From 8febae85a96c87c846a4f41682bc7950b7adc51e Mon Sep 17 00:00:00 2001
From: Roger Kowalewski <kowalewski@nm.ifi.lmu.de>
Date: Mon, 27 Mar 2017 11:06:33 +0200
Subject: [PATCH] Bug #346: Fix Copy CArray to Dash Array

---
 dash/include/dash/algorithm/Copy.h | 52 ++++++++++++++++++++++--------
 dash/test/algorithm/CopyTest.cc    | 19 +++++++++++
 2 files changed, 58 insertions(+), 13 deletions(-)

diff --git a/dash/include/dash/algorithm/Copy.h b/dash/include/dash/algorithm/Copy.h
index 368d58ed9..0fca05125 100644
--- a/dash/include/dash/algorithm/Copy.h
+++ b/dash/include/dash/algorithm/Copy.h
@@ -493,17 +493,42 @@ GlobOutputIt copy_impl(
                  "l_in_last:",   in_last,
                  "g_out_first:", out_first.pos());
 
+
+
   auto num_elements = std::distance(in_first, in_last);
-  dart_storage_t ds = dash::dart_storage<ValueType>(num_elements);
-  DASH_ASSERT_RETURNS(
-    dart_put_blocking(
-      out_first.dart_gptr(),
-      in_first,
-      ds.nelem,
-      ds.dtype),
-    DART_OK);
 
-  auto out_last = out_first + num_elements;
+  if (num_elements < 1) return out_first;
+
+  auto nremaining = static_cast<dash::default_size_t>(num_elements);
+  auto pattern = out_first.pattern();
+
+  while (nremaining) {
+    // global index to local unit and index
+    auto local_pos = pattern.local(out_first.pos());
+    // number of elements in unit
+    auto local_size = pattern.local_extents(team_unit_t{local_pos.unit});
+
+    auto ncopy = std::min(
+        std::initializer_list<dash::default_size_t>{local_size[0],
+        nremaining});
+
+    dart_storage_t ds = dash::dart_storage<ValueType>(ncopy);
+
+    DASH_ASSERT_RETURNS(
+        dart_put_blocking(
+          out_first.dart_gptr(),
+          in_first,
+          ds.nelem,
+          ds.dtype),
+        DART_OK);
+
+    std::advance(in_first, ncopy);
+    std::advance(out_first, ncopy);
+
+    nremaining = std::distance(in_first, in_last);
+  }
+
+  auto out_last = out_first;
   DASH_LOG_TRACE("dash::copy_impl >",
                  "g_out_last:", out_last.dart_gptr());
 
@@ -867,7 +892,7 @@ ValueType * copy(
   auto total_copy_elem = in_last - in_first;
 
   // Instead of testing in_first.local() and in_last.local(), this test for
-  // a local-only range only requires one call to in_first.local() which 
+  // a local-only range only requires one call to in_first.local() which
   // increases throughput by ~10% for local ranges.
   if (num_local_elem == total_copy_elem) {
     // Entire input range is local:
@@ -1113,10 +1138,11 @@ GlobOutputIt copy(
     // Copy to remote elements succeeding the local subrange:
     if (g_l_offset_end < out_h_last.pos()) {
       DASH_LOG_TRACE("dash::copy", "copy to global succeeding local subrange");
+
       out_last = dash::internal::copy_impl(
-                   in_first + l_elem_offset + num_local_elem,
-                   in_last,
-                   out_first + num_local_elem);
+                 in_first + l_elem_offset + num_local_elem,
+                 in_last,
+                 out_first + num_local_elem);
     }
   } else {
     // All elements in output range are remote
diff --git a/dash/test/algorithm/CopyTest.cc b/dash/test/algorithm/CopyTest.cc
index c27b60f70..12c4013fb 100644
--- a/dash/test/algorithm/CopyTest.cc
+++ b/dash/test/algorithm/CopyTest.cc
@@ -803,6 +803,25 @@ TEST_F(CopyTest, AsyncGlobalToLocalBlock)
   }
 }
 
+TEST_F(CopyTest, CArrayToDashArray)
+{
+  dash::Array<int> arr(100);
+
+  if (dash::myid() == 0) {
+    int buf[100];
+    std::iota(buf, buf + 100, 0);
+    // copy local buffer to global array
+    dash::copy(buf, buf + 100, arr.begin());
+  }
+
+  arr.barrier();
+
+  if (dash::myid() == 0) {
+    for (size_t idx = 0; idx < 100; ++idx) {
+      EXPECT_EQ_U(idx, static_cast<int>(arr[idx]));
+    }
+  }
+}
 #if 0
 // TODO
 TEST_F(CopyTest, AsyncAllToLocalVector)