Merge pull request #18 from ENCCS/fix-issue-16

Fix #16

Author: wikfeldt

content/code/day-1/07_axpy-usm_allocator/axpy.cpp

@@ -24,6 +24,9 @@ axpy(
   assert(x.size() == y.size());
   auto sz = x.size();
 
+  auto ptrX = x.data();
+  auto ptrY = y.data();
+
   // FIXME allocate output vector
   std::vector z(...);
   // NOTE why do we take the address of the output vector?

content/code/day-1/07_axpy-usm_allocator/solution/axpy.cpp

@@ -22,6 +22,9 @@ axpy(
   assert(x.size() == y.size());
   auto sz = x.size();
 
+  auto ptrX = x.data();
+  auto ptrY = y.data();
+
   std::vector z(sz, T { 0.0 }, shared_allocator<T>(Q));
   // get address of z, because we rely on by-copy capture in the kernel lambda.
   // Why?
@@ -33,7 +36,7 @@ axpy(
   Q.submit([&](handler& cgh) {
      cgh.parallel_for(range { sz }, [=](id<1> tid) {
        auto i  = tid[0];
-       ptrZ[i] = alpha * x[i] + y[i];
+       ptrZ[i] = alpha * ptrX[i] + ptrY[i];
      });
    })
     .wait();

content/unified-shared-memory.rst

@@ -334,6 +334,17 @@ Implicit
 
          A working solution is in the ``solution`` subfolder.
 
+         .. note:: 
+             
+            The fact that we use a device-aware allocator for ``std::vector``
+            does not mean that STL objects will *magically* work within the
+            SYCL runtime. Indeed, doing so is absolutely not legal code: it
+            might compile and even work in some cases, but is to be considered
+            a bug.
+            You can read up a complete explanation `here
+            <https://github.com/illuhad/hipSYCL/issues/817#issuecomment-1231922115>`_.
+
+
 
 .. keypoints::