1 files changed, 406 insertions, 635 deletions
diff --git a/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-70.asm b/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-70.asm
index 60492a1..2d7ae09 100644
--- a/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-70.asm
+++ b/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-70.asm
@@ -4,710 +4,481 @@
 @in(r0.w)	in3
 @in(r1.x)	in4
 @in(r1.y)	in5
-@out(r3.x)	out0
-@out(r3.y)	out1
-@out(r3.z)	out2
-@out(r3.w)	out3
-(sy)(ss)bary.f r0.x, 7, r1.x
-bary.f r0.y, 8, r1.x
+@out(r2.x)	out0
+@out(r2.y)	out1
+@out(r2.z)	out2
+@out(r2.w)	out3
+@const(c14.x)	0x3f000000, 0x00000000, 0x43160000, 0x3bdb8bac
+@const(c15.x)	0x41000000, 0x3f600000, 0x3e000000, 0x3d4ccccd
+@const(c16.x)	0x40000000, 0xbf800000, 0xbb449ba6, 0xbf000000
+@const(c17.x)	0x3f800000, 0xbf000000, 0x3de38866, 0x3fb8aa65
+@const(c18.x)	0x3cf5c28f, 0x00000000, 0x00000000, 0x00000000
+@const(c19.x)	0x00000000, 0x3f800000, 0x43000000, 0x00000000
+(sy)(ss)bary.f r0.x, 20, r1.x
+bary.f r0.y, 7, r1.x
+bary.f r1.z, 8, r1.x
 add.f r0.w, r0.w, c14.y
-bary.f r1.z, 9, r1.x
-mov.f32f32 r0.x, r0.x
-bary.f r1.w, 20, r1.x
-add.f r2.x, r0.y, c16.w
-add.f r2.y, r1.z, c16.w
-bary.f r2.z, 18, r1.x
-mul.f r2.w, r1.w, r0.x
-bary.f r3.x, 15, r1.x
-floor.f r3.y, r2.x
+bary.f r1.w, 9, r1.x
+mul.f r2.x, r0.x, r0.y
+bary.f r2.y, 21, r1.x
+bary.f r2.z, 15, r1.x
+add.f r2.w, r1.z, c16.w
+add.f r3.y, r1.w, c16.w
 rcp r0.w, r0.w
 add.f r0.z, r0.z, c14.y
-floor.f r3.z, r2.y
-mov.f32f32 r3.x, r3.x
-bary.f r3.w, 21, r1.x
-add.f r2.x, r2.x, (neg)r3.y
+mad.f32 r2.x, r2.y, r2.z, r2.x
+bary.f r3.z, 22, r1.x
+bary.f r3.w, 3, r1.x
+floor.f r4.x, r2.w
 (ss)mul.f r0.z, r0.z, r0.w
-(ss)add.f r0.w, r2.y, (neg)r3.z
-mad.f32 r2.y, r3.w, r3.x, r2.w
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r0.z, r0.z
-absneg.f r2.w, (neg)c11.x
-mov.f32f32 r2.y, r2.y
-bary.f r3.y, 3, r1.x
-mul.f r3.z, c14.x, r2.x
-mul.f r2.w, r2.w, c11.x
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r3.y, r3.y
-bary.f r4.x, 22, r1.x
-mov.f32f32 r3.z, r3.z
-mul.f r2.w, r2.w, r0.z
-mul.f r4.y, c14.x, r0.w
-mad.f32 r2.y, r4.x, r3.y, r2.y
-add.f r0.y, r0.y, (neg)r3.z
-mov.f32f32 r2.w, r2.w
-mov.f32f32 r3.z, r4.y
-mov.f32f32 r2.y, r2.y
-bary.f r4.y, 12, r1.x
-mov.f32f32 r0.y, r0.y
-mul.f r0.z, r2.w, r0.z
-add.f r1.z, r1.z, (neg)r3.z
-mul.f r2.w, r1.w, r4.y
-bary.f r3.z, 13, r1.x
-add.f r4.z, c17.y, r0.y
+(ss)absneg.f r0.w, (neg)c11.x
+mad.f32 r2.x, r3.z, r3.w, r2.x
+add.f r2.w, r2.w, (neg)r4.x
+floor.f r4.x, r3.y
+mul.f r0.w, r0.w, c11.x
+mov.f32f32 r4.y, r2.x
+bary.f r4.z, 12, r1.x
+mov.f32f32 r4.w, r2.w
+mul.f r0.w, r0.w, r0.z
 mov.f32f32 r0.z, r0.z
-add.f r0.y, c17.x, r0.y
-mad.f32 r2.w, r3.w, r3.z, r2.w
-mov.f32f32 r4.z, r4.z
+mul.f r5.x, r0.x, r4.z
+bary.f r5.y, 13, r1.x
+mul.f r5.z, c14.x, r4.w
+mul.f r0.z, r0.w, r0.z
+add.f r0.w, r3.y, (neg)r4.x
+mad.f32 r3.y, r2.y, r5.y, r5.x
+bary.f r4.x, 14, r1.x
+add.f r1.z, r1.z, (neg)r5.z
 mul.f r0.z, r0.z, c17.w
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r2.w, r2.w
-bary.f r4.w, 14, r1.x
-mul.f r4.z, r4.z, c5.z
-mov.f32f32 r0.z, r0.z
-mul.f r0.y, r0.y, c5.z
-mad.f32 r2.w, r4.x, r4.w, r2.w
-mov.f32f32 r5.x, r4.z
-mov.f32f32 r4.z, r4.z
-mov.f32f32 r5.y, r0.y
-mul.f r5.z, r2.w, r2.w
-mov.f32f32 r5.w, r5.x
-mad.f32 r5.x, r2.y, r2.y, r5.z
-mov.f32f32 r1.z, r1.z
+mov.f32f32 r5.x, r0.w
+mad.f32 r3.y, r3.z, r4.x, r3.y
+mov.f32f32 r5.z, r1.z
+add.f r1.z, c17.x, r1.z
+mul.f r5.w, c14.x, r5.x
+mul.f r6.x, r3.y, r3.y
+add.f r5.z, c17.y, r5.z
+mad.f32 r2.x, r2.x, r4.y, r6.x
+bary.f r6.x, 4, r1.x
 exp2 r0.z, r0.z
-(ss)mov.f32f32 r0.z, r0.z
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r5.x, r5.x
-bary.f r5.z, 4, r1.x
-add.f r6.x, c17.y, r1.z
-add.f r6.y, c19.y, (neg)r0.z
-mov.f32f32 r0.y, r0.y
-mul.f r6.z, r1.w, r5.z
-bary.f r6.w, 5, r1.x
-mov.f32f32 r6.x, r6.x
+(ss)mov.f32f32 r6.y, r0.z
+mul.f r6.z, r5.z, c5.z
+add.f r1.w, r1.w, (neg)r5.w
+mul.f r5.z, r0.x, r6.x
+bary.f r5.w, 5, r1.x
+mov.f32f32 r7.y, r6.z
+add.f r6.y, c19.y, (neg)r6.y
+mov.f32f32 r6.w, r1.w
+mad.f32 r5.z, r2.y, r5.w, r5.z
+bary.f r8.x, 6, r1.x
 mul.f r6.y, r6.y, c11.y
-mul.f r0.z, r0.z, c17.x
-mad.f32 r6.z, r3.w, r6.w, r6.z
-mul.f r7.x, r6.x, c5.w
-mov.f32f32 r7.y, r0.y
-add.f r0.y, c17.x, r1.z
-mov.f32f32 r1.z, r6.z
-bary.f r6.z, 6, r1.x
-mov.f32f32 r6.x, r7.x
+(ss)mul.f r0.z, r0.z, c17.x
+add.f r6.w, c17.y, r6.w
+mad.f32 r5.z, r3.z, r8.x, r5.z
+mul.f r8.y, r1.z, c5.z
+add.f r1.z, c17.x, r1.w
 add.f r0.z, r0.z, r6.y
-mov.f32f32 r0.y, r0.y
-mad.f32 r1.z, r4.x, r6.z, r1.z
-mov.f32f32 r6.x, r6.x
-bary.f r6.y, 10, r1.x
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r1.z, r1.z
-mul.f r0.y, r0.y, c5.w
-add.f r8.x, r6.y, c16.z
-mov.f32f32 r8.y, r5.y
-mad.f32 r5.x, r1.z, r1.z, r5.x
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r5.y, r8.x
-mov.f32f32 r7.z, r0.y
-mov.f32f32 r7.x, r7.x
-mov.f32f32 r9.x, r4.z
-mov.f32f32 r6.y, r5.y
-rsq r4.z, r5.x
-(ss)mov.f32f32 r4.z, r4.z
+mov.f32f32 r1.w, r5.z
+mul.f r9.y, r6.w, c5.w
+mov.f32f32 r9.x, r8.y
+mul.f r8.z, r1.z, c5.w
+mad.f32 r1.z, r5.z, r1.w, r2.x
+mov.f32f32 r7.z, r9.y
 max.f r0.z, r0.z, c14.y
-(ss)mov.f32f32 r5.x, r7.z
-mov.f32f32 r8.z, r7.x
-mul.f r1.z, r1.z, r4.z
+bary.f r2.x, 10, r1.x
+mov.f32f32 r6.w, r8.z
+add.f r4.w, c16.x, (neg)r4.w
+mul.f r0.x, r0.x, r0.x
+rsq r1.z, r1.z
+(ss)mov.f32f32 r5.z, r1.z
+add.f r8.w, r2.x, c16.z
 min.f r0.z, r0.z, c17.x
-sam.s (f32)(x)r5.y, r5.w, s#4, t#4
-(sy)mov.f32f32 r5.y, r5.y
-mov.f32f32 r7.z, r5.x
-mov.f32f32 r1.z, r1.z
-add.f r5.x, c19.y, (neg)r0.z
-(ss)add.f r5.w, c19.y, (neg)r0.z
-add.f r6.x, c19.y, (neg)r0.z
-mul.f r1.z, r1.z, c15.x
-mov.f32f32 r5.y, r5.y
-add.f r6.y, c16.x, (neg)r2.x
-mul.f r5.x, r5.x, c10.z
-mov.f32f32 r1.z, r1.z
-mul.f r5.w, r5.w, c10.y
-mul.f r6.x, r6.x, c10.x
-mov.f32f32 r6.y, r6.y
-add.f r7.x, c16.x, (neg)r0.w
-mov.f32f32 r7.w, r8.x
-mul.f r2.y, r2.y, r4.z
-rcp r1.z, r1.z
-(ss)mov.f32f32 r1.z, r1.z
-mul.f r2.w, r2.w, r4.z
-mov.f32f32 r4.z, r7.x
-mov.f32f32 r2.y, r2.y
-mov.f32f32 r7.w, r7.w
-mov.f32f32 r2.w, r2.w
-mul.f r7.x, r6.y, r4.z
-mov.f32f32 r2.y, r2.y
-mul.f r1.w, r1.w, r1.w
-absneg.f r2.w, (neg)r2.w
-mad.f32 r1.w, r3.w, r3.w, r1.w
-mul.f r3.w, r7.x, r5.y
-mov.f32f32 r5.y, r8.x
-sam.s (f32)(x)r7.x, r7.y, s#4, t#4
-(sy)mov.f32f32 r7.x, r7.x
-mov.f32f32 r1.w, r1.w
-mov.f32f32 r0.y, r0.y
-mad.f32 r1.w, r4.x, r4.x, r1.w
-mov.f32f32 r8.w, r5.y
-mov.f32f32 r4.x, r2.z
-add.f r2.x, r2.x, c17.x
+(ss)mul.f r1.z, r3.y, r1.z
+mul.f r1.w, r1.w, r5.z
+mov.f32f32 r7.w, r8.w
+add.f r2.x, c19.y, (neg)r0.z
+add.f r3.y, c19.y, (neg)r0.z
+mul.f r1.w, r1.w, c15.x
+add.f r6.y, c19.y, (neg)r0.z
+mov.f32f32 r9.z, r8.w
+mov.f32f32 r7.x, r8.w
+nop
+sam.s (f32)(x)r9.w, r7.y, s#4, t#4
+(ss)mul.f r7.y, r2.x, c10.z
+mov.f32f32 r2.x, r4.w
+rcp r1.w, r1.w
+(ss)mov.f32f32 r7.z, r1.w
+mul.f r4.y, r4.y, r5.z
+mad.f32 r0.x, r2.y, r2.y, r0.x
+add.f r2.y, c16.x, (neg)r5.x
+mad.f32 r0.x, r3.z, r3.z, r0.x
+mul.f r3.y, r3.y, c10.y
+mul.f r3.z, r6.y, c10.x
+mov.f32f32 r5.x, r2.y
+absneg.f r1.z, (neg)r1.z
+sam.s (f32)(x)r10.x, r9.x, s#4, t#4
+sam.s (f32)(x)r6.y, r6.z, s#4, t#4
+sam.s (f32)(x)r8.y, r8.y, s#4, t#4
+add.f r2.w, r2.w, c17.x
 add.f r0.w, r0.w, c17.x
-bary.f r5.y, 16, r1.x
-(ss)mov.f32f32 r7.y, r4.x
-sqrt r1.w, r1.w
-(ss)mov.f32f32 r1.w, r1.w
-sam.s (f32)(x)r4.x, r8.y, s#4, t#4
-(sy)mov.f32f32 r4.x, r4.x
-mul.f r4.z, r2.x, r4.z
-mov.f32f32 r9.y, r0.y
-add.f r0.y, c14.z, (neg)r1.w
-mov.f32f32 r1.w, r8.x
-bary.f r7.w, 19, r1.x
-mul.f r6.y, r6.y, r0.w
-mov.f32f32 r0.y, r0.y
-mad.f32 r3.w, r4.z, r4.x, r3.w
-mov.f32f32 r9.z, r1.w
-mov.f32f32 r1.w, r7.w
-mul.f r0.y, c12.z, r0.y
-mov.f32f32 r3.w, r3.w
-mov.f32f32 r4.x, r5.y
-mov.f32f32 r7.z, r1.w
-mov.f32f32 r0.y, r0.y
-sam.s (f32)(x)r1.w, r9.x, s#4, t#4
-(sy)mov.f32f32 r1.w, r1.w
-mov.f32f32 r8.x, r4.x
-bary.f r4.x, 17, r1.x
-mul.f r0.y, r0.y, c14.w
-mad.f32 r1.w, r6.y, r1.w, r3.w
-(ss)nop
-sam (f32)(w)r8.y, r7.y, s#2, t#2
-(sy)cmps.f.lt r3.w, r9.x, c15.y
-mov.f32f32 r4.x, r4.x
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r1.w, r1.w
-mul.f r0.w, r2.x, r0.w
-cov.u32f32 r2.x, r3.w
-mul.f r2.w, r2.w, r0.y
-mul.f r0.y, r2.y, r0.y
-mad.f32 r0.w, r0.w, r7.x, r1.w
-cmps.f.ne r1.w, r2.x, c14.y
-mov.f32f32 r2.x, r2.w
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r2.y, c15.z
-mul.f r2.x, r2.x, r1.z
-mul.f r0.y, r0.y, r1.z
+sqrt r0.x, r0.x
+(ss)add.f r0.x, c14.z, (neg)r0.x
+mul.f r2.x, r2.x, r5.x
+mul.f r2.y, r2.w, r2.y
+mul.f r4.w, r4.w, r0.w
+mul.f r0.x, c12.z, r0.x
+(sy)mul.f r2.x, r2.x, r9.w
+bary.f r6.z, 18, r1.x
+mul.f r0.w, r2.w, r0.w
+mul.f r0.x, r0.x, c14.w
+mad.f32 r2.x, r2.y, r10.x, r2.x
+bary.f r6.w, 19, r1.x
+bary.f r8.z, 16, r1.x
+mov.f32f32 r2.y, r0.x
+mad.f32 r2.x, r4.w, r6.y, r2.x
+mul.f r0.x, r1.z, r0.x
+mad.f32 r0.w, r0.w, r8.y, r2.x
+mul.f r1.z, r4.y, r2.y
+sam (f32)(w)r9.x, r6.z, s#2, t#2
+bary.f r8.w, 17, r1.x
+mul.f r0.x, r0.x, r1.w
+(sy)cmps.f.lt r1.w, r9.w, c15.y
+mul.f r1.z, r1.z, r7.z
 mul.f r0.w, c17.z, r0.w
-mov.f32f32 r1.z, c14.y
-mov.f32f32 r2.x, r2.x
+mov.f32f32 r2.x, r0.x
+cov.u32f32 r1.w, r1.w
+mov.f32f32 r2.y, r1.z
+cmps.f.lt r2.w, r9.w, c15.y
+mov.f32f32 r4.y, r0.w
+cmps.f.ne r1.w, r1.w, c14.y
+mov.f32f32 r4.w, c14.y
+mov.f32f32 r5.x, c14.y
+cov.u32f32 r2.w, r2.w
+sam (f32)(w)r8.y, r8.z, s#1, t#1
+(sy)cmps.f.lt r5.z, r9.x, c18.x
+sel.b32 r1.z, r1.z, r1.w, r4.w
+sel.b32 r0.x, r0.x, r1.w, r5.x
+(rpt1)nop
+add.f r1.w, r6.w, r1.z
+add.f r1.z, r6.z, r0.x
+cmps.f.ne r0.x, r2.w, c14.y
+cov.u32f32 r2.w, r5.z
+mov.f32f32 r4.w, c15.z
+mov.f32f32 r5.x, c14.y
 mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.w, r0.w
-sel.b32 r1.z, r2.y, r1.w, r1.z
-mov.f32f32 r1.w, r2.x
-mov.f32f32 r2.y, r0.y
-mov.f32f32 r2.w, r0.y
-mov.f32f32 r3.w, r0.y
-mov.f32f32 r1.w, r1.w
-cmps.f.lt r4.z, r9.x, c15.y
-mov.f32f32 r2.y, r2.y
-mov.f32f32 r2.w, r2.w
+mov.f32f32 r5.z, r2.z
+sam (f32)(w)r6.y, r1.z, s#2, t#2
+cmps.f.ne r2.z, r2.w, c14.y
+sel.b32 r0.x, r4.w, r0.x, r5.x
+mov.f32f32 r2.w, c14.y
 mov.f32f32 r3.w, r3.w
-cov.u32f32 r4.z, r4.z
-mov.f32f32 r5.y, r2.x
-mov.f32f32 r6.y, r0.y
-mov.f32f32 r7.x, r2.x
-cmps.f.ne r4.z, r4.z, c14.y
-(ss)mov.f32f32 r7.y, c14.y
-mov.f32f32 r7.z, c14.y
-mov.f32f32 r5.y, r5.y
-mov.f32f32 r6.y, r6.y
-sel.b32 r1.w, r1.w, r4.z, r7.y
-sel.b32 r2.y, r2.y, r4.z, r7.z
-mov.f32f32 r4.z, r7.x
-mov.f32f32 r7.x, r0.y
-add.f r1.w, r2.z, r1.w
-add.f r2.y, r7.w, r2.y
-mov.f32f32 r2.z, r2.x
-mov.f32f32 r7.y, r0.y
-mov.f32f32 r7.z, r1.w
-mov.f32f32 r7.w, r2.y
-mov.f32f32 r7.x, r7.x
-mov.f32f32 r2.z, r2.z
-mov.f32f32 r8.y, r7.z
-mov.f32f32 r8.z, r7.w
-mov.f32f32 r7.y, r7.y
-mov.f32f32 r7.z, r2.x
-mov.f32f32 r7.w, r0.y
-mov.f32f32 r8.w, r2.x
-mov.f32f32 r9.x, r0.y
-mov.f32f32 r9.y, r2.x
-sam (f32)(w)r9.z, r8.y, s#2, t#2
-add.f r1.z, c15.y, (neg)r1.z
-mov.f32f32 r7.z, r7.z
-mov.f32f32 r7.w, r7.w
-(ss)mov.f32f32 r8.z, r8.w
-(sy)cmps.f.lt r8.y, r10.y, r1.z
-cmps.f.lt r8.w, r10.y, r1.z
-mov.f32f32 r9.x, r9.x
-mov.f32f32 r9.y, r9.y
+mov.f32f32 r4.w, (0.000000)
+add.f r0.x, c15.y, (neg)r0.x
+sel.b32 r5.x, r2.w, r2.z, r9.x
+bary.f r6.y, 2, r1.x
+(ss)bary.f r6.z, 1, r1.x
+(sy)cmps.f.lt r2.z, r7.x, r0.x
+cmps.f.lt r2.w, r7.x, r0.x
+bary.f r6.w, 23, r1.x
+bary.f r7.x, 24, r1.x
+cov.u32f32 r2.z, r2.z
+cov.u32f32 r2.w, r2.w
+bary.f r7.z, 25, r1.x
+bary.f (ei)r1.x, 0, r1.x
+cmps.f.ne r1.y, r2.z, c14.y
+mov.f32f32 r2.z, c14.y
+cmps.f.ne r2.w, r2.w, c14.y
+mov.f32f32 r7.w, c14.y
+mov.f32f32 r8.y, c15.z
+sel.b32 r2.z, r2.x, r1.y, r2.z
+mov.f32f32 r8.z, c14.y
+sel.b32 r1.y, r2.y, r1.y, r7.w
+mov.f32f32 r7.w, c14.y
+add.f r1.z, r1.z, r2.z
+sel.b32 r2.z, r8.y, r2.w, r8.z
+add.f r8.z, r1.w, r1.y
+mov.f32f32 r1.y, c14.y
+mov.f32f32 r8.y, r1.z
+add.f r0.x, r0.x, (neg)r2.z
+mov.f32f32 r1.w, r8.z
+mov.f32f32 r2.z, c14.y
+mov.f32f32 r2.w, c15.z
+mov.f32f32 r8.w, c14.y
+mov.f32f32 r9.x, c14.y
+sam (f32)(w)r9.y, r8.y, s#2, t#2
+(sy)(ss)cmps.f.lt r8.y, r10.x, r0.x
+mov.f32f32 r0.x, r0.x
+mov.f32f32 r8.z, c14.y
+mov.f32f32 r9.y, c15.z
 cov.u32f32 r8.y, r8.y
-cov.u32f32 r8.w, r8.w
-mov.f32f32 r9.z, r2.x
-nop
-mov.f32f32 r8.y, r8.y
-cmps.f.ne r8.w, r8.w, c14.y
-mov.f32f32 r9.w, c15.z
+cmps.f.lt r9.z, r10.x, r0.x
+mov.f32f32 r9.w, c14.y
 mov.f32f32 r10.x, c14.y
-cmps.f.ne r10.y, r8.y, c14.y
-mov.f32f32 r9.z, r9.z
-mov.f32f32 r10.z, c14.y
-mov.f32f32 r10.w, c14.y
-sel.b32 r8.w, r9.w, r8.w, r10.x
-mov.f32f32 r8.y, r4.x
-sel.b32 r4.x, r9.z, r10.y, r10.z
-sel.b32 r9.x, r9.x, r10.y, r10.w
-add.f r1.z, r1.z, (neg)r8.w
-bary.f r8.w, 23, r1.x
-add.f r1.w, r1.w, r4.x
-add.f r2.y, r2.y, r9.x
-mov.f32f32 r1.z, r1.z
-sam (f32)(w)r9.z, r8.x, s#1, t#1
-(sy)cmps.f.lt r4.x, r10.y, c18.x
-mov.f32f32 r1.w, r1.w
-mov.f32f32 r2.y, r2.y
-(ss)mov.f32f32 r8.x, r10.y
-cov.u32f32 r4.x, r4.x
-mov.f32f32 r8.y, r1.w
-mov.f32f32 r9.x, r2.y
-mov.f32f32 r8.x, r8.x
-mov.f32f32 r8.w, r8.w
-mov.f32f32 r9.z, r8.y
-mov.f32f32 r9.w, r9.x
-cmps.f.ne r4.x, r4.x, c14.y
-(rpt1)nop
-mov.f32f32 r8.y, c14.y
-bary.f r9.x, 24, r1.x
-bary.f r10.x, 25, r1.x
-sam (f32)(w)r10.y, r9.z, s#2, t#2
-(sy)(ss)cmps.f.lt r9.z, r11.x, r1.z
-cmps.f.lt r9.w, r11.x, r1.z
-sel.b32 r4.x, r8.y, r4.x, r8.x
-mov.f32f32 r8.x, r9.x
-cov.u32f32 r8.y, r9.z
-cov.u32f32 r9.x, r9.w
-mov.f32f32 r9.z, r10.x
-mov.f32f32 r9.w, (0.000000)
-mov.f32f32 r8.y, r8.y
-cmps.f.ne r9.x, r9.x, c14.y
-mov.f32f32 r10.x, c15.z
-mov.f32f32 r10.y, c14.y
 cmps.f.ne r8.y, r8.y, c14.y
+mov.f32f32 r10.y, c14.y
+cov.u32f32 r9.z, r9.z
 mov.f32f32 r10.z, c14.y
 mov.f32f32 r10.w, c14.y
-sel.b32 r9.x, r10.x, r9.x, r10.y
-bary.f r10.x, 2, r1.x
-sel.b32 r9.y, r9.y, r8.y, r10.z
-sel.b32 r7.w, r7.w, r8.y, r10.w
-add.f r1.z, r1.z, (neg)r9.x
-bary.f r8.y, 1, r1.x
-add.f r1.w, r1.w, r9.y
-add.f r2.y, r2.y, r7.w
-mov.f32f32 r1.z, r1.z
-bary.f (ei)r1.x, 0, r1.x
-mov.f32f32 r1.y, r1.w
-mov.f32f32 r1.w, r2.y
-mov.f32f32 r2.y, c14.y
-mov.f32f32 r7.w, c14.y
-mov.f32f32 r9.x, r1.y
-mov.f32f32 r9.y, r1.w
-mov.f32f32 r10.y, c14.y
+sel.b32 r10.y, r2.x, r8.y, r10.y
+cmps.f.ne r9.z, r9.z, c14.y
+sel.b32 r8.y, r2.y, r8.y, r10.z
 mov.f32f32 r10.z, c15.z
-mov.f32f32 r10.w, r9.x
-mov.f32f32 r11.x, r9.y
-mov.f32f32 r9.x, c14.y
-mov.f32f32 r9.y, c14.y
-mov.f32f32 r11.y, c14.y
-mov.f32f32 r11.z, c15.z
-mov.f32f32 r11.w, c14.y
-mov.f32f32 r12.x, c14.y
-sam (f32)(w)r12.y, r10.w, s#2, t#2
-(sy)(ss)cmps.f.lt r10.w, r13.x, r1.z
-cmps.f.lt r11.x, r13.x, r1.z
-mov.f32f32 r12.y, c14.y
-mov.f32f32 r12.z, c15.z
-cov.u32f32 r10.w, r10.w
-cov.u32f32 r11.x, r11.x
-mov.f32f32 r12.w, c14.y
-mov.f32f32 r13.x, c14.y
-mov.f32f32 r10.w, r10.w
-cmps.f.ne r11.x, r11.x, c14.y
-mov.f32f32 r13.y, c15.z
-mov.f32f32 r13.z, c14.y
-cmps.f.ne r10.w, r10.w, c14.y
-mov.f32f32 r13.w, c14.y
-mov.f32f32 r14.x, c14.y
-sel.b32 r11.x, r13.y, r11.x, r13.z
-nop
-sel.b32 r8.z, r8.z, r10.w, r13.w
-sel.b32 r7.y, r7.y, r10.w, r14.x
-add.f r1.z, r1.z, (neg)r11.x
-nop
-add.f r1.y, r1.y, r8.z
-add.f r1.w, r1.w, r7.y
-mov.f32f32 r1.z, r1.z
+add.f r1.z, r1.z, r10.y
+mov.f32f32 r10.y, c14.y
+add.f r11.y, r1.w, r8.y
+mov.f32f32 r1.w, c15.z
+mov.f32f32 r11.x, r1.z
+sel.b32 r8.y, r10.z, r9.z, r10.y
+mov.f32f32 r9.z, r11.y
+mov.f32f32 r10.y, c14.y
+mov.f32f32 r10.z, c14.y
+mov.f32f32 r11.z, c14.y
+mov.f32f32 r11.w, c15.z
+sam (f32)(w)r12.x, r11.x, s#2, t#2
+add.f r0.x, r0.x, (neg)r8.y
+mov.f32f32 r8.y, c14.y
+(ss)mov.f32f32 r11.x, c14.y
 nop
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r1.w, r1.w
-(rpt1)nop
-mov.f32f32 r7.y, r1.y
-mov.f32f32 r8.z, r1.w
-(rpt1)nop
-mov.f32f32 r10.w, r7.y
-mov.f32f32 r11.x, r8.z
-(rpt5)nop
-sam (f32)(w)r13.y, r10.w, s#2, t#2
-(sy)cmps.f.lt r7.y, r14.x, r1.z
-cmps.f.lt r8.z, r14.x, r1.z
+(sy)cmps.f.lt r11.y, r12.w, r0.x
+mov.f32f32 r0.x, r0.x
 (rpt1)nop
-cov.u32f32 r7.y, r7.y
-cov.u32f32 r8.z, r8.z
+cov.u32f32 r11.y, r11.y
+cmps.f.lt r12.x, r12.w, r0.x
 (rpt1)nop
-mov.f32f32 r7.y, r7.y
-cmps.f.ne r8.z, r8.z, c14.y
+cmps.f.ne r11.y, r11.y, c14.y
+cov.u32f32 r12.x, r12.x
 (rpt1)nop
-cmps.f.ne r7.y, r7.y, c14.y
-sel.b32 r8.z, r12.z, r8.z, r12.y
-(rpt1)nop
-sel.b32 r7.z, r7.z, r7.y, r13.x
-sel.b32 r7.x, r7.x, r7.y, r12.w
-add.f r1.z, r1.z, (neg)r8.z
+sel.b32 r11.x, r2.x, r11.y, r11.x
+cmps.f.ne r12.x, r12.x, c14.y
+sel.b32 r8.y, r2.y, r11.y, r8.y
 nop
-add.f r1.y, r1.y, r7.z
-add.f r1.w, r1.w, r7.x
-mov.f32f32 r1.z, r1.z
+add.f r1.z, r1.z, r11.x
+sel.b32 r11.x, r11.w, r12.x, r11.z
+add.f r11.z, r9.z, r8.y
 nop
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r1.w, r1.w
-(rpt1)nop
-mov.f32f32 r7.x, r1.y
-mov.f32f32 r7.y, r1.w
-(rpt1)nop
-(ss)mov.f32f32 r10.w, r7.x
-mov.f32f32 r11.x, r7.y
-(rpt5)nop
-sam (f32)(w)r12.y, r10.w, s#2, t#2
-(sy)cmps.f.lt r7.x, r13.x, r1.z
-cmps.f.lt r7.y, r13.x, r1.z
-(rpt1)nop
-cov.u32f32 r7.x, r7.x
-cov.u32f32 r7.y, r7.y
+mov.f32f32 r11.y, r1.z
+add.f r0.x, r0.x, (neg)r11.x
+mov.f32f32 r8.y, r11.z
+(rpt3)nop
+sam (f32)(w)r11.x, r11.y, s#2, t#2
+(sy)cmps.f.lt r9.z, r11.w, r0.x
+mov.f32f32 r0.x, r0.x
 (rpt1)nop
-mov.f32f32 r7.x, r7.x
-cmps.f.ne r7.y, r7.y, c14.y
+cov.u32f32 r9.z, r9.z
+cmps.f.lt r11.x, r11.w, r0.x
 (rpt1)nop
-cmps.f.ne r7.x, r7.x, c14.y
-sel.b32 r7.y, r11.z, r7.y, r11.y
+cmps.f.ne r9.z, r9.z, c14.y
+cov.u32f32 r11.x, r11.x
 (rpt1)nop
-sel.b32 r2.z, r2.z, r7.x, r12.x
-sel.b32 r6.y, r6.y, r7.x, r11.w
-add.f r1.z, r1.z, (neg)r7.y
+sel.b32 r10.z, r2.x, r9.z, r10.z
+cmps.f.ne r11.x, r11.x, c14.y
+sel.b32 r9.z, r2.y, r9.z, r10.y
 nop
-add.f r1.y, r1.y, r2.z
-add.f r1.w, r1.w, r6.y
-mov.f32f32 r1.z, r1.z
+add.f r1.z, r1.z, r10.z
+sel.b32 r1.w, r1.w, r11.x, r10.w
+add.f r10.z, r8.y, r9.z
 nop
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r1.w, r1.w
+mov.f32f32 r10.y, r1.z
+add.f r0.x, r0.x, (neg)r1.w
+mov.f32f32 r1.w, r10.z
+(rpt3)nop
+sam (f32)(w)r10.y, r10.y, s#2, t#2
+(sy)cmps.f.lt r8.y, r11.x, r0.x
+mov.f32f32 r0.x, r0.x
 (rpt1)nop
-mov.f32f32 r2.z, r1.y
-mov.f32f32 r6.y, r1.w
+cov.u32f32 r8.y, r8.y
+cmps.f.lt r9.z, r11.x, r0.x
 (rpt1)nop
-mov.f32f32 r7.x, r2.z
-mov.f32f32 r7.y, r6.y
-(rpt5)nop
-sam (f32)(w)r10.w, r7.x, s#2, t#2
-(sy)cmps.f.lt r2.z, r11.z, r1.z
-cmps.f.lt r6.y, r11.z, r1.z
+cmps.f.ne r8.y, r8.y, c14.y
+cov.u32f32 r9.z, r9.z
 (rpt1)nop
-cov.u32f32 r2.z, r2.z
-cov.u32f32 r6.y, r6.y
+sel.b32 r10.x, r2.x, r8.y, r10.x
+cmps.f.ne r9.z, r9.z, c14.y
+sel.b32 r8.y, r2.y, r8.y, r9.w
+nop
+add.f r1.z, r1.z, r10.x
+sel.b32 r8.z, r9.y, r9.z, r8.z
+add.f r9.z, r1.w, r8.y
+nop
+mov.f32f32 r9.y, r1.z
+add.f r0.x, r0.x, (neg)r8.z
+mov.f32f32 r1.w, r9.z
+(rpt3)nop
+sam (f32)(w)r9.y, r9.y, s#2, t#2
+(sy)cmps.f.lt r8.y, r10.x, r0.x
+mov.f32f32 r0.x, r0.x
 (rpt1)nop
-mov.f32f32 r2.z, r2.z
-cmps.f.ne r6.y, r6.y, c14.y
+cov.u32f32 r8.y, r8.y
+cmps.f.lt r8.z, r10.x, r0.x
 (rpt1)nop
-cmps.f.ne r2.z, r2.z, c14.y
-sel.b32 r6.y, r10.z, r6.y, r10.y
+cmps.f.ne r8.y, r8.y, c14.y
+cov.u32f32 r8.z, r8.z
 (rpt1)nop
-sel.b32 r4.z, r4.z, r2.z, r9.y
-sel.b32 r2.z, r3.w, r2.z, r9.x
-add.f r1.z, r1.z, (neg)r6.y
+sel.b32 r9.x, r2.x, r8.y, r9.x
+cmps.f.ne r8.z, r8.z, c14.y
+sel.b32 r8.y, r2.y, r8.y, r8.w
 nop
-add.f r1.y, r1.y, r4.z
-add.f r1.w, r1.w, r2.z
-mov.f32f32 r1.z, r1.z
+add.f r1.z, r1.z, r9.x
+sel.b32 r2.z, r2.w, r8.z, r2.z
+add.f r8.z, r1.w, r8.y
 nop
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r1.w, r1.w
-(rpt1)nop
-mov.f32f32 r2.z, r1.y
-mov.f32f32 r3.w, r1.w
+mov.f32f32 r8.y, r1.z
+add.f r0.x, r0.x, (neg)r2.z
+mov.f32f32 r1.w, r8.z
+(rpt3)nop
+sam (f32)(w)r8.y, r8.y, s#2, t#2
+(sy)cmps.f.lt r2.z, r9.x, r0.x
+mov.f32f32 r0.x, r0.x
 (rpt1)nop
-(ss)mov.f32f32 r7.x, r2.z
-mov.f32f32 r7.y, r3.w
-(rpt5)nop
-sam (f32)(w)r10.y, r7.x, s#2, t#2
-(sy)cmps.f.lt r2.z, r11.x, r1.z
-(rpt2)nop
 cov.u32f32 r2.z, r2.z
 (rpt2)nop
 cmps.f.ne r2.z, r2.z, c14.y
 (rpt2)nop
-sel.b32 r3.w, r5.y, r2.z, r7.w
-sel.b32 r2.y, r2.w, r2.z, r2.y
+sel.b32 r1.y, r2.x, r2.z, r1.y
+sel.b32 r2.z, r2.y, r2.z, r7.w
 (rpt1)nop
-add.f r1.y, r1.y, r3.w
-add.f r1.w, r1.w, r2.y
+add.f r1.y, r1.z, r1.y
+add.f r1.w, r1.w, r2.z
 (rpt1)nop
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r1.w, r1.w
+mov.f32f32 r1.z, r1.y
+mov.f32f32 r2.z, r1.w
 (rpt1)nop
-add.f r2.y, r1.y, (neg)r2.x
-mov.f32f32 r2.z, r1.y
-add.f r2.w, r1.w, (neg)r0.y
-mov.f32f32 r3.w, r1.w
-mov.f32f32 r2.y, r2.y
-(ss)mov.f32f32 r7.x, r2.z
-mov.f32f32 r2.z, r2.w
-mov.f32f32 r7.y, r3.w
-mov.f32f32 r2.y, r2.y
-(rpt2)nop
-mov.f32f32 r7.z, r2.y
-mov.f32f32 r2.y, r2.z
-sam (f32)(w)r10.y, r7.x, s#2, t#2
-(sy)add.f r2.z, c15.z, r11.x
-add.f r1.z, r11.x, (neg)r1.z
-nop
-mov.f32f32 r7.w, r2.y
-mov.f32f32 r2.y, r2.z
-mov.f32f32 r1.z, r1.z
-(rpt3)nop
-sam (f32)(w)r7.x, r7.z, s#2, t#2
-(sy)add.f r2.y, r2.y, (neg)r7.w
-(rpt2)nop
-mov.f32f32 r2.y, r2.y
+(ss)add.f r8.y, r1.z, (neg)r2.x
+add.f r8.z, r2.z, (neg)r2.y
+(rpt1)nop
+sam (f32)(w)r8.w, r1.z, s#2, t#2
+(sy)(ss)add.f r1.z, c15.z, r9.z
+add.f r0.x, r9.z, (neg)r0.x
+(rpt1)nop
+sam (f32)(w)r8.y, r8.y, s#2, t#2
+(sy)add.f r1.z, r1.z, (neg)r9.x
 (rpt5)nop
-rcp r2.y, r2.y
-(ss)mov.f32f32 r2.y, r2.y
-(rpt2)nop
-mul.f r1.z, r1.z, r2.y
-(rpt2)nop
-mov.f32f32 r1.z, r1.z
+rcp r1.z, r1.z
+(ss)mul.f r0.x, r0.x, r1.z
 (rpt2)nop
-mul.f r2.x, r2.x, r1.z
-mul.f r0.y, r0.y, r1.z
-(rpt1)nop
-mov.f32f32 r1.z, r2.x
-mov.f32f32 r0.y, r0.y
+(ss)mov.f32f32 r1.z, r0.x
+mul.f r0.x, r2.x, r0.x
 (rpt1)nop
-add.f r1.y, r1.y, (neg)r1.z
-add.f r0.y, r1.w, (neg)r0.y
+mul.f r1.z, r2.y, r1.z
+add.f r1.w, r1.y, (neg)r0.x
 (rpt1)nop
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r0.y, r0.y
+add.f r2.x, r2.z, (neg)r1.z
+mov.f32f32 r1.y, r1.w
 (rpt1)nop
-mov.f32f32 r1.z, r1.y
-mov.f32f32 r1.w, r1.y
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r2.x, r0.y
-mov.f32f32 r2.y, r1.z
-mov.f32f32 r1.z, r0.y
-mov.f32f32 r7.x, r1.w
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r2.z, r1.z
 mov.f32f32 r1.z, r2.x
-mov.f32f32 r7.y, r0.y
-nop
-mov.f32f32 r1.w, r1.y
-mov.f32f32 r2.x, r1.z
+mov.f32f32 r8.y, r1.y
 nop
-sam (f32)(xyz)r10.y, r2.y, s#2, t#2
-(sy)mad.f32 r0.y, c16.x, r10.z, c16.y
-mad.f32 r1.y, c16.x, r10.y, c16.y
-sam (f32)(xyzw)r7.x, r7.x, s#0, t#0
-(sy)cmps.f.lt r1.z, r7.w, c15.w
-(ss)mov.f32f32 r2.y, r7.w
-mov.f32f32 r0.y, r0.y
-sam (f32)(xyz)r11.x, r1.w, s#3, t#3
-(sy)(ss)mul.f r1.w, c8.y, r11.y
-mul.f r2.x, c8.x, r11.x
-cov.u32f32 r1.z, r1.z
-absneg.f r0.y, (neg)r0.y
-mov.f32f32 r1.y, r1.y
-mul.f r2.z, c8.z, r11.z
-mov.f32f32 r1.w, r1.w
-mov.f32f32 r2.x, r2.x
-mul.f r2.w, r4.y, r1.y
-mul.f r3.z, r3.z, r1.y
-mad.f32 r0.x, r0.x, r0.y, r2.w
-mad.f32 r2.w, r3.x, r0.y, r3.z
-mul.f r1.y, r4.w, r1.y
-cmps.f.ne p0.x, r1.z, r9.w
-mov.f32f32 r0.x, r0.x
-mad.f32 r1.z, c16.x, r10.w, c16.y
-mov.f32f32 r2.w, r2.w
-mad.f32 r0.y, r3.y, r0.y, r1.y
-mov.f32f32 r1.y, r2.z
-mov.f32f32 r1.z, r1.z
+sam (f32)(xyzw)r2.x, r1.w, s#0, t#0
+(sy)cmps.f.lt r0.x, r2.w, c15.w
+mov.f32f32 r8.z, r1.z
+(rpt1)nop
+sam (f32)(xyz)r8.w, r1.y, s#2, t#2
+(sy)(ss)mad.f32 r1.y, c16.x, r8.w, c16.y
+cov.u32f32 r0.x, r0.x
+mad.f32 r1.z, c16.x, r9.x, c16.y
+mad.f32 r1.w, c16.x, r9.y, c16.y
+mov.f32f32 r7.w, r1.y
+cmps.f.ne p0.x, r0.x, r4.w
+absneg.f r0.x, (neg)r1.z
+mul.f r1.y, r4.x, r1.y
+mul.f r1.z, r4.z, r7.w
+mul.f r4.x, r5.y, r7.w
+mad.f32 r0.y, r0.y, r0.x, r1.z
+mov.f32f32 r1.z, r1.w
+mad.f32 r4.x, r5.z, r0.x, r4.x
+mad.f32 r0.x, r3.w, r0.x, r1.y
 kill p0.x
-mov.f32f32 r3.w, r2.y
-mov.f32f32 r0.y, r0.y
-mad.f32 r0.x, r5.z, r1.z, r0.x
-mad.f32 r2.y, r6.w, r1.z, r2.w
-mad.f32 r0.y, r6.z, r1.z, r0.y
+mad.f32 r0.y, r6.x, r1.z, r0.y
+mad.f32 r1.y, r5.w, r1.z, r4.x
+mad.f32 r0.x, r8.x, r1.w, r0.x
 nop
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r1.z, r2.y
-mov.f32f32 r0.y, r0.y
-nop
-mul.f r2.y, r0.x, r0.x
-nop
-mad.f32 r2.y, r1.z, r1.z, r2.y
-(rpt2)nop
-mov.f32f32 r2.y, r2.y
-nop
-mad.f32 r2.y, r0.y, r0.y, r2.y
+mov.f32f32 r1.z, r0.y
+mov.f32f32 r1.w, r1.y
+mov.f32f32 r3.w, r0.x
+sam (f32)(xyz)r5.y, r8.y, s#3, t#3
+(sy)mul.f r4.x, c8.z, r5.w
+mul.f r0.y, r0.y, r1.z
+mul.f r4.z, c8.y, r5.z
+mad.f32 r0.y, r1.y, r1.w, r0.y
+mul.f r1.y, c8.x, r5.y
+mad.f32 r0.y, r3.w, r3.w, r0.y
 (rpt5)nop
-rsq r2.y, r2.y
-(ss)mov.f32f32 r2.y, r2.y
-(rpt2)nop
-mul.f r0.x, r0.x, r2.y
-mul.f r1.z, r1.z, r2.y
-mul.f r0.y, r0.y, r2.y
-nop
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r0.y, r0.y
-nop
-mul.f r2.y, r0.x, r0.x
-mul.f r2.w, (neg)c9.x, r0.x
-mad.f32 r2.y, r1.z, r1.z, r2.y
-mad.f32 r2.w, (neg)c9.y, r1.z, r2.w
-(rpt1)nop
-mov.f32f32 r2.y, r2.y
-mov.f32f32 r2.w, r2.w
-mad.f32 r2.y, r0.y, r0.y, r2.y
-mad.f32 r2.w, (neg)c9.z, r0.y, r2.w
-(rpt4)nop
-rsq r2.y, r2.y
-(ss)mov.f32f32 r2.y, r2.y
-max.f r2.w, r2.w, c14.y
+rsq r0.y, r0.y
+(ss)mov.f32f32 r3.w, r0.y
+mul.f r0.x, r0.x, r0.y
 (rpt1)nop
-mul.f r0.x, r0.x, r2.y
-mov.f32f32 r2.w, r2.w
-mul.f r1.z, r1.z, r2.y
-mul.f r0.y, r0.y, r2.y
-mov.f32f32 r0.x, r0.x
-mad.f32 r2.y, c8.z, r2.w, (neg)r10.x
-mad.f32 r4.y, c8.y, r2.w, (neg)r8.y
-mad.f32 r2.w, c8.x, r2.w, (neg)r1.x
-mul.f r0.x, r0.x, r8.w
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r2.y, r2.y
-mov.f32f32 r4.y, r4.y
-mov.f32f32 r2.w, r2.w
-mad.f32 r0.x, r1.z, r8.x, r0.x
-mad.f32 r1.z, c12.x, r2.y, r10.x
-mad.f32 r2.y, c12.x, r4.y, r8.y
-mad.f32 r1.x, c12.x, r2.w, r1.x
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r2.y, r2.y
-mov.f32f32 r1.x, r1.x
-mad.f32 r0.x, r0.y, r9.z, r0.x
-(rpt2)nop
-mov.f32f32 r0.x, r0.x
-(rpt2)nop
-max.f r0.x, c14.y, r0.x
-(rpt2)nop
-mov.f32f32 r0.x, r0.x
+(ss)mul.f r0.y, r1.z, r3.w
+mul.f r1.z, r1.w, r3.w
+mov.f32f32 r1.w, r0.x
+nop
+mov.f32f32 r3.w, r0.y
+mul.f r0.y, (neg)c9.x, r0.y
+mov.f32f32 r4.w, r1.z
+nop
+mul.f r5.y, r3.w, r3.w
+mad.f32 r0.y, (neg)c9.y, r1.z, r0.y
+mad.f32 r1.z, r4.w, r4.w, r5.y
+mad.f32 r0.x, (neg)c9.z, r0.x, r0.y
+mad.f32 r0.y, r1.w, r1.w, r1.z
 (rpt5)nop
-log2 r0.x, r0.x
-(ss)mul.f r0.x, c12.y, r0.x
-(rpt2)nop
-mov.f32f32 r0.x, r0.x
+rsq r0.y, r0.y
+(ss)mov.f32f32 r1.z, r0.y
+max.f r0.x, r0.x, c14.y
+(ss)mul.f r0.y, r1.w, r0.y
+nop
+mul.f r1.w, r3.w, r1.z
+mov.f32f32 r3.w, r0.x
+mul.f r1.z, r4.w, r1.z
+mad.f32 r0.x, c8.x, r0.x, (neg)r1.x
+mul.f r1.w, r1.w, r6.w
+mad.f32 r4.w, c8.z, r3.w, (neg)r6.y
+mad.f32 r1.z, r1.z, r7.x, r1.w
+mad.f32 r1.w, c8.y, r3.w, (neg)r6.z
+mad.f32 r0.y, r0.y, r7.z, r1.z
+mad.f32 r1.z, c12.x, r4.w, r6.y
+mad.f32 r0.x, c12.x, r0.x, r1.x
+nop
+max.f r0.y, c14.y, r0.y
+mad.f32 r1.x, c12.x, r1.w, r6.z
+(rpt4)nop
+log2 r0.y, r0.y
+(ss)mul.f r0.y, c12.y, r0.y
 (rpt5)nop
-exp2 r0.x, r0.x
-(ss)mul.f r0.y, r1.y, r0.x
-mul.f r1.y, r1.w, r0.x
-mad.f32 r0.y, r7.z, r1.z, r0.y
-mad.f32 r1.y, r7.y, r2.y, r1.y
-(ss)mul.f r0.x, r2.x, r0.x
-nop
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r1.y, r1.y
-mad.f32 r0.x, r7.x, r1.x, r0.x
-nop
-mul.f r0.y, r0.y, r0.w
-mul.f r1.x, r1.y, r0.w
-mov.f32f32 r0.x, r0.x
-nop
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r1.x, r1.x
-mad.f32 r0.y, c7.z, r7.z, r0.y
-mad.f32 r1.x, c7.y, r7.y, r1.x
+exp2 r0.y, r0.y
+(ss)mul.f r1.w, r4.x, r0.y
+mul.f r3.w, r4.z, r0.y
+mad.f32 r1.z, r2.z, r1.z, r1.w
+mad.f32 r1.x, r2.y, r1.x, r3.w
+(ss)mul.f r0.y, r1.y, r0.y
+nop
+mul.f r1.y, r1.z, r4.y
+mul.f r1.x, r1.x, r4.y
+mad.f32 r1.y, c7.z, r2.z, r1.y
+mad.f32 r1.x, c7.y, r2.y, r1.x
+mad.f32 r0.x, r2.x, r0.x, r0.y
+nop
+mul.f r0.y, r0.z, r1.y
+mul.f r1.x, r0.z, r1.x
 mul.f r0.x, r0.x, r0.w
 nop
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.w, r1.x
-mov.f32f32 r0.x, r0.x
-nop
-mul.f r0.y, r0.z, r0.y
-mul.f r0.w, r0.z, r0.w
-mad.f32 r0.x, c7.x, r7.x, r0.x
-nop
-add.f r0.y, r0.y, r5.x
-add.f r0.w, r0.w, r5.w
-mov.f32f32 r0.x, r0.x
+add.f r0.y, r0.y, r7.y
+add.f r0.w, r1.x, r3.y
+mad.f32 r0.x, c7.x, r2.x, r0.x
 nop
-mul.f r0.y, r0.y, r4.x
-mul.f r0.w, r0.w, r4.x
+mul.f r0.y, r0.y, r5.x
+mul.f r0.w, r0.w, r5.x
 mul.f r0.x, r0.z, r0.x
 nop
-mul.f r0.y, r0.y, c6.z
-mul.f r0.z, r0.w, c6.y
-add.f r0.x, r0.x, r6.x
-nop
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.z, r0.z
-mul.f r0.x, r0.x, r4.x
-nop
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.z, r0.z
-mul.f r0.x, r0.x, c6.x
-nop
-mov.f32f32 r3.z, r0.y
-mov.f32f32 r3.y, r0.z
-mov.f32f32 r0.x, r0.x
+mul.f r2.z, r0.y, c6.z
+mul.f r2.y, r0.w, c6.y
+add.f r0.x, r0.x, r3.z
 (rpt2)nop
-mov.f32f32 r0.x, r0.x
+mul.f r0.x, r0.x, r5.x
 (rpt2)nop
-mov.f32f32 r3.x, r0.x
+mul.f r2.x, r0.x, c6.x
 end
 nop
 nop
-nop
-; FRAG: outputs: r3.x (1:0)
-; FRAG: inputs: r0.x (0:0,cm=f,il=8,b=0) r2.z (5:9,cm=f,il=8,b=1) r63.y (5:10,cm=f,il=12,b=1) r3.x (5:11,cm=f,il=16,b=1) r2.y (5:12,cm=f,il=20,b=1) r7.x (5:13,cm=f,il=24,b=1) r8.x (5:14,cm=f,il=28,b=1) r7.x (5:15,cm=f,il=32,b=1)
-; FRAG: 802 instructions, 0 half, 65 full
-; pos (bary): r1.x
-; color: r3.x
-; fragcoord: r0.x
+; FRAG: outputs: r2.x (1:0)
+; FRAG: inputs: r0.x (0:0,cm=f,il=8,b=0) r3.x (5:9,cm=f,il=8,b=1) r63.z (5:10,cm=f,il=12,b=1) r3.x (5:11,cm=f,il=16,b=1) r1.w (5:12,cm=f,il=20,b=1) r6.x (5:13,cm=f,il=24,b=1) r6.x (5:14,cm=f,il=28,b=1) r7.x (5:15,cm=f,il=32,b=1)
+; FRAG: 532 instructions, 0 half, 13 full