docs

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

Author: 0ax1

vortex-cuda/src/dynamic_dispatch/mod.rs

@@ -53,11 +53,11 @@ include!(concat!(env!("OUT_DIR"), "/dynamic_dispatch.rs"));
 ///
 /// # Safety
 ///
-/// The caller must ensure `T` is a `#[repr(C)]` type with no padding that
-/// contains uninitialised bytes.  All the types we serialise (`PlanHeader`,
-/// `PackedStage`, `ScalarOp`) satisfy this because they are bindgen-generated
-/// `#[repr(C)]` structs whose padding bytes are always written before
-/// serialisation.
+/// The caller must ensure `T` is a `#[repr(C)]` type whose layout is
+/// compatible with the C ABI.  All the types we serialise (`PlanHeader`,
+/// `PackedStage`, `ScalarOp`) are bindgen-generated `#[repr(C)]` structs.
+/// Padding bytes may be uninitialised on the Rust side, but the C reader
+/// never inspects them, so the values are irrelevant.
 fn as_bytes<T: Sized>(val: &T) -> &[u8] {
     unsafe { from_raw_parts(std::ptr::addr_of!(*val).cast(), size_of::<T>()) }
 }
@@ -514,7 +514,7 @@ mod tests {
     #[crate::test]
     fn test_plan_structure() {
         // Stage 0: input dict values (BP→FoR) into smem[0..256)
-        // Stage 1: output codes (BP→FoR→DICT) into smem[256..2304), gather from smem[0]
+        // Stage 1: output codes (BP→FoR→DICT) into smem[256..1280), gather from smem[0]
         let plan = CudaDispatchPlan::new([
             MaterializedStage::new(
                 0xAAAA,

vortex-cuda/src/dynamic_dispatch/plan_builder.rs

@@ -108,7 +108,7 @@ impl Stage {
 }
 
 type SmemOffset = u32;
-type StageLen = u32;
+type OutputLen = u32;
 
 /// A dispatch plan before device materialization.
 ///
@@ -118,7 +118,6 @@ pub enum DispatchPlan {
     /// Entire encoding tree is fusable into a single kernel launch.
     Fused(FusedPlan),
     /// Some subtrees need separate execution before the fused plan can run.
-    /// Shared memory has already been validated.
     PartiallyFused {
         /// The fused plan (with placeholder buffer slots for pending subtrees).
         plan: FusedPlan,
@@ -165,11 +164,10 @@ pub enum DispatchPlan {
 /// exceed `stage.len` by up to 1023 elements. This overflow is absorbed by
 /// the scratch region (`SMEM_TILE_SIZE` ≥ `FL_CHUNK_SIZE`).
 pub struct FusedPlan {
-    /// Stages in kernel execution order. All stages except the last decode
-    /// fully into persistent shared memory; the final stage produces the
-    /// output.
-    stages: Vec<(Stage, SmemOffset, StageLen)>,
-    /// Shared memory elements reserved by the preceding (non-output) stages.
+    /// Stages in kernel execution order; all but the last decode into
+    /// shared memory, the last decodes into global memeory.
+    stages: Vec<(Stage, SmemOffset, OutputLen)>,
+    /// Shared memory reserved by the non-output stages.
     smem_cursor: SmemOffset,
     /// Source buffers. `None` entries are placeholder slots for pending subtrees,
     /// filled by [`materialize_with_subtrees`] before device copy.