eic-opticks/qudarap/QU.hh

0001 #pragma once
0002 #include "QUDARAP_API_EXPORT.hh"
0003 #include "plog/Severity.h"
0004 #include <vector>
0005 #include <cstdint>
0006
0007 struct NP ;
0008 struct dim3 ;
0009 struct salloc ;
0010
0011 struct QUDARAP_API QU
0012 {
0013     static const plog::Severity LEVEL ;
0014
0015     static constexpr const char* _MEMCHECK = "QU__MEMCHECK" ;
0016     static bool MEMCHECK ;
0017
0018     static salloc* alloc ;
0019
0020     static void alloc_add(const char* label, uint64_t num_items, uint64_t sizeof_item );
0021
0022     template <typename T>
0023     static char typecode() ;
0024
0025     template <typename T>
0026     static std::string rng_sequence_reldir(const char* prefix, unsigned ni, unsigned nj, unsigned nk, unsigned ni_tranche_size );
0027
0028     template <typename T>
0029     static std::string rng_sequence_name(const char* prefix, unsigned ni, unsigned nj, unsigned nk, unsigned ioffset );
0030
0031     template <typename T>
0032     static T* UploadArray(const T* array, unsigned num_items, const char* label ) ;
0033
0034     template <typename T>
0035     static T* DownloadArray(const T* array, unsigned num_items ) ;
0036
0037     template <typename T>
0038     static void Download(std::vector<T>& vec, const T* d_array, unsigned num_items) ;
0039
0040
0041     static const char* _cudaMalloc_OOM_NOTES ;
0042     static void _cudaMalloc( void** p2p, size_t size, const char* label );
0043
0044
0045     template <typename T>
0046     static T*   device_alloc( unsigned num_items, const char* label ) ;
0047
0048     template <typename T>
0049     static T*   device_alloc_zero( unsigned num_items, const char* label ) ;
0050
0051     template <typename T>
0052     static void device_memset( T* d, int value, unsigned num_items );
0053
0054     template <typename T>
0055     static void device_free( T* d ) ;
0056
0057     template <typename T>
0058     static void device_free_and_alloc(T** dd, unsigned num_items );  // dd : pointer-to-device-pointer
0059
0060     template <typename T>
0061     static int copy_device_to_host( T* h, T* d,  unsigned num_items);
0062
0063     template <typename T>
0064     static void copy_device_to_host_and_free( T* h, T* d,  unsigned num_items, const char* label );
0065
0066     template <typename T>
0067     static void copy_host_to_device( T* d, const T* h,  unsigned num_items);
0068
0069     template <typename T>
0070     static unsigned NumItems( const NP* a );
0071
0072     template <typename T>
0073     static unsigned copy_host_to_device( T* d, const NP* a);
0074
0075
0076     static void ConfigureLaunch16( dim3& numBlocks, dim3& threadsPerBlock );
0077     static void ConfigureLaunch( dim3& numBlocks, dim3& threadsPerBlock, unsigned width, unsigned height );
0078     static void ConfigureLaunch2D( dim3& numBlocks, dim3& threadsPerBlock, unsigned width, unsigned height );
0079     static void ConfigureLaunch1D( dim3& numBlocks, dim3& threadsPerBlock, unsigned num, unsigned threads_per_block );
0080
0081     static std::string Desc(const dim3& d, int w);
0082     static std::string DescLaunch( const dim3& numBlocks, const dim3& threadsPerBlock );
0083 };
0084