From 33d6dafd61b91dda97e17e5af46674a9d592365c Mon Sep 17 00:00:00 2001
From: Denis Demidov <dennis.demidov@gmail.com>
Date: Fri, 18 Mar 2016 10:09:18 +0300
Subject: [PATCH] Rewrite operator<<(std::ostream, vex::vector) with mapped
 pointers

Also introduce const version of vector::map().
This fixes #197
---
 vexcl/backend/compute/device_vector.hpp | 12 +++++++++
 vexcl/backend/cuda/device_vector.hpp    |  6 +++++
 vexcl/backend/opencl/device_vector.hpp  | 12 +++++++++
 vexcl/vector.hpp                        | 35 ++++++++++++++++---------
 4 files changed, 53 insertions(+), 12 deletions(-)

diff --git a/vexcl/backend/compute/device_vector.hpp b/vexcl/backend/compute/device_vector.hpp
index f8095e871..b1977c119 100644
--- a/vexcl/backend/compute/device_vector.hpp
+++ b/vexcl/backend/compute/device_vector.hpp
@@ -131,6 +131,18 @@ class device_vector {
                     );
         }
 
+        mapped_array map(boost::compute::command_queue q) const {
+            return mapped_array(
+                    static_cast<T*>(
+                        q.enqueue_map_buffer(
+                            buffer, CL_MAP_READ,
+                            0, size() * sizeof(T)
+                            )
+                        ),
+                    buffer_unmapper(q, buffer)
+                    );
+        }
+
         cl_mem raw() const {
             return buffer.get();
         }
diff --git a/vexcl/backend/cuda/device_vector.hpp b/vexcl/backend/cuda/device_vector.hpp
index 7c91fe6fd..d8ef83a4b 100644
--- a/vexcl/backend/cuda/device_vector.hpp
+++ b/vexcl/backend/cuda/device_vector.hpp
@@ -171,6 +171,12 @@ class device_vector {
             return ptr;
         }
 
+        mapped_array map(const command_queue &q) const {
+            mapped_array ptr(new T[n], buffer_unmapper(q, *this));
+            read(q, 0, n, ptr.get(), true);
+            return ptr;
+        }
+
         /// Returns raw CUdeviceptr handle.
         CUdeviceptr raw() const {
             return static_cast<CUdeviceptr>(reinterpret_cast<size_t>(buffer.get()));
diff --git a/vexcl/backend/opencl/device_vector.hpp b/vexcl/backend/opencl/device_vector.hpp
index 8f1e985d2..6b0bbd317 100644
--- a/vexcl/backend/opencl/device_vector.hpp
+++ b/vexcl/backend/opencl/device_vector.hpp
@@ -116,6 +116,18 @@ class device_vector {
                     );
         }
 
+        mapped_array map(const cl::CommandQueue &q) const {
+            return mapped_array(
+                    static_cast<T*>(
+                        q.enqueueMapBuffer(
+                            buffer, CL_TRUE, CL_MAP_READ,
+                            0, size() * sizeof(T)
+                            )
+                        ),
+                    buffer_unmapper(q, buffer)
+                    );
+        }
+
         cl_mem raw() const {
             return buffer();
         }
diff --git a/vexcl/vector.hpp b/vexcl/vector.hpp
index fe3d16d46..2f6603cf2 100644
--- a/vexcl/vector.hpp
+++ b/vexcl/vector.hpp
@@ -626,6 +626,15 @@ class vector : public vector_terminal_expression {
             return buf[d].map(queue[d]);
         }
 
+        /// Maps vector part located on the given device to a host array.
+        /**
+         * This returns a smart pointer that will be unmapped automatically
+         * upon destruction */
+        typename backend::device_vector<T>::mapped_array
+        map(unsigned d = 0) const {
+            return buf[d].map(queue[d]);
+        }
+
         /// Copy assignment
         const vector& operator=(const vector &x) {
             if (&x != this)
@@ -1007,19 +1016,21 @@ std::ostream &operator<<(std::ostream &o, const vex::vector<T> &t) {
     boost::io::ios_all_saver stream_state(o);
     const size_t chunk = std::is_integral<T>::value ? 10 : 5;
 
-    std::vector<T> data(t.size());
-    copy(t, data);
-
     o << "{" << std::setprecision(6);
-    for(size_t i = 0 ; i < data.size() ; i++) {
-        if (i % chunk == 0) o << "\n" << std::setw(6) << i << ":";
-
-        if (std::is_integral<T>::value)
-            o << " " << std::setw(6) << data[i];
-        else if (std::is_arithmetic<T>::value)
-            o << std::scientific << std::setw(14) << data[i];
-        else
-            o << " " << data[i];
+    for(unsigned p = 0; p < t.nparts(); ++p) {
+        size_t ps = t.part_size(p);
+        auto ptr = t.map(p);
+
+        for(size_t i = t.part_start(p), j = 0; j < ps; ++j, ++i) {
+            if (i % chunk == 0) o << "\n" << std::setw(6) << i << ":";
+
+            if (std::is_integral<T>::value)
+                o << " " << std::setw(6) << ptr[j];
+            else if (std::is_arithmetic<T>::value)
+                o << std::scientific << std::setw(14) << ptr[j];
+            else
+                o << " " << ptr[j];
+        }
     }
     return o << "\n}\n";
 }