Merge branch 'threadpool' into 'master'

Implement thread pool.

See merge request !727

mace/core/BUILD

@@ -64,6 +64,7 @@ cc_library(
         "//mace/codegen:generated_version",
         "//mace/proto:mace_cc",
         "//mace/utils",
+        "@gemmlowp",
     ] + if_opencl_enabled([
         ":opencl_headers",
         "//mace/codegen:generated_opencl",
@@ -73,6 +74,26 @@ cc_library(
     ]),
 )
 
+cc_test(
+    name = "core_test",
+    srcs = glob(["*_test.cc"]),
+    copts = [
+        "-Werror",
+        "-Wextra",
+        "-Wno-missing-field-initializers",
+    ],
+    linkopts = ["-ldl"] + if_android([
+        "-pie",
+        "-lm",
+    ]),
+    deps = [
+        ":core",
+        "//mace/utils",
+        "@gtest",
+        "@gtest//:gtest_main",
+    ],
+)
+
 cc_library(
     name = "opencl_headers",
     hdrs = glob([

mace/core/runtime/cpu/cpu_runtime.cc

@@ -75,6 +75,9 @@ int GetCPUMaxFreq(int cpu_id) {
   return freq;
 }
 
+}  // namespace
+
+
 MaceStatus SetThreadAffinity(cpu_set_t mask) {
 #if defined(__ANDROID__)
   pid_t pid = gettid();
@@ -90,8 +93,6 @@ MaceStatus SetThreadAffinity(cpu_set_t mask) {
   }
 }
 
-}  // namespace
-
 MaceStatus GetCPUBigLittleCoreIDs(std::vector<int> *big_core_ids,
                                   std::vector<int> *little_core_ids) {
   MACE_CHECK_NOTNULL(big_core_ids);

mace/core/runtime/cpu/cpu_runtime.h

@@ -15,6 +15,7 @@
 #ifndef MACE_CORE_RUNTIME_CPU_CPU_RUNTIME_H_
 #define MACE_CORE_RUNTIME_CPU_CPU_RUNTIME_H_
 
+#include <sched.h>
 #include <vector>
 
 #include "mace/public/mace.h"
@@ -31,6 +32,8 @@ MaceStatus SetOpenMPThreadsAndAffinityCPUs(int omp_num_threads,
 MaceStatus SetOpenMPThreadsAndAffinityPolicy(int omp_num_threads_hint,
                                              CPUAffinityPolicy policy);
 
+MaceStatus SetThreadAffinity(cpu_set_t mask);
+
 }  // namespace mace
 
 #endif  // MACE_CORE_RUNTIME_CPU_CPU_RUNTIME_H_

mace/core/testing/test_benchmark_main.cc

@@ -16,6 +16,7 @@
 
 #include "gflags/gflags.h"
 #include "mace/core/runtime/cpu/cpu_runtime.h"
+#include "mace/core/threadpool.h"
 #include "mace/core/runtime/opencl/opencl_runtime.h"
 #include "mace/core/testing/test_benchmark.h"
 #include "mace/public/mace.h"
@@ -36,6 +37,10 @@ int main(int argc, char **argv) {
   gflags::ParseCommandLineFlags(&argc, &argv, true);
 
   // config runtime
+  // TODO(liyin): Uncomment the following after removing openmp.
+//  mace::ThreadPoolRegister::ConfigThreadPool(FLAGS_omp_num_threads,
+//      static_cast<mace::CPUAffinityPolicy>(FLAGS_cpu_affinity_policy));
+
   mace::MaceStatus status = mace::SetOpenMPThreadsAndAffinityPolicy(
       FLAGS_omp_num_threads,
       static_cast<mace::CPUAffinityPolicy>(FLAGS_cpu_affinity_policy));

mace/core/threadpool.cc

+// Copyright 2018 Xiaomi, Inc.  All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <utility>
+#include <algorithm>
+#include <functional>
+#include <vector>
+
+#include "mace/core/threadpool.h"
+#include "mace/utils/logging.h"
+
+namespace mace {
+
+namespace {
+class InitTask : public gemmlowp::Task {
+ public:
+  explicit InitTask(const std::vector<int> &cpu_ids)
+      : Task(), cpu_ids_(cpu_ids) {}
+  void Run() {
+    cpu_set_t mask;
+    CPU_ZERO(&mask);
+    for (auto cpu_id : cpu_ids_) {
+      CPU_SET(cpu_id, &mask);
+    }
+    SetThreadAffinity(mask);
+  }
+
+ private:
+  const std::vector<int> cpu_ids_;
+};
+
+class BlockTask : public gemmlowp::Task {
+ public:
+  BlockTask(const int64_t start,
+            const int64_t end,
+            const std::function<void(const int64_t, const int64_t)> &fn)
+      : Task(),
+        start_(start),
+        end_(end),
+        fn_(fn) {}
+
+  void Run() {
+    fn_(start_, end_);
+  }
+
+ private:
+  const int64_t start_;
+  const int64_t end_;
+  const std::function<void(int64_t, int64_t)> fn_;
+};
+}  // namespace
+
+ThreadPool::ThreadPool(int num_threads, CPUAffinityPolicy policy)
+    : num_threads_(num_threads), busy_(false) {
+  std::vector<int> big_core_ids;
+  std::vector<int> little_core_ids;
+  std::vector<int> use_cpu_ids;
+  MaceStatus res = GetCPUBigLittleCoreIDs(&big_core_ids, &little_core_ids);
+  if (res == MaceStatus::MACE_SUCCESS) {
+    if (policy == CPUAffinityPolicy::AFFINITY_BIG_ONLY) {
+      use_cpu_ids = std::move(big_core_ids);
+    } else {
+      use_cpu_ids = std::move(little_core_ids);
+    }
+
+    if (num_threads <= 0 ||
+        num_threads > static_cast<int>(use_cpu_ids.size())) {
+      num_threads_ = use_cpu_ids.size();
+    }
+  } else {
+    LOG(WARNING) << "Failed to get cpu big little cores info";
+  }
+
+  // if set mask failed and user set num_threads as -1, use single thread
+  if (num_threads_ < 0) {
+    num_threads_ = 1;
+  }
+
+  VLOG(2) << "Use thread count: " << num_threads_;
+
+  std::vector<gemmlowp::Task *> init_tasks(num_threads_);
+  for (int i = 0; i < num_threads_; ++i) {
+    init_tasks[i] = new InitTask(use_cpu_ids);
+  }
+  workers_pool_.Execute(init_tasks);
+}
+
+ThreadPool::~ThreadPool() {}
+
+void ThreadPool::ParallelRun(const int64_t total,
+                             std::function<void(const int64_t,
+                                                const int64_t)> fn) {
+  if (total <= 0) {
+    return;
+  } else if (total == 1 || busy_) {
+    fn(0, total);
+    return;
+  }
+  busy_ = true;
+
+  const int64_t
+      shards = std::min(total, static_cast<const int64_t>(num_threads_));
+  const int64_t work_size_per_shard = total / shards;
+  const int64_t remain = total - shards * work_size_per_shard;
+  const int64_t work_size_per_shard_plus_one = work_size_per_shard + 1;
+
+  std::vector<gemmlowp::Task *> tasks(shards);
+
+  int64_t start = 0;
+  int64_t end = 0;
+  for (int64_t i = 0; i < remain; ++i) {
+    end = start + work_size_per_shard_plus_one;
+    tasks[i] = new BlockTask(start, end, fn);
+    start = end;
+  }
+  for (int64_t i = remain; i < shards; ++i) {
+    end = start + work_size_per_shard;
+    tasks[i] = new BlockTask(start, end, fn);
+    start = end;
+  }
+
+  workers_pool_.Execute(tasks);
+  busy_ = false;
+}
+
+void ThreadPool::ParallelFor(const int64_t start,
+                             const int64_t end,
+                             const int64_t step,
+                             std::function<void(const int64_t)> fn) {
+  MACE_CHECK(start <= end && step > 0, "start must be le end and step gt 0");
+  const int64_t total = (end - start + step - 1) / step;
+  ParallelRun(total, [&](const int64_t s, const int64_t t) {
+    const int64_t start_i = start + s * step;
+    const int64_t end_i = std::min(end, start + t * step);
+    for (int64_t i = start_i; i < end_i; i += step) {
+      fn(i);
+    }
+  });
+}
+
+void ThreadPool::ParallelFor(const int64_t start1,
+                             const int64_t end1,
+                             const int64_t step1,
+                             const int64_t start2,
+                             const int64_t end2,
+                             const int64_t step2,
+                             std::function<void(const int64_t,
+                                                const int64_t)> fn) {
+  MACE_CHECK(start1 <= end1 && step1 > 0 && start2 <= end2 && step2 > 0,
+             "start must be le end and step gt 0");
+
+  const int64_t total1 = ((end1 - start1 + step1 - 1) / step1);
+  const int64_t total2 = ((end2 - start2 + step2 - 1) / step2);
+  const int64_t total = total1 * total2;
+
+  if (total == 0) {
+    return;
+  } else if (total1 == 1) {
+    ParallelFor(start2,
+                end2,
+                step2,
+                [&](const int64_t arg2) {
+                  fn(start1, arg2);
+                });
+  } else if (total1 % num_threads_ == 0) {
+    ParallelFor(start1,
+                end1,
+                step1,
+                [&](const int64_t arg1) {
+                  for (int64_t i = start2; i < end2; i += step2) {
+                    fn(arg1, i);
+                  }
+                });
+  } else {
+    ParallelRun(total, [&](const int64_t s, const int64_t t) {
+      for (int64_t idx = s; idx < t; ++idx) {
+        const int64_t i = idx / total2;
+        const int64_t j = idx - i * total2;
+        fn(start1 + step1 * i, start2 + step2 * j);
+      }
+    });
+  }
+}
+
+void ThreadPool::ParallelFor(const int64_t start1,
+                             const int64_t end1,
+                             const int64_t step1,
+                             const int64_t start2,
+                             const int64_t end2,
+                             const int64_t step2,
+                             const int64_t start3,
+                             const int64_t end3,
+                             const int64_t step3,
+                             std::function<void(const int64_t,
+                                                const int64_t,
+                                                const int64_t)> fn) {
+  MACE_CHECK(start1 <= end1 && step1 > 0 && start2 <= end2 && step2 > 0
+                 && start3 <= end3 && step3 > 0,
+             "start must be le end and step gt 0");
+
+  const int64_t total1 = ((end1 - start1 + step1 - 1) / step1);
+  const int64_t total2 = ((end2 - start2 + step2 - 1) / step2);
+  const int64_t total3 = ((end3 - start3 + step3 - 1) / step3);
+  const int64_t total23 = total2 * total3;
+  const int64_t total = total1 * total23;
+
+  if (total == 0) {
+    return;
+  } else if (total1 == 1) {
+    ParallelFor(start2,
+                end2,
+                step2,
+                start3,
+                end3,
+                step3,
+                [&](const int64_t arg2, const int64_t arg3) {
+                  fn(start1, arg2, arg3);
+                });
+  } else if ((total1 * total2) % num_threads_ == 0) {
+    ParallelFor(start1,
+                end1,
+                step1,
+                start2,
+                end2,
+                step2,
+                [&](const int64_t arg1, const int64_t arg2) {
+                  for (int64_t i = start3; i < end3; i += step3) {
+                    fn(arg1, arg2, i);
+                  }
+                });
+  } else {
+    ParallelRun(total, [&](int64_t s, int64_t t) {
+      for (int64_t idx = s; idx < t; ++idx) {
+        const int64_t i = idx / total23;
+        const int64_t mod_i = idx - i * total23;
+        const int64_t j = mod_i / total3;
+        const int64_t k = mod_i - j * total3;
+        fn(start1 + step1 * i, start2 + step2 * j, start3 + step3 * k);
+      }
+    });
+  }
+}
+
+ThreadPool *ThreadPoolRegister::thread_pool = nullptr;
+
+ThreadPool *ThreadPoolRegister::GetThreadPool() {
+  if (thread_pool == nullptr) {
+    ConfigThreadPool(-1, CPUAffinityPolicy::AFFINITY_NONE);
+  }
+  return thread_pool;
+}
+
+void ThreadPoolRegister::ConfigThreadPool(int num_threads,
+                                          CPUAffinityPolicy policy) {
+  MACE_CHECK(thread_pool == nullptr,
+             "ThreadPool has already been initialized.");
+  thread_pool = new ThreadPool(num_threads, policy);
+}
+
+}  // namespace mace

mace/core/threadpool.h

+// Copyright 2018 Xiaomi, Inc.  All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef MACE_CORE_THREADPOOL_H_
+#define MACE_CORE_THREADPOOL_H_
+
+#include "mace/core/runtime/cpu/cpu_runtime.h"
+
+// Remove the following macros after removing openmp dependency
+#ifdef GEMMLOWP_USE_OPENMP
+#define MACE_GEMMLOWP_USE_OPENMP
+#undef GEMMLOWP_USE_OPENMP
+#endif
+#include "internal/multi_thread_gemm.h"
+#ifdef MACE_GEMMLOWP_USE_OPENMP
+#define GEMMLOWP_USE_OPENMP
+#undef MACE_GEMMLOWP_USE_OPENMP
+#endif
+
+namespace mace {
+
+class ThreadPool {
+ public:
+  explicit ThreadPool(int num_threads, CPUAffinityPolicy policy);
+
+  ~ThreadPool();
+
+  void ParallelRun(const int64_t total,
+                   std::function<void(const int64_t, const int64_t)> fn);
+
+  // Parallel for
+  void ParallelFor(const int64_t start, const int64_t end, const int64_t step,
+                   std::function<void(const int64_t)> fn);
+
+  // Parallel for collapse(2)
+  void ParallelFor(const int64_t start1,
+                   const int64_t end1,
+                   const int64_t step1,
+                   const int64_t start2,
+                   const int64_t end2,
+                   const int64_t step2,
+                   std::function<void(const int64_t, const int64_t)> fn);
+
+  // Parallel for collapse(3)
+  void ParallelFor(const int64_t start1,
+                   const int64_t end1,
+                   const int64_t step1,
+                   const int64_t start2,
+                   const int64_t end2,
+                   const int64_t step2,
+                   const int64_t start3,
+                   const int64_t end3,
+                   const int64_t step3,
+                   std::function<void(const int64_t,
+                                      const int64_t,
+                                      const int64_t)> fn);
+
+ private:
+  int num_threads_;
+  bool busy_;
+  gemmlowp::WorkersPool workers_pool_;
+};
+
+class ThreadPoolRegister {
+ public:
+  static ThreadPool *GetThreadPool();
+
+  static void ConfigThreadPool(int num_threads, CPUAffinityPolicy policy);
+
+ private:
+  static ThreadPool *thread_pool;
+};
+
+}  // namespace mace
+
+
+
+#endif  // MACE_CORE_THREADPOOL_H_

mace/core/threadpool_test.cc

+// Copyright 2018 Xiaomi, Inc.  All rights reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <gtest/gtest.h>
+#include <vector>
+
+#include "mace/core/threadpool.h"
+#include "mace/utils/logging.h"
+
+namespace mace {
+namespace kernels {
+namespace test {
+
+namespace {
+
+void TestParallelRun(int64_t size) {
+  ThreadPool thread_pool(4, CPUAffinityPolicy::AFFINITY_NONE);
+  std::vector<int64_t> input(size);
+  thread_pool.ParallelRun(size, [&](const int64_t start, const int64_t end) {
+    for (int64_t i = start; i < end; ++i) {
+      input[i] = i;
+    }
+  });
+
+  for (int64_t i = 0; i < size; ++i) {
+    EXPECT_EQ(input[i], i);
+  }
+}
+
+void TestParallelForLoop1(int64_t start, int64_t end, int64_t step) {
+  ThreadPool thread_pool(4, CPUAffinityPolicy::AFFINITY_NONE);
+  std::vector<int64_t> input(end);
+  thread_pool.ParallelFor(start, end, step, [&](const int64_t i) {
+    input[i] = i;
+  });
+
+  for (int64_t i = start; i < end; i += step) {
+    EXPECT_EQ(input[i], i);
+  }
+}
+
+void TestParallelForLoop2(int64_t start1, int64_t end1, int64_t step1,
+                          int64_t start2, int64_t end2, int64_t step2) {
+  ThreadPool thread_pool(4, CPUAffinityPolicy::AFFINITY_BIG_ONLY);
+  std::vector<std::vector<int64_t>> input(end1, std::vector<int64_t>(end2));
+  thread_pool.ParallelFor(start1, end1, step1,
+                          start2, end2, step2,
+                          [&](const int64_t i, const int64_t j) {
+                            input[i][j] = i * j;
+                          });
+
+  for (int64_t i = start1; i < end1; i += step1) {
+    for (int64_t j = start2; j < end2; j += step2) {
+      EXPECT_EQ(input[i][j], i * j);
+    }
+  }
+}
+
+void TestParallelForLoop3(int64_t start1, int64_t end1, int64_t step1,
+                          int64_t start2, int64_t end2, int64_t step2,
+                          int64_t start3, int64_t end3, int64_t step3) {
+  ThreadPool thread_pool(4, CPUAffinityPolicy::AFFINITY_BIG_ONLY);
+  std::vector<std::vector<std::vector<int64_t>>>
+      input(end1,
+            std::vector<std::vector<int64_t>>(end2,
+                                              std::vector<int64_t>(end3)));
+  thread_pool.ParallelFor(start1, end1, step1,
+                          start2, end2, step2,
+                          start3, end3, step3,
+                          [&](const int64_t i,
+                              const int64_t j,
+                              const int64_t k) {
+                            input[i][j][k] = i * j * k;
+                          });
+
+  for (int64_t i = start1; i < end1; i += step1) {
+    for (int64_t j = start2; j < end2; j += step2) {
+      for (int64_t k = start3; k < end3; k += step3) {
+        EXPECT_EQ(input[i][j][k], i * j * k);
+      }
+    }
+  }
+}
+
+}  // namespace
+
+TEST(ThreadPoolTest, TestParallelRun) {
+  TestParallelRun(102);
+}
+
+TEST(ThreadPoolTest, TestParallelFor1) {
+  TestParallelForLoop1(1, 102, 2);
+}
+
+TEST(ThreadPoolTest, TestParallelFor2) {
+  TestParallelForLoop2(1, 1, 2, 2, 53, 4);
+  TestParallelForLoop2(1, 102, 2, 2, 53, 4);
+  TestParallelForLoop2(1, 2, 2, 2, 53, 4);
+  TestParallelForLoop2(1, 101, 1, 2, 53, 4);
+}
+
+TEST(ThreadPoolTest, TestParallelFor3) {
+  TestParallelForLoop3(1, 1, 2, 2, 53, 4, 3, 31, 3);
+  TestParallelForLoop3(1, 102, 2, 2, 53, 4, 3, 31, 3);
+  TestParallelForLoop3(1, 2, 2, 2, 53, 4, 3, 31, 3);
+  TestParallelForLoop3(1, 5, 1, 2, 53, 4, 3, 31, 3);
+  TestParallelForLoop3(1, 5, 2, 2, 50, 24, 3, 31, 3);
+}
+
+}  // namespace test
+}  // namespace kernels
+}  // namespace mace