• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
2 //
3 // Licensed under the Apache License, Version 2.0 (the "License");
4 // you may not use this file except in compliance with the License.
5 // You may obtain a copy of the License at
6 //
7 //     http://www.apache.org/licenses/LICENSE-2.0
8 //
9 // Unless required by applicable law or agreed to in writing, software
10 // distributed under the License is distributed on an "AS IS" BASIS,
11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 // See the License for the specific language governing permissions and
13 // limitations under the License.
14 
15 // multi_thread_common.h: Multithreading code shared by different meta gemm
16 // versions.
17 
18 #ifndef GEMMLOWP_META_MULTI_THREAD_COMMON_H_
19 #define GEMMLOWP_META_MULTI_THREAD_COMMON_H_
20 
21 #include "../internal/multi_thread_gemm.h"
22 
23 namespace gemmlowp {
24 namespace meta {
25 namespace internal {
26 
27 const std::int32_t kMinTaskSize = 16000;
28 const std::int32_t kMinTaskDimension = 4;
29 
30 struct TaskRect {
31   std::int32_t m_offset;
32   std::int32_t m;
33   std::int32_t n_offset;
34   std::int32_t n;
35 
TaskRectTaskRect36   TaskRect(std::int32_t m_offset, std::int32_t m, std::int32_t n_offset,
37            std::int32_t n)
38       : m_offset(m_offset), m(m), n_offset(n_offset), n(n) {}
39 };
40 
41 template <typename IN_TYPE, typename OUT_TYPE, typename F>
42 struct MetaTask : gemmlowp::Task {
43   std::uint8_t* scratch;
44   const IN_TYPE* lhs;
45   const IN_TYPE* rhs;
46   TaskRect task_rect;
47   std::int32_t k;
48   OUT_TYPE* result;
49   std::int32_t result_stride;
50   const F& operation;
51 
MetaTaskMetaTask52   MetaTask(std::uint8_t* scratch, const IN_TYPE* lhs, const IN_TYPE* rhs,
53            const TaskRect& task_rect, std::int32_t k, OUT_TYPE* result,
54            std::int32_t result_stride, const F& operation)
55       : scratch(scratch),
56         lhs(lhs),
57         rhs(rhs),
58         task_rect(task_rect),
59         k(k),
60         result(result),
61         result_stride(result_stride),
62         operation(operation) {}
63 
RunMetaTask64   void Run() override {
65     const IN_TYPE* task_lhs = lhs + task_rect.m_offset * k;
66     const IN_TYPE* task_rhs = rhs + task_rect.n_offset * k;
67     OUT_TYPE* task_result =
68         result + task_rect.m_offset * result_stride + task_rect.n_offset;
69     operation.ExecuteMatrixMatrix(scratch, task_lhs, task_rhs, task_rect.m,
70                                   task_rect.n, k, task_result, result_stride);
71   }
72 };
73 
ResolveMaxThreads(std::int32_t max_threads)74 std::int32_t ResolveMaxThreads(std::int32_t max_threads) {
75   if (max_threads == 0) {
76     static const int hardware_threads_count =
77         static_cast<int>(sysconf(_SC_NPROCESSORS_CONF));
78     return hardware_threads_count;
79   }
80   return max_threads;
81 }
82 
PrepareTasks(std::int32_t max_tasks,std::int32_t m,std::int32_t n,std::int32_t k,std::vector<internal::TaskRect> * tasks)83 void PrepareTasks(std::int32_t max_tasks, std::int32_t m, std::int32_t n,
84                   std::int32_t k, std::vector<internal::TaskRect>* tasks) {
85   const std::int32_t max_tasks_by_size = (m * n * k) / kMinTaskSize;
86   const std::int32_t max_tasks_m = m / kMinTaskDimension;
87   const std::int32_t max_tasks_n = n / kMinTaskDimension;
88   const std::int32_t max_tasks_dimension = std::max(max_tasks_m, max_tasks_n);
89 
90   std::int32_t real_tasks = std::max(
91       1, std::min(max_tasks, std::min(max_tasks_by_size, max_tasks_dimension)));
92 
93   if (real_tasks == 1) {
94     tasks->push_back(TaskRect(0, m, 0, n));
95     return;
96   }
97 
98   if (max_tasks_m > max_tasks_n) {
99     const std::int32_t m_chunk = m / real_tasks;
100     for (int i = 0; i < real_tasks - 1; ++i) {
101       tasks->push_back(TaskRect(i * m_chunk, m_chunk, 0, n));
102     }
103     const std::int32_t last_m_offset = (real_tasks - 1) * m_chunk;
104     tasks->push_back(TaskRect(last_m_offset, m - last_m_offset, 0, n));
105   } else {
106     const std::int32_t n_chunk = n / real_tasks;
107     for (int i = 0; i < real_tasks - 1; ++i) {
108       tasks->push_back(TaskRect(0, m, i * n_chunk, n_chunk));
109     }
110     const std::int32_t last_n_offset = (real_tasks - 1) * n_chunk;
111     tasks->push_back(TaskRect(0, m, last_n_offset, n - last_n_offset));
112   }
113 }
114 
115 template <typename IN_TYPE, typename OUT_TYPE, typename F>
MultiThreadedMatrixMatrix(gemmlowp::WorkersPool * pool,std::int32_t max_threads,std::uint8_t * scratch,const IN_TYPE * lhs,const IN_TYPE * rhs,std::int32_t m,std::int32_t n,std::int32_t k,OUT_TYPE * result,std::int32_t result_stride,const F & operation)116 void MultiThreadedMatrixMatrix(gemmlowp::WorkersPool* pool,
117                                std::int32_t max_threads, std::uint8_t* scratch,
118                                const IN_TYPE* lhs, const IN_TYPE* rhs,
119                                std::int32_t m, std::int32_t n, std::int32_t k,
120                                OUT_TYPE* result, std::int32_t result_stride,
121                                const F& operation) {
122   max_threads = internal::ResolveMaxThreads(max_threads);
123 
124   std::vector<internal::TaskRect> task_rects;
125   internal::PrepareTasks(max_threads, m, n, k, &task_rects);
126 
127   if (task_rects.size() == 1) {
128     operation.ExecuteMatrixMatrix(scratch, lhs, rhs, m, n, k, result,
129                                   result_stride);
130     return;
131   }
132 
133   std::uint8_t* task_scratch = scratch;
134   std::int32_t scratch_per_thread = operation.ScratchPerThread(m, n, k);
135   std::vector<Task*> tasks;
136   std::for_each(
137       task_rects.begin(), task_rects.end(),
138       [&tasks, &task_scratch, lhs, rhs, k, result, result_stride, operation,
139        scratch_per_thread](internal::TaskRect& rect) {
140         tasks.push_back(new internal::MetaTask<IN_TYPE, OUT_TYPE, F>(
141             task_scratch, lhs, rhs, rect, k, result, result_stride, operation));
142         task_scratch += scratch_per_thread;
143       });
144   pool->Execute(tasks);
145 }
146 
147 }  // namespace internal
148 }  // namespace meta
149 }  // namespace gemmlowp
150 
151 #endif  // GEMMLOWP_META_MULTI_THREAD_COMMON_H_
152