MIOpen

MIOpen copied to clipboard

MIOpen cache default location issue

15

[Issue] For some cluster environments, $HOME are accessible to log-in nodes but not compute nodes [Proposal] - [ ] 1. emit warning messages when $HOME/.cache/miopen is not accessible - [x]...

junliume

[ImplicitGemm] Revise workarounds

3

I guess it's time to find out if these W/A are still necessary in 4.3.1 (which we currently run CI on) and disable them. This is the key point for...

atamazov

tensortransform operation -- doc is missing

8

l can't find the corresponding API to `cudnnTransformTensor()` in miopen library, can anyone please help me?

MOLOjl

Do we need many members in PerformanceConfigAsmImplicitGemmGTC?

3

Originated from https://github.com/ROCmSoftwarePlatform/MIOpen/pull/1230#discussion_r737878197 (see the whole thread). Synopsis: > As far as I see, for PerformanceConfigAsmImplicitGemmGTCFwdXdlopsNHWC we only need to store table index and `gemm_k_global_split`. The rest of data can...

atamazov

Fused FP16 `miopenGcnAsmConv1x1U` verification issue

1

- rocm 4.0.0 - gfx906 ```bash export MIOPEN_FIND_MODE=1 export MIOPEN_LOG_LEVEL=7 export MIOPEN_DEBUG_CONV_DIRECT=1 export MIOPEN_DEBUG_CONV_WINOGRAD=0 export MIOPEN_DEBUG_AMD_FUSED_WINOGRAD=0 export MIOPEN_DEBUG_CONV_FFT=0 export MIOPEN_DEBUG_CONV_GEMM=0 export MIOPEN_DEBUG_CONV_SCGEMM=0 export MIOPEN_DEBUG_CONV_IMPLICIT_GEMM=0 ./MIOpenDriver CBAInferfp16 -W 4 -H 4...

Slimakanzer

[gfx90a][COMGR][Fp16][Bf16] Verification errors of ImplicitGemm kernels (no errors with hip-clang)

15

Bf16 and Fp16 Full tests fail on gfx90a when library is built with COMGR. This prevents from switching fully to COMGR builds by default. Some tests PASS with hip-clang, FAIL...

atamazov

Get rid of W/A: HIP kernels disabled for embedded builds

1

There were times when HIP compilation didn't work in embedded MIOpen, because precompiled HIP headers weren't available. https://github.com/ROCmSoftwarePlatform/MIOpen/blob/5cb2e548131d777d589917db8a0d8669b3ae8c5d/CMakeLists.txt#L132-L135 This workaround leads to performance drop of embedded MIOpen. We should remove...

atamazov

Excessive number of file handles opened and held onto during Tuning

1

As reported in [SWDEV-304151](https://ontrack-internal.amd.com/browse/SWDEV-304151) when tuning using MIGraphX as the driver framework, a large number (more than 10k) of file handles are created and never released. One of the possible...

JehandadKhan

"Unable to read system database file" warning

2

If the user doesn't have the MIOpen kernels package installed, they see a message such as below: `MIOpen(HIP): Warning [SQLiteBase] Unable to read system database file:gfx906_60.kdb Performance may degrade` This...

jithunnair-amd

Dedicated primitive (e.g. `Fused`) for fused convolutions?

2

https://github.com/ROCmSoftwarePlatform/MIOpen/blob/4e61a3ebdfe2b07b4d331cc46832eb4a6b49941c/src/solver.cpp#L264 Can we introduce some dedicated primitive (e.g. `Fused`) for fused convolutions and use it in the registry? This would help to avoid hacks like this: https://github.com/ROCmSoftwarePlatform/MIOpen/pull/1212/files/db05b058f77cc838f5bbafc93fe08fd8e20c713d..f26346d12ba5f2ee472d26d7e6c207325bbe6233#r723662796

atamazov