diff options
| author | Giuseppe Rossini <giuseppe.rossini@arm.com> | 2019-11-06 14:57:49 +0000 |
|---|---|---|
| committer | Pablo Marquez <pablo.tello@arm.com> | 2019-12-04 12:45:12 +0000 |
| commit | f01201abec0a102f6e7a517971f83fef1eaffd50 (patch) | |
| tree | adf844c3c9c8e0e96af9c56de27a094fab515e35 /src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp | |
| parent | 6e1791b1bfabc81f08d3117939f6eb5264ed4edf (diff) | |
| download | ComputeLibrary-f01201abec0a102f6e7a517971f83fef1eaffd50.tar.gz | |
COMPMID-2305: NEDepthwiseConvolution 3x3: support for QUANT8_PER_CHANNEL_SYMM
Change-Id: I9a917cff6a089ce6ae16fb4e6066a4194e2e9487
Signed-off-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2241
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Marquez <pablo.tello@arm.com>
Diffstat (limited to 'src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp')
| -rw-r--r-- | src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp | 9 |
1 files changed, 5 insertions, 4 deletions
diff --git a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp index 6cf7b97e66..5e47dd56ae 100644 --- a/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEDepthwiseConvolutionLayer.cpp @@ -40,7 +40,10 @@ Status validate_arguments_optimized(const ITensorInfo *input, const ITensorInfo { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); + if(!is_data_type_quantized_per_channel(weights->data_type())) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); + } ARM_COMPUTE_RETURN_ERROR_ON(input->data_layout() == DataLayout::UNKNOWN); ARM_COMPUTE_RETURN_ERROR_ON(dilation.x() < 1 || dilation.y() < 1); const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH); @@ -55,7 +58,7 @@ Status validate_arguments_optimized(const ITensorInfo *input, const ITensorInfo ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(channel_idx)); } - const bool is_quantized = is_data_type_quantized_asymmetric(input->data_type()); + const bool is_quantized = (!is_data_type_quantized_per_channel(weights->data_type())) && is_data_type_quantized_asymmetric(input->data_type()); if(is_quantized) { @@ -67,7 +70,6 @@ Status validate_arguments_optimized(const ITensorInfo *input, const ITensorInfo ARM_COMPUTE_UNUSED(multiplier); ARM_COMPUTE_RETURN_ERROR_ON(multiplier > 1.0f); } - if(!NEDepthwiseConvolutionAssemblyDispatch::is_optimized_supported(input, weights, conv_info, depth_multiplier, dilation)) { TensorInfo accumulator = TensorInfo(output->clone()->set_is_resizable(true).reset_padding().set_data_type(DataType::S32)); @@ -88,7 +90,6 @@ Status validate_arguments_optimized(const ITensorInfo *input, const ITensorInfo { ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(output, nullptr, act_info)); } - return Status{}; } } // namespace |