Date: 2022-07-01
| software | description |
|---|---|
| OS | Ubuntu-20.04.4 |
| Python | 3.8.10 |
| ROCm | 5.2.0 |
| hardware | Product Name | ISA | CHIP IP |
|---|---|---|---|
| CPU | Xeon 2620v3 | ||
| GPU | RX580 8G | gfx803(Polaris10) | 0x67df |
The ubuntu-22.04 patches put in the patch directory, with ubuntu2204 in file name. https://github.com/xuhuisheng/rocm-build/blob/develop/patch
ROCm-5.1.3 cannot install properly on ubuntu-22.04. We need some patches. Some patches comes from ROCm-5.2 develop branch.
The main different parts.
| # | ubuntu-20.04 | ubuntu-22.04 | |
|---|---|---|---|
| 1 | gcc-9 | gcc-11 | many components |
| 2 | python-3.8 | python-3.10 | tensorflow, pytorch |
| 3 | fmt-6 | fmt-8 | rocSOLVER |
- comment out thread header. No idea why this caused miopen native_conv.cpp error
- comment out complex, new header. No idea why this caused miopen native_conv.cpp error
- add thread header. because of hip modification
- (ROCm-5.2.0 already fix this)
- replace pthread_yield to sched_yield for strict check
- (ROCm-5.2.0 already fix this)
- update for fmt8 https://github.com/ROCmSoftwarePlatform/rocSOLVER/commit/2bbfb8976f6e4d667499c77e41a6433850063e88
- upgrade googletest 1.11.0 ROCm/ROCmValidationSuite#543
miopen depends boost-1.72.0. We need remove PTHREAD_STACK_MIN related lines from boost/thread/pthread/thread_data.hpp
The breakpad from tensorflow throw a compiling error about max(int, long int) type cast.
The pytorch need compile with USE_ROCM=1 USE_NUMPY=1 ROCM_VERSION=50103 python3 setup.py bdist_wheel