From 6af5ac7e2749bdea7a31323855ef3b4333b91c3e Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 4 Oct 2016 08:52:13 -0700
Subject: [PATCH] Cleanup the cuda executor code.

---
 unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h | 7 +------
 1 file changed, 1 insertion(+), 6 deletions(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index a116bf17f..0cac7b179 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -234,16 +234,11 @@ struct EigenMetaKernelEval<Evaluator, Index, true> {
 template <typename Evaluator, typename Index>
 __global__ void
 __launch_bounds__(1024)
-EigenMetaKernel(Evaluator memcopied_eval, Index size) {
+EigenMetaKernel(Evaluator eval, Index size) {
 
   const Index first_index = blockIdx.x * blockDim.x + threadIdx.x;
   const Index step_size = blockDim.x * gridDim.x;
 
-  // Cuda memcopies the kernel arguments. That's fine for POD, but for more
-  // complex types such as evaluators we should really conform to the C++
-  // standard and call a proper copy constructor.
-  Evaluator eval(memcopied_eval);
-
   const bool vectorizable = Evaluator::PacketAccess & Evaluator::IsAligned;
   EigenMetaKernelEval<Evaluator, Index, vectorizable>::run(eval, first_index, size, step_size);
 }