yum-mirror/glslang

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

Jeff BolzImplement GL_EXT_bfloat16ba164044

main
branchesmainall refs →
3.5 KiB146 linesraw
1#version 450 core
2
3#extension GL_EXT_bfloat16 : require
4#extension GL_KHR_cooperative_matrix : enable
5#extension GL_KHR_memory_scope_semantics : enable
6#extension GL_EXT_shader_explicit_arithmetic_types : enable
7#extension GL_EXT_scalar_block_layout : enable
8
9layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
10
11const bfloat16_t bc10 = bfloat16_t(10);
12const bf16vec2 b2c20 = bf16vec2(20);
13layout(constant_id = 1) const bfloat16_t bsc10 = bfloat16_t(10);
14
15struct S { bfloat16_t b; };
16const S sc = S(bc10);
17
18const bfloat16_t bca[2] = {bc10, bc10};
19
20shared bfloat16_t bfs[10];
21
22layout(scalar) buffer B {
23    bf16vec3 b3;
24    bf16vec2 b2;
25    bfloat16_t b1;
26} buf;
27
28bfloat16_t funcbf16(bfloat16_t x)
29{
30    return x;
31}
32
33bfloat16_t funcf32(float32_t x)
34{
35    return bfloat16_t(x);
36}
37
38bfloat16_t funcf64(float64_t x)
39{
40    return bfloat16_t(x);
41}
42
43void main()
44{
45    float f = 2.0;
46    bfloat16_t b = bfloat16_t(1.0);
47    bfloat16_t(f);
48    bf16vec2 b2 = bf16vec2(f);
49
50    uint8_t u8 = uint8_t(5);
51    uint16_t u16 = uint16_t(5);
52    uint32_t u32 = 5;
53    uint64_t u64 = 5;
54    int8_t i8 = int8_t(6);
55    int16_t i16 = int16_t(6);
56    int32_t i32 = 6;
57    int64_t i64 = 6;
58    bfloat16_t bf16 = bfloat16_t(7);
59    float16_t f16 = float16_t(7);
60    float32_t f32 = 7;
61    float64_t f64 = 7;
62    b = bfloat16_t(u8);
63    b = bfloat16_t(u16);
64    b = bfloat16_t(u32);
65    b = bfloat16_t(u64);
66    b = bfloat16_t(i8);
67    b = bfloat16_t(i16);
68    b = bfloat16_t(i32);
69    b = bfloat16_t(i64);
70    b = bfloat16_t(bf16);
71    b = bfloat16_t(f16);
72    b = bfloat16_t(f32);
73    b = bfloat16_t(f64);
74    u8 = uint8_t(b);
75    u16 = uint16_t(b);
76    u32 = uint32_t(b);
77    u64 = uint64_t(b);
78    i8 = int8_t(b);
79    i16 = int16_t(b);
80    i32 = int32_t(b);
81    i64 = int64_t(b);
82    bf16 = bfloat16_t(b);
83    f16 = float16_t(b);
84    f32 = float32_t(b);
85    f64 = float64_t(b);
86
87    f32 = b;
88    f64 = b;
89
90    b = buf.b1;
91    b2 = buf.b2;
92    bf16vec3 b3 = buf.b3;
93
94    dot(b2, b2);
95
96    coopmat<bfloat16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseA> cmA = coopmat<bfloat16_t, gl_ScopeSubgroup, 16, 16, gl_MatrixUseA>(3.0);
97
98    coopmat<float, gl_ScopeSubgroup, 16, 16, gl_MatrixUseA> cmAf = coopmat<float, gl_ScopeSubgroup, 16, 16, gl_MatrixUseA>(cmA);
99
100    funcbf16(b);
101    funcf32(b);
102    funcf64(b);
103
104    int16_t i16_1 = bfloat16BitsToIntEXT(b);
105    i16vec2 i16_2 = bfloat16BitsToIntEXT(b2);
106    i16vec3 i16_3 = bfloat16BitsToIntEXT(bf16vec3(b2, b));
107    i16vec4 i16_4 = bfloat16BitsToIntEXT(bf16vec4(b2, b2));
108
109    uint16_t u16_1 = bfloat16BitsToUintEXT(b);
110    u16vec2  u16_2 = bfloat16BitsToUintEXT(b2);
111    u16vec3  u16_3 = bfloat16BitsToUintEXT(bf16vec3(b2, b));
112    u16vec4  u16_4 = bfloat16BitsToUintEXT(bf16vec4(b2, b2));
113
114    bfloat16_t b16_1 = intBitsToBFloat16EXT(i16_1);
115    bf16vec2   b16_2 = intBitsToBFloat16EXT(i16_2);
116    bf16vec3   b16_3 = intBitsToBFloat16EXT(i16_3);
117    bf16vec4   b16_4 = intBitsToBFloat16EXT(i16_4);
118
119    b16_1 = uintBitsToBFloat16EXT(u16_1);
120    b16_2 = uintBitsToBFloat16EXT(u16_2);
121    b16_3 = uintBitsToBFloat16EXT(u16_3);
122    b16_4 = uintBitsToBFloat16EXT(u16_4);
123
124}
125
126bfloat16_t func2(bf16vec4 v[2], int i)
127{
128    bfloat16_t b;
129    b = (bf16vec2(v[0]))[i];
130    b = (bf16vec2(v[1])).y;
131
132    v[i].z = v[i][i];
133
134    S s;
135    s.b = b;
136
137    b = (i != 0) ? b : bfloat16_t(2);
138
139    if (i == 2) {
140        b = v[0].x;
141    } else {
142        b = v[1].y;
143    }
144
145    return b;
146}