aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp12
-rw-r--r--src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp6
-rw-r--r--tests/validation/CL/Winograd.cpp30
-rw-r--r--tests/validation/NEON/ActivationLayer.cpp5
-rw-r--r--tests/validation/NEON/ConvolutionLayer.cpp21
-rw-r--r--tests/validation/fixtures/WinogradConvolutionLayerFixture.h26
6 files changed, 61 insertions, 39 deletions
diff --git a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp b/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp
index cfd53d7082..50e69a8adf 100644
--- a/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp
+++ b/src/core/NEON/kernels/NEWinogradConvolutionLayerKernel.cpp
@@ -182,7 +182,6 @@ Status validate_arguments_winograd_output_trans(const ITensorInfo *input, const
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output);
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
- ARM_COMPUTE_RETURN_ERROR_ON(winograd_info.output_data_layout != DataLayout::NCHW);
ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) != num_tiles.area());
ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != 3U && kernel_dims.width != 5U), "Winograd output transform only supports 3x3 and 5x5 kernels");
ARM_COMPUTE_RETURN_ERROR_ON_MSG((kernel_dims.width != kernel_dims.height), "Winograd output transform only supports 3x3 and 5x5 kernels");
@@ -529,12 +528,15 @@ void NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, Ker
_num_rows = num_rows;
_num_cols = num_cols;
_num_channels = num_channels;
-
// We don't have the biases buffer at this stage as it hasn't been allocated, we pass in nullptr OutputTransform is only used here to compute the window
OutputTransform output_transform(_output_workspace, _matrix_stride, _matrix_row_stride, nullptr, nullptr, _num_batches, _num_rows, _num_cols, _num_channels);
- Window win;
- auto win_last = output_transform.get_window();
+
+ Window win;
+ auto win_last = output_transform.get_window();
win.set(Window::DimX, Window::Dimension(0, win_last, 1));
+
+ _output_nhwc->info()->set_valid_region(ValidRegion(Coordinates(), _output_nhwc->info()->tensor_shape()));
+
INEKernel::configure(win);
}
@@ -548,7 +550,7 @@ void NEWinogradLayerTransformOutputKernel<T, OutputTileRows, OutputTileCols, Ker
OutputTransform output_transform(_output_workspace, _matrix_stride, _matrix_row_stride,
(_biases ? reinterpret_cast<T *>(_biases->buffer()) : nullptr), reinterpret_cast<T *>(_output_nhwc->buffer()),
- _num_batches, _num_rows, _num_cols, _num_channels);
+ _num_batches, _num_rows, _num_cols, _num_channels, 0, _output_nhwc->info()->strides_in_bytes()[2] / sizeof(T), _output_nhwc->info()->strides_in_bytes()[1] / sizeof(T));
// The code below cannot be moved to configure because biases hasn't been allocated at that point
const size_t fst = window.x().start();
diff --git a/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp b/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp
index d9f6c0e0f8..39175c26c6 100644
--- a/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp
+++ b/src/runtime/NEON/functions/NEWinogradConvolutionLayer.cpp
@@ -61,7 +61,6 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *weights,
ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32);
ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights);
ARM_COMPUTE_RETURN_ERROR_ON_MSG(weights->dimension(width_idx) != 3 && weights->dimension(height_idx) != 5, "Only 3 and 5 kernels are supported");
- ARM_COMPUTE_RETURN_ERROR_ON(data_layout != DataLayout::NCHW); // COMPMID-1287
ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 4);
ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.stride().first != 1 || conv_info.stride().second != 1, "Winograd layer only supports unit strides.");
@@ -325,7 +324,7 @@ void NEWinogradConvolutionLayer::configure(const ITensor *input, const ITensor *
//Configure Activation Layer
_is_activationlayer_enabled = act_info.enabled();
- if(data_layout == DataLayout::NCHW && _is_activationlayer_enabled)
+ if(_is_activationlayer_enabled)
{
_activationlayer_function.configure(_output, nullptr, act_info);
}
@@ -363,6 +362,7 @@ void NEWinogradConvolutionLayer::run()
{
_activationlayer_function.run();
}
+
_memory_group.release();
}
@@ -396,6 +396,7 @@ Status NEWinogradConvolutionLayer::validate(const ITensorInfo *input, const ITen
// Validate input transform
const TensorShape input0_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, winograd_info);
const TensorInfo input0 = input->clone()->set_tensor_shape(input0_shape);
+
switch(weights->dimension(idx_width))
{
case 3:
@@ -482,7 +483,6 @@ Status NEWinogradConvolutionLayer::validate(const ITensorInfo *input, const ITen
break;
}
}
-
// Validate Activation Layer
if(act_info.enabled())
{
diff --git a/tests/validation/CL/Winograd.cpp b/tests/validation/CL/Winograd.cpp
index 59fe25db73..e0b4b5f795 100644
--- a/tests/validation/CL/Winograd.cpp
+++ b/tests/validation/CL/Winograd.cpp
@@ -379,18 +379,20 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(
using CLWinogradConvolutionLayerFastMathFixture = WinogradConvolutionLayerFastMathValidationFixture<CLTensor, CLAccessor, CLWinogradConvolutionLayer, float>;
TEST_SUITE(Conv3x3)
FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::PRECOMMIT,
- combine(combine(datasets::SmallWinogradConvolutionLayer3x3Dataset(),
- framework::dataset::make("DataType", { DataType::F32 })),
- framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() })))
+ combine(combine(combine(datasets::SmallWinogradConvolutionLayer3x3Dataset(),
+ 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(datasets::LargeWinogradConvolutionLayer3x3Dataset(),
- framework::dataset::make("DataType", { DataType::F32 })),
- framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() })))
+ combine(combine(combine(datasets::LargeWinogradConvolutionLayer3x3Dataset(),
+ 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);
@@ -399,18 +401,22 @@ TEST_SUITE_END() // Conv3x3
TEST_SUITE(Conv5x5)
FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::PRECOMMIT,
- combine(combine(datasets::SmallWinogradConvolutionLayer5x5Dataset(),
- framework::dataset::make("DataType", { DataType::F32 })),
- framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() })))
+ combine(combine(combine(datasets::SmallWinogradConvolutionLayer5x5Dataset(),
+ 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(datasets::LargeWinogradConvolutionLayer5x5Dataset(),
- framework::dataset::make("DataType", { DataType::F32 })),
- framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() })))
+ combine(combine(combine(datasets::LargeWinogradConvolutionLayer5x5Dataset(),
+ 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);
diff --git a/tests/validation/NEON/ActivationLayer.cpp b/tests/validation/NEON/ActivationLayer.cpp
index e77fe5a295..868683259c 100644
--- a/tests/validation/NEON/ActivationLayer.cpp
+++ b/tests/validation/NEON/ActivationLayer.cpp
@@ -193,12 +193,13 @@ TEST_SUITE_END()
TEST_SUITE(FP32)
FIXTURE_DATA_TEST_CASE(RunSmall, NEActivationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ActivationDataset), framework::dataset::make("DataType",
DataType::F32)))
+
{
// Validate output
validate(Accessor(_target), _reference, tolerance(_data_type, _function));
}
-FIXTURE_DATA_TEST_CASE(RunLarge, NEActivationLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), ActivationDataset), framework::dataset::make("DataType",
- DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEActivationLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), ActivationDataset),
+ framework::dataset::make("DataType", DataType::F32)))
{
// Validate output
validate(Accessor(_target), _reference, tolerance(_data_type, _function));
diff --git a/tests/validation/NEON/ConvolutionLayer.cpp b/tests/validation/NEON/ConvolutionLayer.cpp
index b286eb61b6..2c17487107 100644
--- a/tests/validation/NEON/ConvolutionLayer.cpp
+++ b/tests/validation/NEON/ConvolutionLayer.cpp
@@ -121,21 +121,26 @@ template <typename T>
using NEWinogradConvolutionLayerNoBiasFixture = WinogradConvolutionLayerFastMathValidationFixture<Tensor, Accessor, NEWinogradConvolutionLayer, T, false>;
TEST_SUITE(FP32)
+
FIXTURE_DATA_TEST_CASE(RunSmall, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT,
- combine(combine(framework::dataset::concat(datasets::SmallWinogradConvolutionLayer3x3Dataset(),
- datasets::SmallWinogradConvolutionLayer5x5Dataset()),
- framework::dataset::make("DataType", { DataType::F32 })),
- ActivationFunctionsDataset))
+ combine(combine(combine(framework::dataset::concat(datasets::SmallWinogradConvolutionLayer3x3Dataset(),
+ datasets::SmallWinogradConvolutionLayer5x5Dataset()),
+ framework::dataset::make("DataType", { DataType::F32 })),
+ ActivationFunctionsDataset),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
+
{
// Validate output
validate(Accessor(_target), _reference, abs_tolerance_f32);
}
FIXTURE_DATA_TEST_CASE(RunSmallNoBias, NEWinogradConvolutionLayerNoBiasFixture<float>, framework::DatasetMode::PRECOMMIT,
- combine(combine(framework::dataset::concat(datasets::SmallWinogradConvolutionLayer3x3Dataset(),
- datasets::SmallWinogradConvolutionLayer5x5Dataset()),
- framework::dataset::make("DataType", { DataType::F32 })),
- ActivationFunctionsDataset))
+ combine(combine(combine(framework::dataset::concat(datasets::SmallWinogradConvolutionLayer3x3Dataset(),
+ datasets::SmallWinogradConvolutionLayer5x5Dataset()),
+ framework::dataset::make("DataType", { DataType::F32 })),
+ ActivationFunctionsDataset),
+
+ framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })))
{
// Validate output
validate(Accessor(_target), _reference, abs_tolerance_f32);
diff --git a/tests/validation/fixtures/WinogradConvolutionLayerFixture.h b/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
index 07795c2361..aca24f13ae 100644
--- a/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
+++ b/tests/validation/fixtures/WinogradConvolutionLayerFixture.h
@@ -160,11 +160,12 @@ class WinogradConvolutionLayerFastMathValidationFixture : public framework::Fixt
{
public:
template <typename...>
- void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, DataType data_type, ActivationLayerInfo act_info)
+ void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation,
+ DataType data_type, ActivationLayerInfo act_info, const DataLayout &data_layout)
+
{
ARM_COMPUTE_UNUSED(dilation);
-
- _target = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, data_type, act_info);
+ _target = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, data_type, act_info, data_layout);
_reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, data_type, act_info);
}
@@ -189,14 +190,21 @@ protected:
}
}
- TensorType compute_target(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const PadStrideInfo &info,
- DataType data_type, ActivationLayerInfo act_info)
+ TensorType compute_target(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, const PadStrideInfo &info,
+ DataType data_type, ActivationLayerInfo act_info, const DataLayout data_layout)
{
+ if(data_layout == DataLayout::NHWC)
+ {
+ permute(input_shape, PermutationVector(2U, 0U, 1U));
+ permute(weights_shape, PermutationVector(2U, 0U, 1U));
+ permute(output_shape, PermutationVector(2U, 0U, 1U));
+ }
+
// Create tensors
- TensorType src = create_tensor<TensorType>(input_shape, data_type, 1);
- TensorType weights = create_tensor<TensorType>(weights_shape, data_type, 1);
- TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1);
- TensorType dst = create_tensor<TensorType>(output_shape, data_type, 1);
+ TensorType src = create_tensor<TensorType>(input_shape, data_type, 1, 0, QuantizationInfo(), data_layout);
+ TensorType weights = create_tensor<TensorType>(weights_shape, data_type, 1, 0, QuantizationInfo(), data_layout);
+ TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1, 0, QuantizationInfo(), data_layout);
+ TensorType dst = create_tensor<TensorType>(output_shape, data_type, 1, 0, QuantizationInfo(), data_layout);
// Create and configure function
FunctionType conv;
generated by cgit v1.2.1 (git 2.23.0) at 2024-05-17 10:36:39 +0000