lammps cuda arch=21?

LAMMPS LAMMPS Mailing List Mirror
1

Hello,

I noticed when making the lammps cuda library that if I run:
$ make arch=21
the arch reverts to 20.

This is in line with what is in “lammps/lib/cuda/makefile.common” (see below).

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
CUDA_FLAGS += -DCUDA_ARCH=99
SMVERSIONFLAGS := -arch sm_13
endif
endif
endif

I noticed this for both the 13Dec11 and 17Jan12 versions of lammps.
Is there any reason for this?

Thanks,

– WM

2

Hello,

I noticed when making the lammps cuda library that if I run:
$ make arch=21
the arch reverts to 20.

This is in line with what is in "lammps/lib/cuda/makefile.common" (see
below).

# 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
CUDA_FLAGS += -DCUDA_ARCH=99
SMVERSIONFLAGS := -arch sm_13
endif
endif
endif

I noticed this for both the 13Dec11 and 17Jan12 versions of lammps.
Is there any reason for this?

they obviously don't need differentiation at the source code level.

axel.

3

Hi

Axel is correct, there is no source code level difference for CC21 and CC20. Hence the only difference is the actual compiler flag for producing code for devices with CC21 ( the -arch_sm=21 flag, for nvcc), while the source code level path is that for all fermi level GPUs.

Cheers
Christian

-------- Original-Nachricht --------

4

Christian can comment if this is a typo in the makefile,
or as Axel says, that there is no need to distinguish
between the two.

Steve