summaryrefslogtreecommitdiff
path: root/reference/es2gears-vert.asm
diff options
context:
space:
mode:
Diffstat (limited to 'reference/es2gears-vert.asm')
-rw-r--r--reference/es2gears-vert.asm112
1 files changed, 43 insertions, 69 deletions
diff --git a/reference/es2gears-vert.asm b/reference/es2gears-vert.asm
index bc04597..b3981b3 100644
--- a/reference/es2gears-vert.asm
+++ b/reference/es2gears-vert.asm
@@ -14,6 +14,8 @@
@out(r1.y) out5
@out(r1.z) out6
@out(r1.w) out7
+@const(c10.x) 0x00000000, 0x00000000, 0x00000000, 0x00000000
+@const(c11.x) 0x00000000, 0x3f800000, 0x43000000, 0x00000000
(sy)(ss)mul.f r0.w, c4.x, r0.x
mov.f32f32 r1.w, c8.x
mad.f32 r0.w, c5.x, r0.y, r0.w
@@ -25,88 +27,60 @@ mad.f32 r2.x, c5.y, r0.y, r2.x
add.f r0.w, r0.w, c7.x
mad.f32 r2.x, c6.y, r0.z, r2.x
mul.f r0.x, c4.z, r0.x
-mul.f r2.z, c0.w, r1.x
-mov.f32f32 r2.w, r0.w
-mad.f32 r0.w, c8.y, r2.y, r1.w
-add.f r1.w, r2.x, c7.y
+mad.f32 r1.w, c8.y, r2.y, r1.w
+mov.f32f32 r2.y, r0.w
+mov.f32f32 r2.z, c8.z
+add.f r2.x, r2.x, c7.y
mad.f32 r0.x, c5.z, r0.y, r0.x
-mul.f r0.y, r2.w, r2.w
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r1.w, r1.w
-mov.f32f32 r2.x, c8.z
+mul.f r0.y, r0.w, r2.y
+mad.f32 r0.w, c8.z, r2.z, r1.w
+mov.f32f32 r1.w, r2.x
mad.f32 r0.x, c6.z, r0.z, r0.x
-mad.f32 r0.z, c1.w, r1.y, r2.z
-mad.f32 r0.y, r1.w, r1.w, r0.y
-mad.f32 r0.w, c8.z, r2.x, r0.w
+mul.f r0.z, c0.w, r1.x
+mul.f r2.z, c0.z, r1.x
+mad.f32 r0.y, r2.x, r1.w, r0.y
add.f r0.x, r0.x, c7.z
-mad.f32 r0.z, c2.w, r1.z, r0.z
-mov.f32f32 r0.y, r0.y
-mul.f r2.x, c0.z, r1.x
-mov.f32f32 r0.x, r0.x
-add.f r0.z, r0.z, c3.w
-mad.f32 r2.x, c1.z, r1.y, r2.x
-mul.f r2.y, c0.y, r1.x
-mad.f32 r0.y, r0.x, r0.x, r0.y
rsq r0.w, r0.w
-(ss)mov.f32f32 r2.z, r0.w
-(ss)mov.f32f32 r0.w, r0.z
-mad.f32 r0.z, c2.z, r1.z, r2.x
-mad.f32 r2.x, c1.y, r1.y, r2.y
+(ss)mov.f32f32 r2.x, r0.w
+mul.f r2.w, c8.z, r0.w
+mad.f32 r0.z, c1.w, r1.y, r0.z
+(ss)mov.f32f32 r0.w, r0.x
+mul.f r3.x, c8.x, r2.x
+mul.f r2.x, c8.y, r2.x
+mad.f32 r0.z, c2.w, r1.z, r0.z
+mad.f32 r0.y, r0.w, r0.w, r0.y
+mad.f32 r0.w, c1.z, r1.y, r2.z
+mul.f r2.z, c0.y, r1.x
mul.f r1.x, c0.x, r1.x
-mul.f r2.y, c8.x, r2.z
+mad.f32 r3.y, c2.z, r1.z, r0.w
+mad.f32 r2.z, c1.y, r1.y, r2.z
+mad.f32 r1.x, c1.x, r1.y, r1.x
rsq r0.y, r0.y
-(ss)mov.f32f32 r0.y, r0.y
-mul.f r3.x, c8.y, r2.z
-mul.f r2.z, c8.z, r2.z
-nop
-mul.f r2.w, r2.w, r0.y
-mul.f r1.w, r1.w, r0.y
+(ss)mov.f32f32 r1.y, r0.y
mul.f r0.x, r0.x, r0.y
-nop
-mov.f32f32 r0.y, r2.w
-mov.f32f32 r2.y, r2.y
-mov.f32f32 r1.w, r1.w
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r2.w, r3.x
-mul.f r0.y, r0.y, r2.y
-mov.f32f32 r2.y, r2.z
-add.f r0.z, r0.z, c3.z
-mad.f32 r0.y, r1.w, r2.w, r0.y
-mad.f32 r1.w, c2.y, r1.z, r2.x
-mad.f32 r1.x, c1.x, r1.y, r1.x
-nop
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.z, r0.z
-mad.f32 r0.x, r0.x, r2.y, r0.y
-add.f r0.y, r1.w, c3.y
+add.f r0.w, r0.z, c3.w
+add.f r0.z, r3.y, c3.z
+(ss)mul.f r0.y, r2.y, r1.y
+mul.f r1.y, r1.w, r1.y
+mad.f32 r1.w, c2.y, r1.z, r2.z
mad.f32 r1.x, c2.x, r1.z, r1.x
+mul.f r0.y, r0.y, r3.x
nop
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r0.y, r0.y
-add.f r1.x, r1.x, c3.x
-nop
-max.f r0.x, r0.x, c10.x
+mad.f32 r1.y, r1.y, r2.x, r0.y
+add.f r0.y, r1.w, c3.y
+mad.f32 r1.y, r0.x, r2.w, r1.y
+add.f r0.x, r1.x, c3.x
+(rpt1)nop
+max.f r1.x, r1.y, c10.x
(rpt2)nop
-mov.f32f32 r1.y, r0.x
-mov.f32f32 r0.x, r1.x
+mov.f32f32 r1.y, r1.x
+mul.f r1.x, r1.x, c9.x
(rpt1)nop
-mul.f r1.x, r1.y, c9.w
+mul.f r1.w, r1.y, c9.w
mul.f r1.z, r1.y, c9.z
-mul.f r1.w, r1.y, c9.y
-mul.f r1.y, r1.y, c9.x
-mov.f32f32 r1.x, r1.x
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r2.x, r1.w
-mov.f32f32 r2.y, r1.y
-mov.f32f32 r1.w, r1.x
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r1.y, r2.x
-mov.f32f32 r1.x, r2.y
+mul.f r1.y, r1.y, c9.y
end
-nop
-nop
-nop
; VERT: outputs: r0.x (0:0) r1.x (5:19)
; VERT: inputs: r1.x (0:0,cm=7,il=8,b=0) r0.x (0:0,cm=7,il=12,b=0)
-; VERT: 92 instructions, 0 half, 4 full
+; VERT: 68 instructions, 0 half, 4 full