Commits
Chi Lo authored and GitHub committed 7572e6055ce
[TensorRT EP] Back out the PerThreadContext (#17690) Current TRT EP's PerthreadContext allows more than one IExecutionContext instance to be created by one engine instance. But, it's possible to hit an error that caused by TRT API context.setBindingDimensions() in our TRT EP code [here](https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2Fmicrosoft%2Fonnxruntime%2Fblob%2Fmain%2Fonnxruntime%2Fcore%2Fproviders%2Ftensorrt%2Ftensorrt_execution_provider.cc%23L2775&data=05%7C01%7CChi.Lo%40microsoft.com%7Cd8b23c3a4c0b4dcce9b408dbbd9309de%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C638312211465211140%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=5EZoAoXgWFSuz%2BIRMH%2FXZaO%2BfKNP%2FZDZYEZg3W%2Ff30w%3D&reserved=0) under the case of the input shape changes ( meaning engine being rebuilt) with multithreading. From the [doc](https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdocs.nvidia.com%2Fdeeplearning%2Ftensorrt%2Fapi%2Fc_api%2Fclassnvinfer1_1_1_i_execution_context.html%23ada050e88320bcc40987b0acadc2ef962&data=05%7C01%7CChi.Lo%40microsoft.com%7Cd8b23c3a4c0b4dcce9b408dbbd9309de%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C638312211465211140%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=%2BmVZU5iLD97B3YBPdHZP7jOQ2dGoleI3R0mSMVgopG4%3D&reserved=0) and the [discussion](https://nam06.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgithub.com%2FNVIDIA%2FTensorRT%2Fissues%2F846&data=05%7C01%7CChi.Lo%40microsoft.com%7Cd8b23c3a4c0b4dcce9b408dbbd9309de%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%7C638312211465211140%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=c8v%2FK2UkQ%2FNbf8w1sHNDGsB2kxw4sSmkyQ2QuCs8Fs8%3D&reserved=0), it seems we should have different OptimizationProfile for different IExecutionContext which our current TRT EP doesn’t support regardless of using PerThreadContext implementation. Back out the PerThreadContext until we completely solve this issue.