[FINUFFT] Build CUDA library v2.5.0

C/cufinufft/build_tarballs.jl

@@ -8,12 +8,12 @@ include(joinpath(YGGDRASIL_DIR, "fancy_toys.jl"))
 include(joinpath(YGGDRASIL_DIR, "platforms", "cuda.jl"))
 
 # Build script for the CUDA part of FINUFFT
-
 # Builds for all compatible CUDA platforms, but without microarchitecture expansion (not
 # needed for CUDA cuda, and would produce a giant amount of artifacts)
+
 name = "cufinufft"
-version = v"2.4.1"
-commit_hash = "e7144a5c08cbaf3e3b344a4fdd92bc3c7e468ff2" # v2.4.1
+version = v"2.5.0"
+commit_hash = "8b7ac66e617ea317cd42366ac27c536793131696"
 preferred_gcc_version=v"11"
 
 # Collection of sources required to complete build
@@ -24,8 +24,10 @@ sources = [
 # Build script: cufinufft, all possible archs available for each CUDA version
 # - CMake toolchain looks for compiler in CUDA_PATH/bin/nvcc
 #   and libs in CUDA_PATH/lib64, so create link.
+# - Need higher version of CMake, so remove existing
 script = raw"""
 cd $WORKSPACE/srcdir/finufft*/
+apk del cmake
 
 export CUDA_PATH="$prefix/cuda"
 ln -s $prefix/cuda/lib $prefix/cuda/lib64
@@ -47,25 +49,22 @@ cmake --install .
 unlink $prefix/cuda/lib64
 """
 
-# Build for all supported CUDA > v11
-platforms = expand_cxxstring_abis(CUDA.supported_platforms(min_version=v"11.0"))
+# Build for all supported CUDA >= v11.8. Only one v11.x should be sufficient.
+platforms = expand_cxxstring_abis(CUDA.supported_platforms(min_version=v"11.8"))
+
 # Cmake toolchain breaks on aarch64, so only x86_64 for now
 filter!(p -> arch(p)=="x86_64", platforms)
 
-# cuFINUFFT does not compile with CUDA 12.5, so exclude
-filter!(p -> VersionNumber(p["cuda"]) != v"12.5", platforms)
-
-# CUDA 13 doesn't seem to build (yet)
-filter!(p -> VersionNumber(p["cuda"]) < v"13", platforms)
-
 # The products that we will ensure are always built
 products = [
     LibraryProduct("libcufinufft", :libcufinufft)
 ]
 
 # Dependencies that must be installed before this package can be built
 # NVTX_jll is needed for nvToolsExt. (tested with v3.1.0+2)
-dependencies = [Dependency("NVTX_jll")]
+# CMake needs higher version than what is bundled
+dependencies = [Dependency("NVTX_jll"),
+                HostBuildDependency(PackageSpec(; name="CMake_jll", version = v"3.24.3+0"))]
 
 for platform in platforms
     should_build_platform(triplet(platform)) || continue
@@ -75,14 +74,18 @@ for platform in platforms
 
     # Build for all major archs supported by SDK
     # See https://en.wikipedia.org/wiki/CUDA
-    # sm_90 works for CUDA v12.1 and up, due to use of atomic operaitons
-    # sm_52 required for alloca from CUDA v12.0 and up
-    if VersionNumber(platform["cuda"]) < v"12.0"
-        cuda_archs = "50;60;70;80"
-    elseif VersionNumber(platform["cuda"]) < v"12.1"
-        cuda_archs = "60;70;80"
-    else
+    cuda_ver = VersionNumber(platform["cuda"])
+    if cuda_ver >= v"13" # 13.0+
+        # Skip 110,121 to not run out of space
+        cuda_archs = "75;80;90;100;120"
+    elseif cuda_ver >= v"12.1" # 12.1-12.9
+        # sm_90 works for CUDA v12.1 and up, due to use of atomic operations
         cuda_archs = "60;70;80;90"
+    elseif cuda_ver >= v"12.0" # 12.0
+        # sm_52 required for alloca from CUDA v12.0 and up
+        cuda_archs = "60;70;80"
+    else # < 12.0
+        cuda_archs = "50;60;70;80"
     end
     arch_line = "export CUDA_ARCHS=\"$cuda_archs\"\n"
     platform_script = arch_line * script