Double to Half Error when Making
Issue #12
resolved
I am attempting to build MAGMA and getting the following error (see below). Is CUDA 9.2 required to build MAGMA now that half precision solvers have been added?
...
icpc -O3 -fPIC -qopenmp -DNDEBUG -DADD_ -Wall -Wshadow -DMAGMA_WITH_MKL -std=c++11 -DHAVE_CUBLAS -DMIN_CUDA_ARCH=300 -I/curc/sw/cuda/9.1.85/include -I/curc/sw/intel/17.4/mkl/include -I./include -I./control -c -o src/xhsgetrf_gpu.o src/xhsgetrf_gpu.cpp
src/xhsgetrf_gpu.cpp(109): error: no suitable constructor exists to convert from "double" to "__half"
const magmaHalf h_one = (magmaHalf) 1.0;
^
src/xhsgetrf_gpu.cpp(110): error: no suitable constructor exists to convert from "double" to "__half"
const magmaHalf h_neg_one = (magmaHalf)-1.0;
^
compilation aborted for src/xhsgetrf_gpu.cpp (code 2)
Comments (5)
-
-
Bump.
I’m having the same problem.
My setup is the same in software. Has a solution been pushed out?
-
Built MAGMA from source using the Bitbucket repo and I also built OpenBLAS from source (https://github.com/xianyi/OpenBLAS).
GCC version 6.5.0
CUDA 10 (since the hardware is an NVidia RTX 2070 card)
I had to set CUDADIR to /usr/local/cuda in my /etc/environment to source it in for root access when installing.
-
This should be fixed in MAGMA 2.5.1 alpha 1, or in the Bitbucket repo.
-
- changed status to resolved
Fixed in commit 0ed4a8b, etc.
- Log in to comment
I have the same problem. Ubuntu 16 64bit CUDA version 8 and 9, 3 TITAN GPU processors.
Being carefull in order to select the correct version of gcc and g++, the installation process goes on until at the end of 40 minutes it crashes out with this message
¿Could anyone help us? Thanks in advance!!