yum-mirror/glslang

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

Jeff BolzImplement GL_NV_cooperative_matrix_decode_vector (#4275)3ceed991

main
branchesmainall refs →
1.8 KiB56 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_NV_cooperative_matrix2 : enable
6#extension GL_NV_cooperative_matrix_decode_vector : enable
7#extension GL_EXT_buffer_reference : enable
8
9layout (local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
10
11buffer BufType {
12   float16_t x[];
13} Buf;
14
15layout(buffer_reference, std430, buffer_reference_align = 4) buffer dwordBuf {
16    uint32_t d;
17};
18
19float16_t decodeF16Scalar(const in dwordBuf b, const in uint32_t blockCoords[2], const in uint32_t coordInBlock[2])
20{
21    return float16_t(b.d);
22}
23
24f16vec2 decodeF16x2(const in dwordBuf b, const in uint32_t blockCoords[2], const in uint32_t coordInBlock[2])
25{
26    return unpackFloat2x16(b.d);
27}
28
29// Vector return with V == 3, not in {2, 4, 8}.
30f16vec3 decodeF16x3_badLength(const in dwordBuf b, const in uint32_t blockCoords[2], const in uint32_t coordInBlock[2])
31{
32    return f16vec3(0);
33}
34
35// Vector return whose component type doesn't match the matrix type.
36vec2 decodeF32x2_badType(const in dwordBuf b, const in uint32_t blockCoords[2], const in uint32_t coordInBlock[2])
37{
38    return vec2(0);
39}
40
41void main()
42{
43    coopmat<float16_t, gl_ScopeWorkgroup, 64, 32, gl_MatrixUseA> A;
44
45    tensorLayoutNV<2> t  = createTensorLayoutNV(2);
46
47    // Single-arg form must be scalar; a vector return is only legal in
48    // the second function position now.
49    coopMatLoadTensorNV(A, Buf.x, 0, t, decodeF16x2);
50
51    // V == 3, not in {2, 4, 8}.
52    coopMatLoadTensorNV(A, Buf.x, 0, t, decodeF16Scalar, decodeF16x3_badLength);
53
54    // Vector component type doesn't match the matrix component type.
55    coopMatLoadTensorNV(A, Buf.x, 0, t, decodeF16Scalar, decodeF32x2_badType);
56}