/**
 * Copyright (c) 2023 Nomic, Inc. All rights reserved.
 *
 * This software is licensed under the terms of the Software for Open Models License (SOM),
 * version 1.0, as detailed in the LICENSE_SOM.txt file. A copy of this license should accompany
 * this software. Except as expressly granted in the SOM license, all rights are reserved by Nomic, Inc.
 */

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

    for (int j = 0; j < pcs.ne00; j++) {
        out_[i*pcs.nb1 + j + pcs.outOff] = inA[r*pcs.nb01/2+j + pcs.inAOff];
    }
}