diff options
author | Michalis Spyrou <michalis.spyrou@arm.com> | 2019-08-22 16:52:00 +0100 |
---|---|---|
committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2019-08-30 13:39:54 +0000 |
commit | 3f632f3f16e29ebeb7065b30008060fd4bfd09f1 (patch) | |
tree | ce15e27cf559bd1828b302e1257c187fece0b477 /src/runtime/CL | |
parent | 351bd137e48c5276963274ac741b172483e98d21 (diff) | |
download | ComputeLibrary-3f632f3f16e29ebeb7065b30008060fd4bfd09f1.tar.gz |
COMPMID-2418: CLDequantizationLayer support for QASYMM8_PER_CHANNEL
Add support for QASYMM8_PER_CHANNEL in CLDequantiazationLayer.
Added tests for NHWC and also updated NEON code to work with NHWC
data layout.
Cleaned up the reference implementation.
Change-Id: Ic1d51f16f7f625503fffdbbb66f6487aa588f08c
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1828
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Diffstat (limited to 'src/runtime/CL')
-rw-r--r-- | src/runtime/CL/CLTensorAllocator.cpp | 12 |
1 files changed, 10 insertions, 2 deletions
diff --git a/src/runtime/CL/CLTensorAllocator.cpp b/src/runtime/CL/CLTensorAllocator.cpp index 028a764fc2..51caf69297 100644 --- a/src/runtime/CL/CLTensorAllocator.cpp +++ b/src/runtime/CL/CLTensorAllocator.cpp @@ -79,8 +79,6 @@ void clear_quantization_arrays(CLFloatArray &scale, CLInt32Array &offset) * @param[in, out] offset Quantization offset array * @param[in] qinfo Quantization info * @param[in] pad_size Pad size to use in case array needs to be padded for computation purposes - * - * @return A pair (scale, offset) containing the respective allocated and filled arrays */ void populate_quantization_info(CLFloatArray &scale, CLInt32Array &offset, const QuantizationInfo &qinfo, size_t pad_size) { @@ -93,6 +91,16 @@ void populate_quantization_info(CLFloatArray &scale, CLInt32Array &offset, const scale = CLFloatArray(num_elements + pad_size); scale.resize(num_elements); CLScheduler::get().queue().enqueueWriteBuffer(scale.cl_buffer(), CL_TRUE, 0, num_elements * element_size, qinfo.scale().data()); + + if(!qinfo.offset().empty()) + { + // Create offset array + const std::vector<int32_t> &qoffset = qinfo.offset(); + const size_t offset_element_size = sizeof(std::remove_reference<decltype(qoffset)>::type::value_type); + offset = CLInt32Array(num_elements + pad_size); + offset.resize(num_elements); + CLScheduler::get().queue().enqueueWriteBuffer(offset.cl_buffer(), CL_TRUE, 0, num_elements * offset_element_size, qinfo.offset().data()); + } } } // namespace |