char-fork/llama.cpp

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

Jeff Bolzvulkan: Programmatically add RoundingModeRTE to all shaders when the device supports it (#21572)1f30ac0ce

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
2.2 KiB65 linesraw
1#extension GL_EXT_shader_16bit_storage : require
2#extension GL_EXT_control_flow_attributes : require
3
4#include "utils.glsl"
5#if RMS_NORM_ROPE_FUSION
6#include "rope_params.glsl"
7#endif
8
9layout (push_constant) uniform parameter
10{
11    uint ne;
12    uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03;
13    uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
14    uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
15    uint misalign_offsets;
16    float param1; float param2; int param3;
17#if RMS_NORM_ROPE_FUSION
18    rope_params rope;
19#endif
20} p;
21
22#if !RMS_NORM_ROPE_FUSION
23layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
24#if defined(A_TYPE_PACKED16)
25layout (binding = 0) readonly buffer A_PACKED16 {A_TYPE_PACKED16 data_a_packed16[];};
26#endif
27#if defined(A_TYPE_PACKED32)
28layout (binding = 0) readonly buffer A_PACKED32 {A_TYPE_PACKED32 data_a_packed32[];};
29#endif
30
31layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
32layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
33#endif
34
35// true if src0/src1 are the same shape and the indices can be reused without additional modulus
36layout(constant_id = 0) const bool norepeat = false;
37
38uint get_idx() {
39    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
40}
41
42uint get_aoffset() { return p.misalign_offsets >> 16; }
43uint get_boffset() { return (p.misalign_offsets >> 8) & 0xFF; }
44uint get_doffset() { return p.misalign_offsets & 0xFF; }
45
46
47void get_indices(uint idx, out uint i00, out uint i01, out uint i02, out uint i03) {
48    get_indices(idx, i00, i01, i02, i03, p.ne00, p.ne01, p.ne02, p.ne03);
49}
50
51uint src0_idx(uint i00, uint i01, uint i02, uint i03) {
52    return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00;
53}
54
55uint src1_idx(uint i00, uint i01, uint i02, uint i03) {
56    if (norepeat) {
57        return i03*p.nb13 + i02*p.nb12 + i01*p.nb11 + i00*p.nb10;
58    } else {
59        return fastmod(i03, p.ne13)*p.nb13 + fastmod(i02, p.ne12)*p.nb12 + fastmod(i01, p.ne11)*p.nb11 + fastmod(i00, p.ne10)*p.nb10;
60    }
61}
62
63uint dst_idx(uint i00, uint i01, uint i02, uint i03) {
64    return i03*p.nb23 + i02*p.nb22 + i01*p.nb21 + i00*p.nb20;
65}