trsm_kernel_ln.s

Optimized GotoBLAS libraries

	;;
	FNMA	f72  = f120, f38, f72
	;;
	FNMA	f64  = f120, f39, f64
	;;
	FMPY	f112 = f112, f40
	;;
	FNMA	f104 = f112, f41, f104
	;;
	FNMA	f96  = f112, f42, f96
	;;
	FNMA	f88  = f112, f43, f88
	;;
	FNMA	f80  = f112, f44, f80
	;;
	FNMA	f72  = f112, f45, f72
	;;
	FNMA	f64  = f112, f46, f64
	;;
	FMPY	f104 = f104, f47
	;;
	FNMA	f96  = f104, f48, f96
	;;
	FNMA	f88  = f104, f49, f88
	;;
	FNMA	f80  = f104, f50, f80
	;;
	FNMA	f72  = f104, f51, f72
	;;
	FNMA	f64  = f104, f52, f64
	;;
	FMPY	f96  = f96,  f53
	;;
	FNMA	f88  = f96,  f54, f88
	;;
	FNMA	f80  = f96,  f55, f80
	;;
	FNMA	f72  = f96,  f56, f72
	;;
	FNMA	f64  = f96,  f57, f64
	;;
	FMPY	f88  = f88,  f58
	;;
	FNMA	f80  = f88,  f59, f80
	;;
	FNMA	f72  = f88,  f60, f72
	;;
	FNMA	f64  = f88,  f61, f64
	;;
	FMPY	f80  = f80,  f16
	;;
	FNMA	f72  = f80,  f17, f72
	;;
	FNMA	f64  = f80,  f18, f64
	;;
	FMPY	f72  = f72,  f19
	;;
	FNMA	f64  = f72,  f20, f64
	;;
	FMPY	f64  = f64,  f21
	;;
	STFD	[AOFFSET]  = f64, SIZE
	STFD	[AOFFSET2] = f96, SIZE
	;;
	STFD	[AOFFSET]  = f72, SIZE
	STFD	[AOFFSET2] = f104, SIZE
	;;
	STFD	[AOFFSET]  = f80, SIZE
	STFD	[AOFFSET2] = f112, SIZE
	;;
	STFD	[AOFFSET]  = f88,  - 3 * SIZE
	STFD	[AOFFSET2] = f120, - 3 * SIZE
	;;

#endif

#ifndef LN
	STFD	[C1 ] = f64, SIZE
#else
	STFD	[C1 ] = f64
#endif
#ifndef LN
	STFD	[C2 ] = f72, SIZE
#else
	STFD	[C2 ] = f72
#endif
#ifndef LN
	STFD	[C3 ] = f80, SIZE
#else
	STFD	[C3 ] = f80
#endif
#ifndef LN
	STFD	[C4 ] = f88, SIZE
#else
	STFD	[C4 ] = f88
#endif
#ifndef LN
	STFD	[C5 ] = f96,  SIZE
#else
	STFD	[C5 ] = f96
#endif
#ifndef LN
	STFD	[C6 ] = f104, SIZE
#else
	STFD	[C6 ] = f104
#endif
#ifndef LN
	STFD	[C7 ] = f112, SIZE
#else
	STFD	[C7 ] = f112
#endif
#ifndef LN
	STFD	[C8 ] = f120, SIZE
#else
	STFD	[C8 ] = f120
#endif
	;;

	mov	f64  = f0
	mov	f72  = f0
	mov	f80  = f0
	mov	f88  = f0
	mov	f96  = f0
	mov	f104 = f0
	mov	f112 = f0
	mov	f120 = f0
	;;
	shladd	r2 = K, BASE_SHIFT, r0
	;;
	sub	L = K, KK
	;;
#ifdef RT
	add	AORIG = r2, AORIG
#else
	nop	__LINE__
#endif
	;;
#if defined(LT) || defined(RN)
	shladd	L = L, BASE_SHIFT, r0
#else
	nop	__LINE__
#endif
	;;
#if defined(LT) || defined(RN)
	add	AOFFSET = L, AOFFSET
#else
	nop	__LINE__
#endif
	;;
#if defined(LT) || defined(RN)
	shladd	BOFFSET = L, 3, BOFFSET
#else
	nop	__LINE__
#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 8

.L030:
	{ .mib
	sub	L = K, KK
	tbit.z	p6, p0 = M, 1
	(p6)	br.cond.dptk .L020
	}
	;;
	;;
	{ .mmi
	cmp.ne	p7, p0 = r0, L
	adds	BOFFSET = 0 * SIZE, B
	shl	r2 = K, 1 + BASE_SHIFT
	}
	{ .mmi
	shladd	r3 = KK, BASE_SHIFT, r0
	nop	__LINE__
	nop	__LINE__
	}
	;;
#if defined(LT) || defined(RN)
	{ .mmf
	(p7) LDFPD	f48, f49 = [BOFFSET], 2 * SIZE
	setf.d	f73  = r0
	mov	f65  = f0
	}
	;;
#else
	{ .mfi
	shladd	BOFFSET = r3, 3, B
	mov	f65  = f0
#ifdef LN
	sub	AORIG = AORIG, r2
#else
	nop	__LINE__
#endif
	}
	;;
	{ .mfi
	(p7) LDFPD	f48, f49 = [BOFFSET], 2 * SIZE
	mov	f73  = f0
	shladd	AOFFSET = r3, 1, AORIG
	}
	;;
#endif
	{ .mfi
	setf.d	f105 = r0
	mov	f81  = f0
	adds	L =  1, L
	}
	{ .mfi
	adds	PREA = (PREFETCHSIZE + 8) * SIZE, AOFFSET
	mov	f89  = f0
	cmp.eq	p3, p0 = r0, r0
	}
	;;
	{ .mfi
	(p7) LDFPD	f50, f51 = [BOFFSET], 2 * SIZE
	mov	f113 = f0
	tbit.z	p12, p0 = L, 0
	}
	{ .mfi
	setf.d	f97  = r0
	mov	f121 = f0
	shr	L = L, 1
	}
	;;
	{ .mmf
	(p7) LDFPD	f52, f53 = [BOFFSET], 2 * SIZE
	adds	L =  -1, L
	}
	;;
	{ .mmf
	(p7) LDFPD	f54, f55 = [BOFFSET], 2 * SIZE
	cmp.eq  p6, p0 = -1, L
	}
	;;
	{ .mib
	(p7) LDFPD	f32, f33 = [AOFFSET], 2 * SIZE
	mov	ar.lc = L
	(p6) br.cond.dpnt   .L038
	}
	;;

.L032:
	{ .mfb
	lfetch.nt1	[PREA],  4 * 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],  16 * SIZE
	FMA	f80   = f32, f50, f80	// A1 * B3
	cmp.ne	p4, p5 =  0, L
	}
	{ .mfb
	nop	__LINE__
	FMA	f88   = f32, f51, f88	// A1 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	(p3) LDFPD	f56, f57 = [BOFFSET],   2 * SIZE
	FMA	f96   = f32, f52, f96	// A1 * B5
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA	f104  = f32, f53, f104	// A1 * B6
	nop	__LINE__
	}
	;;
	{ .mfb
	(p3) LDFPD	f40, f41 = [AOFFSET], 2 * SIZE
	FMA	f112  = f32, f54, f112	// A1 * B7
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA	f120  = f32, f55, f120	// 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	f73   = f33, f49, f73	// 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	f89   = f33, f51, f89	// 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	f105  = f33, f53, f105	// A2 * B6
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	FMA	f113  = f33, f54, f113	// A2 * B7
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	FMA	f121  = f33, f55, f121	// A2 * B8
	nop	__LINE__
	}
	;;
	{ .mfb
	(p4) LDFPD	f32, f33 = [AOFFSET],   2 * SIZE
	(p3) FMA	f64   = f40, f56, f64	// A1 * B1
	nop	__LINE__
	}
	{ .mfb
	(p4) LDFPD	f48, f49 = [BOFFSET],   2 * SIZE
	(p3) FMA	f72   = f40, f57, f72	// A1 * B2
	nop	__LINE__
	}
	;;
	{ .mfb
	(p4) LDFPD	f50, f51 = [BOFFSET],  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	f52, f53 = [BOFFSET], 2 * SIZE
	(p3) FMA	f96   = f40, f60, f96	// A1 * B5
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA	f104  = f40, f61, f104	// A1 * B6
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	(p3) FMA	f112  = f40, f62, f112	// A1 * B7
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA	f120  = f40, f63, f120	// 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	f73   = f41, f57, f73	// A2 * B2
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA	f81   = f41, f58, f81	// A2 * B3
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA	f89   = f41, f59, f89	// A2 * B4
	nop	__LINE__
	}
	;;
	{ .mfb
	nop	__LINE__
	(p3) FMA	f97   = f41, f60, f97	// A2 * B5
	nop	__LINE__
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA	f105  = f41, f61, f105	// A2 * B6
	nop	__LINE__
	}
	;;
	{ .mfi
	nop	__LINE__
	(p3) FMA	f113  = f41, f62, f113	// A2 * B7
	adds	L = -1, L
	}
	{ .mfb
	nop	__LINE__
	(p3) FMA	f121  = f41, f63, f121	// A2 * B8
	br.cloop.sptk.few .L032
	}
	;;

.L038:
#if defined(LN) || defined(RT)
#ifdef LN
	adds	r2 = -2, KK
#else
	adds	r2 = -8, KK
#endif
	;;
	shladd	r2 = r2, BASE_SHIFT, r0
	;;
	shladd	AOFFSET = r2, 1, AORIG
	shladd	BOFFSET = r2, 3, B
	;;	
#endif
 	adds	AOFFSET2 = 4 * SIZE, AOFFSET
 	adds	BOFFSET2 = 4 * SIZE, BOFFSET
	;;

#if defined(LN) || defined(LT)
	LDFPD	f32, f33 = [BOFFSET], 2 * SIZE
	;;
	LDFPD	f34, f35 = [BOFFSET], 2 * SIZE
	;;
	LDFPD	f36, f37 = [BOFFSET], 2 * SIZE
	;;
	LDFPD	f38, f39 = [BOFFSET], 2 * SIZE
	;;
	LDFPD	f40, f41 = [BOFFSET], 2 * SIZE
	;;
	LDFPD	f42, f43 = [BOFFSET], 2 * SIZE
	;;
	LDFPD	f44, f45 = [BOFFSET], 2 * SIZE
	;;
	LDFPD	f46, f47 = [BOFFSET]
	adds	BOFFSET = -14 * SIZE, BOFFSET
	;;
	{ .mfi
	FSUB	f64  = f32, f64
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB	f72  = f33, f72
	nop	__LINE__
	}
	;;
	{ .mfi
	FSUB	f80  = f34, f80
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB	f88  = f35, f88
	nop	__LINE__
	}
	;;
	{ .mfi
	FSUB	f96  = f36, f96
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB	f104 = f37, f104
	nop	__LINE__
	}
	;;
	{ .mfi
	FSUB	f112 = f38, f112
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB	f120 = f39, f120
	nop	__LINE__
	}
	;;
	{ .mfi
	FSUB	f65  = f40, f65
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB	f73  = f41, f73
	nop	__LINE__
	}
	;;
	{ .mfi
	FSUB	f81  = f42, f81
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB	f89  = f43, f89
	nop	__LINE__
	}
	;;
	{ .mfi
	FSUB	f97  = f44, f97
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FSUB	f105 = f45, f105
	nop	__LINE__
	}
	;;
	{ .mfi
	FSUB	f113 = f46, f113
	}
	{ .mfi
	nop	__LINE__
	FSUB	f121 = f47, f121
	nop	__LINE__
	}
	;;
#else
	LDFPD	f32, f33 = [AOFFSET], 2 * SIZE
	;;
	LDFPD	f34, f35 = [AOFFSET], 2 * SIZE
	;;
	LDFPD	f36, f37 = [AOFFSET], 2 * SIZE
	;;
	LDFPD	f38, f39 = [AOFFSET], 2 * SIZE
	;;
	LDFPD	f40, f41 = [AOFFSET], 2 * SIZE
	;;
	LDFPD	f42, f43 = [AOFFSET], 2 * SIZE
	;;
	LDFPD	f44, f45 = [AOFFSET], 2 * SIZE
	;;
	LDFPD	f46, f47 = [AOFFSET]
	adds	AOFFSET = -14 * SIZE, AOFFSET
	;;
	FSUB	f64  = f32, f64
	FSUB	f65  = f33, f65

	FSUB	f72  = f34, f72
	FSUB	f73  = f35, f73

	FSUB	f80  = f36, f80
	FSUB	f81  = f37, f81

	FSUB	f88  = f38, f88
	FSUB	f89  = f39, f89
	;;
	FSUB	f96  = f40, f96
	FSUB	f97  = f41, f97
	;;
	FSUB	f104 = f42, f104
	FSUB	f105 = f43, f105
	;;
	FSUB	f112 = f44, f112
	FSUB	f113 = f45, f113
	;;
	FSUB	f120 = f46, f120
	FSUB	f121 = f47, f121
	;;
#endif

#ifdef LN
	adds	AOFFSET = 2 * SIZE, AOFFSET
	;;
	LDFPD	f33, f32 = [AOFFSET]
	adds	AOFFSET = - 2 * SIZE, AOFFSET
	;;
	LDFD	f34 = [AOFFSET]
	;;
	FMPY	f65  = f65,  f32
	FMPY	f97  = f97,  f32
	FMPY	f73  = f73,  f32
 	FMPY	f105 = f105, f32
	FMPY	f81  = f81,  f32
	FMPY	f113 = f113, f32
	FMPY	f89  = f89,  f32
	FMPY	f121 = f121, f32
	;;
	FNMA	f64  = f65,  f33, f64
	FNMA	f96  = f97,  f33, f96
	FNMA	f72  = f73,  f33, f72
	FNMA	f104 = f105, f33, f104
	FNMA	f80  = f81,  f33, f80
	FNMA	f112 = f113, f33, f112
	FNMA	f88  = f89,  f33, f88
	FNMA	f120 = f121, f33, f120
	;;
	FMPY	f64  = f64,  f34
	FMPY	f96  = f96,  f34
	FMPY	f72  = f72,  f34
	FMPY	f104 = f104, f34
	FMPY	f80  = f80,  f34
 	FMPY	f112 = f112, f34
	FMPY	f88  = f88,  f34
	FMPY	f120 = f120, f34
	;;
	adds	BOFFSET  =  8 * SIZE, BOFFSET
	adds	BOFFSET2 =  8 * SIZE, BOFFSET2
	;;
	{ .mfi
	STFD	[BOFFSET]  = f65, SIZE
	}
	{ .mfi
	STFD	[BOFFSET2] = f97, SIZE
	}
	;;
	{ .mfi
	STFD	[BOFFSET]  = f73, SIZE
	}
	{ .mfi
	STFD	[BOFFSET2] = f105, SIZE
	}
	;;
	{ .mfi
	STFD	[BOFFSET]  = f81, SIZE
	}
	{ .mfi
	STFD	[BOFFSET2] = f113, SIZE
	}
	;;
	{ .mfi
	STFD	[BOFFSET]  = f89,  - 11 * SIZE
	}
	{ .mfi
	STFD	[BOFFSET2] = f121, - 11 * SIZE
	}
	;;
	{ .mmi
	STFD	[BOFFSET]  = f64, SIZE
	STFD	[BOFFSET2] = f96, SIZE
	adds	C1 = -2 * SIZE, C1
	}
	;;
	{ .mmi
	STFD	[BOFFSET]  = f72, SIZE
	STFD	[BOFFSET2] = f104, SIZE
	adds	C2 = -2 * SIZE, C2
	}
	;;
	{ .mmi
	STFD	[BOFFSET]  = f80, SIZE
	STFD	[BOFFSET2] = f112, SIZE
	nop	__LINE__
	}
	;;
	{ .mmi
	STFD	[BOFFSET]  = f88,  - 3 * SIZE
	STFD	[BOFFSET2] = f120, - 3 * SIZE
	}
	;;
	adds	C3 = -2 * SIZE, C3
	adds	C4 = -2 * SIZE, C4
	adds	C5 = -2 * SIZE, C5
	adds	C6 = -2 * SIZE, C6
	adds	C7 = -2 * SIZE, C7
	adds	C8 = -2 * SIZE, C8
	;;
#endif

#ifdef LT
	LDFPD	f32, f33 = [AOFFSET]
	adds	AOFFSET = 3 * SIZE, AOFFSET
	;;
	LDFD	f34 = [AOFFSET], - 3 * SIZE
	;;
	{ .mfi
	FMPY	f64  = f64,  f32
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FMPY	f96  = f96,  f32
	nop	__LINE__
	}
	;;
	{ .mfi
	FMPY	f72  = f72,  f32
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FMPY	f104 = f104, f32
	nop	__LINE__
	}
	;;
	{ .mfi
	FMPY	f80  = f80,  f32
	}
	{ .mfi
	nop	__LINE__
	FMPY	f112 = f112, f32
	nop	__LINE__
	}
	;;
	{ .mfi
	FMPY	f88  = f88,  f32
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FMPY	f120 = f120, f32
	nop	__LINE__
	}
	;;
	{ .mfi
	FNMA	f65  = f64,  f33, f65
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FNMA	f97  = f96,  f33, f97
	nop	__LINE__
	}
	;;
	{ .mfi
	FNMA	f73  = f72,  f33, f73
	nop	__LINE__
	}
	{ .mfi
	nop	__LINE__
	FNMA	f105 = f104, f33, f105
	nop	__LINE__