summaryrefslogtreecommitdiff
path: root/reference/0ad-alpine-valley/0ad-61.asm
diff options
context:
space:
mode:
Diffstat (limited to 'reference/0ad-alpine-valley/0ad-61.asm')
-rw-r--r--reference/0ad-alpine-valley/0ad-61.asm296
1 files changed, 110 insertions, 186 deletions
diff --git a/reference/0ad-alpine-valley/0ad-61.asm b/reference/0ad-alpine-valley/0ad-61.asm
index e890c72..95d5c75 100644
--- a/reference/0ad-alpine-valley/0ad-61.asm
+++ b/reference/0ad-alpine-valley/0ad-61.asm
@@ -8,211 +8,135 @@
@out(r2.y) out1
@out(r2.z) out2
@out(r2.w) out3
+@const(c9.x) 0x3f000000, 0x00000000, 0x3f800000, 0xba03126f
+@const(c10.x) 0xbf000000, 0x40000000, 0x3f800000, 0x3fb8aa65
+@const(c11.x) 0x3de38866, 0x3cf5c28f, 0x00000000, 0x00000000
+@const(c12.x) 0x00000000, 0x3f800000, 0x43000000, 0x00000000
(sy)(ss)bary.f r0.x, 0, r1.x
add.f r0.y, r0.w, c9.y
bary.f r0.w, 1, r1.x
-bary.f r1.z, 4, r1.x
-add.f r1.w, r0.x, c10.x
-bary.f r2.x, 6, r1.x
+bary.f r1.z, 6, r1.x
+add.f r2.x, r0.x, c10.x
+bary.f r1.w, 7, r1.x
add.f r2.y, r0.w, c10.x
-mov.f32f32 r1.z, r1.z
-floor.f r2.z, r1.w
+bary.f r2.z, 2, r1.x
+floor.f r2.w, r2.x
rcp r0.y, r0.y
add.f r0.z, r0.z, c9.y
-floor.f r2.w, r2.y
-mov.f32f32 r3.x, r1.z
-add.f r1.z, r1.w, (neg)r2.z
+floor.f r3.x, r2.y
+add.f r3.w, r2.z, c9.w
+add.f r2.x, r2.x, (neg)r2.w
(ss)mul.f r0.y, r0.z, r0.y
-add.f r0.z, r2.y, (neg)r2.w
-bary.f r1.w, 5, r1.x
-mov.f32f32 r1.z, r1.z
+absneg.f r0.z, (neg)c7.x
+add.f r2.y, r2.y, (neg)r3.x
+mov.f32f32 r2.z, r2.x
+add.f r2.x, r2.x, c9.z
+mul.f r0.z, r0.z, c7.x
+sam (f32)(w)r4.x, r1.z, s#2, t#2
+(ss)mov.f32f32 r1.z, r2.y
+mul.f r1.w, c9.x, r2.z
+add.f r2.z, c10.y, (neg)r2.z
+mul.f r0.z, r0.z, r0.y
mov.f32f32 r0.y, r0.y
-absneg.f r2.y, (neg)c7.x
-mov.f32f32 r0.z, r0.z
-mul.f r2.z, c9.x, r1.z
-mov.f32f32 r1.w, r1.w
-add.f r2.w, c10.y, (neg)r1.z
-add.f r3.z, c10.y, (neg)r0.z
-mov.f32f32 r2.z, r2.z
-mul.f r2.y, r2.y, c7.x
-mul.f r3.w, c9.x, r0.z
-mov.f32f32 r3.y, r1.w
-add.f r0.x, r0.x, (neg)r2.z
-mul.f r1.w, r2.y, r0.y
-mov.f32f32 r2.y, r3.w
-mov.f32f32 r2.z, r2.w
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r1.w, r1.w
-add.f r0.w, r0.w, (neg)r2.y
-sam (f32)(w)r3.w, r3.x, s#1, t#1
-(sy)add.f r2.y, c9.z, (neg)r4.z
-add.f r2.w, c10.x, r0.x
-mul.f r0.y, r1.w, r0.y
+add.f r0.x, r0.x, (neg)r1.w
+mul.f r1.w, c9.x, r1.z
+mov.f32f32 r2.w, r2.z
+mul.f r0.y, r0.z, r0.y
+mov.f32f32 r0.z, r0.x
+add.f r0.w, r0.w, (neg)r1.w
add.f r0.x, c10.z, r0.x
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r1.w, r2.w
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.x, r0.x
-add.f r2.w, c10.z, r0.w
-mul.f r1.w, r1.w, c4.z
mul.f r0.y, r0.y, c10.w
-mul.f r0.x, r0.x, c4.z
-mov.f32f32 r2.w, r2.w
-(ss)mov.f32f32 r3.x, r1.w
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r3.y, r0.x
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r3.w, r3.x
-add.f r0.w, c10.x, r0.w
-mov.f32f32 r3.x, r3.y
-mov.f32f32 r4.z, r0.x
-mov.f32f32 r0.x, r1.w
-mov.f32f32 r0.w, r0.w
-exp2 r0.y, r0.y
-(ss)mov.f32f32 r0.y, r0.y
-mov.f32f32 r5.y, r3.x
-mul.f r1.w, r2.w, c4.w
-mul.f r0.w, r0.w, c4.w
-add.f r2.w, c12.y, (neg)r0.y
-mov.f32f32 r6.x, r0.x
-mov.f32f32 r0.x, r1.w
-mov.f32f32 r3.x, r0.w
-mul.f r2.w, r2.w, c7.y
-mul.f r0.y, r0.y, c9.z
-mov.f32f32 r1.w, r1.w
-mov.f32f32 r4.x, r3.x
-bary.f r3.x, 2, r1.x
-add.f r0.y, r0.y, r2.w
-mov.f32f32 r1.w, r1.w
-mov.f32f32 r0.w, r0.w
-add.f r2.w, r3.x, c9.w
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r5.z, r1.w
-mov.f32f32 r4.w, r0.w
-mov.f32f32 r0.w, r2.w
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r1.w, r2.w
-mov.f32f32 r3.x, r2.w
-mov.f32f32 r4.y, r0.w
-max.f r0.y, r0.y, c9.y
-mov.f32f32 r5.w, r1.w
-mov.f32f32 r5.x, r3.x
-mov.f32f32 r6.y, r0.x
-mov.f32f32 r0.x, r2.w
-mov.f32f32 r0.w, r2.y
-sam.s (f32)(x)r6.z, r3.w, s#3, t#3
-(sy)mov.f32f32 r1.w, r6.z
-min.f r0.y, r0.y, c9.z
-sam.s (f32)(x)r6.z, r5.y, s#3, t#3
-(sy)mov.f32f32 r2.y, r6.z
+add.f r0.z, c10.x, r0.z
+mov.f32f32 r1.w, r0.w
+mul.f r3.y, r0.x, c4.z
+add.f r0.x, c10.z, r0.w
+mul.f r4.x, r0.z, c4.z
+add.f r0.z, c10.x, r1.w
+mov.f32f32 r5.x, r3.y
+mul.f r3.z, r0.x, c4.w
+mov.f32f32 r5.w, r4.x
+mul.f r5.y, r0.z, c4.w
+exp2 r0.x, r0.y
+(ss)mov.f32f32 r0.y, r0.x
+mov.f32f32 r5.z, r3.w
+mov.f32f32 r4.y, r3.z
+mov.f32f32 r6.x, r5.y
+mov.f32f32 r6.y, r3.w
+add.f r0.y, c12.y, (neg)r0.y
+mov.f32f32 r4.z, r3.w
+sam.s (f32)(x)r3.x, r3.y, s#3, t#3
+add.f r0.z, c10.y, (neg)r1.z
+sam.s (f32)(x)r6.z, r5.x, s#3, t#3
+mul.f r0.x, r0.x, c9.z
+add.f r0.w, r2.y, c9.z
+mul.f r0.y, r0.y, c7.y
(ss)nop
-sam.s (f32)(x)r3.w, r4.z, s#3, t#3
-(sy)mov.f32f32 r2.w, r3.w
-mov.f32f32 r1.w, r1.w
-mov.f32f32 r3.x, r3.z
-add.f r3.y, c12.y, (neg)r0.y
-add.f r3.z, c12.y, (neg)r0.y
-add.f r3.w, c12.y, (neg)r0.y
-mul.f r4.x, r2.z, r3.x
-mul.f r3.y, r3.y, c6.z
-mul.f r3.z, r3.z, c6.y
-mul.f r3.w, r3.w, c6.x
-mul.f r1.w, r4.x, r1.w
-add.f r1.z, r1.z, c9.z
-mov.f32f32 r6.z, r0.x
-mov.f32f32 r0.x, r0.w
-add.f r0.z, r0.z, c9.z
-mul.f r0.w, r1.z, r3.x
-mov.f32f32 r2.x, r2.x
-bary.f r3.x, 8, r1.x
-mov.f32f32 r0.x, r0.x
-mad.f32 r0.w, r0.w, r2.w, r1.w
-sam.s (f32)(x)r4.x, r6.x, s#3, t#3
+sam.s (f32)(x)r5.x, r5.w, s#3, t#3
+mov.f32f32 r1.z, r0.z
+mul.f r0.z, r2.x, r0.z
+mul.f r1.w, r2.z, r0.w
+add.f r0.x, r0.x, r0.y
+mul.f r0.y, r2.w, r1.z
+sam.s (f32)(x)r3.y, r4.x, s#3, t#3
+mul.f r0.w, r2.x, r0.w
+(sy)cmps.f.lt r1.z, r4.w, c11.y
+bary.f r3.z, 4, r1.x
+mul.f r0.y, r0.y, r5.x
+max.f r0.x, r0.x, c9.y
+mad.f32 r0.y, r0.z, r6.z, r0.y
+cov.u32f32 r0.z, r1.z
+mad.f32 r0.y, r1.w, r3.y, r0.y
+min.f r0.x, r0.x, c9.z
+mad.f32 r0.y, r0.w, r3.x, r0.y
+cmps.f.ne r0.z, r0.z, c9.y
(rpt1)nop
-(sy)mov.f32f32 r1.w, r4.x
-mov.f32f32 r0.w, r0.w
-mul.f r2.z, r2.z, r0.z
-mov.f32f32 r2.w, r0.x
-mov.f32f32 r4.x, r2.x
-mul.f r0.x, r1.z, r0.z
-mad.f32 r0.z, r2.z, r1.w, r0.w
-bary.f r0.w, 7, r1.x
-mov.f32f32 r1.z, r3.x
+mul.f r0.y, c11.x, r0.y
+bary.f r1.z, 8, r1.x
bary.f r1.w, 9, r1.x
-mov.f32f32 r0.z, r0.z
-mov.f32f32 r0.w, r0.w
-mad.f32 r0.x, r0.x, r2.y, r0.z
-mov.f32f32 r2.x, r1.z
-mov.f32f32 r0.z, r1.w
-mov.f32f32 r4.y, r0.w
-mov.f32f32 r0.x, r0.x
-bary.f r0.w, 12, r1.x
-bary.f r1.z, 11, r1.x
-bary.f (ei)r1.x, 10, r1.x
-mul.f r0.x, c11.x, r0.x
-mov.f32f32 r2.y, r0.z
+add.f r0.w, c12.y, (neg)r0.x
+mov.f32f32 r2.x, r0.y
+add.f r2.y, c12.y, (neg)r0.x
+add.f r2.z, c12.y, (neg)r0.x
+(rpt1)nop
(ss)nop
-sam (f32)(w)r4.x, r4.x, s#2, t#2
-(sy)cmps.f.lt r0.z, r4.w, c11.y
-mov.f32f32 r1.y, r4.w
-mov.f32f32 r0.x, r0.x
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r1.x, r1.x
-sam (f32)(xyz)r1.w, r2.x, s#0, t#0
-cov.u32f32 r0.z, r0.z
-(sy)mul.f r0.w, r2.y, r0.w
-mul.f r1.z, r2.x, r1.z
-mul.f r1.x, r1.w, r1.x
-cmps.f.ne r0.z, r0.z, c9.y
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r1.z, r1.z
-mov.f32f32 r1.x, r1.x
-mov.f32f32 r1.y, r1.y
-mul.f r0.w, r0.w, r0.x
-mul.f r1.z, r1.z, r0.x
-mul.f r0.x, r1.x, r0.x
-mov.f32f32 r1.x, c9.y
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r1.z, r1.z
-mad.f32 r0.w, c5.z, r2.y, r0.w
-mad.f32 r1.z, c5.y, r2.x, r1.z
-mov.f32f32 r0.x, r0.x
-sel.b32 r0.z, r1.x, r0.z, r1.y
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r1.x, r1.z
-mad.f32 r0.x, c5.x, r1.w, r0.x
-nop
-mul.f r0.w, r0.y, r0.w
-mul.f r1.x, r0.y, r1.x
-mov.f32f32 r0.x, r0.x
-nop
-add.f r0.w, r0.w, r3.y
-add.f r1.x, r1.x, r3.z
-mul.f r0.x, r0.y, r0.x
-nop
-mul.f r0.y, r0.w, r0.z
-mul.f r0.w, r1.x, r0.z
-add.f r0.x, r0.x, r3.w
-nop
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.w, r0.w
-mul.f r0.x, r0.x, r0.z
+sam (f32)(xyz)r3.w, r1.z, s#0, t#0
+(ss)bary.f r1.z, 12, r1.x
+bary.f r1.w, 11, r1.x
+bary.f r2.w, 10, r1.x
+mul.f r0.w, r0.w, c6.z
+(sy)mul.f r1.z, r4.y, r1.z
+mul.f r1.w, r4.x, r1.w
+mul.f r2.w, r3.w, r2.w
+mul.f r2.y, r2.y, c6.y
+mul.f r1.z, r1.z, r2.x
+mul.f r1.w, r1.w, r2.x
+mad.f32 r1.z, c5.z, r4.y, r1.z
+mad.f32 r1.w, c5.y, r4.x, r1.w
+mul.f r0.y, r2.w, r0.y
+mul.f r2.x, r2.z, c6.x
+mul.f r1.z, r0.x, r1.z
+mul.f r1.w, r0.x, r1.w
+mad.f32 r0.y, c5.x, r3.w, r0.y
+mov.f32f32 r2.z, c9.y
+add.f r0.w, r1.z, r0.w
+add.f r1.z, r1.w, r2.y
nop
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.z, r0.w
-mov.f32f32 r0.x, r0.x
+sel.b32 r0.z, r2.z, r0.z, r4.w
+mul.f r0.x, r0.x, r0.y
+bary.f (ei)r3.w, 5, r1.x
nop
-mov.f32f32 r2.z, r0.y
-(ss)mov.f32f32 r2.y, r0.z
-mov.f32f32 r0.x, r0.x
-(rpt2)nop
-mov.f32f32 r2.x, r0.x
+mul.f r2.z, r0.w, r0.z
+mul.f r2.y, r1.z, r0.z
+add.f r0.x, r0.x, r2.x
+(rpt1)nop
+sam (f32)(w)r0.w, r3.z, s#1, t#1
+(sy)add.f r2.w, c9.z, (neg)r1.z
+mul.f r2.x, r0.x, r0.z
end
nop
nop
nop
; FRAG: outputs: r2.x (1:0)
-; FRAG: inputs: r0.x (0:0,cm=f,il=8,b=0) r1.x (5:9,cm=f,il=8,b=1) r0.x (5:10,cm=f,il=12,b=1) r0.w (5:11,cm=f,il=16,b=1) r4.x (5:12,cm=f,il=20,b=1)
-; FRAG: 204 instructions, 0 half, 7 full
+; FRAG: inputs: r0.x (0:0,cm=f,il=8,b=0) r1.x (5:9,cm=f,il=8,b=1) r1.x (5:10,cm=f,il=12,b=1) r1.x (5:11,cm=f,il=16,b=1) r4.x (5:12,cm=f,il=20,b=1)
+; FRAG: 124 instructions, 0 half, 7 full