KBaba7/llama.cpp
0
1#version 4502 3#include "dequant_head.comp"4 5layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;6 7layout (binding = 0) readonly buffer A {block_q4_1 data_a[];};8layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};9 10void main() {11 const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;12 13 const uint tid = gl_LocalInvocationID.x % 64;14 const uint il = tid/32;15 const uint ir = tid%32;16 const uint ib = 32*i + ir;17 if (ib >= p.nel / 32) {18 return;19 }20 21 const uint b_idx = 1024*i + 32*ir + 8*il;22 23 const float d = float(data_a[ib].d);24 const float m = float(data_a[ib].m);25 26 const uint q_idx = 8*il;27 28 [[unroll]] for (uint l = 0; l < 8; ++l) {29 data_b[b_idx + l + 0] = D_TYPE(d * (data_a[ib].qs[q_idx + l] & 0xF) + m);30 data_b[b_idx + l + 16] = D_TYPE(d * (data_a[ib].qs[q_idx + l] >> 4) + m);31 }32}33 