USER-CUDA questions

LAMMPS LAMMPS Mailing List Mirror
1

When I compile using Make.py and specifiy mode=single [or double] arch=35 I still observe this flying by CUDA_ARCH=20?

mpicxx -g -O3 -DLAMMPS_GZIP -I…/…/lib/cuda -DLMP_USER_CUDA -DMPICH_SKIP_MPICXX -DOMPI_SKIP_MPICXX=1 -I/cm/shared/apps/cuda50/toolkit/5.0.35/include -DUNIX -DFFT_CUFFT -DCUDA_PRECISION=1 -DCUDA_ARCH=20 -M …/pair.cpp > pair.d

(btw what are the keywords or precision 3&4?)

When I edit the lib/cudaMakefile.defaults and compile using make auto; make yes-user-cuda;make mpi I observe the same. Is there a way to verify what arch and precision modes where set?

-Henk

2

See doc/accelerate_cuda.html for details on this.

If you build the lib/cuda manually you do something like this;

make precision=1,2,4 arch=20,35,etc

You should see those choices reflected in the compile lines (for the lib files)
and written into 2 files in lib/cuda - Makefile.defaults and Makefile.lammps

If you then build LAMMPS itself with the USER-CUDA package, it takes
those precision/arch settings from the lib/cuda/Makefile.lammps file, so
you should also see them reflected in the LAMMPS compile lines.

Make.py is simply doing both steps for you. Underneath it is doing
the same thing.

Steve

3

Sir, I am also facing the same problem. i think there is bug in Makefile.common:

given below is line 83-118 of file Makefile.common in lib/cuda dir
As you can see irrespective of choice given, code below will assign -DCUDA_ARCH=20 flag

make architecture settings

ifeq ((strip (arch)), 13)
CUDA_FLAGS += -DCUDA_ARCH=13
SMVERSIONFLAGS := -arch sm_13
else
ifeq ((strip (arch)), 20)
CUDA_FLAGS += -DCUDA_ARCH=20
#NVCC_FLAGS += -ftz=false -prec-div=true -prec-sqrt=true
NVCC_FLAGS += -ftz=true -prec-div=false -prec-sqrt=false
SMVERSIONFLAGS := -arch sm_20
else
ifeq ((strip (arch)), 21)
CUDA_FLAGS += -DCUDA_ARCH=20
#NVCC_FLAGS += -ftz=false -prec-div=true -prec-sqrt=true
NVCC_FLAGS += -ftz=true -prec-div=false -prec-sqrt=false
SMVERSIONFLAGS := -arch sm_21
else
ifeq ((strip (arch)), 30)
CUDA_FLAGS += -DCUDA_ARCH=20
#NVCC_FLAGS += -ftz=false -prec-div=true -prec-sqrt=true
NVCC_FLAGS += -ftz=true -prec-div=false -prec-sqrt=false
SMVERSIONFLAGS := -arch sm_30
else
ifeq ((strip (arch)), 35)
CUDA_FLAGS += -DCUDA_ARCH=20
#NVCC_FLAGS += -ftz=false -prec-div=true -prec-sqrt=true
NVCC_FLAGS += -ftz=true -prec-div=false -prec-sqrt=false
SMVERSIONFLAGS := -arch sm_35
else
CUDA_FLAGS += -DCUDA_ARCH=99
SMVERSIONFLAGS := -arch sm_13
endif

Once i edited it, its fine for me. (line 107, change 20 to 35)

4

ah, will try that…now I get confirmation upon compilation that the right arch is picked, thanks, btw there are other errors in that code block

to respond to Steve, the page is a bit incomplete, for example the doc states ‘mode=single’ but ‘mode’ does not appear as option to switch -cuda, so is ‘single’ precision=1? when I use mode=4 or precision=4 with Make.py I receive invalid args error…ah, the mode option are ‘single,double,mixed’ (buried in code)

also in the manual section the first command ‘make’ will just display the help page, I believe it should be ‘make auto’

Thanks all, I’m able to compile now.

-Henk