summaryrefslogtreecommitdiff
path: root/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-32.asm
diff options
context:
space:
mode:
Diffstat (limited to 'reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-32.asm')
-rw-r--r--reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-32.asm838
1 files changed, 330 insertions, 508 deletions
diff --git a/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-32.asm b/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-32.asm
index a1623b7..dc38031 100644
--- a/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-32.asm
+++ b/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-32.asm
@@ -4,546 +4,368 @@
@in(r0.w) in3
@in(r1.x) in4
@in(r1.y) in5
-@out(r4.w) out0
-@out(r5.x) out1
-@out(r5.y) out2
-@out(r5.z) out3
+@out(r2.x) out0
+@out(r2.y) out1
+@out(r2.z) out2
+@out(r2.w) out3
+@const(c13.x) 0x3f000000, 0x00000000, 0x3f800000, 0xbeef9097
+@const(c14.x) 0x3cff9724, 0x40000000, 0xbf800000, 0xba03126f
+@const(c15.x) 0xbf000000, 0x3f800000, 0x3fb8aa65, 0x3de38866
+@const(c16.x) 0x3cf5c28f, 0x00000000, 0x00000000, 0x00000000
+@const(c17.x) 0x00000000, 0x3f800000, 0x43000000, 0x00000000
(sy)(ss)bary.f r0.x, 19, r1.x
bary.f r0.y, 0, r1.x
add.f r0.w, r0.w, c13.y
bary.f r1.z, 1, r1.x
-mov.f32f32 r0.x, r0.x
-add.f r1.w, r0.y, c15.x
-bary.f r2.x, 23, r1.x
-bary.f r2.y, 24, r1.x
-mul.f r2.z, r0.x, r0.x
+mov.f32f32 r1.w, r0.x
+add.f r2.x, r0.y, c15.x
+bary.f r2.y, 8, r1.x
+bary.f r2.z, 23, r1.x
+mul.f r0.x, r0.x, r1.w
bary.f r2.w, 20, r1.x
-floor.f r3.x, r1.w
+floor.f r3.x, r2.x
rcp r0.w, r0.w
add.f r0.z, r0.z, c13.y
add.f r3.y, r1.z, c15.x
-mov.f32f32 r2.w, r2.w
-add.f r1.w, r1.w, (neg)r3.x
+mov.f32f32 r3.z, r2.w
+add.f r2.x, r2.x, (neg)r3.x
(ss)mul.f r0.z, r0.z, r0.w
-(ss)floor.f r0.w, r3.y
-mad.f32 r2.z, r2.w, r2.w, r2.z
-mov.f32f32 r1.w, r1.w
+(ss)absneg.f r0.w, (neg)c10.x
+mad.f32 r0.x, r2.w, r3.z, r0.x
+bary.f r2.w, 21, r1.x
+mov.f32f32 r3.x, r2.x
+mul.f r0.w, r0.w, c10.x
+floor.f r3.w, r3.y
+mov.f32f32 r4.x, r2.w
+mul.f r4.y, c13.x, r3.x
+mul.f r0.w, r0.w, r0.z
mov.f32f32 r0.z, r0.z
-absneg.f r3.x, (neg)c10.x
-mov.f32f32 r2.z, r2.z
-bary.f r3.z, 21, r1.x
-mul.f r3.w, c13.x, r1.w
-mul.f r3.x, r3.x, c10.x
-add.f r0.w, r3.y, (neg)r0.w
-mov.f32f32 r3.y, r3.z
-mov.f32f32 r3.z, r3.w
-mul.f r3.x, r3.x, r0.z
-mov.f32f32 r0.w, r0.w
-mad.f32 r2.z, r3.y, r3.y, r2.z
-add.f r0.y, r0.y, (neg)r3.z
-mov.f32f32 r3.x, r3.x
-mul.f r3.z, c13.x, r0.w
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r2.y, r2.y
-mov.f32f32 r0.y, r0.y
-rsq r3.w, r2.z
-(ss)mul.f r4.x, r0.x, r3.w
-rsq r4.y, r2.z
-(ss)mov.f32f32 r4.y, r4.y
-(ss)rsq r2.z, r2.z
-(ss)mul.f r4.z, r0.x, r2.z
-add.f r4.w, c15.x, r0.y
-mov.f32f32 r4.x, r4.x
-mul.f r5.x, r0.x, r4.y
-mov.f32f32 r4.z, r4.z
-mov.f32f32 r4.w, r4.w
-absneg.f r4.x, (abs)r4.x
-mov.f32f32 r5.x, r5.x
+mad.f32 r0.x, r4.x, r4.x, r0.x
+add.f r0.y, r0.y, (neg)r4.y
+add.f r3.y, r3.y, (neg)r3.w
+mul.f r0.z, r0.w, r0.z
+add.f r0.w, c14.y, (neg)r3.x
+mov.f32f32 r3.x, r0.y
+mov.f32f32 r3.w, r3.y
+rsq r4.y, r0.x
+(ss)mul.f r4.z, r1.w, r4.y
+rsq r4.w, r0.x
+(ss)mov.f32f32 r5.x, r4.w
+(ss)rsq r0.x, r0.x
+(ss)mul.f r5.y, r1.w, r0.x
+add.f r3.x, c15.x, r3.x
absneg.f r4.z, (abs)r4.z
-mul.f r4.w, r4.w, c5.z
-mov.f32f32 r4.x, r4.x
-absneg.f r5.x, (abs)r5.x
-mov.f32f32 r4.z, r4.z
-mov.f32f32 r5.y, r4.w
-add.f r4.x, r4.x, c13.w
-mov.f32f32 r5.x, r5.x
+mul.f r5.z, r1.w, r5.x
+absneg.f r5.y, (abs)r5.y
+mul.f r5.w, r3.x, c5.z
+add.f r3.x, r4.z, c13.w
+absneg.f r4.z, (abs)r5.z
+add.f r5.y, r5.y, c13.w
+mov.f32f32 r6.x, r5.w
+max.f r3.x, r3.x, c13.y
+mul.f r5.z, r3.z, r4.y
add.f r4.z, r4.z, c13.w
-mov.f32f32 r5.y, r5.y
-mov.f32f32 r4.x, r4.x
-add.f r5.x, r5.x, c13.w
-mov.f32f32 r4.z, r4.z
-mov.f32f32 r3.z, r3.z
-max.f r4.x, r4.x, c13.y
-mov.f32f32 r5.x, r5.x
-max.f r4.z, r4.z, c13.y
-add.f r1.z, r1.z, (neg)r3.z
-mov.f32f32 r3.z, r4.x
-mul.f r4.x, r2.w, r3.w
-max.f r5.x, r5.x, c13.y
-mov.f32f32 r4.z, r4.z
-mul.f r5.z, r2.w, r2.z
-mov.f32f32 r4.x, r4.x
-mov.f32f32 r5.x, r5.x
-mul.f r5.w, r2.w, r4.y
-mov.f32f32 r5.z, r5.z
-absneg.f r4.x, (abs)r4.x
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r5.w, r5.w
+max.f r5.y, r5.y, c13.y
+mul.f r6.y, r3.z, r0.x
absneg.f r5.z, (abs)r5.z
-mov.f32f32 r4.x, r4.x
-add.f r6.x, c15.x, r1.z
-absneg.f r5.w, (abs)r5.w
-mov.f32f32 r5.z, r5.z
-add.f r4.x, r4.x, c13.w
-mov.f32f32 r6.x, r6.x
-mov.f32f32 r5.w, r5.w
+max.f r4.z, r4.z, c13.y
+mul.f r5.x, r3.z, r5.x
+absneg.f r6.y, (abs)r6.y
add.f r5.z, r5.z, c13.w
-mov.f32f32 r4.x, r4.x
-mul.f r6.x, r6.x, c5.w
-add.f r5.w, r5.w, c13.w
-mov.f32f32 r5.z, r5.z
-max.f r4.x, r4.x, c13.y
-mov.f32f32 r6.y, r6.x
-mov.f32f32 r5.w, r5.w
-max.f r6.z, r5.z, c13.y
-mov.f32f32 r4.x, r4.x
-mov.f32f32 r5.z, r6.y
-max.f r5.w, r5.w, c13.y
-mov.f32f32 r6.y, r6.z
-add.f r6.z, r3.z, r4.x
-mul.f r3.w, r3.y, r3.w
-mov.f32f32 r6.w, r5.w
-add.f r7.x, r4.z, r6.y
-(ss)mul.f r2.z, r3.y, r2.z
-mov.f32f32 r3.w, r3.w
-add.f r7.y, r5.x, r6.w
-mul.f r4.y, r3.y, r4.y
-mov.f32f32 r2.z, r2.z
-absneg.f r3.w, (abs)r3.w
-bary.f r5.w, 2, r1.x
-mov.f32f32 r4.y, r4.y
-absneg.f r2.z, (abs)r2.z
-mov.f32f32 r3.w, r3.w
-add.f r7.z, r5.w, c14.w
+mul.f r6.z, c13.x, r3.w
+mul.f r0.z, r0.z, c15.z
+absneg.f r5.x, (abs)r5.x
+max.f r5.z, r5.z, c13.y
+add.f r6.y, r6.y, c13.w
+add.f r1.z, r1.z, (neg)r6.z
+add.f r5.x, r5.x, c13.w
+add.f r6.z, r3.x, r5.z
+mul.f r4.y, r4.x, r4.y
+max.f r6.w, r6.y, c13.y
+max.f r5.x, r5.x, c13.y
+mov.f32f32 r6.y, r1.z
+absneg.f r4.y, (abs)r4.y
+add.f r7.x, r5.y, r6.w
+add.f r7.y, r4.z, r5.x
+(ss)mul.f r0.x, r2.w, r0.x
+add.f r2.w, r4.y, c13.w
+mul.f r4.y, r4.x, r4.w
+add.f r4.w, c15.x, r6.y
+absneg.f r0.x, (abs)r0.x
+max.f r2.w, r2.w, c13.y
absneg.f r4.y, (abs)r4.y
-mov.f32f32 r2.z, r2.z
-add.f r3.w, r3.w, c13.w
-mov.f32f32 r5.w, r7.z
-mov.f32f32 r4.y, r4.y
-add.f r2.z, r2.z, c13.w
-mov.f32f32 r3.w, r3.w
-mov.f32f32 r5.w, r5.w
+mul.f r7.w, r4.w, c5.w
+add.f r0.x, r0.x, c13.w
+mov.f32f32 r4.w, r2.w
add.f r4.y, r4.y, c13.w
-mov.f32f32 r2.z, r2.z
-max.f r3.w, r3.w, c13.y
-mul.f r0.z, r3.x, r0.z
-mov.f32f32 r3.x, r4.y
-max.f r2.z, r2.z, c13.y
-mov.f32f32 r3.w, r3.w
-sam.s (f32)(x)r4.y, r5.y, s#4, t#4
-(sy)mov.f32f32 r4.y, r4.y
-max.f r3.x, r3.x, c13.y
-mov.f32f32 r2.z, r2.z
-(ss)add.f r5.y, r6.z, r3.w
-mov.f32f32 r4.y, r4.y
+mov.f32f32 r6.y, r7.w
+max.f r0.x, r0.x, c13.y
+add.f r4.w, r6.z, r4.w
+max.f r4.y, r4.y, c13.y
+bary.f r6.z, 2, r1.x
+mov.f32f32 r7.z, r0.x
+mov.f32f32 r8.x, r4.w
+mov.f32f32 r8.y, r4.y
+add.f r9.x, r6.z, c14.w
+add.f r7.x, r7.x, r7.z
+exp2 r0.z, r0.z
+(ss)mov.f32f32 r7.z, r0.z
+add.f r7.y, r7.y, r8.y
+mov.f32f32 r6.z, r9.x
+rcp r8.y, r8.x
mov.f32f32 r3.x, r3.x
-add.f r5.z, r7.x, r2.z
-mov.f32f32 r5.y, r5.y
-add.f r5.w, c14.y, (neg)r1.w
-add.f r6.z, r7.y, r3.x
-mov.f32f32 r5.z, r5.z
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r5.w, r5.w
-add.f r7.x, c14.y, (neg)r0.w
-rcp r7.y, r5.y
-(ss)mov.f32f32 r7.y, r7.y
-mov.f32f32 r6.z, r6.z
-mul.f r0.z, r0.z, c15.z
-rcp r7.w, r5.y
-nop
-rcp r8.x, r5.z
-mul.f r3.z, r3.z, r7.y
-(ss)mov.f32f32 r7.y, r8.x
-mov.f32f32 r7.x, r7.x
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r3.z, r3.z
-mul.f r2.x, r2.x, c14.x
-rcp r8.x, r6.z
-(ss)mov.f32f32 r8.x, r8.x
-mul.f r4.z, r4.z, r7.y
-mul.f r7.y, r5.w, r7.x
-mov.f32f32 r2.x, r2.x
-mul.f r5.x, r5.x, r8.x
+mov.f32f32 r8.z, r7.x
+mov.f32f32 r8.w, r7.y
+add.f r7.z, c17.y, (neg)r7.z
+(ss)mul.f r3.x, r3.x, r8.y
+mul.f r9.y, r2.z, c14.x
+sam.s (f32)(x)r9.w, r6.x, s#4, t#4
+rcp r2.z, r7.y
+(ss)mov.f32f32 r6.x, r0.w
+mul.f r6.y, r7.z, c10.y
+mov.f32f32 r6.z, r3.x
+mov.f32f32 r7.y, r9.y
+rcp r7.z, r8.w
mov.f32f32 r4.z, r4.z
-mul.f r4.y, r7.y, r4.y
-mov.f32f32 r7.y, r2.x
-mov.f32f32 r5.x, r5.x
-mov.f32f32 r8.x, r2.x
-mov.f32f32 r8.y, r2.x
-mov.f32f32 r8.z, r7.y
-mul.f r2.y, r2.y, c14.x
-mov.f32f32 r7.y, r8.x
-mov.f32f32 r8.x, r8.y
-add.f r0.y, c15.y, r0.y
-mov.f32f32 r2.y, r2.y
-mov.f32f32 r9.x, r7.y
-rcp r7.y, r6.z
-nop
-(ss)rcp r6.z, r6.z
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r8.y, r2.y
-mov.f32f32 r9.y, r2.y
-mov.f32f32 r9.z, r2.y
-mul.f r0.y, r0.y, c5.z
-mov.f32f32 r8.w, r8.y
-mov.f32f32 r9.y, r9.y
-mov.f32f32 r8.y, r9.z
-mov.f32f32 r9.z, r0.y
-(ss)mov.f32f32 r7.y, r7.y
-(ss)mov.f32f32 r6.z, r6.z
-mov.f32f32 r9.y, r9.y
-sam (f32)(xyzw)r9.w, r8.z, s#2, t#2
-(sy)(ss)mov.f32f32 r8.z, r9.w
-add.f r8.w, c13.z, (neg)r10.x
-mov.f32f32 r9.w, r10.y
-sam (f32)(xyzw)r10.x, r8.x, s#0, t#0
-(sy)(ss)mul.f r8.x, r10.z, r4.z
-mul.f r8.y, r8.z, r3.z
-mov.f32f32 r7.w, r7.w
-sam (f32)(xyzw)r10.z, r9.x, s#3, t#3
-(sy)mul.f r8.z, r10.w, r5.x
-(ss)mul.f r9.x, r10.z, r5.x
-mul.f r5.x, r11.x, r5.x
-mul.f r4.x, r4.x, r7.w
-mul.f r6.w, r6.w, r7.y
-mov.f32f32 r7.y, r8.w
-mul.f r7.w, r9.w, r3.z
-mov.f32f32 r4.x, r4.x
-mov.f32f32 r8.w, r2.y
-mov.f32f32 r6.w, r6.w
-mov.f32f32 r9.y, r2.y
-mov.f32f32 r7.y, r7.y
-mov.f32f32 r10.z, r8.w
-bary.f r8.w, 22, r1.x
-mov.f32f32 r11.x, r9.y
-mul.f r3.z, r7.y, r3.z
-rcp r7.y, r5.z
-(ss)mov.f32f32 r7.y, r7.y
-mov.f32f32 r8.w, r8.w
-mov.f32f32 r9.y, r9.z
-mul.f r10.y, r10.y, r4.z
-mul.f r4.z, r10.x, r4.z
-mul.f r8.w, r8.w, c14.x
-mul.f r6.y, r6.y, r7.y
-mov.f32f32 r6.x, r6.x
-mul.f r3.x, r3.x, r6.z
-mov.f32f32 r6.z, r8.w
-mov.f32f32 r6.y, r6.y
-mov.f32f32 r2.y, r2.y
-mov.f32f32 r9.z, r6.x
-mov.f32f32 r6.x, r6.z
-mov.f32f32 r7.y, r6.z
-mov.f32f32 r11.z, r2.y
-mov.f32f32 r2.y, r6.z
-mov.f32f32 r10.w, r6.x
-mov.f32f32 r11.y, r7.y
-mov.f32f32 r6.x, r7.z
-mov.f32f32 r11.w, r2.y
-mov.f32f32 r2.y, r3.x
-exp2 r0.z, r0.z
-(ss)mov.f32f32 r0.z, r0.z
-mov.f32f32 r9.w, r6.x
-sam (f32)(xyzw)r12.x, r10.z, s#2, t#2
-(sy)mov.f32f32 r3.x, r12.y
-(ss)nop
-sam (f32)(xyzw)r10.z, r11.x, s#3, t#3
-(sy)mad.f32 r6.x, r10.w, r6.w, r8.z
-mad.f32 r7.y, r10.z, r6.w, r9.x
-mad.f32 r5.x, r11.x, r6.w, r5.x
-mad.f32 r3.x, r3.x, r4.x, r8.y
-rcp r5.y, r5.y
-(ss)mov.f32f32 r5.y, r5.y
-mov.f32f32 r6.w, r2.x
-add.f r8.y, c13.z, (neg)r12.x
-mov.f32f32 r8.z, r12.z
-mul.f r3.w, r3.w, r5.y
-mov.f32f32 r8.w, r6.w
-mov.f32f32 r5.y, r6.z
-mov.f32f32 r6.w, r8.y
-mov.f32f32 r3.w, r3.w
-mov.f32f32 r8.y, r2.x
-mov.f32f32 r9.x, r5.y
-mov.f32f32 r5.y, r6.w
-mad.f32 r6.w, r8.z, r4.x, r7.w
-mov.f32f32 r8.y, r8.y
-mov.f32f32 r7.w, r6.z
-mad.f32 r3.z, r5.y, r4.x, r3.z
-sam (f32)(xyzw)r10.z, r11.z, s#0, t#0
-(sy)mad.f32 r4.x, r10.w, r6.y, r10.y
-(ss)nop
-sam (f32)(xyzw)r11.y, r8.w, s#3, t#3
-(sy)mad.f32 r5.x, r11.w, r2.y, r5.x
-mov.f32f32 r8.z, r7.w
-mad.f32 r5.y, r11.z, r2.y, r6.x
-mad.f32 r2.y, r11.y, r2.y, r7.y
-mov.f32f32 r5.x, r5.x
-mad.f32 r6.x, r11.x, r6.y, r8.x
-rcp r5.z, r5.z
+rcp r7.w, r8.z
mov.f32f32 r5.y, r5.y
-mov.f32f32 r2.y, r2.y
-sam (f32)(xyzw)r7.w, r8.y, s#2, t#2
-(sy)mov.f32f32 r7.y, r8.x
-mul.f r5.x, c7.z, r5.x
-mul.f r5.y, c7.y, r5.y
-mul.f r2.y, c7.x, r2.y
-mad.f32 r3.x, r7.y, r3.w, r3.x
-add.f r7.y, c13.z, (neg)r7.w
-mov.f32f32 r7.w, r8.y
-(ss)mov.f32f32 r5.z, r5.z
-mov.f32f32 r3.x, r3.x
-mov.f32f32 r7.y, r7.y
-mad.f32 r6.w, r7.w, r3.w, r6.w
-mul.f r2.z, r2.z, r5.z
-mad.f32 r3.x, c14.y, r3.x, c14.z
-mov.f32f32 r5.z, r7.y
-mov.f32f32 r6.w, r6.w
-mov.f32f32 r2.z, r2.z
-mov.f32f32 r3.x, r3.x
-bary.f r7.y, 4, r1.x
-bary.f r7.w, 5, r1.x
-bary.f r8.x, 6, r1.x
-mad.f32 r3.z, r5.z, r3.w, r3.z
-mul.f r3.w, r7.y, r3.x
-mul.f r5.z, r7.w, r3.x
-mul.f r3.x, r8.x, r3.x
-mov.f32f32 r3.z, r3.z
-mad.f32 r6.w, c14.y, r6.w, c14.z
-mov.f32f32 r2.x, r2.x
-mad.f32 r4.z, r10.z, r6.y, r4.z
-mad.f32 r3.z, c14.y, r3.z, c14.z
-mov.f32f32 r6.y, r6.w
-mov.f32f32 r7.w, r2.x
-mov.f32f32 r2.x, r6.z
-mov.f32f32 r3.z, r3.z
-bary.f r6.z, 10, r1.x
-bary.f r6.w, 11, r1.x
-bary.f r7.y, 12, r1.x
-mov.f32f32 r8.x, r2.x
-mov.f32f32 r2.x, r6.z
-bary.f r6.z, 7, r1.x
+rcp r8.y, r8.w
+mov.f32f32 r5.x, r5.x
+mov.f32f32 r10.x, r7.y
+(ss)bary.f r8.w, 24, r1.x
+(ss)mul.f r4.z, r4.z, r7.z
+mul.f r5.y, r5.y, r7.w
+mov.f32f32 r10.z, r7.y
+mul.f r11.x, r8.w, c14.x
+mov.f32f32 r7.w, r4.z
+mov.f32f32 r8.w, r5.y
+mul.f r5.x, r5.x, r8.y
+mov.f32f32 r7.z, r11.x
+add.f r3.w, c14.y, (neg)r3.w
+mul.f r0.z, r0.z, c13.z
+mul.f r2.z, r4.y, r2.z
+mov.f32f32 r10.y, r7.z
+mov.f32f32 r10.w, r7.z
+mov.f32f32 r4.y, r5.x
+sam (f32)(xyzw)r11.y, r7.y, s#0, t#0
+(sy)mul.f r8.y, r11.w, r8.w
+rcp r8.z, r8.z
mov.f32f32 r6.w, r6.w
-mov.f32f32 r7.y, r7.y
-sam.s (f32)(x)r8.y, r9.y, s#4, t#4
-(sy)mov.f32f32 r8.y, r8.y
-mul.f r2.x, r2.x, (neg)r6.z
-mul.f r6.w, r6.w, (neg)r6.z
-mul.f r6.z, r7.y, (neg)r6.z
+mul.f r8.w, r11.z, r8.w
+mul.f r5.y, r11.y, r5.y
+sam (f32)(xyzw)r11.y, r10.x, s#2, t#2
+(sy)(ss)mul.f r10.x, r11.y, r6.z
+rcp r8.x, r8.x
+mov.f32f32 r5.z, r5.z
+sam (f32)(xyzw)r12.x, r10.z, s#3, t#3
+(sy)mul.f r10.y, r12.y, r7.w
+add.f r9.z, c13.z, (neg)r11.z
+(ss)mul.f r10.z, r12.z, r7.w
+(ss)mul.f r5.z, r5.z, r8.x
+mov.f32f32 r12.y, r7.z
+bary.f r7.w, 22, r1.x
+mul.f r6.z, r9.z, r6.z
+mov.f32f32 r8.x, r5.z
+mul.f r6.w, r6.w, r8.z
+mul.f r9.z, r7.w, c14.x
+mul.f r4.z, r12.x, r4.z
+mul.f r3.x, r11.w, r3.x
+mov.f32f32 r8.z, r6.w
+mov.f32f32 r11.y, r9.z
+mov.f32f32 r10.w, r3.w
+add.f r0.z, r0.z, r6.y
+mov.f32f32 r6.y, r2.z
+mov.f32f32 r7.w, r11.y
+mov.f32f32 r12.z, r11.y
+mul.f r6.x, r6.x, r10.w
+sam (f32)(xyzw)r12.w, r11.x, s#0, t#0
+(sy)mad.f32 r8.y, r13.y, r8.z, r8.y
+rcp r7.x, r7.x
+(ss)mul.f r0.x, r0.x, r7.x
+(ss)mad.f32 r7.x, r13.x, r8.z, r8.w
+mad.f32 r5.y, r12.w, r6.w, r5.y
+sam (f32)(xyzw)r12.w, r7.z, s#2, t#2
+(sy)mad.f32 r6.w, r13.x, r8.x, r10.x
+rcp r4.w, r4.w
+(ss)mul.f r2.w, r2.w, r4.w
+sam (f32)(xyzw)r11.z, r12.y, s#3, t#3
+(sy)(ss)mad.f32 r4.w, r11.w, r4.y, r10.y
+add.f r7.z, c13.z, (neg)r12.w
+mad.f32 r4.y, r12.x, r4.y, r10.z
+mov.f32f32 r8.z, r2.w
+mov.f32f32 r10.x, r7.y
+mov.f32f32 r10.y, r11.y
+mov.f32f32 r10.z, r7.y
+mov.f32f32 r10.w, r11.y
+mad.f32 r6.z, r7.z, r8.x, r6.z
+mov.f32f32 r7.y, r0.x
+sam (f32)(xyzw)r11.w, r9.y, s#0, t#0
+mad.f32 r4.z, r11.z, r5.x, r4.z
+mad.f32 r3.x, r13.y, r5.z, r3.x
+sam (f32)(xyzw)r12.z, r10.x, s#2, t#2
+(sy)mad.f32 r5.x, r12.w, r8.z, r6.w
+mad.f32 r5.z, r12.y, r7.y, r8.y
(ss)nop
-sam (f32)(xyzw)r8.z, r7.w, s#0, t#0
-(sy)mad.f32 r6.x, r9.x, r2.z, r6.x
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r6.w, r6.w
-mov.f32f32 r6.z, r6.z
-mov.f32f32 r6.x, r6.x
-mad.f32 r2.x, r2.x, r3.z, r3.w
-mad.f32 r3.w, r6.w, r3.z, r5.z
-mad.f32 r3.x, r6.z, r3.z, r3.x
-mad.f32 r3.z, r8.w, r2.z, r4.x
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r3.w, r3.w
-mad.f32 r0.x, r0.x, r6.y, r2.x
-mad.f32 r2.x, r2.w, r6.y, r3.w
-mov.f32f32 r2.w, r3.x
-mov.f32f32 r3.x, r3.z
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r2.x, r2.x
-mad.f32 r2.w, r3.y, r6.y, r2.w
-nop
-mul.f r3.y, r0.x, r0.x
-mad.f32 r2.z, r8.z, r2.z, r4.z
-mad.f32 r3.y, r2.x, r2.x, r3.y
-mov.f32f32 r2.w, r2.w
-add.f r1.w, r1.w, c13.z
-add.f r3.z, c17.y, (neg)r0.z
-mov.f32f32 r3.y, r3.y
-mov.f32f32 r2.z, r2.z
-mad.f32 r3.y, r2.w, r2.w, r3.y
-mul.f r3.w, r1.w, r7.x
-mul.f r3.z, r3.z, c10.y
-mul.f r0.z, r0.z, c13.z
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r4.x, r4.w
-add.f r1.z, c15.y, r1.z
-rsq r3.y, r3.y
-(ss)mov.f32f32 r3.y, r3.y
-mad.f32 r3.w, r3.w, r8.y, r4.y
-add.f r0.z, r0.z, r3.z
-mov.f32f32 r0.y, r0.y
-mul.f r0.x, r0.x, r3.y
-mul.f r2.x, r2.x, r3.y
-mul.f r2.w, r2.w, r3.y
-nop
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r2.w, r2.w
-mov.f32f32 r3.y, r3.w
-mul.f r3.z, r0.x, r0.x
-mul.f r3.w, (neg)c8.x, r0.x
-mad.f32 r3.z, r2.x, r2.x, r3.z
-mad.f32 r3.w, (neg)c8.y, r2.x, r3.w
-mov.f32f32 r4.x, r4.x
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r3.z, r3.z
-mov.f32f32 r3.w, r3.w
-mad.f32 r3.z, r2.w, r2.w, r3.z
-mad.f32 r3.w, (neg)c8.z, r2.w, r3.w
-mul.f r1.z, r1.z, c5.w
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r6.y, r0.y
-add.f r0.y, r0.w, c13.z
-bary.f r0.w, 8, r1.x
-rsq r3.z, r3.z
-(ss)mov.f32f32 r3.z, r3.z
-mov.f32f32 r3.w, r3.w
-mov.f32f32 r4.y, r1.z
-mov.f32f32 r0.z, r0.z
-mul.f r0.x, r0.x, r3.z
-max.f r3.w, r3.w, c13.y
-mul.f r2.x, r2.x, r3.z
-mul.f r2.w, r2.w, r3.z
-mov.f32f32 r0.x, r0.x
-bary.f r3.z, 13, r1.x
-mov.f32f32 r3.w, r3.w
-bary.f r4.z, 17, r1.x
-bary.f r4.w, 16, r1.x
-mul.f r0.x, r0.x, r3.z
-mov.f32f32 r2.x, r2.x
-bary.f r3.z, 14, r1.x
-bary.f r5.z, 18, r1.x
-mad.f32 r6.z, c7.y, r3.w, (neg)r4.z
-mad.f32 r6.w, c7.x, r3.w, (neg)r4.w
-mad.f32 r0.x, r2.x, r3.z, r0.x
-mad.f32 r2.x, c7.z, r3.w, (neg)r5.z
-mov.f32f32 r3.z, r6.z
-mov.f32f32 r3.w, r6.w
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r2.w, r2.w
-bary.f r6.z, 15, r1.x
-mov.f32f32 r2.x, r2.x
-mad.f32 r3.z, c11.x, r3.z, r4.z
-mad.f32 r3.w, c11.x, r3.w, r4.w
-mad.f32 r0.x, r2.w, r6.z, r0.x
-mad.f32 r2.x, c11.x, r2.x, r5.z
-mov.f32f32 r2.w, r3.z
-mov.f32f32 r3.z, r3.w
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r4.y, r4.y
-mov.f32f32 r3.w, r7.z
-max.f r0.x, c13.y, r0.x
+sam (f32)(xyzw)r10.x, r10.z, s#3, t#3
+(sy)mad.f32 r4.y, r10.z, r6.y, r4.y
+mad.f32 r4.w, r10.y, r6.y, r4.w
+mad.f32 r5.x, c14.y, r5.x, c14.z
+bary.f r6.y, 6, r1.x
+mul.f r4.y, c7.z, r4.y
+mul.f r4.w, c7.y, r4.w
+mov.f32f32 r6.w, r5.x
+bary.f r7.z, 4, r1.x
+bary.f r8.x, 5, r1.x
+mul.f r5.x, r6.y, r5.x
+add.f r6.y, c13.z, (neg)r12.z
+mul.f r7.z, r7.z, r6.w
+mul.f r6.w, r8.x, r6.w
+mov.f32f32 r8.y, r5.z
+mad.f32 r6.y, r6.y, r8.z, r6.z
+mad.f32 r6.z, r12.x, r7.y, r7.x
+mad.f32 r2.z, r10.x, r2.z, r4.z
+mad.f32 r2.w, r13.x, r2.w, r3.x
+mad.f32 r3.x, c14.y, r6.y, c14.z
+bary.f r4.z, 12, r1.x
+bary.f r6.y, 7, r1.x
+mov.f32f32 r7.x, r6.z
+mov.f32f32 r7.y, r3.x
+bary.f r8.x, 10, r1.x
+bary.f r8.z, 11, r1.x
+mul.f r4.z, r4.z, (neg)r6.y
+mul.f r9.y, c7.x, r2.z
+mul.f r2.z, r8.x, (neg)r6.y
+mul.f r6.y, r8.z, (neg)r6.y
+mad.f32 r3.x, r4.z, r3.x, r5.x
+mad.f32 r2.w, c14.y, r2.w, c14.z
+mad.f32 r2.z, r2.z, r7.y, r7.z
+mad.f32 r4.z, r6.y, r7.y, r6.w
+mad.f32 r0.x, r11.w, r0.x, r5.y
+mov.f32f32 r5.x, r2.w
+mad.f32 r2.w, r4.x, r2.w, r3.x
+(rpt1)nop
+mad.f32 r1.w, r1.w, r5.x, r2.z
+mad.f32 r2.z, r3.z, r5.x, r4.z
+(rpt1)nop
+mov.f32f32 r3.x, r1.w
+mov.f32f32 r3.z, r2.z
+mov.f32f32 r4.x, r2.w
+mov.f32f32 r4.z, r0.x
+mul.f r1.w, r1.w, r3.x
+mul.f r5.x, r6.x, r9.w
+mad.f32 r1.w, r2.z, r3.z, r1.w
+add.f r0.y, c15.y, r0.y
+mad.f32 r1.w, r4.x, r4.x, r1.w
max.f r0.z, r0.z, c13.y
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r4.z, r3.w
-mov.f32f32 r0.x, r0.x
+add.f r1.z, c15.y, r1.z
+mul.f r8.z, r0.y, c5.z
+add.f r0.y, r2.x, c13.z
+add.f r2.x, r3.y, c13.z
+bary.f r2.z, 9, r1.x
+rsq r1.w, r1.w
+(ss)mov.f32f32 r3.y, r1.w
+(ss)mul.f r1.w, r2.w, r1.w
+mov.f32f32 r7.z, r8.z
min.f r0.z, r0.z, c13.z
-mov.f32f32 r1.z, r1.z
-mul.f r3.w, r5.w, r0.y
-mov.f32f32 r0.w, r0.w
-mul.f r0.y, r1.w, r0.y
-sam.s (f32)(x)r1.w, r4.x, s#4, t#4
-(sy)mov.f32f32 r1.w, r1.w
-log2 r0.x, r0.x
-(ss)mul.f r0.x, c11.y, r0.x
-add.f r4.x, c17.y, (neg)r0.z
-add.f r4.y, c17.y, (neg)r0.z
-add.f r4.z, c17.y, (neg)r0.z
-mov.f32f32 r0.x, r0.x
-mad.f32 r1.w, r3.w, r1.w, r3.y
-mul.f r3.y, r4.x, c9.z
-mul.f r3.w, r4.y, c9.y
-mul.f r4.x, r4.z, c9.x
-mov.f32f32 r1.w, r1.w
+mul.f r2.w, r3.x, r3.y
+mul.f r3.x, r3.z, r3.y
+mov.f32f32 r3.y, r1.w
+mov.f32f32 r8.x, r9.x
+mov.f32f32 r3.z, r2.w
+mul.f r2.w, (neg)c8.x, r2.w
+mov.f32f32 r4.x, r3.x
+mad.f32 r2.w, (neg)c8.y, r3.x, r2.w
+mul.f r3.x, r3.z, r3.z
+mad.f32 r1.w, (neg)c8.z, r1.w, r2.w
+mad.f32 r2.w, r4.x, r4.x, r3.x
+sam.s (f32)(x)r7.y, r7.z, s#4, t#4
+mul.f r3.x, r0.y, r3.w
+mad.f32 r2.w, r3.y, r3.y, r2.w
+max.f r1.w, r1.w, c13.y
+bary.f r3.w, 16, r1.x
+(sy)mad.f32 r3.x, r3.x, r7.y, r5.x
+add.f r5.x, c17.y, (neg)r0.z
+add.f r5.y, c17.y, (neg)r0.z
+add.f r6.x, c17.y, (neg)r0.z
+rsq r2.w, r2.w
+(ss)mov.f32f32 r6.y, r2.w
+mov.f32f32 r6.w, r1.w
+bary.f r7.y, 17, r1.x
+(ss)bary.f r7.z, 18, r1.x
+mul.f r3.z, r3.z, r6.y
+bary.f r7.w, 13, r1.x
+mad.f32 r8.x, c7.z, r6.w, (neg)r7.z
+mad.f32 r6.w, c7.y, r6.w, (neg)r7.y
+mul.f r4.x, r4.x, r6.y
+mul.f r3.z, r3.z, r7.w
+bary.f r6.y, 14, r1.x
+mad.f32 r7.z, c11.x, r8.x, r7.z
+mad.f32 r6.w, c11.x, r6.w, r7.y
+mul.f r2.w, r3.y, r2.w
+mad.f32 r3.y, r4.x, r6.y, r3.z
+bary.f (ei)r1.x, 15, r1.x
+mad.f32 r1.y, c7.x, r1.w, (neg)r3.w
+(rpt1)nop
+mad.f32 r1.x, r2.w, r1.x, r3.y
+mad.f32 r1.y, c11.x, r1.y, r3.w
+mul.f r8.w, r1.z, c5.w
+mul.f r1.z, r5.x, c9.z
+max.f r1.x, c13.y, r1.x
+mul.f r1.w, r5.y, c9.y
+mul.f r3.y, r6.x, c9.x
+mov.f32f32 r6.x, r8.w
+mov.f32f32 r6.y, r9.x
nop
-exp2 r0.x, r0.x
-(ss)mul.f r4.y, r5.x, r0.x
-mul.f r4.z, r5.y, r0.x
-mad.f32 r2.x, r6.x, r2.x, r4.y
-mad.f32 r2.w, r3.x, r2.w, r4.z
-(ss)mul.f r0.x, r2.y, r0.x
-mov.f32f32 r6.z, r1.z
-mov.f32f32 r1.z, r2.x
-mov.f32f32 r2.x, r7.z
-mov.f32f32 r2.y, r2.w
-mad.f32 r0.x, r2.z, r3.z, r0.x
-mov.f32f32 r4.y, r0.w
-mov.f32f32 r6.w, r2.x
-bary.f (ei)r0.w, 9, r1.x
-mov.f32f32 r1.x, c13.z
-mov.f32f32 r1.y, c13.y
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r5.z, r1.x
-sam.s (f32)(x)r1.x, r6.y, s#4, t#4
-(sy)mov.f32f32 r1.x, r1.x
-(rpt2)nop
-mad.f32 r0.y, r0.y, r1.x, r1.w
-mov.f32f32 r4.z, r0.w
+sam.s (f32)(x)r9.z, r8.z, s#4, t#4
+mul.f r0.w, r0.w, r2.x
+log2 r1.x, r1.x
+(ss)mul.f r1.x, c11.y, r1.x
+mul.f r0.y, r0.y, r2.x
+sam (f32)(w)r9.w, r2.y, s#1, t#1
+(sy)cmps.f.lt r2.x, r10.z, c16.x
+mov.f32f32 r2.w, c13.z
+(ss)mov.f32f32 r2.y, c13.y
(rpt1)nop
-mov.f32f32 r0.y, r0.y
-(rpt2)nop
-mul.f r0.y, c15.w, r0.y
-sam (f32)(w)r6.y, r4.y, s#1, t#1
-(sy)mov.f32f32 r0.w, r7.x
-cmps.f.lt r1.x, r7.x, c16.x
+exp2 r1.x, r1.x
+(ss)mul.f r2.z, r4.y, r1.x
+mul.f r3.z, r4.w, r1.x
+mad.f32 r2.z, r8.y, r7.z, r2.z
+sam.s (f32)(x)r7.y, r5.w, s#4, t#4
+(sy)mad.f32 r0.w, r0.w, r7.y, r3.x
+mad.f32 r3.x, r7.x, r6.w, r3.z
+mad.f32 r0.y, r0.y, r9.z, r0.w
+mul.f r0.w, r9.y, r1.x
+(ss)cov.u32f32 r1.x, r2.x
nop
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.w, r0.w
-cov.u32f32 r1.x, r1.x
+mul.f r0.y, c15.w, r0.y
+mad.f32 r0.w, r4.z, r1.y, r0.w
+cmps.f.ne r1.x, r1.x, c13.y
nop
-mul.f r1.z, r1.z, r0.y
-mul.f r1.w, r2.y, r0.y
-mul.f r0.x, r0.x, r0.y
-cmps.f.ne r0.y, r1.x, c13.y
-mov.f32f32 r1.x, r1.z
-mov.f32f32 r1.z, r1.w
-mad.f32 r1.x, c6.z, r6.x, r1.x
-mad.f32 r1.z, c6.y, r3.x, r1.z
-mov.f32f32 r0.x, r0.x
-sel.b32 r0.y, r1.y, r0.y, r0.w
-mov.f32f32 r0.w, r1.x
-mov.f32f32 r1.x, r1.z
-mad.f32 r0.x, c6.x, r2.z, r0.x
+mov.f32f32 r1.y, r0.y
+mul.f r0.y, r0.w, r0.y
+sel.b32 r0.w, r2.y, r1.x, r10.z
nop
-mul.f r0.w, r0.z, r0.w
-mul.f r1.x, r0.z, r1.x
-mov.f32f32 r0.x, r0.x
+mul.f r1.x, r2.z, r1.y
+mul.f r1.y, r3.x, r1.y
+mad.f32 r1.x, c6.z, r5.z, r1.x
+mad.f32 r1.y, c6.y, r6.z, r1.y
+mad.f32 r0.x, c6.x, r0.x, r0.y
nop
-add.f r0.w, r0.w, r3.y
-add.f r1.x, r1.x, r3.w
+mul.f r0.y, r0.z, r1.x
+mul.f r1.x, r0.z, r1.y
mul.f r0.x, r0.z, r0.x
nop
-mul.f r0.z, r0.w, r0.y
-mul.f r0.w, r1.x, r0.y
-add.f r0.x, r0.x, r4.x
+add.f r0.y, r0.y, r1.z
+add.f r0.z, r1.x, r1.w
+add.f r0.x, r0.x, r3.y
nop
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r0.w, r0.w
-mul.f r0.x, r0.x, r0.y
+mul.f r2.z, r0.y, r0.w
+mul.f r2.y, r0.z, r0.w
+mul.f r2.x, r0.x, r0.w
+end
nop
-mov.f32f32 r0.y, r0.z
-mov.f32f32 r0.z, r0.w
-mov.f32f32 r0.x, r0.x
nop
-mov.f32f32 r5.y, r0.y
-mov.f32f32 r5.x, r0.z
-mov.f32f32 r0.x, r0.x
-(rpt2)nop
-mov.f32f32 r4.w, r0.x
-end
-; FRAG: outputs: r4.w (1:0)
-; FRAG: inputs: r0.x (0:0,cm=f,il=8,b=0) r1.x (5:9,cm=f,il=8,b=1) r5.w (5:10,cm=f,il=12,b=1) r6.x (5:11,cm=f,il=16,b=1) r5.w (5:12,cm=f,il=20,b=1) r63.y (5:13,cm=f,il=24,b=1) r1.y (5:14,cm=f,il=28,b=1) r7.x (5:15,cm=f,il=32,b=1)
-; FRAG: 539 instructions, 0 half, 65 full
-; pos (bary): r1.x
-; color: r4.w
-; fragcoord: r0.x
+; FRAG: outputs: r2.x (1:0)
+; FRAG: inputs: r0.x (0:0,cm=f,il=8,b=0) r1.x (5:9,cm=f,il=8,b=1) r5.z (5:10,cm=f,il=12,b=1) r7.w (5:11,cm=f,il=16,b=1) r0.y (5:12,cm=f,il=20,b=1) r63.y (5:13,cm=f,il=24,b=1) r1.w (5:14,cm=f,il=28,b=1) r7.x (5:15,cm=f,il=32,b=1)
+; FRAG: 354 instructions, 0 half, 14 full