COMPMID-1226 Extend CLMeanStdDev to support FP32 / FP16

- Extend support for FP16 in CLReduction.
- For F16/F32 MeanStdDev we perform one reduction operation for mean
and one for stddev and we calculate the final result in the host CPU.

Change-Id: Iad2099f26c0ba7969737d22f00c6c275634d875c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/135870
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

6 files changed, 84 insertions, 17 deletions

diff --git a/tests/validation/CL/MeanStdDev.cpp b/tests/validation/CL/MeanStdDev.cpp
index 92d87e09f2..8ccb757364 100644
--- a/tests/validation/CL/MeanStdDev.cpp
+++ b/tests/validation/CL/MeanStdDev.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -39,12 +39,17 @@ namespace
 {
 RelativeTolerance<float> tolerance_rel_high_error(0.05f);
 RelativeTolerance<float> tolerance_rel_low_error(0.0005f);
+RelativeTolerance<float> tolerance_rel_high_error_f32(0.001f);
+RelativeTolerance<float> tolerance_rel_low_error_f32(0.00001f);
+RelativeTolerance<float> tolerance_rel_high_error_f16(0.1f);
+RelativeTolerance<float> tolerance_rel_low_error_f16(0.01f);
 } // namespace
 
 TEST_SUITE(CL)
 TEST_SUITE(MeanStdDev)
 
-DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(concat(datasets::Small2DShapes(), datasets::Large2DShapes()), framework::dataset::make("DataType", DataType::U8)), shape, data_type)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(concat(datasets::Small2DShapes(), datasets::Large2DShapes()), framework::dataset::make("DataType", { DataType::U8 })), shape,
+               data_type)
 {
     // Create tensors
     CLTensor src = create_tensor<CLTensor>(shape, data_type);
@@ -71,6 +76,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(concat(datase
 template <typename T>
 using CLMeanStdDevFixture = MeanStdDevValidationFixture<CLTensor, CLAccessor, CLMeanStdDev, T>;
 
+TEST_SUITE(U8)
 FIXTURE_DATA_TEST_CASE(RunSmall, CLMeanStdDevFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), framework::dataset::make("DataType",
                                                                                                           DataType::U8)))
 {
@@ -89,9 +95,43 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLMeanStdDevFixture<uint8_t>, framework::Datase
     // Validate std_dev output
     validate(_target.second, _reference.second, tolerance_rel_high_error);
 }
+TEST_SUITE_END() // U8
 
-TEST_SUITE_END()
-TEST_SUITE_END()
+TEST_SUITE(F16)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLMeanStdDevFixture<half>, framework::DatasetMode::ALL, combine(datasets::Small2DShapes(), framework::dataset::make("DataType",
+                                                                                                 DataType::F16)))
+{
+    // Validate mean output
+    validate(_target.first, _reference.first, tolerance_rel_low_error_f16);
+
+    // Validate std_dev output
+    validate(_target.second, _reference.second, tolerance_rel_high_error_f16);
+}
+TEST_SUITE_END() // F16
+
+TEST_SUITE(F32)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLMeanStdDevFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), framework::dataset::make("DataType",
+                                                                                                        DataType::F32)))
+{
+    // Validate mean output
+    validate(_target.first, _reference.first, tolerance_rel_low_error_f32);
+
+    // Validate std_dev output
+    validate(_target.second, _reference.second, tolerance_rel_high_error_f32);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLMeanStdDevFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::Large2DShapes(), framework::dataset::make("DataType",
+                                                                                                      DataType::F32)))
+{
+    // Validate mean output
+    validate(_target.first, _reference.first, tolerance_rel_low_error_f32);
+
+    // Validate std_dev output
+    validate(_target.second, _reference.second, tolerance_rel_high_error_f32);
+}
+TEST_SUITE_END() // F32
+
+TEST_SUITE_END() // MeanStdDev
+TEST_SUITE_END() // CL
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/CL/ReductionOperation.cpp b/tests/validation/CL/ReductionOperation.cpp
index a48e2f9d5f..ca0988f955 100644
--- a/tests/validation/CL/ReductionOperation.cpp
+++ b/tests/validation/CL/ReductionOperation.cpp
@@ -45,6 +45,7 @@ namespace
 {
 /** Tolerance for float operations */
 RelativeTolerance<float> tolerance_f32(0.00001f);
+RelativeTolerance<float> tolerance_f16(0.1f);
 } // namespace
 
 TEST_SUITE(CL)
@@ -55,7 +56,7 @@ TEST_SUITE(ReductionOperation)
 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), 1, DataType::S16), // DataType != F32
+                                                     TensorInfo(TensorShape(128U, 64U), 1, DataType::S16), // DataType != F16/F32
                                                      TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), // Axis >= num_max_dimensions
                                                      TensorInfo(TensorShape(128U, 64U), 1, DataType::F32), // Axis > 0
                                                      TensorInfo(TensorShape(128U, 64U), 1, DataType::F32)
@@ -84,9 +85,23 @@ template <typename T>
 using CLReductionOperationFixture = ReductionOperationValidationFixture<CLTensor, CLAccessor, CLReductionOperation, T>;
 
 TEST_SUITE(Float)
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLReductionOperationFixture<half>, framework::DatasetMode::PRECOMMIT,
+                       combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("Axis", { 0 })), datasets::ReductionOperations()))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLReductionOperationFixture<half>, framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("Axis", { 0 })), datasets::ReductionOperations()))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+TEST_SUITE_END() // F16
 TEST_SUITE(FP32)
 FIXTURE_DATA_TEST_CASE(RunSmall, CLReductionOperationFixture<float>, framework::DatasetMode::PRECOMMIT,
-                       combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("Axis", { 0 })), datasets::ReductionOperations()))
+                       combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("Axis", { 0 })), datasets::ReductionOperations()))
 {
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f32);
@@ -97,11 +112,11 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLReductionOperationFixture<float>, framework::
     // Validate output
     validate(CLAccessor(_target), _reference, tolerance_f32);
 }
-TEST_SUITE_END()
-TEST_SUITE_END()
+TEST_SUITE_END() // F32
+TEST_SUITE_END() // Float
 
-TEST_SUITE_END()
-TEST_SUITE_END()
+TEST_SUITE_END() // Reduction
+TEST_SUITE_END() // CL
 } // namespace validation
 } // namespace test
 } // namespace arm_compute
diff --git a/tests/validation/fixtures/MeanStdDevFixture.h b/tests/validation/fixtures/MeanStdDevFixture.h
index 17dfe78dbd..58d4644069 100644
--- a/tests/validation/fixtures/MeanStdDevFixture.h
+++ b/tests/validation/fixtures/MeanStdDevFixture.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -50,7 +50,15 @@ protected:
     template <typename U>
     void fill(U &&tensor)
     {
-        library->fill_tensor_uniform(tensor, 0);
+        if(is_data_type_float(tensor.data_type()))
+        {
+            std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+            library->fill(tensor, distribution, 0);
+        }
+        else
+        {
+            library->fill_tensor_uniform(tensor, 0);
+        }
     }
 
     std::pair<float, float> compute_target(const TensorShape &shape, DataType data_type)
diff --git a/tests/validation/fixtures/ReductionOperationFixture.h b/tests/validation/fixtures/ReductionOperationFixture.h
index 6fa5f0c44f..0dee7eb707 100644
--- a/tests/validation/fixtures/ReductionOperationFixture.h
+++ b/tests/validation/fixtures/ReductionOperationFixture.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -56,7 +56,8 @@ protected:
     template <typename U>
     void fill(U &&tensor)
     {
-        library->fill_tensor_uniform(tensor, 0);
+        std::uniform_real_distribution<> distribution(-1.0f, 1.0f);
+        library->fill(tensor, distribution, 0);
     }
 
     TensorType compute_target(const TensorShape &src_shape, const TensorShape &dst_shape, DataType data_type, unsigned int axis, ReductionOperation op)
diff --git a/tests/validation/reference/MeanStdDev.cpp b/tests/validation/reference/MeanStdDev.cpp
index 4a39b13d56..f48fcb11d2 100644
--- a/tests/validation/reference/MeanStdDev.cpp
+++ b/tests/validation/reference/MeanStdDev.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -51,6 +51,8 @@ std::pair<float, float> mean_and_standard_deviation(const SimpleTensor<T> &in)
 }
 
 template std::pair<float, float> mean_and_standard_deviation(const SimpleTensor<uint8_t> &in);
+template std::pair<float, float> mean_and_standard_deviation(const SimpleTensor<half> &in);
+template std::pair<float, float> mean_and_standard_deviation(const SimpleTensor<float> &in);
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/reference/ReductionOperation.cpp b/tests/validation/reference/ReductionOperation.cpp
index acfcc09cea..871a761b1a 100644
--- a/tests/validation/reference/ReductionOperation.cpp
+++ b/tests/validation/reference/ReductionOperation.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -53,7 +53,7 @@ T reduce_operation(T *ptr, int reduce_elements, ReductionOperation op)
     switch(op)
     {
         case ReductionOperation::SUM_SQUARE:
-            return std::accumulate(ptr, ptr + reduce_elements, 0.f, square<T>());
+            return std::accumulate(ptr, ptr + reduce_elements, static_cast<T>(0), square<T>());
         default:
             ARM_COMPUTE_ERROR("Unsupported reduction operation");
     }
@@ -87,6 +87,7 @@ SimpleTensor<T> reduction_operation(const SimpleTensor<T> &src, const TensorShap
 }
 
 template SimpleTensor<float> reduction_operation(const SimpleTensor<float> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op);
+template SimpleTensor<half> reduction_operation(const SimpleTensor<half> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op);
 } // namespace reference
 } // namespace validation
 } // namespace test