Generated Code Layout

When Hybridizer processes your project, it generates several files. Understanding this layout helps with debugging and customization.

Output Directory Structure

bin/Release/
├── YourProject.dll              ← Your .NET assembly
├── YourProject_CUDA.cu          ← Generated CUDA source
├── YourProject_CUDA.dll         ← Compiled GPU binary
├── YourProject_CUDA.config.xml  ← Hybridizer configuration
├── YourProject_CUDA.ptx         ← PTX intermediate (if kept)
└── ...

The Generated .cu File

This is the most informative file. It contains:

1. Struct Definitions

Your C# structs are translated to C++ structs:

// From: public struct MyData { public float x; public int count; }
struct MyData {
    float x;
    int count;
};

2. Device Functions

Methods marked [Kernel]:

// From: [Kernel] public static float Compute(float x) { ... }
__device__ float Compute(float x) {
    // ...
}

3. Global Kernels

Methods marked [EntryPoint]:

// From: [EntryPoint] public static void Run(float[] a, int N) { ... }
__global__ void Run(float* a, int N) {
    // Grid-stride loop generated from Parallel.For or manual loop
}

4. Wrapper Functions

Host-callable functions that handle marshalling:

extern "C" __declspec(dllexport)
void Run_wrapper(float* a, int N) {
    Run<<<gridDim, blockDim, sharedMem, stream>>>(a, N);
}

The Configuration File

YourProject_CUDA.config.xml contains:

• Entry point list
• Type mappings
• Generic specializations ([HybridRegisterTemplate])
• Intrinsic mappings

Each generated source file includes a sub-configuration comment for reproducibility.

Customizing the Build

MSBuild Properties

Add to your .csproj:

<PropertyGroup>
  <!-- Emit C# line info in generated CUDA -->
  <HybridizerEmitLineInfo>true</HybridizerEmitLineInfo>

  <!-- Target specific GPU architecture -->
  <HybridizerCudaArch>sm_86</HybridizerCudaArch>

  <!-- Keep intermediate PTX -->
  <HybridizerKeepPTX>true</HybridizerKeepPTX>

  <!-- Additional nvcc flags -->
  <HybridizerNvccFlags>--use_fast_math</HybridizerNvccFlags>
</PropertyGroup>

Common Architecture Codes

GPU Series	Compute Capability	sm_ Value
GTX 10xx	6.1	sm_61
RTX 20xx	7.5	sm_75
RTX 30xx	8.6	sm_86
RTX 40xx	8.9	sm_89
H100	9.0	sm_90

Reading Generated Code for Debugging

When a kernel gives wrong results:

1. Open the .cu file
2. Find your entry point function
3. Check:
   • Are loop bounds correct?
   • Are threadIdx/blockIdx used properly?
   • Are shared memory declarations present?
   • Are [In]/[Out] reflected in the wrapper?

tip

The generated code is valid, readable CUDA C++. If you know CUDA, reading it is the fastest way to understand what Hybridizer does with your code.