char-fork/llama.cpp

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

Jeff Bolzvulkan: support GGML_OP_SET (#19584)53aef25a8

char/aliceai-t5-moe
1.1 KiB37 linesraw
1#version 450
2
3#include "types.glsl"
4#include "generic_binary_head.glsl"
5
6// false for SET, true for ACC
7layout(constant_id = 1) const bool ACC = true;
8
9layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
10
11void main() {
12    const uint idx = gl_GlobalInvocationID.x;
13    if (idx >= p.ne) {
14        return;
15    }
16
17    const uint offset = p.param3;
18    const uint src1_i = idx - offset;
19    const uint i3 = src1_i / p.nb03;
20    const uint rem2 = src1_i - i3 * p.nb03;
21    const uint i2 = rem2 / p.nb02;
22    const uint rem1 = rem2 - i2 * p.nb02;
23    const uint i1 = rem1 / p.nb01;
24    const uint i0 = rem1 % p.nb01;
25
26    uint i00, i01, i02, i03;
27
28    if (i0 < p.ne10 && i1 < p.ne11 && i2 < p.ne12 && i3 < p.ne13) {
29        if (ACC) {
30            data_d[get_doffset() + idx] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + idx]) + FLOAT_TYPE(data_b[get_boffset() + src1_idx(i0, i1, i2, i3)]));
31        } else {
32            data_d[get_doffset() + idx] = D_TYPE(FLOAT_TYPE(data_b[get_boffset() + src1_idx(i0, i1, i2, i3)]));
33        }
34    } else {
35        data_d[get_doffset() + idx] = D_TYPE(FLOAT_TYPE(data_a[get_aoffset() + idx]));
36    }
37}