Merge pull request #290 from intel/develop

Develop
intel · Apr 25, 2024 · cb75540 · cb75540
2 parents abb66ec + 4105e91
commit cb75540
Show file tree

Hide file tree

Showing 9 changed files with 32 additions and 18 deletions.
diff --git a/src/common/common_utils.cpp b/src/common/common_utils.cpp
@@ -47,7 +47,7 @@ namespace yask {
     // https://semver.org/.
 
     // Format: "major.minor.patch[-alpha|-beta]".
-    const string version = "4.05.00";
+    const string version = "4.05.01";
 
     string yask_get_version_string() {
         return version;

diff --git a/src/examples/swe_main.cpp b/src/examples/swe_main.cpp
@@ -320,12 +320,12 @@ int main(int argc, char** argv) {
         double t_export = 0.02;
 
         // Some 1D buffers used to copy slices of data to/from YASK vars.
-        double hbuf[ysz];
-        double ebuf[ysz];
-        double ubuf[ysz];
-        double vbuf[ysz];
-        double keHbuf[ysz];
-        double pebuf[ysz];
+        auto* hbuf = new double[ysz];
+        auto* ebuf = new double[ysz];
+        auto* ubuf = new double[ysz];
+        auto* vbuf = new double[ysz];
+        auto* keHbuf = new double[ysz];
+        auto* pebuf = new double[ysz];
 
         // Init whole vars to default values.
         h->set_all_elements_same(1.0);
@@ -601,6 +601,14 @@ int main(int argc, char** argv) {
         }
         #endif
 
+        // Free buffers.
+        delete[] hbuf;
+        delete[] ebuf;
+        delete[] ubuf;
+        delete[] vbuf;
+        delete[] keHbuf;
+        delete[] pebuf;
+
         // Report global error.
         if (rank_num == msg_rank) {
             gerr_L2 = sqrt(gerr_L2);

diff --git a/src/examples/wave_eq_main.cpp b/src/examples/wave_eq_main.cpp
@@ -249,7 +249,7 @@ int main(int argc, char** argv) {
         soln->get_var("depth")->set_element(h, {});
 
         // Buffer used to copy slices of data to/from YASK var.
-        double ebuf[ysz];
+        auto* ebuf = new double[ysz];
 
         // Init vars.
         e->set_all_elements_same(0.0);
@@ -392,6 +392,9 @@ int main(int argc, char** argv) {
         }
         #endif
 
+        // Free buffer.
+        delete[] ebuf;
+
         // Report global stats.
         if (rank_num == msg_rank) {
             oerr_L2 = sqrt(oerr_L2);

diff --git a/src/kernel/Makefile b/src/kernel/Makefile
@@ -881,7 +881,6 @@ help:
 	@echo "Example debug builds of kernel cmd-line tool:"; \
 	echo " $(MAKE) clean; $(MAKE) -j stencil=iso3dfd mpi=0 omp=0 YK_CXXOPT='-O0' check=1 # No optimization, OpenMP, or MPI; use internal checking"; \
 	echo " $(MAKE) clean; $(MAKE) -j arch=avx2 stencil=test_2d YK_CXXOPT='-O0' trace=1   # Enable tracing; run with '-trace' to get trace"; \
-	#echo " $(MAKE) clean; $(MAKE) -j arch=intel64 stencil=ssg radius=0 fold='x=1,y=1,z=1' YK_CXXOPT='-O0' trace_mem=1 # Trace all mem accesses with '-trace'"; \
 	echo " "
 	@echo "Example regression tests (run before any git push or pull request):"; \
 	echo " $(MAKE) -j all                           # Normal full API and stencil tests"; \
@@ -1090,7 +1089,7 @@ single-stencil-tests:
 	$(MAKE) clean; $(STENCIL_TEST) stencil=test_empty_2d
 	$(MAKE) clean; $(STENCIL_TEST) stencil=test_2d YK_STENCIL_SUFFIX=-t1 $(call FOLD,x=2 y=2) inner_loop_dim=1
 	$(MAKE) clean; $(STENCIL_TEST) stencil=test_2d YK_STENCIL_SUFFIX=-t2 $(call FOLD,x=2 y=2) inner_loop_dim=2
-	$(MAKE) clean; $(STENCIL_TEST) stencil=test_2d YK_STENCIL_SUFFIX=-t3 arch=intel64 $(call FOLD,x=3 y=2) domain_dims=y,x
+	$(MAKE) clean; $(STENCIL_TEST) stencil=test_2d YK_STENCIL_SUFFIX=-t3 arch=intel64 domain_dims=y,x
 	$(MAKE) clean; $(STENCIL_TEST) stencil=test_reverse_2d YK_STENCIL_SUFFIX=-t1 radius=1
 	$(MAKE) clean; $(STENCIL_TEST) stencil=test_stages_2d $(call FOLD,x=2 y=2)
 	$(MAKE) clean; $(STENCIL_TEST) stencil=test_misc_2d YK_STENCIL_SUFFIX=-t1 $(call FOLD,x=2 y=2) inner_misc_layout=0 outer_domain_layout=0

diff --git a/src/kernel/lib/generic_var.cpp b/src/kernel/lib/generic_var.cpp
@@ -146,7 +146,6 @@ namespace yask {
                                   const T v = val;
                                   T* RESTRICT e = elems;
 
-                                  #pragma omp simd
                                   for (idx_t i = start; i < stop; i++)
                                       e[i] = v;
                               });
@@ -179,9 +178,11 @@ namespace yask {
                                   T* RESTRICT e = elems;
                                   const T s = seed;
 
-                                  #pragma omp simd
-                                  for (idx_t i = start; i < stop; i++)
-                                      e[i] = s * T(imod_flr(i, wrap) + 1);
+                                  for (idx_t i = start; i < stop; i++) {
+                                      T v = s * T(imod_flr(i, wrap) + 1);
+                                      e[i] = v;
+                                      //cout << "e["<<i<<"] = "<< v <<endl;
+                                  }
                               });
 
             // Also update the version on the device to the same sequence.

diff --git a/src/kernel/lib/yask.hpp b/src/kernel/lib/yask.hpp
@@ -74,6 +74,8 @@ typedef int MPI_Status;
 #include <unordered_set>
 #include <sstream>
 #include <stdexcept>
+#include <limits>
+
 #include <malloc.h>
 #include <stddef.h>
 #include <time.h>

diff --git a/src/kernel/lib/yk_var.cpp b/src/kernel/lib/yk_var.cpp
@@ -90,7 +90,7 @@ namespace yask {
             }
         }
         assert(dim_names.size() ==
-               _num_step_dims + _num_domain_dims + _num_misc_dims);
+               size_t(_num_step_dims + _num_domain_dims + _num_misc_dims));
     }
 
 

diff --git a/src/kernel/lib/yk_var.hpp b/src/kernel/lib/yk_var.hpp
@@ -575,8 +575,8 @@ namespace yask {
             double _sum = 0.0;
             double _sum_sq = 0.0;
             double _prod = 1.0;
-            double _max = DBL_MIN;
-            double _min = DBL_MAX;
+            double _max = std::numeric_limits<double>::min();
+            double _min = std::numeric_limits<double>::max();
 
             virtual ~red_res() { }
 
@@ -1120,7 +1120,7 @@ namespace yask {
                       make_index_string(range));
             if (ne <= 0)
                 return 0;
-            if (buffer_size < ne)
+            if (buffer_size < size_t(ne))
                 THROW_YASK_EXCEPTION(std::string("call to '") +
                                      Visitor::fname() + "' with buffer of size " +
                                      std::to_string(buffer_size) + "; " +

diff --git a/src/kernel/yask.sh b/src/kernel/yask.sh
@@ -394,6 +394,7 @@ if [[ $nranks > 1 || $force_mpi == 1 ]]; then
 
     # Add default Intel MPI settings.
     envs+=" I_MPI_PRINT_VERSION=1 I_MPI_DEBUG=5"
+    envs+=" I_MPI_PIN_DOMAIN=numa"
 
     # Check whether HBM policy setting is allowed.
     if [[ `I_MPI_HBW_POLICY=hbw_preferred,hbw_preferred mpirun -np 1 /bin/date |& grep -c 'Unknown memory policy'` == 0 ]]; then