ztrsm_kernel_ln.s

Optimized GotoBLAS libraries

	fmr	f7, f0

	srawi.	r0, KK,  2
	mr	BO,  B
	mtspr	CTR, r0
#else

#ifdef LN
	slwi	r0,   K,  0 + ZBASE_SHIFT
	sub	AORIG, AORIG, r0
#endif

	slwi	r0,   KK, 0 + ZBASE_SHIFT

	add	AO, AORIG, r0
	add	BO, B,     r0

	sub	TEMP, K, KK

	LFD	f16,  0 * SIZE(AO)
	LFD	f17,  1 * SIZE(AO)
	LFD	f18,  2 * SIZE(AO)
	LFD	f19,  3 * SIZE(AO)

	LFD	f20,  0 * SIZE(BO)
	LFD	f21,  1 * SIZE(BO)
	LFD	f22,  2 * SIZE(BO)
	LFD	f23,  3 * SIZE(BO)

	lfs	f0, FZERO
	fmr	f1, f0
	fmr	f2, f0
	fmr	f3, f0
	fmr	f4, f0
	fmr	f5, f0
	fmr	f6, f0
	fmr	f7, f0

	srawi.	r0, TEMP,  2
	mtspr	CTR, r0
#endif
	ble	LL(45)
	.align 4

LL(42):
	fmadd	f0,  f16,  f20,  f0
	fmadd	f1,  f17,  f21,  f1
	fmadd	f2,  f17,  f20,  f2
	fmadd	f3,  f16,  f21,  f3

	LFD	f16,  4 * SIZE(AO)
	LFD	f17,  5 * SIZE(AO)
	LFD	f20,  4 * SIZE(BO)
	LFD	f21,  5 * SIZE(BO)

	fmadd	f4,  f18,  f22,  f4
	fmadd	f5,  f19,  f23,  f5
	fmadd	f6,  f19,  f22,  f6
	fmadd	f7,  f18,  f23,  f7

	LFD	f18,  6 * SIZE(AO)
	LFD	f19,  7 * SIZE(AO)
	LFD	f22,  6 * SIZE(BO)
	LFD	f23,  7 * SIZE(BO)

	fmadd	f0,  f16,  f20,  f0
	fmadd	f1,  f17,  f21,  f1
	fmadd	f2,  f17,  f20,  f2
	fmadd	f3,  f16,  f21,  f3

	LFD	f16,  8 * SIZE(AO)
	LFD	f17,  9 * SIZE(AO)
	LFD	f20,  8 * SIZE(BO)
	LFD	f21,  9 * SIZE(BO)

	fmadd	f4,  f18,  f22,  f4
	fmadd	f5,  f19,  f23,  f5
	fmadd	f6,  f19,  f22,  f6
	fmadd	f7,  f18,  f23,  f7

	LFD	f18, 10 * SIZE(AO)
	LFD	f19, 11 * SIZE(AO)
	LFD	f22, 10 * SIZE(BO)
	LFD	f23, 11 * SIZE(BO)

	addi	AO, AO,  8 * SIZE
	addi	BO, BO,  8 * SIZE
	bdnz	LL(42)
	.align 4

LL(45):
	fadd	f0, f0, f4
	fadd	f1, f1, f5
	fadd	f2, f2, f6
	fadd	f3, f3, f7

#if defined(LT) || defined(RN)
	andi.	r0, KK,  3
#else
	andi.	r0, TEMP, 3
#endif
	mtspr	CTR,r0
	ble	LL(47)
	.align 4

LL(46):
	fmadd	f0,  f16,  f20,  f0
	fmadd	f1,  f17,  f21,  f1
	fmadd	f2,  f17,  f20,  f2
	fmadd	f3,  f16,  f21,  f3

	LFD	f16, 2 * SIZE(AO)
	LFD	f17, 3 * SIZE(AO)
	LFD	f20, 2 * SIZE(BO)
	LFD	f21, 3 * SIZE(BO)

	addi	AO, AO, 2 * SIZE
	addi	BO, BO, 2 * SIZE

	bdnz	LL(46)
	.align 4

LL(47):
#ifndef CONJ
	FSUB	  f0,  f0,  f1
	FADD	  f1,  f2,  f3
#else
	FADD	  f0,  f0,  f1
	FSUB	  f1,  f3,  f2
#endif

#if defined(LN) || defined(RT)
	subi	r0, KK, 1
	slwi	r0,   r0, 0 + ZBASE_SHIFT

	add	AO, AORIG, r0
	add	BO, B,     r0
#endif

#if defined(LN) || defined(LT)
	LFD	f16,  0 * SIZE(BO)
	LFD	f17,  1 * SIZE(BO)
	FSUB	f0,  f16, f0
	FSUB	f1,  f17, f1
#else
	LFD	f16,  0 * SIZE(AO)
	LFD	f17,  1 * SIZE(AO)
#ifndef CONJ
	FSUB	f0,  f16, f0
	FSUB	f1,  f17, f1
#else
	FSUB	f0,  f16, f0
	FADD	f1,  f17, f1
#endif
#endif

#ifdef LN
	LFD	f20,  0 * SIZE(AO)
	LFD	f21,  1 * SIZE(AO)

	FMUL	f4,  f21, f1
	FMUL	f5,  f21, f0
#ifndef CONJ
	FMSUB	f0,  f20, f0,  f4
	FMADD	f1,  f20, f1,  f5
#else
	FMADD	f0,  f20, f0,  f4
	FMSUB	f1,  f20, f1,  f5
#endif
#endif

#ifdef LT
	LFD	f16,  0 * SIZE(AO)
	LFD	f17,  1 * SIZE(AO)

	FMUL	f4,  f17, f1
	FMUL	f5,  f17, f0

#ifndef CONJ
	FMSUB	f0,  f16, f0,  f4
	FMADD	f1,  f16, f1,  f5
#else
	FMADD	f0,  f16, f0,  f4
	FMSUB	f1,  f16, f1,  f5
#endif
#endif

#ifdef RN
	LFD	f16,  0 * SIZE(BO)
	LFD	f17,  1 * SIZE(BO)

	FMUL	f4,  f17, f1
	FMUL	f5,  f17, f0
#ifndef CONJ
	FMSUB	f0,  f16, f0,  f4
	FMADD	f1,  f16, f1,  f5
#else
	FMADD	f0,  f16, f0,  f4
	FMSUB	f1,  f16, f1,  f5
#endif
#endif

#ifdef RT
	LFD	f20,  0 * SIZE(BO)
	LFD	f21,  1 * SIZE(BO)

	FMUL	f4,  f21, f1
	FMUL	f5,  f21, f0
#ifndef CONJ
	FMSUB	f0,  f20, f0,  f4
	FMADD	f1,  f20, f1,  f5
#else
	FMADD	f0,  f20, f0,  f4
	FMSUB	f1,  f20, f1,  f5
#endif
#endif

#ifdef LN
	subi	CO1, CO1, 2 * SIZE
#endif

#if defined(LN) || defined(LT)
	STFD	f0,   0 * SIZE(BO)
	STFD	f1,   1 * SIZE(BO)
#else
	STFD	f0,   0 * SIZE(AO)
	STFD	f1,   1 * SIZE(AO)
#endif

	STFD	f0,   0 * SIZE(CO1)
	STFD	f1,   1 * SIZE(CO1)

#ifndef LN
	addi	CO1, CO1, 2 * SIZE
#endif

#ifdef RT
	slwi	r0, K, 0 + ZBASE_SHIFT
	add	AORIG, AORIG, r0
#endif

#if defined(LT) || defined(RN)
	sub	TEMP, K, KK
	slwi	TEMP, TEMP, 0 + ZBASE_SHIFT
	add	AO, AO, TEMP
	add	BO, BO, TEMP
#endif

#ifdef LT
	addi	KK, KK, 1
#endif

#ifdef LN
	subi	KK, KK, 1
#endif
	.align 4

LL(40):
	srawi.	I,  M,  1
	ble	LL(49)
	.align 4

LL(31):
#if defined(LT) || defined(RN)
	LFD	f20,  0 * SIZE(AO)
	LFD	f21,  1 * SIZE(AO)
	LFD	f22,  2 * SIZE(AO)
	LFD	f23,  3 * SIZE(AO)

	LFD	f24,  4 * SIZE(AO)
	LFD	f25,  5 * SIZE(AO)
	LFD	f26,  6 * SIZE(AO)
	LFD	f27,  7 * SIZE(AO)

	LFD	f16, 0 * SIZE(B)
	LFD	f17, 1 * SIZE(B)
	LFD	f18, 2 * SIZE(B)
	LFD	f19, 3 * SIZE(B)

	lfs	f0, FZERO
	fmr	f1, f0
	fmr	f2, f0
	fmr	f3, f0
	fmr	f4, f0
	fmr	f5, f0
	fmr	f6, f0
	fmr	f7, f0

	dcbt	CO1, PREC

	srawi.	r0, KK,  2
	mr	BO,  B
	mtspr	CTR, r0
#else

#ifdef LN
	slwi	r0,   K,  1 + ZBASE_SHIFT
	sub	AORIG, AORIG, r0
#endif

	slwi	r0,   KK, 1 + ZBASE_SHIFT
	slwi	TEMP, KK, 0 + ZBASE_SHIFT

	add	AO, AORIG, r0
	add	BO, B,     TEMP

	sub	TEMP, K, KK

	LFD	f20,  0 * SIZE(AO)
	LFD	f21,  1 * SIZE(AO)
	LFD	f22,  2 * SIZE(AO)
	LFD	f23,  3 * SIZE(AO)

	LFD	f24,  4 * SIZE(AO)
	LFD	f25,  5 * SIZE(AO)
	LFD	f26,  6 * SIZE(AO)
	LFD	f27,  7 * SIZE(AO)

	LFD	f16, 0 * SIZE(BO)
	LFD	f17, 1 * SIZE(BO)
	LFD	f18, 2 * SIZE(BO)
	LFD	f19, 3 * SIZE(BO)

	lfs	f0, FZERO
	fmr	f1, f0
	fmr	f2, f0
	fmr	f3, f0
	fmr	f4, f0
	fmr	f5, f0
	fmr	f6, f0
	fmr	f7, f0

	srawi.	r0, TEMP,  2
	mtspr	CTR, r0
#endif
	ble	LL(35)
	.align 4

LL(32):
	fmadd	f0,  f16, f20, f0
	fmadd	f1,  f16, f21, f1
	fmadd	f2,  f16, f22, f2
	fmadd	f3,  f16, f23, f3

	fmadd	f4,  f17, f20, f4
	fmadd	f5,  f17, f21, f5
	fmadd	f6,  f17, f22, f6
	fmadd	f7,  f17, f23, f7

	LFD	f20,  8 * SIZE(AO)
	LFD	f21,  9 * SIZE(AO)
	LFD	f22, 10 * SIZE(AO)
	LFD	f23, 11 * SIZE(AO)

	fmadd	f0,  f18, f24, f0
	fmadd	f1,  f18, f25, f1
	fmadd	f2,  f18, f26, f2
	fmadd	f3,  f18, f27, f3

	fmadd	f4,  f19, f24, f4
	fmadd	f5,  f19, f25, f5
	fmadd	f6,  f19, f26, f6
	fmadd	f7,  f19, f27, f7

	LFD	f24, 12 * SIZE(AO)
	LFD	f25, 13 * SIZE(AO)
	LFD	f26, 14 * SIZE(AO)
	LFD	f27, 15 * SIZE(AO)

	LFD	f16,  4 * SIZE(BO)
	LFD	f17,  5 * SIZE(BO)
	LFD	f18,  6 * SIZE(BO)
	LFD	f19,  7 * SIZE(BO)

	fmadd	f0,  f16, f20, f0
	fmadd	f1,  f16, f21, f1
	fmadd	f2,  f16, f22, f2
	fmadd	f3,  f16, f23, f3

	fmadd	f4,  f17, f20, f4
	fmadd	f5,  f17, f21, f5
	fmadd	f6,  f17, f22, f6
	fmadd	f7,  f17, f23, f7

	LFD	f20, 16 * SIZE(AO)
	LFD	f21, 17 * SIZE(AO)
	LFD	f22, 18 * SIZE(AO)
	LFD	f23, 19 * SIZE(AO)

	fmadd	f0,  f18, f24, f0
	fmadd	f1,  f18, f25, f1
	fmadd	f2,  f18, f26, f2
	fmadd	f3,  f18, f27, f3

	fmadd	f4,  f19, f24, f4
	fmadd	f5,  f19, f25, f5
	fmadd	f6,  f19, f26, f6
	fmadd	f7,  f19, f27, f7

	LFD	f24, 20 * SIZE(AO)
	LFD	f25, 21 * SIZE(AO)
	LFD	f26, 22 * SIZE(AO)
	LFD	f27, 23 * SIZE(AO)

	LFD	f16,  8 * SIZE(BO)
	LFD	f17,  9 * SIZE(BO)
	LFD	f18, 10 * SIZE(BO)
	LFD	f19, 11 * SIZE(BO)

	addi	AO, AO, 16 * SIZE
	addi	BO, BO,  8 * SIZE
 	dcbt	PREA, AO
 	dcbt	PREA, BO
	bdnz	LL(32)
	.align 4

LL(35):
#if defined(LT) || defined(RN)
	andi.	r0, KK,  3
#else
	andi.	r0, TEMP, 3
#endif
	mtspr	CTR, r0
	ble	LL(37)
	.align 4

LL(36):
	fmadd	f0,  f16, f20, f0
	fmadd	f1,  f16, f21, f1
	fmadd	f2,  f16, f22, f2
	fmadd	f3,  f16, f23, f3

	fmadd	f4,  f17, f20, f4
	fmadd	f5,  f17, f21, f5
	fmadd	f6,  f17, f22, f6
	fmadd	f7,  f17, f23, f7

	LFD	f20, 4 * SIZE(AO)
	LFD	f21, 5 * SIZE(AO)
	LFD	f22, 6 * SIZE(AO)
	LFD	f23, 7 * SIZE(AO)

	LFD	f16, 2 * SIZE(BO)
	LFD	f17, 3 * SIZE(BO)

	addi	BO,  BO, 2 * SIZE
	addi	AO,  AO, 4 * SIZE
	bdnz	LL(36)
	.align 4

LL(37):
#ifndef CONJ
	FSUB	  f0,  f0,  f5
	FADD	  f1,  f1,  f4
	FSUB	  f2,  f2,  f7
	FADD	  f3,  f3,  f6
#else
	FADD	  f0,  f0,  f5
	FSUB	  f1,  f4,  f1
	FADD	  f2,  f2,  f7
	FSUB	  f3,  f6,  f3
#endif

#if defined(LN) || defined(RT)
#ifdef LN
	subi	r0, KK, 2
#else
	subi	r0, KK, 1
#endif
	slwi	TEMP, r0, 1 + ZBASE_SHIFT
	slwi	r0,   r0, 0 + ZBASE_SHIFT

	add	AO, AORIG, TEMP
	add	BO, B,     r0
#endif

#if defined(LN) || defined(LT)
	LFD	f16,  0 * SIZE(BO)
	LFD	f17,  1 * SIZE(BO)
	LFD	f18,  2 * SIZE(BO)
	LFD	f19,  3 * SIZE(BO)

	FSUB	f0,  f16, f0
	FSUB	f1,  f17, f1
	FSUB	f2,  f18, f2
	FSUB	f3,  f19, f3
#else
	LFD	f16,  0 * SIZE(AO)
	LFD	f17,  1 * SIZE(AO)
	LFD	f18,  2 * SIZE(AO)
	LFD	f19,  3 * SIZE(AO)

#ifndef CONJ
	FSUB	f0,  f16, f0
	FSUB	f1,  f17, f1
	FSUB	f2,  f18, f2
	FSUB	f3,  f19, f3
#else
	FSUB	f0,  f16, f0
	FADD	f1,  f17, f1
	FSUB	f2,  f18, f2
	FADD	f3,  f19, f3
#endif
#endif

#ifdef LN
	LFD	f16,  6 * SIZE(AO)
	LFD	f17,  7 * SIZE(AO)
	LFD	f18,  4 * SIZE(AO)
	LFD	f19,  5 * SIZE(AO)
	LFD	f20,  0 * SIZE(AO)
	LFD	f21,  1 * SIZE(AO)

	FMUL	f6,  f17, f3
	FMUL	f7,  f17, f2

#ifndef CONJ
	FMSUB	f2,  f16, f2,  f6
	FMADD	f3,  f16, f3,  f7

	FMADD	f0,  f19, f3,  f0
	FNMSUB	f1,  f19, f2,  f1
	FNMSUB	f0,  f18, f2,  f0
	FNMSUB	f1,  f18, f3,  f1

	FMUL	f4,  f21, f1
	FMUL	f5,  f21, f0
	FMSUB	f0,  f20, f0,  f4
	FMADD	f1,  f20, f1,  f5
#else
	FMADD	f2,  f16, f2,  f6
	FMSUB	f3,  f16, f3,  f7

	FMSUB	f0,  f19, f3,  f0
	FNMADD	f1,  f19, f2,  f1
	FNMADD	f0,  f18, f2,  f0
	FNMADD	f1,  f18, f3,  f1

	FMUL	f4,  f21, f1
	FMUL	f5,  f21, f0
	FMADD	f0,  f20, f0,  f4
	FMSUB	f1,  f20, f1,  f5
#endif
#endif

#ifdef LT
	LFD	f16,  0 * SIZE(AO)
	LFD	f17,  1 * SIZE(AO)
	LFD	f18,  2 * SIZE(AO)
	LFD	f19,  3 * SIZE(AO)
	LFD	f20,  6 * SIZE(AO)
	LFD	f21,  7 * SIZE(AO)

	FMUL	f4,  f17, f1
	FMUL	f5,  f17, f0

#ifndef CONJ
	FMSUB	f0,  f16, f0,  f4
	FMADD	f1,  f16, f1,  f5

	FMADD	f2,  f19, f1, f2
	FNMSUB	f3,  f19, f0, f3
	FNMSUB	f2,  f18, f0,  f2
	FNMSUB	f3,  f18, f1,  f3

	FMUL	f4,  f21, f3
	FMUL	f5,  f21, f2
	FMSUB	f2,  f20, f2,  f4
	FMADD	f3,  f20, f3,  f5
#else
	FMADD	f0,  f16, f0,  f4
	FMSUB	f1,  f16, f1,  f5

	FMSUB	f2,  f19, f1, f2
	FNMADD	f3,  f19, f0, f3
	FNMADD	f2,  f18, f0,  f2
	FNMADD	f3,  f18, f1,  f3

	FMUL	f4,  f21, f3
	FMUL	f5,  f21, f2
	FMADD	f2,  f20, f2,  f4
	FMSUB	f3,  f20, f3,  f5
#endif
#endif

#ifdef RN
	LFD	f16,  0 * SIZE(BO)
	LFD	f17,  1 * SIZE(BO)

	FMUL	f4,  f17, f1
	FMUL	f5,  f17, f0
	FMUL	f6,  f17, f3
	FMUL	f7,  f17, f2

#ifndef CONJ
	FMSUB	f0,  f16, f0,  f4
	FMADD	f1,  f16, f1,  f5
	FMSUB	f2,  f16, f2,  f6
	FMADD	f3,  f16, f3,  f7
#else
	FMADD	f0,  f16, f0,  f4
	FMSUB	f1,  f16, f1,  f5
	FMADD	f2,  f16, f2,  f6
	FMSUB	f3,  f16, f3,  f7
#endif
#endif

#ifdef RT
	LFD	f20,  0 * SIZE(BO)
	LFD	f21,  1 * SIZE(BO)

	FMUL	f4,  f21, f1
	FMUL	f5,  f21, f0
	FMUL	f6,  f21, f3
	FMUL	f7,  f21, f2

#ifndef CONJ
	FMSUB	f0,  f20, f0,  f4
	FMADD	f1,  f20, f1,  f5
	FMSUB	f2,  f20, f2,  f6
	FMADD	f3,  f20, f3,  f7

#else
	FMADD	f0,  f20, f0,  f4
	FMSUB	f1,  f20, f1,  f5
	FMADD	f2,  f20, f2,  f6
	FMSUB	f3,  f20, f3,  f7
#endif
#endif

#ifdef LN
	subi	CO1, CO1, 4 * SIZE
#endif

#if defined(LN) || defined(LT)
	STFD	f0,   0 * SIZE(BO)
	STFD	f1,   1 * SIZE(BO)
	STFD	f2,   2 * SIZE(BO)
	STFD	f3,   3 * SIZE(BO)
#else
	STFD	f0,   0 * SIZE(AO)
	STFD	f1,   1 * SIZE(AO)
	STFD	f2,   2 * SIZE(AO)
	STFD	f3,   3 * SIZE(AO)
#endif

	STFD	f0,   0 * SIZE(CO1)
	STFD	f1,   1 * SIZE(CO1)
	STFD	f2,   2 * SIZE(CO1)
	STFD	f3,   3 * SIZE(CO1)

#ifndef LN
	addi	CO1, CO1, 4 * SIZE
#endif

#ifdef RT
	slwi	r0, K, 1 + ZBASE_SHIFT
	add	AORIG, AORIG, r0
#endif

#if defined(LT) || defined(RN)
	sub	TEMP, K, KK
	slwi	r0,   TEMP, 1 + ZBASE_SHIFT
	slwi	TEMP, TEMP, 0 + ZBASE_SHIFT
	add	AO, AO, r0
	add	BO, BO, TEMP
#endif

#ifdef LT
	addi	KK, KK, 2
#endif

#ifdef LN
	subi	KK, KK, 2
#endif

	addic.	I, I, -1
	bgt	LL(31)
	.align 4

LL(49):
#ifdef LN
	slwi	r0, K, 0 + ZBASE_SHIFT
	add	B, B, r0
#endif

#if defined(LT) || defined(RN)
	mr	B,  BO
#endif

#ifdef RN
	addi	KK, KK, 1
#endif

#ifdef RT
	subi	KK, KK, 1
#endif
	.align 4

LL(999):
	addi	r3, 0, 0

	lfd	f14,    0(SP)
	lfd	f15,    8(SP)
	lfd	f16,   16(SP)
	lfd	f17,   24(SP)

	lfd	f18,   32(SP)
	lfd	f19,   40(SP)
	lfd	f20,   48(SP)
	lfd	f21,   56(SP)

	lfd	f22,   64(SP)
	lfd	f23,   72(SP)
	lfd	f24,   80(SP)
	lfd	f25,   88(SP)

	lfd	f26,   96(SP)
	lfd	f27,  104(SP)
	lfd	f28,  112(SP)
	lfd	f29,  120(SP)

	lfd	f30,  128(SP)
	lfd	f31,  136(SP)

#ifdef __64BIT__
	ld	r31,  144(SP)
	ld	r30,  152(SP)
	ld	r29,  160(SP)
	ld	r28,  168(SP)
	ld	r27,  176(SP)
	ld	r26,  184(SP)
	ld	r25,  192(SP)
	ld	r24,  200(SP)
	ld	r23,  208(SP)
	ld	r22,  216(SP)
	ld	r21,  224(SP)
#else
	lwz	r31,  144(SP)
	lwz	r30,  148(SP)
	lwz	r29,  152(SP)
	lwz	r28,  156(SP)
	lwz	r27,  160(SP)
	lwz	r26,  164(SP)
	lwz	r25,  168(SP)
	lwz	r24,  172(SP)
	lwz	r23,  176(SP)
	lwz	r22,  180(SP)
	lwz	r21,  184(SP)
#endif

	addi	SP, SP, STACKSIZE

	blr

	EPILOGUE
#endif