char-fork/llama.cpp

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

Aclyvulkan : incremental shader builds (#16341)e29acf74f

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
796 B31 linesraw
1#version 450
2
3#extension GL_EXT_control_flow_attributes : enable
4
5#include "types.glsl"
6#include "generic_head.glsl"
7
8layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
9
10layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
11layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
12layout (binding = 2) buffer D {D_TYPE data_d[];};
13
14const uint CHUNK_SIZE = 512;
15
16void main() {
17    const uint base = gl_WorkGroupID.x * CHUNK_SIZE;
18    const uint col = gl_LocalInvocationID.x;
19
20    uint count = 0;
21    [[unroll]]
22    for (uint i = 0; i < CHUNK_SIZE; i += gl_WorkGroupSize.x) {
23        const uint idx = base + i + col;
24        if (idx >= p.KX) {
25            break;
26        }
27        count += uint(data_a[idx] == data_b[idx]);
28    }
29
30    atomicAdd(data_d[0], D_TYPE(count));
31}