Imported Upstream version 1.65.0upstream/1.65.0

Change-Id: Icf8400b375482cb11bcf77440a6934ba360d6ba4
Signed-off-by: DongHun Kwak <dh0128.kwak@samsung.com>

10 files changed, 26 insertions, 15 deletions

diff --git a/boost/compute/algorithm/detail/copy_on_device.hpp b/boost/compute/algorithm/detail/copy_on_device.hpp
index 8738c8c0b4..034b3bc212 100644
--- a/boost/compute/algorithm/detail/copy_on_device.hpp
+++ b/boost/compute/algorithm/detail/copy_on_device.hpp
@@ -144,7 +144,7 @@ inline future<OutputIterator> copy_on_device_async(InputIterator first,
     return make_future(result + std::distance(first, last), event_);
 }
 
-#ifdef CL_VERSION_2_0
+#ifdef BOOST_COMPUTE_CL_VERSION_2_0
 // copy_on_device() specialization for svm_ptr
 template<class T>
 inline svm_ptr<T> copy_on_device(svm_ptr<T> first,
@@ -181,7 +181,7 @@ inline future<svm_ptr<T> > copy_on_device_async(svm_ptr<T> first,
 
     return make_future(result + count, event_);
 }
-#endif // CL_VERSION_2_0
+#endif // BOOST_COMPUTE_CL_VERSION_2_0
 
 } // end detail namespace
 } // end compute namespace
diff --git a/boost/compute/algorithm/detail/copy_to_device.hpp b/boost/compute/algorithm/detail/copy_to_device.hpp
index bce5975f53..8601bb20ec 100644
--- a/boost/compute/algorithm/detail/copy_to_device.hpp
+++ b/boost/compute/algorithm/detail/copy_to_device.hpp
@@ -124,7 +124,7 @@ inline future<DeviceIterator> copy_to_device_async(HostIterator first,
     return make_future(result + static_cast<difference_type>(count), event_);
 }
 
-#ifdef CL_VERSION_2_0
+#ifdef BOOST_COMPUTE_CL_VERSION_2_0
 // copy_to_device() specialization for svm_ptr
 template<class HostIterator, class T>
 inline svm_ptr<T> copy_to_device(HostIterator first,
@@ -184,7 +184,7 @@ inline svm_ptr<T> copy_to_device_map(HostIterator first,
 
     return result + count;
 }
-#endif // CL_VERSION_2_0
+#endif // BOOST_COMPUTE_CL_VERSION_2_0
 
 } // end detail namespace
 } // end compute namespace
diff --git a/boost/compute/algorithm/detail/copy_to_host.hpp b/boost/compute/algorithm/detail/copy_to_host.hpp
index d770a996ef..89b57174fa 100644
--- a/boost/compute/algorithm/detail/copy_to_host.hpp
+++ b/boost/compute/algorithm/detail/copy_to_host.hpp
@@ -125,7 +125,7 @@ inline future<HostIterator> copy_to_host_async(DeviceIterator first,
     return make_future(iterator_plus_distance(result, count), event_);
 }
 
-#ifdef CL_VERSION_2_0
+#ifdef BOOST_COMPUTE_CL_VERSION_2_0
 // copy_to_host() specialization for svm_ptr
 template<class T, class HostIterator>
 inline HostIterator copy_to_host(svm_ptr<T> first,
@@ -189,7 +189,7 @@ inline HostIterator copy_to_host_map(svm_ptr<T> first,
 
     return iterator_plus_distance(result, count);
 }
-#endif // CL_VERSION_2_0
+#endif // BOOST_COMPUTE_CL_VERSION_2_0
 
 } // end detail namespace
 } // end compute namespace
diff --git a/boost/compute/algorithm/detail/find_extrema.hpp b/boost/compute/algorithm/detail/find_extrema.hpp
index eef2e36c3c..734b75aa90 100644
--- a/boost/compute/algorithm/detail/find_extrema.hpp
+++ b/boost/compute/algorithm/detail/find_extrema.hpp
@@ -56,7 +56,7 @@ inline InputIterator find_extrema(InputIterator first,
 
     // use serial method for OpenCL version 1.0 due to
     // problems with atomic_cmpxchg()
-    #ifndef CL_VERSION_1_1
+    #ifndef BOOST_COMPUTE_CL_VERSION_1_1
         return serial_find_extrema(first, last, compare, find_minimum, queue);
     #endif
 
diff --git a/boost/compute/algorithm/detail/find_extrema_with_reduce.hpp b/boost/compute/algorithm/detail/find_extrema_with_reduce.hpp
index 8f2a83c38b..515d7cc6da 100644
--- a/boost/compute/algorithm/detail/find_extrema_with_reduce.hpp
+++ b/boost/compute/algorithm/detail/find_extrema_with_reduce.hpp
@@ -246,6 +246,7 @@ inline void find_extrema_with_reduce(InputIterator input,
     );
 }
 
+// Space complexity: \Omega(2 * work-group-size * work-groups-per-compute-unit)
 template<class InputIterator, class Compare>
 InputIterator find_extrema_with_reduce(InputIterator first,
                                        InputIterator last,
diff --git a/boost/compute/algorithm/detail/find_if_with_atomics.hpp b/boost/compute/algorithm/detail/find_if_with_atomics.hpp
index 112c34cf00..e14fd12ae0 100644
--- a/boost/compute/algorithm/detail/find_if_with_atomics.hpp
+++ b/boost/compute/algorithm/detail/find_if_with_atomics.hpp
@@ -153,6 +153,7 @@ inline InputIterator find_if_with_atomics_multiple_vpt(InputIterator first,
     return first + static_cast<difference_type>(index.read(queue));
 }
 
+// Space complexity: O(1)
 template<class InputIterator, class UnaryPredicate>
 inline InputIterator find_if_with_atomics(InputIterator first,
                                           InputIterator last,
diff --git a/boost/compute/algorithm/detail/merge_sort_on_gpu.hpp b/boost/compute/algorithm/detail/merge_sort_on_gpu.hpp
index e62c6beb8d..d5e1a2d8c9 100644
--- a/boost/compute/algorithm/detail/merge_sort_on_gpu.hpp
+++ b/boost/compute/algorithm/detail/merge_sort_on_gpu.hpp
@@ -91,6 +91,7 @@ inline size_t bitonic_block_sort(KeyIterator keys_first,
                                  command_queue &queue)
 {
     typedef typename std::iterator_traits<KeyIterator>::value_type key_type;
+    typedef typename std::iterator_traits<ValueIterator>::value_type value_type;
 
     meta_kernel k("bitonic_block_sort");
     size_t count_arg = k.add_arg<const uint_>("count");
@@ -249,8 +250,11 @@ inline size_t bitonic_block_sort(KeyIterator keys_first,
             k.var<key_type>("my_key") << ";\n";
     if(sort_by_key)
     {
-        k << values_first[k.var<const uint_>("gid")] << " = " <<
-                values_first[k.var<const uint_>("offset + my_index")] << ";\n";
+        k <<
+            k.decl<value_type>("my_value") << " = " <<
+                values_first[k.var<const uint_>("offset + my_index")] << ";\n" <<
+            "barrier(CLK_GLOBAL_MEM_FENCE);\n" <<
+            values_first[k.var<const uint_>("gid")] << " = my_value;\n";
     }
     k <<
         // end if
@@ -418,7 +422,7 @@ inline void merge_blocks_on_gpu(KeyIterator keys_first,
                     ");\n" <<
                 "left_idx = equal ? mid_idx + 1 : left_idx + 1;\n" <<
                 "right_idx = equal ? right_idx : mid_idx;\n" <<
-                "upper_key = equal ? upper_key : " <<
+                "upper_key = " <<
                     keys_first[k.var<const uint_>("left_idx")] << ";\n" <<
             "}\n" <<
         "}\n" <<
diff --git a/boost/compute/algorithm/detail/radix_sort.hpp b/boost/compute/algorithm/detail/radix_sort.hpp
index 8e6d5f9c0a..53b1205c70 100644
--- a/boost/compute/algorithm/detail/radix_sort.hpp
+++ b/boost/compute/algorithm/detail/radix_sort.hpp
@@ -17,6 +17,9 @@
 #include <boost/type_traits/is_signed.hpp>
 #include <boost/type_traits/is_floating_point.hpp>
 
+#include <boost/mpl/and.hpp>
+#include <boost/mpl/not.hpp>
+
 #include <boost/compute/kernel.hpp>
 #include <boost/compute/program.hpp>
 #include <boost/compute/command_queue.hpp>
@@ -305,9 +308,12 @@ inline void radix_sort_impl(const buffer_iterator<T> first,
         options << " -DASC";
     }
 
+    // get type definition if it is a custom struct
+    std::string custom_type_def = boost::compute::type_definition<T2>() + "\n";
+
     // load radix sort program
     program radix_sort_program = cache->get_or_build(
-        cache_key, options.str(), radix_sort_source, context
+       cache_key, options.str(), custom_type_def + radix_sort_source, context
     );
 
     kernel count_kernel(radix_sort_program, "count");
diff --git a/boost/compute/algorithm/detail/serial_reduce.hpp b/boost/compute/algorithm/detail/serial_reduce.hpp
index 53aaf140fe..8b121274b9 100644
--- a/boost/compute/algorithm/detail/serial_reduce.hpp
+++ b/boost/compute/algorithm/detail/serial_reduce.hpp
@@ -20,6 +20,7 @@ namespace boost {
 namespace compute {
 namespace detail {
 
+// Space complexity: O(1)
 template<class InputIterator, class OutputIterator, class BinaryFunction>
 inline void serial_reduce(InputIterator first,
                           InputIterator last,
diff --git a/boost/compute/algorithm/detail/serial_reduce_by_key.hpp b/boost/compute/algorithm/detail/serial_reduce_by_key.hpp
index f9bda8e476..6fb04baa6d 100644
--- a/boost/compute/algorithm/detail/serial_reduce_by_key.hpp
+++ b/boost/compute/algorithm/detail/serial_reduce_by_key.hpp
@@ -55,11 +55,9 @@ inline size_t serial_reduce_by_key(InputKeyIterator keys_first,
     size_t result_size_arg = k.add_arg<uint_ *>(memory_object::global_memory,
                                                 "result_size");
 
-    convert<result_type> to_result_type;
-
     k <<
         k.decl<result_type>("result") <<
-            " = " << to_result_type(values_first[0]) << ";\n" <<
+            " = " << values_first[0] << ";\n" <<
         k.decl<key_type>("previous_key") << " = " << keys_first[0] << ";\n" <<
         k.decl<result_type>("value") << ";\n" <<
         k.decl<key_type>("key") << ";\n" <<
@@ -70,7 +68,7 @@ inline size_t serial_reduce_by_key(InputKeyIterator keys_first,
         values_result[0] << " = result;\n" <<
 
         "for(ulong i = 1; i < count; i++) {\n" <<
-        "    value = " << to_result_type(values_first[k.var<uint_>("i")]) << ";\n" <<
+        "    value = " << values_first[k.var<uint_>("i")] << ";\n" <<
         "    key = " << keys_first[k.var<uint_>("i")] << ";\n" <<
         "    if (" << predicate(k.var<key_type>("previous_key"),
                                 k.var<key_type>("key")) << ") {\n" <<