Regards #681: Added a fatbin builder based on NVIDIA's libnvfatbin an…

…d bumped version to 0.8.0; but...

* not tested it yet
* need to add the nvFatbin documentation example program, rewritten to use the APIs

.github/action-scripts/install-cuda-ubuntu.sh

@@ -30,6 +30,7 @@ CUDA_PACKAGES_IN=(
 	"cudart-dev"
 	"nvcc"
 	"profiler-api"
+	"libnvfatbin"
 )
 
 ## -------------------

.github/action-scripts/install-cuda-windows.ps1

@@ -120,12 +120,9 @@ $VISUAL_STUDIO_YEAR = $VISUAL_STUDIO.Substring($VISUAL_STUDIO.Length-4)
 
 $CUDA_PACKAGES = ""
 
-# for CUDA >= 11 cudart is a required package.
-# if([version]$CUDA_VERSION_FULL -ge [version]"11.0") {
-#     if(-not $CUDA_PACKAGES_IN -contains "cudart") {
-#         $CUDA_PACKAGES_IN += 'cudart'
-#     }
-# }
+if([version]$CUDA_VERSION_FULL -ge [version]"12.4") {
+    $CUDA_PACKAGES_IN += "nvfatbin"
+}
 
 Foreach ($package in $CUDA_PACKAGES_IN) {
     # Make sure the correct package name is used for nvcc.

CMakeLists.txt

@@ -7,7 +7,7 @@
 cmake_minimum_required(VERSION 3.25 FATAL_ERROR)
 
 PROJECT(cuda-api-wrappers
-	VERSION 0.7.2
+	VERSION 0.8.0
 	DESCRIPTION "Thin C++-flavored wrappers for the CUDA Runtime API"
 	HOMEPAGE_URL https://github.com/eyalroz/cuda-api-wrappers
 )
@@ -25,8 +25,8 @@ if(libm_exists)
 	set(c_math_lib m)
 endif()
 
-if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL 11.1)
-	foreach(tgt in nvptxcompiler nvptxcompiler_static)
+if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL 12.4)
+	foreach(tgt in nvfatbin nvfatbin_static)
 		if (NOT TARGET ${tgt})
 			_CUDAToolkit_find_and_add_import_lib(${tgt})
 		endif()
@@ -35,7 +35,7 @@ endif()
 set(CMAKE_THREAD_PREFER_PTHREAD TRUE)
 
 
-set(targets runtime-and-driver nvtx rtc)
+set(targets runtime-and-driver nvtx rtc fatbin)
 set(prefixed-targets "")
 set(caw_namespace "cuda-api-wrappers")
 
@@ -79,13 +79,29 @@ add_library("${caw_namespace}::driver-and-runtime" ALIAS caw_runtime-and-driver)
 target_link_libraries(caw_rtc INTERFACE cuda-api-wrappers::runtime-and-driver CUDA::nvrtc)
 if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL 11.1)
 	if (TARGET CUDA::nvptxcompiler)
+		target_link_libraries(caw_rtc INTERFACE CUDA::nvptxcompiler)
 		set(ptx_compiler_target nvptxcompiler)
+	elseif (TARGET CUDA::nvptxcompiler_static)
+		target_link_libraries(caw_rtc INTERFACE CUDA::nvptxcompiler_static)
+	else()
+		message(WARNING "No valid NVIDIA PTX Compiler target is available")
+	endif()
+endif()
+if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL 12.4)
+	if (TARGET CUDA::nvfatbin)
+		target_link_libraries(caw_fatbin INTERFACE CUDA::nvfatbin)
+	elseif (TARGET CUDA::nvfatbin)
+		target_link_libraries(caw_fatbin INTERFACE CUDA::nvfatbin_static)
+	elseif(EXISTS "${CUDA_nvfatbin_LIBRARY}") 
+		target_link_libraries(caw_fatbin INTERFACE "${CUDA_nvfatbin_LIBRARY}")
+	elseif(EXISTS "${CUDA_nvfatbin_static_LIBRARY}") 
+		target_link_libraries(caw_fatbin INTERFACE "${CUDA_nvfatbin_static_LIBRARY}")
 	else()
-		set(ptx_compiler_target nvptxcompiler_static)
+		message(WARNING "Could not locate a valid NVIDIA fatbin creator target or library file")
 	endif()
-	target_link_libraries(caw_rtc INTERFACE CUDA::${ptx_compiler_target})
 endif()
 
+target_link_libraries(caw_fatbin INTERFACE cuda-api-wrappers::runtime-and-driver)
 target_link_libraries(caw_nvtx INTERFACE cuda-api-wrappers::runtime-and-driver)
 
 if(CUDAToolkit_VERSION VERSION_GREATER_EQUAL 10.0)
@@ -99,8 +115,10 @@ endif()
 # but for two settings I won't bother creating one of those
 if(DEFINED CMAKE_USE_PTHREADS_INIT)
 	target_compile_definitions(caw_nvtx INTERFACE "" "CUDA_API_WRAPPERS_USE_PTHREADS")
+	target_compile_definitions(caw_fatbin INTERFACE "" "CUDA_API_WRAPPERS_USE_PTHREADS")
 elseif(DEFINED CMAKE_USE_WIN32_THREADS_INIT)
 	target_compile_definitions(caw_nvtx INTERFACE "" "CUDA_API_WRAPPERS_USE_WIN32_THREADS")
+	target_compile_definitions(caw_fatbin INTERFACE "" "CUDA_API_WRAPPERS_USE_WIN32_THREADS")
 endif()
 
 # --------

README.md

@@ -1,4 +1,4 @@
-# Thin C++-flavored wrappers for the CUDA APIs:<br> Runtime, Driver, NVRTC, PTX compiler and NVTX
+# Thin C++-flavored wrappers for the CUDA APIs:<br> Runtime, Driver, NVRTC and more
 
 <!--Branch Build Status: Master [![Master Build Status](https://api.travis-ci.com/eyalroz/cuda-api-wrappers.svg?branch=master)](https://travis-ci.com/eyalroz/cuda-api-wrappers) | Development: [![Development Build Status](https://api.travis-ci.com/eyalroz/cuda-api-wrappers.svg?branch=development)](https://travis-ci.com/eyalroz/cuda-api-wrappers) -->
 
@@ -16,6 +16,7 @@ This is a header-only library of integrated wrappers around the core parts of NV
 * The slightly higher-level CUDA [Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
 * NVIDIA's dynamic CUDA code compilation library, [NVRTC](http://docs.nvidia.com/cuda/nvrtc/index.html)
 * NVIDIA's out-of-driver, full-featured [PTX compiler library](https://docs.nvidia.com/cuda/ptx-compiler-api/index.html) (available since CUDA 11.1)
+* NVIDIA's fat binary creation library [nvFatbin](https://docs.nvidia.com/cuda/nvfatbin/index.html) (available since CUDA 12.4)
 * The NVIDIA profiler in-program API, also known as [NVTX](https://docs.nvidia.com/cuda/profiler-users-guide/index.html#nvtx) (the NVIDIA Toolkit Extensions library).
 
 It is intended for those who would otherwise use these APIs directly, to make working with them be more intuitive and consistent, making use of modern C++ language capabilities, programming idioms and best practices. In a nutshell - making CUDA API work more fun :-)

examples/CMakeLists.txt

@@ -216,7 +216,7 @@ if( NOT MSVC )
 			endif()
 			set(tgt "cpp_${std_version}")
 			add_executable(${tgt} other/new_cpp_standard/main.cpp)
-			target_link_libraries(${tgt} cuda-api-wrappers::rtc cuda-api-wrappers::nvtx)
+			target_link_libraries(${tgt} cuda-api-wrappers::rtc cuda-api-wrappers::nvtx cuda-api-wrappers::fatbin)
 			if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
 				target_link_libraries(${tgt} stdc++fs)
 			endif()

examples/other/new_cpp_standard/main.cpp

@@ -13,6 +13,7 @@
 #include <cuda/nvtx.hpp>
 #endif
 #include <cuda/rtc.hpp>
+#include <cuda/fatbin.hpp>
 
 #include <cstdlib>
 #include <iostream>
@@ -42,6 +43,9 @@ int main()
 	auto nvrtc_version = cuda::version_numbers::nvrtc();
 	(void) nvrtc_version;
 
+	auto fatbin_version = cuda::version_numbers::fatbin();
+	(void) fatbin_version;
+
 #ifndef _MSC_VER
 	auto nvtx_color_yellow = cuda::profiling::color_t::from_hex(0x0FFFF00);
 	(void) nvtx_color_yellow;

src/cuda/api/types.hpp

@@ -19,6 +19,11 @@
 #ifndef CUDA_API_WRAPPERS_COMMON_TYPES_HPP_
 #define CUDA_API_WRAPPERS_COMMON_TYPES_HPP_
 
+#ifdef _MSC_VER
+// See @url https://stackoverflow.com/q/4913922/1593077
+#define NOMINMAX
+#endif
+
 #if (__cplusplus < 201103L && (!defined(_MSVC_LANG) || _MSVC_LANG < 201103L))
 #error "The CUDA API headers can only be compiled with C++11 or a later version of the C++ language standard"
 #endif

src/cuda/api/versions.hpp

@@ -12,8 +12,14 @@
 
 #include "error.hpp"
 
+#if CUDA_VERSION >= 12040
+#include <nvFatbin.h>
+#endif
+
 #include <ostream>
 #include <utility>
+#include <limits>
+
 
 namespace cuda {
 
@@ -161,6 +167,24 @@ inline version_t runtime() {
 	return version_t::from_single_number(version);
 }
 
+#if CUDA_VERSION >= 12040
+
+inline version_t fatbin() {
+	unsigned int major { 0 }, minor { 0 };
+
+	auto status = nvFatbinVersion(&major, &minor);
+	throw_if_error_lazy(status, "Failed obtaining the nvfatbin library version");
+#ifndef NDEBUG
+	if ((major == 0) or (major > ::std::numeric_limits<int>::max())
+		or (minor == 0) or (minor > ::std::numeric_limits<int>::max())) {
+		throw ::std::logic_error("Invalid version encountered: ("
+			+ ::std::to_string(major) + ", " + ::std::to_string(minor) + ')' );
+	}
+#endif
+	return version_t{ static_cast<int>(major), static_cast<int>(minor) };
+}
+#endif // CUDA_VERSION >= 12040
+
 } // namespace version_numbers
 } // namespace cuda
 

src/cuda/fatbin.hpp

@@ -0,0 +1,20 @@
+/**
+ * @file
+ *
+ * @brief A single file which includes, in turn, the CUDA
+ * fatbin creator library API wrappers and related headers.
+ */
+#pragma once
+#ifndef CUDA_FATBIN_WRAPPERS_HPP_
+#define CUDA_FATBIN_WRAPPERS_HPP_
+
+#if CUDA_VERSION >= 12040
+
+#include "fatbin/types.hpp"
+#include "fatbin/error.hpp"
+#include "fatbin/builder_options.hpp"
+#include "fatbin/builder.hpp"
+
+#endif // CUDA_VERSION >= 12040
+
+#endif // CUDA_FATBIN_WRAPPERS_HPP_