zgemm3m_kernel.s

Optimized GotoBLAS libraries

	(p5) LDFD	f11  = [C2], -SIZE
	FMA	f120  = f32, f55, f120	// A1 * B8
	}
	;;
	{ .mmf
	(p3) LDFPD	f62, f63 = [BOFFSET], 2 * SIZE
	(p4) LDFD	f32 = [AOFFSET],   1 * SIZE
	(p3) FMA	f64   = f40, f56, f64	// A1 * B1
	}
	{ .mmf
	(p5) LDFD	f12  = [C3], SIZE
	(p5) LDFD	f14  = [C4], SIZE
	(p3) FMA	f72   = f40, f57, f72	// A1 * B2
	}
	;;
	{ .mfi
	(p4) LDFPD	f48, f49 = [BOFFSET],   2 * SIZE
	(p3) FMA	f80   = f40, f58, f80	// A1 * B3
	nop	__LINE__
	}
	{ .mmf
	(p5) LDFD	f13  = [C3], -SIZE
	(p5) LDFD	f15  = [C4], -SIZE
	(p3) FMA	f88   = f40, f59, f88	// A1 * B4
	}
	;;
	{ .mfi
	(p4) LDFPD	f50, f51 = [BOFFSET],  2 * SIZE
	(p3) FMA	f96   = f40, f60, f96	// A1 * B5
	nop	__LINE__
	}
	{ .mmf
	(p5) LDFD	f16  = [C5], SIZE
	(p5) LDFD	f18  = [C6], SIZE
	(p3) FMA	f104  = f40, f61, f104	// A1 * B6
	}
	;;
	{ .mfi
	(p4) LDFPD	f52, f53 = [BOFFSET], 2 * SIZE
	(p3) FMA	f112  = f40, f62, f112	// A1 * B7
	adds	L = -1, L
	}
	{ .mmb
	(p5) LDFD	f17 = [C5], -SIZE
	(p5) LDFD	f19 = [C6], -SIZE
	nop	__LINE__
	}
	;;
	{ .mfb
	(p4) LDFPD	f54, f55 = [BOFFSET], 2 * SIZE
	(p3) FMA	f120  = f40, f63, f120	// A1 * B8
	nop	__LINE__
	}
	{ .mmb
	(p5) LDFD	f20 = [C7], SIZE
	(p5) LDFD	f22 = [C8], SIZE
	br.cloop.sptk.few .L042
	}
	;;
	{ .mmf
	LDFD	f21 = [C7], -SIZE
	LDFD	f23 = [C8], -SIZE
	FMA	f6  = ALPHA_R, f64, f6
	}
	{ .mmf
	nop	__LINE__
	nop	__LINE__
	FMA	f10 = ALPHA_R, f72, f10
	}
	;;
	{ .mmf
	nop	__LINE__
	nop	__LINE__
	FMA	f7  = ALPHA_I, f64, f7
	}
	{ .mmf
	nop	__LINE__
	nop	__LINE__
	FMA	f11 = ALPHA_I, f72, f11
	}
	;;
	{ .mmf
	nop	__LINE__
	nop	__LINE__
	FMA	f12 = ALPHA_R, f80, f12
	}
	{ .mmf
	nop	__LINE__
	nop	__LINE__
	FMA	f14 = ALPHA_R, f88, f14
	}
	;;
	{ .mmf
	nop	__LINE__
	nop	__LINE__
	FMA	f13 = ALPHA_I, f80, f13
	}
	{ .mmf
	nop	__LINE__
	nop	__LINE__
	FMA	f15 = ALPHA_I, f88, f15
	}
	;;
	{ .mmf
	STFD	[C1 ] = f6,  SIZE
	STFD	[C2 ] = f10, SIZE
	FMA	f16 = ALPHA_R, f96,  f16
	}
	{ .mmf
	nop	__LINE__
	nop	__LINE__
	FMA	f18 = ALPHA_R, f104, f18
	}
	;;
	{ .mmf
	STFD	[C1 ] = f7,  SIZE
	STFD	[C2 ] = f11, SIZE
	FMA	f17 = ALPHA_I, f96,  f17
	}
	{ .mmf
	nop	__LINE__
	nop	__LINE__
	FMA	f19 = ALPHA_I, f104, f19
	}
	;;
	{ .mmf
	STFD	[C3 ] = f12, SIZE
	STFD	[C4 ] = f14, SIZE
	FMA	f20 = ALPHA_R, f112, f20
	}
	{ .mmf
	nop	__LINE__
	nop	__LINE__
	FMA	f22 = ALPHA_R, f120, f22
	}
	;;
	{ .mmf
	STFD	[C3 ] = f13, SIZE
	STFD	[C4 ] = f15, SIZE
	FMA	f21 = ALPHA_I, f112, f21
	}
	{ .mmf
	nop	__LINE__
	nop	__LINE__
	FMA	f23 = ALPHA_I, f120, f23
	}
	;;
	{ .mmi
	STFD	[C5 ] = f16, SIZE
	STFD	[C6 ] = f18, SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	STFD	[C5 ] = f17, SIZE
	STFD	[C6 ] = f19, SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	STFD	[C7 ] = f20, SIZE
	STFD	[C8 ] = f22, SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	STFD	[C7 ] = f21, SIZE
	STFD	[C8 ] = f23, SIZE
	nop	__LINE__
	}
	;;
	.align 32

.L049:
	{ .mmi
	mov	B = BOFFSET
	mov	AOFFSET = A
	nop	__LINE__
	}
	;;
	{ .mmb
	nop	__LINE__
	cmp.lt	p6, p0 = 0, J
	(p6)	br.cond.dptk .L010
	}
	;;
	.align 32

.L050:
	{ .mfi
	mov	C1 = C
	mov	f64  = f0
	tbit.z	p6, p0 = N, 2
	}
	{ .mfi
	add	C2 = LDC, C
	mov	f72  = f0
	shr	I  = M, 3
	}
	;;
	{ .mfi
	shladd	C3 = LDC, 1, C
	mov	f80  = f0
	nop	__LINE__
	}
	{ .mfb
	mov	AOFFSET = A
	mov	f88  = f0
	(p6)	br.cond.dpnt .L090
	}
	;;
	{ .mfi
	cmp.eq	p6, p7 = 0, I
	mov	f65  = f0
	nop	__LINE__
	} 
	{ .mfi
	shladd	C4 = LDC, 1, C2
	mov	f73  = f0
	nop	__LINE__
	}
	;;
	{ .mfi
	nop	__LINE__
	mov	f81  = f0
	nop	__LINE__
	}
	{ .mfb
	shladd	C = LDC, 2, C
	mov	f89  = f0
	(p6)	br.cond.dpnt .L060
	}
	;;
	.align 32

.L052:
	{ .mfb
	LDFPD	f48, f49 = [B]
	mov	f66  = f0
	nop	__LINE__
	}
	{ .mfb
	adds	BOFFSET = 2 * SIZE, B
	mov	f74  = f0
	nop	__LINE__
	}
	;;
	{ .mfi
	LDFPD	f32, f33 = [AOFFSET], 2 * SIZE
	mov	f82  = f0
	nop	__LINE__
	}
	{ .mfi
	setf.d	f84  = r0
	mov	f90  = f0
	nop	__LINE__
	}
	;;
	{ .mfi
	LDFPD	f50, f51 = [BOFFSET], 2 * SIZE
	mov	f67  = f0
	adds	PREC = CPREFETCHSIZE * SIZE, C1
	}
	{ .mfi
	LDFPD	f34, f35 = [AOFFSET], 2 * SIZE
	mov	f75  = f0
	adds	L =  1, K
	}
	;;
	{ .mfi
	LDFPD	f36, f37 = [AOFFSET], 2 * SIZE
	mov	f83  = f0
	tbit.z	p12, p0 = L, 0
	}
	{ .mfi
	setf.d	f91  = r0
	mov	f68  = f0
	adds	PREB = (PREFETCHSIZE + 0) * SIZE, BOFFSET
	}
	;;
	{ .mfi
	CPREFETCH [PREC], LDC
	mov	f76  = f0
	adds	PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
	}
	{ .mfi
	LDFPD	f38, f39 = [AOFFSET], 2 * SIZE
	mov	f92  = f0
	cmp.eq	p3, p0 = r0, r0
	}
	;;
	{ .mfi
	CPREFETCH [PREC], LDC
	mov	f69  = f0
	shr	L = L, 1
	}
	{ .mmf
	setf.d	f77  = r0
	setf.d	f85  = r0
	mov	f93  = f0
	}
	;;
	{ .mfi
	CPREFETCH [PREC], LDC
	mov	f70  = f0
	adds	L =  -1, L
	}
	{ .mmf
	setf.d	f78  = r0
	setf.d	f86  = r0
	mov	f94  = f0
	}
	;;
	{ .mfi
	CPREFETCH [PREC]
	mov	f71  = f0
	mov	ar.lc = L
	}
	{ .mmf
	setf.d	f79  = r0
	setf.d	f87  = r0
	mov	f95  = f0
	}
	;;
	.align 32

.L053:
	{ .mfb
	lfetch.nt1	[PREA],  16 * SIZE
	FMA	f64   = f32, f48, f64	// A1 * B1
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FMA	f72   = f32, f49, f72	// A1 * B2
	(p12) cmp.ne p3, p0 =  0, L
	}
	;;
	{ .mfi
	lfetch.nt1	[PREB],   8 * SIZE
	FMA	f80   = f32, f50, f80	// A1 * B3
	cmp.ne	p4, p5 =  0, L
	}
	{ .mfi
	nop	__LINE__
	FMA	f88   = f32, f51, f88	// A1 * B4
	adds	C9  = 4 * SIZE, C1
	}
	;;
	{ .mfi
	(p3) LDFPD	f40, f41 = [AOFFSET], 2 * SIZE
	FMA	f65   = f33, f48, f65	// A2 * B1
	adds	C10 = 4 * SIZE, C2
	}
	{ .mfi
	nop	__LINE__
	FMA	f73   = f33, f49, f73	// A2 * B2
	adds	C11 = 4 * SIZE, C3
	}
	;;
	{ .mfi
	(p3) LDFPD	f56, f57 = [BOFFSET],  2 * SIZE
	FMA	f81   = f33, f50, f81	// A2 * B3
	adds	C12 = 4 * SIZE, C4
	}
	{ .mfb
	nop	__LINE__
	FMA	f89   = f33, f51, f89	// A2 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	(p3) LDFPD	f58, f59 = [BOFFSET],  2 * SIZE
	FMA	f66   = f34, f48, f66	// A3 * B1
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA	f74   = f34, f49, f74	// A3 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	(p3) LDFPD	f42, f43 = [AOFFSET], 2 * SIZE
	FMA	f82   = f34, f50, f82	// A3 * B3
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA	f90   = f34, f51, f90	// A3 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	(p3) LDFPD	f44, f45 = [AOFFSET], 2 * SIZE
	FMA	f67   = f35, f48, f67	// A4 * B1
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA	f75   = f35, f49, f75	// A4 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	(p3) LDFPD	f46, f47 = [AOFFSET], 2 * SIZE
	FMA	f83   = f35, f50, f83	// A4 * B3
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA	f91   = f35, f51, f91	// A4 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	FMA	f68   = f36, f48, f68	// A5 * B1
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA	f76   = f36, f49, f76	// A5 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	FMA	f84   = f36, f50, f84	// A5 * B3
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA	f92   = f36, f51, f92	// A5 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	FMA	f69   = f37, f48, f69	// A6 * B1
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA	f77   = f37, f49, f77	// A6 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	FMA	f85   = f37, f50, f85	// A6 * B3
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA	f93   = f37, f51, f93	// A6 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	FMA	f70   = f38, f48, f70	// A7 * B1
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA	f78   = f38, f49, f78	// A7 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	FMA	f86   = f38, f50, f86	// A7 * B3
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA	f94   = f38, f51, f94	// A7 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	(p4) LDFPD	f32, f33 = [AOFFSET],   2 * SIZE
	FMA	f71   = f39, f48, f71	// A8 * B1
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA	f79   = f39, f49, f79	// A8 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	(p4) LDFPD	f48, f49 = [BOFFSET],  2 * SIZE
	FMA	f87   = f39, f50, f87	// A8 * B3
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA	f95   = f39, f51, f95	// A8 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	(p4) LDFPD	f50, f51 = [BOFFSET],  2 * SIZE
	(p3) FMA	f64   = f40, f56, f64	// A1 * B1
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA	f72   = f40, f57, f72	// A1 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	(p4) LDFPD	f34, f35 = [AOFFSET], 2 * SIZE
	(p3) FMA	f80   = f40, f58, f80	// A1 * B3
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA	f88   = f40, f59, f88	// A1 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	(p4) LDFPD	f36, f37 = [AOFFSET], 2 * SIZE
	(p3) FMA	f65   = f41, f56, f65	// A2 * B1
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA	f73   = f41, f57, f73	// A2 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	(p4) LDFPD	f38, f39 = [AOFFSET], 2 * SIZE
	(p3) FMA	f81   = f41, f58, f81	// A2 * B3
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA	f89   = f41, f59, f89	// A2 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	(p5) LDFD	f6  = [C1 ], SIZE
	(p3) FMA	f66   = f42, f56, f66	// A3 * B1
	nop	__LINE__
	}
	{ .mfb
	(p5) LDFD	f7  = [C9 ], SIZE
	(p3) FMA	f74   = f42, f57, f74	// A3 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	(p5) LDFD	f10 = [C1 ], SIZE
	(p3) FMA	f82   = f42, f58, f82	// A3 * B3
	nop	__LINE__
	}
	{ .mfb
	(p5) LDFD	f11 = [C9 ], SIZE
	(p3) FMA	f90   = f42, f59, f90	// A3 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	(p5) LDFD	f12 = [C1 ], SIZE
	(p3) FMA	f67   = f43, f56, f67	// A4 * B1
	nop	__LINE__
	}
	{ .mfb
	(p5) LDFD	f13 = [C9 ], SIZE
	(p3) FMA	f75   = f43, f57, f75	// A4 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	(p5) LDFD	f14 = [C1 ], 5 * SIZE
	(p3) FMA	f83   = f43, f58, f83	// A4 * B3
	nop	__LINE__
	}
	{ .mfb
	(p5) LDFD	f15 = [C9 ], 5 * SIZE
	(p3) FMA	f91   = f43, f59, f91	// A4 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	(p5) LDFD	f16 = [C1 ], SIZE
	(p3) FMA	f68   = f44, f56, f68	// A5 * B1
	nop	__LINE__
	}
	{ .mfb
	(p5) LDFD	f17 = [C9], SIZE
	(p3) FMA	f76   = f44, f57, f76	// A5 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	(p5) LDFD	f18 = [C1 ], SIZE
	(p3) FMA	f84   = f44, f58, f84	// A5 * B3
	nop	__LINE__
	}
	{ .mfb
	(p5) LDFD	f19 = [C9], SIZE
	(p3) FMA	f92   = f44, f59, f92	// A5 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	(p5) LDFD	f20 = [C1 ], SIZE
	(p3) FMA	f69   = f45, f56, f69	// A6 * B1
	nop	__LINE__
	}
	{ .mfb
	(p5) LDFD	f21 = [C9], SIZE
	(p3) FMA	f77   = f45, f57, f77	// A6 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	(p5) LDFD	f22 = [C1 ], -11 * SIZE
	(p3) FMA	f85   = f45, f58, f85	// A6 * B3
	nop	__LINE__
	}
	{ .mfb
	(p5) LDFD	f23 = [C9 ], -11 * SIZE
	(p3) FMA	f93   = f45, f59, f93	// A6 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	(p5) LDFD	f24 = [C2 ], SIZE
	(p3) FMA	f70   = f46, f56, f70	// A7 * B1
	nop	__LINE__
	}
	{ .mfb
	(p5) LDFD	f25 = [C10], SIZE
	(p3) FMA	f78   = f46, f57, f78	// A7 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	(p5) LDFD	f26 = [C2 ], SIZE
	(p3) FMA	f86   = f46, f58, f86	// A7 * B3
	nop	__LINE__
	}
	{ .mfb
	(p5) LDFD	f27 = [C10], SIZE
	(p3) FMA	f94   = f46, f59, f94	// A7 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	(p5) LDFD	f28 = [C2 ], SIZE
	(p3) FMA	f71   = f47, f56, f71	// A8 * B1
	nop	__LINE__
	}
	{ .mfb
	(p5) LDFD	f29 = [C10], SIZE
	(p3) FMA	f79   = f47, f57, f79	// A8 * B2
	nop	__LINE__
	}
	;;
	{ .mfi
	(p5) LDFD	f30 = [C2 ], 5 * SIZE
	(p3) FMA	f87   = f47, f58, f87	// A8 * B3
	adds	L = -1, L
	}
	{ .mfb
	(p5) LDFD	f31 = [C10], 5 * SIZE
	(p3) FMA	f95   = f47, f59, f95	// A8 * B4
	br.cloop.sptk.few .L053
	}
	;;
	.align 32

.L058:
	{ .mmf
	LDFD	f32 = [C2 ], SIZE
	LDFD	f33 = [C10], SIZE
	FMA	f6   = ALPHA_R, f64, f6
	}
	{ .mmf
	nop	__LINE__
	nop	__LINE__
	FMA	f7   = ALPHA_R, f66, f7
	}
	;;
	{ .mmf
	LDFD	f34 = [C2 ], SIZE
	LDFD	f35 = [C10], SIZE
	FMA	f10  = ALPHA_I, f64, f10
	}
	{ .mmf
	nop	__LINE__
	nop	__LINE__
	FMA	f11  = ALPHA_I, f66, f11
	}
	;;
	{ .mmf
	LDFD	f36 = [C2 ], SIZE
	LDFD	f37 = [C10], SIZE
	FMA	f12  = ALPHA_R, f65, f12
	}
	{ .mmf
	nop	__LINE__
	nop	__LINE__
	FMA	f13  = ALPHA_R, f67, f13
	}
	;;
	{ .mmf
	LDFD	f38 = [C2 ], - 11 * SIZE
	LDFD	f39 = [C10], - 11 * SIZE
	FMA	f14  = ALPHA_I, f65, f14
	}
	{ .mmf
	nop	__LINE__
	nop	__LINE__
	FMA	f15  = ALPHA_I, f67, f15
	}
	;;
	{ .mmf
	STFD	[C1 ] = f6, SIZE
	STFD	[C9 ] = f7, SIZE
	FMA	f16  = ALPHA_R, f68, f16
	}
	{ .mmf
	LDFD	f48 = [C3 ], SIZE
	LDFD	f49 = [C11], SIZE
	FMA	f17  = ALPHA_R, f70, f17
	}
	;;
	{ .mmf
	STFD	[C1 ] = f10, SIZE
	STFD	[C9 ] = f11, SIZE
	FMA	f18  = ALPHA_I, f68, f18
	}
	{ .mmf
	LDFD	f50 = [C3 ], SIZE
	LDFD	f51 = [C11], SIZE
	FMA	f19  = ALPHA_I, f70, f19
	}
	;;
	{ .mmf
	STFD	[C1 ] = f12, SIZE
	STFD	[C9 ] = f13, SIZE
	FMA	f20  = ALPHA_R, f69, f20
	}
	{ .mmf
	LDFD	f52 = [C3 ], SIZE
	LDFD	f53 = [C11], SIZE
	FMA	f21  = ALPHA_R, f71, f21
	}
	;;
	{ .mmf
	STFD	[C1 ] = f14, 5 * SIZE
	STFD	[C9 ] = f15, 5 * SIZE
	FMA	f22  = ALPHA_I, f69, f22
	}
	{ .mmf
	LDFD	f54 = [C3 ], 5 * SIZE
	LDFD	f55 = [C11], 5 * SIZE
	FMA	f23  = ALPHA_I, f71, f23
	}
	;;
	{ .mmf
	STFD	[C1 ] = f16, SIZE
	STFD	[C9 ] = f17, SIZE
	FMA	f24  = ALPHA_R, f72, f24
	}
	{ .mmf
	LDFD	f40 = [C3 ], SIZE
	LDFD	f41 = [C11], SIZE
	FMA	f25  = ALPHA_R, f74, f25
	}
	;;
	{ .mmf
	STFD	[C1 ] = f18, SIZE
	STFD	[C9 ] = f19, SIZE
	FMA	f26  = ALPHA_I, f72, f26
	}
	{ .mmf
	LDFD	f42 = [C3 ], SIZE
	LDFD	f43 = [C11], SIZE
	FMA	f27  = ALPHA_I, f74, f27
	}
	;;
	{ .mmf
	STFD	[C1 ] = f20, SIZE
	STFD	[C9 ] = f21, SIZE
	FMA	f28  = ALPHA_R, f73, f28
	}
	{ .mmf
	LDFD	f44 = [C3 ], SIZE
	LDFD	f45 = [C11], SIZE
	FMA	f29  = ALPHA_R, f75, f29
	}
	;;
	{ .mmf
	STF