Misc improvements and cleanups

unsupported/test/CMakeLists.txt

@@ -110,6 +110,7 @@ if(EIGEN_TEST_CXX11)
 #  ei_add_test(cxx11_tensor_fixed_size "-std=c++0x")
   ei_add_test(cxx11_tensor_const "-std=c++0x")
   ei_add_test(cxx11_tensor_of_const_values "-std=c++0x")
+  ei_add_test(cxx11_tensor_of_complex "-std=c++0x")
   ei_add_test(cxx11_tensor_of_strings "-std=c++0x")
   ei_add_test(cxx11_tensor_intdiv "-std=c++0x")
   ei_add_test(cxx11_tensor_lvalue "-std=c++0x")

unsupported/test/cxx11_tensor_assign.cpp

@@ -253,6 +253,39 @@ static void test_auto_resize()
 }
 
 
+static void test_compound_assign()
+{
+  Tensor<int, 1> start_tensor(10);
+  Tensor<int, 1> offset_tensor(10);
+  start_tensor.setRandom();
+  offset_tensor.setRandom();
+
+  Tensor<int, 1> tensor = start_tensor;
+  tensor += offset_tensor;
+  for (int i = 0; i < 10; ++i) {
+    VERIFY_IS_EQUAL(tensor(i), start_tensor(i) + offset_tensor(i));
+  }
+
+  tensor = start_tensor;
+  tensor -= offset_tensor;
+  for (int i = 0; i < 10; ++i) {
+    VERIFY_IS_EQUAL(tensor(i), start_tensor(i) - offset_tensor(i));
+  }
+
+  tensor = start_tensor;
+  tensor *= offset_tensor;
+  for (int i = 0; i < 10; ++i) {
+    VERIFY_IS_EQUAL(tensor(i), start_tensor(i) * offset_tensor(i));
+  }
+
+  tensor = start_tensor;
+  tensor /= offset_tensor;
+  for (int i = 0; i < 10; ++i) {
+    VERIFY_IS_EQUAL(tensor(i), start_tensor(i) / offset_tensor(i));
+  }
+}
+
+
 void test_cxx11_tensor_assign()
 {
   CALL_SUBTEST(test_1d());
@@ -260,5 +293,5 @@ void test_cxx11_tensor_assign()
   CALL_SUBTEST(test_3d());
   CALL_SUBTEST(test_same_type());
   CALL_SUBTEST(test_auto_resize());
-
+  CALL_SUBTEST(test_compound_assign());
 }

unsupported/test/cxx11_tensor_convolution.cpp

@@ -64,8 +64,78 @@ static void test_expr()
 }
 
 
+static void test_modes() {
+  Tensor<float, 1> input(3);
+  Tensor<float, 1> kernel(3);
+  input(0) = 1.0f;
+  input(1) = 2.0f;
+  input(2) = 3.0f;
+  kernel(0) = 0.5f;
+  kernel(1) = 1.0f;
+  kernel(2) = 0.0f;
+
+  const Eigen::array<ptrdiff_t, 1> dims{{0}};
+  Eigen::array<std::pair<ptrdiff_t, ptrdiff_t>, 1> padding;
+
+  // Emulate VALID mode (as defined in
+  // http://docs.scipy.org/doc/numpy/reference/generated/numpy.convolve.html).
+  padding[0] = std::make_pair(0, 0);
+  Tensor<float, 1> valid(1);
+  valid = input.pad(padding).convolve(kernel, dims);
+  VERIFY_IS_EQUAL(valid.dimension(0), 1);
+  VERIFY_IS_APPROX(valid(0), 2.5f);
+
+  // Emulate SAME mode (as defined in
+  // http://docs.scipy.org/doc/numpy/reference/generated/numpy.convolve.html).
+  padding[0] = std::make_pair(1, 1);
+  Tensor<float, 1> same(3);
+  same = input.pad(padding).convolve(kernel, dims);
+  VERIFY_IS_EQUAL(same.dimension(0), 3);
+  VERIFY_IS_APPROX(same(0), 1.0f);
+  VERIFY_IS_APPROX(same(1), 2.5f);
+  VERIFY_IS_APPROX(same(2), 4.0f);
+
+  // Emulate FULL mode (as defined in
+  // http://docs.scipy.org/doc/numpy/reference/generated/numpy.convolve.html).
+  padding[0] = std::make_pair(2, 2);
+  Tensor<float, 1> full(5);
+  full = input.pad(padding).convolve(kernel, dims);
+  VERIFY_IS_EQUAL(full.dimension(0), 5);
+  VERIFY_IS_APPROX(full(0), 0.0f);
+  VERIFY_IS_APPROX(full(1), 1.0f);
+  VERIFY_IS_APPROX(full(2), 2.5f);
+  VERIFY_IS_APPROX(full(3), 4.0f);
+  VERIFY_IS_APPROX(full(4), 1.5f);
+}
+
+
+static void test_strides() {
+  Tensor<float, 1> input(13);
+  Tensor<float, 1> kernel(3);
+  input.setRandom();
+  kernel.setRandom();
+
+  const Eigen::array<ptrdiff_t, 1> dims{{0}};
+  const Eigen::array<ptrdiff_t, 1> stride_of_3{{3}};
+  const Eigen::array<ptrdiff_t, 1> stride_of_2{{2}};
+
+  Tensor<float, 1> result;
+  result = input.stride(stride_of_3).convolve(kernel, dims).stride(stride_of_2);
+
+  VERIFY_IS_EQUAL(result.dimension(0), 2);
+  VERIFY_IS_APPROX(result(0), (input(0)*kernel(0) + input(3)*kernel(1) +
+                               input(6)*kernel(2)));
+  VERIFY_IS_APPROX(result(1), (input(6)*kernel(0) + input(9)*kernel(1) +
+                               input(12)*kernel(2)));
+}
+
+
+
+
 void test_cxx11_tensor_convolution()
 {
   CALL_SUBTEST(test_evals());
   CALL_SUBTEST(test_expr());
+  CALL_SUBTEST(test_modes());
+  CALL_SUBTEST(test_strides());
 }

unsupported/test/cxx11_tensor_device.cpp

@@ -123,6 +123,14 @@ static void test_forced_contextual_eval(Context* context)
   context->out().device(context->device()) = (context->in1() + context->in2()).eval() * 3.14f + context->in1().constant(2.718f);
 }
 
+template <typename Context>
+static void test_compound_assignment(Context* context)
+{
+  context->out().device(context->device()) = context->in1().constant(2.718f);
+  context->out().device(context->device()) += context->in1() + context->in2() * 3.14f;
+}
+
+
 template <typename Context>
 static void test_contraction(Context* context)
 {
@@ -197,6 +205,15 @@ static void test_cpu() {
     }
   }
 
+  test_compound_assignment(&context);
+  for (int i = 0; i < 40; ++i) {
+    for (int j = 0; j < 50; ++j) {
+      for (int k = 0; k < 70; ++k) {
+        VERIFY_IS_APPROX(out(Eigen::array<int, 3>(i,j,k)), in1(Eigen::array<int, 3>(i,j,k)) + in2(Eigen::array<int, 3>(i,j,k)) * 3.14f + 2.718f);
+      }
+    }
+  }
+
   test_contraction(&context);
   for (int i = 0; i < 40; ++i) {
     for (int j = 0; j < 40; ++j) {
@@ -299,6 +316,16 @@ static void test_gpu() {
     }
   }
 
+  test_compound_assignment(&context);
+  assert(cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost) == cudaSuccess);
+  for (int i = 0; i < 40; ++i) {
+    for (int j = 0; j < 50; ++j) {
+      for (int k = 0; k < 70; ++k) {
+        VERIFY_IS_APPROX(out(Eigen::array<int, 3>(i,j,k)), in1(Eigen::array<int, 3>(i,j,k)) + in2(Eigen::array<int, 3>(i,j,k)) * 3.14f + 2.718f);
+      }
+    }
+  }
+
   test_contraction(&context);
   assert(cudaMemcpy(out.data(), d_out, out_bytes, cudaMemcpyDeviceToHost) == cudaSuccess);
   for (int i = 0; i < 40; ++i) {

unsupported/test/cxx11_tensor_morphing.cpp

@@ -12,6 +12,7 @@
 #include <Eigen/CXX11/Tensor>
 
 using Eigen::Tensor;
+using Eigen::IndexPair;
 
 static void test_simple_reshape()
 {
@@ -52,7 +53,7 @@ static void test_reshape_in_expr() {
   TensorMap<Tensor<float, 5>> tensor2(m2.data(), 3,5,7,11,13);
   Tensor<float, 2>::Dimensions newDims1{{2,3*5*7*11}};
   Tensor<float, 2>::Dimensions newDims2{{3*5*7*11,13}};
-  array<Tensor<float, 1>::DimensionPair, 1> contract_along{{std::make_pair(1, 0)}};
+  Eigen::array<IndexPair<DenseIndex>, 1> contract_along{{IndexPair<DenseIndex>(1, 0)}};
   Tensor<float, 2> tensor3(2,13);
   tensor3 = tensor1.reshape(newDims1).contract(tensor2.reshape(newDims2), contract_along);
 
@@ -125,7 +126,7 @@ static void test_slice_in_expr() {
   TensorMap<Tensor<float, 2>> tensor1(m1.data(), 7, 7);
   TensorMap<Tensor<float, 2>> tensor2(m2.data(), 3, 3);
   Tensor<float, 2> tensor3(3,1);
-  array<Tensor<float, 1>::DimensionPair, 1> contract_along{{std::make_pair(1, 0)}};
+  array<IndexPair<DenseIndex>, 1> contract_along{{IndexPair<DenseIndex>(1, 0)}};
 
   Eigen::DSizes<ptrdiff_t, 2> indices1(1,2);
   Eigen::DSizes<ptrdiff_t, 2> sizes1(3,3);

unsupported/test/cxx11_tensor_of_complex.cpp

@@ -0,0 +1,64 @@
+// 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/.
+
+#include "main.h"
+
+#include <Eigen/CXX11/Tensor>
+
+using Eigen::Tensor;
+using Eigen::TensorMap;
+
+
+
+static void test_additions()
+{
+  Tensor<std::complex<float>, 1> data1(3);
+  Tensor<std::complex<float>, 1> data2(3);
+  for (int i = 0; i < 3; ++i) {
+    data1(i) = std::complex<float>(i, -i);
+    data2(i) = std::complex<float>(i, 7 * i);
+  }
+
+  Tensor<std::complex<float>, 1> sum = data1 + data2;
+  for (int i = 0; i < 3; ++i) {
+    VERIFY_IS_EQUAL(sum(i),  std::complex<float>(2*i, 6*i));
+  }
+}
+
+
+static void test_contractions()
+{
+  Tensor<std::complex<float>, 4> t_left(30, 50, 8, 31);
+  Tensor<std::complex<float>, 5> t_right(8, 31, 7, 20, 10);
+  Tensor<std::complex<float>, 5> t_result(30, 50, 7, 20, 10);
+
+  t_left.setRandom();
+  t_right.setRandom();
+
+  typedef Map<Matrix<std::complex<float>, Dynamic, Dynamic>> MapXcf;
+  MapXcf m_left(t_left.data(), 1500, 248);
+  MapXcf m_right(t_right.data(), 248, 1400);
+  Matrix<std::complex<float>, Dynamic, Dynamic> m_result(1500, 1400);
+
+  // This contraction should be equivalent to a regular matrix multiplication
+  typedef Tensor<float, 1>::DimensionPair DimPair;
+  Eigen::array<DimPair, 2> dims({{DimPair(2, 0), DimPair(3, 1)}});
+  t_result = t_left.contract(t_right, dims);
+  m_result = m_left * m_right;
+  for (size_t i = 0; i < t_result.dimensions().TotalSize(); i++) {
+    VERIFY_IS_APPROX(t_result.data()[i], m_result.data()[i]);
+  }
+}
+
+
+void test_cxx11_tensor_of_complex()
+{
+  CALL_SUBTEST(test_additions());
+  CALL_SUBTEST(test_contractions());
+}

unsupported/test/cxx11_tensor_thread_pool.cpp

@@ -9,22 +9,23 @@
 
 #define EIGEN_USE_THREADS
 
-
+#include <iostream>
 #include "main.h"
 #include <Eigen/CXX11/Tensor>
 
+
 using Eigen::Tensor;
 
-void test_cxx11_tensor_thread_pool()
+static void test_multithread_elementwise()
 {
-  Eigen::Tensor<float, 3> in1(2,3,7);
-  Eigen::Tensor<float, 3> in2(2,3,7);
-  Eigen::Tensor<float, 3> out(2,3,7);
+  Tensor<float, 3> in1(2,3,7);
+  Tensor<float, 3> in2(2,3,7);
+  Tensor<float, 3> out(2,3,7);
 
   in1.setRandom();
   in2.setRandom();
 
-  Eigen::ThreadPoolDevice thread_pool_device(3);
+  Eigen::ThreadPoolDevice thread_pool_device(internal::random<int>(3, 11));
   out.device(thread_pool_device) = in1 + in2 * 3.14f;
 
   for (int i = 0; i < 2; ++i) {
@@ -35,3 +36,222 @@ void test_cxx11_tensor_thread_pool()
     }
   }
 }
+
+
+static void test_multithread_compound_assignment()
+{
+  Tensor<float, 3> in1(2,3,7);
+  Tensor<float, 3> in2(2,3,7);
+  Tensor<float, 3> out(2,3,7);
+
+  in1.setRandom();
+  in2.setRandom();
+
+  Eigen::ThreadPoolDevice thread_pool_device(internal::random<int>(3, 11));
+  out.device(thread_pool_device) = in1;
+  out.device(thread_pool_device) += in2 * 3.14f;
+
+  for (int i = 0; i < 2; ++i) {
+    for (int j = 0; j < 3; ++j) {
+      for (int k = 0; k < 7; ++k) {
+        VERIFY_IS_APPROX(out(i,j,k), in1(i,j,k) + in2(i,j,k) * 3.14f);
+      }
+    }
+  }
+}
+
+
+static void test_multithread_contraction()
+{
+  Tensor<float, 4> t_left(30, 50, 37, 31);
+  Tensor<float, 5> t_right(37, 31, 70, 2, 10);
+  Tensor<float, 5> t_result(30, 50, 70, 2, 10);
+
+  t_left.setRandom();
+  t_right.setRandom();
+
+  // this contraction should be equivalent to a single matrix multiplication
+  typedef Tensor<float, 1>::DimensionPair DimPair;
+  Eigen::array<DimPair, 2> dims({{DimPair(2, 0), DimPair(3, 1)}});
+
+
+  typedef Map<MatrixXf> MapXf;
+  MapXf m_left(t_left.data(), 1500, 1147);
+  MapXf m_right(t_right.data(), 1147, 1400);
+  MatrixXf m_result(1500, 1400);
+
+  Eigen::ThreadPoolDevice thread_pool_device(4);
+
+  // compute results by separate methods
+  t_result.device(thread_pool_device) = t_left.contract(t_right, dims);
+  m_result = m_left * m_right;
+
+ for (ptrdiff_t i = 0; i < t_result.size(); i++) {
+    VERIFY(&t_result.data()[i] != &m_result.data()[i]);
+    if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4) {
+      std::cout << "mismatch detected: " << t_result.data()[i] << " vs " <<  m_result.data()[i] << std::endl;
+      assert(false);
+    }
+  }
+}
+
+
+static void test_contraction_corner_cases()
+{
+  Tensor<float, 2> t_left(32, 500);
+  Tensor<float, 2> t_right(32, 28*28);
+  Tensor<float, 2> t_result(500, 28*28);
+
+  t_left = (t_left.constant(-0.5f) + t_left.random()) * 2.0f;
+  t_right = (t_right.constant(-0.6f) + t_right.random()) * 2.0f;
+  t_result = t_result.constant(NAN);
+
+  // this contraction should be equivalent to a single matrix multiplication
+  typedef Tensor<float, 1>::DimensionPair DimPair;
+  Eigen::array<DimPair, 1> dims{{DimPair(0, 0)}};
+
+  typedef Map<MatrixXf> MapXf;
+  MapXf m_left(t_left.data(), 32, 500);
+  MapXf m_right(t_right.data(), 32, 28*28);
+  MatrixXf m_result(500, 28*28);
+
+  Eigen::ThreadPoolDevice thread_pool_device(12);
+
+  // compute results by separate methods
+  t_result.device(thread_pool_device) = t_left.contract(t_right, dims);
+  m_result = m_left.transpose() * m_right;
+
+  for (ptrdiff_t i = 0; i < t_result.size(); i++) {
+    assert(!isnan(t_result.data()[i]));
+    if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4) {
+      std::cout << "mismatch detected at index " << i << " : " << t_result.data()[i] << " vs " <<  m_result.data()[i] << std::endl;
+      assert(false);
+    }
+  }
+
+  t_left.resize(32, 1);
+  t_left = (t_left.constant(-0.5f) + t_left.random()) * 2.0f;
+  t_result.resize (1, 28*28);
+  t_result = t_result.constant(NAN);
+  t_result.device(thread_pool_device) = t_left.contract(t_right, dims);
+  new(&m_left) MapXf(t_left.data(), 32, 1);
+  m_result = m_left.transpose() * m_right;
+  for (ptrdiff_t i = 0; i < t_result.size(); i++) {
+    assert(!isnan(t_result.data()[i]));
+    if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4) {
+      std::cout << "mismatch detected: " << t_result.data()[i] << " vs " <<  m_result.data()[i] << std::endl;
+      assert(false);
+    }
+  }
+
+  t_left.resize(32, 500);
+  t_right.resize(32, 4);
+  t_left = (t_left.constant(-0.5f) + t_left.random()) * 2.0f;
+  t_right = (t_right.constant(-0.6f) + t_right.random()) * 2.0f;
+  t_result.resize (500, 4);
+  t_result = t_result.constant(NAN);
+  t_result.device(thread_pool_device) = t_left.contract(t_right, dims);
+  new(&m_left) MapXf(t_left.data(), 32, 500);
+  new(&m_right) MapXf(t_right.data(), 32, 4);
+  m_result = m_left.transpose() * m_right;
+  for (ptrdiff_t i = 0; i < t_result.size(); i++) {
+    assert(!isnan(t_result.data()[i]));
+    if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4) {
+      std::cout << "mismatch detected: " << t_result.data()[i] << " vs " <<  m_result.data()[i] << std::endl;
+      assert(false);
+    }
+  }
+
+  t_left.resize(32, 1);
+  t_right.resize(32, 4);
+  t_left = (t_left.constant(-0.5f) + t_left.random()) * 2.0f;
+  t_right = (t_right.constant(-0.6f) + t_right.random()) * 2.0f;
+  t_result.resize (1, 4);
+  t_result = t_result.constant(NAN);
+  t_result.device(thread_pool_device) = t_left.contract(t_right, dims);
+  new(&m_left) MapXf(t_left.data(), 32, 1);
+  new(&m_right) MapXf(t_right.data(), 32, 4);
+  m_result = m_left.transpose() * m_right;
+  for (ptrdiff_t i = 0; i < t_result.size(); i++) {
+    assert(!isnan(t_result.data()[i]));
+    if (fabs(t_result.data()[i] - m_result.data()[i]) >= 1e-4) {
+      std::cout << "mismatch detected: " << t_result.data()[i] << " vs " <<  m_result.data()[i] << std::endl;
+      assert(false);
+    }
+  }
+}
+
+
+static void test_multithread_contraction_agrees_with_singlethread() {
+  int contract_size = internal::random<int>(1, 5000);
+
+  Tensor<float, 3> left(internal::random<int>(1, 80),
+                        contract_size,
+                        internal::random<int>(1, 100));
+
+  Tensor<float, 4> right(internal::random<int>(1, 25),
+                         internal::random<int>(1, 37),
+                         contract_size,
+                         internal::random<int>(1, 51));
+
+  left.setRandom();
+  right.setRandom();
+
+  // add constants to shift values away from 0 for more precision
+  left += left.constant(1.5f);
+  right += right.constant(1.5f);
+
+  typedef Tensor<float, 1>::DimensionPair DimPair;
+  Eigen::array<DimPair, 1> dims({{DimPair(1, 2)}});
+
+  Eigen::ThreadPoolDevice thread_pool_device(internal::random<int>(2, 11));
+
+  Tensor<float, 5> st_result;
+  st_result = left.contract(right, dims);
+
+  Tensor<float, 5> tp_result(st_result.dimensions());
+  tp_result.device(thread_pool_device) = left.contract(right, dims);
+
+  VERIFY(internal::dimensions_match(st_result.dimensions(), tp_result.dimensions()));
+  for (ptrdiff_t i = 0; i < st_result.size(); i++) {
+    // if both of the values are very small, then do nothing (because the test will fail
+    // due to numerical precision issues when values are small)
+    if (fabs(st_result.data()[i] - tp_result.data()[i]) >= 1e-4) {
+      VERIFY_IS_APPROX(st_result.data()[i], tp_result.data()[i]);
+    }
+  }
+}
+
+
+static void test_memcpy() {
+
+  for (int i = 0; i < 5; ++i) {
+    const int num_threads = internal::random<int>(3, 11);
+    Eigen::ThreadPoolDevice thread_pool_device(num_threads);
+
+    const int size = internal::random<int>(13, 7632);
+    Tensor<float, 1> t1(size);
+    t1.setRandom();
+    std::vector<float> result(size);
+    thread_pool_device.memcpy(&result[0], t1.data(), size*sizeof(float));
+    for (int i = 0; i < size; i++) {
+      VERIFY_IS_EQUAL(t1(i), result[i]);
+    }
+  }
+}
+
+
+void test_cxx11_tensor_thread_pool()
+{
+  CALL_SUBTEST(test_multithread_elementwise());
+  CALL_SUBTEST(test_multithread_compound_assignment());
+
+  CALL_SUBTEST(test_multithread_contraction());
+
+  CALL_SUBTEST(test_multithread_contraction_agrees_with_singlethread());
+
+  // Exercise various cases that have been problematic in the past.
+  CALL_SUBTEST(test_contraction_corner_cases());
+
+  CALL_SUBTEST(test_memcpy());
+}