diff options
author | Michele Di Giorgio <michele.digiorgio@arm.com> | 2021-01-22 09:47:04 +0000 |
---|---|---|
committer | Michele Di Giorgio <michele.digiorgio@arm.com> | 2021-06-18 10:33:48 +0000 |
commit | d02d5edfa15ba6c04a9986a8a362a945cb38ac31 (patch) | |
tree | ced4f49691d6c7038e347a8709b315bff59c64cf /src/core/utils/quantization/AsymmHelpers.cpp | |
parent | b014c27ba6db9840e4a72519760d51a87a2af7e7 (diff) | |
download | ComputeLibrary-d02d5edfa15ba6c04a9986a8a362a945cb38ac31.tar.gz |
Integrate improved CPU depthwise convolution kernels
* Replace assembly kernels for depthwise convolution with more optimized
ones.
* Add int8 assembly kernels.
* Fix implicit padding on optimized kernels
Resolves: COMPMID-3867, COMPMID-4361
Change-Id: I0b0867e05f61be4f368f62190d55e14d0ab3ebf2
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5622
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Diffstat (limited to 'src/core/utils/quantization/AsymmHelpers.cpp')
-rw-r--r-- | src/core/utils/quantization/AsymmHelpers.cpp | 7 |
1 files changed, 3 insertions, 4 deletions
diff --git a/src/core/utils/quantization/AsymmHelpers.cpp b/src/core/utils/quantization/AsymmHelpers.cpp index 49e39f663f..81e813caff 100644 --- a/src/core/utils/quantization/AsymmHelpers.cpp +++ b/src/core/utils/quantization/AsymmHelpers.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -185,16 +185,15 @@ std::pair<int, int> get_min_max_values_from_quantized_data_type(DataType data_ty void compute_quantized_multipliers_and_shifts(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, - unsigned int idx_ofms, int32_t *output_multipliers_ptr, int32_t *output_shifts_ptr) { - const unsigned int num_filters = is_data_type_quantized_per_channel(weights->data_type()) ? weights->dimension(idx_ofms) : 1; - const UniformQuantizationInfo iq_info = input->quantization_info().uniform(); const QuantizationInfo wq_info = weights->quantization_info(); const UniformQuantizationInfo oq_info = output->quantization_info().uniform(); + const unsigned int num_filters = wq_info.scale().size(); + for(unsigned int i = 0; i < num_filters; ++i) { int32_t output_multiplier = 0; |