h264dsputil.s

君正早期ucos系统(只有早期的才不没有打包成库),MPLAYER,文件系统,图片解码,浏览,电子书,录音,想学ucos,识货的人就下吧 russblock fmradio explore set

	srl	$12,$3,24
	srl	$7,$2,8
	srl	$9,$3,8
	srl	$10,$2,16
	srl	$11,$3,16
	sb	$2,-1($8)
	sb	$3,0($8)
	sb	$7,-1($4)
	sb	$9,0($4)
	sb	$10,-1($6)
	sb	$11,0($6)
	sb	$12,0($5)
	.set	noreorder
	.set	nomacro
	j	$31
	sb	$13,-1($5)
	.set	macro
	.set	reorder

	.end	h264_h_loop_filter_chroma_mxu
	.section	.text.h264_v_loop_filter_chroma_intra_mxu,"ax",@progbits
	.align	2
	.align	5
	.ent	h264_v_loop_filter_chroma_intra_mxu
	.type	h264_v_loop_filter_chroma_intra_mxu, @function
h264_v_loop_filter_chroma_intra_mxu:
	.frame	$sp,0,$31		# vars= 0, regs= 0/0, args= 0, gp= 0
	.mask	0x00000000,0
	.fmask	0x00000000,0
#APP
	S32I2M xr14,$7
	S32SFL xr0,xr14,xr14,xr14,ptn0
	S32SFL xr0,xr14,xr14,xr14,ptn3
	S32I2M xr13,$6
	S32SFL xr0,xr13,xr13,xr13,ptn0
	S32SFL xr0,xr13,xr13,xr13,ptn3
#NO_APP
	li	$2,16842752			# 0x1010000
	ori	$2,$2,0x101
#APP
	S32I2M xr8,$2
#NO_APP
	sll	$3,$5,1
	subu	$2,$4,$3
#APP
	S32LDD xr11,$2,0
	S32LDIV xr10,$2,$5,0
	S32LDIV xr9,$2,$5,0
	S32LDIV xr1,$2,$5,0
	Q8ABD xr3,xr10,xr9
	Q8ABD xr4,xr11,xr10
	Q8ABD xr5,xr1,xr9
	Q8ADDE xr6,xr3,xr13,xr7,SS
	Q8ADDE xr3,xr4,xr14,xr4,SS
	Q16SLR xr6,xr6,xr7,xr7,15
	Q16SLR xr3,xr3,xr4,xr4,15
	Q16SAT xr4,xr3,xr4
	Q16SAT xr3,xr6,xr7
	Q8ADDE xr6,xr5,xr14,xr7,SS
	Q16SLR xr6,xr6,xr7,xr7,15
	Q16SAT xr5,xr6,xr7
	Q8MADL xr0,xr3,xr4,xr3,AA
	Q8ADDE xr12,xr11,xr8,xr7,AA
	Q16SLL xr12,xr12,xr7,xr7,1
	Q8ACCE xr12,xr10,xr1,xr7,AA
	Q8MADL xr0,xr3,xr5,xr3,AA
	Q8ADDE xr2,xr1,xr8,xr6,AA
	Q16SLL xr2,xr2,xr6,xr6,1
	Q8ACCE xr2,xr9,xr11,xr6,AA
	Q16SAR xr12,xr12,xr7,xr7,2
	S32SFL xr0,xr12,xr7,xr12,ptn1
	Q8MADL xr0,xr3,xr12,xr12,AA
	Q16SAR xr2,xr2,xr6,xr6,2
	S32SFL xr0,xr2,xr6,xr2,ptn1
	Q8MADL xr0,xr3,xr2,xr2,AA
	Q8ADD xr4,xr8,xr3,SS
	Q8MADL xr0,xr4,xr10,xr10,AA
	Q8MADL xr0,xr4,xr9,xr9,AA
#NO_APP
	subu	$2,$4,$5
#APP
	Q8ADD xr10,xr10,xr12,AA
	S32STD xr10,$2,0
	Q8ADD xr9,xr9,xr2,AA
	S32STDV xr9,$2,$5,0
#NO_APP
	addiu	$4,$4,4
	subu	$3,$4,$3
#APP
	S32LDD xr11,$3,0
	S32LDIV xr10,$3,$5,0
	S32LDIV xr9,$3,$5,0
	S32LDIV xr1,$3,$5,0
	Q8ABD xr3,xr10,xr9
	Q8ABD xr4,xr11,xr10
	Q8ABD xr5,xr1,xr9
	Q8ADDE xr6,xr3,xr13,xr7,SS
	Q8ADDE xr3,xr4,xr14,xr4,SS
	Q16SLR xr6,xr6,xr7,xr7,15
	Q16SLR xr3,xr3,xr4,xr4,15
	Q16SAT xr4,xr3,xr4
	Q16SAT xr3,xr6,xr7
	Q8ADDE xr6,xr5,xr14,xr7,SS
	Q16SLR xr6,xr6,xr7,xr7,15
	Q16SAT xr5,xr6,xr7
	Q8MADL xr0,xr3,xr4,xr3,AA
	Q8ADDE xr12,xr11,xr8,xr7,AA
	Q16SLL xr12,xr12,xr7,xr7,1
	Q8ACCE xr12,xr10,xr1,xr7,AA
	Q8MADL xr0,xr3,xr5,xr3,AA
	Q8ADDE xr2,xr1,xr8,xr6,AA
	Q16SLL xr2,xr2,xr6,xr6,1
	Q8ACCE xr2,xr9,xr11,xr6,AA
	Q16SAR xr12,xr12,xr7,xr7,2
	S32SFL xr0,xr12,xr7,xr12,ptn1
	Q8MADL xr0,xr3,xr12,xr12,AA
	Q16SAR xr2,xr2,xr6,xr6,2
	S32SFL xr0,xr2,xr6,xr2,ptn1
	Q8MADL xr0,xr3,xr2,xr2,AA
	Q8ADD xr4,xr8,xr3,SS
	Q8MADL xr0,xr4,xr10,xr10,AA
	Q8MADL xr0,xr4,xr9,xr9,AA
#NO_APP
	subu	$4,$4,$5
#APP
	Q8ADD xr10,xr10,xr12,AA
	S32STD xr10,$4,0
	Q8ADD xr9,xr9,xr2,AA
	S32STDV xr9,$4,$5,0
#NO_APP
	j	$31
	.end	h264_v_loop_filter_chroma_intra_mxu
	.section	.text.h264_h_loop_filter_chroma_intra_mxu,"ax",@progbits
	.align	2
	.align	5
	.ent	h264_h_loop_filter_chroma_intra_mxu
	.type	h264_h_loop_filter_chroma_intra_mxu, @function
h264_h_loop_filter_chroma_intra_mxu:
	.frame	$sp,0,$31		# vars= 0, regs= 0/0, args= 0, gp= 0
	.mask	0x00000000,0
	.fmask	0x00000000,0
#APP
	S32I2M xr14,$7
	S32SFL xr0,xr14,xr14,xr14,ptn0
	S32SFL xr0,xr14,xr14,xr14,ptn3
	S32I2M xr13,$6
	S32SFL xr0,xr13,xr13,xr13,ptn0
	S32SFL xr0,xr13,xr13,xr13,ptn3
#NO_APP
	li	$2,16842752			# 0x1010000
	ori	$2,$2,0x101
#APP
	S32I2M xr8,$2
	S32LDD xr2,$4,0
	S32LDD xr1,$4,-4
#NO_APP
	move	$2,$4
#APP
	S32LDIV xr4,$2,$5,0
	S32LDD xr3,$2,-4
	S32SFL xr6,xr4,xr2,xr12,ptn0
	S32SFL xr5,xr3,xr1,xr7,ptn0
	S32LDIV xr2,$2,$5,0
	S32LDD xr1,$2,-4
	S32LDIV xr4,$2,$5,0
	S32LDD xr3,$2,-4
	S32SFL xr4,xr4,xr2,xr2,ptn0
	S32SFL xr3,xr3,xr1,xr1,ptn0
	S32SFL xr10,xr3,xr5,xr11,ptn3
	S32SFL xr1,xr2,xr12,xr9,ptn3
	Q8ABD xr3,xr10,xr9
	Q8ABD xr4,xr11,xr10
	Q8ABD xr5,xr1,xr9
	Q8ADDE xr6,xr3,xr13,xr7,SS
	Q8ADDE xr3,xr4,xr14,xr4,SS
	Q16SLR xr6,xr6,xr7,xr7,15
	Q16SLR xr3,xr3,xr4,xr4,15
	Q16SAT xr4,xr3,xr4
	Q16SAT xr3,xr6,xr7
	Q8ADDE xr6,xr5,xr14,xr7,SS
	Q16SLR xr6,xr6,xr7,xr7,15
	Q16SAT xr5,xr6,xr7
	Q8MADL xr0,xr3,xr4,xr3,AA
	Q8ADDE xr12,xr11,xr8,xr7,AA
	Q16SLL xr12,xr12,xr7,xr7,1
	Q8ACCE xr12,xr10,xr1,xr7,AA
	Q8MADL xr0,xr3,xr5,xr3,AA
	Q8ADDE xr2,xr1,xr8,xr6,AA
	Q16SLL xr2,xr2,xr6,xr6,1
	Q8ACCE xr2,xr9,xr11,xr6,AA
	Q16SAR xr12,xr12,xr7,xr7,2
	S32SFL xr0,xr12,xr7,xr12,ptn1
	Q8MADL xr0,xr3,xr12,xr12,AA
	Q16SAR xr2,xr2,xr6,xr6,2
	S32SFL xr0,xr2,xr6,xr2,ptn1
	Q8MADL xr0,xr3,xr2,xr2,AA
	Q8ADD xr4,xr8,xr3,SS
	Q8MADL xr0,xr4,xr10,xr10,AA
	Q8MADL xr0,xr4,xr9,xr9,AA
	Q8ADD xr10,xr10,xr12,AA
	Q8ADD xr9,xr9,xr2,AA
	S32M2I xr10, $2
	S32M2I xr9, $3
#NO_APP
	addu	$6,$5,$4
	addu	$7,$5,$6
	addu	$8,$5,$7
	srl	$13,$2,24
	srl	$14,$3,24
	srl	$9,$2,8
	srl	$10,$3,8
	srl	$11,$2,16
	srl	$12,$3,16
	sb	$2,-1($4)
	sb	$3,0($4)
	sb	$9,-1($6)
	sb	$10,0($6)
	sb	$11,-1($7)
	sb	$12,0($7)
	sb	$13,-1($8)
	sb	$14,0($8)
	addu	$8,$5,$8
#APP
	S32LDD xr2,$8,0
	S32LDD xr1,$8,-4
#NO_APP
	move	$2,$8
#APP
	S32LDIV xr4,$2,$5,0
	S32LDD xr3,$2,-4
	S32SFL xr6,xr4,xr2,xr12,ptn0
	S32SFL xr5,xr3,xr1,xr7,ptn0
	S32LDIV xr2,$2,$5,0
	S32LDD xr1,$2,-4
	S32LDIV xr4,$2,$5,0
	S32LDD xr3,$2,-4
	S32SFL xr4,xr4,xr2,xr2,ptn0
	S32SFL xr3,xr3,xr1,xr1,ptn0
	S32SFL xr10,xr3,xr5,xr11,ptn3
	S32SFL xr1,xr2,xr12,xr9,ptn3
	Q8ABD xr3,xr10,xr9
	Q8ABD xr4,xr11,xr10
	Q8ABD xr5,xr1,xr9
	Q8ADDE xr6,xr3,xr13,xr7,SS
	Q8ADDE xr3,xr4,xr14,xr4,SS
	Q16SLR xr6,xr6,xr7,xr7,15
	Q16SLR xr3,xr3,xr4,xr4,15
	Q16SAT xr4,xr3,xr4
	Q16SAT xr3,xr6,xr7
	Q8ADDE xr6,xr5,xr14,xr7,SS
	Q16SLR xr6,xr6,xr7,xr7,15
	Q16SAT xr5,xr6,xr7
	Q8MADL xr0,xr3,xr4,xr3,AA
	Q8ADDE xr12,xr11,xr8,xr7,AA
	Q16SLL xr12,xr12,xr7,xr7,1
	Q8ACCE xr12,xr10,xr1,xr7,AA
	Q8MADL xr0,xr3,xr5,xr3,AA
	Q8ADDE xr2,xr1,xr8,xr6,AA
	Q16SLL xr2,xr2,xr6,xr6,1
	Q8ACCE xr2,xr9,xr11,xr6,AA
	Q16SAR xr12,xr12,xr7,xr7,2
	S32SFL xr0,xr12,xr7,xr12,ptn1
	Q8MADL xr0,xr3,xr12,xr12,AA
	Q16SAR xr2,xr2,xr6,xr6,2
	S32SFL xr0,xr2,xr6,xr2,ptn1
	Q8MADL xr0,xr3,xr2,xr2,AA
	Q8ADD xr4,xr8,xr3,SS
	Q8MADL xr0,xr4,xr10,xr10,AA
	Q8MADL xr0,xr4,xr9,xr9,AA
	Q8ADD xr10,xr10,xr12,AA
	Q8ADD xr9,xr9,xr2,AA
	S32M2I xr10, $2
	S32M2I xr9, $3
#NO_APP
	addu	$4,$5,$8
	addu	$6,$5,$4
	addu	$5,$5,$6
	srl	$13,$2,24
	srl	$12,$3,24
	srl	$7,$2,8
	srl	$9,$3,8
	srl	$10,$2,16
	srl	$11,$3,16
	sb	$2,-1($8)
	sb	$3,0($8)
	sb	$7,-1($4)
	sb	$9,0($4)
	sb	$10,-1($6)
	sb	$11,0($6)
	sb	$12,0($5)
	.set	noreorder
	.set	nomacro
	j	$31
	sb	$13,-1($5)
	.set	macro
	.set	reorder

	.end	h264_h_loop_filter_chroma_intra_mxu
	.section	.text.put_h264_qpel4_hv_lowpass_mxu,"ax",@progbits
	.align	2
	.align	5
	.ent	put_h264_qpel4_hv_lowpass_mxu
	.type	put_h264_qpel4_hv_lowpass_mxu, @function
put_h264_qpel4_hv_lowpass_mxu:
	.frame	$sp,16,$31		# vars= 16, regs= 0/0, args= 0, gp= 0
	.mask	0x00000000,0
	.fmask	0x00000000,0
	addiu	$sp,$sp,-16
	lw	$12,36($sp)
	addiu	$9,$6,-2
	sll	$8,$12,1
	subu	$8,$6,$8
	andi	$9,$9,0x3
	andi	$6,$6,0x3
	li	$3,4			# 0x4
	subu	$11,$3,$6
	li	$10,-4			# 0xfffffffffffffffc
	subu	$3,$3,$9
	addiu	$2,$8,-2
	and	$2,$2,$10
	and	$8,$8,$10
	sw	$3,8($sp)
	li	$3,17104896			# 0x1050000
	sw	$2,0($sp)
	sw	$8,4($sp)
	move	$9,$4
	sw	$11,12($sp)
	ori	$3,$3,0x105
	lw	$6,32($sp)
#APP
	S32I2M xr15,$3
#NO_APP
	li	$2,336855040			# 0x14140000
	ori	$2,$2,0x501
#APP
	S32I2M xr14,$2
#NO_APP
	move	$8,$0
	.set	noreorder
	.set	nomacro
	j	$L43
	addiu	$4,$5,-4
	.set	macro
	.set	reorder

$L48:
	addiu	$4,$4,-8
$L43:
	lw	$2,0($sp)
#APP
	S32LDD xr1,$2,0
	S32LDD xr2,$2,4
	S32LDD xr3,$2,8
#NO_APP
	addu	$2,$12,$2
	sw	$2,0($sp)
#APP
	pref 0,0($2)
#NO_APP
	lw	$3,8($sp)
#APP
	S32ALN xr4,xr2,xr1,$3
	S32ALN xr5,xr3,xr2,$3
#NO_APP
	li	$2,3			# 0x3
#APP
	S32ALN xr6,xr5,xr4,$2
	D32SLR xr1,xr5,xr0,xr0,8
	S32SFL xr0,xr1,xr5,xr7,ptn3
	Q8MUL xr2,xr4,xr14,xr1
	Q8MUL xr8,xr7,xr15,xr7
	Q8MUL xr4,xr6,xr14,xr3
	Q16ADD xr2,xr2,xr2,xr0,AA,XW
	Q16ADD xr0,xr1,xr1,xr1,SS,XW
	Q16ADD xr0,xr7,xr7,xr7,SS,XW
	S32SFL xr1,xr0,xr1,xr0,ptn3
	Q16ACC xr0,xr2,xr7,xr1,AA
	Q16ADD xr4,xr4,xr4,xr0,AA,XW
	Q16ADD xr3,xr3,xr3,xr0,SS,XW
	Q16ADD xr0,xr8,xr8,xr8,SS,XW
	S32SFL xr3,xr0,xr3,xr0,ptn3
	Q16ACC xr0,xr4,xr8,xr3,AA
	S32SFL xr0,xr3,xr1,xr1,ptn3
	S32SDI xr1,$4,4
#NO_APP
	lw	$2,4($sp)
#APP
	S32LDD xr1,$2,0
	S32LDD xr2,$2,4
	S32LDD xr3,$2,8
#NO_APP
	addu	$2,$12,$2
	sw	$2,4($sp)
#APP
	pref 0,0($2)
	S32ALN xr4,xr2,xr1,$11
	S32ALN xr5,xr3,xr2,$11
#NO_APP
	li	$2,3			# 0x3
#APP
	S32ALN xr6,xr5,xr4,$2
	D32SLR xr1,xr5,xr0,xr0,8
	S32SFL xr0,xr1,xr5,xr7,ptn3
	Q8MUL xr2,xr4,xr14,xr1
	Q8MUL xr8,xr7,xr15,xr7
	Q8MUL xr4,xr6,xr14,xr3
	Q16ADD xr2,xr2,xr2,xr0,AA,XW
	Q16ADD xr0,xr1,xr1,xr1,SS,XW
	Q16ADD xr0,xr7,xr7,xr7,SS,XW
	S32SFL xr1,xr0,xr1,xr0,ptn3
	Q16ACC xr0,xr2,xr7,xr1,AA
	Q16ADD xr4,xr4,xr4,xr0,AA,XW
	Q16ADD xr3,xr3,xr3,xr0,SS,XW
	Q16ADD xr0,xr8,xr8,xr8,SS,XW
	S32SFL xr3,xr0,xr3,xr0,ptn3
	Q16ACC xr0,xr4,xr8,xr3,AA
	S32SFL xr0,xr3,xr1,xr1,ptn3
	S32SDI xr1,$4,4
#NO_APP
	li	$2,9			# 0x9
	addiu	$8,$8,1
	.set	noreorder
	.set	nomacro
	bne	$8,$2,$L48
	addu	$4,$6,$4
	.set	macro
	.set	reorder

	li	$2,1310720			# 0x140000
	ori	$2,$2,0x14
#APP
	S32I2M xr15,$2
#NO_APP
	li	$3,327680			# 0x50000
	ori	$3,$3,0x5
#APP
	S32I2M xr14,$3
#NO_APP
	li	$4,512			# 0x200
#APP
	S32I2M xr13,$4
#NO_APP
	li	$2,65536			# 0x10000
	ori	$2,$2,0x1
#APP
	S32I2M xr11,$2
#NO_APP
	li	$8,4			# 0x4
	subu	$4,$5,$6
	move	$5,$0
$L45:
	move	$2,$4
#APP
	S32LDIV xr1,$2,$6,0
	S32LDIV xr2,$2,$6,0
	S32LDIV xr3,$2,$6,0
	D16MUL xr10,xr2,xr14,xr9,WW
	D16MUL xr8,xr3,xr15,xr7,WW
	S32LDIV xr4,$2,$6,0
	S32LDIV xr5,$2,$6,0
	S32LDDV xr12,$2,$6,0
	D16MUL xr3,xr5,xr14,xr2,WW
	D16MUL xr5,xr4,xr15,xr4,WW
	Q16ADD xr12,xr1,xr12,xr0,AA,WW
	D16MUL xr12,xr12,xr11,xr1,WW
	D32ACC xr12,xr8,xr10,xr0,SS
	D32ACC xr1,xr7,xr9,xr0,SS
	D32ACC xr12,xr5,xr3,xr0,SS
	D32ACC xr1,xr4,xr2,xr0,SS
	D32ADD xr12,xr12,xr13,xr0,AA
	D32ADD xr1,xr1,xr13,xr0,AA
	D32SAR xr2,xr12,xr1,xr1,10
	S32SFL xr0,xr2,xr1,xr1,ptn3
	Q16SAT xr1,xr0,xr1
	S32M2I xr1, $2
#NO_APP
	sh	$2,0($9)
	move	$3,$9
	addiu	$2,$4,4
#APP
	S32LDIV xr1,$2,$6,0
	S32LDIV xr2,$2,$6,0
	S32LDIV xr3,$2,$6,0
	D16MUL xr10,xr2,xr14,xr9,WW
	D16MUL xr8,xr3,xr15,xr7,WW
	S32LDIV xr4,$2,$6,0
	S32LDIV xr5,$2,$6,0
	S32LDDV xr12,$2,$6,0
	D16MUL xr3,xr5,xr14,xr2,WW
	D16MUL xr5,xr4,xr15,xr4,WW
	Q16ADD xr12,xr1,xr12,xr0,AA,WW
	D16MUL xr12,xr12,xr11,xr1,WW
	D32ACC xr12,xr8,xr10,xr0,SS
	D32ACC xr1,xr7,xr9,xr0,SS
	D32ACC xr12,xr5,xr3,xr0,SS
	D32ACC xr1,xr4,xr2,xr0,SS
	D32ADD xr12,xr12,xr13,xr0,AA
	D32ADD xr1,xr1,xr13,xr0,AA
	D32SAR xr2,xr12,xr1,xr1,10
	S32SFL xr0,xr2,xr1,xr1,ptn3
	Q16SAT xr1,xr0,xr1
	S32M2I xr1, $2
#NO_APP
	addiu	$5,$5,1
	addu	$9,$9,$7
	sh	$2,2($3)
	.set	noreorder
	.set	nomacro
	bne	$5,$8,$L45
	addu	$4,$4,$6
	.set	macro
	.set	reorder

	.set	noreorder
	.set	nomacro
	j	$31
	addiu	$sp,$sp,16
	.set	macro
	.set	reorder

	.end	put_h264_qpel4_hv_lowpass_mxu
	.section	.text.put_h264_qpel8_hv_lowpass_mxu,"ax",@progbits
	.align	2
	.align	5
	.ent	put_h264_qpel8_hv_lowpass_mxu
	.type	put_h264_qpel8_hv_lowpass_mxu, @function
put_h264_qpel8_hv_lowpass_mxu:
	.frame	$sp,32,$31		# vars= 32, regs= 0/0, args= 0, gp= 0
	.mask	0x00000000,0
	.fmask	0x00000000,0
	addiu	$sp,$sp,-32
	lw	$11,52($sp)
	li	$8,-4			# 0xfffffffffffffffc
	sll	$3,$11,1
	subu	$3,$6,$3
	addiu	$2,$3,-2
	and	$9,$2,$8
	addiu	$2,$6,-2
	and	$3,$3,$8
	andi	$2,$2,0x3
	andi	$6,$6,0x3
	li	$8,4			# 0x4
	subu	$10,$8,$6
	move	$13,$5
	subu	$8,$8,$2
	addiu	$5,$3,4
	addiu	$2,$9,4
	sw	$3,4($sp)
	li	$3,17104896			# 0x1050000
	sw	$9,0($sp)
	sw	$2,8($sp)
	sw	$5,12($sp)
	sw	$8,24($sp)
	move	$12,$4
	sw	$8,16($sp)
	sw	$10,20($sp)
	sw	$10,28($sp)
	ori	$3,$3,0x105
	lw	$6,48($sp)
#APP
	S32I2M xr15,$3
#NO_APP
	li	$2,336855040			# 0x14140000
	ori	$2,$2,0x501
#APP
	S32I2M xr14,$2
#NO_APP
	move	$5,$0
	.set	noreorder
	.set	nomacro
	j	$L52
	addiu	$4,$13,-4
	.set	macro
	.set	reorder

$L57:
	lw	$9,0($sp)
	addiu	$4,$2,-16
$L52:
#APP
	S32LDD xr1,$9,0
	S32LDD xr2,$9,4
	S32LDD xr3,$9,8
#NO_APP
	addu	$2,$11,$9
	sw	$2,0($sp)
#APP
	pref 0,0($2)
#NO_APP
	lw	$3,16($sp)
#APP
	S32ALN xr4,xr2,xr1,$3
	S32ALN xr5,xr3,xr2,$3
#NO_APP
	li	$2,3			# 0x3
#APP
	S32ALN xr6,xr5,xr4,$2
	D32SLR xr1,xr5,xr0,xr0,8
	S32SFL xr0,xr1,xr5,xr7,ptn3
	Q8MUL xr2,xr4,xr14,xr1
	Q8MUL xr8,xr7,xr15,xr7
	Q8MUL xr4,xr6,xr14,xr3
	Q16ADD xr2,xr2,xr2,xr0,AA,XW
	Q16ADD xr0,xr1,xr1,xr1,SS,XW
	Q16ADD xr0,xr7,xr7,xr7,SS,XW
	S32SFL xr1,xr0,xr1,xr0,ptn3
	Q16ACC xr0,xr2,xr7,xr1,AA
	Q16ADD xr4,xr4,xr4,xr0,AA,XW
	Q16ADD xr3,xr3,xr3,xr0,SS,XW
	Q16ADD xr0,xr8,xr8,xr8,SS,XW
	S32SFL xr3,xr0,xr3,xr0,ptn3
	Q16ACC xr0,xr4,xr8,xr3,AA
	S32SFL xr0,xr3,xr1,xr1,ptn3
	S32SDI xr1,$4,4
#NO_APP
	lw	$2,4($sp)
#APP
	S32LDD xr1,$2,0
	S32LDD xr2,$2,4
	S32LDD xr3,$2,8
#NO_APP
	addu	$2,$11,$2
	sw	$2,4($sp)
#APP
	pref 0,0($2)