COMPMID-1691: Optimize CLDepthwiseConvolutionKernel (QASYMM8/NHWC) for 3x3 kernels (stride=1 and stride=2)
Change-Id: I7d0d2dc350feeb40d253d17f9ffd5051a8fb42ef
Reviewed-on: https://review.mlplatform.org/511
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
index 594d0b6..5e5a35c 100644
--- a/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayer3x3NHWCKernel.cpp
@@ -37,9 +37,8 @@
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
-using namespace arm_compute;
-using namespace arm_compute::misc::shape_calculator;
-
+namespace arm_compute
+{
namespace
{
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier,
@@ -54,11 +53,24 @@
"For QASYMM8 only logistic, relu, lower bounded relu and lower-upper bounded relu are supported"); //COMPMID-1317 add fused activation for F32
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
ARM_COMPUTE_RETURN_ERROR_ON(depth_multiplier > 1); // COMPMID-1071 Add depth multiplier support for NHWC
- ARM_COMPUTE_RETURN_ERROR_ON(weights->dimension(1) != 3 || weights->dimension(2) != 3);
- ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1 || conv_info.stride().first > 2);
+
+ ARM_COMPUTE_RETURN_ERROR_ON(conv_info.stride().first < 1);
ARM_COMPUTE_RETURN_ERROR_ON(std::max(conv_info.pad_top(), conv_info.pad_bottom()) > 1);
- const bool is_qasymm = is_data_type_quantized_asymmetric(input->data_type());
+ const bool is_qasymm = is_data_type_quantized_asymmetric(input->data_type());
+ const size_t weights_width = 3;
+ const size_t weights_height = 3;
+
+ if(is_qasymm)
+ {
+ DepthwiseConvolutionReshapeInfo info;
+ info.c0 = 4;
+ ARM_COMPUTE_RETURN_ERROR_ON((weights->dimension(0) / info.c0) != weights_width * weights_height);
+ }
+ else
+ {
+ ARM_COMPUTE_RETURN_ERROR_ON((weights->dimension(1) != weights_width) || (weights->dimension(2) != weights_height));
+ }
if(biases != nullptr)
{
@@ -68,15 +80,16 @@
}
else
{
+ ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(0));
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases);
}
- ARM_COMPUTE_RETURN_ERROR_ON(biases->dimension(0) != weights->dimension(0));
+
ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
}
if(output->total_size() != 0)
{
- const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, conv_info, depth_multiplier);
+ const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*input, weights_width, weights_height, conv_info, depth_multiplier);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), output_shape);
}
@@ -84,10 +97,13 @@
}
std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *output,
- const PadStrideInfo &conv_info)
+ const PadStrideInfo &conv_info, unsigned int depth_multiplier)
{
+ const size_t weights_width = 3;
+ const size_t weights_height = 3;
+
// Get convolved dimensions
- const TensorShape output_shape = compute_depthwise_convolution_shape(*input, *weights, conv_info, 1 /* depth_multiplier */);
+ const TensorShape output_shape = arm_compute::misc::shape_calculator::compute_depthwise_convolution_shape(*input, weights_width, weights_height, conv_info, depth_multiplier);
// Output auto inizialitation if not yet initialized
auto_init_if_empty(*output,
@@ -113,9 +129,18 @@
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);
- AccessWindowStatic weights_access(weights, 0, 0, ceil_to_multiple(weights->dimension(0), num_elems_accessed_per_iteration), weights->dimension(1));
- bool window_changed = update_window_and_padding(win, input_access, weights_access, output_access);
+ bool window_changed = false;
+
+ if(is_qasymm)
+ {
+ window_changed = update_window_and_padding(win, input_access, output_access);
+ }
+ else
+ {
+ AccessWindowStatic weights_access(weights, 0, 0, ceil_to_multiple(weights->dimension(0), num_elems_accessed_per_iteration), weights->dimension(1));
+ window_changed = update_window_and_padding(win, input_access, weights_access, output_access);
+ }
if(bias != nullptr)
{
@@ -144,18 +169,9 @@
unsigned int depth_multiplier, ActivationLayerInfo act_info)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, weights, output);
-
- // Get convolved dimensions
- const TensorShape output_shape = compute_depthwise_convolution_shape(*input->info(), *weights->info(), conv_info, depth_multiplier);
-
- // Output auto inizialitation if not yet initialized
- auto_init_if_empty(*output->info(),
- output_shape,
- 1,
- input->info()->data_type(),
- input->info()->quantization_info());
-
ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), weights->info(), (biases != nullptr) ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier, act_info));
+ auto win_config = validate_and_configure_window(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), conv_info, depth_multiplier);
+ ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
const bool is_qasymm = is_data_type_quantized_asymmetric(input->info()->data_type());
const bool is_stride_1 = ((conv_info.stride().first == conv_info.stride().second) && (conv_info.stride().first == 1));
@@ -250,13 +266,8 @@
// Create kernel
std::string kernel_name = std::string("depthwise_convolution_3x3") + (is_qasymm ? std::string("_quantized") + ((is_dot8_supported
&& is_stride_1) ? "_dot8" : "") : "") + "_nhwc" + (is_stride_1 ? "_stride1" : "");
-
- _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
-
- // Configure kernel window
- auto win_config = validate_and_configure_window(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), conv_info);
- ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
ICLKernel::configure_internal(win_config.second);
+ _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name, build_opts.options()));
// Set config_id for enabling LWS tuning
_config_id = kernel_name;
@@ -281,7 +292,7 @@
ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, weights, biases, output, conv_info, depth_multiplier, act_info));
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)
+ output->clone().get(), conv_info, depth_multiplier)
.first);
return Status{};
@@ -295,6 +306,7 @@
// Collapse window
Window window_collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ);
const size_t total_batches = _input->info()->tensor_shape().total_size_upper(3);
+ const bool is_qasymm = is_data_type_quantized_asymmetric(_input->info()->data_type());
Window win = window_collapsed;
win.set(Window::DimZ, Window::Dimension(0, std::ceil(_output->info()->dimension(2) / static_cast<float>(_num_planes_processed_per_iteration)) * total_batches, 1));
@@ -309,7 +321,7 @@
Window slice_in = win_in.first_slice_window_4D();
Window slice_out = win.first_slice_window_4D();
- unsigned int idx = 2 * num_arguments_per_4D_tensor() + num_arguments_per_3D_tensor();
+ unsigned int idx = 2 * num_arguments_per_4D_tensor() + (is_qasymm ? num_arguments_per_2D_tensor() : num_arguments_per_3D_tensor());
if(_biases != nullptr)
{
@@ -328,9 +340,16 @@
unsigned int idx = 0;
add_4D_tensor_argument(idx, _input, slice_in);
add_4D_tensor_argument(idx, _output, slice_out);
- add_3D_tensor_argument(idx, _weights, slice_out);
-
+ if(is_qasymm)
+ {
+ add_2D_tensor_argument(idx, _weights, slice_out);
+ }
+ else
+ {
+ add_3D_tensor_argument(idx, _weights, slice_out);
+ }
enqueue(queue, *this, slice_out, lws_hint());
}
while(win.slide_window_slice_4D(slice_out) && win_in.slide_window_slice_4D(slice_in));
}
+} // namespace arm_compute
diff --git a/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.cpp
similarity index 86%
rename from src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp
rename to src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.cpp
index 683dda8..4432ce5 100644
--- a/src/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.cpp
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.cpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
*
* SPDX-License-Identifier: MIT
*
@@ -21,7 +21,7 @@
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
-#include "arm_compute/core/CL/kernels/CLDepthwiseWeightsReshapeKernel.h"
+#include "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel.h"
#include "arm_compute/core/CL/CLHelpers.h"
#include "arm_compute/core/CL/CLKernelLibrary.h"
@@ -61,12 +61,12 @@
}
} // namespace
-CLDepthwiseWeightsReshapeKernel::CLDepthwiseWeightsReshapeKernel()
+CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel::CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel()
: _input(nullptr), _biases(nullptr), _output(nullptr)
{
}
-void CLDepthwiseWeightsReshapeKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *biases)
+void CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *biases)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), (biases != nullptr) ? biases->info() : nullptr));
@@ -88,23 +88,23 @@
build_opts.emplace("-DHAS_BIAS");
}
- _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_weights_reshape", build_opts));
+ _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("depthwise_convolution_reshape_weights_generic", build_opts));
// Configure kernel window
Window win = calculate_max_window(*input->info(), Steps());
- // The CLDepthwiseWeightsReshapeKernel doesn't need padding so update_window_and_padding() can be skipped
+ // The CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel doesn't need padding so update_window_and_padding() can be skipped
output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
ICLKernel::configure_internal(win);
}
-Status CLDepthwiseWeightsReshapeKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *biases)
+Status CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *biases)
{
ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, biases));
return Status{};
}
-void CLDepthwiseWeightsReshapeKernel::run(const Window &window, cl::CommandQueue &queue)
+void CLDepthwiseConvolutionLayerReshapeWeightsGenericKernel::run(const Window &window, cl::CommandQueue &queue)
{
ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window);
diff --git a/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.cpp b/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.cpp
new file mode 100644
index 0000000..608181d
--- /dev/null
+++ b/src/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.cpp
@@ -0,0 +1,125 @@
+/*
+ * Copyright (c) 2019 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 "arm_compute/core/CL/kernels/CLDepthwiseConvolutionLayerReshapeWeightsKernel.h"
+
+#include "arm_compute/core/AccessWindowStatic.h"
+#include "arm_compute/core/CL/CLHelpers.h"
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/CL/CLValidate.h"
+#include "arm_compute/core/CL/ICLKernel.h"
+#include "arm_compute/core/CL/ICLTensor.h"
+#include "arm_compute/core/Error.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/Utils.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "arm_compute/core/utils/quantization/AsymmHelpers.h"
+
+namespace arm_compute
+{
+namespace
+{
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const DepthwiseConvolutionReshapeInfo &info)
+{
+ const size_t idx_w = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH);
+ const size_t idx_h = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT);
+
+ ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
+ ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32);
+ 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);
+
+ 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);
+ }
+
+ 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);
+
+ output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
+ 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)
+{
+ 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("-DDATA_TYPE=" + get_cl_type_from_data_type(_input->info()->data_type()));
+ 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");
+
+ _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("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);
+}
+} // namespace arm_compute