llama.cpp/kompute-shaders/op_cpy_f32_f32.comp

#version 450

#include "common.comp"

#define IN_TYPE float
#define IN_TYPE_SIZE 4
#define OUT_TYPE float
#define OUT_TYPE_SIZE 4

layout(local_size_x = 1024) in;

layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };

layout (push_constant) uniform parameter {
    uint inOff;
    uint outOff;
    int ne00;
    int ne01;
    int ne02;
    uint nb00;
    uint nb01;
    uint nb02;
    uint nb03;
    int ne0;
    int ne1;
    int ne2;
    uint nb0;
    uint nb1;
    uint nb2;
    uint nb3;
} pcs;

void main() {
    const uint i03 = gl_WorkGroupID.z;
    const uint i02 = gl_WorkGroupID.y;
    const uint i01 = gl_WorkGroupID.x;

    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;

    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);

    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_

    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
        out_[dst_data+i00] = OUT_TYPE(in_[src]);
    }
}