robert-kalmar

@fengwang what SDK you use to compile the TIM-VX? Can you please check if there is the `VX_BATCH_GEMM_API_SUPPORT` makro is defined and set in any header file? This macro should...

Reviewers this is https://github.com/google/XNNPACK/issues/1559 is related to the PR.

@terryheo , imagine this situation when the cross compilation SDK set up the target processor with (`--mcpu` switch in GNU Toolchain) and one cross-compiles the TensorFlow Lite with CMake, using...

Hi @gbaned, @terryheo, any progress/comments on this PR?

Hi @gbaned, @terryheo, any progress/comments on this PR? I updated the PR based on terry's comment. Any other comments?

Hi @sunshinemyson , I still see the problem. My configuration: Tensorflow Lite 2.9.1 VX Delegate: Revision e04ebf7f55f69ca193c025eace7a8c6d6880d339 TIM-VX: 3e8d5e3493cc8bbd6908b72d2b7fe2d508a397c4 + "Supoort layout inference for PRule" patch ``` [----------] Global test...

@sunshinemyson , I don't see too much difference btw 2.9.1 and 2.7.0 (original finding). Can you replicate those issues (and the non-determinicity) on your side?

@sunshinemyson I run some iteration of the gather_nd_test and identified the test cases, which are non-deterministic - that means they fail/pass randomly (see [gather_nd_test.xlsx](https://github.com/VeriSilicon/tflite-vx-delegate/files/9101734/gather_nd_test.xlsx) ) , hopefully it helps you....

@sunshinemyson , I am using the standalone OVXLIB, which shall correspond to patch release _rel_6.4.3.p4_20220609_ from our GPU driver delivery.

@sunshinemyson I did another thest with stadalone OVXLIB, corresponding to patch release rel_6.4.3.p4_20220714, problem still exist. I can reproduce the problem even with your revisions:   What might be...