COMPMID-1580 Implement ReduceMean in NEON
Change-Id: Id974efad304c2513b8824a6561ad45ee60b9e7fb
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/153763
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Reviewed-by: Isabella Gottardi <isabella.gottardi@arm.com>
Tested-by: bsgcomp <bsgcomp@arm.com>
diff --git a/tests/validation/CL/ReductionOperation.cpp b/tests/validation/CL/ReductionOperation.cpp
index 35cb0c5..516a134 100644
--- a/tests/validation/CL/ReductionOperation.cpp
+++ b/tests/validation/CL/ReductionOperation.cpp
@@ -84,7 +84,7 @@
// *INDENT-ON*
template <typename T>
-using CLReductionOperationFixture = ReductionOperationValidationFixture<CLTensor, CLAccessor, CLReductionOperation, T>;
+using CLReductionOperationFixture = ReductionOperationFixture<CLTensor, CLAccessor, CLReductionOperation, T>;
TEST_SUITE(Float)
TEST_SUITE(FP16)
diff --git a/tests/validation/NEON/ReduceMean.cpp b/tests/validation/NEON/ReduceMean.cpp
new file mode 100644
index 0000000..3cd7ce3
--- /dev/null
+++ b/tests/validation/NEON/ReduceMean.cpp
@@ -0,0 +1,176 @@
+/*
+ * Copyright (c) 2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/Types.h"
+#include "arm_compute/runtime/NEON/functions/NEReduceMean.h"
+#include "arm_compute/runtime/Tensor.h"
+#include "arm_compute/runtime/TensorAllocator.h"
+
+#include "tests/NEON/Accessor.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/datasets/SplitDataset.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/ReduceMeanFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+namespace
+{
+constexpr AbsoluteTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for 32-bit floating-point type */
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+constexpr AbsoluteTolerance<float> tolerance_f16(0.03f); /**< Tolerance value for comparing reference's output against implementation's output for 16-bit floating-point type */
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); /**< Tolerance value for comparing reference's output against implementation's output for 8-bit asymmetric quantized type */
+
+const auto axis_keep = combine(framework::dataset::make("Axis", { Coordinates(0), Coordinates(1, 0), Coordinates(1, 2), Coordinates(0, 2), Coordinates(1, 3), Coordinates(0, 1, 2, 3) }),
+ framework::dataset::make("KeepDims", { true }));
+const auto axis_drop = combine(framework::dataset::make("Axis", { Coordinates(0), Coordinates(1), Coordinates(3) }), framework::dataset::make("KeepDims", { false }));
+} // namespace
+TEST_SUITE(NEON)
+TEST_SUITE(ReduceMean)
+
+// *INDENT-OFF*
+// clang-format off
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
+ framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 3U, 16U, 2U), 1, DataType::F32), // Invalid axis
+ TensorInfo(TensorShape(27U, 3U, 16U, 2U), 1, DataType::F32), // Invalid output shape
+ TensorInfo(TensorShape(32U, 16U, 16U, 2U), 1, DataType::F32)
+ }),
+ framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(27U, 3U, 1U, 2U), 1, DataType::F32),
+ TensorInfo(TensorShape(27U, 3U, 1U, 2U), 1, DataType::F32),
+ TensorInfo(TensorShape(32U, 16U, 1U, 2U), 1, DataType::F32)
+ })),
+ framework::dataset::make("Axis", { Coordinates(4), Coordinates(0,2), Coordinates(2) })),
+ framework::dataset::make("Expected", { false, false, true })),
+ input_info, output_info, axis, expected)
+{
+ const Status status = NEReduceMean::validate(&input_info.clone()->set_is_resizable(false), axis, true, &output_info.clone()->set_is_resizable(false));
+ ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS);
+}
+// clang-format on
+// *INDENT-ON*
+
+DATA_TEST_CASE(Configuration,
+ framework::DatasetMode::ALL,
+ combine(datasets::SmallShapes(), framework::dataset::make("DataType", { DataType::F32 })),
+ shape, data_type)
+{
+ // Create tensors
+ Tensor ref_src = create_tensor<Tensor>(shape, data_type);
+ Tensor dst;
+
+ Coordinates axis(1);
+
+ // Create and Configure function
+ NEReduceMean reduce_mean;
+ reduce_mean.configure(&ref_src, axis, true, &dst);
+
+ // Validate valid region
+ TensorShape output_shape = shape;
+ output_shape.set(1, 1);
+ const ValidRegion valid_region = shape_to_valid_region(output_shape);
+ validate(dst.info()->valid_region(), valid_region);
+}
+
+template <typename T>
+using NEReduceMeanFixture = ReduceMeanFixture<Tensor, Accessor, NEReduceMean, T>;
+
+TEST_SUITE(Float)
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall,
+ NEReduceMeanFixture<half>,
+ framework::DatasetMode::PRECOMMIT,
+ combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::F16)), concat(axis_keep, axis_drop)))
+{
+ // Validate output
+ validate(Accessor(_target), _reference, tolerance_f16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge,
+ NEReduceMeanFixture<half>,
+ framework::DatasetMode::NIGHTLY,
+ combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::F16)), concat(axis_keep, axis_drop)))
+{
+ // Validate output
+ validate(Accessor(_target), _reference, tolerance_f16);
+}
+TEST_SUITE_END() // FP16
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall,
+ NEReduceMeanFixture<float>,
+ framework::DatasetMode::PRECOMMIT,
+ combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::F32)), concat(axis_keep, axis_drop)))
+{
+ // Validate output
+ validate(Accessor(_target), _reference, tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge,
+ NEReduceMeanFixture<float>,
+ framework::DatasetMode::NIGHTLY,
+ combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::F32)), concat(axis_keep, axis_drop)))
+{
+ // Validate output
+ validate(Accessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END() // FP32
+TEST_SUITE_END() // Float
+
+template <typename T>
+using NEReduceMeanQuantizedFixture = ReduceMeanQuantizedFixture<Tensor, Accessor, NEReduceMean, T>;
+
+TEST_SUITE(Quantized)
+TEST_SUITE(QASYMM8)
+FIXTURE_DATA_TEST_CASE(RunSmall,
+ NEReduceMeanQuantizedFixture<uint8_t>,
+ framework::DatasetMode::PRECOMMIT,
+ combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), concat(axis_keep, axis_drop)), framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255, 0) })))
+{
+ // Validate output
+ validate(Accessor(_target), _reference, tolerance_qasymm8);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge,
+ NEReduceMeanQuantizedFixture<uint8_t>,
+ framework::DatasetMode::NIGHTLY,
+ combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), concat(axis_keep, axis_drop)), framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255, 0) })))
+{
+ // Validate output
+ validate(Accessor(_target), _reference, tolerance_qasymm8);
+}
+TEST_SUITE_END() // QASYMM8
+TEST_SUITE_END() // Quantized
+TEST_SUITE_END() // ReduceMean
+TEST_SUITE_END() // NEON
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/validation/NEON/ReductionOperation.cpp b/tests/validation/NEON/ReductionOperation.cpp
index b0480b0..2a381bf 100644
--- a/tests/validation/NEON/ReductionOperation.cpp
+++ b/tests/validation/NEON/ReductionOperation.cpp
@@ -45,6 +45,8 @@
{
/** Tolerance for float operations */
RelativeTolerance<float> tolerance_f32(0.00001f);
+/** Tolerance for quantized operations */
+RelativeTolerance<float> tolerance_qasymm8(1);
} // namespace
TEST_SUITE(NEON)
@@ -81,25 +83,47 @@
// *INDENT-ON*
template <typename T>
-using NEReductionOperationFixture = ReductionOperationValidationFixture<Tensor, Accessor, NEReductionOperation, T>;
+using NEReductionOperationFixture = ReductionOperationFixture<Tensor, Accessor, NEReductionOperation, T>;
TEST_SUITE(FP32)
FIXTURE_DATA_TEST_CASE(RunSmall, NEReductionOperationFixture<float>, framework::DatasetMode::PRECOMMIT,
- combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("Axis", { 0 })), framework::dataset::make("Op", { ReductionOperation::SUM_SQUARE })))
+ combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), datasets::ReductionOperations()))
{
// Validate output
validate(Accessor(_target), _reference, tolerance_f32);
}
FIXTURE_DATA_TEST_CASE(RunLarge, NEReductionOperationFixture<float>, framework::DatasetMode::NIGHTLY,
- combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("Axis", { 0 })), framework::dataset::make("Op", { ReductionOperation::SUM_SQUARE })))
+ combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), datasets::ReductionOperations()))
{
// Validate output
validate(Accessor(_target), _reference, tolerance_f32);
}
-TEST_SUITE_END()
+TEST_SUITE_END() // FP32
-TEST_SUITE_END()
-TEST_SUITE_END()
+template <typename T>
+using NEReductionOperationQuantizedFixture = ReductionOperationQuantizedFixture<Tensor, Accessor, NEReductionOperation, T>;
+
+TEST_SUITE(QASYMM8)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEReductionOperationQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT,
+ combine(combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("Axis", { 0, 1, 2, 3 })),
+ datasets::ReductionOperations()),
+ framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255, 0) })))
+{
+ // Validate output
+ validate(Accessor(_target), _reference, tolerance_qasymm8);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEReductionOperationQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY,
+ combine(combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("Axis", { 0, 1, 2, 3 })),
+ datasets::ReductionOperations()),
+ framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255, 0) })))
+{
+ // Validate output
+ validate(Accessor(_target), _reference, tolerance_qasymm8);
+}
+TEST_SUITE_END() // QASYMM8
+
+TEST_SUITE_END() // ReductionOperation
+TEST_SUITE_END() // NEON
} // namespace validation
} // namespace test
} // namespace arm_compute
diff --git a/tests/validation/fixtures/ReductionOperationFixture.h b/tests/validation/fixtures/ReductionOperationFixture.h
index 0dee7eb..9079b47 100644
--- a/tests/validation/fixtures/ReductionOperationFixture.h
+++ b/tests/validation/fixtures/ReductionOperationFixture.h
@@ -45,26 +45,36 @@
{
public:
template <typename...>
- void setup(TensorShape shape, DataType data_type, unsigned int axis, ReductionOperation op)
+ void setup(TensorShape shape, DataType data_type, unsigned int axis, ReductionOperation op, QuantizationInfo quantization_info)
{
const TensorShape output_shape = get_output_shape(shape, axis);
- _target = compute_target(shape, output_shape, data_type, axis, op);
- _reference = compute_reference(shape, output_shape, data_type, axis, op);
+ _target = compute_target(shape, output_shape, data_type, axis, op, quantization_info);
+ _reference = compute_reference(shape, output_shape, data_type, axis, op, quantization_info);
}
protected:
template <typename U>
void fill(U &&tensor)
{
- std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
- library->fill(tensor, distribution, 0);
+ if(!is_data_type_quantized(tensor.data_type()))
+ {
+ std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+ library->fill(tensor, distribution, 0);
+ }
+ else
+ {
+ std::pair<int, int> bounds = get_quantized_bounds(tensor.quantization_info(), -1.0f, 1.0f);
+ std::uniform_int_distribution<uint8_t> distribution(bounds.first, bounds.second);
+
+ library->fill(tensor, distribution, 0);
+ }
}
- TensorType compute_target(const TensorShape &src_shape, const TensorShape &dst_shape, DataType data_type, unsigned int axis, ReductionOperation op)
+ TensorType compute_target(const TensorShape &src_shape, const TensorShape &dst_shape, DataType data_type, unsigned int axis, ReductionOperation op, QuantizationInfo quantization_info)
{
// Create tensors
- TensorType src = create_tensor<TensorType>(src_shape, data_type);
- TensorType dst = create_tensor<TensorType>(dst_shape, data_type);
+ TensorType src = create_tensor<TensorType>(src_shape, data_type, 1, quantization_info);
+ TensorType dst = create_tensor<TensorType>(dst_shape, data_type, 1, quantization_info);
// Create and configure function
FunctionType reduction_func;
@@ -89,10 +99,10 @@
return dst;
}
- SimpleTensor<T> compute_reference(const TensorShape &src_shape, const TensorShape &dst_shape, DataType data_type, unsigned int axis, ReductionOperation op)
+ SimpleTensor<T> compute_reference(const TensorShape &src_shape, const TensorShape &dst_shape, DataType data_type, unsigned int axis, ReductionOperation op, QuantizationInfo quantization_info)
{
// Create reference
- SimpleTensor<T> src{ src_shape, data_type };
+ SimpleTensor<T> src{ src_shape, data_type, 1, quantization_info };
// Fill reference
fill(src);
@@ -111,6 +121,28 @@
return output_shape;
}
};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ReductionOperationQuantizedFixture : public ReductionOperationValidationFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+ template <typename...>
+ void setup(TensorShape shape, DataType data_type, unsigned int axis, ReductionOperation op, QuantizationInfo quantization_info = QuantizationInfo())
+ {
+ ReductionOperationValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, axis, op, quantization_info);
+ }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class ReductionOperationFixture : public ReductionOperationValidationFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+ template <typename...>
+ void setup(TensorShape shape, DataType data_type, unsigned int axis, ReductionOperation op)
+ {
+ ReductionOperationValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, axis, op, QuantizationInfo());
+ }
+};
} // namespace validation
} // namespace test
} // namespace arm_compute
diff --git a/tests/validation/reference/ReductionOperation.cpp b/tests/validation/reference/ReductionOperation.cpp
index 11947bd..499263f 100644
--- a/tests/validation/reference/ReductionOperation.cpp
+++ b/tests/validation/reference/ReductionOperation.cpp
@@ -76,7 +76,7 @@
SimpleTensor<T> reduction_operation(const SimpleTensor<T> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op)
{
// Create reference
- SimpleTensor<T> dst{ dst_shape, src.data_type() };
+ SimpleTensor<T> dst{ dst_shape, src.data_type(), 1, src.quantization_info() };
const unsigned int src_width = src.shape().x();
const unsigned int src_height = src.shape().y();
const unsigned int src_depth = src.shape().z();
@@ -102,7 +102,7 @@
{
res /= src_width;
}
- dst[du] = static_cast<uint8_t>(res);
+ dst[du] = saturate_cast<uint8_t>(res);
}
else
{
@@ -136,7 +136,7 @@
{
res /= src_height;
}
- dst[du * src_width + x] = static_cast<uint8_t>(res);
+ dst[du * src_width + x] = saturate_cast<uint8_t>(res);
}
else
{
@@ -175,7 +175,7 @@
{
res /= src_depth;
}
- dst[du * src_width * src_height + y * src_width + x] = static_cast<uint8_t>(res);
+ dst[du * src_width * src_height + y * src_width + x] = saturate_cast<uint8_t>(res);
}
else
{
@@ -218,7 +218,7 @@
res /= src_batch;
}
- dst[du * src_depth * src_height * src_width + z * src_width * src_height + y * src_width + x] = static_cast<uint8_t>(res);
+ dst[du * src_depth * src_height * src_width + z * src_width * src_height + y * src_width + x] = saturate_cast<uint8_t>(res);
}
else
{