From d12e07c0e9f96aca2f54652f76743c5e488c8db5 Mon Sep 17 00:00:00 2001
From: Jakub Szuppe <j.szuppe@gmail.com>
Date: Sat, 21 May 2016 16:58:15 +0200
Subject: [PATCH] Fix type-safety issues for async copying host->device

---
 include/boost/compute/algorithm/copy.hpp | 57 ++++++++++++++++++++++--
 test/test_copy_type_mismatch.cpp         | 26 +++++++++++
 2 files changed, 80 insertions(+), 3 deletions(-)

diff --git a/include/boost/compute/algorithm/copy.hpp b/include/boost/compute/algorithm/copy.hpp
index 52fe5de5..655485a6 100644
--- a/include/boost/compute/algorithm/copy.hpp
+++ b/include/boost/compute/algorithm/copy.hpp
@@ -209,9 +209,14 @@ dispatch_copy_async(InputIterator first,
                     InputIterator last,
                     OutputIterator result,
                     command_queue &queue,
-                    typename boost::enable_if_c<
-                        !is_device_iterator<InputIterator>::value &&
-                        is_device_iterator<OutputIterator>::value
+                    typename boost::enable_if<
+                        mpl::and_<
+                            mpl::not_<
+                                is_device_iterator<InputIterator>
+                            >,
+                            is_device_iterator<OutputIterator>,
+                            is_same_value_type<InputIterator, OutputIterator>
+                        >
                     >::type* = 0)
 {
     BOOST_STATIC_ASSERT_MSG(
@@ -222,6 +227,52 @@ dispatch_copy_async(InputIterator first,
     return copy_to_device_async(first, last, result, queue);
 }
 
+// host -> device (async)
+// Type mismatch between InputIterator and OutputIterator value_types
+template<class InputIterator, class OutputIterator>
+inline future<OutputIterator>
+dispatch_copy_async(InputIterator first,
+                    InputIterator last,
+                    OutputIterator result,
+                    command_queue &queue,
+                    typename boost::enable_if<
+                        mpl::and_<
+                            mpl::not_<
+                                is_device_iterator<InputIterator>
+                            >,
+                            is_device_iterator<OutputIterator>,
+                            mpl::not_<
+                                is_same_value_type<InputIterator, OutputIterator>
+                            >
+                        >
+                    >::type* = 0)
+{
+    BOOST_STATIC_ASSERT_MSG(
+        is_contiguous_iterator<InputIterator>::value,
+        "copy_async() is only supported for contiguous host iterators"
+    );
+
+    typedef typename std::iterator_traits<InputIterator>::value_type input_type;
+
+    const context &context = queue.get_context();
+    size_t input_size = iterator_range_size(first, last);
+
+    // map [first; last) to device and run copy kernel
+    // on device for copying & casting
+    ::boost::compute::mapped_view<input_type> mapped_host(
+        // make sure it's a pointer to constant data
+        // to force read only mapping
+        const_cast<const input_type*>(
+            ::boost::addressof(*first)
+        ),
+        input_size,
+        context
+    );
+    return copy_on_device_async(
+        mapped_host.begin(), mapped_host.end(), result, queue
+    );
+}
+
 // device -> host
 template<class InputIterator, class OutputIterator>
 inline OutputIterator
diff --git a/test/test_copy_type_mismatch.cpp b/test/test_copy_type_mismatch.cpp
index 885f969f..f84b0ccc 100644
--- a/test/test_copy_type_mismatch.cpp
+++ b/test/test_copy_type_mismatch.cpp
@@ -238,4 +238,30 @@ BOOST_AUTO_TEST_CASE(copy_host_float_to_device_int_with_transform)
     parameters->set(cache_key, "direct_copy_threshold", direct_copy_threshold);
 }
 
+BOOST_AUTO_TEST_CASE(copy_async_host_float_to_device_int)
+{
+    using compute::int_;
+    using compute::float_;
+
+    float_ host[] = { 6.1f, -10.2f, 19.3f, 25.4f };
+    bc::vector<int_> device_vector(4, context);
+
+    // copy host float data to int device vector
+    compute::future<void> future =
+        bc::copy_async(host, host + 4, device_vector.begin(), queue);
+    future.wait();
+
+    CHECK_RANGE_EQUAL(
+        int_,
+        4,
+        device_vector,
+        (
+            static_cast<int_>(6.1f),
+            static_cast<int_>(-10.2f),
+            static_cast<int_>(19.3f),
+            static_cast<int_>(25.4f)
+        )
+    );
+}
+
 BOOST_AUTO_TEST_SUITE_END()