perf(cuda): migrate consumers to vortex-cuda-driver (#10722)

## Problem

Vortex's CUDA consumers still use cudarc and cannot use the new driver's access tracking.

## Solution

Migrate consumers to `vortex-cuda-driver`, move `i256` support into its optional feature, and remove cudarc.

## Assessment

Changes public CUDA types and removes `vortex-array`'s `cudarc` feature. The driver stays unpublished. Builds on the merged CUDA PRs.

Signed-off-by: Alexander Droste <alexander.droste@protonmail.com>

Author: 0ax1

Cargo.lock

@@ -132,12 +132,6 @@ clap = "4.5"
 criterion = { package = "codspeed-criterion-compat-walltime", version = "5.0.0" }
 crossbeam-channel = "0.5.16"
 crossterm = "0.29"
-cudarc = { version = "0.19.0", features = [
-    # The NSight Compute version available on lambda.ai hosts does not inject a symbol for
-    # several functions that cudarc expects to load when this is set to "cuda-12080". We
-    # don't use any CUDA 12.8 specific features currently so this is fine.
-    "cuda-12050",
-] }
 custom-labels = "0.4.4"
 dashmap = "6.2.1"
 datafusion = { version = "55.0.0", default-features = false, features = [

Cargo.toml

@@ -28,7 +28,6 @@ arcref = { workspace = true }
 arrow-buffer = { workspace = true }
 bytes = { workspace = true }
 cfg-if = { workspace = true }
-cudarc = { workspace = true, optional = true }
 enum-iterator = { workspace = true }
 fearless_simd = { workspace = true }
 fearless_simd_macros = { workspace = true }
@@ -76,7 +75,6 @@ vortex-utils = { workspace = true, features = ["dyn-traits"] }
 [features]
 arbitrary = ["dep:arbitrary", "dep:primitive-types"]
 canonical_counter = []
-cudarc = ["dep:cudarc"]
 table-display = ["dep:tabled"]
 _test-harness = ["dep:goldenfile", "dep:rstest", "dep:rstest_reuse"]
 serde = ["dep:serde", "vortex-buffer/serde", "vortex-mask/serde"]

vortex-array/Cargo.toml

@@ -39,9 +39,6 @@ use vortex_error::VortexExpect;
 #[derive(Debug, Copy, Clone, Default, Eq, PartialEq, Hash, PartialOrd, Ord)]
 pub struct i256(arrow_buffer::i256);
 
-#[cfg(feature = "cudarc")]
-unsafe impl cudarc::driver::DeviceRepr for i256 {}
-
 #[expect(
     clippy::same_name_method,
     reason = "inherent methods intentionally shadow arrow_buffer::i256 methods"

vortex-array/src/dtype/bigint/mod.rs

@@ -23,7 +23,6 @@ _test-harness = []
 [dependencies]
 arrow-schema = { workspace = true, features = ["ffi"] }
 async-trait = { workspace = true }
-cudarc = { workspace = true, features = ["f16"] }
 futures = { workspace = true, features = ["executor"] }
 itertools = { workspace = true }
 kanal = { workspace = true }
@@ -36,9 +35,10 @@ prost = { workspace = true }
 tokio = { workspace = true, features = ["fs"] }
 tracing = { workspace = true, features = ["std", "attributes"] }
 vortex = { workspace = true, features = ["zstd"] }
-vortex-array = { workspace = true, features = ["cudarc"] }
+vortex-array = { workspace = true }
 vortex-arrow = { workspace = true }
 vortex-cub = { path = "cub" }
+vortex-cuda-driver = { workspace = true, features = ["f16", "i256"] }
 vortex-cuda-macros = { workspace = true }
 vortex-error = { workspace = true, features = ["object_store"] }
 vortex-nvcomp = { path = "nvcomp" }

vortex-cuda/Cargo.toml

@@ -19,7 +19,6 @@ use std::time::Duration;
 use criterion::BenchmarkId;
 use criterion::Criterion;
 use criterion::Throughput;
-use cudarc::driver::DeviceRepr;
 use futures::executor::block_on;
 use mimalloc::MiMalloc;
 use vortex::array::IntoArray;
@@ -43,6 +42,7 @@ use vortex_array::session::ArraySession;
 use vortex_cuda::CudaDispatchMode;
 use vortex_cuda::CudaSession;
 use vortex_cuda::executor::CudaArrayExt;
+use vortex_cuda_driver::DeviceRepr;
 use vortex_cuda_macros::cuda_available;
 use vortex_cuda_macros::cuda_not_available;
 

vortex-cuda/benches/alp_cuda.rs

@@ -19,7 +19,6 @@ use std::time::Duration;
 use criterion::BenchmarkId;
 use criterion::Criterion;
 use criterion::Throughput;
-use cudarc::driver::DeviceRepr;
 use futures::executor::block_on;
 use mimalloc::MiMalloc;
 use vortex::array::IntoArray;
@@ -42,6 +41,7 @@ use vortex_array::session::ArraySession;
 use vortex_cuda::CudaDispatchMode;
 use vortex_cuda::CudaSession;
 use vortex_cuda::executor::CudaArrayExt;
+use vortex_cuda_driver::DeviceRepr;
 use vortex_cuda_macros::cuda_available;
 use vortex_cuda_macros::cuda_not_available;
 

vortex-cuda/benches/bitpacked_cuda.rs

@@ -15,15 +15,15 @@ use criterion::BatchSize;
 use criterion::BenchmarkId;
 use criterion::Criterion;
 use criterion::Throughput;
-use cudarc::driver::CudaContext;
-use cudarc::driver::CudaStream;
-use cudarc::driver::HostSlice;
-use cudarc::driver::SyncOnDrop;
-use cudarc::driver::result;
-use cudarc::driver::sys;
-use cudarc::driver::sys::CU_MEMHOSTALLOC_WRITECOMBINED;
-use cudarc::driver::sys::CUmemPool_attribute_enum;
 use mimalloc::MiMalloc;
+use vortex_cuda_driver::CudaContext;
+use vortex_cuda_driver::CudaStream;
+use vortex_cuda_driver::HostSlice;
+use vortex_cuda_driver::SyncOnDrop;
+use vortex_cuda_driver::result;
+use vortex_cuda_driver::sys;
+use vortex_cuda_driver::sys::CU_MEMHOSTALLOC_WRITECOMBINED;
+use vortex_cuda_driver::sys::CUmemPool_attribute_enum;
 use vortex_cuda_macros::cuda_available;
 use vortex_cuda_macros::cuda_not_available;
 
@@ -92,7 +92,7 @@ impl HostSlice<u8> for CudaHostBuffer {
     ) -> (&'a [u8], SyncOnDrop<'a>) {
         (
             unsafe { std::slice::from_raw_parts(self.ptr, self.len) },
-            SyncOnDrop::Sync(None),
+            SyncOnDrop::none(),
         )
     }
 
@@ -102,7 +102,7 @@ impl HostSlice<u8> for CudaHostBuffer {
     ) -> (&'a mut [u8], SyncOnDrop<'a>) {
         (
             unsafe { std::slice::from_raw_parts_mut(self.ptr, self.len) },
-            SyncOnDrop::Sync(None),
+            SyncOnDrop::none(),
         )
     }
 }

vortex-cuda/benches/core_primitives.rs

@@ -21,7 +21,6 @@ use std::time::Duration;
 use criterion::BenchmarkId;
 use criterion::Criterion;
 use criterion::Throughput;
-use cudarc::driver::DeviceRepr;
 use futures::executor::block_on;
 use mimalloc::MiMalloc;
 use vortex::array::IntoArray;
@@ -43,6 +42,7 @@ use vortex_array::session::ArraySession;
 use vortex_cuda::CudaDispatchMode;
 use vortex_cuda::CudaSession;
 use vortex_cuda::executor::CudaArrayExt;
+use vortex_cuda_driver::DeviceRepr;
 use vortex_cuda_macros::cuda_available;
 use vortex_cuda_macros::cuda_not_available;
 

vortex-cuda/benches/delta_cuda.rs

@@ -18,7 +18,6 @@ use std::time::Duration;
 use criterion::BenchmarkId;
 use criterion::Criterion;
 use criterion::Throughput;
-use cudarc::driver::DeviceRepr;
 use futures::executor::block_on;
 use mimalloc::MiMalloc;
 use vortex::array::IntoArray;
@@ -31,6 +30,7 @@ use vortex::error::VortexExpect;
 use vortex_cuda::CudaDispatchMode;
 use vortex_cuda::CudaSession;
 use vortex_cuda::executor::CudaArrayExt;
+use vortex_cuda_driver::DeviceRepr;
 use vortex_cuda_macros::cuda_available;
 use vortex_cuda_macros::cuda_not_available;