From cf343e3798d2a8c2ad2fcac488e4b78e2b5c968d Mon Sep 17 00:00:00 2001 From: morgolock Date: Mon, 12 Oct 2020 14:00:43 +0100 Subject: COMPMID-3719: Remove OpenCL padding: CLGEMMLowpMatrixMultiplyNativeKernel Change-Id: Iee28abcbba1e7b9e2f3aaa55685936dce815d5a3 Signed-off-by: morgolock Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4141 Reviewed-by: Gian Marco Iodice Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins --- .../CLGEMMLowpMatrixMultiplyNativeKernel.cpp | 38 +++++++--------------- 1 file changed, 12 insertions(+), 26 deletions(-) (limited to 'src/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.cpp') diff --git a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.cpp b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.cpp index 9a2918d12f..d30a9e5d18 100644 --- a/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.cpp +++ b/src/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.cpp @@ -111,8 +111,6 @@ std::pair validate_and_configure_window(ITensorInfo *input0, ITe bool reinterpret_output_as_3d = (gemm_info.depth_output_gemm3d() != 0); Window win{}; - Window win_out{}; - bool window_changed = false; // In case both input and output have to be reinterpreted as 3D tensors, // force reinterpret_output_as_3d to be false. @@ -139,28 +137,8 @@ std::pair validate_and_configure_window(ITensorInfo *input0, ITe num_elems_processed_per_iteration_x = rhs_info.n0; num_elems_processed_per_iteration_y = lhs_info.m0; - // Note: bottom paddings are calculated manually as the output can be reinterpreted as 3D tensor - // The only way to set properly the paddings, it is to set those explicitly through the AccessWindowStatic - const int m = reinterpret_output_as_3d ? gemm_info.m() : input0->dimension(1); - const int bottom_pad = (num_elems_processed_per_iteration_y - (m % num_elems_processed_per_iteration_y)) % num_elems_processed_per_iteration_y; - win = calculate_max_window(tmp_info, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); - win_out = calculate_max_window(*output, Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); - - AccessWindowStatic input0_access(input0, 0, 0, - input0->dimension(0), - input0->dimension(1) + bottom_pad); - AccessWindowStatic input1_access(input1, 0, 0, - ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration_x), - input1->dimension(1)); - AccessWindowStatic output_access(output, 0, 0, - ceil_to_multiple(output->dimension(0), num_elems_processed_per_iteration_x), - output->dimension(1) + bottom_pad); - - window_changed = update_window_and_padding(win, input0_access, input1_access) || // window used by the execute_window_loop - update_window_and_padding(win_out, output_access); // window used to update the padding requirements of output tensor - - output_access.set_valid_region(win_out, ValidRegion(Coordinates(), output->tensor_shape())); + output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape())); // Collapse along the Z direction // This collapse needs to be here in order to tune the Z dimension of LWS @@ -168,8 +146,7 @@ std::pair validate_and_configure_window(ITensorInfo *input0, ITe const unsigned int dimension_to_collapse = std::min(static_cast(output->num_dimensions()), 2u); collapsed = win.collapse(win, dimension_to_collapse); - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - return std::make_pair(err, collapsed); + return std::make_pair(Status(), collapsed); } } // namespace @@ -218,6 +195,14 @@ void CLGEMMLowpMatrixMultiplyNativeKernel::configure(const CLCompileContext &com ARM_COMPUTE_ERROR_THROW_ON(win_config.first); ICLKernel::configure_internal(win_config.second); + // If _reinterpret_input_as_3d = _reinterpret_output_as_3d = true, + // we will dispatch a batched-GEMM to reduce the complexity of the address calculation within the OpenCL kernel. + // This means that the actual m used by the kernel is given by output->info()->dimension(1) and not by gemm_info.m + const unsigned int internal_m = input0->info()->dimension(1); + // Calculate partial (store instead of load) M0 and partial N0 for the partial blocks at the end of a row/column if any. This is to avoid padding. + const unsigned int partial_store_m0 = internal_m % lhs_info.m0; + const unsigned int partial_store_n0 = gemm_info.n() % rhs_info.n0; + // Create build options CLBuildOptions build_opts; build_opts.add_option_if(_reinterpret_input_as_3d, "-DREINTERPRET_INPUT_AS_3D"); @@ -234,7 +219,8 @@ void CLGEMMLowpMatrixMultiplyNativeKernel::configure(const CLCompileContext &com build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0)); build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type())); build_opts.add_option("-DACC_DATA_TYPE=" + get_cl_dot8_acc_type_from_data_type(input0->info()->data_type())); - + build_opts.add_option("-DPARTIAL_STORE_M0=" + support::cpp11::to_string(partial_store_m0)); + build_opts.add_option("-DPARTIAL_STORE_N0=" + support::cpp11::to_string(partial_store_n0)); std::string kernel_name("gemmlowp_mm_native"); // Create kernel -- cgit v1.2.1