Fix reshape dq operator

mcr229 · facebook-github-bot · commit 689796499024 · 2024-03-11T13:30:34.000-07:00
Summary:
This seems to be the fix.

My guess is that we have been reshaping the convert_qd8 operator using output's batch size. However the output batch size is not known yet because we are reshaping in order of computation. This is why we could fix it by running reshape runtime twice. We need to reshape the first time so that the output batch size is set correctly, and then we reshape again.

Will push an external PR for this with test.

bypass-github-export-checks

Reviewed By: digantdesai

Differential Revision: D54737642

fbshipit-source-id: 6cf5918195a02a587d3658f7277633ff087e160d
diff --git a/backends/xnnpack/runtime/XNNExecutor.cpp b/backends/xnnpack/runtime/XNNExecutor.cpp
@@ -100,7 +100,6 @@ __ET_NODISCARD Error XNNExecutor::prepare_args(EValue** args) {
   }
   // // Propagate Input Shape and Memory Plan for increased allocation
   status = xnn_reshape_runtime(runtime_.get());
-  status = xnn_reshape_runtime(runtime_.get());
 
   ET_CHECK_OR_RETURN_ERROR(
       status == xnn_status_success,

Original file line number	Diff line number	Diff line change
`@@ -100,7 +100,6 @@ __ET_NODISCARD Error XNNExecutor::prepare_args(EValue** args) {`
`100`	`100`	`}`
`101`	`101`	`// // Propagate Input Shape and Memory Plan for increased allocation`
`102`	`102`	`status = xnn_reshape_runtime(runtime_.get());`
`103`		`- status = xnn_reshape_runtime(runtime_.get());`
`104`	`103`
`105`	`104`	`ET_CHECK_OR_RETURN_ERROR(`
`106`	`105`	`status == xnn_status_success,`