COMPMID-2235: Extend type support for CL/NEON DequantizationLayer.
Adds support for:
- QSYMM8
Change-Id: Ia0b839fc844ce0f968dad1b69a001f9a660dbcd5
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1378
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>
diff --git a/tests/validation/CL/DequantizationLayer.cpp b/tests/validation/CL/DequantizationLayer.cpp
index b1b0d81..2ef8c60 100644
--- a/tests/validation/CL/DequantizationLayer.cpp
+++ b/tests/validation/CL/DequantizationLayer.cpp
@@ -27,6 +27,7 @@
#include "arm_compute/runtime/CL/functions/CLDequantizationLayer.h"
#include "tests/CL/CLAccessor.h"
#include "tests/PaddingCalculator.h"
+#include "tests/datasets/DatatypeDataset.h"
#include "tests/datasets/ShapeDatasets.h"
#include "tests/framework/Asserts.h"
#include "tests/framework/Macros.h"
@@ -96,16 +97,14 @@
using CLDequantizationLayerFixture = DequantizationValidationFixture<CLTensor, CLAccessor, CLDequantizationLayer, T>;
TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLDequantizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(),
- framework::dataset::make("DataType", DataType::F16)),
- framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.1f, 128.0f) })))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDequantizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), datasets::QuantizedTypes()),
+ framework::dataset::make("DataType", DataType::F16)))
{
// Validate output
validate(CLAccessor(_target), _reference);
}
-FIXTURE_DATA_TEST_CASE(RunLarge, CLDequantizationLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(),
- framework::dataset::make("DataType", DataType::F16)),
- framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.1f, 128.0f) })))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDequantizationLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), datasets::QuantizedTypes()),
+ framework::dataset::make("DataType", DataType::F16)))
{
// Validate output
validate(CLAccessor(_target), _reference);
@@ -113,16 +112,14 @@
TEST_SUITE_END() // FP16
TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLDequantizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(),
- framework::dataset::make("DataType", DataType::F32)),
- framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.1f, 128.0f) })))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLDequantizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), datasets::QuantizedTypes()),
+ framework::dataset::make("DataType", DataType::F32)))
{
// Validate output
validate(CLAccessor(_target), _reference);
}
-FIXTURE_DATA_TEST_CASE(RunLarge, CLDequantizationLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(),
- framework::dataset::make("DataType", DataType::F32)),
- framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.1f, 128.0f) })))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLDequantizationLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), datasets::QuantizedTypes()),
+ framework::dataset::make("DataType", DataType::F32)))
{
// Validate output
validate(CLAccessor(_target), _reference);
diff --git a/tests/validation/CL/UNIT/TensorAllocator.cpp b/tests/validation/CL/UNIT/TensorAllocator.cpp
index 4b8e105..d91f4dd 100644
--- a/tests/validation/CL/UNIT/TensorAllocator.cpp
+++ b/tests/validation/CL/UNIT/TensorAllocator.cpp
@@ -249,9 +249,9 @@
// Check quantization information
ARM_COMPUTE_EXPECT(!tensor.info()->quantization_info().empty(), framework::LogLevel::ERRORS);
- ARM_COMPUTE_EXPECT(!tensor.info()->quantization_info().scale.empty(), framework::LogLevel::ERRORS);
- ARM_COMPUTE_EXPECT(tensor.info()->quantization_info().scale.size() == scale.size(), framework::LogLevel::ERRORS);
- ARM_COMPUTE_EXPECT(tensor.info()->quantization_info().offset.empty(), framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(!tensor.info()->quantization_info().scale().empty(), framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(tensor.info()->quantization_info().scale().size() == scale.size(), framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(tensor.info()->quantization_info().offset().empty(), framework::LogLevel::ERRORS);
CLQuantization quantization = tensor.quantization();
ARM_COMPUTE_ASSERT(quantization.scale != nullptr);
diff --git a/tests/validation/NEON/DequantizationLayer.cpp b/tests/validation/NEON/DequantizationLayer.cpp
index 0ae20b7..a4606fe 100644
--- a/tests/validation/NEON/DequantizationLayer.cpp
+++ b/tests/validation/NEON/DequantizationLayer.cpp
@@ -27,6 +27,7 @@
#include "arm_compute/runtime/TensorAllocator.h"
#include "tests/NEON/Accessor.h"
#include "tests/PaddingCalculator.h"
+#include "tests/datasets/DatatypeDataset.h"
#include "tests/datasets/ShapeDatasets.h"
#include "tests/framework/Asserts.h"
#include "tests/framework/Macros.h"
@@ -106,16 +107,14 @@
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEDequantizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(),
- framework::dataset::make("DataType", DataType::F16)),
- framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.1f, 128.0f) })))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEDequantizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), datasets::QuantizedTypes()),
+ framework::dataset::make("DataType", DataType::F16)))
{
// Validate output
validate(Accessor(_target), _reference);
}
-FIXTURE_DATA_TEST_CASE(RunLarge, NEDequantizationLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(),
- framework::dataset::make("DataType", DataType::F16)),
- framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.1f, 128.0f) })))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEDequantizationLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), datasets::QuantizedTypes()),
+ framework::dataset::make("DataType", DataType::F16)))
{
// Validate output
validate(Accessor(_target), _reference);
@@ -124,16 +123,14 @@
#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEDequantizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(),
- framework::dataset::make("DataType", DataType::F32)),
- framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.1f, 128.0f) })))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEDequantizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), datasets::QuantizedTypes()),
+ framework::dataset::make("DataType", DataType::F32)))
{
// Validate output
validate(Accessor(_target), _reference);
}
-FIXTURE_DATA_TEST_CASE(RunLarge, NEDequantizationLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(),
- framework::dataset::make("DataType", DataType::F32)),
- framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.1f, 128.0f) })))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEDequantizationLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), datasets::QuantizedTypes()),
+ framework::dataset::make("DataType", DataType::F32)))
{
// Validate output
validate(Accessor(_target), _reference);
diff --git a/tests/validation/UNIT/TensorInfo.cpp b/tests/validation/UNIT/TensorInfo.cpp
index 96d07da..009c757 100644
--- a/tests/validation/UNIT/TensorInfo.cpp
+++ b/tests/validation/UNIT/TensorInfo.cpp
@@ -141,9 +141,9 @@
// Check quantization information
ARM_COMPUTE_EXPECT(!info.quantization_info().empty(), framework::LogLevel::ERRORS);
- ARM_COMPUTE_EXPECT(!info.quantization_info().scale.empty(), framework::LogLevel::ERRORS);
- ARM_COMPUTE_EXPECT(info.quantization_info().scale.size() == 1, framework::LogLevel::ERRORS);
- ARM_COMPUTE_EXPECT(info.quantization_info().offset.empty(), framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(!info.quantization_info().scale().empty(), framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(info.quantization_info().scale().size() == 1, framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(info.quantization_info().offset().empty(), framework::LogLevel::ERRORS);
UniformQuantizationInfo qinfo = info.quantization_info().uniform();
ARM_COMPUTE_EXPECT(qinfo.scale == scale, framework::LogLevel::ERRORS);
@@ -160,10 +160,10 @@
// Check quantization information
ARM_COMPUTE_EXPECT(!info.quantization_info().empty(), framework::LogLevel::ERRORS);
- ARM_COMPUTE_EXPECT(!info.quantization_info().scale.empty(), framework::LogLevel::ERRORS);
- ARM_COMPUTE_EXPECT(info.quantization_info().scale.size() == 1, framework::LogLevel::ERRORS);
- ARM_COMPUTE_EXPECT(!info.quantization_info().offset.empty(), framework::LogLevel::ERRORS);
- ARM_COMPUTE_EXPECT(info.quantization_info().offset.size() == 1, framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(!info.quantization_info().scale().empty(), framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(info.quantization_info().scale().size() == 1, framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(!info.quantization_info().offset().empty(), framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(info.quantization_info().offset().size() == 1, framework::LogLevel::ERRORS);
UniformQuantizationInfo qinfo = info.quantization_info().uniform();
ARM_COMPUTE_EXPECT(qinfo.scale == scale, framework::LogLevel::ERRORS);
@@ -179,9 +179,9 @@
// Check quantization information
ARM_COMPUTE_EXPECT(!info.quantization_info().empty(), framework::LogLevel::ERRORS);
- ARM_COMPUTE_EXPECT(!info.quantization_info().scale.empty(), framework::LogLevel::ERRORS);
- ARM_COMPUTE_EXPECT(info.quantization_info().scale.size() == scale.size(), framework::LogLevel::ERRORS);
- ARM_COMPUTE_EXPECT(info.quantization_info().offset.empty(), framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(!info.quantization_info().scale().empty(), framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(info.quantization_info().scale().size() == scale.size(), framework::LogLevel::ERRORS);
+ ARM_COMPUTE_EXPECT(info.quantization_info().offset().empty(), framework::LogLevel::ERRORS);
}
TEST_SUITE_END() // TensorInfoValidation
diff --git a/tests/validation/fixtures/DequantizationLayerFixture.h b/tests/validation/fixtures/DequantizationLayerFixture.h
index 2e3712d..15f3711 100644
--- a/tests/validation/fixtures/DequantizationLayerFixture.h
+++ b/tests/validation/fixtures/DequantizationLayerFixture.h
@@ -47,10 +47,11 @@
{
public:
template <typename...>
- void setup(TensorShape shape, DataType data_type, QuantizationInfo qinfo)
+ void setup(TensorShape shape, DataType src_data_type, DataType dst_datatype)
{
- _target = compute_target(shape, data_type, qinfo);
- _reference = compute_reference(shape, data_type, qinfo);
+ _quantization_info = generate_quantization_info(src_data_type);
+ _target = compute_target(shape, src_data_type, dst_datatype);
+ _reference = compute_reference(shape, src_data_type);
}
protected:
@@ -60,11 +61,11 @@
library->fill_tensor_uniform(tensor, 0);
}
- TensorType compute_target(const TensorShape &shape, DataType data_type, QuantizationInfo qinfo)
+ TensorType compute_target(const TensorShape &shape, DataType src_data_type, DataType dst_datatype)
{
// Create tensors
- TensorType src = create_tensor<TensorType>(shape, DataType::QASYMM8, 1, qinfo);
- TensorType dst = create_tensor<TensorType>(shape, data_type);
+ TensorType src = create_tensor<TensorType>(shape, src_data_type, 1, _quantization_info);
+ TensorType dst = create_tensor<TensorType>(shape, dst_datatype);
// Create and configure function
FunctionType dequantization_layer;
@@ -89,19 +90,43 @@
return dst;
}
- SimpleTensor<T> compute_reference(const TensorShape &shape, DataType data_type, QuantizationInfo qinfo)
+ SimpleTensor<T> compute_reference(const TensorShape &shape, DataType src_data_type)
{
- // Create reference
- SimpleTensor<uint8_t> src{ shape, DataType::QASYMM8, 1, qinfo };
-
- // Fill reference
- fill(src);
-
- return reference::dequantization_layer<T>(src);
+ if(is_data_type_quantized_asymmetric(src_data_type))
+ {
+ SimpleTensor<uint8_t> src{ shape, src_data_type, 1, _quantization_info };
+ fill(src);
+ return reference::dequantization_layer<T>(src);
+ }
+ else
+ {
+ SimpleTensor<int8_t> src{ shape, src_data_type, 1, _quantization_info };
+ fill(src);
+ return reference::dequantization_layer<T>(src);
+ }
}
- TensorType _target{};
- SimpleTensor<T> _reference{};
+protected:
+ QuantizationInfo generate_quantization_info(DataType data_type)
+ {
+ std::uniform_int_distribution<> distribution(1, 127);
+ std::mt19937 gen(library.get()->seed());
+
+ switch(data_type)
+ {
+ case DataType::QSYMM8:
+ return QuantizationInfo(1.f / distribution(gen));
+ case DataType::QASYMM8:
+ return QuantizationInfo(1.f / distribution(gen), distribution(gen));
+ default:
+ ARM_COMPUTE_ERROR("Unsupported data type");
+ }
+ }
+
+protected:
+ TensorType _target{};
+ SimpleTensor<T> _reference{};
+ QuantizationInfo _quantization_info{};
};
} // namespace validation
} // namespace test
diff --git a/tests/validation/reference/DequantizationLayer.cpp b/tests/validation/reference/DequantizationLayer.cpp
index 286a609..d07371c 100644
--- a/tests/validation/reference/DequantizationLayer.cpp
+++ b/tests/validation/reference/DequantizationLayer.cpp
@@ -23,6 +23,8 @@
*/
#include "DequantizationLayer.h"
+#include "Permute.h"
+
namespace arm_compute
{
namespace test
@@ -31,24 +33,82 @@
{
namespace reference
{
-template <typename T>
-SimpleTensor<T> dequantization_layer(const SimpleTensor<uint8_t> &src)
+namespace
{
- const DataType dst_data_type = std::is_same<T, float>::value ? DataType::F32 : DataType::F16;
- const UniformQuantizationInfo &quantization_info = src.quantization_info().uniform();
+template <typename TOut>
+TOut dequantize(int8_t val, const UniformQuantizationInfo qinfo)
+{
+ return static_cast<TOut>(dequantize_qsymm8(val, qinfo));
+}
+template <typename TOut>
+TOut dequantize(uint8_t val, const UniformQuantizationInfo qinfo)
+{
+ return static_cast<TOut>(dequantize_qasymm8(val, qinfo));
+}
- SimpleTensor<T> dst{ src.shape(), dst_data_type };
+template <typename TOut, typename TIn>
+SimpleTensor<TOut> dequantization_layer_nchw(const SimpleTensor<TIn> &src)
+{
+ const DataType src_data_type = src.data_type();
+ const DataType dst_data_type = std::is_same<TOut, float>::value ? DataType::F32 : DataType::F16;
- for(int i = 0; i < src.num_elements(); ++i)
+ SimpleTensor<TOut> dst{ src.shape(), dst_data_type };
+
+ if(src_data_type == DataType::QSYMM8_PER_CHANNEL)
{
- dst[i] = static_cast<T>(dequantize_qasymm8(src[i], quantization_info));
+ const int WH = src.shape().x() * src.shape().y();
+ const int C = src.shape().z();
+ const int N = src.shape().total_size() / (WH * C);
+
+ const std::vector<float> qscales = src.quantization_info().scale();
+
+ for(int n = 0; n < N; ++n)
+ {
+ for(int c = 0; c < C; ++c)
+ {
+ const size_t idx = n * C * WH + c * WH;
+ const UniformQuantizationInfo channel_qinfo = { qscales[c], 0 };
+
+ // Dequantize slice
+ for(int s = 0; s < WH; ++s)
+ {
+ dst[idx + s] = dequantize<TOut>(src[idx + s], channel_qinfo);
+ }
+ }
+ }
+ }
+ else
+ {
+ const UniformQuantizationInfo &quantization_info = src.quantization_info().uniform();
+ ARM_COMPUTE_ERROR_ON(quantization_info.offset != 0 && src_data_type == DataType::QSYMM8);
+
+ for(int i = 0; i < src.num_elements(); ++i)
+ {
+ dst[i] = static_cast<TOut>(dequantize<TOut>(src[i], quantization_info));
+ }
}
return dst;
}
+} // namespace
+template <typename TOut, typename TIn>
+SimpleTensor<TOut> dequantization_layer(const SimpleTensor<TIn> &src)
+{
+ if(src.data_layout() == DataLayout::NHWC && src.data_type() == DataType::QSYMM8_PER_CHANNEL)
+ {
+ SimpleTensor<TIn> src_nchw = reference::permute<TIn>(src, PermutationVector(1U, 2U, 0U));
+ return reference::permute<TOut>(dequantization_layer_nchw<TOut>(src_nchw), PermutationVector(2U, 0U, 1U));
+ }
+ else
+ {
+ return dequantization_layer_nchw<TOut>(src);
+ }
+}
template SimpleTensor<half> dequantization_layer(const SimpleTensor<uint8_t> &src);
template SimpleTensor<float> dequantization_layer(const SimpleTensor<uint8_t> &src);
+template SimpleTensor<half> dequantization_layer(const SimpleTensor<int8_t> &src);
+template SimpleTensor<float> dequantization_layer(const SimpleTensor<int8_t> &src);
} // namespace reference
} // namespace validation
} // namespace test
diff --git a/tests/validation/reference/DequantizationLayer.h b/tests/validation/reference/DequantizationLayer.h
index 1d0e54b..8c78084 100644
--- a/tests/validation/reference/DequantizationLayer.h
+++ b/tests/validation/reference/DequantizationLayer.h
@@ -35,8 +35,8 @@
{
namespace reference
{
-template <typename T>
-SimpleTensor<T> dequantization_layer(const SimpleTensor<uint8_t> &src);
+template <typename TOut, typename TIn>
+SimpleTensor<TOut> dequantization_layer(const SimpleTensor<TIn> &src);
} // namespace reference
} // namespace validation
} // namespace test