devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Asynchronous expression evaluation with TensorAsyncDevice

Browse Source
This commit is contained in:
Eugene Zhulenev
2019-08-30 14:49:40 -07:00
parent f6c51d9209
commit 66665e7e76
8 changed files with 414 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files

107
unsupported/test/cxx11_tensor_executor.cpp
View File

@@ -562,37 +562,112 @@ static void test_execute_reverse_rvalue(Device d)
}
}
template <typename T, int NumDims, typename Device, bool Vectorizable,
bool Tileable, int Layout>
static void test_async_execute_unary_expr(Device d)
{
static constexpr int Options = 0 | Layout;
// Pick a large enough tensor size to bypass small tensor block evaluation
// optimization.
auto dims = RandomDims<NumDims>(50 / NumDims, 100 / NumDims);
Tensor<T, NumDims, Options, Index> src(dims);
Tensor<T, NumDims, Options, Index> dst(dims);
src.setRandom();
const auto expr = src.square();
using Assign = TensorAssignOp<decltype(dst), const decltype(expr)>;
using Executor = internal::TensorAsyncExecutor<const Assign, Device,
Vectorizable, Tileable>;
Eigen::Barrier done(1);
Executor::runAsync(Assign(dst, expr), d, [&done]() { done.Notify(); });
done.Wait();
for (Index i = 0; i < dst.dimensions().TotalSize(); ++i) {
T square = src.coeff(i) * src.coeff(i);
VERIFY_IS_EQUAL(square, dst.coeff(i));
}
}
template <typename T, int NumDims, typename Device, bool Vectorizable,
bool Tileable, int Layout>
static void test_async_execute_binary_expr(Device d)
{
static constexpr int Options = 0 | Layout;
// Pick a large enough tensor size to bypass small tensor block evaluation
// optimization.
auto dims = RandomDims<NumDims>(50 / NumDims, 100 / NumDims);
Tensor<T, NumDims, Options, Index> lhs(dims);
Tensor<T, NumDims, Options, Index> rhs(dims);
Tensor<T, NumDims, Options, Index> dst(dims);
lhs.setRandom();
rhs.setRandom();
const auto expr = lhs + rhs;
using Assign = TensorAssignOp<decltype(dst), const decltype(expr)>;
using Executor = internal::TensorAsyncExecutor<const Assign, Device,
Vectorizable, Tileable>;
Eigen::Barrier done(1);
Executor::runAsync(Assign(dst, expr), d, [&done]() { done.Notify(); });
done.Wait();
for (Index i = 0; i < dst.dimensions().TotalSize(); ++i) {
T sum = lhs.coeff(i) + rhs.coeff(i);
VERIFY_IS_EQUAL(sum, dst.coeff(i));
}
}
#ifdef EIGEN_DONT_VECTORIZE
#define VECTORIZABLE(VAL) !EIGEN_DONT_VECTORIZE && VAL
#else
#else
#define VECTORIZABLE(VAL) VAL
#endif
#define CALL_SUBTEST_PART(PART) \
CALL_SUBTEST_##PART
#define CALL_SUBTEST_COMBINATIONS(PART, NAME, T, NUM_DIMS) \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice, false, false, ColMajor>(default_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice, false, true, ColMajor>(default_device))); \
#define CALL_SUBTEST_COMBINATIONS(PART, NAME, T, NUM_DIMS) \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice, false, false, ColMajor>(default_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice, false, true, ColMajor>(default_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice, VECTORIZABLE(true), false, ColMajor>(default_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice, VECTORIZABLE(true), true, ColMajor>(default_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice, false, false, RowMajor>(default_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice, false, true, RowMajor>(default_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice, false, false, RowMajor>(default_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice, false, true, RowMajor>(default_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice, VECTORIZABLE(true), false, RowMajor>(default_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, DefaultDevice, VECTORIZABLE(true), true, RowMajor>(default_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, false, false, ColMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, false, true, ColMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, false, false, ColMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, false, true, ColMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, VECTORIZABLE(true), false, ColMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, VECTORIZABLE(true), true, ColMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, false, false, RowMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, false, true, RowMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, false, false, RowMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, false, true, RowMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, VECTORIZABLE(true), false, RowMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, VECTORIZABLE(true), true, RowMajor>(tp_device)))
// NOTE: Currently only ThreadPoolDevice supports async expression evaluation.
#define CALL_ASYNC_SUBTEST_COMBINATIONS(PART, NAME, T, NUM_DIMS) \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, false, false, ColMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, false, true, ColMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, VECTORIZABLE(true), false, ColMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, VECTORIZABLE(true), true, ColMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, false, false, RowMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, false, true, RowMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, VECTORIZABLE(true), false, RowMajor>(tp_device))); \
CALL_SUBTEST_PART(PART)((NAME<T, NUM_DIMS, ThreadPoolDevice, VECTORIZABLE(true), true, RowMajor>(tp_device)))
EIGEN_DECLARE_TEST(cxx11_tensor_executor) {
Eigen::DefaultDevice default_device;
// Default device is unused in ASYNC tests.
EIGEN_UNUSED_VARIABLE(default_device);
const auto num_threads = internal::random<int>(1, 24);
const auto num_threads = internal::random<int>(20, 24);
Eigen::ThreadPool tp(num_threads);
Eigen::ThreadPoolDevice tp_device(&tp, num_threads);
@@ -660,8 +735,16 @@ EIGEN_DECLARE_TEST(cxx11_tensor_executor) {
CALL_SUBTEST_COMBINATIONS(14, test_execute_reverse_rvalue, float, 4);
CALL_SUBTEST_COMBINATIONS(14, test_execute_reverse_rvalue, float, 5);
CALL_ASYNC_SUBTEST_COMBINATIONS(15, test_async_execute_unary_expr, float, 3);
CALL_ASYNC_SUBTEST_COMBINATIONS(15, test_async_execute_unary_expr, float, 4);
CALL_ASYNC_SUBTEST_COMBINATIONS(15, test_async_execute_unary_expr, float, 5);
CALL_ASYNC_SUBTEST_COMBINATIONS(16, test_async_execute_binary_expr, float, 3);
CALL_ASYNC_SUBTEST_COMBINATIONS(16, test_async_execute_binary_expr, float, 4);
CALL_ASYNC_SUBTEST_COMBINATIONS(16, test_async_execute_binary_expr, float, 5);
// Force CMake to split this test.
// EIGEN_SUFFIXES;1;2;3;4;5;6;7;8;9;10;11;12;13;14
// EIGEN_SUFFIXES;1;2;3;4;5;6;7;8;9;10;11;12;13;14;15;16
}
#undef CALL_SUBTEST_COMBINATIONS

39
unsupported/test/cxx11_tensor_thread_pool.cpp
View File

@@ -38,9 +38,9 @@ class TestAllocator : public Allocator {
void test_multithread_elementwise()
{
Tensor<float, 3> in1(2,3,7);
Tensor<float, 3> in2(2,3,7);
Tensor<float, 3> out(2,3,7);
Tensor<float, 3> in1(200, 30, 70);
Tensor<float, 3> in2(200, 30, 70);
Tensor<float, 3> out(200, 30, 70);
in1.setRandom();
in2.setRandom();
@@ -49,15 +49,39 @@ void test_multithread_elementwise()
Eigen::ThreadPoolDevice thread_pool_device(&tp, internal::random<int>(3, 11));
out.device(thread_pool_device) = in1 + 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);
for (int i = 0; i < 200; ++i) {
for (int j = 0; j < 30; ++j) {
for (int k = 0; k < 70; ++k) {
VERIFY_IS_APPROX(out(i, j, k), in1(i, j, k) + in2(i, j, k) * 3.14f);
}
}
}
}
void test_async_multithread_elementwise()
{
Tensor<float, 3> in1(200, 30, 70);
Tensor<float, 3> in2(200, 30, 70);
Tensor<float, 3> out(200, 30, 70);
in1.setRandom();
in2.setRandom();
Eigen::ThreadPool tp(internal::random<int>(3, 11));
Eigen::ThreadPoolDevice thread_pool_device(&tp, internal::random<int>(3, 11));
Eigen::Barrier b(1);
out.device(thread_pool_device, [&b]() { b.Notify(); }) = in1 + in2 * 3.14f;
b.Wait();
for (int i = 0; i < 200; ++i) {
for (int j = 0; j < 30; ++j) {
for (int k = 0; k < 70; ++k) {
VERIFY_IS_APPROX(out(i, j, k), in1(i, j, k) + in2(i, j, k) * 3.14f);
}
}
}
}
void test_multithread_compound_assignment()
{
@@ -516,6 +540,7 @@ void test_threadpool_allocate(TestAllocator* allocator)
EIGEN_DECLARE_TEST(cxx11_tensor_thread_pool)
{
CALL_SUBTEST_1(test_multithread_elementwise());
CALL_SUBTEST_1(test_async_multithread_elementwise());
CALL_SUBTEST_1(test_multithread_compound_assignment());
CALL_SUBTEST_2(test_multithread_contraction<ColMajor>());