COMPMID-1959: Implements 2D FFT on OpenCL

Change-Id: I73cf3984a5463acc854c8a59dc2bd9a5234cd99c
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/936
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

5 files changed, 247 insertions, 16 deletions

diff --git a/tests/validation/CL/ConvolutionLayer.cpp b/tests/validation/CL/ConvolutionLayer.cpp
index 41d2b7bb5e..f1f9b59330 100644
--- a/tests/validation/CL/ConvolutionLayer.cpp
+++ b/tests/validation/CL/ConvolutionLayer.cpp
@@ -46,7 +46,7 @@ namespace validation
 namespace
 {
 constexpr AbsoluteTolerance<float>  absolute_tolerance_float(0.0001f);    /**< Absolute Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
-RelativeTolerance<float>            tolerance_f32(0.05f);                 /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
+RelativeTolerance<float>            tolerance_f32(0.1f);                  /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */
 RelativeTolerance<half_float::half> tolerance_f16(half_float::half(0.2)); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */
 constexpr AbsoluteTolerance<float>  tolerance_qasymm8(1);                 /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */
 constexpr float                     tolerance_num = 0.07f;                /**< Tolerance number */
diff --git a/tests/validation/CL/FFT.cpp b/tests/validation/CL/FFT.cpp
index 0d29532c29..9fdd85b604 100644
--- a/tests/validation/CL/FFT.cpp
+++ b/tests/validation/CL/FFT.cpp
@@ -24,7 +24,10 @@
 #include "arm_compute/core/Types.h"
 #include "arm_compute/runtime/CL/CLTensor.h"
 #include "arm_compute/runtime/CL/functions/CLFFT1D.h"
+#include "arm_compute/runtime/CL/functions/CLFFT2D.h"
+#include "arm_compute/runtime/CL/functions/CLFFTConvolutionLayer.h"
 #include "tests/CL/CLAccessor.h"
+#include "tests/datasets/SmallConvolutionLayerDataset.h"
 #include "tests/framework/Asserts.h"
 #include "tests/framework/Macros.h"
 #include "tests/framework/datasets/Datasets.h"
@@ -40,7 +43,7 @@ namespace validation
 namespace
 {
 const auto data_types = framework::dataset::make("DataType", { DataType::F32 });
-const auto shapes     = framework::dataset::make("TensorShape", { TensorShape(2U, 2U, 3U), TensorShape(3U, 2U, 3U),
+const auto shapes_1d  = framework::dataset::make("TensorShape", { TensorShape(2U, 2U, 3U), TensorShape(3U, 2U, 3U),
                                                                   TensorShape(4U, 2U, 3U), TensorShape(5U, 2U, 3U),
                                                                   TensorShape(7U, 2U, 3U), TensorShape(8U, 2U, 3U),
                                                                   TensorShape(9U, 2U, 3U), TensorShape(25U, 2U, 3U),
@@ -48,11 +51,27 @@ const auto shapes     = framework::dataset::make("TensorShape", { TensorShape(2U
                                                                   TensorShape(16U, 2U, 3U), TensorShape(32U, 2U, 3U),
                                                                   TensorShape(96U, 2U, 2U)
                                                                 });
+const auto shapes_2d = framework::dataset::make("TensorShape", { TensorShape(2U, 2U, 3U), TensorShape(3U, 6U, 3U),
+                                                                 TensorShape(4U, 5U, 3U), TensorShape(5U, 7U, 3U),
+                                                                 TensorShape(7U, 25U, 3U), TensorShape(8U, 2U, 3U),
+                                                                 TensorShape(9U, 16U, 3U), TensorShape(25U, 32U, 3U),
+                                                                 TensorShape(192U, 128U, 2U)
+                                                               });
+
+const auto ActivationFunctionsSmallDataset = framework::dataset::make("ActivationInfo",
+{
+    ActivationLayerInfo(),
+    ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.5f)
+});
+
+RelativeTolerance<float> tolerance_f32(0.1f);   /**< Relative tolerance value for FP32 */
+constexpr float          tolerance_num = 0.07f; /**< Tolerance number */
+
 } // namespace
 TEST_SUITE(CL)
 TEST_SUITE(FFT1D)
 
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(shapes, data_types),
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(shapes_1d, data_types),
                shape, data_type)
 {
     // Create tensors
@@ -81,19 +100,19 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(shapes, data_
 DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
         framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 2, DataType::F32), // Mismatching data types
                                                 TensorInfo(TensorShape(32U, 13U, 2U), 2, DataType::F32), // Mismatching shapes
-                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Invalid channels
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 3, DataType::F32), // Invalid channels
                                                 TensorInfo(TensorShape(32U, 13U, 2U), 2, DataType::F32), // Unsupported axis
                                                 TensorInfo(TensorShape(11U, 13U, 2U), 2, DataType::F32), // Undecomposable FFT
                                                 TensorInfo(TensorShape(25U, 13U, 2U), 2, DataType::F32),
         }),
         framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 2, DataType::F16),
                                                 TensorInfo(TensorShape(16U, 13U, 2U), 2, DataType::F32),
-                                                TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 2, DataType::F32),
                                                 TensorInfo(TensorShape(32U, 13U, 2U), 2, DataType::F32),
                                                 TensorInfo(TensorShape(11U, 13U, 2U), 2, DataType::F32),
                                                 TensorInfo(TensorShape(25U, 13U, 2U), 2, DataType::F32),
         })),
-        framework::dataset::make("Axis", { 0, 0, 0, 1, 0, 0 })),
+        framework::dataset::make("Axis", { 0, 0, 0, 2, 0, 0 })),
         framework::dataset::make("Expected", { false, false, false, false, false, true })),
         input_info, output_info, axis, expected)
 {
@@ -106,19 +125,103 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
 // *INDENT-ON*
 
 template <typename T>
-using CLFFT1DFixture = FFTValidationFixture<CLTensor, CLAccessor, CLFFT1D, T>;
+using CLFFT1DFixture = FFTValidationFixture<CLTensor, CLAccessor, CLFFT1D, FFT1DInfo, T>;
 
 TEST_SUITE(Float)
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLFFT1DFixture<float>, framework::DatasetMode::ALL, combine(shapes, framework::dataset::make("DataType", DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLFFT1DFixture<float>, framework::DatasetMode::ALL, combine(shapes_1d, framework::dataset::make("DataType", DataType::F32)))
 {
     // Validate output
-    validate(CLAccessor(_target), _reference, RelativeTolerance<float>(0.1f), 0.05f);
+    validate(CLAccessor(_target), _reference, tolerance_f32, tolerance_num);
 }
 TEST_SUITE_END() // FP32
 TEST_SUITE_END() // Float
 
 TEST_SUITE_END() // FFT1D
+
+TEST_SUITE(FFT2D)
+
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(shapes_2d, data_types),
+               shape, data_type)
+{
+    // Create tensors
+    CLTensor src = create_tensor<CLTensor>(shape, data_type, 2);
+    CLTensor dst = create_tensor<CLTensor>(shape, data_type, 2);
+
+    ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
+    ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+    // Create and configure function
+    CLFFT2D fft2d;
+    fft2d.configure(&src, &dst, FFT2DInfo());
+
+    // Validate valid region
+    const ValidRegion valid_region = shape_to_valid_region(shape);
+    validate(src.info()->valid_region(), valid_region);
+    validate(dst.info()->valid_region(), valid_region);
+
+    // Validate padding
+    validate(src.info()->padding(), PaddingSize());
+    validate(dst.info()->padding(), PaddingSize());
+}
+
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(
+        framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 25U, 2U), 2, DataType::F32), // Mismatching data types
+                                                TensorInfo(TensorShape(32U, 25U, 2U), 2, DataType::F32), // Mismatching shapes
+                                                TensorInfo(TensorShape(32U, 25U, 2U), 3, DataType::F32), // Invalid channels
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 2, DataType::F32), // Undecomposable FFT
+                                                TensorInfo(TensorShape(32U, 25U, 2U), 2, DataType::F32),
+        }),
+        framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 25U, 2U), 2, DataType::F16),
+                                                TensorInfo(TensorShape(16U, 25U, 2U), 2, DataType::F32),
+                                                TensorInfo(TensorShape(32U, 25U, 2U), 1, DataType::F32),
+                                                TensorInfo(TensorShape(32U, 13U, 2U), 2, DataType::F32),
+                                                TensorInfo(TensorShape(32U, 25U, 2U), 2, DataType::F32),
+        })),
+        framework::dataset::make("Expected", { false, false, false, false, true })),
+               input_info, output_info, expected)
+{
+    const Status s = CLFFT2D::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), FFT2DInfo());
+    ARM_COMPUTE_EXPECT(bool(s) == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
+template <typename T>
+using CLFFT2DFixture = FFTValidationFixture<CLTensor, CLAccessor, CLFFT2D, FFT2DInfo, T>;
+
+TEST_SUITE(Float)
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLFFT2DFixture<float>, framework::DatasetMode::ALL, combine(shapes_2d, framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32, tolerance_num);
+}
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // Float
+TEST_SUITE_END() // FFT2D
+
+TEST_SUITE(FFTConvolutionLayer)
+
+template <typename T>
+using CLFFTConvolutionLayerFixture = FFTConvolutionValidationFixture<CLTensor, CLAccessor, CLFFTConvolutionLayer, T>;
+
+TEST_SUITE(Float)
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLFFTConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallFFTConvolutionLayerDataset(),
+                                                                                                                 framework::dataset::make("DataType", DataType::F32)),
+                                                                                                                 framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })),
+                                                                                                                 ActivationFunctionsSmallDataset))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32, tolerance_num);
+}
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // Float
+TEST_SUITE_END() // FFTConvolutionLayer
+
 TEST_SUITE_END() // CL
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/CL/ReductionOperation.cpp b/tests/validation/CL/ReductionOperation.cpp
index c8474e97e6..79308c8229 100644
--- a/tests/validation/CL/ReductionOperation.cpp
+++ b/tests/validation/CL/ReductionOperation.cpp
@@ -63,7 +63,7 @@ TEST_SUITE(ReductionOperation)
 // clang-format off
 DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
     framework::dataset::make("InputInfo",          { TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), // Mismatching data type input/output
-                                                     TensorInfo(TensorShape(128U, 64U), 2, DataType::F32), // Number of Input channels != 1
+                                                     TensorInfo(TensorShape(128U, 64U), 3, DataType::F32), // Number of Input channels != 1
                                                      TensorInfo(TensorShape(128U, 64U), 1, DataType::S16), // DataType != QASYMM8/F16/F32
                                                      TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), // Axis >= num_max_dimensions
                                                      TensorInfo(TensorShape(128U, 64U), 1, DataType::QASYMM8), // Axis == 0 and SUM_SQUARE and QASYMM8
diff --git a/tests/validation/fixtures/ConvolutionLayerFixture.h b/tests/validation/fixtures/ConvolutionLayerFixture.h
index 795b9de6cd..52fa8da60b 100644
--- a/tests/validation/fixtures/ConvolutionLayerFixture.h
+++ b/tests/validation/fixtures/ConvolutionLayerFixture.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
diff --git a/tests/validation/fixtures/FFTFixture.h b/tests/validation/fixtures/FFTFixture.h
index 8e3c01eaff..1aaa5965b2 100644
--- a/tests/validation/fixtures/FFTFixture.h
+++ b/tests/validation/fixtures/FFTFixture.h
@@ -31,6 +31,8 @@
 #include "tests/IAccessor.h"
 #include "tests/framework/Asserts.h"
 #include "tests/framework/Fixture.h"
+#include "tests/validation/reference/ActivationLayer.h"
+#include "tests/validation/reference/ConvolutionLayer.h"
 #include "tests/validation/reference/DFT.h"
 
 #include <random>
@@ -41,7 +43,7 @@ namespace test
 {
 namespace validation
 {
-template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+template <typename TensorType, typename AccessorType, typename FunctionType, typename InfoType, typename T>
 class FFTValidationFixture : public framework::Fixture
 {
 public:
@@ -68,8 +70,8 @@ protected:
         TensorType dst = create_tensor<TensorType>(shape, data_type, 2);
 
         // Create and configure function
-        FunctionType fft1d;
-        fft1d.configure(&src, &dst, FFT1DInfo());
+        FunctionType fft;
+        fft.configure(&src, &dst, InfoType());
 
         ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
         ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -85,7 +87,7 @@ protected:
         fill(AccessorType(src));
 
         // Compute function
-        fft1d.run();
+        fft.run();
 
         return dst;
     }
@@ -97,12 +99,138 @@ protected:
 
         // Fill reference
         fill(src);
+        if(std::is_same<InfoType, FFT1DInfo>::value)
+        {
+            return reference::dft_1d(src, reference::FFTDirection::Forward);
+        }
+        else
+        {
+            return reference::dft_2d(src, reference::FFTDirection::Forward);
+        }
+    }
+
+    TensorType      _target{};
+    SimpleTensor<T> _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class FFTConvolutionValidationGenericFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation,
+               DataType data_type, DataLayout data_layout, ActivationLayerInfo act_info)
+    {
+        _data_type   = data_type;
+        _data_layout = data_layout;
+
+        _target    = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, dilation, act_info);
+        _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, dilation, act_info);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        switch(tensor.data_type())
+        {
+            case DataType::F32:
+            {
+                std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+                library->fill(tensor, distribution, i);
+                break;
+            }
+            default:
+                library->fill_tensor_uniform(tensor, i);
+        }
+    }
+
+    TensorType compute_target(TensorShape input_shape, TensorShape weights_shape, const TensorShape &bias_shape, TensorShape output_shape, const PadStrideInfo &info,
+                              const Size2D &dilation, const ActivationLayerInfo act_info)
+    {
+        ARM_COMPUTE_UNUSED(dilation);
+        ARM_COMPUTE_ERROR_ON((input_shape[2] % weights_shape[2]) != 0);
+
+        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, QuantizationInfo(), _data_layout);
+        TensorType weights = create_tensor<TensorType>(weights_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+        TensorType bias    = create_tensor<TensorType>(bias_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+        TensorType dst     = create_tensor<TensorType>(output_shape, _data_type, 1, QuantizationInfo(), _data_layout);
+
+        // Create and configure function
+        FunctionType conv;
+        conv.configure(&src, &weights, &bias, &dst, info, act_info);
+
+        ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Allocate tensors
+        src.allocator()->allocate();
+        weights.allocator()->allocate();
+        bias.allocator()->allocate();
+        dst.allocator()->allocate();
+
+        ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!weights.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS);
+        ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS);
+
+        // Fill tensors
+        fill(AccessorType(src), 0);
+        fill(AccessorType(weights), 1);
+        fill(AccessorType(bias), 2);
+
+        // Compute convolution function
+        conv.run();
+
+        return dst;
+    }
 
-        return reference::dft_1d(src, reference::FFTDirection::Forward);
+    SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const PadStrideInfo &info,
+                                      const Size2D &dilation, const ActivationLayerInfo act_info)
+    {
+        ARM_COMPUTE_ERROR_ON((input_shape[2] % weights_shape[2]) != 0);
+
+        // Create reference
+        SimpleTensor<T> src{ input_shape, _data_type, 1 };
+        SimpleTensor<T> weights{ weights_shape, _data_type, 1 };
+        SimpleTensor<T> bias{ bias_shape, _data_type, 1 };
+
+        // Fill reference
+        fill(src, 0);
+        fill(weights, 1);
+        fill(bias, 2);
+
+        return (act_info.enabled()) ? reference::activation_layer<T>(reference::convolution_layer<T>(src, weights, bias, output_shape, info, dilation), act_info) : reference::convolution_layer<T>(src,
+                weights, bias, output_shape, info, dilation);
     }
 
     TensorType      _target{};
     SimpleTensor<T> _reference{};
+    DataType        _data_type{};
+    DataLayout      _data_layout{};
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class FFTConvolutionValidationFixture : public FFTConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation,
+               DataType data_type, DataLayout data_layout, ActivationLayerInfo act_info)
+    {
+        FFTConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, dilation,
+                                                                                                 data_type, data_layout, act_info);
+    }
 };
 } // namespace validation
 } // namespace test