summaryrefslogtreecommitdiff
path: root/reference/xon7.asm
diff options
context:
space:
mode:
Diffstat (limited to 'reference/xon7.asm')
-rw-r--r--reference/xon7.asm327
1 files changed, 131 insertions, 196 deletions
diff --git a/reference/xon7.asm b/reference/xon7.asm
index 3846190..faa1d33 100644
--- a/reference/xon7.asm
+++ b/reference/xon7.asm
@@ -6,215 +6,150 @@
@out(r1.y) out1
@out(r1.z) out2
@out(r1.w) out3
+@const(c10.x) 0x3f000000, 0xbf000000, 0x40000000, 0xbf800000
+@const(c11.x) 0x00000000, 0x3f800000, 0x3e800000, 0x00000000
+@const(c12.x) 0x00000000, 0x3f800000, 0x43000000, 0x00000000
(sy)(ss)bary.f r0.z, 2, r0.x
-bary.f r0.w, 0, r0.x
-bary.f r1.x, 4, r0.x
-bary.f r1.y, 3, r0.x
-mov.f32f32 r1.z, r0.z
-mov.f32f32 r1.w, r0.w
-mov.f32f32 r2.y, r0.w
-mul.f r2.z, r1.x, r1.x
-mov.f32f32 r2.w, r1.z
-mov.f32f32 r1.z, r1.y
-mov.f32f32 r3.y, r1.w
-bary.f r3.w, 1, r0.x
-mov.f32f32 r4.x, r2.y
-mov.f32f32 r3.x, r1.z
-bary.f r1.z, 5, r0.x
-mov.f32f32 r1.w, r3.w
-mov.f32f32 r2.y, r3.w
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r0.z, r0.z
-mad.f32 r2.z, r1.z, r1.z, r2.z
-sam (f32)(xyz)r4.z, r2.w, s#4, t#4
-(sy)(ss)mad.f32 r2.w, c10.z, r4.z, c10.w
-mov.f32f32 r3.z, r1.w
-mov.f32f32 r4.y, r2.y
-mov.f32f32 r1.w, r2.z
-mov.f32f32 r2.y, r2.w
-bary.f r2.z, 12, r0.x
-bary.f r2.w, 16, r0.x
-bary.f r3.x, 8, r0.x
-sam (f32)(xyz)r5.y, r3.y, s#0, t#0
-(sy)(ss)add.f r3.y, r5.y, c10.y
-mul.f r2.z, r2.y, r2.z
-mad.f32 r3.z, c10.z, r4.w, c10.w
-mul.f r2.w, r2.y, r2.w
-mov.f32f32 r3.y, r3.y
-mul.f r2.y, r2.y, r3.x
-mov.f32f32 r3.x, r3.z
-bary.f r3.z, 13, r0.x
-mul.f r4.z, r3.y, r3.y
-add.f r4.w, r5.z, c10.y
-bary.f r5.y, 17, r0.x
-mad.f32 r2.z, r3.x, r3.z, r2.z
-bary.f r3.z, 9, r0.x
-mov.f32f32 r4.w, r4.w
-mad.f32 r2.w, r3.x, r5.y, r2.w
-mov.f32f32 r2.z, r2.z
-mad.f32 r5.x, c10.z, r5.x, c10.w
-mad.f32 r4.z, r4.w, r4.w, r4.z
-mov.f32f32 r2.w, r2.w
-mad.f32 r2.y, r3.x, r3.z, r2.y
-mov.f32f32 r3.x, r5.x
-bary.f r3.z, 14, r0.x
-mov.f32f32 r4.z, r4.z
-add.f r5.x, r5.w, c10.y
-bary.f r5.y, 18, r0.x
-mad.f32 r2.z, r3.x, r3.z, r2.z
-mov.f32f32 r2.y, r2.y
-mov.f32f32 r3.z, r5.x
-bary.f r5.x, 10, r0.x
-mov.f32f32 r2.z, r2.z
-mad.f32 r2.w, r3.x, r5.y, r2.w
-mad.f32 r4.z, r3.z, r3.z, r4.z
-mad.f32 r2.y, r3.x, r5.x, r2.y
+bary.f r0.w, 3, r0.x
+bary.f r1.x, 0, r0.x
+bary.f r1.z, 4, r0.x
+mov.f32f32 r2.y, r0.z
+mov.f32f32 r2.z, r0.w
+mov.f32f32 r2.w, r1.x
+bary.f r1.y, 1, r0.x
+sam (f32)(xyz)r3.y, r0.z, s#4, t#4
+(sy)(ss)mad.f32 r0.z, c10.z, r3.y, c10.w
+mov.f32f32 r4.x, r1.x
+bary.f r0.w, 16, r0.x
+mul.f r1.w, r1.z, r1.z
+mov.f32f32 r3.y, r0.z
+bary.f r4.z, 12, r0.x
+mov.f32f32 r4.y, r1.y
+mov.f32f32 r3.x, r1.y
+bary.f r4.w, 8, r0.x
+mul.f r4.z, r3.y, r4.z
+mad.f32 r3.z, c10.z, r3.z, c10.w
+mul.f r0.z, r0.z, r0.w
+mul.f r0.w, r3.y, r4.w
+bary.f r3.y, 17, r0.x
+mov.f32f32 r4.w, r3.z
+bary.f r5.x, 13, r0.x
+sam (f32)(xyz)r5.y, r4.x, s#0, t#0
+(sy)(ss)add.f r4.x, r5.y, c10.y
+sam (f32)(xyzw)r6.x, r2.w, s#1, t#1
+(ss)bary.f r2.w, 9, r0.x
+(sy)cmps.f.lt r3.x, r6.w, c10.x
+mad.f32 r4.y, r4.w, r5.x, r4.z
+mad.f32 r3.w, c10.z, r3.w, c10.w
+mov.f32f32 r4.z, r4.x
+cov.u32f32 r3.x, r3.x
+mad.f32 r0.w, r4.w, r2.w, r0.w
+mov.f32f32 r2.w, r3.w
+bary.f r4.w, 14, r0.x
+mul.f r4.x, r4.x, r4.z
+add.f r5.x, r5.z, c10.y
+mov.f32f32 r5.y, (0.000000)
+mad.f32 r4.y, r2.w, r4.w, r4.y
+bary.f r4.w, 10, r0.x
+mad.f32 r0.z, r3.z, r3.y, r0.z
+bary.f r3.y, 5, r0.x
+mov.f32f32 r3.z, r4.y
+mad.f32 r0.w, r2.w, r4.w, r0.w
+mov.f32f32 r2.w, r5.x
+cmps.f.ne p0.x, r3.x, r5.y
+bary.f r3.x, 18, r0.x
+mul.f r4.w, r0.w, r0.w
+mad.f32 r4.x, r5.x, r2.w, r4.x
+mad.f32 r4.y, r4.y, r3.z, r4.w
+mad.f32 r0.z, r3.w, r3.x, r0.z
+add.f r3.x, r5.w, c10.y
bary.f (ei)r0.x, 6, r0.x
-sam (f32)(xyzw)r5.x, r4.x, s#1, t#1
-mov.f32f32 r0.y, r2.w
-(sy)cmps.f.lt r2.w, r5.w, c10.x
-mul.f r3.x, r2.y, r2.y
-mad.f32 r1.w, r0.x, r0.x, r1.w
-mad.f32 r3.x, r2.z, r2.z, r3.x
-(ss)rsq r4.x, r4.z
-(ss)mov.f32f32 r4.x, r4.x
+kill p0.x
+mov.f32f32 r0.y, r0.z
+mov.f32f32 r3.w, r3.x
+mad.f32 r4.w, r3.y, r3.y, r1.w
+mul.f r1.w, r6.w, c9.x
mov.f32f32 r0.y, r0.y
-cov.u32f32 r2.w, r2.w
-mov.f32f32 r3.x, r3.x
-mul.f r3.y, r3.y, r4.x
-mad.f32 r3.x, r0.y, r0.y, r3.x
-mul.f r4.y, r4.w, r4.x
-mul.f r3.z, r3.z, r4.x
-(rpt3)nop
-rsq r3.x, r3.x
-(ss)mov.f32f32 r3.x, r3.x
-mov.f32f32 r3.y, r3.y
-mov.f32f32 r4.x, r4.y
+mad.f32 r3.w, r3.w, r3.w, r4.x
+mad.f32 r4.x, r0.x, r0.x, r4.w
+mul.f r4.w, r6.z, c5.z
+mad.f32 r0.y, r0.y, r0.y, r4.y
+mul.f r4.y, r6.z, c6.z
+mul.f r5.x, r6.y, c6.y
+mul.f r5.y, r6.x, c6.x
+rsq r3.w, r3.w
+(ss)mov.f32f32 r5.z, r3.w
+mul.f r3.x, r3.x, r3.w
+(ss)rsq r3.w, r4.x
+(ss)mov.f32f32 r4.x, r3.w
+rsq r0.y, r0.y
+(ss)mov.f32f32 r5.w, r0.y
+(ss)mul.f r0.y, r0.z, r0.y
+mul.f r0.z, r4.z, r5.z
+mul.f r2.w, r2.w, r5.z
+mul.f r0.w, r0.w, r5.w
+mov.f32f32 r4.z, r0.y
+mul.f r3.z, r3.z, r5.w
+mov.f32f32 r5.z, r0.z
+mul.f r0.z, r0.z, r0.w
+mov.f32f32 r0.w, r0.w
+mad.f32 r0.z, r2.w, r3.z, r0.z
+max.f r5.w, c11.z, r4.z
+mad.f32 r0.z, r3.x, r4.z, r0.z
+mul.f r6.z, r5.z, r0.w
mov.f32f32 r3.z, r3.z
-mul.f r2.y, r2.y, r3.x
-mul.f r0.y, r0.y, r3.x
-mul.f r2.z, r2.z, r3.x
-mov.f32f32 r3.x, (0.000000)
-mov.f32f32 r2.y, r2.y
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r2.z, r2.z
-cmps.f.ne p0.x, r2.w, r3.x
-mul.f r2.w, r3.y, r2.y
-mul.f r3.x, r3.y, r2.y
-mad.f32 r2.w, r4.x, r2.z, r2.w
-mad.f32 r3.x, r4.x, r2.z, r3.x
-max.f r4.y, c11.z, r0.y
-rsq r1.w, r1.w
-(ss)mov.f32f32 r1.w, r1.w
mov.f32f32 r2.w, r2.w
+mul.f r5.z, r0.z, r5.z
mov.f32f32 r3.x, r3.x
-mad.f32 r2.w, r3.z, r0.y, r2.w
-mad.f32 r3.x, r3.z, r0.y, r3.x
-mov.f32f32 r4.y, r4.y
-kill p0.x
-mul.f r3.y, r2.w, r3.y
-mul.f r4.x, r2.w, r4.x
-mul.f r2.w, r2.w, r3.z
-mov.f32f32 r3.x, r3.x
-mov.f32f32 r3.y, r3.y
-mov.f32f32 r3.z, r4.x
-mov.f32f32 r2.w, r2.w
-max.f r3.x, r3.x, c11.x
-mul.f r3.y, c10.z, r3.y
-mul.f r3.z, c10.z, r3.z
+rcp r5.w, r5.w
+(ss)mov.f32f32 r6.w, r5.w
+mad.f32 r6.z, r2.w, r3.z, r6.z
+mul.f r5.z, c10.z, r5.z
+mad.f32 r4.z, r3.x, r4.z, r6.z
+mul.f r2.w, r0.z, r2.w
+mul.f r0.z, r0.z, r3.x
+add.f r0.w, r0.w, (neg)r5.z
+mul.f r1.z, r1.z, r4.x
+max.f r3.x, r4.z, c11.x
mul.f r2.w, c10.z, r2.w
-mov.f32f32 r3.x, r3.x
-mov.f32f32 r3.y, r3.y
-mov.f32f32 r3.z, r3.z
-mov.f32f32 r2.w, r2.w
-mul.f r4.x, r5.z, c6.z
-add.f r2.y, r2.y, (neg)r3.y
-add.f r2.z, r2.z, (neg)r3.z
-add.f r0.y, r0.y, (neg)r2.w
-mul.f r2.w, r4.x, r3.x
-mov.f32f32 r2.y, r2.y
-mul.f r1.x, r1.x, r1.w
-mov.f32f32 r2.z, r2.z
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r2.w, r2.w
-mov.f32f32 r1.x, r1.x
-mul.f r3.y, r5.y, c6.y
-mul.f r3.z, r5.x, c6.x
-rcp r4.x, r4.y
-(ss)mov.f32f32 r4.x, r4.x
-mul.f r1.x, r2.y, r1.x
-mul.f r1.z, r1.z, r1.w
-mul.f r2.y, r3.y, r3.x
-mul.f r3.x, r3.z, r3.x
-nop
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r2.y, r2.y
-mov.f32f32 r3.x, r3.x
-mov.f32f32 r3.y, r0.z
-mad.f32 r0.z, r2.z, r1.z, r1.x
-mov.f32f32 r1.x, r1.y
-mul.f r0.x, r0.x, r1.w
-nop
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r3.z, r1.x
-mov.f32f32 r0.x, r0.x
-mul.f r1.x, r5.w, c9.x
-mul.f r1.y, r5.z, c5.z
-mul.f r1.z, r5.y, c5.y
-mad.f32 r0.x, r0.y, r0.x, r0.z
-mov.f32f32 r0.y, r1.x
-nop
-sam (f32)(xyz)r5.y, r3.y, s#3, t#3
-(sy)mul.f r0.z, r5.w, r4.x
-mov.f32f32 r0.x, r0.x
-mul.f r1.x, r5.z, r4.x
-mul.f r1.w, r5.y, r4.x
-nop
+mul.f r0.z, c10.z, r0.z
+mul.f r0.w, r0.w, r1.z
+mov.f32f32 r1.z, r3.x
+add.f r2.w, r3.z, (neg)r2.w
+mul.f r3.y, r3.y, r4.x
+mul.f r3.x, r5.y, r3.x
+add.f r0.y, r0.y, (neg)r0.z
+sam (f32)(xyz)r7.x, r2.y, s#3, t#3
+mul.f r0.z, r4.y, r1.z
+mad.f32 r0.w, r2.w, r3.y, r0.w
+mul.f r0.x, r0.x, r3.w
+mul.f r1.z, r5.x, r1.z
+(sy)(ss)mul.f r2.y, r7.z, r6.w
+mul.f r2.z, r7.y, r6.w
+mad.f32 r0.x, r0.y, r0.x, r0.w
+mul.f r0.y, r6.y, c5.y
+mul.f r0.w, r6.x, c5.x
+mul.f r2.w, r7.x, r5.w
max.f r0.x, (neg)r0.x, c11.x
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r1.x, r1.x
-mov.f32f32 r2.z, r1.w
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r1.w, r0.y
-mov.f32f32 r0.y, r1.y
-mov.f32f32 r1.y, r1.z
-mul.f r1.z, r5.x, c5.x
-(ss)mov.f32f32 r3.y, r0.w
-mov.f32f32 r0.w, r3.w
-log2 r0.x, r0.x
-(rpt2)nop
-mov.f32f32 r3.z, r0.w
-mov.f32f32 r0.w, r1.z
-(rpt4)nop
-sam (f32)(xyzw)r3.y, r3.y, s#2, t#2
-(sy)mad.f32 r1.z, c8.x, r4.x, c11.y
-mul.f r3.w, r3.w, c7.z
-(ss)mul.f r3.z, r3.z, c7.y
+sam (f32)(xyzw)r3.y, r1.x, s#2, t#2
+(sy)(ss)mul.f r1.x, r3.w, c7.z
+mul.f r1.y, r3.z, c7.y
+mad.f32 r3.z, c8.x, r4.x, c11.y
mul.f r3.y, r3.y, c7.x
-mov.f32f32 r1.z, r1.z
-(rpt2)nop
-(ss)mul.f r0.x, r1.z, r0.x
-(rpt2)nop
-mov.f32f32 r0.x, r0.x
+(rpt1)nop
+log2 r0.x, r0.x
+(ss)mul.f r0.x, r3.z, r0.x
(rpt5)nop
exp2 r0.x, r0.x
-(ss)mad.f32 r1.z, r3.w, r0.x, r2.w
-mad.f32 r2.y, r3.z, r0.x, r2.y
+(ss)mad.f32 r0.z, r1.x, r0.x, r0.z
+mad.f32 r1.x, r1.y, r0.x, r1.z
(ss)mad.f32 r0.x, r3.y, r0.x, r3.x
nop
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r2.y, r2.y
-mov.f32f32 r0.x, r0.x
-nop
-mad.f32 r0.y, r1.z, r0.z, r0.y
-mad.f32 r0.z, r2.y, r1.x, r1.y
-mad.f32 r0.x, r0.x, r2.z, r0.w
-nop
-mov.f32f32 r1.z, r0.y
-mov.f32f32 r1.y, r0.z
-mov.f32f32 r1.x, r0.x
+mad.f32 r1.z, r0.z, r2.y, r4.w
+mad.f32 r1.y, r1.x, r2.z, r0.y
+mad.f32 r1.x, r0.x, r2.w, r0.w
end
; FRAG: outputs: r1.x (1:0)
-; FRAG: inputs: r0.z (5:20,cm=f,il=8,b=1) r1.x (5:21,cm=f,il=12,b=1) r2.x (5:22,cm=f,il=16,b=1) r3.x (5:23,cm=f,il=20,b=1) r4.x (5:24,cm=f,il=24,b=1)
-; FRAG: 226 instructions, 0 half, 6 full
+; FRAG: inputs: r0.x (5:20,cm=f,il=8,b=1) r1.x (5:21,cm=f,il=12,b=1) r2.x (5:22,cm=f,il=16,b=1) r3.x (5:23,cm=f,il=20,b=1) r4.x (5:24,cm=f,il=24,b=1)
+; FRAG: 146 instructions, 0 half, 8 full