char-fork/llama.cpp
git clone https://git.t4t.associates/char-fork/llama.cpp
e29acf74f
char/aliceai-t5-moe
1#version 450 2 3#extension GL_EXT_control_flow_attributes : require 4 5#include "types.glsl" 6 7layout (push_constant) uniform parameter 8{ 9 uint ne0; 10 uint ne1; 11 uint s01; 12 uint s02; 13 uint s11; 14 uint s21; 15} p; 16 17#define BLOCK_SIZE 512 18 19layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in; 20 21layout (binding = 0) readonly buffer X {A_TYPE data_a[];}; 22layout (binding = 1) readonly buffer Y {B_TYPE data_b[];}; 23layout (binding = 2) readonly buffer Z {int32_t data_c[];}; 24layout (binding = 3) writeonly buffer D {D_TYPE data_d[];}; 25 26void main() { 27 const uint i1 = gl_WorkGroupID.x; 28 const uint i2 = gl_WorkGroupID.y; 29 30 const uint i11 = data_c[i1 + i2 * p.s21]; 31 32 const uint s1 = p.ne0; 33 const uint s2 = p.ne0 * p.ne1; 34 35 const uint d0 = i1 * s1 + i2 * s2; 36 const uint a0 = i1 * p.s01 + i2 * p.s02; 37 const uint b0 = i11 * p.s11; 38 39 for (uint i0 = gl_LocalInvocationID.x; i0 < p.ne0; i0 += BLOCK_SIZE) { 40 data_d[d0 + i0] = data_a[a0 + i0] + data_b[b0 + i0]; 41 } 42}