char-fork/llama.cpp

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

Jeff Bolzvulkan: support non-contig unary/glu ops (#24215)1a7718b4c

char/aliceai-t5-moe
921 B28 linesraw
1#version 450
2
3#include "types.glsl"
4#include "generic_unary_head.glsl"
5
6layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
7
8void main() {
9    const uint idx = get_idx();
10
11    if (idx >= p.ne) {
12        return;
13    }
14
15    const uint i13 = fastdiv(idx, p.ne1_012mp, fastdiv_L(p.ne1_Ls, 0));
16    const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10;
17    const uint i12 = fastdiv(idx - i13_offset, p.ne1_01mp, fastdiv_L(p.ne1_Ls, 1));
18    const uint i12_offset = i12*p.ne11*p.ne10;
19    const uint i11 = fastdiv(idx - i13_offset - i12_offset, p.ne1_0mp, fastdiv_L(p.ne1_Ls, 2));
20    const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10;
21
22    if (i10 == i11) {
23        const float val = float(data_a[get_aoffset() + i13*p.nb03 + i12*p.nb02 + 0*p.nb01 + i10*p.nb00]);
24        data_d[get_doffset() + dst_idx(idx)] = D_TYPE(val);
25    } else {
26        data_d[get_doffset() + dst_idx(idx)] = D_TYPE(0);
27    }
28}