arm_compute v18.08
diff --git a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
index 2888b43..92e641e 100644
--- a/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEGEMMConvolutionLayer.cpp
@@ -23,10 +23,10 @@
*/
#include "arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h"
-#include "arm_compute/core/PixelValue.h"
#include "arm_compute/core/Size2D.h"
#include "arm_compute/core/Utils.h"
#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
#include "arm_compute/runtime/NEON/NEScheduler.h"
#include "support/ToolchainSupport.h"
@@ -34,98 +34,50 @@
#include <cmath>
#include <tuple>
-namespace
-{
-arm_compute::TensorShape get_reshaped_weights_shape(const arm_compute::ITensorInfo *weights, bool append_bias)
-{
- const unsigned int mat_weights_cols = weights->dimension(3);
- const unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (append_bias ? 1 : 0);
- return arm_compute::TensorShape(mat_weights_cols, mat_weights_rows);
-}
-} // namespace
+using namespace arm_compute;
+using namespace arm_compute::misc::shape_calculator;
-namespace arm_compute
-{
-NEConvolutionLayerReshapeWeights::NEConvolutionLayerReshapeWeights(std::shared_ptr<IMemoryManager> memory_manager)
- : _memory_group(std::move(memory_manager)), _weights_reshape_kernel(), _weights_transposed_kernel(), _weights_reshaped(), _transpose1xW(false)
+NEConvolutionLayerReshapeWeights::NEConvolutionLayerReshapeWeights()
+ : _weights_reshape_kernel()
{
}
-void NEConvolutionLayerReshapeWeights::configure(const ITensor *weights, const ITensor *biases, ITensor *output, bool transpose1xW)
+void NEConvolutionLayerReshapeWeights::configure(const ITensor *weights, const ITensor *biases, ITensor *output)
{
// Perform validation step
ARM_COMPUTE_ERROR_ON_NULLPTR(weights, output);
ARM_COMPUTE_ERROR_THROW_ON(NEConvolutionLayerReshapeWeights::validate(weights->info(),
(biases != nullptr) ? biases->info() : nullptr,
- output->info(),
- transpose1xW));
+ output->info()));
- // Check if bias are present, if yes they will be embedded to the weights matrix
- const bool append_biases = (biases != nullptr) && !is_data_type_quantized_asymmetric(weights->info()->data_type());
- //const unsigned bias_element = (append_biases) ? 1 : 0;
+ const bool append_biases = (biases != nullptr) && !is_data_type_quantized_asymmetric(weights->info()->data_type());
const ITensor *biases_to_use = (append_biases) ? biases : nullptr;
- _transpose1xW = transpose1xW;
-
- if(transpose1xW)
- {
- // Create tensor to store the reshaped weights
- TensorInfo info_wr = weights->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(get_reshaped_weights_shape(weights->info(), append_biases));
-
- _weights_reshaped.allocator()->init(info_wr);
- _memory_group.manage(&_weights_reshaped);
-
- _weights_reshape_kernel.configure(weights, biases, &_weights_reshaped);
- _weights_transposed_kernel.configure(&_weights_reshaped, output);
-
- _weights_reshaped.allocator()->allocate();
- }
- else
- {
- _weights_reshape_kernel.configure(weights, biases_to_use, output);
- }
+ _weights_reshape_kernel.configure(weights, biases_to_use, output);
output->info()->set_quantization_info(weights->info()->quantization_info());
}
-Status NEConvolutionLayerReshapeWeights::validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, bool transpose1xW)
+Status NEConvolutionLayerReshapeWeights::validate(const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output)
{
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QS8, DataType::QASYMM8, DataType::QS16, DataType::F16, DataType::F32);
+ ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(weights);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4);
- if(!is_data_type_quantized_asymmetric(weights->data_type()))
- {
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output);
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output);
- }
- // Check if bias are present, if yes they will be embedded to the weights matrix
- const bool append_bias = (biases != nullptr);
- if(append_bias)
+ if(biases != nullptr)
{
+ const int idx_kernels = get_data_layout_dimension_index(weights->data_layout(), DataLayoutDimension::BATCHES);
ARM_COMPUTE_RETURN_ERROR_ON(is_data_type_quantized_asymmetric(weights->data_type()));
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(weights, biases);
- ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(3));
+ ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(idx_kernels));
ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
}
- // Checks performed when biases are present
- if(append_bias)
+ if((output != nullptr) && (output->total_size() != 0))
{
- 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);
- }
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output);
- if(transpose1xW)
- {
- TensorInfo weights_reshaped = weights->clone()->set_tensor_shape(get_reshaped_weights_shape(weights, append_bias));
- ARM_COMPUTE_RETURN_ON_ERROR(NEWeightsReshapeKernel::validate(weights, biases, &weights_reshaped));
- ARM_COMPUTE_RETURN_ON_ERROR(NEGEMMTranspose1xWKernel::validate(&weights_reshaped, output));
- }
- else
- {
- ARM_COMPUTE_RETURN_ON_ERROR(NEWeightsReshapeKernel::validate(weights, biases, output));
+ NEWeightsReshapeKernel::validate(weights, biases, output);
}
return Status{};
@@ -133,110 +85,21 @@
void NEConvolutionLayerReshapeWeights::run()
{
- _memory_group.acquire();
-
NEScheduler::get().schedule(&_weights_reshape_kernel, 3);
-
- if(_transpose1xW)
- {
- NEScheduler::get().schedule(&_weights_transposed_kernel, Window::DimY);
- }
-
- _memory_group.release();
}
-namespace
-{
-TensorShape get_reshaped_weights_shape_conv(const ITensorInfo *weights, bool append_bias, bool is_fully_connected_convolution)
-{
- unsigned int mat_weights_cols = weights->dimension(3);
- unsigned int mat_weights_rows = weights->dimension(0) * weights->dimension(1) * weights->dimension(2) + (append_bias ? 1 : 0);
-
- if(is_fully_connected_convolution)
- {
- // Create tensor to store the reshaped weights
- return TensorShape(mat_weights_cols, mat_weights_rows);
- }
- else
- {
- // Create tensor to store transposed weights
- const float transpose_width = 16.0f / weights->element_size();
- return TensorShape(mat_weights_rows * static_cast<unsigned int>(transpose_width), static_cast<unsigned int>(std::ceil(mat_weights_cols / transpose_width)));
- }
-}
-
-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 &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,
- unsigned int &conv_w, unsigned int &conv_h, const Size2D &dilation)
-{
- 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_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);
-
- if(biases != nullptr)
- {
- if(is_quantized)
- {
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
- }
- else
- {
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
- }
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT(input, biases);
- ARM_COMPUTE_RETURN_ERROR_ON(!weights_info.are_reshaped() && biases->dimension(0) != weights->dimension(3));
- 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(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(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(idx_width), input->dimension(idx_height), kernel_width, kernel_height,
- conv_info, dilation);
-
- // Check if its a "fully connected" convolution
- is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1));
- is_interleaved = (!is_fully_connected_convolution && !is_quantized);
- is_activationlayer_enabled = act_info.enabled();
-
- return Status{};
-}
-} // namespace
-
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(), _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)
+ : _memory_group(memory_manager), _reshape_weights(), _im2col_kernel(), _mm_gemm(), _mm_gemmlowp(memory_manager), _gemmlowp_output_stage(), _col2im_kernel(), _activationlayer_function(),
+ _add_bias_kernel(), _reshape_layer(), _original_weights(nullptr), _im2col_output(), _weights_reshaped(), _gemm_output(), _tmp_output(), _data_layout(DataLayout::NCHW), _append_bias(false),
+ _skip_im2col(false), _skip_col2im(false), _is_quantized(false), _is_activationlayer_enabled(false), _is_prepared(false)
{
}
-void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *weights, ITensor *output, bool is_interleaved, const GEMMReshapeInfo &reshape_info)
+void NEGEMMConvolutionLayer::configure_mm(const ITensor *input, const ITensor *weights, ITensor *output, int gemm_3d_depth)
{
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights);
+ ARM_COMPUTE_ERROR_THROW_ON(validate_mm(input->info(), weights->info(), output->info(), gemm_3d_depth, _skip_im2col));
+
if(_is_quantized)
{
// Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
@@ -255,128 +118,145 @@
}
else
{
- _mm_kernel.configure(input, weights, output, 1.f, is_interleaved, reshape_info);
+ // Configure matrix multiply function
+ _mm_gemm.configure(input, weights, nullptr, output, 1.0f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run*/, gemm_3d_depth,
+ _skip_im2col /* Reinterpret the input as 3D if im2col is skipped */));
}
}
-void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info,
- const Size2D &dilation, const ActivationLayerInfo &act_info)
+Status NEGEMMConvolutionLayer::validate_mm(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, int gemm_3d_depth, bool skip_im2col)
{
- // Perform validate step
- ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
+ const bool is_quantized = is_data_type_quantized_asymmetric(input->data_type());
- DataType dt{};
- unsigned int kernel_width = 0;
- unsigned int kernel_height = 0;
- unsigned int mat_weights_cols = 0;
- unsigned int mat_weights_rows = 0;
- unsigned int conv_w = 0;
- unsigned int conv_h = 0;
-
- _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,
- mat_weights_cols, mat_weights_rows, conv_w, conv_h, dilation);
-
- ARM_COMPUTE_ERROR_THROW_ON(status);
-
- _original_weights = weights;
- const unsigned int fixed_point_position = input->info()->fixed_point_position();
- const ITensor *biases_to_use = (_append_bias) ? biases : nullptr;
-
- bool run_optimised = dt == DataType::F32;
-
- // Reshape weights if needed
- if(run_optimised)
+ const GEMMInfo gemm_info = GEMMInfo(false, false, true /* Reshape weights only for the first run */, gemm_3d_depth, skip_im2col);
+ if(is_quantized)
{
- TensorShape reshaped_weights_shape{ mat_weights_cols, mat_weights_rows };
+ // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
+ // Extract and negate input and weights offset
+ const QuantizationInfo input_quantization_info = input->quantization_info();
+ const QuantizationInfo weights_quantization_info = weights->quantization_info();
- // 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;
+ std::unique_ptr<ITensorInfo> input_qa = input->clone();
+ std::unique_ptr<ITensorInfo> weights_qa = weights->clone();
+ input_qa->set_quantization_info(QuantizationInfo(input_quantization_info.scale, -input_quantization_info.offset));
+ weights_qa->set_quantization_info(QuantizationInfo(weights_quantization_info.scale, -weights_quantization_info.offset));
+
+ // Perform validation step on GEMMLowp
+ return NEGEMMLowpMatrixMultiplyCore::validate(input_qa.get(), weights_qa.get(), output, gemm_info);
}
else
{
- if(_are_weights_reshaped)
- {
- if(_is_fully_connected_convolution || _is_quantized)
- {
- mat_weights_cols = weights_info.num_kernels();
- 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(idx_channel) + (_append_bias ? 1 : 0);
- }
- }
- else
- {
- TensorShape reshaped_weights_shape;
+ // Perform validation step on Matrix multiply function
+ return NEGEMM::validate(input, weights, nullptr, output, 1.0f, 0.0f, gemm_info);
+ }
+}
- if(_is_fully_connected_convolution || _is_quantized)
- {
- reshaped_weights_shape = TensorShape{ mat_weights_cols, mat_weights_rows };
- }
- else
- {
- // Create tensor to store transposed weights
- const float transpose_width = 16.0f / input->info()->element_size();
- reshaped_weights_shape = TensorShape{ mat_weights_rows *static_cast<unsigned int>(transpose_width),
- static_cast<unsigned int>(std::ceil(mat_weights_cols / transpose_width)) };
- }
+Status NEGEMMConvolutionLayer::validate_gemm3d(DataType data_type, int gemm_3d_depth, bool skip_im2col)
+{
+ const bool is_quantized = is_data_type_quantized_asymmetric(data_type);
+ const DataType output_gemm_data_type = is_quantized ? DataType::S32 : data_type;
+ const unsigned int mult_y = skip_im2col ? 1U : gemm_3d_depth;
+ const unsigned int mult_z = skip_im2col ? gemm_3d_depth : 1U;
- // 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_to_use, &_weights_reshaped, _is_interleaved /* 1xW transpose */);
- weights = &_weights_reshaped;
+ // Set dummy tensor shapes for the validation
+ const TensorInfo dummy_input_info(TensorShape(4U, 4U * mult_y, 1U * mult_z), 1, data_type);
+ const TensorInfo dummy_weights_info(TensorShape(4U, 4U), 1, data_type);
+ const TensorInfo dummy_output_info(TensorShape(4U, 4U, gemm_3d_depth), 1, output_gemm_data_type);
+
+ return validate_mm(&dummy_input_info, &dummy_weights_info, &dummy_output_info, gemm_3d_depth, skip_im2col);
+}
+
+void NEGEMMConvolutionLayer::configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info,
+ const Size2D &dilation, const ActivationLayerInfo &act_info, unsigned int num_groups)
+{
+ ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
+ ARM_COMPUTE_UNUSED(num_groups);
+ ARM_COMPUTE_ERROR_THROW_ON(NEGEMMConvolutionLayer::validate(input->info(),
+ weights->info(),
+ biases != nullptr ? biases->info() : nullptr,
+ output->info(),
+ conv_info,
+ weights_info,
+ dilation,
+ act_info,
+ num_groups));
+
+ const DataType data_type = input->info()->data_type();
+ const 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 int idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES);
+
+ const unsigned int kernel_width = weights->info()->dimension(idx_width);
+ const unsigned int kernel_height = weights->info()->dimension(idx_height);
+
+ _is_prepared = weights_info.retain_internal_weights();
+ _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 && conv_info.stride().first == 1 && conv_info.stride().second == 1);
+ _skip_col2im = data_layout == DataLayout::NHWC;
+ _append_bias = (biases != nullptr) && (!_is_quantized);
+
+ const ITensor *gemm_input_to_use = input;
+ ITensor *gemm_output_to_use = output;
+ ITensor *gemm_output_staged_to_use = output;
+
+ // Get convolved dimensions
+ unsigned int conv_w = 0;
+ unsigned int conv_h = 0;
+ std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(idx_width),
+ input->info()->dimension(idx_height),
+ kernel_width,
+ kernel_height,
+ conv_info,
+ dilation);
+
+ // Check if GEMM3D is supported
+ if(_skip_col2im)
+ {
+ // If not supported, we need to perform im2col and col2im (or reshape layer)
+ if(!bool(validate_gemm3d(input->info()->data_type(), conv_h, _skip_im2col)))
+ {
+ _skip_im2col = false;
+ _skip_col2im = false;
}
}
- // In case we skip im2col we have to add bias
+ const unsigned bias_element = (_append_bias && !_skip_im2col) ? 1 : 0;
+ const ITensor *biases_to_use = (_append_bias && !_skip_im2col) ? biases : nullptr;
+
+ // Get parameters from conv_info
+ unsigned int stride_x = 0;
+ unsigned int stride_y = 0;
+ std::tie(stride_x, stride_y) = conv_info.stride();
+
+ unsigned int mat_weights_cols = weights->info()->dimension(idx_kernels);
+ unsigned int mat_weights_rows = weights->info()->dimension(idx_width) * weights->info()->dimension(idx_height) * weights->info()->dimension(idx_channel) + bias_element;
+
+ // _weights_reshaped will be auto configured in the kernel.
+ // Just append biases and do not transpose 1xW as it will be reshaped in NEGEMM
+ _reshape_weights.configure(weights, biases_to_use, &_weights_reshaped);
+
+ // Create tensor to store im2col reshaped inputs
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);
+ // Calculate im2col shape
+ // For NEON the batch size is on the fourth dimension
+ TensorShape shape_im2col = input->info()->tensor_shape();
+ shape_im2col.set(0, mat_weights_rows);
+ shape_im2col.set(1, conv_w * conv_h);
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);
- }
+ _im2col_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_im2col));
+ _memory_group.manage(&_im2col_output);
- // 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);
+ // Configure
+ _im2col_kernel.configure(input, &_im2col_output, Size2D(kernel_width, kernel_height), conv_info, _append_bias, dilation);
- // Configure im2col
- _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _append_bias, false, false, dilation);
+ // Update GEMM input
+ gemm_input_to_use = &_im2col_output;
}
else if(_append_bias)
{
@@ -384,129 +264,187 @@
_add_bias_kernel.configure(output, biases, output, ConvertPolicy::SATURATE);
}
- // Configure matrix multiply
- if(run_optimised)
+ // Create temporary GEMM output tensor in case we cannot skip col2im
+ if(!_skip_col2im)
{
- 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.");
- }
- }
- else
- {
- if(_is_interleaved)
- {
- // Configure GEMMInterleave4x4. _input_interleaved_reshaped will be auto configured in the kernel
- _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped);
+ // Calculate GEMM output shape
+ TensorShape shape_gemm = _im2col_output.info()->tensor_shape();
+ shape_gemm.set(0, mat_weights_cols);
+ shape_gemm.set(1, conv_w * conv_h);
- // Configure GEMM
- 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
- {
- configure_mm(&_input_im2col_reshaped, weights, &_gemm_output, _is_interleaved);
- }
+ // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input.
+ const DataType gemm_data_type = _is_quantized ? DataType::S32 : data_type;
+ TensorInfo info_gemm(shape_gemm, 1, gemm_data_type);
+ info_gemm.set_quantization_info(output->info()->quantization_info());
+ _gemm_output.allocator()->init(info_gemm);
+ _memory_group.manage(&_gemm_output);
+
+ // Update GEMM output
+ gemm_output_to_use = &_gemm_output;
}
+ // Configure GEMM
+ configure_mm(gemm_input_to_use, &_weights_reshaped, gemm_output_to_use, _skip_col2im ? conv_h : 1);
+
if(!_skip_im2col)
{
- _input_im2col_reshaped.allocator()->allocate();
+ _im2col_output.allocator()->allocate();
+ }
- // Configure output stage for quantized case
- if(_is_quantized)
+ // 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();
+
+ 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);
+ gemm_output_staged_to_use = &_tmp_output;
+
+ _gemmlowp_output_stage.configure(gemm_output_to_use, biases, gemm_output_staged_to_use, output_multiplier, output_shift, output_quant_info.offset);
+ }
+
+ if(!_skip_col2im)
+ {
+ if(_data_layout == DataLayout::NCHW)
{
- const QuantizationInfo output_quant_info = (output->info()->total_size() == 0) ? input->info()->quantization_info() : output->info()->quantization_info();
-
- 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
+ _col2im_kernel.configure(_is_quantized ? gemm_output_staged_to_use : gemm_output_to_use, output, Size2D(conv_w, conv_h));
}
-
- // Configure Col2Im
- if(!is_nhwc)
+ else
{
- _output_col2im_kernel.configure(_is_quantized ? &_tmp_output : &_gemm_output, output, Size2D(conv_w, conv_h));
+ // Configure reshape layer
+ _reshape_layer.configure(_is_quantized ? gemm_output_staged_to_use : gemm_output_to_use, output);
}
+ }
- if(_is_quantized)
- {
- _tmp_output.allocator()->allocate();
- }
+ if(_is_quantized)
+ {
+ _tmp_output.allocator()->allocate();
+ }
+
+ if(!_skip_col2im)
+ {
_gemm_output.allocator()->allocate();
}
- 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)
- {
- _weights_reshaped.allocator()->allocate();
- }
+ 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");
//Configure Activation Layer
+ _is_activationlayer_enabled = act_info.enabled();
+
if(_is_activationlayer_enabled)
{
_activationlayer_function.configure(output, nullptr, act_info);
}
+
+ ARM_COMPUTE_UNUSED(weights_info);
}
Status NEGEMMConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
- const WeightsInfo &weights_info, const Size2D &dilation, const ActivationLayerInfo &act_info)
+ const WeightsInfo &weights_info, const Size2D &dilation, const ActivationLayerInfo &act_info, unsigned int num_groups)
{
- ARM_COMPUTE_UNUSED(output);
-
- DataType dt{};
- bool append_bias{};
- bool skip_im2col{};
- bool are_weights_reshaped{};
- bool is_fully_connected_convolution{};
- bool is_interleaved{};
- bool is_quantized{};
- bool is_activationlayer_enabled{};
- unsigned int kernel_width = 0;
- unsigned int kernel_height = 0;
- unsigned int mat_weights_cols = 0;
- unsigned int mat_weights_rows = 0;
- unsigned int conv_w = 0;
- unsigned int conv_h = 0;
+ ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, weights, output);
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights_info.are_reshaped(), "Weights already reshaped are not supported!");
+ 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);
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, weights);
+ ARM_COMPUTE_RETURN_ERROR_ON_MSG(num_groups > 1, "Grouping (num_groups != 1) is not supported on NEON");
const DataLayout data_layout = input->data_layout();
- const bool is_nhwc = data_layout == DataLayout::NHWC;
+ const DataType data_type = input->data_type();
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 int idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES);
- 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);
+ const unsigned int kernel_width = weights->dimension(idx_width);
+ const unsigned int kernel_height = weights->dimension(idx_height);
- const Size2D kernel_weights = Size2D(kernel_width, kernel_height);
+ TensorInfo im2col_reshaped_info, info_gemm, tmp_info, weights_reshaped_info;
+ const ITensorInfo *gemm_input_to_use = input;
+ const ITensorInfo *gemm_output_to_use = output;
+ const ITensorInfo *gemm_output_staged_to_use = output;
+ const ITensorInfo *weights_to_use = weights;
- ARM_COMPUTE_RETURN_ON_ERROR(status);
+ const bool is_quantized = is_data_type_quantized_asymmetric(data_type);
+ const bool append_bias = (biases != nullptr) && (!is_quantized);
+ bool skip_im2col = (data_layout == DataLayout::NHWC && kernel_width == 1 && kernel_height == 1 && conv_info.stride().first == 1 && conv_info.stride().second == 1);
+ bool skip_col2im = data_layout == DataLayout::NHWC;
- std::unique_ptr<ITensorInfo> reshaped_weights = weights->clone();
- bool optimised_kernel = false;
+ // Get convolved dimensions
+ unsigned int conv_w = 0;
+ unsigned int conv_h = 0;
- if(dt == DataType::F32)
+ 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 GEMM3D is supported
+ if(skip_col2im)
{
- optimised_kernel = true;
+ // If not supported, we need to perform im2col and col2im (or reshape layer)
+ if(!bool(validate_gemm3d(input->data_type(), conv_h, skip_im2col)))
+ {
+ skip_im2col = false;
+ skip_col2im = false;
+ }
}
- 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();
- shape_im2col.set(0, mat_input_cols);
- shape_im2col.set(1, mat_input_rows);
- shape_im2col.set(2, 1);
- TensorInfo im2_col_info = input->clone()->set_tensor_shape(shape_im2col);
+ const unsigned bias_element = (append_bias && !skip_im2col) ? 1 : 0;
+ const ITensorInfo *biases_to_use = (append_bias && !skip_im2col) ? biases : nullptr;
+
+ ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(idx_channel) != input->dimension(idx_channel));
+ ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4);
+
+ // Validate biases
+ if(biases != nullptr)
+ {
+ if(is_quantized)
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
+ }
+ else
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases);
+ }
+ ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(idx_kernels));
+ ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
+ }
+
+ if(act_info.enabled())
+ {
+ ARM_COMPUTE_ERROR_ON(act_info.b() > act_info.a());
+ }
+
+ unsigned int mat_weights_cols = weights->dimension(idx_kernels);
+ unsigned int mat_weights_rows = weights->dimension(idx_width) * weights->dimension(idx_height) * weights->dimension(idx_channel) + bias_element;
+
+ // Output tensor auto inizialization if not yet initialized
+ ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases_to_use, nullptr));
+ weights_reshaped_info = TensorInfo(compute_weights_reshaped_shape(*weights, (append_bias && !skip_im2col)), 1, data_type);
+ weights_to_use = &weights_reshaped_info;
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));
+ // Create tensor info for im2col reshaped inputs
+ // For NEON the batch size is on the fourth dimension
+ TensorShape shape_im2col = input->tensor_shape();
+ shape_im2col.set(0, mat_weights_rows);
+ shape_im2col.set(1, conv_w * conv_h);
+ shape_im2col.set(2, 1);
+
+ im2col_reshaped_info = TensorInfo(shape_im2col, 1, data_type);
+ im2col_reshaped_info.set_quantization_info(input->quantization_info());
+
+ ARM_COMPUTE_RETURN_ON_ERROR(NEIm2ColKernel::validate(input, &im2col_reshaped_info, Size2D(kernel_width, kernel_height), conv_info, append_bias, dilation));
+ gemm_input_to_use = &im2col_reshaped_info;
}
else if(append_bias)
{
@@ -514,66 +452,45 @@
ARM_COMPUTE_RETURN_ON_ERROR(NEArithmeticAdditionKernel::validate(output, biases, output, ConvertPolicy::SATURATE));
}
- // Create GEMM output tensor
- TensorShape shape_gemm(im2_col_info.tensor_shape());
- shape_gemm.set(0, mat_weights_cols);
- shape_gemm.set(1, mat_input_rows);
- TensorInfo gemm_output_info = input->clone()->set_tensor_shape(shape_gemm);
-
- // Reshape weights if needed
- if(optimised_kernel)
+ // Create temporary GEMM output tensor in case we cannot skip col2im
+ if(!skip_col2im)
{
- ARM_COMPUTE_RETURN_ERROR_ON(are_weights_reshaped);
+ TensorShape shape_gemm = gemm_input_to_use->tensor_shape();
+ shape_gemm.set(0, mat_weights_cols);
+ shape_gemm.set(1, conv_w * conv_h);
+ const DataType gemm_data_type = is_quantized ? DataType::S32 : data_type;
+ // GEMM output should be S32 for acquiring raw integer accumulator without quantized postprocessing for quantized asymmetric input.
+ info_gemm = TensorInfo(shape_gemm, 1, gemm_data_type);
+ info_gemm.set_quantization_info(output->quantization_info());
- // Create tensor to store the reshaped weights
- reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, append_bias, is_fully_connected_convolution));
- ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */));
- }
- else if(!is_quantized)
- {
- TensorShape reshaped_weights_shape;
-
- if(is_fully_connected_convolution || is_quantized)
- {
- reshaped_weights_shape = TensorShape{ mat_weights_cols, mat_weights_rows };
- }
- else
- {
- // Create tensor to store transposed weights
- const float transpose_width = 16.0f / input->element_size();
- reshaped_weights_shape = TensorShape{ mat_weights_rows *static_cast<unsigned int>(transpose_width),
- static_cast<unsigned int>(std::ceil(mat_weights_cols / transpose_width)) };
- }
-
- // Create tensor to store the reshaped weights
- reshaped_weights->set_tensor_shape(get_reshaped_weights_shape_conv(weights, append_bias, is_fully_connected_convolution));
- ARM_COMPUTE_RETURN_ON_ERROR(NEConvolutionLayerReshapeWeights::validate(weights, biases, reshaped_weights.get(), !is_fully_connected_convolution /* 1xW transpose */));
- weights = reshaped_weights.get();
-
- // Validate GEMM interleave and multiply
- if(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.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
- weights->tensor_shape()[0], // n
- shape_im2col[0]) /* k */));
- }
- else
- {
- 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)));
+ gemm_output_to_use = &info_gemm;
}
- 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");
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(gemm_input_to_use, weights_to_use, gemm_output_to_use, skip_col2im ? conv_h : 1, skip_im2col));
+ if(is_quantized)
+ {
+ float multiplier = input->quantization_info().scale * weights_to_use->quantization_info().scale / output->quantization_info().scale;
+ int output_multiplier, output_shift;
+ quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+
+ tmp_info = TensorInfo(gemm_output_to_use->tensor_shape(), 1, DataType::QASYMM8);
+ tmp_info.set_quantization_info(output->quantization_info());
+ gemm_output_staged_to_use = &tmp_info;
+
+ // Validate output stage for quantized case
+ NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(gemm_output_to_use, biases, gemm_output_staged_to_use, output->quantization_info().offset);
+ }
+
+ // Validate Col2Im/ReshapeLayer
+ if(!skip_col2im && (data_layout == DataLayout::NCHW))
+ {
+ ARM_COMPUTE_RETURN_ON_ERROR(NECol2ImKernel::validate(is_quantized ? gemm_output_staged_to_use : gemm_output_to_use,
+ output,
+ Size2D(conv_w, conv_h)));
+ }
+
+ //Validate Activation Layer
if(act_info.enabled())
{
ARM_COMPUTE_RETURN_ON_ERROR(NEActivationLayer::validate(output, nullptr, act_info));
@@ -584,54 +501,30 @@
void NEGEMMConvolutionLayer::run()
{
- // Run weights reshaping (Runs once for every configure)
- if(!_are_weights_reshaped)
- {
- ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
-
- _are_weights_reshaped = true;
- _reshape_weights.run();
-
- // Mark original weights tensor as unused
- _original_weights->mark_as_unused();
- }
+ prepare();
_memory_group.acquire();
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);
+ unsigned int y_dim = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT);
+ NEScheduler::get().schedule(&_im2col_kernel, y_dim);
}
- // Runs matrix multiply on reshaped matrices
- if(_asm_glue._optimised_kernel != nullptr)
+ // Runs NEGEMM or NEGEMMLowpMatrixMultiplyCore functions
+ if(_is_quantized)
{
- _asm_glue.run();
- // Release weights in case buffer is pretransposed
- if(!_weights_reshaped.is_used())
- {
- _weights_reshaped.allocator()->free();
- }
+ // Run gemmlowp
+ _mm_gemmlowp.run();
+
+ // Run output stage
+ _gemmlowp_output_stage.run();
}
else
{
- if(_is_interleaved)
- {
- // Run interleave
- NEScheduler::get().schedule(&_input_interleave_kernel, Window::DimY);
- }
-
- // Runs matrix multiply on reshaped matrices
- if(_is_quantized)
- {
- _mm_gemmlowp.run();
- }
- else
- {
- NEScheduler::get().schedule(&_mm_kernel, Window::DimY);
- }
+ // Run gemm
+ _mm_gemm.run();
}
if(_skip_im2col && _append_bias)
@@ -639,16 +532,17 @@
NEScheduler::get().schedule(&_add_bias_kernel, Window::DimY);
}
- // Run output stage for quantized case
- if(_is_quantized)
- {
- _gemmlowp_output_stage.run();
- }
-
// Reshape output matrix
- if(_data_layout == DataLayout::NCHW)
+ if(!_skip_col2im)
{
- NEScheduler::get().schedule(&_output_col2im_kernel, Window::DimY);
+ if(_data_layout == DataLayout::NCHW)
+ {
+ NEScheduler::get().schedule(&_col2im_kernel, Window::DimY);
+ }
+ else
+ {
+ _reshape_layer.run();
+ }
}
if(_is_activationlayer_enabled)
@@ -658,4 +552,25 @@
_memory_group.release();
}
-} // namespace arm_compute
+
+void NEGEMMConvolutionLayer::prepare()
+{
+ if(!_is_prepared)
+ {
+ ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
+
+ // Run weights reshaping and mark original weights tensor as unused
+ _weights_reshaped.allocator()->allocate();
+ _reshape_weights.run();
+ _original_weights->mark_as_unused();
+
+ // Prepare GEMM
+ _is_quantized ? _mm_gemmlowp.prepare() : _mm_gemm.prepare();
+ if(!_weights_reshaped.is_used())
+ {
+ _weights_reshaped.allocator()->free();
+ }
+
+ _is_prepared = true;
+ }
+}