Artem Tamazov

Add support for package configuration files

1

This is required for https://github.com/ROCmSoftwarePlatform/MIOpen. Why we need rocm_smi_lib: - Beginning from ROCm 4.1 RC, MIOpen needs to inform offline compilers about target features (XNACK and SRAMECC). - Otherwise HIP/OCL...

comgr lacks MT safety

1

It seems like comgr is not working properly in multithreaded contexts. This prevents it from full-blown usage in MIOpen. ## How to reproduce Prerequisites: A linux machine with gfx900/906/908 GPU...

File does not exists: /usr/local/etc/cget/recipes/pcre/package.txt

1

Installation of MIOpen deps fail due to the above. `cget install pfultz2/cget-recipes` works but does not help. ```bash # cmake -P install_deps.cmake /usr/local/pip install cget Requirement already satisfied: cget in...

Generate .amdhsa_reserve_xnack_mask 0 in all assembly kernels.

4

This will prevent rolling back changes from https://github.com/ROCmSoftwarePlatform/MIOpen/pull/505 when regenerating kernels. /cc @daniellowell @shurale-nkn

[conv] miopenFindConvolution: Officially allow smaller workspace than GWSS returns.

2

This essentially re-implements "_Consider workspace constraints when loading solutions from DB (#2888)_" but in a bit stricter manner and with (hopefully) almost all necessary bells and whistles. ### Functional differences...

[softmax] Store full info about unpacked tensors in the problem descriptor

1

_Originated from: https://github.com/ROCm/MIOpen/pull/2782#discussion_r1515323515_ Currently we do not store full information about unpacked tensors, only the flag. Also we do not store information about tensor layouts, at all. The above means...

The original CMakeLists.txt has the detection of the "NO GPU case" and then ``` if(MIOPEN_NO_GPU) set(SKIP_ALL_EXCEPT_TESTS test_include_inliner test_kernel_build_params test_test_errors test_type_name test_tensor_test test_sqlite_perfdb test_sequences test_pooling3d test_perfdb) endif() ``` If we want...

Resolve post-merge review comments of PR #2320

5

- https://github.com/ROCmSoftwarePlatform/MIOpen/labels/urgency_high https://github.com/ROCmSoftwarePlatform/MIOpen/labels/request_for_comments Finally decide how to enable experimental API features - [ ] https://github.com/ROCmSoftwarePlatform/MIOpen/pull/2320#discussion_r1316419544 - https://github.com/ROCmSoftwarePlatform/MIOpen/labels/urgency_blocker Let's enable and properly test the experimental feature everywhere. - [x] https://github.com/ROCmSoftwarePlatform/MIOpen/pull/2320#discussion_r1326549529 -...

[MI200][MI100][MI300] UB of ConvHipImplicitGemmBwdXdlops

8

_Originated from: https://github.com/ROCmSoftwarePlatform/MIOpen/pull/1911#issuecomment-1533884703_ > ...it failed in `smoke_solver_ConvHipImplicitGemmBwdXdlops` test for the stage `Fp32 Hip Debug gfx90a`. Here is the error message: > > ``` > [2023-05-03T19:26:30.939Z] /home/jenkins/workspace/MLLibs_MIOpen_PR-1911/build/bin/test_conv2d --float --cmode conv...

System Find-db: 4 solvers limitation

1

@cderb To continue https://github.com/ROCmSoftwarePlatform/MIOpen/pull/2112#issuecomment-1544613299 (merged as #2151) The artificial limitation by 4 solvers reduces the useful effect of #2065. What we have now is better, but still similar as before,...