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/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.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/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp')
| -rw-r--r-- | src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp | 25 |
1 files changed, 21 insertions, 4 deletions
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp index 61c8d90f78..d1f0338739 100644 --- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp +++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayerNativeKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -211,8 +211,25 @@ void CLDepthwiseConvolutionLayerNativeKernel::configure(const CLCompileContext & arm_compute::opencl::kernels::gemm::update_padding_for_cl_image(weights->info()); } - build_opts.add_option("-cl-fast-relaxed-math"); - build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(conv_info.act_info.activation()))); + // Conditions of -cl-fast-relaxed-math causing accuracy issues can be traced from COMPMID-5324 + const GPUTarget gpu_target = get_target(); + const auto act_function = conv_info.act_info.activation(); + const auto dst_data_type = _output->info()->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_opts.add_option("-cl-unsafe-math-optimizations"); + } + else + { + build_opts.add_option("-cl-fast-relaxed-math"); + } + + build_opts.add_option("-DACTIVATION_TYPE=" + lower_string(string_from_activation_func(act_function))); build_opts.add_option("-DDEPTH_MULTIPLIER=" + support::cpp11::to_string(conv_info.depth_multiplier)); build_opts.add_option("-DSRC_TENSOR_TYPE=BUFFER"); // Note: SRC_DATA_TYPE must have the same data type of WEI_DATA_TYPE. In quantized, we could @@ -220,7 +237,7 @@ void CLDepthwiseConvolutionLayerNativeKernel::configure(const CLCompileContext & // only works when both have same data type, we have to change the offset to take into account this aspect build_opts.add_option("-DSRC_DATA_TYPE=" + get_cl_type_from_data_type(_input->info()->data_type())); build_opts.add_option("-DDST_TENSOR_TYPE=BUFFER"); - build_opts.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(_output->info()->data_type())); + build_opts.add_option("-DDST_DATA_TYPE=" + get_cl_type_from_data_type(dst_data_type)); build_opts.add_option_if_else(_export_to_cl_image, "-DWEI_TENSOR_TYPE=IMAGE", "-DWEI_TENSOR_TYPE=BUFFER"); build_opts.add_option("-DWEI_WIDTH=" + support::cpp11::to_string(_weights->info()->dimension(1))); build_opts.add_option("-DWEI_HEIGHT=" + support::cpp11::to_string(_weights->info()->dimension(2))); |