diff options
Diffstat (limited to 'src/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.cpp')
-rw-r--r-- | src/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.cpp | 30 |
1 files changed, 17 insertions, 13 deletions
diff --git a/src/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.cpp b/src/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.cpp index d5a52845a..da57aa447 100644 --- a/src/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.cpp +++ b/src/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2020 ARM Limited. + * Copyright (c) 2019-2020 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -69,6 +69,7 @@ Status validate_arguments(const ITensorInfo *input0, const ITensorInfo *input1, && (!gemm_info.broadcast_bias), "Bias addition only supported with broadcast mode in case the input or output has to be reinterpreted as 3D"); ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.fp_mixed_precision, "Mixed precision not supported"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(rhs_info.export_to_cl_image, "Export to CLImage not supported for GEMM native"); const unsigned int m = gemm_info.m; const unsigned int n = gemm_info.n; @@ -154,33 +155,26 @@ std::pair<Status, Window> 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 unsigned int m = reinterpret_output_as_3d ? gemm_info.m : output->dimension(1); - const unsigned 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); + input0->dimension(1)); 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); + output->dimension(0), + output->dimension(1)); if(input2 != nullptr) { const int bias_processed_per_iteration_x = num_elems_processed_per_iteration_x; - const int bias_processed_per_iteration_y = gemm_info.broadcast_bias ? 1 : num_elems_processed_per_iteration_y; - AccessWindowStatic input2_access(input2, 0, 0, ceil_to_multiple(input2->dimension(0), bias_processed_per_iteration_x), - ceil_to_multiple(input2->dimension(1), bias_processed_per_iteration_y)); + input2->dimension(1)); window_changed = update_window_and_padding(win, input0_access, input1_access, input2_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 @@ -263,6 +257,14 @@ void CLGEMMMatrixMultiplyNativeKernel::configure(const CLCompileContext &compile const unsigned int h_gemm_3d = _reinterpret_output_as_3d ? output->info()->dimension(1) : input0->info()->dimension(1); const unsigned int d_gemm_3d = _reinterpret_output_as_3d ? output->info()->dimension(2) : input0->info()->dimension(2); + // 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; + + // Shrink M0 to be always <= M (internal_m) to prevent out-of-bounds reads. + // NOTE: This might have implications on heuristics and performance + const unsigned int internal_m0 = std::min(internal_m, lhs_info.m0); + // Create build options CLBuildOptions build_opts; build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input0->info()->data_type())); @@ -279,9 +281,11 @@ void CLGEMMMatrixMultiplyNativeKernel::configure(const CLCompileContext &compile build_opts.add_option("-DM=" + support::cpp11::to_string(internal_m)); build_opts.add_option("-DN=" + support::cpp11::to_string(gemm_info.n)); build_opts.add_option("-DK=" + support::cpp11::to_string(gemm_info.k)); - build_opts.add_option("-DM0=" + support::cpp11::to_string(lhs_info.m0)); + build_opts.add_option("-DM0=" + support::cpp11::to_string(internal_m0)); build_opts.add_option("-DN0=" + support::cpp11::to_string(rhs_info.n0)); build_opts.add_option("-DK0=" + support::cpp11::to_string(rhs_info.k0)); + 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)); build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(gemm_info.activation_info.activation()))); build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.a())); build_opts.add_option_if(gemm_info.activation_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(gemm_info.activation_info.b())); |