yum-mirror/glslang

git clone https://git.yummers.dev/yum-mirror/glslang

alelenvFix bug in handling slice of array being passed cooperative matrix load/store functions and strengthen error checks. (#4007)8e12361e

main
branchesmainall refs →
4.9 KiB125 linesraw
1#version 450 core
2#extension GL_KHR_memory_scope_semantics : enable
3#extension GL_KHR_cooperative_matrix : enable
4#extension GL_EXT_shader_explicit_arithmetic_types : enable
5#extension GL_EXT_buffer_reference : enable
6
7layout (local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
8
9const int X = 8;
10layout(constant_id = 0) const int Y = 2;
11const int Z = X*Y;
12
13coopmat<float16_t, gl_ScopeSubgroup, Z, 8, gl_MatrixUseAccumulator> mC;
14coopmat<float16_t, gl_ScopeSubgroup, Z, 8, gl_MatrixUseAccumulator> mC2[3];
15
16layout(constant_id = 1) const float F = 3.0;
17
18const coopmat<float, gl_ScopeSubgroup, Z, 8, gl_MatrixUseAccumulator> mD = coopmat<float, gl_ScopeSubgroup, Z, 8, gl_MatrixUseAccumulator>(0.0);
19const coopmat<float16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseAccumulator> mD2 = coopmat<float16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseAccumulator>(1);
20
21struct S { int a; int b; int c; };
22
23const S s = S(12, 23, 34);
24
25layout(set = 0, binding = 0, buffer_reference) coherent buffer Block {
26    float y[1024*1024];
27    float x[];
28} block;
29
30layout(set = 0, binding = 0) coherent buffer Block16 {
31    float16_t arr[256][256];
32    float16_t y[1024*1024];
33    float16_t x[];
34
35    Block b;
36} block16;
37
38coopmat<float16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseAccumulator> f16(coopmat<float16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseAccumulator> m) { return -m; }
39coopmat<float, gl_ScopeSubgroup, 8, 8, gl_MatrixUseAccumulator> f32(coopmat<float, gl_ScopeSubgroup, 8, 8, gl_MatrixUseAccumulator> m) { return -m; }
40
41layout(constant_id = 2) const int SC = 1;
42coopmat<float16_t, gl_ScopeSubgroup, SC, SC, gl_MatrixUseAccumulator> scm[SC][SC];
43
44// sized for coopmat<float16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseAccumulator>
45shared uvec4 shmatrix[16*16*2/16];
46
47void main()
48{
49    coopmat<float, gl_ScopeSubgroup, 16, (2>1?8:4), gl_MatrixUseAccumulator> m = coopmat<float, gl_ScopeSubgroup, 16, (2>1?8:4), gl_MatrixUseAccumulator>(0.0);
50
51    m = m + m;
52    m = m - m;
53    m = -m;
54    m = 2.0*m;
55    m = m*2.0;
56
57    coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> m2 = coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator>(m);
58
59    float x = m[1];
60    m[0] = x;
61
62    coopMatLoad(m, block.x, 16, 128, gl_CooperativeMatrixLayoutRowMajor);
63    coopMatStore(m, block.x, 16, 128, gl_CooperativeMatrixLayoutRowMajor);
64    coopMatLoad(m2, block16.x, 16, 128, gl_CooperativeMatrixLayoutRowMajor);
65    coopMatStore(m2, block16.x, 16, 128, gl_CooperativeMatrixLayoutRowMajor);
66    coopMatLoad(m, block16.b.x, 16, 128, gl_CooperativeMatrixLayoutRowMajor);
67    coopMatStore(m, block16.b.x, 16, 128, gl_CooperativeMatrixLayoutRowMajor);
68
69    coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseA> A;
70    coopmat<float16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseB> B;
71    coopmat<float, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> C;
72    coopmat<float, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> D;
73    D = coopMatMulAdd(A, B, C);
74
75    int l = D.length();
76
77    coopmat<float16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseAccumulator> E;
78
79    coopmat<float16_t, gl_ScopeSubgroup, Z, Z, gl_MatrixUseAccumulator> F = coopmat<float16_t, gl_ScopeSubgroup, Z, Z, gl_MatrixUseAccumulator>(0.0);
80
81    coopmat<float, gl_ScopeSubgroup, 16, (2>1?8:4), gl_MatrixUseAccumulator> a[5];
82    a[3][0] = 1.0;
83
84    float md1 = mD[1];
85
86    md1 += (m += m)[1234];
87
88    mC2[1] = mC2[2];
89
90    coopMatLoad(m, block.y, 16, 128, gl_CooperativeMatrixLayoutRowMajor);
91    coopMatStore(m, block.y, 16, 128, gl_CooperativeMatrixLayoutRowMajor);
92    coopMatLoad(m2, block16.y, 16, 128, gl_CooperativeMatrixLayoutRowMajor);
93    coopMatStore(m2, block16.y, 16, 128, gl_CooperativeMatrixLayoutRowMajor);
94
95    coopmat<float16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseAccumulator> p1;
96    coopmat<float, gl_ScopeSubgroup, 8, 8, gl_MatrixUseAccumulator> p2;
97
98    p1 = f16(p1);
99    p2 = f32(p2);
100
101    p1 = coopmat<float16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseAccumulator>(0.0);
102    p2 = coopmat<float, gl_ScopeSubgroup, 8, 8, gl_MatrixUseAccumulator>(0.0);
103
104    p1 /= p1;
105
106    p1 *= float16_t(2.0);
107    p2 *= 4.0;
108
109    coopmat<float16_t, gl_ScopeSubgroup, 16, 8, gl_MatrixUseAccumulator> ms;
110    coopMatLoad(ms, shmatrix, 1, 2, gl_CooperativeMatrixLayoutRowMajor);
111    coopMatStore(ms, shmatrix, 1, 2, gl_CooperativeMatrixLayoutRowMajor);
112
113    coopMatLoad(ms, block16.arr[8], 1, 2, gl_CooperativeMatrixLayoutRowMajor);
114    coopMatStore(ms, block16.arr[16], 1, 2, gl_CooperativeMatrixLayoutRowMajor);
115
116    coopmat<int8_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseA> ms8A;
117    coopmat<int8_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseB> ms8B;
118    coopmat<int8_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseAccumulator> ms8C;
119    coopMatMulAdd(ms8A, ms8B, ms8C);
120    coopMatMulAdd(ms8A, ms8B, ms8C, 0);
121    coopMatMulAdd(ms8A, ms8B, ms8C, gl_MatrixOperandsSaturatingAccumulation);
122
123    coopmat<int16_t, gl_ScopeSubgroup, 8, 8, gl_MatrixUseA> m16;
124    coopMatStore(m16, shmatrix, 1, 2, gl_CooperativeMatrixLayoutRowMajor);
125}