zgemv_t.s

Optimized GotoBLAS libraries

	add	A2, 6 * SIZE, A2
	FMUL	a5, b1, t1
	nop

	FADDX	c6, t2, c6
	nop
	FMUL	a5, b2, t2
	LDF	[A3 + 4 * SIZE], a5

	FADD	c7, t3, c7
	nop
	FMUL	a6, b1, t3
	nop

	FADD	c8, t4, c8
	nop
	FMUL	a6, b2, t4
	LDF	[A3 + 5 * SIZE], a6

	FADD	c9, t1, c9
	add	A3, 6 * SIZE, A3
	FMUL	a7, b1, t1
	nop

	FADDX	c10, t2, c10
	nop
	FMUL	a7, b2, t2
	LDF	[A4 + 4 * SIZE], a7

	FADD	c11, t3, c11
	nop
	FMUL	a8, b1, t3
	nop

	FADD	c12, t4, c12
	nop
	FMUL	a8, b2, t4
	LDF	[A4 + 5 * SIZE], a8

	FADD	c13, t1, c13
	add	A4, 6 * SIZE, A4
	FMUL	a1, b3, t1
	nop

	FADDX	c14, t2, c14
	nop
	FMUL	a1, b4, t2
	nop

	FADD	c15, t3, c15
	FMUL	a2, b3, t3
	FADD	c16, t4, c16
	FMUL	a2, b4, t4

	FADD	c1, t1, c1
	FMUL	a3, b3, t1
	FADDX	c2, t2, c2
	FMUL	a3, b4, t2
	FADD	c3, t3, c3
	FMUL	a4, b3, t3
	FADD	c4, t4, c4
	FMUL	a4, b4, t4

	FADD	c5, t1, c5
	FMUL	a5, b3, t1
	FADDX	c6, t2, c6
	FMUL	a5, b4, t2
	FADD	c7, t3, c7
	FMUL	a6, b3, t3
	FADD	c8, t4, c8
	FMUL	a6, b4, t4

	FADD	c9, t1, c9
	FMUL	a7, b3, t1
	FADDX	c10, t2, c10
	FMUL	a7, b4, t2
	FADD	c11, t3, c11
	FMUL	a8, b3, t3
	FADD	c12, t4, c12
	FMUL	a8, b4, t4

.LL115:
	andcc	MIN_M, 3, I
	LDF	ALPHA_R, b3
	mov	Y1, Y2
	ble,pn	%icc, .LL119
	LDF	ALPHA_I, b4

.L116:
	LDF	[A1 + 0 * SIZE], a1
	LDF	[A1 + 1 * SIZE], a2
	add	A1, 2 * SIZE, A1
	LDF	[X1 + 0 * SIZE], b1
	LDF	[X1 + 1 * SIZE], b2
	add	X1, 2 * SIZE, X1
	LDF	[A2 + 0 * SIZE], a3
	LDF	[A2 + 1 * SIZE], a4
	add	A2, 2 * SIZE, A2
	LDF	[A3 + 0 * SIZE], a5
	LDF	[A3 + 1 * SIZE], a6
	add	A3, 2 * SIZE, A3
	LDF	[A4 + 0 * SIZE], a7
	LDF	[A4 + 1 * SIZE], a8
	add	A4, 2 * SIZE, A4

	FADD	c13, t1, c13
	FMUL	a1, b1, t1
	FADDX	c14, t2, c14
	FMUL	a1, b2, t2
	FADD	c15, t3, c15
	FMUL	a2, b1, t3
	FADD	c16, t4, c16
	FMUL	a2, b2, t4

	FADD	c1, t1, c1
	FMUL	a3, b1, t1
	FADDX	c2, t2, c2
	FMUL	a3, b2, t2
	FADD	c3, t3, c3
	FMUL	a4, b1, t3
	FADD	c4, t4, c4
	FMUL	a4, b2, t4

	FADD	c5, t1, c5
	FMUL	a5, b1, t1
	FADDX	c6, t2, c6
	FMUL	a5, b2, t2
	FADD	c7, t3, c7
	FMUL	a6, b1, t3
	FADD	c8, t4, c8
	FMUL	a6, b2, t4

	FADD	c9, t1, c9
	FMUL	a7, b1, t1
	FADDX	c10, t2, c10
	FMUL	a7, b2, t2
	FADD	c11, t3, c11
	FMUL	a8, b1, t3
	FADD	c12, t4, c12
	FMUL	a8, b2, t4

	deccc	I
	bg	%icc, .L116
	nop

.LL119:
	FADD	c13, t1, c13
	LDF	[Y1 + 0 * SIZE], a1
	FADDX	c14, t2, c14
	LDF	[Y1 + 1 * SIZE] ,a2
	add	Y1, INCY, Y1
	FADD	c15, t3, c15
	LDF	[Y1 + 0 * SIZE], a3
	FADD	c16, t4, c16
	LDF	[Y1 + 1 * SIZE] ,a4
	add	Y1, INCY, Y1

#if (!defined(XCONJ) && !defined(CONJ)) || (defined(XCONJ) && defined(CONJ))
	FSUB	c1,  c4,  c1
	LDF	[Y1 + 0 * SIZE], a5
	FSUB	c5,  c8,  c5
	LDF	[Y1 + 1 * SIZE] ,a6
	add	Y1, INCY, Y1
	FSUB	c9,  c12, c9
	LDF	[Y1 + 0 * SIZE], a7
	FSUB	c13, c16, c13
	LDF	[Y1 + 1 * SIZE] ,a8
	add	Y1, INCY, Y1
#else
	FADD	c1,  c4,  c1
	LDF	[Y1 + 0 * SIZE], a5
	FADD	c5,  c8,  c5
	LDF	[Y1 + 1 * SIZE] ,a6
	add	Y1, INCY, Y1
	FADD	c9,  c12, c9
	LDF	[Y1 + 0 * SIZE], a7
	FADD	c13, c16, c13
	LDF	[Y1 + 1 * SIZE] ,a8
	add	Y1, INCY, Y1
#endif

#ifndef CONJ
	FADD	c2,  c3,  c2
	LDF	FZERO, t1
	FADD	c6,  c7,  c6
	FADD	c10, c11, c10
	FADD	c14, c15, c14
#else
	FSUB	c2,  c3,  c2
	LDF	FZERO, t1
	FSUB	c6,  c7,  c6
	FSUB	c10, c11, c10
	FSUB	c14, c15, c14
#endif

	FMUL	b3, c1, c3
	FMOV	t1, t2
	FMUL	b4, c1, c4
	FMOV	t1, t3
	FMUL	b4, c2, c1
	FMOV	t1, t4
	FMUL	b3, c2, c2

	FMUL	b3, c5, c7
	FMUL	b4, c5, c8
	FMUL	b4, c6, c5
	FMUL	b3, c6, c6

	FMUL	b3, c9,  c11
	FMUL	b4, c9,  c12
	FMUL	b4, c10, c9
	FMUL	b3, c10, c10

	FMUL	b3, c13, c15
	FSUB	c3,  c1,  c1
	FMUL	b4, c13, c16
	FADD	c2,  c4,  c2
	FMUL	b4, c14, c13
	FSUB	c7,  c5,  c5
	FMUL	b3, c14, c14
	FADD	c6,  c8,  c6

	FSUB	c11, c9,  c9
	FADD	c10, c12, c10
	FSUB	c15, c13, c13
	FADD	c14, c16, c14

	FADD	a1, c1, a1
	FADD	a2, c2, a2
	FADD	a3, c5, a3
	FADD	a4, c6, a4

	STF	a1, [Y2 + 0 * SIZE]
	FADD	a5, c9,  a5
	STF	a2, [Y2 + 1 * SIZE]
	FADD	a6, c10, a6
	add	Y2, INCY, Y2
	STF	a3, [Y2 + 0 * SIZE]
	FADD	a7, c13, a7
	STF	a4, [Y2 + 1 * SIZE]
	FADD	a8, c14, a8
	add	Y2, INCY, Y2

	STF	a5, [Y2 + 0 * SIZE]
	FMOV	t1, c1
	add	J, -1, J
	STF	a6, [Y2 + 1 * SIZE]
	FMOV	t1, c2
	cmp	J, 0
	add	Y2, INCY, Y2
	STF	a7, [Y2 + 0 * SIZE]
	FMOV	t1, c3
	STF	a8, [Y2 + 1 * SIZE]
	FMOV	t1, c4
	add	Y2, INCY, Y2

	FMOV	t1, c5
	bg	%icc, .LL110
	FMOV	t1, c6

.LL200:
	LDF	FZERO, t1

	and	N, 2, J
	cmp	J, 0
	FMOV	t1, c1
	ble	%icc, .LL300

	FMOV	t1, c2
	sra	MIN_M, 2, I
	FMOV	t1, t2
	add	A,  LDA, A2
	FMOV	t1, c3
	mov	A,  A1
	FMOV	t1, t3
	cmp	I, 0
	FMOV	t1, c4

	FMOV	t1, c5
	FMOV	t1, c6
	FMOV	t1, c7
	FMOV	t1, c8

	add	A2, LDA, A
	FMOV	t1, t4
	ble	%icc, .LL215
	mov	XP, X1

	LDF	[A1 + 0 * SIZE], a1
	LDF	[A1 + 1 * SIZE], a2
	LDF	[A1 + 2 * SIZE], a5
	LDF	[A1 + 3 * SIZE], a6
	add	A1, 4 * SIZE, A1

	LDF	[A2 + 0 * SIZE], a3
	LDF	[A2 + 1 * SIZE], a4
	LDF	[A2 + 2 * SIZE], a7
	LDF	[A2 + 3 * SIZE], a8
	add	A2, 4 * SIZE, A2

	LDF	[X1 + 0 * SIZE], b1
	add	I, -1, I
	LDF	[X1 + 1 * SIZE], b2
	cmp	I, 0
	LDF	[X1 + 2 * SIZE], b3
	LDF	[X1 + 3 * SIZE], b4
	ble	%icc, .LL212
	add	X1, 4 * SIZE, X1

.LL211:
	prefetch [A1 + PREFETCHSIZE * SIZE], 1

	FADD	c5, t1, c5
	FMUL	a1, b1, t1
	FADDX	c6, t2, c6
	FMUL	a1, b2, t2
	LDF	[A1 + 0 * SIZE], a1
	FADD	c7, t3, c7
	FMUL	a2, b1, t3
	FADD	c8, t4, c8
	FMUL	a2, b2, t4
	LDF	[A1 + 1 * SIZE], a2

	FADD	c1, t1, c1
	FMUL	a3, b1, t1
	FADDX	c2, t2, c2
	FMUL	a3, b2, t2
	LDF	[A2 + 0 * SIZE], a3
	FADD	c3, t3, c3
	FMUL	a4, b1, t3
	LDF	[X1 + 0 * SIZE], b1
	FADD	c4, t4, c4
	FMUL	a4, b2, t4
	LDF	[A2 + 1 * SIZE], a4

	FADD	c5, t1, c5
	LDF	[X1 + 1 * SIZE], b2
	FMUL	a5, b3, t1
	FADDX	c6, t2, c6
	FMUL	a5, b4, t2
	LDF	[A1 + 2 * SIZE], a5
	FADD	c7, t3, c7
	add	I, -1, I
	FMUL	a6, b3, t3
	FADD	c8, t4, c8
	cmp	I, 0
	FMUL	a6, b4, t4
	LDF	[A1 + 3 * SIZE], a6

	FADD	c1, t1, c1
	FMUL	a7, b3, t1
	FADDX	c2, t2, c2
	FMUL	a7, b4, t2
	LDF	[A2 + 2 * SIZE], a7
	FADD	c3, t3, c3
	FMUL	a8, b3, t3
	LDF	[X1 + 2 * SIZE], b3
	FADD	c4, t4, c4
	FMUL	a8, b4, t4
	LDF	[A2 + 3 * SIZE], a8

	prefetch [A2 + PREFETCHSIZE * SIZE], 1
	FADD	c5, t1, c5
	LDF	[X1 + 3 * SIZE], b4
	FMUL	a1, b1, t1
	FADDX	c6, t2, c6
	FMUL	a1, b2, t2
	LDF	[A1 + 4 * SIZE], a1
	FADD	c7, t3, c7
	FMUL	a2, b1, t3
	FADD	c8, t4, c8
	FMUL	a2, b2, t4
	LDF	[A1 + 5 * SIZE], a2

	FADD	c1, t1, c1
	FMUL	a3, b1, t1
	FADDX	c2, t2, c2
	FMUL	a3, b2, t2
	LDF	[A2 + 4 * SIZE], a3
	FADD	c3, t3, c3
	FMUL	a4, b1, t3
	LDF	[X1 + 4 * SIZE], b1
	FADD	c4, t4, c4
	FMUL	a4, b2, t4
	LDF	[A2 + 5 * SIZE], a4

	FADD	c5, t1, c5
	LDF	[X1 + 5 * SIZE], b2
	FMUL	a5, b3, t1
	FADDX	c6, t2, c6
	FMUL	a5, b4, t2
	LDF	[A1 + 6 * SIZE], a5
	FADD	c7, t3, c7
	FMUL	a6, b3, t3
	FADD	c8, t4, c8
	FMUL	a6, b4, t4
	LDF	[A1 + 7 * SIZE], a6
	add	A1, 8 * SIZE, A1

	FADD	c1, t1, c1
	FMUL	a7, b3, t1
	FADDX	c2, t2, c2
	FMUL	a7, b4, t2
	LDF	[A2 + 6 * SIZE], a7
	FADD	c3, t3, c3
	FMUL	a8, b3, t3
	LDF	[X1 + 6 * SIZE], b3
	FADD	c4, t4, c4
	add	X1, 8 * SIZE, X1
	FMUL	a8, b4, t4
	LDF	[A2 + 7 * SIZE], a8
	add	A2, 8 * SIZE, A2
	bg,pn	%icc, .LL211
	LDF	[X1 - 1 * SIZE], b4

.LL212:
	FADD	c5, t1, c5
	FMUL	a1, b1, t1
	FADDX	c6, t2, c6
	FMUL	a1, b2, t2
	LDF	[A1 + 0 * SIZE], a1
	FADD	c7, t3, c7
	FMUL	a2, b1, t3
	FADD	c8, t4, c8
	FMUL	a2, b2, t4
	LDF	[A1 + 1 * SIZE], a2

	FADD	c1, t1, c1
	FMUL	a3, b1, t1
	FADDX	c2, t2, c2
	FMUL	a3, b2, t2
	LDF	[A2 + 0 * SIZE], a3
	FADD	c3, t3, c3
	FMUL	a4, b1, t3
	LDF	[X1 + 0 * SIZE], b1
	FADD	c4, t4, c4
	FMUL	a4, b2, t4
	LDF	[A2 + 1 * SIZE], a4

	FADD	c5, t1, c5
	LDF	[X1 + 1 * SIZE], b2
	FMUL	a5, b3, t1
	FADDX	c6, t2, c6
	FMUL	a5, b4, t2
	LDF	[A1 + 2 * SIZE], a5
	FADD	c7, t3, c7
	FMUL	a6, b3, t3
	FADD	c8, t4, c8
	FMUL	a6, b4, t4
	LDF	[A1 + 3 * SIZE], a6
	add	A1, 4 * SIZE, A1

	FADD	c1, t1, c1
	FMUL	a7, b3, t1
	FADDX	c2, t2, c2
	FMUL	a7, b4, t2
	LDF	[A2 + 2 * SIZE], a7
	FADD	c3, t3, c3
	FMUL	a8, b3, t3
	LDF	[X1 + 2 * SIZE], b3
	FADD	c4, t4, c4
	FMUL	a8, b4, t4
	LDF	[A2 + 3 * SIZE], a8
	add	A2, 4 * SIZE, A2

	FADD	c5, t1, c5
	LDF	[X1 + 3 * SIZE], b4
	add	X1, 4 * SIZE, X1
	FMUL	a1, b1, t1
	FADDX	c6, t2, c6
	FMUL	a1, b2, t2
	FADD	c7, t3, c7
	FMUL	a2, b1, t3
	FADD	c8, t4, c8
	FMUL	a2, b2, t4

	FADD	c1, t1, c1
	FMUL	a3, b1, t1
	FADDX	c2, t2, c2
	FMUL	a3, b2, t2
	FADD	c3, t3, c3
	FMUL	a4, b1, t3
	FADD	c4, t4, c4
	FMUL	a4, b2, t4

	FADD	c5, t1, c5
	FMUL	a5, b3, t1
	FADDX	c6, t2, c6
	FMUL	a5, b4, t2
	FADD	c7, t3, c7
	FMUL	a6, b3, t3
	FADD	c8, t4, c8
	FMUL	a6, b4, t4

	FADD	c1, t1, c1
	FMUL	a7, b3, t1
	FADDX	c2, t2, c2
	FMUL	a7, b4, t2
	FADD	c3, t3, c3
	FMUL	a8, b3, t3
	FADD	c4, t4, c4
	FMUL	a8, b4, t4

.LL215:
	andcc	MIN_M, 3, I
	LDF	ALPHA_R, b3
	mov	Y1, Y2
	ble	%icc, .LL219
	LDF	ALPHA_I, b4

	LDF	[A1 + 0 * SIZE], a1
	add	I, -1, I
	LDF	[A1 + 1 * SIZE], a2
	cmp	I, 0
	add	A1, 2 * SIZE, A1

	LDF	[A2 + 0 * SIZE], a3
	LDF	[A2 + 1 * SIZE], a4
	add	A2, 2 * SIZE, A2

	LDF	[X1 + 0 * SIZE], b1
	LDF	[X1 + 1 * SIZE], b2
	ble	%icc, .LL217
	add	X1, 2 * SIZE, X1

.LL216:
	FADD	c5, t1, c5
	FMUL	a1, b1, t1
	FADDX	c6, t2, c6
	FMUL	a1, b2, t2
	LDF	[A1 + 0 * SIZE], a1
	FADD	c7, t3, c7
	add	I, -1, I
	FMUL	a2, b1, t3
	FADD	c8, t4, c8
	cmp	I, 0
	FMUL	a2, b2, t4
	LDF	[A1 + 1 * SIZE], a2
	add	A1, 2 * SIZE, A1

	FADD	c1, t1, c1
	FMUL	a3, b1, t1
	FADDX	c2, t2, c2
	FMUL	a3, b2, t2
	LDF	[A2 + 0 * SIZE], a3
	FADD	c3, t3, c3
	FMUL	a4, b1, t3
	LDF	[X1 + 0 * SIZE], b1
	FADD	c4, t4, c4
	add	X1, 2 * SIZE, X1
	FMUL	a4, b2, t4
	LDF	[A2 + 1 * SIZE], a4
	add	A2, 2 * SIZE, A2
	bg,pn	%icc, .LL216
	LDF	[X1 - 1 * SIZE], b2

.LL217:
	FADD	c5, t1, c5
	FMUL	a1, b1, t1
	FADDX	c6, t2, c6
	FMUL	a1, b2, t2
	FADD	c7, t3, c7
	FMUL	a2, b1, t3
	FADD	c8, t4, c8
	FMUL	a2, b2, t4

	FADD	c1, t1, c1
	FMUL	a3, b1, t1
	FADDX	c2, t2, c2
	FMUL	a3, b2, t2
	FADD	c3, t3, c3
	FMUL	a4, b1, t3
	FADD	c4, t4, c4
	FMUL	a4, b2, t4

.LL219:
	FADD	c5, t1, c5
	LDF	[Y1 + 0 * SIZE], a1
	FADDX	c6, t2, c6
	LDF	[Y1 + 1 * SIZE] ,a2
	add	Y1, INCY, Y1
	FADD	c7, t3, c7
	LDF	[Y1 + 0 * SIZE], a3
	FADD	c8, t4, c8
	LDF	[Y1 + 1 * SIZE] ,a4
	add	Y1, INCY, Y1

#if (!defined(XCONJ) && !defined(CONJ)) || (defined(XCONJ) && defined(CONJ))
	FSUB	c1, c4, c1
	FSUB	c5, c8, c5
#else
	FADD	c1, c4, c1
	FADD	c5, c8, c5
#endif

#ifndef CONJ
	FADD	c2, c3, c2
	FADD	c6, c7, c6
#else
	FSUB	c2, c3, c2
	FSUB	c6, c7, c6
#endif

	FMUL	b3, c1, c3
	FMUL	b4, c1, c4
	FMUL	b4, c2, c1
	FMUL	b3, c2, c2

	FMUL	b3, c5, c7
	FMUL	b4, c5, c8
	FMUL	b4, c6, c5
	FMUL	b3, c6, c6

	FSUB	c3, c1, c1
	FADD	c2, c4, c2
	FSUB	c7, c5, c5
	FADD	c6, c8, c6

	FADD	a1, c1, a1
	FADD	a2, c2, a2
	FADD	a3, c5, a3
	FADD	a4, c6, a4

	STF	a1, [Y2 + 0 * SIZE]
	STF	a2, [Y2 + 1 * SIZE]
	add	Y2, INCY, Y2
	STF	a3, [Y2 + 0 * SIZE]
	STF	a4, [Y2 + 1 * SIZE]
	
.LL300:
	andcc	N, 1, J
	LDF	FZERO, t1
	ble	%icc, .LL400
	FMOV	t1, c1

.LL310:
	sra	MIN_M, 2, I
	FMOV	t1, c2
	FMOV	t1, c3
	FMOV	t1, c4
	mov	A, A1
	FMOV	t1, t2
	add	A, LDA, A
	FMOV	t1, t3
	cmp	I, 0
	FMOV	t1, t4
	ble	%icc, .LL315
	mov	XP, X1

	LDF	[A1 + 0 * SIZE], a1
	LDF	[A1 + 1 * SIZE], a2
	LDF	[A1 + 2 * SIZE], a3
	LDF	[A1 + 3 * SIZE], a4
	LDF	[A1 + 4 * SIZE], a5
	LDF	[A1 + 5 * SIZE], a6
	LDF	[A1 + 6 * SIZE], a7
	LDF	[A1 + 7 * SIZE], a8
	add	A1, 8 * SIZE, A1

	LDF	[X1 + 0 * SIZE], c9
	add	I, -1, I
	LDF	[X1 + 1 * SIZE], c10
	cmp	I, 0
	LDF	[X1 + 2 * SIZE], c11
	LDF	[X1 + 3 * SIZE], c12
	LDF	[X1 + 4 * SIZE], c13
	LDF	[X1 + 5 * SIZE], c14
	LDF	[X1 + 6 * SIZE], c15
	LDF	[X1 + 7 * SIZE], c16
	ble	%icc, .LL312
	add	X1, 8 * SIZE, X1

.LL311:
	prefetch [A1 + PREFETCHSIZE * SIZE], 1

	FADD	c1, t1, c1
	FMUL	a1, c9,  t1
	FADDX	c2, t2, c2
	FMUL	a1, c10, t2
	LDF	[A1 + 0 * SIZE], a1
	FADD	c3, t3, c3
	FMUL	a2, c9,  t3
	LDF	[X1 + 0 * SIZE], c9
	FADD	c4, t4, c4
	FMUL	a2, c10, t4
	LDF	[A1 + 1 * SIZE], a2
	LDF	[X1 + 1 * SIZE], c10

	FADD	c1, t1, c1
	FMUL	a3, c11, t1
	FADDX	c2, t2, c2
	FMUL	a3, c12, t2
	LDF	[A1 + 2 * SIZE], a3
	FADD	c3, t3, c3
	add	I, -1, I
	FMUL	a4, c11, t3
	LDF	[X1 + 2 * SIZE], c11
	FADD	c4, t4, c4
	cmp	I, 0
	FMUL	a4, c12, t4
	LDF	[A1 + 3 * SIZE], a4
	LDF	[X1 + 3 * SIZE], c12

	FADD	c1, t1, c1
	FMUL	a5, c13, t1
	FADDX	c2, t2, c2
	FMUL	a5, c14, t2
	LDF	[A1 + 4 * SIZE], a5
	FADD	c3, t3, c3
	FMUL	a6, c13, t3
	LDF	[X1 + 4 * SIZE], c13
	FADD	c4, t4, c4
	FMUL	a6, c14, t4
	LDF	[A1 + 5 * SIZE], a6
	LDF	[X1 + 5 * SIZE], c14

	FADD	c1, t1, c1
	FMUL	a7, c15, t1
	FADDX	c2, t2, c2
	FMUL	a7, c16, t2
	LDF	[A1 + 6 * SIZE], a7

	FADD	c3, t3, c3
	FMUL	a8, c15, t3
	LDF	[X1 + 6 * SIZE], c15
	FADD	c4, t4, c4
	add	X1, 8 * SIZE, X1
	FMUL	a8, c16, t4
	LDF	[A1 + 7 * SIZE], a8
	add	A1, 8 * SIZE, A1
	bg,pn	%icc, .LL311
	LDF	[X1 - 1 * SIZE], c16

.LL312:
	FADD	c1, t1, c1
	FMUL	a1, c9,  t1
	FADDX	c2, t2, c2
	FMUL	a1, c10, t2
	FADD	c3, t3, c3
	FMUL	a2, c9,  t3
	FADD	c4, t4, c4
	FMUL	a2, c10, t4

	FADD	c1, t1, c1
	FMUL	a3, c11, t1
	FADDX	c2, t2, c2
	FMUL	a3, c12, t2
	FADD	c3, t3, c3
	FMUL	a4, c11, t3
	FADD	c4, t4, c4
	FMUL	a4, c12, t4

	FADD	c1, t1, c1
	FMUL	a5, c13, t1
	FADDX	c2, t2, c2
	FMUL	a5, c14, t2
	FADD	c3, t3, c3
	FMUL	a6, c13, t3
	FADD	c4, t4, c4
	FMUL	a6, c14, t4

	FADD	c1, t1, c1
	FMUL	a7, c15, t1
	FADDX	c2, t2, c2
	FMUL	a7, c16, t2
	FADD	c3, t3, c3
	FMUL	a8, c15, t3
	FADD	c4, t4, c4
	FMUL	a8, c16, t4

.LL315:
	andcc	MIN_M, 3, I
	LDF	ALPHA_R, b3
	mov	Y1, Y2
	ble	%icc, .LL319
	LDF	ALPHA_I, b4

	LDF	[A1 + 0 * SIZE], a1
	add	I, -1, I
	LDF	[A1 + 1 * SIZE], a2
	add	A1, 2 * SIZE, A1
	LDF	[X1 + 0 * SIZE], b1
	cmp	I, 0
	LDF	[X1 + 1 * SIZE], b2
	ble	%icc, .LL317
	add	X1, 2 * SIZE, X1

.LL316:
	FADD	c1, t1, c1
	add	I, -1, I
	FMUL	a1, b1, t1
	FADDX	c2, t2, c2
	FMUL	a1, b2, t2
	LDF	[A1 + 0 * SIZE], a1
	FADD	c3, t3, c3
	cmp	I, 0
	FMUL	a2, b1, t3
	LDF	[X1 + 0 * SIZE], b1
	FADD	c4, t4, c4
	add	X1, 2 * SIZE, X1
	FMUL	a2, b2, t4
	LDF	[A1 + 1 * SIZE], a2
	add	A1, 2 * SIZE, A1

	bg,pn	%icc, .LL316
	LDF	[X1 - 1 * SIZE], b2

.LL317:
	FADD	c1, t1, c1
	FMUL	a1, b1, t1
	FADDX	c2, t2, c2
	FMUL	a1, b2, t2
	FADD	c3, t3, c3
	FMUL	a2, b1, t3
	FADD	c4, t4, c4
	FMUL	a2, b2, t4

.LL319:
	FADD	c1, t1, c1
	LDF	[Y1 + 0 * SIZE], a1
	FADDX	c2, t2, c2
	LDF	[Y1 + 1 * SIZE] ,a2
	add	Y1, INCY, Y1
	FADD	c3, t3, c3
	FADD	c4, t4, c4

#if (!defined(XCONJ) && !defined(CONJ)) || (defined(XCONJ) && defined(CONJ))
	FSUB	c1, c4, c1
#else
	FADD	c1, c4, c1
#endif

#ifndef CONJ
	FADD	c2, c3, c2
#else
	FSUB	c2, c3, c2
#endif

	FMUL	b3, c1, c3
	FMUL	b4, c1, c4
	FMUL	b4, c2, c1
	FMUL	b3, c2, c2

	FSUB	c3, c1, c1
	FADD	c2, c4, c2
	FADD	a1, c1, a1
	FADD	a2, c2, a2

	STF	a1, [Y2 + 0 * SIZE]
	STF	a2, [Y2 + 1 * SIZE]

.LL400:
	mov	P, I
	add	IS, I, IS
	cmp	IS, M
	bl	%icc, .LL10
	add	A, PNLDA, A
		
.LL999:
	return	%i7 + 8
	clr	%o0

	EPILOGUE