llama.cpp/kompute-shaders/op_getrows_f16.comp

#version 450

#include "common.comp"

layout(local_size_x = 1) in;

layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
layout (binding = 1) readonly buffer tensorInB { int inB[]; };
layout (binding = 2) writeonly buffer tensorOut { float out_[]; };

layout (push_constant) uniform parameter {
    uint inAOff;
    uint inBOff;
    uint outOff;
    int ne00;
    int nb01;
    int nb1;
} pcs;

void dequantize_row_f16(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
    for (int j = 0; j < k; j++) {
        out_[y + j] = inA[x + j];
    }
}

void main() {
    const uint i = gl_WorkGroupID.x;
    const int r = inB[i + pcs.inBOff];

    dequantize_row_f16(r*pcs.nb01/2/*bytes for float16*/ + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
}