From afa5d817b1d083837cd7ea30d32f845d82620c12 Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Thu, 30 Nov 2017 14:25:57 +0000 Subject: COMPMID-617 Add validation methods to Kernels - NEActivationLayer - NESoftmax - NEDirectConvolutionLayer - NENormalizationLayer - NEPoolingLayer Change-Id: Ib279f1c1b7f9247679b0d6593aed7393da8fe87b Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/111335 Tested-by: BSG Visual Compute Jenkins server to access repositories on http://mpd-gerrit.cambridge.arm.com Reviewed-by: Georgios Pinitas --- src/runtime/NEON/functions/NEActivationLayer.cpp | 5 +++++ .../NEON/functions/NEDirectConvolutionLayer.cpp | 24 ++++++++++++++++++++++ .../NEON/functions/NENormalizationLayer.cpp | 13 +++++++++++- src/runtime/NEON/functions/NEPoolingLayer.cpp | 5 +++++ src/runtime/NEON/functions/NESoftmaxLayer.cpp | 19 ++++++++++++++++- 5 files changed, 64 insertions(+), 2 deletions(-) (limited to 'src/runtime/NEON') diff --git a/src/runtime/NEON/functions/NEActivationLayer.cpp b/src/runtime/NEON/functions/NEActivationLayer.cpp index 57a1738f85..cdf1b54659 100644 --- a/src/runtime/NEON/functions/NEActivationLayer.cpp +++ b/src/runtime/NEON/functions/NEActivationLayer.cpp @@ -34,3 +34,8 @@ void NEActivationLayer::configure(ITensor *input, ITensor *output, ActivationLay k->configure(input, output, activation_info); _kernel = std::move(k); } + +Status NEActivationLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const ActivationLayerInfo &act_info) +{ + return NEActivationLayerKernel::validate(input, output, act_info); +} diff --git a/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp b/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp index 52a4cc158f..2eabe459a5 100644 --- a/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEDirectConvolutionLayer.cpp @@ -85,6 +85,30 @@ void NEDirectConvolutionLayer::configure(ITensor *input, const ITensor *weights, _input_border_handler.configure(input, _conv_kernel.border_size(), BorderMode::CONSTANT, PixelValue(static_cast(0.f))); } +Status NEDirectConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *bias, const ITensorInfo *output, const PadStrideInfo &conv_info) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, bias, output); + + DataType data_type = output->data_type(); + if(is_data_type_fixed_point(data_type)) + { + // Promote data type in case of fixed point + data_type = ((data_type == DataType::QS8) ? DataType::QS16 : DataType::QS32); + } + TensorInfo accumulator(output->clone()->set_is_resizable(true).reset_padding().set_data_type(data_type)); + + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, bias); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(bias->dimension(0) != weights->dimension(3), + "Biases size and number of input feature maps should match"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(bias->num_dimensions() > 1, + "Biases should be one dimensional"); + + ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerKernel::validate(input, weights, &accumulator, conv_info)); + ARM_COMPUTE_RETURN_ON_ERROR(NEDirectConvolutionLayerBiasAccumulateKernel::validate(&accumulator, bias, output)); + + return Status{}; +} + void NEDirectConvolutionLayer::run() { NEScheduler::get().schedule(&_input_border_handler, Window::DimZ); diff --git a/src/runtime/NEON/functions/NENormalizationLayer.cpp b/src/runtime/NEON/functions/NENormalizationLayer.cpp index da4314b5ed..af98ac1f17 100644 --- a/src/runtime/NEON/functions/NENormalizationLayer.cpp +++ b/src/runtime/NEON/functions/NENormalizationLayer.cpp @@ -39,7 +39,7 @@ NENormalizationLayer::NENormalizationLayer(std::shared_ptr memor void NENormalizationLayer::configure(const ITensor *input, ITensor *output, const NormalizationLayerInfo &norm_info) { - ARM_COMPUTE_ERROR_ON(input == nullptr); + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); TensorInfo tensor_info(input->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position()); _input_squared.allocator()->init(tensor_info); @@ -56,6 +56,17 @@ void NENormalizationLayer::configure(const ITensor *input, ITensor *output, cons _input_squared.allocator()->allocate(); } +Status NENormalizationLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const NormalizationLayerInfo &norm_info) +{ + // Perform validation step + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); + + ARM_COMPUTE_RETURN_ON_ERROR(NENormalizationLayerKernel::validate(input, input, output, norm_info)); + ARM_COMPUTE_RETURN_ON_ERROR(NEPixelWiseMultiplicationKernel::validate(input, input, output, 1.0f, ConvertPolicy::SATURATE, RoundingPolicy::TO_ZERO)); + + return Status{}; +} + void NENormalizationLayer::run() { _memory_group.acquire(); diff --git a/src/runtime/NEON/functions/NEPoolingLayer.cpp b/src/runtime/NEON/functions/NEPoolingLayer.cpp index f8a85b9897..530c7fca4a 100644 --- a/src/runtime/NEON/functions/NEPoolingLayer.cpp +++ b/src/runtime/NEON/functions/NEPoolingLayer.cpp @@ -48,6 +48,11 @@ void NEPoolingLayer::configure(ITensor *input, ITensor *output, const PoolingLay _border_handler.configure(input, _pooling_layer_kernel.border_size(), border_mode, PixelValue(static_cast(0.f))); } +Status NEPoolingLayer::validate(const ITensorInfo *input, const ITensorInfo *output, const PoolingLayerInfo &pool_info) +{ + return NEPoolingLayerKernel::validate(input, output, pool_info); +} + void NEPoolingLayer::run() { // Fill border diff --git a/src/runtime/NEON/functions/NESoftmaxLayer.cpp b/src/runtime/NEON/functions/NESoftmaxLayer.cpp index 84ecfdaf33..8e6773c5b1 100644 --- a/src/runtime/NEON/functions/NESoftmaxLayer.cpp +++ b/src/runtime/NEON/functions/NESoftmaxLayer.cpp @@ -38,7 +38,7 @@ NESoftmaxLayer::NESoftmaxLayer(std::shared_ptr memory_manager) void NESoftmaxLayer::configure(ITensor *input, ITensor *output, float beta) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QS16, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); // Create intermediate tensors shapes TensorInfo tensor_info_tmp(input->info()->tensor_shape(), input->info()->num_channels(), input->info()->data_type(), input->info()->fixed_point_position()); @@ -67,6 +67,23 @@ void NESoftmaxLayer::configure(ITensor *input, ITensor *output, float beta) _sum.allocator()->allocate(); } +Status NESoftmaxLayer::validate(const ITensorInfo *input, const ITensorInfo *output, float beta) +{ + // Perform validation step + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); + + TensorShape max_sum_shape = input->tensor_shape(); + max_sum_shape.set(0, 1); + + TensorInfo tensor_info_max_sum(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(max_sum_shape)); + + ARM_COMPUTE_RETURN_ON_ERROR(NELogits1DMaxKernel::validate(input, &tensor_info_max_sum)); + ARM_COMPUTE_RETURN_ON_ERROR(NELogits1DShiftExpSumKernel::validate(input, &tensor_info_max_sum, input, &tensor_info_max_sum, beta)); + ARM_COMPUTE_RETURN_ON_ERROR(NELogits1DNormKernel::validate(input, &tensor_info_max_sum, output)); + + return Status{}; +} + void NESoftmaxLayer::run() { _memory_group.acquire(); -- cgit v1.2.1