summaryrefslogtreecommitdiff
path: root/reference/maniadrive/maniadrive-10.asm
diff options
context:
space:
mode:
Diffstat (limited to 'reference/maniadrive/maniadrive-10.asm')
-rw-r--r--reference/maniadrive/maniadrive-10.asm360
1 files changed, 167 insertions, 193 deletions
diff --git a/reference/maniadrive/maniadrive-10.asm b/reference/maniadrive/maniadrive-10.asm
index cdcbf04..0babf82 100644
--- a/reference/maniadrive/maniadrive-10.asm
+++ b/reference/maniadrive/maniadrive-10.asm
@@ -4,10 +4,10 @@
@in(r6.y) in1
@in(r6.z) in2
@in(r6.w) in3
-@in(r7.x) in4
-@in(r7.y) in5
-@in(r7.z) in6
-@in(r7.w) in7
+@in(r3.x) in4
+@in(r3.y) in5
+@in(r3.z) in6
+@in(r3.w) in7
@out(r0.x) out0
@out(r0.y) out1
@out(r0.z) out2
@@ -32,219 +32,193 @@
@out(r5.y) out21
@out(r5.z) out22
@out(r5.w) out23
+@const(c25.x) 0x00000000, 0x3f800000, 0x00000000, 0x00000000
+@const(c26.x) 0x00000000, 0x3f800000, 0x43000000, 0x00000000
(sy)(ss)mul.f r0.x, r6.x, c14.x
-mov.f32f32 r0.y, c25.y
+mul.f r0.y, r6.x, c14.y
mad.f32 r0.x, c15.x, r6.y, r0.x
-mov.f32f32 r0.z, c12.w
+mad.f32 r0.y, c15.y, r6.y, r0.y
mad.f32 r0.x, c16.x, r6.z, r0.x
-mov.f32f32 r0.y, r0.y
+mad.f32 r0.y, c16.y, r6.z, r0.y
mad.f32 r0.x, c17.x, r6.w, r0.x
-mov.f32f32 r0.z, r0.z
+mad.f32 r0.y, c17.y, r6.w, r0.y
+mul.f r0.z, r6.x, c14.z
mov.f32f32 r0.w, c25.y
-mov.f32f32 r1.x, c9.w
add.f r0.x, c13.x, (neg)r0.x
-max.f r0.y, r0.y, c25.x
-max.f r0.z, r0.z, c25.x
-mov.f32f32 r0.w, r0.w
-mul.f r1.y, r0.x, r0.x
-mul.f r1.z, r6.x, c14.y
-min.f r5.w, r0.y, c25.y
-mad.f32 r0.y, c15.y, r6.y, r1.z
-min.f r4.w, r0.z, c25.y
-mad.f32 r0.y, c16.y, r6.z, r0.y
-max.f r0.z, r0.w, c25.x
-mad.f32 r0.y, c17.y, r6.w, r0.y
-mov.f32f32 r0.w, r1.x
-mov.f32f32 r1.x, c4.x
-mov.f32f32 r1.z, c4.x
add.f r0.y, c13.y, (neg)r0.y
-min.f r2.w, r0.z, c25.y
-max.f r0.z, r0.w, c25.x
-mul.f r0.w, r1.x, c5.x
-mad.f32 r1.x, r0.y, r0.y, r1.y
-mov.f32f32 r1.y, c5.y
-mul.f r1.z, r1.z, c6.x
-mov.f32f32 r2.x, c4.x
-mov.f32f32 r1.x, r1.x
-mul.f r2.y, r6.x, c14.z
-min.f r1.w, r0.z, c25.y
-mad.f32 r0.z, c15.z, r6.y, r2.y
-mad.f32 r0.w, c4.y, r1.y, r0.w
+mad.f32 r0.z, c15.z, r6.y, r0.z
+max.f r0.w, r0.w, c25.x
+mul.f r1.x, r0.x, r0.x
mad.f32 r0.z, c16.z, r6.z, r0.z
-mov.f32f32 r1.y, c6.y
+mad.f32 r1.x, r0.y, r0.y, r1.x
mad.f32 r0.z, c17.z, r6.w, r0.z
-mov.f32f32 r0.w, r0.w
-mul.f r2.x, r2.x, c7.x
-mad.f32 r1.y, c4.y, r1.y, r1.z
+min.f r5.w, r0.w, c25.y
+mov.f32f32 r0.w, c12.w
+mov.f32f32 r1.y, c25.y
add.f r0.z, c13.z, (neg)r0.z
-mov.f32f32 r1.z, c5.z
-mov.f32f32 r2.y, c7.y
-mov.f32f32 r1.y, r1.y
+mov.f32f32 r1.z, c9.w
+mov.f32f32 r1.w, c4.x
+mov.f32f32 r2.x, c4.x
mad.f32 r1.x, r0.z, r0.z, r1.x
-mad.f32 r0.w, c4.z, r1.z, r0.w
-mov.f32f32 r1.z, c6.z
-mad.f32 r2.x, c4.y, r2.y, r2.x
-mul.f r2.y, r6.x, c0.w
-mul.f r2.z, r6.x, c0.z
-mul.f r3.x, r6.x, c0.y
+max.f r0.w, r0.w, c25.x
+max.f r1.y, r1.y, c25.x
+max.f r1.z, r1.z, c25.x
+mul.f r1.w, r1.w, c5.x
+mul.f r2.x, r2.x, c6.x
+mov.f32f32 r2.y, c4.x
rsq r1.x, (abs)r1.x
(ss)mul.f r0.x, r0.x, r1.x
-rcp r3.y, r1.x
-(ss)mov.f32f32 r3.y, r3.y
-rcp r3.z, r1.x
-(ss)mov.f32f32 r3.z, r3.z
-rsq r3.w, (abs)r1.x
-(ss)mul.f r0.y, r0.y, r3.w
-add.f r3.w, r0.x, c25.x
-mul.f r4.x, r1.x, r3.y
-mul.f r0.x, r0.w, r0.x
-mad.f32 r1.y, c4.z, r1.z, r1.y
-mul.f r1.z, r3.w, r3.w
-add.f r4.y, r0.y, c25.x
-mov.f32f32 r4.x, r4.x
-mad.f32 r0.x, r1.y, r0.y, r0.x
-mul.f r0.y, r3.z, r3.y
-mad.f32 r1.z, r4.y, r4.y, r1.z
-mul.f r3.z, c18.x, r4.x
-mov.f32f32 r0.x, r0.x
-(ss)rsq r1.x, (abs)r1.x
-mad.f32 r3.y, c18.y, r3.y, r3.z
-mov.f32f32 r1.z, r1.z
-(ss)mul.f r0.z, r0.z, r1.x
-(ss)mov.f32f32 r1.x, r2.x
-mov.f32f32 r2.x, c7.z
-mov.f32f32 r3.y, r3.y
-add.f r3.z, r0.z, c25.y
-mov.f32f32 r0.y, r0.y
-mad.f32 r1.x, c4.z, r2.x, r1.x
-mad.f32 r2.x, c1.w, r6.y, r2.y
-mad.f32 r1.z, r3.z, r3.z, r1.z
-mad.f32 r0.y, c18.z, r0.y, r3.y
-mad.f32 r0.x, r1.x, r0.z, r0.x
-mad.f32 r0.z, c2.w, r6.z, r2.x
-mad.f32 r3.y, c1.z, r6.y, r2.z
-mad.f32 r3.x, c1.y, r6.y, r3.x
-mul.f r6.x, r6.x, c0.x
-rsq r1.z, (abs)r1.z
-(ss)mov.f32f32 r1.z, r1.z
-mov.f32f32 r0.y, r0.y
-max.f r2.x, (neg)r0.x, c26.x
-max.f r2.y, r0.x, c26.x
-mul.f r2.z, r3.w, r1.z
-mul.f r3.w, r4.y, r1.z
-mul.f r1.z, r3.z, r1.z
-mov.f32f32 r2.x, r2.x
-mul.f r0.w, r0.w, r2.z
-rcp r2.z, r0.y
-mov.f32f32 r3.z, c26.y
-mad.f32 r0.w, r1.y, r3.w, r0.w
-mov.f32f32 r1.y, c26.y
-rcp r3.w, r0.y
-mov.f32f32 r2.y, r2.y
-(ss)mul.f r3.z, r3.z, r2.z
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r4.x, c12.z
-mad.f32 r0.w, r1.x, r1.z, r0.w
-mad.f32 r1.x, c22.z, r3.z, r4.x
-mov.f32f32 r1.z, c12.y
-mov.f32f32 r4.x, c12.x
-max.f r4.y, (neg)r0.w, c26.x
+rcp r2.z, r1.x
+(ss)mov.f32f32 r2.w, r2.z
+mov.f32f32 r4.x, c5.y
+rsq r4.y, (abs)r1.x
+(ss)mul.f r0.y, r0.y, r4.y
+add.f r4.y, r0.x, c25.x
+mul.f r4.z, r1.x, r2.w
+mad.f32 r1.w, c4.y, r4.x, r1.w
+mov.f32f32 r4.x, c5.z
+mul.f r4.w, r4.y, r4.y
+add.f r5.x, r0.y, c25.x
+mul.f r4.z, c18.x, r4.z
+mad.f32 r4.x, c4.z, r4.x, r1.w
+mad.f32 r1.w, c18.y, r2.w, r4.z
+mad.f32 r4.z, r5.x, r5.x, r4.w
+rsq r2.w, (abs)r1.x
+(ss)mul.f r0.z, r0.z, r2.w
+(ss)rcp r1.x, r1.x
+(ss)mul.f r1.x, r1.x, r2.z
+mul.f r0.x, r4.x, r0.x
+mov.f32f32 r2.z, c6.y
+add.f r5.y, r0.z, c25.y
+mad.f32 r1.x, c18.z, r1.x, r1.w
+min.f r4.w, r0.w, c25.y
+min.f r2.w, r1.y, c25.y
+mad.f32 r0.w, r5.y, r5.y, r4.z
+mov.f32f32 r1.y, r1.x
+mad.f32 r2.x, c4.y, r2.z, r2.x
+mov.f32f32 r2.z, c6.z
+min.f r1.w, r1.z, c25.y
+mul.f r1.z, r2.y, c7.x
+mov.f32f32 r2.y, c7.y
+rsq r0.w, (abs)r0.w
+(ss)mov.f32f32 r4.z, r0.w
+rcp r5.z, r1.y
+mov.f32f32 r7.x, c26.y
+mov.f32f32 r7.y, c26.y
+(ss)rcp r1.y, r1.y
+nop
+rcp r1.x, r1.x
+mul.f r4.y, r4.y, r4.z
+(ss)mul.f r7.x, r7.x, r5.z
+mov.f32f32 r7.z, c12.y
+mov.f32f32 r7.w, c12.x
+mul.f r4.x, r4.x, r4.y
+mul.f r4.y, r5.x, r4.z
+mad.f32 r2.x, c4.z, r2.z, r2.x
+mov.f32f32 r2.z, c12.z
+mad.f32 r4.z, c22.y, r7.x, r7.z
+mad.f32 r5.x, c22.x, r7.x, r7.w
+mad.f32 r4.x, r2.x, r4.y, r4.x
+mul.f r0.w, r5.y, r0.w
+mad.f32 r1.z, c4.y, r2.y, r1.z
+mov.f32f32 r2.y, c7.z
+mad.f32 r2.z, c22.z, r7.x, r2.z
+mad.f32 r0.x, r2.x, r0.y, r0.x
+mul.f r0.y, r7.y, r5.z
+mad.f32 r1.z, c4.z, r2.y, r1.z
+mov.f32f32 r2.x, c9.z
+mov.f32f32 r2.y, c9.y
+mov.f32f32 r4.y, c9.x
+mad.f32 r0.w, r1.z, r0.w, r4.x
+mad.f32 r0.x, r1.z, r0.z, r0.x
+mad.f32 r0.z, c19.z, r0.y, r2.x
+mad.f32 r1.z, c19.y, r0.y, r2.y
+max.f r2.x, (neg)r0.w, c26.x
max.f r0.w, r0.w, c26.x
-mul.f r2.x, r2.x, r3.w
-mad.f32 r1.z, c22.y, r3.z, r1.z
-mad.f32 r3.z, c22.x, r3.z, r4.x
-mul.f r1.y, r1.y, r2.z
-mov.f32f32 r2.z, c9.y
-log2 r4.x, r4.y
-(ss)mov.f32f32 r4.x, r4.x
-(ss)absneg.f r4.y, (neg)c11.x
-log2 r0.w, r0.w
-(ss)mov.f32f32 r0.w, r0.w
-mov.f32f32 r4.z, c8.x
-mad.f32 r1.x, c23.z, r2.x, r1.x
-min.f r4.y, (neg)r4.y, c26.z
-mad.f32 r1.z, c23.y, r2.x, r1.z
-mad.f32 r2.x, c23.x, r2.x, r3.z
-mov.f32f32 r3.z, c9.z
-mul.f r4.x, r4.y, r4.x
-min.f r4.y, r4.z, c26.z
-max.f r1.x, r1.x, c25.x
-max.f r1.z, r1.z, c25.x
-mov.f32f32 r4.x, r4.x
-mul.f r0.w, r4.y, r0.w
-min.f r4.z, r1.x, c25.y
-min.f r4.y, r1.z, c25.y
-max.f r1.x, r2.x, c25.x
-mad.f32 r1.z, c19.z, r1.y, r3.z
-mad.f32 r2.x, c19.y, r1.y, r2.z
-exp2 r2.z, r4.x
-(ss)mov.f32f32 r2.z, r2.z
-cmps.f.lt r3.z, (neg)r0.x, c25.x
-mov.f32f32 r0.w, r0.w
-(ss)min.f r4.x, r1.x, c25.y
-mul.f r1.x, r2.y, r3.w
-sel.b32 r2.y, r2.z, r3.z, c26.x
-mov.f32f32 r2.z, c9.x
-rcp r0.y, r0.y
+max.f r2.y, (neg)r0.x, c26.x
+max.f r4.x, r0.x, c26.x
+mad.f32 r0.y, c19.x, r0.y, r4.y
+cmps.f.lt r4.y, (neg)r0.x, c25.x
cmps.f.lt r0.x, (neg)r0.x, c25.x
-mad.f32 r1.z, c20.z, r1.x, r1.z
-mov.f32f32 r2.y, r2.y
+log2 r2.x, r2.x
+absneg.f r5.y, (neg)c11.x
+log2 r0.w, r0.w
+mov.f32f32 r5.z, c8.x
+mul.f r2.y, r2.y, r1.x
+(ss)mul.f r1.x, r4.x, r1.x
+min.f r4.x, (neg)r5.y, c26.z
+min.f r5.y, r5.z, c26.z
+mad.f32 r2.z, c23.z, r2.y, r2.z
+mad.f32 r4.z, c23.y, r2.y, r4.z
+(ss)mul.f r2.x, r4.x, r2.x
+mul.f r0.w, r5.y, r0.w
+max.f r2.z, r2.z, c25.x
+max.f r4.x, r4.z, c25.x
+mad.f32 r2.y, c23.x, r2.y, r5.x
+mad.f32 r0.z, c20.z, r1.x, r0.z
+mad.f32 r7.x, c20.y, r1.x, r1.z
+exp2 r1.z, r2.x
+(ss)sel.b32 r1.z, r1.z, r4.y, c26.x
exp2 r0.w, r0.w
-(ss)mov.f32f32 r0.w, r0.w
-mad.f32 r2.x, c20.y, r1.x, r2.x
-max.f r1.z, r1.z, c25.x
-mov.f32f32 r2.y, r2.y
-sel.b32 r0.x, r0.w, r0.x, c26.x
-max.f r0.w, r2.x, c25.x
-mad.f32 r3.z, c19.x, r1.y, r2.z
-mul.f r1.y, r2.y, r0.y
+(ss)sel.b32 r0.x, r0.w, r0.x, c26.x
+min.f r4.z, r2.z, c25.y
+min.f r4.y, r4.x, c25.y
+(ss)mul.f r0.w, r1.z, r1.y
+mov.f32f32 r1.z, c25.x
mov.f32f32 r2.x, c25.x
-mov.f32f32 r2.y, c25.x
mov.f32f32 r2.z, c25.x
-mad.f32 r2.x, c24.z, r1.y, r2.x
-mad.f32 r2.y, c24.y, r1.y, r2.y
-mad.f32 r1.y, c24.x, r1.y, r2.z
-mov.f32f32 r0.x, r0.x
-max.f r2.x, r2.x, c25.x
-max.f r2.y, r2.y, c25.x
-max.f r1.y, r1.y, c25.x
-mov.f32f32 r0.x, r0.x
-min.f r5.z, r2.x, c25.y
-min.f r5.y, r2.y, c25.y
-min.f r5.x, r1.y, c25.y
-mul.f r0.x, r0.x, r0.y
-mov.f32f32 r0.y, c25.x
-mov.f32f32 r1.y, c25.x
+mad.f32 r1.z, c24.z, r0.w, r1.z
+mad.f32 r2.x, c24.y, r0.w, r2.x
+mad.f32 r0.w, c24.x, r0.w, r2.z
+mul.f r0.x, r0.x, r1.y
+max.f r1.y, r1.z, c25.x
+max.f r1.z, r2.x, c25.x
+max.f r0.w, r0.w, c25.x
mov.f32f32 r2.x, c25.x
-mad.f32 r0.y, c21.z, r0.x, r0.y
+min.f r5.z, r1.y, c25.y
+min.f r5.y, r1.z, c25.y
+min.f r5.x, r0.w, c25.y
+mad.f32 r0.w, c21.z, r0.x, r2.x
+mov.f32f32 r1.y, c25.x
+mov.f32f32 r1.z, c25.x
mad.f32 r1.y, c21.y, r0.x, r1.y
-mad.f32 r0.x, c21.x, r0.x, r2.x
-min.f r1.z, r1.z, c25.y
-max.f r0.y, r0.y, c25.x
-max.f r1.y, r1.y, c25.x
+max.f r0.w, r0.w, c25.x
+mad.f32 r0.x, c21.x, r0.x, r1.z
+max.f r1.z, r2.y, c25.x
+max.f r0.z, r0.z, c25.x
+min.f r2.z, r0.w, c25.y
+max.f r0.w, r1.y, c25.x
max.f r0.x, r0.x, c25.x
-nop
-min.f r2.z, r0.y, c25.y
-min.f r2.y, r1.y, c25.y
+(rpt1)nop
+min.f r2.y, r0.w, c25.y
min.f r2.x, r0.x, c25.y
-min.f r1.y, r0.w, c25.y
-mad.f32 r0.x, c20.x, r1.x, r3.z
-mad.f32 r0.w, c3.w, r6.w, r0.z
-mad.f32 r0.y, c2.z, r6.z, r3.y
-mad.f32 r1.x, c2.y, r6.z, r3.x
-max.f r0.x, r0.x, c25.x
-mad.f32 r0.z, c3.z, r6.w, r0.y
-mad.f32 r0.y, c3.y, r6.w, r1.x
-mad.f32 r3.x, c1.x, r6.y, r6.x
+min.f r4.x, r1.z, c25.y
+min.f r1.z, r0.z, c25.y
+max.f r0.x, r7.x, c25.x
+mad.f32 r0.y, c20.x, r1.x, r0.y
+mul.f r0.z, r6.x, c0.w
+mul.f r0.w, r6.x, c0.z
+min.f r1.y, r0.x, c25.y
+max.f r0.x, r0.y, c25.x
+mad.f32 r0.y, c1.w, r6.y, r0.z
+mad.f32 r0.z, c1.z, r6.y, r0.w
+mul.f r7.x, r6.x, c0.y
min.f r1.x, r0.x, c25.y
-mad.f32 r0.x, c2.x, r6.z, r3.x
-mov.f32f32 r3.w, r7.w
+mad.f32 r0.x, c2.w, r6.z, r0.y
+mad.f32 r0.y, c2.z, r6.z, r0.z
+mad.f32 r0.w, c3.w, r6.w, r0.x
+mad.f32 r0.z, c3.z, r6.w, r0.y
+mad.f32 r0.x, c1.y, r6.y, r7.x
+mul.f r0.y, r6.x, c0.x
+mad.f32 r0.x, c2.y, r6.z, r0.x
+mad.f32 r6.x, c1.x, r6.y, r0.y
+mad.f32 r0.y, c3.y, r6.w, r0.x
+mad.f32 r0.x, c2.x, r6.z, r6.x
+nop
mad.f32 r0.x, c3.x, r6.w, r0.x
-mov.f32f32 r3.z, r7.z
-mov.f32f32 r3.y, r7.y
-mov.f32f32 r3.x, r7.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: r6.x (0:0,cm=f,il=8,b=0) r7.x (0:0,cm=f,il=12,b=0)
-; VERT: 212 instructions, 0 half, 8 full
+; VERT: inputs: r6.x (0:0,cm=f,il=8,b=0) r3.x (0:0,cm=f,il=12,b=0)
+; VERT: 184 instructions, 0 half, 8 full