1 files changed, 120 insertions, 183 deletions
diff --git a/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-76.asm b/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-76.asm
index 0274fe9..be30c1a 100644
--- a/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-76.asm
+++ b/reference/0ad-cycladic-archipelago/0ad-cycladic-archipelago-76.asm
@@ -4,206 +4,143 @@
 @in(r0.w)	in3
 @in(r1.x)	in4
 @in(r1.y)	in5
-@out(r1.y)	out0
-@out(r1.z)	out1
-@out(r1.w)	out2
-@out(r2.x)	out3
+@out(r2.x)	out0
+@out(r2.y)	out1
+@out(r2.z)	out2
+@out(r2.w)	out3
+@const(c9.x)	0x3f000000, 0x00000000, 0x3d4ccccd, 0xbb449ba6
+@const(c10.x)	0xbf000000, 0x3f800000, 0x40000000, 0x3de38866
+@const(c11.x)	0x3fb8aa65, 0x3cf5c28f, 0x00000000, 0x00000000
+@const(c12.x)	0x00000000, 0x3f800000, 0x43000000, 0x00000000
 (sy)(ss)bary.f r0.x, 4, r1.x
 add.f r0.y, r0.w, c9.y
 bary.f r0.w, 5, r1.x
 bary.f r1.z, 10, r1.x
-add.f r1.w, r0.x, c10.x
-bary.f r2.x, 8, r1.x
+add.f r2.x, r0.x, c10.x
+bary.f r1.w, 11, r1.x
 add.f r2.y, r0.w, c10.x
-mov.f32f32 r1.z, r1.z
-floor.f r2.z, r1.w
+bary.f r3.y, 8, r1.x
+floor.f r2.z, 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
+bary.f r3.z, 9, r1.x
+add.f r3.w, r2.x, (neg)r2.z
 (ss)mul.f r0.y, r0.z, r0.y
-add.f r0.z, r2.y, (neg)r2.w
-bary.f r1.w, 11, r1.x
-mov.f32f32 r1.z, r1.z
+absneg.f r0.z, (neg)c7.x
+add.f r4.x, r2.y, (neg)r2.w
+mov.f32f32 r4.y, r3.w
+sam (f32)(xyzw)r2.x, r1.z, s#0, t#0
+(ss)add.f r1.z, r3.w, c10.y
+mul.f r0.z, r0.z, c7.x
+mov.f32f32 r1.w, r4.x
+mul.f r3.w, c9.x, r4.y
+add.f r4.y, c10.z, (neg)r4.y
+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.z, (neg)r1.z
-add.f r3.z, c10.z, (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)(xyzw)r3.w, r3.x, s#0, t#0
-(sy)cmps.f.lt r2.y, r4.z, c9.z
-add.f r2.w, c10.x, r0.x
-mul.f r0.y, r1.w, r0.y
+add.f r0.x, r0.x, (neg)r3.w
+mul.f r3.w, c9.x, r1.w
+mov.f32f32 r4.z, r4.y
+mul.f r0.y, r0.z, r0.y
+mov.f32f32 r0.z, r0.x
+add.f r0.w, r0.w, (neg)r3.w
 add.f r0.x, c10.y, 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.y, r0.w
-mul.f r1.w, r1.w, c3.z
 mul.f r0.y, r0.y, c11.x
-mul.f r0.x, r0.x, c3.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 r4.w, r3.x
-add.f r0.w, c10.x, r0.w
-mov.f32f32 r3.x, r3.y
-mov.f32f32 r5.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 r6.y, r3.x
-mul.f r1.w, r2.w, c3.w
-mul.f r0.w, r0.w, c3.w
-add.f r2.w, c12.y, (neg)r0.y
-mov.f32f32 r7.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, c10.y
-mov.f32f32 r1.w, r1.w
-mov.f32f32 r5.x, r3.x
-bary.f r3.x, 6, 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 r6.z, r1.w
-mov.f32f32 r5.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 r5.y, r0.w
-max.f r0.y, r0.y, c9.y
-mov.f32f32 r6.w, r1.w
-mov.f32f32 r6.x, r3.x
-mov.f32f32 r7.y, r0.x
-mov.f32f32 r0.x, r2.w
-cov.u32f32 r0.w, r2.y
-sam.s (f32)(x)r1.w, r4.w, s#2, t#2
-(sy)mov.f32f32 r1.w, r1.w
-min.f r0.y, r0.y, c10.y
-sam.s (f32)(x)r2.y, r6.y, s#2, t#2
-(sy)mov.f32f32 r2.y, r2.y
-sam.s (f32)(x)r2.w, r5.z, s#2, t#2
-(sy)mov.f32f32 r2.w, r2.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
-(ss)add.f r4.w, c12.y, (neg)r0.y
-mul.f r5.x, r2.z, r3.x
-mul.f r3.y, r3.y, c6.z
-mul.f r3.z, r3.z, c6.y
-mul.f r4.w, r4.w, c6.x
-mul.f r1.w, r5.x, r1.w
-add.f r1.z, r1.z, c10.y
-mov.f32f32 r7.z, r0.x
-mov.f32f32 r0.x, (0.000000)
-add.f r0.z, r0.z, c10.y
-mul.f r3.x, r1.z, r3.x
-mov.f32f32 r2.x, r2.x
-cmps.f.ne p0.x, r0.w, r0.x
-mul.f r0.x, r2.z, r0.z
-mad.f32 r0.w, r3.x, r2.w, r1.w
-sam.s (f32)(x)r1.w, r7.x, s#2, t#2
+add.f r0.z, c10.x, r0.z
+mov.f32f32 r3.w, r0.w
+mul.f r4.w, r0.x, c3.z
+add.f r0.x, c10.y, r0.w
+mul.f r5.z, r0.z, c3.z
+add.f r0.z, c10.x, r3.w
+mov.f32f32 r6.y, r4.w
+mul.f r5.x, r0.x, c3.w
+mov.f32f32 r7.x, r5.z
+mul.f r6.z, r0.z, c3.w
+exp2 r0.x, r0.y
+(ss)mov.f32f32 r0.y, r0.x
+bary.f r0.z, 6, r1.x
+mov.f32f32 r5.w, r5.x
+mov.f32f32 r7.y, r6.z
+add.f r0.y, c12.y, (neg)r0.y
+add.f r5.y, r0.z, c9.w
+add.f r0.z, c10.z, (neg)r1.w
+mul.f r0.x, r0.x, c10.y
+add.f r0.w, r4.x, c10.y
+mov.f32f32 r7.z, r5.y
+mul.f r0.y, r0.y, c7.y
+mov.f32f32 r6.w, r5.y
+mov.f32f32 r6.x, r5.y
 (rpt1)nop
-(sy)mov.f32f32 r1.w, r1.w
-mov.f32f32 r0.w, r0.w
-bary.f r2.z, 0, r1.x
-mov.f32f32 r2.w, r2.x
-mad.f32 r0.x, r0.x, r1.w, r0.w
-bary.f r0.w, 1, r1.x
-bary.f r1.w, 9, r1.x
+add.f r0.x, r0.x, r0.y
+sam.s (f32)(x)r7.x, r7.x, s#2, t#2
+mov.f32f32 r0.y, r0.z
+nop
+(ss)nop
+sam.s (f32)(x)r7.y, r6.y, s#2, t#2
+sam.s (f32)(x)r7.z, r4.w, s#2, t#2
 mul.f r0.z, r1.z, r0.z
-mov.f32f32 r0.x, r0.x
-bary.f (ei)r1.x, 2, r1.x
-mov.f32f32 r1.y, r1.w
-mad.f32 r0.x, r0.z, r2.y, r0.x
+mul.f r1.w, r4.y, r0.w
+mul.f r0.y, r4.z, r0.y
+max.f r0.x, r0.x, c9.y
+sam.s (f32)(x)r3.w, r5.z, s#2, t#2
+(sy)cmps.f.lt r4.x, r2.w, c9.z
+mul.f r0.w, r1.z, r0.w
+mul.f r0.y, r0.y, r7.x
+min.f r0.x, r0.x, c10.y
+mad.f32 r0.y, r0.z, r7.y, r0.y
+cov.u32f32 r0.z, r4.x
+mad.f32 r0.y, r1.w, r3.w, r0.y
+add.f r1.z, c12.y, (neg)r0.x
+mad.f32 r0.y, r0.w, r7.z, r0.y
+add.f r0.w, c12.y, (neg)r0.x
+add.f r1.w, c12.y, (neg)r0.x
+mov.f32f32 r3.w, (0.000000)
+mul.f r0.y, c10.w, r0.y
+bary.f r4.x, 0, r1.x
+mul.f r1.z, r1.z, c6.z
+mul.f r0.w, r0.w, c6.y
+mov.f32f32 r4.y, r0.y
+bary.f r4.z, 2, r1.x
+bary.f (ei)r1.x, 1, r1.x
+mul.f r1.y, r2.x, r4.x
+mul.f r1.w, r1.w, c6.x
+mul.f r4.x, r2.z, r4.z
+mul.f r1.x, r2.y, r1.x
+mul.f r0.y, r1.y, r0.y
+cmps.f.ne p0.x, r0.z, r3.w
+mul.f r0.z, r4.x, r4.y
+mul.f r1.x, r1.x, r4.y
+mad.f32 r0.z, c5.z, r2.z, r0.z
+mad.f32 r1.x, c5.y, r2.y, r1.x
+mad.f32 r0.y, c5.x, r2.x, r0.y
+sam (f32)(w)r3.y, r3.y, s#1, t#1
+(sy)cmps.f.lt r1.y, r4.x, c11.y
+mul.f r0.z, r0.x, r0.z
+mul.f r1.x, r0.x, r1.x
+mul.f r0.x, r0.x, r0.y
 kill p0.x
-mov.f32f32 r1.z, r4.z
-mov.f32f32 r3.x, r1.y
-mov.f32f32 r0.x, r0.x
-mul.f r1.x, r4.y, r1.x
-mul.f r0.w, r4.x, r0.w
-mov.f32f32 r2.x, r1.z
-mul.f r0.x, c10.w, r0.x
-mul.f r2.y, r3.w, r2.z
-sam (f32)(w)r5.x, r2.w, s#1, t#1
-nop
-(sy)cmps.f.lt r2.z, r5.w, c11.y
-mov.f32f32 r0.x, r0.x
-(ss)mov.f32f32 r2.w, r5.w
-mov.f32f32 r3.x, c9.y
-nop
-mul.f r1.x, r1.x, r0.x
-mul.f r0.w, r0.w, r0.x
-mul.f r0.x, r2.y, r0.x
-cov.u32f32 r2.y, r2.z
-mov.f32f32 r1.x, r1.x
-mov.f32f32 r0.w, r0.w
-mad.f32 r1.x, c5.z, r4.y, r1.x
-mad.f32 r0.w, c5.y, r4.x, r0.w
-mov.f32f32 r0.x, r0.x
-cmps.f.ne r2.y, r2.y, c9.y
-mov.f32f32 r1.x, r1.x
-mov.f32f32 r0.w, r0.w
-mad.f32 r0.x, c5.x, r3.w, r0.x
-mov.f32f32 r2.z, r2.w
-mul.f r1.x, r0.y, r1.x
-mul.f r0.w, r0.y, r0.w
-mov.f32f32 r0.x, r0.x
-sel.b32 r2.y, r3.x, r2.y, r2.z
-add.f r1.x, r1.x, r3.y
-add.f r0.w, r0.w, r3.z
-mul.f r0.x, r0.y, r0.x
-nop
-mul.f r0.y, r1.x, r2.y
-mul.f r0.w, r0.w, r2.y
-add.f r0.x, r0.x, r4.w
-nop
-mul.f r0.y, r0.y, c4.z
-mul.f r0.w, r0.w, c4.y
-mul.f r0.x, r0.x, r2.y
+add.f r0.y, r0.z, r1.z
+cov.u32f32 r0.z, r1.y
+add.f r0.w, r1.x, r0.w
+add.f r0.x, r0.x, r1.w
+mov.f32f32 r1.x, c9.y
+cmps.f.ne r0.z, r0.z, c9.y
+(rpt2)nop
+sel.b32 r0.z, r1.x, r0.z, r4.x
+(rpt2)nop
+mul.f r0.y, r0.y, r0.z
+mul.f r0.w, r0.w, r0.z
+(rpt1)nop
+mul.f r2.z, r0.y, c4.z
+mul.f r2.y, r0.w, c4.y
+mul.f r0.x, r0.x, r0.z
+(rpt2)nop
+mul.f r2.x, r0.x, c4.x
+end
 nop
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.w, r0.w
-mul.f r0.x, r0.x, c4.x
 nop
-mov.f32f32 r0.y, r0.y
-mov.f32f32 r0.w, r0.w
-mov.f32f32 r0.x, r0.x
 nop
-mov.f32f32 r1.w, r0.y
-mov.f32f32 r1.z, r0.w
-mov.f32f32 r0.x, r0.x
-(rpt2)nop
-mov.f32f32 r1.y, r0.x
-end
-; FRAG: outputs: r1.y (1:0)
+; 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) r2.x (5:10,cm=f,il=12,b=1) r1.x (5:11,cm=f,il=16,b=1)
-; FRAG: 195 instructions, 0 half, 8 full
-; pos (bary): r1.x
-; color: r1.y
-; fragcoord: r0.x
+; FRAG: 133 instructions, 0 half, 8 full