diff options
| author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2018-08-02 16:43:24 +0100 |
|---|---|---|
| committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:54:54 +0000 |
| commit | ebbb7f2ee00bdebad5da3629bbc78dc3a65fe0c5 (patch) | |
| tree | ece1bc256de86ea6fa25803e30976b1eed7b3ffd /src/runtime | |
| parent | 4cb39096ded770bfc4cf9712b85ed38e66c0e3f7 (diff) | |
| download | ComputeLibrary-ebbb7f2ee00bdebad5da3629bbc78dc3a65fe0c5.tar.gz | |
COMPMID-1188 - Fixed CLGEMMConvolutionLayer/NEGEMMConvolutionLayer for NHWC
We skipped im2col also without unit strides
Change-Id: I04c63a6dda8553b3890e832a56ff6854349c829a
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/142520
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/runtime')
| -rw-r--r-- | src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp | 4 | ||||
| -rw-r--r-- | src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp | 2 |
2 files changed, 3 insertions, 3 deletions
diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp index de628291eb..fde15613d6 100644 --- a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp @@ -181,7 +181,7 @@ void CLGEMMConvolutionLayer::configure(const ICLTensor *input, const ICLTensor * _original_weights = weights; _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type()); _data_layout = data_layout; - _skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1) && !_is_quantized; + _skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1) && !_is_quantized; _append_bias = (biases != nullptr) && (!_is_quantized); // Set the GPU target for im2col and col2im @@ -350,7 +350,7 @@ Status CLGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI const bool is_nhwc = data_layout == DataLayout::NHWC; const bool is_quantized = is_data_type_quantized_asymmetric(data_type); - const bool skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1) && !is_quantized; + const bool skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1) && !is_quantized; const bool append_bias = (biases != nullptr) && (!is_quantized); const unsigned bias_element = (append_bias && !skip_im2col) ? 1 : 0; diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp index 25e8d9e60b..aace261e32 100644 --- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp @@ -209,7 +209,7 @@ Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInf kernel_height = (are_weights_reshaped) ? weights_info.kernel_size().second : weights->dimension(idx_height); mat_weights_cols = weights->dimension(3); mat_weights_rows = weights->dimension(idx_width) * weights->dimension(idx_height) * weights->dimension(idx_channel) + ((append_bias && !skip_im2col) ? 1 : 0); - skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1); + skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1); std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(idx_width), input->dimension(idx_height), kernel_width, kernel_height, conv_info, dilation); |