Integrate improved CPU depthwise convolution kernels
* Replace assembly kernels for depthwise convolution with more optimized
ones.
* Add int8 assembly kernels.
* Fix implicit padding on optimized kernels
Resolves: COMPMID-3867, COMPMID-4361
Change-Id: I0b0867e05f61be4f368f62190d55e14d0ab3ebf2
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5622
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
diff --git a/tests/datasets/DepthwiseConvolutionLayerDataset.h b/tests/datasets/DepthwiseConvolutionLayerDataset.h
index 4fd461d..a19e7ee 100644
--- a/tests/datasets/DepthwiseConvolutionLayerDataset.h
+++ b/tests/datasets/DepthwiseConvolutionLayerDataset.h
@@ -220,8 +220,6 @@
add_config(TensorShape(9U, 9U, 32U), Size2D(3U, 3U), PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL));
add_config(TensorShape(9U, 9U, 32U), Size2D(3U, 3U), PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), Size2D(2U, 2U));
add_config(TensorShape(9U, 9U, 32U), Size2D(3U, 3U), PadStrideInfo(2, 2, 1, 1, DimensionRoundingType::CEIL));
- // TODO(COMPMID-2464): Enable once dilated conv with stride 2 is supported
- // add_config(TensorShape(9U, 9U, 1U), Size2D(3U, 3U), PadStrideInfo(2, 2, 2, 2, DimensionRoundingType::CEIL), Size2D(2U, 2U));
}
};
/** Dataset containing optimized, 3x3 depthwise convolution shapes. */
@@ -259,11 +257,9 @@
add_config(TensorShape(7U, 7U, 16U), Size2D(5U, 5U), PadStrideInfo(1, 1, 4, 4, DimensionRoundingType::CEIL), Size2D(2U, 2U));
// Stride 2
add_config(TensorShape(9U, 9U, 32U), Size2D(5U, 5U), PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL));
- // TODO(COMPMID-2464): Enable once dilated conv with stride 2 is supported
- // add_config(TensorShape(9U, 9U, 32U), Size2D(5U, 5U), PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), Size2D(2U, 2U));
+ add_config(TensorShape(9U, 9U, 32U), Size2D(5U, 5U), PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL), Size2D(2U, 2U));
add_config(TensorShape(9U, 9U, 32U), Size2D(5U, 5U), PadStrideInfo(2, 2, 2, 2, 2, 2, DimensionRoundingType::CEIL));
- // TODO(COMPMID-2464): Enable once dilated conv with stride 2 is supported
- // add_config(TensorShape(9U, 9U, 32U), Size2D(5U, 5U), PadStrideInfo(2, 2, 4, 4, 4, 4, DimensionRoundingType::CEIL), Size2D(2U, 2U));
+ add_config(TensorShape(9U, 9U, 32U), Size2D(5U, 5U), PadStrideInfo(2, 2, 4, 4, 4, 4, DimensionRoundingType::CEIL), Size2D(2U, 2U));
}
};
} // namespace datasets
diff --git a/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h b/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h
index 2c94373..19ec6b2 100644
--- a/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h
+++ b/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h
@@ -117,8 +117,7 @@
void allocate_and_run_target()
{
- // TODO: uncomment after COMPMID-4361
- // add_padding_x({ &_src, &_weights, &_biases, &_target }, _data_layout);
+ add_padding_x({ &_src, &_weights, &_biases, &_target }, _data_layout);
// Allocate tensors
_src.allocator()->allocate();