Method configure

src/gpu/cl/kernels/ClDirectConv3dKernel.cpp:119–248

}

void ClDirectConv3dKernel::configure(const CLCompileContext &compile_context,
                                     const ITensorInfo      *src0,
                                     const ITensorInfo      *src1,
                                     const ITensorInfo      *src2,
                                     ITensorInfo            *dst,
                                     const Conv3dInfo       &conv3d_info)
{
    ARM_COMPUTE_ERROR_ON_NULLPTR(src0, src1, dst);

    // Perform validation
    ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src0, src1, src2, dst, conv3d_info));

    // Create window and update padding
    const DataType data_type      = src0->data_type();
    const size_t   src_width      = src0->dimension(1);
    const size_t   src_height     = src0->dimension(2);
    const size_t   src_depth      = src0->dimension(3);
    const size_t   src_channels   = src0->dimension(0);
    const size_t   dst_width      = dst->dimension(1);
    const size_t   dst_height     = dst->dimension(2);
    const size_t   dst_depth      = dst->dimension(3);
    const size_t   dst_channels   = dst->dimension(0);
    const size_t   weights_width  = src1->dimension(2);
    const size_t   weights_height = src1->dimension(3);
    const size_t   weights_depth  = src1->dimension(4);
    const size_t   pad_left       = conv3d_info.padding.left;
    const size_t   pad_top        = conv3d_info.padding.top;
    const size_t   pad_front      = conv3d_info.padding.front;
    const size_t   conv_stride_x  = conv3d_info.stride.x();
    const size_t   conv_stride_y  = conv3d_info.stride.y();
    const size_t   conv_stride_z  = conv3d_info.stride.z();

    const size_t n0               = std::min(dst->dimension(0), static_cast<size_t>(4u));
    const size_t m0               = (dst->tensor_shape()[0] > 16) ? ((data_type == DataType::F32) ? 2U : 4U) : 1U;
    const size_t k0               = adjust_vec_size(8u, src0->dimension(0));
    const size_t partial_store_n0 = dst->dimension(0) % n0;

    CLBuildOptions build_options;
    build_options.add_option("-cl-fast-relaxed-math");
    build_options.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
    build_options.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(src_width));
    build_options.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(src_height));
    build_options.add_option("-DSRC_DEPTH=" + support::cpp11::to_string(src_depth));
    build_options.add_option("-DSRC_CHANNELS=" + support::cpp11::to_string(src_channels));
    build_options.add_option("-DDST_WIDTH=" + support::cpp11::to_string(dst_width));
    build_options.add_option("-DDST_HEIGHT=" + support::cpp11::to_string(dst_height));
    build_options.add_option("-DDST_DEPTH=" + support::cpp11::to_string(dst_depth));
    build_options.add_option("-DDST_CHANNELS=" + support::cpp11::to_string(dst_channels));
    build_options.add_option("-DWEI_WIDTH=" + support::cpp11::to_string(weights_width));
    build_options.add_option("-DWEI_HEIGHT=" + support::cpp11::to_string(weights_height));
    build_options.add_option("-DWEI_DEPTH=" + support::cpp11::to_string(weights_depth));
    build_options.add_option("-DSTRIDE_X=" + support::cpp11::to_string(conv_stride_x));
    build_options.add_option("-DSTRIDE_Y=" + support::cpp11::to_string(conv_stride_y));
    build_options.add_option("-DSTRIDE_Z=" + support::cpp11::to_string(conv_stride_z));
    build_options.add_option("-DPAD_LEFT=" + support::cpp11::to_string(pad_left));
    build_options.add_option("-DPAD_TOP=" + support::cpp11::to_string(pad_top));
    build_options.add_option("-DPAD_FRONT=" + support::cpp11::to_string(pad_front));
    build_options.add_option("-DN0=" + support::cpp11::to_string(n0));