summaryrefslogtreecommitdiff
path: root/reference/bump/bump-12.asm
diff options
context:
space:
mode:
Diffstat (limited to 'reference/bump/bump-12.asm')
-rw-r--r--reference/bump/bump-12.asm258
1 files changed, 108 insertions, 150 deletions
diff --git a/reference/bump/bump-12.asm b/reference/bump/bump-12.asm
index 0f8742e..9112ce0 100644
--- a/reference/bump/bump-12.asm
+++ b/reference/bump/bump-12.asm
@@ -1,16 +1,16 @@
; options:
; VERT: new compiler
-@in(r0.x) in0
-@in(r0.y) in1
-@in(r0.z) in2
+@in(r3.w) in0
+@in(r4.x) in1
+@in(r4.y) in2
@in(r1.x) in4
@in(r1.y) in5
-@in(r1.z) in8
-@in(r1.w) in9
-@in(r2.x) in10
-@in(r3.w) in12
-@in(r4.x) in13
-@in(r4.y) in14
+@in(r0.x) in8
+@in(r0.y) in9
+@in(r0.z) in10
+@in(r4.z) in12
+@in(r4.w) in13
+@in(r5.x) in14
@out(r0.x) out0
@out(r0.y) out1
@out(r0.z) out2
@@ -27,151 +27,109 @@
@out(r3.y) out13
@out(r3.z) out14
@out(r3.w) out15
-(sy)(ss)mul.f r0.w, r2.x, r4.x
-mul.f r2.y, c4.x, r1.z
-mad.f32 r0.w, r1.w, r4.y, (neg)r0.w
-mad.f32 r2.y, c5.x, r1.w, r2.y
-mul.f r2.z, c4.x, r3.w
-mad.f32 r2.y, c6.x, r2.x, r2.y
-mov.f32f32 r0.w, r0.w
-mad.f32 r2.z, c5.x, r4.x, r2.z
-mul.f r2.w, c4.y, r3.w
-mul.f r3.x, c4.z, r3.w
-mul.f r3.y, c4.x, r0.w
-mul.f r3.z, r1.z, r4.y
-add.f r2.y, r2.y, c7.x
-mad.f32 r3.z, r2.x, r3.w, (neg)r3.z
-mad.f32 r2.z, c6.x, r4.y, r2.z
-mul.f r4.z, c4.y, r0.w
-mul.f r0.w, c4.z, r0.w
-mov.f32f32 r3.z, r3.z
-mov.f32f32 r4.w, r2.y
-add.f r2.y, r2.z, c7.x
-mad.f32 r2.z, c5.y, r4.x, r2.w
-mad.f32 r2.w, c5.x, r3.z, r3.y
-mul.f r3.y, r1.w, r3.w
-mul.f r5.x, r4.w, r4.w
-mad.f32 r3.y, r1.z, r4.x, (neg)r3.y
-mul.f r5.y, c4.y, r1.z
-mov.f32f32 r2.y, r2.y
-mad.f32 r5.y, c5.y, r1.w, r5.y
-mov.f32f32 r3.y, r3.y
-mad.f32 r5.y, c6.y, r2.x, r5.y
-mul.f r5.z, r2.y, r2.y
-mad.f32 r2.z, c6.y, r4.y, r2.z
-mad.f32 r2.w, c6.x, r3.y, r2.w
-add.f r5.y, r5.y, c7.y
-mad.f32 r4.z, c5.y, r3.z, r4.z
-mad.f32 r0.w, c5.z, r3.z, r0.w
-mov.f32f32 r2.w, r2.w
-mov.f32f32 r5.y, r5.y
-add.f r2.z, r2.z, c7.y
-mad.f32 r3.z, c6.y, r3.y, r4.z
-add.f r2.w, r2.w, c7.x
-mad.f32 r4.z, r5.y, r5.y, r5.x
-mov.f32f32 r2.z, r2.z
-mov.f32f32 r3.z, r3.z
-mov.f32f32 r2.w, r2.w
-mov.f32f32 r4.z, r4.z
-mad.f32 r5.x, r2.z, r2.z, r5.z
-mul.f r1.z, c4.z, r1.z
-mul.f r5.z, r2.w, r2.w
-add.f r3.z, r3.z, c7.y
-mad.f32 r1.z, c5.z, r1.w, r1.z
-mov.f32f32 r1.w, r5.x
-mad.f32 r3.x, c5.z, r4.x, r3.x
-mov.f32f32 r3.z, r3.z
-mad.f32 r1.z, c6.z, r2.x, r1.z
-mad.f32 r2.x, c6.z, r4.y, r3.x
-mad.f32 r0.w, c6.z, r3.y, r0.w
-mad.f32 r3.x, r3.z, r3.z, r5.z
-add.f r1.z, r1.z, c7.z
-add.f r2.x, r2.x, c7.z
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r3.x, r3.x
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r2.x, r2.x
-add.f r0.w, r0.w, c7.z
-mul.f r3.y, c0.w, r0.x
-mul.f r5.x, c0.z, r0.x
-mul.f r5.z, c0.y, r0.x
-mov.f32f32 r0.w, r0.w
-mad.f32 r4.z, r1.z, r1.z, r4.z
-mad.f32 r1.w, r2.x, r2.x, r1.w
-mad.f32 r3.y, c1.w, r0.y, r3.y
-mad.f32 r3.x, r0.w, r0.w, r3.x
-mad.f32 r3.y, c2.w, r0.z, r3.y
-mad.f32 r5.x, c1.z, r0.y, r5.x
-mad.f32 r5.z, c1.y, r0.y, r5.z
-mul.f r0.x, c0.x, r0.x
-mul.f r3.w, c4.w, r3.w
-add.f r5.w, r3.y, c3.w
-rsq r3.x, r3.x
-(ss)mov.f32f32 r3.x, r3.x
-rsq r3.y, r4.z
-nop
-rsq r1.w, r1.w
-(ss)mov.f32f32 r4.z, r3.y
-mov.f32f32 r1.w, r1.w
-mul.f r0.w, r0.w, r3.x
-mul.f r3.y, r3.z, r3.x
-mul.f r2.w, r2.w, r3.x
-mul.f r1.z, r1.z, r4.z
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r3.x, r3.y
-mov.f32f32 r2.w, r2.w
-nop
-mov.f32f32 r3.z, r0.w
-mov.f32f32 r3.y, r3.x
-mov.f32f32 r3.x, r2.w
-mov.f32f32 r0.w, r1.z
-mul.f r1.z, r2.x, r1.w
-mul.f r2.x, r2.z, r1.w
-mul.f r1.w, r2.y, r1.w
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r2.x, r2.x
-mov.f32f32 r1.w, r1.w
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r2.y, r2.x
-mov.f32f32 r1.w, r1.w
+@const(c8.x) 0x3f800000, 0x00000000, 0x00000000, 0x00000000
+@const(c9.x) 0x00000000, 0x3f800000, 0x43000000, 0x00000000
+(sy)(ss)mul.f r0.w, r0.z, r4.w
+mul.f r1.z, c4.x, r4.z
+mad.f32 r0.w, r0.y, r5.x, (neg)r0.w
+mad.f32 r1.z, c5.x, r4.w, r1.z
+mul.f r1.w, c4.x, r0.x
+mad.f32 r1.z, c6.x, r5.x, r1.z
mov.f32f32 r2.x, r0.w
+mad.f32 r1.w, c5.x, r0.y, r1.w
+mul.f r0.w, c4.x, r0.w
+add.f r1.z, r1.z, c7.x
+mul.f r2.y, c4.y, r2.x
+mul.f r2.z, r0.x, r5.x
+mad.f32 r1.w, c6.x, r0.z, r1.w
+mad.f32 r2.z, r0.z, r4.z, (neg)r2.z
mov.f32f32 r2.w, r1.z
-mov.f32f32 r2.z, r2.y
-mov.f32f32 r2.y, r1.w
-mul.f r1.z, r5.y, r4.z
-mul.f r1.w, r4.w, r4.z
-mov.f32f32 r0.w, r5.w
-mad.f32 r4.z, c2.z, r0.z, r5.x
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r1.w, r1.w
-mad.f32 r4.w, c2.y, r0.z, r5.z
-mad.f32 r0.x, c1.x, r0.y, r0.x
-mov.f32f32 r0.y, r1.z
-mov.f32f32 r1.z, r1.w
-(rpt1)nop
-mov.f32f32 r1.w, r0.y
-mov.f32f32 r1.z, r1.z
-add.f r0.y, r4.z, c3.z
-add.f r4.z, r4.w, c3.y
-mad.f32 r0.x, c2.x, r0.z, r0.x
-mad.f32 r3.w, c5.w, r4.x, r3.w
-mov.f32f32 r0.z, r0.y
-mov.f32f32 r0.y, r4.z
+mul.f r2.x, c4.z, r2.x
+add.f r1.w, r1.w, c7.x
+mov.f32f32 r3.x, r2.z
+mul.f r2.w, r2.w, r2.w
+mul.f r3.y, c4.y, r4.z
+mov.f32f32 r3.z, r1.w
+mad.f32 r2.y, c5.y, r3.x, r2.y
+mul.f r5.y, r0.y, r4.z
+mad.f32 r3.y, c5.y, r4.w, r3.y
+mad.f32 r5.y, r0.x, r4.w, (neg)r5.y
+mad.f32 r3.y, c6.y, r5.x, r3.y
+mul.f r3.z, r3.z, r3.z
+mul.f r5.z, c4.y, r0.x
+mov.f32f32 r5.w, r5.y
+add.f r3.y, r3.y, c7.y
+mad.f32 r5.z, c5.y, r0.y, r5.z
+mad.f32 r0.w, c5.x, r2.z, r0.w
+mad.f32 r2.y, c6.y, r5.w, r2.y
+mov.f32f32 r2.z, r3.y
+mad.f32 r5.z, c6.y, r0.z, r5.z
+mad.f32 r0.w, c6.x, r5.y, r0.w
+add.f r2.y, r2.y, c7.y
+mad.f32 r2.w, r3.y, r2.z, r2.w
+mul.f r3.y, c4.z, r4.z
+add.f r5.y, r5.z, c7.y
+mov.f32f32 r5.z, r2.y
+add.f r0.w, r0.w, c7.x
+mad.f32 r3.y, c5.z, r4.w, r3.y
+mov.f32f32 r6.x, r5.y
+mad.f32 r3.y, c6.z, r5.x, r3.y
+mov.f32f32 r6.y, r0.w
+mad.f32 r2.x, c5.z, r3.x, r2.x
+mad.f32 r5.y, r5.y, r6.x, r3.z
+mad.f32 r2.x, c6.z, r5.w, r2.x
+mul.f r3.x, r6.y, r6.y
+add.f r3.y, r3.y, c7.z
+mad.f32 r2.y, r2.y, r5.z, r3.x
+add.f r2.x, r2.x, c7.z
+mul.f r0.x, c4.z, r0.x
+mov.f32f32 r5.w, r3.y
+mad.f32 r0.x, c5.z, r0.y, r0.x
+mov.f32f32 r0.y, r2.x
+mad.f32 r0.x, c6.z, r0.z, r0.x
+mad.f32 r0.z, r3.y, r5.w, r2.w
+mul.f r2.w, c0.w, r3.w
+mad.f32 r2.x, r2.x, r0.y, r2.y
+add.f r0.x, r0.x, c7.z
+mad.f32 r6.y, c1.w, r4.x, r2.w
+mul.f r6.z, c0.z, r3.w
+mul.f r6.w, c0.y, r3.w
+mul.f r3.w, c0.x, r3.w
+mul.f r4.z, c4.w, r4.z
+rsq r2.x, r2.x
+(ss)mov.f32f32 r2.y, r2.x
+mul.f r3.x, r0.w, r2.x
+rsq r0.z, r0.z
+(ss)mov.f32f32 r0.w, r0.z
+(ss)mov.f32f32 r2.x, r0.x
+mul.f r3.z, r0.y, r2.y
+mul.f r3.y, r5.z, r2.y
+mul.f r2.w, r5.w, r0.w
+mul.f r2.z, r2.z, r0.w
+mad.f32 r0.x, r0.x, r2.x, r5.y
+mul.f r2.y, r1.z, r0.z
+mad.f32 r0.y, c2.w, r4.y, r6.y
+mad.f32 r0.z, c1.z, r4.x, r6.z
+mad.f32 r5.y, c1.y, r4.x, r6.w
+mad.f32 r3.w, c1.x, r4.x, r3.w
+mad.f32 r4.x, c5.w, r4.w, r4.z
+rsq r0.x, r0.x
+(ss)mov.f32f32 r4.z, r0.x
+mul.f r1.z, r1.w, r0.x
+add.f r0.w, r0.y, c3.w
+(ss)mad.f32 r0.x, c2.z, r4.y, r0.z
+mul.f r2.x, r2.x, r4.z
+mul.f r1.w, r6.x, r4.z
+mad.f32 r0.y, c2.y, r4.y, r5.y
+add.f r0.z, r0.x, c3.z
+mad.f32 r0.x, c2.x, r4.y, r3.w
+mad.f32 r3.w, c6.w, r5.x, r4.x
+add.f r0.y, r0.y, c3.y
+nop
add.f r0.x, r0.x, c3.x
-mad.f32 r3.w, c6.w, r4.y, r3.w
-(rpt1)nop
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r3.w, r3.w
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r1.x, r1.x
-(rpt1)nop
-mov.f32f32 r1.y, r1.y
-mov.f32f32 r1.x, r1.x
end
nop
; VERT: outputs: r0.x (0:0) r1.x (5:20) r2.x (5:21) r3.x (5:22)
-; VERT: inputs: r0.x (0:0,cm=7,il=8,b=0) r1.x (0:0,cm=3,il=12,b=0) r1.z (0:0,cm=7,il=16,b=0) r3.w (0:0,cm=7,il=20,b=0)
-; VERT: 146 instructions, 0 half, 6 full
+; VERT: inputs: r3.w (0:0,cm=7,il=8,b=0) r1.x (0:0,cm=3,il=12,b=0) r0.x (0:0,cm=7,il=16,b=0) r4.z (0:0,cm=7,il=20,b=0)
+; VERT: 99 instructions, 0 half, 7 full