zgemm_kernel_g4.s

Optimized GotoBLAS libraries

	fmadd	f0,  f18, f24, f0
	LFD	f17,  1 * SIZE(AO)
	fmadd	f1,  f18, f25, f1
	nop
	fmadd	f2,  f18, f26, f2
	nop
	fmadd	f3,  f18, f27, f3
	LFD	f18,  2 * SIZE(AO)

	fmadd	f4,  f19, f24, f4
 	LFD	f24,  4 * SIZE(BO)
	fmadd	f5,  f19, f25, f5
	LFD	f25,  5 * SIZE(BO)
	fmadd	f6,  f19, f26, f6
	LFD	f26,  6 * SIZE(BO)
	fmadd	f7,  f19, f27, f7
	bdnz	.L22
	.align 4

.L25:
#ifndef TRMMKERNEL
	andi.	r0,  K,  3
	lfd	f30, ALPHA_R
	lfd	f31, ALPHA_I
	mtspr	CTR, r0
	ble	.L27
#else
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	sub	TEMP, K, KK
#elif defined(LEFT)
	addi	TEMP, KK, 1
#else
	addi	TEMP, KK, 2
#endif
	andi.	TEMP,  TEMP,  3
	lfd	f30, ALPHA_R
	lfd	f31, ALPHA_I
	mtspr	CTR, TEMP
	ble	.L27
#endif
	.align 4

.L26:
	fmadd	f0,  f16, f20, f0
	fmadd	f1,  f16, f21, f1
	fmadd	f2,  f16, f22, f2
	fmadd	f3,  f16, f23, f3
	LFDU	f16,  2 * SIZE(AO)

	fmadd	f4,  f17, f20, f4
 	LFDU	f20,  4 * SIZE(BO)
	fmadd	f5,  f17, f21, f5
	LFD	f21,  1 * SIZE(BO)
	fmadd	f6,  f17, f22, f6
	LFD	f22,  2 * SIZE(BO)
	fmadd	f7,  f17, f23, f7
	LFD	f23,  3 * SIZE(BO)
	LFD	f17,  1 * SIZE(AO)
	bdnz	.L26
	.align 4

.L27:
#if   defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
      defined(CC) || defined(CR) || defined(RC) || defined(RR)

	FSUB	  f0,  f0,  f5
	FADD	  f1,  f1,  f4
	FSUB	  f2,  f2,  f7
	FADD	  f3,  f3,  f6

#elif defined(CN) || defined(CT) || defined(RN) || defined(RT)

	FADD	  f0,  f0,  f5
	FSUB	  f1,  f4,  f1
	FADD	  f2,  f2,  f7
	FSUB	  f3,  f6,  f3

#else /* defined(NC) || defined(TC) || defined(NR) || defined(TR) */

	FADD	  f0,  f0,  f5
	FSUB	  f1,  f1,  f4
	FADD	  f2,  f2,  f7
	FSUB	  f3,  f3,  f6

#endif

#ifndef TRMMKERNEL
	LFD	f16, 0 * SIZE(CO1)
	LFD	f17, 1 * SIZE(CO1)

	LFD	f18, 0 * SIZE(CO2)
	LFD	f19, 1 * SIZE(CO2)
#endif

#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)

#ifndef TRMMKERNEL
	FMADD	f16, f30, f0,  f16
	FMADD	f17, f30, f1,  f17
	FMADD	f18, f30, f2,  f18
	FMADD	f19, f30, f3,  f19
#else
	FMUL	f16, f30, f0
	FMUL	f17, f30, f1
	FMUL	f18, f30, f2
	FMUL	f19, f30, f3
#endif

	FNMSUB	f16, f31, f1,  f16
	FMADD	f17, f31, f0,  f17
	FNMSUB	f18, f31, f3,  f18
	FMADD	f19, f31, f2,  f19


#else /* defined(CN)||defined(CT)||defined(NC)||defined(TC)||defined(CC) */
      /* defined(RN)||defined(RT)||defined(NR)||defined(TR)||defined(CR) */
      /* defined(RC)|| defined(RR) */

#ifndef TRMMKERNEL
	FMADD	f16, f30, f0,  f16
	FNMSUB	f17, f30, f1,  f17
	FMADD	f18, f30, f2,  f18
	FNMSUB	f19, f30, f3,  f19

	FMADD	f16, f31, f1,  f16
	FMADD	f17, f31, f0,  f17
	FMADD	f18, f31, f3,  f18
	FMADD	f19, f31, f2,  f19
#else
	FMUL	f16, f30, f0
	FMUL	f17, f30, f1
	FMUL	f18, f30, f2
	FMUL	f19, f30, f3

	FMADD	f16, f31, f1,  f16
	FNMADD	f17, f31, f0,  f17
	FMADD	f18, f31, f3,  f18
	FNMADD	f19, f31, f2,  f19
#endif
#endif

	STFD	f16, 0 * SIZE(CO1)
	STFD	f17, 1 * SIZE(CO1)
	STFD	f18, 0 * SIZE(CO2)
	STFD	f19, 1 * SIZE(CO2)

	addi	CO1, CO1, 2 * SIZE
	addi	CO2, CO2, 2 * SIZE

#ifdef TRMMKERNEL
#if ( defined(LEFT) &&  defined(TRANSA)) || \
    (!defined(LEFT) && !defined(TRANSA))
	sub	TEMP, K, KK
#ifdef LEFT
	addi	TEMP, TEMP, -1
#else
	addi	TEMP, TEMP, -2
#endif
	slwi	r0,   TEMP, 0 + ZBASE_SHIFT
	slwi	TEMP, TEMP, 1 + ZBASE_SHIFT
	add	AO, AO, r0
	add	BO, BO, TEMP
#endif

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

.L29:
#if defined(TRMMKERNEL) && !defined(LEFT)
	addi	KK, KK, 2
#endif

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

.L30:
	andi.	J, N,  1
	ble	.L999

#if defined(TRMMKERNEL) && defined(LEFT)
	mr	KK, OFFSET
#endif

	srawi.	I,  M,  1
	mr	CO1, C
	add	C, C, LDC
	mr	AO, A
	ble	.L40
	.align 4

.L31:
#ifndef TRMMKERNEL
	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

	srawi.	r0,  K,  2
	mr	BO, B
	mtspr	CTR, r0
	ble	.L35
#else
#if	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	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

	mr	BO,  B
#else
	slwi	r0,   KK, 1 + ZBASE_SHIFT
	slwi	TEMP, KK, 0 + ZBASE_SHIFT
	add	AO, AO, r0
	add	BO, B,  TEMP

	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
#endif

#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	sub	TEMP, K, KK
#elif defined(LEFT)
	addi	TEMP, KK, 2
#else
	addi	TEMP, KK, 1
#endif
	srawi.	TEMP,  TEMP,  2
	mtspr	CTR, TEMP
	ble	.L35
#endif
	.align 4

.L32:
	fmadd	f0,  f16, f20, f0
	LFD	f27,  7 * SIZE(AO)
	fmadd	f1,  f16, f21, f1
	LFD	f19,  3 * SIZE(BO)
	fmadd	f2,  f16, f22, f2
	nop
	fmadd	f3,  f16, f23, f3
	LFD	f16,  4 * SIZE(BO)

	fmadd	f4,  f17, f20, f4
	LFD	f20,  8 * SIZE(AO)
	fmadd	f5,  f17, f21, f5
	LFD	f21,  9 * SIZE(AO)
	fmadd	f6,  f17, f22, f6
	LFD	f22, 10 * SIZE(AO)
	fmadd	f7,  f17, f23, f7
	LFD	f23, 11 * SIZE(AO)

	fmadd	f0,  f18, f24, f0
	LFD	f17,  5 * SIZE(BO)
	fmadd	f1,  f18, f25, f1
	nop
	fmadd	f2,  f18, f26, f2
	nop
	fmadd	f3,  f18, f27, f3
	LFD	f18,  6 * SIZE(BO)

	fmadd	f4,  f19, f24, f4
	LFD	f24, 12 * SIZE(AO)
	fmadd	f5,  f19, f25, f5
	LFD	f25, 13 * SIZE(AO)
	fmadd	f6,  f19, f26, f6
	LFD	f26, 14 * SIZE(AO)
	fmadd	f7,  f19, f27, f7
	LFD	f27, 15 * SIZE(AO)

	fmadd	f0,  f16, f20, f0
	LFD	f19,  7 * SIZE(BO)
	fmadd	f1,  f16, f21, f1
	nop
	fmadd	f2,  f16, f22, f2
	nop
	fmadd	f3,  f16, f23, f3
	LFDU	f16,  8 * SIZE(BO)

	fmadd	f4,  f17, f20, f4
	LFDU	f20, 16 * SIZE(AO)
	fmadd	f5,  f17, f21, f5
	LFD	f21,  1 * SIZE(AO)
	fmadd	f6,  f17, f22, f6
	LFD	f22,  2 * SIZE(AO)
	fmadd	f7,  f17, f23, f7
	LFD	f23,  3 * SIZE(AO)

	fmadd	f0,  f18, f24, f0
	LFD	f17,  1 * SIZE(BO)
	fmadd	f1,  f18, f25, f1
	nop
	fmadd	f2,  f18, f26, f2
	nop
	fmadd	f3,  f18, f27, f3
	LFD	f18,  2 * SIZE(BO)

	fmadd	f4,  f19, f24, f4
	LFD	f24,  4 * SIZE(AO)
	fmadd	f5,  f19, f25, f5
	LFD	f25,  5 * SIZE(AO)
	fmadd	f6,  f19, f26, f6
	LFD	f26,  6 * SIZE(AO)
	fmadd	f7,  f19, f27, f7
	bdnz	.L32
	.align 4

.L35:
#ifndef TRMMKERNEL
	andi.	r0,  K,  3
	lfd	f30, ALPHA_R
	lfd	f31, ALPHA_I
	mtspr	CTR, r0
	ble	.L37
#else
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	sub	TEMP, K, KK
#elif defined(LEFT)
	addi	TEMP, KK, 2
#else
	addi	TEMP, KK, 1
#endif
	andi.	TEMP,  TEMP,  3
	lfd	f30, ALPHA_R
	lfd	f31, ALPHA_I
	mtspr	CTR, TEMP
	ble	.L37
#endif
	.align 4

.L36:
	fmadd	f0,  f16, f20, f0
	fmadd	f1,  f16, f21, f1
	fmadd	f2,  f16, f22, f2
	fmadd	f3,  f16, f23, f3
	LFDU	f16, 2 * SIZE(BO)

	fmadd	f4,  f17, f20, f4
	LFDU	f20, 4 * SIZE(AO)
	fmadd	f5,  f17, f21, f5
	LFD	f21, 1 * SIZE(AO)
	fmadd	f6,  f17, f22, f6
	LFD	f22, 2 * SIZE(AO)
	fmadd	f7,  f17, f23, f7
	LFD	f23, 3 * SIZE(AO)
	LFD	f17, 1 * SIZE(BO)
	bdnz	.L36
	.align 4

.L37:
#if   defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
      defined(CC) || defined(CR) || defined(RC) || defined(RR)

	FSUB	  f0,  f0,  f5
	FADD	  f1,  f1,  f4
	FSUB	  f2,  f2,  f7
	FADD	  f3,  f3,  f6

#elif defined(CN) || defined(CT) || defined(RN) || defined(RT)

	FADD	  f0,  f0,  f5
	FSUB	  f1,  f1,  f4
	FADD	  f2,  f2,  f7
	FSUB	  f3,  f3,  f6

#else /* defined(NC) || defined(TC) || defined(NR) || defined(TR) */

	FADD	  f0,  f0,  f5
	FSUB	  f1,  f4,  f1
	FADD	  f2,  f2,  f7
	FSUB	  f3,  f6,  f3

#endif

#ifndef TRMMKERNEL
	LFD	f16, 0 * SIZE(CO1)
	LFD	f17, 1 * SIZE(CO1)
	LFD	f18, 2 * SIZE(CO1)
	LFD	f19, 3 * SIZE(CO1)
#endif

#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)

#ifndef TRMMKERNEL
	FMADD	f16, f30, f0,  f16
	FMADD	f17, f30, f1,  f17
	FMADD	f18, f30, f2,  f18
	FMADD	f19, f30, f3,  f19
#else
	FMUL	f16, f30, f0
	FMUL	f17, f30, f1
	FMUL	f18, f30, f2
	FMUL	f19, f30, f3
#endif

	FNMSUB	f16, f31, f1,  f16
	FMADD	f17, f31, f0,  f17
	FNMSUB	f18, f31, f3,  f18
	FMADD	f19, f31, f2,  f19

#else /* defined(CN)||defined(CT)||defined(NC)||defined(TC)||defined(CC) */
      /* defined(RN)||defined(RT)||defined(NR)||defined(TR)||defined(CR) */
      /* defined(RC)|| defined(RR) */

#ifndef TRMMKERNEL
	FMADD	f16, f30, f0,  f16
	FNMSUB	f17, f30, f1,  f17
	FMADD	f18, f30, f2,  f18
	FNMSUB	f19, f30, f3,  f19

	FMADD	f16, f31, f1,  f16
	FMADD	f17, f31, f0,  f17
	FMADD	f18, f31, f3,  f18
	FMADD	f19, f31, f2,  f19
#else
	FMUL	f16, f30, f0
	FMUL	f17, f30, f1
	FMUL	f18, f30, f2
	FMUL	f19, f30, f3

	FMADD	f16, f31, f1,  f16
	FNMADD	f17, f31, f0,  f17
	FMADD	f18, f31, f3,  f18
	FNMADD	f19, f31, f2,  f19
#endif

#endif

	STFD	f16, 0 * SIZE(CO1)
	STFD	f17, 1 * SIZE(CO1)
	STFD	f18, 2 * SIZE(CO1)
	STFD	f19, 3 * SIZE(CO1)

	addi	CO1, CO1, 4 * SIZE

#ifdef TRMMKERNEL
#if ( defined(LEFT) &&  defined(TRANSA)) || \
    (!defined(LEFT) && !defined(TRANSA))
	sub	TEMP, K, KK
#ifdef LEFT
	addi	TEMP, TEMP, -2
#else
	addi	TEMP, TEMP, -1
#endif
	slwi	r0,   TEMP, 1 + ZBASE_SHIFT
	slwi	TEMP, TEMP, 0 + ZBASE_SHIFT
	add	AO, AO, r0
	add	BO, BO, TEMP
#endif

#ifdef LEFT
	addi	KK, KK, 2
#endif
#endif

	addic.	I, I, -1
	bgt	.L31
	.align 4

.L40:
	andi.	I,  M,  1
	ble	.L999

#ifndef TRMMKERNEL
	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)

	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,  K,  2
	mr	BO, B
	mtspr	CTR, r0
	ble	.L45
#else
#if	(defined(TRMMKERNEL) &&  defined(LEFT) &&  defined(TRANSA)) || \
	(defined(TRMMKERNEL) && !defined(LEFT) && !defined(TRANSA))

	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)

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

	mr	BO,  B
#else
	slwi	r0,   KK, 0 + ZBASE_SHIFT
	slwi	TEMP, KK, 0 + ZBASE_SHIFT
	add	AO, AO, r0
	add	BO, B,  TEMP

	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
#endif

#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	sub	TEMP, K, KK
#elif defined(LEFT)
	addi	TEMP, KK, 1
#else
	addi	TEMP, KK, 1
#endif
	srawi.	TEMP,  TEMP,  2
	mtspr	CTR, TEMP
	ble	.L45
#endif
	.align 4

.L42:
	fmadd	f0,  f16,  f20,  f0
	LFD	f23,  3 * SIZE(BO)
	fmadd	f3,  f16,  f21,  f3
	LFD	f16,  4 * SIZE(AO)
	fmadd	f2,  f17,  f20,  f2
	LFD	f20,  4 * SIZE(BO)
	fmadd	f1,  f17,  f21,  f1
	LFD	f17,  5 * SIZE(AO)

	fmadd	f4,  f18,  f22,  f4
	LFD	f21,  5 * SIZE(BO)
	fmadd	f7,  f18,  f23,  f7
	LFD	f18,  6 * SIZE(AO)
	fmadd	f6,  f19,  f22,  f6
	LFD	f22,  6 * SIZE(BO)
	fmadd	f5,  f19,  f23,  f5
	LFD	f19,  7 * SIZE(AO)

	fmadd	f0,  f16,  f20,  f0
	LFD	f23,  7 * SIZE(BO)
	fmadd	f3,  f16,  f21,  f3
	LFDU	f16,  8 * SIZE(AO)
	fmadd	f2,  f17,  f20,  f2
	LFDU	f20,  8 * SIZE(BO)
	fmadd	f1,  f17,  f21,  f1
	LFD	f17,  1 * SIZE(AO)

	fmadd	f4,  f18,  f22,  f4
	LFD	f21,  1 * SIZE(BO)
	fmadd	f7,  f18,  f23,  f7
	LFD	f18,  2 * SIZE(AO)
	fmadd	f6,  f19,  f22,  f6
	LFD	f22,  2 * SIZE(BO)
	fmadd	f5,  f19,  f23,  f5
	LFD	f19,  3 * SIZE(AO)
	bdnz	.L42
	.align 4

.L45:
	fadd	f0, f0, f4
	fadd	f1, f1, f5
	fadd	f2, f2, f6
	fadd	f3, f3, f7

#ifndef TRMMKERNEL
	andi.	r0,  K,  3
	lfd	f30, ALPHA_R
	lfd	f31, ALPHA_I
	mtspr	CTR,r0
	ble	.L47
#else
#if (defined(LEFT) && !defined(TRANSA)) || (!defined(LEFT) && defined(TRANSA))
	sub	TEMP, K, KK
#elif defined(LEFT)
	addi	TEMP, KK, 1
#else
	addi	TEMP, KK, 1
#endif
	andi.	TEMP,  TEMP,  3
	lfd	f30, ALPHA_R
	lfd	f31, ALPHA_I
	mtspr	CTR,TEMP
	ble	.L47
#endif
	.align 4

.L46:
	fmadd	f0,  f16,  f20,  f0
	fmadd	f3,  f16,  f21,  f3
	LFDU	f16, 2 * SIZE(AO)
	fmadd	f2,  f17,  f20,  f2
	LFDU	f20, 2 * SIZE(BO)
	fmadd	f1,  f17,  f21,  f1
	LFD	f17, 1 * SIZE(AO)
	LFD	f21, 1 * SIZE(BO)
	bdnz	.L46
	.align 4

.L47:
#if   defined(NN) || defined(NT) || defined(TN) || defined(TT) || \
      defined(CC) || defined(CR) || defined(RC) || defined(RR)
	fsub	f0, f0, f1
	fadd	f2, f2, f3
#elif defined(CN) || defined(CT) || defined(RN) || defined(RT)
	fadd	f0, f0, f1
	fsub	f2, f2, f3
#else
	fadd	f0, f0, f1
	fsub	f2, f3, f2
#endif

#ifndef TRMMKERNEL
	LFD	f16,  0 * SIZE(CO1)
	LFD	f17,  1 * SIZE(CO1)
#endif

#if   defined(NN) || defined(NT) || defined(TN) || defined(TT)
#ifndef TRMMKERNEL
	FMADD	f16, f30, f0, f16
	FMADD	f17, f30, f2, f17
#else
	FMUL	f16, f30, f0
	FMUL	f17, f30, f2
#endif

	FNMSUB	f16, f31, f2, f16
	FMADD	f17, f31, f0, f17

#else /* defined(CN)||defined(CT)||defined(NC)||defined(TC)||defined(CC) */
      /* defined(RN)||defined(RT)||defined(NR)||defined(TR)||defined(CR) */
      /* defined(RC) || defined(RR) */

#ifndef TRMMKERNEL
	FMADD	f16, f30, f0, f16
	FNMSUB	f17, f30, f2, f17

	FMADD	f16, f31, f2, f16
	FMADD	f17, f31, f0, f17
#else
	FMUL	f16, f30, f0
	FMUL	f17, f30, f2

	FMADD	f16, f31, f2, f16
	FNMADD	f17, f31, f0, f17
#endif

#endif
	STFD	f16,  0 * SIZE(CO1)
	STFD	f17,  1 * SIZE(CO1)
	.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)
#ifdef TRMMKERNEL
	ld	r23,  208(SP)
	ld	r22,  216(SP)
#endif
#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)
#ifdef TRMMKERNEL
	lwz	r23,  176(SP)
	lwz	r22,  180(SP)
#endif
#endif

	addi	SP, SP, STACKSIZE

	blr

	EPILOGUE
#endif