feat(cuda): support column projection and dictionary decoding in file scans

vortex-cuda/ffi/Cargo.toml

@@ -22,8 +22,12 @@ vortex-cuda = { path = ".." }
 vortex-ffi = { path = "../../vortex-ffi" }
 
 [dev-dependencies]
+tempfile = { workspace = true }
 vortex-cuda-macros = { workspace = true }
 
+[build-dependencies]
+cbindgen = { workspace = true }
+
 [lib]
 name = "vortex_cuda_ffi"
 crate-type = ["rlib", "staticlib", "cdylib"]

vortex-cuda/ffi/README.md

@@ -27,3 +27,8 @@ pool and CUDA state are reused as well.
 On Linux, use `vx_cuda_scan_path_arrow_device_stream_with_options` with
 `vx_cuda_scan_options.flags = VX_CUDA_SCAN_FLAG_DIRECT_IO` to bypass the operating system page
 cache for pooled data-plane reads. Footer and zone-map reads remain buffered on the host.
+
+`build.rs` generates `cinclude/vortex_cuda.h` with cbindgen on stable Rust, without macro
+expansion. Edit the API and docs in `src/lib.rs`, not the generated header; commit
+regenerated headers with API changes. `cbindgen.toml` supplies the standard Arrow Device
+interface compatibility preamble.

vortex-cuda/ffi/build.rs

@@ -0,0 +1,29 @@
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright the Vortex contributors
+
+use std::error::Error;
+use std::process::Command;
+
+fn main() -> Result<(), Box<dyn Error>> {
+    println!("cargo:rerun-if-changed=src");
+    println!("cargo:rerun-if-changed=cbindgen.toml");
+    println!("cargo:rerun-if-changed=build.rs");
+
+    let header = "cinclude/vortex_cuda.h";
+    // The CUDA API needs no macro expansion or dependency parsing, so generate on stable Rust
+    // without recursively building the CUDA implementation.
+    cbindgen::Builder::new()
+        .with_src("src/lib.rs")
+        .with_config(cbindgen::Config::from_file("cbindgen.toml")?)
+        .generate()?
+        .write_to_file(header);
+    if !Command::new("clang-format")
+        .args(["--style=file", "-i"])
+        .arg(header)
+        .status()
+        .is_ok_and(|status| status.success())
+    {
+        println!("cargo:warning=clang-format unavailable or failed; CUDA header left unformatted");
+    }
+    Ok(())
+}

vortex-cuda/ffi/cbindgen.toml

@@ -0,0 +1,72 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright the Vortex contributors
+
+language = "C"
+braces = "SameLine"
+cpp_compat = true
+usize_is_size_t = true
+style = "both"
+no_includes = true
+
+header = """
+// SPDX-License-Identifier: Apache-2.0
+// SPDX-FileCopyrightText: Copyright the Vortex contributors
+#pragma once
+
+// THIS FILE IS AUTO-GENERATED, DO NOT MAKE EDITS DIRECTLY
+
+#include <stddef.h>
+#include <stdint.h>
+
+#include "vortex.h"
+
+/* Link against the CUDA-enabled FFI library that provides both the base Vortex FFI and these CUDA
+ * entry points. Do not pass Vortex handles between independently linked Rust FFI libraries. */
+
+/* Definitions from the Arrow C Device data interface. Define USE_OWN_ARROW_DEVICE to skip them.
+ * https://arrow.apache.org/docs/format/CDeviceDataInterface.html */
+#if !defined(ARROW_C_DEVICE_DATA_INTERFACE) && !defined(USE_OWN_ARROW_DEVICE)
+#define ARROW_C_DEVICE_DATA_INTERFACE
+
+typedef int32_t ArrowDeviceType;
+#define ARROW_DEVICE_CPU          1
+#define ARROW_DEVICE_CUDA         2
+#define ARROW_DEVICE_CUDA_HOST    3
+#define ARROW_DEVICE_OPENCL       4
+#define ARROW_DEVICE_VULKAN       7
+#define ARROW_DEVICE_METAL        8
+#define ARROW_DEVICE_VPI          9
+#define ARROW_DEVICE_ROCM         10
+#define ARROW_DEVICE_ROCM_HOST    11
+#define ARROW_DEVICE_EXT_DEV      12
+#define ARROW_DEVICE_CUDA_MANAGED 13
+#define ARROW_DEVICE_ONEAPI       14
+#define ARROW_DEVICE_WEBGPU       15
+#define ARROW_DEVICE_HEXAGON      16
+
+struct ArrowDeviceArray {
+    struct ArrowArray array;
+    int64_t device_id;
+    ArrowDeviceType device_type;
+    void *sync_event;
+    int64_t reserved[3];
+};
+#endif
+
+#if !defined(ARROW_C_DEVICE_STREAM_INTERFACE) && !defined(USE_OWN_ARROW_DEVICE)
+#define ARROW_C_DEVICE_STREAM_INTERFACE
+struct ArrowDeviceArrayStream {
+    ArrowDeviceType device_type;
+    int (*get_schema)(struct ArrowDeviceArrayStream *, struct ArrowSchema *out);
+    int (*get_next)(struct ArrowDeviceArrayStream *, struct ArrowDeviceArray *out);
+    const char *(*get_last_error)(struct ArrowDeviceArrayStream *);
+    void (*release)(struct ArrowDeviceArrayStream *);
+    void *private_data;
+};
+#endif
+"""
+
+# These externally defined Arrow ABI types use struct tags, not typedef names.
+[export.rename]
+"ArrowDeviceArray" = "struct ArrowDeviceArray"
+"ArrowDeviceArrayStream" = "struct ArrowDeviceArrayStream"