diff --git a/lib/codegen/selection/generator.cc b/lib/codegen/selection/generator.cc
index f0534b5ce..8bcbefdf8 100644
--- a/lib/codegen/selection/generator.cc
+++ b/lib/codegen/selection/generator.cc
@@ -713,12 +713,10 @@ void generator::visit_atomic_add_inst(ir::atomic_add_inst* add) {
     
     // vector size
     int vector_size = 1;
-    /*
     int ld = ptrs->get_order()[0];
     unsigned alignment = alignment_->get(ptr, ld);
     vector_size = gcd(ptrs->axis(ld).contiguous, alignment);
     vector_size = std::min(vector_size, val->get_type()->get_tile_element_ty()->is_half_ty() ? 2 : 1);
-    vector_size = 1;
 
     std::map<unsigned, Value*> packets;
     for_each(val, [&](indices_t idx){
@@ -729,19 +727,18 @@ void generator::visit_atomic_add_inst(ir::atomic_add_inst* add) {
         packets[id] = UndefValue::get(VectorType::get(in_value->getType(), vector_size));
       packets[id] = builder_->CreateInsertElement(packets.at(id), in_value, linear % vector_size);
     });
-    */
 
     for_each(ptr, [&](indices_t idx){
       unsigned linear = vals->get_linear_index(idx);
       unsigned id = linear / vector_size;
-      //if(linear % vector_size != 0)
-      //  return;
+      if(linear % vector_size != 0)
+        return;
       // num bytes
       Value *rmw_ptr = ptrs->get_value(idx);
       Value *rmw_msk = msks->get_value(idx);
-      Value *rmw_val = vals->get_value(idx);
-      //if(vector_size == 1)
-      //  rmw_val = builder_->CreateExtractElement(rmw_val, builder_->getInt32(0));
+      Value *rmw_val = packets[id];
+      if(vector_size == 1)
+        rmw_val = builder_->CreateExtractElement(rmw_val, builder_->getInt32(0));
       Type* ty = rmw_val->getType();
       size_t nbits = ty->getScalarSizeInBits();
       // extract pointer offset
diff --git a/lib/driver/stream.cc b/lib/driver/stream.cc
index fd581b085..8d6762767 100755
--- a/lib/driver/stream.cc
+++ b/lib/driver/stream.cc
@@ -82,6 +82,7 @@ void host_stream::synchronize() {
 
 void host_stream::enqueue(driver::kernel* kernel, std::array<size_t, 3> grid, std::array<size_t, 3> block, std::vector<event> const *, event* event, void **args, size_t args_size) {
   auto hst = kernel->module()->hst();
+  hst_->futures->reserve(hst_->futures->size() + grid[0]*grid[1]*grid[2]);
   char* params = new char[args_size];
   std::memcpy((void*)params, (void*)args, args_size);
   for(size_t i = 0; i < grid[0]; i++)
diff --git a/python/setup.py b/python/setup.py
index 69e3aa58a..9346e73fe 100644
--- a/python/setup.py
+++ b/python/setup.py
@@ -55,7 +55,7 @@ class CMakeBuild(build_ext):
             self.build_extension(ext)
 
     def build_extension(self, ext):
-        self.debug = True
+        #self.debug = True
         extdir = os.path.abspath(os.path.dirname(self.get_ext_fullpath(ext.path)))
         # python directories
         python_include_dirs = distutils.sysconfig.get_python_inc()
@@ -75,7 +75,7 @@ class CMakeBuild(build_ext):
                       '-DLLVM_CONFIG=' + find_llvm()]
         # configuration
         cfg = 'Debug' if self.debug else 'Release'
-        cfg = 'Debug'
+        cfg = 'Release'
         build_args = ['--config', cfg]
 
         if platform.system() == "Windows":