char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Aclyvulkan : incremental shader builds (#16341)e29acf74f

char/aliceai-t5-moe
1.0 KiB42 linesraw
1#version 450
2
3#extension GL_EXT_shader_16bit_storage : require
4
5layout (push_constant) uniform parameter
6{
7    uint nb1;
8    uint dim;
9    uint max_period;
10} p;
11
12#include "types.glsl"
13
14#extension GL_EXT_control_flow_attributes : enable
15#define BLOCK_SIZE 256
16
17layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
18
19layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
20layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
21
22void main() {
23    const uint i = gl_WorkGroupID.y;
24    const uint j = gl_GlobalInvocationID.x;
25    const uint d_offset = i * p.nb1;
26
27    const uint half_dim = p.dim / 2;
28
29    if (p.dim % 2 != 0 && j == half_dim) {
30        data_d[d_offset + 2 * half_dim] = 0.f;
31    }
32
33    if (j >= half_dim) {
34        return;
35    }
36
37    const float timestep = float(data_a[i]);
38    const float freq = float(exp(-log(p.max_period) * j / half_dim));
39    const float arg = timestep * freq;
40    data_d[d_offset + j] = D_TYPE(cos(arg));
41    data_d[d_offset + j + half_dim] = D_TYPE(sin(arg));
42}