ztrsm_kernel_ln.s

Optimized GotoBLAS libraries

	FMA_A	f65  = f107, f80, f65
	;;
	FMA_B	f64  = f107, f81, f64
	FNMA	f65  = f106, f81, f65
	;;
	FMPY	f32  = f120, f64
	FMPY	f33  = f121, f64
	;;
	FMA_C	f64  = f121, f65, f32
	FMA_D	f65  = f120, f65, f33
	;;
#endif

#if defined(LN) || defined(LT)
	adds	BOFFSET2 = 4 * SIZE, BOFFSET
	;;
	STFD	[BOFFSET]  = f64,  SIZE
	STFD	[BOFFSET2] = f96,  SIZE
	;;
	STFD	[BOFFSET]  = f65,  SIZE
	STFD	[BOFFSET2] = f97,  SIZE
	;;
	STFD	[BOFFSET]  = f80,  SIZE
	STFD	[BOFFSET2] = f112, SIZE
	;;
	STFD	[BOFFSET]  = f81,  5 * SIZE
	STFD	[BOFFSET2] = f113, 5 * SIZE
	;;
	adds	BOFFSET = - 8 * SIZE, BOFFSET
	;;
#else
	adds	AOFFSET2 = 4 * SIZE, AOFFSET
	;;
	STFD	[AOFFSET]  = f64,  SIZE
	STFD	[AOFFSET2] = f96,  SIZE
	;;
	STFD	[AOFFSET]  = f65,  SIZE
	STFD	[AOFFSET2] = f97,  SIZE
	;;
	STFD	[AOFFSET]  = f80,  SIZE
	STFD	[AOFFSET2] = f112, SIZE
	;;
	STFD	[AOFFSET]  = f81,  5 * SIZE
	STFD	[AOFFSET2] = f113, 5 * SIZE
	;;
	adds	AOFFSET = - 8 * SIZE, AOFFSET
	;;
#endif

#ifdef LN
	adds	C1 = -2 * SIZE, C1
	adds	C2 = -2 * SIZE, C2
	adds	C3 = -2 * SIZE, C3
	adds	C4 = -2 * SIZE, C4
#endif
	;;
	STFD	[C1 ] = f64, SIZE
	;;
	STFD	[C1 ] = f65, SIZE
	;;
	STFD	[C2 ] = f80, SIZE
	;;
	STFD	[C2 ] = f81, SIZE
	;;
	STFD	[C3 ] = f96, SIZE
	;;
	STFD	[C3 ] = f97, SIZE
	;;
	STFD	[C4 ] = f112, SIZE
	;;
	STFD	[C4 ] = f113, SIZE
	;;
	mov	f64 = f0
	mov	f65 = f0
	mov	f80 = f0
	mov	f81 = f0
	mov	f96 = f0
	mov	f97 = f0
	mov	f112 = f0
	mov	f113 = f0
	;;
#ifdef LN
	adds	C1 = -2 * SIZE, C1
	adds	C2 = -2 * SIZE, C2
	adds	C3 = -2 * SIZE, C3
	adds	C4 = -2 * SIZE, C4
#endif
	;;
	cmp.ne	p6, p0 = 1, I
	;;
	adds	I = -1, I
	;;
	shladd	r2 = K, ZBASE_SHIFT, r0
	;;
	sub	L = K, KK
	;;
#ifdef RT
	add	AORIG = r2, AORIG
#endif
	;;
#if defined(LT) || defined(RN)
	shladd	L = L, ZBASE_SHIFT, r0
	;;
	add	AOFFSET = L,    AOFFSET
	shladd	BOFFSET = L, 2, BOFFSET
#endif
	;;
#ifdef LT
	adds	KK =  1, KK
#elif defined LN
	adds	KK = -1, KK
#else
	nop	__LINE__
#endif
	;;
#if defined(LT) || defined(RN)
	mov	L = KK
#else
	sub	L = K, KK
#endif
	;;
	.align 16

.L020:
	{ .mib
#if defined(LT) || defined(RN)
	mov	L = KK
#else
	sub	L = K, KK
#endif
	tbit.z	p6, p7 = M, 1
	(p6)	br.cond.dptk .L010x
	}
	;;
	{ .mmi
	cmp.ne	p7, p0 = r0, L
	adds	BOFFSET = 0 * SIZE, B
	shl	r2 = K, 1 + ZBASE_SHIFT
	}
	{ .mmi
	shladd	r3 = KK, ZBASE_SHIFT, r0
	nop	__LINE__
	nop	__LINE__
	}
	;;
#if defined(LT) || defined(RN)
	{ .mfb
	(p7) LDFPD	f48, f49 = [BOFFSET], 2 * SIZE
	mov	f66  = f0
	nop	__LINE__
	}
	{ .mmf
	nop	__LINE__
	nop	__LINE__
	mov	f67  = f0
	}
	;;
#else
	{ .mfi
	shladd	BOFFSET = r3, 2, B
	mov	f66  = f0
#ifdef LN
	sub	AORIG = AORIG, r2
#else
	nop	__LINE__
#endif
	}
	;;
	{ .mfi
	(p7) LDFPD	f48, f49 = [BOFFSET], 2 * SIZE
	mov	f67  = f0
	shladd	AOFFSET = r3, 1, AORIG
	}
	;;
#endif
	;;
	adds	L =  1, L
	;;
	{ .mfi
	(p7) LDFPD	f32, f33 = [AOFFSET], 2 * SIZE
	mov	f82 = f0
	tbit.z	p12, p0 = L, 0
	}
	{ .mfi	
	(p7) LDFPD	f50, f51 = [BOFFSET], 2 * SIZE
	mov	f83 = f0
	shr	L = L, 1
	}
	;;
	{ .mfi
	(p7) LDFPD	f34, f35 = [AOFFSET], 2 * SIZE
	mov	f98 = f0
	adds	L =  -1, L
	}
	{ .mfi
	(p7) LDFPD	f52, f53 = [BOFFSET], 2 * SIZE
	mov	f99 = f0
	cmp.eq	p3, p0 = r0, r0
	}
	;;
	{ .mfi
	(p7) LDFPD	f54, f55 = [BOFFSET], 2 * SIZE
	mov	f114 = f0
	mov	ar.lc = L
	}
	{ .mfi	
	adds	PREA = (PREFETCHSIZE + 0) * SIZE, AOFFSET
	mov	f115 = f0
	nop	__LINE__
	}
	;;
	cmp.eq  p6, p0 = -1, L
	(p6) br.cond.dpnt   .L028
	;;
	.align 16

.L022:
	{ .mfi
	lfetch.nt1	[PREA],  8 * SIZE
	FMA	f64   = f32, f48, f64	// A1 * B1
	adds	PREB = (PREFETCHSIZE + 0) * SIZE, BOFFSET
	}
	{ .mfi
	nop	__LINE__
	FMA_B	f65   = f32, f49, f65	// A1 * B2
	(p12) cmp.ne p3, p0 =  0, L
	}
	;;
	{ .mfi
	lfetch.nt1	[PREB],  16 * SIZE
	FMA	f80   = f32, f50, f80	// A1 * B3
	cmp.ne	p4, p5 =  0, L
	}
	{ .mfb
	nop	__LINE__
	FMA_B	f81   = f32, f51, f81	// A1 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	(p3) LDFPD	f40, f41 = [AOFFSET], 2 * SIZE
	FMA	f96   = f32, f52, f96	// A1 * B5
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA_B	f97  = f32, f53, f97	// A1 * B6
	nop	__LINE__
	}
	;;
	{ .mfb
	(p3) LDFPD	f56, f57 = [BOFFSET],   2 * SIZE
	FMA	f112  = f32, f54, f112	// A1 * B7
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA_B	f113  = f32, f55, f113	// A1 * B8
	nop	__LINE__
	}
	;;
	{ .mfb
	(p3) LDFPD	f58, f59 = [BOFFSET],  2 * SIZE
	FMA	f65   = f33, f48, f65	// A2 * B1
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA_A	f64   = f33, f49, f64	// A2 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	(p3) LDFPD	f60, f61 = [BOFFSET], 2 * SIZE
	FMA	f81   = f33, f50, f81	// A2 * B3
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA_A	f80   = f33, f51, f80	// A2 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	(p3) LDFPD	f62, f63 = [BOFFSET], 2 * SIZE
	FMA	f97   = f33, f52, f97	// A2 * B5
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA_A	f96  = f33, f53, f96	// A2 * B6
	nop	__LINE__
	}
	;;
	{ .mfb
	(p3) LDFPD	f42, f43 = [AOFFSET], 2 * SIZE
	FMA	f113  = f33, f54, f113	// A2 * B7
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA_A	f112  = f33, f55, f112	// A2 * B8
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	FMA	f66   = f34, f48, f66	// A3 * B1
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA_B	f67   = f34, f49, f67	// A3 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	FMA	f82   = f34, f50, f82	// A3 * B3
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA_B	f83   = f34, f51, f83	// A3 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	FMA	f98   = f34, f52, f98	// A3 * B5
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA_B	f99  = f34, f53, f99	// A3 * B6
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	FMA	f114  = f34, f54, f114	// A3 * B7
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA_B	f115  = f34, f55, f115	// A3 * B8
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	FMA	f67   = f35, f48, f67	// A4 * B1
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA_A	f66   = f35, f49, f66	// A4 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	FMA	f83   = f35, f50, f83	// A4 * B3
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA_A	f82   = f35, f51, f82	// A4 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	(p4) LDFPD	f32, f33 = [AOFFSET],   2 * SIZE
	FMA	f99   = f35, f52, f99	// A4 * B5
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA_A	f98  = f35, f53, f98	// A4 * B6
	nop	__LINE__
	}
	;;
	{ .mfb
	(p4) LDFPD	f48, f49 = [BOFFSET],   2 * SIZE
	FMA	f115  = f35, f54, f115	// A4 * B7
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA_A	f114  = f35, f55, f114	// A4 * B8
	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_B	f65   = f40, f57, f65	// A1 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	(p4) LDFPD	f52, f53 = [BOFFSET], 2 * SIZE
	(p3) FMA	f80   = f40, f58, f80	// A1 * B3
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA_B	f81   = f40, f59, f81	// A1 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	(p3) FMA	f96   = f40, f60, f96	// A1 * B5
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA_B	f97  = f40, f61, f97	// A1 * B6
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	(p3) FMA	f112  = f40, f62, f112	// A1 * B7
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA_B	f113  = f40, f63, f113	// A1 * B8
	nop	__LINE__
	}
	;;
	{ .mfb
	(p4) LDFPD	f54, f55 = [BOFFSET], 2 * SIZE
	(p3) FMA	f65   = f41, f56, f65	// A2 * B1
	nop	__LINE__
	}
	{ .mfb
	(p3) FMA_A	f64   = f41, f57, f64	// A2 * B2
	nop	__LINE__
	}
	{ .mfb
	(p4) LDFPD	f34, f35 = [AOFFSET], 2 * SIZE
	(p3) FMA	f81   = f41, f58, f81	// A2 * B3
	nop	__LINE__
	}
	{ .mfb
	(p3) FMA_A	f80   = f41, f59, f80	// A2 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	(p3) FMA	f97   = f41, f60, f97	// A2 * B5
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA_A	f96  = f41, f61, f96	// A2 * B6
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	(p3) FMA	f113  = f41, f62, f113	// A2 * B7
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA_A	f112  = f41, f63, f112	// A2 * B8
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	(p3) FMA	f66   = f42, f56, f66	// A3 * B1
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA_B	f67   = f42, f57, f67	// A3 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	(p3) FMA	f82   = f42, f58, f82	// A3 * B3
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA_B	f83   = f42, f59, f83	// A3 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	(p3) FMA	f98   = f42, f60, f98	// A3 * B5
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA_B	f99  = f42, f61, f99	// A3 * B6
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	(p3) FMA	f114  = f42, f62, f114	// A3 * B7
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA_B	f115  = f42, f63, f115	// A3 * B8
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	(p3) FMA	f67   = f43, f56, f67	// A4 * B1
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA_A	f66   = f43, f57, f66	// A4 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	(p3) FMA	f83   = f43, f58, f83	// A4 * B3
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA_A	f82   = f43, f59, f82	// A4 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	(p3) FMA	f99   = f43, f60, f99	// A4 * B5
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA_A	f98  = f43, f61, f98	// A4 * B6
	nop	__LINE__
	}
	;;
	{ .mfi
	nop	__LINE__
	(p3) FMA	f115  = f43, f62, f115	// A4 * B7
	adds	L = -1, L
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA_A	f114  = f43, f63, f114	// A4 * B8
	br.cloop.sptk.few .L022
	}
	;;
.L028:
#if defined(LN) || defined(RT)
#ifdef LN
	adds	r2 = -2, KK
#else
	adds	r2 = -4, KK
#endif
	;;
	shladd	r2 = r2, ZBASE_SHIFT, r0
	;;
	shladd	AOFFSET = r2, 1, AORIG
	shladd	BOFFSET = r2, 2, B
	;;	
#endif

#if defined(LN) || defined(LT)
	LDFPD	f72, f73 = [BOFFSET], 2 * SIZE
	;;
	LDFPD	f74, f75 = [BOFFSET], 2 * SIZE
	;;
	LDFPD	f88, f89 = [BOFFSET], 2 * SIZE
	;;
	LDFPD	f90, f91 = [BOFFSET], 2 * SIZE
	;;
	LDFPD	f104, f105 = [BOFFSET], 2 * SIZE
	;;
	LDFPD	f106, f107 = [BOFFSET], 2 * SIZE
	;;
	{ .mfi
	LDFPD	f120, f121 = [BOFFSET], 2 * SIZE
	FSUB	f64  = f72, f64
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB_A	f65  = f73, f65
	nop	__LINE__
	}
	;;
	{ .mfi
	LDFPD	f122, f123 = [BOFFSET]
	FSUB	f80  = f74, f80
	adds	BOFFSET = -14 * SIZE, BOFFSET
	}
	{ .mfi
	nop	__LINE__
	FSUB_A	f81  = f75, f81
	nop	__LINE__
	}
	;;
	{ .mfi
	nop	__LINE__
	FSUB	f96  = f88, f96
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB_A	f97  = f89, f97
	nop	__LINE__
	}
	;;
	{ .mfi
	nop	__LINE__
	FSUB	f112 = f90, f112
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB_A	f113 = f91, f113
	nop	__LINE__
	}
	;;
	{ .mfi
	nop	__LINE__
	FSUB	f66  = f104, f66
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB_A	f67  = f105, f67
	nop	__LINE__
	}
	;;
	{ .mfi
	nop	__LINE__
	FSUB	f82  = f106, f82
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB_A	f83  = f107, f83
	nop	__LINE__
	}
	;;
	{ .mfi
	nop	__LINE__
	FSUB	f98  = f120, f98
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB_A	f99  = f121, f99
	nop	__LINE__
	}
	;;
	{ .mfi
	nop	__LINE__
	FSUB	f114 = f122, f114
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB_A	f115 = f123, f115
	nop	__LINE__
	}
	;;
#else
	LDFPD	f72, f73 = [AOFFSET], 2 * SIZE
	;;
	LDFPD	f74, f75 = [AOFFSET], 2 * SIZE
	;;
	LDFPD	f88, f89 = [AOFFSET], 2 * SIZE
	;;
	LDFPD	f90, f91 = [AOFFSET], 2 * SIZE
	;;
	LDFPD	f104, f105 = [AOFFSET], 2 * SIZE
	;;
	LDFPD	f106, f107 = [AOFFSET], 2 * SIZE
	;;
	{ .mfi
	LDFPD	f120, f121 = [AOFFSET], 2 * SIZE
	FSUB	f64  = f72, f64
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB	f65  = f73, f65
	nop	__LINE__
	}
	;;
	{ .mfi
	LDFPD	f122, f123 = [AOFFSET]
	FSUB	f66  = f74, f66
	adds	AOFFSET = -14 * SIZE, AOFFSET
	}
	{ .mfi
	nop	__LINE__
	FSUB	f67  = f75, f67
	nop	__LINE__
	}
	;;
	{ .mfi
	nop	__LINE__
	FSUB	f80  = f88, f80
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB	f81  = f89, f81
	nop	__LINE__
	}
	;;
	{ .mfi
	nop	__LINE__
	FSUB	f82  = f90, f82
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB	f83  = f91, f83
	nop	__LINE__
	}
	;;
	{ .mfi
	nop	__LINE__
	FSUB	f96  = f104, f96
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB	f97  = f105, f97
	nop	__LINE__
	}
	;;
	{ .mfi
	nop	__LINE__
	FSUB	f98  = f106, f98
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB	f99  = f107, f99
	nop	__LINE__
	}
	;;
	{ .mfi
	nop	__LINE__
	FSUB	f112 = f120, f112
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB	f113 = f121, f113
	nop	__LINE__
	}
	;;
	{ .mfi
	nop	__LINE__
	FSUB	f114 = f122, f114
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB	f115 = f123, f115
	nop	__LINE__
	}
	;;
#endif

#ifdef LN
	adds	AOFFSET = 6 * SIZE, AOFFSET
	;;
	LDFPD	f104, f105 = [AOFFSET]
	adds	AOFFSET = - 2 * SIZE, AOFFSET
	;;
	LDFPD	f106, f107 = [AOFFSET]
	adds	AOFFSET = - 4 * SIZE, AOFFSET
	;;
	LDFPD	f120, f121 = [AOFFSET]
	;;
	FMPY	f32  = f104, f66
	FMPY	f33  = f105, f66
	FMPY	f34  = f104, f82
	FMPY	f35  = f105, f82
	FMPY	f36  = f104, f98
	FMPY	f37  = f105, f98
	FMPY	f38  = f104, f114
	FMPY	f39  = f105, f114
	;;
	FMA_C	f66  = f105, f67,  f32
	FMA_D	f67  = f104, f67,  f33
	FMA_C	f82  = f105, f83,  f34
	FMA_D	f83  = f104, f83,  f35
	FMA_C	f98  = f105, f99,  f36
	FMA_D	f99  = f104, f99,  f37