diff options
| author | Gunes Bayir <gunes.bayir@arm.com> | 2022-05-25 10:09:39 +0100 |
|---|---|---|
| committer | Gunes Bayir <gunes.bayir@arm.com> | 2022-05-26 08:31:57 +0000 |
| commit | fdb534293b84182a89a78b5086cf05c07ef970e8 (patch) | |
| tree | d27f9d0e50cec63ad3b39c4f31cff4c6c9077ea4 /src/gpu/cl/kernels/ClDirectConv2dKernel.cpp | |
| parent | 5fcf22dadf092efd7aafb359f9229aa270eb1129 (diff) | |
| download | ComputeLibrary-fdb534293b84182a89a78b5086cf05c07ef970e8.tar.gz | |
Disable unsafe FP optimizations causing accuracy issues
Resolves: COMPMID-5324
Change-Id: I289b1bb42296c562cb90b918c20def8c6c1825d2
Signed-off-by: Gunes Bayir <gunes.bayir@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7639
Reviewed-by: SiCong Li <sicong.li@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/gpu/cl/kernels/ClDirectConv2dKernel.cpp')
| -rw-r--r-- | src/gpu/cl/kernels/ClDirectConv2dKernel.cpp | 24 |
1 files changed, 20 insertions, 4 deletions
diff --git a/src/gpu/cl/kernels/ClDirectConv2dKernel.cpp b/src/gpu/cl/kernels/ClDirectConv2dKernel.cpp index ff8c2c32a0..be4c8ef5b7 100644 --- a/src/gpu/cl/kernels/ClDirectConv2dKernel.cpp +++ b/src/gpu/cl/kernels/ClDirectConv2dKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -253,11 +253,27 @@ void ClDirectConv2dKernel::configure(const CLCompileContext &compile_context, IT build_options.add_option(std::string("-DBIA_DATA_TYPE=" + get_cl_type_from_data_type(biases->data_type()))); } - build_options.add_option("-cl-fast-relaxed-math"); + // Conditions of -cl-fast-relaxed-math causing accuracy issues can be traced from COMPMID-5324 + const auto act_function = act_info.activation(); + const auto dst_data_type = dst->data_type(); + + if((gpu_target != GPUTarget::G71 && (gpu_target & GPUTarget::GPU_ARCH_MASK) == GPUTarget::BIFROST) + && (act_function == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU || act_function == ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) + && (dst_data_type == DataType::F32 || dst_data_type == DataType::F16)) + { + // -cl-fast-relaxed-math also sets -cl-finite-math-only and -cl-unsafe-math-optimizations + // to disable -cl-finite-math-only, we only include -cl-unsafe-math-optimizations + build_options.add_option("-cl-unsafe-math-optimizations"); + } + else + { + build_options.add_option("-cl-fast-relaxed-math"); + } + build_options.add_option("-DSRC_TENSOR_TYPE=BUFFER"); build_options.add_option("-DSRC_DATA_TYPE=" + get_cl_type_from_data_type(src->data_type())); build_options.add_option("-DDST_TENSOR_TYPE=BUFFER"); - build_options.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(dst->data_type())); + build_options.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(dst_data_type)); build_options.add_option_if_else(export_to_cl_image, "-DWEI_TENSOR_TYPE=IMAGE", "-DWEI_TENSOR_TYPE=BUFFER"); build_options.add_option("-DWEI_WIDTH=" + support::cpp11::to_string(weights->dimension(width_idx))); build_options.add_option("-DWEI_HEIGHT=" + support::cpp11::to_string(weights->dimension(height_idx))); @@ -271,7 +287,7 @@ void ClDirectConv2dKernel::configure(const CLCompileContext &compile_context, IT build_options.add_option("-DK0=" + support::cpp11::to_string(k0)); build_options.add_option("-DPARTIAL_N0=" + support::cpp11::to_string(partial_store_n0)); build_options.add_option_if((src->dimension(channel_idx) % k0) != 0, "-DLEFTOVER_LOOP"); - build_options.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(act_info.activation()))); + build_options.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(act_function))); if(is_data_type_quantized(data_type)) { |