Merge branch 'develop' into nx_delay_rank

QMCPACK · Jun 3, 2021 · 20254e2 · 20254e2
2 parents c7aa84f + 6ab5543
commit 20254e2
Show file tree

Hide file tree

Showing 8 changed files with 70 additions and 7 deletions.
diff --git a/CMake/PGICompilers.cmake b/CMake/PGICompilers.cmake
@@ -23,6 +23,13 @@ ADD_DEFINITIONS( -Drestrict=__restrict__ )
 SET(CMAKE_C_FLAGS     "${CMAKE_C_FLAGS} ")
 SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -D__forceinline=inline")
 
+# Suppress compile warnings
+# 177 variable "XX" was declared but never referenced
+# 550 variable "XX" was set but never used
+# 612 overloaded virtual function "AA" is only partially overridden in class "BB"
+# 998 function "AA" is hidden by "BB" -- virtual function override intended?
+SET(CMAKE_C_FLAGS     "${CMAKE_C_FLAGS} --display_error_number --diag_suppress 177 --diag_suppress 550")
+SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} --display_error_number --diag_suppress 177 --diag_suppress 550 --diag_suppress 612 --diag_suppress 998")
 
 # Set extra optimization specific flags
 SET( CMAKE_C_FLAGS_RELEASE     "${CMAKE_C_FLAGS_RELEASE} -fast" )

diff --git a/src/Particle/SoaDistanceTableABOMPTarget.h b/src/Particle/SoaDistanceTableABOMPTarget.h
@@ -254,9 +254,9 @@ class SoaDistanceTableABOMPTarget : public DTD_BConds<T, D, SC>, public Distance
     const size_t total_targets = count_targets;
 
     const int N_sources_padded = getAlignedSize<T>(N_sources);
-    const int stride_size      = N_sources_padded * (D + 1);
 
 #ifndef NDEBUG
+    const int stride_size = N_sources_padded * (D + 1);
     count_targets = 0;
     for (size_t iw = 0; iw < dt_list.size(); iw++)
     {

diff --git a/src/Platforms/CMakeLists.txt b/src/Platforms/CMakeLists.txt
@@ -35,7 +35,7 @@ SUBDIRS(CPU)
 TARGET_LINK_LIBRARIES(platform_device PUBLIC platform_cpu)
 
 SUBDIRS(OMPTarget)
-TARGET_LINK_LIBRARIES(platform_device PRIVATE platform_omptarget)
+TARGET_LINK_LIBRARIES(platform_device PUBLIC platform_omptarget)
 
 IF(QMC_CUDA)
   SUBDIRS(CUDA_legacy)

diff --git a/src/Platforms/OMPTarget/CMakeLists.txt b/src/Platforms/OMPTarget/CMakeLists.txt
@@ -2,7 +2,7 @@
 #// This file is distributed under the University of Illinois/NCSA Open Source License.
 #// See LICENSE file in top directory for details.
 #//
-#// Copyright (c) 2020 QMCPACK developers.
+#// Copyright (c) 2021 QMCPACK developers.
 #//
 #// File developed by: Ye Luo, yeluo@anl.gov, Argonne National Laboratory
 #//
@@ -22,5 +22,35 @@ ELSE()
 ENDIF()
 
 IF(NOT QMC_COMPLEX)
-  TARGET_COMPILE_DEFINITIONS(platform_omptarget PRIVATE -DOPENMP_NO_COMPLEX)
+  TARGET_COMPILE_DEFINITIONS(platform_omptarget PUBLIC OPENMP_NO_COMPLEX)
+else()
+  # Test if C++ compiler supports OpenMP user defined reduction on complex type
+  try_compile(OMP_UDR_COMPLEX_OKAY ${CMAKE_CURRENT_BINARY_DIR} ${CMAKE_CURRENT_SOURCE_DIR}/test_user_defined_reduction_complex.cpp
+              CMAKE_FLAGS "${CMAKE_CXX_FLAGS}"
+              OUTPUT_VARIABLE COMPILE_OUTPUT)
+
+  if (NOT OMP_UDR_COMPLEX_OKAY)
+    set(COMPILE_FAIL_OUTPUT omp_UDR_complex_fail.txt)
+    file(WRITE "${CMAKE_BINARY_DIR}/${COMPILE_FAIL_OUTPUT}" "${COMPILE_OUTPUT}")
+    message(STATUS "OpenMP user-defined reduction functionality compilation check failed!"
+                   "See compiler output at ${COMPILE_FAIL_OUTPUT}")
+
+    # Test if C++ compiler supports OpenMP implicit reduction on complex type
+    try_compile(OMP_IMPLICIT_REDUCTION_COMPLEX_OKAY ${CMAKE_CURRENT_BINARY_DIR} ${CMAKE_CURRENT_SOURCE_DIR}/test_user_defined_reduction_complex.cpp
+                CMAKE_FLAGS "${CMAKE_CXX_FLAGS}"
+                COMPILE_DEFINITIONS "-DOPENMP_NO_UDR"
+                OUTPUT_VARIABLE COMPILE_OUTPUT)
+
+    if (NOT OMP_IMPLICIT_REDUCTION_COMPLEX_OKAY)
+      set(COMPILE_FAIL_OUTPUT omp_implicit_reduction_complex_fail.txt)
+      file(WRITE "${CMAKE_BINARY_DIR}/${COMPILE_FAIL_OUTPUT}" "${COMPILE_OUTPUT}")
+      message(FATAL_ERROR "OpenMP implicit reduction on complex functionality compilation check failed!"
+                          "See compiler output at ${COMPILE_FAIL_OUTPUT}")
+    else()
+      TARGET_COMPILE_DEFINITIONS(platform_omptarget PUBLIC OPENMP_NO_UDR)
+      message(STATUS "OpenMP implicit reduction functionality compilation check pass")
+    endif()
+  else()
+    message(STATUS "OpenMP user-defined reduction functionality compilation check pass")
+  endif()
 ENDIF()
diff --git a/src/Platforms/OMPTarget/ompReduction.hpp b/src/Platforms/OMPTarget/ompReduction.hpp
@@ -16,7 +16,9 @@
 #include <complex>
 #include "config.h"
 
+#if !defined(OPENMP_NO_COMPLEX) && !defined(OPENMP_NO_UDR)
 PRAGMA_OFFLOAD("omp declare reduction(+: std::complex<float>: omp_out += omp_in)")
 PRAGMA_OFFLOAD("omp declare reduction(+: std::complex<double>: omp_out += omp_in)")
+#endif
 
 #endif // QMCPLUSPLUS_OMPREDUCTION_H
diff --git a/src/Platforms/OMPTarget/test_user_defined_reduction_complex.cpp b/src/Platforms/OMPTarget/test_user_defined_reduction_complex.cpp
@@ -0,0 +1,25 @@
+#include <vector>
+#include <complex>
+#include <cassert>
+
+#if !defined(OPENMP_NO_UDR)
+#pragma omp declare reduction(+: std::complex<float>: omp_out += omp_in)
+#endif
+
+int main()
+{
+  const int N = 100;
+  std::vector<std::complex<float>> array(N);
+
+  auto array_ptr = array.data();
+  for (int i = 0; i < N; i++)
+    array_ptr[i] = std::complex<float>(i);
+
+  std::complex<float> sum;
+  #pragma omp parallel for reduction(+: sum)
+  for (int i = 0; i < N; i++)
+    sum += array_ptr[i];
+
+  assert(std::real(sum) == 4950);
+  assert(std::imag(sum) == 0);
+}
diff --git a/src/QMCHamiltonians/NonLocalECPComponent.cpp b/src/QMCHamiltonians/NonLocalECPComponent.cpp
@@ -238,7 +238,6 @@ void NonLocalECPComponent::mw_evaluateOne(const RefVectorWithLeader<NonLocalECPC
     for (size_t i = 0; i < p_list.size(); i++)
     {
       NonLocalECPComponent& component(ecp_component_list[i]);
-      auto* VP = component.VP;
       ParticleSet& W(p_list[i]);
       TrialWaveFunction& psi(psi_list[i]);
       const NLPPJob<RealType>& job = joblist[i];

diff --git a/src/QMCWaveFunctions/BsplineFactory/HybridRepSetReader.h b/src/QMCWaveFunctions/BsplineFactory/HybridRepSetReader.h
@@ -534,7 +534,7 @@ struct HybridRepSetReader : public SplineSetReader<SA>
         {
           auto& mycenter = centers[mygroup[idx]];
           aligned_vector<double> splineData_r(spline_npoints);
-          UBspline_1d_d* atomic_spline_r;
+          UBspline_1d_d* atomic_spline_r = nullptr;
           for (size_t ip = 0; ip < spline_npoints; ip++)
             splineData_r[ip] = all_vals[idx][ip][lm];
           atomic_spline_r = einspline::create(atomic_spline_r, 0.0, spline_radius, spline_npoints, splineData_r.data(),
@@ -547,7 +547,7 @@ struct HybridRepSetReader : public SplineSetReader<SA>
           else
           {
             aligned_vector<double> splineData_i(spline_npoints);
-            UBspline_1d_d* atomic_spline_i;
+            UBspline_1d_d* atomic_spline_i = nullptr;
             for (size_t ip = 0; ip < spline_npoints; ip++)
               splineData_i[ip] = all_vals[idx][ip][lm + lm_tot];
             atomic_spline_i = einspline::create(atomic_spline_i, 0.0, spline_radius, spline_npoints,