fuel_dispatch/

compiled.rs

// SPDX-License-Identifier: MIT OR Apache-2.0
//! Compiled-node representation. Phase 7.5 B4 (interface only).
//!
//! A [`CompiledNode`] is the output of dispatch resolution: a Node
//! plus its resolved kernel function pointer plus its `OpParams`.
//! The executor consumes these and runs them.
//!
//! ## Design intent: pipelined compile + execute (deferred)
//!
//! The original B-plan called for pipelined compilation: a compiler
//! thread that walks the graph in topological order, resolves
//! kernels per node, and pushes [`CompiledNode`] entries onto a
//! `crossbeam::channel`. An executor thread consumes the channel,
//! looks up inputs from the graph's storage map (blocking when
//! inputs aren't yet realized), calls the kernel, and stores the
//! output back. Time-to-first-token wins because the compiler runs
//! ahead of the executor.
//!
//! Today's "compile" step is just a `KernelBindingTable` lookup —
//! roughly nanoseconds per node. There's no meaningful work to
//! overlap with execution. Pipelining adds `std::thread` +
//! `crossbeam::channel` + lifecycle handling for zero current win.
//!
//! When compile grows beyond trivial — residency-aware planning,
//! transfer-cost minimization, kernel auto-tuning, dynamic-shape
//! specialization — pipelining lands as a follow-up. The
//! interfaces in this module ([`CompiledNode`] +
//! [`compile_node`] + [`execute_compiled`]) are designed so the
//! threaded variant slots in without changing call sites.

use std::sync::{Arc, RwLock};

#[cfg_attr(not(feature = "cuda"), allow(unused_imports))]
use fuel_ir::Error;
use fuel_ir::dispatch::OpKind;
use fuel_ir::probe::BackendId;
use fuel_ir::{DType, Layout, Result};

use crate::kernel::{KernelBindingTable, KernelCaps, KernelDTypes, KernelRef, OpParams};
use fuel_graph::{Graph, NodeId};
use fuel_memory::Storage;

/// A graph node plus its resolved kernel function pointer and
/// op-specific parameters. Produced by [`compile_node`]; consumed
/// by [`execute_compiled`].
#[derive(Debug, Clone)]
pub struct CompiledNode {
    /// The op family this node implements.
    pub op: OpKind,
    /// Per-operand dtypes used as the binding-table lookup key.
    /// Inputs in order, then outputs. For uniform-precision ops the
    /// list is `[T, T, ..., T]`; for mixed-precision ops it carries
    /// the distinguishing combo. The output dtype is always the last
    /// entry.
    pub dtypes: KernelDTypes,
    /// Which backend's kernel was selected.
    pub backend: BackendId,
    /// Resolved kernel function pointer. Looked up once at compile
    /// time; the executor calls this directly.
    pub kernel: KernelRef,
    /// Capability flags registered alongside the kernel. The executor
    /// consults `caps.strided_input` to decide whether to skip
    /// auto-Contiguize for non-contiguous inputs.
    pub caps: KernelCaps,
    /// Op-specific parameters. Most ops use `OpParams::None`;
    /// reductions / conv / slice carry their config here.
    pub op_params: OpParams,
}

impl CompiledNode {
    /// The output dtype — last entry in `dtypes`. Convenience accessor
    /// for callers that only need the output type (e.g. shape inference,
    /// allocation).
    pub fn output_dtype(&self) -> DType {
        *self
            .dtypes
            .last()
            .expect("CompiledNode::dtypes must have at least one entry (the output)")
    }
}

/// Opt-in per-`(OpKind, BackendId)` host-time accounting for kernel launches
/// ([`execute_compiled_with_wait_hint`]). Off by default (one relaxed atomic load
/// per launch). What it measures: wall time of the kernel call on the HOST. A
/// kernel that only enqueues work (Vulkan batches into a command buffer, CUDA
/// launches on a stream) is charged its enqueue cost, NOT its device time; a
/// kernel that blocks (CPU math, a host-to-device upload, a forced flush) is
/// charged all of it. Device time lands wherever the realize finally waits, so
/// compare the per-op sum with the caller's own wall clock around a step.
pub mod op_profile {
    use std::collections::HashMap;
    use std::sync::Mutex;
    use std::sync::atomic::{AtomicBool, Ordering};
    use std::time::Duration;

    use fuel_ir::dispatch::OpKind;
    use fuel_ir::probe::BackendId;

    static ENABLED: AtomicBool = AtomicBool::new(false);
    static PHASES: Mutex<Option<HashMap<&'static str, (u64, Duration)>>> = Mutex::new(None);

    /// Times a named phase of a realize from `start` until it is dropped, when
    /// accounting is on (a no-op otherwise: one relaxed atomic load). Phases are
    /// hierarchical by NAME only: an outer phase's time includes the inner ones
    /// nested inside it, so rows of [`take_phases`] are not additive.
    pub struct PhaseTimer(Option<(&'static str, std::time::Instant)>);

    impl PhaseTimer {
        pub fn start(name: &'static str) -> Self {
            Self(enabled().then(|| (name, std::time::Instant::now())))
        }
    }

    impl Drop for PhaseTimer {
        fn drop(&mut self) {
            if let Some((name, t0)) = self.0.take() {
                let mut g = PHASES.lock().unwrap_or_else(|p| p.into_inner());
                let e = g
                    .get_or_insert_with(HashMap::new)
                    .entry(name)
                    .or_insert((0, Duration::ZERO));
                e.0 += 1;
                e.1 += t0.elapsed();
            }
        }
    }

    /// Drain the phase totals: `(phase, entries, total time)`, longest first.
    pub fn take_phases() -> Vec<(&'static str, u64, Duration)> {
        let mut g = PHASES.lock().unwrap_or_else(|p| p.into_inner());
        let mut v: Vec<_> = g
            .take()
            .unwrap_or_default()
            .into_iter()
            .map(|(k, (n, d))| (k, n, d))
            .collect();
        v.sort_by_key(|row| std::cmp::Reverse(row.2));
        v
    }

    static TRACE: AtomicBool = AtomicBool::new(false);
    static TRACE_ROWS: Mutex<Vec<LaunchRow>> = Mutex::new(Vec::new());

    /// One kernel launch, in launch order (see [`set_trace`]).
    #[derive(Clone, Debug)]
    pub struct LaunchRow {
        /// `OpKind` rendered with `Debug`.
        pub op: String,
        /// The kernel's backend, rendered with `Debug`.
        pub backend: String,
        /// Where the first input lives (`Cpu`, `Vulkan`, `Cuda`, `Metal`, `?`).
        pub src: &'static str,
        /// Where the first output lives.
        pub dst: &'static str,
        /// Byte length of the first input.
        pub bytes: usize,
        /// Host wall time of the launch.
        pub dt: Duration,
    }

    fn kind(s: &std::sync::Arc<std::sync::RwLock<fuel_memory::Storage>>) -> (&'static str, usize) {
        let Ok(g) = s.try_read() else {
            return ("?", 0);
        };
        let k = match &g.inner {
            fuel_memory::BackendStorage::Cpu(_) => "Cpu",
            #[cfg(feature = "cuda")]
            fuel_memory::BackendStorage::Cuda(_) => "Cuda",
            #[cfg(feature = "vulkan")]
            fuel_memory::BackendStorage::Vulkan(_) => "Vulkan",
            #[allow(unreachable_patterns)]
            _ => "Other",
        };
        (k, g.len_bytes())
    }
    /// `(op, backend)` rendered with `Debug` -> (launches, total host time).
    type Totals = HashMap<(String, String), (u64, Duration)>;

    static TOTALS: Mutex<Option<Totals>> = Mutex::new(None);

    /// Turn accounting on or off. Does not clear what was recorded.
    pub fn set_enabled(on: bool) {
        ENABLED.store(on, Ordering::Relaxed);
    }

    /// Also keep every launch as a [`LaunchRow`], in order (implies the
    /// accounting path). For a one-step anatomy, not for long runs: one row per
    /// launch. Does not clear what was recorded.
    pub fn set_trace(on: bool) {
        TRACE.store(on, Ordering::Relaxed);
    }

    /// Drain the launch trace.
    pub fn take_trace() -> Vec<LaunchRow> {
        std::mem::take(&mut *TRACE_ROWS.lock().unwrap_or_else(|p| p.into_inner()))
    }

    /// Whether accounting (or tracing) is on.
    pub fn enabled() -> bool {
        ENABLED.load(Ordering::Relaxed) || TRACE.load(Ordering::Relaxed)
    }

    pub(super) fn record_launch(
        compiled: &super::CompiledNode,
        inputs: &[std::sync::Arc<std::sync::RwLock<fuel_memory::Storage>>],
        outputs: &[std::sync::Arc<std::sync::RwLock<fuel_memory::Storage>>],
        dt: Duration,
    ) {
        record(compiled.op, compiled.backend, dt);
        if TRACE.load(Ordering::Relaxed) {
            let (src, bytes) = inputs.first().map_or(("?", 0), kind);
            let dst = outputs.first().map_or("?", |o| kind(o).0);
            TRACE_ROWS
                .lock()
                .unwrap_or_else(|p| p.into_inner())
                .push(LaunchRow {
                    op: format!("{:?}", compiled.op),
                    backend: format!("{:?}", compiled.backend),
                    src,
                    dst,
                    bytes,
                    dt,
                });
        }
    }

    pub(super) fn record(op: OpKind, backend: BackendId, dt: Duration) {
        let mut g = TOTALS.lock().unwrap_or_else(|p| p.into_inner());
        let e = g
            .get_or_insert_with(HashMap::new)
            .entry((format!("{op:?}"), format!("{backend:?}")))
            .or_insert((0, Duration::ZERO));
        e.0 += 1;
        e.1 += dt;
    }

    /// Drain the totals: `(op, backend, launches, total host time)`, longest first.
    pub fn take() -> Vec<(String, String, u64, Duration)> {
        let mut g = TOTALS.lock().unwrap_or_else(|p| p.into_inner());
        let mut v: Vec<_> = g
            .take()
            .unwrap_or_default()
            .into_iter()
            .map(|((op, be), (n, d))| (op, be, n, d))
            .collect();
        v.sort_by_key(|row| std::cmp::Reverse(row.3));
        v
    }
}

/// Resolve a node's kernel from the binding table and return a
/// [`CompiledNode`] ready for execution. Production-correct: surfaces
/// `NoBackendForOp` if the table doesn't have a binding for the
/// requested triple.
///
/// Today this is one HashMap lookup. When dispatch grows more
/// sophisticated (residency-aware, cost-aware, auto-tuning), the
/// added work happens here. The interface stays the same.
pub fn compile_node(
    op: OpKind,
    dtypes: &[DType],
    backend: BackendId,
    op_params: OpParams,
    bindings: &KernelBindingTable,
) -> Result<CompiledNode> {
    let (kernel, caps) = bindings.lookup_with_caps(op, dtypes, backend)?;
    Ok(CompiledNode {
        op,
        dtypes: KernelDTypes::from_slice(dtypes),
        backend,
        kernel,
        caps,
        op_params,
    })
}

/// The `(kernel_source, kernel_revision_hash)` identity of the kernel the
/// executor dispatches for `node`, derived from the SAME graph stamp +
/// registry the executor resolves through: `op_to_op_kind` +
/// `build_lookup_dtypes` + `graph.target_backend` → the first-registered
/// binding at `(op, dtypes, backend)` (the entry [`compile_node`]'s
/// `lookup_with_caps` picks). `None` when the node has no dispatch mapping
/// (view/structural op) or no registered binding.
///
/// Both fields come from the ONE `find()` — the same
/// [`BindingEntry`](crate::kernel::BindingEntry) — so
/// the tag and the revision describe the same kernel by construction; a
/// second lookup keyed on the tag could not disambiguate variant siblings
/// that share it, which is precisely the collapse the revision exists to
/// prevent. This is the post-realize attribution the bridge reports for the
/// Judge. Step D moved it off the plan's `AlternativeSet::winner()`: the
/// production realize path dispatches via the binding-table lookup (no plan),
/// so the first-registered binding IS the matching attribution.
pub fn dispatched_kernel_ident(
    graph: &Graph,
    node: NodeId,
    bindings: &KernelBindingTable,
) -> Option<(&'static str, u64)> {
    let n = graph.node(node);
    let op = crate::pipelined::op_to_op_kind(&n.op)?;
    let dtypes = crate::pipelined::build_lookup_dtypes(graph, n);
    let backend = graph.target_backend(node)?;
    // Mirror `lookup_with_caps`'s dense-safe pick: skip baked-broadcast
    // siblings so the attribution matches the kernel the executor actually
    // dispatches (path 1a).
    bindings
        .lookup_alternatives(op, &dtypes, backend)
        .iter()
        .find(|e| !e.caps.requires_broadcast)
        .map(|e| (e.kernel_source, e.kernel_revision_hash))
}

/// Run a compiled node against the given inputs/outputs. The output
/// `Storage`s must be pre-allocated (the executor's responsibility,
/// using the node's shape + dtype).
///
/// `layouts` parallels `inputs.append(outputs)` and carries the
/// stride/shape/offset metadata that stride-aware kernels consume.
/// Today's contiguous-only kernels ignore it; the executor's
/// auto-Contiguize pass guarantees every input layout is
/// `Layout::contiguous(shape)` before this is called.
///
/// Production-correct: surfaces kernel errors as `Result`; never
/// panics on dispatch mismatch (the wrapper functions handle that).
pub fn execute_compiled(
    compiled: &CompiledNode,
    inputs: &[Arc<RwLock<Storage>>],
    outputs: &mut [Arc<RwLock<Storage>>],
    layouts: &[Layout],
) -> Result<CompletionHandle> {
    // Conservative default: `true` reproduces the pre-"event only where
    // waited" behavior byte-for-byte (every CUDA node gets an Event). Callers
    // that know a node will never be individually waited mid-walk (the
    // `pipelined` executor's wait-set pre-scan) should call
    // [`execute_compiled_with_wait_hint`] instead.
    execute_compiled_with_wait_hint(compiled, inputs, outputs, layouts, true)
}

/// Step E A4b open question #4 — "event only where waited". Identical to
/// [`execute_compiled`] except the caller supplies `will_be_waited`: whether
/// ANY later step in this realize will individually wait this node's
/// completion handle before the realize-end drain (today: the cross-device
/// `Copy`/`Move` producer wait, `wait_producer_handle` in `pipelined.rs`).
///
/// `will_be_waited == false` lets `produce_pending` skip the CUDA
/// `Event`/in-flight-counter machinery entirely and return
/// [`CompletionHandle::Ready`] immediately — the kernel is still launched on
/// the device stream exactly as before; only the completion-observation
/// bookkeeping is elided. Same-device consumers stay correct because CUDA has
/// one stream per device: stream order alone guarantees a later same-device
/// kernel doesn't start before this one's writes land, with no host-side wait
/// required. Cutting the event is safe PROVIDED the caller's wait-set is
/// conservative (a producer that turns out to be waited must always pass
/// `true`) — `pipelined.rs`'s pre-scan is exactly that: built from the SAME
/// Copy/Move nodes the mid-walk wait consults.
///
/// `will_be_waited == true` is byte-identical to the pre-elision path.
pub fn execute_compiled_with_wait_hint(
    compiled: &CompiledNode,
    inputs: &[Arc<RwLock<Storage>>],
    outputs: &mut [Arc<RwLock<Storage>>],
    layouts: &[Layout],
    will_be_waited: bool,
) -> Result<CompletionHandle> {
    let profile_start = op_profile::enabled().then(std::time::Instant::now);
    let launched = (compiled.kernel)(inputs, outputs, layouts, &compiled.op_params);
    if let Some(t0) = profile_start {
        op_profile::record_launch(compiled, inputs, outputs, t0.elapsed());
    }
    launched?;
    // Step E A4b-1: the kernel has *enqueued* its work (CUDA: launched on the
    // device stream and returned without a per-op sync, per A3; CPU/Vulkan: see
    // below). Produce a completion handle FROM THE OUTPUT STORAGE — the kernel
    // reaches the device via the `Storage`, not an executor-held backend, so the
    // output's `BackendStorage` is where we find the device/stream to record a
    // signal on. The executor stores this handle and defers the wait to a
    // dependency / realize-end boundary instead of `wait`ing inline here.
    produce_pending(compiled.backend, outputs, will_be_waited)
}

/// Step E A4b-1: build the [`CompletionHandle`] for a just-launched kernel by
/// inspecting `outputs[0]`'s backend storage.
///
/// - **CUDA, `will_be_waited`** → [`CompletionHandle::Pending`] wrapping an
///   [`Event`] recorded on the output's device stream (signals when this
///   kernel + all prior stream work completes; one stream per device makes a
///   per-node event a sufficient stream marker — A4b-1 §1.2).
/// - **CUDA, NOT `will_be_waited`** (A4b open question #4 — "event only
///   where waited") → [`CompletionHandle::Ready`], with NO `cuEventCreate` /
///   `cuEventRecord` and NO in-flight-counter increment. The kernel already
///   launched (line above); only the completion-observation bookkeeping is
///   skipped, because nothing will ever individually wait THIS node's
///   handle — the realize-end drain now does one blanket
///   `cuStreamSynchronize` per active CUDA device instead
///   (`pipelined::sync_active_cuda_devices`), which flushes elided work too
///   (single stream per device ⇒ stream order already makes it complete by
///   then).
/// - **CPU** → [`CompletionHandle::Ready`] (the kernel computed in-process; the
///   work is already done).
/// - **Vulkan** → [`CompletionHandle::Ready`]. The per-op compute path keeps
///   RECORDING into the deferred batch and returns `Ready` during the walk (no
///   per-op handle, no eager submit — eager submission at sub-DAG boundaries is
///   A4b-4). A4b-2's async split lives at the executor's realize-end drain
///   (`drain_vulkan_pending`): it `submit_pending`s the open batch (the per-batch
///   fence) then waits it through a `VulkanCompletion` handle — byte-identical to
///   A2's atomic submit+wait, just split. The A2 lazy-batch model still carries
///   intra-walk Vulkan correctness (same-queue order + the eviction `force_flush`).
///
/// Multi-output kernels share one backing buffer (one `BackendStorage`), so
/// `outputs[0]` is the device handle for every slot — a single event covers them.
fn produce_pending(
    _backend: BackendId,
    outputs: &[Arc<RwLock<Storage>>],
    will_be_waited: bool,
) -> Result<CompletionHandle> {
    // `_backend` (= `compiled.backend`) records the *intent*; we read the
    // device truth off the output storage's variant instead, which is robust
    // even where the backend stamp and the realized storage could ever diverge.
    #[cfg(not(feature = "cuda"))]
    let _ = (outputs, will_be_waited);
    #[cfg(feature = "cuda")]
    {
        if will_be_waited {
            if let Some(out0) = outputs.first() {
                let guard = out0.read().map_err(|_| {
                    Error::Msg("output storage lock poisoned recording completion".into()).bt()
                })?;
                if let fuel_memory::BackendStorage::Cuda(cuda_bytes) = &guard.inner {
                    // Step E B1: this is the single CUDA async-submit site — the kernel
                    // was just enqueued on the device stream. The completion handle's
                    // lifetime IS the in-flight window, so `CudaCompletion::new`
                    // increments the per-device in-flight counter here and the handle's
                    // `Drop` decrements it (covering wait, drain, AND the eviction-drop
                    // path uniformly — see the type docs). The device location keys the
                    // counter.
                    let loc = cuda_bytes.device().location();
                    let ev = cuda_bytes.record_completion_event()?;
                    return Ok(CompletionHandle::Pending(Box::new(CudaCompletion::new(
                        ev, loc,
                    ))));
                }
            }
        }
    }
    Ok(CompletionHandle::Ready)
}

/// Step E A4b-1: a CUDA [`Completion`] over a recorded [`Event`]. `wait` is a
/// blocking host wait (`cuEventSynchronize`) on the event — which fires after
/// the producing kernel and every prior op on the (single per-device) stream.
///
/// Step E B1: carries the producing op's [`DeviceLocation`] and owns one slot
/// of the process-wide in-flight counter. [`CudaCompletion::new`] (called at the
/// async-submit site, `produce_pending`) increments; [`Drop`] decrements. Putting
/// the decrement in `Drop` — NOT in `wait` — is what makes the counter BALANCE:
/// every path that ends this handle's life decrements exactly once and only once:
///   - `wait(self: Box<Self>)` synchronizes the event (a `&self` read) then the
///     `Box` drops → one dec;
///   - the executor's `drain_handles` / `wait_producer_handle` consume the handle
///     via `wait` → the same one dec;
///   - the eviction path `handles.remove(&destroyed)` DROPS the handle WITHOUT
///     waiting → still one dec (via `Drop`);
///   - an error (`?`) unwinding the realize loop drops the still-owned handles →
///     one dec each.
/// `wait` deliberately does not touch the counter, so no path double-counts.
#[cfg(feature = "cuda")]
struct CudaCompletion {
    ev: fuel_cuda_backend::Event,
    loc: fuel_ir::DeviceLocation,
}

#[cfg(feature = "cuda")]
impl CudaCompletion {
    /// Build the handle for a just-enqueued CUDA op and register it as one
    /// in-flight op on `loc`. The matching decrement is in [`Drop`].
    fn new(ev: fuel_cuda_backend::Event, loc: fuel_ir::DeviceLocation) -> Self {
        crate::dispatch::inflight_inc(loc);
        Self { ev, loc }
    }
}

#[cfg(feature = "cuda")]
impl Drop for CudaCompletion {
    fn drop(&mut self) {
        // The single in-flight decrement — fires on wait-consume, drain, OR
        // eviction-drop, so the counter returns to baseline once every handle
        // is gone. Relaxed + saturating (a hint, never a correctness gate).
        crate::dispatch::inflight_dec(self.loc);
    }
}

#[cfg(feature = "cuda")]
impl Completion for CudaCompletion {
    fn wait(self: Box<Self>) -> Result<()> {
        use fuel_cuda_backend::WrapErr;
        // cuEventSynchronize: blocks the host until the event completes; surfaces
        // a sticky async kernel fault on this stream as the mapped `Err`. The
        // in-flight decrement is NOT here — it rides the `Box`'s `Drop` (which
        // runs as this method returns), so the eviction-drop path also decrements.
        self.ev.synchronize().w()
    }
}

/// A node's completion signal, returned by [`execute_compiled`]. Step E
/// Phase A1 foundation: today every dispatch is synchronous, so this is always
/// [`CompletionHandle::Ready`] (the work finished before the call returned) and
/// callers `wait` it immediately — byte-identical to the prior `Result<()>`.
/// Phases A2/A3 add an async GPU path that returns [`CompletionHandle::Pending`]
/// carrying a device fence/event the executor defers the wait on.
///
/// `#[must_use]`: dropping a `Pending` handle without `wait`ing would leak
/// un-awaited async work (harmless for today's `Ready`, a correctness bug once
/// A2/A3 land) — callers must `wait` it or thread it to the executor.
#[must_use = "a CompletionHandle must be waited on (or threaded to the executor) — dropping a Pending handle leaks async work"]
pub enum CompletionHandle {
    /// The work finished synchronously before [`execute_compiled`] returned —
    /// [`wait`](CompletionHandle::wait) is a no-op.
    Ready,
    /// The work was enqueued asynchronously; [`wait`](CompletionHandle::wait)
    /// blocks until the carried device signal fires. (No producers until A2/A3.)
    Pending(Box<dyn Completion>),
}

impl CompletionHandle {
    /// Block until this node's work has finished. No-op for [`Ready`].
    ///
    /// [`Ready`]: CompletionHandle::Ready
    pub fn wait(self) -> Result<()> {
        match self {
            CompletionHandle::Ready => Ok(()),
            CompletionHandle::Pending(c) => c.wait(),
        }
    }
}

/// An async-dispatch completion signal — a backend's wrapper over a CUDA event
/// / Vulkan fence so the executor can defer the wait past the submit point
/// (Step E Phase A2/A3). No implementors today; all dispatch is synchronous.
pub trait Completion: Send {
    /// Block until the enqueued work this handle represents has finished.
    fn wait(self: Box<Self>) -> Result<()>;
}

#[cfg(test)]
mod op_profile_tests {
    use super::op_profile;
    use fuel_ir::dispatch::OpKind;
    use fuel_ir::probe::BackendId;
    use std::time::Duration;

    /// The only test that touches the process-global totals.
    #[test]
    fn totals_sum_per_key_sort_longest_first_and_drain() {
        let _ = op_profile::take();
        op_profile::record(
            OpKind::AddElementwise,
            BackendId::Cpu,
            Duration::from_millis(2),
        );
        op_profile::record(
            OpKind::AddElementwise,
            BackendId::Cpu,
            Duration::from_millis(3),
        );
        op_profile::record(OpKind::MatMul, BackendId::Cpu, Duration::from_millis(1));
        op_profile::record(OpKind::MatMul, BackendId::Vulkan, Duration::from_millis(9));
        let rows = op_profile::take();
        let got: Vec<(&str, &str, u64, u128)> = rows
            .iter()
            .map(|(o, b, n, d)| (o.as_str(), b.as_str(), *n, d.as_millis()))
            .collect();
        assert_eq!(
            got,
            vec![
                ("MatMul", "Vulkan", 1, 9),
                ("AddElementwise", "Cpu", 2, 5),
                ("MatMul", "Cpu", 1, 1),
            ]
        );
        assert!(op_profile::take().is_empty(), "take() drains");
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::dispatch::register_cpu_kernels;

    /// E2E with the compile/execute split: register CPU kernels,
    /// compile a node, execute it. Same shape as B5's e2e test but
    /// goes through the CompiledNode interface.
    #[test]
    fn compile_then_execute_add_f32() {
        let mut bindings = KernelBindingTable::new();
        register_cpu_kernels(&mut bindings);

        let compiled = compile_node(
            OpKind::AddElementwise,
            &[DType::F32, DType::F32, DType::F32],
            BackendId::Cpu,
            OpParams::None,
            &bindings,
        )
        .expect("compile");

        assert_eq!(compiled.op, OpKind::AddElementwise);
        assert_eq!(compiled.output_dtype(), DType::F32);
        assert_eq!(compiled.backend, BackendId::Cpu);

        let lhs = fuel_memory::from_slice_cpu(&[5.0_f32, 6.0, 7.0]);
        let rhs = fuel_memory::from_slice_cpu(&[1.0_f32, 2.0, 3.0]);
        let out = fuel_memory::alloc_cpu_zeroed(DType::F32, 3).unwrap();
        let inputs = vec![Arc::new(RwLock::new(lhs)), Arc::new(RwLock::new(rhs))];
        let mut outputs = vec![Arc::new(RwLock::new(out))];

        let layout_3 = Layout::contiguous(fuel_ir::Shape::from(vec![3]));
        let layouts = vec![layout_3.clone(), layout_3.clone(), layout_3];
        execute_compiled(&compiled, &inputs, &mut outputs, &layouts)
            .expect("execute")
            .wait()
            .expect("wait");

        let out_guard = outputs[0].read().unwrap();
        if let fuel_memory::BackendStorage::Cpu(c) = &out_guard.inner {
            let typed: &[f32] = c.as_slice().unwrap();
            assert_eq!(typed, &[6.0, 8.0, 10.0]);
        }
    }

    /// Step E A4b-1 behavior-preservation: a CPU kernel's `execute_compiled`
    /// returns `CompletionHandle::Ready` — `produce_pending` must NOT turn the
    /// (synchronous, in-process) CPU path into a `Pending` handle. The CUDA arm
    /// (Pending) is covered by the live `cuda_async_realize_live` gate.
    #[test]
    fn execute_compiled_cpu_is_ready_not_pending() {
        let mut bindings = KernelBindingTable::new();
        register_cpu_kernels(&mut bindings);
        let compiled = compile_node(
            OpKind::AddElementwise,
            &[DType::F32, DType::F32, DType::F32],
            BackendId::Cpu,
            OpParams::None,
            &bindings,
        )
        .expect("compile");

        let lhs = fuel_memory::from_slice_cpu(&[1.0_f32, 2.0]);
        let rhs = fuel_memory::from_slice_cpu(&[3.0_f32, 4.0]);
        let out = fuel_memory::alloc_cpu_zeroed(DType::F32, 2).unwrap();
        let inputs = vec![Arc::new(RwLock::new(lhs)), Arc::new(RwLock::new(rhs))];
        let mut outputs = vec![Arc::new(RwLock::new(out))];
        let l = Layout::contiguous(fuel_ir::Shape::from(vec![2]));
        let layouts = vec![l.clone(), l.clone(), l];

        let handle = execute_compiled(&compiled, &inputs, &mut outputs, &layouts).expect("execute");
        assert!(
            matches!(handle, CompletionHandle::Ready),
            "CPU execute_compiled must return Ready (A4b-1 behavior-preserving)",
        );
        handle.wait().expect("wait");
    }

    /// A4b open question #4 — "event only where waited": the wait-hint entry
    /// point computes and returns Ready on CPU under BOTH hints (the hint only
    /// gates the CUDA event machinery), and the elided (`false`) hint must not
    /// change the computed bytes.
    #[test]
    fn execute_compiled_with_wait_hint_cpu_ready_and_correct_under_both_hints() {
        let mut bindings = KernelBindingTable::new();
        register_cpu_kernels(&mut bindings);
        let compiled = compile_node(
            OpKind::AddElementwise,
            &[DType::F32, DType::F32, DType::F32],
            BackendId::Cpu,
            OpParams::None,
            &bindings,
        )
        .expect("compile");

        for hint in [true, false] {
            let lhs = fuel_memory::from_slice_cpu(&[1.0_f32, 2.0]);
            let rhs = fuel_memory::from_slice_cpu(&[3.0_f32, 4.0]);
            let out = fuel_memory::alloc_cpu_zeroed(DType::F32, 2).unwrap();
            let inputs = vec![Arc::new(RwLock::new(lhs)), Arc::new(RwLock::new(rhs))];
            let mut outputs = vec![Arc::new(RwLock::new(out))];
            let l = Layout::contiguous(fuel_ir::Shape::from(vec![2]));
            let layouts = vec![l.clone(), l.clone(), l];

            let handle =
                execute_compiled_with_wait_hint(&compiled, &inputs, &mut outputs, &layouts, hint)
                    .expect("execute");
            assert!(
                matches!(handle, CompletionHandle::Ready),
                "CPU path returns Ready regardless of the wait hint ({hint})",
            );
            handle.wait().expect("wait");

            let out_guard = outputs[0].read().unwrap();
            if let fuel_memory::BackendStorage::Cpu(c) = &out_guard.inner {
                let typed: &[f32] = c.as_slice().unwrap();
                assert_eq!(typed, &[4.0, 6.0], "bytes identical under hint={hint}");
            }
        }
    }

    /// Step E A1: the completion-handle contract — `Ready.wait()` is a no-op
    /// success; `Pending(c).wait()` delegates to the boxed `Completion`.
    #[test]
    fn completion_handle_ready_is_noop_pending_delegates() {
        use std::sync::Arc as StdArc;
        use std::sync::atomic::{AtomicBool, Ordering};

        CompletionHandle::Ready.wait().expect("Ready.wait() is Ok");

        struct Flag(StdArc<AtomicBool>);
        impl Completion for Flag {
            fn wait(self: Box<Self>) -> Result<()> {
                self.0.store(true, Ordering::SeqCst);
                Ok(())
            }
        }
        let waited = StdArc::new(AtomicBool::new(false));
        CompletionHandle::Pending(Box::new(Flag(waited.clone())))
            .wait()
            .expect("Pending.wait() is Ok");
        assert!(
            waited.load(Ordering::SeqCst),
            "Pending.wait() must run the boxed Completion",
        );
    }

    /// compile_node surfaces NoBackendForOp on missing binding.
    #[test]
    fn compile_node_errors_on_missing_binding() {
        let bindings = KernelBindingTable::new(); // empty
        let result = compile_node(
            OpKind::AddElementwise,
            &[DType::F32, DType::F32, DType::F32],
            BackendId::Cpu,
            OpParams::None,
            &bindings,
        );
        assert!(result.is_err());
    }
}