experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[tf.unique()] Optimize the hash table implementation in UniqueOp::Compute().

Browse Source 
This change makes two improvements to the `UniqueOp` implementation:

1. Use `absl::flat_hash_map` instead of `std::unordered_map`.
2. For the `tstring` implementation, use `StringPiece` as the key instead of `tstring`, which avoids copying the strings into the map.

In addition, this change switches the microbenchmarks in unique_op_test.cc to use the SINGLE_THREADED_EXECUTOR, which removes thread scheduling overhead from the microbenchmark, and reduces noise in the results.

Microbenchmark results show a saving of between 0% and 65% on BM_Unique_INT32, between 8% and 26% on BM_Unique_INT32_Repeat, and between 17% and 40% on BM_Unique_STRING.

PiperOrigin-RevId: 307647292
Change-Id: If4367df37b856bf1c4cf91fcb34eea479014077f
This commit is contained in:
Derek Murray 2020-04-21 11:35:19 -07:00 committed by TensorFlower Gardener
parent e2bb4b2acd
commit 37d16f759a
3 changed files with 61 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
tensorflow/core/kernels/BUILD
View File

@ -1371,7 +1371,9 @@ tf_kernel_library(
tf_kernel_library( tf_kernel_library(
name = "unique_op", name = "unique_op",
prefix = "unique_op", prefix = "unique_op",
deps = ARRAY_DEPS, deps = ARRAY_DEPS + [
"@com_google_absl//absl/container:flat_hash_map",
],
) )
tf_kernel_library( tf_kernel_library(
@ -2335,6 +2337,7 @@ tf_cc_test(
"//tensorflow/core:test", "//tensorflow/core:test",
"//tensorflow/core:test_main", "//tensorflow/core:test_main",
"//tensorflow/core:testlib", "//tensorflow/core:testlib",
"//tensorflow/core/kernels/data:single_threaded_executor",
], ],
) )

70
tensorflow/core/kernels/unique_op.cc
View File

@ -17,6 +17,7 @@ limitations under the License.
#include <unordered_map> #include <unordered_map>
#include <utility> #include <utility>
#include "absl/container/flat_hash_map.h"
#include "tensorflow/core/framework/bounds_check.h" #include "tensorflow/core/framework/bounds_check.h"
#include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/register_types.h" #include "tensorflow/core/framework/register_types.h"
@ -26,10 +27,19 @@ limitations under the License.
#include "tensorflow/core/lib/hash/hash.h" #include "tensorflow/core/lib/hash/hash.h"
namespace tensorflow { namespace tensorflow {
namespace {
typedef Eigen::ThreadPoolDevice CPUDevice; typedef Eigen::ThreadPoolDevice CPUDevice;
template <typename T, typename TIndex> // `UniqueOp` computes the unique elements in the input tensor.
//
// * `T` is the element type.
// * `TKey` is the key type used in a local hash map. It must be explicitly
// convertible from `T`. For POD inputs, `TKey = T`. For `tstring` inputs,
// `TKey = absl::string_view` avoids copying the input strings into the map.
// * `TIndex` is the type used to represent indices in the output, either
// `int32` or `int64`.
template <typename T, typename TKey, typename TIndex>
class UniqueOp : public OpKernel { class UniqueOp : public OpKernel {
public: public:
explicit UniqueOp(OpKernelConstruction* context) : OpKernel(context) {} explicit UniqueOp(OpKernelConstruction* context) : OpKernel(context) {}
@ -106,10 +116,10 @@ class UniqueOp : public OpKernel {
auto Tin = input.flat<T>(); auto Tin = input.flat<T>();
const int64 N = static_cast<int64>(Tin.size()); const int64 N = static_cast<int64>(Tin.size());
std::unordered_map<T, TIndex> uniq; absl::flat_hash_map<TKey, TIndex> uniq;
uniq.reserve(2 * N); uniq.reserve(2 * N);
for (Eigen::Index i = 0, j = 0; i < N; ++i) { for (Eigen::Index i = 0, j = 0; i < N; ++i) {
auto it = uniq.insert(std::make_pair(Tin(i), j)); auto it = uniq.emplace(TKey(Tin(i)), j);
idx_vec(i) = it.first->second; idx_vec(i) = it.first->second;
if (it.second) { if (it.second) {
++j; ++j;
@ -153,13 +163,14 @@ class UniqueOp : public OpKernel {
return true; return true;
}; };
std::unordered_map<int64, int64, decltype(hash_fn), decltype(equal_to_fn)> absl::flat_hash_map<int64, int64, decltype(hash_fn),
decltype(equal_to_fn)>
uniq(0, hash_fn, equal_to_fn); uniq(0, hash_fn, equal_to_fn);
uniq.reserve(2 * Tin.dimension(1)); uniq.reserve(2 * Tin.dimension(1));
for (int64 i = 0, j = 0; i < Tin.dimension(1); ++i) { for (int64 i = 0, j = 0; i < Tin.dimension(1); ++i) {
auto it = uniq.insert(std::make_pair(i, j)); auto it = uniq.emplace(i, j);
idx_vec(i) = it.first->second; idx_vec(i) = it.first->second;
if (it.second) { if (it.second) {
++j; ++j;
@ -194,51 +205,56 @@ class UniqueOp : public OpKernel {
} }
}; };
#define REGISTER_UNIQUE(type) \ #define REGISTER_UNIQUE_WITH_KEY_TYPE(type, key_type) \
REGISTER_KERNEL_BUILDER(Name("Unique") \ REGISTER_KERNEL_BUILDER(Name("Unique") \
.Device(DEVICE_CPU) \ .Device(DEVICE_CPU) \
.TypeConstraint<type>("T") \ .TypeConstraint<type>("T") \
.TypeConstraint<int32>("out_idx"), \ .TypeConstraint<int32>("out_idx"), \
UniqueOp<type, int32>); \ UniqueOp<type, key_type, int32>); \
REGISTER_KERNEL_BUILDER(Name("Unique") \ REGISTER_KERNEL_BUILDER(Name("Unique") \
.Device(DEVICE_CPU) \ .Device(DEVICE_CPU) \
.TypeConstraint<type>("T") \ .TypeConstraint<type>("T") \
.TypeConstraint<int64>("out_idx"), \ .TypeConstraint<int64>("out_idx"), \
UniqueOp<type, int64>); \ UniqueOp<type, key_type, int64>); \
REGISTER_KERNEL_BUILDER(Name("UniqueV2") \ REGISTER_KERNEL_BUILDER(Name("UniqueV2") \
.Device(DEVICE_CPU) \ .Device(DEVICE_CPU) \
.TypeConstraint<type>("T") \ .TypeConstraint<type>("T") \
.TypeConstraint<int32>("out_idx"), \ .TypeConstraint<int32>("out_idx"), \
UniqueOp<type, int32>); \ UniqueOp<type, key_type, int32>); \
REGISTER_KERNEL_BUILDER(Name("UniqueV2") \ REGISTER_KERNEL_BUILDER(Name("UniqueV2") \
.Device(DEVICE_CPU) \ .Device(DEVICE_CPU) \
.TypeConstraint<type>("T") \ .TypeConstraint<type>("T") \
.TypeConstraint<int64>("out_idx"), \ .TypeConstraint<int64>("out_idx"), \
UniqueOp<type, int64>); \ UniqueOp<type, key_type, int64>); \
REGISTER_KERNEL_BUILDER(Name("UniqueWithCounts") \ REGISTER_KERNEL_BUILDER(Name("UniqueWithCounts") \
.Device(DEVICE_CPU) \ .Device(DEVICE_CPU) \
.TypeConstraint<type>("T") \ .TypeConstraint<type>("T") \
.TypeConstraint<int32>("out_idx"), \ .TypeConstraint<int32>("out_idx"), \
UniqueOp<type, int32>) \ UniqueOp<type, key_type, int32>) \
REGISTER_KERNEL_BUILDER(Name("UniqueWithCounts") \ REGISTER_KERNEL_BUILDER(Name("UniqueWithCounts") \
.Device(DEVICE_CPU) \ .Device(DEVICE_CPU) \
.TypeConstraint<type>("T") \ .TypeConstraint<type>("T") \
.TypeConstraint<int64>("out_idx"), \ .TypeConstraint<int64>("out_idx"), \
UniqueOp<type, int64>); \ UniqueOp<type, key_type, int64>); \
REGISTER_KERNEL_BUILDER(Name("UniqueWithCountsV2") \ REGISTER_KERNEL_BUILDER(Name("UniqueWithCountsV2") \
.Device(DEVICE_CPU) \ .Device(DEVICE_CPU) \
.TypeConstraint<type>("T") \ .TypeConstraint<type>("T") \
.TypeConstraint<int32>("out_idx"), \ .TypeConstraint<int32>("out_idx"), \
UniqueOp<type, int32>) \ UniqueOp<type, key_type, int32>) \
REGISTER_KERNEL_BUILDER(Name("UniqueWithCountsV2") \ REGISTER_KERNEL_BUILDER(Name("UniqueWithCountsV2") \
.Device(DEVICE_CPU) \ .Device(DEVICE_CPU) \
.TypeConstraint<type>("T") \ .TypeConstraint<type>("T") \
.TypeConstraint<int64>("out_idx"), \ .TypeConstraint<int64>("out_idx"), \
UniqueOp<type, int64>) UniqueOp<type, key_type, int64>)
TF_CALL_REAL_NUMBER_TYPES(REGISTER_UNIQUE); #define REGISTER_UNIQUE_WITH_SAME_KEY_TYPE(type) \
REGISTER_UNIQUE(tstring) REGISTER_UNIQUE_WITH_KEY_TYPE(type, type)
REGISTER_UNIQUE(bool)
#undef REGISTER_UNIQUE TF_CALL_REAL_NUMBER_TYPES(REGISTER_UNIQUE_WITH_SAME_KEY_TYPE);
REGISTER_UNIQUE_WITH_SAME_KEY_TYPE(bool)
#undef REGISTER_UNIQUE_WITH_SAME_KEY_TYPE
REGISTER_UNIQUE_WITH_KEY_TYPE(tstring, absl::string_view)
#undef REGISTER_UNIQUE_WITH_KEY_TYPE
// Fake integer GPU kernels so that the use of Unique in optimizers (to // Fake integer GPU kernels so that the use of Unique in optimizers (to
// de-duplicate sparse gradient indices) does not conflict with gradients being // de-duplicate sparse gradient indices) does not conflict with gradients being
@ -251,7 +267,7 @@ REGISTER_KERNEL_BUILDER(Name("Unique")
.HostMemory("x") .HostMemory("x")
.HostMemory("y") .HostMemory("y")
.HostMemory("idx"), .HostMemory("idx"),
UniqueOp<int32, int32>); UniqueOp<int32, int32, int32>);
REGISTER_KERNEL_BUILDER(Name("Unique") REGISTER_KERNEL_BUILDER(Name("Unique")
.Device(DEVICE_GPU) .Device(DEVICE_GPU)
.TypeConstraint<int32>("T") .TypeConstraint<int32>("T")
@ -259,7 +275,7 @@ REGISTER_KERNEL_BUILDER(Name("Unique")
.HostMemory("x") .HostMemory("x")
.HostMemory("y") .HostMemory("y")
.HostMemory("idx"), .HostMemory("idx"),
UniqueOp<int32, int64>); UniqueOp<int32, int32, int64>);
REGISTER_KERNEL_BUILDER(Name("Unique") REGISTER_KERNEL_BUILDER(Name("Unique")
.Device(DEVICE_GPU) .Device(DEVICE_GPU)
.TypeConstraint<int64>("T") .TypeConstraint<int64>("T")
@ -267,7 +283,7 @@ REGISTER_KERNEL_BUILDER(Name("Unique")
.HostMemory("x") .HostMemory("x")
.HostMemory("y") .HostMemory("y")
.HostMemory("idx"), .HostMemory("idx"),
UniqueOp<int64, int32>); UniqueOp<int64, int64, int32>);
REGISTER_KERNEL_BUILDER(Name("Unique") REGISTER_KERNEL_BUILDER(Name("Unique")
.Device(DEVICE_GPU) .Device(DEVICE_GPU)
.TypeConstraint<int64>("T") .TypeConstraint<int64>("T")
@ -275,7 +291,7 @@ REGISTER_KERNEL_BUILDER(Name("Unique")
.HostMemory("x") .HostMemory("x")
.HostMemory("y") .HostMemory("y")
.HostMemory("idx"), .HostMemory("idx"),
UniqueOp<int64, int64>); UniqueOp<int64, int64, int64>);
#ifdef TENSORFLOW_USE_SYCL #ifdef TENSORFLOW_USE_SYCL
REGISTER_KERNEL_BUILDER(Name("Unique") REGISTER_KERNEL_BUILDER(Name("Unique")
@ -285,7 +301,7 @@ REGISTER_KERNEL_BUILDER(Name("Unique")
.HostMemory("x") .HostMemory("x")
.HostMemory("y") .HostMemory("y")
.HostMemory("idx"), .HostMemory("idx"),
UniqueOp<int32, int32>); UniqueOp<int32, int32, int32>);
REGISTER_KERNEL_BUILDER(Name("Unique") REGISTER_KERNEL_BUILDER(Name("Unique")
.Device(DEVICE_SYCL) .Device(DEVICE_SYCL)
.TypeConstraint<int64>("T") .TypeConstraint<int64>("T")
@ -293,7 +309,7 @@ REGISTER_KERNEL_BUILDER(Name("Unique")
.HostMemory("x") .HostMemory("x")
.HostMemory("y") .HostMemory("y")
.HostMemory("idx"), .HostMemory("idx"),
UniqueOp<int64, int32>); UniqueOp<int64, int64, int32>);
REGISTER_KERNEL_BUILDER(Name("Unique") REGISTER_KERNEL_BUILDER(Name("Unique")
.Device(DEVICE_SYCL) .Device(DEVICE_SYCL)
.TypeConstraint<int32>("T") .TypeConstraint<int32>("T")
@ -301,7 +317,7 @@ REGISTER_KERNEL_BUILDER(Name("Unique")
.HostMemory("x") .HostMemory("x")
.HostMemory("y") .HostMemory("y")
.HostMemory("idx"), .HostMemory("idx"),
UniqueOp<int32, int64>); UniqueOp<int32, int32, int64>);
REGISTER_KERNEL_BUILDER(Name("Unique") REGISTER_KERNEL_BUILDER(Name("Unique")
.Device(DEVICE_SYCL) .Device(DEVICE_SYCL)
.TypeConstraint<int64>("T") .TypeConstraint<int64>("T")
@ -309,6 +325,8 @@ REGISTER_KERNEL_BUILDER(Name("Unique")
.HostMemory("x") .HostMemory("x")
.HostMemory("y") .HostMemory("y")
.HostMemory("idx"), .HostMemory("idx"),
UniqueOp<int64, int64>); UniqueOp<int64, int64, int64>);
#endif // TENSORFLOW_USE_SYCL #endif // TENSORFLOW_USE_SYCL
} // namespace
} // namespace tensorflow } // namespace tensorflow

16
tensorflow/core/kernels/unique_op_test.cc
View File

@ -22,6 +22,7 @@ limitations under the License.
#include "tensorflow/core/framework/tensor_shape.pb.h" #include "tensorflow/core/framework/tensor_shape.pb.h"
#include "tensorflow/core/framework/types.h" #include "tensorflow/core/framework/types.h"
#include "tensorflow/core/framework/types.pb.h" #include "tensorflow/core/framework/types.pb.h"
#include "tensorflow/core/graph/algorithm.h"
#include "tensorflow/core/graph/node_builder.h" #include "tensorflow/core/graph/node_builder.h"
#include "tensorflow/core/graph/testlib.h" #include "tensorflow/core/graph/testlib.h"
#include "tensorflow/core/kernels/ops_testutil.h" #include "tensorflow/core/kernels/ops_testutil.h"
@ -75,11 +76,14 @@ static void BM_Unique_INT32(int iters, int dim, int max_int) {
.Input(test::graph::Constant(g, input)) .Input(test::graph::Constant(g, input))
.Attr("T", DT_INT32) .Attr("T", DT_INT32)
.Finalize(g, &node)); .Finalize(g, &node));
FixupSourceAndSinkEdges(g);
testing::BytesProcessed(static_cast<int64>(iters) * dim * sizeof(int32)); testing::BytesProcessed(static_cast<int64>(iters) * dim * sizeof(int32));
testing::UseRealTime(); testing::UseRealTime();
testing::StartTiming(); testing::StartTiming();
test::Benchmark("cpu", g).Run(iters); test::Benchmark("cpu", g, nullptr, nullptr, nullptr,
"SINGLE_THREADED_EXECUTOR")
.Run(iters);
} }
static void BM_Unique_INT32_Repeat(int iters, int dim, int max_int) { static void BM_Unique_INT32_Repeat(int iters, int dim, int max_int) {
@ -95,12 +99,15 @@ static void BM_Unique_INT32_Repeat(int iters, int dim, int max_int) {
.Input(test::graph::Constant(g, input)) .Input(test::graph::Constant(g, input))
.Attr("T", DT_INT32) .Attr("T", DT_INT32)
.Finalize(g, &node)); .Finalize(g, &node));
FixupSourceAndSinkEdges(g);
testing::BytesProcessed(static_cast<int64>(iters) * dim * 200 * testing::BytesProcessed(static_cast<int64>(iters) * dim * 200 *
sizeof(int32)); sizeof(int32));
testing::UseRealTime(); testing::UseRealTime();
testing::StartTiming(); testing::StartTiming();
test::Benchmark("cpu", g).Run(iters); test::Benchmark("cpu", g, nullptr, nullptr, nullptr,
"SINGLE_THREADED_EXECUTOR")
.Run(iters);
} }
TensorProto GetRandomStringsTensorProto(int dim, int max_str_len) { TensorProto GetRandomStringsTensorProto(int dim, int max_str_len) {
@ -132,11 +139,14 @@ static void BM_Unique_STRING(int iters, int dim) {
.Input(test::graph::Constant(g, input)) .Input(test::graph::Constant(g, input))
.Attr("T", DT_STRING) .Attr("T", DT_STRING)
.Finalize(g, &node)); .Finalize(g, &node));
FixupSourceAndSinkEdges(g);
testing::BytesProcessed(static_cast<int64>(iters) * dim * sizeof(tstring)); testing::BytesProcessed(static_cast<int64>(iters) * dim * sizeof(tstring));
testing::UseRealTime(); testing::UseRealTime();
testing::StartTiming(); testing::StartTiming();
test::Benchmark("cpu", g).Run(iters); test::Benchmark("cpu", g, nullptr, nullptr, nullptr,
"SINGLE_THREADED_EXECUTOR")
.Run(iters);
} }
BENCHMARK(BM_Unique_INT32) BENCHMARK(BM_Unique_INT32)