set CMAKE_CUDA_ARCHITECTURES

recipe/build.sh

@@ -5,6 +5,16 @@ set -ex
 mkdir build
 cd build
 
+ARCHES=(53 62 72 75 80)
+LATEST_ARCH=86
+for arch in "${ARCHES[@]}"; do
+    CMAKE_CUDA_ARCHS="${CMAKE_CUDA_ARCHS+${CMAKE_CUDA_ARCHS};}${arch}-real"
+done
+# for -real vs. -virtual, see cmake.org/cmake/help/latest/prop_tgt/CUDA_ARCHITECTURES.html
+# this is to support PTX JIT compilation; see first link above or cf.
+# devblogs.nvidia.com/cuda-pro-tip-understand-fat-binaries-jit-caching
+CMAKE_CUDA_ARCHS="${CMAKE_CUDA_ARCHS+${CMAKE_CUDA_ARCHS};}${LATEST_ARCH}"
+
 # workaround for cmake-vs-nvcc: make sure we pick up the our own c-compiler
 ln -s $BUILD_PREFIX/bin/x86_64-conda-linux-gnu-cc $BUILD_PREFIX/bin/gcc
 ln -s $BUILD_PREFIX/bin/x86_64-conda-linux-gnu-c++ $BUILD_PREFIX/bin/c++
@@ -15,6 +25,7 @@ ln -s $BUILD_PREFIX/bin/x86_64-conda-linux-gnu-g++ $BUILD_PREFIX/bin/g++
 cmake \
     ${CMAKE_ARGS} \
     -DCMAKE_BUILD_TYPE=Release \
+    -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHS} \
     -DCMAKE_LIBRARY_OUTPUT_DIRECTORY="lib" \
     -DCUTLASS_ENABLE_F16C=ON \
     -DCUTLASS_ENABLE_CUDNN=ON \