Merged in rryan/eigen/tensorfunctors (pull request PR-233)

Fully support complex types in SumReducer and MeanReducer when building for CUDA by using scalar_sum_op and scalar_product_op instead of operator+ and operator*.

2 files changed, 46 insertions, 4 deletions

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
index 7164e8d60..d73f6dc68 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
@@ -124,7 +124,8 @@ template <typename T> struct SumReducer
   }
   template <typename Packet>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizeBoth(const T saccum, const Packet& vaccum) const {
-    return saccum + predux(vaccum);
+    internal::scalar_sum_op<T> sum_op;
+    return sum_op(saccum, predux(vaccum));
   }
 };
 
@@ -173,7 +174,8 @@ template <typename T> struct MeanReducer
   }
   template <typename Packet>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizeBoth(const T saccum, const Packet& vaccum) const {
-    return (saccum + predux(vaccum)) / (scalarCount_ + packetCount_ * unpacket_traits<Packet>::size);
+    internal::scalar_sum_op<T> sum_op;
+    return sum_op(saccum, predux(vaccum)) / (scalarCount_ + packetCount_ * unpacket_traits<Packet>::size);
   }
 
   protected:
@@ -304,7 +306,8 @@ template <typename T> struct ProdReducer
   static const bool IsStateful = false;
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reduce(const T t, T* accum) const {
-    (*accum) *= t;
+    internal::scalar_product_op<T> prod_op;
+    (*accum) = prod_op(*accum, t);
   }
   template <typename Packet>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void reducePacket(const Packet& p, Packet* accum) const {
@@ -328,7 +331,8 @@ template <typename T> struct ProdReducer
   }
   template <typename Packet>
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T finalizeBoth(const T saccum, const Packet& vaccum) const {
-    return saccum * predux_mul(vaccum);
+    internal::scalar_product_op<T> prod_op;
+    return prod_op(saccum, predux_mul(vaccum));
   }
 };
 
diff --git a/unsupported/test/cxx11_tensor_complex_cuda.cu b/unsupported/test/cxx11_tensor_complex_cuda.cu
index f895efd01..d4e111f5d 100644
--- a/unsupported/test/cxx11_tensor_complex_cuda.cu
+++ b/unsupported/test/cxx11_tensor_complex_cuda.cu
@@ -108,8 +108,46 @@ static void test_cuda_sum_reductions() {
 }
 
 
+static void test_cuda_product_reductions() {
+
+  Eigen::CudaStreamDevice stream;
+  Eigen::GpuDevice gpu_device(&stream);
+
+  const int num_rows = internal::random<int>(1024, 5*1024);
+  const int num_cols = internal::random<int>(1024, 5*1024);
+
+  Tensor<std::complex<float>, 2> in(num_rows, num_cols);
+  in.setRandom();
+
+  Tensor<std::complex<float>, 0> full_redux;
+  full_redux = in.prod();
+
+  std::size_t in_bytes = in.size() * sizeof(std::complex<float>);
+  std::size_t out_bytes = full_redux.size() * sizeof(std::complex<float>);
+  std::complex<float>* gpu_in_ptr = static_cast<std::complex<float>*>(gpu_device.allocate(in_bytes));
+  std::complex<float>* gpu_out_ptr = static_cast<std::complex<float>*>(gpu_device.allocate(out_bytes));
+  gpu_device.memcpyHostToDevice(gpu_in_ptr, in.data(), in_bytes);
+
+  TensorMap<Tensor<std::complex<float>, 2> > in_gpu(gpu_in_ptr, num_rows, num_cols);
+  TensorMap<Tensor<std::complex<float>, 0> > out_gpu(gpu_out_ptr);
+
+  out_gpu.device(gpu_device) = in_gpu.prod();
+
+  Tensor<std::complex<float>, 0> full_redux_gpu;
+  gpu_device.memcpyDeviceToHost(full_redux_gpu.data(), gpu_out_ptr, out_bytes);
+  gpu_device.synchronize();
+
+  // Check that the CPU and GPU reductions return the same result.
+  VERIFY_IS_APPROX(full_redux(), full_redux_gpu());
+
+  gpu_device.deallocate(gpu_in_ptr);
+  gpu_device.deallocate(gpu_out_ptr);
+}
+
+
 void test_cxx11_tensor_complex()
 {
   CALL_SUBTEST(test_cuda_nullary());
   CALL_SUBTEST(test_cuda_sum_reductions());
+  CALL_SUBTEST(test_cuda_product_reductions());
 }