include/Gaudi/CUDAAsynchronousAlgHelper.cuh

0001 /***********************************************************************************\
0002 * (c) Copyright 2023-2024 CERN for the benefit of the LHCb and ATLAS collaborations *
0003 *                                                                                   *
0004 * This software is distributed under the terms of the Apache version 2 licence,     *
0005 * copied verbatim in the file "LICENSE".                                            *
0006 *                                                                                   *
0007 * In applying this licence, CERN does not waive the privileges and immunities       *
0008 * granted to it by virtue of its status as an Intergovernmental Organization        *
0009 * or submit itself to any jurisdiction.                                             *
0010 \***********************************************************************************/
0011
0012 #include <fmt/format.h>
0013
0014 #include <atomic>
0015 #include <memory_resource>
0016 #include <string>
0017
0018 namespace Gaudi {
0019   namespace CUDA {
0020     inline std::string SI( double number, std::string_view unit ) {
0021       static const std::string prefix[] = { "q", "r", "y", "z", "a", "f", "p", "n", "µ", "m", "",
0022                                             "k", "M", "G", "T", "P", "E", "Z", "Y", "R", "Q" };
0023       int                      idx      = std::floor( ( std::log10( number ) + 30 ) / 3 );
0024       if ( idx < 0 ) { idx = 0; }
0025       if ( idx > 20 ) { idx = 20; }
0026       // splitting 30 into 36 - 6 gives us 6 digits after the decimal point
0027       double reduced = std::round( number / std::pow( 10, 3 * idx - 36 ) ) / 1e6;
0028       if ( unit.size() <= 1 ) { return fmt::format( "{} {}{}", reduced, prefix[idx], unit ); }
0029       return fmt::format( "{} {}·{}", reduced, prefix[idx], unit );
0030     }
0031     std::string                err_fmt( cudaError_t err, std::string file, int line );
0032     cudaError_t                cuda_stream_await( cudaStream_t stream );
0033     std::pmr::memory_resource* get_pinned_memory_resource();
0034     namespace Detail {
0035       class PinnedMemoryResource : public std::pmr::memory_resource {
0036         void*              do_allocate( std::size_t bytes, std::size_t alignment ) override;
0037         void               do_deallocate( void* p, std::size_t bytes, std::size_t alignment ) override;
0038         bool               do_is_equal( const std::pmr::memory_resource& other ) const noexcept override;
0039         std::atomic_size_t num_allocs   = 0;
0040         std::atomic_size_t num_deallocs = 0;
0041
0042       public:
0043         ~PinnedMemoryResource() {
0044 #ifndef NDEBUG
0045           fmt::print( "Allocated {} times and deallocated {} times\n", num_allocs.load(), num_deallocs.load() );
0046 #endif
0047         }
0048       };
0049     } // namespace Detail
0050
0051   } // namespace CUDA
0052 } // namespace Gaudi