cudarc/driver/safe/

core.rs

use crate::driver::{
    result::{self, DriverError},
    sys::{self, CUfunc_cache_enum, CUfunction_attribute_enum},
};

use std::{
    ffi::CString,
    marker::PhantomData,
    ops::{Bound, RangeBounds},
    string::String,
    sync::atomic::{AtomicBool, AtomicU32, AtomicUsize, Ordering},
    sync::Arc,
    vec::Vec,
};

/// Represents a CUDA context on a certain device.
///
/// - [`CudaContext::new()`] retains the device's primary context.
/// - [`CudaContext::new_non_primary()`] creates an independent non-primary context.
/// - [`CudaContext::new_cig()`] creates a non-primary context with CiG (CUDA in Graphics) parameters (CUDA 12.050+).
/// - [`CudaContext::from_raw_context()`] wraps a pre-existing raw `CUcontext`.
///
/// This is the entrypoint to using any cuda calls, all objects maintain a pointer to `Arc<CudaContext>`
/// to ensure proper lifetimes.
///
/// # On thread safety
///
/// This object is thread safe and can be shared/used on multiple threads. All safe apis call
/// [CudaContext::bind_to_thread()] before doing work in a certain context.
#[derive(Debug)]
pub struct CudaContext {
    pub(crate) cu_device: sys::CUdevice,
    pub(crate) cu_ctx: sys::CUcontext,
    pub(crate) ordinal: usize,
    pub(crate) has_async_alloc: bool,
    /// Whether this wraps a primary context (true) or a non-primary context (false).
    /// Primary contexts are released via `cuDevicePrimaryCtxRelease`, while non-primary
    /// contexts are destroyed via `cuCtxDestroy_v2`.
    pub(crate) is_primary: bool,
    pub(crate) num_streams: AtomicUsize,
    pub(crate) event_tracking: AtomicBool,
    pub(crate) error_state: AtomicU32,
}

unsafe impl Send for CudaContext {}
unsafe impl Sync for CudaContext {}

impl Drop for CudaContext {
    fn drop(&mut self) {
        self.record_err(self.bind_to_thread());
        let ctx = std::mem::replace(&mut self.cu_ctx, std::ptr::null_mut());
        if !ctx.is_null() {
            if self.is_primary {
                self.record_err(unsafe { result::primary_ctx::release(self.cu_device) });
            } else {
                // Non-primary contexts (e.g., CiG) are destroyed directly.
                self.record_err(unsafe { sys::cuCtxDestroy_v2(ctx).result() });
            }
        }
    }
}

impl PartialEq for CudaContext {
    fn eq(&self, other: &Self) -> bool {
        self.cu_device == other.cu_device
            && self.cu_ctx == other.cu_ctx
            && self.ordinal == other.ordinal
    }
}
impl Eq for CudaContext {}

impl CudaContext {
    /// Creates a new context on the specified device ordinal.
    pub fn new(ordinal: usize) -> Result<Arc<Self>, DriverError> {
        result::init()?;
        let cu_device = result::device::get(ordinal as i32)?;
        let cu_ctx = unsafe { result::primary_ctx::retain(cu_device) }?;
        let has_async_alloc = unsafe {
            let memory_pools_supported = result::device::get_attribute(
                cu_device,
                sys::CUdevice_attribute_enum::CU_DEVICE_ATTRIBUTE_MEMORY_POOLS_SUPPORTED,
            )?;
            memory_pools_supported > 0
        };
        let ctx = Arc::new(CudaContext {
            cu_device,
            cu_ctx,
            ordinal,
            has_async_alloc,
            is_primary: true,
            num_streams: AtomicUsize::new(0),
            event_tracking: AtomicBool::new(true),
            error_state: AtomicU32::new(0),
        });
        ctx.bind_to_thread()?;
        Ok(ctx)
    }

    /// Creates a new non-primary CUDA context on the specified device ordinal.
    ///
    /// Unlike [`CudaContext::new()`] which retains the device's primary context,
    /// this creates an independent context via `cuCtxCreate_v4` (CUDA 12.050+)
    /// or `cuCtxCreate_v3` (CUDA 11.040–12.040). On drop, the context is
    /// destroyed via `cuCtxDestroy_v2`.
    ///
    /// `flags` controls scheduling policy and other options — use 0 for defaults
    /// (`CU_CTX_SCHED_AUTO`). See [`sys::CUctx_flags`] for available flags.
    #[cfg(any(
        feature = "cuda-11040",
        feature = "cuda-11050",
        feature = "cuda-11060",
        feature = "cuda-11070",
        feature = "cuda-11080",
        feature = "cuda-12000",
        feature = "cuda-12010",
        feature = "cuda-12020",
        feature = "cuda-12030",
        feature = "cuda-12040",
        feature = "cuda-12050",
        feature = "cuda-12060",
        feature = "cuda-12080",
        feature = "cuda-12090",
        feature = "cuda-13000",
        feature = "cuda-13010"
    ))]
    pub fn new_non_primary(ordinal: usize, flags: u32) -> Result<Arc<Self>, DriverError> {
        result::init()?;
        let cu_device = result::device::get(ordinal as i32)?;

        #[cfg(any(
            feature = "cuda-12050",
            feature = "cuda-12060",
            feature = "cuda-12080",
            feature = "cuda-12090",
            feature = "cuda-13000",
            feature = "cuda-13010"
        ))]
        let cu_ctx = unsafe { result::ctx::create_v4(std::ptr::null_mut(), flags, cu_device) }?;

        #[cfg(not(any(
            feature = "cuda-12050",
            feature = "cuda-12060",
            feature = "cuda-12080",
            feature = "cuda-12090",
            feature = "cuda-13000",
            feature = "cuda-13010"
        )))]
        let cu_ctx = unsafe { result::ctx::create_v3(flags, cu_device) }?;

        let has_async_alloc = unsafe {
            let memory_pools_supported = result::device::get_attribute(
                cu_device,
                sys::CUdevice_attribute_enum::CU_DEVICE_ATTRIBUTE_MEMORY_POOLS_SUPPORTED,
            )?;
            memory_pools_supported > 0
        };
        let ctx = Arc::new(CudaContext {
            cu_device,
            cu_ctx,
            ordinal,
            has_async_alloc,
            is_primary: false,
            num_streams: AtomicUsize::new(0),
            event_tracking: AtomicBool::new(true),
            error_state: AtomicU32::new(0),
        });
        ctx.bind_to_thread()?;
        Ok(ctx)
    }

    /// Creates a new CUDA context with CiG (CUDA in Graphics) parameters.
    ///
    /// This uses `cuCtxCreate_v4` to create a non-primary context that shares
    /// resources with a graphics API (e.g., D3D12). Requires CUDA 12.050+.
    ///
    /// `flags` controls scheduling policy and other options — use 0 for defaults.
    /// `cig_params` specifies the CiG shared data type and pointer.
    ///
    /// On drop, the context is destroyed via `cuCtxDestroy_v2`.
    #[cfg(any(
        feature = "cuda-12050",
        feature = "cuda-12060",
        feature = "cuda-12080",
        feature = "cuda-12090",
        feature = "cuda-13000",
        feature = "cuda-13010"
    ))]
    pub fn new_cig(
        ordinal: usize,
        flags: u32,
        cig_params: &mut sys::CUctxCigParam,
    ) -> Result<Arc<Self>, DriverError> {
        result::init()?;
        let cu_device = result::device::get(ordinal as i32)?;
        let mut ctx_create_params = sys::CUctxCreateParams_st {
            execAffinityParams: std::ptr::null_mut(),
            numExecAffinityParams: 0,
            cigParams: cig_params as *mut sys::CUctxCigParam,
        };
        let cu_ctx = unsafe { result::ctx::create_v4(&mut ctx_create_params, flags, cu_device) }?;
        let has_async_alloc = unsafe {
            let memory_pools_supported = result::device::get_attribute(
                cu_device,
                sys::CUdevice_attribute_enum::CU_DEVICE_ATTRIBUTE_MEMORY_POOLS_SUPPORTED,
            )?;
            memory_pools_supported > 0
        };
        let ctx = Arc::new(CudaContext {
            cu_device,
            cu_ctx,
            ordinal,
            has_async_alloc,
            is_primary: false,
            num_streams: AtomicUsize::new(0),
            event_tracking: AtomicBool::new(true),
            error_state: AtomicU32::new(0),
        });
        ctx.bind_to_thread()?;
        Ok(ctx)
    }

    /// Wrap a pre-existing raw CUcontext (e.g., a CiG context created via `cuCtxCreate_v4`).
    ///
    /// The context must already be valid and will be made current on the calling thread.
    /// On drop, calls `cuCtxDestroy_v2` instead of `cuDevicePrimaryCtxRelease`.
    ///
    /// # Safety
    ///
    /// - `cu_ctx` must be a valid CUDA context that was created (not yet destroyed).
    /// - `cu_device` must be the device the context was created for.
    /// - The caller must not destroy or release the context after calling this function;
    ///   ownership is transferred to the returned `Arc<CudaContext>`.
    pub unsafe fn from_raw_context(
        ordinal: usize,
        cu_device: sys::CUdevice,
        cu_ctx: sys::CUcontext,
    ) -> Result<Arc<Self>, DriverError> {
        let has_async_alloc = {
            let memory_pools_supported = result::device::get_attribute(
                cu_device,
                sys::CUdevice_attribute_enum::CU_DEVICE_ATTRIBUTE_MEMORY_POOLS_SUPPORTED,
            )?;
            memory_pools_supported > 0
        };
        let ctx = Arc::new(CudaContext {
            cu_device,
            cu_ctx,
            ordinal,
            has_async_alloc,
            is_primary: false,
            num_streams: AtomicUsize::new(0),
            event_tracking: AtomicBool::new(true),
            error_state: AtomicU32::new(0),
        });
        ctx.bind_to_thread()?;
        Ok(ctx)
    }

    /// Returns whether this context wraps a primary context.
    ///
    /// Primary contexts are created via `cuDevicePrimaryCtxRetain` and released on drop.
    /// Non-primary contexts (e.g., CiG) are destroyed via `cuCtxDestroy_v2` on drop.
    pub fn is_primary(&self) -> bool {
        self.is_primary
    }

    /// Returns whether this context supports asynchronous memory allocation.
    ///
    /// By default, the value of this parameter is filled by querying the
    /// `CU_DEVICE_ATTRIBUTE_MEMORY_POOLS_SUPPORTED` attribute and checking if the number of pools
    /// is greater than 0.
    /// Memory allocations performed through the default [CudaStream] will use `cuMemAllocAsync`
    /// over `cuMemAlloc` if this method returns `true`.
    pub fn has_async_alloc(&self) -> bool {
        self.has_async_alloc
    }

    /// The number of devices available.
    pub fn device_count() -> Result<i32, DriverError> {
        result::init()?;
        result::device::get_count()
    }

    /// Get the `ordinal` index of the device this is on.
    pub fn ordinal(&self) -> usize {
        self.ordinal
    }

    /// Get the name of this device.
    pub fn name(&self) -> Result<String, result::DriverError> {
        self.check_err()?;
        result::device::get_name(self.cu_device)
    }

    /// Get the UUID of this device.
    pub fn uuid(&self) -> Result<sys::CUuuid, result::DriverError> {
        self.check_err()?;
        result::device::get_uuid(self.cu_device)
    }

    /// Get the compute capability of this device as a (major,minor) tuple
    pub fn compute_capability(&self) -> Result<(i32, i32), result::DriverError> {
        self.check_err()?;
        let capability_major =
            self.attribute(sys::CUdevice_attribute::CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR)?;
        let capability_minor =
            self.attribute(sys::CUdevice_attribute::CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR)?;

        Ok((capability_major, capability_minor))
    }

    /// Get the total memory available on this device, in bytes.
    pub fn total_mem(&self) -> Result<usize, DriverError> {
        self.check_err()?;
        unsafe { result::device::total_mem(self.cu_device) }
    }

    /// Returns the free and total device memory in bytes as a `(free, total)` tuple.
    /// Note: this calls [CudaContext::bind_to_thread()] to ensure the query
    /// runs against this device's context.
    pub fn mem_get_info(&self) -> Result<(usize, usize), DriverError> {
        self.bind_to_thread()?;
        result::mem_get_info()
    }
    /// Get the underlying [sys::CUdevice] of this [CudaContext].
    ///
    /// # Safety
    /// While this function is marked as safe, actually using the
    /// returned object is unsafe.
    ///
    /// **You must not free/release the device pointer**, as it is still
    /// owned by the [CudaContext].
    pub fn cu_device(&self) -> sys::CUdevice {
        self.cu_device
    }

    /// Get the underlying [sys::CUcontext] of this [CudaContext].
    ///
    /// # Safety
    /// While this function is marked as safe, actually using the
    /// returned object is unsafe.
    ///
    /// **You must not free/release the context pointer**, as it is still
    /// owned by the [CudaContext].
    pub fn cu_ctx(&self) -> sys::CUcontext {
        self.cu_ctx
    }

    /// Binds this context to the calling thread. Calling this is key for thread safety.
    pub fn bind_to_thread(&self) -> Result<(), DriverError> {
        self.check_err()?;
        if match result::ctx::get_current()? {
            Some(curr_ctx) => curr_ctx != self.cu_ctx,
            None => true,
        } {
            unsafe { result::ctx::set_current(self.cu_ctx) }?;
        }
        Ok(())
    }

    /// Get the value of the specified attribute of the device in [CudaContext].
    pub fn attribute(&self, attrib: sys::CUdevice_attribute) -> Result<i32, result::DriverError> {
        self.check_err()?;
        unsafe { result::device::get_attribute(self.cu_device, attrib) }
    }

    /// Synchronize this context. Will only block CPU if you call [CudaContext::set_flags()] with
    /// [sys::CUctx_flags::CU_CTX_SCHED_BLOCKING_SYNC].
    pub fn synchronize(&self) -> Result<(), DriverError> {
        self.bind_to_thread()?;
        result::ctx::synchronize()
    }

    /// Ensures calls to [CudaContext::synchronize()] block the calling thread.
    ///
    /// Sets [sys::CUctx_flags::CU_CTX_SCHED_BLOCKING_SYNC]
    #[cfg(not(any(
        feature = "cuda-11040",
        feature = "cuda-11050",
        feature = "cuda-11060",
        feature = "cuda-11070",
        feature = "cuda-11080",
        feature = "cuda-12000"
    )))]
    pub fn set_blocking_synchronize(&self) -> Result<(), DriverError> {
        self.set_flags(sys::CUctx_flags::CU_CTX_SCHED_BLOCKING_SYNC)
    }

    /// Set flags for this context
    #[cfg(not(any(
        feature = "cuda-11040",
        feature = "cuda-11050",
        feature = "cuda-11060",
        feature = "cuda-11070",
        feature = "cuda-11080",
        feature = "cuda-12000"
    )))]
    pub fn set_flags(&self, flags: sys::CUctx_flags) -> Result<(), DriverError> {
        self.bind_to_thread()?;
        result::ctx::set_flags(flags)
    }

    /// Gets the value of a context limit.
    ///
    /// See [cuda docs](https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__CTX.html#group__CUDA__CTX_1g9f2d47d1745752aa16da7ed0d111b6a8)
    pub fn get_limit(&self, limit: sys::CUlimit) -> Result<usize, DriverError> {
        self.bind_to_thread()?;
        result::ctx::get_limit(limit)
    }

    /// Sets the value of a context limit.
    ///
    /// Common limits:
    /// - `CU_LIMIT_STACK_SIZE` - Stack size for each thread
    /// - `CU_LIMIT_PRINTF_FIFO_SIZE` - Size of printf buffer
    /// - `CU_LIMIT_MALLOC_HEAP_SIZE` - Heap size for malloc() in kernels
    ///
    /// See [cuda docs](https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__CTX.html#group__CUDA__CTX_1g0651954dfb9788173e60a9af7201e65a)
    pub fn set_limit(&self, limit: sys::CUlimit, value: usize) -> Result<(), DriverError> {
        self.bind_to_thread()?;
        result::ctx::set_limit(limit, value)
    }

    /// Gets the L1/shared memory cache configuration preference.
    ///
    /// See [cuda docs](https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__CTX.html#group__CUDA__CTX_1g40b6b141698f76744dea6e39b9a25360)
    pub fn get_cache_config(&self) -> Result<sys::CUfunc_cache, DriverError> {
        self.bind_to_thread()?;
        result::ctx::get_cache_config()
    }

    /// Sets the L1/shared memory cache configuration preference.
    ///
    /// Options:
    /// - `CU_FUNC_CACHE_PREFER_NONE` - No preference
    /// - `CU_FUNC_CACHE_PREFER_SHARED` - Prefer larger shared memory, smaller L1
    /// - `CU_FUNC_CACHE_PREFER_L1` - Prefer larger L1, smaller shared memory
    /// - `CU_FUNC_CACHE_PREFER_EQUAL` - Equal split between L1 and shared
    ///
    /// See [cuda docs](https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__CTX.html#group__CUDA__CTX_1g54699acf7e2ef27279d013ca2095f4a3)
    pub fn set_cache_config(&self, config: sys::CUfunc_cache) -> Result<(), DriverError> {
        self.bind_to_thread()?;
        result::ctx::set_cache_config(config)
    }

    /// Whether multiple streams have been created in this context. If so,
    /// the [CudaSlice::read] and [CudaSlice::write] events will be activated.
    ///
    /// This only get's set to true by [CudaContext::new_stream()].
    pub fn is_in_multi_stream_mode(&self) -> bool {
        self.num_streams.load(Ordering::Relaxed) > 0
    }

    /// Whether event tracking is being managed by this context
    /// (via [CudaContext::enable_event_tracking()], which is the default behavior),
    /// or `false` if the user is manually managing stream synchronization
    /// (via [CudaContext::disable_event_tracking()]).
    pub fn is_event_tracking(&self) -> bool {
        self.event_tracking.load(Ordering::Relaxed)
    }

    /// Whether the context is automatically managing multiple stream synchronization.
    /// Both of these must be true:
    /// - [CudaContext::is_in_multi_stream_mode()]
    /// - [CudaContext::is_event_tracking()]
    pub fn is_managing_stream_synchronization(&self) -> bool {
        self.is_in_multi_stream_mode() && self.is_event_tracking()
    }

    /// When turned on, all [CudaSlice] **created after calling this function** will
    /// record usages using [CudaEvent] to ensure proper synchronization between streams.
    ///
    /// # Safety
    ///
    /// If [CudaContext::disable_event_tracking()] was called previously, then any
    /// [CudaSlice] created after that and before this current call won't have [CudaEvent]
    /// tracking their uses. Those [CudaSlice] will not manage their synchronization, even
    /// after this call.
    pub unsafe fn enable_event_tracking(&self) {
        self.event_tracking.store(true, Ordering::Relaxed);
    }

    /// When turned on, all [CudaSlice] **created after calling this function** will
    /// not track uses via [CudaEvent]s.
    ///
    /// # Safety
    ///
    /// It is up to the user to ensure proper synchronization between multiple streams:
    /// - Ensure that no [CudaSlice] is freed before a use on another stream is finished.
    /// - Ensure that a [CudaSlice] is not used on another stream before allocation on the
    ///   allocating stream finishes.
    /// - Ensure that a [CudaSlice] is not written two concurrently by multiple streams.
    pub unsafe fn disable_event_tracking(&self) {
        self.event_tracking.store(false, Ordering::Relaxed);
    }

    /// Checks to see if there have been any calls that stored an Err in a function
    /// that couldn't return a result (e.g. Drop calls).
    ///
    /// If there are any errors stored, this method will return the Err value, and
    /// then clear the stored error state.
    pub fn check_err(&self) -> Result<(), DriverError> {
        let error_state = self.error_state.swap(0, Ordering::Relaxed);
        if error_state == 0 {
            Ok(())
        } else {
            Err(result::DriverError(unsafe {
                std::mem::transmute::<u32, sys::cudaError_enum>(error_state)
            }))
        }
    }

    /// Records a result for later inspection when a Result can be returned.
    pub fn record_err<T>(&self, result: Result<T, DriverError>) {
        if let Err(err) = result {
            self.error_state.store(err.0 as u32, Ordering::Relaxed)
        }
    }
}

/// A lightweight synchronization primitive used to synchronize between [CudaStream]s.
///
/// - Create using [CudaContext::new_event()].
/// - Record a point of time in a stream using [CudaEvent::record()].
/// - Either call [CudaEvent::synchronize()] or [CudaStream::wait()] to use.
///
/// Note that calls to [CudaEvent::record()] will not change any **previous calls** to [CudaStream::wait()].
///
/// # Thread safety
/// This object is thread safe
#[derive(Debug)]
pub struct CudaEvent {
    pub(crate) cu_event: sys::CUevent,
    pub(crate) ctx: Arc<CudaContext>,
}

unsafe impl Send for CudaEvent {}
unsafe impl Sync for CudaEvent {}

impl Drop for CudaEvent {
    fn drop(&mut self) {
        self.ctx.record_err(self.ctx.bind_to_thread());
        self.ctx
            .record_err(unsafe { result::event::destroy(self.cu_event) });
    }
}

impl CudaContext {
    /// Creates a new [CudaEvent] with no work recorded. If `flags` is None, the event is created with
    /// [sys::CUevent_flags::CU_EVENT_DISABLE_TIMING].
    pub fn new_event(
        self: &Arc<Self>,
        flags: Option<sys::CUevent_flags>,
    ) -> Result<CudaEvent, DriverError> {
        let flags = flags.unwrap_or(sys::CUevent_flags::CU_EVENT_DISABLE_TIMING);
        self.bind_to_thread()?;
        let cu_event = result::event::create(flags)?;
        Ok(CudaEvent {
            cu_event,
            ctx: self.clone(),
        })
    }
}

impl CudaEvent {
    /// The underlying cu_event object.
    ///
    /// # Safety
    /// Do not destroy this value
    pub fn cu_event(&self) -> sys::CUevent {
        self.cu_event
    }

    /// The context this was created in.
    pub fn context(&self) -> &Arc<CudaContext> {
        &self.ctx
    }

    /// Records the current amount of work in [CudaStream] into this event.
    ///
    /// **This does not affect any previous calls to [CudaStream::wait()]**
    ///
    /// If `stream` belongs to a different [CudaContext], this will fail with
    /// [sys::cudaError_enum::CUDA_ERROR_INVALID_CONTEXT].
    ///
    /// See [cuda docs](https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__EVENT.html#group__CUDA__EVENT_1g95424d3be52c4eb95d83861b70fb89d1)
    pub fn record(&self, stream: &CudaStream) -> Result<(), DriverError> {
        if self.ctx != stream.ctx {
            return Err(DriverError(sys::cudaError_enum::CUDA_ERROR_INVALID_CONTEXT));
        }
        self.ctx.bind_to_thread()?;
        unsafe { result::event::record(self.cu_event, stream.cu_stream) }
    }

    /// Will only block CPU thraed if [sys::CUevent_flags::CU_EVENT_BLOCKING_SYNC] was used to create this event.
    pub fn synchronize(&self) -> Result<(), DriverError> {
        self.ctx.bind_to_thread()?;
        unsafe { result::event::synchronize(self.cu_event) }
    }

    /// The time between two events. `self` is the start event, and `end` is the end event.
    /// This is effectively `end - self`.
    pub fn elapsed_ms(&self, end: &Self) -> Result<f32, DriverError> {
        if self.ctx != end.ctx {
            return Err(DriverError(sys::cudaError_enum::CUDA_ERROR_INVALID_CONTEXT));
        }
        self.ctx.bind_to_thread()?;
        self.synchronize()?;
        end.synchronize()?;
        unsafe { result::event::elapsed(self.cu_event, end.cu_event) }
    }

    /// Returns `true` if all recorded work has been completed, `false` otherwise.
    pub fn is_complete(&self) -> bool {
        unsafe { result::event::query(self.cu_event) }.is_ok()
    }
}

/// A wrapper around [sys::CUstream] that you can schedule work on.
///
/// - Create with [CudaContext::new_stream()], [CudaContext::default_stream()], or [CudaStream::fork()].
///
/// **Work done on this is asynchronous with respect to the host.**
///
/// See [CUDA C/C++ Streams and Concurrency](https://developer.download.nvidia.com/CUDA/training/StreamsAndConcurrencyWebinar.pdf)
/// See [3. Stream synchronization behavior](https://docs.nvidia.com/cuda/cuda-runtime-api/stream-sync-behavior.html)
/// See [6.6. Event Management](https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__EVENT.html)
/// See [Out-of-order execution](https://en.wikipedia.org/wiki/Out-of-order_execution)
/// See [Dependence analysis](https://en.wikipedia.org/wiki/Dependence_analysis)
#[derive(Debug, PartialEq, Eq)]
pub struct CudaStream {
    pub(crate) cu_stream: sys::CUstream,
    pub(crate) ctx: Arc<CudaContext>,
}

unsafe impl Send for CudaStream {}
unsafe impl Sync for CudaStream {}

impl Drop for CudaStream {
    fn drop(&mut self) {
        self.ctx.record_err(self.ctx.bind_to_thread());
        let cu_stream = std::mem::replace(&mut self.cu_stream, std::ptr::null_mut());
        if !cu_stream.is_null() && cu_stream != (0x2 as _) {
            self.ctx.num_streams.fetch_sub(1, Ordering::Relaxed);
            self.ctx
                .record_err(unsafe { result::stream::destroy(cu_stream) });
        }
    }
}

impl CudaContext {
    /// Get's the default stream for this context (the null ptr stream). Note that context's
    /// on the same device can all submit to the same default stream from separate context objects.
    pub fn default_stream(self: &Arc<Self>) -> Arc<CudaStream> {
        Arc::new(CudaStream {
            cu_stream: std::ptr::null_mut(),
            ctx: self.clone(),
        })
    }

    /// Get's the per-thread stream handle. See https://docs.nvidia.com/cuda/cuda-runtime-api/stream-sync-behavior.html#stream-sync-behavior
    pub fn per_thread_stream(self: &Arc<Self>) -> Arc<CudaStream> {
        Arc::new(CudaStream {
            // See https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html#group__CUDART__TYPES_1g7b7129befd6f52708309acafd1c46197
            cu_stream: 0x2 as _,
            ctx: self.clone(),
        })
    }

    /// Create a new [sys::CUstream_flags::CU_STREAM_NON_BLOCKING] stream.
    ///
    /// This will swap the calling context to multi stream mode [CudaContext::is_in_multi_stream_mode()].
    /// If the context is not already in multiple stream mode, then this function will also call [CudaContext::synchronize()].
    pub fn new_stream(self: &Arc<Self>) -> Result<Arc<CudaStream>, DriverError> {
        self.bind_to_thread()?;
        let prev_num_streams = self.num_streams.fetch_add(1, Ordering::Relaxed);
        if prev_num_streams == 0 && self.is_event_tracking() {
            self.synchronize()?;
        }
        let cu_stream = result::stream::create(result::stream::StreamKind::NonBlocking)?;
        Ok(Arc::new(CudaStream {
            cu_stream,
            ctx: self.clone(),
        }))
    }

    /// Create a new [sys::CUstream_flags::CU_STREAM_NON_BLOCKING] stream with the
    /// specified priority.
    ///
    /// Lower numerical values indicate higher priority. Use
    /// [`result::stream::get_priority_range`] to query the valid range.
    ///
    /// This will swap the calling context to multi stream mode
    /// [`CudaContext::is_in_multi_stream_mode()`].
    pub fn new_stream_with_priority(
        self: &Arc<Self>,
        priority: i32,
    ) -> Result<Arc<CudaStream>, DriverError> {
        self.bind_to_thread()?;
        let prev_num_streams = self.num_streams.fetch_add(1, Ordering::Relaxed);
        if prev_num_streams == 0 && self.is_event_tracking() {
            self.synchronize()?;
        }
        let cu_stream = result::stream::create_with_priority(
            result::stream::StreamKind::NonBlocking,
            priority,
        )?;
        Ok(Arc::new(CudaStream {
            cu_stream,
            ctx: self.clone(),
        }))
    }
}

impl CudaStream {
    /// Create's a new stream and then makes the new stream wait on `self`
    pub fn fork(&self) -> Result<Arc<Self>, DriverError> {
        self.ctx.bind_to_thread()?;
        self.ctx.num_streams.fetch_add(1, Ordering::Relaxed);
        let cu_stream = result::stream::create(result::stream::StreamKind::NonBlocking)?;
        let stream = Arc::new(CudaStream {
            cu_stream,
            ctx: self.ctx.clone(),
        });
        stream.join(self)?;
        Ok(stream)
    }

    /// The underlying cuda stream object
    /// # Safety
    /// Do not destroy this value.
    pub fn cu_stream(&self) -> sys::CUstream {
        self.cu_stream
    }

    /// The context the stream belongs to.
    pub fn context(&self) -> &Arc<CudaContext> {
        &self.ctx
    }

    /// Will only block CPU if you call [CudaContext::set_flags()] with
    /// [sys::CUctx_flags::CU_CTX_SCHED_BLOCKING_SYNC].
    ///
    /// See [cuda docs](https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__STREAM.html#group__CUDA__STREAM_1g15e49dd91ec15991eb7c0a741beb7dad)
    pub fn synchronize(&self) -> Result<(), DriverError> {
        self.ctx.bind_to_thread()?;
        unsafe { result::stream::synchronize(self.cu_stream) }
    }

    /// Creates a new [CudaEvent] and records the current work in the stream to the event.
    pub fn record_event(
        &self,
        flags: Option<sys::CUevent_flags>,
    ) -> Result<CudaEvent, DriverError> {
        let event = self.ctx.new_event(flags)?;
        event.record(self)?;
        Ok(event)
    }

    /// Waits for the work recorded in [CudaEvent] to be completed.
    ///
    /// You can record new work in `event` after calling this method without
    /// affecting this call.
    ///
    /// See [cuda docs](https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__STREAM.html#group__CUDA__STREAM_1g6a898b652dfc6aa1d5c8d97062618b2f)
    pub fn wait(&self, event: &CudaEvent) -> Result<(), DriverError> {
        self.ctx.bind_to_thread()?;
        unsafe {
            result::stream::wait_event(
                self.cu_stream,
                event.cu_event,
                sys::CUevent_wait_flags::CU_EVENT_WAIT_DEFAULT,
            )
        }
    }

    /// Ensures this stream waits for the current workload in `other` to complete.
    /// This is shorthand for `self.wait(other.record_event())`
    pub fn join(&self, other: &CudaStream) -> Result<(), DriverError> {
        self.wait(&other.record_event(None)?)
    }
}

/// `Vec<T>` on a cuda device. You can allocate and modify this with [CudaStream].
///
/// This object is thread safe.
#[derive(Debug)]
pub struct CudaSlice<T> {
    pub(crate) cu_device_ptr: sys::CUdeviceptr,
    pub(crate) len: usize,
    pub(crate) read: Option<CudaEvent>,
    pub(crate) write: Option<CudaEvent>,
    pub(crate) stream: Arc<CudaStream>,
    pub(crate) marker: PhantomData<*const T>,
}

unsafe impl<T> Send for CudaSlice<T> {}
unsafe impl<T> Sync for CudaSlice<T> {}

impl<T> Drop for CudaSlice<T> {
    fn drop(&mut self) {
        let ctx = &self.stream.ctx;
        if let Some(read) = self.read.as_ref() {
            ctx.record_err(self.stream.wait(read));
        }
        if let Some(write) = self.write.as_ref() {
            ctx.record_err(self.stream.wait(write));
        }
        if ctx.has_async_alloc {
            ctx.record_err(unsafe {
                result::free_async(self.cu_device_ptr, self.stream.cu_stream)
            });
        } else {
            ctx.record_err(self.stream.synchronize());
            ctx.record_err(unsafe { result::free_sync(self.cu_device_ptr) });
        }
    }
}

impl<T> CudaSlice<T> {
    /// The number of elements of `T` in this object.
    pub fn len(&self) -> usize {
        self.len
    }

    /// The number of bytes in this object.
    pub fn num_bytes(&self) -> usize {
        self.len * std::mem::size_of::<T>()
    }

    /// True if there are no elements in the object.
    pub fn is_empty(&self) -> bool {
        self.len == 0
    }

    /// The device ordinal this belongs to
    pub fn ordinal(&self) -> usize {
        self.stream.ctx.ordinal
    }

    /// The context this belongs to
    pub fn context(&self) -> &Arc<CudaContext> {
        &self.stream.ctx
    }

    /// The stream this object was allocated on and later will be dropped on.
    pub fn stream(&self) -> &Arc<CudaStream> {
        &self.stream
    }
}

impl<T: DeviceRepr> CudaSlice<T> {
    /// Allocates copy of self and schedules a device to device copy of memory.
    pub fn try_clone(&self) -> Result<Self, result::DriverError> {
        self.stream.clone_dtod(self)
    }
}

impl<T: DeviceRepr> Clone for CudaSlice<T> {
    fn clone(&self) -> Self {
        self.try_clone().unwrap()
    }
}

impl<T: Clone + Default + DeviceRepr> TryFrom<CudaSlice<T>> for Vec<T> {
    type Error = result::DriverError;
    fn try_from(value: CudaSlice<T>) -> Result<Self, Self::Error> {
        value.stream.clone_dtoh(&value)
    }
}

/// `&[T]` on a cuda device. An immutable sub-view into a [CudaSlice] created by [CudaSlice::as_view()]/[CudaSlice::slice()].
#[derive(Debug)]
pub struct CudaView<'a, T> {
    pub(crate) ptr: sys::CUdeviceptr,
    pub(crate) len: usize,
    pub(crate) read: &'a Option<CudaEvent>,
    pub(crate) write: &'a Option<CudaEvent>,
    pub(crate) stream: &'a Arc<CudaStream>,
    marker: PhantomData<&'a [T]>,
}

impl<T> CudaSlice<T> {
    pub fn as_view(&self) -> CudaView<'_, T> {
        CudaView {
            ptr: self.cu_device_ptr,
            len: self.len,
            read: &self.read,
            write: &self.write,
            stream: &self.stream,
            marker: PhantomData,
        }
    }
}

impl<T> CudaView<'_, T> {
    /// The number of elements `T` in this view.
    pub fn len(&self) -> usize {
        self.len
    }

    pub fn is_empty(&self) -> bool {
        self.len == 0
    }

    fn resize(&self, start: usize, end: usize) -> Self {
        assert!(start <= end && end <= self.len);
        Self {
            ptr: self.ptr + (start * std::mem::size_of::<T>()) as u64,
            len: end - start,
            read: self.read,
            write: self.write,
            stream: self.stream,
            marker: PhantomData,
        }
    }
}

/// `&mut [T]` on a cuda device. A mutable sub-view into a [CudaSlice] created by [CudaSlice::as_view_mut()]/[CudaSlice::slice_mut()].
#[derive(Debug)]
pub struct CudaViewMut<'a, T> {
    pub(crate) ptr: sys::CUdeviceptr,
    pub(crate) len: usize,
    pub(crate) read: &'a Option<CudaEvent>,
    pub(crate) write: &'a Option<CudaEvent>,
    pub(crate) stream: &'a Arc<CudaStream>,
    marker: PhantomData<&'a mut [T]>,
}

impl<T> CudaSlice<T> {
    pub fn as_view_mut(&mut self) -> CudaViewMut<'_, T> {
        CudaViewMut {
            ptr: self.cu_device_ptr,
            len: self.len,
            read: &self.read,
            write: &self.write,
            stream: &self.stream,
            marker: PhantomData,
        }
    }
}

impl<T> CudaViewMut<'_, T> {
    /// Number of elements `T` that are in this view.
    pub fn len(&self) -> usize {
        self.len
    }
    pub fn is_empty(&self) -> bool {
        self.len == 0
    }

    /// Downgrade this to a `&[T]`
    pub fn as_view<'b>(&'b self) -> CudaView<'b, T> {
        CudaView {
            ptr: self.ptr,
            len: self.len,
            read: self.read,
            write: self.write,
            stream: self.stream,
            marker: PhantomData,
        }
    }
}

/// Marker trait to indicate that the type is valid
/// when all of its bits are set to 0.
///
/// # Safety
/// Not all types are valid when all bits are set to 0.
/// Be very sure when implementing this trait!
pub unsafe trait ValidAsZeroBits {}
unsafe impl ValidAsZeroBits for bool {}
unsafe impl ValidAsZeroBits for i8 {}
unsafe impl ValidAsZeroBits for i16 {}
unsafe impl ValidAsZeroBits for i32 {}
unsafe impl ValidAsZeroBits for i64 {}
unsafe impl ValidAsZeroBits for i128 {}
unsafe impl ValidAsZeroBits for isize {}
unsafe impl ValidAsZeroBits for u8 {}
unsafe impl ValidAsZeroBits for u16 {}
unsafe impl ValidAsZeroBits for u32 {}
unsafe impl ValidAsZeroBits for u64 {}
unsafe impl ValidAsZeroBits for u128 {}
unsafe impl ValidAsZeroBits for usize {}
unsafe impl ValidAsZeroBits for f32 {}
unsafe impl ValidAsZeroBits for f64 {}
#[cfg(feature = "f16")]
unsafe impl ValidAsZeroBits for half::f16 {}
#[cfg(feature = "f16")]
unsafe impl ValidAsZeroBits for half::bf16 {}
unsafe impl<T: ValidAsZeroBits, const M: usize> ValidAsZeroBits for [T; M] {}
/// Implement `ValidAsZeroBits` for tuples if all elements are `ValidAsZeroBits`,
///
/// # Note
/// This will also implement `ValidAsZeroBits` for a tuple with one element
macro_rules! impl_tuples {
    ($t:tt) => {
        impl_tuples!(@ $t);
    };
    // the $l is in front of the reptition to prevent parsing ambiguities
    ($l:tt $(,$t:tt)+) => {
        impl_tuples!($($t),+);
        impl_tuples!(@ $l $(,$t)+);
    };
    (@ $($t:tt),+) => {
        unsafe impl<$($t: ValidAsZeroBits,)+> ValidAsZeroBits for ($($t,)+) {}
    };
}
impl_tuples!(A, B, C, D, E, F, G, H, I, J, K, L);

/// Something that can be copied to device memory and
/// turned into a parameter for [result::launch_kernel].
///
/// # Safety
///
/// This is unsafe because a struct should likely
/// be `#[repr(C)]` to be represented in cuda memory,
/// and not all types are valid.
pub unsafe trait DeviceRepr {}
unsafe impl DeviceRepr for bool {}
unsafe impl DeviceRepr for i8 {}
unsafe impl DeviceRepr for i16 {}
unsafe impl DeviceRepr for i32 {}
unsafe impl DeviceRepr for i64 {}
unsafe impl DeviceRepr for i128 {}
unsafe impl DeviceRepr for isize {}
unsafe impl DeviceRepr for u8 {}
unsafe impl DeviceRepr for u16 {}
unsafe impl DeviceRepr for u32 {}
unsafe impl DeviceRepr for u64 {}
unsafe impl DeviceRepr for u128 {}
unsafe impl DeviceRepr for usize {}
unsafe impl DeviceRepr for f32 {}
unsafe impl DeviceRepr for f64 {}
#[cfg(feature = "f16")]
unsafe impl DeviceRepr for half::f16 {}
#[cfg(feature = "f16")]
unsafe impl DeviceRepr for half::bf16 {}

#[cfg(feature = "f8")]
unsafe impl DeviceRepr for float8::F8E4M3 {}
#[cfg(feature = "f8")]
unsafe impl ValidAsZeroBits for float8::F8E4M3 {}

#[cfg(feature = "f8")]
unsafe impl DeviceRepr for float8::F8E5M2 {}
#[cfg(feature = "f8")]
unsafe impl ValidAsZeroBits for float8::F8E5M2 {}

#[cfg(feature = "f4")]
unsafe impl DeviceRepr for float4::F4E2M1 {}
#[cfg(feature = "f4")]
unsafe impl ValidAsZeroBits for float4::F4E2M1 {}

#[cfg(feature = "f4")]
unsafe impl DeviceRepr for float4::E8M0 {}
#[cfg(feature = "f4")]
unsafe impl ValidAsZeroBits for float4::E8M0 {}

#[cfg(feature = "f4")]
unsafe impl DeviceRepr for float4::F4E2M1x2 {}
#[cfg(feature = "f4")]
unsafe impl ValidAsZeroBits for float4::F4E2M1x2 {}

unsafe impl<const N: usize, T> DeviceRepr for [T; N] where T: DeviceRepr {}

/// Base trait for abstracting over [CudaSlice]/[CudaView]/[CudaViewMut].
///
/// Don't use this directly - use [DevicePtr]/[DevicePtrMut].
pub trait DeviceSlice<T> {
    fn len(&self) -> usize;
    fn num_bytes(&self) -> usize {
        self.len() * std::mem::size_of::<T>()
    }
    fn is_empty(&self) -> bool {
        self.len() == 0
    }
    fn stream(&self) -> &Arc<CudaStream>;
}

impl<T> DeviceSlice<T> for CudaSlice<T> {
    fn len(&self) -> usize {
        self.len
    }
    fn stream(&self) -> &Arc<CudaStream> {
        &self.stream
    }
}

impl<T> DeviceSlice<T> for CudaView<'_, T> {
    fn len(&self) -> usize {
        self.len
    }
    fn stream(&self) -> &Arc<CudaStream> {
        self.stream
    }
}

impl<T> DeviceSlice<T> for CudaViewMut<'_, T> {
    fn len(&self) -> usize {
        self.len
    }
    fn stream(&self) -> &Arc<CudaStream> {
        self.stream
    }
}

/// A synchronization primitive to enable stream & event synchronization.
/// Primarily used with [DevicePtr] and [DevicePtrMut]
#[derive(Debug)]
#[must_use]
pub enum SyncOnDrop<'a> {
    /// Will record the stream's workload to the event on drop.
    Record(Option<(&'a CudaEvent, &'a CudaStream)>),
    /// Will call stream synchronize on drop.
    Sync(Option<&'a CudaStream>),
}

impl<'a> SyncOnDrop<'a> {
    /// Construct a [SyncOnDrop::Record] variant
    pub fn record_event(event: &'a Option<CudaEvent>, stream: &'a CudaStream) -> Self {
        SyncOnDrop::Record(event.as_ref().map(|e| (e, stream)))
    }
    /// Construct a [SyncOnDrop::Sync] variant
    pub fn sync_stream(stream: &'a CudaStream) -> Self {
        SyncOnDrop::Sync(Some(stream))
    }
}

impl Drop for SyncOnDrop<'_> {
    fn drop(&mut self) {
        match self {
            SyncOnDrop::Record(target) => {
                if let Some((event, stream)) = std::mem::take(target) {
                    stream.ctx.record_err(event.record(stream));
                }
            }
            SyncOnDrop::Sync(target) => {
                if let Some(stream) = std::mem::take(target) {
                    stream.ctx.record_err(stream.synchronize());
                }
            }
        }
    }
}

/// Abstraction over [CudaSlice]/[CudaView]
pub trait DevicePtr<T>: DeviceSlice<T> {
    /// Retrieve the device pointer with the intent to read the device memory
    /// associated with it.
    ///
    /// Implementations of this method should ensure `stream` waits for any previous
    /// writes of this memory before continuing (do not need to wait for any previous reads).
    ///
    /// The [SyncOnDrop] item of the return tuple should be dropped **after** the read of
    /// the [sys::CUdeviceptr] is scheduled.
    ///
    /// In most cases you can use like:
    /// ```ignore
    /// let (src, _record_src) = src.device_ptr(&stream);
    /// ```
    /// Which will drop the [SyncOnDrop] at the end of the scope.
    fn device_ptr<'a>(&'a self, stream: &'a CudaStream) -> (sys::CUdeviceptr, SyncOnDrop<'a>);
}

impl<T> DevicePtr<T> for CudaSlice<T> {
    fn device_ptr<'a>(&'a self, stream: &'a CudaStream) -> (sys::CUdeviceptr, SyncOnDrop<'a>) {
        if self.stream.context().is_managing_stream_synchronization() {
            if let Some(write) = self.write.as_ref() {
                stream.ctx.record_err(stream.wait(write));
            }
        }
        (
            self.cu_device_ptr,
            SyncOnDrop::record_event(&self.read, stream),
        )
    }
}

impl<T> DevicePtr<T> for CudaView<'_, T> {
    fn device_ptr<'a>(&'a self, stream: &'a CudaStream) -> (sys::CUdeviceptr, SyncOnDrop<'a>) {
        if self.stream.context().is_managing_stream_synchronization() {
            if let Some(write) = self.write.as_ref() {
                stream.ctx.record_err(stream.wait(write));
            }
        }
        (self.ptr, SyncOnDrop::record_event(self.read, stream))
    }
}

impl<'a, T> CudaView<'a, T> {
    /// Identical behavior to [DevicePtr::device_ptr()], but the lifetime on the returned
    /// [SyncOnDrop], matches the lifetime of the view.
    pub fn view_ptr(self, stream: &'a CudaStream) -> (sys::CUdeviceptr, SyncOnDrop<'a>) {
        if self.stream.context().is_managing_stream_synchronization() {
            if let Some(write) = self.write.as_ref() {
                stream.ctx.record_err(stream.wait(write));
            }
        }
        (self.ptr, SyncOnDrop::record_event(self.read, stream))
    }
}

impl<T> DevicePtr<T> for CudaViewMut<'_, T> {
    fn device_ptr<'a>(&'a self, stream: &'a CudaStream) -> (sys::CUdeviceptr, SyncOnDrop<'a>) {
        if self.stream.context().is_managing_stream_synchronization() {
            if let Some(write) = self.write.as_ref() {
                stream.ctx.record_err(stream.wait(write));
            }
        }
        (self.ptr, SyncOnDrop::record_event(self.read, stream))
    }
}

impl<'a, T> CudaViewMut<'a, T> {
    /// Identical behavior to [DevicePtr::device_ptr()], but the lifetime on the returned
    /// [SyncOnDrop], matches the lifetime of the view.
    pub fn view_ptr(self, stream: &'a CudaStream) -> (sys::CUdeviceptr, SyncOnDrop<'a>) {
        if self.stream.context().is_managing_stream_synchronization() {
            if let Some(write) = self.write.as_ref() {
                stream.ctx.record_err(stream.wait(write));
            }
        }
        (self.ptr, SyncOnDrop::record_event(self.read, stream))
    }
}

/// Abstraction over [CudaSlice]/[CudaViewMut]
pub trait DevicePtrMut<T>: DeviceSlice<T> {
    /// Retrieve the device pointer with the intent to modify the device memory
    /// associated with it.
    ///
    /// Implementations of this method should ensure `stream` waits for any previous
    /// reads/writes of this memory before continuing.
    ///
    /// The [SyncOnDrop] item of the return tuple should be dropped **after** the write of
    /// the [sys::CUdeviceptr] is scheduled.
    ///
    /// In most cases you can use like:
    /// ```ignore
    /// let (src, _record_src) = src.device_ptr_mut(&stream);
    /// ```
    /// Which will drop the [SyncOnDrop] at the end of the scope.
    fn device_ptr_mut<'a>(
        &'a mut self,
        stream: &'a CudaStream,
    ) -> (sys::CUdeviceptr, SyncOnDrop<'a>);
}

impl<T> DevicePtrMut<T> for CudaSlice<T> {
    fn device_ptr_mut<'a>(
        &'a mut self,
        stream: &'a CudaStream,
    ) -> (sys::CUdeviceptr, SyncOnDrop<'a>) {
        if self.stream.context().is_managing_stream_synchronization() {
            if let Some(read) = self.read.as_ref() {
                stream.ctx.record_err(stream.wait(read));
            }
            if let Some(write) = self.write.as_ref() {
                stream.ctx.record_err(stream.wait(write));
            }
        }
        (
            self.cu_device_ptr,
            SyncOnDrop::record_event(&self.write, stream),
        )
    }
}

impl<T> DevicePtrMut<T> for CudaViewMut<'_, T> {
    fn device_ptr_mut<'a>(
        &'a mut self,
        stream: &'a CudaStream,
    ) -> (sys::CUdeviceptr, SyncOnDrop<'a>) {
        if self.stream.context().is_managing_stream_synchronization() {
            if let Some(read) = self.read.as_ref() {
                stream.ctx.record_err(stream.wait(read));
            }
            if let Some(write) = self.write.as_ref() {
                stream.ctx.record_err(stream.wait(write));
            }
        }
        (self.ptr, SyncOnDrop::record_event(self.write, stream))
    }
}

impl<'a, T> CudaViewMut<'a, T> {
    /// Identical behavior to [DevicePtrMut::device_ptr_mut()], but the lifetime on the returned
    /// [SyncOnDrop], matches the lifetime of the view.
    pub fn view_ptr_mut(self, stream: &'a CudaStream) -> (sys::CUdeviceptr, SyncOnDrop<'a>) {
        if self.stream.context().is_managing_stream_synchronization() {
            if let Some(read) = self.read.as_ref() {
                stream.ctx.record_err(stream.wait(read));
            }
            if let Some(write) = self.write.as_ref() {
                stream.ctx.record_err(stream.wait(write));
            }
        }
        (self.ptr, SyncOnDrop::record_event(self.write, stream))
    }
}

/// Abstraction over `&[T]`, `&Vec<T>` and [`PinnedHostSlice<T>`].
pub trait HostSlice<T> {
    fn len(&self) -> usize;
    fn is_empty(&self) -> bool {
        self.len() == 0
    }

    /// # Safety
    /// This is **only** safe if the resulting slice is used with `stream`. Otherwise
    /// You may run into device synchronization errors
    unsafe fn stream_synced_slice<'a>(
        &'a self,
        stream: &'a CudaStream,
    ) -> (&'a [T], SyncOnDrop<'a>);

    /// # Safety
    /// This is **only** safe if the resulting slice is used with `stream`. Otherwise
    /// You may run into device synchronization errors
    unsafe fn stream_synced_mut_slice<'a>(
        &'a mut self,
        stream: &'a CudaStream,
    ) -> (&'a mut [T], SyncOnDrop<'a>);
}

impl<T, const N: usize> HostSlice<T> for [T; N] {
    fn len(&self) -> usize {
        N
    }
    unsafe fn stream_synced_slice<'a>(
        &'a self,
        _stream: &'a CudaStream,
    ) -> (&'a [T], SyncOnDrop<'a>) {
        (self, SyncOnDrop::Sync(None))
    }
    unsafe fn stream_synced_mut_slice<'a>(
        &'a mut self,
        _stream: &'a CudaStream,
    ) -> (&'a mut [T], SyncOnDrop<'a>) {
        (self, SyncOnDrop::Sync(None))
    }
}

impl<T> HostSlice<T> for [T] {
    fn len(&self) -> usize {
        self.len()
    }
    unsafe fn stream_synced_slice<'a>(
        &'a self,
        _stream: &'a CudaStream,
    ) -> (&'a [T], SyncOnDrop<'a>) {
        (self, SyncOnDrop::Sync(None))
    }
    unsafe fn stream_synced_mut_slice<'a>(
        &'a mut self,
        _stream: &'a CudaStream,
    ) -> (&'a mut [T], SyncOnDrop<'a>) {
        (self, SyncOnDrop::Sync(None))
    }
}

impl<T> HostSlice<T> for Vec<T> {
    fn len(&self) -> usize {
        self.len()
    }
    unsafe fn stream_synced_slice<'a>(
        &'a self,
        _stream: &'a CudaStream,
    ) -> (&'a [T], SyncOnDrop<'a>) {
        (self, SyncOnDrop::Sync(None))
    }
    unsafe fn stream_synced_mut_slice<'a>(
        &'a mut self,
        _stream: &'a CudaStream,
    ) -> (&'a mut [T], SyncOnDrop<'a>) {
        (self, SyncOnDrop::Sync(None))
    }
}

/// Rust side data that the `cuda` driver knows is pinned. This is different
/// than `Pin<Vec<T>>` mainly because cuda driver manages this memory and ensures
/// it is page locked.
///
/// Allocate this with [CudaContext::alloc_pinned()], and do device copies with
/// [CudaStream::clone_htod()]/[CudaStream::memcpy_htod()]/[CudaStream::memcpy_dtoh()]
#[derive(Debug)]
pub struct PinnedHostSlice<T> {
    pub(crate) ptr: *mut T,
    pub(crate) len: usize,
    pub(crate) event: CudaEvent,
}

unsafe impl<T> Send for PinnedHostSlice<T> {}
unsafe impl<T> Sync for PinnedHostSlice<T> {}

impl<T> Drop for PinnedHostSlice<T> {
    fn drop(&mut self) {
        let ctx = &self.event.ctx;
        ctx.record_err(self.event.synchronize());
        ctx.record_err(unsafe { result::free_host(self.ptr as _) });
    }
}

impl CudaContext {
    /// Allocates page locked host memory with [sys::CU_MEMHOSTALLOC_WRITECOMBINED] flags.
    ///
    /// See [cuda docs](https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__MEM.html#group__CUDA__MEM_1g572ca4011bfcb25034888a14d4e035b9)
    ///
    /// # Safety
    /// 1. This is unsafe because the memory is unset after this call.
    pub unsafe fn alloc_pinned<T: DeviceRepr>(
        self: &Arc<Self>,
        len: usize,
    ) -> Result<PinnedHostSlice<T>, DriverError> {
        self.bind_to_thread()?;
        let ptr = result::malloc_host(
            len * std::mem::size_of::<T>(),
            sys::CU_MEMHOSTALLOC_WRITECOMBINED,
        )?;
        let ptr = ptr as *mut T;
        assert!(!ptr.is_null());
        assert!(len * std::mem::size_of::<T>() < isize::MAX as usize);
        assert!(ptr.is_aligned());
        let event = self.new_event(Some(sys::CUevent_flags::CU_EVENT_BLOCKING_SYNC))?;
        Ok(PinnedHostSlice { ptr, len, event })
    }
}

impl<T> PinnedHostSlice<T> {
    /// The context this was created in.
    pub fn context(&self) -> &Arc<CudaContext> {
        &self.event.ctx
    }

    /// The number of elements `T` in this slice.
    pub fn len(&self) -> usize {
        self.len
    }

    /// The number of bytes in this slice.
    pub fn num_bytes(&self) -> usize {
        self.len * std::mem::size_of::<T>()
    }

    pub fn is_empty(&self) -> bool {
        self.len() == 0
    }
}

impl<T: ValidAsZeroBits> PinnedHostSlice<T> {
    /// Waits for any scheduled work to complete and then returns a refernce
    /// to the host side data.
    pub fn as_ptr(&self) -> Result<*const T, DriverError> {
        self.event.synchronize()?;
        Ok(self.ptr)
    }

    /// Waits for any scheduled work to complete and then returns a refernce
    /// to the host side data.
    pub fn as_mut_ptr(&mut self) -> Result<*mut T, DriverError> {
        self.event.synchronize()?;
        Ok(self.ptr)
    }

    /// Waits for any scheduled work to complete and then returns a refernce
    /// to the host side data.
    pub fn as_slice(&self) -> Result<&[T], DriverError> {
        self.event.synchronize()?;
        Ok(unsafe { std::slice::from_raw_parts(self.ptr, self.len) })
    }

    /// Waits for any scheduled work to complete and then returns a refernce
    /// to the host side data.
    pub fn as_mut_slice(&mut self) -> Result<&mut [T], DriverError> {
        self.event.synchronize()?;
        Ok(unsafe { std::slice::from_raw_parts_mut(self.ptr, self.len) })
    }
}

impl<T> HostSlice<T> for PinnedHostSlice<T> {
    fn len(&self) -> usize {
        self.len
    }

    unsafe fn stream_synced_slice<'a>(
        &'a self,
        stream: &'a CudaStream,
    ) -> (&'a [T], SyncOnDrop<'a>) {
        stream.ctx.record_err(stream.wait(&self.event));
        (
            std::slice::from_raw_parts(self.ptr, self.len),
            SyncOnDrop::Record(Some((&self.event, stream))),
        )
    }
    unsafe fn stream_synced_mut_slice<'a>(
        &'a mut self,
        stream: &'a CudaStream,
    ) -> (&'a mut [T], SyncOnDrop<'a>) {
        stream.ctx.record_err(stream.wait(&self.event));
        (
            std::slice::from_raw_parts_mut(self.ptr, self.len),
            SyncOnDrop::Record(Some((&self.event, stream))),
        )
    }
}

impl CudaStream {
    /// Allocates an empty [CudaSlice] with 0 length.
    pub fn null<T>(self: &Arc<Self>) -> Result<CudaSlice<T>, result::DriverError> {
        self.ctx.bind_to_thread()?;
        let cu_device_ptr = if self.ctx.has_async_alloc {
            unsafe { result::malloc_async(self.cu_stream, 0) }?
        } else {
            unsafe { result::malloc_sync(0) }?
        };
        Ok(CudaSlice {
            cu_device_ptr,
            len: 0,
            read: None,
            write: None,
            stream: self.clone(),
            marker: PhantomData,
        })
    }

    /// Allocates a [CudaSlice] with `len` elements of type `T`.
    /// # Safety
    /// This is unsafe because the memory is unset.
    pub unsafe fn alloc<T: DeviceRepr>(
        self: &Arc<Self>,
        len: usize,
    ) -> Result<CudaSlice<T>, DriverError> {
        self.ctx.bind_to_thread()?;
        let cu_device_ptr = if self.ctx.has_async_alloc {
            result::malloc_async(self.cu_stream, len * std::mem::size_of::<T>())?
        } else {
            result::malloc_sync(len * std::mem::size_of::<T>())?
        };
        let (read, write) = if self.ctx.is_event_tracking() {
            (
                Some(self.ctx.new_event(None)?),
                Some(self.ctx.new_event(None)?),
            )
        } else {
            (None, None)
        };
        Ok(CudaSlice {
            cu_device_ptr,
            len,
            read,
            write,
            stream: self.clone(),
            marker: PhantomData,
        })
    }

    /// Allocates a [CudaSlice] with `len` elements of type `T`. All values are zero'd out.
    pub fn alloc_zeros<T: DeviceRepr + ValidAsZeroBits>(
        self: &Arc<Self>,
        len: usize,
    ) -> Result<CudaSlice<T>, DriverError> {
        let mut dst = unsafe { self.alloc(len) }?;
        self.memset_zeros(&mut dst)?;
        Ok(dst)
    }

    /// Set's all the memory in `dst` to 0. `dst` can be a [CudaSlice] or [CudaViewMut]
    pub fn memset_zeros<T: DeviceRepr + ValidAsZeroBits, Dst: DevicePtrMut<T>>(
        self: &Arc<Self>,
        dst: &mut Dst,
    ) -> Result<(), DriverError> {
        self.ctx.bind_to_thread()?;
        let num_bytes = dst.num_bytes();
        let (dptr, _record) = dst.device_ptr_mut(self);
        unsafe { result::memset_d8_async(dptr, 0, num_bytes, self.cu_stream) }?;
        Ok(())
    }

    /// Copy a `[T]`/`Vec<T>`/[`PinnedHostSlice<T>`] to a new [`CudaSlice`].
    #[deprecated = "Use clone_htod"]
    pub fn memcpy_stod<T: DeviceRepr, Src: HostSlice<T> + ?Sized>(
        self: &Arc<Self>,
        src: &Src,
    ) -> Result<CudaSlice<T>, DriverError> {
        let mut dst = unsafe { self.alloc(src.len()) }?;
        self.memcpy_htod(src, &mut dst)?;
        Ok(dst)
    }

    /// Copy a `[T]`/`Vec<T>`/[`PinnedHostSlice<T>`] to a new [`CudaSlice`].
    pub fn clone_htod<T: DeviceRepr, Src: HostSlice<T> + ?Sized>(
        self: &Arc<Self>,
        src: &Src,
    ) -> Result<CudaSlice<T>, DriverError> {
        let mut dst = unsafe { self.alloc(src.len()) }?;
        self.memcpy_htod(src, &mut dst)?;
        Ok(dst)
    }

    /// Copy a `[T]`/`Vec<T>`/[`PinnedHostSlice<T>`] into an existing [`CudaSlice`]/[`CudaViewMut`].
    pub fn memcpy_htod<T: DeviceRepr, Src: HostSlice<T> + ?Sized, Dst: DevicePtrMut<T>>(
        self: &Arc<Self>,
        src: &Src,
        dst: &mut Dst,
    ) -> Result<(), DriverError> {
        assert!(dst.len() >= src.len());
        self.ctx.bind_to_thread()?;
        let (src, _record_src) = unsafe { src.stream_synced_slice(self) };
        let (dst, _record_dst) = dst.device_ptr_mut(self);
        unsafe { result::memcpy_htod_async(dst, src, self.cu_stream) }
    }

    /// Copy a [`CudaSlice`]/[`CudaView`] to a new [`Vec<T>`].
    #[deprecated = "Use clone_dtoh"]
    pub fn memcpy_dtov<T: DeviceRepr, Src: DevicePtr<T>>(
        self: &Arc<Self>,
        src: &Src,
    ) -> Result<Vec<T>, DriverError> {
        let mut dst = Vec::with_capacity(src.len());
        #[allow(clippy::uninit_vec)]
        unsafe {
            dst.set_len(src.len())
        };
        self.memcpy_dtoh(src, &mut dst)?;
        Ok(dst)
    }

    /// Copy a [`CudaSlice`]/[`CudaView`] to a new [`Vec<T>`].
    pub fn clone_dtoh<T: DeviceRepr, Src: DevicePtr<T>>(
        self: &Arc<Self>,
        src: &Src,
    ) -> Result<Vec<T>, DriverError> {
        let mut dst = Vec::with_capacity(src.len());
        #[allow(clippy::uninit_vec)]
        unsafe {
            dst.set_len(src.len())
        };
        self.memcpy_dtoh(src, &mut dst)?;
        Ok(dst)
    }

    /// Copy a [`CudaSlice`]/[`CudaView`] to a existing `[T]`/[`Vec<T>`]/[`PinnedHostSlice<T>`].
    pub fn memcpy_dtoh<T: DeviceRepr, Src: DevicePtr<T>, Dst: HostSlice<T> + ?Sized>(
        self: &Arc<Self>,
        src: &Src,
        dst: &mut Dst,
    ) -> Result<(), DriverError> {
        assert!(dst.len() >= src.len());
        self.ctx.bind_to_thread()?;
        let (src, _record_src) = src.device_ptr(self);
        let (dst, _record_dst) = unsafe { dst.stream_synced_mut_slice(self) };
        unsafe { result::memcpy_dtoh_async(dst, src, self.cu_stream) }
    }

    /// Copy a [`CudaSlice`]/[`CudaView`] to a existing [`CudaSlice`]/[`CudaViewMut`].
    pub fn memcpy_dtod<T, Src: DevicePtr<T>, Dst: DevicePtrMut<T>>(
        self: &Arc<Self>,
        src: &Src,
        dst: &mut Dst,
    ) -> Result<(), DriverError> {
        assert!(dst.len() >= src.len());
        self.ctx.bind_to_thread()?;

        let num_bytes = src.num_bytes();

        let src_ctx = src.stream().context();
        let dst_ctx = self.context();

        if src_ctx == dst_ctx {
            let (src_ptr, _record_src) = src.device_ptr(self);
            let (dst_ptr, _record_dst) = dst.device_ptr_mut(self);
            unsafe { result::memcpy_dtod_async(dst_ptr, src_ptr, num_bytes, self.cu_stream) }
        } else {
            // NOTE: Although we want the current stream to wait on src to be ready,
            // we can't use src.device_ptr(self). When `_record_src` is dropped,
            // we record an event from the src_stream onto dst_stream (i.e., self). This is not
            // allowed in CUDA, and will return a CUDA_ERROR_INVALID_HANDLE
            // https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__CTX.html#group__CUDA__CTX_1gf3ee63561a7a371fa9d4dc0e31f94afd
            let (src_ptr, _record_src) = src.device_ptr(src.stream());
            let (dst_ptr, _record_dst) = dst.device_ptr_mut(self);
            // NOTE: Although we can't record events on streams they weren't created on,
            // we can *wait* on events from any stream. We can leverage this and wait on
            // a src event.
            // OPTIM: Ideally we could wait on the src write_events, but we can artificially
            // insert an event which guarantees src is available.
            self.wait(&src.stream().record_event(None)?)?;
            unsafe {
                result::memcpy_peer_async(
                    dst_ctx.cu_ctx,
                    dst_ptr,
                    src_ctx.cu_ctx,
                    src_ptr,
                    num_bytes,
                    self.cu_stream,
                )
            }
        }
    }

    /// Copy a [`CudaSlice`]/[`CudaView`] to a new [`CudaSlice`].
    pub fn clone_dtod<T: DeviceRepr, Src: DevicePtr<T>>(
        self: &Arc<Self>,
        src: &Src,
    ) -> Result<CudaSlice<T>, DriverError> {
        let mut dst = unsafe { self.alloc(src.len()) }?;
        self.memcpy_dtod(src, &mut dst)?;
        Ok(dst)
    }
}

impl<T> CudaSlice<T> {
    /// Creates a [CudaView] at the specified offset from the start of `self`.
    ///
    /// Panics if `range.start >= self.len`.
    ///
    /// # Example
    ///
    /// ```rust
    /// # use cudarc::driver::safe::{CudaContext, CudaSlice, CudaView};
    /// # fn do_something(view: &CudaView<u8>) {}
    /// # let ctx = CudaContext::new(0).unwrap();
    /// # let stream = ctx.default_stream();
    /// let mut slice = stream.alloc_zeros::<u8>(100).unwrap();
    /// let mut view = slice.slice(0..50);
    /// do_something(&view);
    /// ```
    ///
    /// Like a normal slice, borrow checking prevents the underlying [CudaSlice] from being dropped.
    /// ```rust,compile_fail
    /// # use cudarc::driver::safe::{CudaContext, CudaSlice, CudaView};
    /// # fn do_something(view: &CudaView<u8>) {}
    /// # let ctx = CudaContext::new(0).unwrap();
    /// # let stream = ctx.default_stream();
    /// let view = {
    ///     let mut slice = stream.alloc_zeros::<u8>(100).unwrap();
    ///     // cannot return view, since it borrows from slice
    ///     slice.slice(0..50)
    /// };
    /// do_something(&view);
    /// ```
    pub fn slice(&self, bounds: impl RangeBounds<usize>) -> CudaView<'_, T> {
        self.as_view().slice(bounds)
    }

    /// Fallible version of [CudaSlice::slice()].
    pub fn try_slice(&self, bounds: impl RangeBounds<usize>) -> Option<CudaView<'_, T>> {
        self.as_view().try_slice(bounds)
    }

    /// Creates a [CudaViewMut] at the specified offset from the start of `self`.
    ///
    /// Panics if `range` and `0...self.len()` are not overlapping.
    ///
    /// # Example
    ///
    /// ```rust
    /// # use cudarc::driver::safe::{CudaContext, CudaSlice, CudaViewMut};
    /// # fn do_something(view: &mut CudaViewMut<u8>) {}
    /// # let ctx = CudaContext::new(0).unwrap();
    /// # let stream = ctx.default_stream();
    /// let mut slice = stream.alloc_zeros::<u8>(100).unwrap();
    /// let mut view = slice.slice_mut(0..50);
    /// do_something(&mut view);
    /// ```
    ///
    /// Like a normal mutable slice, borrow checking prevents the underlying [CudaSlice] from being dropped.
    /// ```rust,compile_fail
    /// # use cudarc::driver::safe::{CudaContext, CudaSlice, CudaViewMut};
    /// # fn do_something(view: &mut CudaViewMut<u8>) {}
    /// # let ctx = CudaContext::new(0).unwrap();
    /// # let stream = ctx.default_stream();
    /// let mut view = {
    ///     let mut slice = stream.alloc_zeros::<u8>(100).unwrap();
    ///     // cannot return view, since it borrows from slice
    ///     slice.slice_mut(0..50)
    /// };
    /// do_something(&mut view);
    /// ```
    ///
    /// Like with normal mutable slices, one cannot mutably slice twice into the same [CudaSlice]:
    /// ```rust,compile_fail
    /// # use cudarc::driver::safe::{CudaContext, CudaSlice, CudaViewMut};
    /// # fn do_something(view: CudaViewMut<u8>, view2: CudaViewMut<u8>) {}
    /// # let ctx = CudaContext::new(0).unwrap();
    /// # let stream = ctx.default_stream();
    /// let mut slice = stream.alloc_zeros::<u8>(100).unwrap();
    /// let mut view1 = slice.slice_mut(0..50);
    /// // cannot borrow twice from slice
    /// let mut view2 = slice.slice_mut(50..100);
    /// do_something(view1, view2);
    /// ```
    /// If you need non-overlapping mutable views into a [CudaSlice], you can use [CudaSlice::split_at_mut()].
    pub fn slice_mut(&mut self, bounds: impl RangeBounds<usize>) -> CudaViewMut<'_, T> {
        self.try_slice_mut(bounds).unwrap()
    }

    /// Fallible version of [CudaSlice::slice_mut]
    pub fn try_slice_mut(&mut self, bounds: impl RangeBounds<usize>) -> Option<CudaViewMut<'_, T>> {
        to_range(bounds, self.len).map(|(start, end)| CudaViewMut {
            ptr: self.cu_device_ptr + (start * std::mem::size_of::<T>()) as u64,
            len: end - start,
            read: &self.read,
            write: &self.write,
            stream: &self.stream,
            marker: PhantomData,
        })
    }

    /// Reinterprets the slice of memory into a different type. `len` is the number
    /// of elements of the new type `S` that are expected. If not enough bytes
    /// are allocated in `self` for the view, then this returns `None`.
    ///
    /// # Safety
    /// This is unsafe because not the memory for the view may not be a valid interpretation
    /// for the type `S`.
    pub unsafe fn transmute<S>(&self, len: usize) -> Option<CudaView<'_, S>> {
        self.as_view().transmute(len)
    }

    /// Reinterprets the slice of memory into a different type. `len` is the number
    /// of elements of the new type `S` that are expected. If not enough bytes
    /// are allocated in `self` for the view, then this returns `None`.
    ///
    /// # Safety
    /// This is unsafe because not the memory for the view may not be a valid interpretation
    /// for the type `S`.
    pub unsafe fn transmute_mut<S>(&mut self, len: usize) -> Option<CudaViewMut<'_, S>> {
        (len * std::mem::size_of::<S>() <= self.len * std::mem::size_of::<T>()).then_some(
            CudaViewMut {
                ptr: self.cu_device_ptr,
                len,
                read: &self.read,
                write: &self.write,
                stream: &self.stream,
                marker: PhantomData,
            },
        )
    }

    pub fn split_at(&self, mid: usize) -> (CudaView<'_, T>, CudaView<'_, T>) {
        self.as_view().split_at(mid)
    }

    /// Fallible version of [CudaSlice::split_at]. Returns `None` if `mid > self.len`.
    pub fn try_split_at(&self, mid: usize) -> Option<(CudaView<'_, T>, CudaView<'_, T>)> {
        self.as_view().try_split_at(mid)
    }

    /// Splits the [CudaSlice] into two at the given index, returning two [CudaViewMut] for the two halves.
    ///
    /// Panics if `mid > self.len`.
    ///
    /// This method can be used to create non-overlapping mutable views into a [CudaSlice].
    /// ```rust
    /// # use cudarc::driver::safe::{CudaContext, CudaSlice, CudaViewMut};
    /// # fn do_something(view: CudaViewMut<u8>, view2: CudaViewMut<u8>) {}
    /// # let ctx = CudaContext::new(0).unwrap();
    /// # let stream = ctx.default_stream();
    /// let mut slice = stream.alloc_zeros::<u8>(100).unwrap();
    /// // split the slice into two non-overlapping, mutable views
    /// let (mut view1, mut view2) = slice.split_at_mut(50);
    /// do_something(view1, view2);
    /// ```
    pub fn split_at_mut(&mut self, mid: usize) -> (CudaViewMut<'_, T>, CudaViewMut<'_, T>) {
        self.try_split_at_mut(mid).unwrap()
    }

    /// Fallible version of [CudaSlice::split_at_mut].
    ///
    /// Returns `None` if `mid > self.len`.
    pub fn try_split_at_mut(
        &mut self,
        mid: usize,
    ) -> Option<(CudaViewMut<'_, T>, CudaViewMut<'_, T>)> {
        let length = self.len;
        (mid <= length).then(|| {
            let a = CudaViewMut {
                ptr: self.cu_device_ptr,
                len: mid,
                read: &self.read,
                write: &self.write,
                stream: &self.stream,
                marker: PhantomData,
            };
            let b = CudaViewMut {
                ptr: self.cu_device_ptr + (mid * std::mem::size_of::<T>()) as u64,
                len: length - mid,
                read: &self.read,
                write: &self.write,
                stream: &self.stream,
                marker: PhantomData,
            };
            (a, b)
        })
    }
}

impl<'a, T> CudaView<'a, T> {
    /// Creates a [CudaView] at the specified offset from the start of `self`.
    ///
    /// Panics if `range.start >= self.len`.
    ///
    /// # Example
    ///
    /// ```rust
    /// # use cudarc::driver::safe::{CudaContext, CudaSlice, CudaView};
    /// # fn do_something(view: &CudaView<u8>) {}
    /// # let ctx = CudaContext::new(0).unwrap();
    /// # let stream = ctx.default_stream();
    /// let mut slice = stream.alloc_zeros::<u8>(100).unwrap();
    /// let mut view = slice.slice(0..50);
    /// let mut view2 = view.slice(0..25);
    /// do_something(&view);
    /// ```
    pub fn slice(&self, bounds: impl RangeBounds<usize>) -> Self {
        self.try_slice(bounds).unwrap()
    }

    /// Fallible version of [CudaView::slice]
    pub fn try_slice(&self, bounds: impl RangeBounds<usize>) -> Option<Self> {
        to_range(bounds, self.len).map(|(start, end)| self.resize(start, end))
    }

    /// Reinterprets the slice of memory into a different type. `len` is the number
    /// of elements of the new type `S` that are expected. If not enough bytes
    /// are allocated in `self` for the view, then this returns `None`.
    ///
    /// # Safety
    /// This is unsafe because not the memory for the view may not be a valid interpretation
    /// for the type `S`.
    pub unsafe fn transmute<S>(&self, len: usize) -> Option<CudaView<'a, S>> {
        (len * std::mem::size_of::<S>() <= self.len * std::mem::size_of::<T>()).then_some(
            CudaView {
                ptr: self.ptr,
                len,
                read: self.read,
                write: self.write,
                stream: self.stream,
                marker: PhantomData,
            },
        )
    }

    pub fn split_at(&self, mid: usize) -> (Self, Self) {
        self.try_split_at(mid).unwrap()
    }

    /// Fallible version of [CudaSlice::split_at].
    ///
    /// Returns `None` if `mid > self.len`.
    pub fn try_split_at(&self, mid: usize) -> Option<(Self, Self)> {
        (mid <= self.len()).then(|| (self.resize(0, mid), self.resize(mid, self.len)))
    }

    /// Returns an iterarow over subviews of size `chunk_size`. Differs from [std::slice::ChunksExact],
    /// in that it asserts that the chunk_size must divide evenly into the length, instead of returning
    /// a remainder.
    pub fn chunks_exact(&self, chunk_size: usize) -> impl Iterator<Item = CudaView<'a, T>> + '_ {
        assert!(self.len.is_multiple_of(chunk_size));
        let num_chunks = self.len / chunk_size;
        (0..num_chunks).map(move |i| self.resize(i * chunk_size, (i + 1) * chunk_size))
    }
}

impl<'a, T> CudaViewMut<'a, T> {
    /// Creates a [CudaView] at the specified offset from the start of `self`.
    ///
    /// Panics if `range` and `0...self.len()` are not overlapping.
    ///
    /// # Example
    ///
    /// ```rust
    /// # use cudarc::driver::safe::{CudaContext, CudaSlice, CudaViewMut};
    /// # fn do_something(view: &mut CudaViewMut<u8>) {}
    /// # let ctx = CudaContext::new(0).unwrap();
    /// # let stream = ctx.default_stream();
    /// let mut slice = stream.alloc_zeros::<u8>(100).unwrap();
    /// let mut view = slice.slice_mut(0..50);
    /// let mut view2 = view.slice_mut(0..25);
    /// do_something(&mut view2);
    /// ```
    ///
    /// One cannot slice twice into the same [CudaViewMut]:
    /// ```rust,compile_fail
    /// # use cudarc::driver::safe::{CudaContext, CudaSlice, CudaViewMut};
    /// # fn do_something(view: CudaViewMut<u8>, view2: CudaViewMut<u8>) {}
    /// # let ctx = CudaContext::new(0).unwrap();
    /// # let stream = ctx.default_stream();
    /// let mut slice = stream.alloc_zeros::<u8>(100).unwrap();
    /// let mut view = slice.slice_mut(0..50);
    /// // cannot borrow twice from same view
    /// let mut view1 = slice.slice_mut(0..25);
    /// let mut view2 = slice.slice_mut(25..50);
    /// do_something(view1, view2);
    /// ```
    /// If you need non-overlapping mutable views into a [CudaViewMut], you can use [CudaViewMut::split_at_mut()].
    pub fn slice<'b>(&'b self, bounds: impl RangeBounds<usize>) -> CudaView<'b, T> {
        self.try_slice(bounds).unwrap()
    }

    /// Fallible version of [CudaViewMut::slice]
    pub fn try_slice<'b>(&'b self, bounds: impl RangeBounds<usize>) -> Option<CudaView<'b, T>> {
        to_range(bounds, self.len).map(move |(start, end)| self.as_view().resize(start, end))
    }

    /// Reinterprets the slice of memory into a different type. `len` is the number
    /// of elements of the new type `S` that are expected. If not enough bytes
    /// are allocated in `self` for the view, then this returns `None`.
    ///
    /// # Safety
    /// This is unsafe because not the memory for the view may not be a valid interpretation
    /// for the type `S`.
    pub unsafe fn transmute<'b, S>(&'b self, len: usize) -> Option<CudaView<'b, S>> {
        (len * std::mem::size_of::<S>() <= self.len * std::mem::size_of::<T>()).then_some(
            CudaView {
                ptr: self.ptr,
                len,
                read: self.read,
                write: self.write,
                stream: self.stream,
                marker: PhantomData,
            },
        )
    }

    /// Creates a [CudaViewMut] at the specified offset from the start of `self`.
    ///
    /// Panics if `range` and `0...self.len()` are not overlapping.
    pub fn slice_mut<'b>(&'b mut self, bounds: impl RangeBounds<usize>) -> CudaViewMut<'b, T> {
        self.try_slice_mut(bounds).unwrap()
    }

    /// Fallible version of [CudaViewMut::slice_mut]
    pub fn try_slice_mut<'b>(
        &'b mut self,
        bounds: impl RangeBounds<usize>,
    ) -> Option<CudaViewMut<'b, T>> {
        to_range(bounds, self.len).map(|(start, end)| CudaViewMut {
            ptr: self.ptr + (start * std::mem::size_of::<T>()) as u64,
            len: end - start,
            read: self.read,
            write: self.write,
            stream: self.stream,
            marker: PhantomData,
        })
    }

    /// Splits the [CudaViewMut] into two at the given index.
    ///
    /// Panics if `mid > self.len`.
    ///
    /// This method can be used to create non-overlapping mutable views into a [CudaViewMut].
    /// ```rust
    /// # use cudarc::driver::safe::{CudaContext, CudaSlice, CudaViewMut};
    /// # fn do_something(view: CudaViewMut<u8>, view2: CudaViewMut<u8>) {}
    /// # let ctx = CudaContext::new(0).unwrap();
    /// # let stream = ctx.default_stream();
    /// let mut slice = stream.alloc_zeros::<u8>(100).unwrap();
    /// let mut view = slice.slice_mut(0..50);
    /// // split the view into two non-overlapping, mutable views
    /// let (mut view1, mut view2) = view.split_at_mut(25);
    /// do_something(view1, view2);
    pub fn split_at_mut<'b>(&'b mut self, mid: usize) -> (CudaViewMut<'b, T>, CudaViewMut<'b, T>) {
        self.try_split_at_mut(mid).unwrap()
    }

    /// Fallible version of [CudaViewMut::split_at_mut].
    ///
    /// Returns `None` if `mid > self.len`
    pub fn try_split_at_mut<'b>(
        &'b mut self,
        mid: usize,
    ) -> Option<(CudaViewMut<'b, T>, CudaViewMut<'b, T>)> {
        let length = self.len;
        (mid <= length).then(|| {
            let a = CudaViewMut {
                ptr: self.ptr,
                len: mid,
                read: self.read,
                write: self.write,
                stream: self.stream,
                marker: PhantomData,
            };
            let b = CudaViewMut {
                ptr: self.ptr + (mid * std::mem::size_of::<T>()) as u64,
                len: length - mid,
                read: self.read,
                write: self.write,
                stream: self.stream,
                marker: PhantomData,
            };
            (a, b)
        })
    }

    /// Returns an iterarow over subviews of size `chunk_size`. Differs from [std::slice::ChunksExactMut],
    /// in that it asserts that the chunk_size must divide evenly into the length, instead of returning
    /// a remainder.
    pub fn chunks_exact_mut(self, chunk_size: usize) -> impl Iterator<Item = CudaViewMut<'a, T>> {
        assert!(self.len.is_multiple_of(chunk_size));
        let num_chunks = self.len / chunk_size;
        (0..num_chunks).map(move |i| CudaViewMut {
            ptr: self.ptr + (i * chunk_size * std::mem::size_of::<T>()) as u64,
            len: chunk_size,
            read: self.read,
            write: self.write,
            stream: self.stream,
            marker: PhantomData,
        })
    }

    /// Reinterprets the slice of memory into a different type. `len` is the number
    /// of elements of the new type `S` that are expected. If not enough bytes
    /// are allocated in `self` for the view, then this returns `None`.
    ///
    /// # Safety
    /// This is unsafe because not the memory for the view may not be a valid interpretation
    /// for the type `S`.
    pub unsafe fn transmute_mut<'b, S>(&'b mut self, len: usize) -> Option<CudaViewMut<'b, S>> {
        (len * std::mem::size_of::<S>() <= self.len * std::mem::size_of::<T>()).then_some(
            CudaViewMut {
                ptr: self.ptr,
                len,
                read: self.read,
                write: self.write,
                stream: self.stream,
                marker: PhantomData,
            },
        )
    }
}

pub(super) fn to_range(range: impl RangeBounds<usize>, len: usize) -> Option<(usize, usize)> {
    let start = match range.start_bound() {
        Bound::Included(&n) => n,
        Bound::Excluded(&n) => n + 1,
        Bound::Unbounded => 0,
    };
    let end = match range.end_bound() {
        Bound::Included(&n) => n + 1,
        Bound::Excluded(&n) => n,
        Bound::Unbounded => len,
    };
    (start <= end && end <= len).then_some((start, end))
}

/// Wrapper around [sys::CUmodule]. Create with [CudaContext::load_module()].
///
/// Call [CudaModule::load_function] to load a [CudaFunction].
#[derive(Debug)]
pub struct CudaModule {
    pub(crate) cu_module: sys::CUmodule,
    pub(crate) ctx: Arc<CudaContext>,
}

unsafe impl Send for CudaModule {}
unsafe impl Sync for CudaModule {}

impl Drop for CudaModule {
    fn drop(&mut self) {
        self.ctx.record_err(self.ctx.bind_to_thread());
        self.ctx
            .record_err(unsafe { result::module::unload(self.cu_module) });
    }
}

impl CudaContext {
    /// Dynamically load a compiled ptx into this context.
    ///
    /// - `ptx` contains the compiled ptx
    #[cfg(feature = "nvrtc")]
    pub fn load_module(
        self: &Arc<Self>,
        ptx: crate::nvrtc::Ptx,
    ) -> Result<Arc<CudaModule>, result::DriverError> {
        self.bind_to_thread()?;

        let cu_module = match ptx.0 {
            crate::nvrtc::PtxKind::Image(image) => unsafe {
                result::module::load_data(image.as_ptr() as *const _)
            },
            crate::nvrtc::PtxKind::Src(src) => {
                let c_src = CString::new(src).unwrap();
                unsafe { result::module::load_data(c_src.as_ptr() as *const _) }
            }
            crate::nvrtc::PtxKind::File(path) => {
                let name_c = CString::new(path.to_str().unwrap()).unwrap();
                result::module::load(name_c)
            }
            crate::nvrtc::PtxKind::Binary(data) => unsafe {
                result::module::load_data(data.as_ptr() as *const _)
            },
        }?;
        Ok(Arc::new(CudaModule {
            cu_module,
            ctx: self.clone(),
        }))
    }
}

/// Wrapper around [sys::CUfunction]. Used by [CudaStream::launch_builder] to execute kernels.
#[derive(Debug, Clone)]
pub struct CudaFunction {
    pub(crate) cu_function: sys::CUfunction,
    #[allow(unused)]
    pub(crate) module: Arc<CudaModule>,
}

unsafe impl Send for CudaFunction {}
unsafe impl Sync for CudaFunction {}

impl CudaModule {
    /// Loads a function from the loaded module with the given name.
    pub fn load_function(self: &Arc<Self>, fn_name: &str) -> Result<CudaFunction, DriverError> {
        let fn_name_c = CString::new(fn_name).unwrap();
        let cu_function = unsafe { result::module::get_function(self.cu_module, fn_name_c) }?;
        Ok(CudaFunction {
            cu_function,
            module: self.clone(),
        })
    }

    /// Gets a global/constant symbol from the loaded module as a [CudaSlice<u8>].
    ///
    /// This can be used to access `__constant__` memory declared in CUDA kernels.
    /// The returned slice can be transmuted to the appropriate type via views.
    ///
    /// # Example
    ///
    /// ```ignore
    /// // In CUDA: __constant__ float my_const[4];
    /// let symbol = module.get_global("my_const", &stream)?;
    /// let mut symbol_view = symbol.as_view_mut();
    /// let mut symbol_f32 = unsafe { symbol_view.transmute_mut::<f32>(4).unwrap() };
    /// stream.memcpy_htod(&[1.0f32, 2.0, 3.0, 4.0], &mut symbol_f32)?;
    /// ```
    pub fn get_global<'a>(
        self: &'a Arc<Self>,
        name: &str,
        stream: &'a Arc<CudaStream>,
    ) -> Result<CudaViewMut<'a, u8>, DriverError> {
        let name_c =
            CString::new(name).map_err(|_| DriverError(sys::CUresult::CUDA_ERROR_INVALID_VALUE))?;
        let (cu_device_ptr, bytes) = unsafe { result::module::get_global(self.cu_module, name_c) }?;
        Ok(CudaViewMut {
            ptr: cu_device_ptr,
            len: bytes,
            read: &None,
            write: &None,
            stream,
            marker: PhantomData,
        })
    }
}

impl CudaFunction {
    pub fn occupancy_available_dynamic_smem_per_block(
        &self,
        num_blocks: u32,
        block_size: u32,
    ) -> Result<usize, result::DriverError> {
        let mut dynamic_smem_size: usize = 0;

        unsafe {
            sys::cuOccupancyAvailableDynamicSMemPerBlock(
                &mut dynamic_smem_size,
                self.cu_function,
                num_blocks as std::ffi::c_int,
                block_size as std::ffi::c_int,
            )
            .result()?
        };

        Ok(dynamic_smem_size)
    }

    pub fn occupancy_max_active_blocks_per_multiprocessor(
        &self,
        block_size: u32,
        dynamic_smem_size: usize,
        flags: Option<sys::CUoccupancy_flags_enum>,
    ) -> Result<u32, result::DriverError> {
        let mut num_blocks: std::ffi::c_int = 0;
        let flags = flags.unwrap_or(sys::CUoccupancy_flags_enum::CU_OCCUPANCY_DEFAULT);

        unsafe {
            sys::cuOccupancyMaxActiveBlocksPerMultiprocessorWithFlags(
                &mut num_blocks,
                self.cu_function,
                block_size as std::ffi::c_int,
                dynamic_smem_size,
                flags as std::ffi::c_uint,
            )
            .result()?
        };

        Ok(num_blocks as u32)
    }

    #[cfg(not(any(
        feature = "cuda-11070",
        feature = "cuda-11060",
        feature = "cuda-11050",
        feature = "cuda-11040"
    )))]
    pub fn occupancy_max_active_clusters(
        &self,
        config: crate::driver::LaunchConfig,
        stream: &CudaStream,
    ) -> Result<u32, result::DriverError> {
        let mut num_clusters: std::ffi::c_int = 0;

        let cfg = sys::CUlaunchConfig {
            gridDimX: config.grid_dim.0,
            gridDimY: config.grid_dim.1,
            gridDimZ: config.grid_dim.2,
            blockDimX: config.block_dim.0,
            blockDimY: config.block_dim.1,
            blockDimZ: config.block_dim.2,
            sharedMemBytes: config.shared_mem_bytes,
            hStream: stream.cu_stream,
            attrs: std::ptr::null_mut(),
            numAttrs: 0,
        };

        unsafe {
            sys::cuOccupancyMaxActiveClusters(&mut num_clusters, self.cu_function, &cfg).result()?
        };

        Ok(num_clusters as u32)
    }

    pub fn occupancy_max_potential_block_size(
        &self,
        block_size_to_dynamic_smem_size: extern "C" fn(block_size: std::ffi::c_int) -> usize,
        dynamic_smem_size: usize,
        block_size_limit: u32,
        flags: Option<sys::CUoccupancy_flags_enum>,
    ) -> Result<(u32, u32), result::DriverError> {
        let mut min_grid_size: std::ffi::c_int = 0;
        let mut block_size: std::ffi::c_int = 0;
        let flags = flags.unwrap_or(sys::CUoccupancy_flags_enum::CU_OCCUPANCY_DEFAULT);

        unsafe {
            sys::cuOccupancyMaxPotentialBlockSizeWithFlags(
                &mut min_grid_size,
                &mut block_size,
                self.cu_function,
                Some(block_size_to_dynamic_smem_size),
                dynamic_smem_size,
                block_size_limit as std::ffi::c_int,
                flags as std::ffi::c_uint,
            )
            .result()?
        };

        Ok((min_grid_size as u32, block_size as u32))
    }

    #[cfg(not(any(
        feature = "cuda-11070",
        feature = "cuda-11060",
        feature = "cuda-11050",
        feature = "cuda-11040"
    )))]
    pub fn occupancy_max_potential_cluster_size(
        &self,
        config: crate::driver::LaunchConfig,
        stream: &CudaStream,
    ) -> Result<u32, result::DriverError> {
        let mut cluster_size: std::ffi::c_int = 0;

        let cfg = sys::CUlaunchConfig {
            gridDimX: config.grid_dim.0,
            gridDimY: config.grid_dim.1,
            gridDimZ: config.grid_dim.2,
            blockDimX: config.block_dim.0,
            blockDimY: config.block_dim.1,
            blockDimZ: config.block_dim.2,
            sharedMemBytes: config.shared_mem_bytes,
            hStream: stream.cu_stream,
            attrs: std::ptr::null_mut(),
            numAttrs: 0,
        };

        unsafe {
            sys::cuOccupancyMaxPotentialClusterSize(&mut cluster_size, self.cu_function, &cfg)
                .result()?
        };

        Ok(cluster_size as u32)
    }

    /// Get the value of a specific attribute of this [CudaFunction].
    ///
    /// See [CUDA docs](https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__EXEC.html#group__CUDA__EXEC_1g5e92a1b0d8d1b82cb00dcfb2de15961b)
    pub fn get_attribute(
        &self,
        attribute: CUfunction_attribute_enum,
    ) -> Result<i32, result::DriverError> {
        self.module.ctx.bind_to_thread()?;
        unsafe { result::function::get_function_attribute(self.cu_function, attribute) }
    }

    /// Get the number of registers used per thread.
    pub fn num_regs(&self) -> Result<i32, result::DriverError> {
        self.get_attribute(CUfunction_attribute_enum::CU_FUNC_ATTRIBUTE_NUM_REGS)
    }

    /// Get the size of statically-allocated shared memory in bytes.
    pub fn shared_size_bytes(&self) -> Result<i32, result::DriverError> {
        self.get_attribute(CUfunction_attribute_enum::CU_FUNC_ATTRIBUTE_SHARED_SIZE_BYTES)
    }

    /// Get the size of constant memory in bytes used by this function.
    pub fn const_size_bytes(&self) -> Result<i32, result::DriverError> {
        self.get_attribute(CUfunction_attribute_enum::CU_FUNC_ATTRIBUTE_CONST_SIZE_BYTES)
    }

    /// Get the size of local memory in bytes used per thread.
    pub fn local_size_bytes(&self) -> Result<i32, result::DriverError> {
        self.get_attribute(CUfunction_attribute_enum::CU_FUNC_ATTRIBUTE_LOCAL_SIZE_BYTES)
    }

    /// Get the maximum number of threads per block for this function.
    pub fn max_threads_per_block(&self) -> Result<i32, result::DriverError> {
        self.get_attribute(CUfunction_attribute_enum::CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK)
    }

    /// Get the PTX virtual architecture version for which the function was compiled.
    pub fn ptx_version(&self) -> Result<i32, result::DriverError> {
        self.get_attribute(CUfunction_attribute_enum::CU_FUNC_ATTRIBUTE_PTX_VERSION)
    }

    /// Get the binary architecture version for which the function was compiled.
    pub fn binary_version(&self) -> Result<i32, result::DriverError> {
        self.get_attribute(CUfunction_attribute_enum::CU_FUNC_ATTRIBUTE_BINARY_VERSION)
    }

    /// Set the value of a specific attribute of this [CudaFunction].
    pub fn set_attribute(
        &self,
        attribute: CUfunction_attribute_enum,
        value: i32,
    ) -> Result<(), result::DriverError> {
        unsafe { result::function::set_function_attribute(self.cu_function, attribute, value) }
    }

    /// Set the cache config of this [CudaFunction].
    pub fn set_function_cache_config(
        &self,
        attribute: CUfunc_cache_enum,
    ) -> Result<(), result::DriverError> {
        unsafe { result::function::set_function_cache_config(self.cu_function, attribute) }
    }
}

impl<T> CudaSlice<T> {
    /// Takes ownership of the underlying [sys::CUdeviceptr]. **It is up
    /// to the owner to free this value**.
    ///
    /// Drops the underlying host_buf if there is one.
    pub fn leak(self) -> sys::CUdeviceptr {
        let mut s = std::mem::ManuallyDrop::new(self);
        let ptr = s.cu_device_ptr;

        // Ensure pending operations are complete before resources are released.
        if let Some(read) = s.read.as_ref() {
            s.stream.ctx.record_err(s.stream.wait(read));
        }
        if let Some(write) = s.write.as_ref() {
            s.stream.ctx.record_err(s.stream.wait(write));
        }

        // Manually drop fields that own resources.
        unsafe {
            std::ptr::drop_in_place(&mut s.read);
            std::ptr::drop_in_place(&mut s.write);
            std::ptr::drop_in_place(&mut s.stream);
        }

        ptr
    }
}

impl CudaStream {
    /// Creates a [CudaSlice] from a [sys::CUdeviceptr]. Useful in conjunction with
    /// [`CudaSlice::leak()`].
    ///
    /// # Safety
    /// - `cu_device_ptr` must be a valid allocation
    /// - `cu_device_ptr` must space for `len * std::mem::size_of<T>()` bytes
    /// - The memory may not be valid for type `T`, so some sort of memset operation
    ///   should be called on the memory.
    pub unsafe fn upgrade_device_ptr<T>(
        self: &Arc<Self>,
        cu_device_ptr: sys::CUdeviceptr,
        len: usize,
    ) -> CudaSlice<T> {
        let (read, write) = if self.ctx.is_event_tracking() {
            (
                Some(self.ctx.new_event(None).unwrap()),
                Some(self.ctx.new_event(None).unwrap()),
            )
        } else {
            (None, None)
        };
        CudaSlice {
            cu_device_ptr,
            len,
            read,
            write,
            stream: self.clone(),
            marker: PhantomData,
        }
    }
}

#[cfg(test)]
mod tests {
    use std::time::Instant;

    use super::*;

    #[test]
    fn test_transmutes() {
        let ctx = CudaContext::new(0).unwrap();
        let stream = ctx.default_stream();
        let mut slice = stream.alloc_zeros::<u8>(100).unwrap();
        assert!(unsafe { slice.transmute::<f32>(25) }.is_some());
        assert!(unsafe { slice.transmute::<f32>(26) }.is_none());
        assert!(unsafe { slice.transmute_mut::<f32>(25) }.is_some());
        assert!(unsafe { slice.transmute_mut::<f32>(26) }.is_none());

        {
            let view = slice.slice(0..100);
            assert!(unsafe { view.transmute::<f32>(25) }.is_some());
            assert!(unsafe { view.transmute::<f32>(26) }.is_none());
        }

        {
            let mut view_mut = slice.slice_mut(0..100);
            assert!(unsafe { view_mut.transmute::<f32>(25) }.is_some());
            assert!(unsafe { view_mut.transmute::<f32>(26) }.is_none());
            assert!(unsafe { view_mut.transmute_mut::<f32>(25) }.is_some());
            assert!(unsafe { view_mut.transmute_mut::<f32>(26) }.is_none());
        }
    }

    #[test]
    fn test_threading() {
        let ctx1 = CudaContext::new(0).unwrap();
        let ctx2 = ctx1.clone();

        let thread1 = std::thread::spawn(move || {
            ctx1.bind_to_thread()?;
            ctx1.default_stream().alloc_zeros::<f32>(10)
        });
        let thread2 = std::thread::spawn(move || {
            ctx2.bind_to_thread()?;
            ctx2.default_stream().alloc_zeros::<f32>(10)
        });

        let _: crate::driver::CudaSlice<f32> = thread1.join().unwrap().unwrap();
        let _: crate::driver::CudaSlice<f32> = thread2.join().unwrap().unwrap();
    }

    #[test]
    fn test_post_build_arc_count() {
        let ctx = CudaContext::new(0).unwrap();
        assert_eq!(Arc::strong_count(&ctx), 1);
    }

    #[test]
    fn test_post_alloc_arc_counts() {
        let ctx = CudaContext::new(0).unwrap();
        assert_eq!(Arc::strong_count(&ctx), 1);
        let stream = ctx.default_stream();
        assert_eq!(Arc::strong_count(&ctx), 2);
        let t = stream.alloc_zeros::<f32>(1).unwrap();
        assert_eq!(Arc::strong_count(&ctx), 4);
        assert_eq!(Arc::strong_count(&stream), 2);
        drop(t);
        assert_eq!(Arc::strong_count(&ctx), 2);
        assert_eq!(Arc::strong_count(&stream), 1);
        drop(stream);
        assert_eq!(Arc::strong_count(&ctx), 1);
    }

    #[test]
    #[ignore = "must be executed by itself"]
    fn test_post_alloc_memory() {
        let ctx = CudaContext::new(0).unwrap();
        let stream = ctx.default_stream();

        let (free1, total1) = ctx.mem_get_info().unwrap();

        let t = stream.clone_htod(&[0.0f32; 5]).unwrap();
        let (free2, total2) = ctx.mem_get_info().unwrap();
        assert_eq!(total1, total2);
        assert!(free2 < free1);

        drop(t);
        ctx.synchronize().unwrap();

        let (free3, total3) = ctx.mem_get_info().unwrap();
        assert_eq!(total2, total3);
        assert!(free3 > free2);
        assert_eq!(free3, free1);
    }

    #[test]
    fn test_ctx_copy_to_views() {
        let ctx = CudaContext::new(0).unwrap();
        let stream = ctx.default_stream();

        let smalls = [
            stream.clone_htod(&[-1.0f32, -0.8]).unwrap(),
            stream.clone_htod(&[-0.6, -0.4]).unwrap(),
            stream.clone_htod(&[-0.2, 0.0]).unwrap(),
            stream.clone_htod(&[0.2, 0.4]).unwrap(),
            stream.clone_htod(&[0.6, 0.8]).unwrap(),
        ];
        let mut big = stream.alloc_zeros::<f32>(10).unwrap();

        let mut offset = 0;
        for small in smalls.iter() {
            let mut sub = big.slice_mut(offset..offset + small.len());
            stream.memcpy_dtod(small, &mut sub).unwrap();
            offset += small.len();
        }

        assert_eq!(
            stream.clone_dtoh(&big).unwrap(),
            [-1.0, -0.8, -0.6, -0.4, -0.2, 0.0, 0.2, 0.4, 0.6, 0.8]
        );
    }

    #[test]
    fn test_leak_and_upgrade() {
        let ctx = CudaContext::new(0).unwrap();
        let stream = ctx.default_stream();

        let a = stream.clone_htod(&[1.0f32, 2.0, 3.0, 4.0, 5.0]).unwrap();

        let ptr = a.leak();
        let b = unsafe { stream.upgrade_device_ptr::<f32>(ptr, 3) };
        assert_eq!(stream.clone_dtoh(&b).unwrap(), &[1.0, 2.0, 3.0]);

        let ptr = b.leak();
        let c = unsafe { stream.upgrade_device_ptr::<f32>(ptr, 5) };
        assert_eq!(stream.clone_dtoh(&c).unwrap(), &[1.0, 2.0, 3.0, 4.0, 5.0]);
    }

    /// See https://github.com/chelsea0x3b/cudarc/issues/160
    #[test]
    fn test_slice_is_freed_with_correct_context() {
        let ctx0 = CudaContext::new(0).unwrap();
        let slice = ctx0.default_stream().clone_htod(&[1.0; 10]).unwrap();
        let ctx1 = CudaContext::new(0).unwrap();
        ctx1.bind_to_thread().unwrap();
        drop(ctx0);
        drop(slice);
        drop(ctx1);
    }

    /// See https://github.com/chelsea0x3b/cudarc/issues/161
    #[test]
    fn test_copy_uses_correct_context() {
        let ctx0 = CudaContext::new(0).unwrap();
        let _ctx1 = CudaContext::new(0).unwrap();
        let slice = ctx0.default_stream().clone_htod(&[1.0; 10]).unwrap();
        let _out = ctx0.default_stream().clone_dtoh(&slice).unwrap();
    }

    #[test]
    fn test_htod_copy_pinned() {
        let truth = [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0];
        let ctx = CudaContext::new(0).unwrap();
        let stream = ctx.default_stream();
        let mut pinned = unsafe { ctx.alloc_pinned::<f32>(10) }.unwrap();
        pinned.as_mut_slice().unwrap().clone_from_slice(&truth);
        assert_eq!(pinned.as_slice().unwrap(), &truth);
        let dst = stream.clone_htod(&pinned).unwrap();
        let host = stream.clone_dtoh(&dst).unwrap();
        assert_eq!(&host, &truth);
    }

    #[test]
    fn test_pinned_copy_is_faster() {
        let ctx = CudaContext::new(0).unwrap();
        let stream = ctx.new_stream().unwrap();

        let n = 100_000;
        let n_samples = 5;
        let not_pinned = std::vec![0.0f32; n];

        let start = Instant::now();
        for _ in 0..n_samples {
            let _ = stream.clone_htod(&not_pinned).unwrap();
            stream.synchronize().unwrap();
        }
        let unpinned_elapsed = start.elapsed() / n_samples;

        let pinned = unsafe { ctx.alloc_pinned::<f32>(n) }.unwrap();

        let start = Instant::now();
        for _ in 0..n_samples {
            let _ = stream.clone_htod(&pinned).unwrap();
            stream.synchronize().unwrap();
        }
        let pinned_elapsed = start.elapsed() / n_samples;

        // pinned memory transfer speed should be at least 2x faster, but this depends
        // on device
        assert!(
            pinned_elapsed.as_secs_f32() * 1.5 < unpinned_elapsed.as_secs_f32(),
            "{unpinned_elapsed:?} vs {pinned_elapsed:?}"
        );
    }

    #[test]
    fn test_primary_context_is_primary() {
        let ctx = CudaContext::new(0).unwrap();
        assert!(ctx.is_primary());
    }

    /// Helper to create a non-primary context for testing `from_raw_context`.
    /// Uses `cuCtxCreate_v4` (CUDA 12.050+) or `cuCtxCreate_v3` (CUDA 11.040–12.040).
    #[cfg(any(
        feature = "cuda-11040",
        feature = "cuda-11050",
        feature = "cuda-11060",
        feature = "cuda-11070",
        feature = "cuda-11080",
        feature = "cuda-12000",
        feature = "cuda-12010",
        feature = "cuda-12020",
        feature = "cuda-12030",
        feature = "cuda-12040",
        feature = "cuda-12050",
        feature = "cuda-12060",
        feature = "cuda-12080",
        feature = "cuda-12090",
    ))]
    fn create_non_primary_context() -> (sys::CUdevice, sys::CUcontext) {
        result::init().unwrap();
        let cu_device = result::device::get(0).unwrap();

        #[cfg(any(
            feature = "cuda-12050",
            feature = "cuda-12060",
            feature = "cuda-12080",
            feature = "cuda-12090",
            feature = "cuda-13000",
            feature = "cuda-13010",
        ))]
        let cu_ctx = unsafe { result::ctx::create_v4(std::ptr::null_mut(), 0, cu_device) }
            .expect("cuCtxCreate_v4 failed");

        #[cfg(not(any(
            feature = "cuda-12050",
            feature = "cuda-12060",
            feature = "cuda-12080",
            feature = "cuda-12090",
            feature = "cuda-13000",
            feature = "cuda-13010",
        )))]
        let cu_ctx =
            unsafe { result::ctx::create_v3(0, cu_device) }.expect("cuCtxCreate_v3 failed");

        assert!(!cu_ctx.is_null());
        (cu_device, cu_ctx)
    }

    #[test]
    #[cfg(any(
        feature = "cuda-11040",
        feature = "cuda-11050",
        feature = "cuda-11060",
        feature = "cuda-11070",
        feature = "cuda-11080",
        feature = "cuda-12000",
        feature = "cuda-12010",
        feature = "cuda-12020",
        feature = "cuda-12030",
        feature = "cuda-12040",
        feature = "cuda-12050",
        feature = "cuda-12060",
        feature = "cuda-12080",
        feature = "cuda-12090",
    ))]
    fn test_from_raw_context_creates_and_destroys() {
        let (cu_device, cu_ctx) = create_non_primary_context();

        let ctx = unsafe { CudaContext::from_raw_context(0, cu_device, cu_ctx) }.unwrap();
        assert!(!ctx.is_primary());
        // Verify the context is bound and usable.
        ctx.bind_to_thread().unwrap();
        // Drop should call cuCtxDestroy_v2, not primary_ctx::release.
        drop(ctx);
    }

    #[test]
    #[cfg(any(
        feature = "cuda-11040",
        feature = "cuda-11050",
        feature = "cuda-11060",
        feature = "cuda-11070",
        feature = "cuda-11080",
        feature = "cuda-12000",
        feature = "cuda-12010",
        feature = "cuda-12020",
        feature = "cuda-12030",
        feature = "cuda-12040",
        feature = "cuda-12050",
        feature = "cuda-12060",
        feature = "cuda-12080",
        feature = "cuda-12090",
    ))]
    fn test_from_raw_context_bind_to_thread() {
        let (cu_device, cu_ctx) = create_non_primary_context();

        let ctx = unsafe { CudaContext::from_raw_context(0, cu_device, cu_ctx) }.unwrap();

        // Verify bind_to_thread works from another thread.
        let ctx2 = ctx.clone();
        let handle = std::thread::spawn(move || {
            ctx2.bind_to_thread().unwrap();
            let stream = ctx2.default_stream();
            let data = stream.clone_htod(&[1.0f32, 2.0, 3.0]).unwrap();
            let result = stream.clone_dtoh(&data).unwrap();
            assert_eq!(result, std::vec![1.0f32, 2.0, 3.0]);
        });
        handle.join().unwrap();
    }

    #[test]
    #[cfg(any(
        feature = "cuda-11040",
        feature = "cuda-11050",
        feature = "cuda-11060",
        feature = "cuda-11070",
        feature = "cuda-11080",
        feature = "cuda-12000",
        feature = "cuda-12010",
        feature = "cuda-12020",
        feature = "cuda-12030",
        feature = "cuda-12040",
        feature = "cuda-12050",
        feature = "cuda-12060",
        feature = "cuda-12080",
        feature = "cuda-12090",
        feature = "cuda-13000",
        feature = "cuda-13010",
    ))]
    fn test_new_non_primary_creates_and_destroys() {
        let ctx = CudaContext::new_non_primary(0, 0).unwrap();
        assert!(!ctx.is_primary());
        ctx.bind_to_thread().unwrap();
        drop(ctx);
    }

    #[test]
    #[cfg(any(
        feature = "cuda-11040",
        feature = "cuda-11050",
        feature = "cuda-11060",
        feature = "cuda-11070",
        feature = "cuda-11080",
        feature = "cuda-12000",
        feature = "cuda-12010",
        feature = "cuda-12020",
        feature = "cuda-12030",
        feature = "cuda-12040",
        feature = "cuda-12050",
        feature = "cuda-12060",
        feature = "cuda-12080",
        feature = "cuda-12090",
        feature = "cuda-13000",
        feature = "cuda-13010",
    ))]
    fn test_new_non_primary_htod_dtoh() {
        let ctx = CudaContext::new_non_primary(0, 0).unwrap();
        let stream = ctx.default_stream();
        let data = stream.clone_htod(&[1.0f32, 2.0, 3.0]).unwrap();
        let result = stream.clone_dtoh(&data).unwrap();
        assert_eq!(result, std::vec![1.0f32, 2.0, 3.0]);
    }

    #[test]
    #[cfg(any(
        feature = "cuda-11040",
        feature = "cuda-11050",
        feature = "cuda-11060",
        feature = "cuda-11070",
        feature = "cuda-11080",
        feature = "cuda-12000",
        feature = "cuda-12010",
        feature = "cuda-12020",
        feature = "cuda-12030",
        feature = "cuda-12040",
        feature = "cuda-12050",
        feature = "cuda-12060",
        feature = "cuda-12080",
        feature = "cuda-12090",
        feature = "cuda-13000",
        feature = "cuda-13010",
    ))]
    fn test_new_non_primary_cross_thread() {
        let ctx = CudaContext::new_non_primary(0, 0).unwrap();
        let ctx2 = ctx.clone();
        let handle = std::thread::spawn(move || {
            ctx2.bind_to_thread().unwrap();
            let stream = ctx2.default_stream();
            let data = stream.clone_htod(&[4.0f32, 5.0, 6.0]).unwrap();
            let result = stream.clone_dtoh(&data).unwrap();
            assert_eq!(result, std::vec![4.0f32, 5.0, 6.0]);
        });
        handle.join().unwrap();
    }
}