Cortex_m backend: Add quantizer + avoid linear decomp #15459
Conversation
Changes to_edge_and_transform to to_edge which supports the preserver_ops arg of the EdgeCompileConfig to avoid decomposing of the linear op. This significantly simplifies lowering the linear operator as it does not have to be re-fused. Adds a cortex_m quantizer, with the intention to be general enough to be used for a general MCU. It is implemented as a ComposableQuantizer using multiple instances of a new OperatorConfigQuantizer class. This gives a number of abstraction levels for configuration - McuQuantizer - ComposableQuantizer - OperatorConfig - QuantizerConfig - QuantizationSpec The new quantizer also adds a transform_for_annotation pass pipeline which allows to fix scalar + tensor operations. Old test_quantize_op_fusion_pass test is removed since it is not relevant anymore after the add implementation has been redone. Change-Id: Ic1d5b48623a14d220cb1ba472948db6a1406e0b7 Signed-off-by: Adrian Lundell <adrian.lundell@arm.com>
AdrianLundell
added
partner: arm
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/15459
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ✅ No FailuresAs of commit 364159b with merge base a11d555 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
| # LICENSE file in the root directory of this source tree. | ||
|
|
||
|
|
||
| from executorch.backends.arm._passes import ScalarsToAttributePass |
There was a problem hiding this comment.
Do you plan to add "EdgeCompileConfig(preserve_ops...." step in CortexMPassManager to avoid decomposition ?
There was a problem hiding this comment.
Or "Do you plan to implement a CortexMPartitioner with ops_to_not_decompose() method to work with to_edge_transform_and_lower(), similar to how the Cadence backend handles operation preservation?"
executorch/backends/cadence/aot/compiler.py
Lines 275 to 277 in 007ccc6
There was a problem hiding this comment.
I am rather neutral towards how the list of ops to not preserve should be implemented, do you have a preference? From what I noticed preserve_ops in EdgeCompileConfig does not do anything in to_edge_transfrom_and_lower however, only in to_edge.
Changes to_edge_and_transform to to_edge which supports the preserver_ops arg of the EdgeCompileConfig to avoid decomposing of the linear op. This significantly simplifies lowering the linear operator as it does not have to be re-fused. Adds a cortex_m quantizer, with the intention to be general enough to be used for a general MCU. It is implemented as a ComposableQuantizer using multiple instances of a new OperatorConfigQuantizer class. This gives a number of abstraction levels for configuration - McuQuantizer - ComposableQuantizer - OperatorConfig - QuantizerConfig - QuantizationSpec The new quantizer also adds a transform_for_annotation pass pipeline which allows to fix scalar + tensor operations. Signed-off-by: Adrian Lundell <adrian.lundell@arm.com>
2 participants
Changes to_edge_and_transform to to_edge which supports the preserver_ops arg of the EdgeCompileConfig to avoid decomposing of the linear op. This significantly simplifies lowering the linear operator as it does not have to be re-fused.
Adds a cortex_m quantizer, with the intention to be general enough to be used for a general MCU. It is implemented as a ComposableQuantizer using multiple instances of a new OperatorConfigQuantizer class. This gives a number of abstraction levels for configuration
The new quantizer also adds a transform_for_annotation pass pipeline which allows to fix scalar + tensor operations.
cc @freddan80 @per @zingo @oscarandersson8218 @digantdesai