Files
ANSCORE/CMakeLists.txt
T
nghianguyenandClaude Opus 5 e76d3b81e2 fix(cuda): spell the fast-math flag the way nvcc spells it
Every .cu file failed on CUDA 13 with "nvcc fatal: Unknown option
'--use-fast-math'". The flag is --use_fast_math, with underscores; older
toolkits accepted the hyphenated form as well and this one does not, so
the engine layer could not compile a single kernel on a current CUDA.

Found on the Windows machine with CUDA 13.0, where it stopped the build
at the first two kernels.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-09-13 07:58:49 +10:00

210 lines
9.9 KiB
CMake
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
cmake_minimum_required(VERSION 3.28)
# CMP0167 (FindBoost removed from CMake) only exists from 3.30, and this
# project asks for 3.28 — so setting it unconditionally makes every older
# CMake fail at line two with "policy is not known to this version". Guarded,
# a newer CMake still gets the quiet FindBoost it was added for.
if(POLICY CMP0167)
cmake_policy(SET CMP0167 NEW)
endif()
# ── CUDA is optional (not available on macOS) ───────────────────
include(CheckLanguage)
check_language(CUDA)
if(CMAKE_CUDA_COMPILER)
project(ANSCORE LANGUAGES C CXX CUDA)
set(ANSCORE_HAS_CUDA ON)
set(CMAKE_CUDA_STANDARD 20)
set(CMAKE_CUDA_STANDARD_REQUIRED ON)
# SASS (-real) for each shipping arch from Turing (sm_75) through Blackwell
# consumer (sm_120), plus compute_75 PTX (-virtual) so any newer/unlisted GPU
# JIT-compiles at first run (forward compatibility). NOTE: this only governs
# this project's own .cu files; prebuilt deps (odhublib, TensorRT, ORT, cuBLAS)
# carry their own arch coverage independently.
set(CMAKE_CUDA_ARCHITECTURES
75-real 80-real 86-real 87-real 88-real 89-real 90-real 100-real 120-real
75-virtual)
# --use_fast_math, with underscores: that is nvcc's own spelling. Older
# toolkits also accepted the hyphenated form; CUDA 13 does not, and every
# .cu file fails with "Unknown option '--use-fast-math'" before it compiles
# a line.
if(MSVC)
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --use_fast_math -Xcompiler=/utf-8")
else()
set(CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS} --use_fast_math")
endif()
else()
project(ANSCORE LANGUAGES C CXX)
set(ANSCORE_HAS_CUDA OFF)
message(WARNING "CUDA not found — GPU-accelerated modules will be limited")
endif()
# ── Platform detection ──────────────────────────────────────────
if(WIN32)
set(ANSCORE_PLATFORM "windows")
elseif(APPLE)
set(ANSCORE_PLATFORM "macos")
else()
set(ANSCORE_PLATFORM "linux")
endif()
message(STATUS "ANSCORE platform: ${ANSCORE_PLATFORM}, CUDA: ${ANSCORE_HAS_CUDA}")
# ── Global Settings ──────────────────────────────────────────────
set(CMAKE_CXX_STANDARD 20)
set(CMAKE_CXX_STANDARD_REQUIRED ON)
set(CMAKE_POSITION_INDEPENDENT_CODE ON)
# ANSCORE uses no C++20 modules. CMake auto-enables module dependency scanning
# for C++20 + Ninja, which adds a fragile per-TU scan/dyndep (.ddi/.modmap) step
# that can fail (empty .ddi -> "cannot open ...modmap" / dyndep parse errors),
# especially after an interrupted build. Disable it: faster builds, no dyndep.
set(CMAKE_CXX_SCAN_FOR_MODULES OFF)
# Output all binaries to a single directory
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
# ── Compiler Configuration ──────────────────────────────────────
if(MSVC)
# Dynamic CRT (/MD) — matches original VS projects and prebuilt libs like HWiNFO
set(CMAKE_MSVC_RUNTIME_LIBRARY "MultiThreaded$<$<CONFIG:Debug>:Debug>DLL")
# Apply MSVC flags only to C/C++ (not CUDA — nvcc doesn't understand them)
add_compile_options(
"$<$<COMPILE_LANGUAGE:C,CXX>:/W3>"
"$<$<COMPILE_LANGUAGE:C,CXX>:/utf-8>"
"$<$<AND:$<COMPILE_LANGUAGE:C,CXX>,$<NOT:$<CONFIG:MINSIZEREL>>>:/Zi>"
# RelWithDebInfo: keep /O2 but disable inlining so debuggers can land
# breakpoints on small dispatch functions (e.g. avframeToCVMat).
"$<$<AND:$<COMPILE_LANGUAGE:C,CXX>,$<CONFIG:RelWithDebInfo>>:/Ob0>"
)
add_link_options(
"$<$<NOT:$<CONFIG:DEBUG>>:/DEBUG:FULL>"
# /OPT:REF and /OPT:ICF improve Release size/perf but confuse the
# debugger (folds identical functions, strips unused ones). Apply
# them to Release only, not RelWithDebInfo.
"$<$<CONFIG:Release>:/OPT:REF>"
"$<$<CONFIG:Release>:/OPT:ICF>"
"$<$<CONFIG:RelWithDebInfo>:/OPT:NOREF>"
"$<$<CONFIG:RelWithDebInfo>:/OPT:NOICF>"
)
add_compile_definitions(_CRT_SECURE_NO_WARNINGS _WINSOCK_DEPRECATED_NO_WARNINGS)
elseif(CMAKE_CXX_COMPILER_ID MATCHES "Clang|GNU")
add_compile_options(-Wall -Wextra -Wno-unused-parameter)
if(NOT APPLE)
add_link_options(-Wl,--as-needed)
endif()
endif()
# ── DebugView logging toggle (runtime-gated) ────────────────────
# ANS_DBG is compiled into every build by default and gated at runtime by
# ANSCENTER_IsDebugViewEnabled() (see ANSLicense.h/.cpp). Users toggle
# logging without a rebuild by creating or deleting
# C:\ProgramData\ANSCENTER\ansvisdebugview.txt (Windows), or
# /tmp/ansvisdebugview.txt (POSIX),
# or by setting env var ANSCENTER_DBGVIEW=1/0. Effect propagates within ~2 s.
#
# This CMake option is a HARD KILL-SWITCH only — set it ON to strip ANS_DBG
# to ((void)0) at compile time. Use it only when even the cached atomic-load
# cost is unacceptable (tight inner loops). Default OFF = runtime-gated.
#
# Usage:
# cmake -B build # default: runtime-gated
# cmake -B build -DANSCORE_DEBUGVIEW_STRIP=ON # hard-strip all ANS_DBG
option(ANSCORE_DEBUGVIEW_STRIP "Hard-strip ANS_DBG calls at compile time (removes runtime toggle)" OFF)
if(ANSCORE_DEBUGVIEW_STRIP)
add_compile_definitions(ANSCORE_DEBUGVIEW=0)
message(STATUS "ANSCORE_DEBUGVIEW_STRIP = ON — ANS_DBG compiled OUT (no runtime toggle)")
else()
message(STATUS "ANSCORE_DEBUGVIEW_STRIP = OFF — ANS_DBG compiled IN, gated at runtime (sentinel file / env var)")
endif()
# ── Vendored libyuv (submodule: 3rdparty/libyuv) ────────────────
# SIMD-optimized YUV conversion library. Only genuinely fast on toolchains
# that understand GCC inline assembly (Clang/clang-cl/GCC) — 144 of its
# SIMD row functions live in row_gcc.cc. MSVC (cl.exe) can only compile the
# 8 routines in row_win.cc, so on MSVC libyuv silently falls back to the
# scalar C code in row_common.cc and runs ~10× slower than OpenCV+IPP's
# cv::cvtColor(COLOR_YUV2BGR_I420). Do NOT enable on MSVC builds.
#
# Default: OFF. Enable with `-DANSCORE_USE_LIBYUV=ON` only after switching
# the project's compiler to clang-cl / clang / gcc.
option(ANSCORE_USE_LIBYUV "Use libyuv for YUV→BGR conversion (only effective on Clang/GCC, NOT MSVC)" OFF)
# Prebuilt libyuv path — built separately with clang-cl so its SIMD rows compile,
# then linked into the MSVC main build (clang-cl output is MSVC-ABI-compatible).
# Override with -DANSCORE_LIBYUV_LIB=<path> if the .lib lives elsewhere.
set(ANSCORE_LIBYUV_LIB "${CMAKE_SOURCE_DIR}/3rdparty/libyuv/build-clangcl/Release/yuv.lib"
CACHE FILEPATH "Path to the prebuilt libyuv static library (built with clang-cl)")
if(ANSCORE_USE_LIBYUV AND EXISTS "${ANSCORE_LIBYUV_LIB}")
add_library(yuv STATIC IMPORTED GLOBAL)
set_target_properties(yuv PROPERTIES
IMPORTED_LOCATION "${ANSCORE_LIBYUV_LIB}"
INTERFACE_INCLUDE_DIRECTORIES "${CMAKE_SOURCE_DIR}/3rdparty/libyuv/include")
set(ANSCORE_HAS_LIBYUV ON)
message(STATUS "libyuv: ENABLED (prebuilt static lib at ${ANSCORE_LIBYUV_LIB})")
elseif(ANSCORE_USE_LIBYUV)
set(ANSCORE_HAS_LIBYUV OFF)
message(WARNING "ANSCORE_USE_LIBYUV=ON but prebuilt libyuv not found at "
"${ANSCORE_LIBYUV_LIB}. Build it first with clang-cl:\n"
" cmake -S 3rdparty/libyuv -B 3rdparty/libyuv/build-clangcl "
"-G \"Visual Studio 17 2022\" -T ClangCL -A x64\n"
" cmake --build 3rdparty/libyuv/build-clangcl --config Release\n"
"Falling back to cv::cvtColor path.")
else()
set(ANSCORE_HAS_LIBYUV OFF)
message(STATUS "libyuv: DISABLED — using cv::cvtColor+IPP path (fast on MSVC)")
endif()
# ── External Dependencies ───────────────────────────────────────
include(cmake/Dependencies.cmake)
# ── Subprojects (ordered by dependency) ─────────────────────────
# Core — licensing & infrastructure
add_subdirectory(core/anslicensing)
add_subdirectory(core/ANSLicensingSystem)
# ANSLibsLoader dynamically binds cudart/cuBLAS/nvJPEG at runtime, but it still
# compiles against <cuda_runtime.h> for their types and links CUDA::cudart_static.
if(ANSCORE_HAS_CUDA)
add_subdirectory(core/ANSLibsLoader)
endif()
add_subdirectory(core/ANSModelPackage)
# Modules — feature DLLs (cross-platform)
add_subdirectory(modules/ANSMOT)
add_subdirectory(modules/ANSUtilities)
add_subdirectory(modules/ANSTrainingEngine)
add_subdirectory(modules/ANSLLM)
# Integrations — hardware & protocol DLLs
if(WIN32)
add_subdirectory(integrations/ANNHUB)
add_subdirectory(integrations/ANSIO)
add_subdirectory(integrations/ANSONVIF)
endif()
add_subdirectory(integrations/ANSPulsar)
add_subdirectory(integrations/ANSRabbitMQ)
# Modules — engine layer (CUDA + inference frameworks)
if(ANSCORE_HAS_CUDA)
add_subdirectory(modules/ANSODEngine)
add_subdirectory(modules/ANSOCR)
add_subdirectory(modules/ANSFR)
add_subdirectory(modules/ANSLPR)
# ANSCV hosts NVDEC decode, the NV12 GPU frame pool and the nvJPEG encoder:
# ANSGpuFrameOps.h includes <cuda_runtime.h> unconditionally and calls
# cudaMemcpy2D/cudaMallocPitch directly, so it needs the toolkit like the
# engine modules do.
add_subdirectory(modules/ANSCV)
endif()
# Facade
add_subdirectory(ANSLIB)
# Tests
# enable_testing() (not include(CTest), which would add Nightly/Experimental
# targets) so the add_test() calls under tests/ reach ctest.
enable_testing()
add_subdirectory(tests)