subgroup_threads + download sync fixes

crates/krnl/src/context/device/backend/vulkan.rs

@@ -1280,15 +1280,15 @@ impl super::Slice for Slice {
                 self.range.start + offset,
                 chunk.len(),
             )?;
-            for event in std::mem::take(&mut *buffer_events) {
-                event.wait()?;
-            }
             if let Some(event) = host_buffer.event.take() {
                 event.wait()?;
             }
             unsafe {
                 (&mut *host_buffer.raw.mapped_slice.unwrap())[..chunk.len()].copy_from_slice(chunk);
             }
+            for event in std::mem::take(&mut *buffer_events) {
+                event.wait()?;
+            }
             let event = command_buffer.submit()?;
             offset += chunk.len();
             if offset < bytes.len() {
@@ -1307,7 +1307,7 @@ impl super::Slice for Slice {
             return Ok(());
         }
         let device = &self.buffer.device;
-        let mut buffer_events = self.buffer.events.read().clone();
+        let buffer_events = self.buffer.events.read().clone();
         if device.transfer_queue.is_none() {
             if let Some(mapped_slice) = self.buffer.raw.as_ref().unwrap().mapped_slice {
                 for event in buffer_events {
@@ -1340,6 +1340,9 @@ impl super::Slice for Slice {
             0,
             host_buffer.len,
         )?;
+        for event in buffer_events {
+            event.wait()?;
+        }
         host_buffer.event.replace(command_buffer.submit()?);
         let mut chunk_iter = bytes.chunks_mut(HOST_BUFFER_LEN).peekable();
         while let Some(chunk) = chunk_iter.next() {
@@ -1358,9 +1361,6 @@ impl super::Slice for Slice {
             } else {
                 None
             };
-            for event in std::mem::take(&mut buffer_events) {
-                event.wait()?;
-            }
             let host_buffer = &mut host_buffers.first;
             if let Some(event) = host_buffer.event.take() {
                 event.wait()?;

crates/krnl/src/kernel.rs

@@ -69,10 +69,10 @@ host_only! {
                 #[cfg(feature = "device")]
                 Context::Device(device) => {
                     let threads = self.threads.unwrap_or(256) as u32;
-                    let subgroup_threads = if let Some(subgroup_threads) = self.subgroup_threads {
+                    let subgroup_threads = if let Some(subgroup_threads) = self.subgroup_threads && subgroup_threads != device.default_subgroup_threads() {
                         Some(subgroup_threads as u32)
                     } else {
-                        Some(device.default_subgroup_threads() as u32)
+                        None
                     };
                     let key = KernelKey::new::<T>(&self.args, threads, subgroup_threads);
                     let raw = RawKernel::get_or_create(device, key, || KernelCreateInfo::new::<T>(&self.args, threads as u32, subgroup_threads)).unwrap();