Skip to main content

OMP Backend

The OMP flavor is a plain C/C++ flavor enabling support for OpenMP features. While this flavor does not offer dramatic performance improvements over JIT-compiled code, it provides valuable capabilities for development and testing.

Use Cases​

1. Testing and Debugging​

The OMP flavor is an excellent testing environment to disambiguate issues between:

  • Input code core processing
  • Flavor source-code generation
tip

If generated code provides accurate values for OMP but not for CUDA, the issue is likely that the code is not parallelization-safe (incorrect work distribution).

2. Sequential Execution​

OMP code, if compiled without any OpenMP library/compiler, provides plain sequential execution of the algorithm. This is useful for:

  • Validating algorithmic correctness
  • Debugging without parallel complexity
  • Running on systems without OpenMP support

3. Optimizing Compiler Benefits​

OMP code outputs plain C++ that can be used as input for an optimizing compiler, which might generate better machine code than what the JIT would provide.

Example: Vector Addition​

[EntryPoint]
public void VectorAdd(int n, float[] a, float[] b, float[] c)
{
    for (int i = threadIdx.x + blockDim.x * blockIdx.x; 
         i < n; 
         i += blockDim.x * gridDim.x)
    {
        c[i] = a[i] + b[i];
    }
}

When compiled to OMP flavor:

// Work distribution maps to OpenMP threads
#pragma omp parallel for
for (int i = 0; i < n; i++)
{
    c[i] = a[i] + b[i];
}

Concept Mapping​

CUDA ConceptOMP Equivalent
GridTotal iteration space
BlockOpenMP thread
ThreadLoop iteration

When to Use OMP​

ScenarioRecommendation
No GPU available✅ Use OMP
Algorithm validation✅ Use OMP first
Debugging race conditions✅ Run sequential OMP
Production with GPU❌ Use CUDA
Maximum CPU performance⚠️ Consider AVX

Build Configuration​

To compile OMP flavor output:

# With OpenMP (parallel)
g++ -fopenmp -O3 -o program program.cpp

# Without OpenMP (sequential)
g++ -O3 -o program program.cpp

Requirements​

  • GCC, Clang, or MSVC with OpenMP support
  • No special hardware required

Next Steps​