Finish converting to OpenCL

8 files changed, 308 insertions, 176 deletions

diff --git a/zluda/src/impl/context.rs b/zluda/src/impl/context.rs
index 18a2bd6..e8de477 100644
--- a/zluda/src/impl/context.rs
+++ b/zluda/src/impl/context.rs
@@ -1,5 +1,5 @@
 use super::{device, stream::Stream, stream::StreamData, HasLivenessCookie, LiveCheck};
-use super::{CUresult, GlobalState};
+use super::{transmute_lifetime_mut, CUresult, GlobalState};
 use crate::{cuda::CUcontext, cuda_impl};
 use l0::sys::ze_result_t;
 use std::{cell::RefCell, num::NonZeroU32, os::raw::c_uint, ptr, sync::atomic::AtomicU32};
@@ -98,14 +98,11 @@ pub struct ContextData {
 
 impl ContextData {
     pub fn new(
-        l0_ctx: &'static l0::Context,
-        l0_dev: l0::Device,
         flags: c_uint,
         is_primary: bool,
-        host_event: (l0::Event<'static>, u64),
         dev: *mut device::Device,
     ) -> Result<Self, CUresult> {
-        let default_stream = StreamData::new_unitialized(l0_ctx, l0_dev, host_event)?;
+        let default_stream = StreamData::new_unitialized()?;
         Ok(ContextData {
             flags: AtomicU32::new(flags),
             device: dev,
@@ -121,8 +118,15 @@ impl ContextData {
 
 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 _;
+        let ctx_data: &'static mut _ = {
+            let this = self.as_option_mut().unwrap();
+            let result = { unsafe { transmute_lifetime_mut(this) } };
+            drop(this);
+            result
+        };
+        { self.as_option_mut().unwrap() }
+            .default_stream
+            .late_init(ctx_data);
     }
 }
 
@@ -137,11 +141,8 @@ pub fn create_v2(
     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(
-            &dev.ocl_context,
-            dev.base,
             flags,
             false,
-            dev.host_event_pool.get(dev.base, &dev.ocl_context)?,
             dev_ptr as *mut _,
         )?));
         ctx_box.late_init();
diff --git a/zluda/src/impl/device.rs b/zluda/src/impl/device.rs
index c356bda..3b43c49 100644
--- a/zluda/src/impl/device.rs
+++ b/zluda/src/impl/device.rs
@@ -1,9 +1,11 @@
 use super::{context, transmute_lifetime, transmute_lifetime_mut, CUresult, GlobalState};
 use crate::cuda;
 use cuda::{CUdevice_attribute, CUuuid_st};
-use ocl_core::DeviceType;
+use ocl_core::{ClDeviceIdPtr, ContextProperties, DeviceType};
 use std::{
-    cmp, mem,
+    cmp,
+    ffi::c_void,
+    mem,
     os::raw::{c_char, c_int, c_uint},
     ptr,
     sync::atomic::{AtomicU32, Ordering},
@@ -22,6 +24,7 @@ pub struct Device {
     pub ocl_base: ocl_core::DeviceId,
     pub default_queue: ocl_core::CommandQueue,
     pub ocl_context: ocl_core::Context,
+    pub(crate) ocl_ext: OpenCLExtensions,
     pub primary_context: context::Context,
     properties: Option<Box<l0::sys::ze_device_properties_t>>,
     image_properties: Option<Box<l0::sys::ze_device_image_properties_t>>,
@@ -29,19 +32,185 @@ pub struct Device {
     compute_properties: Option<Box<l0::sys::ze_device_compute_properties_t>>,
 }
 
+type cl_mem_properties_intel = ocl_core::ffi::cl_bitfield;
+
+pub(crate) struct OpenCLExtensions {
+    pub clDeviceMemAllocINTEL: unsafe extern "system" fn(
+        ocl_core::ffi::cl_context,
+        ocl_core::ffi::cl_device_id,
+        *const cl_mem_properties_intel,
+        usize,
+        ocl_core::ffi::cl_uint,
+        *mut ocl_core::ffi::cl_int,
+    ) -> *mut c_void,
+    pub clEnqueueMemcpyINTEL: unsafe extern "system" fn(
+        ocl_core::ffi::cl_command_queue,
+        ocl_core::ffi::cl_bool,
+        *mut c_void,
+        *const c_void,
+        usize,
+        ocl_core::ffi::cl_uint,
+        *const ocl_core::ffi::cl_event,
+        *mut ocl_core::ffi::cl_event,
+    ) -> ocl_core::ffi::cl_int,
+    pub clMemBlockingFreeINTEL:
+        unsafe extern "system" fn(ocl_core::ffi::cl_context, *mut c_void) -> ocl_core::ffi::cl_int,
+    pub clEnqueueMemFillINTEL: unsafe extern "system" fn(
+        ocl_core::ffi::cl_command_queue,
+        *mut c_void,
+        *const c_void,
+        usize,
+        usize,
+        ocl_core::ffi::cl_uint,
+        *const ocl_core::ffi::cl_event,
+        *mut ocl_core::ffi::cl_event,
+    ) -> ocl_core::ffi::cl_int,
+}
+
+impl OpenCLExtensions {
+    fn new(plat: &ocl_core::PlatformId) -> Result<Self, CUresult> {
+        let clDeviceMemAllocINTEL = unsafe {
+            ocl_core::get_extension_function_address_for_platform(
+                plat,
+                "clDeviceMemAllocINTEL",
+                None,
+            )?
+        };
+        let clEnqueueMemcpyINTEL = unsafe {
+            ocl_core::get_extension_function_address_for_platform(
+                plat,
+                "clEnqueueMemcpyINTEL",
+                None,
+            )?
+        };
+        let clMemBlockingFreeINTEL = unsafe {
+            ocl_core::get_extension_function_address_for_platform(
+                plat,
+                "clMemBlockingFreeINTEL",
+                None,
+            )?
+        };
+        let clEnqueueMemFillINTEL = unsafe {
+            ocl_core::get_extension_function_address_for_platform(
+                plat,
+                "clEnqueueMemFillINTEL",
+                None,
+            )?
+        };
+        Ok(Self {
+            clDeviceMemAllocINTEL: unsafe { mem::transmute(clDeviceMemAllocINTEL) },
+            clEnqueueMemcpyINTEL: unsafe { mem::transmute(clEnqueueMemcpyINTEL) },
+            clMemBlockingFreeINTEL: unsafe { mem::transmute(clMemBlockingFreeINTEL) },
+            clEnqueueMemFillINTEL: unsafe { mem::transmute(clEnqueueMemFillINTEL) },
+        })
+    }
+
+    pub unsafe fn device_mem_alloc(
+        &self,
+        ctx: &ocl_core::Context,
+        device: &ocl_core::DeviceId,
+        size: usize,
+        alignment: ocl_core::ffi::cl_uint,
+    ) -> Result<*mut c_void, CUresult> {
+        let mut error = 0;
+        let result = (self.clDeviceMemAllocINTEL)(
+            ctx.as_ptr(),
+            device.as_ptr(),
+            ptr::null(),
+            size,
+            alignment,
+            &mut error,
+        );
+        if error == 0 {
+            Ok(result)
+        } else {
+            Err(CUresult::CUDA_ERROR_UNKNOWN)
+        }
+    }
+
+    pub unsafe fn enqueue_memcpy(
+        &self,
+        queue: &ocl_core::CommandQueue,
+        blocking: bool,
+        dst: *mut c_void,
+        src: *const c_void,
+        size: usize,
+    ) -> Result<(), CUresult> {
+        let error = (self.clEnqueueMemcpyINTEL)(
+            queue.as_ptr(),
+            if blocking { 1 } else { 0 },
+            dst,
+            src,
+            size,
+            0,
+            ptr::null(),
+            ptr::null_mut(),
+        );
+        if error == 0 {
+            Ok(())
+        } else {
+            Err(CUresult::CUDA_ERROR_UNKNOWN)
+        }
+    }
+
+    pub unsafe fn mem_blocking_free(
+        &self,
+        ctx: &ocl_core::Context,
+        mem_ptr: *mut c_void,
+    ) -> Result<(), CUresult> {
+        let error = (self.clMemBlockingFreeINTEL)(ctx.as_ptr(), mem_ptr);
+        if error == 0 {
+            Ok(())
+        } else {
+            Err(CUresult::CUDA_ERROR_UNKNOWN)
+        }
+    }
+
+    pub unsafe fn enqueue_memfill(
+        &self,
+        queue: &ocl_core::CommandQueue,
+        dst: *mut c_void,
+        pattern: *const c_void,
+        patternSize: usize,
+        size: usize,
+    ) -> Result<ocl_core::Event, CUresult> {
+        let mut signal: ocl_core::ffi::cl_event = ptr::null_mut();
+        let error = (self.clEnqueueMemFillINTEL)(
+            queue.as_ptr(),
+            dst,
+            pattern,
+            patternSize,
+            size,
+            0,
+            ptr::null(),
+            &mut signal,
+        );
+        if error == 0 {
+            Ok(ocl_core::Event::from_raw(signal))
+        } else {
+            Err(CUresult::CUDA_ERROR_UNKNOWN)
+        }
+    }
+}
+
 unsafe impl Send for Device {}
 
 impl Device {
     pub fn new(
-        drv: &l0::Driver,
         l0_dev: l0::Device,
+        platform: ocl_core::PlatformId,
         ocl_dev: ocl_core::DeviceId,
         idx: usize,
     ) -> Result<Self, CUresult> {
-        let ctx = ocl_core::create_context(None, &[ocl_dev], None, None)?;
+        let ocl_ext = OpenCLExtensions::new(&platform)?;
+        let mut props = ocl_core::ContextProperties::new();
+        props.set_platform(platform);
+        let ctx = ocl_core::create_context(Some(&props), &[ocl_dev], None, None)?;
         let queue = ocl_core::create_command_queue(&ctx, ocl_dev, None)?;
-        let primary_context = context::Context::new(context::ContextData::new());
+        let primary_context =
+            context::Context::new(context::ContextData::new(0, true, ptr::null_mut())?);
         Ok(Self {
+            ocl_ext,
             index: Index(idx as c_int),
             base: l0_dev,
             ocl_base: ocl_dev,
@@ -55,6 +224,10 @@ impl Device {
         })
     }
 
+    pub fn late_init(&mut self) {
+        self.primary_context.as_option_mut().unwrap().device = self as *mut _;
+    }
+
     fn get_properties<'a>(&'a mut self) -> l0::Result<&'a l0::sys::ze_device_properties_t> {
         if let Some(ref prop) = self.properties {
             return Ok(prop);
@@ -207,7 +380,7 @@ pub fn get_attribute(
                     & l0::sys::ze_device_property_flags_t::ZE_DEVICE_PROPERTY_FLAG_INTEGRATED)
                     == l0::sys::ze_device_property_flags_t::ZE_DEVICE_PROPERTY_FLAG_INTEGRATED
                 {
-                    Ok(1)
+                    Ok::<_, CUresult>(1)
                 } else {
                     Ok(0)
                 }
diff --git a/zluda/src/impl/function.rs b/zluda/src/impl/function.rs
index 2658d27..05f864b 100644
--- a/zluda/src/impl/function.rs
+++ b/zluda/src/impl/function.rs
@@ -27,7 +27,7 @@ impl HasLivenessCookie for FunctionData {
 
 pub struct FunctionData {
     pub base: ocl_core::Kernel,
-    pub arg_size: Vec<usize>,
+    pub arg_size: Vec<(usize, bool)>,
     pub use_shared_mem: bool,
     pub legacy_args: LegacyArguments,
 }
@@ -73,14 +73,28 @@ pub fn launch_kernel(
     GlobalState::lock_enqueue(hstream, |queue| {
         let func: &mut FunctionData = unsafe { &mut *f }.as_result_mut()?;
         if kernel_params != ptr::null_mut() {
-            for (i, arg_size) in func.arg_size.iter().enumerate() {
-                unsafe {
-                    ocl_core::set_kernel_arg(
-                        &func.base,
-                        i as u32,
-                        ocl_core::ArgVal::from_raw(*arg_size, *kernel_params.add(i), false),
-                    )?;
-                };
+            for (i, &(arg_size, is_mem)) in func.arg_size.iter().enumerate() {
+                if is_mem {
+                    let error = 0;
+                    unsafe {
+                        ocl_core::ffi::clSetKernelArgSVMPointer(
+                            func.base.as_ptr(),
+                            i as u32,
+                            *(*kernel_params.add(i) as *const _),
+                        )
+                    };
+                    if error != 0 {
+                        panic!("clSetKernelArgSVMPointer");
+                    }
+                } else {
+                    unsafe {
+                        ocl_core::set_kernel_arg(
+                            &func.base,
+                            i as u32,
+                            ocl_core::ArgVal::from_raw(arg_size, *kernel_params.add(i), is_mem),
+                        )?;
+                    };
+                }
             }
         } else {
             let mut offset = 0;
@@ -102,27 +116,27 @@ pub fn launch_kernel(
             match (buffer_size, buffer_ptr) {
                 (Some(buffer_size), Some(buffer_ptr)) => {
                     let sum_of_kernel_argument_sizes =
-                        func.arg_size.iter().fold(0, |offset, size_of_arg| {
-                            size_of_arg + round_up_to_multiple(offset, *size_of_arg)
+                        func.arg_size.iter().fold(0, |offset, &(size_of_arg, _)| {
+                            size_of_arg + round_up_to_multiple(offset, size_of_arg)
                         });
                     if buffer_size < sum_of_kernel_argument_sizes {
                         return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
                     }
                     let mut offset = 0;
-                    for (i, arg_size) in func.arg_size.iter().enumerate() {
-                        let buffer_offset = round_up_to_multiple(offset, *arg_size);
+                    for (i, &(arg_size, is_mem)) in func.arg_size.iter().enumerate() {
+                        let buffer_offset = round_up_to_multiple(offset, arg_size);
                         unsafe {
                             ocl_core::set_kernel_arg(
                                 &func.base,
                                 i as u32,
                                 ocl_core::ArgVal::from_raw(
-                                    *arg_size,
+                                    arg_size,
                                     buffer_ptr.add(buffer_offset) as *const _,
-                                    false,
+                                    is_mem,
                                 ),
                             )?;
                         };
-                        offset = buffer_offset + *arg_size;
+                        offset = buffer_offset + arg_size;
                     }
                 }
                 _ => return Err(CUresult::CUDA_ERROR_INVALID_VALUE),
diff --git a/zluda/src/impl/memory.rs b/zluda/src/impl/memory.rs
index 5919690..3e96a8c 100644
--- a/zluda/src/impl/memory.rs
+++ b/zluda/src/impl/memory.rs
@@ -1,60 +1,77 @@
 use super::{stream, CUresult, GlobalState};

-use std::{ffi::c_void, mem};

+use std::{

+    ffi::c_void,

+    mem::{self, size_of},

+};

 

 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>(dev.ocl_context.mem_alloc_device(bytesize, 0, dev.base)?)

+        Ok::<_, CUresult>(unsafe {

+            dev.ocl_ext

+                .device_mem_alloc(&dev.ocl_context, &dev.ocl_base, bytesize, 0)?

+        })

     })??;

     unsafe { *dptr = ptr };

     Ok(())

 }

 

 pub fn copy_v2(dst: *mut c_void, src: *const c_void, bytesize: usize) -> Result<(), CUresult> {

-    GlobalState::lock_enqueue(stream::CU_STREAM_LEGACY, |cmd_list, signal, wait| {

-        unsafe { cmd_list.append_memory_copy_raw(dst, src, bytesize, Some(signal), wait)? };

+    GlobalState::lock_stream(stream::CU_STREAM_LEGACY, |stream_data| {

+        let dev = unsafe { &*(*stream_data.context).device };

+        let queue = stream_data.cmd_list.as_ref().unwrap();

+        unsafe {

+            dev.ocl_ext

+                .enqueue_memcpy(queue, true, dst, src, bytesize)?

+        };

         Ok(())

-    })

+    })?

 }

 

 pub fn free_v2(ptr: *mut c_void) -> Result<(), CUresult> {

     GlobalState::lock_current_context(|ctx| {

         let dev = unsafe { &mut *ctx.device };

-        Ok::<_, CUresult>(dev.ocl_context.mem_free(ptr)?)

-    })

-    .map_err(|_| CUresult::CUDA_ERROR_INVALID_VALUE)?

+        unsafe { dev.ocl_ext.mem_blocking_free(&dev.ocl_context, ptr)? };

+        Ok(())

+    })?

 }

 

 pub(crate) fn set_d32_v2(dst: *mut c_void, mut ui: u32, n: usize) -> Result<(), CUresult> {

-    GlobalState::lock_enqueue(stream::CU_STREAM_LEGACY, |cmd_list, signal, wait| {

-        unsafe {

-            cmd_list.append_memory_fill_raw(

+    GlobalState::lock_stream(stream::CU_STREAM_LEGACY, move |stream_data| {

+        let dev = unsafe { &*(*stream_data.context).device };

+        let queue = stream_data.cmd_list.as_ref().unwrap();

+        let pattern_size = mem::size_of_val(&ui);

+        let event = unsafe {

+            dev.ocl_ext.enqueue_memfill(

+                queue,

                 dst,

-                &mut ui as *mut _ as *mut _,

-                mem::size_of::<u32>(),

-                mem::size_of::<u32>() * n,

-                Some(signal),

-                wait,

-            )

-        }?;

+                &ui as *const _ as *const _,

+                pattern_size,

+                pattern_size * n,

+            )?

+        };

+        ocl_core::wait_for_event(&event)?;

         Ok(())

-    })

+    })?

 }

 

 pub(crate) fn set_d8_v2(dst: *mut c_void, mut uc: u8, n: usize) -> Result<(), CUresult> {

-    GlobalState::lock_enqueue(stream::CU_STREAM_LEGACY, |cmd_list, signal, wait| {

-        unsafe {

-            cmd_list.append_memory_fill_raw(

+    GlobalState::lock_stream(stream::CU_STREAM_LEGACY, move |stream_data| {

+        let dev = unsafe { &*(*stream_data.context).device };

+        let queue = stream_data.cmd_list.as_ref().unwrap();

+        let pattern_size = mem::size_of_val(&uc);

+        let event = unsafe {

+            dev.ocl_ext.enqueue_memfill(

+                queue,

                 dst,

-                &mut uc as *mut _ as *mut _,

-                mem::size_of::<u8>(),

-                mem::size_of::<u8>() * n,

-                Some(signal),

-                wait,

-            )

-        }?;

+                &uc as *const _ as *const _,

+                pattern_size,

+                pattern_size * n,

+            )?

+        };

+        ocl_core::wait_for_event(&event)?;

         Ok(())

-    })

+    })?

 }

 

 #[cfg(test)]

diff --git a/zluda/src/impl/mod.rs b/zluda/src/impl/mod.rs
index d410554..4b7a761 100644
--- a/zluda/src/impl/mod.rs
+++ b/zluda/src/impl/mod.rs
@@ -290,15 +290,7 @@ impl GlobalState {
             let l0_dev = unsafe { (*(*stream_data.context).device).base };
             let l0_ctx = unsafe { &mut (*(*stream_data.context).device).ocl_context };
             let cmd_list = unsafe { transmute_lifetime(&stream_data.cmd_list) };
-            // TODO: make new_marker drop-safe
-            let (new_event, new_marker) = stream_data.get_event(l0_dev, l0_ctx)?;
-            stream_data.try_reuse_finished_events()?;
-            let prev_event = stream_data.get_last_event();
-            let prev_event_array = prev_event.map(|e| [e]);
-            let empty = [];
-            let prev_event_slice = prev_event_array.as_ref().map_or(&empty[..], |arr| &arr[..]);
-            f(cmd_list, &new_event, prev_event_slice)?;
-            stream_data.push_event((new_event, new_marker));
+            f(&stream_data.cmd_list.as_ref().unwrap())?;
             Ok(())
         })?
     }
@@ -350,15 +342,19 @@ pub fn init() -> Result<(), CUresult> {
         })
         .ok_or(CUresult::CUDA_ERROR_UNKNOWN)?;
     let drivers = l0::Driver::get()?;
-    let devices = match drivers.into_iter().find(is_intel_gpu_driver) {
+    let mut devices = match drivers.into_iter().find(is_intel_gpu_driver) {
         None => return Err(CUresult::CUDA_ERROR_UNKNOWN),
         Some(driver) => driver
             .devices()?
             .into_iter()
             .enumerate()
-            .map(|(idx, l0_dev)| device::Device::new(&driver, l0_dev, device, idx).unwrap())
+            .map(|(idx, l0_dev)| device::Device::new(l0_dev, platform, device, idx).unwrap())
             .collect::<Vec<_>>(),
     };
+    for d in devices.iter_mut() {
+        d.late_init();
+        d.primary_context.late_init();
+    }
     let global_heap = unsafe { os::heap_create() };
     if global_heap == ptr::null_mut() {
         return Err(CUresult::CUDA_ERROR_OUT_OF_MEMORY);
diff --git a/zluda/src/impl/module.rs b/zluda/src/impl/module.rs
index 88d85c4..a1fa9dd 100644
--- a/zluda/src/impl/module.rs
+++ b/zluda/src/impl/module.rs
@@ -100,32 +100,36 @@ impl SpirvModule {
             )
         };
         let main_module = ocl_core::create_program_with_il(ctx, byte_il, None)?;
-        match self.should_link_ptx_impl {
+        let main_module = match self.should_link_ptx_impl {
             None => {
-                ocl_core::compile_program(
+                ocl_core::build_program(
                     &main_module,
                     Some(&[dev]),
                     &self.build_options,
-                    &[],
-                    &[],
-                    None,
                     None,
                     None,
                 )?;
+                main_module
             }
             Some(ptx_impl) => {
                 let ptx_impl_prog = ocl_core::create_program_with_il(ctx, ptx_impl, None)?;
-                ocl_core::build_program(
+                ocl_core::compile_program(
                     &main_module,
                     Some(&[dev]),
                     &self.build_options,
+                    &[],
+                    &[],
+                    None,
                     None,
                     None,
                 )?;
-                ocl_core::build_program(
+                ocl_core::compile_program(
                     &ptx_impl_prog,
                     Some(&[dev]),
                     &self.build_options,
+                    &[],
+                    &[],
+                    None,
                     None,
                     None,
                 )?;
@@ -137,7 +141,7 @@ impl SpirvModule {
                     None,
                     None,
                     None,
-                )?;
+                )?
             }
         };
         Ok(main_module)
diff --git a/zluda/src/impl/ocl_ext.rs b/zluda/src/impl/ocl_ext.rs
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/zluda/src/impl/ocl_ext.rs
diff --git a/zluda/src/impl/stream.rs b/zluda/src/impl/stream.rs
index f3910d6..0231cd8 100644
--- a/zluda/src/impl/stream.rs
+++ b/zluda/src/impl/stream.rs
@@ -34,118 +34,45 @@ impl HasLivenessCookie for StreamData {
 pub struct StreamData {
     pub context: *mut ContextData,
     // Immediate CommandList
-    pub cmd_list: l0::CommandList<'static>,
-    pub busy_events: VecDeque<(l0::Event<'static>, u64)>,
-    // This could be a Vec, but I'd rather reuse earliest enqueued event not the one recently enqueued
-    pub free_events: VecDeque<(l0::Event<'static>, u64)>,
-    pub synchronization_event: (l0::Event<'static>, u64),
+    pub cmd_list: Option<ocl_core::CommandQueue>,
 }
 
 impl StreamData {
-    pub fn new_unitialized(
-        ctx: &'static l0::Context,
-        device: l0::Device,
-        host_event: (l0::Event<'static>, u64),
-    ) -> Result<Self, CUresult> {
+    pub fn new_unitialized() -> Result<Self, CUresult> {
         Ok(StreamData {
             context: ptr::null_mut(),
-            cmd_list: l0::CommandList::new_immediate(ctx, device)?,
-            busy_events: VecDeque::new(),
-            free_events: VecDeque::new(),
-            synchronization_event: host_event,
+            cmd_list: None,
         })
     }
+
     pub fn new(ctx: &mut ContextData) -> Result<Self, CUresult> {
-        let l0_ctx = &mut unsafe { &mut *ctx.device }.ocl_context;
-        let device = unsafe { &*ctx.device }.base;
-        let synchronization_event = unsafe { &mut *ctx.device }
-            .host_event_pool
-            .get(device, l0_ctx)?;
+        let ocl_ctx = &unsafe { &*ctx.device }.ocl_context;
+        let device = unsafe { &*ctx.device }.ocl_base;
         Ok(StreamData {
             context: ctx as *mut _,
-            cmd_list: l0::CommandList::new_immediate(l0_ctx, device)?,
-            busy_events: VecDeque::new(),
-            free_events: VecDeque::new(),
-            synchronization_event,
+            cmd_list: Some(ocl_core::create_command_queue::<
+                &ocl_core::Context,
+                ocl_core::DeviceId,
+            >(ocl_ctx, device, None)?),
         })
     }
 
-    pub fn try_reuse_finished_events(&mut self) -> l0::Result<()> {
-        loop {
-            match self.busy_events.get(0) {
-                None => return Ok(()),
-                Some((ev, _)) => {
-                    if ev.is_ready()? {
-                        let (ev, marker) = self.busy_events.pop_front().unwrap();
-                        ev.host_reset()?;
-                        self.free_events.push_back((ev, marker));
-                    } else {
-                        return Ok(());
-                    }
-                }
-            }
-        }
-    }
-
-    pub fn reuse_all_finished_events(&mut self) -> l0::Result<()> {
-        self.free_events.reserve(self.busy_events.len());
-        for (ev, marker) in self.busy_events.drain(..) {
-            ev.host_reset()?;
-            self.free_events.push_back((ev, marker));
-        }
-        Ok(())
-    }
-
-    pub fn get_last_event(&self) -> Option<&l0::Event<'static>> {
-        self.busy_events.iter().next_back().map(|(ev, _)| ev)
-    }
-
-    pub fn push_event(&mut self, ev: (l0::Event<'static>, u64)) {
-        self.busy_events.push_back(ev);
+    pub fn late_init(&mut self, ctx: &mut ContextData) {
+        let ocl_ctx = &unsafe { &*ctx.device }.ocl_context;
+        let device = unsafe { &*ctx.device }.ocl_base;
+        self.context = ctx as *mut _;
+        self.cmd_list = Some(
+            ocl_core::create_command_queue::<&ocl_core::Context, ocl_core::DeviceId>(
+                ocl_ctx, device, None,
+            )
+            .unwrap(),
+        );
     }
 
-    pub fn synchronize(&mut self) -> l0::Result<()> {
-        let empty = [];
-        let busy_event_arr = self.busy_events.back().map(|(ev, _)| [ev]);
-        let wait_events = busy_event_arr.as_ref().map_or(&empty[..], |arr| &arr[..]);
-        unsafe {
-            self.cmd_list
-                .append_barrier(Some(&self.synchronization_event.0), wait_events)?
-        };
-        self.synchronization_event
-            .0
-            .host_synchronize(u64::max_value())?;
-        self.synchronization_event.0.host_reset()?;
-        self.reuse_all_finished_events()?;
+    pub fn synchronize(&mut self) -> Result<(), CUresult> {
+        ocl_core::finish(self.cmd_list.as_ref().unwrap())?;
         Ok(())
     }
-
-    pub fn get_event(
-        &mut self,
-        l0_dev: l0::Device,
-        l0_ctx: &'static l0::Context,
-    ) -> l0::Result<(l0::Event<'static>, u64)> {
-        self.free_events
-            .pop_front()
-            .map(|x| Ok(x))
-            .unwrap_or_else(|| {
-                let event_pool = unsafe { &mut (*(*self.context).device).device_event_pool };
-                event_pool.get(l0_dev, l0_ctx)
-            })
-    }
-}
-
-impl Drop for StreamData {
-    fn drop(&mut self) {
-        if self.context == ptr::null_mut() {
-            return;
-        }
-        for (_, marker) in self.busy_events.iter().chain(self.free_events.iter()) {
-            let device_event_pool = unsafe { &mut (*(*self.context).device).device_event_pool };
-            device_event_pool.mark_as_free(*marker);
-        }
-        unsafe { (&mut *self.context).streams.remove(&(&mut *self as *mut _)) };
-    }
 }
 
 pub(crate) fn get_ctx(hstream: *mut Stream, pctx: *mut *mut Context) -> Result<(), CUresult> {