char-fork/llama.cpp
git clone https://git.t4t.associates/char-fork/llama.cpp
92e854ab8
char/aliceai-t5-moe
1#version 450 2 3#include "types.glsl" 4#include "generic_binary_head.glsl" 5 6layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in; 7 8void main() { 9 const uint col = gl_GlobalInvocationID.x; 10 11 if (col >= p.ne20) { 12 return; 13 } 14 15 for (uint row = gl_GlobalInvocationID.y; row < p.ne21; row += gl_WorkGroupSize.y * gl_NumWorkGroups.y) { 16 float sum = 0.0f; 17 for (uint i = 0; i < p.ne10; ++i) { 18 if (data_b[get_boffset() + i*p.nb10] == int(row)) { 19 sum += data_a[get_aoffset() + i*p.nb01 + col*p.nb00]; 20 } 21 } 22 23 data_d[get_doffset() + row*p.nb21 + col*p.nb20] = sum; 24 } 25}