summaryrefslogtreecommitdiff
path: root/reference/xon9.asm
diff options
context:
space:
mode:
Diffstat (limited to 'reference/xon9.asm')
-rw-r--r--reference/xon9.asm124
1 files changed, 45 insertions, 79 deletions
diff --git a/reference/xon9.asm b/reference/xon9.asm
index 47cf0de..6189f11 100644
--- a/reference/xon9.asm
+++ b/reference/xon9.asm
@@ -6,91 +6,57 @@
@out(r1.y) out1
@out(r1.z) out2
@out(r1.w) out3
+@const(c9.x) 0x3f000000, 0x00000000, 0x00000000, 0x00000000
+@const(c10.x) 0x00000000, 0x3f800000, 0x43000000, 0x00000000
(sy)(ss)bary.f r0.z, 0, r0.x
-bary.f r0.w, 4, r0.x
-mov.f32f32 r1.x, c9.y
-bary.f r1.y, 1, r0.x
+bary.f r0.w, 1, r0.x
+bary.f r1.x, 4, r0.x
+mov.f32f32 r1.y, (0.000000)
mov.f32f32 r1.z, r0.z
-mul.f r0.w, r0.w, r0.w
-bary.f r1.w, 5, r0.x
-mov.f32f32 r2.y, r0.z
-mov.f32f32 r2.z, r1.z
-mov.f32f32 r1.z, r1.y
-mad.f32 r0.w, r1.w, r1.w, r0.w
-mov.f32f32 r3.x, r2.y
-mov.f32f32 r1.w, r1.y
-mov.f32f32 r2.w, r1.z
-mov.f32f32 r0.w, r0.w
+mov.f32f32 r1.w, r0.w
+mul.f r1.x, r1.x, r1.x
+bary.f r2.y, 5, r0.x
+sam (f32)(xyzw)r2.z, r0.z, s#0, t#0
+(sy)cmps.f.lt r3.z, r3.y, c9.x
+mov.f32f32 r3.w, r0.z
+mov.f32f32 r4.x, r0.w
+(ss)mad.f32 r0.z, r2.y, r2.y, r1.x
+sam (f32)(xyz)r4.y, r1.z, s#1, t#1
+(sy)mul.f r0.w, r4.w, c3.z
+mul.f r1.x, r4.z, c3.y
+(ss)mul.f r1.z, r4.y, c3.x
+cov.u32f32 r1.w, r3.z
bary.f (ei)r0.x, 6, r0.x
-mov.f32f32 r3.y, r1.w
-mov.f32f32 r0.y, r0.z
-mov.f32f32 r0.z, r1.x
-mad.f32 r0.x, r0.x, r0.x, r0.w
-sam (f32)(xyz)r3.z, r2.z, s#1, t#1
-(sy)mul.f r0.w, r4.x, c3.z
-mul.f r1.x, r3.w, c3.y
-mul.f r1.z, r3.z, c3.x
-(ss)nop
-sam (f32)(xyzw)r2.y, r3.x, s#0, t#0
-(sy)cmps.f.lt r1.w, r3.x, c9.x
-mov.f32f32 r0.w, r0.w
-(ss)mov.f32f32 r3.y, r0.y
-mov.f32f32 r0.y, r1.y
-mov.f32f32 r1.x, r1.x
-mov.f32f32 r1.y, r1.z
-sqrt r0.x, r0.x
-(ss)mov.f32f32 r0.x, r0.x
-mov.f32f32 r3.z, r0.y
-cov.u32f32 r0.y, r1.w
-mov.f32f32 r1.z, (0.000000)
-mul.f r1.w, r3.x, c7.x
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r0.z, r0.z
-cmps.f.ne p0.x, r0.y, r1.z
-sam (f32)(xyz)r3.x, r3.y, s#2, t#2
-(sy)mad.f32 r0.y, c4.z, r3.z, r0.w
-mad.f32 r0.w, c4.y, r3.y, r1.x
-mad.f32 r1.x, c4.x, r3.x, r1.y
-mov.f32f32 r1.y, r0.x
-mov.f32f32 r0.x, r0.y
-mov.f32f32 r0.y, r0.w
-mov.f32f32 r0.w, r1.x
-mov.f32f32 r1.z, r0.z
-add.f r0.x, r2.w, r0.x
-add.f r0.y, r2.z, r0.y
-add.f r0.z, r2.y, r0.w
+sam (f32)(xyz)r3.z, r3.w, s#2, t#2
+(sy)mad.f32 r0.y, c4.z, r4.x, r0.w
+mad.f32 r0.w, c4.y, r3.w, r1.x
+mad.f32 r1.x, c4.x, r3.z, r1.z
+cmps.f.ne p0.x, r1.w, r1.y
+add.f r0.y, r3.x, r0.y
+add.f r0.w, r2.w, r0.w
+add.f r1.x, r2.z, r1.x
+mad.f32 r0.x, r0.x, r0.x, r0.z
+mul.f r0.y, r0.y, c6.z
+mul.f r0.z, r0.w, c6.y
+mul.f r0.w, r1.x, c6.x
kill p0.x
-mul.f r0.x, r0.x, c6.z
-mul.f r0.y, r0.y, c6.y
-mul.f r0.z, r0.z, c6.x
-sam (f32)(x)r0.w, r1.y, s#3, t#3
-mov.f32f32 r1.x, r1.w
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r1.w, r1.x
-mul.f r0.x, r0.x, c8.z
-mul.f r0.y, r0.y, c8.y
-mul.f r0.z, r0.z, c8.x
-nop
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.z, r0.z
-nop
-(sy)mul.f r0.x, r0.x, r0.w
-mul.f r0.y, r0.y, r0.w
-mul.f r0.z, r0.z, r0.w
-nop
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.z, r0.z
-nop
-(ss)mov.f32f32 r1.z, r0.x
-mov.f32f32 r1.y, r0.y
-mov.f32f32 r1.x, r0.z
+mul.f r0.y, r0.y, c8.z
+mul.f r0.z, r0.z, c8.y
+sqrt r1.x, r0.x
+mov.f32f32 r1.y, c9.y
+(ss)mul.f r0.x, r0.w, c8.x
+(rpt1)nop
+mul.f r1.w, r3.y, c7.x
+(rpt1)nop
+(ss)nop
+sam (f32)(x)r2.y, r1.x, s#3, t#3
+(sy)mul.f r1.z, r0.y, r2.y
+(ss)mul.f r1.y, r0.z, r2.y
+mul.f r1.x, r0.x, r2.y
end
nop
+nop
; FRAG: outputs: r1.x (1:0)
; FRAG: inputs: r0.x (5:20,cm=f,il=8,b=1) r1.x (5:21,cm=f,il=12,b=1)
-; FRAG: 83 instructions, 0 half, 5 full
+; FRAG: 48 instructions, 0 half, 5 full