16) FMA	f9  = f100, f44, f9
	;;
	(p16) FMA	f8  = f101, f37, f8 
	(p16) FMA	f9  = f101, f45, f9 
	;;
	(p16) FMA	f8  = f102, f38, f8
	(p16) FMA	f9  = f102, f46, f9
	;;
	(p16) FMA	f8  = f103, f39, f8 
	(p16) FMA	f9  = f103, f47, f9 
	br.ctop.sptk.few .L32
	;;
	.align 16

.L35:
	tbit.nz	p13, p11 = M, 2
	tbit.nz	p14, p12 = M, 1
	;;
	{ .mmi
	(p11) adds	AO5  = - 4 * SIZE, AO5
	}
	{ .mbb
	(p11) adds	AO7  = - 4 * SIZE, AO7
	}
	;;
	{ .mmi
	(p13) LDFD	f32  = [AO1], LDA
	(p13) LDFD	f34  = [AO3], LDA
	tbit.nz	p15, p0  = M, 0
	}
	{ .mmi
	(p14) LDFD	f36  = [AO5], LDA
	(p11) adds	AO6  = - 4 * SIZE, AO6
	(p12) adds	AO7  = - 2 * SIZE, AO7
	}
	;;
	(p13) LDFD	f33  = [AO2], LDA
	(p13) LDFD	f35  = [AO4], LDA
	(p14) LDFD	f37  = [AO6], LDA
	(p15) LDFD	f38  = [AO7], LDA
	;;
	(p13) LDFD	f40  = [AO1]
	(p13) LDFD	f42  = [AO3]
	(p14) LDFD	f44  = [AO5]
	(p15) LDFD	f46  = [AO7]
	;;
	(p13) LDFD	f41  = [AO2]
	(p13) LDFD	f43  = [AO4]
	(p14) LDFD	f45  = [AO6]
	;;
	(p13) LDFD	f96  = [X1], 1 * SIZE
	(p13) LDFD	f98  = [X2], 1 * SIZE
	;;
	(p13) LDFD	f97  = [X1], 3 * SIZE
	(p13) LDFD	f99  = [X2], 3 * SIZE
	;;
	(p14) LDFD	f100 = [X1], 1 * SIZE
	;;
	(p14) LDFD	f101 = [X1], 1 * SIZE
	;;
	(p15) LDFD	f102 = [X1], 1 * SIZE
	;;
	(p13) FMA	f8  = f96,  f32, f8
	(p13) FMA	f9  = f96,  f40, f9
	;;
	(p13) FMA	f8  = f97,  f33, f8 
	(p13) FMA	f9  = f97,  f41, f9 
	;;
	(p13) FMA	f8  = f98,  f34, f8
	(p13) FMA	f9  = f98,  f42, f9
	;;
	(p13) FMA	f8  = f99,  f35, f8 
	(p13) FMA	f9  = f99,  f43, f9 
	;;
	(p14) FMA	f8  = f100, f36, f8
	(p14) FMA	f9  = f100, f44, f9
	;;
	(p14) FMA	f8  = f101, f37, f8 
	(p14) FMA	f9  = f101, f45, f9 
	;;
	(p15) FMA	f8  = f102, f38, f8
	(p15) FMA	f9  = f102, f46, f9
	;;
	LDFD	f32 = [Y1], INCY
	;;
	LDFD	f33 = [Y1], INCY
	;;
	FMA	f32 = ALPHA, f8,  f32
	FMA	f33 = ALPHA, f9,  f33
	;;
	STFD [YY1] = f32
	add	YY1 = YY1, INCY
	;;
	STFD [YY1] = f33
	add	YY1 = YY1, INCY
	;;
	.align 16

.L40:
	tbit.z	p6, p0  = N, 0
	;;
	(p6) br.cond.dpnt .L999
	;;
	mov	AO1 = A
	adds	AO2 = 1 * SIZE, A
	adds	AO3 = 2 * SIZE, A
	adds	AO4 = 3 * SIZE, A
	adds	AO5 = 4 * SIZE, A
	adds	AO6 = 5 * SIZE, A
	adds	AO7 = 6 * SIZE, A
	adds	AO8 = 7 * SIZE, A
	add	A   = LDA, A
	;;
	mov	f8  = f0
	mov	f9  = f0
	mov	f10 = f0
	mov	f11 = f0
	mov	f12 = f0
	mov	f13 = f0
	mov	f14 = f0
	mov	f15 = f0

	mov	pr.rot= 0
	shr	I = M, 3
	mov	ar.ec = 2
	;;
	mov	X1  = BUFFER
	adds	X2  = 2 * SIZE, BUFFER
	;;
	cmp.eq	p16, p0 = r0, r0
	;;
	adds	I = -1, I
	;;
	mov	ar.lc = I
	cmp.eq	p6, p0 = -1, I
	(p6) br.cond.dpnt .L45
	;;
	.align 16

.L42:
	(p16) LDFD	f32  = [AO1], 8 * SIZE
	(p16) LDFD	f34  = [AO3], 8 * SIZE
	(p16) LDFD	f36  = [AO5], 8 * SIZE
	(p16) LDFD	f38  = [AO7], 8 * SIZE
	;;
	(p16) LDFD	f33  = [AO2], 8 * SIZE
	(p16) LDFD	f35  = [AO4], 8 * SIZE
	(p16) LDFD	f37  = [AO6], 8 * SIZE
	(p16) LDFD	f39  = [AO8], 8 * SIZE
	;;
	(p16) LDFD	f96  = [X1], 1 * SIZE
	(p16) LDFD	f98  = [X2], 1 * SIZE
	;;
	(p16) LDFD	f97  = [X1], 3 * SIZE
	(p16) LDFD	f99  = [X2], 3 * SIZE
	;;
	(p16) LDFD	f100 = [X1], 1 * SIZE
	(p16) LDFD	f102 = [X2], 1 * SIZE
	;;
	(p16) LDFD	f101 = [X1], 3 * SIZE
	(p16) LDFD	f103 = [X2], 3 * SIZE
	;;
	(p16) FMA	f8  = f96,  f32, f8
	;;
	(p16) FMA	f8  = f97,  f33, f8 
	;;
	(p16) FMA	f8  = f98,  f34, f8
	;;
	(p16) FMA	f8  = f99,  f35, f8 
	;;
	(p16) FMA	f8  = f100, f36, f8
	;;
	(p16) FMA	f8  = f101, f37, f8 
	;;
	(p16) FMA	f8  = f102, f38, f8
	;;
	(p16) FMA	f8  = f103, f39, f8 
	br.ctop.sptk.few .L42
	;;
	.align 16

.L45:
	tbit.nz	p13, p11 = M, 2
	tbit.nz	p14, p12 = M, 1
	;;
	{ .mmi
	(p11) adds	AO5  = - 4 * SIZE, AO5
	}
	{ .mbb
	(p11) adds	AO7  = - 4 * SIZE, AO7
	}
	;;
	{ .mmi
	(p13) LDFD	f32  = [AO1]
	(p13) LDFD	f34  = [AO3]
	tbit.nz	p15, p0  = M, 0
	}
	{ .mmi
	(p14) LDFD	f36  = [AO5]
	(p11) adds	AO6  = - 4 * SIZE, AO6
	(p12) adds	AO7  = - 2 * SIZE, AO7
	}
	;;
	(p13) LDFD	f33  = [AO2]
	(p13) LDFD	f35  = [AO4]
	(p14) LDFD	f37  = [AO6]
	(p15) LDFD	f38  = [AO7]
	;;
	(p13) LDFD	f96  = [X1], 1 * SIZE
	(p13) LDFD	f98  = [X2], 1 * SIZE
	;;
	(p13) LDFD	f97  = [X1], 3 * SIZE
	(p13) LDFD	f99  = [X2], 3 * SIZE
	;;
	(p14) LDFD	f100 = [X1], 1 * SIZE
	;;
	(p14) LDFD	f101 = [X1], 1 * SIZE
	;;
	(p15) LDFD	f102 = [X1], 1 * SIZE
	;;
	(p13) FMA	f8  = f96,  f32, f8
	;;
	(p13) FMA	f8  = f97,  f33, f8 
	;;
	(p13) FMA	f8  = f98,  f34, f8
	;;
	(p13) FMA	f8  = f99,  f35, f8 
	;;
	(p14) FMA	f8  = f100, f36, f8
	;;
	(p14) FMA	f8  = f101, f37, f8 
	;;
	(p15) FMA	f8  = f102, f38, f8
	;;
	LDFD	f32 = [Y1], INCY
	;;
	FMA	f32 = ALPHA, f8,  f32
	;;
	STFD [YY1] = f32
	.align 16

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