openmp

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

OpenMP

Purpose

目的

Guide agents through OpenMP shared-memory parallelism:

#pragma omp parallel for

with scheduling clauses, reductions, data-sharing attributes, SIMD hints, task parallelism, OpenMP 5.x GPU

target

offloading, common pitfalls (false sharing, data races), environment tuning, and profiling with Score-P or TAU.

指导 Agent 进行 OpenMP 共享内存并行编程：带调度子句的

#pragma omp parallel for

、reduction、数据共享属性、SIMD 提示、任务并行、OpenMP 5.x GPU

target

卸载、常见陷阱（false sharing、数据竞争）、环境调优，以及使用 Score-P 或 TAU 进行性能分析。

When to Use

使用场景

Parallelizing C/C++/Fortran loops on multicore CPUs
Implementing reductions (sum, max, custom)
Task parallelism for irregular workloads
Offloading compute to GPU with OpenMP target directives
Diagnosing scaling failures (false sharing, load imbalance)
Tuning thread count and spin behavior

在多核 CPU 上并行化 C/C++/Fortran 循环
实现 reduction（求和、最大值、自定义）
针对不规则工作负载使用任务并行
使用 OpenMP target 指令将计算卸载到 GPU
诊断并行扩展性能问题（false sharing、负载不均衡）
调整线程数量和自旋行为

Workflow

工作流

1. Basic parallel for

1. 基础并行 for 循环

#include <omp.h>
#include <stdio.h>

int main(void) {
    const int n = 1000000;
    double sum = 0.0;

    #pragma omp parallel for reduction(+:sum)
    for (int i = 0; i < n; i++)
        sum += i * 0.001;

    printf("sum = %f, threads = %d\n", sum, omp_get_max_threads());
    return 0;
}

bash

gcc -fopenmp -O3 -o omp_sum omp_sum.c
export OMP_NUM_THREADS=8
./omp_sum

#include <omp.h>
#include <stdio.h>

int main(void) {
    const int n = 1000000;
    double sum = 0.0;

    #pragma omp parallel for reduction(+:sum)
    for (int i = 0; i < n; i++)
        sum += i * 0.001;

    printf("sum = %f, threads = %d\n", sum, omp_get_max_threads());
    return 0;
}

bash

gcc -fopenmp -O3 -o omp_sum omp_sum.c
export OMP_NUM_THREADS=8
./omp_sum

2. Schedule clauses

2. 调度子句

#pragma omp parallel for schedule(static)          // equal chunks, low overhead
#pragma omp parallel for schedule(dynamic, 64)     // dynamic chunks of 64
#pragma omp parallel for schedule(guided)          // decreasing chunk size
#pragma omp parallel for schedule(auto)            // compiler/runtime decides

Schedule	Best for
`static`	Uniform work per iteration
`dynamic`	Variable iteration cost
`guided`	Decreasing iteration cost
`static,1`	Cache blocking with interleaved chunks

#pragma omp parallel for schedule(static)          // equal chunks, low overhead
#pragma omp parallel for schedule(dynamic, 64)     // dynamic chunks of 64
#pragma omp parallel for schedule(guided)          // decreasing chunk size
#pragma omp parallel for schedule(auto)            // compiler/runtime decides

Schedule	适用场景
`static`	每次迭代的工作量均匀
`dynamic`	每次迭代的成本可变
`guided`	每次迭代的成本递减
`static,1`	使用交错式块进行缓存阻塞

3. Data sharing attributes

3. 数据共享属性

int shared_var = 0;
#pragma omp parallel private(i) shared(shared_var)
{
    int i = omp_get_thread_num();
    #pragma omp atomic
    shared_var += i;
}

// firstprivate — copy in; lastprivate — copy out after loop
#pragma omp parallel for firstprivate(offset) lastprivate(result)
for (int i = 0; i < n; i++) { ... }

Clause	Meaning
`private`	Uninitialized per-thread copy
`shared`	One variable, all threads
`reduction(op:var)`	Combine at end (+, *, max, &&, \|\|)
`firstprivate`	Initialize from master
`lastprivate`	Master gets last iteration value

int shared_var = 0;
#pragma omp parallel private(i) shared(shared_var)
{
    int i = omp_get_thread_num();
    #pragma omp atomic
    shared_var += i;
}

// firstprivate — copy in; lastprivate — copy out after loop
#pragma omp parallel for firstprivate(offset) lastprivate(result)
for (int i = 0; i < n; i++) { ... }

子句	含义
`private`	每线程一个未初始化的副本
`shared`	一个变量，由所有线程共享
`reduction(op:var)`	在结束时合并（+, *, max, &&, \|\|）
`firstprivate`	从 master 初始化
`lastprivate`	master 获取最后一次迭代的值

4. SIMD vectorization hint

4. SIMD 向量化提示

#pragma omp simd
for (int i = 0; i < n; i++)
    c[i] = a[i] + b[i];

// SIMD + parallel
#pragma omp parallel for simd
for (int i = 0; i < n; i++)
    c[i] = a[i] * b[i];

Requires

-fopenmp-simd

-fopenmp

with compiler SIMD support. Check with

-fopt-info-vec

#pragma omp simd
for (int i = 0; i < n; i++)
    c[i] = a[i] + b[i];

// SIMD + parallel
#pragma omp parallel for simd
for (int i = 0; i < n; i++)
    c[i] = a[i] * b[i];

需要使用

-fopenmp-simd

，或使用支持 SIMD 的编译器并启用

-fopenmp

。可通过

-fopt-info-vec

检查。

5. Task parallelism

5. 任务并行

#pragma omp parallel
{
    #pragma omp single
    {
        for (int i = 0; i < 10; i++) {
            #pragma omp task firstprivate(i)
            process_subtree(i);
        }
        #pragma omp taskwait
    }
}

Tasks suit recursive algorithms (quicksort, tree traversal) where loop parallelism doesn't fit.

#pragma omp parallel
{
    #pragma omp single
    {
        for (int i = 0; i < 10; i++) {
            #pragma omp task firstprivate(i)
            process_subtree(i);
        }
        #pragma omp taskwait
    }
}

Task 适合递归算法（quicksort、树遍历）等不适合使用循环并行的场景。

6. Timing

6. 计时

double start = omp_get_wtime();
#pragma omp parallel for
for (int i = 0; i < n; i++) work(i);
double elapsed = omp_get_wtime() - start;
printf("elapsed: %f s\n", elapsed);

double start = omp_get_wtime();
#pragma omp parallel for
for (int i = 0; i < n; i++) work(i);
double elapsed = omp_get_wtime() - start;
printf("elapsed: %f s\n", elapsed);

7. GPU target offloading (OpenMP 5.x)

7. GPU target 卸载（OpenMP 5.x）

#pragma omp target teams distribute parallel for map(to:a[0:n]) map(from:c[0:n])
for (int i = 0; i < n; i++)
    c[i] = a[i] * 2.0f;

bash

undefined

#pragma omp target teams distribute parallel for map(to:a[0:n]) map(from:c[0:n])
for (int i = 0; i < n; i++)
    c[i] = a[i] * 2.0f;

bash

undefined

NVIDIA offload

gcc -fopenmp -foffload=-march=sm_80 -o offload offload.c

Check device

export OMP_DEFAULT_TARGET_DEVICE=1


Requires compiler offload support (GCC offload, Clang/OpenMP, NVIDIA HPC SDK).

export OMP_DEFAULT_TARGET_DEVICE=1


需要编译器支持 offload（GCC offload、Clang/OpenMP、NVIDIA HPC SDK）。

8. Environment variables

8. 环境变量

bash

export OMP_NUM_THREADS=16
export OMP_PROC_BIND=close        # bind threads to nearby cores
export OMP_PLACES=cores
export GOMP_SPINCOUNT=2000        # spin before sleep
export OMP_WAIT_POLICY=active     # active vs passive waiting
export OMP_DISPLAY_ENV=true       # print config at startup

bash

export OMP_NUM_THREADS=16
export OMP_PROC_BIND=close        # bind threads to nearby cores
export OMP_PLACES=cores
export GOMP_SPINCOUNT=2000        # spin before sleep
export OMP_WAIT_POLICY=active     # active vs passive waiting
export OMP_DISPLAY_ENV=true       # print config at startup

9. Profiling

9. 性能分析

bash

undefined

bash

undefined

Score-P (compile with wrapper)

scorep gcc -fopenmp -o app app.c export SCOREP_METRIC_MANAGER=1 scorep ./app scorep-score -f scorep_*/profile.cubex

TAU

tau_cc.sh -fopenmp -o app app.c export TAU_TRACE=1 ./app pprof app profile.*

undefined

tau_cc.sh -fopenmp -o app app.c export TAU_TRACE=1 ./app pprof app profile.*

undefined

10. Pitfalls

10. 常见陷阱

False sharing: threads modify adjacent cache lines.

// Bad: sum_array[tid] on same cache line
#pragma omp parallel
{
    int tid = omp_get_thread_num();
    sum_array[tid] += local_sum;  // pad to 64 bytes between elements
}

// Fix: padding
double sum_padded[MAX_THREADS][8];  // 8 doubles = 64 bytes

Nested parallelism:

bash

export OMP_MAX_ACTIVE_LEVELS=2
export OMP_NESTED=true   # deprecated; use MAX_ACTIVE_LEVELS

False sharing：线程修改相邻的缓存行。

// Bad: sum_array[tid] on same cache line
#pragma omp parallel
{
    int tid = omp_get_thread_num();
    sum_array[tid] += local_sum;  // pad to 64 bytes between elements
}

// Fix: padding
double sum_padded[MAX_THREADS][8];  // 8 doubles = 64 bytes

Nested parallelism：

bash

export OMP_MAX_ACTIVE_LEVELS=2
export OMP_NESTED=true   # deprecated; use MAX_ACTIVE_LEVELS

Common Problems

常见问题

Symptom	Cause	Fix
No speedup	Loop too small	Increase work; check `if` clause threshold
Wrong reduction result	Race on non-reduction var	Use `reduction` or `atomic`
Slower with more threads	False sharing	Pad per-thread arrays
GPU offload fails	No target device	Check `-foffload` ; `nvidia-smi`
Threads not bound	Default spread	`OMP_PROC_BIND=close`
Nested deadlock	Oversubscription	Limit `OMP_NUM_THREADS` per level

症状	原因	修复方式
没有加速	循环过小	增大工作量；检查 `if` 子句阈值
reduction 结果错误	非 reduction 变量存在竞争	使用 `reduction` 或 `atomic`
线程越多反而越慢	False sharing	为每线程数组添加填充
GPU offload 失败	没有 target 设备	检查 `-foffload` ； `nvidia-smi`
线程未绑定	默认分散策略	`OMP_PROC_BIND=close`
嵌套死锁	过度订阅	限制每一层的 `OMP_NUM_THREADS`

openmp

Original

Translation

OpenMP

OpenMP

Purpose

目的

When to Use

使用场景

Workflow

工作流

1. Basic parallel for

1. 基础并行 for 循环

2. Schedule clauses

2. 调度子句

3. Data sharing attributes

3. 数据共享属性

4. SIMD vectorization hint

4. SIMD 向量化提示

5. Task parallelism

5. 任务并行

6. Timing

6. 计时

7. GPU target offloading (OpenMP 5.x)

7. GPU target 卸载（OpenMP 5.x）

NVIDIA offload

NVIDIA offload

Check device

Check device

8. Environment variables

8. 环境变量

9. Profiling

9. 性能分析

Score-P (compile with wrapper)

Score-P (compile with wrapper)

TAU

TAU

10. Pitfalls

10. 常见陷阱

Common Problems

常见问题

Related Skills

相关 Skill