feat(cuda): add minimal CUDA driver bindings

Cargo.toml

@@ -29,6 +29,7 @@ members = [
     "vortex-duckdb",
     "vortex-cuda",
     "vortex-cuda/cub",
+    "vortex-cuda/driver",
     "vortex-cuda/ffi",
     "vortex-cuda/gpu-scan-cli",
     "vortex-cuda/macros",
@@ -348,6 +349,7 @@ vortex-zstd = { version = "0.1.0", path = "./encodings/zstd", default-features =
 vortex-bench = { path = "./vortex-bench", default-features = false }
 vortex-bench-support = { path = "./benchmarks/bench-support" }
 vortex-cuda = { path = "./vortex-cuda", default-features = false }
+vortex-cuda-driver = { path = "./vortex-cuda/driver" }
 vortex-cuda-macros = { path = "./vortex-cuda/macros" }
 vortex-duckdb = { path = "./vortex-duckdb", default-features = false }
 vortex-tui = { path = "./vortex-tui" }

vortex-cuda/driver/Cargo.toml

@@ -0,0 +1,23 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright the Vortex contributors
+
+[package]
+name = "vortex-cuda-driver"
+authors.workspace = true
+description = "Minimal CUDA driver API bindings for Vortex"
+edition = { workspace = true }
+categories = { workspace = true }
+include = { workspace = true }
+keywords = { workspace = true }
+license = { workspace = true }
+readme = { workspace = true }
+repository = { workspace = true }
+rust-version = { workspace = true }
+version = { workspace = true }
+publish = false
+
+[lints]
+workspace = true
+
+[dependencies]
+libloading = { workspace = true }

vortex-cuda/driver/src/lib.rs

@@ -0,0 +1,12 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright the Vortex contributors
+
+//! Minimal CUDA driver API bindings for Vortex.
+//!
+//! The driver library is loaded lazily, so dependents build and run without CUDA until
+//! they first use it. The raw bindings and result wrappers operate on CUDA driver handles.
+
+pub mod result;
+pub mod sys;
+
+pub use result::DriverError;

vortex-cuda/driver/src/result.rs

@@ -0,0 +1,257 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright the Vortex contributors
+
+//! Driver calls returning [`Result`]s, for callers that need the raw handles.
+
+use std::ffi::CStr;
+use std::ffi::c_uint;
+use std::ffi::c_void;
+use std::fmt;
+use std::mem::MaybeUninit;
+
+use crate::sys;
+use crate::sys::CUresult;
+
+/// A failed CUDA driver call.
+#[derive(Clone, Copy, PartialEq, Eq, Hash)]
+pub struct DriverError(pub CUresult);
+
+impl DriverError {
+    /// The driver's name for the error, such as `CUDA_ERROR_INVALID_VALUE`.
+    pub fn error_name(&self) -> Result<&'static CStr, DriverError> {
+        let mut name = std::ptr::null();
+        // SAFETY: The driver writes a pointer to a static string.
+        unsafe { sys::cuGetErrorName(self.0, &raw mut name) }.result()?;
+        // SAFETY: Successful calls return a NUL-terminated static string.
+        Ok(unsafe { CStr::from_ptr(name) })
+    }
+
+    /// The driver's description of the error.
+    pub fn error_string(&self) -> Result<&'static CStr, DriverError> {
+        let mut description = std::ptr::null();
+        // SAFETY: The driver writes a pointer to a static string.
+        unsafe { sys::cuGetErrorString(self.0, &raw mut description) }.result()?;
+        // SAFETY: Successful calls return a NUL-terminated static string.
+        Ok(unsafe { CStr::from_ptr(description) })
+    }
+}
+
+impl fmt::Debug for DriverError {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        match self.error_name() {
+            Ok(name) => write!(f, "DriverError({}, {})", self.0.0, name.to_string_lossy()),
+            Err(_) => write!(f, "DriverError({})", self.0.0),
+        }
+    }
+}
+
+impl fmt::Display for DriverError {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        match (self.error_name(), self.error_string()) {
+            (Ok(name), Ok(description)) => write!(
+                f,
+                "{}: {}",
+                name.to_string_lossy(),
+                description.to_string_lossy()
+            ),
+            _ => write!(f, "CUDA driver error {}", self.0.0),
+        }
+    }
+}
+
+impl std::error::Error for DriverError {}
+
+impl CUresult {
+    /// `Ok` for [`CUresult::CUDA_SUCCESS`], otherwise the error.
+    #[inline]
+    pub fn result(self) -> Result<(), DriverError> {
+        if self == CUresult::CUDA_SUCCESS {
+            Ok(())
+        } else {
+            Err(DriverError(self))
+        }
+    }
+}
+
+/// Initializes the driver; repeated calls are cheap.
+pub fn init() -> Result<(), DriverError> {
+    // SAFETY: cuInit has no preconditions.
+    unsafe { sys::cuInit(0) }.result()
+}
+
+/// Allocates `bytes` of page-locked host memory with the given `cuMemHostAlloc` flags.
+///
+/// # Safety
+///
+/// A CUDA context must be current; the memory is uninitialized.
+pub unsafe fn malloc_host(bytes: usize, flags: c_uint) -> Result<*mut c_void, DriverError> {
+    let mut pointer = MaybeUninit::uninit();
+    // SAFETY: The driver writes the allocation into `pointer` on success.
+    unsafe { sys::cuMemHostAlloc(pointer.as_mut_ptr(), bytes, flags) }.result()?;
+    // SAFETY: Initialized by the successful call.
+    Ok(unsafe { pointer.assume_init() })
+}
+
+/// Frees memory allocated by [`malloc_host`].
+///
+/// # Safety
+///
+/// `pointer` must come from [`malloc_host`] and no pending copy may still use it.
+pub unsafe fn free_host(pointer: *mut c_void) -> Result<(), DriverError> {
+    // SAFETY: Under the caller's guarantees.
+    unsafe { sys::cuMemFreeHost(pointer) }.result()
+}
+
+/// Synchronously copies `target.len()` elements from device memory at `source`.
+///
+/// # Safety
+///
+/// `source` must address at least `size_of_val(target)` readable device bytes.
+pub unsafe fn memcpy_dtoh_sync<T>(
+    target: &mut [T],
+    source: sys::CUdeviceptr,
+) -> Result<(), DriverError> {
+    // SAFETY: Under the caller's guarantees; `target` is writable for its whole size.
+    unsafe { sys::cuMemcpyDtoH_v2(target.as_mut_ptr().cast(), source, size_of_val(target)) }
+        .result()
+}
+
+/// Device queries.
+pub mod device {
+    use super::*;
+
+    /// The device with the given ordinal.
+    pub fn get(ordinal: i32) -> Result<sys::CUdevice, DriverError> {
+        let mut device = MaybeUninit::uninit();
+        // SAFETY: The driver writes the device into `device` on success.
+        unsafe { sys::cuDeviceGet(device.as_mut_ptr(), ordinal) }.result()?;
+        // SAFETY: Initialized by the successful call.
+        Ok(unsafe { device.assume_init() })
+    }
+
+    /// The number of devices.
+    pub fn get_count() -> Result<i32, DriverError> {
+        let mut count = 0;
+        // SAFETY: The driver writes the count.
+        unsafe { sys::cuDeviceGetCount(&raw mut count) }.result()?;
+        Ok(count)
+    }
+
+    /// A device attribute.
+    ///
+    /// # Safety
+    ///
+    /// `device` must be a valid device handle.
+    pub unsafe fn get_attribute(
+        device: sys::CUdevice,
+        attribute: sys::CUdevice_attribute,
+    ) -> Result<i32, DriverError> {
+        let mut value = 0;
+        // SAFETY: Under the caller's guarantees.
+        unsafe { sys::cuDeviceGetAttribute(&raw mut value, attribute, device) }.result()?;
+        Ok(value)
+    }
+
+    /// The device's current memory pool.
+    ///
+    /// # Safety
+    ///
+    /// `device` must be a valid device handle.
+    pub unsafe fn get_mem_pool(device: sys::CUdevice) -> Result<sys::CUmemoryPool, DriverError> {
+        let mut pool = MaybeUninit::uninit();
+        // SAFETY: Under the caller's guarantees.
+        unsafe { sys::cuDeviceGetMemPool(pool.as_mut_ptr(), device) }.result()?;
+        // SAFETY: Initialized by the successful call.
+        Ok(unsafe { pool.assume_init() })
+    }
+}
+
+/// Memory pool attributes.
+pub mod mem_pool {
+    use super::*;
+
+    /// Reads a pool attribute into `value`.
+    ///
+    /// # Safety
+    ///
+    /// `value` must point to storage of the attribute's type.
+    pub unsafe fn get_attribute(
+        pool: sys::CUmemoryPool,
+        attribute: sys::CUmemPool_attribute,
+        value: *mut c_void,
+    ) -> Result<(), DriverError> {
+        // SAFETY: Under the caller's guarantees.
+        unsafe { sys::cuMemPoolGetAttribute(pool, attribute, value) }.result()
+    }
+
+    /// Sets a pool attribute from `value`.
+    ///
+    /// # Safety
+    ///
+    /// `value` must point to a value of the attribute's type.
+    pub unsafe fn set_attribute(
+        pool: sys::CUmemoryPool,
+        attribute: sys::CUmemPool_attribute,
+        value: *mut c_void,
+    ) -> Result<(), DriverError> {
+        // SAFETY: Under the caller's guarantees.
+        unsafe { sys::cuMemPoolSetAttribute(pool, attribute, value) }.result()
+    }
+}
+
+/// Stream operations on raw handles.
+pub mod stream {
+    use super::*;
+
+    /// The (least, greatest) stream priorities of the current context.
+    pub fn get_priority_range() -> Result<(i32, i32), DriverError> {
+        let (mut least, mut greatest) = (0, 0);
+        // SAFETY: The driver writes both values.
+        unsafe { sys::cuCtxGetStreamPriorityRange(&raw mut least, &raw mut greatest) }.result()?;
+        Ok((least, greatest))
+    }
+
+    /// Makes `stream` wait for the work recorded in `event`.
+    ///
+    /// # Safety
+    ///
+    /// Both handles must be valid and belong to the current context.
+    pub unsafe fn wait_event(
+        stream: sys::CUstream,
+        event: sys::CUevent,
+        flags: sys::CUevent_wait_flags,
+    ) -> Result<(), DriverError> {
+        // SAFETY: Under the caller's guarantees.
+        unsafe { sys::cuStreamWaitEvent(stream, event, flags.0) }.result()
+    }
+
+    /// Enqueues a host function on `stream`.
+    ///
+    /// # Safety
+    ///
+    /// `stream` must be valid, and `function` must be safe to call once with `argument` from a
+    /// driver thread, without making CUDA calls.
+    pub unsafe fn launch_host_function(
+        stream: sys::CUstream,
+        function: unsafe extern "C" fn(*mut c_void),
+        argument: *mut c_void,
+    ) -> Result<(), DriverError> {
+        // SAFETY: Under the caller's guarantees.
+        unsafe { sys::cuLaunchHostFunc(stream, Some(function), argument) }.result()
+    }
+}
+
+/// Event operations on raw handles.
+pub mod event {
+    use super::*;
+
+    /// Blocks until the work recorded in `event` completes.
+    ///
+    /// # Safety
+    ///
+    /// `event` must be a valid event handle.
+    pub unsafe fn synchronize(event: sys::CUevent) -> Result<(), DriverError> {
+        // SAFETY: Under the caller's guarantees.
+        unsafe { sys::cuEventSynchronize(event) }.result()
+    }
+}