Replace std::vector with our own implementation, as using the stl when compiling with nvcc and avx enabled leads to many issues.

unsupported/Eigen/CXX11/Core

@@ -33,6 +33,7 @@
 #include <vector>
 
 #include "src/Core/util/EmulateArray.h"
+#include "src/Core/util/MaxSizeVector.h"
 
 // Emulate the cxx11 functionality that we need if the compiler doesn't support it.
 // Visual studio 2015 doesn't advertise itself as cxx11 compliant, although it

unsupported/Eigen/CXX11/src/Core/util/MaxSizeVector.h

@@ -0,0 +1,130 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_FIXEDSIZEVECTOR_H
+#define EIGEN_FIXEDSIZEVECTOR_H
+
+namespace Eigen {
+
+/** \class MaxSizeVector
+  * \ingroup Core
+  *
+  * \brief The MaxSizeVector class.
+  *
+  * The %MaxSizeVector provides a subset of std::vector functionality.
+  *
+  * The goal is to provide basic std::vector operations when using
+  * std::vector is not an option (e.g. on GPU or when compiling using
+  * FMA/AVX, as this can cause either compilation failures or illegal
+  * instruction failures).
+  *
+  * Beware: The constructors are not API compatible with these of
+  * std::vector.
+  */
+template <typename T>
+class MaxSizeVector {
+ public:
+  // Construct a new MaxSizeVector, reserve n elements.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  explicit MaxSizeVector(size_t n)
+      : reserve_(n), size_(0),
+        data_(static_cast<T*>(internal::aligned_malloc(n * sizeof(T)))) {
+    for (size_t i = 0; i < n; ++i) { new (&data_[i]) T; }
+  }
+
+  // Construct a new MaxSizeVector, reserve and resize to n.
+  // Copy the init value to all elements.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  explicit MaxSizeVector(size_t n, const T& init)
+      : reserve_(n), size_(n),
+        data_(static_cast<T*>(internal::aligned_malloc(n * sizeof(T)))) {
+    for (size_t i = 0; i < n; ++i) { new (&data_[i]) T(init); }
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  ~MaxSizeVector() {
+    for (size_t i = 0; i < size_; ++i) {
+      data_[i].~T();
+    }
+    internal::aligned_free(data_);
+  }
+
+  // Append new elements (up to reserved size).
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void push_back(const T& t) {
+    eigen_assert(size_ < reserve_);
+    data_[size_++] = t;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const T& operator[] (size_t i) const {
+    eigen_assert(i < size_);
+    return data_[i];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  T& operator[] (size_t i) {
+    eigen_assert(i < size_);
+    return data_[i];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  T& back() {
+    eigen_assert(size_ > 0);
+    return data_[size_ - 1];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const T& back() const {
+    eigen_assert(size_ > 0);
+    return data_[size_ - 1];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  void pop_back() {
+    // NOTE: This does not destroy the value at the end the way
+    // std::vector's version of pop_back() does.  That happens when
+    // the Vector is destroyed.
+    eigen_assert(size_ > 0);
+    size_--;
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  size_t size() const { return size_; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  bool empty() const { return size_ == 0; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  T* data() { return data_; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const T* data() const { return data_; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  T* begin() { return data_; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  T* end() { return data_ + size_; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const T* begin() const { return data_; }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const T* end() const { return data_ + size_; }
+
+ private:
+  size_t reserve_;
+  size_t size_;
+  T* data_;
+};
+
+}  // namespace Eigen
+
+#endif  // EIGEN_FIXEDSIZEVECTOR_H

unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h

@@ -28,7 +28,7 @@ struct packLhsArg {
 
 template<typename LhsScalar, typename RhsScalar, typename RhsMapper, typename OutputMapper, typename Index>
 struct packRhsAndKernelArg {
-  const std::vector<LhsScalar*>* blockAs;
+  const MaxSizeVector<LhsScalar*>* blockAs;
   RhsScalar* blockB;
   const RhsMapper& rhs;
   OutputMapper& output;
@@ -46,8 +46,8 @@ struct packRhsAndKernelArg {
   const Index n_block_idx;
   const Index m_blocks;
   const Index n_blocks;
-  std::vector<Notification*>* kernel_notifications;
+  MaxSizeVector<Notification*>* kernel_notifications;
-  const std::vector<Notification*>* lhs_notifications;
+  const MaxSizeVector<Notification*>* lhs_notifications;
   const bool need_to_pack;
 };
 
@@ -202,8 +202,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
     //       the alignment requirements with the assumption that
     //       (Traits::mr * sizeof(ResScalar)) % 16 == 0
     const Index numBlockAs = numext::mini(num_threads, m_blocks);
-    std::vector<LhsScalar *> blockAs;
+    MaxSizeVector<LhsScalar *> blockAs(num_threads);
-    blockAs.reserve(num_threads);
     for (int i = 0; i < num_threads; i++) {
       blockAs.push_back(static_cast<LhsScalar *>(this->m_device.allocate(sizeA * sizeof(LhsScalar))));
     }
@@ -212,18 +211,17 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
     // TODO: is this too much memory to allocate? This simplifies coding a lot, but is wasteful.
     //       Other options: (1) reuse memory when a thread finishes. con: tricky
     //                      (2) allocate block B memory in each thread. con: overhead
-    std::vector<RhsScalar *> blockBs;
+    MaxSizeVector<RhsScalar *> blockBs(n_blocks);
-    blockBs.reserve(n_blocks);
     for (int i = 0; i < n_blocks; i++) {
       blockBs.push_back(static_cast<RhsScalar *>(this->m_device.allocate(sizeB * sizeof(RhsScalar))));
     }
 
     // lhs_notifications starts with all null Notifications
-    std::vector<Notification*> lhs_notifications(num_threads, nullptr);
+    MaxSizeVector<Notification*> lhs_notifications(num_threads, nullptr);
 
     // this should really be numBlockAs * n_blocks;
     const Index num_kernel_notifications = num_threads * n_blocks;
-    std::vector<Notification*> kernel_notifications(num_kernel_notifications,
+    MaxSizeVector<Notification*> kernel_notifications(num_kernel_notifications,
                                                     nullptr);
 
     for (Index k_block_idx = 0; k_block_idx < k_blocks; k_block_idx++) {

unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h

@@ -127,8 +127,7 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable>
       const Index blocksize = numext::maxi<Index>(PacketSize, (blocksz - (blocksz % PacketSize)));
       const Index numblocks = size / blocksize;
 
-      std::vector<Notification*> results;
+      MaxSizeVector<Notification*> results(numblocks);
-      results.reserve(numblocks);
       for (int i = 0; i < numblocks; ++i) {
         results.push_back(device.enqueue(&EvalRange<Evaluator, Index, Vectorizable>::run, evaluator, i*blocksize, (i+1)*blocksize));
       }

unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h

@@ -256,9 +256,8 @@ struct FullReducer<Self, Op, ThreadPoolDevice, false> {
       const Index numblocks = blocksize > 0 ? num_coeffs / blocksize : 0;
       eigen_assert(num_coeffs >= numblocks * blocksize);
 
-      std::vector<Notification*> results;
+      MaxSizeVector<Notification*> results(numblocks);
-      results.reserve(numblocks);
+      MaxSizeVector<typename Self::CoeffReturnType> shards(numblocks, reducer.initialize());
-      std::vector<typename Self::CoeffReturnType> shards(numblocks, reducer.initialize());
       for (Index i = 0; i < numblocks; ++i) {
         results.push_back(
             device.enqueue(&FullReducerShard<Self, Op, false>::run, self,
@@ -308,9 +307,8 @@ struct FullReducer<Self, Op, ThreadPoolDevice, true> {
     const Index numblocks = blocksize > 0 ? num_coeffs / blocksize : 0;
     eigen_assert(num_coeffs >= numblocks * blocksize);
 
-    std::vector<Notification*> results;
+    MaxSizeVector<Notification*> results(numblocks);
-    results.reserve(numblocks);
+    MaxSizeVector<typename Self::CoeffReturnType> shards(numblocks, reducer.initialize());
-    std::vector<typename Self::CoeffReturnType> shards(numblocks, reducer.initialize());
     for (Index i = 0; i < numblocks; ++i) {
       results.push_back(device.enqueue(&FullReducerShard<Self, Op, true>::run,
                                        self, i * blocksize, blocksize, reducer,