diff options
| -rw-r--r-- | arm_compute/runtime/NEON/AssemblyHelper.h | 9 | ||||
| -rw-r--r-- | arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h | 16 | ||||
| -rw-r--r-- | src/core/NEON/kernels/NEWeightsReshapeKernel.cpp | 27 | ||||
| -rw-r--r-- | src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp | 242 | ||||
| -rw-r--r-- | tests/validation/CL/ConvolutionLayer.cpp | 12 | ||||
| -rw-r--r-- | tests/validation/CL/DilatedConvolutionLayer.cpp | 12 | ||||
| -rw-r--r-- | tests/validation/GLES_COMPUTE/ConvolutionLayer.cpp | 16 | ||||
| -rw-r--r-- | tests/validation/NEON/ConvolutionLayer.cpp | 28 | ||||
| -rw-r--r-- | tests/validation/NEON/DilatedConvolutionLayer.cpp | 28 | ||||
| -rw-r--r-- | tests/validation/fixtures/ConvolutionLayerFixture.h | 158 | ||||
| -rw-r--r-- | tests/validation/reference/ConvolutionLayer.cpp | 28 | ||||
| -rw-r--r-- | tests/validation/reference/Permute.cpp | 4 |
12 files changed, 299 insertions, 281 deletions
diff --git a/arm_compute/runtime/NEON/AssemblyHelper.h b/arm_compute/runtime/NEON/AssemblyHelper.h index 3db419e148..ecaf35ac3e 100644 --- a/arm_compute/runtime/NEON/AssemblyHelper.h +++ b/arm_compute/runtime/NEON/AssemblyHelper.h @@ -84,7 +84,12 @@ public: const int ldb = _b->info()->strides_in_bytes().y() / sizeof(TypeInput); const int ldd = _d->info()->strides_in_bytes().y() / sizeof(TypeOutput); - const int batch_stride_a = _a->info()->strides_in_bytes().z() / sizeof(TypeInput); + // In the case of NHWC we want to interpret the output shape as 3D. Thus, the batch stride for A is + // the relevant multiple of the row stride. + const bool is_nhwc = _a->info()->data_layout() == DataLayout::NHWC; + const int stride_in_bytes_a = is_nhwc ? _a->info()->strides_in_bytes().y() * _d->info()->dimension(1) : _a->info()->strides_in_bytes().z(); + + const int batch_stride_a = stride_in_bytes_a / sizeof(TypeInput); const int batch_stride_d = _d->info()->strides_in_bytes().z() / sizeof(TypeOutput); const int multi_stride_a = _a->info()->strides_in_bytes()[3] / sizeof(TypeInput); @@ -158,7 +163,7 @@ inline bool setup_assembly_kernel(const ITensor *a, const ITensor *b, ITensor *d const int M = d->info()->tensor_shape().y(); const int N = d->info()->tensor_shape().x(); const int K = a->info()->tensor_shape().x(); - const int batches = a->info()->tensor_shape().total_size_upper(2); + const int batches = d->info()->tensor_shape().total_size_upper(2); const int multis = b->info()->tensor_shape().z(); unsigned int num_threads = NEScheduler::get().num_threads(); diff --git a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h index 752693188c..d64fd9e771 100644 --- a/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h +++ b/arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h @@ -26,6 +26,7 @@ #include "arm_compute/runtime/IFunction.h" +#include "arm_compute/core/NEON/kernels/NEArithmeticAdditionKernel.h" #include "arm_compute/core/NEON/kernels/NECol2ImKernel.h" #include "arm_compute/core/NEON/kernels/NEFillBorderKernel.h" #include "arm_compute/core/NEON/kernels/NEGEMMAssemblyBaseKernel.h" @@ -176,6 +177,7 @@ private: NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint _gemmlowp_output_stage; NECol2ImKernel _output_col2im_kernel; NEActivationLayer _activationlayer_function; + NEArithmeticAdditionKernel _add_bias_kernel; const ITensor *_original_weights; @@ -187,12 +189,14 @@ private: Tensor _workspace; Tensor _B_pretransposed; - bool _append_bias; - bool _is_fully_connected_convolution; - bool _are_weights_reshaped; - bool _is_quantized; - bool _is_interleaved; - bool _is_activationlayer_enabled; + DataLayout _data_layout; + bool _append_bias; + bool _is_fully_connected_convolution; + bool _are_weights_reshaped; + bool _is_quantized; + bool _is_interleaved; + bool _is_activationlayer_enabled; + bool _skip_im2col; }; } #endif /* __ARM_COMPUTE_NECONVOLUTIONGEMMLAYER_H__ */ diff --git a/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp b/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp index 150140271d..3031a87637 100644 --- a/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp +++ b/src/core/NEON/kernels/NEWeightsReshapeKernel.cpp @@ -34,12 +34,16 @@ using namespace arm_compute; namespace { -template <typename T> +template <typename T, bool is_nhwc> void weights_reshape(const ITensor *input, const ITensor *bias, ITensor *output, const Window &window) { - const unsigned int kernel_size_x = input->info()->dimension(0); - const unsigned int kernel_size_y = input->info()->dimension(1); - const unsigned int kernel_depth = input->info()->dimension(2); + DataLayout data_layout = input->info()->data_layout(); + const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + const int idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); + const unsigned int kernel_size_x = input->info()->dimension(idx_width); + const unsigned int kernel_size_y = input->info()->dimension(idx_height); + const unsigned int kernel_depth = input->info()->dimension(idx_channel); const unsigned int input_stride_x = input->info()->strides_in_bytes().x(); const unsigned int input_stride_y = input->info()->strides_in_bytes().y(); const unsigned int input_stride_z = input->info()->strides_in_bytes().z(); @@ -67,13 +71,13 @@ void weights_reshape(const ITensor *input, const ITensor *bias, ITensor *output, for(unsigned int i = 0; i < kernel_size_x; ++i) { *(reinterpret_cast<T *>(tmp_output_ptr)) = *(reinterpret_cast<const T *>(tmp_input_ptr)); - tmp_input_ptr += input_stride_x; + tmp_input_ptr += is_nhwc ? input_stride_y : input_stride_x; tmp_output_ptr += output_stride_y; } - curr_input_row_ptr += input_stride_y; + curr_input_row_ptr += is_nhwc ? input_stride_z : input_stride_y; tmp_input_ptr = curr_input_row_ptr; } - curr_input_depth_ptr += input_stride_z; + curr_input_depth_ptr += is_nhwc ? input_stride_x : input_stride_z; curr_input_row_ptr = curr_input_depth_ptr; tmp_input_ptr = curr_input_depth_ptr; } @@ -161,21 +165,24 @@ void NEWeightsReshapeKernel::configure(const ITensor *input, const ITensor *bias _bias = bias; _output = output; + const DataLayout data_layout = input->info()->data_layout(); + const bool is_nhwc = data_layout == DataLayout::NHWC; + switch(_input->info()->element_size()) { case 4: { - _func = &weights_reshape<uint32_t>; + _func = is_nhwc ? &weights_reshape<uint32_t, true> : &weights_reshape<uint32_t, false>; break; } case 2: { - _func = &weights_reshape<uint16_t>; + _func = is_nhwc ? &weights_reshape<uint16_t, true> : &weights_reshape<uint16_t, false>; break; } case 1: { - _func = &weights_reshape<uint8_t>; + _func = is_nhwc ? &weights_reshape<uint8_t, true> : &weights_reshape<uint8_t, false>; break; } default: diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp index 5a35463365..a5f30557a0 100644 --- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp +++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp @@ -109,6 +109,14 @@ Status NEConvolutionLayerReshapeWeights::validate(const ITensorInfo *weights, co ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); } + // Checks performed when biases are present + if(append_bias) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); + ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3)); + ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); + } + if(transpose1xW) { TensorInfo weights_reshaped = weights->clone()->set_tensor_shape(get_reshaped_weights_shape(weights, append_bias)); @@ -159,7 +167,7 @@ TensorShape get_reshaped_weights_shape_conv(const ITensorInfo *weights, bool app Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const PadStrideInfo &conv_info, const WeightsInfo &weights_info, const ActivationLayerInfo &act_info, DataType &dt, - bool &append_bias, + bool &append_bias, bool &skip_im2col, bool &are_weights_reshaped, unsigned int &kernel_width, unsigned int &kernel_height, bool &is_fully_connected_convolution, bool &is_interleaved, bool &is_quantized, bool &is_activationlayer_enabled, unsigned int &mat_weights_cols, unsigned int &mat_weights_rows, @@ -168,9 +176,17 @@ Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInf ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, weights); - ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && weights->dimension(2) != input->dimension(2)); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, weights); + + DataLayout data_layout = input->data_layout(); + const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + const int idx_channel = get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL); + + ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && weights->dimension(idx_channel) != input->dimension(idx_channel)); ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4); ARM_COMPUTE_RETURN_ERROR_ON(weights_info.are_reshaped() && is_data_type_quantized_asymmetric(input->data_type())); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(data_layout == DataLayout::NHWC && input->data_type() != DataType::F32, "NHWC is only supported for FP32 data type."); dt = input->data_type(); is_quantized = is_data_type_quantized_asymmetric(dt); @@ -190,14 +206,16 @@ Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInf ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1); } + // If we have 1x1 convolution and data layout is NHWC we can disable im2col append_bias = (biases != nullptr) && (!is_quantized); are_weights_reshaped = weights_info.are_reshaped(); - kernel_width = (are_weights_reshaped) ? weights_info.kernel_size().first : weights->dimension(0); - kernel_height = (are_weights_reshaped) ? weights_info.kernel_size().second : weights->dimension(1); + kernel_width = (are_weights_reshaped) ? weights_info.kernel_size().first : weights->dimension(idx_width); + 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(0) * weights->dimension(1) * weights->dimension(2) + (append_bias ? 1 : 0); + 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); - std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(0), input->dimension(1), kernel_width, kernel_height, + std::tie(conv_w, conv_h) = scaled_dimensions(input->dimension(idx_width), input->dimension(idx_height), kernel_width, kernel_height, conv_info, dilation); // Check if its a "fully connected" convolution @@ -211,9 +229,9 @@ Status validate_and_initialize_values(const ITensorInfo *input, const ITensorInf NEGEMMConvolutionLayer::NEGEMMConvolutionLayer(const std::shared_ptr<IMemoryManager> &memory_manager) : _asm_glue(), _memory_group(memory_manager), _input_im2col_kernel(), _input_interleave_kernel(), _reshape_weights(), _mm_kernel(), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), - _output_col2im_kernel(), _activationlayer_function(), _original_weights(nullptr), _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), _tmp_output(), - _workspace(), _B_pretransposed(), _append_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false), _is_quantized(false), _is_interleaved(false), - _is_activationlayer_enabled(false) + _output_col2im_kernel(), _activationlayer_function(), _add_bias_kernel(), _original_weights(nullptr), _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), _gemm_output(), + _tmp_output(), _workspace(), _B_pretransposed(), _data_layout(DataLayout::NCHW), _append_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false), _is_quantized(false), + _is_interleaved(false), _is_activationlayer_enabled(false), _skip_im2col(false) { } @@ -255,7 +273,13 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig unsigned int conv_w = 0; unsigned int conv_h = 0; - Status status = validate_and_initialize_values(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(), conv_info, weights_info, act_info, dt, _append_bias, + _data_layout = input->info()->data_layout(); + const bool is_nhwc = _data_layout == DataLayout::NHWC; + const int idx_width = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT); + const int idx_channel = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::CHANNEL); + + Status status = validate_and_initialize_values(input->info(), weights->info(), (biases == nullptr) ? nullptr : biases->info(), conv_info, weights_info, act_info, dt, _append_bias, _skip_im2col, _are_weights_reshaped, kernel_width, kernel_height, _is_fully_connected_convolution, _is_interleaved, _is_quantized, _is_activationlayer_enabled, @@ -272,20 +296,12 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig // Reshape weights if needed if(run_optimised) { - if(_are_weights_reshaped) - { - mat_weights_cols = weights_info.num_kernels(); - mat_weights_rows = weights->info()->dimension(1); - } - else - { - TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows }; + TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows }; - // Create tensor to store the reshaped weights - _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position)); - _reshape_weights.configure(weights, biases, &_weights_reshaped, false /* 1xW transpose */); - weights = &_weights_reshaped; - } + // Create tensor to store the reshaped weights + _weights_reshaped.allocator()->init(TensorInfo(reshaped_weights_shape, 1, dt, fixed_point_position)); + _reshape_weights.configure(weights, biases, &_weights_reshaped, false /* 1xW transpose */); + weights = &_weights_reshaped; } else { @@ -294,12 +310,12 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig if(_is_fully_connected_convolution || _is_quantized) { mat_weights_cols = weights_info.num_kernels(); - mat_weights_rows = weights->info()->dimension(1); + mat_weights_rows = weights->info()->dimension(idx_height); } else { mat_weights_cols = weights_info.num_kernels(); - mat_weights_rows = weights_info.kernel_size().first * weights_info.kernel_size().second * input->info()->dimension(2) + (_append_bias ? 1 : 0); + mat_weights_rows = weights_info.kernel_size().first * weights_info.kernel_size().second * input->info()->dimension(idx_channel) + (_append_bias ? 1 : 0); } } else @@ -325,48 +341,56 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig } } - // Create tensor to store im2col reshaped inputs - const unsigned int mat_input_cols = mat_weights_rows; - const unsigned int mat_input_rows = conv_w * conv_h; - - TensorShape shape_im2col(input->info()->tensor_shape()); - shape_im2col.set(0, mat_input_cols); - shape_im2col.set(1, mat_input_rows); - shape_im2col.set(2, 1); - _input_im2col_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col)); - _memory_group.manage(&_input_im2col_reshaped); + // In case we skip im2col we have to add bias + if(!_skip_im2col) + { + const unsigned int mat_input_cols = mat_weights_rows; + const unsigned int mat_input_rows = conv_w * conv_h; + + // Create tensor to store im2col reshaped inputs + TensorShape shape_im2col(input->info()->tensor_shape()); + shape_im2col.set(0, mat_input_cols); + shape_im2col.set(1, mat_input_rows); + shape_im2col.set(2, 1); + _input_im2col_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col)); + _memory_group.manage(&_input_im2col_reshaped); + + // Create tensor (interleave) to prepare input tensor for GEMM + if(!_is_fully_connected_convolution && !run_optimised && _is_interleaved) + { + TensorShape shape_interleaved(shape_im2col); + shape_interleaved.set(idx_width, shape_interleaved.x() * 4); + shape_interleaved.set(idx_height, std::ceil(shape_interleaved[idx_height] / 4.f)); + _input_interleaved_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_interleaved)); + _memory_group.manage(&_input_interleaved_reshaped); + } - // Create tensor (interleave) to prepare input tensor for GEMM - if(!_is_fully_connected_convolution && !run_optimised && _is_interleaved) + // Create GEMM output tensor + TensorShape shape_gemm(_input_im2col_reshaped.info()->tensor_shape()); + shape_gemm.set(0, mat_weights_cols); + shape_gemm.set(1, mat_input_rows); + const DataType gemm_data_type = _is_quantized ? DataType::S32 : dt; + // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input. + TensorInfo info_gemm(shape_gemm, 1, gemm_data_type, input->info()->fixed_point_position()); + info_gemm.set_quantization_info(output->info()->quantization_info()); + _gemm_output.allocator()->init(info_gemm); + + // FIXME: enabling memory manager for _gemm_output gives incorrect results (maybe bound to the assembly kernel in GEMMLowp?) + // _memory_group.manage(&_gemm_output); + + // Configure im2col + _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _append_bias, false, false, dilation); + } + else if(_append_bias) { - TensorShape shape_interleaved(shape_im2col); - shape_interleaved.set(0, shape_interleaved.x() * 4); - shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f)); - _input_interleaved_reshaped.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_interleaved)); - _memory_group.manage(&_input_interleaved_reshaped); + // Configure add bias kernel + _add_bias_kernel.configure(output, biases, output, ConvertPolicy::SATURATE); } - // Create GEMM output tensor - TensorShape shape_gemm(_input_im2col_reshaped.info()->tensor_shape()); - shape_gemm.set(0, mat_weights_cols); - shape_gemm.set(1, mat_input_rows); - const DataType gemm_data_type = _is_quantized ? DataType::S32 : dt; - // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input. - TensorInfo info_gemm(shape_gemm, 1, gemm_data_type, input->info()->fixed_point_position()); - info_gemm.set_quantization_info(output->info()->quantization_info()); - _gemm_output.allocator()->init(info_gemm); - - // FIXME: enabling memory manager for _gemm_output gives incorrect results (maybe bound to the assembly kernel in GEMMLowp?) - // _memory_group.manage(&_gemm_output); - - // Configure kernels - // Configure im2col - _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _append_bias, false, false, dilation); - // Configure matrix multiply if(run_optimised) { - if(!setup_assembly_kernel(&_input_im2col_reshaped, weights, &_gemm_output, 1.f, 0.f, true, _workspace, _B_pretransposed, _memory_group, _asm_glue)) + if(!setup_assembly_kernel(_skip_im2col ? input : &_input_im2col_reshaped, weights, is_nhwc ? output : &_gemm_output, 1.f, 0.f, true, _workspace, _B_pretransposed, _memory_group, _asm_glue)) { ARM_COMPUTE_ERROR("setup_assembly_kernel failed."); } @@ -379,8 +403,8 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped); // Configure GEMM - configure_mm(&_input_interleaved_reshaped, weights, &_gemm_output, _is_interleaved, GEMMReshapeInfo(_input_im2col_reshaped.info()->dimension(1), 0 /* no transpose */, - _input_im2col_reshaped.info()->dimension(0))); + configure_mm(&_input_interleaved_reshaped, weights, &_gemm_output, _is_interleaved, GEMMReshapeInfo(_input_im2col_reshaped.info()->dimension(idx_height), 0 /* no transpose */, + _input_im2col_reshaped.info()->dimension(idx_width))); _input_interleaved_reshaped.allocator()->allocate(); } else @@ -389,29 +413,36 @@ void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weig } } - _input_im2col_reshaped.allocator()->allocate(); - - // Configure output stage for quantized case - if(_is_quantized) + if(!_skip_im2col) { - const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info(); + _input_im2col_reshaped.allocator()->allocate(); - float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale; - int output_multiplier, output_shift; - quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); - _memory_group.manage(&_tmp_output); - _gemmlowp_output_stage.configure(&_gemm_output, biases, &_tmp_output, output_multiplier, output_shift, output_quant_info.offset); - } + // Configure output stage for quantized case + if(_is_quantized) + { + const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info(); - // Configure Col2Im - _output_col2im_kernel.configure(_is_quantized ? &_tmp_output : &_gemm_output, output, Size2D(conv_w, conv_h)); - if(_is_quantized) - { - _tmp_output.allocator()->allocate(); + float multiplier = input->info()->quantization_info().scale * weights->info()->quantization_info().scale / output_quant_info.scale; + int output_multiplier, output_shift; + quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); + _memory_group.manage(&_tmp_output); + _gemmlowp_output_stage.configure(&_gemm_output, biases, &_tmp_output, output_multiplier, output_shift, output_quant_info.offset); + } + + // Configure Col2Im + if(!is_nhwc) + { + _output_col2im_kernel.configure(_is_quantized ? &_tmp_output : &_gemm_output, output, Size2D(conv_w, conv_h)); + } + + if(_is_quantized) + { + _tmp_output.allocator()->allocate(); + } + _gemm_output.allocator()->allocate(); } - _gemm_output.allocator()->allocate(); - ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one"); + ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(idx_width) != conv_w) || (output->info()->dimension(idx_height) != conv_h), "Output shape does not match the expected one"); // Allocate intermediate tensor if(!_are_weights_reshaped) @@ -433,6 +464,7 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI DataType dt{}; bool append_bias{}; + bool skip_im2col{}; bool are_weights_reshaped{}; bool is_fully_connected_convolution{}; bool is_interleaved{}; @@ -445,7 +477,12 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI unsigned int conv_w = 0; unsigned int conv_h = 0; - Status status = validate_and_initialize_values(input, weights, biases, conv_info, weights_info, act_info, dt, append_bias, are_weights_reshaped, kernel_width, kernel_height, + const DataLayout data_layout = input->data_layout(); + const bool is_nhwc = data_layout == DataLayout::NHWC; + const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); + + Status status = validate_and_initialize_values(input, weights, biases, conv_info, weights_info, act_info, dt, append_bias, skip_im2col, are_weights_reshaped, kernel_width, kernel_height, is_fully_connected_convolution, is_interleaved, is_quantized, is_activationlayer_enabled, mat_weights_cols, mat_weights_rows, conv_w, conv_h, dilation); @@ -461,7 +498,6 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI optimised_kernel = true; } - // Validate im2col const unsigned int mat_input_cols = mat_weights_rows; const unsigned int mat_input_rows = conv_w * conv_h; TensorShape shape_im2col = input->tensor_shape(); @@ -469,7 +505,17 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI shape_im2col.set(1, mat_input_rows); shape_im2col.set(2, 1); TensorInfo im2_col_info = input->clone()->set_tensor_shape(shape_im2col); - ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, &im2_col_info, kernel_weights, conv_info, append_bias, false, false, dilation)); + + if(!skip_im2col) + { + // Validate im2col + ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, &im2_col_info, kernel_weights, conv_info, append_bias, false, false, dilation)); + } + else if(append_bias) + { + // Validate add bias kernel + ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAdditionKernel::validate(output, biases, output, ConvertPolicy::SATURATE)); + } // Create GEMM output tensor TensorShape shape_gemm(im2_col_info.tensor_shape()); @@ -511,8 +557,8 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI if(is_interleaved) { TensorShape shape_interleaved = shape_im2col; - shape_interleaved.set(0, shape_interleaved.x() * 4); - shape_interleaved.set(1, std::ceil(shape_interleaved.y() / 4.f)); + shape_interleaved.set(idx_width, shape_interleaved.x() * 4); + shape_interleaved.set(idx_height, std::ceil(shape_interleaved.y() / 4.f)); TensorInfo input_interleaved_info = input->clone()->set_tensor_shape(shape_interleaved); ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMInterleave4x4Kernel::validate(&im2_col_info, &input_interleaved_info)); ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&input_interleaved_info, weights, &gemm_output_info, 1.f, is_interleaved, GEMMReshapeInfo(shape_im2col[1], // m @@ -524,10 +570,12 @@ Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorI ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMMatrixMultiplyKernel::validate(&im2_col_info, weights, &gemm_output_info, 1.f, is_interleaved, GEMMReshapeInfo())); } } + if(!is_nhwc) + { + ARM_COMPUTE_RETURN_ON_ERROR(NECol2ImKernel::validate(&gemm_output_info, output, Size2D(conv_w, conv_h))); + } - ARM_COMPUTE_RETURN_ON_ERROR(NECol2ImKernel::validate(&gemm_output_info, output, Size2D(conv_w, conv_h))); - - ARM_COMPUTE_RETURN_ERROR_ON_MSG((output->dimension(0) != conv_w) || (output->dimension(1) != conv_h), "Output shape does not match the expected one"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG((output->dimension(idx_width) != conv_w) || (output->dimension(idx_height) != conv_h), "Output shape does not match the expected one"); if(act_info.enabled()) { @@ -553,8 +601,12 @@ void NEGEMMConvolutionLayer::run() _memory_group.acquire(); - // Run input reshaping - NEScheduler::get().schedule(&_input_im2col_kernel, Window::DimY); + if(!_skip_im2col) + { + // Run input reshaping + unsigned int _y_dim = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT); + NEScheduler::get().schedule(&_input_im2col_kernel, _y_dim); + } // Runs matrix multiply on reshaped matrices if(_asm_glue._optimised_kernel != nullptr) @@ -585,6 +637,11 @@ void NEGEMMConvolutionLayer::run() } } + if(_skip_im2col && _append_bias) + { + NEScheduler::get().schedule(&_add_bias_kernel, Window::DimY); + } + // Run output stage for quantized case if(_is_quantized) { @@ -592,7 +649,10 @@ void NEGEMMConvolutionLayer::run() } // Reshape output matrix - NEScheduler::get().schedule(&_output_col2im_kernel, Window::DimY); + if(_data_layout == DataLayout::NCHW) + { + NEScheduler::get().schedule(&_output_col2im_kernel, Window::DimY); + } if(_is_activationlayer_enabled) { diff --git a/tests/validation/CL/ConvolutionLayer.cpp b/tests/validation/CL/ConvolutionLayer.cpp index a2b55a8555..935a6ebefa 100644 --- a/tests/validation/CL/ConvolutionLayer.cpp +++ b/tests/validation/CL/ConvolutionLayer.cpp @@ -198,20 +198,22 @@ using CLGEMMConvolutionLayerFixture = ConvolutionValidationFixture<CLTensor, CLA TEST_SUITE(Float) TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", { DataLayout::NCHW })), ActivationFunctionsDataset)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", { DataLayout::NCHW })), ActivationFunctionsDataset)) { // Validate output @@ -221,20 +223,22 @@ TEST_SUITE_END() TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", { DataLayout::NCHW })), ActivationFunctionsDataset)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", { DataLayout::NCHW })), ActivationFunctionsDataset)) { // Validate output diff --git a/tests/validation/CL/DilatedConvolutionLayer.cpp b/tests/validation/CL/DilatedConvolutionLayer.cpp index 9ee002cc5a..d02497d853 100644 --- a/tests/validation/CL/DilatedConvolutionLayer.cpp +++ b/tests/validation/CL/DilatedConvolutionLayer.cpp @@ -164,17 +164,19 @@ using CLGEMMDilatedConvolutionLayerFixture = ConvolutionValidationFixture<CLTens TEST_SUITE(Float) TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMDilatedConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallDilatedConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMDilatedConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallDilatedConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", { DataLayout::NCHW })), framework::dataset::make("ActivationLayerInfo", ActivationLayerInfo()))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMDilatedConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeDilatedConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMDilatedConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDilatedConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", { DataLayout::NCHW })), framework::dataset::make("ActivationLayerInfo", ActivationLayerInfo()))) { // Validate output @@ -183,17 +185,19 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMDilatedConvolutionLayerFixture<half>, fra TEST_SUITE_END() TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMDilatedConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallDilatedConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMDilatedConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallDilatedConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", { DataLayout::NCHW })), framework::dataset::make("ActivationLayerInfo", ActivationLayerInfo()))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMDilatedConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeDilatedConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMDilatedConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDilatedConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", { DataLayout::NCHW })), framework::dataset::make("ActivationLayerInfo", ActivationLayerInfo()))) { // Validate output diff --git a/tests/validation/GLES_COMPUTE/ConvolutionLayer.cpp b/tests/validation/GLES_COMPUTE/ConvolutionLayer.cpp index bc0170fa06..0f8151278a 100644 --- a/tests/validation/GLES_COMPUTE/ConvolutionLayer.cpp +++ b/tests/validation/GLES_COMPUTE/ConvolutionLayer.cpp @@ -117,19 +117,23 @@ using GCConvolutionLayerFixture = ConvolutionValidationFixture<GCTensor, GCAcces TEST_SUITE(Float) TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, GCConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, GCConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", + DataLayout::NCHW)), ActivationFunctionsDataset)) { // Validate output validate(GCAccessor(_target), _reference, tolerance_f16, tolerance_num); } -FIXTURE_DATA_TEST_CASE(RunLarge, GCConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunLarge, GCConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", + DataLayout::NCHW)), ActivationFunctionsDataset)) { // Validate output @@ -138,17 +142,21 @@ FIXTURE_DATA_TEST_CASE(RunLarge, GCConvolutionLayerFixture<half>, framework::Dat TEST_SUITE_END() TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, GCConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, GCConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", + DataLayout::NCHW)), ActivationFunctionsDataset)) { // Validate output validate(GCAccessor(_target), _reference, tolerance_f32, tolerance_num); } -FIXTURE_DATA_TEST_CASE(RunLarge, GCConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunLarge, GCConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", + DataLayout::NCHW)), ActivationFunctionsDataset)) { // Validate output diff --git a/tests/validation/NEON/ConvolutionLayer.cpp b/tests/validation/NEON/ConvolutionLayer.cpp index 8b2e21e796..4f59345f6c 100644 --- a/tests/validation/NEON/ConvolutionLayer.cpp +++ b/tests/validation/NEON/ConvolutionLayer.cpp @@ -194,17 +194,19 @@ using NEGEMMConvolutionLayerFixture = ConvolutionValidationFixture<Tensor, Acces TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", { DataLayout::NCHW })), ActivationFunctionsDataset)) { // Validate output validate(Accessor(_target), _reference, tolerance_f16); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", { DataLayout::NCHW })), ActivationFunctionsDataset)) { // Validate output @@ -214,17 +216,19 @@ TEST_SUITE_END() #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsDataset)) { // Validate output validate(Accessor(_target), _reference, tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsDataset)) { // Validate output @@ -240,7 +244,7 @@ TEST_SUITE(FixedPoint) TEST_SUITE(QS8) // We test for fixed point precision [4,6] FIXTURE_DATA_TEST_CASE(RunTiny, NEGEMMConvolutionLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::TinyConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 4, 7)), ActivationFunctionsDataset)) @@ -249,7 +253,7 @@ FIXTURE_DATA_TEST_CASE(RunTiny, NEGEMMConvolutionLayerFixedPointFixture<int8_t>, validate(Accessor(_target), _reference, tolerance_q); } FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 4, 7)), ActivationFunctionsDataset)) @@ -262,7 +266,7 @@ TEST_SUITE_END() TEST_SUITE(QS16) // Testing for fixed point position [1,14) FIXTURE_DATA_TEST_CASE(RunTiny, NEGEMMConvolutionLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::TinyConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14)), ActivationFunctionsDataset)) @@ -271,7 +275,7 @@ FIXTURE_DATA_TEST_CASE(RunTiny, NEGEMMConvolutionLayerFixedPointFixture<int16_t> validate(Accessor(_target), _reference, tolerance_q); } FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14)), ActivationFunctionsDataset)) diff --git a/tests/validation/NEON/DilatedConvolutionLayer.cpp b/tests/validation/NEON/DilatedConvolutionLayer.cpp index 358cec3d6f..d9fd093c8e 100644 --- a/tests/validation/NEON/DilatedConvolutionLayer.cpp +++ b/tests/validation/NEON/DilatedConvolutionLayer.cpp @@ -157,17 +157,19 @@ using NEGEMMDilatedConvolutionLayerFixture = ConvolutionValidationFixture<Tensor TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMDilatedConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallDilatedConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMDilatedConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallDilatedConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", { DataLayout::NCHW })), framework::dataset::make("ActivationLayerInfo", ActivationLayerInfo()))) { // Validate output validate(Accessor(_target), _reference, tolerance_f16); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMDilatedConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeDilatedConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMDilatedConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDilatedConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", { DataLayout::NCHW })), framework::dataset::make("ActivationLayerInfo", ActivationLayerInfo()))) { // Validate output @@ -177,17 +179,19 @@ TEST_SUITE_END() #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMDilatedConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallDilatedConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMDilatedConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallDilatedConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), framework::dataset::make("ActivationLayerInfo", ActivationLayerInfo()))) { // Validate output validate(Accessor(_target), _reference, tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMDilatedConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeDilatedConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMDilatedConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDilatedConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), framework::dataset::make("ActivationLayerInfo", ActivationLayerInfo()))) { // Validate output @@ -204,7 +208,7 @@ TEST_SUITE(QS8) // We test for fixed point precision [4,6] FIXTURE_DATA_TEST_CASE(RunTiny, NEGEMMDilatedConvolutionLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::TinyDilatedConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 4, 7)), framework::dataset::make("ActivationLayerInfo", ActivationLayerInfo()))) @@ -214,7 +218,7 @@ FIXTURE_DATA_TEST_CASE(RunTiny, NEGEMMDilatedConvolutionLayerFixedPointFixture<i } FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMDilatedConvolutionLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::SmallDilatedConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 4, 7)), framework::dataset::make("ActivationLayerInfo", ActivationLayerInfo()))) @@ -228,7 +232,7 @@ TEST_SUITE(QS16) // Testing for fixed point position [1,14) FIXTURE_DATA_TEST_CASE(RunTiny, NEGEMMDilatedConvolutionLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::TinyDilatedConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14)), framework::dataset::make("ActivationLayerInfo", ActivationLayerInfo()))) @@ -238,7 +242,7 @@ FIXTURE_DATA_TEST_CASE(RunTiny, NEGEMMDilatedConvolutionLayerFixedPointFixture<i } FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMDilatedConvolutionLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::SmallDilatedConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14)), framework::dataset::make("ActivationLayerInfo", ActivationLayerInfo()))) diff --git a/tests/validation/fixtures/ConvolutionLayerFixture.h b/tests/validation/fixtures/ConvolutionLayerFixture.h index 1bcffed526..93de24d1bd 100644 --- a/tests/validation/fixtures/ConvolutionLayerFixture.h +++ b/tests/validation/fixtures/ConvolutionLayerFixture.h @@ -35,6 +35,7 @@ #include "tests/validation/Helpers.h" #include "tests/validation/reference/ActivationLayer.h" #include "tests/validation/reference/ConvolutionLayer.h" +#include "tests/validation/reference/Permute.h" #include "tests/validation/reference/Utils.h" #include <random> @@ -56,13 +57,14 @@ public: public: template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, bool reshape_weights, - DataType data_type, int fractional_bits, QuantizationInfo quantization_info, ActivationLayerInfo act_info) + DataType data_type, DataLayout data_layout, int fractional_bits, QuantizationInfo quantization_info, ActivationLayerInfo act_info) { _data_type = data_type; _is_quantized = is_data_type_quantized_asymmetric(data_type); _bias_data_type = _is_quantized ? DataType::S32 : data_type; _fractional_bits = fractional_bits; _quantization_info = quantization_info; + _data_layout = data_layout; _target = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, reshape_weights, dilation, act_info); _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, dilation, act_info); @@ -98,46 +100,27 @@ protected: } } - TensorType compute_target(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const PadStrideInfo &info, + TensorType compute_target(TensorShape input_shape, TensorShape weights_shape, const TensorShape &bias_shape, TensorShape output_shape, const PadStrideInfo &info, bool reshape_weights, const Size2D &dilation, const ActivationLayerInfo act_info) { - const bool is_optimised = std::is_same<FunctionType, NEConvolutionLayer>::value && _data_type == DataType::F32; - - WeightsInfo weights_info(!reshape_weights, weights_shape.x(), weights_shape.y(), weights_shape[3]); - TensorShape reshaped_weights_shape(weights_shape); - - if(!reshape_weights) + if(_data_layout == DataLayout::NHWC) { - // Check if its a "fully connected" convolution - const bool is_fully_connected_convolution = (output_shape.x() == 1 && output_shape.y() == 1); - - reshaped_weights_shape.collapse(3); + permute(input_shape, PermutationVector(2U, 0U, 1U)); + permute(weights_shape, PermutationVector(2U, 0U, 1U)); + permute(output_shape, PermutationVector(2U, 0U, 1U)); + } - if(bias_shape.total_size() > 0 && !_is_quantized) - { - // Add bias to the weights reshaped matrix - reshaped_weights_shape.set(0, reshaped_weights_shape.x() + 1); - } + const int idx_width = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT); - if(is_fully_connected_convolution || is_optimised) - { - const size_t shape_x = reshaped_weights_shape.x(); - reshaped_weights_shape.set(0, reshaped_weights_shape.y()); - reshaped_weights_shape.set(1, shape_x); - } - else - { - const int interleave_width = 16 / data_size_from_type(_data_type); - reshaped_weights_shape.set(0, reshaped_weights_shape.x() * interleave_width); - reshaped_weights_shape.set(1, static_cast<unsigned int>(std::ceil(reshaped_weights_shape.y() / static_cast<float>(interleave_width)))); - } - } + WeightsInfo weights_info(!reshape_weights, weights_shape[idx_width], weights_shape[idx_height], weights_shape[3]); + TensorShape reshaped_weights_shape(weights_shape); // Create tensors - TensorType src = create_tensor<TensorType>(input_shape, _data_type, 1, _fractional_bits, _quantization_info); - TensorType weights = create_tensor<TensorType>(reshaped_weights_shape, _data_type, 1, _fractional_bits, _quantization_info); - TensorType bias = create_tensor<TensorType>(bias_shape, _bias_data_type, 1, _fractional_bits, _quantization_info); - TensorType dst = create_tensor<TensorType>(output_shape, _data_type, 1, _fractional_bits, _quantization_info); + TensorType src = create_tensor<TensorType>(input_shape, _data_type, 1, _fractional_bits, _quantization_info, _data_layout); + TensorType weights = create_tensor<TensorType>(reshaped_weights_shape, _data_type, 1, _fractional_bits, _quantization_info, _data_layout); + TensorType bias = create_tensor<TensorType>(bias_shape, _bias_data_type, 1, _fractional_bits, _quantization_info, _data_layout); + TensorType dst = create_tensor<TensorType>(output_shape, _data_type, 1, _fractional_bits, _quantization_info, _data_layout); // Create and configure function FunctionType conv; @@ -161,48 +144,8 @@ protected: // Fill tensors fill(AccessorType(src), 0); - - if(!reshape_weights) - { - const bool is_fully_connected_convolution = (output_shape.x() == 1 && output_shape.y() == 1); - TensorShape tmp_weights_shape(weights_shape); - SimpleTensor<T> tmp_weights(tmp_weights_shape, _data_type, 1, _fractional_bits, _quantization_info); - - // Fill with original shape - fill(tmp_weights, 1); - - if(_is_quantized) - { - fill(AccessorType(bias), 2); - tmp_weights = linearise_weights(tmp_weights); - } - else - { - SimpleTensor<T> tmp_bias(bias_shape, _bias_data_type, 1, _fractional_bits, _quantization_info); - fill(tmp_bias, 2); - tmp_weights = linearise_weights(tmp_weights, &tmp_bias); - } - - if(!is_fully_connected_convolution && !is_optimised) - { - // Transpose with interleave - const int interleave_size = 16 / tmp_weights.element_size(); - tmp_weights = transpose(std::move(tmp_weights), interleave_size); - } - - AccessorType weights_accessor(weights); - - for(int i = 0; i < tmp_weights.num_elements(); ++i) - { - Coordinates coord = index2coord(tmp_weights.shape(), i); - std::copy_n(static_cast<const T *>(tmp_weights(coord)), 1, static_cast<T *>(weights_accessor(coord))); - } - } - else - { - fill(AccessorType(weights), 1); - fill(AccessorType(bias), 2); - } + fill(AccessorType(weights), 1); + fill(AccessorType(bias), 2); // Compute NEConvolutionLayer function conv.run(); @@ -232,53 +175,10 @@ protected: SimpleTensor<T> _reference{}; DataType _data_type{}; DataType _bias_data_type{}; + DataLayout _data_layout{}; int _fractional_bits{}; QuantizationInfo _quantization_info{}; bool _is_quantized = false; - -private: - template <typename U> - SimpleTensor<U> linearise_weights(const SimpleTensor<U> &weights, const SimpleTensor<U> *biases = nullptr) - { - TensorShape dst_shape(weights.shape()); - dst_shape.collapse(3); - - if(biases != nullptr) - { - dst_shape.set(0, dst_shape.x() + 1); - } - - const size_t shape_x = dst_shape.x(); - dst_shape.set(0, dst_shape.y()); - dst_shape.set(1, shape_x); - - SimpleTensor<U> dst(dst_shape, weights.data_type()); - - // Don't iterate over biases yet - for(int weights_idx = 0; weights_idx < weights.num_elements(); ++weights_idx) - { - Coordinates weights_coord = index2coord(weights.shape(), weights_idx); - const int dst_row = weights_idx % weights.shape().total_size_lower(3); - Coordinates dst_coord{ weights_coord[3], dst_row, weights_coord[4] }; - const int dst_idx = coord2index(dst.shape(), dst_coord); - - dst[dst_idx] = weights[weights_idx]; - } - if(biases != nullptr) - { - // Fill last row with biases - for(int bias_idx = 0; bias_idx < biases->num_elements(); ++bias_idx) - { - Coordinates bias_coord = index2coord(biases->shape(), bias_idx); - Coordinates dst_coord{ bias_coord.x(), static_cast<int>(dst.shape().y()) - 1, bias_coord.y() }; - int dst_idx = coord2index(dst.shape(), dst_coord); - - dst[dst_idx] = (*biases)[bias_idx]; - } - } - - return dst; - } }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T> @@ -287,11 +187,10 @@ class ConvolutionValidationFixture : public ConvolutionValidationGenericFixture< public: template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, bool reshape_weights, DataType data_type, - ActivationLayerInfo act_info) + DataLayout data_layout, ActivationLayerInfo act_info) { - ConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, dilation, reshape_weights, data_type, 0, - QuantizationInfo(), - act_info); + ConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, dilation, reshape_weights, data_type, data_layout, 0, + QuantizationInfo(), act_info); } }; @@ -301,11 +200,11 @@ class ConvolutionValidationFixedPointFixture : public ConvolutionValidationGener public: template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, bool reshape_weights, DataType data_type, - int fractional_bits, - ActivationLayerInfo act_info) + int fractional_bits, ActivationLayerInfo act_info) { - ConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, dilation, reshape_weights, data_type, fractional_bits, - QuantizationInfo(), act_info); + ConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, dilation, reshape_weights, data_type, + DataLayout::NCHW, + fractional_bits, QuantizationInfo(), act_info); } }; @@ -317,7 +216,8 @@ public: void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, bool reshape_weights, DataType data_type, QuantizationInfo quantization_info, ActivationLayerInfo act_info) { - ConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, dilation, reshape_weights, data_type, 0, + ConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, dilation, reshape_weights, data_type, + DataLayout::NCHW, 0, quantization_info, act_info); } }; diff --git a/tests/validation/reference/ConvolutionLayer.cpp b/tests/validation/reference/ConvolutionLayer.cpp index 617e85c8c2..fe558ba4af 100644 --- a/tests/validation/reference/ConvolutionLayer.cpp +++ b/tests/validation/reference/ConvolutionLayer.cpp @@ -26,6 +26,7 @@ #include "tests/validation/FixedPoint.h" #include "tests/validation/Helpers.h" #include "tests/validation/reference/Convolution3d.h" +#include "tests/validation/reference/Permute.h" #include "tests/validation/reference/Utils.h" #include "tests/validation/reference/UtilsQuantizedAsymm.h" @@ -46,12 +47,9 @@ namespace } // namespace template <typename T, typename TB> -SimpleTensor<T> convolution_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, const TensorShape &output_shape, const PadStrideInfo &info, - const Size2D &dilation) +SimpleTensor<T> convolution_layer_nchw(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, SimpleTensor<T> &dst, const PadStrideInfo &info, + const Size2D &dilation) { - // Create reference - SimpleTensor<T> dst{ output_shape, src.data_type(), 1, src.fixed_point_position(), src.quantization_info() }; - // Compute reference const int width_in = src.shape().x(); const int height_in = src.shape().y(); @@ -105,6 +103,26 @@ SimpleTensor<T> convolution_layer(const SimpleTensor<T> &src, const SimpleTensor return dst; } +template <typename T, typename TB> +SimpleTensor<T> convolution_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, const TensorShape &output_shape, const PadStrideInfo &info, + const Size2D &dilation) +{ + // Create reference + SimpleTensor<T> dst{ output_shape, src.data_type(), 1, src.fixed_point_position(), src.quantization_info() }; + + if(src.data_layout() == DataLayout::NHWC) + { + SimpleTensor<T> src_nchw = reference::permute<T>(src, PermutationVector(1U, 2U, 0U)); + SimpleTensor<T> weights_nchw = reference::permute<T>(weights, PermutationVector(1U, 2U, 0U)); + SimpleTensor<T> dst_nchw = reference::permute<T>(dst, PermutationVector(1U, 2U, 0U)); + + return reference::permute<T>(convolution_layer_nchw(src_nchw, weights_nchw, bias, dst_nchw, info, dilation), PermutationVector(2U, 0U, 1U)); + } + else + { + return convolution_layer_nchw(src, weights, bias, dst, info, dilation); + } +} template SimpleTensor<float> convolution_layer(const SimpleTensor<float> &src, const SimpleTensor<float> &weights, const SimpleTensor<float> &bias, const TensorShape &output_shape, const PadStrideInfo &info, const Size2D &dilation); diff --git a/tests/validation/reference/Permute.cpp b/tests/validation/reference/Permute.cpp index c670c3ea6e..bbb2e8d4d7 100644 --- a/tests/validation/reference/Permute.cpp +++ b/tests/validation/reference/Permute.cpp @@ -57,11 +57,11 @@ SimpleTensor<T> permute(const SimpleTensor<T> &src, PermutationVector perm) return dst; } +template SimpleTensor<int8_t> permute(const SimpleTensor<int8_t> &src, PermutationVector perm); template SimpleTensor<uint8_t> permute(const SimpleTensor<uint8_t> &src, PermutationVector perm); +template SimpleTensor<int16_t> permute(const SimpleTensor<int16_t> &src, PermutationVector perm); template SimpleTensor<uint16_t> permute(const SimpleTensor<uint16_t> &src, PermutationVector perm); template SimpleTensor<uint32_t> permute(const SimpleTensor<uint32_t> &src, PermutationVector perm); -template SimpleTensor<int8_t> permute(const SimpleTensor<int8_t> &src, PermutationVector perm); -template SimpleTensor<int16_t> permute(const SimpleTensor<int16_t> &src, PermutationVector perm); template SimpleTensor<float> permute(const SimpleTensor<float> &src, PermutationVector perm); template SimpleTensor<half> permute(const SimpleTensor<half> &src, PermutationVector perm); } // namespace reference |