COMPMID-661: softmax-uint8 implementation (#16)
Change-Id: Iad11ce70a8a0878a48e445a092035c49c926cece
Reviewed-on: http://mpd-gerrit.cambridge.arm.com/94855
Tested-by: Kaizen <jeremy.johnson+kaizengerrit@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
diff --git a/tests/validation/CL/SoftmaxLayer.cpp b/tests/validation/CL/SoftmaxLayer.cpp
index 7842c5c..a06aa7b 100644
--- a/tests/validation/CL/SoftmaxLayer.cpp
+++ b/tests/validation/CL/SoftmaxLayer.cpp
@@ -50,9 +50,13 @@
/** Tolerance for fixed point operations */
constexpr AbsoluteTolerance<int16_t> tolerance_fixed_point(2);
+/** Tolerance for quantized operations */
+constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1);
+
/** CNN data types */
const auto CNNDataTypes = framework::dataset::make("DataType",
{
+ DataType::QASYMM8,
DataType::F16,
DataType::F32,
DataType::QS8,
@@ -65,12 +69,13 @@
DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(concat(datasets::SoftmaxLayerSmallShapes(), datasets::SoftmaxLayerLargeShapes()), CNNDataTypes), shape, data_type)
{
- // Set fixed point position data type allowed
- const int fixed_point_position = is_data_type_fixed_point(data_type) ? 3 : 0;
+ // Set fixed point position and quantization info if is allowed
+ const int fixed_point_position = is_data_type_fixed_point(data_type) ? 3 : 0;
+ const QuantizationInfo quantization_info = is_data_type_quantized_asymmetric(data_type) ? QuantizationInfo(1.f / 255.f, 0) : QuantizationInfo();
// Create tensors
- CLTensor src = create_tensor<CLTensor>(shape, data_type, 1, fixed_point_position);
- CLTensor dst = create_tensor<CLTensor>(shape, data_type, 1, fixed_point_position);
+ CLTensor src = create_tensor<CLTensor>(shape, data_type, 1, fixed_point_position, quantization_info);
+ CLTensor dst = create_tensor<CLTensor>(shape, data_type, 1, fixed_point_position, QuantizationInfo(1.f / 256.f, 0));
ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS);
ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
@@ -88,8 +93,17 @@
CLLogits1DMaxShiftExpSumKernel::ParallelReductionInfo reduction_info = CLLogits1DMaxShiftExpSumKernel::is_parallel_reduction(shape.x());
// Validate src padding
- const PaddingSize padding_src = PaddingCalculator(shape.x(), std::get<1>(reduction_info)).required_padding();
- validate(src.info()->padding(), padding_src);
+ // Legacy path used only by quantized asymmetric data type TODO(COMPMID-661) : Remove when port to new path
+ if(is_data_type_quantized_asymmetric(data_type))
+ {
+ const PaddingSize padding_src = PaddingCalculator(shape.x(), 16).required_padding();
+ validate(src.info()->padding(), padding_src);
+ }
+ else
+ {
+ const PaddingSize padding_src = PaddingCalculator(shape.x(), std::get<1>(reduction_info)).required_padding();
+ validate(src.info()->padding(), padding_src);
+ }
// Validate dst padding
const PaddingSize padding_dst = PaddingCalculator(shape.x(), 16).required_padding();
@@ -101,7 +115,7 @@
TEST_SUITE(Float)
TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLSoftmaxLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(datasets::SoftmaxLayerSmallShapes(), framework::dataset::make("DataType", DataType::F16)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLSoftmaxLayerFixture<half>, framework::DatasetMode::ALL, combine(datasets::SoftmaxLayerSmallShapes(), framework::dataset::make("DataType", DataType::F16)))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_f16);
@@ -114,7 +128,7 @@
TEST_SUITE_END()
TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLSoftmaxLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::SoftmaxLayerSmallShapes(), framework::dataset::make("DataType", DataType::F32)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLSoftmaxLayerFixture<float>, framework::DatasetMode::ALL, combine(datasets::SoftmaxLayerSmallShapes(), framework::dataset::make("DataType", DataType::F32)))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_f32);
@@ -130,12 +144,12 @@
template <typename T>
using CLSoftmaxLayerFixedPointFixture = SoftmaxValidationFixedPointFixture<CLTensor, CLAccessor, CLSoftmaxLayer, T>;
-TEST_SUITE(Quantized)
+TEST_SUITE(FixedPoint)
TEST_SUITE(QS8)
// Testing for fixed point position [1,6) as reciprocal limits the maximum fixed point position to 5
-FIXTURE_DATA_TEST_CASE(RunSmall, CLSoftmaxLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SoftmaxLayerSmallShapes(), framework::dataset::make("DataType",
- DataType::QS8)),
- framework::dataset::make("FractionalBits", 1, 6)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLSoftmaxLayerFixedPointFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(datasets::SoftmaxLayerSmallShapes(), framework::dataset::make("DataType",
+ DataType::QS8)),
+ framework::dataset::make("FractionalBits", 1, 6)))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_fixed_point);
@@ -151,10 +165,10 @@
TEST_SUITE(QS16)
// Testing for fixed point position [1,14) as reciprocal limits the maximum fixed point position to 14
-FIXTURE_DATA_TEST_CASE(RunSmall, CLSoftmaxLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SoftmaxLayerSmallShapes(),
- framework::dataset::make("DataType",
- DataType::QS16)),
- framework::dataset::make("FractionalBits", 1, 14)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLSoftmaxLayerFixedPointFixture<int16_t>, framework::DatasetMode::ALL, combine(combine(datasets::SoftmaxLayerSmallShapes(),
+ framework::dataset::make("DataType",
+ DataType::QS16)),
+ framework::dataset::make("FractionalBits", 1, 14)))
{
// Validate output
validate(CLAccessor(_target), _reference, tolerance_fixed_point);
@@ -170,6 +184,30 @@
TEST_SUITE_END()
TEST_SUITE_END()
+template <typename T>
+using CLSoftmaxLayerQuantizedFixture = SoftmaxValidationQuantizedFixture<CLTensor, CLAccessor, CLSoftmaxLayer, T>;
+
+TEST_SUITE(Quantized)
+TEST_SUITE(QASYMM8)
+FIXTURE_DATA_TEST_CASE(RunSmall, CLSoftmaxLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(datasets::SoftmaxLayerSmallShapes(),
+ framework::dataset::make("DataType",
+ DataType::QASYMM8)),
+ framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, -10) })))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_qasymm8);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, CLSoftmaxLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SoftmaxLayerLargeShapes(),
+ framework::dataset::make("DataType",
+ DataType::QASYMM8)),
+ framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, -10) })))
+{
+ // Validate output
+ validate(CLAccessor(_target), _reference, tolerance_qasymm8);
+}
+TEST_SUITE_END()
+TEST_SUITE_END()
+
TEST_SUITE_END()
TEST_SUITE_END()
} // namespace validation
diff --git a/tests/validation/CPP/SoftmaxLayer.cpp b/tests/validation/CPP/SoftmaxLayer.cpp
index eb76550..8e8cc1b 100644
--- a/tests/validation/CPP/SoftmaxLayer.cpp
+++ b/tests/validation/CPP/SoftmaxLayer.cpp
@@ -112,6 +112,18 @@
return dst;
}
+template <>
+SimpleTensor<uint8_t> softmax_layer<uint8_t>(const SimpleTensor<uint8_t> &src)
+{
+ // Note: Output quantization info should always have scale = 1/256 and offset = 0
+ const QuantizationInfo output_quantization_info = QuantizationInfo(1.f / 256, 0);
+
+ SimpleTensor<float> src_tmp = convert_from_asymmetric(src);
+ SimpleTensor<float> dst_tmp = softmax_layer<float>(src_tmp);
+ SimpleTensor<uint8_t> dst = convert_to_asymmetric(dst_tmp, output_quantization_info);
+ return dst;
+}
+
template SimpleTensor<float> softmax_layer(const SimpleTensor<float> &src);
template SimpleTensor<half> softmax_layer(const SimpleTensor<half> &src);
template SimpleTensor<qint8_t> softmax_layer(const SimpleTensor<qint8_t> &src);
diff --git a/tests/validation/fixtures/SoftmaxLayerFixture.h b/tests/validation/fixtures/SoftmaxLayerFixture.h
index 9c8f044..9836502 100644
--- a/tests/validation/fixtures/SoftmaxLayerFixture.h
+++ b/tests/validation/fixtures/SoftmaxLayerFixture.h
@@ -43,27 +43,33 @@
namespace validation
{
template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
-class SoftmaxValidationFixedPointFixture : public framework::Fixture
+class SoftmaxValidationGenericFixture : public framework::Fixture
{
public:
template <typename...>
- void setup(TensorShape shape, DataType data_type, int fractional_bits)
+ void setup(TensorShape shape, DataType data_type, int fractional_bits, QuantizationInfo quantization_info)
{
- _fractional_bits = fractional_bits;
+ _fractional_bits = fractional_bits;
+ _quantization_info = quantization_info;
- _target = compute_target(shape, data_type, fractional_bits);
- _reference = compute_reference(shape, data_type, fractional_bits);
+ _target = compute_target(shape, data_type, fractional_bits, quantization_info);
+ _reference = compute_reference(shape, data_type, fractional_bits, quantization_info);
}
protected:
template <typename U>
void fill(U &&tensor)
{
- if(_fractional_bits == 0)
+ if(!is_data_type_quantized(tensor.data_type()))
{
std::uniform_real_distribution<> distribution(-1000.f, 1000.f);
library->fill(tensor, distribution, 0);
}
+ else if(is_data_type_quantized_asymmetric(tensor.data_type()))
+ {
+ std::uniform_int_distribution<> distribution(0, 100);
+ library->fill(tensor, distribution, 0);
+ }
else
{
const int one_fixed = 1 << _fractional_bits;
@@ -72,11 +78,11 @@
}
}
- TensorType compute_target(const TensorShape &shape, DataType data_type, int fixed_point_position = 0)
+ TensorType compute_target(const TensorShape &shape, DataType data_type, int fixed_point_position, QuantizationInfo quantization_info)
{
// Create tensors
- TensorType src = create_tensor<TensorType>(shape, data_type, 1, fixed_point_position);
- TensorType dst = create_tensor<TensorType>(shape, data_type, 1, fixed_point_position);
+ TensorType src = create_tensor<TensorType>(shape, data_type, 1, fixed_point_position, quantization_info);
+ TensorType dst = create_tensor<TensorType>(shape, data_type, 1, fixed_point_position, QuantizationInfo(1.f / 256, 0));
// Create and configure function
FunctionType smx_layer;
@@ -101,10 +107,10 @@
return dst;
}
- SimpleTensor<T> compute_reference(const TensorShape &shape, DataType data_type, int fixed_point_position = 0)
+ SimpleTensor<T> compute_reference(const TensorShape &shape, DataType data_type, int fixed_point_position, QuantizationInfo quantization_info)
{
// Create reference
- SimpleTensor<T> src{ shape, data_type, 1, fixed_point_position };
+ SimpleTensor<T> src{ shape, data_type, 1, fixed_point_position, quantization_info };
// Fill reference
fill(src);
@@ -112,19 +118,42 @@
return reference::softmax_layer<T>(src);
}
- TensorType _target{};
- SimpleTensor<T> _reference{};
- int _fractional_bits{};
+ TensorType _target{};
+ SimpleTensor<T> _reference{};
+ int _fractional_bits{};
+ QuantizationInfo _quantization_info{};
};
template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
-class SoftmaxValidationFixture : public SoftmaxValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>
+class SoftmaxValidationFixture : public SoftmaxValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
{
public:
template <typename...>
void setup(TensorShape shape, DataType data_type)
{
- SoftmaxValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, 0);
+ SoftmaxValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, 0, QuantizationInfo());
+ }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class SoftmaxValidationFixedPointFixture : public SoftmaxValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+ template <typename...>
+ void setup(TensorShape shape, DataType data_type, int fixed_point_position)
+ {
+ SoftmaxValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, fixed_point_position, QuantizationInfo());
+ }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class SoftmaxValidationQuantizedFixture : public SoftmaxValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+ template <typename...>
+ void setup(TensorShape shape, DataType data_type, QuantizationInfo quantization_info)
+ {
+ SoftmaxValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, 0, quantization_info);
}
};
} // namespace validation