search

ROCm / MIOpen

AMD's Machine Intelligence Library
https://rocm.docs.amd.com/projects/MIOpen/en/latest/
Other
1.07k stars 224 forks source link

Issue with int8 datatype for miopenBatchNormalizationForwardInference #2010

Closed RambabuSwargam closed 3 weeks ago

RambabuSwargam commented 1 year ago

For int8 inference case we are facing some issues and we tried to confirm with the reproducer. And we checked with different combinations, all are passing in the float32 cases and for int8 some are giving core dump and some are failing. Attached the examples for float32 and int8 and also the results.

Could some one help us regarding this, what could be the issue in the example and is there any limitations with int8 datatype

And is NHWC format is supported for miopenBatchNormalizationForwardInference and miopenBatchNormalizationForwardTraining ?

Note : Compilation command: clang++ -I/opt/rocm/include -L/opt/rocm/lib -DHIP_PLATFORM_HCC= -DHIP_PLATFORM_AMD= -I/opt/rocm/hip/include -I/opt/rocm/llvm/bin/../lib/clang/14.0.0 -I/opt/rocm/hsa/include -L/opt/rocm/rocblas/lib -lMIOpen -lamdhip64 -v -lrocblas

s8IssueCase.log f32Case.log FWD_Inference_f32Pass_Code.txt

FWD_Inference_s8_CoreFailCaes_Code.txt

Attached code files in .txt since here it not allowing to attach cpp file.

ppanchad-amd commented 4 months ago

@RambabuSwargam Can you please test with latest ROCm 6.1.2 and see if issue still exists? Thanks!

ppanchad-amd commented 3 weeks ago

@RambabuSwargam Closing ticket for now. Please feel free to re-open ticket if you still require assistance. Thanks!