vkml 0.0.2

High-level Vulkan-based machine learning library
[vk::constant_id(0)]
const int WORKGROUP_SIZE_X = 1;
[vk::constant_id(1)]
const int WORKGROUP_SIZE_Y = 1;
[vk::constant_id(2)]
const int WORKGROUP_SIZE_Z = 1;

struct PushConstants
{
    uint n;
    uint c;
    uint m;
    uint in_len;
    uint out_len;
    uint kernel;
    uint stride;
    uint dilation;
    uint pad_begin;
    uint group;
    uint has_bias;
}

[[vk::push_constant]]
PushConstants pc;

[shader("compute")]
[numthreads(WORKGROUP_SIZE_X, WORKGROUP_SIZE_Y, WORKGROUP_SIZE_Z)]
void main<T : IArithmetic>(
    StructuredBuffer<T> src,
    StructuredBuffer<T> weights,
    RWStructuredBuffer<T> dst,
    StructuredBuffer<T> bias,
    uint3 threadId: SV_DispatchThreadID)
{
    uint ox = threadId.x;
    uint oc = threadId.y;
    uint batch = threadId.z;

    if (ox >= pc.out_len || oc >= pc.m || batch >= pc.n)
        return;

    T acc = T(0);

    uint m_per_group = pc.m / pc.group;
    uint c_per_group = pc.c / pc.group;
    uint group_id = oc / m_per_group;
    uint c_start = group_id * c_per_group;

    for (uint ic = c_start; ic < c_start + c_per_group; ++ic)
    {
        for (uint k = 0u; k < pc.kernel; ++k)
        {
            int in_x_i = int(ox) * int(pc.stride) - int(pc.pad_begin) + int(k) * int(pc.dilation);
            if (in_x_i < 0)
                continue;
            uint in_x = uint(in_x_i);
            if (in_x >= pc.in_len)
                continue;

            uint src_off = ((batch * pc.c + ic) * pc.in_len) + in_x;
            uint ic_in_group = ic - c_start;
            uint w_off = ((oc * c_per_group + ic_in_group) * pc.kernel) + k;

            acc = acc + src[src_off] * weights[w_off];
        }
    }

    if (pc.has_bias != 0u)
    {
        acc = acc + bias[oc];
    }

    uint dst_off = ((batch * pc.m + oc) * pc.out_len) + ox;
    dst[dst_off] = acc;
}