COMPMID-2708 NEDepthwiseConvolution Generic: support for QUANT8_PER_CHANNEL_SYMM
COMPMID-2470 Implement a new and generic depthwise convolution for NEON QASYMM8 NHWC
COMPMID-2477 Enable FP16 data type for the new generic convolution on NEON for NHWC
COMPMID-2625 Remove old implementation files for the generic NEDepthwiseConvolution
Change-Id: I8f6deda4fc69dd7e472fba3228b1ed5dad172f3e
Signed-off-by: Giorgio Arena <giorgio.arena@arm.com>
Reviewed-on: https://review.mlplatform.org/c/2094
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
diff --git a/tests/validation/NEON/DepthwiseConvolutionLayer.cpp b/tests/validation/NEON/DepthwiseConvolutionLayer.cpp
index 6392906..0b482ff 100644
--- a/tests/validation/NEON/DepthwiseConvolutionLayer.cpp
+++ b/tests/validation/NEON/DepthwiseConvolutionLayer.cpp
@@ -163,16 +163,16 @@
}
DATA_TEST_CASE(ValidateGeneric, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zip(
- framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Mismatching data type input/weights
- TensorInfo(TensorShape(27U, 13U, 3U), 1, DataType::F32), // Mismatching input feature maps
- TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Mismatching depth multiplier
- TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid biases size
- TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid biases dimensions
- TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid output size
- TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32), // patch size bigger than input width
- TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32), // dilation < 1
- TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32),
- TensorInfo(TensorShape(32U, 13U, 8U), 1, DataType::QASYMM8),
+ framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Mismatching data type input/weights
+ TensorInfo(TensorShape(27U, 13U, 3U), 1, DataType::F32), // Mismatching input feature maps
+ TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Mismatching depth multiplier
+ TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid biases size
+ TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid biases dimensions
+ TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid output size
+ TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32), // Patch size bigger than input width
+ TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32), // Dilation < 1
+ TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32), // Window shrinking
+ TensorInfo(TensorShape(32U, 13U, 8U), 1, DataType::QASYMM8), // Window shrinking
}),
framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F16),
TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32),
@@ -240,7 +240,7 @@
Size2D(1U, 1U),
Size2D(1U, 1U),
})),
- framework::dataset::make("Expected", { false, false, false, false, false, false,false, false, true, true })),
+ framework::dataset::make("Expected", { false, false, false, false, false, false,false, false, false, false })),
input_info, weights_info, biases_info, output_info, conv_info, depth_multiplier,dilation, expected)
{
bool is_valid = bool(NEDepthwiseConvolutionLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &biases_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv_info, depth_multiplier, ActivationLayerInfo(), dilation));
@@ -512,7 +512,8 @@
template <typename T>
using NEDepthwiseConvolutionLayerQuantizedFixtureOptimized = DepthwiseConvolutionLayerValidationQuantizedFixture<Tensor, Accessor, NEDepthwiseConvolutionLayerOptimized, T>;
template <typename T>
-using NEDepthwiseConvolutionLayerQuantizedFixture = DepthwiseConvolutionLayerValidationQuantizedFixture<Tensor, Accessor, NEDepthwiseConvolutionLayer, T>;
+using NEDepthwiseConvolutionLayerQuantizedFixture = DepthwiseConvolutionLayerValidationQuantizedFixture<Tensor, Accessor, NEDepthwiseConvolutionLayer, T>;
+using NEDepthwiseConvolutionLayerQuantizedSymmetricPerChannelFixture = DepthwiseConvolutionLayerValidationQuantizedPerChannelFixture<Tensor, Accessor, NEDepthwiseConvolutionLayer, uint8_t, int8_t>;
TEST_SUITE(Quantized)
TEST_SUITE(QASYMM8)
@@ -642,6 +643,49 @@
}
TEST_SUITE_END() // Optimized
TEST_SUITE_END() // QASYMM8
+TEST_SUITE(QSYMM8_PER_CHANNEL)
+TEST_SUITE(Generic)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthwiseConvolutionLayerQuantizedSymmetricPerChannelFixture, framework::DatasetMode::ALL,
+ combine(combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(),
+ depth_multipliers),
+ framework::dataset::make("InputDataType", DataType::QASYMM8)),
+ framework::dataset::make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)),
+ framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10) })),
+ framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })),
+ ActivationFunctionsDataset))
+{
+ validate(Accessor(_target), _reference, tolerance_qasymm8);
+}
+
+TEST_SUITE(Dilation)
+FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthwiseConvolutionLayerQuantizedSymmetricPerChannelFixture, framework::DatasetMode::ALL,
+ combine(combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(),
+ depth_multipliers),
+ framework::dataset::make("InputDataType", DataType::QASYMM8)),
+ framework::dataset::make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)),
+ framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10) })),
+ framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })),
+ ActivationFunctionsDataset))
+{
+ validate(Accessor(_target), _reference, tolerance_qasymm8);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthwiseConvolutionLayerQuantizedSymmetricPerChannelFixture, framework::DatasetMode::NIGHTLY,
+ combine(combine(combine(combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset(),
+ depth_multipliers),
+ framework::dataset::make("InputDataType", DataType::QASYMM8)),
+ framework::dataset::make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)),
+ framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10) })),
+ framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })),
+ framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })),
+ ActivationFunctionsDataset))
+{
+ validate(Accessor(_target), _reference, tolerance_qasymm8);
+}
+TEST_SUITE_END() // Dilation
+TEST_SUITE_END() // Generic
+TEST_SUITE_END() // QSYMM8_PER_CHANNEL
TEST_SUITE_END() // Quantized
TEST_SUITE_END() // DepthwiseConvLayer
diff --git a/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h b/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h
index 85930eb..f909885 100644
--- a/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h
+++ b/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h
@@ -449,12 +449,22 @@
public:
template <typename...>
void setup(TensorShape in_shape, Size2D kernel_size, PadStrideInfo pad_stride_info, Size2D dilation, unsigned int depth_multiplier, DataType input_data_type, DataType weights_data_type,
- QuantizationInfo input_quantization_info, QuantizationInfo weights_quantization_info, QuantizationInfo output_quantization_info,
- DataLayout data_layout, ActivationLayerInfo act_info)
+ QuantizationInfo input_quantization_info, QuantizationInfo output_quantization_info, DataLayout data_layout, ActivationLayerInfo act_info)
{
+ const float out_scale = output_quantization_info.uniform().scale;
+ const float in_scale = input_quantization_info.uniform().scale;
+
+ std::vector<float> weights_scales{};
+ std::mt19937 gen(library->seed());
+ std::uniform_real_distribution<> dis(0.01f, out_scale / in_scale);
+ for(size_t i = 0; i < in_shape.z() * depth_multiplier; ++i)
+ {
+ weights_scales.push_back(dis(gen));
+ }
+
DepthwiseConvolutionLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T, TW>::setup(in_shape, kernel_size, pad_stride_info, dilation, depth_multiplier,
input_data_type, weights_data_type,
- input_quantization_info, weights_quantization_info, output_quantization_info,
+ input_quantization_info, QuantizationInfo(weights_scales), output_quantization_info,
data_layout, act_info);
}
};
diff --git a/tests/validation/reference/DepthwiseConvolutionLayer.cpp b/tests/validation/reference/DepthwiseConvolutionLayer.cpp
index 7458f81..608093d 100644
--- a/tests/validation/reference/DepthwiseConvolutionLayer.cpp
+++ b/tests/validation/reference/DepthwiseConvolutionLayer.cpp
@@ -188,17 +188,17 @@
{
for(int z = 0; z < input_depth; ++z)
{
- int output_multiplier = 0;
- int output_shift = 0;
- const float weights_scale = (is_quantized_per_channel) ? weights_scale_vec[z] : weights_scale_vec[0];
- const float multiplier = input_scale * weights_scale / output_scale;
- arm_compute::quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
-
for(unsigned int m = 0; m < depth_multiplier; ++m)
{
const int out_z = z * depth_multiplier + m;
const int32_t bias_val = *static_cast<const int32_t *>(biases(Coordinates(out_z)));
+ int output_multiplier = 0;
+ int output_shift = 0;
+ const float weights_scale = (is_quantized_per_channel) ? weights_scale_vec[out_z] : weights_scale_vec[0];
+ const float multiplier = input_scale * weights_scale / output_scale;
+ arm_compute::quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift);
+
for(int y = minimum_y; y <= minimum_y + maximum_y; y += conv_info.stride().second)
{
for(int x = minimum_x; x <= minimum_x + maximum_x; x += conv_info.stride().first)