From b5d2078c4a9cdb81416586cca5658e38b059148d Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Tue, 22 Dec 2015 15:06:17 -0800
Subject: [PATCH] Optimized outer reduction on GPUs.

---
 .../Eigen/CXX11/src/Tensor/TensorReduction.h  | 39 +++++++++++++-
 .../CXX11/src/Tensor/TensorReductionCuda.h    | 54 +++++++++++++++++++
 2 files changed, 92 insertions(+), 1 deletion(-)

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index aaa877185..c30980a49 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -337,9 +337,23 @@ struct FullReducer<Self, Op, ThreadPoolDevice, true> {
 #endif
 
 
+// Default outer reducer
+template <typename Self, typename Op, typename Device>
+struct OuterReducer {
+  static const bool HasOptimizedImplementation = false;
+
+  static EIGEN_DEVICE_FUNC void run(const Self&, Op&, const Device&, typename Self::CoeffReturnType*, typename Self::Index, typename Self::Index) {
+    assert(false && "Not implemented");
+  }
+};
+
+
 #if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
 template <int B, int N, typename S, typename R, typename I>
 __global__ void FullReductionKernel(R, const S, I, typename S::CoeffReturnType*);
+
+template <int NPT, typename S, typename R, typename I>
+__global__ void OuterReductionKernel(R, const S, I, I, typename S::CoeffReturnType*);
 #endif
 
 }  // end namespace internal
@@ -439,7 +453,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
 	}
       }
     }
-    
+
     // Precompute input strides.
     if (NumInputDims > 0) {
       array<Index, NumInputDims> input_strides;
@@ -498,6 +512,28 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
       internal::FullReducer<Self, Op, Device>::run(*this, reducer, m_device, data);
       return need_assign;
     }
+
+    // Attempt to use an optimized reduction.
+#if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
+    else if (RunningOnGPU && data && (m_device.majorDeviceVersion() >= 3)) {
+      bool preserving_inner_dims = true;
+      for (int i = 0; i < NumReducedDims; ++i) {
+        if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+          preserving_inner_dims &= m_reducedDims[NumInputDims - 1 - i];
+        } else {
+          preserving_inner_dims &= m_reducedDims[i];
+        }
+      }
+      if (internal::OuterReducer<Self, Op, GpuDevice>::HasOptimizedImplementation &&
+          preserving_inner_dims) {
+        const Index num_values_to_reduce = internal::array_prod(m_reducedDims);
+        const Index num_coeffs_to_preserve = internal::array_prod(m_dimensions);
+        Op reducer(m_reducer);
+        internal::OuterReducer<Self, Op, GpuDevice>::run(*this, reducer, m_device, data, num_values_to_reduce, num_coeffs_to_preserve);
+        return false;
+      }
+    }
+#endif
     return true;
   }
 
@@ -579,6 +615,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
 #endif
 #if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
   template <int B, int N, typename S, typename R, typename I> friend void internal::FullReductionKernel(R, const S, I, typename S::CoeffReturnType*);
+  template <int NPT, typename S, typename R, typename I> friend void internal::OuterReductionKernel(R, const S, I, I, typename S::CoeffReturnType*);
 #endif
 
   // Returns the Index in the input tensor of the first value that needs to be
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
index af1b9432c..f0e9d528e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
@@ -131,6 +131,60 @@ struct FullReducer<Self, Op, GpuDevice, Vectorizable> {
   }
 };
 
+
+template <int NumPerThread, typename Self,
+          typename Reducer, typename Index>
+__global__ void OuterReductionKernel(Reducer reducer, const Self input, Index num_coeffs_to_reduce, Index num_preserved_coeffs,
+                                     typename Self::CoeffReturnType* output) {
+  const Index num_threads = blockDim.x * gridDim.x;
+  const Index thread_id = blockIdx.x * blockDim.x + threadIdx.x;
+  // Initialize the output values
+  for (Index i = thread_id; i < num_preserved_coeffs; i += num_threads) {
+    output[i] = reducer.initialize();
+  }
+
+  // Do the reduction.
+  const Index max_iter = DIVUP(num_coeffs_to_reduce, NumPerThread) * num_preserved_coeffs;
+  for (Index i = thread_id; i < max_iter; i += num_threads) {
+    const Index input_col = i % num_preserved_coeffs;
+    const Index input_row = (i / num_preserved_coeffs) * NumPerThread;
+    typename Self::CoeffReturnType reduced_val = reducer.initialize();
+    const Index max_row = numext::mini(input_row + NumPerThread, num_coeffs_to_reduce);
+    for (Index j = input_row; j < max_row; j++) {
+      typename Self::CoeffReturnType val = input.m_impl.coeff(j * num_preserved_coeffs + input_col);
+      reducer.reduce(val, &reduced_val);
+    }
+    atomicReduce(&(output[input_col]), reduced_val, reducer);
+  }
+}
+
+
+template <typename Self, typename Op>
+struct OuterReducer<Self, Op, GpuDevice> {
+  // Unfortunately nvidia doesn't support well exotic types such as complex,
+  // so reduce the scope of the optimized version of the code to the simple case
+  // of floats.
+  static const bool HasOptimizedImplementation = !Op::IsStateful &&
+                                                 internal::is_same<typename Self::CoeffReturnType, float>::value;
+
+  template <typename Device, typename OutputType>
+  static void run(const Self&, Op&, const Device&, OutputType*, typename Self::Index, typename Self::Index) {
+    assert(false && "Should only be called to reduce floats on a gpu device");
+  }
+
+  static void run(const Self& self, Op& reducer, const GpuDevice& device, float* output, typename Self::Index num_coeffs_to_reduce, typename Self::Index num_preserved_vals) {
+    typedef typename Self::Index Index;
+
+    const Index num_coeffs = num_coeffs_to_reduce * num_preserved_vals;
+    const int block_size = 256;
+    const int num_per_thread = 16;
+    const int num_blocks = std::ceil(static_cast<float>(num_coeffs) / (block_size * num_per_thread));
+
+    LAUNCH_CUDA_KERNEL((OuterReductionKernel<num_per_thread>),
+                       num_blocks, block_size, 0, device, reducer, self, num_coeffs_to_reduce, num_preserved_vals, output);
+  }
+};
+
 #endif