1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
|
#version 450 core
#extension GL_KHR_memory_scope_semantics : enable
#extension GL_KHR_cooperative_matrix : enable
#extension GL_EXT_shader_explicit_arithmetic_types : enable
#extension GL_NV_cooperative_matrix2 : enable
#extension GL_EXT_buffer_reference : enable
layout (local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
buffer BufType {
float16_t x[];
} Buf;
layout(buffer_reference, std430, buffer_reference_align = 2) buffer fp16Buf {
float16_t f;
};
float16_t decode0(const in fp16Buf b, const in uint32_t blockCoords[2], const in uint32_t coordInBlock[2]) { return b.f; }
float16_t decode1(const fp16Buf b, const in uint32_t blockCoords[2], const in uint32_t coordInBlock[2]) { return b.f; }
float16_t decode2(in fp16Buf b, const in uint32_t blockCoords[2], const in uint32_t coordInBlock[2]) { return b.f; }
float16_t decode3(fp16Buf b, const in uint32_t blockCoords[2], const in uint32_t coordInBlock[2]) { return b.f; }
float16_t decode4(const in fp16Buf b, const uint32_t blockCoords[2], const in uint32_t coordInBlock[2]) { return b.f; }
float16_t decode5(const in fp16Buf b, in uint32_t blockCoords[2], const in uint32_t coordInBlock[2]) { return b.f; }
float16_t decode6(const in fp16Buf b, uint32_t blockCoords[2], const in uint32_t coordInBlock[2]) { return b.f; }
float16_t decode7(const in fp16Buf b, const in uint32_t blockCoords[2], const uint32_t coordInBlock[2]) { return b.f; }
float16_t decode8(const in fp16Buf b, const in uint32_t blockCoords[2], in uint32_t coordInBlock[2]) { return b.f; }
float16_t decode9(const in fp16Buf b, const in uint32_t blockCoords[2], uint32_t coordInBlock[2]) { return b.f; }
float16_t decode10(const in uint32_t b, const in uint16_t blockCoords[2], const in uint16_t coordInBlock[2]) { return float16_t(0); }
float16_t decode11(const in fp16Buf b, const in uint32_t blockCoords, const in uint32_t coordInBlock) { return float16_t(0); }
struct S {
f16vec2 x;
};
float16_t combineSum(const in float16_t a, const in float16_t b) { return a + b; }
float16_t combineSum2(float16_t a, float16_t b) { return a + b; }
layout(constant_id = 0) const uint32_t Clamp = gl_CooperativeMatrixClampModeConstantNV;
float16_t relu(const in uint32_t row, const in uint32_t col, const in float16_t x) { return max(x, float16_t(0)); }
float16_t add(const in uint32_t row, const in uint32_t col, const in float16_t x, const in float16_t y) { return x+y; }
float32_t perelemf32(const in uint32_t row, const in uint32_t col, const in float16_t x) { return float32_t(x); }
void main()
{
coopmat<float16_t, gl_ScopeWorkgroup, 64, 32, gl_MatrixUseAccumulator> A;
tensorLayoutNV<2> t = createTensorLayoutNV(2);
coopMatLoadTensorNV(A, Buf.x, 0, t, decode0);
coopMatLoadTensorNV(A, Buf.x, 0, t, decode1);
coopMatLoadTensorNV(A, Buf.x, 0, t, decode2);
coopMatLoadTensorNV(A, Buf.x, 0, t, decode3);
coopMatLoadTensorNV(A, Buf.x, 0, t, decode4);
coopMatLoadTensorNV(A, Buf.x, 0, t, decode5);
coopMatLoadTensorNV(A, Buf.x, 0, t, decode6);
coopMatLoadTensorNV(A, Buf.x, 0, t, decode7);
coopMatLoadTensorNV(A, Buf.x, 0, t, decode8);
coopMatLoadTensorNV(A, Buf.x, 0, t, decode9);
coopMatLoadTensorNV(A, Buf.x, 0, t, decode10);
coopMatLoadTensorNV(A, Buf.x, 0, t, decode11);
coopmat<float32_t, gl_ScopeWorkgroup, 64, 32, gl_MatrixUseAccumulator> Af32;
coopMatReduceNV(A, A, gl_CooperativeMatrixReduceRowNV, combineSum);
coopMatReduceNV(A, A, gl_CooperativeMatrixReduceRowNV, combineSum2);
coopMatReduceNV(Af32, Af32, gl_CooperativeMatrixReduceRowNV, combineSum);
coopMatPerElementNV(A, A, relu);
coopMatPerElementNV(A, A, add, float16_t(1.0));
coopMatPerElementNV(A, A, add, coopmat<float16_t, gl_ScopeWorkgroup, 64, 32, gl_MatrixUseAccumulator>(1.0));
coopMatPerElementNV(A, A, add, float32_t(1.0));
coopMatPerElementNV(A, A, add, coopmat<float32_t, gl_ScopeWorkgroup, 64, 32, gl_MatrixUseAccumulator>(1.0));
coopMatPerElementNV(A, A, add);
coopMatPerElementNV(A, A, add, float16_t(1.0), float16_t(1.0));
coopMatPerElementNV(Af32, A, perelemf32);
coopMatPerElementNV(Af32, A, relu);
float16_t larr[16];
coopMatLoadTensorNV(A, larr, 0, t);
coopMatStoreTensorNV(A, larr, 0, t);
}
|
