zsymv_l.s

Optimized GotoBLAS libraries

	FMADD1	xsum3, xtemp4, a8,  xsum3
	LFD	xtemp4,  3 * SIZE(XX)
	FNMSUB	y07, atemp4, a8, y07
	NOP2

	FMADD2	xsum4, xtemp3, a8,  xsum4
	LFD	xtemp3,  2 * SIZE(XX)
	FMADD	y08, atemp3, a8, y08
	LFD	a8,  3 * SIZE(AO2)

	STFD	y05, -4 * SIZE(YY)
	STFD	y06, -3 * SIZE(YY)
	STFD	y07, -2 * SIZE(YY)
	STFD	y08, -1 * SIZE(YY)
	.align 4

LL(15):
	andi.	r0,  TEMP, 4
	ble	LL(16)

	FMADD	xsum1, xtemp1, a1,  xsum1
	NOP1
	FMADD	y01, atemp1, a1, y01
	NOP2

	FMADD	xsum2, xtemp2, a1,  xsum2
	NOP1
	FMADD	y02, atemp2, a1, y02
	LFD	a1,  4 * SIZE(AO1)

	FMADD	xsum3, xtemp1, a5,  xsum3
	NOP1
	FMADD	y03, atemp1, a3, y03
	NOP2

	FMADD	xsum4, xtemp2, a5,  xsum4
	NOP1
	FMADD	y04, atemp2, a3, y04
	NOP2

	FMADD1	xsum1, xtemp2, a2,  xsum1
	LFD	y05,  4 * SIZE(YY)
	FNMSUB	y01, atemp2, a2, y01
	NOP2

	FMADD2	xsum2, xtemp1, a2,  xsum2
	LFD	y06,  5 * SIZE(YY)
	FMADD	y02, atemp1, a2, y02
	LFD	a2,  5 * SIZE(AO1)

	FMADD1	xsum3, xtemp2, a6,  xsum3
	LFD	xtemp2,  5 * SIZE(XX)
	FNMSUB	y03, atemp2, a4, y03
	NOP2

	FMADD2	xsum4, xtemp1, a6,  xsum4
	LFD	xtemp1,  4 * SIZE(XX)
	FMADD	y04, atemp1, a4, y04
	NOP2

	FMADD	xsum1, xtemp3, a3,  xsum1
	LFD	y07,  6 * SIZE(YY)
	FMADD	y01, atemp3, a5, y01
	NOP2

	FMADD	xsum2, xtemp4, a3,  xsum2
	LFD	a3,  6 * SIZE(AO1)
	FMADD	y02, atemp4, a5, y02
	LFD	a5,  4 * SIZE(AO2)

	FMADD	xsum3, xtemp3, a7,  xsum3
	LFD	y08,  7 * SIZE(YY)
	FMADD	y03, atemp3, a7, y03
	NOP2

	FMADD	xsum4, xtemp4, a7,  xsum4
	NOP1
	FMADD	y04, atemp4, a7, y04
	LFD	a7,  6 * SIZE(AO2)

	FMADD1	xsum1, xtemp4, a4,  xsum1
	NOP1
	FNMSUB	y01, atemp4, a6, y01
	NOP2

	FMADD2	xsum2, xtemp3, a4,  xsum2
	LFD	a4,  7 * SIZE(AO1)
	FMADD	y02, atemp3, a6, y02
	LFD	a6,  5 * SIZE(AO2)

	FMADD1	xsum3, xtemp4, a8,  xsum3
	LFD	xtemp4,  7 * SIZE(XX)
	FNMSUB	y03, atemp4, a8, y03
	NOP2

	FMADD2	xsum4, xtemp3, a8,  xsum4
	LFD	xtemp3,  6 * SIZE(XX)
	FMADD	y04, atemp3, a8, y04
	LFD	a8,  7 * SIZE(AO2)

	FMADD	xsum1, xtemp1, a1,  xsum1
	STFD	y01,  0 * SIZE(YY)
	FMADD	y05, atemp1, a1, y05
	NOP2

	FMADD	xsum2, xtemp2, a1,  xsum2
	STFD	y02,  1 * SIZE(YY)
	FMADD	y06, atemp2, a1, y06
	LFD	a1,  8 * SIZE(AO1)

	FMADD	xsum3, xtemp1, a5,  xsum3
	STFD	y03,  2 * SIZE(YY)
	FMADD	y07, atemp1, a3, y07
	NOP2

	FMADD	xsum4, xtemp2, a5,  xsum4
	STFD	y04,  3 * SIZE(YY)
	FMADD	y08, atemp2, a3, y08
	NOP2

	FMADD1	xsum1, xtemp2, a2,  xsum1
	LFD	y01,  8 * SIZE(YY)
	FNMSUB	y05, atemp2, a2, y05
	NOP2

	FMADD2	xsum2, xtemp1, a2,  xsum2
	LFD	y02,  9 * SIZE(YY)
	FMADD	y06, atemp1, a2, y06
	LFD	a2,  9 * SIZE(AO1)

	FMADD1	xsum3, xtemp2, a6,  xsum3
	LFD	xtemp2,  9 * SIZE(XX)
	FNMSUB	y07, atemp2, a4, y07
	NOP2

	FMADD2	xsum4, xtemp1, a6,  xsum4
	LFD	xtemp1,  8 * SIZE(XX)
	FMADD	y08, atemp1, a4, y08
	NOP2

	FMADD	xsum1, xtemp3, a3,  xsum1
	LFD	y03, 10 * SIZE(YY)
	FMADD	y05, atemp3, a5, y05
	NOP2

	FMADD	xsum2, xtemp4, a3,  xsum2
	LFD	a3, 10 * SIZE(AO1)
	FMADD	y06, atemp4, a5, y06
	LFD	a5,  8 * SIZE(AO2)

	FMADD	xsum3, xtemp3, a7,  xsum3
	LFD	y04, 11 * SIZE(YY)
	FMADD	y07, atemp3, a7, y07
	NOP2

	FMADD	xsum4, xtemp4, a7,  xsum4
	addi	YY, YY, 8 * SIZE
	FMADD	y08, atemp4, a7, y08
	LFD	a7, 10 * SIZE(AO2)

	FMADD1	xsum1, xtemp4, a4,  xsum1
	addi	AO2, AO2, 8 * SIZE
	FNMSUB	y05, atemp4, a6, y05
	addi	XX, XX, 8 * SIZE

	FMADD2	xsum2, xtemp3, a4,  xsum2
	LFD	a4, 11 * SIZE(AO1)
	FMADD	y06, atemp3, a6, y06
	LFD	a6,  1 * SIZE(AO2)

	FMADD1	xsum3, xtemp4, a8,  xsum3
	LFD	xtemp4,  3 * SIZE(XX)
	FNMSUB	y07, atemp4, a8, y07
	addi	AO1, AO1, 8 * SIZE

	FMADD2	xsum4, xtemp3, a8,  xsum4
	LFD	xtemp3,  2 * SIZE(XX)
	FMADD	y08, atemp3, a8, y08
	LFD	a8,  3 * SIZE(AO2)

	STFD	y05, -4 * SIZE(YY)
	STFD	y06, -3 * SIZE(YY)
	STFD	y07, -2 * SIZE(YY)
	STFD	y08, -1 * SIZE(YY)
	.align 4

LL(16):
	andi.	r0,  TEMP, 2
	ble	LL(17)

	FMADD	xsum1, xtemp1, a1,  xsum1
	NOP1
	FMADD	y01, atemp1, a1, y01
	NOP2

	FMADD	xsum2, xtemp2, a1,  xsum2
	NOP1
	FMADD	y02, atemp2, a1, y02
	LFD	a1,  4 * SIZE(AO1)

	FMADD	xsum3, xtemp1, a5,  xsum3
	FMADD	y03, atemp1, a3, y03
	FMADD	xsum4, xtemp2, a5,  xsum4
	FMADD	y04, atemp2, a3, y04

	FMADD1	xsum1, xtemp2, a2,  xsum1
	NOP1
	FNMSUB	y01, atemp2, a2, y01
	NOP2

	FMADD2	xsum2, xtemp1, a2,  xsum2
	NOP1
	FMADD	y02, atemp1, a2, y02
	LFD	a2,  5 * SIZE(AO1)

	FMADD1	xsum3, xtemp2, a6,  xsum3
	LFD	xtemp2,  5 * SIZE(XX)
	FNMSUB	y03, atemp2, a4, y03
	NOP2

	FMADD2	xsum4, xtemp1, a6,  xsum4
	LFD	xtemp1,  4 * SIZE(XX)
	FMADD	y04, atemp1, a4, y04
	NOP2

	FMADD	xsum1, xtemp3, a3,  xsum1
	NOP1
	FMADD	y01, atemp3, a5, y01
	NOP2

	FMADD	xsum2, xtemp4, a3,  xsum2
	NOP1
	FMADD	y02, atemp4, a5, y02
	LFD	a5,  4 * SIZE(AO2)

	FMADD	xsum3, xtemp3, a7,  xsum3
	FMADD	y03, atemp3, a7, y03
	FMADD	xsum4, xtemp4, a7,  xsum4
	FMADD	y04, atemp4, a7, y04

	FMADD1	xsum1, xtemp4, a4,  xsum1
	NOP1
	FNMSUB	y01, atemp4, a6, y01
	NOP2

	FMADD2	xsum2, xtemp3, a4,  xsum2
	NOP1
	FMADD	y02, atemp3, a6, y02
	LFD	a6,  5 * SIZE(AO2)

	FMADD1	xsum3, xtemp4, a8,  xsum3
	addi	AO1, AO1, 4 * SIZE
	FNMSUB	y03, atemp4, a8, y03
	addi	AO2, AO2, 4 * SIZE
	FMADD2	xsum4, xtemp3, a8,  xsum4
	addi	YY,  YY, 4 * SIZE
	FMADD	y04, atemp3, a8, y04
	NOP2

	STFD	y01, -4 * SIZE(YY)
	LFD	y01,  0 * SIZE(YY)
	STFD	y02, -3 * SIZE(YY)
	LFD	y02,  1 * SIZE(YY)

	STFD	y03, -2 * SIZE(YY)
	STFD	y04, -1 * SIZE(YY)
	.align 4

LL(17):
	andi.	r0,  M, 1
	ble	LL(18)

	FMADD	xsum1, xtemp1, a1,  xsum1
	FMADD	y01, atemp1, a1, y01
	FMADD	xsum2, xtemp2, a1,  xsum2
	FMADD	y02, atemp2, a1, y02
	FMADD	xsum3, xtemp1, a5,  xsum3
	FNMSUB	y01, atemp2, a2, y01
	FMADD	xsum4, xtemp2, a5,  xsum4
	FMADD	y02, atemp1, a2, y02

	FMADD1	xsum1, xtemp2, a2,  xsum1
	FMADD	y01, atemp3, a5, y01
	FMADD2	xsum2, xtemp1, a2,  xsum2
	FMADD	y02, atemp4, a5, y02
	FMADD1	xsum3, xtemp2, a6,  xsum3
	FNMSUB	y01, atemp4, a6, y01
	FMADD2	xsum4, xtemp1, a6,  xsum4
	FMADD	y02, atemp3, a6, y02

	STFD	y01,  0 * SIZE(YY)
	STFD	y02,  1 * SIZE(YY)
	STFD	y03,  2 * SIZE(YY)
	STFD	y04,  3 * SIZE(YY)
	.align 4

LL(18):
	LFD	y05, ALPHA_R
	LFD	y06, ALPHA_I

	slwi	TEMP,  IS,  ZBASE_SHIFT
	add	YY, NEW_Y, TEMP

	LFD	y01,  0 * SIZE(YY)
	LFD	y02,  1 * SIZE(YY)
	LFD	y03,  2 * SIZE(YY)
	LFD	y04,  3 * SIZE(YY)

	FMUL	xtemp1, y05, xsum1
	FMUL	xtemp2, y06, xsum1
	FMUL	xtemp3, y05, xsum3
	FMUL	xtemp4, y06, xsum3

	FNMSUB	xsum1, y06, xsum2, xtemp1
	FMADD	xsum2, y05, xsum2, xtemp2
	FNMSUB	xsum3, y06, xsum4, xtemp3
	FMADD	xsum4, y05, xsum4, xtemp4

	FADD	y01, y01, xsum1
	FADD	y02, y02, xsum2
	FADD	y03, y03, xsum3
	FADD	y04, y04, xsum4

	STFD	y01,  0 * SIZE(YY)
	addi	TEMP, IS, 4
	STFD	y02,  1 * SIZE(YY)
	addi	IS,   IS, 2
	STFD	y03,  2 * SIZE(YY)
	cmpw	cr0, TEMP, M
	STFD	y04,  3 * SIZE(YY)
	ble	LL(11)
	.align 4
	
LL(20):
	andi.	TEMP, M, 1
	ble	LL(990)

	slwi	TEMP,  IS,  ZBASE_SHIFT
	add	XX, X,     TEMP
	add	YY, NEW_Y, TEMP

	LFD	y05, ALPHA_R
	LFD	y06, ALPHA_I

	LFD	atemp1, 0 * SIZE(XX)
	LFD	atemp2, 1 * SIZE(XX)

	LFD	a1,  0 * SIZE(A)
	LFD	a2,  1 * SIZE(A)

	FMUL	xsum1, atemp1, a1
	FMUL	xsum2, atemp2, a1

#ifndef HEMV
	FNMSUB	xsum1, atemp2, a2, xsum1
	FMADD	xsum2, atemp1, a2, xsum2
#endif

	FMUL	xtemp1, y05, atemp1
	FMUL	xtemp2, y06, atemp1

	FNMSUB	atemp1, y06, atemp2, xtemp1
	FMADD	atemp2, y05, atemp2, xtemp2

	LFD	y05, ALPHA_R
	LFD	y06, ALPHA_I

	LFD	y01,  0 * SIZE(YY)
	LFD	y02,  1 * SIZE(YY)

	FMUL	xtemp1, y05, xsum1
	FMUL	xtemp2, y06, xsum1

	FNMSUB	xsum1, y06, xsum2, xtemp1
	FMADD	xsum2, y05, xsum2, xtemp2

	FADD	y01, y01, xsum1
	FADD	y02, y02, xsum2

	STFD	y01,  0 * SIZE(YY)
	STFD	y02,  1 * SIZE(YY)
	.align 4
	
LL(990):
	cmpwi	cr0, INCY, 2 * SIZE
	beq	LL(999)

	mr	YY, Y

	srawi.	r0, M, 2
	mtspr	CTR, r0
	ble	LL(995)
	.align 4

LL(991):
	LFD	f0,  0 * SIZE(Y)
	LFD	f1,  1 * SIZE(Y)
	add	Y, Y, INCY
	LFD	f2,  0 * SIZE(Y)
	LFD	f3,  1 * SIZE(Y)
	add	Y, Y, INCY
	LFD	f4,  0 * SIZE(Y)
	LFD	f5,  1 * SIZE(Y)
	add	Y, Y, INCY
	LFD	f6,  0 * SIZE(Y)
	LFD	f7,  1 * SIZE(Y)
	add	Y, Y, INCY

	LFD	f8,   0 * SIZE(NEW_Y)
	LFD	f9,   1 * SIZE(NEW_Y)
	LFD	f10,  2 * SIZE(NEW_Y)
	LFD	f11,  3 * SIZE(NEW_Y)
	LFD	f12,  4 * SIZE(NEW_Y)
	LFD	f13,  5 * SIZE(NEW_Y)
	LFD	f14,  6 * SIZE(NEW_Y)
	LFD	f15,  7 * SIZE(NEW_Y)
	addi	NEW_Y, NEW_Y, 8 * SIZE

	FADD	f8,  f8,  f0
	FADD	f9,  f9,  f1
	FADD	f10, f10, f2
	FADD	f11, f11, f3
	FADD	f12, f12, f4
	FADD	f13, f13, f5
	FADD	f14, f14, f6
	FADD	f15, f15, f7

	STFD	f8,  0 * SIZE(YY)
	STFD	f9,  1 * SIZE(YY)
	add	YY, YY, INCY
	STFD	f10, 0 * SIZE(YY)
	STFD	f11, 1 * SIZE(YY)
	add	YY, YY, INCY
	STFD	f12, 0 * SIZE(YY)
	STFD	f13, 1 * SIZE(YY)
	add	YY, YY, INCY
	STFD	f14, 0 * SIZE(YY)
	STFD	f15, 1 * SIZE(YY)
	add	YY, YY, INCY
	bdnz	LL(991)
	.align 4

LL(995):
	andi.	J, M, 2
	ble	LL(996)

	LFD	f0,  0 * SIZE(Y)
	LFD	f1,  1 * SIZE(Y)
	add	Y, Y, INCY
	LFD	f2,  0 * SIZE(Y)
	LFD	f3,  1 * SIZE(Y)
	add	Y, Y, INCY

	LFD	f8,   0 * SIZE(NEW_Y)
	LFD	f9,   1 * SIZE(NEW_Y)
	LFD	f10,  2 * SIZE(NEW_Y)
	LFD	f11,  3 * SIZE(NEW_Y)
	addi	NEW_Y, NEW_Y, 4 * SIZE

	FADD	f8,  f8,  f0
	FADD	f9,  f9,  f1
	FADD	f10, f10, f2
	FADD	f11, f11, f3

	STFD	f8,  0 * SIZE(YY)
	STFD	f9,  1 * SIZE(YY)
	add	YY, YY, INCY
	STFD	f10, 0 * SIZE(YY)
	STFD	f11, 1 * SIZE(YY)
	add	YY, YY, INCY
	.align 4

LL(996):
	andi.	J, M, 1
	ble	LL(999)

	LFD	f0,  0 * SIZE(Y)
	LFD	f1,  1 * SIZE(Y)

	LFD	f8,   0 * SIZE(NEW_Y)
	LFD	f9,   1 * SIZE(NEW_Y)

	FADD	f8,  f8,  f0
	FADD	f9,  f9,  f1

	STFD	f8,  0 * SIZE(YY)
	STFD	f9,  1 * SIZE(YY)
	.align 4

LL(999):
	li	r3, 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	r14,   144(SP)
	ld	r15,   152(SP)
	ld	r16,   160(SP)
	ld	r17,   168(SP)
	ld	r18,   176(SP)
	ld	r19,   184(SP)
	ld	r20,   192(SP)
	ld	r21,   200(SP)
	ld	r22,   208(SP)
	ld	r23,   216(SP)
	ld	r24,   224(SP)
	ld	r25,   232(SP)
	ld	r26,   240(SP)
	ld	r27,   248(SP)
#else
	lwz	r14,   144(SP)
	lwz	r15,   148(SP)
	lwz	r16,   152(SP)
	lwz	r17,   156(SP)
	lwz	r18,   160(SP)
	lwz	r19,   164(SP)
	lwz	r20,   168(SP)
	lwz	r21,   172(SP)
	lwz	r22,   176(SP)
	lwz	r23,   180(SP)
	lwz	r24,   184(SP)
	lwz	r25,   188(SP)
	lwz	r26,   192(SP)
	lwz	r27,   196(SP)
#endif

	addi	SP, SP, STACKSIZE
	blr

	EPILOGUE
#endif