GPU Scan (#5012)

GPU specific readers

Signed-off-by: Robert Kruszewski <[email protected]>

Author: robert3005

Cargo.lock

@@ -19,6 +19,7 @@ all-features = true
 [dependencies]
 async-trait = { workspace = true }
 bytes = { workspace = true }
+cudarc = { workspace = true, optional = true }
 flatbuffers = { workspace = true }
 futures = { workspace = true, features = ["std", "async-await"] }
 # Needed to pickup the "wasm_js" feature for wasm targets from the workspace configuration
@@ -44,6 +45,7 @@ vortex-expr = { workspace = true }
 vortex-fastlanes = { workspace = true }
 vortex-flatbuffers = { workspace = true, features = ["file"] }
 vortex-fsst = { workspace = true }
+vortex-gpu = { workspace = true, optional = true }
 vortex-io = { workspace = true }
 vortex-layout = { workspace = true }
 vortex-metrics = { workspace = true }
@@ -76,3 +78,10 @@ tokio = [
     "vortex-scan/tokio",
 ]
 zstd = ["dep:vortex-zstd", "vortex-layout/zstd"]
+gpu = [
+    "dep:cudarc",
+    "dep:vortex-gpu",
+    "vortex-gpu/cuda",
+    "vortex-layout/gpu",
+    "vortex-scan/gpu",
+]

java/testfiles/Cargo.lock

@@ -90,6 +90,20 @@ impl VortexFile {
         Ok(ScanBuilder::new(self.layout_reader()?).with_metrics(self.metrics.clone()))
     }
 
+    #[cfg(feature = "gpu")]
+    pub fn gpu_scan(
+        &self,
+        ctx: Arc<cudarc::driver::CudaContext>,
+    ) -> VortexResult<vortex_scan::gpu::GpuScanBuilder<vortex_gpu::GpuVector>> {
+        let segment_source = self.segment_source();
+        let gpu_reader = self
+            .footer
+            .layout()
+            .new_gpu_reader("".into(), segment_source, ctx)?;
+
+        Ok(vortex_scan::gpu::GpuScanBuilder::new(gpu_reader))
+    }
+
     /// Returns true if the expression will never match any rows in the file.
     pub fn can_prune(&self, filter: &ExprRef) -> VortexResult<bool> {
         let Some((stats, fields)) = self

vortex-file/Cargo.toml

@@ -204,8 +204,8 @@ fn benchmark_gpu_for_bp_jit_decompress_kernel_only(c: &mut Criterion) {
                 let mut total_time = Duration::ZERO;
                 for _ in 0..iters {
                     // This only measures kernel execution time, not memory transfers
-                    let (_result, kernel_time) = create_run_jit_kernel(ctx.clone(), array).unwrap();
-                    total_time += kernel_time;
+                    let (_result, kernel_time) = create_run_jit_kernel(&ctx, array).unwrap();
+                    total_time += kernel_time.elapsed().unwrap();
                 }
                 total_time
             });