summaryrefslogtreecommitdiff
path: root/reference/maniadrive/maniadrive-15.asm
diff options
context:
space:
mode:
Diffstat (limited to 'reference/maniadrive/maniadrive-15.asm')
-rw-r--r--reference/maniadrive/maniadrive-15.asm374
1 files changed, 174 insertions, 200 deletions
diff --git a/reference/maniadrive/maniadrive-15.asm b/reference/maniadrive/maniadrive-15.asm
index 7441bc0..a69b40d 100644
--- a/reference/maniadrive/maniadrive-15.asm
+++ b/reference/maniadrive/maniadrive-15.asm
@@ -1,16 +1,16 @@
; options:
; VERT: new compiler
-@in(r3.x) in0
-@in(r3.y) in1
-@in(r3.z) in2
-@in(r3.w) in3
+@in(r6.x) in0
+@in(r6.y) in1
+@in(r6.z) in2
+@in(r6.w) in3
@in(r0.x) in4
@in(r0.y) in5
@in(r0.z) in6
-@in(r6.x) in8
-@in(r6.y) in9
-@in(r6.z) in10
-@in(r6.w) in11
+@in(r3.x) in8
+@in(r3.y) in9
+@in(r3.z) in10
+@in(r3.w) in11
@out(r0.x) out0
@out(r0.y) out1
@out(r0.z) out2
@@ -35,211 +35,185 @@
@out(r5.y) out21
@out(r5.z) out22
@out(r5.w) out23
-(sy)(ss)mul.f r0.w, r3.x, c13.x
-mov.f32f32 r1.x, c24.y
-mad.f32 r0.w, c14.x, r3.y, r0.w
-mov.f32f32 r1.y, c11.w
-mad.f32 r0.w, c15.x, r3.z, r0.w
-mov.f32f32 r1.x, r1.x
-mad.f32 r0.w, c16.x, r3.w, r0.w
-mov.f32f32 r1.y, r1.y
+@const(c24.x) 0x00000000, 0x3f800000, 0x00000000, 0x00000000
+@const(c25.x) 0x00000000, 0x3f800000, 0x43000000, 0x00000000
+(sy)(ss)mul.f r0.w, r6.x, c13.x
+mul.f r1.x, r6.x, c13.y
+mad.f32 r0.w, c14.x, r6.y, r0.w
+mad.f32 r1.x, c14.y, r6.y, r1.x
+mad.f32 r0.w, c15.x, r6.z, r0.w
+mad.f32 r1.x, c15.y, r6.z, r1.x
+mad.f32 r0.w, c16.x, r6.w, r0.w
+mad.f32 r1.x, c16.y, r6.w, r1.x
+mul.f r1.y, r6.x, c13.z
mov.f32f32 r1.z, c24.y
-mov.f32f32 r1.w, c8.w
add.f r0.w, c12.x, (neg)r0.w
-max.f r1.x, r1.x, c24.x
-max.f r1.y, r1.y, c24.x
-mov.f32f32 r1.z, r1.z
-mul.f r2.x, r0.w, r0.w
-mul.f r2.y, r3.x, c13.y
-min.f r5.w, r1.x, c24.y
-mad.f32 r1.x, c14.y, r3.y, r2.y
-min.f r4.w, r1.y, c24.y
-mad.f32 r1.x, c15.y, r3.z, r1.x
-max.f r1.y, r1.z, c24.x
-mad.f32 r1.x, c16.y, r3.w, r1.x
-mov.f32f32 r1.z, r1.w
-mul.f r1.w, r3.x, c13.z
-mul.f r2.y, r0.x, c4.x
add.f r1.x, c12.y, (neg)r1.x
-min.f r2.w, r1.y, c24.y
-max.f r1.y, r1.z, c24.x
-mad.f32 r1.z, c14.z, r3.y, r1.w
-mad.f32 r1.w, r1.x, r1.x, r2.x
-mad.f32 r1.z, c15.z, r3.z, r1.z
-mad.f32 r2.x, c4.y, r0.y, r2.y
-mul.f r2.y, r0.x, c5.x
-mov.f32f32 r2.z, r1.w
-mad.f32 r1.z, c16.z, r3.w, r1.z
-min.f r1.w, r1.y, c24.y
-mov.f32f32 r1.y, r2.x
-mad.f32 r2.x, c5.y, r0.y, r2.y
-add.f r1.z, c12.z, (neg)r1.z
-mad.f32 r1.y, c4.z, r0.z, r1.y
-mul.f r0.x, r0.x, c6.x
-mov.f32f32 r2.x, r2.x
-mad.f32 r2.y, r1.z, r1.z, r2.z
-mad.f32 r2.x, c5.z, r0.z, r2.x
-mad.f32 r0.x, c6.y, r0.y, r0.x
-mul.f r0.y, r3.x, c0.w
-mul.f r2.z, r3.x, c0.z
-mul.f r4.x, r3.x, c0.y
-mul.f r3.x, r3.x, c0.x
-rsq r2.y, (abs)r2.y
-(ss)mul.f r0.w, r0.w, r2.y
-rcp r4.y, r2.y
-(ss)mov.f32f32 r4.y, r4.y
-rcp r4.z, r2.y
-(ss)mov.f32f32 r4.z, r4.z
-rsq r5.x, (abs)r2.y
+mad.f32 r1.y, c14.z, r6.y, r1.y
+max.f r1.z, r1.z, c24.x
+mul.f r1.w, r0.w, r0.w
+mad.f32 r1.y, c15.z, r6.z, r1.y
+mad.f32 r1.w, r1.x, r1.x, r1.w
+mad.f32 r1.y, c16.z, r6.w, r1.y
+min.f r5.w, r1.z, c24.y
+mov.f32f32 r1.z, c11.w
+mov.f32f32 r2.x, c24.y
+add.f r1.y, c12.z, (neg)r1.y
+mov.f32f32 r2.y, c8.w
+mul.f r2.z, r6.x, c0.w
+mul.f r2.w, r6.x, c0.z
+mad.f32 r1.w, r1.y, r1.y, r1.w
+max.f r1.z, r1.z, c24.x
+max.f r2.x, r2.x, c24.x
+max.f r2.y, r2.y, c24.x
+mad.f32 r2.z, c1.w, r6.y, r2.z
+mad.f32 r4.x, c1.z, r6.y, r2.w
+mul.f r4.y, r6.x, c0.y
+rsq r1.w, (abs)r1.w
+(ss)mul.f r0.w, r0.w, r1.w
+rcp r2.w, r1.w
+(ss)mov.f32f32 r4.z, r2.w
+mul.f r4.w, r0.x, c4.x
+rsq r5.x, (abs)r1.w
(ss)mul.f r1.x, r1.x, r5.x
add.f r5.x, r0.w, c24.x
-mul.f r5.y, r2.y, r4.y
-mul.f r0.w, r1.y, r0.w
-mul.f r4.z, r4.z, r4.y
-mul.f r5.z, r5.x, r5.x
-add.f r7.x, r1.x, c24.x
-mov.f32f32 r5.y, r5.y
-mad.f32 r0.w, r2.x, r1.x, r0.w
-mov.f32f32 r1.x, r4.z
-mad.f32 r4.z, r7.x, r7.x, r5.z
+mul.f r5.y, r1.w, r4.z
+mad.f32 r4.w, c4.y, r0.y, r4.w
+add.f r5.z, r1.x, c24.x
+mul.f r7.x, r5.x, r5.x
mul.f r5.y, c17.x, r5.y
-mov.f32f32 r0.w, r0.w
-(ss)rsq r2.y, (abs)r2.y
-mad.f32 r4.y, c17.y, r4.y, r5.y
-mov.f32f32 r4.z, r4.z
-(ss)mul.f r1.z, r1.z, r2.y
-mov.f32f32 r0.x, r0.x
-(ss)mov.f32f32 r2.y, r4.y
+mad.f32 r7.y, c4.z, r0.z, r4.w
+mad.f32 r7.x, r5.z, r5.z, r7.x
+rsq r4.w, (abs)r1.w
+(ss)mul.f r1.y, r1.y, r4.w
+mad.f32 r4.z, c17.y, r4.z, r5.y
+(ss)rcp r1.w, r1.w
+(ss)mul.f r1.w, r1.w, r2.w
+mul.f r0.w, r7.y, r0.w
+add.f r5.y, r1.y, c24.y
+mul.f r7.z, r0.x, c5.x
+min.f r4.w, r1.z, c24.y
+min.f r2.w, r2.x, c24.y
+mad.f32 r1.z, r5.y, r5.y, r7.x
+mad.f32 r2.x, c17.z, r1.w, r4.z
+mad.f32 r4.z, c5.y, r0.y, r7.z
+min.f r1.w, r2.y, c24.y
+mad.f32 r7.x, c2.w, r6.z, r2.z
+mad.f32 r7.z, c2.z, r6.z, r4.x
+mad.f32 r7.w, c1.y, r6.y, r4.y
+rsq r1.z, (abs)r1.z
+(ss)mov.f32f32 r2.y, r1.z
+mov.f32f32 r2.z, r2.x
+rcp r2.x, r2.x
+mad.f32 r4.x, c5.z, r0.z, r4.z
+(ss)mul.f r1.z, r5.y, r1.z
+mul.f r4.y, r5.x, r2.y
+mul.f r2.y, r5.z, r2.y
+mad.f32 r0.w, r4.x, r1.x, r0.w
+mul.f r0.x, r0.x, c6.x
+mul.f r1.x, r7.y, r4.y
+rcp r4.y, r2.z
+mov.f32f32 r4.z, c25.y
+mad.f32 r1.x, r4.x, r2.y, r1.x
+mad.f32 r0.x, c6.y, r0.y, r0.x
+mov.f32f32 r0.y, c25.y
mad.f32 r0.x, c6.z, r0.z, r0.x
-add.f r0.z, r1.z, c24.y
-mad.f32 r1.x, c17.z, r1.x, r2.y
-mad.f32 r0.y, c1.w, r3.y, r0.y
-mad.f32 r2.y, c1.z, r3.y, r2.z
-mad.f32 r2.z, r0.z, r0.z, r4.z
-mov.f32f32 r1.x, r1.x
-mad.f32 r0.w, r0.x, r1.z, r0.w
-mad.f32 r0.y, c2.w, r3.z, r0.y
-mad.f32 r2.y, c2.z, r3.z, r2.y
-mad.f32 r7.y, c1.y, r3.y, r4.x
-mad.f32 r3.x, c1.x, r3.y, r3.x
-rsq r1.z, (abs)r2.z
-(ss)mov.f32f32 r1.z, r1.z
-(ss)rcp r2.z, r1.x
-mov.f32f32 r3.y, c25.y
-mov.f32f32 r4.x, c25.y
-rcp r5.y, r1.x
-max.f r4.y, (neg)r0.w, c25.x
-mul.f r4.z, r5.x, r1.z
-(ss)mul.f r3.y, r3.y, r2.z
-mov.f32f32 r5.x, c11.y
-mov.f32f32 r5.z, c11.x
-mul.f r1.y, r1.y, r4.z
-mul.f r4.z, r7.x, r1.z
-mov.f32f32 r7.x, c11.z
-mad.f32 r5.x, c21.y, r3.y, r5.x
-mad.f32 r5.z, c21.x, r3.y, r5.z
-mad.f32 r1.y, r2.x, r4.z, r1.y
-mad.f32 r2.x, c21.z, r3.y, r7.x
-mov.f32f32 r3.y, r4.y
-mul.f r2.z, r4.x, r2.z
-mov.f32f32 r1.y, r1.y
-mul.f r0.z, r0.z, r1.z
-mul.f r1.z, r3.y, r5.y
-mov.f32f32 r3.y, c8.z
-mov.f32f32 r4.x, c8.y
-mad.f32 r0.x, r0.x, r0.z, r1.y
-mad.f32 r0.z, c22.z, r1.z, r2.x
-mad.f32 r1.y, c22.y, r1.z, r5.x
-mad.f32 r1.z, c22.x, r1.z, r5.z
-max.f r2.x, (neg)r0.x, c25.x
-max.f r0.x, r0.x, c25.x
+(ss)mul.f r0.z, r4.z, r4.y
+mov.f32f32 r2.y, c11.y
+mov.f32f32 r4.x, c11.x
+mad.f32 r1.x, r0.x, r1.z, r1.x
+mov.f32f32 r1.z, c11.z
+mad.f32 r2.y, c21.y, r0.z, r2.y
+mad.f32 r4.x, c21.x, r0.z, r4.x
+max.f r4.z, (neg)r1.x, c25.x
+max.f r1.x, r1.x, c25.x
+mad.f32 r0.z, c21.z, r0.z, r1.z
+mad.f32 r0.x, r0.x, r1.y, r0.w
+mul.f r0.y, r0.y, r4.y
+mov.f32f32 r0.w, c8.y
+mov.f32f32 r1.y, c8.x
+log2 r1.z, r4.z
+absneg.f r4.y, (neg)c10.x
+log2 r1.x, r1.x
+(ss)mov.f32f32 r4.z, c7.x
+max.f r5.x, (neg)r0.x, c25.x
+mov.f32f32 r5.y, c8.z
+min.f r4.y, (neg)r4.y, c25.z
+min.f r4.z, r4.z, c25.z
+mul.f r5.x, r5.x, r2.x
+mad.f32 r7.y, c18.z, r0.y, r5.y
+(ss)mul.f r1.z, r4.y, r1.z
+mul.f r1.x, r4.z, r1.x
+mad.f32 r0.z, c22.z, r5.x, r0.z
+mad.f32 r2.y, c22.y, r5.x, r2.y
+mad.f32 r4.x, c22.x, r5.x, r4.x
+max.f r8.x, r0.x, c25.x
+mad.f32 r0.w, c18.y, r0.y, r0.w
+exp2 r1.z, r1.z
+cmps.f.lt r4.y, (neg)r0.x, c24.x
+exp2 r1.x, r1.x
+cmps.f.lt r0.x, (neg)r0.x, c24.x
max.f r0.z, r0.z, c24.x
-max.f r1.y, r1.y, c24.x
-max.f r1.z, r1.z, c24.x
-mad.f32 r3.y, c18.z, r2.z, r3.y
-mad.f32 r5.x, c18.y, r2.z, r4.x
-log2 r2.x, r2.x
-(ss)mov.f32f32 r2.x, r2.x
-absneg.f r4.x, (neg)c10.x
-log2 r0.x, r0.x
-(ss)mov.f32f32 r0.x, r0.x
-mov.f32f32 r5.z, c7.x
+max.f r2.y, r2.y, c24.x
+(ss)sel.b32 r1.z, r1.z, r4.y, c25.x
+rcp r2.z, r2.z
+sel.b32 r0.x, r1.x, r0.x, c25.x
min.f r4.z, r0.z, c24.y
-min.f r0.z, (neg)r4.x, c25.z
-min.f r4.y, r1.y, c24.y
-min.f r4.x, r1.z, c24.y
-max.f r1.y, r0.w, c25.x
-mul.f r0.z, r0.z, r2.x
-min.f r1.z, r5.z, c25.z
-mov.f32f32 r2.x, c8.x
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r0.z, r0.z
-mul.f r0.x, r1.z, r0.x
-mad.f32 r1.z, c18.x, r2.z, r2.x
-mul.f r1.y, r1.y, r5.y
-rcp r2.x, r1.x
-(ss)cmps.f.lt r1.x, (neg)r0.w, c24.x
-cmps.f.lt r0.w, (neg)r0.w, c24.x
-mov.f32f32 r0.x, r0.x
-exp2 r0.z, r0.z
-(ss)mov.f32f32 r0.z, r0.z
-mad.f32 r2.z, c19.z, r1.y, r3.y
-mad.f32 r3.y, c19.y, r1.y, r5.x
-mad.f32 r1.y, c19.x, r1.y, r1.z
-sel.b32 r0.z, r0.z, r1.x, c25.x
-max.f r1.x, r2.z, c24.x
-max.f r2.z, r3.y, c24.x
-max.f r3.y, r1.y, c24.x
-mov.f32f32 r0.z, r0.z
-exp2 r0.x, r0.x
-(ss)mov.f32f32 r0.x, r0.x
-min.f r1.z, r1.x, c24.y
-min.f r1.y, r2.z, c24.y
-mov.f32f32 r0.z, r0.z
-sel.b32 r0.x, r0.x, r0.w, c25.x
-min.f r1.x, r3.y, c24.y
-mad.f32 r0.w, c3.w, r3.w, r0.y
-mul.f r0.y, r0.z, r2.x
-mov.f32f32 r0.z, c24.x
-mov.f32f32 r2.z, c24.x
-mov.f32f32 r3.y, c24.x
-mad.f32 r0.z, c23.z, r0.y, r0.z
-mad.f32 r2.z, c23.y, r0.y, r2.z
-mad.f32 r0.y, c23.x, r0.y, r3.y
-mov.f32f32 r0.x, r0.x
+min.f r4.y, r2.y, c24.y
+(ss)mul.f r0.z, r1.z, r2.z
+mov.f32f32 r1.x, c24.x
+mov.f32f32 r1.z, c24.x
+mov.f32f32 r2.y, c24.x
+mad.f32 r1.x, c23.z, r0.z, r1.x
+mad.f32 r1.z, c23.y, r0.z, r1.z
+mad.f32 r0.z, c23.x, r0.z, r2.y
+mul.f r0.x, r0.x, r2.z
+max.f r1.x, r1.x, c24.x
+max.f r1.z, r1.z, c24.x
+max.f r0.z, r0.z, c24.x
+mov.f32f32 r2.y, c24.x
+min.f r5.z, r1.x, c24.y
+min.f r5.y, r1.z, c24.y
+min.f r5.x, r0.z, c24.y
+mad.f32 r0.z, c20.z, r0.x, r2.y
+mov.f32f32 r1.x, c24.x
+mov.f32f32 r1.z, c24.x
+mad.f32 r1.x, c20.y, r0.x, r1.x
max.f r0.z, r0.z, c24.x
-max.f r2.z, r2.z, c24.x
-max.f r0.y, r0.y, c24.x
-mov.f32f32 r0.x, r0.x
-min.f r5.z, r0.z, c24.y
-min.f r5.y, r2.z, c24.y
-min.f r5.x, r0.y, c24.y
-mul.f r0.x, r0.x, r2.x
-mov.f32f32 r0.y, c24.x
-mov.f32f32 r0.z, c24.x
-mov.f32f32 r2.x, c24.x
-mad.f32 r0.y, c20.z, r0.x, r0.y
-mad.f32 r2.z, c20.y, r0.x, r0.z
-mad.f32 r0.x, c20.x, r0.x, r2.x
-mad.f32 r0.z, c3.z, r3.w, r2.y
-max.f r0.y, r0.y, c24.x
-max.f r2.x, r2.z, c24.x
+mad.f32 r0.x, c20.x, r0.x, r1.z
+max.f r1.z, r4.x, c24.x
+mul.f r8.x, r8.x, r2.x
+(ss)min.f r2.z, r0.z, c24.y
+max.f r0.z, r1.x, c24.x
max.f r0.x, r0.x, c24.x
-nop
-min.f r2.z, r0.y, c24.y
-min.f r2.y, r2.x, c24.y
+(rpt1)nop
+min.f r2.y, r0.z, c24.y
min.f r2.x, r0.x, c24.y
-mad.f32 r0.x, c2.y, r3.z, r7.y
-mad.f32 r3.x, c2.x, r3.z, r3.x
-mad.f32 r0.y, c3.y, r3.w, r0.x
-mad.f32 r0.x, c3.x, r3.w, r3.x
-mov.f32f32 r3.w, r6.w
-mov.f32f32 r3.z, r6.z
-mov.f32f32 r3.y, r6.y
-mov.f32f32 r3.x, r6.x
+min.f r4.x, r1.z, c24.y
+mad.f32 r0.x, c19.z, r8.x, r7.y
+mad.f32 r0.z, c19.y, r8.x, r0.w
+mad.f32 r0.y, c18.x, r0.y, r1.y
+mad.f32 r0.w, c3.w, r6.w, r7.x
+max.f r0.x, r0.x, c24.x
+max.f r1.x, r0.z, c24.x
+mad.f32 r0.y, c19.x, r8.x, r0.y
+mad.f32 r0.z, c3.z, r6.w, r7.z
+min.f r1.z, r0.x, c24.y
+min.f r1.y, r1.x, c24.y
+max.f r0.x, r0.y, c24.x
+mad.f32 r0.y, c2.y, r6.z, r7.w
+mul.f r6.x, r6.x, c0.x
+nop
+min.f r1.x, r0.x, c24.y
+mad.f32 r0.y, c3.y, r6.w, r0.y
+mad.f32 r0.x, c1.x, r6.y, r6.x
+nop
+mad.f32 r0.x, c2.x, r6.z, r0.x
+nop
+mad.f32 r0.x, c3.x, r6.w, r0.x
end
nop
; VERT: outputs: r0.x (0:0) r1.x (1:0) r2.x (1:1) r3.x (5:0) r4.x (2:0) r5.x (2:1)
-; VERT: inputs: r3.x (0:0,cm=f,il=8,b=0) r0.x (0:0,cm=7,il=12,b=0) r6.x (0:0,cm=f,il=16,b=0)
-; VERT: 203 instructions, 0 half, 8 full
+; VERT: inputs: r6.x (0:0,cm=f,il=8,b=0) r0.x (0:0,cm=7,il=12,b=0) r3.x (0:0,cm=f,il=16,b=0)
+; VERT: 176 instructions, 0 half, 9 full