From 3f8aac4474b245b20c07b3a5384577a83f4950a7 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Mon, 24 Dec 2018 13:09:02 +0000 Subject: COMPMID-1860: Invalid arguments in CLDepthwiseConvolution3x3 for NHWC -Alters the kernel/function selection process to use validate for selection. -Fixes border kernel input in case of permutation. Change-Id: Ia61df3a0ed661349114dc125f33ad53ee40d9c76 Reviewed-on: https://review.mlplatform.org/443 Reviewed-by: Anthony Barbier Tested-by: Arm Jenkins --- .../CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp | 8 +++--- .../CL/functions/CLDepthwiseConvolutionLayer.cpp | 30 +++++++--------------- 2 files changed, 12 insertions(+), 26 deletions(-) diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp index 1fce14f215..4f3636b081 100644 --- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp +++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp @@ -55,6 +55,8 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier > 1); // COMPMID-1071 Add depth multiplier support for NHWC ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(1) != 3 || weights->dimension(2) != 3); + ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1 || conv_info.stride().first > 2); + ARM_COMPUTE_RETURN_ERROR_ON(std::max(conv_info.pad_top(), conv_info.pad_bottom()) > 1); const bool is_qasymm = is_data_type_quantized_asymmetric(input->data_type()); @@ -155,10 +157,6 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input, ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, act_info)); - const unsigned int conv_stride_x = conv_info.stride().first; - ARM_COMPUTE_ERROR_ON(conv_stride_x < 1 || conv_stride_x > 2); - ARM_COMPUTE_ERROR_ON(std::max(conv_info.pad_top(), conv_info.pad_bottom()) > 1); - const bool is_qasymm = is_data_type_quantized_asymmetric(input->info()->data_type()); const bool is_stride_1 = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1)); const bool is_dot8_supported = dot8_supported(CLKernelLibrary::get().get_device()); @@ -243,7 +241,7 @@ void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input, } else { - build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(conv_stride_x)); + build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(conv_info.stride().first)); build_opts.add_option("-DCONV_STRIDE_Y=" + support::cpp11::to_string(_conv_stride_y)); } build_opts.add_option_if(_input->info()->tensor_shape().total_size_upper(3) > 1, diff --git a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp index c2782aaa89..be13f500ea 100644 --- a/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp +++ b/src/runtime/CL/functions/CLDepthwiseConvolutionLayer.cpp @@ -95,7 +95,7 @@ void CLDepthwiseConvolutionLayer3x3::configure(ICLTensor *input, const ICLTensor if(_needs_permute) { // Configure the function to transform the convoluted output to ACL's native ordering format NCHW - _permuted_output.info()->set_data_layout(DataLayout::NHWC); + _permuted_output.info()->set_data_layout(DataLayout::NCHW); _permute_output_to_nhwc.configure(&_permuted_output, output, PermutationVector(2U, 0U, 1U)); // Allocate tensors @@ -109,7 +109,7 @@ void CLDepthwiseConvolutionLayer3x3::configure(ICLTensor *input, const ICLTensor { zero_value = PixelValue(static_cast(input->info()->quantization_info().offset)); } - _border_handler.configure(input, _kernel->border_size(), BorderMode::CONSTANT, zero_value); + _border_handler.configure(input_to_use, _kernel->border_size(), BorderMode::CONSTANT, zero_value); } Status CLDepthwiseConvolutionLayer3x3::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, @@ -187,19 +187,6 @@ void CLDepthwiseConvolutionLayer3x3::prepare() } } -namespace -{ -inline bool can_run_optimised_3x3_kernel(const ITensorInfo *weights, unsigned int depth_multiplier) -{ - const DataLayout data_layout = weights->data_layout(); - const size_t idx_w = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - const size_t idx_h = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); - const Size2D weights_size(weights->dimension(idx_w), weights->dimension(idx_h)); - return weights_size == Size2D(3, 3) && (data_layout == DataLayout::NHWC && depth_multiplier <= 1); -} - -} // namespace - CLDepthwiseConvolutionLayer::CLDepthwiseConvolutionLayer() : _im2col_kernel(), _weights_reshape_kernel(), _v2mm_kernel(), _vector_to_tensor_kernel(), _output_stage_kernel(), _activationlayer_function(), _v2mm_input_fill_border(), _v2mm_weights_fill_border(), _input_reshaped(), _weights_reshaped(), _v2mm_output(), _output_reshaped(), _is_prepared(false), _is_quantized(false), _is_activationlayer_enabled(false), _original_weights(nullptr), @@ -214,7 +201,12 @@ void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *w ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output); - if(can_run_optimised_3x3_kernel(weights->info(), depth_multiplier)) + const Status can_run_optimised_3x3_kernel = CLDepthwiseConvolutionLayer3x3::validate(input->info(), + weights->info(), + biases != nullptr ? biases->info() : nullptr, + output->info(), + conv_info, depth_multiplier, act_info); + if(bool(can_run_optimised_3x3_kernel)) { auto f = arm_compute::support::cpp14::make_unique(); f->configure(input, weights, biases, output, conv_info, depth_multiplier, act_info); @@ -323,11 +315,7 @@ void CLDepthwiseConvolutionLayer::configure(ICLTensor *input, const ICLTensor *w Status CLDepthwiseConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier, const ActivationLayerInfo &act_info) { - if(can_run_optimised_3x3_kernel(weights, depth_multiplier)) - { - ARM_COMPUTE_RETURN_ON_ERROR(CLDepthwiseConvolutionLayer3x3::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info)); - } - else + if(!bool(CLDepthwiseConvolutionLayer3x3::validate(input, weights, biases, output, conv_info, depth_multiplier, act_info))) { const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH); const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT); -- cgit v1.2.1