summaryrefslogtreecommitdiff
path: root/reference/flow.asm
diff options
context:
space:
mode:
Diffstat (limited to 'reference/flow.asm')
-rw-r--r--reference/flow.asm315
1 files changed, 121 insertions, 194 deletions
diff --git a/reference/flow.asm b/reference/flow.asm
index e9b7b37..c61e896 100644
--- a/reference/flow.asm
+++ b/reference/flow.asm
@@ -6,227 +6,154 @@
@out(r1.y) out1
@out(r1.z) out2
@out(r1.w) out3
+@const(c5.x) 0x40000000, 0xbf800000, 0x40800000, 0x3f000000
+@const(c6.x) 0x00000000, 0x3f333333, 0x3e99999a, 0x3d4ccccd
+@const(c7.x) 0x00000000, 0x3f800000, 0x43000000, 0x00000000
(sy)(ss)bary.f r0.z, 0, r0.x
mov.f32f32 r0.w, c3.x
bary.f r1.x, 1, r0.x
mov.f32f32 r1.y, c6.x
-add.f r1.z, r0.z, c3.x
-add.f r1.w, r0.z, c4.x
+add.f r1.z, r0.z, c4.x
+add.f r2.x, r0.z, c3.x
+add.f r1.w, r1.x, c4.y
+add.f r2.y, r1.x, c3.y
add.f r0.z, r0.z, r0.w
add.f r0.w, r1.x, r1.y
-mov.f32f32 r1.y, r1.z
-mov.f32f32 r1.z, r1.w
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r1.w, r1.y
-add.f r1.y, r1.x, c3.y
-mov.f32f32 r2.y, r1.z
-add.f r1.x, r1.x, c4.y
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r1.x, r1.x
-bary.f r1.z, 7, r0.x
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r2.w, r0.z
-mov.f32f32 r3.x, r0.w
-mov.f32f32 r0.z, r1.x
-mov.f32f32 r2.x, r1.y
-mov.f32f32 r0.w, r1.z
-bary.f r1.x, 8, r0.x
-bary.f r1.y, 9, r0.x
-mov.f32f32 r2.z, r0.z
-sam (f32)(xyzw)r2.w, r2.w, s#2, t#2
-mov.f32f32 r0.z, c6.x
-mov.f32f32 r1.z, c6.x
-sam (f32)(xyz)r3.w, r1.w, s#0, t#0
-(sy)(ss)mad.f32 r1.w, c5.x, r4.x, c5.y
-mad.f32 r2.x, c5.x, r3.w, c5.y
-mad.f32 r3.w, c5.x, r4.y, c5.y
-mov.f32f32 r1.x, r1.x
-mov.f32f32 r1.w, r1.w
-sam (f32)(xyz)r4.x, r2.y, s#1, t#1
-(sy)(ss)mad.f32 r2.y, c5.x, r4.y, c5.y
-mov.f32f32 r2.x, r2.x
-mad.f32 r2.z, c5.x, r4.x, c5.y
+mov.f32f32 r1.x, c6.x
+mov.f32f32 r1.y, c6.x
+bary.f r2.z, 7, r0.x
+sam (f32)(xyz)r2.w, r1.z, s#1, t#1
+(sy)(ss)mad.f32 r1.z, c5.x, r3.x, c5.y
+sam (f32)(xyz)r3.z, r2.x, s#0, t#0
+(sy)mad.f32 r1.w, c5.x, r3.w, c5.y
+(ss)mad.f32 r2.x, c5.x, r2.w, c5.y
+mad.f32 r2.y, c5.x, r3.z, c5.y
+mul.f r1.z, c5.z, r1.z
mul.f r1.w, c5.z, r1.w
-mov.f32f32 r2.y, r2.y
-mul.f r4.x, r2.x, r2.x
-mov.f32f32 r2.z, r2.z
-mov.f32f32 r1.w, r1.w
-mul.f r2.y, c5.z, r2.y
-mov.f32f32 r3.w, r3.w
-mov.f32f32 r2.z, r2.z
-mad.f32 r4.x, r1.w, r1.w, r4.x
-mov.f32f32 r2.y, r2.y
-mad.f32 r4.y, c5.x, r4.z, c5.y
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r4.x, r4.x
-mov.f32f32 r2.y, r2.y
-mad.f32 r4.x, r3.w, r3.w, r4.x
-mul.f r4.z, r2.z, r2.z
-mov.f32f32 r4.y, r4.y
-(rpt3)nop
-rsq r4.x, r4.x
-(ss)mov.f32f32 r4.x, r4.x
-mad.f32 r4.z, r2.y, r2.y, r4.z
-mov.f32f32 r4.y, r4.y
-bary.f r4.w, 4, r0.x
-mul.f r2.x, r2.x, r4.x
-mul.f r1.w, r1.w, r4.x
-mul.f r3.w, r3.w, r4.x
-mov.f32f32 r4.x, r4.z
-mov.f32f32 r2.x, r2.x
-mad.f32 r4.x, r4.y, r4.y, r4.x
-mov.f32f32 r1.w, r1.w
-mov.f32f32 r3.w, r3.w
-bary.f r4.z, 5, r0.x
+mov.f32f32 r2.w, r2.x
+mov.f32f32 r3.x, r2.y
+mov.f32f32 r3.z, r1.z
+mov.f32f32 r3.w, r1.w
+mov.f32f32 r2.w, r2.w
+mul.f r2.y, r2.y, r3.x
+mov.f32f32 r3.z, r3.z
+mad.f32 r1.w, r1.w, r3.w, r2.y
+mul.f r2.x, r2.x, r2.w
+mad.f32 r2.y, c5.x, r4.x, c5.y
+mad.f32 r1.z, r1.z, r3.z, r2.x
+mad.f32 r2.x, c5.x, r3.y, c5.y
+sam (f32)(xyzw)r4.x, r0.z, s#2, t#2
+(ss)bary.f r0.z, 8, r0.x
+bary.f r0.w, 9, r0.x
+bary.f r3.y, 4, r0.x
+mov.f32f32 r5.x, r2.x
+mov.f32f32 r5.y, r2.y
+bary.f r5.z, 5, r0.x
bary.f (ei)r0.x, 6, r0.x
+mov.f32f32 r0.y, r5.x
+mad.f32 r1.w, r5.y, r5.y, r1.w
(rpt1)nop
-rsq r0.y, r4.x
-(ss)mov.f32f32 r0.y, r0.y
-(rpt2)nop
-mad.f32 r2.x, r2.z, r0.y, r2.x
-mad.f32 r1.w, r2.y, r0.y, r1.w
-mad.f32 r0.y, r4.y, r0.y, r3.w
-nop
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r1.w, r1.w
-mov.f32f32 r0.y, r0.y
+mad.f32 r0.y, r0.y, r0.y, r1.z
+(rpt5)nop
+rsq r0.y, r0.y
+(ss)mov.f32f32 r1.z, r0.y
+rsq r1.w, r1.w
+(ss)mov.f32f32 r5.x, r1.w
+(ss)mul.f r1.w, r2.y, r1.w
+(rpt1)nop
+mul.f r2.y, r3.x, r5.x
+mul.f r3.x, r3.w, r5.x
+mad.f32 r2.y, r2.w, r1.z, r2.y
+mad.f32 r1.z, r3.z, r1.z, r3.x
+mad.f32 r0.y, r2.x, r0.y, r1.w
nop
-mul.f r2.x, r2.x, c5.w
-mul.f r1.w, r1.w, c5.w
+mul.f r1.w, r2.y, c5.w
+mul.f r1.z, r1.z, c5.w
mul.f r0.y, r0.y, c5.w
nop
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r1.w, r1.w
-mov.f32f32 r0.y, r0.y
-nop
-mul.f r2.y, r2.x, r4.w
-mul.f r2.z, r4.w, r2.x
-mad.f32 r2.y, r1.w, r4.z, r2.y
-mad.f32 r2.z, r4.z, r1.w, r2.z
-(rpt1)nop
-mov.f32f32 r2.y, r2.y
-mov.f32f32 r2.z, r2.z
-mad.f32 r2.y, r0.y, r0.x, r2.y
-mad.f32 r2.z, r0.x, r0.y, r2.z
-(rpt1)nop
-mul.f r2.x, r2.y, r2.x
-max.f r2.z, r2.z, c6.x
-mul.f r1.w, r2.y, r1.w
-mul.f r0.y, r2.y, r0.y
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r2.y, r2.z
-mov.f32f32 r1.w, r1.w
-mov.f32f32 r0.y, r0.y
-mul.f r2.x, c5.x, r2.x
-mad.f32 r2.y, c6.y, r2.y, c6.z
-mul.f r1.w, c5.x, r1.w
-mul.f r0.y, c5.x, r0.y
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r2.y, r2.y
-mov.f32f32 r1.w, r1.w
-mov.f32f32 r0.y, r0.y
-add.f r2.x, r4.w, (neg)r2.x
-mul.f r2.z, r3.z, r2.y
-mul.f r3.y, r3.y, r2.y
-mul.f r3.x, r3.x, r2.y
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r3.z, r2.z
-mov.f32f32 r3.w, r3.y
-(ss)mov.f32f32 r4.x, r3.x
-mul.f r4.y, r2.x, r2.x
-add.f r1.w, r4.z, (neg)r1.w
-add.f r1.z, r3.z, r1.z
-add.f r0.x, r0.x, (neg)r0.y
-mul.f r0.y, r2.w, r2.y
-mov.f32f32 r1.w, r1.w
-add.f r0.z, r1.z, r0.z
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r1.z, r0.y
-mad.f32 r2.y, r1.w, r1.w, r4.y
+mov.f32f32 r2.x, r1.w
+mul.f r1.w, r3.y, r1.w
+mov.f32f32 r2.y, r1.z
+mov.f32f32 r2.w, r0.y
+mul.f r3.x, r2.x, r3.y
+mad.f32 r1.z, r5.z, r1.z, r1.w
+mad.f32 r1.w, r2.y, r5.z, r3.x
+mad.f32 r0.y, r0.x, r0.y, r1.z
+mad.f32 r1.z, r2.w, r0.x, r1.w
(rpt2)nop
-mov.f32f32 r2.y, r2.y
-nop
-mad.f32 r2.y, r0.x, r0.x, r2.y
-(rpt5)nop
-rsq r2.y, r2.y
-(ss)mov.f32f32 r2.y, r2.y
+mul.f r1.w, r1.z, r2.x
+max.f r0.y, r0.y, c6.x
+mul.f r2.x, r1.z, r2.y
+mul.f r1.z, r1.z, r2.w
+mul.f r1.w, c5.x, r1.w
+mad.f32 r0.y, c6.y, r0.y, c6.z
+mul.f r2.x, c5.x, r2.x
+mul.f r1.z, c5.x, r1.z
+add.f r1.w, r3.y, (neg)r1.w
+mov.f32f32 r2.y, r0.y
+add.f r2.x, r5.z, (neg)r2.x
+add.f r0.x, r0.x, (neg)r1.z
+mov.f32f32 r1.z, r1.w
+(sy)mul.f r2.w, r4.w, r2.y
+mov.f32f32 r3.x, r2.x
+mov.f32f32 r3.y, r0.x
+mul.f r1.w, r1.w, r1.z
+add.f r1.y, r2.w, r1.y
+mad.f32 r1.w, r2.x, r3.x, r1.w
+mul.f r2.x, r4.z, r2.y
+mad.f32 r1.w, r3.y, r3.y, r1.w
+add.f r1.x, r1.y, r1.x
+mul.f r1.y, r4.y, r2.y
+mul.f r0.y, r4.x, r0.y
(rpt2)nop
-mul.f r2.x, r2.x, r2.y
-mul.f r1.w, r1.w, r2.y
-mul.f r0.x, r0.x, r2.y
-nop
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r1.w, r1.w
-mov.f32f32 r0.x, r0.x
+rsq r1.w, r1.w
+(ss)mov.f32f32 r2.y, r1.w
+mul.f r0.x, r0.x, r1.w
+(rpt1)nop
+mul.f r1.z, r1.z, r2.y
+(ss)mul.f r1.w, r3.x, r2.y
+(rpt1)nop
+mul.f r1.z, r1.z, r2.z
nop
-mul.f r0.w, r2.x, r0.w
+mad.f32 r0.z, r1.w, r0.z, r1.z
nop
-mad.f32 r0.w, r1.w, r1.x, r0.w
-(rpt2)nop
-mov.f32f32 r0.w, r0.w
-nop
-mad.f32 r0.x, r0.x, r1.y, r0.w
-(rpt2)nop
-mov.f32f32 r0.x, r0.x
+mad.f32 r0.x, r0.x, r0.w, r0.z
(rpt2)nop
max.f r0.x, r0.x, c6.x
(rpt2)nop
-mov.f32f32 r0.x, r0.x
+mov.f32f32 r0.z, r0.x
(rpt2)nop
-mov.f32f32 r0.w, r0.x
-mov.f32f32 r1.x, r0.x
-cmps.f.lt r0.x, c6.x, r0.x
+mul.f r0.x, r0.x, r0.z
+cmps.f.lt r0.z, c6.x, r0.z
(rpt1)nop
-mul.f r0.w, r0.w, r1.x
-cov.u32f32 r0.x, r0.x
+mov.f32f32 r0.w, r0.x
+cov.u32f32 r0.z, r0.z
(rpt1)nop
-mov.f32f32 r0.w, r0.w
-cmps.f.ne r0.x, r0.x, c6.x
+mul.f r0.x, r0.x, r0.w
+cmps.f.ne r0.z, r0.z, c6.x
(rpt1)nop
+mov.f32f32 r0.w, r0.x
+mul.f r0.x, r0.x, c6.w
+sel.b32 r1.w, r1.x, r0.z, r2.w
+nop
mul.f r0.w, r0.w, r0.w
-sel.b32 r0.z, r0.z, r0.x, r2.z
-(rpt1)nop
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r1.w, r0.z
+add.f r1.x, r2.x, r0.x
+add.f r1.z, r1.y, r0.x
+add.f r0.x, r0.y, r0.x
+mul.f r0.w, r0.w, c6.z
+(rpt2)nop
+mov.f32f32 r2.y, r0.w
+add.f r0.x, r0.x, r0.w
(rpt1)nop
-mul.f r0.z, r0.w, r0.w
-mul.f r0.w, r0.w, c6.w
+add.f r0.w, r1.x, r2.y
+add.f r1.x, r1.z, r2.y
(rpt1)nop
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r1.x, r0.w
-mov.f32f32 r1.y, r0.w
-mov.f32f32 r0.w, r0.w
-mul.f r0.z, r0.z, c6.z
-add.f r1.x, r3.w, r1.x
-add.f r1.y, r4.x, r1.y
-add.f r0.w, r1.z, r0.w
-mov.f32f32 r0.z, r0.z
-(rpt2)nop
-mov.f32f32 r1.z, r0.z
-mov.f32f32 r2.x, r0.z
-mov.f32f32 r0.z, r0.z
-nop
-add.f r1.x, r1.x, r1.z
-add.f r1.y, r1.y, r2.x
-add.f r0.z, r0.w, r0.z
-nop
-sel.b32 r0.w, r1.x, r0.x, r3.y
-sel.b32 r1.x, r1.y, r0.x, r3.x
-sel.b32 r0.x, r0.z, r0.x, r0.y
-nop
-mov.f32f32 r1.z, r0.w
-mov.f32f32 r1.y, r1.x
-mov.f32f32 r1.x, r0.x
+sel.b32 r1.z, r0.w, r0.z, r2.x
+sel.b32 r1.y, r1.x, r0.z, r1.y
+sel.b32 r1.x, r0.x, r0.z, r0.y
end
nop
; FRAG: outputs: r1.x (1:0)
-; FRAG: inputs: r0.x (5:0,cm=f,il=8,b=1) r0.w (5:19,cm=f,il=12,b=1) r2.x (5:20,cm=f,il=16,b=1)
-; FRAG: 254 instructions, 0 half, 5 full
+; FRAG: inputs: r0.x (5:0,cm=f,il=8,b=1) r1.w (5:19,cm=f,il=12,b=1) r2.x (5:20,cm=f,il=16,b=1)
+; FRAG: 169 instructions, 0 half, 6 full