add new executable runTFCSStdPar to test stdpar features (#23)

FastCaloSimAnalyzer/FastCaloGpu/FastCaloGpu/CountingIterator.h

@@ -1,3 +1,10 @@
+#ifndef COUNTING_ITERATOR_H
+#define COUNTING_ITERATOR_H 1
+
+#include <cstddef>
+#include <type_traits>
+#include <iterator>
+
 struct counting_iterator {
 
   typedef size_t Index_t;
@@ -67,3 +74,6 @@ struct counting_iterator {
 private:
   value_type value;
 };
+
+
+#endif

FastCaloSimAnalyzer/FastCaloGpu/FastCaloGpu/TestStdPar.h

@@ -0,0 +1,16 @@
+#ifndef FCS_TEST_STDPAR
+#define FCS_TEST_STDPAR 1
+
+class TestStdPar {
+
+public:
+
+  void testAll(unsigned long);
+
+  void test_vector(unsigned long);
+  void test_atomicAdd_int(unsigned long);
+  void test_atomicAdd_float(unsigned long);
+
+};
+
+#endif

FastCaloSimAnalyzer/FastCaloGpu/src/CMakeLists.txt

@@ -57,7 +57,7 @@ endif()
 
 # Sources
 if(USE_STDPAR)
-  set(FastCaloGpu_Srcs GeoLoadGpu.cxx KernelWrapper_sp.cxx Rand4Hits_sp.cxx)
+  set(FastCaloGpu_Srcs GeoLoadGpu.cxx KernelWrapper_sp.cxx Rand4Hits_sp.cxx TestStdPar.cxx)
   if ( ${STDPAR_TARGET} STREQUAL "gpu" )
     set(FastCaloGpu_Srcs ${FastCaloGpu_Srcs} gpuQ.cu )
   endif()

FastCaloSimAnalyzer/FastCaloGpu/src/TestStdPar.cxx

@@ -0,0 +1,85 @@
+#include "TestStdPar.h"
+#include "CountingIterator.h"
+
+#include <algorithm>
+#include <execution>
+#include <iostream>
+#include <vector>
+#include <atomic>
+
+void TestStdPar::testAll(unsigned long num) {
+  test_vector(num);
+  test_atomicAdd_int(num);
+  test_atomicAdd_float(num);
+}
+
+void TestStdPar::test_vector(unsigned long num) {
+
+    std::cout << "---------- test_vec( " << num << " ) -------------\n";
+
+    std::vector<float>* pvec = new std::vector<float>;
+    pvec->resize(num);
+
+    float* pdat = pvec->data();
+
+    for (int i=0; i<num; ++i) {
+      pdat[i] = float(i) / 100.;
+      if (i<10) { printf("vec test CPU: %d %f\n", i, pdat[i]); }
+    }
+
+    std::for_each_n(std::execution::par_unseq, counting_iterator(0), num,
+                    [=](int i) {
+                      if (i<10) { printf("vec test GPU: %d %f\n", i, pdat[i]); }
+                      pdat[i] += 1.;
+                    } );
+
+    delete pvec;
+
+    std::cout << "------- done test_vec(num) -------------\n";
+
+  }
+
+
+void TestStdPar::test_atomicAdd_int(unsigned long num) {
+    std::cout << "---------- test_atomic<int>_add -------------\n";
+    std::atomic<int> *ii = new std::atomic<int>{0};
+    std::for_each_n(std::execution::par_unseq, counting_iterator(0), num,
+                    [=](int i) {
+                      int j = (*ii)++;
+                      printf("%d %d\n",i,j);
+                    } );
+    std::cout << "   after loop: " << *ii << " (should be " << num << ")" <<std::endl;
+    std::cout << "---------- done test_atomic<int>_add -------------\n\n";
+  }
+
+  /* * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * */
+
+void TestStdPar::test_atomicAdd_float(unsigned long N) {
+    std::cout << "---------- test_atomicAdd_float -------------\n";
+
+    float ta[N]{0.}, tc[N]{0.};
+    for (int i=0; i<N; ++i) {
+      ta[i%2] += i;
+      tc[i] += i;
+    }
+
+
+    float *fa = new float[N];
+    float *fb = new float[N];
+    float *fc = new float[N];
+    std::for_each_n(std::execution::par_unseq, counting_iterator(0), N,
+                    [=] (int i) {
+                      fb[i%2] += i;
+#if defined ( _NVHPC_STDPAR_NONE ) || defined ( _NVHPC_STDPAR_MULTICORE )
+                      fa[i % 2] += i;
+                      fc[i] += i;
+#else
+                      atomicAdd(&fa[i%2],float(i));
+                      atomicAdd(&fc[i],float(i));
+#endif
+                    });
+    for (int i=0; i<N; ++i) {
+      printf("%d : %2g [%2g] %g  %g [%g]\n",i, fa[i], ta[i], fb[i], fc[i], tc[i]);
+    }
+    std::cout << "---------- done test_atomicAdd_float -------------\n\n";
+  }

FastCaloSimAnalyzer/FastCaloSimAnalyzer/TFCSStdParTest.h

@@ -0,0 +1,10 @@
+#ifndef TFCSSTDPARTEST_H
+#define TFCSSTDPARTEST_H 1
+
+class TFCSStdParTest {
+public:
+  void test(bool doAtomic, bool doVector, unsigned long num);
+};
+
+
+#endif

FastCaloSimAnalyzer/Root/CMakeLists.txt

@@ -22,6 +22,10 @@ set(FastCaloSimAnalyzer_Srcs
   TFCSWriteCellsToTree.cxx
 )
 
+if(USE_STDPAR)
+  set(FastCaloSimAnalyzer_Srcs ${FastCaloSimAnalyzer_Srcs} TFCSStdParTest.cxx)
+endif()
+
 # # Global include is needed for dictionary generation to work
 include_directories(../)
 

FastCaloSimAnalyzer/Root/TFCSStdParTest.cxx

@@ -0,0 +1,17 @@
+#include "FastCaloSimAnalyzer/TFCSStdParTest.h"
+#include "FastCaloGpu/TestStdPar.h"
+
+#include <iostream>
+
+void TFCSStdParTest::test(bool doAtomic, bool doVector, unsigned long num) {
+
+  TestStdPar tst;
+
+  if (doAtomic) {
+    tst.test_atomicAdd_int(num);
+    tst.test_atomicAdd_float(num);
+  }
+
+  if (doVector) tst.test_vector(num);
+
+}

FastCaloSimAnalyzer/macro/CMakeLists.txt

@@ -47,3 +47,13 @@ fcs_make_task(runTFCSSimulation
   DEPENDENCY ${AthenaStandalone_LIB}
   DEPENDENCY ${FastCaloSimAnalyzer_LIB}
 )
+
+if(USE_STDPAR)
+  message(STATUS "Building StdPar tests")
+  fcs_make_task(runTFCSStdParTest
+    SOURCE runTFCSStdParTest.cxx
+    DEPENDENCY ${FastCaloSimCommon_LIB}
+    DEPENDENCY ${AthenaStandalone_LIB}
+    DEPENDENCY ${FastCaloSimAnalyzer_LIB}
+    )
+endif()

FastCaloSimAnalyzer/macro/runTFCSStdParTest.cxx

@@ -0,0 +1,37 @@
+#include <cstdio>
+#include <map>
+#include <iostream>
+#include "citer.h"
+#include <docopt/docopt.h>
+
+#include "FastCaloSimAnalyzer/TFCSStdParTest.h"
+
+static const char* USAGE =
+    R"(Run test for stdpar
+
+Usage:
+  runTFCSStdParTest [--doAtomicTest] [--doVectorTest] [-n <size> | --num <size>]
+  runTFCSStdParTest (-h | --help)
+
+Options:
+  -h --help                    Show help screen.
+  --doAtomicTest               Do test for atomic increments [default: false].
+  --doVectorTest               Do test for allocating/accessing vector<float> [default: false].
+  -n <size> --num <size>       Size of array to allocate for tests [default: 10].
+)";
+
+
+int main( int argc, char** argv ) {
+
+  std::map<std::string, docopt::value> args = docopt::docopt( USAGE, {argv + 1, argv + argc}, true );
+
+  bool doAtomic = args["--doAtomicTest"].asBool();
+  bool doVec    = args["--doVectorTest"].asBool();
+  int  num      = args["--num"].asLong();
+
+  TFCSStdParTest test;
+
+  test.test(doAtomic,doVec,num);
+
+  return 0;
+}