COMPMID-1201 - Implementing Winograd Convolution Layer 1x3 and 3x1 kernels on OpenCL
Change-Id: I39667bab49daa4da009694163274a59fd3574c73
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/137595
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
diff --git a/tests/validation/CL/Winograd.cpp b/tests/validation/CL/Winograd.cpp
index b869f4c..f68ec8c 100644
--- a/tests/validation/CL/Winograd.cpp
+++ b/tests/validation/CL/Winograd.cpp
@@ -23,6 +23,7 @@
*/
#include "arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h"
#include "arm_compute/core/CL/kernels/CLWinogradOutputTransformKernel.h"
+#include "arm_compute/core/Helpers.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "arm_compute/runtime/CL/CLTensor.h"
@@ -51,12 +52,66 @@
{
namespace
{
+// *INDENT-OFF*
+// clang-format off
constexpr AbsoluteTolerance<float> tolerance_f32(0.001f);
constexpr AbsoluteTolerance<float> tolerance_convolution_layer_f32(0.1f);
-const auto SmallWinogradInputTransformDataset = framework::dataset::concat(datasets::SmallWinogradInputTransformDataset2x2_3x3(),
- framework::dataset::concat(datasets::SmallWinogradInputTransformDataset4x4_3x3(), datasets::SmallWinogradInputTransformDataset4x4_5x5()));
-const auto LargeWinogradInputTransformDataset = framework::dataset::concat(datasets::LargeWinogradInputTransformDataset2x2_3x3(),
- framework::dataset::concat(datasets::LargeWinogradInputTransformDataset4x4_3x3(), datasets::LargeWinogradInputTransformDataset4x4_5x5()));
+
+// Input transform
+const auto SmallWinogradInputTransformDatasetNCHW =
+ framework::dataset::concat(datasets::SmallWinogradInputTransformDataset2x2_3x3(),
+ framework::dataset::concat(datasets::SmallWinogradInputTransformDataset2x1_3x1(),
+ framework::dataset::concat(datasets::SmallWinogradInputTransformDataset1x2_1x3(),
+ framework::dataset::concat(datasets::SmallWinogradInputTransformDataset4x4_3x3(),
+ framework::dataset::concat(datasets::SmallWinogradInputTransformDataset4x1_3x1(),
+ framework::dataset::concat(datasets::SmallWinogradInputTransformDataset1x4_1x3(),
+ datasets::SmallWinogradInputTransformDataset4x4_5x5()))))));
+
+const auto SmallWinogradInputTransformDatasetNHWC = framework::dataset::concat(datasets::SmallWinogradInputTransformDataset4x4_3x3(),
+ datasets::SmallWinogradInputTransformDataset4x4_5x5());
+
+const auto LargeWinogradInputTransformDatasetNCHW =
+ framework::dataset::concat(datasets::LargeWinogradInputTransformDataset2x2_3x3(),
+ framework::dataset::concat(datasets::LargeWinogradInputTransformDataset2x1_3x1(),
+ framework::dataset::concat(datasets::LargeWinogradInputTransformDataset1x2_1x3(),
+ framework::dataset::concat(datasets::LargeWinogradInputTransformDataset4x4_3x3(),
+ framework::dataset::concat(datasets::LargeWinogradInputTransformDataset4x1_3x1(),
+ framework::dataset::concat(datasets::LargeWinogradInputTransformDataset1x4_1x3(),
+ datasets::LargeWinogradInputTransformDataset4x4_5x5()))))));
+
+const auto LargeWinogradInputTransformDatasetNHWC =
+ framework::dataset::concat(datasets::LargeWinogradInputTransformDataset4x4_3x3(),
+ datasets::LargeWinogradInputTransformDataset4x4_5x5());
+
+// Filter transform
+const auto SmallWinogradFilterTransformDatasetNCHW =
+ framework::dataset::concat(combine(datasets::Small3x3Shapes(), framework::dataset::make("OutputTile", { Size2D(2U, 2U), Size2D(4U, 4U) })),
+ framework::dataset::concat(combine(datasets::Small3x1Shapes(), framework::dataset::make("OutputTile", { Size2D(2U, 1U), Size2D(4U, 1U) })),
+ framework::dataset::concat(combine(datasets::Small1x3Shapes(), framework::dataset::make("OutputTile", { Size2D(1U, 2U), Size2D(1U, 4U) })),
+ combine(datasets::Small5x5Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 4U) })))));
+
+const auto SmallWinogradFilterTransformDatasetNHWC =
+ framework::dataset::concat(combine(datasets::Small3x3Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 4U) })),
+ combine(datasets::Small5x5Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 4U) })));
+
+const auto LargeWinogradFilterTransformDatasetNCHW =
+ framework::dataset::concat(combine(datasets::Large3x3Shapes(), framework::dataset::make("OutputTile", { Size2D(2U, 2U), Size2D(4U, 4U) })),
+ framework::dataset::concat(combine(datasets::Large3x1Shapes(), framework::dataset::make("OutputTile", { Size2D(2U, 1U), Size2D(4U, 1U) })),
+ framework::dataset::concat(combine(datasets::Large1x3Shapes(), framework::dataset::make("OutputTile", { Size2D(1U, 2U), Size2D(1U, 4U) })),
+ combine(datasets::Large5x5Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 4U) })))));
+
+const auto LargeWinogradFilterTransformDatasetNHWC =
+ framework::dataset::concat(combine(datasets::Large3x3Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 4U) })),
+ combine(datasets::Large5x5Shapes(), framework::dataset::make("OutputTile", { Size2D(4U, 4U) })));
+
+// Output transform
+const auto SmallWinogradOutputTransformDatasetNCHW = datasets::SmallWinogradOutputTransformDatasetNCHW();
+
+const auto SmallWinogradOutputTransformDatasetNHWC = datasets::SmallWinogradOutputTransformDatasetNHWC();
+
+const auto LargeWinogradOutputTransformDatasetNCHW = datasets::LargeWinogradOutputTransformDatasetNCHW();
+
+const auto LargeWinogradOutputTransformDatasetNHWC = datasets::LargeWinogradOutputTransformDatasetNHWC();
} // namespace
using namespace arm_compute::misc::shape_calculator;
@@ -65,9 +120,6 @@
TEST_SUITE(Winograd)
TEST_SUITE(InputTransform)
-
-// *INDENT-OFF*
-// clang-format off
DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
framework::dataset::make("InputInfo",{
TensorInfo(TensorShape(53U, 21U, 5U, 3U), 1, DataType::F16), // F16 not supported
@@ -101,17 +153,20 @@
{
ARM_COMPUTE_EXPECT(bool(CLWinogradInputTransform::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), winograd_info)) == expected, framework::LogLevel::ERRORS);
}
-// clang-format on
-// *INDENT-ON*
using CLWinogradInputTransformFixture = WinogradInputTransformValidationFixture<CLTensor, CLAccessor, CLWinogradInputTransform, float>;
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(framework::dataset::concat(SmallWinogradInputTransformDataset, LargeWinogradInputTransformDataset),
+TEST_SUITE(NCHW)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(framework::dataset::concat(SmallWinogradInputTransformDatasetNCHW,
+ LargeWinogradInputTransformDatasetNCHW),
framework::dataset::make("DataLayout", { DataLayout::NCHW })),
- framework::dataset::make("DataType", { DataType::F32 })),
+ framework::dataset::make("DataType", { DataType::F32 })),
shape_in, winograd_info, data_layout, data_type)
{
- TensorShape shape_out = compute_winograd_input_transform_shape(TensorInfo(shape_in, 1, data_type), winograd_info);
+ TensorInfo tensor_info_in(shape_in, 1, data_type);
+ tensor_info_in.set_data_layout(data_layout);
+
+ TensorShape shape_out = compute_winograd_input_transform_shape(tensor_info_in, winograd_info);
// Create tensors
CLTensor in = create_tensor<CLTensor>(shape_in, data_type, 1, 0, QuantizationInfo(), data_layout);
@@ -127,28 +182,70 @@
winograd_input_transform.configure(&in, &out, winograd_info);
}
-FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradInputTransformFixture, framework::DatasetMode::PRECOMMIT, combine(framework::dataset::concat(combine(SmallWinogradInputTransformDataset,
- framework::dataset::make("DataLayout", { DataLayout::NCHW })),
- combine(framework::dataset::concat(datasets::SmallWinogradInputTransformDataset4x4_3x3(), datasets::SmallWinogradInputTransformDataset4x4_5x5()),
- framework::dataset::make("DataLayout", { DataLayout::NHWC }))),
- framework::dataset::make("DataType", { DataType::F32 })))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradInputTransformFixture, framework::DatasetMode::PRECOMMIT, combine(combine(SmallWinogradInputTransformDatasetNCHW,
+ framework::dataset::make("DataLayout", { DataLayout::NCHW })),
+ framework::dataset::make("DataType", { DataType::F32 })))
{
validate(CLAccessor(_target), _reference, tolerance_f32);
}
-FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradInputTransformFixture, framework::DatasetMode::NIGHTLY, combine(framework::dataset::concat(combine(LargeWinogradInputTransformDataset,
- framework::dataset::make("DataLayout", { DataLayout::NCHW })),
- combine(framework::dataset::concat(datasets::LargeWinogradInputTransformDataset4x4_3x3(), datasets::LargeWinogradInputTransformDataset4x4_5x5()),
- framework::dataset::make("DataLayout", { DataLayout::NHWC }))),
- framework::dataset::make("DataType", { DataType::F32 })))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradInputTransformFixture, framework::DatasetMode::NIGHTLY, combine(combine(LargeWinogradInputTransformDatasetNCHW,
+ framework::dataset::make("DataLayout", { DataLayout::NCHW })),
+ framework::dataset::make("DataType", { DataType::F32 })))
{
validate(CLAccessor(_target), _reference, tolerance_f32);
}
+TEST_SUITE_END() // NCHW
+
+TEST_SUITE(NHWC)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(framework::dataset::concat(SmallWinogradInputTransformDatasetNHWC,
+ LargeWinogradInputTransformDatasetNHWC),
+ framework::dataset::make("DataLayout", { DataLayout::NHWC })),
+ framework::dataset::make("DataType", { DataType::F32 })),
+ shape_in, winograd_info, data_layout, data_type)
+{
+ TensorShape shape_in_nhwc(shape_in);
+
+ // Convert the shape to NHWC
+ permute(shape_in_nhwc, PermutationVector(2U, 0U, 1U));
+
+ // TensorInfo
+ TensorInfo tensor_info_in(shape_in_nhwc, 1, data_type);
+ tensor_info_in.set_data_layout(data_layout);
+
+ TensorShape shape_out = compute_winograd_input_transform_shape(tensor_info_in, winograd_info);
+
+ // Create tensors
+ CLTensor in = create_tensor<CLTensor>(shape_in_nhwc, data_type, 1, 0, QuantizationInfo(), data_layout);
+ CLTensor out = create_tensor<CLTensor>(shape_out, data_type);
+
+ ARM_COMPUTE_EXPECT(in.info()->is_resizable(), framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(out.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+ // Create and configure function
+ CLWinogradInputTransform winograd_input_transform;
+
+ // Configure the function
+ winograd_input_transform.configure(&in, &out, winograd_info);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradInputTransformFixture, framework::DatasetMode::PRECOMMIT, combine(combine(SmallWinogradInputTransformDatasetNHWC,
+ framework::dataset::make("DataLayout", { DataLayout::NHWC })),
+ framework::dataset::make("DataType", { DataType::F32 })))
+{
+ validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradInputTransformFixture, framework::DatasetMode::NIGHTLY, combine(combine(LargeWinogradInputTransformDatasetNHWC,
+ framework::dataset::make("DataLayout", { DataLayout::NHWC })),
+ framework::dataset::make("DataType", { DataType::F32 })))
+{
+ validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END() // NHWC
TEST_SUITE_END() // InputTransform
TEST_SUITE(FilterTransform)
-// *INDENT-OFF*
-// clang-format off
DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
framework::dataset::make("InputInfo",{
TensorInfo(TensorShape(3U, 3U, 5U, 3U), 1, DataType::F16), // F16 not supported
@@ -182,19 +279,19 @@
{
ARM_COMPUTE_EXPECT(bool(CLWinogradFilterTransformKernel::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), winograd_info)) == expected, framework::LogLevel::ERRORS);
}
-// clang-format on
-// *INDENT-ON*
using CLWinogradFilterTransform = CLSynthetizeFunctionWithZeroConstantBorder<CLWinogradFilterTransformKernel, 0>;
using CLWinogradFilterTransformFixture = WinogradFilterTransformValidationFixture<CLTensor, CLAccessor, CLWinogradFilterTransform, float>;
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(framework::dataset::concat(datasets::Small3x3Shapes(), datasets::Large3x3Shapes()),
- framework::dataset::make("OutputTile", { Size2D(2U, 2U), Size2D(4U, 4U) })),
- framework::dataset::make("DataLayout", { DataLayout::NCHW })),
- framework::dataset::make("DataType", { DataType::F32 })),
+TEST_SUITE(NCHW)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL,
+ combine(combine(framework::dataset::concat(SmallWinogradFilterTransformDatasetNCHW,
+ LargeWinogradFilterTransformDatasetNCHW),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW })),
+ framework::dataset::make("DataType", { DataType::F32 })),
shape_a, output_tile, data_layout, data_type)
{
- WinogradInfo winograd_info(output_tile, Size2D(shape_a[0], shape_a[1]), Size2D() /* Not needed */, PadStrideInfo() /* Not needed */, DataLayout::NCHW /* Not needed */);
+ WinogradInfo winograd_info(output_tile, Size2D(shape_a[0], shape_a[1]), Size2D() /* Not needed */, PadStrideInfo() /* Not needed */, data_layout /* Not needed */);
TensorShape shape_b = compute_winograd_filter_transform_shape(TensorInfo(shape_a, 1, data_type), winograd_info);
@@ -210,37 +307,79 @@
winograd_filter_transform.configure(&a, &b, winograd_info);
}
-FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradFilterTransformFixture, framework::DatasetMode::ALL,
- combine(framework::dataset::concat(combine(framework::dataset::concat(framework::dataset::concat(combine(datasets::Small3x3Shapes(), framework::dataset::make("OutputTile", Size2D(2U, 2U))),
- combine(datasets::Small3x3Shapes(),
- framework::dataset::make("OutputTile", Size2D(4U, 4U)))),
- combine(datasets::Small5x5Shapes(), framework::dataset::make("OutputTile", Size2D(4U, 4U)))),
- framework::dataset::make("DataLayout", { DataLayout::NCHW })),
- combine(combine(framework::dataset::concat(datasets::Small3x3Shapes(), datasets::Small5x5Shapes()), framework::dataset::make("OutputTile", Size2D(4U, 4U))), framework::dataset::make("DataLayout", { DataLayout::NHWC }))),
- framework::dataset::make("DataType", { DataType::F32 })))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradFilterTransformFixture, framework::DatasetMode::PRECOMMIT,
+ combine(combine(SmallWinogradFilterTransformDatasetNCHW,
+ framework::dataset::make("DataLayout", { DataLayout::NCHW })),
+ framework::dataset::make("DataType", { DataType::F32 })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_f32);
}
FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradFilterTransformFixture, framework::DatasetMode::NIGHTLY,
- combine(framework::dataset::concat(combine(framework::dataset::concat(framework::dataset::concat(combine(datasets::Large3x3Shapes(), framework::dataset::make("OutputTile", Size2D(2U, 2U))),
- combine(datasets::Large3x3Shapes(),
- framework::dataset::make("OutputTile", Size2D(4U, 4U)))),
- combine(datasets::Large5x5Shapes(), framework::dataset::make("OutputTile", Size2D(4U, 4U)))),
- framework::dataset::make("DataLayout", { DataLayout::NCHW })),
- combine(combine(framework::dataset::concat(datasets::Large3x3Shapes(), datasets::Large5x5Shapes()), framework::dataset::make("OutputTile", Size2D(4U, 4U))), framework::dataset::make("DataLayout", { DataLayout::NHWC }))),
- framework::dataset::make("DataType", { DataType::F32 })))
+ combine(combine(LargeWinogradFilterTransformDatasetNCHW,
+ framework::dataset::make("DataLayout", { DataLayout::NCHW })),
+ framework::dataset::make("DataType", { DataType::F32 })))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END() // NCHW
+
+TEST_SUITE(NHWC)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL,
+ combine(combine(framework::dataset::concat(SmallWinogradFilterTransformDatasetNHWC,
+ LargeWinogradFilterTransformDatasetNHWC),
+ framework::dataset::make("DataLayout", { DataLayout::NHWC })),
+ framework::dataset::make("DataType", { DataType::F32 })),
+ shape_in, output_tile, data_layout, data_type)
+{
+ TensorShape shape_in_nhwc(shape_in);
+
+ // Convert the shape to NHWC
+ permute(shape_in_nhwc, PermutationVector(2U, 0U, 1U));
+
+ // TensorInfo
+ TensorInfo tensor_info_in(shape_in_nhwc, 1, data_type);
+ tensor_info_in.set_data_layout(data_layout);
+
+ WinogradInfo winograd_info(output_tile, Size2D(shape_in[0], shape_in[1]), Size2D() /* Not needed */, PadStrideInfo() /* Not needed */, data_layout /* Not needed */);
+
+ TensorShape shape_b = compute_winograd_filter_transform_shape(tensor_info_in, winograd_info);
+
+ // Create tensors
+ CLTensor a = create_tensor<CLTensor>(shape_in_nhwc, data_type, 1, 0, QuantizationInfo(), data_layout);
+ CLTensor b = create_tensor<CLTensor>(shape_b, data_type, 1, 0, QuantizationInfo(), data_layout);
+
+ ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+ // Create and configure function
+ CLWinogradFilterTransform winograd_filter_transform;
+ winograd_filter_transform.configure(&a, &b, winograd_info);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradFilterTransformFixture, framework::DatasetMode::PRECOMMIT,
+ combine(combine(SmallWinogradFilterTransformDatasetNHWC,
+ framework::dataset::make("DataLayout", { DataLayout::NHWC })),
+ framework::dataset::make("DataType", { DataType::F32 })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_f32);
}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradFilterTransformFixture, framework::DatasetMode::NIGHTLY,
+ combine(combine(LargeWinogradFilterTransformDatasetNHWC,
+ framework::dataset::make("DataLayout", { DataLayout::NHWC })),
+ framework::dataset::make("DataType", { DataType::F32 })))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END() // NHWC
TEST_SUITE_END() // FilterTransform
TEST_SUITE(OutputTransform)
-// *INDENT-OFF*
-// clang-format off
DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(
framework::dataset::make("InputInfo",{
TensorInfo(TensorShape(512U, 49U, 16U, 5U), 1, DataType::F16), // F16 not supported
@@ -291,14 +430,14 @@
{
ARM_COMPUTE_EXPECT(bool(CLWinogradOutputTransformKernel::validate(&input_info.clone()->set_is_resizable(false), &bias_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), winograd_info)) == expected, framework::LogLevel::ERRORS);
}
-// clang-format on
-// *INDENT-ON*
using CLWinogradOutputTransform = CLSynthetizeFunctionWithZeroConstantBorder<CLWinogradOutputTransformKernel, 0>;
using CLWinogradOutputTransformFixture = WinogradOutputTransformValidationFixture<CLTensor, CLAccessor, CLWinogradOutputTransform, float>;
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallWinogradOutputTransformDataset(), datasets::LargeWinogradOutputTransformDataset()),
- framework::dataset::make("DataType", { DataType::F32 })),
+TEST_SUITE(NCHW)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(SmallWinogradOutputTransformDatasetNCHW,
+ LargeWinogradOutputTransformDatasetNCHW),
+ framework::dataset::make("DataType", { DataType::F32 })),
shape_a, winograd_info, data_type)
{
TensorShape shape_b = compute_winograd_output_transform_shape(TensorInfo(shape_a, 1, data_type), winograd_info);
@@ -315,23 +454,62 @@
winograd_output_transform.configure(&a, nullptr, &b, winograd_info);
}
-FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradOutputTransformFixture, framework::DatasetMode::ALL, combine(datasets::SmallWinogradOutputTransformDataset(), framework::dataset::make("DataType", { DataType::F32 })))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradOutputTransformFixture, framework::DatasetMode::ALL,
+ combine(SmallWinogradOutputTransformDatasetNCHW,
+ framework::dataset::make("DataType", { DataType::F32 })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_f32);
}
-FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradOutputTransformFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeWinogradOutputTransformDataset(), framework::dataset::make("DataType", { DataType::F32 })))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradOutputTransformFixture, framework::DatasetMode::NIGHTLY,
+ combine(LargeWinogradOutputTransformDatasetNCHW,
+ framework::dataset::make("DataType", { DataType::F32 })))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END() // NCHW
+
+TEST_SUITE(NHWC)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(SmallWinogradOutputTransformDatasetNHWC,
+ LargeWinogradOutputTransformDatasetNHWC),
+ framework::dataset::make("DataType", { DataType::F32 })),
+ shape_a, winograd_info, data_type)
+{
+ TensorShape shape_b = compute_winograd_output_transform_shape(TensorInfo(shape_a, 1, data_type), winograd_info);
+
+ // Create tensors
+ CLTensor a = create_tensor<CLTensor>(shape_a, data_type);
+ CLTensor b = create_tensor<CLTensor>(shape_b, data_type, 1, 0, QuantizationInfo(), winograd_info.output_data_layout);
+
+ ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+ // Create and configure function
+ CLWinogradOutputTransform winograd_output_transform;
+ winograd_output_transform.configure(&a, nullptr, &b, winograd_info);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradOutputTransformFixture, framework::DatasetMode::ALL,
+ combine(SmallWinogradOutputTransformDatasetNHWC,
+ framework::dataset::make("DataType", { DataType::F32 })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_f32);
}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradOutputTransformFixture, framework::DatasetMode::NIGHTLY,
+ combine(LargeWinogradOutputTransformDatasetNHWC,
+ framework::dataset::make("DataType", { DataType::F32 })))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END() // NHWC
TEST_SUITE_END() // OutputTransform
TEST_SUITE(ConvolutionLayer)
-// *INDENT-OFF*
-// clang-format off
DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(
framework::dataset::make("InputInfo", {
TensorInfo(TensorShape(17U, 31U, 2U), 1, DataType::F16), // FP16 not supported
@@ -373,16 +551,14 @@
{
ARM_COMPUTE_EXPECT(bool(CLWinogradConvolutionLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &bias_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv_info)) == expected, framework::LogLevel::ERRORS);
}
-// clang-format on
-// *INDENT-ON*
using CLWinogradConvolutionLayerFastMathFixture = WinogradConvolutionLayerFastMathValidationFixture<CLTensor, CLAccessor, CLWinogradConvolutionLayer, float>;
TEST_SUITE(Conv3x3)
FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::PRECOMMIT,
combine(combine(combine(datasets::SmallWinogradConvolutionLayer3x3Dataset(),
framework::dataset::make("DataType", { DataType::F32 })),
- framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() })),
- framework::dataset::make("DataLayout", { DataLayout::NCHW })))
+ framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() })),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32);
@@ -391,20 +567,64 @@
FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::NIGHTLY,
combine(combine(combine(datasets::LargeWinogradConvolutionLayer3x3Dataset(),
framework::dataset::make("DataType", { DataType::F32 })),
- framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() })),
- framework::dataset::make("DataLayout", { DataLayout::NCHW })))
+ framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() })),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW })))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32);
}
TEST_SUITE_END() // Conv3x3
+TEST_SUITE(Conv3x1)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::PRECOMMIT,
+ combine(combine(combine(datasets::SmallWinogradConvolutionLayer3x1Dataset(),
+ framework::dataset::make("DataType", { DataType::F32 })),
+ framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() })),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW })))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::NIGHTLY,
+ combine(combine(combine(datasets::LargeWinogradConvolutionLayer3x1Dataset(),
+ framework::dataset::make("DataType", { DataType::F32 })),
+ framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() })),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW })))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32);
+}
+TEST_SUITE_END() // Conv3x1
+
+TEST_SUITE(Conv1x3)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::PRECOMMIT,
+ combine(combine(combine(datasets::SmallWinogradConvolutionLayer1x3Dataset(),
+ framework::dataset::make("DataType", { DataType::F32 })),
+ framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() })),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW })))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::NIGHTLY,
+ combine(combine(combine(datasets::LargeWinogradConvolutionLayer1x3Dataset(),
+ framework::dataset::make("DataType", { DataType::F32 })),
+ framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() })),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW })))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32);
+}
+TEST_SUITE_END() // Conv1x3
+
TEST_SUITE(Conv5x5)
FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::PRECOMMIT,
combine(combine(combine(datasets::SmallWinogradConvolutionLayer5x5Dataset(),
framework::dataset::make("DataType", { DataType::F32 })),
- framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() })),
- framework::dataset::make("DataLayout", { DataLayout::NCHW })))
+ framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() })),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW })))
{
// Validate output
@@ -414,8 +634,8 @@
FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::NIGHTLY,
combine(combine(combine(datasets::LargeWinogradConvolutionLayer5x5Dataset(),
framework::dataset::make("DataType", { DataType::F32 })),
- framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() })),
- framework::dataset::make("DataLayout", { DataLayout::NCHW })))
+ framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() })),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW })))
{
// Validate output
@@ -424,7 +644,6 @@
TEST_SUITE_END() // Conv5x5
TEST_SUITE_END() // ConvolutionLayer
-
TEST_SUITE_END() // Winograd
TEST_SUITE_END() // CL
} // namespace validation
diff --git a/tests/validation/Helpers.cpp b/tests/validation/Helpers.cpp
index e2415a2..ff69b1c 100644
--- a/tests/validation/Helpers.cpp
+++ b/tests/validation/Helpers.cpp
@@ -215,7 +215,7 @@
template <typename T>
void get_tile(const SimpleTensor<T> &in, SimpleTensor<T> &tile, const Coordinates &coord)
{
- ARM_COMPUTE_ERROR_ON(tile.shape().num_dimensions() != 2);
+ ARM_COMPUTE_ERROR_ON(tile.shape().num_dimensions() > 2);
const int w_tile = tile.shape()[0];
const int h_tile = tile.shape()[1];
@@ -272,7 +272,36 @@
}
}
+template <typename T>
+void zeros(SimpleTensor<T> &in, const Coordinates &anchor, const TensorShape &shape)
+{
+ ARM_COMPUTE_ERROR_ON(anchor.num_dimensions() != shape.num_dimensions());
+ ARM_COMPUTE_ERROR_ON(in.shape().num_dimensions() > 2);
+ ARM_COMPUTE_ERROR_ON(shape.num_dimensions() > 2);
+
+ // Check if with the dimensions greater than 2 we could have out-of-bound reads
+ for(size_t d = 0; d < Coordinates::num_max_dimensions; ++d)
+ {
+ if(anchor[d] < 0 || ((anchor[d] + shape[d]) > in.shape()[d]))
+ {
+ ARM_COMPUTE_ERROR("anchor[d] < 0 || (anchor[d] + shape[d]) > in.shape()[d]");
+ }
+ }
+
+ // Get input pointer
+ auto in_ptr = static_cast<T *>(in(anchor[0] + anchor[1] * in.shape()[0]));
+
+ const unsigned int n = in.shape()[0];
+
+ for(unsigned int y = 0; y < shape[1]; ++y)
+ {
+ std::fill(in_ptr, in_ptr + shape[0], 0);
+ in_ptr += n;
+ }
+}
+
template void get_tile(const SimpleTensor<float> &in, SimpleTensor<float> &roi, const Coordinates &coord);
+template void zeros(SimpleTensor<float> &in, const Coordinates &anchor, const TensorShape &shape);
} // namespace validation
} // namespace test
} // namespace arm_compute
diff --git a/tests/validation/Helpers.h b/tests/validation/Helpers.h
index 49432d6..88262d5 100644
--- a/tests/validation/Helpers.h
+++ b/tests/validation/Helpers.h
@@ -259,6 +259,15 @@
*/
template <typename T>
void get_tile(const SimpleTensor<T> &in, SimpleTensor<T> &tile, const Coordinates &coord);
+
+/** Fill with zeros the input tensor in the area defined by anchor and shape
+ *
+ * @param[in] in Input tensor to fill with zeros
+ * @param[out] anchor Starting point of the zeros area
+ * @param[in] shape Ending point of the zeros area
+ */
+template <typename T>
+void zeros(SimpleTensor<T> &in, const Coordinates &anchor, const TensorShape &shape);
} // namespace validation
} // namespace test
} // namespace arm_compute
diff --git a/tests/validation/fixtures/WinogradConvolutionLayerFixture.h b/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
index aca24f1..ac168eb 100644
--- a/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
+++ b/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
@@ -259,7 +259,18 @@
fill(bias, 2, 0.f, 0.f);
}
- WinogradInfo winograd_info(Size2D(4U, 4U),
+ // Set output tile
+ Size2D output_tile(4U, 4U);
+ if(weights_shape[0] == 1)
+ {
+ output_tile.width = 1;
+ }
+ else if(weights_shape[1] == 1)
+ {
+ output_tile.height = 1;
+ }
+
+ WinogradInfo winograd_info(output_tile,
Size2D(weights_shape[0], weights_shape[1]),
Size2D(input_shape[0], input_shape[1]),
info,
diff --git a/tests/validation/reference/Winograd.cpp b/tests/validation/reference/Winograd.cpp
index 197d218..5be4fe2 100644
--- a/tests/validation/reference/Winograd.cpp
+++ b/tests/validation/reference/Winograd.cpp
@@ -29,6 +29,7 @@
#include "arm_compute/core/Types.h"
#include <algorithm>
+#include <cmath>
namespace arm_compute
{
@@ -142,12 +143,24 @@
{
{ WinogradKey(std::pair<int, int>(2, 2), std::pair<int, int>(3, 3), WinogradTransformType::INPUT), imatrix2x2_3x3 },
{ WinogradKey(std::pair<int, int>(4, 4), std::pair<int, int>(3, 3), WinogradTransformType::INPUT), imatrix4x4_3x3 },
+ { WinogradKey(std::pair<int, int>(2, 1), std::pair<int, int>(3, 1), WinogradTransformType::INPUT), imatrix2x2_3x3 },
+ { WinogradKey(std::pair<int, int>(4, 1), std::pair<int, int>(3, 1), WinogradTransformType::INPUT), imatrix4x4_3x3 },
+ { WinogradKey(std::pair<int, int>(1, 2), std::pair<int, int>(1, 3), WinogradTransformType::INPUT), imatrix2x2_3x3 },
+ { WinogradKey(std::pair<int, int>(1, 4), std::pair<int, int>(1, 3), WinogradTransformType::INPUT), imatrix4x4_3x3 },
{ WinogradKey(std::pair<int, int>(4, 4), std::pair<int, int>(5, 5), WinogradTransformType::INPUT), imatrix4x4_5x5 },
{ WinogradKey(std::pair<int, int>(2, 2), std::pair<int, int>(3, 3), WinogradTransformType::FILTER), fmatrix2x2_3x3 },
{ WinogradKey(std::pair<int, int>(4, 4), std::pair<int, int>(3, 3), WinogradTransformType::FILTER), fmatrix4x4_3x3 },
+ { WinogradKey(std::pair<int, int>(2, 1), std::pair<int, int>(3, 1), WinogradTransformType::FILTER), fmatrix2x2_3x3 },
+ { WinogradKey(std::pair<int, int>(4, 1), std::pair<int, int>(3, 1), WinogradTransformType::FILTER), fmatrix4x4_3x3 },
+ { WinogradKey(std::pair<int, int>(1, 2), std::pair<int, int>(1, 3), WinogradTransformType::FILTER), fmatrix2x2_3x3 },
+ { WinogradKey(std::pair<int, int>(1, 4), std::pair<int, int>(1, 3), WinogradTransformType::FILTER), fmatrix4x4_3x3 },
{ WinogradKey(std::pair<int, int>(4, 4), std::pair<int, int>(5, 5), WinogradTransformType::FILTER), fmatrix4x4_5x5 },
{ WinogradKey(std::pair<int, int>(2, 2), std::pair<int, int>(3, 3), WinogradTransformType::OUTPUT), omatrix2x2_3x3 },
{ WinogradKey(std::pair<int, int>(4, 4), std::pair<int, int>(3, 3), WinogradTransformType::OUTPUT), omatrix4x4_3x3 },
+ { WinogradKey(std::pair<int, int>(2, 1), std::pair<int, int>(3, 1), WinogradTransformType::OUTPUT), omatrix2x2_3x3 },
+ { WinogradKey(std::pair<int, int>(4, 1), std::pair<int, int>(3, 1), WinogradTransformType::OUTPUT), omatrix4x4_3x3 },
+ { WinogradKey(std::pair<int, int>(1, 2), std::pair<int, int>(1, 3), WinogradTransformType::OUTPUT), omatrix2x2_3x3 },
+ { WinogradKey(std::pair<int, int>(1, 4), std::pair<int, int>(1, 3), WinogradTransformType::OUTPUT), omatrix4x4_3x3 },
{ WinogradKey(std::pair<int, int>(4, 4), std::pair<int, int>(5, 5), WinogradTransformType::OUTPUT), omatrix4x4_5x5 },
};
@@ -175,6 +188,20 @@
} // namespace
template <typename T>
+void print_tile(SimpleTensor<T> &in)
+{
+ for(int y = 0; y < in.shape()[1]; y++)
+ {
+ for(int x = 0; x < in.shape()[0]; x++)
+ {
+ std::cout << in[x + y * in.shape()[0]] << " ";
+ }
+
+ std::cout << std::endl;
+ }
+}
+
+template <typename T>
SimpleTensor<T> winograd_input_transform(const SimpleTensor<T> &in, const TensorShape &output_shape, const WinogradInfo &winograd_info)
{
ARM_COMPUTE_ERROR_ON(in.data_layout() != DataLayout::NCHW);
@@ -189,7 +216,10 @@
const unsigned int tile_w = output_tile_size.width + kernel_size.width - 1;
const unsigned int tile_h = output_tile_size.height + kernel_size.height - 1;
- TensorShape tile_dims(tile_w, tile_h);
+ // Get the maximum dimension from the tile size
+ const unsigned int tile_max_dim = std::max(tile_w, tile_h);
+
+ TensorShape tile_dims(tile_max_dim, tile_max_dim);
// Simple tensor for the input tile
SimpleTensor<T> src_tile{ tile_dims, in.data_type() };
@@ -217,11 +247,46 @@
const int in_d = in.shape().z();
const int out_d = out.shape().z();
const int num_batches = in.shape().total_size() / (in_w * in_h * in_d);
- const int num_tiles_x = std::ceil((in_w - (kernel_size.width - 1) + conv_info.pad_left() + conv_info.pad_right()) / static_cast<float>(output_tile_size.width));
- const int num_tiles_y = std::ceil((in_h - (kernel_size.height - 1) + conv_info.pad_top() + conv_info.pad_bottom()) / static_cast<float>(output_tile_size.height));
const int step_x = output_tile_size.width;
const int step_y = output_tile_size.height;
+ // Compute the number of output tiles along the x and y direction of size "output_tile_size"
+ const Size2D num_tiles = compute_winograd_convolution_tiles(Size2D(in_w, in_h),
+ kernel_size,
+ output_tile_size,
+ conv_info);
+
+ const int num_tiles_x = num_tiles.width;
+ const int num_tiles_y = num_tiles.height;
+
+ // In case of 1D convolution, the input tile has to be partially filled with zeros
+ int start_x_zero = 0;
+ int start_y_zero = 0;
+ int end_x_zero = 0;
+ int end_y_zero = 0;
+
+ if(output_tile_size.width == 1)
+ {
+ start_x_zero = 1;
+ start_y_zero = 0;
+ end_x_zero = tile_max_dim - 1;
+ end_y_zero = tile_max_dim;
+ }
+ else if(output_tile_size.height == 1)
+ {
+ start_x_zero = 0;
+ start_y_zero = 1;
+ end_x_zero = tile_max_dim;
+ end_y_zero = tile_max_dim - 1;
+ }
+
+ // Set the anchor and shape of the zeros area
+ const Coordinates anchor_zeros(start_x_zero, start_y_zero);
+ const TensorShape shape_zeros(end_x_zero, end_y_zero);
+
+ // If we have a vertical filter (i.e. 1x3, 1x5,..), we need to take the elements along the y direction (step = width of the output tile)
+ const int step_y_transf_tile = kernel_size.width == 1 ? tile_max_dim : 1;
+
ARM_COMPUTE_ERROR_ON((num_tiles_x * num_tiles_y) != static_cast<int>(out.shape().y()));
for(int b = 0; b < num_batches; ++b)
@@ -238,6 +303,9 @@
// Get the tile from the input tensor
get_tile(in, src_tile, Coordinates(xi, yi, z, b));
+ // Fill partially with zeros in case of 1D convolution
+ zeros(src_tile, anchor_zeros, shape_zeros);
+
// Compute the transformation
matrix_multiply(matrix, src_tile, tmp_tile);
matrix_multiply(tmp_tile, matrix_transposed, dst_tile);
@@ -247,7 +315,7 @@
{
int xo = z;
int yo = x + y * num_tiles_x;
- out[coords2index(out.shape(), Coordinates(xo, yo, i, b))] = dst_tile[i];
+ out[coords2index(out.shape(), Coordinates(xo, yo, i, b))] = dst_tile[i * step_y_transf_tile];
}
}
}
@@ -268,27 +336,31 @@
const Size2D output_tile_size = winograd_info.output_tile_size;
const Size2D kernel_size = winograd_info.kernel_size;
- TensorShape kernel_tile_dims(kernel_size.width, kernel_size.height);
-
// Calculate dimensions for the tile
const unsigned int input_tile_w = output_tile_size.width + kernel_size.width - 1;
const unsigned int input_tile_h = output_tile_size.height + kernel_size.height - 1;
const unsigned int input_tile_area = input_tile_w * input_tile_h;
+ // Get the maximum dimension from the filter size
+ const unsigned int kernel_max_dim = std::max(kernel_size.width, kernel_size.height);
+
+ // Get the maximum dimension from the input tile
+ const unsigned int input_tile_max_dim = std::max(input_tile_w, input_tile_h);
+
// Simple tensor for the input tile
- SimpleTensor<T> input_tile{ kernel_tile_dims, in.data_type(), 1 };
+ SimpleTensor<T> input_tile{ TensorShape(kernel_max_dim, kernel_max_dim), in.data_type(), 1 };
// Simple tensor for the transformation matrix
- SimpleTensor<T> trans_matrix{ TensorShape(kernel_tile_dims[0], input_tile_w), in.data_type(), 1 };
+ SimpleTensor<T> trans_matrix{ TensorShape(kernel_max_dim, input_tile_max_dim), in.data_type(), 1 };
// Simple tensor for the transformation matrix transpose
- SimpleTensor<T> trans_matrix_transposed{ TensorShape(input_tile_w, kernel_tile_dims[0]), in.data_type(), 1 };
+ SimpleTensor<T> trans_matrix_transposed{ TensorShape(input_tile_max_dim, kernel_max_dim), in.data_type(), 1 };
// Simple tensor for the temporary tile
- SimpleTensor<T> tmp_tile{ TensorShape(kernel_tile_dims[0], input_tile_w), in.data_type(), 1 };
+ SimpleTensor<T> tmp_tile{ TensorShape(kernel_max_dim, input_tile_max_dim), in.data_type(), 1 };
// Simple tensor for the output tile
- SimpleTensor<T> transf_tile{ TensorShape(input_tile_w, input_tile_w), in.data_type(), 1 };
+ SimpleTensor<T> transf_tile{ TensorShape(input_tile_max_dim, input_tile_max_dim), in.data_type(), 1 };
// Initialize matrix for the filter transform
initialize_matrix_transform(trans_matrix, output_tile_size, kernel_size, WinogradTransformType::FILTER);
@@ -300,6 +372,9 @@
const int num_filters = in.shape()[3];
const int num_batches = in.shape().total_size() / (kernel_size.area() * num_channels * num_filters);
+ // If we have a vertical filter (i.e. 1x3, 1x5,..), we need to take the elements along the y direction (step_y_transf_tile = width of the output tile)
+ const int step_y_transf_tile = kernel_size.width == 1 ? input_tile_max_dim : 1;
+
for(int n = 0; n < num_batches; ++n)
{
for(int w = 0; w < num_filters; ++w)
@@ -321,7 +396,7 @@
// Store the values across the channels
for(unsigned int i = 0; i < input_tile_area; ++i)
{
- out[output_offset + i * num_filters * num_channels] = transf_tile[i];
+ out[output_offset + i * num_filters * num_channels] = transf_tile[i * step_y_transf_tile];
}
}
}
@@ -350,15 +425,19 @@
ARM_COMPUTE_ERROR_ON(in.shape()[2] != (in_tile_w * in_tile_h));
ARM_COMPUTE_ERROR_ON(in.shape()[0] != out.shape()[get_data_layout_dimension_index(winograd_info.output_data_layout, DataLayoutDimension::CHANNEL)]);
+ // Get the maximum dimension from the tile size
+ const unsigned int in_tile_max_dim = std::max(in_tile_w, in_tile_h);
+ const unsigned int out_tile_max_dim = std::max(output_tile_size.width, output_tile_size.height);
+
// Compute tile dimensions
// Input tile dimensions
- TensorShape in_tile_dims(in_tile_w, in_tile_h);
+ TensorShape in_tile_dims(in_tile_max_dim, in_tile_max_dim);
// Output tile dimensions
- TensorShape out_tile_dims(output_tile_size.width, output_tile_size.height);
+ TensorShape out_tile_dims(out_tile_max_dim, out_tile_max_dim);
// Transformation matrix dimensions
- TensorShape tr_tile_dims(in_tile_w, output_tile_size.width);
+ TensorShape tr_tile_dims(in_tile_max_dim, out_tile_max_dim);
// Create tensors
// Simple tensor for the input tile
@@ -400,15 +479,24 @@
const int stridez_out = stridey_out * h_out;
const int stridew_out = stridez_out * c_out;
- // Compute number of elements to process in the X and Y direction
- const int num_elements_x = input_dimensions.width - (kernel_size.width - 1) + conv_info.pad_left() + conv_info.pad_right();
- const int num_elements_y = input_dimensions.height - (kernel_size.height - 1) + conv_info.pad_top() + conv_info.pad_bottom();
- const int num_tiles_x = std::ceil(num_elements_x / static_cast<float>(output_tile_size.width));
- const int num_tiles_y = std::ceil(num_elements_y / static_cast<float>(output_tile_size.height));
+ // Compute the number of output tiles along the x and y direction of size "output_tile_size"
+ const Size2D num_tiles = compute_winograd_convolution_tiles(Size2D(input_dimensions.width, input_dimensions.height),
+ kernel_size,
+ output_tile_size,
+ conv_info);
+
+ const int num_tiles_x = num_tiles.width;
+ const int num_tiles_y = num_tiles.height;
ARM_COMPUTE_UNUSED(num_tiles_y);
ARM_COMPUTE_ERROR_ON(in.shape()[1] != static_cast<unsigned int>(num_tiles_x * num_tiles_y));
+ // If we have a vertical filter (i.e. 1x3, 1x5,..), we still need to take the elements along the x direction (step_y_transf_tile = 1)
+ const int step_y_transf_tile = kernel_size.width == 1 ? 1 : output_tile.shape()[0];
+
+ // Initialize with zeros the input tile
+ zeros(input_tile, Coordinates(0, 0), input_tile.shape());
+
for(int n = 0; n < num_batches; ++n)
{
for(int y = 0; y < h_in; ++y)
@@ -441,7 +529,7 @@
// Check out-of-bound writes
if((xo + xi < w_out) && (yo + yi < h_out))
{
- out[output_offset + yi * stridey_out + xi] = output_tile[xi + yi * out_tile_w];
+ out[output_offset + yi * stridey_out + xi] = output_tile[xi + yi * step_y_transf_tile];
// Add bias
out[output_offset + yi * stridey_out + xi] += b[zo];