File: README.PPC64le

package info (click to toggle)
pocl 6.0-7
  • links: PTS, VCS
  • area: main
  • in suites: forky, sid
  • size: 25,320 kB
  • sloc: lisp: 149,513; ansic: 103,778; cpp: 54,947; python: 1,513; sh: 949; ruby: 255; pascal: 226; tcl: 180; makefile: 175; java: 72; xml: 49
file content (28 lines) | stat: -rw-r--r-- 997 bytes parent folder | download | duplicates (2)
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.