CLDepthwiseConvolutionLayer rework - Part 1
Remove the reshaped variant for CLDepthwiseConvolutionLayer 3x3 NHWC Quantized
- Remove kernel selection by GPUTarget
- Remove unused quantized support from the NHWC kernel
- Remove CLDepthwiseConvolutionLayerReshapeWeightsKernel
- Remove OpenCL kernels for reshaped dwc 3x3 quantized and weights reshape
- Remove the "_bifrost" suffix in common OpenCL kernel
- Remove the ICLDepthwiseConvolutionLayer3x3Kernel common interface
Resolve COMPMID-3864, COMPMID-3907
Change-Id: Icfac0fb6c00e214985beb05dad7c0cdbbee7d830
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5447
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
index 287a965..dda70d2 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.cpp
@@ -114,20 +114,19 @@
}
std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *output, const PadStrideInfo &conv_info,
- unsigned int depth_multiplier, GPUTarget gpu_target, std::string &kernel_name, const Size2D dilation)
+ unsigned int depth_multiplier, std::string &kernel_name, const Size2D dilation)
{
// Output auto inizialitation if not yet initialized
const ConvolutionInfo info
{
conv_info, depth_multiplier, ActivationLayerInfo(), dilation
};
- const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, info);
+ const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, info);
auto_init_if_empty(*output, input->clone()->set_tensor_shape(output_shape).set_quantization_info(output->quantization_info()));
const unsigned int conv_stride_x = conv_info.stride().first;
const unsigned int conv_stride_y = conv_info.stride().second;
const bool is_qasymm = is_data_type_quantized_asymmetric(input->data_type());
- const bool is_bifrost = get_arch_from_target(gpu_target) == GPUTarget::BIFROST;
// Configure kernel window
unsigned int num_elems_read_per_iteration_x = 0;
@@ -156,31 +155,28 @@
num_elems_read_per_iteration_x = 3 + (num_elems_written_per_iteration_x - 1) * conv_stride_x;
break;
}
- if(is_bifrost)
+ if(conv_stride_x == 1 && conv_stride_y == 1)
{
- if(conv_stride_x == 1 && conv_stride_y == 1)
- {
- kernel_name = "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f16";
- num_elems_read_per_iteration_x = 8;
- num_elems_written_per_iteration_x = 4;
- num_elems_read_per_iteration_y = 6;
- num_elems_written_per_iteration_y = 4;
- }
- else if(conv_stride_x == 2 && conv_stride_y == 2)
- {
- kernel_name = "depthwise_convolution_3x3_stridex2_stridey2_bifrost_f16";
- num_elems_read_per_iteration_x = 10;
- num_elems_written_per_iteration_x = 4;
- num_elems_read_per_iteration_y = 5;
- num_elems_written_per_iteration_y = 2;
- }
+ kernel_name = "depthwise_convolution_3x3_stridex1_stridey1_f16";
+ num_elems_read_per_iteration_x = 8;
+ num_elems_written_per_iteration_x = 4;
+ num_elems_read_per_iteration_y = 6;
+ num_elems_written_per_iteration_y = 4;
+ }
+ else if(conv_stride_x == 2 && conv_stride_y == 2)
+ {
+ kernel_name = "depthwise_convolution_3x3_stridex2_stridey2_f16";
+ num_elems_read_per_iteration_x = 10;
+ num_elems_written_per_iteration_x = 4;
+ num_elems_read_per_iteration_y = 5;
+ num_elems_written_per_iteration_y = 2;
}
}
- else if(input->data_type() == DataType::F32 && is_bifrost)
+ else if(input->data_type() == DataType::F32)
{
if(conv_stride_x == 1 && conv_stride_y == 1)
{
- kernel_name = "depthwise_convolution_3x3_stridex1_stridey1_bifrost_f32";
+ kernel_name = "depthwise_convolution_3x3_stridex1_stridey1_f32";
num_elems_read_per_iteration_x = 4;
num_elems_read_per_iteration_y = 6;
num_elems_written_per_iteration_x = 2;
@@ -188,7 +184,7 @@
}
else if(conv_stride_x == 2 && conv_stride_y == 2)
{
- kernel_name = "depthwise_convolution_3x3_stridex2_stridey2_bifrost_f32";
+ kernel_name = "depthwise_convolution_3x3_stridex2_stridey2_f32";
num_elems_read_per_iteration_x = 6;
num_elems_read_per_iteration_y = 5;
num_elems_written_per_iteration_x = 2;
@@ -239,7 +235,7 @@
} // namespace
CLDepthwiseConvolutionLayer3x3NCHWKernel::CLDepthwiseConvolutionLayer3x3NCHWKernel()
- : _conv_stride_x(0), _conv_pad_top(0), _conv_pad_left(0)
+ : _border_size(0), _input(), _output(), _weights(), _biases(), _conv_stride_y(1), _output_multipliers(), _output_shifts(), _is_quantized(false), _conv_stride_x(0), _conv_pad_top(0), _conv_pad_left(0)
{
}
@@ -278,10 +274,9 @@
_is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
// Configure kernel window
- std::string kernel_name;
- const GPUTarget gpu_target = get_target();
+ std::string kernel_name;
- auto win_config = validate_and_configure_window(input->info(), weights->info(), output->info(), conv_info, depth_multiplier, gpu_target, kernel_name, dilation);
+ auto win_config = validate_and_configure_window(input->info(), weights->info(), output->info(), conv_info, depth_multiplier, kernel_name, dilation);
ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
ICLKernel::configure_internal(win_config.second);
@@ -372,13 +367,13 @@
}
Status CLDepthwiseConvolutionLayer3x3NCHWKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
- const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, GPUTarget gpu_target,
+ const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info,
const Size2D &dilation, const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
{
std::string kernel_name;
ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation, output_multipliers, output_shifts));
ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(), output->clone().get(),
- conv_info, depth_multiplier, gpu_target, kernel_name, dilation)
+ conv_info, depth_multiplier, kernel_name, dilation)
.first);
return Status{};
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h
index 45b5869..c4e475f 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NCHWKernel.h
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -24,7 +24,7 @@
#ifndef ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNCHWKERNEL3x3_H
#define ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNCHWKERNEL3x3_H
-#include "src/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
+#include "src/core/CL/ICLKernel.h"
namespace arm_compute
{
@@ -32,11 +32,19 @@
/** Interface for the kernel to run a 3x3 depthwise convolution on a tensor when the data layout is NCHW.
*/
-class CLDepthwiseConvolutionLayer3x3NCHWKernel : public ICLDepthwiseConvolutionLayer3x3Kernel
+class CLDepthwiseConvolutionLayer3x3NCHWKernel : public ICLKernel
{
public:
/** Default constructor */
CLDepthwiseConvolutionLayer3x3NCHWKernel();
+ /** Prevent instances of this class from being copied (As this class contains pointers) */
+ CLDepthwiseConvolutionLayer3x3NCHWKernel(const CLDepthwiseConvolutionLayer3x3NCHWKernel &) = delete;
+ /** Prevent instances of this class from being copied (As this class contains pointers) */
+ CLDepthwiseConvolutionLayer3x3NCHWKernel &operator=(const CLDepthwiseConvolutionLayer3x3NCHWKernel &) = delete;
+ /** Default Move Constructor. */
+ CLDepthwiseConvolutionLayer3x3NCHWKernel(CLDepthwiseConvolutionLayer3x3NCHWKernel &&) = default;
+ /** Default move assignment operator */
+ CLDepthwiseConvolutionLayer3x3NCHWKernel &operator=(CLDepthwiseConvolutionLayer3x3NCHWKernel &&) = default;
/** Initialize the function's source, destination, conv and border_size.
*
* @param[in] input Source tensor. DataType supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
@@ -56,7 +64,7 @@
*/
void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U),
- const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr) override;
+ const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr);
/** Initialize the function's source, destination, conv and border_size.
*
* @param[in] compile_context The compile context to be used.
@@ -77,7 +85,7 @@
*/
void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U),
- const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr) override;
+ const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr);
/** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3NCHWKernel
*
* @param[in] input Source tensor info. DataType supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
@@ -89,7 +97,6 @@
* @param[in] conv_info Padding and stride information to use for the convolution.
* @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
* @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported.
- * @param[in] gpu_target (Optional) GPU target to validate the kernel for. Defaults to midgard.
* @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
* @param[in] output_multipliers (Optional) Output multipliers tensor info for quantized computations. In case of per-channel quantization,
* the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
@@ -99,13 +106,23 @@
* @return a status
*/
static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
- unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), GPUTarget gpu_target = GPUTarget::MIDGARD,
+ unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(),
const Size2D &dilation = Size2D(1U, 1U), const ITensorInfo *output_multipliers = nullptr, const ITensorInfo *output_shifts = nullptr);
void run(const Window &window, cl::CommandQueue &queue) override;
BorderSize border_size() const override;
private:
+ BorderSize _border_size;
+ const ICLTensor *_input;
+ ICLTensor *_output;
+ const ICLTensor *_weights;
+ const ICLTensor *_biases;
+ unsigned int _conv_stride_y;
+ const ICLTensor *_output_multipliers;
+ const ICLTensor *_output_shifts;
+ bool _is_quantized;
+
unsigned int _conv_stride_x;
unsigned int _conv_pad_top;
unsigned int _conv_pad_left;
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
index f7603e6..2a1365e 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
@@ -30,7 +30,6 @@
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Utils.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
-#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
#include "src/core/AccessWindowStatic.h"
#include "src/core/CL/CLValidate.h"
#include "src/core/CL/ICLKernel.h"
@@ -43,17 +42,11 @@
namespace
{
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
- const PadStrideInfo &conv_info, unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation,
- const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
+ const PadStrideInfo &conv_info, unsigned int depth_multiplier, const ActivationLayerInfo &act_info, const Size2D &dilation)
{
+ ARM_COMPUTE_UNUSED(act_info);
ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32, DataType::QASYMM8, DataType::QASYMM8_SIGNED);
- ARM_COMPUTE_RETURN_ERROR_ON_MSG((act_info.enabled()) && (input->data_type() == DataType::QASYMM8 || input->data_type() == DataType::QASYMM8_SIGNED)
- && (act_info.activation() != ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU)
- && (act_info.activation() != ActivationLayerInfo::ActivationFunction::BOUNDED_RELU)
- && (act_info.activation() != ActivationLayerInfo::ActivationFunction::RELU)
- && (act_info.activation() != ActivationLayerInfo::ActivationFunction::LOGISTIC),
- "For QASYMM8 only logistic, relu, lower bounded relu and lower-upper bounded relu are supported");
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32);
ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier > 1); // COMPMID-1071 Add depth multiplier support for NHWC
ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1);
@@ -61,54 +54,21 @@
ARM_COMPUTE_RETURN_ERROR_ON((dilation.x() < 1) || (dilation.y() < 1));
- const bool is_qasymm = is_data_type_quantized_asymmetric(input->data_type());
const size_t weights_width = 3;
const size_t weights_height = 3;
const ConvolutionInfo info{ conv_info, depth_multiplier, ActivationLayerInfo(), dilation };
- const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(
- *input, TensorInfo(TensorShape(weights_width, weights_height), 1, weights->data_type()).set_data_layout(DataLayout::NCHW), info);
- if(is_qasymm)
- {
- DepthwiseConvolutionReshapeInfo info;
- info.c0 = 4;
- ARM_COMPUTE_RETURN_ERROR_ON((weights->dimension(0) / info.c0) != weights_width * weights_height);
- ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output_multipliers, output_shifts);
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_multipliers, 1, DataType::S32);
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output_shifts, 1, DataType::S32);
- ARM_COMPUTE_RETURN_ERROR_ON(output_multipliers->num_dimensions() > 1);
- ARM_COMPUTE_RETURN_ERROR_ON(output_shifts->num_dimensions() > 1);
+ const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(
+ *input, TensorInfo(TensorShape(weights_width, weights_height), 1, weights->data_type()).set_data_layout(DataLayout::NCHW), info);
- if(is_data_type_quantized_per_channel(weights->data_type()))
- {
- ARM_COMPUTE_RETURN_ERROR_ON(output_shape[0] != output_multipliers->dimension(0));
- ARM_COMPUTE_RETURN_ERROR_ON(output_shape[0] != output_shifts->dimension(0));
- }
- else
- {
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
- ARM_COMPUTE_RETURN_ERROR_ON(1 != output_multipliers->dimension(0));
- ARM_COMPUTE_RETURN_ERROR_ON(1 != output_shifts->dimension(0));
- }
- }
- else
- {
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
- ARM_COMPUTE_RETURN_ERROR_ON((weights->dimension(1) != weights_width) || (weights->dimension(2) != weights_height));
- }
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
+ ARM_COMPUTE_RETURN_ERROR_ON((weights->dimension(1) != weights_width) || (weights->dimension(2) != weights_height));
if(biases != nullptr)
{
ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != output_shape[0]);
- if(is_qasymm)
- {
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::S32);
- }
- else
- {
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
- }
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
}
@@ -122,10 +82,9 @@
}
std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *output,
- const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation,
- ITensorInfo *output_multipliers, ITensorInfo *output_shifts)
+ const PadStrideInfo &conv_info, unsigned int depth_multiplier, const Size2D &dilation)
{
- ARM_COMPUTE_UNUSED(weights);
+ ARM_COMPUTE_UNUSED(weights, bias);
ARM_COMPUTE_UNUSED(depth_multiplier);
const bool is_stride_1_dilation_1 = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1) && dilation.x() == 1 && dilation.y() == 1);
@@ -134,115 +93,46 @@
Window win{};
Status err{};
- if(is_data_type_quantized_asymmetric(input->data_type()))
- {
- const unsigned int num_elems_accessed_per_iteration = 4;
- const unsigned int num_rows_read_per_iteration = num_rows_processed_per_iteration + 2;
- const unsigned int num_rows_written_per_iteration = std::ceil(num_rows_processed_per_iteration / static_cast<float>(conv_info.stride().first));
-
- BorderSize border_size;
- border_size = BorderSize(conv_info.pad_left(), 0, std::max(std::max(conv_info.pad_right(), conv_info.pad_bottom()), conv_info.pad_top()), 0);
-
- // Configure kernel window
- win = calculate_max_window(*output, Steps(num_elems_accessed_per_iteration, num_rows_written_per_iteration));
-
- AccessWindowStatic input_access(input, 0, -border_size.top, ceil_to_multiple(input->dimension(0), num_elems_accessed_per_iteration),
- ceil_to_multiple(input->dimension(1) + border_size.bottom, num_rows_read_per_iteration));
- AccessWindowRectangle output_access(output, 0, 0, num_elems_accessed_per_iteration, num_rows_written_per_iteration);
-
- bool window_changed = false;
-
- if((output_multipliers != nullptr) && (output_shifts != nullptr))
- {
- AccessWindowHorizontal output_multipliers_access(output_multipliers, 0, num_elems_accessed_per_iteration);
- AccessWindowHorizontal output_shifts_access(output_shifts, 0, num_elems_accessed_per_iteration);
- window_changed = window_changed || update_window_and_padding(win, input_access, output_access, output_multipliers_access, output_shifts_access);
- }
- else
- {
- Status err = ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "output_multipliers and output_shifts must be non-nullptr for quantized input");
- return std::make_pair(err, win);
- }
-
- if(bias != nullptr)
- {
- AccessWindowHorizontal bias_access(bias, 0, num_elems_accessed_per_iteration);
- window_changed = window_changed || update_window_and_padding(win, bias_access);
- }
-
- err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
- }
- else
- {
- unsigned int num_elems_accessed_per_iteration = adjust_vec_size(4 / input->element_size(), input->dimension(0));
- win = calculate_max_window(*output, Steps(num_elems_accessed_per_iteration, num_rows_processed_per_iteration));
- }
+ unsigned int num_elems_accessed_per_iteration = adjust_vec_size(4 / input->element_size(), input->dimension(0));
+ win = calculate_max_window(*output, Steps(num_elems_accessed_per_iteration, num_rows_processed_per_iteration));
return std::make_pair(err, win);
}
} // namespace
CLDepthwiseConvolutionLayer3x3NHWCKernel::CLDepthwiseConvolutionLayer3x3NHWCKernel()
- : _num_planes_processed_per_iteration(1)
+ : _input(), _output(), _weights(), _biases(), _num_planes_processed_per_iteration(1)
{
}
-BorderSize CLDepthwiseConvolutionLayer3x3NHWCKernel::border_size() const
-{
- return _border_size;
-}
-
void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
- const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation,
- const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
+ const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
{
- configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation, output_multipliers, output_shifts);
+ configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation);
}
void CLDepthwiseConvolutionLayer3x3NHWCKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
- const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation,
- const ICLTensor *output_multipliers, const ICLTensor *output_shifts)
+ const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(),
- conv_info, depth_multiplier, act_info, dilation,
- (output_multipliers != nullptr) ? output_multipliers->info() : nullptr,
- (output_shifts != nullptr) ? output_shifts->info() : nullptr));
+ conv_info, depth_multiplier, act_info, dilation));
auto padding_info = get_padding_info({ input, weights, biases, output });
auto win_config = validate_and_configure_window(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(),
- conv_info, depth_multiplier, dilation,
- (output_multipliers != nullptr) ? output_multipliers->info() : nullptr,
- (output_shifts != nullptr) ? output_shifts->info() : nullptr);
+ conv_info, depth_multiplier, dilation);
- const bool is_stride_1 = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
- const bool is_stride_1_dilation_1 = (is_stride_1 && dilation.x() == 1 && dilation.y() == 1);
- const bool is_quantized_per_channel = is_data_type_quantized_per_channel(weights->info()->data_type());
- const bool is_dot8_supported = dot8_supported(CLKernelLibrary::get().get_device()) && !is_quantized_per_channel;
+ const bool is_stride_1 = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
+ const bool is_stride_1_dilation_1 = (is_stride_1 && dilation.x() == 1 && dilation.y() == 1);
_input = input;
_output = output;
_weights = weights;
_biases = biases;
- _conv_stride_y = conv_info.stride().second;
_num_planes_processed_per_iteration = is_stride_1_dilation_1 ? 2 : 1;
- _output_multipliers = output_multipliers;
- _output_shifts = output_shifts;
- _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
- if(_is_quantized)
- {
- _border_size = BorderSize(input->info()->padding());
-
- // If QASYMM8 and the 8 bit dot product is available, force _num_planes_processed_per_iteration to 1
- if(is_dot8_supported)
- {
- _num_planes_processed_per_iteration = 1;
- }
- }
-
- unsigned int num_elems_accessed_per_iteration = _is_quantized ? 4 : adjust_vec_size(4 / input->info()->element_size(), input->info()->dimension(0));
+ unsigned int num_elems_accessed_per_iteration = adjust_vec_size(4 / input->info()->element_size(), input->info()->dimension(0));
unsigned int num_rows_processed_per_iteration = is_stride_1_dilation_1 ? 2 : 1;
CLBuildOptions build_opts;
@@ -257,54 +147,8 @@
build_opts.add_option_if(_biases != nullptr, "-DHAS_BIAS");
build_opts.add_option_if(_input->info()->tensor_shape().total_size_upper(3) > 1,
"-DDST_DEPTH=" + support::cpp11::to_string(static_cast<int>(std::ceil(_output->info()->dimension(2) / static_cast<float>(_num_planes_processed_per_iteration)))));
-
- if(_is_quantized)
- {
- const UniformQuantizationInfo iq_info = _input->info()->quantization_info().uniform();
- const UniformQuantizationInfo wq_info = _weights->info()->quantization_info().uniform();
- const UniformQuantizationInfo oq_info = _output->info()->quantization_info().uniform();
-
- build_opts.add_option("-DSRC_DIM_1=" + support::cpp11::to_string(_input->info()->dimension(1)));
- build_opts.add_option("-DINPUT_OFFSET=" + support::cpp11::to_string(-iq_info.offset));
- build_opts.add_option("-DWEIGHTS_OFFSET=" + support::cpp11::to_string(-wq_info.offset));
- build_opts.add_option("-DOUTPUT_OFFSET=" + support::cpp11::to_string(oq_info.offset));
- build_opts.add_option("-DK_OFFSET=" + support::cpp11::to_string(9 * iq_info.offset * wq_info.offset));
- build_opts.add_option_if(is_quantized_per_channel, "-DPER_CHANNEL_QUANTIZATION");
- build_opts.add_option_if(is_dot8_supported, "-DIS_DOT8");
-
- // Compute non-per-channel multiplier and shift anyway to make OpenCL kernel simpler
- float multiplier = iq_info.scale * wq_info.scale / oq_info.scale;
- int output_multiplier = 0;
- int output_shift = 0;
- quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift);
- build_opts.add_option("-DOUTPUT_MULTIPLIER=" + support::cpp11::to_string(output_multiplier));
- build_opts.add_option("-DOUTPUT_SHIFT=" + support::cpp11::to_string(output_shift));
-
- if(act_info.enabled())
- {
- int a_val{};
- int b_val{};
- std::tie(b_val, a_val) = get_quantized_activation_min_max(act_info, input->info()->data_type(), oq_info);
-
- const int o1 = oq_info.offset;
-
- build_opts.add_option("-DA_VAL=" + support::cpp11::to_string(a_val));
- build_opts.add_option("-DB_VAL=" + support::cpp11::to_string(b_val));
- build_opts.add_option("-DCONST_0=" + support::cpp11::to_string(o1));
-
- const float s1 = iq_info.scale;
- build_opts.add_option("-DS1_VAL=" + float_to_string_with_full_precision(s1));
- build_opts.add_option("-DO1_VAL=" + support::cpp11::to_string(o1));
- }
-
- build_opts.add_option("-DWEIGHTS_TYPE=" + get_cl_type_from_data_type(weights->info()->data_type()));
- build_opts.add_option("-DWEIGHTS_PROMOTED_TYPE=" + get_cl_promoted_type_from_data_type(weights->info()->data_type()));
- }
- else
- {
- build_opts.add_option_if(act_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(act_info.a()));
- build_opts.add_option_if(act_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(act_info.b()));
- }
+ build_opts.add_option_if(act_info.enabled(), "-DA_VAL=" + float_to_string_with_full_precision(act_info.a()));
+ build_opts.add_option_if(act_info.enabled(), "-DB_VAL=" + float_to_string_with_full_precision(act_info.b()));
if(is_stride_1_dilation_1)
{
@@ -317,30 +161,20 @@
else
{
build_opts.add_option("-DCONV_STRIDE_X=" + support::cpp11::to_string(conv_info.stride().first));
- build_opts.add_option("-DCONV_STRIDE_Y=" + support::cpp11::to_string(_conv_stride_y));
+ build_opts.add_option("-DCONV_STRIDE_Y=" + support::cpp11::to_string(conv_info.stride().second));
build_opts.add_option("-DDILATION_X=" + support::cpp11::to_string(dilation.x()));
build_opts.add_option("-DDILATION_Y=" + support::cpp11::to_string(dilation.y()));
}
- std::string kernel_name;
// Create kernel
- if(_is_quantized)
- {
- kernel_name = std::string("dwc_3x3_reshaped_quantized8");
- kernel_name += (is_dot8_supported && is_stride_1_dilation_1 ? "_dot8" : "");
- kernel_name += (is_stride_1_dilation_1 ? "_stride1" : "");
- kernel_name += "_nhwc";
- }
- else
- {
- kernel_name = std::string("depthwise_convolution_3x3_nhwc");
- kernel_name += (is_stride_1_dilation_1 ? "_stride1" : "");
- }
+ std::string kernel_name;
+ kernel_name = std::string("depthwise_convolution_3x3_nhwc");
+ kernel_name += (is_stride_1_dilation_1 ? "_stride1" : "");
ICLKernel::configure_internal(win_config.second);
_kernel = create_kernel(compile_context, kernel_name, build_opts.options());
- ARM_COMPUTE_ERROR_ON(!_is_quantized && has_padding_changed(padding_info));
+ ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
// Set config_id for enabling LWS tuning
_config_id = kernel_name;
@@ -359,15 +193,12 @@
}
Status CLDepthwiseConvolutionLayer3x3NHWCKernel::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output,
- const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation,
- const ITensorInfo *output_multipliers, const ITensorInfo *output_shifts)
+ const PadStrideInfo &conv_info, unsigned int depth_multiplier, ActivationLayerInfo act_info, const Size2D &dilation)
{
- ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation, output_multipliers, output_shifts));
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, act_info, dilation));
ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), weights->clone().get(),
biases != nullptr ? biases->clone().get() : nullptr,
- output->clone().get(), conv_info, depth_multiplier, dilation,
- (output_multipliers != nullptr) ? output_multipliers->clone().get() : nullptr,
- (output_shifts != nullptr) ? output_shifts->clone().get() : nullptr)
+ output->clone().get(), conv_info, depth_multiplier, dilation)
.first);
return Status{};
}
@@ -382,16 +213,7 @@
Window win = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
win.set(Window::DimZ, Window::Dimension(0, std::ceil(_output->info()->dimension(2) / static_cast<float>(_num_planes_processed_per_iteration)) * total_batches, 1));
- unsigned int idx = 2 * num_arguments_per_4D_tensor() + (_is_quantized ? num_arguments_per_2D_tensor() : num_arguments_per_3D_tensor());
-
- if(_is_quantized)
- {
- Window slice;
- slice.use_tensor_dimensions(_output_multipliers->info()->tensor_shape());
- slice.set_dimension_step(Window::DimX, window.x().step());
- add_1D_tensor_argument(idx, _output_multipliers, slice);
- add_1D_tensor_argument(idx, _output_shifts, slice);
- }
+ unsigned int idx = 2 * num_arguments_per_4D_tensor() + num_arguments_per_3D_tensor();
if(_biases != nullptr)
{
@@ -401,62 +223,14 @@
add_1D_tensor_argument(idx, _biases, win_biases);
}
- if(_is_quantized)
- {
- // Calculate the max_offset.
- // max_offset is the offset for the last NOT valid value in the Z dimension (spatial dimension Y for NHWC)
- // |******************|
- // | pad_top |
- // |******************|
- // | |
- // | plane0 |
- // | batch0 |
- // |__________________|
- // |******************| Batch 0
- // | pad_bottom |
- // | pad_top |
- // |******************|
- // | |
- // | plane1 |
- // | batch0 |
- // |__________________|-----> max_offset
- // |******************|
- // | pad_bottom |
- // | pad_top |
- // |******************|
- // | |
- // | plane0 |
- // | batch1 |
- // |__________________|
- // |******************| Batch 1
- // | pad_bottom |
- // | pad_top |
- // |******************|
- // | |
- // | plane1 |
- // | batch1 |
- // |__________________|
- // | pad_bottom |
- // |******************|
- const int max_offset = ((_input->info()->dimension(1) * _input->info()->dimension(2)) + (_input->info()->padding().bottom + _input->info()->padding().top) * (_input->info()->dimension(
- 2) - 1)) * _input->info()->strides_in_bytes().y();
- _kernel.setArg(idx, max_offset);
- }
-
Window slice = win.first_slice_window_4D();
do
{
unsigned int idx = 0;
add_4D_tensor_argument(idx, _input, slice);
add_4D_tensor_argument(idx, _output, slice);
- if(_is_quantized)
- {
- add_2D_tensor_argument(idx, _weights, slice);
- }
- else
- {
- add_3D_tensor_argument(idx, _weights, slice);
- }
+ add_3D_tensor_argument(idx, _weights, slice);
+
enqueue(queue, *this, slice, lws_hint());
}
while(win.slide_window_slice_4D(slice));
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h
index ce0bf5c..ee47d98 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.h
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2018-2020 Arm Limited.
+ * Copyright (c) 2018-2021 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -24,7 +24,7 @@
#ifndef ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNHWCKERNEL3x3_H
#define ARM_COMPUTE_CLDEPTHWISECONVOLUTIONNHWCKERNEL3x3_H
-#include "src/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h"
+#include "src/core/CL/ICLKernel.h"
namespace arm_compute
{
@@ -32,81 +32,78 @@
/** Interface for the kernel to run a 3x3 depthwise convolution on a tensor when the data layout is NHWC.
*/
-class CLDepthwiseConvolutionLayer3x3NHWCKernel : public ICLDepthwiseConvolutionLayer3x3Kernel
+class CLDepthwiseConvolutionLayer3x3NHWCKernel : public ICLKernel
{
public:
/** Default constructor */
CLDepthwiseConvolutionLayer3x3NHWCKernel();
+ /** Prevent instances of this class from being copied (As this class contains pointers) */
+ CLDepthwiseConvolutionLayer3x3NHWCKernel(const CLDepthwiseConvolutionLayer3x3NHWCKernel &) = delete;
+ /** Prevent instances of this class from being copied (As this class contains pointers) */
+ CLDepthwiseConvolutionLayer3x3NHWCKernel &operator=(const CLDepthwiseConvolutionLayer3x3NHWCKernel &) = delete;
+ /** Default Move Constructor. */
+ CLDepthwiseConvolutionLayer3x3NHWCKernel(CLDepthwiseConvolutionLayer3x3NHWCKernel &&) = default;
+ /** Default move assignment operator */
+ CLDepthwiseConvolutionLayer3x3NHWCKernel &operator=(CLDepthwiseConvolutionLayer3x3NHWCKernel &&) = default;
/** Default move assignment operator. */
/** Initialize the function's source, destination, conv and border_size.
*
- * @param[in] input Source tensor. DataType supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
- * @param[in] weights Weights tensor. A 3D tensor with dimensions [IFM, 3, 3].
- * Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
- * @param[in] biases Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
- * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
- * @param[out] output Destination tensor. Data type supported: Same as @p input.
- * @param[in] conv_info Padding and stride information to use for the convolution.
- * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
- * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported.
- * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
- * @param[in] output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization,
- * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
- * @param[in] output_shifts (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
- * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
+ * @param[in] input Source tensor. DataType supported: F16/F32.
+ * @param[in] weights Weights tensor. A 3D tensor with dimensions [IFM, 3, 3].
+ * Data type supported: Same as @p input.
+ * @param[in] biases Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
+ * Data type supported: Same as @p input.
+ * @param[out] output Destination tensor. Data type supported: Same as @p input.
+ * @param[in] conv_info Padding and stride information to use for the convolution.
+ * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+ * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported.
+ * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
*/
void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
- unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U),
- const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr) override;
+ unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
/** Initialize the function's source, destination, conv and border_size.
*
- * @param[in] compile_context The compile context to be used.
- * @param[in] input Source tensor. DataType supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
- * @param[in] weights Weights tensor. A 3D tensor with dimensions [IFM, 3, 3].
- * Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
- * @param[in] biases Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
- * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
- * @param[out] output Destination tensor. Data type supported: Same as @p input.
- * @param[in] conv_info Padding and stride information to use for the convolution.
- * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
- * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported.
- * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
- * @param[in] output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization,
- * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
- * @param[in] output_shifts (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
- * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
+ * @param[in] compile_context The compile context to be used.
+ * @param[in] input Source tensor. DataType supported: F16/F32.
+ * @param[in] weights Weights tensor. A 3D tensor with dimensions [IFM, 3, 3].
+ * Data type supported: Same as @p input.
+ * @param[in] biases Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
+ * Data type supported: Same as @p input.
+ * @param[out] output Destination tensor. Data type supported: Same as @p input.
+ * @param[in] conv_info Padding and stride information to use for the convolution.
+ * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+ * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported.
+ * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
*/
void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
- unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U),
- const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr) override;
+ unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
/** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3NHWCKernel
*
- * @param[in] input Source tensor info. DataType supported: QASYMM8/QASYMM8_SIGNED/F16/F32.
- * @param[in] weights Weights tensor info. A 3D tensor with dimensions [IFM, 3, 3].
- * Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8/QASYMM8_SIGNED.
- * @param[in] biases Biases tensor info. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
- * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED.
- * @param[in] output Destination tensor info. Data type supported: Same as @p input.
- * @param[in] conv_info Padding and stride information to use for the convolution.
- * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
- * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported.
- * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
- * @param[in] output_multipliers (Optional) Output multipliers tensor info for quantized computations. In case of per-channel quantization,
- * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
- * @param[in] output_shifts (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
- * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
+ * @param[in] input Source tensor info. DataType supported: F16/F32.
+ * @param[in] weights Weights tensor info. A 3D tensor with dimensions [IFM, 3, 3].
+ * Data type supported: Same as @p input.
+ * @param[in] biases Biases tensor info. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
+ * Data type supported: Same as @p input.
+ * @param[in] output Destination tensor info. Data type supported: Same as @p input.
+ * @param[in] conv_info Padding and stride information to use for the convolution.
+ * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
+ * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported.
+ * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
*
* @return a status
*/
static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info,
- unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U),
- const ITensorInfo *output_multipliers = nullptr, const ITensorInfo *output_shifts = nullptr);
+ unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U));
// Inherited methods overridden:
void run(const Window &window, cl::CommandQueue &queue) override;
- BorderSize border_size() const override;
private:
+ const ICLTensor *_input;
+ ICLTensor *_output;
+ const ICLTensor *_weights;
+ const ICLTensor *_biases;
+
unsigned int _num_planes_processed_per_iteration;
};
} // namespace arm_compute
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.cpp
deleted file mode 100644
index 386d634..0000000
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.cpp
+++ /dev/null
@@ -1,131 +0,0 @@
-/*
- * Copyright (c) 2019-2021 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#include "src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.h"
-
-#include "arm_compute/core/CL/CLHelpers.h"
-#include "arm_compute/core/CL/CLKernelLibrary.h"
-#include "arm_compute/core/CL/ICLTensor.h"
-#include "arm_compute/core/Helpers.h"
-#include "arm_compute/core/TensorInfo.h"
-#include "arm_compute/core/Utils.h"
-#include "arm_compute/core/utils/misc/ShapeCalculator.h"
-#include "src/core/AccessWindowStatic.h"
-#include "src/core/CL/CLValidate.h"
-#include "src/core/CL/ICLKernel.h"
-#include "src/core/helpers/AutoConfiguration.h"
-#include "src/core/helpers/WindowHelpers.h"
-#include "support/StringSupport.h"
-
-namespace arm_compute
-{
-namespace
-{
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const DepthwiseConvolutionReshapeInfo &info)
-{
- ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
- const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
- const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
-
- ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
- ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC);
- ARM_COMPUTE_RETURN_ERROR_ON(info.c0 != 4);
- ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_h) != 3);
- ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(idx_w) != 3);
- ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
-
- if(output->total_size() != 0)
- {
- auto reshaped_weights_shape = arm_compute::misc::shape_calculator::compute_reshaped_depthwise_weights_shape(*input, info);
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), reshaped_weights_shape);
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
- }
-
- return Status{};
-}
-
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const DepthwiseConvolutionReshapeInfo &info)
-{
- auto reshaped_input_shape = arm_compute::misc::shape_calculator::compute_reshaped_depthwise_weights_shape(*input, info);
- auto_init_if_empty(*output, reshaped_input_shape, 1, input->data_type(), input->quantization_info());
-
- Window win = calculate_max_window(*input, Steps(info.c0));
- AccessWindowHorizontal weights_access(input, 0, info.c0);
- const bool window_changed = update_window_and_padding(win, weights_access);
-
- Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
- return std::make_pair(err, win);
-}
-} // namespace
-
-CLDepthwiseConvolutionLayerReshapeWeightsKernel::CLDepthwiseConvolutionLayerReshapeWeightsKernel()
- : _input(nullptr), _output(nullptr)
-{
-}
-
-void CLDepthwiseConvolutionLayerReshapeWeightsKernel::configure(const ICLTensor *input, ICLTensor *output, const DepthwiseConvolutionReshapeInfo &info)
-{
- configure(CLKernelLibrary::get().get_compile_context(), input, output, info);
-}
-
-void CLDepthwiseConvolutionLayerReshapeWeightsKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const DepthwiseConvolutionReshapeInfo &info)
-{
- ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
- ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), info));
- auto win_config = validate_and_configure_window(input->info(), output->info(), info);
- ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
-
- ICLKernel::configure_internal(win_config.second);
-
- _input = input;
- _output = output;
-
- // Build the kernel
- CLBuildOptions build_opts;
- build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(info.c0));
- build_opts.add_option("-DDST_WIDTH=" + support::cpp11::to_string(_output->info()->dimension(0)));
- build_opts.add_option_if(info.transpose, "-DTRANSPOSE");
- build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size()));
-
- _kernel = create_kernel(compile_context, "depthwise_convolution_reshape_weights", build_opts.options());
-}
-
-Status CLDepthwiseConvolutionLayerReshapeWeightsKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const DepthwiseConvolutionReshapeInfo &info)
-{
- ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, info));
- ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), info).first);
- return Status{};
-}
-
-void CLDepthwiseConvolutionLayerReshapeWeightsKernel::run(const Window &window, cl::CommandQueue &queue)
-{
- ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
- ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window);
-
- unsigned int idx = 0;
- add_3D_tensor_argument(idx, _input, window);
- add_2D_tensor_argument(idx, _output, window);
- enqueue(queue, *this, window, lws_hint());
-}
-} // namespace arm_compute
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.h b/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.h
deleted file mode 100644
index 650fe9a..0000000
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.h
+++ /dev/null
@@ -1,85 +0,0 @@
-/*
- * Copyright (c) 2019-2020 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifndef ARM_COMPUTE_CLDEPTHWISECONVOLUTIONLAYERRESHAPEWEIGHTSKERNEL_H
-#define ARM_COMPUTE_CLDEPTHWISECONVOLUTIONLAYERRESHAPEWEIGHTSKERNEL_H
-
-#include "src/core/CL/ICLKernel.h"
-
-namespace arm_compute
-{
-class ICLTensor;
-
-/** Interface for the kernel to reshape the weights of depthwise convolution. */
-class CLDepthwiseConvolutionLayerReshapeWeightsKernel : public ICLKernel
-{
-public:
- /** Default constructor */
- CLDepthwiseConvolutionLayerReshapeWeightsKernel();
- /** Prevent instances of this class from being copied (As this class contains pointers) */
- CLDepthwiseConvolutionLayerReshapeWeightsKernel(const CLDepthwiseConvolutionLayerReshapeWeightsKernel &) = delete;
- /** Prevent instances of this class from being copied (As this class contains pointers) */
- CLDepthwiseConvolutionLayerReshapeWeightsKernel &operator=(const CLDepthwiseConvolutionLayerReshapeWeightsKernel &) = delete;
- /** Default Move Constructor. */
- CLDepthwiseConvolutionLayerReshapeWeightsKernel(CLDepthwiseConvolutionLayerReshapeWeightsKernel &&) = default;
- /** Default move assignment operator */
- CLDepthwiseConvolutionLayerReshapeWeightsKernel &operator=(CLDepthwiseConvolutionLayerReshapeWeightsKernel &&) = default;
-
- /** Initialize the function's source and destination.
- *
- * @param[in] input The input tensor of dimension [IFM, W, H]. Data types supported: All. Data layouts supported: NHWC
- * @param[out] output The output tensor of dimension [W*H*C0, ceil(IFM/C0)]. C0 is the number of channels read by each thread. Data types supported: same as @p weights.
- * @param[in] info Depthwise convolution information to reshape the input tensor.
- */
- void configure(const ICLTensor *input, ICLTensor *output, const DepthwiseConvolutionReshapeInfo &info);
- /** Initialize the function's source and destination.
- *
- * @param[in] compile_context The compile context to be used.
- * @param[in] input The input tensor of dimension [IFM, W, H]. Data types supported: All. Data layouts supported: NHWC
- * @param[out] output The output tensor of dimension [W*H*C0, ceil(IFM/C0)]. C0 is the number of channels read by each thread. Data types supported: same as @p weights.
- * @param[in] info Depthwise convolution information to reshape the input tensor.
- */
- void configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const DepthwiseConvolutionReshapeInfo &info);
-
- /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer3x3NHWCKernel
- *
- * @param[in] input The input tensor info of dimension [IFM, W, H]. Data types supported: All. Data layouts supported: NHWC
- * @param[in] output The output tensor info of dimension [W*H*C0, ceil(IFM/C0)]. C0 is the number of channels read by each thread. Data types supported: same as @p weights.
- * @param[in] info Depthwise convolution information to reshape the input tensor.
- *
- * @return a Status
- */
- static Status validate(const ITensorInfo *input, const ITensorInfo *output, const DepthwiseConvolutionReshapeInfo &info);
-
- // Inherited methods overridden:
- void run(const Window &window, cl::CommandQueue &queue) override;
-
-private:
- const ICLTensor *_input;
- ICLTensor *_output;
-
- void configure_dot_product(const DepthwiseConvolutionReshapeInfo &info);
- void configure_generic(const DepthwiseConvolutionReshapeInfo &info);
-};
-} // namespace arm_compute
-#endif /* ARM_COMPUTE_CLDEPTHWISECONVOLUTIONLAYERRESHAPEWEIGHTSKERNEL_H */
diff --git a/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp b/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp
index d9f293b..c688951 100644
--- a/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp
+++ b/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp
@@ -36,8 +36,6 @@
#include "support/StringSupport.h"
-#include "utils/TypePrinter.h"
-
namespace arm_compute
{
namespace
diff --git a/src/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h b/src/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h
deleted file mode 100644
index 4c92ae4..0000000
--- a/src/core/CL/kernels/ICLDepthwiseConvolutionLayer3x3Kernel.h
+++ /dev/null
@@ -1,105 +0,0 @@
-/*
- * Copyright (c) 2017-2020 Arm Limited.
- *
- * SPDX-License-Identifier: MIT
- *
- * Permission is hereby granted, free of charge, to any person obtaining a copy
- * of this software and associated documentation files (the "Software"), to
- * deal in the Software without restriction, including without limitation the
- * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
- * sell copies of the Software, and to permit persons to whom the Software is
- * furnished to do so, subject to the following conditions:
- *
- * The above copyright notice and this permission notice shall be included in all
- * copies or substantial portions of the Software.
- *
- * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
- * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
- * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
- * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
- * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
- * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
- * SOFTWARE.
- */
-#ifndef ARM_COMPUTE_ICLDEPTHWISECONVOLUTIONKERNEL3x3_H
-#define ARM_COMPUTE_ICLDEPTHWISECONVOLUTIONKERNEL3x3_H
-
-#include "src/core/CL/ICLKernel.h"
-
-namespace arm_compute
-{
-class ICLTensor;
-
-/** Interface for the kernel to run a 3x3 depthwise convolution on a tensor.
- */
-class ICLDepthwiseConvolutionLayer3x3Kernel : public ICLKernel
-{
-public:
- /** Default constructor */
- ICLDepthwiseConvolutionLayer3x3Kernel()
- : _border_size(0), _input(), _output(), _weights(), _biases(), _conv_stride_y(1), _output_multipliers(), _output_shifts(), _is_quantized(false)
- {
- }
- /** Prevent instances of this class from being copied (As this class contains pointers) */
- ICLDepthwiseConvolutionLayer3x3Kernel(const ICLDepthwiseConvolutionLayer3x3Kernel &) = delete;
- /** Prevent instances of this class from being copied (As this class contains pointers) */
- ICLDepthwiseConvolutionLayer3x3Kernel &operator=(const ICLDepthwiseConvolutionLayer3x3Kernel &) = delete;
- /** Default Move Constructor. */
- ICLDepthwiseConvolutionLayer3x3Kernel(ICLDepthwiseConvolutionLayer3x3Kernel &&) = default;
- /** Default move assignment operator */
- ICLDepthwiseConvolutionLayer3x3Kernel &operator=(ICLDepthwiseConvolutionLayer3x3Kernel &&) = default;
- /** Initialize the function's source, destination, conv and border_size.
- *
- * @param[in] input Source tensor. DataType supported: QASYMM8/F16/F32.
- * @param[in] weights Weights tensor. A 3D tensor with dimensions [3, 3, IFM].
- * Data type supported: Same as @p input, QASYMM8/QSYMM8_PER_CHANNEL when input is QASYMM8.
- * @param[in] biases Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
- * Data type supported: Same as @p input, S32 when input is QASYMM8.
- * @param[out] output Destination tensor. Data type supported: Same as @p input.
- * @param[in] conv_info Padding and stride information to use for the convolution.
- * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
- * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported for QASYMM8.
- * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
- * @param[in] output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization,
- * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
- * @param[in] output_shifts (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
- * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
- */
- virtual void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
- unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U),
- const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr) = 0;
- /** Initialize the function's source, destination, conv and border_size.
- *
- * @param[in] compile_context The compile context to be used.
- * @param[in] input Source tensor. DataType supported: QASYMM8/F16/F32.
- * @param[in] weights Weights tensor. A 3D tensor with dimensions [3, 3, IFM].
- * Data type supported: Same as @p input, QASYMM8/QSYMM8_PER_CHANNEL when input is QASYMM8.
- * @param[in] biases Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed.
- * Data type supported: Same as @p input, S32 when input is QASYMM8.
- * @param[out] output Destination tensor. Data type supported: Same as @p input.
- * @param[in] conv_info Padding and stride information to use for the convolution.
- * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1.
- * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU are supported for QASYMM8.
- * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1).
- * @param[in] output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization,
- * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
- * @param[in] output_shifts (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization,
- * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32
- */
- virtual void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info,
- unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U),
- const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr) = 0;
-
-protected:
- BorderSize _border_size;
- const ICLTensor *_input;
- ICLTensor *_output;
- const ICLTensor *_weights;
- const ICLTensor *_biases;
- unsigned int _conv_stride_y;
- const ICLTensor *_output_multipliers;
- const ICLTensor *_output_shifts;
- bool _is_quantized;
-};
-} // namespace arm_compute
-#endif /*ARM_COMPUTE_ICLDEPTHWISECONVOLUTIONKERNEL3x3_H */