trsm_kernel_ppc440_rt.s

Optimized GotoBLAS libraries

	fmr	f5,  f0
	fmr	f6,  f0
	fmr	f7,  f0

	fmr	f8,  f0
	fmr	f9,  f0
	fmr	f10, f0
	fmr	f11, f0

	fmr	f12, f0
	fmr	f13, f0
	fmr	f14, f0
	fmr	f15, f0

#ifndef LN
	addi	CO1, CO1, 4 * SIZE
	addi	CO2, CO2, 4 * SIZE
	addi	CO3, CO3, 4 * SIZE
	addi	CO4, CO4, 4 * SIZE
#endif
	
#ifdef RT
	slwi	r0, K, 2 + BASE_SHIFT
	add	AORIG, AORIG, r0
#endif

#if defined(LT) || defined(RN)
	sub	TEMP, K, KK
	slwi	TEMP, TEMP, 2 + BASE_SHIFT
	add	AO, AO, TEMP
	add	BO, BO, TEMP
#endif

#ifdef LT
	addi	KK, KK, 4
#endif

#ifdef LN
	subi	KK, KK, 4
#endif

	addic.	I, I, -1
	bgt+	.L11
	.align 4

.L20:
	andi.	I,  M,  2
	ble	.L30

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

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

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

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

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

	slwi	r0,   KK, 1 + BASE_SHIFT
	slwi	TEMP, KK, 2 + BASE_SHIFT
	add	AO, AORIG, r0
	add	BO, B,     TEMP

	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)

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

	srawi.	r0, TEMP,  2
	mtspr	CTR, r0
#endif
	ble	.L25
	.align 5

.L22:
	FMADD	f0,  f16, f20, f0
	nop
	FMADD	f1,  f17, f20, f1
	LFD	f20,  8 * SIZE(BO)
	FMADD	f4,  f16, f21, f4
	nop
	FMADD	f5,  f17, f21, f5
	LFD	f21,  9 * SIZE(BO)

	FMADD	f8,  f16, f22, f8
	nop
	FMADD	f9,  f17, f22, f9
	LFD	f22, 10 * SIZE(BO)
	FMADD	f12, f16, f23, f12
	LFD	f16,  4 * SIZE(AO)
	FMADD	f13, f17, f23, f13
	LFD	f23, 11 * SIZE(BO)

	FMADD	f2,  f18, f24, f2
	LFD	f17,  5 * SIZE(AO)
	FMADD	f3,  f19, f24, f3
	LFD	f24, 12 * SIZE(BO)
	FMADD	f6,  f18, f25, f6
	nop
	FMADD	f7,  f19, f25, f7
	LFD	f25, 13 * SIZE(BO)

	FMADD	f10, f18, f26, f10
	nop
	FMADD	f11, f19, f26, f11
	LFD	f26, 14 * SIZE(BO)
	FMADD	f14, f18, f27, f14
	LFD	f18,  6 * SIZE(AO)
	FMADD	f15, f19, f27, f15
	LFD	f27, 15 * SIZE(BO)

	FMADD	f0,  f16, f20, f0
	LFD	f19,  7 * SIZE(AO)
	FMADD	f1,  f17, f20, f1
	LFDU	f20, 16 * SIZE(BO)
	FMADD	f4,  f16, f21, f4
	nop
	FMADD	f5,  f17, f21, f5
	LFD	f21,  1 * SIZE(BO)

	FMADD	f8,  f16, f22, f8
	nop
	FMADD	f9,  f17, f22, f9
	LFD	f22,  2 * SIZE(BO)
	FMADD	f12, f16, f23, f12
	LFDU	f16,  8 * SIZE(AO)
	FMADD	f13, f17, f23, f13
	LFD	f23,  3 * SIZE(BO)

	FMADD	f2,  f18, f24, f2
	LFD	f17,  1 * SIZE(AO)
	FMADD	f3,  f19, f24, f3
	LFD	f24,  4 * SIZE(BO)
	FMADD	f6,  f18, f25, f6
	nop
	FMADD	f7,  f19, f25, f7
	LFD	f25,  5 * SIZE(BO)

	FMADD	f10, f18, f26, f10
	nop
	FMADD	f11, f19, f26, f11
	LFD	f26,  6 * SIZE(BO)
	FMADD	f14, f18, f27, f14
	LFD	f18,  2 * SIZE(AO)
	FMADD	f15, f19, f27, f15
	LFD	f19,  3 * SIZE(AO)
	LFD	f27,  7 * SIZE(BO)
	bdnz	.L22

	fadd	f0,  f2,  f0
	fadd	f1,  f3,  f1
	fadd	f4,  f6,  f4
	fadd	f5,  f7,  f5
	fadd	f8,  f10, f8
	fadd	f9,  f11, f9
	fadd	f12, f14, f12
	fadd	f13, f15, f13
	.align 4

.L25:
#if defined(LT) || defined(RN)
	andi.	r0, KK,  3
#else
	andi.	r0, TEMP, 3
#endif
	mtspr	CTR, r0
	ble+	.L28
	.align 4

.L26:
	FMADD	f0,  f16, f20, f0
	nop
	FMADD	f1,  f17, f20, f1
	LFDU	f20,  4 * SIZE(BO)
	FMADD	f4,  f16, f21, f4
	nop
	FMADD	f5,  f17, f21, f5
	LFD	f21,  1 * SIZE(BO)

	FMADD	f8,  f16, f22, f8
	nop
	FMADD	f9,  f17, f22, f9
	LFD	f22,  2 * SIZE(BO)
	FMADD	f12, f16, f23, f12
	LFDU	f16,  2 * SIZE(AO)
	FMADD	f13, f17, f23, f13
	LFD	f17,  1 * SIZE(AO)
	LFD	f23,  3 * SIZE(BO)
	bdnz	.L26
	.align 4

.L28:
#if defined(LN) || defined(RT)
#ifdef LN
	subi	r0, KK, 2
#else
	subi	r0, KK, 4
#endif
	slwi	TEMP, r0, 1 + BASE_SHIFT
	slwi	r0,   r0, 2 + BASE_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)
 
	LFD	f20,  4 * SIZE(BO)
 	LFD	f21,  5 * SIZE(BO)
	LFD	f22,  6 * SIZE(BO)
	LFD	f23,  7 * SIZE(BO)

	FSUB	f0,  f16, f0
	FSUB	f4,  f17, f4
	FSUB	f8,  f18, f8
	FSUB	f12, f19, f12

	FSUB	f1,  f20, f1
	FSUB	f5,  f21, f5
	FSUB	f9,  f22, f9
	FSUB	f13, f23, f13
#else
	LFD	f16,  0 * SIZE(AO)
	LFD	f17,  1 * SIZE(AO)
	LFD	f20,  2 * SIZE(AO)
 	LFD	f21,  3 * SIZE(AO)

	LFD	f24,  4 * SIZE(AO)
	LFD	f25,  5 * SIZE(AO)
	LFD	f28,  6 * SIZE(AO)
	LFD	f29,  7 * SIZE(AO)

	FSUB	f0,  f16, f0
	FSUB	f1,  f17, f1
	FSUB	f4,  f20, f4
	FSUB	f5,  f21, f5

	FSUB	f8,  f24, f8
	FSUB	f9,  f25, f9
	FSUB	f12, f28, f12
	FSUB	f13, f29, f13
#endif

#ifdef LN
	LFD	f19,  3 * SIZE(AO)
	LFD	f20,  2 * SIZE(AO)
	LFD	f21,  0 * SIZE(AO)

	FMUL	f1,  f19, f1
	FMUL	f5,  f19, f5
	FMUL	f9,  f19, f9
	FMUL	f13, f19, f13

	FNMSUB	f0,  f20, f1,  f0
	FNMSUB	f4,  f20, f5,  f4
	FNMSUB	f8,  f20, f9,  f8
	FNMSUB	f12, f20, f13, f12

	FMUL	f0,  f21, f0
	FMUL	f4,  f21, f4
	FMUL	f8,  f21, f8
	FMUL	f12, f21, f12
#endif

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

	FMUL	f0,  f16, f0
	FMUL	f4,  f16, f4
	FMUL	f8,  f16, f8
	FMUL	f12, f16, f12

	FNMSUB	f1,  f17, f0,  f1
	FNMSUB	f5,  f17, f4,  f5
	FNMSUB	f9,  f17, f8,  f9
	FNMSUB	f13, f17, f12, f13

	LFD	f17,  3 * SIZE(AO)

	FMUL	f1,  f17, f1
	FMUL	f5,  f17, f5
	FMUL	f9,  f17, f9
	FMUL	f13, f17, f13
#endif

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

	FMUL	f0,  f16, f0
	FMUL	f1,  f16, f1
	FNMSUB	f4,  f17, f0,  f4
	FNMSUB	f5,  f17, f1,  f5
	FNMSUB	f8,  f18, f0,  f8
	FNMSUB	f9,  f18, f1,  f9
	FNMSUB	f12, f19, f0,  f12
	FNMSUB	f13, f19, f1,  f13

	LFD	f16,  5 * SIZE(BO)
	LFD	f17,  6 * SIZE(BO)
	LFD	f18,  7 * SIZE(BO)
	LFD	f19, 10 * SIZE(BO)

	LFD	f20, 11 * SIZE(BO)
	LFD	f21, 15 * SIZE(BO)

	FMUL	f4,  f16, f4
	FMUL	f5,  f16, f5
	FNMSUB	f8,  f17, f4,  f8
	FNMSUB	f9,  f17, f5,  f9
	FNMSUB	f12, f18, f4,  f12
	FNMSUB	f13, f18, f5,  f13

	FMUL	f8,  f19, f8
	FMUL	f9,  f19, f9
	FNMSUB	f12, f20, f8,  f12
	FNMSUB	f13, f20, f9,  f13
	FMUL	f12, f21, f12
	FMUL	f13, f21, f13
#endif

#ifdef RT
	LFD	f16, 15 * SIZE(BO)
	LFD	f17, 14 * SIZE(BO)
	LFD	f18, 13 * SIZE(BO)
	LFD	f19, 12 * SIZE(BO)

	FMUL	f12, f16, f12
	FMUL	f13, f16, f13
	FNMSUB	f8,  f17, f12, f8
	FNMSUB	f9,  f17, f13, f9
	FNMSUB	f4,  f18, f12, f4
	FNMSUB	f5,  f18, f13, f5
	FNMSUB	f0,  f19, f12, f0
	FNMSUB	f1,  f19, f13, f1

	LFD	f16, 10 * SIZE(BO)
	LFD	f17,  9 * SIZE(BO)
	LFD	f18,  8 * SIZE(BO)
	LFD	f19,  5 * SIZE(BO)
	LFD	f20,  4 * SIZE(BO)
	LFD	f21,  0 * SIZE(BO)

	FMUL	f8,  f16, f8
	FMUL	f9,  f16, f9
	FNMSUB	f4,  f17, f8,  f4
	FNMSUB	f5,  f17, f9,  f5
	FNMSUB	f0,  f18, f8,  f0
	FNMSUB	f1,  f18, f9,  f1

	FMUL	f4,  f19, f4
	FMUL	f5,  f19, f5
	FNMSUB	f0,  f20, f4,  f0
	FNMSUB	f1,  f20, f5,  f1

	FMUL	f0,  f21, f0
	FMUL	f1,  f21, f1
#endif

#ifdef LN
	subi	CO1, CO1, 2 * SIZE
	subi	CO2, CO2, 2 * SIZE
	subi	CO3, CO3, 2 * SIZE
	subi	CO4, CO4, 2 * SIZE
#endif

#if defined(LN) || defined(LT)
	STFD	f0,   0 * SIZE(BO)
	STFD	f4,   1 * SIZE(BO)
	STFD	f8,   2 * SIZE(BO)
	STFD	f12,  3 * SIZE(BO)

	STFD	f1,   4 * SIZE(BO)
	STFD	f5,   5 * SIZE(BO)
	STFD	f9,   6 * SIZE(BO)
	STFD	f13,  7 * SIZE(BO)
#else
	STFD	f0,   0 * SIZE(AO)
	STFD	f1,   1 * SIZE(AO)
	STFD	f4,   2 * SIZE(AO)
	STFD	f5,   3 * SIZE(AO)

	STFD	f8,   4 * SIZE(AO)
	STFD	f9,   5 * SIZE(AO)
	STFD	f12,  6 * SIZE(AO)
	STFD	f13,  7 * SIZE(AO)
#endif

	STFD	f0,   0 * SIZE(CO1)
	STFD	f1,   1 * SIZE(CO1)
	STFD	f4,   0 * SIZE(CO2)
	STFD	f5,   1 * SIZE(CO2)

	STFD	f8,   0 * SIZE(CO3)
	STFD	f9,   1 * SIZE(CO3)
	STFD	f12,  0 * SIZE(CO4)
	STFD	f13,  1 * SIZE(CO4)

	lfs	f0,  FZERO
 	fmr	f1,  f0
	fmr	f2,  f0
	fmr	f3,  f0

	fmr	f4,  f0
	fmr	f5,  f0
	fmr	f6,  f0
	fmr	f7,  f0

	fmr	f8,  f0
	fmr	f9,  f0
	fmr	f10, f0
	fmr	f11, f0

	fmr	f12, f0
	fmr	f13, f0
	fmr	f14, f0
	fmr	f15, f0

#ifndef LN
	addi	CO1, CO1, 2 * SIZE
	addi	CO2, CO2, 2 * SIZE
	addi	CO3, CO3, 2 * SIZE
	addi	CO4, CO4, 2 * SIZE
#endif

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

#if defined(LT) || defined(RN)
	sub	TEMP, K, KK
	slwi	r0,   TEMP, 1 + BASE_SHIFT
	slwi	TEMP, TEMP, 2 + BASE_SHIFT
	add	AO, AO, r0
	add	BO, BO, TEMP
#endif

#ifdef LN
	subi	KK, KK, 2
#endif

#ifdef LT
	addi	KK, KK, 2
#endif
	.align 4

.L30:
	andi.	I,  M,  1
	ble	.L39

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

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

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

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

#ifdef LN
	slwi	r0,   K,  BASE_SHIFT
	sub	AORIG, AORIG, r0
#endif

	slwi	r0,   KK, 0 + BASE_SHIFT
	slwi	TEMP, KK, 2 + BASE_SHIFT
	add	AO, AORIG, r0
	add	BO, B,     TEMP

	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)

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

	srawi.	r0, TEMP,  2
	mtspr	CTR, r0
#endif
	ble	.L35
	.align 5

.L32:
	FMADD	f0,  f16, f20, f0
	LFD	f20,  8 * SIZE(BO)
	FMADD	f4,  f16, f21, f4
	LFD	f21,  9 * SIZE(BO)
	FMADD	f8,  f16, f22, f8
	LFD	f22, 10 * SIZE(BO)
	FMADD	f12, f16, f23, f12
	LFD	f23, 11 * SIZE(BO)
	LFDU	f16,  4 * SIZE(AO)

	FMADD	f1,  f17, f24, f1
	LFD	f24, 12 * SIZE(BO)
	FMADD	f5,  f17, f25, f5
	LFD	f25, 13 * SIZE(BO)
	FMADD	f9,  f17, f26, f9
	LFD	f26, 14 * SIZE(BO)
	FMADD	f13, f17, f27, f13
	LFD	f27, 15 * SIZE(BO)
	LFD	f17,  1 * SIZE(AO)

	FMADD	f0,  f18, f20, f0
	LFDU	f20, 16 * SIZE(BO)
	FMADD	f4,  f18, f21, f4
	LFD	f21,  1 * SIZE(BO)
	FMADD	f8,  f18, f22, f8
	LFD	f22,  2 * SIZE(BO)
	FMADD	f12, f18, f23, f12
	LFD	f23,  3 * SIZE(BO)
	LFD	f18,  2 * SIZE(AO)

	FMADD	f1,  f19, f24, f1
	LFD	f24,  4 * SIZE(BO)
	FMADD	f5,  f19, f25, f5
	LFD	f25,  5 * SIZE(BO)
	FMADD	f9,  f19, f26, f9
	LFD	f26,  6 * SIZE(BO)
	FMADD	f13, f19, f27, f13
	LFD	f27,  7 * SIZE(BO)
	LFD	f19,  3 * SIZE(AO)
	bdnz	.L32

	fadd	f0,  f1,   f0
	fadd	f4,  f5,   f4
	fadd	f8,  f9,   f8
	fadd	f12, f13, f12
	.align 4

.L35:
#if defined(LT) || defined(RN)
	andi.	r0, KK,  3
#else
	andi.	r0, TEMP, 3
#endif
	mtspr	CTR, r0
	ble+	.L38
	.align 4

.L36:
	FMADD	f0,  f16, f20, f0
	LFDU	f20,  4 * SIZE(BO)
	FMADD	f4,  f16, f21, f4
	LFD	f21,  1 * SIZE(BO)
	FMADD	f8,  f16, f22, f8
	LFD	f22,  2 * SIZE(BO)
	FMADD	f12, f16, f23, f12
	LFDU	f16,  1 * SIZE(AO)
	LFD	f23,  3 * SIZE(BO)
	bdnz	.L36
	.align 4

.L38:
#if defined(LN) || defined(RT)
#ifdef LN
	subi	r0, KK, 1
#else
	subi	r0, KK, 4
#endif
	slwi	TEMP, r0, 0 + BASE_SHIFT
	slwi	r0,   r0, 2 + BASE_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	f4,  f17, f4
	FSUB	f8,  f18, f8
	FSUB	f12, f19, f12
#else
	LFD	f16,  0 * SIZE(AO)
	LFD	f20,  1 * SIZE(AO)
	LFD	f24,  2 * SIZE(AO)
	LFD	f28,  3 * SIZE(AO)

	FSUB	f0,  f16, f0
	FSUB	f4,  f20, f4
	FSUB	f8,  f24, f8
	FSUB	f12, f28, f12
#endif

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

	FMUL	f0,  f21, f0
	FMUL	f4,  f21, f4
	FMUL	f8,  f21, f8
	FMUL	f12, f21, f12
#endif

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

	FMUL	f0,  f16, f0
	FMUL	f4,  f16, f4
	FMUL	f8,  f16, f8
	FMUL	f12, f16, f12
#endif

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

	FMUL	f0,  f16, f0
	FNMSUB	f4,  f17, f0,  f4
	FNMSUB	f8,  f18, f0,  f8
	FNMSUB	f12, f19, f0,  f12

	LFD	f16,  5 * SIZE(BO)
	LFD	f17,  6 * SIZE(BO)
	LFD	f18,  7 * SIZE(BO)
	LFD	f19, 10 * SIZE(BO)

	LFD	f20, 11 * SIZE(BO)
	LFD	f21, 15 * SIZE(BO)

	FMUL	f4,  f16, f4
	FNMSUB	f8,  f17, f4,  f8
	FNMSUB	f12, f18, f4,  f12
	FMUL	f8,  f19, f8
	FNMSUB	f12, f20, f8,  f12
	FMUL	f12, f21, f12
#endif

#ifdef RT
	LFD	f16, 15 * SIZE(BO)
	LFD	f17, 14 * SIZE(BO)
	LFD	f18, 13 * SIZE(BO)
	LFD	f19, 12 * SIZE(BO)

	FMUL	f12, f16, f12
	FNMSUB	f8,  f17, f12, f8
	FNMSUB	f4,  f18, f12, f4
	FNMSUB	f0,  f19, f12, f0

	LFD	f16, 10 * SIZE(BO)
	LFD	f17,  9 * SIZE(BO)
	LFD	f18,  8 * SIZE(BO)
	LFD	f19,  5 * SIZE(BO)

	FMUL	f8,  f16, f8

	LFD	f20,  4 * SIZE(BO)
	LFD	f21,  0 * SIZE(BO)

	FNMSUB	f4,  f17, f8,  f4
	FNMSUB	f0,  f18, f8,  f0

	FMUL	f4,  f19, f4
	FNMSUB	f0,  f20, f4,  f0
	FMUL	f0,  f21, f0
#endif

#ifdef LN
	subi	CO1, CO1, 1 * SIZE
	subi	CO2, CO2, 1 * SIZE
	subi	CO3, CO3, 1 * SIZE
	subi	CO4, CO4, 1 * SIZE
#endif

#if defined(LN) || defined(LT)
	STFD	f0,   0 * SIZE(BO)
	STFD	f4,   1 * SIZE(BO)
	STFD	f8,   2 * SIZE(BO)
	STFD	f12,  3 * SIZE(BO)
#else
	STFD	f0,   0 * SIZE(AO)
	STFD	f4,   1 * SIZE(AO)
	STFD	f8,   2 * SIZE(AO)
	STFD	f12,  3 * SIZE(AO)
#endif

	STFD	f0,   0 * SIZE(CO1)
	STFD	f4,   0 * SIZE(CO2)
	STFD	f8,   0 * SIZE(CO3)
	STFD	f12,  0 * SIZE(CO4)

	lfs	f0,  FZERO
 	fmr	f1,  f0
	fmr	f4,  f0
	fmr	f5,  f0

	fmr	f8,  f0
	fmr	f9,  f0
	fmr	f12, f0
	fmr	f13, f0

#ifndef LN
	addi	CO1, CO1, 1 * SIZE
	addi	CO2, CO2, 1 * SIZE
	addi	CO3, CO3, 1 * SIZE
	addi	CO4, CO4, 1 * SIZE
#endif

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

#if defined(LT) || defined(RN)
	sub	TEMP, K, KK
	slwi	r0,   TEMP, 0 + BASE_SHIFT
	slwi	TEMP, TEMP, 2 + BASE_SHIFT
	add	AO, AO, r0
	add	BO, BO, TEMP
#endif

#ifdef LN
	subi	KK, KK, 1
#endif

#ifdef LT
	addi	KK, KK, 1
#endif
	.align 4

.L39:
#ifdef LN
	slwi	r0, K, 2 + BASE_SHIFT
	add	B, B, r0
#endif

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

#ifdef RN
	addi	KK, KK, 4
#endif

#ifdef RT
	subi	KK, KK, 4
#endif

	addic.	J, J, -1
	lfs	f0, FZERO
	bgt	.L10
	.align 4

.L999:
	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)
	ld	r20,  232(SP)
	ld	r19,  240(SP)
	ld	r18,  248(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)
	lwz	r20,  188(SP)
	lwz	r19,  192(SP)
	lwz	r18,  196(SP)
#endif

	addi	SP, SP, STACKSIZE

	blr

	EPILOGUE