Please add NCCL support for complex data types. I am currently converting complex data I’m using in cuSOLVER and cuBLAS to real numbers to send using NCCL. However, this requires twice the amount of allocated device memory, and I am running out. I am using CUDA Fortran along with its interfaces to the libraries. I have also tried CUDA-aware MPI, but am having troubles with that. NCCL seems to work better. Thank you.
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| LIbrary for complex numbers in device kernel | 8 | 3318 | December 9, 2016 | |
| cuBLAS for lower-end GPUs | 1 | 575 | May 20, 2016 | |
| Is there a NCCL 2.x for Windows? | 4 | 6410 | August 1, 2020 | |
| how to generate a complex double array output in cuda kernel | 3 | 1696 | November 20, 2012 | |
| Current state of Device Extension libraries | 0 | 725 | March 3, 2023 | |
| Using CUDA for (or to replace) Successive over-relaxation solver | 3 | 639 | April 23, 2019 | |
| opencl support for nvidia 8200M | 3 | 903 | August 20, 2011 | |
| How can I use "cuComplex" type? | 1 | 5776 | September 30, 2008 | |
| Complex Sparse Matrix Linear Solver | 0 | 2286 | December 13, 2008 | |
| Fortran, CUDA and OpenCL : Where are we? | 1 | 1140 | June 10, 2009 |