fixing issue #157 (block size/forall)

system/New_loop_tests.cpp

@@ -125,6 +125,64 @@ void test_forall_here() {
 
 }
 
+struct OneBlock {
+  int64_t x;
+} GRAPPA_BLOCK_ALIGNED;
+struct TwoBlocks {
+  int64_t y;
+  char stuff[BLOCK_SIZE-sizeof(int64_t)]; 
+  int64_t a,b,c,d;
+} GRAPPA_BLOCK_ALIGNED;
+
+void test_forall_large_data() {
+  BOOST_MESSAGE("Testing forall on array of large structs");
+  {
+    size_t c = 7;
+    auto arr = global_alloc<OneBlock>(c);
+    // initialize
+    forall(arr, c, [] (int64_t i, OneBlock& two) {
+        two.x = 10*i;
+        });
+
+    // check without forall
+    for (int i=0; i<c; i++) {
+      auto res = delegate::read( arr + i );
+      BOOST_CHECK_EQUAL( res.x, 10*i );
+    }
+  }
+  {
+    size_t c = 7;
+    auto arr = global_alloc<TwoBlocks>(c);
+    // initialize
+/*
+    forall(arr, c, [] (int64_t i, TwoBlocks& two) {
+        two.y = i;
+        two.a = 10*i;
+        two.b = 100*i;
+        two.c = 1000*i;
+        two.d = 10000*i;
+        });
+    */
+
+    for (int i=0; i<c; i++) {
+      TwoBlocks b = { i, "", 10*i, 100*i, 1000*i, 10000*i };
+      delegate::write( arr + i, b );
+    }
+
+
+    // check without forall
+    for (int i=0; i<c; i++) {
+      auto res = delegate::read( arr + i );
+      BOOST_MESSAGE( "check " << i );
+      BOOST_CHECK_EQUAL( res.y, i );
+      BOOST_CHECK_EQUAL( res.a, 10*i );
+      //BOOST_CHECK_EQUAL( res.b, 1000*i );
+    }
+  }
+}
+
+
+
 void test_forall_global_private() {
   BOOST_MESSAGE("Testing forall_global...");
   const int64_t N = 1 << 8;
@@ -317,6 +375,8 @@ BOOST_AUTO_TEST_CASE( test1 ) {
     test_forall_localized();
 
     test_forall_here_async();
+
+    test_forall_large_data();
 
     Metrics::merge_and_dump_to_file();
   });

system/ParallelLoop.hpp

@@ -35,6 +35,8 @@
 #include "Collective.hpp"
 #include "function_traits.hpp"
 
+#include <algorithm>
+
 /// Flag: loop_threshold
 ///
 /// Iterations of `forall` loops *may* be run in parallel. A complete loop is decomposed into
@@ -333,7 +335,7 @@ namespace Grappa {
     auto end = base+nelem;
     if (nelem > 0) { fc = 1; }
 
-    size_t block_elems = block_size / sizeof(T);
+    size_t block_elems = std::max<size_t>(1, block_size / sizeof(T));
     int64_t nfirstcore = base.block_max() - base;
     int64_t n = nelem - nfirstcore;