BUILD: Add configurable CUDA architecture targeting with meson built-in options

README.md

@@ -114,6 +114,11 @@ $ meson setup <name_of_build_dir> \
     -Ducx_path=/path/to/ucx \     # Custom UCX installation path
     -Dinstall_headers=true \      # Install development headers
     -Ddisable_gds_backend=false   # Enable GDS backend
+
+# CUDA GPU architecture targeting (example)
+$ meson setup <name_of_build_dir> \
+    -Dcuda_args="-gencode=arch=compute_80,code=sm_80 -gencode=arch=compute_90,code=sm_90" \
+    -Dcuda_link_args="-gencode=arch=compute_80,code=sm_80 -gencode=arch=compute_90,code=sm_90"
 ```
 
 Common build options:
@@ -123,6 +128,32 @@ Common build options:
 - `disable_gds_backend`: Disable GDS backend (default: false)
 - `cudapath_inc`, `cudapath_lib`: Custom CUDA paths
 - `static_plugins`: Comma-separated list of plugins to build statically
+- `cuda_args`: CUDA compiler arguments (default: targets compute_80 and compute_90)
+- `cuda_link_args`: CUDA linker arguments (default: targets compute_80 and compute_90)
+
+### CUDA GPU Architecture Configuration
+
+NIXL automatically targets CUDA compute capabilities 8.0 (Ampere) and 9.0 (Hopper) by default.
+To target specific GPU architectures or add additional ones, use the `cuda_args` and `cuda_link_args` options:
+
+```bash
+# Target specific architectures
+$ meson setup build \
+    -Dcuda_args="-gencode=arch=compute_75,code=sm_75" \
+    -Dcuda_link_args="-gencode=arch=compute_75,code=sm_75"
+
+# Target multiple architectures
+$ meson setup build \
+    -Dcuda_args="-gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_80,code=sm_80 -gencode=arch=compute_90,code=sm_90" \
+    -Dcuda_link_args="-gencode=arch=compute_75,code=sm_75 -gencode=arch=compute_80,code=sm_80 -gencode=arch=compute_90,code=sm_90"
+```
+
+Common CUDA compute capabilities:
+- `compute_75,code=sm_75`: NVIDIA Turing (RTX 20xx, Tesla T4)
+- `compute_80,code=sm_80`: NVIDIA Ampere (A100, RTX 30xx)
+- `compute_86,code=sm_86`: NVIDIA Ampere (RTX 30xx consumer)
+- `compute_89,code=sm_89`: NVIDIA Ada Lovelace (RTX 40xx)
+- `compute_90,code=sm_90`: NVIDIA Hopper (H100)
 
 ### Building Documentation
 

meson.build

@@ -82,24 +82,37 @@ else
 endif
 
 if cuda_dep.found()
-    add_languages('CUDA')
+    add_languages('cuda')
     cuda = import('unstable-cuda')
     nvcc = meson.get_compiler('cuda')
 
-    # Please extend with your arch if not present in the list
-    nvcc_flags = []
-    nvcc_flags += ['-gencode', 'arch=compute_80,code=sm_80']
-    nvcc_flags += ['-gencode', 'arch=compute_90,code=sm_90']
-    add_project_arguments(nvcc_flags, language: 'cuda')
+    # Set default CUDA architectures if not specified by user
+    # Users can override with cuda_args and cuda_link_args options
+    # See README.md for details and examples
+    cuda_args_option = get_option('cuda_args')
+    cuda_link_args_option = get_option('cuda_link_args')
+
+    cuda_args = [
+        '-gencode=arch=compute_80,code=sm_80',
+        '-gencode=arch=compute_90,code=sm_90'
+    ]
+    cuda_link_args = [
+        '-gencode=arch=compute_80,code=sm_80',
+        '-gencode=arch=compute_90,code=sm_90'
+    ]
+
+    # Apply defaults only if user hasn't specified custom args
+    if cuda_args_option == []
+        add_project_arguments(cuda_args, language: 'cuda')
+    endif
+
+    if cuda_link_args_option == []
+        add_project_link_arguments(cuda_link_args, language: 'cuda')
+    endif
 
     # Refer to https://mesonbuild.com/Cuda-module.html
     add_project_arguments('-forward-unknown-to-host-compiler', language: 'cuda')
     add_project_arguments('-rdc=true', language: 'cuda')
-
-    nvcc_flags_link = []
-    nvcc_flags_link += ['-gencode=arch=compute_80,code=sm_80']
-    nvcc_flags_link += ['-gencode=arch=compute_90,code=sm_90']
-    add_project_link_arguments(nvcc_flags_link, language: 'cuda')
 else
     warning('CUDA not found. UCX backend will be built without CUDA support, and some plugins will be disabled.')
 endif
@@ -177,7 +190,7 @@ if ucx_dep.found() and cuda_dep.found() and nvcc_prog.found()
     have_gpu_side = cuda.compiles('''
             #include <ucp/api/device/ucp_device_impl.h>
             int main() { return 0; }
-        ''', dependencies : ucx_dep, args: nvcc_flags)
+        ''', dependencies : ucx_dep, args: cuda_args)
 
     have_host_side = cpp.compiles('''
             #include <ucp/api/device/ucp_host.h>