diff --git a/level_zero/src/ze.rs b/level_zero/src/ze.rs
index f8a2c3b..4267682 100644
--- a/level_zero/src/ze.rs
+++ b/level_zero/src/ze.rs
@@ -173,6 +173,16 @@ impl Context {
         check!(sys::zeContextCreate(drv.0, &ctx_desc, &mut result));
         Ok(Context(result))
     }
+
+    pub unsafe fn mem_free(&mut self, ptr: *mut c_void) -> Result<()> {
+        check! {
+            sys::zeMemFree(
+                self.0,
+                ptr,
+            )
+        };
+        Ok(())
+    }
 }
 
 impl Drop for Context {
@@ -239,7 +249,7 @@ pub struct Module(sys::ze_module_handle_t);
 
 impl Module {
     // HACK ALERT
-    // We use OpenCL for now to do SPIR-V linking, because Level0 
+    // We use OpenCL for now to do SPIR-V linking, because Level0
     // does not allow linking. Don't let presence of zeModuleDynamicLink fool
     // you, it's not currently possible to create non-compiled modules.
     // zeModuleCreate always compiles (builds and links).
diff --git a/notcuda/build.rs b/notcuda/build.rs
new file mode 100644
index 0000000..3b8999f
--- /dev/null
+++ b/notcuda/build.rs
@@ -0,0 +1,27 @@
+// HACK ALERT
+// This buidl script has been copy-pasted from cl-sys to avoid CUDA libraries
+// overriding path to OpenCL
+
+ fn main() {
+    if cfg!(windows) {
+        let known_sdk = [
+            // E.g. "c:\Program Files (x86)\Intel\OpenCL SDK\lib\x86\"
+            ("INTELOCLSDKROOT", "x64", "x86"),
+            // E.g. "c:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0\lib\Win32\"
+            ("CUDA_PATH", "x64", "Win32"),
+            // E.g. "C:\Program Files (x86)\AMD APP SDK\3.0\lib\x86\"
+            ("AMDAPPSDKROOT", "x86_64", "x86"),            
+        ];
+
+        for info in known_sdk.iter() {
+            if let Ok(sdk) = std::env::var(info.0) {
+                let mut path = std::path::PathBuf::from(sdk);
+                path.push("lib");
+                path.push(if cfg!(target_arch="x86_64") { info.1 } else { info.2 });
+                println!("cargo:rustc-link-search=native={}", path.display());
+            }
+        }
+
+        println!("cargo:rustc-link-search=native=C:\\Program Files (x86)\\OCL_SDK_Light\\lib\\x86_64");        
+    }
+}
\ No newline at end of file
diff --git a/notcuda/src/cuda.rs b/notcuda/src/cuda.rs
index a18ebf9..335da4a 100644
--- a/notcuda/src/cuda.rs
+++ b/notcuda/src/cuda.rs
@@ -2210,12 +2210,12 @@ pub extern "C" fn cuDriverGetVersion(driverVersion: *mut ::std::os::raw::c_int)
 
 #[cfg_attr(not(test), no_mangle)]
 pub extern "C" fn cuDeviceGet(device: *mut CUdevice, ordinal: ::std::os::raw::c_int) -> CUresult {
-    r#impl::device::get(device.decuda(), ordinal)
+    r#impl::device::get(device.decuda(), ordinal).encuda()
 }
 
 #[cfg_attr(not(test), no_mangle)]
 pub extern "C" fn cuDeviceGetCount(count: *mut ::std::os::raw::c_int) -> CUresult {
-    r#impl::device::get_count(count)
+    r#impl::device::get_count(count).encuda()
 }
 
 #[cfg_attr(not(test), no_mangle)]
@@ -2314,7 +2314,6 @@ pub extern "C" fn cuDevicePrimaryCtxReset(dev: CUdevice) -> CUresult {
     cuDevicePrimaryCtxReset_v2(dev)
 }
 
-
 #[cfg_attr(not(test), no_mangle)]
 pub extern "C" fn cuDevicePrimaryCtxReset_v2(dev: CUdevice) -> CUresult {
     r#impl::unimplemented()
@@ -2331,7 +2330,7 @@ pub extern "C" fn cuCtxCreate_v2(
 
 #[cfg_attr(not(test), no_mangle)]
 pub extern "C" fn cuCtxDestroy_v2(ctx: CUcontext) -> CUresult {
-    r#impl::context::destroy_v2(ctx.decuda())
+    r#impl::context::destroy_v2(ctx.decuda()).encuda()
 }
 
 #[cfg_attr(not(test), no_mangle)]
@@ -2356,7 +2355,7 @@ pub extern "C" fn cuCtxGetCurrent(pctx: *mut CUcontext) -> CUresult {
 
 #[cfg_attr(not(test), no_mangle)]
 pub extern "C" fn cuCtxGetDevice(device: *mut CUdevice) -> CUresult {
-    r#impl::context::get_device(device.decuda())
+    r#impl::context::get_device(device.decuda()).encuda()
 }
 
 #[cfg_attr(not(test), no_mangle)]
@@ -2404,7 +2403,7 @@ pub extern "C" fn cuCtxGetApiVersion(
     ctx: CUcontext,
     version: *mut ::std::os::raw::c_uint,
 ) -> CUresult {
-    r#impl::context::get_api_version(ctx.decuda(), version)
+    r#impl::context::get_api_version(ctx.decuda(), version).encuda()
 }
 
 #[cfg_attr(not(test), no_mangle)]
@@ -2422,12 +2421,12 @@ pub extern "C" fn cuCtxResetPersistingL2Cache() -> CUresult {
 
 #[cfg_attr(not(test), no_mangle)]
 pub extern "C" fn cuCtxAttach(pctx: *mut CUcontext, flags: ::std::os::raw::c_uint) -> CUresult {
-    r#impl::unimplemented()
+    r#impl::context::attach(pctx.decuda(), flags).encuda()
 }
 
 #[cfg_attr(not(test), no_mangle)]
 pub extern "C" fn cuCtxDetach(ctx: CUcontext) -> CUresult {
-    r#impl::unimplemented()
+    r#impl::context::detach(ctx.decuda()).encuda()
 }
 
 #[cfg_attr(not(test), no_mangle)]
@@ -2443,7 +2442,7 @@ pub extern "C" fn cuModuleLoadData(
     module: *mut CUmodule,
     image: *const ::std::os::raw::c_void,
 ) -> CUresult {
-    r#impl::unimplemented()
+    r#impl::module::load_data(module.decuda(), image).encuda()
 }
 
 #[cfg_attr(not(test), no_mangle)]
@@ -2564,7 +2563,7 @@ pub extern "C" fn cuMemGetInfo_v2(free: *mut usize, total: *mut usize) -> CUresu
 
 #[cfg_attr(not(test), no_mangle)]
 pub extern "C" fn cuMemAlloc_v2(dptr: *mut CUdeviceptr, bytesize: usize) -> CUresult {
-    r#impl::memory::alloc_v2(dptr.decuda(), bytesize)
+    r#impl::memory::alloc_v2(dptr.decuda(), bytesize).encuda()
 }
 
 #[cfg_attr(not(test), no_mangle)]
@@ -3281,7 +3280,7 @@ pub extern "C" fn cuStreamCreate(
     phStream: *mut CUstream,
     Flags: ::std::os::raw::c_uint,
 ) -> CUresult {
-    r#impl::unimplemented()
+    r#impl::stream::create(phStream.decuda(), Flags).encuda()
 }
 
 #[cfg_attr(not(test), no_mangle)]
@@ -3311,7 +3310,7 @@ pub extern "C" fn cuStreamGetFlags(
 
 #[cfg_attr(not(test), no_mangle)]
 pub extern "C" fn cuStreamGetCtx(hStream: CUstream, pctx: *mut CUcontext) -> CUresult {
-    r#impl::unimplemented()
+    r#impl::stream::get_ctx(hStream.decuda(), pctx.decuda()).encuda()
 }
 
 #[cfg_attr(not(test), no_mangle)]
@@ -3390,7 +3389,7 @@ pub extern "C" fn cuStreamSynchronize(hStream: CUstream) -> CUresult {
 
 #[cfg_attr(not(test), no_mangle)]
 pub extern "C" fn cuStreamDestroy_v2(hStream: CUstream) -> CUresult {
-    r#impl::unimplemented()
+    r#impl::stream::destroy_v2(hStream.decuda()).encuda()
 }
 
 #[cfg_attr(not(test), no_mangle)]
diff --git a/notcuda/src/impl/context.rs b/notcuda/src/impl/context.rs
index 91d4460..9689ecf 100644
--- a/notcuda/src/impl/context.rs
+++ b/notcuda/src/impl/context.rs
@@ -1,18 +1,15 @@
-use super::CUresult;
-use super::{device, HasLivenessCookie, LiveCheck};
+use super::{device, stream::Stream, stream::StreamData, HasLivenessCookie, LiveCheck};
+use super::{CUresult, GlobalState};
 use crate::{cuda::CUcontext, cuda_impl};
 use l0::sys::ze_result_t;
-use std::mem::{self, ManuallyDrop};
+use std::{cell::RefCell, num::NonZeroU32, os::raw::c_uint, ptr, sync::atomic::AtomicU32};
 use std::{
-    cell::RefCell,
-    num::NonZeroU32,
-    os::raw::c_uint,
-    ptr,
-    sync::{atomic::AtomicU32, Mutex},
+    collections::HashSet,
+    mem::{self},
 };
 
 thread_local! {
-    pub static CONTEXT_STACK: RefCell<Vec<*const Context>> = RefCell::new(Vec::new());
+    pub static CONTEXT_STACK: RefCell<Vec<*mut Context>> = RefCell::new(Vec::new());
 }
 
 pub type Context = LiveCheck<ContextData>;
@@ -23,6 +20,17 @@ impl HasLivenessCookie for ContextData {
 
     #[cfg(target_pointer_width = "32")]
     const COOKIE: usize = 0x0b643ffb;
+
+    const LIVENESS_FAIL: CUresult = CUresult::CUDA_ERROR_INVALID_CONTEXT;
+
+    fn try_drop(&mut self) -> Result<(), CUresult> {
+        for stream in self.streams.iter() {
+            let stream = unsafe { &mut **stream };
+            stream.context = ptr::null_mut();
+            Stream::destroy_impl(unsafe { Stream::ptr_from_inner(stream) })?;
+        }
+        Ok(())
+    }
 }
 
 enum ContextRefCount {
@@ -67,26 +75,16 @@ impl ContextRefCount {
             }
         }
     }
-
-    fn is_primary(&self) -> bool {
-        match self {
-            ContextRefCount::Primary => true,
-            ContextRefCount::NonPrimary(_) => false,
-        }
-    }
 }
 
 pub struct ContextData {
     pub flags: AtomicU32,
-    pub device_index: device::Index,
     // This pointer is null only for a moment when constructing primary context
-    pub device: *const Mutex<device::Device>,
-    // The split between mutable / non-mutable is mainly to avoid recursive locking in cuDevicePrimaryCtxGetState
-    pub mutable: Mutex<ContextDataMutable>,
-}
-
-pub struct ContextDataMutable {
+    pub device: *mut device::Device,
     ref_count: ContextRefCount,
+    pub default_stream: StreamData,
+    pub streams: HashSet<*mut StreamData>,
+    // All the fields below are here to support internal CUDA driver API
     pub cuda_manager: *mut cuda_impl::rt::ContextStateManager,
     pub cuda_state: *mut cuda_impl::rt::ContextState,
     pub cuda_dtor_cb: Option<
@@ -100,63 +98,75 @@ pub struct ContextDataMutable {
 
 impl ContextData {
     pub fn new(
+        l0_ctx: &mut l0::Context,
+        l0_dev: &l0::Device,
         flags: c_uint,
         is_primary: bool,
-        dev_index: device::Index,
-        dev: *const Mutex<device::Device>,
-    ) -> Self {
-        ContextData {
+        dev: *mut device::Device,
+    ) -> Result<Self, CUresult> {
+        let default_stream = StreamData::new_unitialized(l0_ctx, l0_dev)?;
+        Ok(ContextData {
             flags: AtomicU32::new(flags),
-            device_index: dev_index,
             device: dev,
-            mutable: Mutex::new(ContextDataMutable {
-                ref_count: ContextRefCount::new(is_primary),
-                cuda_manager: ptr::null_mut(),
-                cuda_state: ptr::null_mut(),
-                cuda_dtor_cb: None,
-            }),
-        }
+            ref_count: ContextRefCount::new(is_primary),
+            default_stream,
+            streams: HashSet::new(),
+            cuda_manager: ptr::null_mut(),
+            cuda_state: ptr::null_mut(),
+            cuda_dtor_cb: None,
+        })
     }
 }
 
-pub fn create_v2(pctx: *mut *mut Context, flags: u32, dev_idx: device::Index) -> CUresult {
+impl Context {
+    pub fn late_init(&mut self) {
+        let ctx_data = self.as_option_mut().unwrap();
+        ctx_data.default_stream.context = ctx_data as *mut _;
+    }
+}
+
+pub fn create_v2(
+    pctx: *mut *mut Context,
+    flags: u32,
+    dev_idx: device::Index,
+) -> Result<(), CUresult> {
     if pctx == ptr::null_mut() {
-        return CUresult::CUDA_ERROR_INVALID_VALUE;
+        return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
     }
-    let dev = device::get_device_ref(dev_idx);
-    let dev = match dev {
-        Ok(d) => d,
-        Err(e) => return e,
-    };
-    let mut ctx = Box::new(LiveCheck::new(ContextData::new(flags, false, dev_idx, dev)));
-    let ctx_ref = ctx.as_mut() as *mut Context;
+    let mut ctx_box = GlobalState::lock_device(dev_idx, |dev| {
+        let dev_ptr = dev as *mut _;
+        let mut ctx_box = Box::new(LiveCheck::new(ContextData::new(
+            &mut dev.l0_context,
+            &dev.base,
+            flags,
+            false,
+            dev_ptr as *mut _,
+        )?));
+        ctx_box.late_init();
+        Ok::<_, CUresult>(ctx_box)
+    })??;
+    let ctx_ref = ctx_box.as_mut() as *mut Context;
     unsafe { *pctx = ctx_ref };
-    mem::forget(ctx);
+    mem::forget(ctx_box);
     CONTEXT_STACK.with(|stack| stack.borrow_mut().push(ctx_ref));
-    CUresult::CUDA_SUCCESS
+    Ok(())
 }
 
-pub fn destroy_v2(ctx: *mut Context) -> CUresult {
+pub fn destroy_v2(ctx: *mut Context) -> Result<(), CUresult> {
     if ctx == ptr::null_mut() {
-        return CUresult::CUDA_ERROR_INVALID_VALUE;
+        return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
     }
     CONTEXT_STACK.with(|stack| {
         let mut stack = stack.borrow_mut();
         let should_pop = match stack.last() {
-            Some(active_ctx) => *active_ctx == (ctx as *const _),
+            Some(active_ctx) => *active_ctx == (ctx as *mut _),
             None => false,
         };
         if should_pop {
             stack.pop();
         }
     });
-    let mut ctx_box = ManuallyDrop::new(unsafe { Box::from_raw(ctx) });
-    if !ctx_box.try_drop() {
-        CUresult::CUDA_ERROR_INVALID_CONTEXT
-    } else {
-        unsafe { ManuallyDrop::drop(&mut ctx_box) };
-        CUresult::CUDA_SUCCESS
-    }
+    GlobalState::lock(|_| Context::destroy_impl(ctx))?
 }
 
 pub fn pop_current_v2(pctx: *mut *mut Context) -> CUresult {
@@ -172,17 +182,6 @@ pub fn pop_current_v2(pctx: *mut *mut Context) -> CUresult {
     CUresult::CUDA_SUCCESS
 }
 
-pub fn with_current<F: FnOnce(&ContextData) -> R, R>(f: F) -> Result<R, CUresult> {
-    CONTEXT_STACK.with(|stack| {
-        stack
-            .borrow()
-            .last()
-            .and_then(|c| unsafe { &**c }.as_ref())
-            .ok_or(CUresult::CUDA_ERROR_INVALID_CONTEXT)
-            .map(f)
-    })
-}
-
 pub fn get_current(pctx: *mut *mut Context) -> l0::Result<()> {
     if pctx == ptr::null_mut() {
         return Err(ze_result_t::ZE_RESULT_ERROR_INVALID_ARGUMENT);
@@ -205,37 +204,53 @@ pub fn set_current(ctx: *mut Context) -> CUresult {
     }
 }
 
-pub fn get_api_version(ctx: *mut Context, version: *mut u32) -> CUresult {
-    let _ctx = match unsafe { ctx.as_mut() } {
-        None => return CUresult::CUDA_ERROR_INVALID_VALUE,
-        Some(ctx) => match ctx.as_mut() {
-            None => return CUresult::CUDA_ERROR_INVALID_CONTEXT,
-            Some(ctx) => ctx,
-        },
-    };
+pub fn get_api_version(ctx: *mut Context, version: *mut u32) -> Result<(), CUresult> {
+    if ctx == ptr::null_mut() {
+        return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
+    }
+    GlobalState::lock(|_| {
+        unsafe { &*ctx }.as_result()?;
+        Ok::<_, CUresult>(())
+    })??;
     //TODO: query device for properties roughly matching CUDA API version
     unsafe { *version = 1100 };
-    CUresult::CUDA_SUCCESS
+    Ok(())
 }
 
-pub fn get_device(dev: *mut device::Index) -> CUresult {
-    let dev_idx = with_current(|ctx| ctx.device_index);
-    match dev_idx {
-        Ok(idx) => {
-            unsafe { *dev = idx }
-            CUresult::CUDA_SUCCESS
-        }
-        Err(err) => err,
+pub fn get_device(dev: *mut device::Index) -> Result<(), CUresult> {
+    let dev_idx = GlobalState::lock_current_context(|ctx| unsafe { &*ctx.device }.index)?;
+    unsafe { *dev = dev_idx };
+    Ok(())
+}
+
+pub fn attach(pctx: *mut *mut Context, _flags: c_uint) -> Result<(), CUresult> {
+    if pctx == ptr::null_mut() {
+        return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
     }
+    let ctx = GlobalState::lock_current_context_unchecked(|unchecked_ctx| {
+        let ctx = unchecked_ctx.as_result_mut()?;
+        ctx.ref_count.incr()?;
+        Ok::<_, CUresult>(unchecked_ctx as *mut _)
+    })??;
+    unsafe { *pctx = ctx };
+    Ok(())
+}
+
+pub fn detach(pctx: *mut Context) -> Result<(), CUresult> {
+    if pctx == ptr::null_mut() {
+        return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
+    }
+    GlobalState::lock_current_context_unchecked(|unchecked_ctx| {
+        let ctx = unchecked_ctx.as_result_mut()?;
+        if ctx.ref_count.decr() {
+            Context::destroy_impl(unchecked_ctx)?;
+        }
+        Ok::<_, CUresult>(())
+    })?
 }
 
 #[cfg(test)]
-pub fn is_context_stack_empty() -> bool {
-    CONTEXT_STACK.with(|stack| stack.borrow().is_empty())
-}
-
-#[cfg(test)]
-mod tests {
+mod test {
     use super::super::test::CudaDriverFns;
     use super::super::CUresult;
     use std::{ffi::c_void, ptr};
diff --git a/notcuda/src/impl/device.rs b/notcuda/src/impl/device.rs
index d4859d3..b8d263d 100644
--- a/notcuda/src/impl/device.rs
+++ b/notcuda/src/impl/device.rs
@@ -1,24 +1,21 @@
-use super::{context, transmute_lifetime, CUresult, Error};
+use super::{context, CUresult, GlobalState};
 use crate::cuda;
 use cuda::{CUdevice_attribute, CUuuid_st};
 use std::{
     cmp, mem,
     os::raw::{c_char, c_int},
     ptr,
-    sync::{
-        atomic::{AtomicU32, Ordering},
-        Mutex, MutexGuard,
-    },
+    sync::atomic::{AtomicU32, Ordering},
 };
 
 const PROJECT_URL_SUFFIX: &'static str = " [github.com/vosen/notCUDA]";
-static mut DEVICES: Option<Vec<Mutex<Device>>> = None;
 
 #[repr(transparent)]
-#[derive(Clone, Copy)]
+#[derive(Clone, Copy, Eq, PartialEq, Hash)]
 pub struct Index(pub c_int);
 
 pub struct Device {
+    pub index: Index,
     pub base: l0::Device,
     pub default_queue: l0::CommandQueue,
     pub l0_context: l0::Context,
@@ -33,17 +30,19 @@ unsafe impl Send for Device {}
 
 impl Device {
     // Unsafe because it does not fully initalize primary_context
-    unsafe fn new(drv: &l0::Driver, d: l0::Device, idx: usize) -> l0::Result<Self> {
+    unsafe fn new(drv: &l0::Driver, l0_dev: l0::Device, idx: usize) -> Result<Self, CUresult> {
         let mut ctx = l0::Context::new(drv)?;
-        let queue = l0::CommandQueue::new(&mut ctx, &d)?;
+        let queue = l0::CommandQueue::new(&mut ctx, &l0_dev)?;
         let primary_context = context::Context::new(context::ContextData::new(
+            &mut ctx,
+            &l0_dev,
             0,
             true,
-            Index(idx as c_int),
-            ptr::null(),
-        ));
+            ptr::null_mut(),
+        )?);
         Ok(Self {
-            base: d,
+            index: Index(idx as c_int),
+            base: l0_dev,
             default_queue: queue,
             l0_context: ctx,
             primary_context: primary_context,
@@ -93,83 +92,53 @@ impl Device {
             Err(e) => Err(e),
         }
     }
+
+    pub fn late_init(&mut self) {
+        self.primary_context.as_option_mut().unwrap().device = self as *mut _;
+    }
 }
 
-pub fn init(driver: &l0::Driver) -> l0::Result<()> {
+pub fn init(driver: &l0::Driver) -> Result<Vec<Device>, CUresult> {
     let ze_devices = driver.devices()?;
     let mut devices = ze_devices
         .into_iter()
         .enumerate()
-        .map(|(idx, d)| unsafe { Device::new(driver, d, idx) }.map(Mutex::new))
+        .map(|(idx, d)| unsafe { Device::new(driver, d, idx) })
         .collect::<Result<Vec<_>, _>>()?;
-    for d in devices.iter_mut() {
-        d.get_mut()
-            .unwrap()
-            .primary_context
-            .as_mut()
-            .unwrap()
-            .device = d;
+    for dev in devices.iter_mut() {
+        dev.late_init();
+        dev.primary_context.late_init();
     }
-    unsafe { DEVICES = Some(devices) };
+    Ok(devices)
+}
+
+pub fn get_count(count: *mut c_int) -> Result<(), CUresult> {
+    let len = GlobalState::lock(|state| state.devices.len())?;
+    unsafe { *count = len as c_int };
     Ok(())
 }
 
-fn devices() -> Result<&'static Vec<Mutex<Device>>, CUresult> {
-    match unsafe { &DEVICES } {
-        Some(devs) => Ok(devs),
-        None => Err(CUresult::CUDA_ERROR_NOT_INITIALIZED),
-    }
-}
-
-pub fn get_device_ref(Index(dev_idx): Index) -> Result<&'static Mutex<Device>, CUresult> {
-    let devs = devices()?;
-    if dev_idx < 0 || dev_idx >= devs.len() as c_int {
-        return Err(CUresult::CUDA_ERROR_INVALID_DEVICE);
-    }
-    Ok(&devs[dev_idx as usize])
-}
-
-pub fn get_device(dev_idx: Index) -> Result<MutexGuard<'static, Device>, CUresult> {
-    let dev = get_device_ref(dev_idx)?;
-    dev.lock().map_err(|_| CUresult::CUDA_ERROR_ILLEGAL_STATE)
-}
-
-pub fn get_count(count: *mut c_int) -> CUresult {
-    let len = devices().map(|d| d.len());
-    match len {
-        Ok(len) => {
-            unsafe { *count = len as c_int };
-            CUresult::CUDA_SUCCESS
-        }
-        Err(e) => e,
-    }
-}
-
-pub fn get(device: *mut Index, ordinal: c_int) -> CUresult {
+pub fn get(device: *mut Index, ordinal: c_int) -> Result<(), CUresult> {
     if device == ptr::null_mut() || ordinal < 0 {
-        return CUresult::CUDA_ERROR_INVALID_VALUE;
+        return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
     }
-    let len = devices().map(|d| d.len());
-    match len {
-        Ok(len) if ordinal < (len as i32) => {
-            unsafe { *device = Index(ordinal) };
-            CUresult::CUDA_SUCCESS
-        }
-        Ok(_) => CUresult::CUDA_ERROR_INVALID_VALUE,
-        Err(e) => e,
+    let len = GlobalState::lock(|state| state.devices.len())?;
+    if ordinal < (len as i32) {
+        unsafe { *device = Index(ordinal) };
+        Ok(())
+    } else {
+        Err(CUresult::CUDA_ERROR_INVALID_VALUE)
     }
 }
 
-pub fn get_name(name: *mut c_char, len: i32, dev: Index) -> Result<(), CUresult> {
+pub fn get_name(name: *mut c_char, len: i32, dev_idx: Index) -> Result<(), CUresult> {
     if name == ptr::null_mut() || len < 0 {
         return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
     }
-    // This is safe because devices are 'static
-    let name_ptr = {
-        let mut dev = get_device(dev)?;
-        let props = dev.get_properties().map_err(Into::<CUresult>::into)?;
-        props.name.as_ptr()
-    };
+    let name_ptr = GlobalState::lock_device(dev_idx, |dev| {
+        let props = dev.get_properties()?;
+        Ok::<_, l0::sys::ze_result_t>(props.name.as_ptr())
+    })??;
     let name_len = (0..256)
         .position(|i| unsafe { *name_ptr.add(i) } == 0)
         .unwrap_or(256);
@@ -189,20 +158,14 @@ pub fn get_name(name: *mut c_char, len: i32, dev: Index) -> Result<(), CUresult>
     Ok(())
 }
 
-pub fn total_mem_v2(bytes: *mut usize, dev: Index) -> Result<(), CUresult> {
+pub fn total_mem_v2(bytes: *mut usize, dev_idx: Index) -> Result<(), CUresult> {
     if bytes == ptr::null_mut() {
         return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
     }
-    // This is safe because devices are 'static
-    let mem_props = {
-        let mut dev = get_device(dev)?;
-        unsafe {
-            transmute_lifetime(
-                dev.get_memory_properties()
-                    .map_err(Into::<CUresult>::into)?,
-            )
-        }
-    };
+    let mem_props = GlobalState::lock_device(dev_idx, |dev| {
+        let mem_props = dev.get_memory_properties()?;
+        Ok::<_, l0::sys::ze_result_t>(mem_props)
+    })??;
     let max_mem = mem_props
         .iter()
         .map(|p| p.totalSize)
@@ -228,56 +191,101 @@ impl CUdevice_attribute {
     }
 }
 
-pub fn get_attribute(pi: *mut i32, attrib: CUdevice_attribute, dev: Index) -> Result<(), Error> {
+pub fn get_attribute(
+    pi: *mut i32,
+    attrib: CUdevice_attribute,
+    dev_idx: Index,
+) -> Result<(), CUresult> {
     if pi == ptr::null_mut() {
-        return Err(Error::Cuda(CUresult::CUDA_ERROR_INVALID_VALUE));
+        return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
     }
     if let Some(value) = attrib.get_static_value() {
         unsafe { *pi = value };
         return Ok(());
     }
-    let mut dev = get_device(dev).map_err(Error::Cuda)?;
     let value = match attrib {
         CUdevice_attribute::CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT => {
-            dev.get_properties().map_err(Error::L0)?.maxHardwareContexts as i32
+            GlobalState::lock_device(dev_idx, |dev| {
+                let props = dev.get_properties()?;
+                Ok::<_, l0::sys::ze_result_t>(props.maxHardwareContexts as i32)
+            })??
         }
         CUdevice_attribute::CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT => {
-            let props = dev.get_properties().map_err(Error::L0)?;
-            (props.numSlices * props.numSubslicesPerSlice * props.numEUsPerSubslice) as i32
+            GlobalState::lock_device(dev_idx, |dev| {
+                let props = dev.get_properties()?;
+                Ok::<_, l0::sys::ze_result_t>(
+                    (props.numSlices * props.numSubslicesPerSlice * props.numEUsPerSubslice) as i32,
+                )
+            })??
+        }
+        CUdevice_attribute::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_WIDTH => {
+            GlobalState::lock_device(dev_idx, |dev| {
+                let props = dev.get_image_properties()?;
+                Ok::<_, l0::sys::ze_result_t>(cmp::min(
+                    props.maxImageDims1D,
+                    c_int::max_value() as u32,
+                ) as c_int)
+            })??
         }
-        CUdevice_attribute::CU_DEVICE_ATTRIBUTE_MAXIMUM_TEXTURE1D_WIDTH => cmp::min(
-            dev.get_image_properties()
-                .map_err(Error::L0)?
-                .maxImageDims1D,
-            c_int::max_value() as u32,
-        ) as c_int,
         CUdevice_attribute::CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_X => {
-            let props = dev.get_compute_properties().map_err(Error::L0)?;
-            cmp::max(i32::max_value() as u32, props.maxGroupCountX) as i32
+            GlobalState::lock_device(dev_idx, |dev| {
+                let props = dev.get_compute_properties()?;
+                Ok::<_, l0::sys::ze_result_t>(cmp::max(
+                    i32::max_value() as u32,
+                    props.maxGroupCountX,
+                ) as i32)
+            })??
         }
         CUdevice_attribute::CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Y => {
-            let props = dev.get_compute_properties().map_err(Error::L0)?;
-            cmp::max(i32::max_value() as u32, props.maxGroupCountY) as i32
+            GlobalState::lock_device(dev_idx, |dev| {
+                let props = dev.get_compute_properties()?;
+                Ok::<_, l0::sys::ze_result_t>(cmp::max(
+                    i32::max_value() as u32,
+                    props.maxGroupCountY,
+                ) as i32)
+            })??
         }
         CUdevice_attribute::CU_DEVICE_ATTRIBUTE_MAX_GRID_DIM_Z => {
-            let props = dev.get_compute_properties().map_err(Error::L0)?;
-            cmp::max(i32::max_value() as u32, props.maxGroupCountZ) as i32
+            GlobalState::lock_device(dev_idx, |dev| {
+                let props = dev.get_compute_properties()?;
+                Ok::<_, l0::sys::ze_result_t>(cmp::max(
+                    i32::max_value() as u32,
+                    props.maxGroupCountZ,
+                ) as i32)
+            })??
         }
         CUdevice_attribute::CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_X => {
-            let props = dev.get_compute_properties().map_err(Error::L0)?;
-            cmp::max(i32::max_value() as u32, props.maxGroupSizeX) as i32
+            GlobalState::lock_device(dev_idx, |dev| {
+                let props = dev.get_compute_properties()?;
+                Ok::<_, l0::sys::ze_result_t>(
+                    cmp::max(i32::max_value() as u32, props.maxGroupSizeX) as i32,
+                )
+            })??
         }
         CUdevice_attribute::CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Y => {
-            let props = dev.get_compute_properties().map_err(Error::L0)?;
-            cmp::max(i32::max_value() as u32, props.maxGroupSizeY) as i32
+            GlobalState::lock_device(dev_idx, |dev| {
+                let props = dev.get_compute_properties()?;
+                Ok::<_, l0::sys::ze_result_t>(
+                    cmp::max(i32::max_value() as u32, props.maxGroupSizeY) as i32,
+                )
+            })??
         }
         CUdevice_attribute::CU_DEVICE_ATTRIBUTE_MAX_BLOCK_DIM_Z => {
-            let props = dev.get_compute_properties().map_err(Error::L0)?;
-            cmp::max(i32::max_value() as u32, props.maxGroupSizeZ) as i32
+            GlobalState::lock_device(dev_idx, |dev| {
+                let props = dev.get_compute_properties()?;
+                Ok::<_, l0::sys::ze_result_t>(
+                    cmp::max(i32::max_value() as u32, props.maxGroupSizeZ) as i32,
+                )
+            })??
         }
         CUdevice_attribute::CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK => {
-            let props = dev.get_compute_properties().map_err(Error::L0)?;
-            cmp::max(i32::max_value() as u32, props.maxTotalGroupSize) as i32
+            GlobalState::lock_device(dev_idx, |dev| {
+                let props = dev.get_compute_properties()?;
+                Ok::<_, l0::sys::ze_result_t>(cmp::max(
+                    i32::max_value() as u32,
+                    props.maxTotalGroupSize,
+                ) as i32)
+            })??
         }
         _ => {
             // TODO: support more attributes for CUDA runtime
@@ -293,14 +301,11 @@ pub fn get_attribute(pi: *mut i32, attrib: CUdevice_attribute, dev: Index) -> Re
     Ok(())
 }
 
-pub fn get_uuid(uuid: *mut CUuuid_st, dev: Index) -> Result<(), Error> {
-    let ze_uuid = {
-        get_device(dev)
-            .map_err(Error::Cuda)?
-            .get_properties()
-            .map_err(Error::L0)?
-            .uuid
-    };
+pub fn get_uuid(uuid: *mut CUuuid_st, dev_idx: Index) -> Result<(), CUresult> {
+    let ze_uuid = GlobalState::lock_device(dev_idx, |dev| {
+        let props = dev.get_properties()?;
+        Ok::<_, l0::sys::ze_result_t>(props.uuid)
+    })??;
     unsafe {
         *uuid = CUuuid_st {
             bytes: mem::transmute(ze_uuid.id),
@@ -309,53 +314,39 @@ pub fn get_uuid(uuid: *mut CUuuid_st, dev: Index) -> Result<(), Error> {
     Ok(())
 }
 
-pub fn with_current_exclusive<F: FnOnce(&mut Device) -> R, R>(f: F) -> Result<R, CUresult> {
-    let dev = super::context::with_current(|ctx| ctx.device);
-    dev.and_then(|dev| {
-        unsafe { &*dev }
-            .try_lock()
-            .map(|mut dev| f(&mut dev))
-            .map_err(|_| CUresult::CUDA_ERROR_ILLEGAL_STATE)
-    })
-}
-
-pub fn with_exclusive<F: FnOnce(&mut Device) -> R, R>(dev: Index, f: F) -> Result<R, CUresult> {
-    let dev = get_device_ref(dev)?;
-    dev.try_lock()
-        .map(|mut dev| f(&mut dev))
-        .map_err(|_| CUresult::CUDA_ERROR_ILLEGAL_STATE)
-}
-
 pub fn primary_ctx_get_state(
-    idx: Index,
+    dev_idx: Index,
     flags: *mut u32,
     active: *mut i32,
 ) -> Result<(), CUresult> {
-    let (ctx_ptr, flags_ptr) = with_exclusive(idx, |dev| {
+    let (is_active, flags_value) = GlobalState::lock_device(dev_idx, |dev| {
         // This is safe because primary context can't be dropped
-        let ctx_ptr = &dev.primary_context as *const _;
+        let ctx_ptr = &mut dev.primary_context as *mut _;
         let flags_ptr =
             (&unsafe { dev.primary_context.as_ref_unchecked() }.flags) as *const AtomicU32;
-        (ctx_ptr, flags_ptr)
-    })?;
-    let is_active = context::CONTEXT_STACK
-        .with(|stack| stack.borrow().last().map(|x| *x))
-        .map(|current| current == ctx_ptr)
-        .unwrap_or(false);
-    let flags_value = unsafe { &*flags_ptr }.load(Ordering::Relaxed);
-    unsafe { *flags = flags_value };
+        let is_active = context::CONTEXT_STACK
+            .with(|stack| stack.borrow().last().map(|x| *x))
+            .map(|current| current == ctx_ptr)
+            .unwrap_or(false);
+        let flags_value = unsafe { &*flags_ptr }.load(Ordering::Relaxed);
+        Ok::<_, l0::sys::ze_result_t>((is_active, flags_value))
+    })??;
     unsafe { *active = if is_active { 1 } else { 0 } };
+    unsafe { *flags = flags_value };
     Ok(())
 }
 
-pub fn primary_ctx_retain(pctx: *mut *mut context::Context, dev: Index) -> Result<(), CUresult> {
-    let ctx_ptr = with_exclusive(dev, |dev| &mut dev.primary_context as *mut _)?;
+pub fn primary_ctx_retain(
+    pctx: *mut *mut context::Context,
+    dev_idx: Index,
+) -> Result<(), CUresult> {
+    let ctx_ptr = GlobalState::lock_device(dev_idx, |dev| &mut dev.primary_context as *mut _)?;
     unsafe { *pctx = ctx_ptr };
     Ok(())
 }
 
 #[cfg(test)]
-mod tests {
+mod test {
     use super::super::test::CudaDriverFns;
     use super::super::CUresult;
 
diff --git a/notcuda/src/impl/export_table.rs b/notcuda/src/impl/export_table.rs
index 562af37..ae9f6e3 100644
--- a/notcuda/src/impl/export_table.rs
+++ b/notcuda/src/impl/export_table.rs
@@ -4,7 +4,7 @@ use crate::{
     cuda_impl,
 };
 
-use super::{context, device, module, Decuda, Encuda};
+use super::{context, context::ContextData, module, Decuda, Encuda, GlobalState};
 use std::mem;
 use std::os::raw::{c_uint, c_ulong, c_ushort};
 use std::{
@@ -110,17 +110,8 @@ static CUDART_INTERFACE_VTABLE: [VTableEntry; CUDART_INTERFACE_LENGTH] = [
     VTableEntry { ptr: ptr::null() },
 ];
 
-unsafe extern "C" fn cudart_interface_fn1(pctx: *mut CUcontext, dev: CUdevice) -> CUresult {
-    cudart_interface_fn1_impl(pctx.decuda(), dev.decuda()).encuda()
-}
-
-fn cudart_interface_fn1_impl(
-    pctx: *mut *mut context::Context,
-    dev: device::Index,
-) -> Result<(), CUresult> {
-    let ctx_ptr = device::with_exclusive(dev, |d| &mut d.primary_context as *mut context::Context)?;
-    unsafe { *pctx = ctx_ptr };
-    Ok(())
+unsafe extern "C" fn cudart_interface_fn1(_pctx: *mut CUcontext, _dev: CUdevice) -> CUresult {
+    super::unimplemented()
 }
 
 /*
@@ -200,7 +191,7 @@ unsafe extern "C" fn get_module_from_cubin(
     ptr1: *mut c_void,
     ptr2: *mut c_void,
 ) -> CUresult {
-    // Not sure what those twoparameters are actually used for,
+    // Not sure what those two parameters are actually used for,
     // they are somehow involved in __cudaRegisterHostVar
     if ptr1 != ptr::null_mut() || ptr2 != ptr::null_mut() {
         return CUresult::CUDA_ERROR_NOT_SUPPORTED;
@@ -234,10 +225,13 @@ unsafe extern "C" fn get_module_from_cubin(
             },
             Err(_) => continue,
         };
-        let module = module::ModuleData::compile_spirv(kernel_text_string);
+        let module = module::SpirvModule::new(kernel_text_string);
         match module {
             Ok(module) => {
-                *result = Box::into_raw(Box::new(module));
+                match module::load_data_impl(result, module) {
+                    Ok(()) => {}
+                    Err(err) => return err,
+                }
                 return CUresult::CUDA_SUCCESS;
             }
             Err(_) => continue,
@@ -309,7 +303,7 @@ unsafe extern "C" fn context_local_storage_ctor(
 }
 
 fn context_local_storage_ctor_impl(
-    mut cu_ctx: *mut context::Context,
+    cu_ctx: *mut context::Context,
     mgr: *mut cuda_impl::rt::ContextStateManager,
     ctx_state: *mut cuda_impl::rt::ContextState,
     dtor_cb: Option<
@@ -320,26 +314,11 @@ fn context_local_storage_ctor_impl(
         ),
     >,
 ) -> Result<(), CUresult> {
-    if cu_ctx == ptr::null_mut() {
-        context::get_current(&mut cu_ctx)?;
-    }
-    if cu_ctx == ptr::null_mut() {
-        return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
-    }
-    unsafe { &*cu_ctx }
-        .as_ref()
-        .ok_or(CUresult::CUDA_ERROR_INVALID_CONTEXT)
-        .and_then(|ctx| {
-            ctx.mutable
-                .try_lock()
-                .map_err(|_| CUresult::CUDA_ERROR_ILLEGAL_STATE)
-                .map(|mut mutable| {
-                    mutable.cuda_manager = mgr;
-                    mutable.cuda_state = ctx_state;
-                    mutable.cuda_dtor_cb = dtor_cb;
-                })
-        })?;
-    Ok(())
+    lock_context(cu_ctx, |ctx: &mut ContextData| {
+        ctx.cuda_manager = mgr;
+        ctx.cuda_state = ctx_state;
+        ctx.cuda_dtor_cb = dtor_cb;
+    })
 }
 
 // some kind of dtor
@@ -357,24 +336,10 @@ unsafe extern "C" fn context_local_storage_get_state(
 
 fn context_local_storage_get_state_impl(
     ctx_state: *mut *mut cuda_impl::rt::ContextState,
-    mut cu_ctx: *mut context::Context,
+    cu_ctx: *mut context::Context,
     _: *mut cuda_impl::rt::ContextStateManager,
 ) -> Result<(), CUresult> {
-    if cu_ctx == ptr::null_mut() {
-        context::get_current(&mut cu_ctx)?;
-    }
-    if cu_ctx == ptr::null_mut() {
-        return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
-    }
-    let cuda_state = unsafe { &*cu_ctx }
-        .as_ref()
-        .ok_or(CUresult::CUDA_ERROR_INVALID_CONTEXT)
-        .and_then(|ctx| {
-            ctx.mutable
-                .try_lock()
-                .map_err(|_| CUresult::CUDA_ERROR_ILLEGAL_STATE)
-                .map(|mutable| mutable.cuda_state)
-        })?;
+    let cuda_state = lock_context(cu_ctx, |ctx: &mut ContextData| ctx.cuda_state)?;
     if cuda_state == ptr::null_mut() {
         Err(CUresult::CUDA_ERROR_INVALID_VALUE)
     } else {
@@ -382,3 +347,17 @@ fn context_local_storage_get_state_impl(
         Ok(())
     }
 }
+
+fn lock_context<T>(
+    cu_ctx: *mut context::Context,
+    fn_impl: impl FnOnce(&mut ContextData) -> T,
+) -> Result<T, CUresult> {
+    if cu_ctx == ptr::null_mut() {
+        GlobalState::lock_current_context(fn_impl)
+    } else {
+        GlobalState::lock(|_| {
+            let ctx = unsafe { &mut *cu_ctx }.as_result_mut()?;
+            Ok(fn_impl(ctx))
+        })?
+    }
+}
diff --git a/notcuda/src/impl/function.rs b/notcuda/src/impl/function.rs
index 0ab3bea..394f806 100644
--- a/notcuda/src/impl/function.rs
+++ b/notcuda/src/impl/function.rs
@@ -1,11 +1,28 @@
 use ::std::os::raw::{c_uint, c_void};
 use std::ptr;
 
-use super::{device, stream::Stream, CUresult};
+use super::{CUresult, GlobalState, HasLivenessCookie, LiveCheck, stream::Stream};
 
-pub struct Function {
+pub type Function = LiveCheck<FunctionData>;
+
+impl HasLivenessCookie for FunctionData {
+    #[cfg(target_pointer_width = "64")]
+    const COOKIE: usize = 0x5e2ab14d5840678e;
+
+    #[cfg(target_pointer_width = "32")]
+    const COOKIE: usize = 0x33e6a1e6;
+
+    const LIVENESS_FAIL: CUresult = CUresult::CUDA_ERROR_INVALID_HANDLE;
+
+    fn try_drop(&mut self) -> Result<(), CUresult> {
+        Ok(())
+    }
+}
+
+pub struct FunctionData {
     pub base: l0::Kernel<'static>,
     pub arg_size: Vec<usize>,
+    pub use_shared_mem: bool,
 }
 
 pub fn launch_kernel(
@@ -17,36 +34,43 @@ pub fn launch_kernel(
     block_dim_y: c_uint,
     block_dim_z: c_uint,
     shared_mem_bytes: c_uint,
-    strean: *mut Stream,
+    hstream: *mut Stream,
     kernel_params: *mut *mut c_void,
     extra: *mut *mut c_void,
 ) -> Result<(), CUresult> {
     if f == ptr::null_mut() {
         return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
     }
-    if shared_mem_bytes != 0 || strean != ptr::null_mut() || extra != ptr::null_mut() {
+    if extra != ptr::null_mut() {
         return Err(CUresult::CUDA_ERROR_NOT_SUPPORTED);
     }
-    let func = unsafe { &*f };
-    for (i, arg_size) in func.arg_size.iter().copied().enumerate() {
-        unsafe {
-            func.base
-                .set_arg_raw(i as u32, arg_size, *kernel_params.add(i))?
-        };
-    }
-    unsafe { &*f }
-        .base
-        .set_group_size(block_dim_x, block_dim_y, block_dim_z)?;
-    device::with_current_exclusive(|dev| {
-        let mut cmd_list = l0::CommandList::new(&mut dev.l0_context, &dev.base)?;
+    GlobalState::lock_stream(hstream, |stream| {
+        let func: &mut FunctionData = unsafe { &mut *f }.as_result_mut()?;
+        for (i, arg_size) in func.arg_size.iter().enumerate() {
+            unsafe {
+                func.base
+                    .set_arg_raw(i as u32, *arg_size, *kernel_params.add(i))?
+            };
+        }
+        if func.use_shared_mem {
+            unsafe {
+                func.base.set_arg_raw(
+                    func.arg_size.len() as u32,
+                    shared_mem_bytes as usize,
+                    ptr::null(),
+                )?
+            };
+        }
+        func.base
+            .set_group_size(block_dim_x, block_dim_y, block_dim_z)?;
+        let mut cmd_list = stream.command_list()?;
         cmd_list.append_launch_kernel(
-            &unsafe { &*f }.base,
+            &mut func.base,
             &[grid_dim_x, grid_dim_y, grid_dim_z],
             None,
             &mut [],
         )?;
-        dev.default_queue.execute(cmd_list)?;
-        l0::Result::Ok(())
-    })??;
-    Ok(())
+        stream.queue.execute(cmd_list)?;
+        Ok(())
+    })?
 }
diff --git a/notcuda/src/impl/memory.rs b/notcuda/src/impl/memory.rs
index 439b26f..62dc1cc 100644
--- a/notcuda/src/impl/memory.rs
+++ b/notcuda/src/impl/memory.rs
@@ -1,57 +1,34 @@
-use super::CUresult;
+use super::{stream, CUresult, GlobalState};
 use std::ffi::c_void;
 
-pub fn alloc_v2(dptr: *mut *mut c_void, bytesize: usize) -> CUresult {
-    let alloc_result = super::device::with_current_exclusive(|dev| unsafe {
-        dev.base.mem_alloc_device(&mut dev.l0_context, bytesize, 0)
-    });
-    match alloc_result {
-        Ok(Ok(alloc)) => {
-            unsafe { *dptr = alloc };
-            CUresult::CUDA_SUCCESS
-        }
-        Ok(Err(e)) => e.into(),
-        Err(e) => e,
-    }
+pub fn alloc_v2(dptr: *mut *mut c_void, bytesize: usize) -> Result<(), CUresult> {
+    let ptr = GlobalState::lock_current_context(|ctx| {
+        let dev = unsafe { &mut *ctx.device };
+        Ok::<_, CUresult>(unsafe { dev.base.mem_alloc_device(&mut dev.l0_context, bytesize, 0) }?)
+    })??;
+    unsafe { *dptr = ptr };
+    Ok(())
 }
 
-pub fn copy_v2(
-    dst: *mut c_void,
-    src: *const c_void,
-    bytesize: usize,
-) -> Result<Result<(), l0::sys::ze_result_t>, CUresult> {
-    super::device::with_current_exclusive(|dev| unsafe {
-        memcpy_impl(
-            &mut dev.l0_context,
-            dst,
-            src,
-            bytesize,
-            &dev.base,
-            &mut dev.default_queue,
-        )
+pub fn copy_v2(dst: *mut c_void, src: *const c_void, bytesize: usize) -> Result<(), CUresult> {
+    GlobalState::lock_stream(stream::CU_STREAM_LEGACY, |stream| {
+        let mut cmd_list = stream.command_list()?;
+        unsafe { cmd_list.append_memory_copy_unsafe(dst, src, bytesize, None, &mut []) }?;
+        stream.queue.execute(cmd_list)?;
+        Ok::<_, CUresult>(())
+    })?
+}
+
+pub fn free_v2(ptr: *mut c_void) -> Result<(), CUresult> {
+    GlobalState::lock_current_context(|ctx| {
+        let dev = unsafe { &mut *ctx.device };
+        Ok::<_, CUresult>(unsafe { dev.l0_context.mem_free(ptr) }?)
     })
-}
-
-unsafe fn memcpy_impl(
-    ctx: &mut l0::Context,
-    dst: *mut c_void,
-    src: *const c_void,
-    bytes_count: usize,
-    dev: &l0::Device,
-    queue: &mut l0::CommandQueue,
-) -> l0::Result<()> {
-    let mut cmd_list = l0::CommandList::new(ctx, &dev)?;
-    cmd_list.append_memory_copy_unsafe(dst, src, bytes_count, None, &mut [])?;
-    queue.execute(cmd_list)?;
-    Ok(())
-}
-
-pub(crate) fn free_v2(_: *mut c_void)-> l0::Result<()>  {
-    Ok(())
+    .map_err(|_| CUresult::CUDA_ERROR_INVALID_VALUE)?
 }
 
 #[cfg(test)]
-mod tests {
+mod test {
     use super::super::test::CudaDriverFns;
     use super::super::CUresult;
     use std::ptr;
@@ -82,4 +59,20 @@ mod tests {
         assert_ne!(mem, ptr::null_mut());
         assert_eq!(T::cuCtxDestroy_v2(ctx), CUresult::CUDA_SUCCESS);
     }
+
+    cuda_driver_test!(free_without_ctx);
+
+    fn free_without_ctx<T: CudaDriverFns>() {
+        assert_eq!(T::cuInit(0), CUresult::CUDA_SUCCESS);
+        let mut ctx = ptr::null_mut();
+        assert_eq!(T::cuCtxCreate_v2(&mut ctx, 0, 0), CUresult::CUDA_SUCCESS);
+        let mut mem = ptr::null_mut();
+        assert_eq!(
+            T::cuMemAlloc_v2(&mut mem, std::mem::size_of::<usize>()),
+            CUresult::CUDA_SUCCESS
+        );
+        assert_ne!(mem, ptr::null_mut());
+        assert_eq!(T::cuCtxDestroy_v2(ctx), CUresult::CUDA_SUCCESS);
+        assert_eq!(T::cuMemFree_v2(mem), CUresult::CUDA_ERROR_INVALID_VALUE);
+    }
 }
diff --git a/notcuda/src/impl/mod.rs b/notcuda/src/impl/mod.rs
index 5a72ce4..770a32b 100644
--- a/notcuda/src/impl/mod.rs
+++ b/notcuda/src/impl/mod.rs
@@ -1,5 +1,15 @@
-use crate::cuda::{CUctx_st, CUdevice, CUdeviceptr, CUfunc_st, CUmod_st, CUresult, CUstream_st};
-use std::{ffi::c_void, mem::{self, ManuallyDrop}, os::raw::c_int, sync::Mutex};
+use crate::{
+    cuda::{CUctx_st, CUdevice, CUdeviceptr, CUfunc_st, CUmod_st, CUresult, CUstream_st},
+    r#impl::device::Device,
+};
+use std::{
+    ffi::c_void,
+    mem::{self, ManuallyDrop},
+    os::raw::c_int,
+    ptr,
+    sync::Mutex,
+    sync::TryLockError,
+};
 
 #[cfg(test)]
 #[macro_use]
@@ -7,9 +17,9 @@ pub mod test;
 pub mod context;
 pub mod device;
 pub mod export_table;
+pub mod function;
 pub mod memory;
 pub mod module;
-pub mod function;
 pub mod stream;
 
 #[cfg(debug_assertions)]
@@ -22,8 +32,11 @@ pub fn unimplemented() -> CUresult {
     CUresult::CUDA_ERROR_NOT_SUPPORTED
 }
 
-pub trait HasLivenessCookie {
+pub trait HasLivenessCookie: Sized {
     const COOKIE: usize;
+    const LIVENESS_FAIL: CUresult;
+
+    fn try_drop(&mut self) -> Result<(), CUresult>;
 }
 
 // This struct is a best-effort check if wrapped value has been dropped,
@@ -42,19 +55,23 @@ impl<T: HasLivenessCookie> LiveCheck<T> {
         }
     }
 
+    fn destroy_impl(this: *mut Self) -> Result<(), CUresult> {
+        let mut ctx_box = ManuallyDrop::new(unsafe { Box::from_raw(this) });
+        ctx_box.try_drop()?;
+        unsafe { ManuallyDrop::drop(&mut ctx_box) };
+        Ok(())
+    }
+
+    unsafe fn ptr_from_inner(this: *mut T) -> *mut Self {
+        let outer_ptr = (this as *mut u8).sub(mem::size_of::<usize>());
+        outer_ptr as *mut Self
+    }
+
     pub unsafe fn as_ref_unchecked(&self) -> &T {
         &self.data
     }
 
-    pub fn as_ref(&self) -> Option<&T> {
-        if self.cookie == T::COOKIE {
-            Some(&self.data)
-        } else {
-            None
-        }
-    }
-
-    pub fn as_mut(&mut self) -> Option<&mut T> {
+    pub fn as_option_mut(&mut self) -> Option<&mut T> {
         if self.cookie == T::COOKIE {
             Some(&mut self.data)
         } else {
@@ -62,14 +79,31 @@ impl<T: HasLivenessCookie> LiveCheck<T> {
         }
     }
 
+    pub fn as_result(&self) -> Result<&T, CUresult> {
+        if self.cookie == T::COOKIE {
+            Ok(&self.data)
+        } else {
+            Err(T::LIVENESS_FAIL)
+        }
+    }
+
+    pub fn as_result_mut(&mut self) -> Result<&mut T, CUresult> {
+        if self.cookie == T::COOKIE {
+            Ok(&mut self.data)
+        } else {
+            Err(T::LIVENESS_FAIL)
+        }
+    }
+
     #[must_use]
-    pub fn try_drop(&mut self) -> bool {
+    pub fn try_drop(&mut self) -> Result<(), CUresult> {
         if self.cookie == T::COOKIE {
             self.cookie = 0;
+            self.data.try_drop()?;
             unsafe { ManuallyDrop::drop(&mut self.data) };
-            return true;
+            return Ok(());
         }
-        false
+        Err(T::LIVENESS_FAIL)
     }
 }
 
@@ -121,6 +155,12 @@ impl From<l0::sys::ze_result_t> for CUresult {
     }
 }
 
+impl<T> From<TryLockError<T>> for CUresult {
+    fn from(_: TryLockError<T>) -> Self {
+        CUresult::CUDA_ERROR_ILLEGAL_STATE
+    }
+}
+
 pub trait Encuda {
     type To: Sized;
     fn encuda(self: Self) -> Self::To;
@@ -157,58 +197,103 @@ impl<T1: Encuda<To = CUresult>, T2: Encuda<To = CUresult>> Encuda for Result<T1,
     }
 }
 
-pub enum Error {
-    L0(l0::sys::ze_result_t),
-    Cuda(CUresult),
-}
-
-impl Encuda for Error {
-    type To = CUresult;
-    fn encuda(self: Self) -> Self::To {
-        match self {
-            Error::L0(e) => e.into(),
-            Error::Cuda(e) => e,
-        }
-    }
-}
-
 lazy_static! {
     static ref GLOBAL_STATE: Mutex<Option<GlobalState>> = Mutex::new(None);
 }
 
 struct GlobalState {
-    driver: l0::Driver,
+    devices: Vec<Device>,
 }
 
 unsafe impl Send for GlobalState {}
 
+impl GlobalState {
+    fn lock<T>(f: impl FnOnce(&mut GlobalState) -> T) -> Result<T, CUresult> {
+        let mut mutex = GLOBAL_STATE
+            .lock()
+            .unwrap_or_else(|poison| poison.into_inner());
+        let global_state = mutex.as_mut().ok_or(CUresult::CUDA_ERROR_ILLEGAL_STATE)?;
+        Ok(f(global_state))
+    }
+
+    fn lock_device<T>(
+        device::Index(dev_idx): device::Index,
+        f: impl FnOnce(&'static mut device::Device) -> T,
+    ) -> Result<T, CUresult> {
+        if dev_idx < 0 {
+            return Err(CUresult::CUDA_ERROR_INVALID_DEVICE);
+        }
+        Self::lock(|global_state| {
+            if dev_idx >= global_state.devices.len() as c_int {
+                Err(CUresult::CUDA_ERROR_INVALID_DEVICE)
+            } else {
+                Ok(f(unsafe {
+                    transmute_lifetime_mut(&mut global_state.devices[dev_idx as usize])
+                }))
+            }
+        })?
+    }
+
+    fn lock_current_context<F: FnOnce(&mut context::ContextData) -> R, R>(
+        f: F,
+    ) -> Result<R, CUresult> {
+        Self::lock_current_context_unchecked(|ctx| Ok(f(ctx.as_result_mut()?)))?
+    }
+
+    fn lock_current_context_unchecked<F: FnOnce(&mut context::Context) -> R, R>(
+        f: F,
+    ) -> Result<R, CUresult> {
+        context::CONTEXT_STACK.with(|stack| {
+            stack
+                .borrow_mut()
+                .last_mut()
+                .ok_or(CUresult::CUDA_ERROR_INVALID_CONTEXT)
+                .map(|ctx| GlobalState::lock(|_| f(unsafe { &mut **ctx })))?
+        })
+    }
+
+    fn lock_stream<T>(
+        stream: *mut stream::Stream,
+        f: impl FnOnce(&mut stream::StreamData) -> T,
+    ) -> Result<T, CUresult> {
+        if stream == ptr::null_mut()
+            || stream == stream::CU_STREAM_LEGACY
+            || stream == stream::CU_STREAM_PER_THREAD
+        {
+            Self::lock_current_context(|ctx| Ok(f(&mut ctx.default_stream)))?
+        } else {
+            Self::lock(|_| {
+                let stream = unsafe { &mut *stream }.as_result_mut()?;
+                Ok(f(stream))
+            })?
+        }
+    }
+}
+
 // TODO: implement
 fn is_intel_gpu_driver(_: &l0::Driver) -> bool {
     true
 }
 
-pub fn init() -> l0::Result<()> {
+pub fn init() -> Result<(), CUresult> {
     let mut global_state = GLOBAL_STATE
         .lock()
-        .map_err(|_| l0::sys::ze_result_t::ZE_RESULT_ERROR_UNKNOWN)?;
+        .map_err(|_| CUresult::CUDA_ERROR_UNKNOWN)?;
     if global_state.is_some() {
         return Ok(());
     }
     l0::init()?;
     let drivers = l0::Driver::get()?;
-    let driver = match drivers.into_iter().find(is_intel_gpu_driver) {
-        None => return Err(l0::sys::ze_result_t::ZE_RESULT_ERROR_UNKNOWN),
-        Some(driver) => {
-            device::init(&driver)?;
-            driver
-        }
+    let devices = match drivers.into_iter().find(is_intel_gpu_driver) {
+        None => return Err(CUresult::CUDA_ERROR_UNKNOWN),
+        Some(driver) => device::init(&driver)?,
     };
-    *global_state = Some(GlobalState { driver });
+    *global_state = Some(GlobalState { devices });
     drop(global_state);
     Ok(())
 }
 
-unsafe fn transmute_lifetime<'a, 'b, T: ?Sized>(t: &'a T) -> &'b T {
+unsafe fn transmute_lifetime_mut<'a, 'b, T: ?Sized>(t: &'a mut T) -> &'b mut T {
     mem::transmute(t)
 }
 
diff --git a/notcuda/src/impl/module.rs b/notcuda/src/impl/module.rs
index 35436c3..4422107 100644
--- a/notcuda/src/impl/module.rs
+++ b/notcuda/src/impl/module.rs
@@ -1,79 +1,90 @@
 use std::{
-    collections::HashMap, ffi::CStr, ffi::CString, mem, os::raw::c_char, ptr, slice, sync::Mutex,
+    collections::hash_map, collections::HashMap, ffi::c_void, ffi::CStr, ffi::CString, mem,
+    os::raw::c_char, ptr, slice,
 };
 
-use super::{function::Function, transmute_lifetime, CUresult};
+use super::{
+    device, function::Function, function::FunctionData, CUresult, GlobalState, HasLivenessCookie,
+    LiveCheck,
+};
 use ptx;
 
-pub type Module = Mutex<ModuleData>;
+pub type Module = LiveCheck<ModuleData>;
+
+impl HasLivenessCookie for ModuleData {
+    #[cfg(target_pointer_width = "64")]
+    const COOKIE: usize = 0xf1313bd46505f98a;
+
+    #[cfg(target_pointer_width = "32")]
+    const COOKIE: usize = 0xbdbe3f15;
+
+    const LIVENESS_FAIL: CUresult = CUresult::CUDA_ERROR_INVALID_HANDLE;
+
+    fn try_drop(&mut self) -> Result<(), CUresult> {
+        Ok(())
+    }
+}
 
 pub struct ModuleData {
-    base: l0::Module,
-    arg_lens: HashMap<CString, Vec<usize>>,
+    pub spirv: SpirvModule,
+    // This should be a Vec<>, but I'm feeling lazy
+    pub device_binaries: HashMap<device::Index, CompiledModule>,
 }
 
-pub enum ModuleCompileError<'a> {
-    Parse(
-        Vec<ptx::ast::PtxError>,
-        Option<ptx::ParseError<usize, ptx::Token<'a>, ptx::ast::PtxError>>,
-    ),
-    Compile(ptx::TranslateError),
-    L0(l0::sys::ze_result_t),
-    CUDA(CUresult),
+pub struct SpirvModule {
+    pub binaries: Vec<u32>,
+    pub kernel_info: HashMap<String, ptx::KernelInfo>,
+    pub should_link_ptx_impl: Option<&'static [u8]>,
+    pub build_options: CString,
 }
 
-impl<'a> ModuleCompileError<'a> {
-    pub fn get_build_log(&self) {}
+pub struct CompiledModule {
+    pub base: l0::Module,
+    pub kernels: HashMap<CString, Box<Function>>,
 }
 
-impl<'a> From<ptx::TranslateError> for ModuleCompileError<'a> {
-    fn from(err: ptx::TranslateError) -> Self {
-        ModuleCompileError::Compile(err)
+impl<L, T, E> From<ptx::ParseError<L, T, E>> for CUresult {
+    fn from(_: ptx::ParseError<L, T, E>) -> Self {
+        CUresult::CUDA_ERROR_INVALID_PTX
     }
 }
 
-impl<'a> From<l0::sys::ze_result_t> for ModuleCompileError<'a> {
-    fn from(err: l0::sys::ze_result_t) -> Self {
-        ModuleCompileError::L0(err)
+impl From<ptx::TranslateError> for CUresult {
+    fn from(_: ptx::TranslateError) -> Self {
+        CUresult::CUDA_ERROR_INVALID_PTX
     }
 }
 
-impl<'a> From<CUresult> for ModuleCompileError<'a> {
-    fn from(err: CUresult) -> Self {
-        ModuleCompileError::CUDA(err)
+impl SpirvModule {
+    pub fn new_raw<'a>(text: *const c_char) -> Result<Self, CUresult> {
+        let u8_text = unsafe { CStr::from_ptr(text) };
+        let ptx_text = u8_text
+            .to_str()
+            .map_err(|_| CUresult::CUDA_ERROR_INVALID_PTX)?;
+        Self::new(ptx_text)
     }
-}
 
-impl ModuleData {
-    pub fn compile_spirv<'a>(ptx_text: &'a str) -> Result<Module, ModuleCompileError<'a>> {
+    pub fn new<'a>(ptx_text: &str) -> Result<Self, CUresult> {
         let mut errors = Vec::new();
-        let ast = ptx::ModuleParser::new().parse(&mut errors, ptx_text);
-        let ast = match ast {
-            Err(e) => return Err(ModuleCompileError::Parse(errors, Some(e))),
-            Ok(_) if errors.len() > 0 => return Err(ModuleCompileError::Parse(errors, None)),
-            Ok(ast) => ast,
-        };
-        let (_, spirv, all_arg_lens) = ptx::to_spirv(ast)?;
+        let ast = ptx::ModuleParser::new().parse(&mut errors, ptx_text)?;
+        let spirv_module = ptx::to_spirv_module(ast)?;
+        Ok(SpirvModule {
+            binaries: spirv_module.assemble(),
+            kernel_info: spirv_module.kernel_info,
+            should_link_ptx_impl: spirv_module.should_link_ptx_impl,
+            build_options: spirv_module.build_options,
+        })
+    }
+
+    pub fn compile(&self, ctx: &mut l0::Context, dev: &l0::Device) -> Result<l0::Module, CUresult> {
         let byte_il = unsafe {
-            slice::from_raw_parts::<u8>(
-                spirv.as_ptr() as *const _,
-                spirv.len() * mem::size_of::<u32>(),
+            slice::from_raw_parts(
+                self.binaries.as_ptr() as *const u8,
+                self.binaries.len() * mem::size_of::<u32>(),
             )
         };
-        let module = super::device::with_current_exclusive(|dev| {
-            l0::Module::build_spirv(&mut dev.l0_context, &dev.base, byte_il, None)
-        });
-        match module {
-            Ok((Ok(module), _)) => Ok(Mutex::new(Self {
-                base: module,
-                arg_lens: all_arg_lens
-                    .into_iter()
-                    .map(|(k, v)| (CString::new(k).unwrap(), v))
-                    .collect(),
-            })),
-            Ok((Err(err), _)) => Err(ModuleCompileError::from(err)),
-            Err(err) => Err(ModuleCompileError::from(err)),
-        }
+        let l0_module = l0::Module::build_spirv(ctx, dev, byte_il, None).0?;
+        Ok(l0_module)
     }
 }
 
@@ -85,34 +96,75 @@ pub fn get_function(
     if hfunc == ptr::null_mut() || hmod == ptr::null_mut() || name == ptr::null() {
         return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
     }
-    let name = unsafe { CStr::from_ptr(name) };
-    let (mut kernel, args_len) = unsafe { &*hmod }
-        .try_lock()
-        .map(|module| {
-            Result::<_, CUresult>::Ok((
-                l0::Kernel::new_resident(unsafe { transmute_lifetime(&module.base) }, name)?,
-                module
-                    .arg_lens
-                    .get(name)
-                    .ok_or(CUresult::CUDA_ERROR_NOT_FOUND)?
-                    .clone(),
-            ))
-        })
-        .map_err(|_| CUresult::CUDA_ERROR_ILLEGAL_STATE)??;
-    kernel.set_indirect_access(
-        l0::sys::ze_kernel_indirect_access_flags_t::ZE_KERNEL_INDIRECT_ACCESS_FLAG_DEVICE
-            | l0::sys::ze_kernel_indirect_access_flags_t::ZE_KERNEL_INDIRECT_ACCESS_FLAG_HOST
-            | l0::sys::ze_kernel_indirect_access_flags_t::ZE_KERNEL_INDIRECT_ACCESS_FLAG_SHARED,
-    )?;
-    unsafe {
-        *hfunc = Box::into_raw(Box::new(Function {
-            base: kernel,
-            arg_size: args_len,
-        }))
-    };
+    let name = unsafe { CStr::from_ptr(name) }.to_owned();
+    let function: *mut Function = GlobalState::lock_current_context(|ctx| {
+        let module = unsafe { &mut *hmod }.as_result_mut()?;
+        let device = unsafe { &mut *ctx.device };
+        let compiled_module = match module.device_binaries.entry(device.index) {
+            hash_map::Entry::Occupied(entry) => entry.into_mut(),
+            hash_map::Entry::Vacant(entry) => {
+                let new_module = CompiledModule {
+                    base: module.spirv.compile(&mut device.l0_context, &device.base)?,
+                    kernels: HashMap::new(),
+                };
+                entry.insert(new_module)
+            }
+        };
+        //let compiled_module = unsafe { transmute_lifetime_mut(compiled_module) };
+        let kernel = match compiled_module.kernels.entry(name) {
+            hash_map::Entry::Occupied(entry) => entry.into_mut().as_mut(),
+            hash_map::Entry::Vacant(entry) => {
+                let kernel_info = module
+                    .spirv
+                    .kernel_info
+                    .get(unsafe {
+                        std::str::from_utf8_unchecked(entry.key().as_c_str().to_bytes())
+                    })
+                    .ok_or(CUresult::CUDA_ERROR_NOT_FOUND)?;
+                let kernel =
+                    l0::Kernel::new_resident(&compiled_module.base, entry.key().as_c_str())?;
+                entry.insert(Box::new(Function::new(FunctionData {
+                    base: kernel,
+                    arg_size: kernel_info.arguments_sizes.clone(),
+                    use_shared_mem: kernel_info.uses_shared_mem,
+                })))
+            }
+        };
+        Ok::<_, CUresult>(kernel as *mut _)
+    })??;
+    unsafe { *hfunc = function };
     Ok(())
 }
 
-pub(crate) fn unload(_: *mut Module) -> Result<(), CUresult> {
+pub(crate) fn load_data(pmod: *mut *mut Module, image: *const c_void) -> Result<(), CUresult> {
+    let spirv_data = SpirvModule::new_raw(image as *const _)?;
+    load_data_impl(pmod, spirv_data)
+}
+
+pub fn load_data_impl(pmod: *mut *mut Module, spirv_data: SpirvModule) -> Result<(), CUresult> {
+    let module = GlobalState::lock_current_context(|ctx| {
+        let device = unsafe { &mut *ctx.device };
+        let l0_module = spirv_data.compile(&mut device.l0_context, &device.base)?;
+        let mut device_binaries = HashMap::new();
+        let compiled_module = CompiledModule {
+            base: l0_module,
+            kernels: HashMap::new(),
+        };
+        device_binaries.insert(device.index, compiled_module);
+        let module_data = ModuleData {
+            spirv: spirv_data,
+            device_binaries,
+        };
+        Ok::<_, CUresult>(module_data)
+    })??;
+    let module_ptr = Box::into_raw(Box::new(Module::new(module)));
+    unsafe { *pmod = module_ptr };
     Ok(())
 }
+
+pub(crate) fn unload(module: *mut Module) -> Result<(), CUresult> {
+    if module == ptr::null_mut() {
+        return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
+    }
+    GlobalState::lock(|_| Module::destroy_impl(module))?
+}
diff --git a/notcuda/src/impl/stream.rs b/notcuda/src/impl/stream.rs
index 1844677..e212dfc 100644
--- a/notcuda/src/impl/stream.rs
+++ b/notcuda/src/impl/stream.rs
@@ -1,36 +1,114 @@
-use std::cell::RefCell;
+use super::{
+    context::{Context, ContextData},
+    CUresult, GlobalState,
+};
+use std::{mem, ptr};
 
-use device::Device;
+use super::{HasLivenessCookie, LiveCheck};
 
-use super::device;
+pub type Stream = LiveCheck<StreamData>;
 
-pub struct Stream {
-    dev: *mut Device,
-}
+pub const CU_STREAM_LEGACY: *mut Stream = 1 as *mut _;
+pub const CU_STREAM_PER_THREAD: *mut Stream = 2 as *mut _;
 
-pub struct DefaultStream {
-    streams: Vec<Option<Stream>>,
-}
+impl HasLivenessCookie for StreamData {
+    #[cfg(target_pointer_width = "64")]
+    const COOKIE: usize = 0x512097354de18d35;
 
-impl DefaultStream {
-    fn new() -> Self {
-        DefaultStream {
-            streams: Vec::new(),
+    #[cfg(target_pointer_width = "32")]
+    const COOKIE: usize = 0x77d5cc0b;
+
+    const LIVENESS_FAIL: CUresult = CUresult::CUDA_ERROR_INVALID_HANDLE;
+
+    fn try_drop(&mut self) -> Result<(), CUresult> {
+        if self.context != ptr::null_mut() {
+            let context = unsafe { &mut *self.context };
+            if !context.streams.remove(&(self as *mut _)) {
+                return Err(CUresult::CUDA_ERROR_UNKNOWN);
+            }
         }
+        Ok(())
     }
 }
 
-thread_local! {
-    pub static DEFAULT_STREAM: RefCell<DefaultStream> = RefCell::new(DefaultStream::new());
+pub struct StreamData {
+    pub context: *mut ContextData,
+    pub queue: l0::CommandQueue,
+}
+
+impl StreamData {
+    pub fn new_unitialized(ctx: &mut l0::Context, dev: &l0::Device) -> Result<Self, CUresult> {
+        Ok(StreamData {
+            context: ptr::null_mut(),
+            queue: l0::CommandQueue::new(ctx, dev)?,
+        })
+    }
+    pub fn new(ctx: &mut ContextData) -> Result<Self, CUresult> {
+        let l0_ctx = &mut unsafe { &mut *ctx.device }.l0_context;
+        let l0_dev = &unsafe { &*ctx.device }.base;
+        Ok(StreamData {
+            context: ctx as *mut _,
+            queue: l0::CommandQueue::new(l0_ctx, l0_dev)?,
+        })
+    }
+
+    pub fn command_list(&self) -> Result<l0::CommandList, l0::sys::_ze_result_t> {
+        let ctx = unsafe { &mut *self.context };
+        let dev = unsafe { &mut *ctx.device };
+        l0::CommandList::new(&mut dev.l0_context, &dev.base)
+    }
+}
+
+impl Drop for StreamData {
+    fn drop(&mut self) {
+        if self.context == ptr::null_mut() {
+            return;
+        }
+        unsafe { (&mut *self.context).streams.remove(&(&mut *self as *mut _)) };
+    }
+}
+
+pub(crate) fn get_ctx(hstream: *mut Stream, pctx: *mut *mut Context) -> Result<(), CUresult> {
+    if pctx == ptr::null_mut() {
+        return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
+    }
+    let ctx_ptr = GlobalState::lock_stream(hstream, |stream| stream.context)?;
+    if ctx_ptr == ptr::null_mut() {
+        return Err(CUresult::CUDA_ERROR_CONTEXT_IS_DESTROYED);
+    }
+    unsafe { *pctx = Context::ptr_from_inner(ctx_ptr) };
+    Ok(())
+}
+
+pub(crate) fn create(phstream: *mut *mut Stream, _flags: u32) -> Result<(), CUresult> {
+    let stream_ptr = GlobalState::lock_current_context(|ctx| {
+        let mut stream_box = Box::new(Stream::new(StreamData::new(ctx)?));
+        let stream_ptr = stream_box.as_mut().as_option_mut().unwrap() as *mut _;
+        if !ctx.streams.insert(stream_ptr) {
+            return Err(CUresult::CUDA_ERROR_UNKNOWN);
+        }
+        mem::forget(stream_box);
+        Ok::<_, CUresult>(stream_ptr)
+    })??;
+    unsafe { *phstream = Stream::ptr_from_inner(stream_ptr) };
+    Ok(())
+}
+
+pub(crate) fn destroy_v2(pstream: *mut Stream) -> Result<(), CUresult> {
+    if pstream == ptr::null_mut() || pstream == CU_STREAM_LEGACY || pstream == CU_STREAM_PER_THREAD
+    {
+        return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
+    }
+    GlobalState::lock(|_| Stream::destroy_impl(pstream))?
 }
 
 #[cfg(test)]
-mod tests {
+mod test {
     use crate::cuda::CUstream;
 
     use super::super::test::CudaDriverFns;
     use super::super::CUresult;
-    use std::ptr;
+    use std::{ptr, thread};
 
     const CU_STREAM_LEGACY: CUstream = 1 as *mut _;
     const CU_STREAM_PER_THREAD: CUstream = 2 as *mut _;
@@ -65,5 +143,100 @@ mod tests {
             CUresult::CUDA_SUCCESS
         );
         assert_eq!(ctx2, stream_ctx2);
+        //  Cleanup
+        assert_eq!(T::cuCtxDestroy_v2(ctx1), CUresult::CUDA_SUCCESS);
+        assert_eq!(T::cuCtxDestroy_v2(ctx2), CUresult::CUDA_SUCCESS);
+    }
+
+    cuda_driver_test!(stream_context_destroyed);
+
+    fn stream_context_destroyed<T: CudaDriverFns>() {
+        assert_eq!(T::cuInit(0), CUresult::CUDA_SUCCESS);
+        let mut ctx = ptr::null_mut();
+        assert_eq!(T::cuCtxCreate_v2(&mut ctx, 0, 0), CUresult::CUDA_SUCCESS);
+        let mut stream = ptr::null_mut();
+        assert_eq!(T::cuStreamCreate(&mut stream, 0), CUresult::CUDA_SUCCESS);
+        let mut stream_ctx1 = ptr::null_mut();
+        assert_eq!(
+            T::cuStreamGetCtx(stream, &mut stream_ctx1),
+            CUresult::CUDA_SUCCESS
+        );
+        assert_eq!(stream_ctx1, ctx);
+        assert_eq!(T::cuCtxDestroy_v2(ctx), CUresult::CUDA_SUCCESS);
+        let mut stream_ctx2 = ptr::null_mut();
+        // When a context gets destroyed, its streams are also destroyed
+        let cuda_result = T::cuStreamGetCtx(stream, &mut stream_ctx2);
+        assert!(
+            cuda_result == CUresult::CUDA_ERROR_INVALID_HANDLE
+                || cuda_result == CUresult::CUDA_ERROR_INVALID_CONTEXT
+                || cuda_result == CUresult::CUDA_ERROR_CONTEXT_IS_DESTROYED
+        );
+        assert_eq!(
+            T::cuStreamDestroy_v2(stream),
+            CUresult::CUDA_ERROR_INVALID_HANDLE
+        );
+        // Check if creating another context is possible
+        let mut ctx2 = ptr::null_mut();
+        assert_eq!(T::cuCtxCreate_v2(&mut ctx2, 0, 0), CUresult::CUDA_SUCCESS);
+        //  Cleanup
+        assert_eq!(T::cuCtxDestroy_v2(ctx2), CUresult::CUDA_SUCCESS);
+    }
+
+    cuda_driver_test!(stream_moves_context_to_another_thread);
+
+    fn stream_moves_context_to_another_thread<T: CudaDriverFns>() {
+        assert_eq!(T::cuInit(0), CUresult::CUDA_SUCCESS);
+        let mut ctx = ptr::null_mut();
+        assert_eq!(T::cuCtxCreate_v2(&mut ctx, 0, 0), CUresult::CUDA_SUCCESS);
+        let mut stream = ptr::null_mut();
+        assert_eq!(T::cuStreamCreate(&mut stream, 0), CUresult::CUDA_SUCCESS);
+        let mut stream_ctx1 = ptr::null_mut();
+        assert_eq!(
+            T::cuStreamGetCtx(stream, &mut stream_ctx1),
+            CUresult::CUDA_SUCCESS
+        );
+        assert_eq!(stream_ctx1, ctx);
+        let stream_ptr = stream as usize;
+        let stream_ctx_on_thread = thread::spawn(move || {
+            let mut stream_ctx2 = ptr::null_mut();
+            assert_eq!(
+                T::cuStreamGetCtx(stream_ptr as *mut _, &mut stream_ctx2),
+                CUresult::CUDA_SUCCESS
+            );
+            stream_ctx2 as usize
+        })
+        .join()
+        .unwrap();
+        assert_eq!(stream_ctx1, stream_ctx_on_thread as *mut _);
+        //  Cleanup
+        assert_eq!(T::cuStreamDestroy_v2(stream), CUresult::CUDA_SUCCESS);
+        assert_eq!(T::cuCtxDestroy_v2(ctx), CUresult::CUDA_SUCCESS);
+    }
+
+    cuda_driver_test!(can_destroy_stream);
+
+    fn can_destroy_stream<T: CudaDriverFns>() {
+        assert_eq!(T::cuInit(0), CUresult::CUDA_SUCCESS);
+        let mut ctx = ptr::null_mut();
+        assert_eq!(T::cuCtxCreate_v2(&mut ctx, 0, 0), CUresult::CUDA_SUCCESS);
+        let mut stream = ptr::null_mut();
+        assert_eq!(T::cuStreamCreate(&mut stream, 0), CUresult::CUDA_SUCCESS);
+        assert_eq!(T::cuStreamDestroy_v2(stream), CUresult::CUDA_SUCCESS);
+        // Cleanup
+        assert_eq!(T::cuCtxDestroy_v2(ctx), CUresult::CUDA_SUCCESS);
+    }
+
+    cuda_driver_test!(cant_destroy_default_stream);
+
+    fn cant_destroy_default_stream<T: CudaDriverFns>() {
+        assert_eq!(T::cuInit(0), CUresult::CUDA_SUCCESS);
+        let mut ctx = ptr::null_mut();
+        assert_eq!(T::cuCtxCreate_v2(&mut ctx, 0, 0), CUresult::CUDA_SUCCESS);
+        assert_ne!(
+            T::cuStreamDestroy_v2(super::CU_STREAM_LEGACY as *mut _),
+            CUresult::CUDA_SUCCESS
+        );
+        // Cleanup
+        assert_eq!(T::cuCtxDestroy_v2(ctx), CUresult::CUDA_SUCCESS);
     }
 }
diff --git a/notcuda/src/impl/test.rs b/notcuda/src/impl/test.rs
index dbd2eff..b6ed926 100644
--- a/notcuda/src/impl/test.rs
+++ b/notcuda/src/impl/test.rs
@@ -1,8 +1,12 @@
 #![allow(non_snake_case)]
 
-use crate::{cuda::CUstream, r#impl as notcuda};
-use crate::r#impl::CUresult;
-use crate::{cuda::CUuuid, r#impl::Encuda};
+use crate::cuda as notcuda;
+use crate::cuda::CUstream;
+use crate::cuda::CUuuid;
+use crate::{
+    cuda::{CUdevice, CUdeviceptr},
+    r#impl::CUresult,
+};
 use ::std::{
     ffi::c_void,
     os::raw::{c_int, c_uint},
@@ -37,48 +41,63 @@ pub trait CudaDriverFns {
     fn cuDeviceGetUuid(uuid: *mut CUuuid, dev: c_int) -> CUresult;
     fn cuDevicePrimaryCtxGetState(dev: c_int, flags: *mut c_uint, active: *mut c_int) -> CUresult;
     fn cuStreamGetCtx(hStream: CUstream, pctx: *mut *mut c_void) -> CUresult;
+    fn cuStreamCreate(stream: *mut CUstream, flags: c_uint) -> CUresult;
+    fn cuMemFree_v2(mem: *mut c_void) -> CUresult;
+    fn cuStreamDestroy_v2(stream: CUstream) -> CUresult;
 }
 
 pub struct NotCuda();
 
 impl CudaDriverFns for NotCuda {
     fn cuInit(_flags: c_uint) -> CUresult {
-        crate::cuda::cuInit(_flags as _)
+        notcuda::cuInit(_flags as _)
     }
 
     fn cuCtxCreate_v2(pctx: *mut *mut c_void, flags: c_uint, dev: c_int) -> CUresult {
-        notcuda::context::create_v2(pctx as *mut _, flags, notcuda::device::Index(dev)).encuda()
+        notcuda::cuCtxCreate_v2(pctx as *mut _, flags, CUdevice(dev))
     }
 
     fn cuCtxDestroy_v2(ctx: *mut c_void) -> CUresult {
-        notcuda::context::destroy_v2(ctx as *mut _)
+        notcuda::cuCtxDestroy_v2(ctx as *mut _)
     }
 
     fn cuCtxPopCurrent_v2(pctx: *mut *mut c_void) -> CUresult {
-        notcuda::context::pop_current_v2(pctx as *mut _)
+        notcuda::cuCtxPopCurrent_v2(pctx as *mut _)
     }
 
     fn cuCtxGetApiVersion(ctx: *mut c_void, version: *mut c_uint) -> CUresult {
-        notcuda::context::get_api_version(ctx as *mut _, version)
+        notcuda::cuCtxGetApiVersion(ctx as *mut _, version)
     }
 
     fn cuCtxGetCurrent(pctx: *mut *mut c_void) -> CUresult {
-        notcuda::context::get_current(pctx as *mut _).encuda()
+        notcuda::cuCtxGetCurrent(pctx as *mut _)
     }
     fn cuMemAlloc_v2(dptr: *mut *mut c_void, bytesize: usize) -> CUresult {
-        notcuda::memory::alloc_v2(dptr as *mut _, bytesize)
+        notcuda::cuMemAlloc_v2(dptr as *mut _, bytesize)
     }
 
     fn cuDeviceGetUuid(uuid: *mut CUuuid, dev: c_int) -> CUresult {
-        notcuda::device::get_uuid(uuid, notcuda::device::Index(dev)).encuda()
+        notcuda::cuDeviceGetUuid(uuid, CUdevice(dev))
     }
 
     fn cuDevicePrimaryCtxGetState(dev: c_int, flags: *mut c_uint, active: *mut c_int) -> CUresult {
-        notcuda::device::primary_ctx_get_state(notcuda::device::Index(dev), flags, active).encuda()
+        notcuda::cuDevicePrimaryCtxGetState(CUdevice(dev), flags, active)
     }
 
     fn cuStreamGetCtx(hStream: CUstream, pctx: *mut *mut c_void) -> CUresult {
-        crate::cuda::cuStreamGetCtx(hStream, pctx as _)
+        notcuda::cuStreamGetCtx(hStream, pctx as _)
+    }
+
+    fn cuStreamCreate(stream: *mut CUstream, flags: c_uint) -> CUresult {
+        notcuda::cuStreamCreate(stream, flags)
+    }
+
+    fn cuMemFree_v2(dptr: *mut c_void) -> CUresult {
+        notcuda::cuMemFree_v2(CUdeviceptr(dptr as _))
+    }
+
+    fn cuStreamDestroy_v2(stream: CUstream) -> CUresult {
+        notcuda::cuStreamDestroy_v2(stream)
     }
 }
 
@@ -123,4 +142,16 @@ impl CudaDriverFns for Cuda {
     fn cuStreamGetCtx(hStream: CUstream, pctx: *mut *mut c_void) -> CUresult {
         unsafe { CUresult(cuda::cuStreamGetCtx(hStream as _, pctx as _) as c_uint) }
     }
+
+    fn cuStreamCreate(stream: *mut CUstream, flags: c_uint) -> CUresult {
+        unsafe { CUresult(cuda::cuStreamCreate(stream as _, flags as _) as c_uint) }
+    }
+
+    fn cuMemFree_v2(mem: *mut c_void) -> CUresult {
+        unsafe { CUresult(cuda::cuMemFree_v2(mem as _) as c_uint) }
+    }
+
+    fn cuStreamDestroy_v2(stream: CUstream) -> CUresult {
+        unsafe { CUresult(cuda::cuStreamDestroy_v2(stream as _) as c_uint) }
+    }
 }
diff --git a/ptx/src/lib.rs b/ptx/src/lib.rs
index 1aac8ab..591428f 100644
--- a/ptx/src/lib.rs
+++ b/ptx/src/lib.rs
@@ -34,8 +34,9 @@ pub use crate::ptx::ModuleParser;
 pub use lalrpop_util::lexer::Token;
 pub use lalrpop_util::ParseError;
 pub use rspirv::dr::Error as SpirvError;
-pub use translate::TranslateError as TranslateError;
-pub use translate::to_spirv;
+pub use translate::to_spirv_module;
+pub use translate::KernelInfo;
+pub use translate::TranslateError;
 
 pub(crate) fn without_none<T>(x: Vec<Option<T>>) -> Vec<T> {
     x.into_iter().filter_map(|x| x).collect()
diff --git a/ptx/src/test/mod.rs b/ptx/src/test/mod.rs
index 0339141..0785f3e 100644
--- a/ptx/src/test/mod.rs
+++ b/ptx/src/test/mod.rs
@@ -12,7 +12,7 @@ fn parse_and_assert(s: &str) {
 fn compile_and_assert(s: &str) -> Result<(), TranslateError> {
     let mut errors = Vec::new();
     let ast = ptx::ModuleParser::new().parse(&mut errors, s).unwrap();
-    crate::to_spirv(ast)?;
+    crate::to_spirv_module(ast)?;
     Ok(())
 }
 
diff --git a/ptx/src/translate.rs b/ptx/src/translate.rs
index c0e15f2..3d0f476 100644
--- a/ptx/src/translate.rs
+++ b/ptx/src/translate.rs
@@ -1,7 +1,7 @@
 use crate::ast;
 use half::f16;
 use rspirv::{binary::Disassemble, dr};
-use std::{borrow::Cow, convert::TryFrom, ffi::CString, hash::Hash, iter, mem};
+use std::{borrow::Cow, ffi::CString, hash::Hash, iter, mem};
 use std::{
     collections::{hash_map, HashMap, HashSet},
     convert::TryInto,
@@ -450,6 +450,11 @@ pub struct Module {
     pub should_link_ptx_impl: Option<&'static [u8]>,
     pub build_options: CString,
 }
+impl Module {
+    pub fn assemble(&self) -> Vec<u32> {
+        self.spirv.assemble()
+    }
+}
 
 pub struct KernelInfo {
     pub arguments_sizes: Vec<usize>,
@@ -1046,8 +1051,12 @@ fn emit_function_header<'a>(
     kernel_info: &mut HashMap<String, KernelInfo>,
 ) -> Result<(), TranslateError> {
     if let MethodName::Kernel(name) = func_decl.name {
-        let args_lens = func_decl
-            .input
+        let input_args = if !func_decl.uses_shared_mem {
+            func_decl.input.as_slice()
+        } else {
+            &func_decl.input[0..func_decl.input.len() - 1]
+        };
+        let args_lens = input_args
             .iter()
             .map(|param| param.v_type.size_of())
             .collect();
@@ -1135,21 +1144,6 @@ fn emit_function_header<'a>(
     Ok(())
 }
 
-pub fn to_spirv<'a>(
-    ast: ast::Module<'a>,
-) -> Result<(Option<&'static [u8]>, Vec<u32>, HashMap<String, Vec<usize>>), TranslateError> {
-    let module = to_spirv_module(ast)?;
-    Ok((
-        module.should_link_ptx_impl,
-        module.spirv.assemble(),
-        module
-            .kernel_info
-            .into_iter()
-            .map(|(k, v)| (k, v.arguments_sizes))
-            .collect(),
-    ))
-}
-
 fn emit_capabilities(builder: &mut dr::Builder) {
     builder.capability(spirv::Capability::GenericPointer);
     builder.capability(spirv::Capability::Linkage);