Skip to content

cpp_cuda_build.md

Latest commit

 

History

History
78 lines (52 loc) · 2.4 KB

cpp_cuda_build.md

File metadata and controls

78 lines (52 loc) · 2.4 KB

C++ and CUDA Build Guide

C++ Library Build

The default build is a shared C++20 library in RelWithDebInfo.

./build_lib.sh
./build_lib.sh -t release -i
./build_lib.sh -D BUILD_SHARED_LIBS=OFF

Use NO_OPTIMIZATION=ON for profiler-friendly debug behavior across build types. It forces -O0, keeps assertions, preserves frame pointers, and disables inlining/sibling-call optimization where the compiler supports it.

./build_lib.sh -D NO_OPTIMIZATION=ON

Toolchains and Cross Builds

Cross builds use CMake toolchain files under cmake/toolchains/defaults/. Native CPU tuning is disabled automatically while cross-compiling so build artifacts stay portable across runner CPUs.

./build_lib.sh --toolchain cmake/toolchains/defaults/aarch64-linux-gnu.cmake --clean \
  -D template_project_BUILD_PROGRAMS=OFF \
  -D template_project_BUILD_EXAMPLES=OFF

Use CPU_EXTRA_OPT_FLAGS for target-specific flags that are safe for the destination CPU.

CUDA

CUDA is optional and enabled explicitly:

./build_lib.sh -D ENABLE_CUDA=ON

CUDA source files live under src/template_src_kernels/.

  • *.cu: regular CUDA translation units
  • *.ptx.cu: PTX inputs compiled into embedded C arrays for OptiX modules

GPU architecture is detected from nvidia-smi on x86_64. On Jetson/Tegra aarch64 systems, the template falls back to device markers for Xavier, Orin, and Thor. If detection is unavailable or ambiguous, configure with CUDA_ARCHITECTURES or CMAKE_CUDA_ARCHITECTURES.

./build_lib.sh -D ENABLE_CUDA=ON \
  -D CUDA_ARCHITECTURES=87 \
  -D CUDA_ENABLE_FMAD=ON \
  -D CUDA_USE_FAST_MATH=OFF \
  -D CUDA_PTX_USE_FAST_MATH=ON

OptiX

ENABLE_OPTIX=ON enables CUDA and requires at least one compiled library source plus one *.ptx.cu source. Header-only OptiX configurations fail at configure time because there is no library artifact that can own the generated PTX integration.

./build_lib.sh -D ENABLE_OPTIX=ON -D CUDA_ARCHITECTURES=87

Set OPTIX_HOME or use the system OptiX SDK layout expected by cmake/HandleOptiX.cmake.

Optional Runtime Libraries

Enable oneTBB when parallel CPU code needs it:

./build_lib.sh -D ENABLE_TBB=ON

Enable profiling only for diagnostic builds:

./build_lib.sh --profile

ENABLE_TCMALLOC stays OFF by default because MATLAB MEX and plugin-style consumers are sensitive to allocator dependencies.