qgemv_n.s

Optimized GotoBLAS libraries

	sub	LDA7M8 = LDA, LDA7M8	
	;;
	adds	LDA7M8 = 8 * SIZE, LDA7M8
	;;	
	mov	YLD1 = YY
	mov	YST1 = YY
	adds	YLD2 = 2 * SIZE, YY
	adds	YST2 = 2 * SIZE, YY
	;;
	LDFD	ALPHA1 = [X], INCX
	;;
	LDFD	ALPHA2 = [X], INCX
	;;
	FMPY	ALPHA1 = ALPHA, ALPHA1
	FMPY	ALPHA2 = ALPHA, ALPHA2
	;;
	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
	;;
	shr	I = M, 3
	mov	pr.rot= 0
	;;
	cmp.eq	p16, p0 = r0, r0
	;;
	adds	I = -1, I
	adds J = -1, J
	;;
	cmp.lt p7, p8 = r0, J
	tbit.nz	p13, p11 = M, 2
	tbit.nz	p14, p12 = M, 1
	mov	ar.ec= 1
	;;
	{ .mfi
	and	II = 7, M
	mov	ar.lc = I
	}
	{ .mfb
	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  = [YLD1], 1 * SIZE
	(p16) LDFD	f98  = [YLD2], 1 * SIZE
	;;
	(p16) LDFD	f97  = [YLD1], 3 * SIZE
	(p16) LDFD	f99  = [YLD2], 3 * SIZE
	;;
	(p16) LDFD	f100 = [YLD1], 1 * SIZE
	(p16) LDFD	f102 = [YLD2], 1 * SIZE
	;;
	(p16) LDFD	f101 = [YLD1], 3 * SIZE
	(p16) LDFD	f103 = [YLD2], 3 * SIZE
	;;
	(p16) FMA	f96  = ALPHA1,  f32, f96
	(p16) FMA	f98  = ALPHA1,  f34, f98
	(p16) FMA	f97  = ALPHA1,  f33, f97
	(p16) FMA	f99  = ALPHA1,  f35, f99
	(p16) FMA	f100 = ALPHA1,  f36, f100
	(p16) FMA	f102 = ALPHA1,  f38, f102
	(p16) FMA	f101 = ALPHA1,  f37, f101
	(p16) FMA	f103 = ALPHA1,  f39, f103
	;;
	(p16) FMA	f16  = ALPHA2,  f40, f96
	(p16) FMA	f18  = ALPHA2,  f42, f98
	(p16) FMA	f17  = ALPHA2,  f41, f97
	(p16) FMA	f19  = ALPHA2,  f43, f99
	(p16) FMA	f20  = ALPHA2,  f44, f100
	(p16) FMA	f22  = ALPHA2,  f46, f102
	(p16) FMA	f21  = ALPHA2,  f45, f101
	(p16) FMA	f23  = ALPHA2,  f47, f103
	;;
	(p16) STFD	[YST1] = f16, 1 * SIZE
	(p16) STFD	[YST2] = f18, 1 * SIZE
	;;
	(p16) STFD	[YST1] = f17, 3 * SIZE
	(p16) STFD	[YST2] = f19, 3 * SIZE
	;;
	(p16) STFD	[YST1] = f20, 1 * SIZE
	(p16) STFD	[YST2] = f22, 1 * SIZE
	;;
	(p16) STFD	[YST1] = f21, 3 * SIZE
	(p16) STFD	[YST2] = f23, 3 * SIZE
	br.ctop.sptk.few .L32
	;;
	.align 16

.L35:
	{ .mmi
	(p8) cmp.eq.unc p10, p0 = r0, II
	(p11) adds	AO5  = - 4 * SIZE, AO5
	}
	{ .mbb
	(p11) adds	AO7  = - 4 * SIZE, AO7
	(p10) br.cond.dptk  .L40
	}
	;;
	{ .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]
	(p13) LDFD	f43  = [AO4]
	(p14) LDFD	f45  = [AO6]
	;;
	(p13) LDFD	f96  = [YLD1], 1 * SIZE
	(p13) LDFD	f98  = [YLD2], 1 * SIZE
	;;
	(p13) LDFD	f97  = [YLD1], 3 * SIZE
	(p13) LDFD	f99  = [YLD2], 3 * SIZE
	;;
	(p14) LDFD	f100 = [YLD1], 1 * SIZE
	;;
	(p14) LDFD	f101 = [YLD1], 1 * SIZE
	;;
	(p15) LDFD	f102 = [YLD1], 1 * SIZE
	;;

	(p13) FMA	f96  = ALPHA1,  f32, f96
	(p13) FMA	f98  = ALPHA1,  f34, f98
	(p13) FMA	f97  = ALPHA1,  f33, f97
	(p13) FMA	f99  = ALPHA1,  f35, f99
	(p14) FMA	f100 = ALPHA1,  f36, f100
	(p15) FMA	f102 = ALPHA1,  f38, f102
	(p14) FMA	f101 = ALPHA1,  f37, f101
	;;
	(p13) FMA	f16  = ALPHA2,  f40, f96
	(p13) FMA	f18  = ALPHA2,  f42, f98
	(p13) FMA	f17  = ALPHA2,  f41, f97
	(p13) FMA	f19  = ALPHA2,  f43, f99
	(p14) FMA	f20  = ALPHA2,  f44, f100
	(p15) FMA	f22  = ALPHA2,  f46, f102
	(p14) FMA	f21  = ALPHA2,  f45, f101
	;;
	{ .mmi
	(p13) STFD	[YST1] = f16, 1 * SIZE
	(p13) STFD	[YST2] = f18, 1 * SIZE
	nop   __LINE__
	}
	;;
	{ .mmi
	(p13) STFD	[YST1] = f17, 3 * SIZE
	(p13) STFD	[YST2] = f19
	nop   __LINE__
	}
	;;
	{ .mmi
	(p14) STFD	[YST1] = f20, 1 * SIZE
	;;
	(p14) STFD	[YST1] = f21, 1 * SIZE
	nop   __LINE__
	}
	;;
	{ .mib
	(p15) STFD	[YST1] = f22
	}
	;;
	.align 16

.L40:
	tbit.z	p6, p0 = N, 0
	;;
	(p6) br.cond.dpnt .L990
	;;
	mov	LDA7M8 = 8 * SIZE
	;;	
	mov	YLD1 = YY
	mov	YST1 = YY
	adds	YLD2 = 2 * SIZE, YY
	adds	YST2 = 2 * SIZE, YY
	;;
	LDFD	ALPHA1 = [X], INCX
	;;
	LDFD	ALPHA2 = [X], INCX
	;;
	FMPY	ALPHA1 = ALPHA, ALPHA1
	FMPY	ALPHA2 = ALPHA, ALPHA2
	;;
	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
	;;
	shr	I = M, 3
	mov	pr.rot= 0
	;;
	cmp.eq	p16, p0 = r0, r0
	;;
	adds	I = -1, I
	adds J = -1, J
	;;
	cmp.lt p7, p8 = r0, J
	tbit.nz	p13, p11 = M, 2
	tbit.nz	p14, p12 = M, 1
	mov	ar.ec= 1
	;;
	{ .mfi
	and	II = 7, M
	mov	ar.lc = I
	}
	{ .mfb
	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  = [YLD1], 1 * SIZE
	(p16) LDFD	f98  = [YLD2], 1 * SIZE
	;;
	(p16) LDFD	f97  = [YLD1], 3 * SIZE
	(p16) LDFD	f99  = [YLD2], 3 * SIZE
	;;
	(p16) LDFD	f100 = [YLD1], 1 * SIZE
	(p16) LDFD	f102 = [YLD2], 1 * SIZE
	;;
	(p16) LDFD	f101 = [YLD1], 3 * SIZE
	(p16) LDFD	f103 = [YLD2], 3 * SIZE
	;;
	(p16) FMA	f16  = ALPHA1,  f32, f96
	(p16) FMA	f18  = ALPHA1,  f34, f98
	(p16) FMA	f17  = ALPHA1,  f33, f97
	(p16) FMA	f19  = ALPHA1,  f35, f99
	(p16) FMA	f20  = ALPHA1,  f36, f100
	(p16) FMA	f22  = ALPHA1,  f38, f102
	(p16) FMA	f21  = ALPHA1,  f37, f101
	(p16) FMA	f23 = ALPHA1,  f39, f103
	;;
	(p16) STFD	[YST1] = f16, 1 * SIZE
	(p16) STFD	[YST2] = f18, 1 * SIZE
	;;
	(p16) STFD	[YST1] = f17, 3 * SIZE
	(p16) STFD	[YST2] = f19, 3 * SIZE
	;;
	(p16) STFD	[YST1] = f20, 1 * SIZE
	(p16) STFD	[YST2] = f22, 1 * SIZE
	;;
	(p16) STFD	[YST1] = f21, 3 * SIZE
	(p16) STFD	[YST2] = f23, 3 * SIZE
	br.ctop.sptk.few .L42
	;;
	.align 16

.L45:
	{ .mmi
	(p8) cmp.eq.unc p10, p0 = r0, II
	(p11) adds	AO5  = - 4 * SIZE, AO5
	}
	{ .mbb
	(p11) adds	AO7  = - 4 * SIZE, AO7
	(p10) br.cond.dptk  .L990
	}
	;;
	{ .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	f96  = [YLD1], 1 * SIZE
	(p13) LDFD	f98  = [YLD2], 1 * SIZE
	;;
	(p13) LDFD	f97  = [YLD1], 3 * SIZE
	(p13) LDFD	f99  = [YLD2], 3 * SIZE
	;;
	(p14) LDFD	f100 = [YLD1], 1 * SIZE
	;;
	(p14) LDFD	f101 = [YLD1], 1 * SIZE
	;;
	(p15) LDFD	f102 = [YLD1], 1 * SIZE
	;;

	(p13) FMA	f16  = ALPHA1,  f32, f96
	(p13) FMA	f18  = ALPHA1,  f34, f98
	(p13) FMA	f17  = ALPHA1,  f33, f97
	(p13) FMA	f19  = ALPHA1,  f35, f99
	(p14) FMA	f20  = ALPHA1,  f36, f100
	(p15) FMA	f22  = ALPHA1,  f38, f102
	(p14) FMA	f21 = ALPHA1,  f37, f101
	;;
	{ .mmi
	(p13) STFD	[YST1] = f16, 1 * SIZE
	(p13) STFD	[YST2] = f18, 1 * SIZE
	nop   __LINE__
	}
	;;
	{ .mmi
	(p13) STFD	[YST1] = f17, 3 * SIZE
	(p13) STFD	[YST2] = f19
	nop   __LINE__
	}
	;;
	{ .mmi
	(p14) STFD	[YST1] = f20, 1 * SIZE
	;;
	(p14) STFD	[YST1] = f21, 1 * SIZE
	nop   __LINE__
	}
	;;
	{ .mib
	(p15) STFD	[YST1] = f22
	}
	;;
	.align 16


.L990:
	cmp.eq	p10, p0 = SIZE, INCY
	;;
	{ .mmi
	mov	YLD1 = YY
	mov	YST1 = Y
	mov	pr.rot= 0
	}
	{ .mib
	mov	YST2 = Y
	shr	J = M, 3
	(p10) br.cond.dptk .L999
	}
	;;
	{ .mmi
	cmp.eq	p6, p0 = r0, J
	adds	J = -1, J
	mov	ar.ec = 4
	}
	{ .mmi
	cmp.eq	p16, p0 = r0, r0
	nop	__LINE__
	tbit.nz	p13, p0 = M, 2
	}
	;;
	{ .mib
	nop	__LINE__
	mov	ar.lc = J
	(p6) br.cond.dpnt .L995
	}
	;;
.L992:
	{ .mfi
	(p19)	STFD	[YST2] = f35
	(p18)	FADD	f34 = f34, f66
	(p19)	add YST2 = YST2, INCY
	}
	{ .mmi
	(p16)	LDFD	f64 = [YLD1], 1 * SIZE
	(p16)	LDFD	f32 = [YST1], INCY
	}
	;;
	{ .mfi
	(p19)	STFD	[YST2] = f39
	(p18)	FADD	f38 = f38, f70
	(p19)	add YST2 = YST2, INCY
	}
	{ .mmi
	(p16)	LDFD	f36 = [YST1], INCY
	(p16)	LDFD	f68 = [YLD1], 1 * SIZE
	}
	;;
	{ .mfi
	(p19)	STFD	[YST2] = f43
	(p18)	FADD	f42 = f42, f74
	(p19)	add YST2 = YST2, INCY
	}
	{ .mmi
	(p16)	LDFD	f72 = [YLD1], 1 * SIZE
	(p16)	LDFD	f40 = [YST1], INCY
	}
	;;
	{ .mfi
	(p19)	STFD	[YST2] = f47
	(p18)	FADD	f46 = f46, f78
	(p19)	add YST2 = YST2, INCY
	}
	{ .mmi
	(p16)	LDFD	f76 = [YLD1], 1 * SIZE
	(p16)	LDFD	f44 = [YST1], INCY
	}
	;;
	{ .mfi
	(p19)	STFD	[YST2] = f51
	(p18)	FADD	f50 = f50, f82
	(p19)	add YST2 = YST2, INCY
	}
	{ .mmi
	(p16)	LDFD	f80 = [YLD1], 1 * SIZE
	(p16)	LDFD	f48 = [YST1], INCY
	}
	;;
	{ .mfi
	(p19)	STFD	[YST2] = f55
	(p18)	FADD	f54 = f54, f86
	(p19)	add YST2 = YST2, INCY
	}
	{ .mmi
	(p16)	LDFD	f84 = [YLD1], 1 * SIZE
	(p16)	LDFD	f52 = [YST1], INCY
	}
	;;
	{ .mfi
	(p19)	STFD	[YST2] = f59
	(p18)	FADD	f58 = f58, f90
	(p19)	add YST2 = YST2, INCY
	}
	{ .mmi
	(p16)	LDFD	f88 = [YLD1], 1 * SIZE
	(p16)	LDFD	f56 = [YST1], INCY
	}
	;;
	{ .mfi
	(p19)	STFD	[YST2] = f63
	(p18)	FADD	f62 = f62, f94
	(p19)	add YST2 = YST2, INCY
	}
	{ .mmb
	(p16)	LDFD	f92 = [YLD1], 1 * SIZE
	(p16)	LDFD	f60 = [YST1], INCY
	br.ctop.sptk.few .L992
	}
	;;

.L995:
	(p13)	LDFD	f32 = [YST1], INCY
	(p13)	LDFD	f40 = [YLD1], 1 * SIZE
	tbit.nz	p14, p0 = M, 1
	;;
	(p13)	LDFD	f33 = [YST1], INCY
	(p13)	LDFD	f41 = [YLD1], 1 * SIZE
	tbit.nz	p15, p0 = M, 0
	;;
	(p13)	LDFD	f34 = [YST1], INCY
	(p13)	LDFD	f42 = [YLD1], 1 * SIZE
	;;
	(p13)	LDFD	f35 = [YST1], INCY
	(p13)	LDFD	f43 = [YLD1], 1 * SIZE
	;;
	(p14)	LDFD	f36 = [YST1], INCY
	(p14)	LDFD	f44 = [YLD1], 1 * SIZE
	;;
	(p14)	LDFD	f37 = [YST1], INCY
	(p14)	LDFD	f45 = [YLD1], 1 * SIZE
	;;
	(p15)	LDFD	f38 = [YST1], INCY
	(p15)	LDFD	f46 = [YLD1], 1 * SIZE
	;;
	(p13)	FADD	f32 = f32, f40
	(p13)	FADD	f33 = f33, f41
	(p13)	FADD	f34 = f34, f42
	(p13)	FADD	f35 = f35, f43
	(p14)	FADD	f36 = f36, f44
	(p14)	FADD	f37 = f37, f45
	(p15)	FADD	f38 = f38, f46
	;;
	(p13)	STFD	[YST2] = f32
	(p13)	add YST2 = YST2, INCY
	;;
	(p13)	STFD	[YST2] = f33
	(p13)	add YST2 = YST2, INCY
	;;
	(p13)	STFD	[YST2] = f34
	(p13)	add YST2 = YST2, INCY
	;;
	(p13)	STFD	[YST2] = f35
	(p13)	add YST2 = YST2, INCY
	;;
	(p14)	STFD	[YST2] = f36
	(p14)	add YST2 = YST2, INCY
	;;
	(p14)	STFD	[YST2] = f37
	(p14)	add YST2 = YST2, INCY
	;;
	(p15)	STFD	[YST2] = f38
	;;

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