Make everything async

level_zero/src/ze.rs

@@ -781,6 +781,26 @@ impl<'a> CommandList<'a> {
         Ok(unsafe { Self::from_ffi(result) })
     }
 
+    pub fn new_immediate(ctx: &'a Context, dev: Device) -> Result<Self> {
+        let queue_desc = sys::ze_command_queue_desc_t {
+            stype: sys::ze_structure_type_t::ZE_STRUCTURE_TYPE_COMMAND_QUEUE_DESC,
+            pNext: ptr::null(),
+            ordinal: 0,
+            index: 0,
+            flags: sys::ze_command_queue_flags_t(0),
+            mode: sys::ze_command_queue_mode_t::ZE_COMMAND_QUEUE_MODE_DEFAULT,
+            priority: sys::ze_command_queue_priority_t::ZE_COMMAND_QUEUE_PRIORITY_NORMAL,
+        };
+        let mut result: sys::ze_command_list_handle_t = ptr::null_mut();
+        check!(sys::zeCommandListCreateImmediate(
+            ctx.as_ffi(),
+            dev.as_ffi(),
+            &queue_desc,
+            &mut result
+        ));
+        Ok(unsafe { Self::from_ffi(result) })
+    }
+
     pub unsafe fn append_memory_copy<
         'dep,
         T: 'a + 'dep + Copy + Sized,

zluda/src/impl/function.rs

@@ -81,7 +81,7 @@ pub fn launch_kernel(
     {
         return Err(CUresult::CUDA_ERROR_INVALID_VALUE);
     }
-    GlobalState::lock_stream(hstream, |stream| {
+    GlobalState::lock_enqueue(hstream, |cmd_list, signal, wait| {
         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() {
@@ -144,19 +144,16 @@ pub fn launch_kernel(
         func.base
             .set_group_size(block_dim_x, block_dim_y, block_dim_z)?;
         func.legacy_args.reset();
-        let cmd_list = stream.command_list()?;
         unsafe {
             cmd_list.append_launch_kernel(
                 &mut func.base,
                 &[grid_dim_x, grid_dim_y, grid_dim_z],
-                None,
+                Some(signal),
-                &mut [],
+                wait,
             )?;
         }
-        cmd_list.close()?;
+        Ok::<_, CUresult>(())
-        stream.queue.execute_and_synchronize(cmd_list)?;
+    })
-        Ok(())
-    })?
 }
 
 fn round_up_to_multiple(x: usize, multiple: usize) -> usize {

zluda/src/impl/memory.rs

@@ -13,7 +13,7 @@ pub fn alloc_v2(dptr: *mut *mut c_void, bytesize: usize) -> Result<(), CUresult>
 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)? };
-        Ok::<_, l0::sys::ze_result_t>(())
+        Ok(())
     })
 }
 
@@ -26,41 +26,35 @@ pub fn free_v2(ptr: *mut c_void) -> Result<(), CUresult> {
 }
 
 pub(crate) fn set_d32_v2(dst: *mut c_void, mut ui: u32, n: usize) -> Result<(), CUresult> {
-    GlobalState::lock_stream(stream::CU_STREAM_LEGACY, |stream| {
+    GlobalState::lock_enqueue(stream::CU_STREAM_LEGACY, |cmd_list, signal, wait| {
-        let cmd_list = stream.command_list()?;
         unsafe {
             cmd_list.append_memory_fill_raw(
                 dst,
                 &mut ui as *mut _ as *mut _,
                 mem::size_of::<u32>(),
                 mem::size_of::<u32>() * n,
-                None,
+                Some(signal),
-                &mut [],
+                wait,
             )
         }?;
-        cmd_list.close()?;
+        Ok(())
-        stream.queue.execute_and_synchronize(cmd_list)?;
+    })
-        Ok::<_, CUresult>(())
-    })?
 }
 
 pub(crate) fn set_d8_v2(dst: *mut c_void, mut uc: u8, n: usize) -> Result<(), CUresult> {
-    GlobalState::lock_stream(stream::CU_STREAM_LEGACY, |stream| {
+    GlobalState::lock_enqueue(stream::CU_STREAM_LEGACY, |cmd_list, signal, wait| {
-        let cmd_list = stream.command_list()?;
         unsafe {
             cmd_list.append_memory_fill_raw(
                 dst,
                 &mut uc as *mut _ as *mut _,
                 mem::size_of::<u8>(),
                 mem::size_of::<u8>() * n,
-                None,
+                Some(signal),
-                &mut [],
+                wait,
             )
         }?;
-        cmd_list.close()?;
+        Ok(())
-        stream.queue.execute_and_synchronize(cmd_list)?;
+    })
-        Ok::<_, CUresult>(())
-    })?
 }
 
 #[cfg(test)]

zluda/src/impl/mod.rs

@@ -275,17 +275,13 @@ impl GlobalState {
 
     fn lock_enqueue(
         stream: *mut stream::Stream,
-        f: impl FnOnce(
+        f: impl FnOnce(&l0::CommandList, &l0::Event<'static>, &[&l0::Event<'static>]) -> Result<(), CUresult>,
-            &mut l0::CommandList,
-            &l0::Event<'static>,
-            &[&l0::Event<'static>],
-        ) -> l0::Result<()>,
     ) -> Result<(), CUresult> {
         Self::lock_stream(stream, |stream_data| {
             let l0_dev = unsafe { (*(*stream_data.context).device).base };
             let l0_ctx = unsafe { &mut (*(*stream_data.context).device).l0_context };
             let event_pool = unsafe { &mut (*(*stream_data.context).device).event_pool };
-            let mut cmd_list = unsafe { mem::transmute(stream_data.command_list()?) };
+            let cmd_list = unsafe { mem::transmute(stream_data.command_list()?) };
             stream_data
                 .process_finished_events(&mut |(_, marker)| event_pool.mark_as_free(marker))?;
             let prev_event = stream_data.get_last_event();
@@ -293,7 +289,7 @@ impl GlobalState {
             let empty = [];
             let prev_event_slice = prev_event_array.as_ref().map_or(&empty[..], |arr| &arr[..]);
             let (new_event, new_marker) = event_pool.get(l0_dev, l0_ctx)?;
-            f(&mut cmd_list, &new_event, prev_event_slice)?;
+            f(&cmd_list, &new_event, prev_event_slice)?;
             cmd_list.close()?;
             unsafe { stream_data.queue.execute(&cmd_list, None)? };
             stream_data.push_event((new_event, new_marker));