/**
 * 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 = 64) in;

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

layout (push_constant) uniform parameter {
    uint inAOff;
    uint inBOff;
    uint outOff;
    int ne00;
    uint nb01;
    uint nb02;
    uint nb11;
    uint nb12;
    int ne0;
    int ne1;
} pcs;

shared float sum[gl_WorkGroupSize.x];

void main() {
    const uint r0 = gl_WorkGroupID.x;
    const uint r1 = gl_WorkGroupID.y;
    const uint im = gl_WorkGroupID.z;

    const uint x = (r0*pcs.nb01 + im*pcs.nb02) / 2 + pcs.inAOff; // Based from inA
    const uint y = (r1*pcs.nb11 + im*pcs.nb12) / 4 + pcs.inBOff; // based from inB

    sum[gl_LocalInvocationID.x] = 0.0;

    for (uint i = gl_LocalInvocationID.x; i < pcs.ne00; i += gl_WorkGroupSize.x) {
        sum[gl_LocalInvocationID.x] += float(inA[x+i]) * float(inB[y+i]);
    }

    // accumulate the sum from all threads in the threadgroup
    barrier();
    memoryBarrierShared();
    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
        if (gl_LocalInvocationID.x < i) {
            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
        }
        barrier();
        memoryBarrierShared();
    }

    if (gl_LocalInvocationID.x == 0) {
        out_[im*pcs.ne1*pcs.ne0 + r1*pcs.ne0 + r0 + pcs.outOff] = sum[0];
    }
}