Implementation:Tencent Ncnn SimpleOMP
| Knowledge Sources | |
|---|---|
| Domains | Concurrency, Embedded Systems |
| Last Updated | 2026-02-09 19:00 GMT |
Overview
Implements a minimal OpenMP runtime replacement that supports the #pragma omp parallel for pattern used throughout ncnn, compatible with both LLVM (libomp) and GCC (libgomp) compiler ABIs.
Description
The simpleomp module consists of a header (simpleomp.h, 42 lines) declaring the public API and an implementation (simpleomp.cpp, 792 lines) providing the thread pool and OpenMP ABI functions. The module is conditional on the NCNN_SIMPLEOMP compile flag.
The implementation builds a thread pool system using ncnn's own Thread, Mutex, and ConditionVariable primitives:
- KMPGlobal: Manages a persistent pool of worker threads and a KMPTaskQueue (a bounded circular buffer for dispatching work).
- KMPTask: Describes a unit of parallel work, including the microtask function pointer, argument count, arguments, thread number, and completion signaling.
- KMPTaskQueue: A thread-safe bounded circular buffer that supports batch dispatch and blocking get operations.
When the compiler-generated OpenMP fork call is invoked (__kmpc_fork_call for Clang or GOMP_parallel for GCC), the runtime creates KMPTask objects and dispatches them to the thread pool. Worker threads pick tasks from the queue, execute the microtask function, and signal completion.
For Clang, the implementation handles up to 31 captured variables via typed function pointer casts (kmpc_micro_0 through kmpc_micro_31). Thread-local storage tracks the current thread's ID within a parallel region. The runtime also implements __kmpc_for_static_init / __kmpc_for_static_fini for static loop scheduling and __kmpc_barrier for barrier synchronization.
The header declares standard OpenMP API functions: omp_get_max_threads(), omp_set_num_threads(), omp_get_dynamic(), omp_set_dynamic(), omp_get_num_threads(), omp_get_thread_num(), and Intel KMP extensions kmp_get_blocktime() and kmp_set_blocktime().
Usage
This module is used automatically when building ncnn with NCNN_SIMPLEOMP=ON on platforms lacking a system OpenMP runtime (e.g., some Android NDK configurations or custom toolchains). Standard #pragma omp parallel for annotations in ncnn layer code work transparently with this backend.
Code Reference
Source Location
- Repository: Tencent_Ncnn
- File: src/simpleomp.cpp
- File: src/simpleomp.h
Signature
// Public OpenMP-compatible API
int omp_get_max_threads();
void omp_set_num_threads(int num_threads);
int omp_get_dynamic();
void omp_set_dynamic(int dynamic);
int omp_get_num_threads();
int omp_get_thread_num();
// Intel KMP extensions
int kmp_get_blocktime();
void kmp_set_blocktime(int blocktime);
Import
#include "simpleomp.h"
I/O Contract
Inputs
| Name | Type | Required | Description |
|---|---|---|---|
| num_threads | int | For omp_set_num_threads | Number of threads to use for parallel regions |
| dynamic | int | For omp_set_dynamic | Enable (1) or disable (0) dynamic thread adjustment |
| blocktime | int | For kmp_set_blocktime | Time in ms threads wait before sleeping after parallel region |
Outputs
| Name | Type | Description |
|---|---|---|
| return (omp_get_max_threads) | int | Maximum number of threads available |
| return (omp_get_num_threads) | int | Number of threads in the current parallel region |
| return (omp_get_thread_num) | int | Thread ID of the calling thread (0-based) |
| return (omp_get_dynamic) | int | Whether dynamic thread adjustment is enabled |
| return (kmp_get_blocktime) | int | Current blocktime value in milliseconds |
Usage Examples
Controlling Thread Count
#include "simpleomp.h"
// Set the number of threads for parallel regions
omp_set_num_threads(4);
// Query current settings
int max_threads = omp_get_max_threads();
Transparent Usage in Layer Code
// Layer implementations use standard OpenMP pragmas.
// simpleomp intercepts the compiler-generated ABI calls.
#pragma omp parallel for num_threads(opt.num_threads)
for (int i = 0; i < channels; i++)
{
// per-channel computation
}