rusticl: add QueueContext to track GPU state

Signed-off-by: Karol Herbst <kherbst@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/25946>

src/gallium/frontends/rusticl/core/event.rs

@@ -4,7 +4,6 @@ use crate::core::context::*;
 use crate::core::queue::*;
 use crate::impl_cl_type_trait;
 
-use mesa_rust::pipe::context::*;
 use mesa_rust::pipe::query::*;
 use mesa_rust_gen::*;
 use mesa_rust_util::static_assert;
@@ -24,7 +23,7 @@ static_assert!(CL_RUNNING == 1);
 static_assert!(CL_SUBMITTED == 2);
 static_assert!(CL_QUEUED == 3);
 
-pub type EventSig = Box<dyn FnOnce(&Arc<Queue>, &PipeContext) -> CLResult<()>>;
+pub type EventSig = Box<dyn FnOnce(&Arc<Queue>, &QueueContext) -> CLResult<()>>;
 
 pub enum EventTimes {
     Queued = CL_PROFILING_COMMAND_QUEUED as isize,
@@ -194,7 +193,7 @@ impl Event {
     // We always assume that work here simply submits stuff to the hardware even if it's just doing
     // sw emulation or nothing at all.
     // If anything requets waiting, we will update the status through fencing later.
-    pub fn call(&self, ctx: &PipeContext) {
+    pub fn call(&self, ctx: &QueueContext) {
         let mut lock = self.state();
         let status = lock.status;
         let queue = self.queue.as_ref().unwrap();

src/gallium/frontends/rusticl/core/queue.rs

@@ -10,6 +10,7 @@ use mesa_rust_util::properties::*;
 use rusticl_opencl_gen::*;
 
 use std::mem;
+use std::ops::Deref;
 use std::sync::mpsc;
 use std::sync::Arc;
 use std::sync::Mutex;
@@ -17,6 +18,33 @@ use std::sync::Weak;
 use std::thread;
 use std::thread::JoinHandle;
 
+/// State tracking wrapper for [PipeContext]
+///
+/// Used for tracking bound GPU state to lower CPU overhead and centralize state tracking
+pub struct QueueContext {
+    ctx: PipeContext,
+}
+
+impl QueueContext {
+    fn new_for(device: &Device) -> CLResult<Self> {
+        Ok(Self {
+            ctx: device
+                .screen()
+                .create_context()
+                .ok_or(CL_OUT_OF_HOST_MEMORY)?,
+        })
+    }
+}
+
+// This should go once we moved all state tracking into QueueContext
+impl Deref for QueueContext {
+    type Target = PipeContext;
+
+    fn deref(&self) -> &Self::Target {
+        &self.ctx
+    }
+}
+
 struct QueueState {
     pending: Vec<Arc<Event>>,
     last: Weak<Event>,
@@ -53,7 +81,7 @@ impl Queue {
     ) -> CLResult<Arc<Queue>> {
         // we assume that memory allocation is the only possible failure. Any other failure reason
         // should be detected earlier (e.g.: checking for CAPs).
-        let pipe = device.screen().create_context().unwrap();
+        let ctx = QueueContext::new_for(device)?;
         let (tx_q, rx_t) = mpsc::channel::<Vec<Arc<Event>>>();
         Ok(Arc::new(Self {
             base: CLObjectBase::new(),
@@ -81,7 +109,7 @@ impl Queue {
                         // If we hit any deps from another queue, flush so we don't risk a dead
                         // lock.
                         if e.deps.iter().any(|ev| ev.queue != e.queue) {
-                            flush_events(&mut flushed, &pipe);
+                            flush_events(&mut flushed, &ctx);
                         }
 
                         // We have to wait on user events or events from other queues.
@@ -98,25 +126,25 @@ impl Queue {
                             continue;
                         }
 
-                        e.call(&pipe);
+                        e.call(&ctx);
 
                         if e.is_user() {
                             // On each user event we flush our events as application might
                             // wait on them before signaling user events.
-                            flush_events(&mut flushed, &pipe);
+                            flush_events(&mut flushed, &ctx);
 
                             // Wait on user events as they are synchronization points in the
                             // application's control.
                             e.wait();
                         } else if Platform::dbg().sync_every_event {
                             flushed.push(e);
-                            flush_events(&mut flushed, &pipe);
+                            flush_events(&mut flushed, &ctx);
                         } else {
                             flushed.push(e);
                         }
                     }
 
-                    flush_events(&mut flushed, &pipe);
+                    flush_events(&mut flushed, &ctx);
                 })
                 .unwrap(),
         }))