Method exec

dnn/src/naive/relayout_format/opr_impl.cpp:486–773

}

void RelayoutFormatImpl::exec(
        _megdnn_tensor_in src, _megdnn_tensor_out dst, _megdnn_workspace workspace) {
#if !MGE_BUILD_WITHOUT_NAIVE_EXEC
    megdnn_assert(
            src.layout.dtype.category() == DTypeCategory::FLOAT ||
            src.layout.dtype.enumv() == DTypeEnum::Int32 ||
            (src.layout.dtype.enumv() == DTypeEnum::Uint8 &&
             dst.layout.dtype.enumv() == DTypeEnum::QuantizedS8) ||
            (src.layout.dtype.enumv() == DTypeEnum::Uint8 &&
             dst.layout.dtype.enumv() == DTypeEnum::Uint8) ||
            src.layout.dtype.category() == DTypeCategory::QUANTIZED);
    check_exec(src.layout, dst.layout, workspace.size);
    HandleImpl* m_handle = static_cast<HandleImpl*>(handle());
    TensorLayout exec_src_layout, exec_dst_layout, exec_workspace_layout;
    deduce_exec_layout(
            src.layout, dst.layout, exec_workspace_layout, exec_src_layout,
            exec_dst_layout);

    // clean dst
    MEGDNN_DISPATCH_CPU_KERN(
            m_handle, memset(dst.raw_ptr(), 0, dst.layout.span().dist_byte()));

    //! construct exec Tensor
    TensorND exec_src_nd{exec_src_layout, src.get_ref_ptr()};
    TensorND exec_dst_nd{exec_dst_layout, dst.get_ref_ptr()};

    // pre
    if (param().mode == Param::Mode::NCHW_NHWCD4I) {
        size_t N = src.layout[0];
        size_t IC = src.layout[1];
        size_t IH = src.layout[2];
        size_t IW = src.layout[3];
        //! ic % 4 != 0
        if ((IC & 0x3)) {
            switch (src.layout.dtype.enumv()) {
#define cb(name, ctype)                                                              \
    case (DTypeEnum::name): {                                                        \
        MIDOUT_BEGIN(                                                                \
                megdnn_naive_relayout_format, ctype,                                 \
                midout_iv(Param::Mode::NCHW_NHWCD4I)) {                              \
            MEGDNN_DISPATCH_CPU_KERN(                                                \
                    m_handle, padding_src_to_workspace<ctype>(                       \
                                      workspace.ptr<ctype>(),                        \
                                      src.compatible_ptr<ctype>(), N, IC, IH, IW);); \
        }                                                                            \
        MIDOUT_END();                                                                \
        break;                                                                       \
    }
                cb(Float32, dt_float32);
                DNN_INC_FLOAT16(cb(Float16, dt_float16));
                cb(Quantized8Asymm, dt_uint8);
                cb(QuantizedS8, dt_int8);
                cb(Uint8, dt_uint8);
#undef cb
                default:
                    megdnn_assert(
                            0, "NCHW_NHWCD4I not support dtype %s",
                            src.layout.dtype.name());