summaryrefslogtreecommitdiff
path: root/reference/stk/stk0601.asm
diff options
context:
space:
mode:
Diffstat (limited to 'reference/stk/stk0601.asm')
-rw-r--r--reference/stk/stk0601.asm239
1 files changed, 112 insertions, 127 deletions
diff --git a/reference/stk/stk0601.asm b/reference/stk/stk0601.asm
index aadaff7..5f20141 100644
--- a/reference/stk/stk0601.asm
+++ b/reference/stk/stk0601.asm
@@ -1,20 +1,20 @@
; options:
; VERT: new compiler
-@in(r2.x) in0
-@in(r2.y) in1
-@in(r2.z) in2
-@in(r2.w) in3
+@in(r3.x) in0
+@in(r3.y) in1
+@in(r3.z) in2
+@in(r3.w) in3
@in(r0.x) in4
@in(r0.y) in5
@in(r0.z) in6
-@in(r0.w) in8
-@in(r1.x) in9
-@in(r1.y) in10
-@in(r1.z) in11
-@in(r3.x) in12
-@in(r3.y) in13
-@in(r3.z) in14
-@in(r3.w) in15
+@in(r1.x) in8
+@in(r1.y) in9
+@in(r1.z) in10
+@in(r1.w) in11
+@in(r2.x) in12
+@in(r2.y) in13
+@in(r2.z) in14
+@in(r2.w) in15
@out(r0.x) out0
@out(r0.y) out1
@out(r0.z) out2
@@ -27,135 +27,120 @@
@out(r2.y) out9
@out(r2.z) out10
@out(r2.w) out11
-(sy)(ss)mul.f r1.w, r2.x, c13.x
+@const(c20.x) 0x00000000, 0x3f800000, 0x43000000, 0x00000000
+(sy)(ss)mul.f r0.w, r3.x, c13.x
mul.f r4.x, r0.x, c4.x
-mad.f32 r1.w, c14.x, r2.y, r1.w
+mad.f32 r0.w, c14.x, r3.y, r0.w
mad.f32 r4.x, c4.y, r0.y, r4.x
-mad.f32 r1.w, c15.x, r2.z, r1.w
-mov.f32f32 r1.z, r1.z
-mad.f32 r4.y, c16.x, r2.w, r1.w
-mov.f32f32 r1.w, r4.x
-mov.f32f32 r4.x, c10.z
+mad.f32 r0.w, c15.x, r3.z, r0.w
+mad.f32 r4.x, c4.z, r0.z, r4.x
+mad.f32 r0.w, c16.x, r3.w, r0.w
+mov.f32f32 r4.y, c10.z
mov.f32f32 r4.z, c10.y
-mul.f r4.w, r4.y, r4.y
-mul.f r5.x, r2.x, c13.y
-mad.f32 r5.y, c4.z, r0.z, r1.w
-mad.f32 r1.w, c14.y, r2.y, r5.x
-mov.f32f32 r1.z, r1.z
-mad.f32 r1.w, c15.y, r2.z, r1.w
-mul.f r5.x, r5.y, c11.x
-mad.f32 r5.z, c16.y, r2.w, r1.w
-mul.f r1.w, r0.x, c5.x
-max.f r1.z, r1.z, c20.x
-mad.f32 r4.x, c8.z, r4.x, c9.z
-mad.f32 r4.w, r5.z, r5.z, r4.w
-mad.f32 r5.w, c5.y, r0.y, r1.w
-min.f r1.w, r1.z, c20.y
-add.f r1.z, c17.z, r4.x
-mov.f32f32 r4.x, r4.w
-mul.f r4.w, r2.x, c13.z
-mov.f32f32 r5.w, r5.w
-mad.f32 r4.w, c14.z, r2.y, r4.w
+mov.f32f32 r4.w, c10.x
+mul.f r5.x, r0.w, r0.w
+mul.f r5.y, r3.x, c13.y
+mul.f r5.z, r4.x, c11.x
+mad.f32 r5.y, c14.y, r3.y, r5.y
+mul.f r5.w, r0.x, c5.x
+mad.f32 r5.y, c15.y, r3.z, r5.y
+mad.f32 r5.w, c5.y, r0.y, r5.w
+mad.f32 r5.y, c16.y, r3.w, r5.y
mad.f32 r5.w, c5.z, r0.z, r5.w
-mad.f32 r4.w, c15.z, r2.z, r4.w
+mad.f32 r4.y, c8.z, r4.y, c9.z
mad.f32 r4.z, c8.y, r4.z, c9.y
-mad.f32 r4.w, c16.z, r2.w, r4.w
-mad.f32 r5.x, c11.y, r5.w, r5.x
-mov.f32f32 r6.x, c10.x
+mad.f32 r5.x, r5.y, r5.y, r5.x
+mul.f r6.x, r3.x, c13.z
+mad.f32 r5.z, c11.y, r5.w, r5.z
+mad.f32 r6.x, c14.z, r3.y, r6.x
mul.f r0.x, r0.x, c6.x
-mad.f32 r4.x, r4.w, r4.w, r4.x
-mov.f32f32 r5.x, r5.x
+mad.f32 r6.x, c15.z, r3.z, r6.x
mad.f32 r0.x, c6.y, r0.y, r0.x
-add.f r0.y, c17.y, r4.z
-mad.f32 r4.z, c8.x, r6.x, c9.x
-mul.f r6.x, r2.x, c0.w
-mul.f r6.y, r2.x, c0.z
-rsq r4.x, (abs)r4.x
-(ss)mov.f32f32 r4.x, r4.x
-mov.f32f32 r0.x, r0.x
-add.f r4.z, c17.x, r4.z
+mad.f32 r0.y, c16.z, r3.w, r6.x
mad.f32 r0.x, c6.z, r0.z, r0.x
-mul.f r0.z, r4.y, r4.x
-mul.f r4.y, r5.z, r4.x
-mul.f r4.x, r4.w, r4.x
-mad.f32 r4.w, c11.z, r0.x, r5.x
-add.f r0.z, c11.x, (neg)r0.z
-add.f r4.y, c11.y, (neg)r4.y
-add.f r4.x, c11.z, (neg)r4.x
-max.f r5.x, r4.w, c20.x
-mul.f r5.z, r0.z, r0.z
-cmps.f.lt r4.w, (neg)r4.w, c20.x
-mad.f32 r5.z, r4.y, r4.y, r5.z
-mov.f32f32 r5.x, r5.x
-mul.f r1.x, c18.y, r1.x
-mul.f r0.w, c18.x, r0.w
-mov.f32f32 r5.z, r5.z
-mul.f r1.y, c18.z, r1.y
-mad.f32 r5.z, r4.x, r4.x, r5.z
-mad.f32 r1.x, r5.x, r1.x, r0.y
-mad.f32 r4.z, r5.x, r0.w, r4.z
-mad.f32 r0.y, c1.w, r2.y, r6.x
-mad.f32 r0.w, c1.z, r2.y, r6.y
-mul.f r6.x, r2.x, c0.y
-mul.f r2.x, r2.x, c0.x
-rsq r5.z, (abs)r5.z
-(ss)mov.f32f32 r5.z, r5.z
-mad.f32 r1.y, r5.x, r1.y, r1.z
-mad.f32 r0.y, c2.w, r2.z, r0.y
-mad.f32 r1.z, c2.z, r2.z, r0.w
-mul.f r0.z, r0.z, r5.z
-mul.f r4.y, r4.y, r5.z
-mul.f r4.x, r4.x, r5.z
-mad.f32 r0.w, c3.w, r2.w, r0.y
-mul.f r0.y, r5.y, r0.z
-mad.f32 r0.z, c3.z, r2.w, r1.z
-mad.f32 r0.y, r5.w, r4.y, r0.y
-mad.f32 r1.z, c1.y, r2.y, r6.x
-mad.f32 r2.x, c1.x, r2.y, r2.x
-mov.f32f32 r2.y, c7.x
-mov.f32f32 r0.y, r0.y
-mad.f32 r1.z, c2.y, r2.z, r1.z
-mad.f32 r0.x, r0.x, r4.x, r0.y
-mad.f32 r0.y, c3.y, r2.w, r1.z
-mad.f32 r1.z, c2.x, r2.z, r2.x
-min.f r4.x, r2.y, c20.z
-max.f r4.y, r0.x, c20.x
-mad.f32 r0.x, c3.x, r2.w, r1.z
-mov.f32f32 r2.w, r3.w
-mov.f32f32 r2.z, r3.z
-mov.f32f32 r2.y, r3.y
-mov.f32f32 r2.x, r3.x
-nop
-log2 r1.z, r4.y
-(ss)mov.f32f32 r1.z, r1.z
-(rpt2)nop
-mul.f r1.z, r4.x, r1.z
+add.f r0.z, c17.z, r4.y
+add.f r4.y, c17.y, r4.z
+mad.f32 r4.z, r0.y, r0.y, r5.x
+mad.f32 r5.x, c11.z, r0.x, r5.z
+mad.f32 r4.w, c8.x, r4.w, c9.x
+mul.f r5.z, r3.x, c0.w
+mul.f r6.x, r3.x, c0.z
+mul.f r6.y, r3.x, c0.y
+mul.f r3.x, r3.x, c0.x
+rsq r4.z, (abs)r4.z
+(ss)mov.f32f32 r6.z, r4.z
+mul.f r0.y, r0.y, r4.z
+(ss)max.f r4.z, r5.x, c20.x
+add.f r4.w, c17.x, r4.w
+mul.f r0.w, r0.w, r6.z
+mul.f r5.y, r5.y, r6.z
+(rpt1)nop
+add.f r0.w, c11.x, (neg)r0.w
+add.f r5.y, c11.y, (neg)r5.y
+add.f r0.y, c11.z, (neg)r0.y
+mov.f32f32 r6.z, r4.z
+mul.f r6.w, r0.w, r0.w
+mul.f r1.z, c18.z, r1.z
+mad.f32 r6.w, r5.y, r5.y, r6.w
+mul.f r1.y, c18.y, r1.y
+mad.f32 r6.w, r0.y, r0.y, r6.w
+mad.f32 r1.z, r6.z, r1.z, r0.z
+mul.f r0.z, c18.x, r1.x
+cmps.f.lt r1.x, (neg)r5.x, c20.x
+mad.f32 r5.x, c1.w, r3.y, r5.z
+mad.f32 r5.z, c1.z, r3.y, r6.x
+mad.f32 r6.x, c1.y, r3.y, r6.y
+rsq r6.y, (abs)r6.w
+(ss)mov.f32f32 r6.w, r6.y
+mul.f r0.y, r0.y, r6.y
+mad.f32 r1.y, r6.z, r1.y, r4.y
+mad.f32 r4.y, r4.z, r0.z, r4.w
+mul.f r0.z, r0.w, r6.w
+mul.f r4.z, r5.y, r6.w
+mad.f32 r0.w, c2.w, r3.z, r5.x
+mad.f32 r4.w, c2.z, r3.z, r5.z
+mul.f r0.z, r4.x, r0.z
+mad.f32 r0.w, c3.w, r3.w, r0.w
+mad.f32 r4.x, r5.w, r4.z, r0.z
+mad.f32 r0.z, c3.z, r3.w, r4.w
+mad.f32 r0.x, r0.x, r0.y, r4.x
+mad.f32 r0.y, c2.y, r3.z, r6.x
+mad.f32 r3.x, c1.x, r3.y, r3.x
+max.f r1.w, r1.w, c20.x
+max.f r0.x, r0.x, c20.x
+mad.f32 r0.y, c3.y, r3.w, r0.y
+mad.f32 r3.x, c2.x, r3.z, r3.x
+min.f r1.w, r1.w, c20.y
+mov.f32f32 r3.y, c7.x
+(rpt1)nop
+log2 r3.z, r0.x
+(ss)mad.f32 r0.x, c3.x, r3.w, r3.x
+min.f r3.x, r3.y, c20.z
(rpt2)nop
-mov.f32f32 r1.z, r1.z
+(ss)mul.f r3.x, r3.x, r3.z
(rpt5)nop
-exp2 r1.z, r1.z
-(ss)mov.f32f32 r1.z, r1.z
-(rpt2)nop
-sel.b32 r1.z, r1.z, r4.w, c20.x
-(rpt2)nop
-mov.f32f32 r1.z, r1.z
-(rpt2)nop
-mov.f32f32 r1.z, r1.z
+exp2 r3.x, r3.x
+(ss)sel.b32 r1.x, r3.x, r1.x, c20.x
(rpt2)nop
-mad.f32 r1.y, c19.z, r1.z, r1.y
-mad.f32 r1.x, c19.y, r1.z, r1.x
-mad.f32 r1.z, c19.x, r1.z, r4.z
-nop
-max.f r1.y, r1.y, c20.x
+(ss)mov.f32f32 r3.x, r1.x
+mad.f32 r1.x, c19.x, r1.x, r4.y
+(rpt1)nop
+mov.f32f32 r3.x, r3.x
max.f r1.x, r1.x, c20.x
-max.f r3.x, r1.z, c20.x
+(rpt1)nop
+mad.f32 r1.z, c19.z, r3.x, r1.z
+mad.f32 r1.y, c19.y, r3.x, r1.y
+min.f r1.x, r1.x, c20.y
nop
-min.f r1.z, r1.y, c20.y
-min.f r1.y, r1.x, c20.y
-min.f r1.x, r3.x, c20.y
+max.f r1.z, r1.z, c20.x
+max.f r1.y, r1.y, c20.x
+(rpt1)nop
+min.f r1.z, r1.z, c20.y
+min.f r1.y, r1.y, c20.y
end
nop
+nop
; VERT: outputs: r0.x (0:0) r1.x (1:0) r2.x (5:0)
-; VERT: inputs: r2.x (0:0,cm=f,il=8,b=0) r0.x (0:0,cm=7,il=12,b=0) r0.w (0:0,cm=f,il=16,b=0) r3.x (0:0,cm=f,il=20,b=0)
-; VERT: 144 instructions, 0 half, 7 full
+; VERT: inputs: r3.x (0:0,cm=f,il=8,b=0) r0.x (0:0,cm=7,il=12,b=0) r1.x (0:0,cm=f,il=16,b=0) r2.x (0:0,cm=f,il=20,b=0)
+; VERT: 124 instructions, 0 half, 7 full