MIOpen
MIOpen copied to clipboard
ROCmSoftwarePlatform
AMD's Machine Intelligence Library
I'm very confused at the experiments conducted below. This config belong to Resnet50 therefore the workload get tested here is fairly important in its performance signature. Note: To reproduce please...
I encountered bn configs failing on gfx908 while using the develop branch. Here are sample configs that failed (these are not all) ./bin/MIOpenDriver bnormfp16 -n 256 -c 64 -H 112...
**./bin/MIOpenDriver conv -n 16 -c 76 -H 9 -W 9 -k 32 -y 1 -x 1 -p 0 -q 0 -u 1 -v 1 -l 1 -j 1 -m conv...
`miopenBatchNormalizationForwardInference` function is failing when the nullptr is passed for mean & variance. https://github.com/GPUOpen-ProfessionalCompute-Libraries/MIVisionX/blob/master/amd_openvx_extensions/amd_nn/src/batch_normalization_layer.cpp#L121 The above function works fine but https://github.com/GPUOpen-ProfessionalCompute-Libraries/MIVisionX/blob/master/amd_openvx_extensions/amd_nn/src/scale_layer.cpp#L103 this function returns Inf values. They both call the...
> > > @atamazov @JehandadKhan @asroy @zjing14 @qianfengz could you take a look at CK related changes? The composable kernel-related changes in this PR might have affected accuracy: > >...
* This PR finds the performance hipMemcpy, hipMemset and SetTensor The performance results for tensor {4, 4, 1115, 1115} in micro sec were : hipMemcpy = 29717 hipMemset = 1...
**[Keywords]:** test; vega20; batchnorm **URGENCY :HIGH** since it is already in the develop branch **[Env]:** vega20 mainly Sometimes gfx908 **[Description]:** This test fails frequently (and usually pass with rerun): http://micimaster.amd.com/blue/organizations/jenkins/MLLibs%2FMIOpen/detail/cpu_conv_acc_tpye_fix/1/pipeline...
This is a debugging ticket brought up from @krzysz00 in MLIR team. In using the MIOpenDriver to help MLIR tune different configs, it is common for MIOpen to pick up...
https://github.com/ROCmSoftwarePlatform/MIOpen/pull/1316#pullrequestreview-869371563 Let's discuss comments there until all those resolved or fixed (we'll need separate PR for the fixes).
