diff options
| author | Diqing Zhong <diqing.zhong@arm.com> | 2020-12-16 16:46:06 +0100 |
|---|---|---|
| committer | patrik.gustavsson <patrik.gustavsson@arm.com> | 2020-12-22 14:44:09 +0000 |
| commit | 016b827ad722aecd4338d1d6c7b1b004760490b7 (patch) | |
| tree | 7a26bfdf670e8885693bc7db531ea647e24d571b /ethosu/vela/npu_performance.py | |
| parent | 3d73717f793100ba6705441fb42514f938780c1e (diff) | |
| download | ethos-u-vela-016b827ad722aecd4338d1d6c7b1b004760490b7.tar.gz | |
MLBEDSW-1493: Optimise strided conv
- Reshape/rearrange IFM and weight tensor for better HW utilization
- Update estimator to cover this case
Change-Id: I4be70a69fa600a1951bf1c247f9973e6cc9b03f4
Signed-off-by: Diqing Zhong <diqing.zhong@arm.com>
Diffstat (limited to 'ethosu/vela/npu_performance.py')
| -rw-r--r-- | ethosu/vela/npu_performance.py | 5 |
1 files changed, 5 insertions, 0 deletions
diff --git a/ethosu/vela/npu_performance.py b/ethosu/vela/npu_performance.py index 4ca46831..c2418d73 100644 --- a/ethosu/vela/npu_performance.py +++ b/ethosu/vela/npu_performance.py @@ -422,6 +422,9 @@ def performance_metrics_for_pass(arch, ps, block_config=None, rewrite_list=None, ifm_tensor, _, weight_tensor, ofm_tensor = ps.get_primary_op_ifm_ifm2_weights_ofm() ifm_tensor_shape = ps.primary_op.ifm_shapes[0].clone() ofm_tensor_shape = ps.primary_op.ofm_shapes[0].clone() + ofm_block.width = min(ofm_block.width, ofm_tensor_shape.width) + ofm_block.height = min(ofm_block.height, ofm_tensor_shape.height) + ofm_block.depth = min(ofm_block.depth, ofm_tensor_shape.depth) if npu_block_type == NpuBlockType.ReduceSum: block_traversal = TensorBlockTraversal.DepthFirst @@ -439,6 +442,8 @@ def performance_metrics_for_pass(arch, ps, block_config=None, rewrite_list=None, ifm_block = arch.get_ifm_block_size( ifm_block_depth, ofm_block, primary_op.kernel, ifm_resampling_mode=ifm_tensor.resampling_mode ) + ifm_block.width = min(ifm_block.width, ifm_tensor_shape.width) + ifm_block.height = min(ifm_block.height, ifm_tensor_shape.height) if npu_block_type in ( NpuBlockType.ConvolutionMxN, |