summaryrefslogtreecommitdiff
path: root/reference/gmaps-frag.asm
diff options
context:
space:
mode:
Diffstat (limited to 'reference/gmaps-frag.asm')
-rw-r--r--reference/gmaps-frag.asm100
1 files changed, 39 insertions, 61 deletions
diff --git a/reference/gmaps-frag.asm b/reference/gmaps-frag.asm
index a658d94..b3a1bb0 100644
--- a/reference/gmaps-frag.asm
+++ b/reference/gmaps-frag.asm
@@ -6,79 +6,57 @@
@out(r1.y) out1
@out(r1.z) out2
@out(r1.w) out3
+@const(c1.x) 0x00000000, 0x3f000000, 0x3f800000, 0x40000000
+@const(c2.x) 0x00000000, 0x3f800000, 0x43000000, 0x00000000
(sy)(ss)bary.f r0.z, 4, r0.x
-bary.f r0.w, 6, r0.x
-bary.f r1.x, 7, r0.x
-bary.f r1.y, 5, r0.x
+bary.f r1.x, 6, r0.x
+bary.f r1.y, 7, r0.x
+bary.f r0.w, 5, r0.x
mul.f r1.z, c0.x, r0.z
-mov.f32f32 r2.x, r0.z
-cmps.f.lt r2.z, c1.x, r0.w
-mul.f r1.w, c0.x, r1.y
-mov.f32f32 r2.y, r1.y
-min.f r0.w, r0.w, r1.x
-cov.u32f32 r2.z, r2.z
-cmps.f.lt r1.x, c1.x, r1.x
-mul.f r2.w, c1.w, r0.z
-add.f r0.w, r0.w, c1.y
-dsy (f32)(xy)r3.x, r1.z
-(sy)(ss)mov.f32f32 r1.z, r3.x
-mov.f32f32 r1.w, r3.y
-dsx (f32)(xy)r3.x, r2.x
-cov.u32f32 r1.x, r1.x
-(sy)(ss)mov.f32f32 r2.x, r3.x
-mad.f32 r1.z, r2.w, r1.z, (neg)r1.w
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r1.w, r3.y
-mul.f r1.x, r2.z, r1.x
-mov.f32f32 r1.z, r1.z
-mad.f32 r1.w, r2.w, r2.x, (neg)r1.w
-min.f r0.w, r0.w, c1.z
-mov.f32f32 r1.x, r1.x
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r1.w, r1.w
+mul.f r2.x, c1.w, r0.z
+cmps.f.lt r2.y, c1.x, r1.x
+mul.f r1.w, c0.x, r0.w
+min.f r1.x, r1.x, r1.y
+cmps.f.lt r1.y, c1.x, r1.y
+dsx (f32)(xy)r2.z, r0.z
+(sy)mad.f32 r2.z, r2.x, r2.z, (neg)r2.w
+cov.u32f32 r2.y, r2.y
+add.f r1.x, r1.x, c1.y
+cov.u32f32 r1.y, r1.y
+dsy (f32)(xy)r2.w, r1.z
+(sy)(ss)mad.f32 r1.z, r2.x, r2.w, (neg)r3.x
+mov.f32f32 r1.w, r2.z
+min.f r1.x, r1.x, c1.z
+mad.f32 r0.z, r0.z, r0.z, (neg)r0.w
+mov.f32f32 r0.w, r1.z
+mul.f r1.w, r2.z, r1.w
+mul.f r1.y, r2.y, r1.y
+nop
mov.f32f32 r0.w, r0.w
-mad.f32 r0.z, r0.z, r0.z, (neg)r1.y
-bary.f r1.y, 3, r0.x
-mul.f r1.w, r1.w, r1.w
-cmps.f.ne r1.x, r1.x, c1.x
-mad.f32 r1.z, r1.z, r1.z, r1.w
-bary.f r1.w, 2, r0.x
-bary.f r2.x, 1, r0.x
+bary.f r2.x, 3, r0.x
+bary.f r2.y, 2, r0.x
+bary.f r2.z, 1, r0.x
+mad.f32 r0.w, r1.z, r0.w, r1.w
+cmps.f.ne r1.y, r1.y, c1.x
bary.f (ei)r0.x, 0, r0.x
-mov.f32f32 r0.y, r1.z
-(rpt5)nop
-rsq r0.y, r0.y
-(ss)mov.f32f32 r0.y, r0.y
-(rpt2)nop
-mul.f r0.y, r0.z, r0.y
-(rpt2)nop
-mov.f32f32 r0.y, r0.y
+(rpt3)nop
+rsq r0.y, r0.w
+(ss)mul.f r0.y, r0.z, r0.y
(rpt2)nop
add.f r0.y, c1.y, (neg)r0.y
(rpt2)nop
-mov.f32f32 r0.y, r0.y
-(rpt2)nop
max.f r0.y, r0.y, c1.x
(rpt2)nop
min.f r0.y, r0.y, c1.z
(rpt2)nop
-mov.f32f32 r0.y, r0.y
+sel.b32 r0.y, r1.x, r1.y, r0.y
(rpt2)nop
-sel.b32 r0.y, r0.w, r1.x, r0.y
-(rpt2)nop
-mul.f r0.z, r1.y, r0.y
-mul.f r0.w, r1.w, r0.y
-mul.f r1.x, r2.x, r0.y
-mul.f r0.x, r0.x, r0.y
-mov.f32f32 r1.w, r0.z
-mov.f32f32 r1.z, r0.w
-mov.f32f32 r1.y, r1.x
-mov.f32f32 r1.x, r0.x
+mul.f r1.w, r2.x, r0.y
+mul.f r1.z, r2.y, r0.y
+mul.f r1.y, r2.z, r0.y
+mul.f r1.x, r0.x, r0.y
end
-nop
-nop
-nop
; FRAG: outputs: r1.x (1:0)
-; FRAG: inputs: r0.z (5:20,cm=f,il=8,b=1) r0.y (5:21,cm=f,il=12,b=1)
-; FRAG: 92 instructions, 0 half, 4 full
+; FRAG: inputs: r1.y (5:20,cm=f,il=8,b=1) r0.z (5:21,cm=f,il=12,b=1)
+; FRAG: 61 instructions, 0 half, 4 full