Add uint8 support for quantized conv2d NHWC and enable per_tensor_out operator#18279
Add uint8 support for quantized conv2d NHWC and enable per_tensor_out operator#18279khazaei wants to merge 1 commit intopytorch:mainfrom
Conversation
… operator Summary: Goal is to get the fallback path working first then add ties. Before this change the tests were not passing. Added support for uint8 (UWORD8) data type in quantized_conv2d_nhwc_out for both standard and depthwise convolutions on HiFi backends. This change: - Extends op_quantized_conv2d_nhwc_out.cpp with uint8 handling for conv2d and depthwise conv2d operations using xa_nn_conv2d_per_chan_sym8sxasym8s and xa_nn_conv2d_depthwise_asym8uxasym8u kernels - Enables the cadence::quantized_conv2d_nhwc.per_tensor_out operator mapping in operator_fallback.bzl for both HiFi and TIE backends - Updates BUCK dependencies for TIE operators to include required exec_aten and kernel_runtime_context libs - Modifies test configuration to run on Artemis_HiFi4_UT_v3 backend - Fixed out_data_format for NHWC (was using NCHW format 1, should be 0) - Added weight transpose for depthwise conv (NHWC weight [OC,KH,KW,1] → nnlib expected [KH,KW,OC]) Differential Revision: D97036131
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/18279
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 Awaiting Approval, 1 Cancelled Job, 2 Unrelated FailuresAs of commit ed52ef6 with merge base 3604d3e ( AWAITING APPROVAL - The following workflows need approval before CI can run:
BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This PR needs a
|
Summary:
Goal is to get the fallback path working first then add ties. Before this change the tests were not passing.
Added support for uint8 (UWORD8) data type in quantized_conv2d_nhwc_out for both standard and depthwise convolutions on HiFi backends. This change:
Differential Revision: D97036131