KBaba7/llama.cpp

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_iq4_nl 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    init_iq_shmem(gl_WorkGroupSize);14 15    const uint tid = gl_LocalInvocationID.x % 64;16    const uint il  = tid/32;17    const uint ir  = tid%32;18    const uint ib = 32*i + ir;19    if (ib >= p.nel / 32) {20        return;21    }22 23    const uint q_idx = 8*il;24    const uint b_idx = 1024*i + 32*ir + q_idx;25 26    const float d = float(data_a[ib].d);27 28    [[unroll]] for (uint l = 0; l < 8; ++l) {29        data_b[b_idx + l +  0] = D_TYPE(d * kvalues_iq4nl[data_a[ib].qs[q_idx + l] & 0xF]);30        data_b[b_idx + l + 16] = D_TYPE(d * kvalues_iq4nl[data_a[ib].qs[q_idx + l] >>  4]);31    }32}33