#version 450

#include "common.comp"

#extension GL_KHR_shader_subgroup_arithmetic : require

layout(local_size_x_id = 0) 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;
    uint ne02;
    uint ne12;
    int ne0;
    int ne1;
} pcs;

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

    uint bc_ab = pcs.ne12 > pcs.ne02 ? im / (pcs.ne12 / pcs.ne02) : im;
    uint bc_ba = pcs.ne02 > pcs.ne12 ? im / (pcs.ne02 / pcs.ne12) : im;

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

    float sumf = 0.0f;
    for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
        sumf += float(inA[x+i]) * float(inB[y+i]);
    }

    const float all_sum = subgroupAdd(sumf);
    if (subgroupElect()) {
        out_[im*pcs.ne1*pcs.ne0 + r1*pcs.ne0 + r0 + pcs.outOff] = all_sum;
    }
}