Skip to content

cuda-mock.md

Latest commit

 

History

History
82 lines (65 loc) · 4.19 KB

cuda-mock.md

File metadata and controls

82 lines (65 loc) · 4.19 KB
Raw

Mock CUDA Library

The mock CUDA library provides a minimal implementation of CUDA Driver and Runtime APIs for container validation workloads. It is built alongside the mock NVML library and deployed via the same DaemonSet.

Status: Early stage -- 15 functions implemented. Sufficient for basic validation workloads (e.g., cuda-sample:vectoradd) but not for complex CUDA applications.

Implemented Functions

Category Functions Notes
Initialization cuInit, cudaDriverGetVersion, cudaRuntimeGetVersion Returns configured driver/runtime versions
Device Management cudaGetDeviceCount, cudaSetDevice, cudaGetDevice, cudaDeviceReset Tracks active device index
Memory cudaMalloc, cudaFree, cudaMemcpy Real host allocation via malloc; device copies are no-ops
Execution cudaLaunchKernel, cudaDeviceSynchronize No-ops (no actual computation)
Error Handling cudaGetErrorString, cudaGetLastError, cudaPeekAtLastError Thread-local error tracking

How It Works

The library is built as libcuda.so.<version> alongside mock NVML. The DaemonSet setup script creates the following symlink chain:

libcuda.so.550.163.01    (versioned library)
libcuda.so.1        ->   libcuda.so.550.163.01
libcuda.so          ->   libcuda.so.1
libcudart.so.12     ->   libcuda.so.1          (runtime API compat)
libcudart.so        ->   libcudart.so.12

The runtime API symlinks (libcudart.so*) point to the same mock library. This is a compatibility workaround -- applications that link against libcudart.so (such as cuda-sample:vectoradd) resolve to the mock without needing a separate runtime library.

Architecture

┌──────────────────────────────────┐
│   CGo Bridge (bridge/cuda.go)    │
│   15 C function exports          │
│   C <-> Go type conversion       │
└──────────────┬───────────────────┘
               │
               v
┌──────────────────────────────────┐
│   Engine (engine/cuda.go)        │
│   Singleton lifecycle            │
│   Device index tracking          │
│   Memory allocation tracking     │
└──────────────────────────────────┘

The CGo bridge layer exports C-ABI functions that the dynamic linker resolves when applications call CUDA APIs. Each exported function delegates to the Go engine, which manages state (active device index, allocated memory pointers, error codes).

Limitations

  • No actual computation -- kernel launches are no-ops.
  • No multi-GPU -- device selection is tracked but has no effect.
  • No streams/events -- async operations are not implemented.
  • Memory copies -- only host-to-host actually copies data; other directions are no-ops.
  • No CUDA context management -- cuCtxCreate, cuCtxDestroy, etc. are not implemented.

Building

cd pkg/gpu/mockcuda
make

Source Code

File Description
pkg/gpu/mockcuda/bridge/cuda.go CGo bridge -- 15 exported C functions
pkg/gpu/mockcuda/bridge/cuda_types.h C type definitions
pkg/gpu/mockcuda/engine/cuda.go Go engine -- device and memory management
pkg/gpu/mockcuda/engine/types.go Error codes and type definitions
pkg/gpu/mockcuda/Makefile Build configuration