diff options
Diffstat (limited to 'reference/stk-mines/stk-mines-37.asm')
-rw-r--r-- | reference/stk-mines/stk-mines-37.asm | 307 |
1 files changed, 130 insertions, 177 deletions
diff --git a/reference/stk-mines/stk-mines-37.asm b/reference/stk-mines/stk-mines-37.asm index 2d84780..726bef5 100644 --- a/reference/stk-mines/stk-mines-37.asm +++ b/reference/stk-mines/stk-mines-37.asm @@ -1,16 +1,16 @@ ; options: ; VERT: new compiler -@in(r1.x) in0 -@in(r1.y) in1 -@in(r1.z) in2 -@in(r1.w) in3 +@in(r4.z) in0 +@in(r4.w) in1 +@in(r5.x) in2 +@in(r5.y) in3 @in(r0.x) in4 @in(r0.y) in5 @in(r0.z) in6 -@in(r5.x) in8 -@in(r5.y) in9 -@in(r5.z) in10 -@in(r5.w) in11 +@in(r1.x) in8 +@in(r1.y) in9 +@in(r1.z) in10 +@in(r1.w) in11 @out(r0.x) out0 @out(r0.y) out1 @out(r0.z) out2 @@ -31,191 +31,144 @@ @out(r4.y) out17 @out(r4.z) out18 @out(r4.w) out19 +@const(c12.x) 0x00000000, 0x3f800000, 0x43000000, 0x00000000 (sy)(ss)mov.f32f32 r0.w, c1.x -mul.f r2.x, c8.x, r1.x +mul.f r2.x, c8.x, r4.z mov.f32f32 r2.y, c1.y mov.f32f32 r2.z, c1.z mul.f r2.w, r0.w, r0.x mov.f32f32 r3.x, c2.x -mad.f32 r2.x, c9.x, r1.y, r2.x +mad.f32 r2.x, c9.x, r4.w, r2.x mul.f r3.y, r0.w, r0.w -mad.f32 r2.x, c10.x, r1.z, r2.x +mad.f32 r2.x, c10.x, r5.x, r2.x mad.f32 r2.w, r3.x, r0.y, r2.w -mad.f32 r2.x, c11.x, r1.w, r2.x -mad.f32 r3.x, r2.y, r2.y, r3.y -mul.f r3.y, r2.y, r0.x -mov.f32f32 r2.w, r2.w -mov.f32f32 r3.z, c3.x -mov.f32f32 r2.x, r2.x -mov.f32f32 r3.x, r3.x +mov.f32f32 r3.x, c3.x +mad.f32 r2.x, c11.x, r5.y, r2.x +mad.f32 r3.y, r2.y, r2.y, r3.y +mul.f r3.z, r2.y, r0.x +mad.f32 r2.w, r3.x, r0.z, r2.w +mov.f32f32 r3.x, r2.x +mad.f32 r3.y, r2.z, r2.z, r3.y mov.f32f32 r3.w, c2.y -mad.f32 r2.w, r3.z, r0.z, r2.w -mul.f r3.z, r2.x, r2.x -mul.f r4.x, c8.y, r1.x -mad.f32 r3.x, r2.z, r2.z, r3.x -mov.f32f32 r4.y, r2.w -mad.f32 r2.w, c9.y, r1.y, r4.x -mad.f32 r3.y, r3.w, r0.y, r3.y -mul.f r0.x, r2.z, r0.x -mul.f r3.w, r4.y, r4.y -mad.f32 r2.w, c10.y, r1.z, r2.w -mov.f32f32 r3.y, r3.y -mov.f32f32 r4.x, c3.y -mad.f32 r2.w, c11.y, r1.w, r2.w -rsq r3.x, r3.x -(ss)mov.f32f32 r3.x, r3.x -mov.f32f32 r4.z, c2.z -mad.f32 r3.y, r4.x, r0.z, r3.y mov.f32f32 r4.x, r2.w -mul.f r0.w, r0.w, r3.x -mul.f r2.y, r2.y, r3.x -mov.f32f32 r3.y, r3.y -mad.f32 r2.w, r4.x, r4.x, r3.z +mul.f r3.x, r3.x, r3.x +mul.f r4.y, c8.y, r4.z +mad.f32 r3.z, r3.w, r0.y, r3.z +mul.f r2.w, r2.w, r4.x +mov.f32f32 r3.w, c3.y +mad.f32 r4.y, c9.y, r4.w, r4.y +rsq r3.y, r3.y +(ss)mov.f32f32 r5.z, r3.y +mad.f32 r4.y, c10.y, r5.x, r4.y +mad.f32 r3.z, r3.w, r0.z, r3.z +mad.f32 r3.w, c11.y, r5.y, r4.y +mul.f r0.w, r0.w, r5.z +mul.f r2.y, r2.y, r5.z +mov.f32f32 r4.y, r3.z +mov.f32f32 r5.z, r3.w +mul.f r5.w, c0.x, r0.w mov.f32f32 r0.w, r0.w -mov.f32f32 r2.y, r2.y -mad.f32 r3.z, r3.y, r3.y, r3.w -mov.f32f32 r2.w, r2.w -mul.f r3.w, c8.z, r1.x -mul.f r4.w, c0.x, r0.w -mov.f32f32 r3.z, r3.z -mad.f32 r0.x, r4.z, r0.y, r0.x -mad.f32 r0.y, c9.z, r1.y, r3.w -mad.f32 r3.w, c0.y, r2.y, r4.w -mad.f32 r0.y, c10.z, r1.z, r0.y -mov.f32f32 r0.x, r0.x -mov.f32f32 r4.z, c3.z -mad.f32 r0.y, c11.z, r1.w, r0.y -mov.f32f32 r3.w, r3.w -mul.f r2.z, r2.z, r3.x -mad.f32 r0.x, r4.z, r0.z, r0.x -mov.f32f32 r0.y, r0.y -mul.f r0.z, c4.w, r1.x -mul.f r3.x, c4.z, r1.x -mov.f32f32 r0.x, r0.x -mad.f32 r2.w, r0.y, r0.y, r2.w +mad.f32 r2.w, r3.z, r4.y, r2.w +mul.f r0.x, r2.z, r0.x +mov.f32f32 r3.z, c2.z +mad.f32 r3.x, r3.w, r5.z, r3.x +mul.f r3.w, c8.z, r4.z +mad.f32 r5.w, c0.y, r2.y, r5.w +mad.f32 r0.x, r3.z, r0.y, r0.x +mov.f32f32 r0.y, c3.z +mad.f32 r3.z, c9.z, r4.w, r3.w +mul.f r2.z, r2.z, r3.y +(ss)mad.f32 r3.y, c10.z, r5.x, r3.z +mad.f32 r0.x, r0.y, r0.z, r0.x +mad.f32 r0.y, c11.z, r5.y, r3.y +mad.f32 r3.y, c0.z, r2.z, r5.w +mov.f32f32 r0.z, r2.y +mov.f32f32 r2.y, r0.x +mov.f32f32 r3.z, r0.y +mov.f32f32 r3.w, r3.y mov.f32f32 r2.z, r2.z -mad.f32 r0.z, c5.w, r1.y, r0.z -mad.f32 r3.z, r0.x, r0.x, r3.z -mad.f32 r0.z, c6.w, r1.z, r0.z -mad.f32 r3.x, c5.z, r1.y, r3.x -mul.f r4.z, c4.y, r1.x -mul.f r1.x, c4.x, r1.x -mad.f32 r0.z, c7.w, r1.w, r0.z -mad.f32 r3.x, c6.z, r1.z, r3.x -rsq r3.z, r3.z -(ss)mov.f32f32 r3.z, r3.z -rsq r2.w, r2.w -(ss)mov.f32f32 r4.w, r2.w -mad.f32 r3.w, c0.z, r2.z, r3.w -(ss)mov.f32f32 r2.w, r0.z -mul.f r0.x, r0.x, r3.z -mul.f r2.x, r2.x, r4.w -mul.f r4.y, r4.y, r3.z -mul.f r3.y, r3.y, r3.z -mov.f32f32 r0.x, r0.x -mov.f32f32 r2.x, r2.x -mov.f32f32 r3.z, r4.y -mov.f32f32 r3.y, r3.y -mul.f r4.y, r0.x, r2.y -absneg.f r2.x, (neg)r2.x -mul.f r6.x, c0.x, r3.z -mad.f32 r4.y, r3.y, r2.z, (neg)r4.y -mad.f32 r6.x, c0.y, r3.y, r6.x -mov.f32f32 r6.y, r2.x -mul.f r2.x, r3.z, r2.z -mov.f32f32 r2.z, r4.y -mov.f32f32 r4.y, r6.x -mul.f r6.x, r6.y, r6.y -mul.f r4.x, r4.x, r4.w -mul.f r2.z, c0.x, r2.z -mad.f32 r2.x, r0.x, r0.w, (neg)r2.x -mad.f32 r0.x, c0.z, r0.x, r4.y +mad.f32 r2.y, r2.y, r2.y, r2.w +mad.f32 r0.y, r0.y, r3.z, r3.x +mul.f r3.x, r3.w, r3.w +mul.f r2.w, c4.w, r4.z +mul.f r3.w, c4.z, r4.z +mul.f r5.w, c4.y, r4.z +mul.f r4.z, c4.x, r4.z +rsq r2.y, r2.y +(ss)mov.f32f32 r6.x, r2.y +rsq r0.y, r0.y +(ss)mov.f32f32 r6.y, r0.y +(ss)mul.f r0.y, r2.x, r0.y +mul.f r0.x, r0.x, r2.y +mul.f r2.x, r4.y, r6.x +mul.f r2.y, r5.z, r6.y +mul.f r4.x, r4.x, r6.x +absneg.f r4.y, (neg)r0.y +mov.f32f32 r0.y, r2.x +mov.f32f32 r5.z, r0.x +absneg.f r2.y, (neg)r2.y +mul.f r6.x, c0.x, r4.x mov.f32f32 r4.x, r4.x -mul.f r0.w, r3.y, r0.w -mov.f32f32 r2.x, r2.x -mov.f32f32 r0.x, r0.x -absneg.f r3.y, (neg)r4.x -mad.f32 r0.w, r3.z, r2.y, (neg)r0.w -mad.f32 r2.x, c0.y, r2.x, r2.z -mov.f32f32 r0.x, r0.x -mov.f32f32 r4.x, r3.y -mov.f32f32 r0.w, r0.w -mov.f32f32 r2.x, r2.x -mov.f32f32 r3.y, r3.w -mul.f r0.y, r0.y, r4.w -mad.f32 r0.w, c0.z, r0.w, r2.x -mad.f32 r2.x, r4.x, r4.x, r6.x -mul.f r2.y, r3.y, r3.y -mov.f32f32 r0.y, r0.y -mov.f32f32 r0.w, r0.w -mov.f32f32 r2.x, r2.x -mad.f32 r3.x, c7.z, r1.w, r3.x -mad.f32 r3.z, c5.y, r1.y, r4.z -mov.f32f32 r3.w, r0.w -absneg.f r0.y, (neg)r0.y -mov.f32f32 r2.z, r3.x -mad.f32 r0.w, c6.y, r1.z, r3.z -mad.f32 r2.y, r3.w, r3.w, r2.y -mov.f32f32 r3.z, r0.y -mad.f32 r0.y, c7.y, r1.w, r0.w -mad.f32 r0.w, c5.x, r1.y, r1.x -mov.f32f32 r1.x, r2.y -mad.f32 r1.y, r3.z, r3.z, r2.x -mad.f32 r1.x, r0.x, r0.x, r1.x +mul.f r6.z, r5.z, r0.z +mov.f32f32 r6.w, r2.y +mad.f32 r6.z, r0.y, r2.z, (neg)r6.z +mov.f32f32 r7.x, r4.y +mad.f32 r2.x, c0.y, r2.x, r6.x +mul.f r2.z, r4.x, r2.z +mul.f r6.x, c0.x, r6.z +mad.f32 r2.z, r5.z, r0.w, (neg)r2.z +mul.f r5.z, r7.x, r7.x +mad.f32 r0.x, c0.z, r0.x, r2.x +mad.f32 r2.x, r2.y, r6.w, r5.z +mad.f32 r2.y, c0.y, r2.z, r6.x +mul.f r0.y, r0.y, r0.w +mul.f r0.w, r3.z, r6.y +mad.f32 r0.y, r4.x, r0.z, (neg)r0.y +mov.f32f32 r0.z, r0.x +mad.f32 r2.z, c5.w, r4.w, r2.w +mad.f32 r2.w, c5.z, r4.w, r3.w +mad.f32 r0.y, c0.z, r0.y, r2.y +absneg.f r2.y, (neg)r0.w +mov.f32f32 r3.z, r0.z +mad.f32 r0.z, c6.w, r5.x, r2.z +mov.f32f32 r2.z, r0.y +mov.f32f32 r4.x, r2.y +mad.f32 r0.w, c7.w, r5.y, r0.z +mad.f32 r0.z, c6.z, r5.x, r2.w +mov.f32f32 r5.z, r2.z +mad.f32 r2.x, r2.y, r4.x, r2.x +mov.f32f32 r2.w, r0.w +mad.f32 r0.z, c7.z, r5.y, r0.z +mad.f32 r0.y, r0.y, r5.z, r3.x +mad.f32 r2.y, c5.y, r4.w, r5.w +mad.f32 r0.x, r0.x, r3.z, r0.y +mov.f32f32 r2.z, r0.z +mad.f32 r0.y, c6.y, r5.x, r2.y +mad.f32 r2.y, c5.x, r4.w, r4.z +mad.f32 r0.y, c7.y, r5.y, r0.y +mad.f32 r4.z, c6.x, r5.x, r2.y +nop +rsq r0.x, r0.x +(ss)mov.f32f32 r2.y, r0.x +mul.f r3.x, r3.y, r0.x +(ss)rsq r0.x, r2.x +(ss)mov.f32f32 r2.x, r0.x +mul.f r3.w, r4.y, r0.x +mul.f r3.z, r3.z, r2.y +mul.f r3.y, r5.z, r2.y +mul.f r4.y, r4.x, r2.x +mul.f r4.x, r6.w, r2.x mov.f32f32 r2.y, r0.y -mad.f32 r1.z, c6.x, r1.z, r0.w -mov.f32f32 r0.w, r0.z -mov.f32f32 r0.z, r3.x -mov.f32f32 r0.y, r0.y -mad.f32 r1.z, c7.x, r1.w, r1.z -rsq r1.x, r1.x -(ss)mov.f32f32 r1.x, r1.x -rsq r1.y, r1.y -(ss)mov.f32f32 r1.y, r1.y -mov.f32f32 r1.w, (0.000000) -mov.f32f32 r2.x, r1.z -mul.f r0.x, r0.x, r1.x -mul.f r3.x, r3.w, r1.x -mul.f r1.x, r3.y, r1.x -mul.f r3.y, r3.z, r1.y -mov.f32f32 r0.x, r0.x -mov.f32f32 r3.x, r3.x -mov.f32f32 r1.x, r1.x -mov.f32f32 r3.w, r3.y -mov.f32f32 r0.x, r0.x -mov.f32f32 r3.x, r3.x -mov.f32f32 r1.x, r1.x -mov.f32f32 r4.y, r3.w -mov.f32f32 r0.x, r0.x -mov.f32f32 r3.x, r3.x -mov.f32f32 r1.x, r1.x +mad.f32 r0.x, c7.x, r5.y, r4.z +mov.f32f32 r4.z, r4.y +mov.f32f32 r4.w, (0.000000) nop -mov.f32f32 r3.z, r0.x -mov.f32f32 r3.y, r3.x -mov.f32f32 r3.x, r1.x -mov.f32f32 r0.x, r4.y -mul.f r1.x, r4.x, r1.y -mul.f r1.y, r6.y, r1.y -mov.f32f32 r4.z, r3.w -mov.f32f32 r4.y, r0.x -mov.f32f32 r1.x, r1.x -mov.f32f32 r1.y, r1.y -mov.f32f32 r0.x, r1.z -mov.f32f32 r4.w, r1.w -mov.f32f32 r1.x, r1.x -mov.f32f32 r1.y, r1.y -(rpt1)nop -mov.f32f32 r1.x, r1.x -mov.f32f32 r3.w, r1.y -mov.f32f32 r1.w, r5.w -mov.f32f32 r1.z, r5.z -mov.f32f32 r4.x, r1.x -mov.f32f32 r1.y, r5.y -mov.f32f32 r1.x, r5.x +mov.f32f32 r2.x, r0.x end nop nop +nop ; VERT: outputs: r0.x (0:0) r1.x (5:0) r2.x (5:20) r3.x (5:21) r4.x (5:22) -; VERT: inputs: r1.x (0:0,cm=f,il=8,b=0) r0.x (0:0,cm=7,il=12,b=0) r5.x (0:0,cm=f,il=16,b=0) -; VERT: 183 instructions, 0 half, 7 full +; VERT: inputs: r4.z (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) +; VERT: 133 instructions, 0 half, 8 full |