You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Notice: You CANNOT use `icpx` compiler for GPU version of ABACUS for now, see discussion here [#2906](https://github.com/deepmodeling/abacus-develop/issues/2906) and [#4976](https://github.com/deepmodeling/abacus-develop/issues/4976)
271
271
272
-
If you wants to use ABACUS GPU-LCAO by `cusolvermp` or `elpa`, please contact the coresponding developer, toolchain do not fully support them now.
272
+
If you wants to use ABACUS GPU-LCAO by "cusolvermp" or "elpa", please compile according to the following usage:
273
+
274
+
Firstly, in cmake, it is necessary to add `-DUSE_CUDA=ON`, which is necessary for compiling NVIDIA GPUs.
275
+
1. For the elpa method, add
276
+
```shell
277
+
--enable-cuda
278
+
--gpu-ver=GPU name
279
+
export CUDA-PATH=/path/to/CUDA
280
+
```
281
+
to the `toolchain_gnu.sh`, and then follow the normal step to install the dependencies using `./toolchain_gnu.sh`.
282
+
Afterwards, add the link files corresponding to
283
+
```shell
284
+
-DUSE_ELPA=ON \
285
+
-DELPA_LINK-NLIBRARIES=/path/to/lib
286
+
```
287
+
in the `build_abacus_gnu.sh` file, just build the abacus executable program by compiling it with `./build_abacus_gnu.sh`.
288
+
289
+
2. For the cusolvermp method, toolchain_gnu.sh does not need to be changed, just follow it directly install dependencies using `./toolchain_gnu.sh`, and then add
0 commit comments