Optimize commutative operations in SkVM.
We now canonicalize commutative operations by ordering their value IDs. The lower-numbered value ID is always placed first into a commutative instruction. In other words, this instruction: bit_and result, v7, v5 Would be silently converted to this: bit_and result, v5, v7 This will allow these two logically-equivalent instructions to be deduplicated: bit_and result, v7, v5 bit_and result, v5, v7 Of course, deduplicating these ops can unlock additional free CSE/DCE. The affected instructions are listed in http://review.skia.org/473238 Change-Id: Ib9beb79d6b72d7903184aaa9a53e8e5a02ae126d Reviewed-on: https://skia-review.googlesource.com/c/skia/+/473239 Commit-Queue: John Stiles <johnstiles@google.com> Auto-Submit: John Stiles <johnstiles@google.com> Reviewed-by: Brian Osman <brianosman@google.com> Reviewed-by: Herb Derby <herb@google.com>
This commit is contained in:
John Stiles
SkCQ
parent
626dbe195a
commit
c809e5ba9f
@ -765,7 +765,7 @@ namespace skvm {
|
||||
return {this, this->push(Op::fma_f32, fProgram[y.id].x, fProgram[y.id].y, x.id)};
|
||||
}
|
||||
}
|
||||
return {this, this->push(Op::add_f32, x.id, y.id)};
|
||||
return {this, this->push(Op::add_f32, std::min(x.id, y.id), std::max(x.id, y.id))};
|
||||
}
|
||||
|
||||
F32 Builder::sub(F32 x, F32 y) {
|
||||
@ -786,7 +786,7 @@ namespace skvm {
|
||||
if (float X,Y; this->allImm(x.id,&X, y.id,&Y)) { return splat(X*Y); }
|
||||
if (this->isImm(y.id, 1.0f)) { return x; } // x*1 == x
|
||||
if (this->isImm(x.id, 1.0f)) { return y; } // 1*y == y
|
||||
return {this, this->push(Op::mul_f32, x.id, y.id)};
|
||||
return {this, this->push(Op::mul_f32, std::min(x.id, y.id), std::max(x.id, y.id))};
|
||||
}
|
||||
|
||||
F32 Builder::fast_mul(F32 x, F32 y) {
|
||||
@ -977,7 +977,7 @@ namespace skvm {
|
||||
if (int X,Y; this->allImm(x.id,&X, y.id,&Y)) { return splat(X+Y); }
|
||||
if (this->isImm(x.id, 0)) { return y; }
|
||||
if (this->isImm(y.id, 0)) { return x; }
|
||||
return {this, this->push(Op::add_i32, x.id, y.id)};
|
||||
return {this, this->push(Op::add_i32, std::min(x.id, y.id), std::max(x.id, y.id))};
|
||||
}
|
||||
SK_ATTRIBUTE(no_sanitize("signed-integer-overflow"))
|
||||
I32 Builder::sub(I32 x, I32 y) {
|
||||
@ -992,7 +992,7 @@ namespace skvm {
|
||||
if (this->isImm(y.id, 0)) { return splat(0); }
|
||||
if (this->isImm(x.id, 1)) { return y; }
|
||||
if (this->isImm(y.id, 1)) { return x; }
|
||||
return {this, this->push(Op::mul_i32, x.id, y.id)};
|
||||
return {this, this->push(Op::mul_i32, std::min(x.id, y.id), std::max(x.id, y.id))};
|
||||
}
|
||||
|
||||
SK_ATTRIBUTE(no_sanitize("shift"))
|
||||
@ -1014,11 +1014,11 @@ namespace skvm {
|
||||
|
||||
I32 Builder:: eq(F32 x, F32 y) {
|
||||
if (float X,Y; this->allImm(x.id,&X, y.id,&Y)) { return splat(X==Y ? ~0 : 0); }
|
||||
return {this, this->push(Op::eq_f32, x.id, y.id)};
|
||||
return {this, this->push(Op::eq_f32, std::min(x.id, y.id), std::max(x.id, y.id))};
|
||||
}
|
||||
I32 Builder::neq(F32 x, F32 y) {
|
||||
if (float X,Y; this->allImm(x.id,&X, y.id,&Y)) { return splat(X!=Y ? ~0 : 0); }
|
||||
return {this, this->push(Op::neq_f32, x.id, y.id)};
|
||||
return {this, this->push(Op::neq_f32, std::min(x.id, y.id), std::max(x.id, y.id))};
|
||||
}
|
||||
I32 Builder::lt(F32 x, F32 y) {
|
||||
if (float X,Y; this->allImm(x.id,&X, y.id,&Y)) { return splat(Y> X ? ~0 : 0); }
|
||||
@ -1040,7 +1040,7 @@ namespace skvm {
|
||||
I32 Builder:: eq(I32 x, I32 y) {
|
||||
if (x.id == y.id) { return splat(~0); }
|
||||
if (int X,Y; this->allImm(x.id,&X, y.id,&Y)) { return splat(X==Y ? ~0 : 0); }
|
||||
return {this, this->push(Op:: eq_i32, x.id, y.id)};
|
||||
return {this, this->push(Op:: eq_i32, std::min(x.id, y.id), std::max(x.id, y.id))};
|
||||
}
|
||||
I32 Builder::neq(I32 x, I32 y) {
|
||||
if (int X,Y; this->allImm(x.id,&X, y.id,&Y)) { return splat(X!=Y ? ~0 : 0); }
|
||||
@ -1065,7 +1065,7 @@ namespace skvm {
|
||||
if (this->isImm(x.id, 0)) { return splat(0); } // (false & y) == false
|
||||
if (this->isImm(y.id,~0)) { return x; } // (x & true) == x
|
||||
if (this->isImm(x.id,~0)) { return y; } // (true & y) == y
|
||||
return {this, this->push(Op::bit_and, x.id, y.id)};
|
||||
return {this, this->push(Op::bit_and, std::min(x.id, y.id), std::max(x.id, y.id))};
|
||||
}
|
||||
I32 Builder::bit_or(I32 x, I32 y) {
|
||||
if (x.id == y.id) { return x; }
|
||||
@ -1074,14 +1074,14 @@ namespace skvm {
|
||||
if (this->isImm(x.id, 0)) { return y; } // (false | y) == y
|
||||
if (this->isImm(y.id,~0)) { return splat(~0); } // (x | true) == true
|
||||
if (this->isImm(x.id,~0)) { return splat(~0); } // (true | y) == true
|
||||
return {this, this->push(Op::bit_or, x.id, y.id)};
|
||||
return {this, this->push(Op::bit_or, std::min(x.id, y.id), std::max(x.id, y.id))};
|
||||
}
|
||||
I32 Builder::bit_xor(I32 x, I32 y) {
|
||||
if (x.id == y.id) { return splat(0); }
|
||||
if (int X,Y; this->allImm(x.id,&X, y.id,&Y)) { return splat(X^Y); }
|
||||
if (this->isImm(y.id, 0)) { return x; } // (x ^ false) == x
|
||||
if (this->isImm(x.id, 0)) { return y; } // (false ^ y) == y
|
||||
return {this, this->push(Op::bit_xor, x.id, y.id)};
|
||||
return {this, this->push(Op::bit_xor, std::min(x.id, y.id), std::max(x.id, y.id))};
|
||||
}
|
||||
|
||||
I32 Builder::bit_clear(I32 x, I32 y) {
|
||||
|
||||
@ -72,17 +72,17 @@ F4 = float index_out_of_bounds_checked()
|
||||
28 r28 = uniform32 ptr0 6C
|
||||
29 r29 = uniform32 ptr0 70
|
||||
30 r30 = splat 0 (0)
|
||||
31 r31 = eq_f32 r1 r0
|
||||
31 r31 = eq_f32 r0 r1
|
||||
32 r32 = splat 3 (4.2038954e-45)
|
||||
33 r33 = splat 2 (2.8025969e-45)
|
||||
34 r34 = add_f32 r9 r8
|
||||
34 r34 = add_f32 r8 r9
|
||||
35 r35 = splat 1 (1.4012985e-45)
|
||||
36 r36 = add_f32 r34 r7
|
||||
37 r37 = add_f32 r36 r6
|
||||
36 r36 = add_f32 r7 r34
|
||||
37 r37 = add_f32 r6 r36
|
||||
38 r38 = splat 3F800000 (1)
|
||||
39 r39 = mul_f32 r10 r10
|
||||
40 r40 = mul_f32 r39 r12
|
||||
41 r41 = mul_f32 r40 r13
|
||||
40 r40 = mul_f32 r12 r39
|
||||
41 r41 = mul_f32 r13 r40
|
||||
42 r42 = splat C0133333 (-2.3)
|
||||
43 r43 = splat 3FB33334 (1.4000001)
|
||||
44 r44 = sub_f32 r30 r15
|
||||
@ -95,7 +95,7 @@ F4 = float index_out_of_bounds_checked()
|
||||
51 r51 = splat 4181999A (16.200001)
|
||||
loop:
|
||||
52 r52 = index
|
||||
53 r53 = eq_f32 r52 r0
|
||||
53 r53 = eq_f32 r0 r52
|
||||
54 r53 = bit_and r53 r31
|
||||
55 trace_var r53 $0 = r2 (F32)
|
||||
56 trace_var r53 $1 = r3 (F32)
|
||||
|
||||
@ -1,4 +1,4 @@
|
||||
17 registers, 56 instructions:
|
||||
11 registers, 23 instructions:
|
||||
0 r0 = uniform32 ptr0 4
|
||||
1 r1 = uniform32 ptr0 8
|
||||
2 r2 = uniform32 ptr0 C
|
||||
@ -7,52 +7,19 @@
|
||||
5 r5 = uniform32 ptr0 18
|
||||
6 r6 = uniform32 ptr0 1C
|
||||
7 r7 = uniform32 ptr0 20
|
||||
8 r8 = uniform32 ptr0 24
|
||||
9 r9 = uniform32 ptr0 28
|
||||
10 r10 = uniform32 ptr0 2C
|
||||
11 r11 = uniform32 ptr0 30
|
||||
12 r12 = splat FFFFFFFF (nan)
|
||||
13 r13 = bit_and r8 r9
|
||||
14 r14 = bit_and r9 r8
|
||||
15 r14 = eq_i32 r13 r14
|
||||
16 r13 = bit_or r8 r9
|
||||
17 r15 = bit_or r9 r8
|
||||
18 r15 = eq_i32 r13 r15
|
||||
19 r15 = bit_and r14 r15
|
||||
20 r14 = bit_xor r8 r9
|
||||
21 r13 = bit_xor r9 r8
|
||||
22 r13 = eq_i32 r14 r13
|
||||
23 r13 = bit_and r15 r13
|
||||
24 r15 = eq_i32 r8 r9
|
||||
25 r14 = eq_i32 r9 r8
|
||||
26 r16 = eq_i32 r15 r14
|
||||
27 r16 = bit_and r13 r16
|
||||
28 r15 = bit_xor r12 r15
|
||||
29 r14 = bit_xor r12 r14
|
||||
30 r14 = eq_i32 r15 r14
|
||||
31 r14 = bit_and r16 r14
|
||||
32 r16 = add_i32 r8 r9
|
||||
33 r15 = add_i32 r9 r8
|
||||
34 r15 = eq_i32 r16 r15
|
||||
35 r15 = bit_and r14 r15
|
||||
36 r14 = add_f32 r10 r11
|
||||
37 r16 = add_f32 r11 r10
|
||||
38 r16 = eq_f32 r14 r16
|
||||
39 r16 = bit_and r15 r16
|
||||
40 r15 = mul_i32 r8 r9
|
||||
41 r8 = mul_i32 r9 r8
|
||||
42 r8 = eq_i32 r15 r8
|
||||
43 r8 = bit_and r16 r8
|
||||
44 r16 = mul_f32 r10 r11
|
||||
45 r10 = mul_f32 r11 r10
|
||||
46 r10 = eq_f32 r16 r10
|
||||
47 r10 = bit_and r8 r10
|
||||
48 r4 = select r10 r0 r4
|
||||
49 r5 = select r10 r1 r5
|
||||
50 r6 = select r10 r2 r6
|
||||
51 r7 = select r10 r3 r7
|
||||
8 r8 = uniform32 ptr0 2C
|
||||
9 r9 = uniform32 ptr0 30
|
||||
10 r10 = add_f32 r8 r9
|
||||
11 r10 = eq_f32 r10 r10
|
||||
12 r9 = mul_f32 r8 r9
|
||||
13 r9 = eq_f32 r9 r9
|
||||
14 r9 = bit_and r10 r9
|
||||
15 r4 = select r9 r0 r4
|
||||
16 r5 = select r9 r1 r5
|
||||
17 r6 = select r9 r2 r6
|
||||
18 r7 = select r9 r3 r7
|
||||
loop:
|
||||
52 store32 ptr1 r4
|
||||
53 store32 ptr2 r5
|
||||
54 store32 ptr3 r6
|
||||
55 store32 ptr4 r7
|
||||
19 store32 ptr1 r4
|
||||
20 store32 ptr2 r5
|
||||
21 store32 ptr3 r6
|
||||
22 store32 ptr4 r7
|
||||
|
||||
@ -6,15 +6,15 @@
|
||||
4 r4 = splat FFFFFFFF (nan)
|
||||
5 r5 = neq_f32 r0 r3
|
||||
6 r6 = splat 3F800000 (1)
|
||||
7 r6 = bit_and r6 r2
|
||||
7 r6 = bit_and r2 r6
|
||||
8 r7 = to_f32 r1
|
||||
9 r7 = select r5 r6 r7
|
||||
10 r3 = eq_i32 r1 r3
|
||||
11 r3 = bit_xor r4 r3
|
||||
12 r4 = splat 1 (1.4012985e-45)
|
||||
13 r2 = bit_and r4 r2
|
||||
13 r4 = bit_and r2 r4
|
||||
14 r0 = trunc r0
|
||||
15 r0 = select r3 r2 r0
|
||||
15 r0 = select r3 r4 r0
|
||||
16 r0 = to_f32 r0
|
||||
17 r0 = mul_f32 r7 r0
|
||||
loop:
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
File diff suppressed because it is too large
Load Diff
@ -156,7 +156,7 @@ F7 = bool lowp_param(half value)
|
||||
8 r8 = uniform32 ptr0 18
|
||||
9 r9 = uniform32 ptr0 1C
|
||||
10 r10 = uniform32 ptr0 20
|
||||
11 r11 = eq_f32 r2 r1
|
||||
11 r11 = eq_f32 r1 r2
|
||||
12 r12 = splat 3F800000 (1)
|
||||
13 r13 = splat 2 (2.8025969e-45)
|
||||
14 r14 = splat 40000000 (2)
|
||||
@ -166,7 +166,7 @@ F7 = bool lowp_param(half value)
|
||||
18 r18 = splat 4 (5.6051939e-45)
|
||||
loop:
|
||||
19 r19 = index
|
||||
20 r20 = eq_f32 r19 r1
|
||||
20 r20 = eq_f32 r1 r19
|
||||
21 r20 = bit_and r20 r11
|
||||
22 trace_var r20 $0 = r3 (F32)
|
||||
23 trace_var r20 $1 = r4 (F32)
|
||||
|
||||
@ -8,7 +8,7 @@
|
||||
loop:
|
||||
6 r5 = index
|
||||
7 r5 = trunc r5
|
||||
8 r5 = mul_i32 r5 r1
|
||||
8 r5 = mul_i32 r1 r5
|
||||
9 r5 = add_i32 r0 r5
|
||||
10 r5 = shl_i32 r5 2
|
||||
11 r6 = gather32 ptr0 4 r5
|
||||
|
||||
@ -18,19 +18,19 @@
|
||||
16 r15 = bit_xor r9 r8
|
||||
17 r8 = bit_and r8 r15
|
||||
18 r16 = splat 1 (1.4012985e-45)
|
||||
19 r10 = eq_i32 r16 r10
|
||||
20 r10 = bit_or r8 r10
|
||||
21 r10 = bit_and r10 r15
|
||||
22 r11 = select r10 r0 r11
|
||||
23 r12 = select r10 r1 r12
|
||||
24 r13 = select r10 r2 r13
|
||||
25 r14 = select r10 r3 r14
|
||||
26 r10 = bit_xor r9 r10
|
||||
27 r10 = bit_and r15 r10
|
||||
28 r11 = select r10 r4 r11
|
||||
29 r12 = select r10 r5 r12
|
||||
30 r13 = select r10 r6 r13
|
||||
31 r14 = select r10 r7 r14
|
||||
19 r16 = eq_i32 r10 r16
|
||||
20 r16 = bit_or r8 r16
|
||||
21 r16 = bit_and r15 r16
|
||||
22 r11 = select r16 r0 r11
|
||||
23 r12 = select r16 r1 r12
|
||||
24 r13 = select r16 r2 r13
|
||||
25 r14 = select r16 r3 r14
|
||||
26 r16 = bit_xor r9 r16
|
||||
27 r16 = bit_and r15 r16
|
||||
28 r11 = select r16 r4 r11
|
||||
29 r12 = select r16 r5 r12
|
||||
30 r13 = select r16 r6 r13
|
||||
31 r14 = select r16 r7 r14
|
||||
loop:
|
||||
32 store32 ptr1 r11
|
||||
33 store32 ptr2 r12
|
||||
|
||||
@ -11,39 +11,39 @@
|
||||
9 r9 = splat FFFFFFFF (nan)
|
||||
10 r10 = trunc r1
|
||||
11 r11 = splat 2 (2.8025969e-45)
|
||||
12 r11 = eq_i32 r11 r10
|
||||
12 r11 = eq_i32 r10 r11
|
||||
13 r12 = bit_xor r9 r11
|
||||
14 r13 = bit_and r11 r12
|
||||
15 r14 = splat 1 (1.4012985e-45)
|
||||
16 r14 = eq_i32 r14 r10
|
||||
16 r14 = eq_i32 r10 r14
|
||||
17 r15 = bit_or r13 r14
|
||||
18 r15 = bit_and r15 r12
|
||||
18 r15 = bit_and r12 r15
|
||||
19 r13 = select r15 r9 r13
|
||||
20 r8 = eq_i32 r8 r10
|
||||
21 r13 = bit_or r13 r8
|
||||
22 r12 = bit_and r13 r12
|
||||
23 r8 = bit_and r12 r8
|
||||
24 r13 = bit_xor r9 r8
|
||||
25 r8 = bit_and r8 r13
|
||||
26 r14 = bit_or r8 r14
|
||||
27 r14 = bit_and r12 r14
|
||||
28 r14 = bit_and r14 r13
|
||||
22 r13 = bit_and r12 r13
|
||||
23 r8 = bit_and r8 r13
|
||||
24 r12 = bit_xor r9 r8
|
||||
25 r8 = bit_and r8 r12
|
||||
26 r14 = bit_or r14 r8
|
||||
27 r14 = bit_and r13 r14
|
||||
28 r14 = bit_and r12 r14
|
||||
29 r8 = select r14 r9 r8
|
||||
30 r11 = bit_or r8 r11
|
||||
31 r11 = bit_and r12 r11
|
||||
32 r11 = bit_and r11 r13
|
||||
30 r11 = bit_or r11 r8
|
||||
31 r11 = bit_and r13 r11
|
||||
32 r11 = bit_and r12 r11
|
||||
33 r8 = select r11 r9 r8
|
||||
34 r9 = splat 3 (4.2038954e-45)
|
||||
35 r10 = eq_i32 r9 r10
|
||||
36 r10 = bit_or r8 r10
|
||||
37 r10 = bit_and r12 r10
|
||||
38 r13 = bit_and r10 r13
|
||||
39 r13 = bit_and r13 r12
|
||||
40 r13 = bit_and r12 r13
|
||||
41 r4 = select r13 r0 r4
|
||||
42 r5 = select r13 r1 r5
|
||||
43 r6 = select r13 r2 r6
|
||||
44 r7 = select r13 r3 r7
|
||||
35 r9 = eq_i32 r10 r9
|
||||
36 r9 = bit_or r8 r9
|
||||
37 r9 = bit_and r13 r9
|
||||
38 r9 = bit_and r12 r9
|
||||
39 r9 = bit_and r13 r9
|
||||
40 r9 = bit_and r13 r9
|
||||
41 r4 = select r9 r0 r4
|
||||
42 r5 = select r9 r1 r5
|
||||
43 r6 = select r9 r2 r6
|
||||
44 r7 = select r9 r3 r7
|
||||
loop:
|
||||
45 store32 ptr1 r4
|
||||
46 store32 ptr2 r5
|
||||
|
||||
@ -10,286 +10,286 @@
|
||||
8 r8 = splat FFFFFFFF (nan)
|
||||
9 r9 = trunc r1
|
||||
10 r10 = splat 1 (1.4012985e-45)
|
||||
11 r9 = eq_i32 r10 r9
|
||||
11 r9 = eq_i32 r9 r10
|
||||
12 r11 = bit_and r10 r9
|
||||
13 r12 = bit_xor r8 r9
|
||||
14 r13 = add_i32 r11 r10
|
||||
14 r13 = add_i32 r10 r11
|
||||
15 r14 = bit_and r9 r12
|
||||
16 r11 = select r14 r13 r11
|
||||
17 r13 = add_i32 r11 r10
|
||||
17 r13 = add_i32 r10 r11
|
||||
18 r11 = select r14 r13 r11
|
||||
19 r14 = bit_xor r8 r14
|
||||
20 r14 = bit_and r12 r14
|
||||
21 r12 = add_i32 r11 r10
|
||||
21 r12 = add_i32 r10 r11
|
||||
22 r13 = bit_and r9 r14
|
||||
23 r11 = select r13 r12 r11
|
||||
24 r12 = splat 2 (2.8025969e-45)
|
||||
25 r15 = add_i32 r11 r10
|
||||
25 r15 = add_i32 r10 r11
|
||||
26 r11 = select r13 r15 r11
|
||||
27 r13 = bit_xor r8 r13
|
||||
28 r13 = bit_and r14 r13
|
||||
29 r14 = add_i32 r11 r10
|
||||
29 r14 = add_i32 r10 r11
|
||||
30 r15 = bit_and r9 r13
|
||||
31 r11 = select r15 r14 r11
|
||||
32 r14 = add_i32 r11 r10
|
||||
32 r14 = add_i32 r10 r11
|
||||
33 r11 = select r15 r14 r11
|
||||
34 r15 = bit_xor r8 r15
|
||||
35 r15 = bit_and r13 r15
|
||||
36 r13 = add_i32 r11 r10
|
||||
36 r13 = add_i32 r10 r11
|
||||
37 r14 = bit_and r9 r15
|
||||
38 r11 = select r14 r13 r11
|
||||
39 r13 = add_i32 r11 r10
|
||||
39 r13 = add_i32 r10 r11
|
||||
40 r11 = select r14 r13 r11
|
||||
41 r14 = bit_xor r8 r14
|
||||
42 r14 = bit_and r15 r14
|
||||
43 r15 = add_i32 r11 r10
|
||||
43 r15 = add_i32 r10 r11
|
||||
44 r13 = bit_and r9 r14
|
||||
45 r11 = select r13 r15 r11
|
||||
46 r15 = add_i32 r11 r10
|
||||
46 r15 = add_i32 r10 r11
|
||||
47 r11 = select r13 r15 r11
|
||||
48 r13 = bit_xor r8 r13
|
||||
49 r13 = bit_and r14 r13
|
||||
50 r14 = add_i32 r11 r10
|
||||
50 r14 = add_i32 r10 r11
|
||||
51 r15 = bit_and r9 r13
|
||||
52 r11 = select r15 r14 r11
|
||||
53 r14 = add_i32 r11 r10
|
||||
53 r14 = add_i32 r10 r11
|
||||
54 r11 = select r15 r14 r11
|
||||
55 r15 = bit_xor r8 r15
|
||||
56 r15 = bit_and r13 r15
|
||||
57 r13 = add_i32 r11 r10
|
||||
57 r13 = add_i32 r10 r11
|
||||
58 r14 = bit_and r9 r15
|
||||
59 r11 = select r14 r13 r11
|
||||
60 r13 = add_i32 r11 r10
|
||||
60 r13 = add_i32 r10 r11
|
||||
61 r11 = select r14 r13 r11
|
||||
62 r14 = bit_xor r8 r14
|
||||
63 r14 = bit_and r15 r14
|
||||
64 r15 = add_i32 r11 r10
|
||||
64 r15 = add_i32 r10 r11
|
||||
65 r13 = bit_and r9 r14
|
||||
66 r11 = select r13 r15 r11
|
||||
67 r15 = add_i32 r11 r10
|
||||
67 r15 = add_i32 r10 r11
|
||||
68 r11 = select r13 r15 r11
|
||||
69 r13 = bit_xor r8 r13
|
||||
70 r13 = bit_and r14 r13
|
||||
71 r14 = add_i32 r11 r10
|
||||
71 r14 = add_i32 r10 r11
|
||||
72 r15 = bit_and r9 r13
|
||||
73 r11 = select r15 r14 r11
|
||||
74 r14 = add_i32 r11 r10
|
||||
74 r14 = add_i32 r10 r11
|
||||
75 r11 = select r15 r14 r11
|
||||
76 r15 = bit_xor r8 r15
|
||||
77 r15 = bit_and r13 r15
|
||||
78 r13 = add_i32 r11 r10
|
||||
78 r13 = add_i32 r10 r11
|
||||
79 r15 = bit_and r9 r15
|
||||
80 r11 = select r15 r13 r11
|
||||
81 r11 = add_i32 r11 r10
|
||||
82 r12 = eq_i32 r11 r12
|
||||
83 r11 = bit_and r12 r9
|
||||
84 r13 = bit_and r10 r11
|
||||
85 r15 = bit_xor r8 r11
|
||||
86 r14 = add_i32 r13 r10
|
||||
87 r16 = bit_and r11 r15
|
||||
81 r11 = add_i32 r10 r11
|
||||
82 r11 = eq_i32 r12 r11
|
||||
83 r12 = bit_and r9 r11
|
||||
84 r13 = bit_and r10 r12
|
||||
85 r15 = bit_xor r8 r12
|
||||
86 r14 = add_i32 r10 r13
|
||||
87 r16 = bit_and r12 r15
|
||||
88 r13 = select r16 r14 r13
|
||||
89 r15 = bit_or r15 r11
|
||||
90 r14 = add_i32 r13 r10
|
||||
91 r16 = bit_and r11 r15
|
||||
89 r15 = bit_or r12 r15
|
||||
90 r14 = add_i32 r10 r13
|
||||
91 r16 = bit_and r12 r15
|
||||
92 r13 = select r16 r14 r13
|
||||
93 r14 = bit_xor r8 r16
|
||||
94 r14 = bit_and r15 r14
|
||||
95 r15 = add_i32 r13 r10
|
||||
96 r17 = bit_and r11 r14
|
||||
95 r15 = add_i32 r10 r13
|
||||
96 r17 = bit_and r12 r14
|
||||
97 r13 = select r17 r15 r13
|
||||
98 r16 = bit_or r14 r16
|
||||
99 r14 = add_i32 r13 r10
|
||||
100 r15 = bit_and r11 r16
|
||||
101 r13 = select r15 r14 r13
|
||||
102 r14 = bit_xor r8 r15
|
||||
103 r14 = bit_and r16 r14
|
||||
104 r16 = add_i32 r13 r10
|
||||
105 r17 = bit_and r11 r14
|
||||
106 r13 = select r17 r16 r13
|
||||
107 r15 = bit_or r14 r15
|
||||
108 r14 = add_i32 r13 r10
|
||||
109 r16 = bit_and r11 r15
|
||||
110 r13 = select r16 r14 r13
|
||||
111 r14 = bit_xor r8 r16
|
||||
112 r14 = bit_and r15 r14
|
||||
113 r15 = add_i32 r13 r10
|
||||
114 r17 = bit_and r11 r14
|
||||
115 r13 = select r17 r15 r13
|
||||
116 r16 = bit_or r14 r16
|
||||
117 r14 = add_i32 r13 r10
|
||||
118 r15 = bit_and r11 r16
|
||||
119 r13 = select r15 r14 r13
|
||||
120 r14 = bit_xor r8 r15
|
||||
121 r14 = bit_and r16 r14
|
||||
122 r16 = add_i32 r13 r10
|
||||
123 r17 = bit_and r11 r14
|
||||
124 r13 = select r17 r16 r13
|
||||
125 r15 = bit_or r14 r15
|
||||
126 r14 = add_i32 r13 r10
|
||||
127 r16 = bit_and r11 r15
|
||||
128 r13 = select r16 r14 r13
|
||||
129 r14 = bit_xor r8 r16
|
||||
130 r14 = bit_and r15 r14
|
||||
131 r15 = add_i32 r13 r10
|
||||
132 r17 = bit_and r11 r14
|
||||
133 r13 = select r17 r15 r13
|
||||
134 r16 = bit_or r14 r16
|
||||
135 r14 = add_i32 r13 r10
|
||||
136 r15 = bit_and r11 r16
|
||||
137 r13 = select r15 r14 r13
|
||||
138 r14 = bit_xor r8 r15
|
||||
139 r14 = bit_and r16 r14
|
||||
140 r16 = add_i32 r13 r10
|
||||
141 r17 = bit_and r11 r14
|
||||
98 r14 = bit_or r16 r14
|
||||
99 r16 = add_i32 r10 r13
|
||||
100 r15 = bit_and r12 r14
|
||||
101 r13 = select r15 r16 r13
|
||||
102 r16 = bit_xor r8 r15
|
||||
103 r16 = bit_and r14 r16
|
||||
104 r14 = add_i32 r10 r13
|
||||
105 r17 = bit_and r12 r16
|
||||
106 r13 = select r17 r14 r13
|
||||
107 r16 = bit_or r15 r16
|
||||
108 r15 = add_i32 r10 r13
|
||||
109 r14 = bit_and r12 r16
|
||||
110 r13 = select r14 r15 r13
|
||||
111 r15 = bit_xor r8 r14
|
||||
112 r15 = bit_and r16 r15
|
||||
113 r16 = add_i32 r10 r13
|
||||
114 r17 = bit_and r12 r15
|
||||
115 r13 = select r17 r16 r13
|
||||
116 r15 = bit_or r14 r15
|
||||
117 r14 = add_i32 r10 r13
|
||||
118 r16 = bit_and r12 r15
|
||||
119 r13 = select r16 r14 r13
|
||||
120 r14 = bit_xor r8 r16
|
||||
121 r14 = bit_and r15 r14
|
||||
122 r15 = add_i32 r10 r13
|
||||
123 r17 = bit_and r12 r14
|
||||
124 r13 = select r17 r15 r13
|
||||
125 r14 = bit_or r16 r14
|
||||
126 r16 = add_i32 r10 r13
|
||||
127 r15 = bit_and r12 r14
|
||||
128 r13 = select r15 r16 r13
|
||||
129 r16 = bit_xor r8 r15
|
||||
130 r16 = bit_and r14 r16
|
||||
131 r14 = add_i32 r10 r13
|
||||
132 r17 = bit_and r12 r16
|
||||
133 r13 = select r17 r14 r13
|
||||
134 r16 = bit_or r15 r16
|
||||
135 r15 = add_i32 r10 r13
|
||||
136 r14 = bit_and r12 r16
|
||||
137 r13 = select r14 r15 r13
|
||||
138 r15 = bit_xor r8 r14
|
||||
139 r15 = bit_and r16 r15
|
||||
140 r16 = add_i32 r10 r13
|
||||
141 r17 = bit_and r12 r15
|
||||
142 r13 = select r17 r16 r13
|
||||
143 r15 = bit_or r14 r15
|
||||
144 r14 = add_i32 r13 r10
|
||||
145 r16 = bit_and r11 r15
|
||||
144 r14 = add_i32 r10 r13
|
||||
145 r16 = bit_and r12 r15
|
||||
146 r13 = select r16 r14 r13
|
||||
147 r14 = bit_xor r8 r16
|
||||
148 r14 = bit_and r15 r14
|
||||
149 r15 = add_i32 r13 r10
|
||||
150 r17 = bit_and r11 r14
|
||||
149 r15 = add_i32 r10 r13
|
||||
150 r17 = bit_and r12 r14
|
||||
151 r13 = select r17 r15 r13
|
||||
152 r16 = bit_or r14 r16
|
||||
153 r14 = add_i32 r13 r10
|
||||
154 r15 = bit_and r11 r16
|
||||
155 r13 = select r15 r14 r13
|
||||
156 r14 = bit_xor r8 r15
|
||||
157 r14 = bit_and r16 r14
|
||||
158 r16 = add_i32 r13 r10
|
||||
159 r17 = bit_and r11 r14
|
||||
160 r13 = select r17 r16 r13
|
||||
161 r15 = bit_or r14 r15
|
||||
162 r14 = add_i32 r13 r10
|
||||
163 r16 = bit_and r11 r15
|
||||
164 r13 = select r16 r14 r13
|
||||
165 r16 = bit_xor r8 r16
|
||||
166 r16 = bit_and r15 r16
|
||||
167 r15 = add_i32 r13 r10
|
||||
168 r16 = bit_and r11 r16
|
||||
169 r13 = select r16 r15 r13
|
||||
170 r15 = add_i32 r13 r10
|
||||
171 r13 = select r12 r15 r13
|
||||
172 r15 = splat B (1.5414283e-44)
|
||||
173 r15 = eq_i32 r13 r15
|
||||
174 r15 = bit_and r15 r12
|
||||
175 r15 = bit_and r12 r15
|
||||
176 r9 = bit_and r15 r9
|
||||
177 r12 = bit_and r10 r9
|
||||
152 r14 = bit_or r16 r14
|
||||
153 r16 = add_i32 r10 r13
|
||||
154 r15 = bit_and r12 r14
|
||||
155 r13 = select r15 r16 r13
|
||||
156 r16 = bit_xor r8 r15
|
||||
157 r16 = bit_and r14 r16
|
||||
158 r14 = add_i32 r10 r13
|
||||
159 r17 = bit_and r12 r16
|
||||
160 r13 = select r17 r14 r13
|
||||
161 r16 = bit_or r15 r16
|
||||
162 r15 = add_i32 r10 r13
|
||||
163 r14 = bit_and r12 r16
|
||||
164 r13 = select r14 r15 r13
|
||||
165 r14 = bit_xor r8 r14
|
||||
166 r14 = bit_and r16 r14
|
||||
167 r16 = add_i32 r10 r13
|
||||
168 r14 = bit_and r12 r14
|
||||
169 r13 = select r14 r16 r13
|
||||
170 r16 = add_i32 r10 r13
|
||||
171 r13 = select r11 r16 r13
|
||||
172 r16 = splat B (1.5414283e-44)
|
||||
173 r16 = eq_i32 r13 r16
|
||||
174 r16 = bit_and r11 r16
|
||||
175 r16 = bit_and r11 r16
|
||||
176 r9 = bit_and r9 r16
|
||||
177 r11 = bit_and r10 r9
|
||||
178 r13 = bit_xor r8 r9
|
||||
179 r13 = bit_and r15 r13
|
||||
180 r16 = add_i32 r12 r10
|
||||
181 r11 = bit_xor r8 r13
|
||||
182 r14 = bit_and r15 r11
|
||||
183 r12 = select r14 r16 r12
|
||||
184 r16 = add_i32 r12 r10
|
||||
185 r14 = bit_and r9 r11
|
||||
186 r12 = select r14 r16 r12
|
||||
187 r14 = bit_xor r8 r14
|
||||
188 r14 = bit_and r15 r14
|
||||
189 r11 = bit_and r14 r11
|
||||
190 r11 = bit_or r13 r11
|
||||
191 r13 = add_i32 r12 r10
|
||||
192 r14 = bit_xor r8 r11
|
||||
193 r16 = bit_and r15 r14
|
||||
194 r12 = select r16 r13 r12
|
||||
195 r13 = add_i32 r12 r10
|
||||
196 r16 = bit_and r9 r14
|
||||
197 r12 = select r16 r13 r12
|
||||
198 r16 = bit_xor r8 r16
|
||||
199 r16 = bit_and r15 r16
|
||||
200 r14 = bit_and r16 r14
|
||||
201 r14 = bit_or r11 r14
|
||||
202 r11 = add_i32 r12 r10
|
||||
203 r16 = bit_xor r8 r14
|
||||
204 r13 = bit_and r15 r16
|
||||
205 r12 = select r13 r11 r12
|
||||
206 r11 = add_i32 r12 r10
|
||||
207 r13 = bit_and r9 r16
|
||||
208 r12 = select r13 r11 r12
|
||||
179 r13 = bit_and r16 r13
|
||||
180 r14 = add_i32 r10 r11
|
||||
181 r12 = bit_xor r8 r13
|
||||
182 r15 = bit_and r16 r12
|
||||
183 r11 = select r15 r14 r11
|
||||
184 r14 = add_i32 r10 r11
|
||||
185 r15 = bit_and r9 r12
|
||||
186 r11 = select r15 r14 r11
|
||||
187 r15 = bit_xor r8 r15
|
||||
188 r15 = bit_and r16 r15
|
||||
189 r15 = bit_and r12 r15
|
||||
190 r15 = bit_or r13 r15
|
||||
191 r13 = add_i32 r10 r11
|
||||
192 r12 = bit_xor r8 r15
|
||||
193 r14 = bit_and r16 r12
|
||||
194 r11 = select r14 r13 r11
|
||||
195 r13 = add_i32 r10 r11
|
||||
196 r14 = bit_and r9 r12
|
||||
197 r11 = select r14 r13 r11
|
||||
198 r14 = bit_xor r8 r14
|
||||
199 r14 = bit_and r16 r14
|
||||
200 r14 = bit_and r12 r14
|
||||
201 r14 = bit_or r15 r14
|
||||
202 r15 = add_i32 r10 r11
|
||||
203 r12 = bit_xor r8 r14
|
||||
204 r13 = bit_and r16 r12
|
||||
205 r11 = select r13 r15 r11
|
||||
206 r15 = add_i32 r10 r11
|
||||
207 r13 = bit_and r9 r12
|
||||
208 r11 = select r13 r15 r11
|
||||
209 r13 = bit_xor r8 r13
|
||||
210 r13 = bit_and r15 r13
|
||||
211 r16 = bit_and r13 r16
|
||||
212 r16 = bit_or r14 r16
|
||||
213 r14 = add_i32 r12 r10
|
||||
214 r13 = bit_xor r8 r16
|
||||
215 r11 = bit_and r15 r13
|
||||
216 r12 = select r11 r14 r12
|
||||
217 r14 = add_i32 r12 r10
|
||||
218 r11 = bit_and r9 r13
|
||||
219 r12 = select r11 r14 r12
|
||||
220 r11 = bit_xor r8 r11
|
||||
221 r11 = bit_and r15 r11
|
||||
222 r13 = bit_and r11 r13
|
||||
223 r13 = bit_or r16 r13
|
||||
224 r16 = add_i32 r12 r10
|
||||
225 r11 = bit_xor r8 r13
|
||||
226 r14 = bit_and r15 r11
|
||||
227 r12 = select r14 r16 r12
|
||||
228 r16 = add_i32 r12 r10
|
||||
229 r14 = bit_and r9 r11
|
||||
230 r12 = select r14 r16 r12
|
||||
210 r13 = bit_and r16 r13
|
||||
211 r13 = bit_and r12 r13
|
||||
212 r13 = bit_or r14 r13
|
||||
213 r14 = add_i32 r10 r11
|
||||
214 r12 = bit_xor r8 r13
|
||||
215 r15 = bit_and r16 r12
|
||||
216 r11 = select r15 r14 r11
|
||||
217 r14 = add_i32 r10 r11
|
||||
218 r15 = bit_and r9 r12
|
||||
219 r11 = select r15 r14 r11
|
||||
220 r15 = bit_xor r8 r15
|
||||
221 r15 = bit_and r16 r15
|
||||
222 r15 = bit_and r12 r15
|
||||
223 r15 = bit_or r13 r15
|
||||
224 r13 = add_i32 r10 r11
|
||||
225 r12 = bit_xor r8 r15
|
||||
226 r14 = bit_and r16 r12
|
||||
227 r11 = select r14 r13 r11
|
||||
228 r13 = add_i32 r10 r11
|
||||
229 r14 = bit_and r9 r12
|
||||
230 r11 = select r14 r13 r11
|
||||
231 r14 = bit_xor r8 r14
|
||||
232 r14 = bit_and r15 r14
|
||||
233 r11 = bit_and r14 r11
|
||||
234 r11 = bit_or r13 r11
|
||||
235 r13 = add_i32 r12 r10
|
||||
236 r14 = bit_xor r8 r11
|
||||
237 r16 = bit_and r15 r14
|
||||
238 r12 = select r16 r13 r12
|
||||
239 r13 = add_i32 r12 r10
|
||||
240 r16 = bit_and r9 r14
|
||||
241 r12 = select r16 r13 r12
|
||||
242 r16 = bit_xor r8 r16
|
||||
243 r16 = bit_and r15 r16
|
||||
244 r14 = bit_and r16 r14
|
||||
245 r14 = bit_or r11 r14
|
||||
246 r11 = add_i32 r12 r10
|
||||
247 r16 = bit_xor r8 r14
|
||||
248 r13 = bit_and r15 r16
|
||||
249 r12 = select r13 r11 r12
|
||||
250 r11 = add_i32 r12 r10
|
||||
251 r13 = bit_and r9 r16
|
||||
252 r12 = select r13 r11 r12
|
||||
253 r13 = bit_xor r8 r13
|
||||
254 r13 = bit_and r15 r13
|
||||
255 r16 = bit_and r13 r16
|
||||
256 r16 = bit_or r14 r16
|
||||
257 r14 = add_i32 r12 r10
|
||||
258 r13 = bit_xor r8 r16
|
||||
259 r11 = bit_and r15 r13
|
||||
260 r12 = select r11 r14 r12
|
||||
261 r14 = add_i32 r12 r10
|
||||
262 r11 = bit_and r9 r13
|
||||
263 r12 = select r11 r14 r12
|
||||
264 r11 = bit_xor r8 r11
|
||||
265 r11 = bit_and r15 r11
|
||||
266 r13 = bit_and r11 r13
|
||||
267 r13 = bit_or r16 r13
|
||||
268 r16 = add_i32 r12 r10
|
||||
269 r11 = bit_xor r8 r13
|
||||
270 r14 = bit_and r15 r11
|
||||
271 r12 = select r14 r16 r12
|
||||
272 r16 = add_i32 r12 r10
|
||||
273 r9 = bit_and r9 r11
|
||||
274 r12 = select r9 r16 r12
|
||||
232 r14 = bit_and r16 r14
|
||||
233 r14 = bit_and r12 r14
|
||||
234 r14 = bit_or r15 r14
|
||||
235 r15 = add_i32 r10 r11
|
||||
236 r12 = bit_xor r8 r14
|
||||
237 r13 = bit_and r16 r12
|
||||
238 r11 = select r13 r15 r11
|
||||
239 r15 = add_i32 r10 r11
|
||||
240 r13 = bit_and r9 r12
|
||||
241 r11 = select r13 r15 r11
|
||||
242 r13 = bit_xor r8 r13
|
||||
243 r13 = bit_and r16 r13
|
||||
244 r13 = bit_and r12 r13
|
||||
245 r13 = bit_or r14 r13
|
||||
246 r14 = add_i32 r10 r11
|
||||
247 r12 = bit_xor r8 r13
|
||||
248 r15 = bit_and r16 r12
|
||||
249 r11 = select r15 r14 r11
|
||||
250 r14 = add_i32 r10 r11
|
||||
251 r15 = bit_and r9 r12
|
||||
252 r11 = select r15 r14 r11
|
||||
253 r15 = bit_xor r8 r15
|
||||
254 r15 = bit_and r16 r15
|
||||
255 r15 = bit_and r12 r15
|
||||
256 r15 = bit_or r13 r15
|
||||
257 r13 = add_i32 r10 r11
|
||||
258 r12 = bit_xor r8 r15
|
||||
259 r14 = bit_and r16 r12
|
||||
260 r11 = select r14 r13 r11
|
||||
261 r13 = add_i32 r10 r11
|
||||
262 r14 = bit_and r9 r12
|
||||
263 r11 = select r14 r13 r11
|
||||
264 r14 = bit_xor r8 r14
|
||||
265 r14 = bit_and r16 r14
|
||||
266 r14 = bit_and r12 r14
|
||||
267 r14 = bit_or r15 r14
|
||||
268 r15 = add_i32 r10 r11
|
||||
269 r12 = bit_xor r8 r14
|
||||
270 r13 = bit_and r16 r12
|
||||
271 r11 = select r13 r15 r11
|
||||
272 r15 = add_i32 r10 r11
|
||||
273 r9 = bit_and r9 r12
|
||||
274 r11 = select r9 r15 r11
|
||||
275 r9 = bit_xor r8 r9
|
||||
276 r9 = bit_and r15 r9
|
||||
277 r11 = bit_and r9 r11
|
||||
278 r11 = bit_or r13 r11
|
||||
279 r10 = add_i32 r12 r10
|
||||
280 r11 = bit_xor r8 r11
|
||||
281 r11 = bit_and r15 r11
|
||||
282 r12 = select r11 r10 r12
|
||||
276 r9 = bit_and r16 r9
|
||||
277 r9 = bit_and r12 r9
|
||||
278 r9 = bit_or r14 r9
|
||||
279 r10 = add_i32 r10 r11
|
||||
280 r9 = bit_xor r8 r9
|
||||
281 r9 = bit_and r16 r9
|
||||
282 r11 = select r9 r10 r11
|
||||
283 r10 = splat 14 (2.8025969e-44)
|
||||
284 r10 = eq_i32 r12 r10
|
||||
285 r11 = bit_and r10 r11
|
||||
286 r11 = bit_and r15 r11
|
||||
287 r4 = select r11 r0 r4
|
||||
288 r5 = select r11 r1 r5
|
||||
289 r6 = select r11 r2 r6
|
||||
290 r7 = select r11 r3 r7
|
||||
284 r10 = eq_i32 r11 r10
|
||||
285 r10 = bit_and r9 r10
|
||||
286 r10 = bit_and r16 r10
|
||||
287 r4 = select r10 r0 r4
|
||||
288 r5 = select r10 r1 r5
|
||||
289 r6 = select r10 r2 r6
|
||||
290 r7 = select r10 r3 r7
|
||||
loop:
|
||||
291 store32 ptr1 r4
|
||||
292 store32 ptr2 r5
|
||||
|
||||
@ -9,13 +9,13 @@
|
||||
7 r7 = uniform32 ptr0 30
|
||||
8 r8 = uniform32 ptr0 40
|
||||
9 r2 = add_f32 r1 r2
|
||||
10 r3 = add_f32 r2 r3
|
||||
11 r4 = add_f32 r3 r4
|
||||
10 r2 = add_f32 r3 r2
|
||||
11 r2 = add_f32 r4 r2
|
||||
12 r5 = mul_f32 r5 r5
|
||||
13 r6 = mul_f32 r5 r6
|
||||
14 r7 = mul_f32 r6 r7
|
||||
13 r5 = mul_f32 r6 r5
|
||||
14 r5 = mul_f32 r7 r5
|
||||
loop:
|
||||
15 store32 ptr1 r0
|
||||
16 store32 ptr2 r4
|
||||
17 store32 ptr3 r7
|
||||
16 store32 ptr2 r2
|
||||
17 store32 ptr3 r5
|
||||
18 store32 ptr4 r8
|
||||
|
||||
Loading…
Reference in New Issue
Block a user