COMPMID-1303: CLDepthConvert : Add support for FP32 -> FP16 and FP16 -> FP32 + validate() function

Change-Id: I6808de0254a7c4bca440322cc14b795b3b32465b
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/142427
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

3 files changed, 123 insertions, 20 deletions

diff --git a/tests/validation/CL/DepthConvertLayer.cpp b/tests/validation/CL/DepthConvertLayer.cpp
index ed1f54ca6e..fe46313568 100644
--- a/tests/validation/CL/DepthConvertLayer.cpp
+++ b/tests/validation/CL/DepthConvertLayer.cpp
@@ -44,15 +44,16 @@ namespace validation
 namespace
 {
 /** Input data sets **/
-const auto DepthConvertLayerU8toU16Dataset             = combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::U16));
-const auto DepthConvertLayerU8toS16Dataset             = combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::S16));
-const auto DepthConvertLayerU8toS32Dataset             = combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::S32));
-const auto DepthConvertLayerU16toU8Dataset             = combine(framework::dataset::make("DataType", DataType::U16), framework::dataset::make("DataType", DataType::U8));
-const auto DepthConvertLayerU16toU32Dataset            = combine(framework::dataset::make("DataType", DataType::U16), framework::dataset::make("DataType", DataType::U32));
-const auto DepthConvertLayerS16toU8Dataset             = combine(framework::dataset::make("DataType", DataType::S16), framework::dataset::make("DataType", DataType::U8));
-const auto DepthConvertLayerS16toS32Dataset            = combine(framework::dataset::make("DataType", DataType::S16), framework::dataset::make("DataType", DataType::S32));
-const auto DepthConvertLayerShiftDataset               = framework::dataset::make("Shift", 0, 7);
-const auto DepthConvertLayerFixedPointQuantizedDataset = framework::dataset::make("FractionalBits", 1, 7);
+const auto DepthConvertLayerU8toU16Dataset  = combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::U16));
+const auto DepthConvertLayerU8toS16Dataset  = combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::S16));
+const auto DepthConvertLayerU8toS32Dataset  = combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::S32));
+const auto DepthConvertLayerU16toU8Dataset  = combine(framework::dataset::make("DataType", DataType::U16), framework::dataset::make("DataType", DataType::U8));
+const auto DepthConvertLayerU16toU32Dataset = combine(framework::dataset::make("DataType", DataType::U16), framework::dataset::make("DataType", DataType::U32));
+const auto DepthConvertLayerS16toU8Dataset  = combine(framework::dataset::make("DataType", DataType::S16), framework::dataset::make("DataType", DataType::U8));
+const auto DepthConvertLayerS16toS32Dataset = combine(framework::dataset::make("DataType", DataType::S16), framework::dataset::make("DataType", DataType::S32));
+const auto DepthConvertLayerF16toF32Dataset = combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F32));
+const auto DepthConvertLayerF32toF16Dataset = combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::F16));
+const auto DepthConvertLayerShiftDataset    = framework::dataset::make("Shift", 0, 7);
 } // namespace
 
 TEST_SUITE(CL)
@@ -67,6 +68,10 @@ template <typename T>
 using CLDepthConvertLayerToU8Fixture = DepthConvertLayerValidationFixture<CLTensor, CLAccessor, CLDepthConvertLayer, T, uint8_t>;
 template <typename T>
 using CLDepthConvertLayerToU32Fixture = DepthConvertLayerValidationFixture<CLTensor, CLAccessor, CLDepthConvertLayer, T, uint32_t>;
+template <typename T>
+using CLDepthConvertLayerToF16Fixture = DepthConvertLayerValidationFixture<CLTensor, CLAccessor, CLDepthConvertLayer, T, half>;
+template <typename T>
+using CLDepthConvertLayerToF32Fixture = DepthConvertLayerValidationFixture<CLTensor, CLAccessor, CLDepthConvertLayer, T, float>;
 
 TEST_SUITE(U8_to_U16)
 DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
@@ -343,6 +348,84 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthConvertLayerToS32Fixture<int16_t>, frame
 }
 TEST_SUITE_END()
 
+TEST_SUITE(F16_to_F32)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+                                                                   DepthConvertLayerShiftDataset),
+               shape, policy, shift)
+{
+    // Create tensors
+    CLTensor src = create_tensor<CLTensor>(shape, DataType::F16, 1);
+    CLTensor dst = create_tensor<CLTensor>(shape, DataType::F32, 1);
+
+    // Create and Configure function
+    CLDepthConvertLayer depth_convert;
+    depth_convert.configure(&src, &dst, policy, shift);
+
+    // Validate valid region
+    const ValidRegion valid_region = shape_to_valid_region(shape);
+    validate(dst.info()->valid_region(), valid_region);
+
+    // Validate padding
+    const PaddingSize padding = PaddingCalculator(shape.x(), 16).required_padding();
+    validate(src.info()->padding(), padding);
+    validate(dst.info()->padding(), padding);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthConvertLayerToF32Fixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), DepthConvertLayerF16toF32Dataset),
+                                                                                                                   framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+                                                                                                                   DepthConvertLayerShiftDataset))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthConvertLayerToF32Fixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), DepthConvertLayerF16toF32Dataset),
+                                                                                                                 framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+                                                                                                                 DepthConvertLayerShiftDataset))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END()
+
+TEST_SUITE(F32_to_F16)
+DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(framework::dataset::concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+                                                                   DepthConvertLayerShiftDataset),
+               shape, policy, shift)
+{
+    // Create tensors
+    CLTensor src = create_tensor<CLTensor>(shape, DataType::F32, 1);
+    CLTensor dst = create_tensor<CLTensor>(shape, DataType::F16, 1);
+
+    // Create and Configure function
+    CLDepthConvertLayer depth_convert;
+    depth_convert.configure(&src, &dst, policy, shift);
+
+    // Validate valid region
+    const ValidRegion valid_region = shape_to_valid_region(shape);
+    validate(dst.info()->valid_region(), valid_region);
+
+    // Validate padding
+    const PaddingSize padding = PaddingCalculator(shape.x(), 16).required_padding();
+    validate(src.info()->padding(), padding);
+    validate(dst.info()->padding(), padding);
+}
+
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthConvertLayerToF16Fixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), DepthConvertLayerF32toF16Dataset),
+                                                                                                                    framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+                                                                                                                    DepthConvertLayerShiftDataset))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthConvertLayerToF16Fixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), DepthConvertLayerF32toF16Dataset),
+                                                                                                                  framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })),
+                                                                                                                  DepthConvertLayerShiftDataset))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference);
+}
+TEST_SUITE_END()
+
 TEST_SUITE_END()
 TEST_SUITE_END()
 } // namespace validation
diff --git a/tests/validation/reference/DepthConvertLayer.cpp b/tests/validation/reference/DepthConvertLayer.cpp
index 6f90963360..fd2e0ae378 100644
--- a/tests/validation/reference/DepthConvertLayer.cpp
+++ b/tests/validation/reference/DepthConvertLayer.cpp
@@ -60,6 +60,33 @@ SimpleTensor<T2> depth_convert(const SimpleTensor<T1> &src, DataType dt_out, Con
     return result;
 }
 
+template < typename T1, typename T2, typename std::enable_if < is_floating_point<T1>::value &&is_floating_point<T2>::value &&!std::is_same<T1, T2>::value, int >::type >
+SimpleTensor<T2> depth_convert(const SimpleTensor<T1> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift)
+{
+    SimpleTensor<T2> result(src.shape(), dt_out);
+
+    const uint32_t scale = 1 << shift;
+
+    // Up-casting
+    if(src.data_type() <= dt_out)
+    {
+        for(int i = 0; i < src.num_elements(); ++i)
+        {
+            result[i] = src[i] * static_cast<T2>(scale);
+        }
+    }
+    // Down-casting
+    else
+    {
+        for(int i = 0; i < src.num_elements(); ++i)
+        {
+            T1 val    = src[i] / static_cast<T1>(scale);
+            result[i] = (policy == ConvertPolicy::SATURATE) ? saturate_cast<T2>(val) : static_cast<T2>(val);
+        }
+    }
+    return result;
+}
+
 template SimpleTensor<uint16_t> depth_convert(const SimpleTensor<uint8_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
 template SimpleTensor<int16_t> depth_convert(const SimpleTensor<uint8_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
 template SimpleTensor<int32_t> depth_convert(const SimpleTensor<uint8_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
@@ -67,6 +94,8 @@ template SimpleTensor<uint8_t> depth_convert(const SimpleTensor<uint16_t> &src,
 template SimpleTensor<uint32_t> depth_convert(const SimpleTensor<uint16_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
 template SimpleTensor<uint8_t> depth_convert(const SimpleTensor<int16_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
 template SimpleTensor<int32_t> depth_convert(const SimpleTensor<int16_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
+template SimpleTensor<half> depth_convert(const SimpleTensor<float> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
+template SimpleTensor<float> depth_convert(const SimpleTensor<half> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
 } // namespace reference
 } // namespace validation
 } // namespace test
diff --git a/tests/validation/reference/DepthConvertLayer.h b/tests/validation/reference/DepthConvertLayer.h
index 1446bfda5b..5d97c73b3c 100644
--- a/tests/validation/reference/DepthConvertLayer.h
+++ b/tests/validation/reference/DepthConvertLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -35,19 +35,10 @@ namespace validation
 {
 namespace reference
 {
-template < typename T1, typename T2, typename std::enable_if < std::is_integral<T1>::value &&std::is_floating_point<T2>::value, int >::type = 0 >
-SimpleTensor<T2> depth_convert(const SimpleTensor<T1> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
-
-template < typename T1, typename T2, typename std::enable_if < std::is_floating_point<T1>::value &&std::is_integral<T2>::value, int >::type = 0 >
-SimpleTensor<T2> depth_convert(const SimpleTensor<T1> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
-
 template < typename T1, typename T2, typename std::enable_if < std::is_integral<T1>::value &&std::is_integral<T2>::value &&!std::is_same<T1, T2>::value, int >::type = 0 >
 SimpleTensor<T2> depth_convert(const SimpleTensor<T1> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
 
-template < typename T1, typename T2, typename std::enable_if < std::is_integral<T1>::value &&std::is_integral<T2>::value &&std::is_same<T1, T2>::value, int >::type = 0 >
-SimpleTensor<T2> depth_convert(const SimpleTensor<T1> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
-
-template < typename T1, typename T2, typename std::enable_if < std::is_floating_point<T1>::value &&is_floating_point<T2>::value, int >::type = 0 >
+template < typename T1, typename T2, typename std::enable_if < is_floating_point<T1>::value &&is_floating_point<T2>::value &&!std::is_same<T1, T2>::value, int >::type = 0 >
 SimpleTensor<T2> depth_convert(const SimpleTensor<T1> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift);
 } // namespace reference
 } // namespace validation