params.h

Go to the documentation of this file.

#ifndef PARAMS_H_
#define PARAMS_H_
/**
 * @file params.h
 * @brief This file contains #defined constants
 * @details
 * These can be overruled in application Makefile, with APP_OPTS := -DPARAMETER=value.
 *
 * There are two types of #define variables, True/False switches or parameter variables.
 *
 * True/False statements can be set with either 0 (False) or 1 (True) as -DPARAMETER=0.
 *
 * Parameter variables are set similary with -DPARAMETER=var where var is the chosen variable
 */
 
#ifdef RELEASE
/**
 * @brief Turn off asserts which are on by default.
 * @details By defining either RELEASE or NDEBUG (No debug) asserts will be turned off.
 * Static asserts naturally remain active
 */
#ifndef NDEBUG
#define NDEBUG
#endif
#endif
 
#ifndef NDIM
/**
 * @brief HILA system dimensionality
 * @details Set's HILA dimensionality for which 4 is default. Options are 2,3,4
 */
#define NDIM 4
#endif
 
#ifndef DEFAULT_OUTPUT_NAME
/**
 * @def DEFAULT_OUTPUT_NAME
 * @brief Default output file name
 */
#define DEFAULT_OUTPUT_NAME "output"
#endif
 
#ifndef EVEN_SITES_FIRST
/**
 * @brief  EVEN_SITES_FIRST is default. To traverse odd sites first set -DEVEN_SITES_FIRST=0
 */
#define EVEN_SITES_FIRST
#elif EVEN_SITES_FIRST == 0
#undef EVEN_SITES_FIRST
#endif
 
/// NODE_LAYOUT_TRIVIAL or NODE_LAYOUT_BLOCK determine how MPI ranks are laid out on logical
/// lattice.  TRIVIAL lays out the lattice on logical order where x-direction runs fastest etc.
/// if NODE_LAYOUT_BLOCK is defined, NODE_LAYOUT_BLOCK consecutive MPI ranks are laid out so that
/// these form a compact "block" of ranks logically close togeter.
/// Define NODE_LAYOUT_BLOCK to be the number of
/// MPI processes within one compute node - tries to maximize the use of fast local communications.
/// Either one of these must be defined.
 
#ifndef NODE_LAYOUT_TRIVIAL
#ifndef NODE_LAYOUT_BLOCK
#define NODE_LAYOUT_BLOCK 4
#endif
#endif
 
// Size of the write buffer in field writes, in bytes
// Larger buffer -> less MPI calls in writing, but more memory
#ifndef WRITE_BUFFER_SIZE
#define WRITE_BUFFER_SIZE 2000000
#endif
 
 
// boundary conditions are "off" by default -- no need to do anything here
// #ifndef SPECIAL_BOUNDARY_CONDITIONS
 
///////////////////////////////////////////////////////////////////////////
// Special defines for GPU targets
#if defined(CUDA) || defined(HIP)
 
// Use gpu memory pool by default
// set off by using -DGPU_MEMORY_POOL=0 in Makefile
#ifndef GPU_MEMORY_POOL
#define GPU_MEMORY_POOL
#elif GPU_MEMORY_POOL == 0
#undef GPU_MEMORY_POOL
#endif
 
// Undef cuda/hip -aware mpi at makefile with -DGPU_AWARE_MPI=0
#ifndef GPU_AWARE_MPI
#define GPU_AWARE_MPI 1
#elif GPU_AWARE_MPI == 0
#undef GPU_AWARE_MPI
#endif
 
// GPU_RNG_THREAD_BLOCKS
// Number of thread blocks (of N_threads threads) to use in onsites()-loops containing random
// numbers. GPU_RNG_THREAD_BLOCKS=0 or undefined means use one RNG on each lattice site, and the
// thread block number is not restricted.  RNG takes about 48 B/generator (with XORWOW). When
// GPU_RNG_THREAD_BLOCKS > 0 only (N_threads * GPU_RNG_THREAD_BLOCKS) generators are in use, which
// reduces the memory footprint substantially (and bandwidth demand) Too small number slows down
// onsites()-loops containing RNGs, because less threads are active. Example:
//     Field<Vector<4,double>> vfield;
//     onsites(ALL) {
//        vfield[X].gaussian_random();      // there's RNG here, so this onsites() is handled by
//                                          // GPU_RNG_THREAD_BLOCKS thread blocks
//     }
// GPU_RNG_THREAD_BLOCKS<0 disables GPU random numbers entirely, and loops like above will crash if
// executed. hilapp will emit a warning, but program is compiled
 
#ifndef GPU_RNG_THREAD_BLOCKS
#define GPU_RNG_THREAD_BLOCKS 32
#endif
 
// GPU_VECTOR_REDUCTION_THREAD_BLOCKS
// # of thread blocks (of N_threads threads) used in ReductionVector (weighted histogram) ops.
// A value > 0 for GPU_VECTOR_REDUCTION_THREAD_BLOCKS means that the onsites-loop where the
// reduction is done is handled by GPU_VECTOR_REDUCTION_THREAD_BLOCKS thread blocks of N_threads
// threads.  Each thread handles its own histogram, thus there are
// (GPU_VECTOR_REDUCTION_THREAD_BLOCKS*N_threads) working copies of the histogram which are then
// combined. Too small value slows the loop where this happens computation, too large uses
// (temporarily) more memory. Example:
//      ReductionVector<double> rv(100);
//      Field<int> index;
//      ... (set index to values 0 .. 99)
//      onsites(ALL) {
//           rv[index[X]] += ..
//           ..
//      }
//
// GPU_VECTOR_REDUCTION_THREAD_BLOCKS = 0 or undefined means that the thread block number is not
// restricted and only a single histogram is used with atomic operations (atomicAdd).  This
// can be slower, but the performance is GPU hardware/driver dependent.  In some
// cases GPU_VECTOR_REDUCTION_THREAD_BLOCKS = 0 turns out to be faster.
//
// Default: 32 is currently OK compromise (32 thread blocks)
 
#ifndef GPU_VECTOR_REDUCTION_THREAD_BLOCKS
#define GPU_VECTOR_REDUCTION_THREAD_BLOCKS 32
#endif
 
// GPUFFT_BATCH_SIZE:
// How many complex fft's in parallel - large value faster, small less memory.
// Performance is reduced if the value is too small, but levels to a ~constant
// when sufficiently large.
#ifndef GPUFFT_BATCH_SIZE
#define GPUFFT_BATCH_SIZE 256
#endif
 
/** @brief GPU_SYNCHRONIZE_TIMERS : if set and !=0 synchronize GPU on timer calls, in order to
 * obtain meaningful timer values
 *
 * @details Because GPU kernel launch is asynchronous process, the timers by default may not measure
 * the actual time used in GPU kernel execution. Defining GPU_SYNCHRONIZE_TIMERS inserts GPU
 * synchronization calls to timers.  This is off by default, because this may slow down GPU code.
 * Turn on in order to measure more accurately the time spent in different parts of the code.
 */
 
#ifdef GPU_SYNCHRONIZE_TIMERS
#if GPU_SYNCHRONIZE_TIMERS == 0
#undef GPU_SYNCHRNONIZE_TIMERS
#endif
#endif
 
 
/** @brief GPU_GLOBAL_ARG_MAX_SIZE : in some __global__functions gives the max size of variable
 * passed directly as an argument of the function call. Larger value sizes are passed with
 * gpuMemcopy() and a pointer. CUDA < 12.1 limits the total parameter size to 4K, >= 12.1 it is 32K.
 * We set the default to 2K. in HIP/rocm I have not found the size. Passing as an arg is faster, but
 * size limit is relevant only for big "matrices"
 */
 
#ifndef GPU_GLOBAL_ARG_MAX_SIZE
#define GPU_GLOBAL_ARG_MAX_SIZE 2048
#endif
 
#endif // CUDA || HIP
 
///////////////////////////////////////////////////////////////////////////
// Special defines for CUDA target
 
#if defined(CUDA)
 
// General # of threads
#ifndef N_threads
#define N_threads 256
#endif
 
 
#ifndef GPU_MEMORY_POOL
 
// CUDA_MALLOC_ASYNC
#ifndef CUDA_MALLOC_ASYNC
// Use async malloc only if version is large enough
// NOTE: does not seem to work with OpenMPI, disable
#if 0 && CUDART_VERSION >= 11020
#define CUDA_MALLOC_ASYNC
#endif
 
#elif CUDA_MALLOC_ASYNC == 0
#undef CUDA_MALLOC_ASYNC
#endif
 
#endif // if not GPU_MEMORY_POOL
 
#endif // CUDA
 
///////////////////////////////////////////////////////////////////////////
// Same for HIP
 
#if defined(HIP)
 
// General # of threads
#ifndef N_threads
#define N_threads 256
#endif
 
 
// End of GPU defines
#endif // HIP
 
#endif