Method exec_internal

dnn/src/cuda/conv_bias/matmul.cpp:138–172

template <typename T>
void ConvBiasForwardImpl::AlgoMatmul::exec_internal(
        const ExecArgs& args, const WorkspaceBundle& bundle) {
    auto&& fm = args.filter_meta;
    size_t N = args.src_layout->shape[0], IC = fm.icpg, IH = args.src_layout->shape[2],
           IW = args.src_layout->shape[3], OC = fm.ocpg,
           OH = args.dst_tensor->layout.shape[2], OW = args.dst_tensor->layout.shape[3],
           FH = fm.spatial[0], FW = fm.spatial[1], PH = fm.padding[0],
           PW = fm.padding[1], SH = fm.stride[0], SW = fm.stride[1],
           DH = fm.dilation[0], DW = fm.dilation[1];
    auto stream = cuda_stream(args.handle);
    T* dst_t = static_cast<T*>(bundle.get(0));
    T* col = static_cast<T*>(bundle.get(1));
    conv_bias::im2col<T>(
            args.src_tensor->ptr<T>(), col, N, args.src_layout->stride[0], IC, IH, IW,
            FH, FW, OH, OW, PH, PW, SH, SW, DH, DW, stream);

    auto config = prepare_sub_opr(args);

    TensorND A(args.filter_tensor->ptr<T>(), config.first[0]), B(col, config.first[1]),
            C(dst_t, config.first[2]);
    size_t matmul_ws_idx = 2;
    if (fm.should_flip) {
        conv_bias::flip_filter(args, bundle.get_workspace(2), A.get_ref_ptr());
        matmul_ws_idx = 3;
    }

    config.second->exec(A, B, C, bundle.get_workspace(matmul_ws_idx));

    TensorLayout C2l({OC * OH * OW, N}, typename DTypeTrait<T>::dtype()), C3l = C2l;
    C3l.stride[0] = 1;
    C3l.stride[1] = args.dst_tensor->layout.stride[0];
    TensorND C2(dst_t, C2l);
    TensorND C3(args.dst_tensor->ptr<T>(), C3l);
    args.handle->relayout_opr()->exec(C2, C3);
}

// vim: syntax=cpp.doxygen