COMPMID-1470 Add auto-init of the output in NECol2im
The output of NECol2Im is already auto-initialized.
This patch is about calling ShapeCalculator instead of computing the shape inside the kernel, adding validate_and_configure_window, and standardize the way convolved dims are passed (now NEON uses Size2D, while CL passes a pair of uint values: using Size2D for both implementations)
Change-Id: I795696e1b6532f57847c3186c1b532c09f5a25da
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/145345
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele DiGiorgio <michele.digiorgio@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
diff --git a/arm_compute/core/CL/kernels/CLCol2ImKernel.h b/arm_compute/core/CL/kernels/CLCol2ImKernel.h
index 2a18ae0..948b412 100644
--- a/arm_compute/core/CL/kernels/CLCol2ImKernel.h
+++ b/arm_compute/core/CL/kernels/CLCol2ImKernel.h
@@ -71,7 +71,7 @@
* @param[in] convolved_dims Output convolved dimensions.
* @param[in] num_groups (Optional) Number of groups when performing a grouped convolution
*/
- void configure(const ICLTensor *input, ICLTensor *output, std::pair<unsigned int, unsigned int> convolved_dims, unsigned int num_groups = 1);
+ void configure(const ICLTensor *input, ICLTensor *output, const Size2D &convolved_dims, unsigned int num_groups = 1);
/** Static function to check if given info will lead to a valid configuration of @ref CLCol2ImKernel
*
* @param[in] input The input tensor to convert. Data types supported: QASYMM8/F16/F32
@@ -82,7 +82,7 @@
*
* @return a status
*/
- static Status validate(const ITensorInfo *input, const ITensorInfo *output, std::pair<unsigned int, unsigned int> convolved_dims, unsigned int num_groups = 1);
+ static Status validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &convolved_dims, unsigned int num_groups = 1);
// Inherited methods overridden:
void run(const Window &window, cl::CommandQueue &queue) override;
@@ -90,7 +90,7 @@
public:
const ICLTensor *_input;
ICLTensor *_output;
- std::pair<unsigned int, unsigned int> _convolved_dims;
+ Size2D _convolved_dims;
};
} // namespace arm_compute
#endif /*__ARM_COMPUTE_CLCOL2IMKERNEL_H__ */
diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h
index c40e711..09f558d 100644
--- a/arm_compute/core/utils/misc/ShapeCalculator.h
+++ b/arm_compute/core/utils/misc/ShapeCalculator.h
@@ -141,18 +141,18 @@
return shape_vector_sum_row;
}
-inline TensorShape compute_col2im_shape(const ITensorInfo &input, std::pair<unsigned int, unsigned int> convolved_dims, unsigned int num_groups = 1)
+inline TensorShape compute_col2im_shape(const ITensorInfo &input, const Size2D &convolved_dims, bool batch_size_on_z, unsigned int num_groups = 1)
{
ARM_COMPUTE_ERROR_ON(num_groups == 0);
- ARM_COMPUTE_ERROR_ON(input.tensor_shape()[1] != (convolved_dims.first * convolved_dims.second));
+ ARM_COMPUTE_ERROR_ON(input.tensor_shape()[1] != (convolved_dims.area()));
ARM_COMPUTE_ERROR_ON((num_groups > 1) && input.tensor_shape()[2] != num_groups);
TensorShape col2im_shape{ input.tensor_shape() };
- col2im_shape.set(0, convolved_dims.first);
- col2im_shape.set(1, convolved_dims.second);
+ col2im_shape.set(0, convolved_dims.width);
+ col2im_shape.set(1, convolved_dims.height);
col2im_shape.set(2, input.tensor_shape()[0] * num_groups);
- const unsigned int batch_idx = (num_groups == 1) ? 2 : 3;
+ const unsigned int batch_idx = (batch_size_on_z && num_groups == 1) ? 2 : 3;
col2im_shape.set(3, input.tensor_shape()[batch_idx]);
return col2im_shape;
diff --git a/src/core/CL/kernels/CLCol2ImKernel.cpp b/src/core/CL/kernels/CLCol2ImKernel.cpp
index 40032f9..74bbb9b 100644
--- a/src/core/CL/kernels/CLCol2ImKernel.cpp
+++ b/src/core/CL/kernels/CLCol2ImKernel.cpp
@@ -40,7 +40,7 @@
namespace
{
-Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, std::pair<unsigned int, unsigned int> convolved_dims, unsigned int num_groups)
+Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Size2D &convolved_dims, unsigned int num_groups)
{
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
@@ -49,7 +49,7 @@
// Checks performed when output is configured
if(output->total_size() != 0)
{
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), compute_col2im_shape(*input, convolved_dims, num_groups));
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), compute_col2im_shape(*input, convolved_dims, true, num_groups));
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
ARM_COMPUTE_RETURN_ERROR_ON_MSG(output->data_layout() != DataLayout::NCHW, "Col2Im output's data layout must always be NCHW");
@@ -58,11 +58,11 @@
return Status{};
}
-std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, std::pair<unsigned int, unsigned int> convolved_dims, unsigned int num_groups)
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const Size2D &convolved_dims, unsigned int num_groups)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
// Output auto inizialitation if not yet initialized
- auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_col2im_shape(*input, convolved_dims, num_groups)).set_data_layout(DataLayout::NCHW));
+ auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_col2im_shape(*input, convolved_dims, true, num_groups)).set_data_layout(DataLayout::NCHW));
const unsigned int num_elems_read_per_iteration = 8;
@@ -87,7 +87,7 @@
{
}
-void CLCol2ImKernel::configure(const ICLTensor *input, ICLTensor *output, std::pair<unsigned int, unsigned int> convolved_dims, unsigned int num_groups)
+void CLCol2ImKernel::configure(const ICLTensor *input, ICLTensor *output, const Size2D &convolved_dims, unsigned int num_groups)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
@@ -105,7 +105,7 @@
build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type));
build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(input->info()->element_size()));
build_opts.add_option("-DWIDTH_INPUT=" + support::cpp11::to_string(input->info()->dimension(0)));
- build_opts.add_option("-DWIDTH_OUTPUT=" + support::cpp11::to_string(_convolved_dims.first));
+ build_opts.add_option("-DWIDTH_OUTPUT=" + support::cpp11::to_string(_convolved_dims.width));
build_opts.add_option_if(num_groups > 1, "-DGROUPING");
_kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("col2im", build_opts.options()));
@@ -130,7 +130,7 @@
_config_id += support::cpp11::to_string(output->info()->dimension(1));
}
-Status CLCol2ImKernel::validate(const ITensorInfo *input, const ITensorInfo *output, std::pair<unsigned int, unsigned int> convolved_dims, unsigned int num_groups)
+Status CLCol2ImKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &convolved_dims, unsigned int num_groups)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, convolved_dims, num_groups));
diff --git a/src/core/NEON/kernels/NECol2ImKernel.cpp b/src/core/NEON/kernels/NECol2ImKernel.cpp
index bb8e758..d6517ac 100644
--- a/src/core/NEON/kernels/NECol2ImKernel.cpp
+++ b/src/core/NEON/kernels/NECol2ImKernel.cpp
@@ -29,26 +29,17 @@
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/Validate.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include <arm_neon.h>
#include <cstddef>
#include <cstdint>
using namespace arm_compute;
+using namespace misc::shape_calculator;
namespace
{
-TensorShape get_output_shape(const ITensorInfo *input, const Size2D &convolved_dims)
-{
- TensorShape output_shape = input->tensor_shape();
- output_shape.set(0, convolved_dims.width);
- output_shape.set(1, convolved_dims.height);
- output_shape.set(2, input->tensor_shape()[0]);
- output_shape.set(3, input->tensor_shape()[3]); // For NEON the batch size is on the fourth dimension of the input tensor
-
- return output_shape;
-}
-
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Size2D &convolved_dims)
{
//Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use NEON FP16 instructions.
@@ -60,12 +51,28 @@
// Validate configured output
if(output->total_size() != 0)
{
- ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), get_output_shape(input, convolved_dims));
+ ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(output->tensor_shape(), compute_col2im_shape(*input, convolved_dims, false));
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
}
return Status{};
}
+
+std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output, const Size2D &convolved_dims)
+{
+ // Output auto inizialitation if not yet initialized
+ auto_init_if_empty(*output, input->clone()->set_tensor_shape(compute_col2im_shape(*input, convolved_dims, false)));
+
+ // Configure kernel window
+ Window win = calculate_max_window(*input, Steps());
+
+ // The NECol2ImKernel doesn't need padding so update_window_and_padding() can be skipped
+ Coordinates coord;
+ coord.set_num_dimensions(output->num_dimensions());
+ output->set_valid_region(ValidRegion(coord, output->tensor_shape()));
+
+ return std::make_pair(Status{}, win);
+}
} // namespace
template <typename T>
@@ -102,11 +109,6 @@
void NECol2ImKernel::configure(const ITensor *input, ITensor *output, const Size2D &convolved_dims)
{
ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
-
- // Output auto inizialitation if not yet initialized
- auto_init_if_empty(*output->info(), input->info()->clone()->set_tensor_shape(get_output_shape(input->info(), convolved_dims)));
-
- // Perform validation step
ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), convolved_dims));
_input = input;
@@ -130,19 +132,15 @@
}
// Configure kernel window
- Window win = calculate_max_window(*input->info(), Steps());
-
- // The NECol2ImKernel doesn't need padding so update_window_and_padding() can be skipped
- Coordinates coord;
- coord.set_num_dimensions(output->info()->num_dimensions());
- output->info()->set_valid_region(ValidRegion(coord, output->info()->tensor_shape()));
-
- INEKernel::configure(win);
+ auto win_config = validate_and_configure_window(input->info(), output->info(), convolved_dims);
+ ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
+ INEKernel::configure(win_config.second);
}
Status NECol2ImKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Size2D &convolved_dims)
{
ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, convolved_dims));
+ ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get(), convolved_dims).first);
return Status{};
}
diff --git a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
index 782fe71..c9daea4 100644
--- a/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
+++ b/src/runtime/CL/functions/CLGEMMConvolutionLayer.cpp
@@ -285,7 +285,7 @@
if(input->info()->data_layout() == DataLayout::NCHW)
{
// Configure and tune Col2Im
- _col2im_kernel.configure(_is_quantized ? gemm_output_staged_to_use : gemm_output_to_use, output, std::make_pair(conv_w, conv_h), num_groups);
+ _col2im_kernel.configure(_is_quantized ? gemm_output_staged_to_use : gemm_output_to_use, output, Size2D(conv_w, conv_h), num_groups);
CLScheduler::get().tune_kernel_static(_col2im_kernel);
}
else
@@ -443,7 +443,7 @@
{
ARM_COMPUTE_RETURN_ON_ERROR(CLCol2ImKernel::validate(is_quantized ? gemm_output_staged_to_use : gemm_output_to_use,
output,
- std::make_pair(conv_w, conv_h), num_groups));
+ Size2D(conv_w, conv_h), num_groups));
}
}
diff --git a/src/runtime/CL/functions/CLLocallyConnectedLayer.cpp b/src/runtime/CL/functions/CLLocallyConnectedLayer.cpp
index 40bf032..5c6bef9 100644
--- a/src/runtime/CL/functions/CLLocallyConnectedLayer.cpp
+++ b/src/runtime/CL/functions/CLLocallyConnectedLayer.cpp
@@ -122,7 +122,7 @@
ARM_COMPUTE_RETURN_ON_ERROR(CLIm2ColKernel::validate(input, &input_im2col_reshaped_info, Size2D(kernel_width, kernel_height), conv_info, has_bias));
ARM_COMPUTE_RETURN_ON_ERROR(CLWeightsReshapeKernel::validate(weights, biases, &weights_reshaped_info));
ARM_COMPUTE_RETURN_ON_ERROR(CLLocallyConnectedMatrixMultiplyKernel::validate(&input_im2col_reshaped_info, &weights_reshaped_info, &gemm_output_info));
- ARM_COMPUTE_RETURN_ON_ERROR(CLCol2ImKernel::validate(&gemm_output_info, output, std::make_pair(conv_w, conv_h)));
+ ARM_COMPUTE_RETURN_ON_ERROR(CLCol2ImKernel::validate(&gemm_output_info, output, Size2D(conv_w, conv_h)));
return Status{};
}
@@ -163,7 +163,7 @@
_input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _has_bias);
_weights_reshape_kernel.configure(weights, biases, &_weights_reshaped);
_mm_kernel.configure(&_input_im2col_reshaped, &_weights_reshaped, &_gemm_output);
- _output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h));
+ _output_col2im_kernel.configure(&_gemm_output, output, Size2D(conv_w, conv_h));
// Allocate intermediate tensors
_input_im2col_reshaped.allocator()->allocate();
diff --git a/src/runtime/CL/tuners/BifrostTuner.cpp b/src/runtime/CL/tuners/BifrostTuner.cpp
index 2d52f33..59d73b4 100644
--- a/src/runtime/CL/tuners/BifrostTuner.cpp
+++ b/src/runtime/CL/tuners/BifrostTuner.cpp
@@ -134,7 +134,7 @@
// via exhaustive autotuning over 30 representative tensor shapes.
if(gpu_target_is_in(gpu_target, GPUTarget::G71, GPUTarget::G72, GPUTarget::G51, GPUTarget::G51BIG, GPUTarget::G51LIT, GPUTarget::G76))
{
- if((k._convolved_dims.first == 7) || (k._convolved_dims.first == 14))
+ if((k._convolved_dims.width == 7) || (k._convolved_dims.width == 14))
{
lws_hint = cl::NDRange(1, 7, 1);
}
diff --git a/tests/validation/CL/Col2Im.cpp b/tests/validation/CL/Col2Im.cpp
index 6f1163c..1fea5c1 100644
--- a/tests/validation/CL/Col2Im.cpp
+++ b/tests/validation/CL/Col2Im.cpp
@@ -63,14 +63,14 @@
framework::dataset::make("Expected", { false, false, false, true })),
input_info, output_info, convolved_width, convolved_height, num_groups, expected)
{
- bool status = bool(CLCol2Im::validate(&input_info, &output_info, std::make_pair(convolved_width, convolved_height), num_groups));
+ bool status = bool(CLCol2Im::validate(&input_info, &output_info, Size2D(convolved_width, convolved_height), num_groups));
ARM_COMPUTE_EXPECT(status == expected, framework::LogLevel::ERRORS);
}
// clang-format on
// *INDENT-ON*
template <typename T>
-using CLCol2ImFixture = Col2ImValidationFixture<CLTensor, CLAccessor, CLCol2Im, T>;
+using CLCol2ImFixture = Col2ImValidationFixture<CLTensor, CLAccessor, CLCol2Im, T, true>;
TEST_SUITE(Float)
TEST_SUITE(FP32)
diff --git a/tests/validation/fixtures/Col2ImFixture.h b/tests/validation/fixtures/Col2ImFixture.h
index ddc78a5..5488f8a 100644
--- a/tests/validation/fixtures/Col2ImFixture.h
+++ b/tests/validation/fixtures/Col2ImFixture.h
@@ -44,16 +44,16 @@
{
using namespace arm_compute::misc::shape_calculator;
-template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool batch_size_on_z>
class Col2ImValidationFixture : public framework::Fixture
{
public:
template <typename...>
void setup(TensorShape input_shape, const unsigned int convolved_width, unsigned int convolved_height, unsigned int num_groups, DataType data_type)
{
- const std::pair<unsigned int, unsigned int> convolved_dims(convolved_width, convolved_height);
+ const Size2D convolved_dims(convolved_width, convolved_height);
- const TensorShape output_shape = compute_col2im_shape(TensorInfo(input_shape, 1, data_type), convolved_dims, num_groups);
+ const TensorShape output_shape = compute_col2im_shape(TensorInfo(input_shape, 1, data_type), convolved_dims, batch_size_on_z, num_groups);
_target = compute_target(input_shape, output_shape, convolved_dims, num_groups, data_type);
_reference = compute_reference(input_shape, output_shape, num_groups, data_type);
@@ -66,7 +66,7 @@
library->fill_tensor_uniform(tensor, seed);
}
- TensorType compute_target(const TensorShape &input_shape, const TensorShape &output_shape, std::pair<unsigned int, unsigned int> convolved_dims, unsigned int num_groups, DataType data_type)
+ TensorType compute_target(const TensorShape &input_shape, const TensorShape &output_shape, const Size2D &convolved_dims, unsigned int num_groups, DataType data_type)
{
// Create tensors
TensorType src = create_tensor<TensorType>(input_shape, data_type);