QUDA Environment Variables - lattice/quda GitHub Wiki

Below is a list of QUDA specific environment variables. We also include a list of useful CUDA specific variables at the bottom of the page.

Variable name	Function
QUDA_RESOURCE_PATH	Path where tune cache and profile files will be output
QUDA_PROFILE_OUTPUT_BASE	Filename prefix for profile output. Setting this will result in the files $(QUDA_PROFILE_OUTPUT_BASE).tsv and $(QUDA_PROFILE_OUTPUT_BASE_async).tsv and being written out (default is simply profile.tsv and profile_async.tsv)
QUDA_ENABLE_P2P	QUDA_ENABLE_P2P=0 # disable all p2p transfersQUDA_ENABLE_P2P=1 # enable only copy enginesQUDA_ENABLE_P2P=2 # enable only remote writingQUDA_ENABLE_P2P=3 # enable both copy engines and remote writing Default is enabling copy engines and remote writing (3)
QUDA_ENABLE_P2P_MAX_ACCESS_RANK	Set limit on which GPUs are peer-to-peer connected (use to disable low-bandwidth connections), e.g., QUDA_ENABLE_P2P_MAX_ACCESS_RANK=0 would limit to only highest bandwidth connections
QUDA_ENABLE_TUNING	Enable / disable kernel autotuning. Default is enabled, disable with QUDA_ENABLE_TUNING=0
QUDA_REORDER_LOCATION	Set where data should be reordered when transferring CPU<->GPU (default is GPU)
QUDA_RANK_VERBOSITY	Set which global ranks are active in printfQuda calls (default is rank 0)
QUDA_ENABLE_DEVICE_MEMORY_POOL	Enable / disable device memory allocator (default is enabled, disable with QUDA_ENABLE_DEVICE_MEMORY_POOL=0
QUDA_ENABLE_PINNED_MEMORY_POOL	Enable / disable device memory allocator (default is enabled, disable with QUDA_ENABLE_PINNED_MEMORY_POOL=0
QUDA_ENABLE_MANAGED_MEMORY	Enable / disable using managed memory for allocations (default is disabled, enable with QUDA_ENABLE_MANAGED_MEMORY=1). Note: managed memory has some limitations for pre-Pascal architectures.
QUDA_ENABLE_MANAGED_PREFETCH	Enable / disable explicit managed memory prefetching calls (default is disabled, enable with QUDA_ENABLE_MANAGED_PREFETCH=1). Does nothing if QUDA_ENABLE_MANAGED_MEMORY isn't enabled.
QUDA_ENABLE_NUMA	Enabled NUMA placement. Default is enabled, if NUMA has been enabled in cmake, disabled with QUDA_ENABLE_NUMA=0
QUDA_MILC_HISQ_RECONSTRUCT	Set the reconstruct type in the MILC interface used for the long links in the HISQ solver. Allowed values are 9/13/18 with 18 the default
QUDA_MILC_HISQ_RECONSTRUCT_SLOPPY	Set the sloppy reconstruct type in the MILC interface used for the long links in the HISQ solver. Allowed values are 9/13/18 with 18 the default
QUDA_ENABLE_GDR	Enable GPU-Direct RDMA. Default is disabled, enabled with QUDA_ENABLE_GDR=1
QUDA_ENABLE_NVSHMEM	Enable NVSHMEM communication policies if QUDA is build with NVSHMEM support. Default is enabled, set to 0 to disable.
QUDA_TEST_GRID_SIZE	Set the process geometry for the unit tests. Overrides the --gridsize parameter if set.
QUDA_TEST_GRID_PARTITION	Set the process grid partition geometry for the unit tests (for split grid). Overrides the --grid-partition parameter if set.
QUDA_ENABLE_MPS	Enable support for MPS in QUDA. Generally not recommended except for testing purposes. Default is disabled, enable with QUDA_ENABLE_MPS=1
QUDA_DEVICE_RESET	Call cudaDeviceReset in endQuda - this legacy behavior can be useful for profiling, but destroys the CUDA context of other CUDA libraries outside of QUDA (e.g., GPU-aware MPI). Default is disabled, enable with QUDA_DEVICE_RESET=1
QUDA_DETERMINISTIC_REDUCE	Perform all MPI reductions deterministically: setting this flag means that post-tuning or no tuning, QUDA will run completely deterministically regardless of the rank order. Default is disabled, enable with QUDA_DETERMINISTIC_REDUCE=1
QUDA_TUNE_VERSION_CHECK	Set QUDA_TUNE_VERSION_CHECK=0 to disable the check that prevents using a tunecache.tsv file from a different QUDA version
QUDA_ENABLE_TUNING_SHARED	Disable shared memory autotuning. Useful for checking the effect of this.
QUDA_TUNING_RANK	Set the global default rank for doing kernel autotuning (default is rank 0)

CUDA environment variables

Variable name	Function
CUDA_LAUNCH_BLOCKING	If set to =0 (default behaviour) this will ensure that all kernels are launched synchronously. If set to =1, kernels are launched asynchronously. This will ensure that that error messages pertain to precisely the last kernel called