-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathREADME.PPC64le
28 lines (22 loc) · 997 Bytes
/
README.PPC64le
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
PoCL is used to provide OpenCL on IBM AC922 computers
featuring IBM Power9 processors and Nvidia Tesla V100 GPU
interconnected with NVlink v2 (up to 72 GByte/s).
This has been tested under debian_11 and Ubuntu_20.04.
Officially, Nvidia does not support OpenCL on this platform
and the driver they are shipping is lacking the compiler part.
## Building tricks (as of 04/2023):
The PPC64le features 128-bit vector unit (Altivec/VSX) which
are easily confused by the C++ compiler with the C++ vector
instruction when using the compile option `-std=c++XX`.
The corresponding code usually fails compiling.
The trick is to pass the option `-std=gnu++XX`.
For example, when configuring pocl:
```
cmake .. -DLLVM_CXXFLAGS="-std=gnu++14 ..." -DENABLE_CUDA=ON
```
The full list of options for the CXXFLAGS is obtained with:
```
llvm-config --cxxflags|sed -e "s/std=c/std=gnu/"
```
Later on, the build continues with `make` ...
Note the CUDA option to enable the GPU support on those computers.