Provide co-resident gcc 8 toolchain for CUDA

README.md

@@ -26,19 +26,14 @@ branch: master
 LAYERSERIES_COMPAT: honister
 
 
-PLEASE NOTE
------------
-
-* CUDA 10.2 supports up through gcc 8 only. Pre-built binaries
-  in the BSP appear to be compatible with gcc 7 and 8 **only**.
-  So use only gcc 7 or gcc 8 if you intend to use CUDA.
-  Recipes for gcc 8 have been imported from the OE-Core warrior branch
-  (the last version of OE-Core to supply gcc 8) to make it easier
-  to use this older toolchain.
-
-  See [this wiki page](https://github.com/OE4T/meta-tegra/wiki/Using-gcc8-from-the-contrib-layer)
-  for information on adding the `meta-tegra/contrib` layer to your
-  builds and configuring them for GCC 8.
+CUDA toolchain compatibility note
+---------------------------------
+
+CUDA 10.2 supports up through gcc 8 only, so recipes are included
+for adding the gcc 8 toolchain to the build for CUDA use, and `cuda.bbclass`
+has been updated to pass the g++ 8 compiler to nvcc for CUDA code compilation.
+This is different from earlier releases/branches, which required setting
+the toolchain used for all builds to a CUDA-compatible version.
 
 
 Getting Help

classes/cuda-gcc.bbclass

@@ -0,0 +1,8 @@
+# CUDA requires gcc/g++ 8, so add that compiler and its runtime as
+# dependencies, and set CC_FOR_CUDA and CXX_FOR_CUDA to point to that compiler.
+DEPENDS:append:cuda = " virtual/${TARGET_PREFIX}cuda-gcc gcc-8-runtime"
+CUDA_HOST_TOOLCHAIN_SUFFIX ??= ""
+CUDA_HOST_TOOLCHAIN_SUFFIX:cuda = "-8.5.0"
+CC_FOR_CUDA ?= "${CCACHE}${HOST_PREFIX}gcc${CUDA_HOST_TOOLCHAIN_SUFFIX} ${HOST_CC_ARCH}${TOOLCHAIN_OPTIONS}"
+CXX_FOR_CUDA ?= "${CCACHE}${HOST_PREFIX}g++${CUDA_HOST_TOOLCHAIN_SUFFIX} ${HOST_CC_ARCH}${TOOLCHAIN_OPTIONS}"
+PACKAGE_ARCH:cuda = "${TEGRA_PKGARCH}"

classes/cuda.bbclass

@@ -1,3 +1,5 @@
+inherit cuda-gcc
+CUDA_ARCHITECTURES ??= ""
 CUDA_NVCC_COMPAT_FLAGS ??= ""
 CUDA_NVCC_PATH_FLAGS ??= "--include-path ${STAGING_DIR_HOST}/usr/local/cuda-${CUDA_VERSION}/include --library-path ${STAGING_DIR_HOST}/usr/local/cuda-${CUDA_VERSION}/${baselib}"
 CUDA_NVCC_EXTRA_FLAGS ??= ""
@@ -19,12 +21,17 @@ def cuda_extract_compiler(compiler, d, prefix='-Xcompiler '):
         return args[1], ' '.join([prefix + arg for arg in args[2:]])
     return args[0], ' '.join([prefix + arg for arg in args[1:]])
 
-export CUDAHOSTCXX = "${@cuda_extract_compiler('CXX', d)[0]}"
-export CUDAFLAGS = "${CUDA_NVCC_FLAGS} ${@cuda_extract_compiler('CXX', d)[1]}"
+export CUDAHOSTCXX = "${@cuda_extract_compiler('CXX_FOR_CUDA', d)[0]}"
+export CUDAFLAGS = "${CUDA_NVCC_FLAGS} ${@cuda_extract_compiler('CXX_FOR_CUDA', d)[1]}"
 OECMAKE_CUDA_COMPILER_LAUNCHER ?= "${CCACHE}"
 OECMAKE_CUDA_COMPILER ?= "nvcc"
 CUDA_CCACHE_COMPILERCHECK ?= "cuda-compiler-check %compiler%"
 
+def cmake_cuda_archs(d):
+    archs = d.getVar('CUDA_ARCHITECTURES')
+    return archs if archs else "OFF"
+OECMAKE_CUDA_ARCHITECTURES ?= "${@cmake_cuda_archs(d)}"
+
 # meson uses 'CUFLAGS' for flags to pass to nvcc
 # and requires all nvcc, compiler, and linker flags to be
 # bundled into that one environment variable.
@@ -41,7 +48,7 @@ def cuda_meson_ldflags(d):
         elif arg.startswith('--sysroot='):
             linkargs.append(arg)
     return '-Xlinker ' + ','.join(linkargs)
-CUFLAGS = "-ccbin ${@cuda_extract_compiler('CXX', d)[0]} ${CUDAFLAGS} ${@cuda_extract_compiler('CXX', d)[1]} ${@cuda_meson_ldflags(d)}"
+CUFLAGS = "-ccbin ${@cuda_extract_compiler('CXX_FOR_CUDA', d)[0]} ${CUDAFLAGS} ${@cuda_extract_compiler('CXX_FOR_CUDA', d)[1]} ${@cuda_meson_ldflags(d)}"
 
 # The following are for the old-style FindCUDA.cmake module (pre-3.8)
 CUDA_EXTRA_OECMAKE = '\
@@ -72,6 +79,7 @@ set(CMAKE_CUDA_TOOLKIT_TARGET_DIR "${STAGING_DIR_HOST}/usr/local/cuda-${CUDA_VER
 set(CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES "\${CMAKE_CUDA_TOOLKIT_ROOT_DIR}/include" "\${CMAKE_CUDA_TOOLKIT_TARGET_DIR}/include" CACHE PATH "" FORCE)
 set(CMAKE_CUDA_COMPILER ${OECMAKE_CUDA_COMPILER})
 set(CMAKE_CUDA_COMPILER_LAUNCHER ${OECMAKE_CUDA_COMPILER_LAUNCHER})
+set(CMAKE_CUDA_ARCHITECTURES ${OECMAKE_CUDA_ARCHITECTURES})
 EOF
 }
 

conf/layer.conf

@@ -55,4 +55,5 @@ SIGGEN_EXCLUDE_SAFE_RECIPE_DEPS += "\
   libdrm->libdrm-nvdc \
   libdrm-nvdc->tegra-libraries \
   tegra-nvs-service->tegra-nvs-base \
+  gcc-8-cross-${TARGET_ARCH}->linux-libc-headers \
 "

conf/machine/include/tegra-common.inc

@@ -31,6 +31,7 @@ PREFERRED_PROVIDER_virtual/secure-os ?= "tos-prebuilt"
 PREFERRED_PROVIDER_virtual/egl = "libglvnd"
 PREFERRED_PROVIDER_virtual/libgles1 = "libglvnd"
 PREFERRED_PROVIDER_virtual/libgles2 = "libglvnd"
+PREFERRED_PROVIDER_virtual/libgles3 = "libglvnd"
 PREFERRED_PROVIDER_virtual/libgl = "libglvnd"
 PREFERRED_PROVIDER_libv4l = "${@'v4l-utils' if 'openembedded-layer' in d.getVar('BBFILE_COLLECTIONS').split() else 'libv4l2-minimal'}"
 PREFERRED_PROVIDER_v4l-utils = "${@'v4l-utils' if 'openembedded-layer' in d.getVar('BBFILE_COLLECTIONS').split() else 'libv4l2-minimal'}"
@@ -55,6 +56,8 @@ PREFERRED_PROVIDER_virtual/xserver-xf86 = "xserver-xorg-tegra"
 XSERVER = "xserver-xorg-tegra xf86-input-evdev xserver-xorg-video-nvidia xserver-xorg-tegra-module-libwfb"
 
 CUDA_VERSION ?= "10.2"
+CUDA_ARCHITECTURES ?= "${TEGRA_CUDA_ARCHITECTURE}"
+CUDA_NVCC_ARCH_FLAGS ?= "--gpu-architecture=compute_${TEGRA_CUDA_ARCHITECTURE} --gpu-code=sm_${TEGRA_CUDA_ARCHITECTURE}"
 # Tegra specific version of libnvidia-container
 PREFERRED_VERSION_libnvidia-container-tools = "0.10.0"
 

conf/machine/include/tegra186.inc

@@ -23,7 +23,7 @@ NVIDIA_CHIP = "0x18"
 
 TEGRA_SIGNING_ENV ?= "BOARDID=${TEGRA_BOARDID} FAB=${TEGRA_FAB}"
 
-CUDA_NVCC_ARCH_FLAGS ?= "--gpu-architecture=compute_62 --gpu-code=sm_62"
+TEGRA_CUDA_ARCHITECTURE ?= "62"
 
 PREFERRED_PROVIDER_tos-trusty ?= "tos-prebuilt"
 PREFERRED_PROVIDER_cboot ?= "cboot-t18x"

conf/machine/include/tegra194.inc

@@ -33,7 +33,7 @@ TEGRA_BUPGEN_SPECS ?= "fab=400;boardsku=0001;boardrev=H.0 \
 
 TEGRA_SIGNING_ENV = "CHIPREV=${TEGRA_CHIPREV} BOARDID=${TEGRA_BOARDID} FAB=${TEGRA_FAB} BOARDSKU=${TEGRA_BOARDSKU} BOARDREV=${TEGRA_BOARDREV}"
 
-CUDA_NVCC_ARCH_FLAGS ?= "--gpu-architecture=compute_72 --gpu-code=sm_72"
+TEGRA_CUDA_ARCHITECTURE = "72"
 
 PREFERRED_PROVIDER_virtual/bootloader ?= "cboot-t19x"
 PREFERRED_PROVIDER_tos-trusty ?= "tos-prebuilt"

conf/machine/include/tegra210.inc

@@ -24,7 +24,7 @@ NVIDIA_CHIP = "0x21"
 
 TEGRA_SIGNING_ENV ?= "BOARDID=${TEGRA_BOARDID} FAB=${TEGRA_FAB} BOARDSKU=${TEGRA_BOARDSKU} BOARDREV=${TEGRA_BOARDREV}"
 
-CUDA_NVCC_ARCH_FLAGS ?= "--gpu-architecture=compute_53 --gpu-code=sm_53"
+TEGRA_CUDA_ARCHITECTURE ?= "53"
 
 require conf/machine/include/tegra-common.inc
 

contrib/README.md

@@ -4,14 +4,6 @@ Contributed recipes for meta-tegra
 This layer contains recipes to support builds with
 meta-tegra that are not directly part of the BSP.
 
-## gcc 8 toolchain
-The main use of this layer is to provide an older
-gcc toolchain that is compatible with the version
-of the CUDA toolkit for the Jetson platforms.
-
-For more information, see
-[this wiki page](https://github.com/OE4T/meta-tegra/wiki/Using-gcc8-from-the-contrib-layer).
-
 ## gstreamer 1.14 recipes
 This layer also includes a port of the gstreamer 1.14
 recipes from the OE-Core warrior branch. When using the