llama.cpp/kompute-shaders/op_diagmask.comp

#version 450

#include "common.comp"

layout(local_size_x = 1) in;

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

layout(push_constant) uniform PushConstants {
    uint inOff;
    uint outOff;
    uint n_past;
    int ne00;
    int ne01;
} pcs;

void main() {
    const uint i02 = gl_WorkGroupID.z;
    const uint i01 = gl_WorkGroupID.y;
    const uint i00 = gl_WorkGroupID.x;

    const uint index = i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00 + i00;

    if (i00 > pcs.n_past + i01) {
        out_[index + pcs.outOff] = uintBitsToFloat(0xFF800000);
    } else {
        out_[index + pcs.outOff] = in_[index + pcs.inOff];
    }
}