#version 450

#extension GL_EXT_control_flow_attributes : enable
#extension GL_EXT_shader_16bit_storage : require

#define BLOCK_SIZE 32
#define FLOAT_TYPE float

layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;

layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
layout (binding = 2) writeonly buffer D {D_TYPE dst[];};

layout (push_constant) uniform parameter
{
    uint ncols_x;
    uint nrows_x;
    uint nchannels_x;
    uint nchannels_y;
    uint b_offset;
    uint d_offset;
} p;

shared FLOAT_TYPE tmp[BLOCK_SIZE];

void main() {
    const uint tid = gl_LocalInvocationID.x;
    const uint row_x = gl_GlobalInvocationID.y;
    const uint channel = gl_GlobalInvocationID.z;
    const uint channel_x = channel / (p.nchannels_y / p.nchannels_x);

    const uint nrows_y = p.ncols_x;
    const uint nrows_dst = p.nrows_x;
    const uint row_dst = row_x;

    tmp[tid] = FLOAT_TYPE(0.0f);

    for (uint col_x0 = 0; col_x0 < p.ncols_x; col_x0 += BLOCK_SIZE) {
        const uint col_x = col_x0 + tid;

        if (col_x >= p.ncols_x) {
            break;
        }

        // x is transposed and permuted
        const uint ix = row_x*p.nchannels_x*p.ncols_x + channel_x*p.ncols_x + col_x;
        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);

        const uint row_y = col_x;

        // y is not transposed but permuted
        const uint iy = channel*nrows_y + row_y;

        tmp[tid] += xi * FLOAT_TYPE(data_b[iy]);
    }

    // dst is not transposed and not permuted
    const uint idst = channel*nrows_dst + row_dst;

    // sum up partial sums and write back result
    barrier();
    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
        if (tid < s) {
            tmp[tid] += tmp[tid + s];
        }
        barrier();
    }

    if (tid == 0) {
        dst[idst] = tmp[0];
    }
}