cpp_thread_test/dgemm_thread_safety.cpp

   1 #include <iostream>
   2 #include <vector>
   3 #include <random>
   4 #include <future>
   5 #include <omp.h>
   6 #include "../cblas.h"
   7 #include "cpp_thread_safety_common.h"
   8
   9 void launch_cblas_dgemm(double* A, double* B, double* C, const blasint randomMatSize){
  10         cblas_dgemm(CblasColMajor, CblasNoTrans, CblasNoTrans, randomMatSize, randomMatSize, randomMatSize, 1.0, A, randomMatSize, B, randomMatSize, 0.1, C, randomMatSize);
  11 }
  12
  13 int main(int argc, char* argv[]){
  14         blasint randomMatSize = 1024; //dimension of the random square matrices used
  15         uint32_t numConcurrentThreads = 52; //number of concurrent calls of the functions being tested
  16         uint32_t numTestRounds = 16; //number of testing rounds before success exit
  17
  18         if (argc > 4){
  19                 std::cout<<"ERROR: too many arguments for thread safety tester"<<std::endl;
  20                 abort();
  21         }
  22
  23         if(argc == 4){
  24                 std::vector<std::string> cliArgs;
  25                 for (int i = 1; i < argc; i++){
  26                         cliArgs.push_back(argv[i]);
  27                         std::cout<<argv[i]<<std::endl;
  28                 }
  29                 randomMatSize = std::stoul(cliArgs[0]);
  30                 numConcurrentThreads = std::stoul(cliArgs[1]);
  31                 numTestRounds = std::stoul(cliArgs[2]);
  32         }
  33
  34         std::uniform_real_distribution<double> rngdist{-1.0, 1.0};
  35         std::vector<std::vector<double>> matBlock(numConcurrentThreads*3);
  36         std::vector<std::future<void>> futureBlock(numConcurrentThreads);
  37
  38         std::cout<<"*----------------------------*\n";
  39         std::cout<<"| DGEMM thread safety tester |\n";
  40         std::cout<<"*----------------------------*\n";
  41         std::cout<<"Size of random matrices(N=M=K): "<<randomMatSize<<'\n';
  42         std::cout<<"Number of concurrent calls into OpenBLAS : "<<numConcurrentThreads<<'\n';
  43         std::cout<<"Number of testing rounds : "<<numTestRounds<<'\n';
  44         std::cout<<"This test will need "<<(static_cast<uint64_t>(randomMatSize*randomMatSize)*numConcurrentThreads*3*8)/static_cast<double>(1024*1024)<<" MiB of RAM\n"<<std::endl;
  45
  46         std::cout<<"Initializing random number generator..."<<std::flush;
  47         std::mt19937_64 PRNG = InitPRNG();
  48         std::cout<<"done\n";
  49
  50         std::cout<<"Preparing to test CBLAS DGEMM thread safety\n";
  51         std::cout<<"Allocating matrices..."<<std::flush;
  52         for(uint32_t i=0; i<(numConcurrentThreads*3); i++){
  53                 matBlock[i].resize(randomMatSize*randomMatSize);
  54         }
  55         std::cout<<"done\n";
  56         //pauser();
  57         std::cout<<"Filling matrices with random numbers..."<<std::flush;
  58         FillMatrices(matBlock, PRNG, rngdist, randomMatSize, numConcurrentThreads, 3);
  59         //PrintMatrices(matBlock, randomMatSize, numConcurrentThreads, 3);
  60         std::cout<<"done\n";
  61         std::cout<<"Testing CBLAS DGEMM thread safety\n";
  62         omp_set_num_threads(numConcurrentThreads);
  63         for(uint32_t R=0; R<numTestRounds; R++){
  64                 std::cout<<"DGEMM round #"<<R<<std::endl;
  65                 std::cout<<"Launching "<<numConcurrentThreads<<" threads simultaneously using OpenMP..."<<std::flush;
  66                 #pragma omp parallel for default(none) shared(futureBlock, matBlock, randomMatSize, numConcurrentThreads)
  67                 for(uint32_t i=0; i<numConcurrentThreads; i++){
  68                         futureBlock[i] = std::async(std::launch::async, launch_cblas_dgemm, &matBlock[i*3][0], &matBlock[i*3+1][0], &matBlock[i*3+2][0], randomMatSize);
  69                         //launch_cblas_dgemm( &matBlock[i][0], &matBlock[i+1][0], &matBlock[i+2][0]);
  70                 }
  71                 std::cout<<"done\n";
  72                 std::cout<<"Waiting for threads to finish..."<<std::flush;
  73                 for(uint32_t i=0; i<numConcurrentThreads; i++){
  74                         futureBlock[i].get();
  75                 }
  76                 std::cout<<"done\n";
  77                 //PrintMatrices(matBlock, randomMatSize, numConcurrentThreads, 3);
  78                 std::cout<<"Comparing results from different threads..."<<std::flush;
  79                 for(uint32_t i=3; i<(numConcurrentThreads*3); i+=3){ //i is the index of matrix A, for a given thread
  80                         for(uint32_t j = 0; j < static_cast<uint32_t>(randomMatSize*randomMatSize); j++){
  81                                 if (std::abs(matBlock[i+2][j] - matBlock[2][j]) > 1.0E-13){ //i+2 is the index of matrix C, for a given thread
  82                                         std::cout<<"ERROR: one of the threads returned a different result! Index : "<<i+2<<std::endl;
  83                                         std::cout<<"CBLAS DGEMM thread safety test FAILED!"<<std::endl;
  84                                         return -1;
  85                                 }
  86                         }
  87                 }
  88                 std::cout<<"OK!\n"<<std::endl;
  89         }
  90         std::cout<<"CBLAS DGEMM thread safety test PASSED!\n"<<std::endl;
  91         return 0;
  92 }