fix(cuda): propagate validity through dynamic dispatch and fix standalone kernel panics

The dynamic dispatch plan builder silently dropped validity (null bitmaps)
from nullable arrays, hardcoding Validity::NonNullable on the output. This
caused silent data corruption when nullable arrays were fused, and
especially when the output fed into downstream kernels like CUB filter.

Dynamic dispatch changes:
- Thread root array validity through FusedPlan -> MaterializedPlan ->
  execute_typed, replacing the hardcoded Validity::NonNullable.
- Guard against Dict with nullable codes (garbage code values could cause
  OOB shared memory reads in the DICT gather scalar op).
- Guard against RunEnd with nullable ends (garbage end values could cause
  unpredictable binary search / forward-scan behavior).
- Skip kernel launch entirely for Validity::AllInvalid arrays.
- Respect nullability in the len==0 early-return path.

Standalone kernel fixes:
- RunEnd: replace unreachable!() with vortex_bail!() when values have
  per-element validity (Validity::Array), allowing graceful CPU fallback
  instead of a panic.
- Zstd: replace unimplemented!() with vortex_bail!() when decompressed
  data contains nulls, allowing graceful CPU fallback instead of a panic.
- ALP: clarify that patch validity does not need scattering (the encoder
  strips null positions from the exception list), remove stale TODO.

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

Author: 0ax1

vortex-cuda/benches/dynamic_dispatch_cuda.rs

@@ -135,6 +135,7 @@ impl BenchRunner {
             dispatch_plan,
             device_buffers,
             shared_mem_bytes,
+            ..
         } = plan.materialize(cuda_ctx).vortex_expect("materialize plan");
 
         let device_plan = Arc::new(

vortex-cuda/src/dynamic_dispatch/mod.rs

@@ -24,14 +24,18 @@ use cudarc::driver::DevicePtr;
 use cudarc::driver::LaunchConfig;
 use cudarc::driver::PushKernelArg;
 use vortex::array::Canonical;
+use vortex::array::IntoArray;
+use vortex::array::arrays::ConstantArray;
 use vortex::array::arrays::PrimitiveArray;
 use vortex::array::buffer::BufferHandle;
 use vortex::array::buffer::DeviceBufferExt;
 use vortex::array::match_each_unsigned_integer_ptype;
+use vortex::array::scalar::Scalar;
 use vortex::array::validity::Validity;
 use vortex::buffer::Alignment;
 use vortex::buffer::ByteBuffer;
 use vortex::buffer::ByteBufferMut;
+use vortex::dtype::DType;
 use vortex::dtype::Nullability;
 use vortex::dtype::PType;
 use vortex::error::VortexResult;
@@ -408,6 +412,15 @@ impl ScalarOp {
 impl MaterializedPlan {
     pub fn execute(self, len: usize, ctx: &mut CudaExecutionCtx) -> VortexResult<Canonical> {
         let output_ptype = self.dispatch_plan.output_ptype();
+
+        // All values are null — no need to touch the GPU.
+        if matches!(self.validity, Validity::AllInvalid) {
+            let dtype = DType::Primitive(output_ptype, Nullability::Nullable);
+            return ConstantArray::new(Scalar::null(dtype), len)
+                .into_array()
+                .to_canonical();
+        }
+
         // The CUDA kernels are instantiated for unsigned integer types only;
         // map signed/float ptypes to their same-width unsigned counterpart.
         let unsigned_ptype = match output_ptype {
@@ -431,9 +444,11 @@ impl MaterializedPlan {
     where
         T: cudarc::driver::DeviceRepr + vortex::dtype::NativePType,
     {
+        let nullability = self.validity.nullability();
+
         if len == 0 {
             return Ok(Canonical::Primitive(PrimitiveArray::empty::<T>(
-                Nullability::NonNullable,
+                nullability,
             )));
         }
 
@@ -467,7 +482,7 @@ impl MaterializedPlan {
         Ok(Canonical::Primitive(PrimitiveArray::from_buffer_handle(
             BufferHandle::new_device(output_buf.slice_typed::<T>(0..len)),
             output_ptype,
-            Validity::NonNullable,
+            self.validity,
         )))
     }
 }
@@ -485,8 +500,10 @@ mod tests {
     use vortex::array::arrays::DictArray;
     use vortex::array::arrays::PrimitiveArray;
     use vortex::array::scalar::Scalar;
+    use vortex::array::validity::Validity;
     use vortex::array::validity::Validity::NonNullable;
     use vortex::buffer::Buffer;
+    use vortex::dtype::Nullability;
     use vortex::dtype::PType;
     use vortex::encodings::alp::ALP;
     use vortex::encodings::alp::ALPArrayExt;
@@ -509,6 +526,7 @@ mod tests {
     use crate::CudaBufferExt;
     use crate::CudaDeviceBuffer;
     use crate::CudaExecutionCtx;
+    use crate::executor::CudaArrayExt;
     use crate::hybrid_dispatch::try_gpu_dispatch;
     use crate::session::CudaSession;
 
@@ -1817,4 +1835,204 @@ mod tests {
 
         Ok(())
     }
+
+    // ═══════════════════════════════════════════════════════════════════
+    // Validity propagation tests
+    // ═══════════════════════════════════════════════════════════════════
+
+    /// Nullable Primitive array — LOAD source with validity propagated.
+    #[crate::test]
+    async fn test_nullable_primitive() -> VortexResult<()> {
+        let mut cuda_ctx = CudaSession::create_execution_ctx(&VortexSession::empty())?;
+
+        let array = PrimitiveArray::from_option_iter(
+            (0..2048u32).map(|i| if i % 3 == 0 { None } else { Some(i) }),
+        );
+        let cpu = array.to_canonical()?.into_array();
+
+        let gpu = try_gpu_dispatch(&array.into_array(), &mut cuda_ctx)
+            .await?
+            .into_host()
+            .await?
+            .into_array();
+
+        vortex::array::assert_arrays_eq!(cpu, gpu);
+        Ok(())
+    }
+
+    /// Nullable FoR(BitPacked) — validity from the root FoR propagated
+    /// through the fused plan.
+    #[crate::test]
+    async fn test_nullable_for_bitpacked() -> VortexResult<()> {
+        let mut cuda_ctx = CudaSession::create_execution_ctx(&VortexSession::empty())?;
+
+        let len = 2048;
+        let reference = 1000u32;
+        let residuals: Vec<u32> = (0..len).map(|i| (i % 64) as u32).collect();
+
+        // Build a nullable source — encode as FoR(BitPacked), then replace
+        // validity on the root FoR array.
+        let prim = PrimitiveArray::from_option_iter(residuals.iter().enumerate().map(|(i, &v)| {
+            if i % 5 == 0 {
+                None
+            } else {
+                Some(v + reference)
+            }
+        }));
+
+        // Build expected from the nullable primitive.
+        let cpu = prim.to_canonical()?.into_array();
+
+        // Encode the non-null residuals into FoR(BitPacked). The validity
+        // lives on the outermost array; inner children are non-nullable.
+        let inner_prim = PrimitiveArray::new(Buffer::from(residuals), NonNullable);
+        let bp = BitPacked::encode(&inner_prim.into_array(), 6)?;
+        let for_arr = FoR::try_new(bp.into_array(), reference.into())?;
+
+        // Re-wrap with the nullable validity from the original primitive.
+        let validity = prim.validity()?;
+        let nullable_for = PrimitiveArray::from_buffer_handle(
+            for_arr.into_array().to_primitive().into_data_parts().buffer,
+            PType::U32,
+            validity,
+        );
+
+        let gpu = try_gpu_dispatch(&nullable_for.into_array(), &mut cuda_ctx)
+            .await?
+            .into_host()
+            .await?
+            .into_array();
+
+        vortex::array::assert_arrays_eq!(cpu, gpu);
+        Ok(())
+    }
+
+    /// AllInvalid array — kernel should be skipped entirely.
+    #[crate::test]
+    async fn test_all_invalid_skips_kernel() -> VortexResult<()> {
+        let mut cuda_ctx = CudaSession::create_execution_ctx(&VortexSession::empty())?;
+
+        let array = PrimitiveArray::new(Buffer::from(vec![0u32; 2048]), Validity::AllInvalid);
+
+        let result = try_gpu_dispatch(&array.into_array(), &mut cuda_ctx)
+            .await?
+            .into_host()
+            .await?;
+
+        let prim = result.into_primitive();
+        assert_eq!(prim.len(), 2048);
+        assert!(matches!(prim.validity()?, Validity::AllInvalid));
+        Ok(())
+    }
+
+    /// AllValid nullable array — should fuse and produce AllValid output.
+    #[crate::test]
+    async fn test_all_valid_nullable() -> VortexResult<()> {
+        let mut cuda_ctx = CudaSession::create_execution_ctx(&VortexSession::empty())?;
+
+        let values: Vec<u32> = (0..2048).collect();
+        let array = PrimitiveArray::new(Buffer::from(values.clone()), Validity::AllValid);
+
+        let cpu = array.to_canonical()?.into_array();
+        let gpu = try_gpu_dispatch(&array.into_array(), &mut cuda_ctx)
+            .await?
+            .into_host()
+            .await?
+            .into_array();
+
+        vortex::array::assert_arrays_eq!(cpu, gpu);
+        Ok(())
+    }
+
+    /// Dict with nullable codes must fall back to Unfused (not fused).
+    #[crate::test]
+    fn test_dict_nullable_codes_rejected() -> VortexResult<()> {
+        use vortex::buffer::buffer;
+
+        let codes = PrimitiveArray::from_option_iter([Some(0u32), None, Some(1), None, Some(2)]);
+        let values = PrimitiveArray::new(buffer![10u32, 20, 30], NonNullable);
+        let dict = DictArray::try_new(codes.into_array(), values.into_array())?;
+
+        let plan = DispatchPlan::new(&dict.into_array())?;
+        assert!(
+            matches!(plan, DispatchPlan::Unfused),
+            "Dict with nullable codes should fall back to Unfused"
+        );
+        Ok(())
+    }
+
+    /// Dict with non-nullable codes but nullable values should still fuse.
+    #[crate::test]
+    async fn test_dict_nullable_values_fuses() -> VortexResult<()> {
+        use vortex::buffer::buffer;
+
+        let mut cuda_ctx = CudaSession::create_execution_ctx(&VortexSession::empty())?;
+
+        let codes = PrimitiveArray::new(buffer![0u32, 1, 2, 2, 1, 0], NonNullable);
+        let values = PrimitiveArray::from_option_iter([Some(10u32), None, Some(30)]);
+        let dict = DictArray::try_new(codes.into_array(), values.into_array())?;
+
+        let cpu = dict.to_canonical()?.into_array();
+        let gpu = dict
+            .into_array()
+            .execute_cuda(&mut cuda_ctx)
+            .await?
+            .into_host()
+            .await?
+            .into_array();
+
+        vortex::array::assert_arrays_eq!(cpu, gpu);
+        Ok(())
+    }
+
+    /// Nullable FoR(BitPacked) through CUB filter — the original bug scenario.
+    /// Validity must survive through fused dispatch and into the filter.
+    #[crate::test]
+    async fn test_nullable_fused_then_filter() -> VortexResult<()> {
+        use vortex::array::arrays::FilterArray;
+        use vortex::mask::Mask;
+
+        let mut cuda_ctx = CudaSession::create_execution_ctx(&VortexSession::empty())?;
+
+        let len = 2048usize;
+        let values: Vec<Option<u32>> = (0..len)
+            .map(|i| {
+                if i % 7 == 0 {
+                    None
+                } else {
+                    Some((i % 64) as u32)
+                }
+            })
+            .collect();
+        let prim = PrimitiveArray::from_option_iter(values.iter().copied());
+
+        // Keep every other element.
+        let mask = Mask::from_iter((0..len).map(|i| i % 2 == 0));
+        let filter_array = FilterArray::try_new(prim.into_array(), mask)?;
+
+        let cpu = filter_array.to_canonical()?.into_array();
+        let gpu = filter_array
+            .into_array()
+            .execute_cuda(&mut cuda_ctx)
+            .await?
+            .into_host()
+            .await?
+            .into_array();
+
+        vortex::array::assert_arrays_eq!(cpu, gpu);
+        Ok(())
+    }
+
+    /// Empty nullable array should preserve nullability.
+    #[crate::test]
+    async fn test_empty_nullable_array() -> VortexResult<()> {
+        let mut cuda_ctx = CudaSession::create_execution_ctx(&VortexSession::empty())?;
+
+        let array = PrimitiveArray::new(Buffer::<u32>::empty(), Validity::AllValid);
+        let result = try_gpu_dispatch(&array.into_array(), &mut cuda_ctx).await?;
+        let prim = result.into_primitive();
+        assert_eq!(prim.len(), 0);
+        assert_eq!(prim.validity()?.nullability(), Nullability::Nullable);
+        Ok(())
+    }
 }

vortex-cuda/src/dynamic_dispatch/plan_builder.rs

@@ -15,6 +15,7 @@ use vortex::array::arrays::Slice;
 use vortex::array::arrays::dict::DictArraySlotsExt;
 use vortex::array::arrays::slice::SliceArrayExt;
 use vortex::array::buffer::BufferHandle;
+use vortex::array::validity::Validity;
 use vortex::dtype::PType;
 use vortex::encodings::alp::ALP;
 use vortex::encodings::alp::ALPArrayExt;
@@ -52,6 +53,8 @@ pub struct MaterializedPlan {
     pub device_buffers: Vec<BufferHandle>,
     /// Dynamic shared memory bytes needed to launch this plan.
     pub shared_mem_bytes: u32,
+    /// Validity of the root array, propagated to the output.
+    pub validity: Validity,
 }
 
 /// Checks whether the encoding of an array can be fused into a dynamic-dispatch plan.
@@ -72,6 +75,11 @@ fn is_dyn_dispatch_compatible(array: &ArrayRef) -> bool {
     }
     if id == Dict::ID {
         let arr = array.as_::<Dict>();
+        // Nullable codes could hold garbage values at null positions, causing
+        // out-of-bounds shared memory reads in the DICT gather scalar op.
+        if arr.codes().dtype().is_nullable() {
+            return false;
+        }
         // Dict codes and values may have different byte widths.
         // The kernel handles mixed widths via widening input stages,
         // but only when codes are no wider than values (the output type).
@@ -85,6 +93,12 @@ fn is_dyn_dispatch_compatible(array: &ArrayRef) -> bool {
     }
     if id == RunEnd::ID {
         let arr = array.as_::<RunEnd>();
+        // Nullable ends could hold garbage values at null positions, causing
+        // unpredictable binary search / forward-scan behavior in the RUNEND
+        // source op.
+        if arr.ends().dtype().is_nullable() {
+            return false;
+        }
         // RunEnd ends and values may have different byte widths.
         // The kernel handles mixed widths via widening input stages,
         // but only when ends are no wider than values (the output type).
@@ -213,14 +227,18 @@ pub struct FusedPlan {
     output_elem_bytes: u32,
     /// PType of the root (output) array, as a C ABI tag.
     output_ptype: PTypeTag,
+    /// Validity of the root array, propagated to the output.
+    validity: Validity,
 }
 
 impl DispatchPlan {
     /// Construct a plan by inspecting the encoding tree in a single pass.
     ///
     /// # Limitations
     ///
-    /// - Validity bitmaps are ignored; only `NonNullable`/`AllValid` is supported.
+    /// - Validity is propagated from the root array to the output. Nullable
+    ///   arrays are supported, but Dict with nullable codes and RunEnd with
+    ///   nullable ends are rejected to guard against out-of bounds access.
     /// - `BitPackedArray` and `ALPArray` with patches are not supported.
     /// - Only f32 ALP is supported (kernel stores multipliers as `float`).
     pub fn new(array: &ArrayRef) -> VortexResult<Self> {
@@ -276,6 +294,7 @@ impl FusedPlan {
         }
         let output_elem_bytes = output_ptype_rust.byte_width() as u32;
         let output_ptype = ptype_to_tag(output_ptype_rust);
+        let validity = array.validity()?;
 
         let mut pending_subtrees: Vec<ArrayRef> = Vec::new();
         let mut plan = Self {
@@ -284,6 +303,7 @@ impl FusedPlan {
             source_buffers: Vec::new(),
             output_elem_bytes,
             output_ptype,
+            validity,
         };
 
         let len = array.len() as u32;
@@ -365,6 +385,7 @@ impl FusedPlan {
             dispatch_plan: CudaDispatchPlan::new(stages, self.output_ptype),
             device_buffers,
             shared_mem_bytes,
+            validity: self.validity,
         })
     }