char-fork/llama.cpp

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

Pascalvulkan: add cpy bf16 -> f32 pipelines (#22677)fcae601e4

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
1.7 KiB53 linesraw
1#version 450
2
3#include "types.glsl"
4#include "generic_unary_head.glsl"
5
6#extension GL_EXT_control_flow_attributes : require
7
8const uint num_threads = 128;
9
10layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
11
12void main() {
13    uint idx = get_idx();
14
15    // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
16    const uint num_iter = 4;
17
18    // fast path for when all four iterations are in-bounds
19    if (idx + (num_iter-1)*num_threads < p.ne) {
20        [[unroll]] for (uint i = 0; i < num_iter; ++i) {
21
22#if defined(DATA_A_BF16)
23            data_d[get_doffset() + idx] = D_TYPE(bf16_to_fp32(uint32_t(data_a[get_aoffset() + idx])));
24#elif defined(DATA_D_BF16)
25            float f = float(data_a[get_aoffset() + idx]);
26            data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f));
27#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
28            data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]);
29#else
30            data_d[get_doffset() + idx] = data_a[get_aoffset() + idx];
31#endif
32            idx += num_threads;
33        }
34    } else {
35        [[unroll]] for (uint i = 0; i < num_iter; ++i) {
36            if (idx >= p.ne) {
37                continue;
38            }
39
40#if defined(DATA_A_BF16)
41            data_d[get_doffset() + idx] = D_TYPE(bf16_to_fp32(uint32_t(data_a[get_aoffset() + idx])));
42#elif defined(DATA_D_BF16)
43            float f = float(data_a[get_aoffset() + idx]);
44            data_d[get_doffset() + idx] = D_TYPE(fp32_to_bf16(f));
45#elif !defined(OPTIMIZATION_ERROR_WORKAROUND)
46            data_d[get_doffset() + idx] = D_TYPE(data_a[get_aoffset() + idx]);
47#else
48            data_d[get_doffset() + idx] = data_a[get_aoffset() + idx];
49#endif
50            idx += num_threads;
51        }
52    }
53}