summaryrefslogtreecommitdiff
path: root/reference/test2.asm
diff options
context:
space:
mode:
Diffstat (limited to 'reference/test2.asm')
-rw-r--r--reference/test2.asm120
1 files changed, 59 insertions, 61 deletions
diff --git a/reference/test2.asm b/reference/test2.asm
index d1549f6..ce47110 100644
--- a/reference/test2.asm
+++ b/reference/test2.asm
@@ -32,99 +32,97 @@
@out(r3.y) out13
@out(r3.z) out14
@out(r3.w) out15
+@const(c25.x) 0x3f000000, 0x00000000, 0x3f800000, 0x00000000
+@const(c26.x) 0x00000000, 0x3f800000, 0x43000000, 0x00000000
(sy)(ss)mul.f r0.x, r1.x, c12.x
-mul.f r0.y, r4.x, c4.x
+mul.f r0.y, r1.x, c12.y
mad.f32 r0.x, c13.x, r1.y, r0.x
-mad.f32 r0.y, c4.y, r4.y, r0.y
+mad.f32 r0.y, c13.y, r1.y, r0.y
mad.f32 r0.x, c14.x, r1.z, r0.x
-mul.f r0.z, r4.x, c5.x
-mad.f32 r0.x, c15.x, r1.w, r0.x
-mov.f32f32 r0.y, r0.y
-mad.f32 r0.z, c5.y, r4.y, r0.z
-mul.f r0.w, r4.x, c6.x
-mul.f r3.x, r0.x, r0.x
-mul.f r3.y, r1.x, c12.y
-mad.f32 r6.x, c4.z, r4.z, r0.y
-mad.f32 r0.y, c13.y, r1.y, r3.y
-mov.f32f32 r0.z, r0.z
mad.f32 r0.y, c14.y, r1.z, r0.y
-mad.f32 r6.y, c5.z, r4.z, r0.z
-mad.f32 r3.y, c15.y, r1.w, r0.y
-mad.f32 r0.y, c6.y, r4.y, r0.w
+mad.f32 r0.x, c15.x, r1.w, r0.x
+mad.f32 r0.y, c15.y, r1.w, r0.y
mul.f r0.z, r1.x, c12.z
mul.f r0.w, r1.x, c0.w
-mad.f32 r3.x, r3.y, r3.y, r3.x
-mov.f32f32 r0.y, r0.y
+mul.f r3.x, r0.x, r0.x
mad.f32 r0.z, c13.z, r1.y, r0.z
-mad.f32 r0.w, c1.w, r1.y, r0.w
-mov.f32f32 r3.x, r3.x
+mad.f32 r3.x, r0.y, r0.y, r3.x
mad.f32 r0.z, c14.z, r1.z, r0.z
-mad.f32 r6.z, c6.z, r4.z, r0.y
-mad.f32 r3.z, c15.z, r1.w, r0.z
-mad.f32 r0.y, c2.w, r1.z, r0.w
-mul.f r0.z, r1.x, c0.z
-mul.f r3.w, r1.x, c0.y
-mad.f32 r3.x, r3.z, r3.z, r3.x
-mad.f32 r0.w, c3.w, r1.w, r0.y
-mad.f32 r0.y, c1.z, r1.y, r0.z
-mad.f32 r0.z, c1.y, r1.y, r3.w
+mad.f32 r0.w, c1.w, r1.y, r0.w
+mad.f32 r0.z, c15.z, r1.w, r0.z
+mad.f32 r0.w, c2.w, r1.z, r0.w
+mul.f r3.y, r1.x, c0.z
+mul.f r3.z, r1.x, c0.y
+mad.f32 r3.x, r0.z, r0.z, r3.x
+mad.f32 r0.w, c3.w, r1.w, r0.w
+mad.f32 r3.y, c1.z, r1.y, r3.y
+mad.f32 r3.z, c1.y, r1.y, r3.z
mul.f r3.w, r1.x, c0.x
-mad.f32 r0.y, c2.z, r1.z, r0.y
-mad.f32 r6.w, c2.y, r1.z, r0.z
-rsq r0.z, (abs)r3.x
-(ss)mov.f32f32 r3.x, r0.z
-mad.f32 r0.z, c3.z, r1.w, r0.y
-mad.f32 r0.y, c3.y, r1.w, r6.w
-mad.f32 r3.w, c1.x, r1.y, r3.w
-mul.f r6.w, r0.x, r3.x
-mul.f r7.x, r3.y, r3.x
-mul.f r7.y, r3.z, r3.x
-mad.f32 r0.x, c2.x, r1.z, r3.w
-mul.f r3.x, r6.x, r6.w
+mul.f r6.x, r4.x, c4.x
+mul.f r6.y, r4.x, c5.x
+rsq r3.x, (abs)r3.x
+(ss)mov.f32f32 r6.z, r3.x
+mul.f r6.w, r0.z, r3.x
+mad.f32 r0.z, c2.z, r1.z, r3.y
+(ss)mad.f32 r3.x, c2.y, r1.z, r3.z
+mul.f r7.x, r0.x, r6.z
+mad.f32 r0.x, c4.y, r4.y, r6.x
+mul.f r6.x, r0.y, r6.z
+mad.f32 r6.z, c4.z, r4.z, r0.x
+mad.f32 r0.z, c3.z, r1.w, r0.z
+mad.f32 r0.y, c3.y, r1.w, r3.x
+mad.f32 r0.x, c1.x, r1.y, r3.w
+mul.f r3.x, r6.z, r7.x
+mad.f32 r3.y, c5.y, r4.y, r6.y
+mad.f32 r0.x, c2.x, r1.z, r0.x
+mad.f32 r6.y, c5.z, r4.z, r3.y
mad.f32 r0.x, c3.x, r1.w, r0.x
-mad.f32 r3.x, r6.y, r7.x, r3.x
+mul.f r3.y, r4.x, c6.x
add.f r3.w, r2.w, r5.w
+mad.f32 r5.w, r6.y, r6.x, r3.x
+mad.f32 r2.w, c6.y, r4.y, r3.y
add.f r3.z, r2.z, r5.z
add.f r3.y, r2.y, r5.y
-mov.f32f32 r2.y, r3.x
+mov.f32f32 r2.y, r5.w
+mad.f32 r5.y, c6.z, r4.z, r2.w
add.f r3.x, r2.x, r5.x
-mad.f32 r2.x, r6.z, r7.y, r2.y
-mad.f32 r2.y, r6.z, r7.y, r2.y
add.f r2.w, r1.w, r4.w
add.f r2.z, r1.z, r4.z
-mad.f32 r1.z, (neg)r2.x, r6.x, r6.w
-mad.f32 r1.w, (neg)r2.y, r6.y, r7.x
+mad.f32 r1.z, r5.y, r6.w, r2.y
+mad.f32 r1.w, r5.y, r6.w, r5.w
add.f r2.y, r1.y, r4.y
add.f r2.x, r1.x, r4.x
-add.f r1.x, r1.z, c25.y
-add.f r1.y, r1.w, c25.y
+mad.f32 r1.x, (neg)r1.z, r6.z, r7.x
+mad.f32 r1.y, (neg)r1.w, r6.y, r6.x
mov.f32f32 r4.x, r4.z
mov.f32f32 r4.y, r4.w
-mul.f r1.x, r1.x, c25.x
-mul.f r1.y, r1.y, c25.x
+add.f r1.z, r1.x, c25.y
+add.f r1.w, r1.y, c25.y
+(rpt1)nop
+mul.f r1.z, r1.z, c25.x
+mul.f r1.w, r1.w, c25.x
(rpt1)nop
-mad.f32 r1.x, r1.z, r1.x, c25.x
-mad.f32 r1.y, r1.w, r1.y, c25.x
+mad.f32 r1.x, r1.x, r1.z, c25.x
+mad.f32 r1.y, r1.y, r1.w, c25.x
(rpt1)nop
mul.f r1.z, r1.x, c21.w
mul.f r1.w, r1.x, c21.z
mad.f32 r1.z, c22.w, r1.y, r1.z
mad.f32 r1.w, c22.z, r1.y, r1.w
mad.f32 r1.z, c23.w, r4.x, r1.z
-mad.f32 r4.z, c23.z, r4.x, r1.w
+mad.f32 r5.x, c23.z, r4.x, r1.w
mad.f32 r1.w, c24.w, r4.y, r1.z
-mad.f32 r1.z, c24.z, r4.y, r4.z
-mul.f r4.z, r1.x, c21.y
+mad.f32 r1.z, c24.z, r4.y, r5.x
+mul.f r5.x, r1.x, c21.y
mul.f r1.x, r1.x, c21.x
-mad.f32 r4.z, c22.y, r1.y, r4.z
+mad.f32 r5.x, c22.y, r1.y, r5.x
mad.f32 r1.x, c22.x, r1.y, r1.x
-mad.f32 r1.y, c23.y, r4.x, r4.z
-mad.f32 r1.x, c23.x, r4.x, r1.x
+mad.f32 r1.y, c23.y, r4.x, r5.x
+mad.f32 r1.x, c23.x, r4.z, r1.x
mad.f32 r1.y, c24.y, r4.y, r1.y
-mad.f32 r1.x, c24.x, r4.y, r1.x
+mad.f32 r1.x, c24.x, r4.w, r1.x
end
-nop
; VERT: outputs: r0.x (0:0) r1.x (5:0) r2.x (5:0) r3.x (5:0)
; VERT: inputs: r1.x (0:0,cm=f,il=8,b=0) r4.x (0:0,cm=f,il=12,b=0) r2.x (0:0,cm=f,il=16,b=0) r5.x (0:0,cm=f,il=20,b=0)
-; VERT: 93 instructions, 0 half, 8 full
+; VERT: 91 instructions, 0 half, 8 full