hal/gles: fences and queries

wgpu-hal/src/gles/adapter.rs

@@ -279,7 +279,7 @@ impl super::Adapter {
 impl crate::Adapter<super::Api> for super::Adapter {
     unsafe fn open(
         &self,
-        features: wgt::Features,
+        _features: wgt::Features,
     ) -> Result<crate::OpenDevice<super::Api>, crate::DeviceError> {
         let gl = &self.shared.context;
         gl.pixel_store_i32(glow::UNPACK_ALIGNMENT, 1);
@@ -294,8 +294,8 @@ impl crate::Adapter<super::Api> for super::Adapter {
             },
             queue: super::Queue {
                 shared: Arc::clone(&self.shared),
-                features,
                 copy_fbo: gl.create_framebuffer().unwrap(),
+                temp_query_results: Vec::new(),
             },
         })
     }

wgpu-hal/src/gles/command.rs

@@ -1,4 +1,3 @@
-use super::Resource; //TEMP
 use super::{conv, Command as C};
 use std::{mem, ops::Range};
 
@@ -6,13 +5,14 @@ impl super::CommandBuffer {
     fn clear(&mut self) {
         self.label = None;
         self.commands.clear();
-        self.data.clear();
+        self.data_bytes.clear();
+        self.data_words.clear();
     }
 
     fn add_marker(&mut self, marker: &str) -> Range<u32> {
-        let start = self.data.len() as u32;
-        self.data.extend(marker.as_bytes());
-        start..self.data.len() as u32
+        let start = self.data_bytes.len() as u32;
+        self.data_bytes.extend(marker.as_bytes());
+        start..self.data_bytes.len() as u32
     }
 }
 
@@ -153,17 +153,39 @@ impl crate::CommandEncoder<super::Api> for super::CommandEncoder {
         }
     }
 
-    unsafe fn begin_query(&mut self, set: &Resource, index: u32) {}
-    unsafe fn end_query(&mut self, set: &Resource, index: u32) {}
-    unsafe fn write_timestamp(&mut self, set: &Resource, index: u32) {}
-    unsafe fn reset_queries(&mut self, set: &Resource, range: Range<u32>) {}
+    unsafe fn begin_query(&mut self, set: &super::QuerySet, index: u32) {
+        let query = set.queries[index as usize];
+        self.cmd_buffer
+            .commands
+            .push(C::BeginQuery(query, set.target));
+    }
+    unsafe fn end_query(&mut self, set: &super::QuerySet, _index: u32) {
+        self.cmd_buffer.commands.push(C::EndQuery(set.target));
+    }
+    unsafe fn write_timestamp(&mut self, _set: &super::QuerySet, _index: u32) {
+        unimplemented!()
+    }
+    unsafe fn reset_queries(&mut self, _set: &super::QuerySet, _range: Range<u32>) {
+        //TODO: what do we do here?
+    }
     unsafe fn copy_query_results(
         &mut self,
-        set: &Resource,
+        set: &super::QuerySet,
         range: Range<u32>,
         buffer: &super::Buffer,
         offset: wgt::BufferAddress,
     ) {
+        let start = self.cmd_buffer.data_words.len();
+        self.cmd_buffer
+            .data_words
+            .extend_from_slice(&set.queries[range.start as usize..range.end as usize]);
+        let query_range = start as u32..self.cmd_buffer.data_words.len() as u32;
+        self.cmd_buffer.commands.push(C::CopyQueryResults {
+            query_range,
+            dst: buffer.raw,
+            dst_target: buffer.target,
+            dst_offset: offset,
+        });
     }
 
     // render

wgpu-hal/src/gles/device.rs

@@ -1,5 +1,4 @@
 use super::conv;
-use super::Resource; //TEMP
 use crate::util::map_naga_stage;
 use glow::HasContext;
 use std::{convert::TryInto, iter, ptr::NonNull, sync::Arc};
@@ -713,28 +712,72 @@ impl crate::Device<super::Api> for super::Device {
     unsafe fn create_query_set(
         &self,
         desc: &wgt::QuerySetDescriptor<crate::Label>,
-    ) -> Result<Resource, crate::DeviceError> {
-        Ok(Resource)
+    ) -> Result<super::QuerySet, crate::DeviceError> {
+        let gl = &self.shared.context;
+
+        let mut queries = Vec::with_capacity(desc.count as usize);
+        for _ in 0..desc.count {
+            let query = gl
+                .create_query()
+                .map_err(|_| crate::DeviceError::OutOfMemory)?;
+            queries.push(query);
+        }
+
+        Ok(super::QuerySet {
+            queries: queries.into_boxed_slice(),
+            target: match desc.ty {
+                wgt::QueryType::Occlusion => glow::ANY_SAMPLES_PASSED,
+                _ => unimplemented!(),
+            },
+        })
+    }
+    unsafe fn destroy_query_set(&self, set: super::QuerySet) {
+        let gl = &self.shared.context;
+        for &query in set.queries.iter() {
+            gl.delete_query(query);
+        }
+    }
+    unsafe fn create_fence(&self) -> Result<super::Fence, crate::DeviceError> {
+        Ok(super::Fence {
+            last_completed: 0,
+            pending: Vec::new(),
+        })
+    }
+    unsafe fn destroy_fence(&self, fence: super::Fence) {
+        let gl = &self.shared.context;
+        for (_, sync) in fence.pending {
+            gl.delete_sync(sync);
         }
-    unsafe fn destroy_query_set(&self, set: Resource) {}
-    unsafe fn create_fence(&self) -> Result<Resource, crate::DeviceError> {
-        Ok(Resource)
     }
-    unsafe fn destroy_fence(&self, fence: Resource) {}
     unsafe fn get_fence_value(
         &self,
-        fence: &Resource,
+        fence: &super::Fence,
     ) -> Result<crate::FenceValue, crate::DeviceError> {
-        Ok(0)
+        Ok(fence.get_latest(&self.shared.context))
     }
     unsafe fn wait(
         &self,
-        fence: &Resource,
-        value: crate::FenceValue,
+        fence: &super::Fence,
+        wait_value: crate::FenceValue,
         timeout_ms: u32,
     ) -> Result<bool, crate::DeviceError> {
+        if fence.last_completed < wait_value {
+            let gl = &self.shared.context;
+            let timeout_ns = (timeout_ms as u64 * 1_000_000).min(!0u32 as u64);
+            let &(_, sync) = fence
+                .pending
+                .iter()
+                .find(|&&(value, _)| value >= wait_value)
+                .unwrap();
+            match gl.client_wait_sync(sync, glow::SYNC_FLUSH_COMMANDS_BIT, timeout_ns as i32) {
+                glow::TIMEOUT_EXPIRED => Ok(false),
+                glow::CONDITION_SATISFIED | glow::ALREADY_SIGNALED => Ok(true),
+                _ => Err(crate::DeviceError::Lost),
+            }
+        } else {
             Ok(true)
         }
+    }
 
     unsafe fn start_capture(&self) -> bool {
         false

wgpu-hal/src/gles/mod.rs

@@ -18,12 +18,9 @@ use std::{ops::Range, sync::Arc};
 
 #[derive(Clone)]
 pub struct Api;
-#[derive(Debug)]
-pub struct Resource;
 
 //Note: we can support more samplers if not every one of them is used at a time,
 // but it probably doesn't worth it.
-const MAX_SAMPLER_SLOTS: usize = 16;
 const MAX_TEXTURE_SLOTS: usize = 16;
 
 impl crate::Api for Api {
@@ -41,8 +38,8 @@ impl crate::Api for Api {
     type SurfaceTexture = Texture;
     type TextureView = TextureView;
     type Sampler = Sampler;
-    type QuerySet = Resource;
-    type Fence = Resource;
+    type QuerySet = QuerySet;
+    type Fence = Fence;
 
     type BindGroupLayout = BindGroupLayout;
     type BindGroup = BindGroup;
@@ -184,8 +181,8 @@ pub struct Device {
 
 pub struct Queue {
     shared: Arc<AdapterShared>,
-    features: wgt::Features,
     copy_fbo: glow::Framebuffer,
+    temp_query_results: Vec<u64>,
 }
 
 #[derive(Debug)]
@@ -344,6 +341,47 @@ pub struct ComputePipeline {
     inner: PipelineInner,
 }
 
+#[derive(Debug)]
+pub struct QuerySet {
+    queries: Box<[glow::Query]>,
+    target: BindTarget,
+}
+
+#[derive(Debug)]
+pub struct Fence {
+    last_completed: crate::FenceValue,
+    pending: Vec<(crate::FenceValue, glow::Fence)>,
+}
+
+unsafe impl Send for Fence {}
+unsafe impl Sync for Fence {}
+
+impl Fence {
+    fn get_latest(&self, gl: &glow::Context) -> crate::FenceValue {
+        let mut max_value = self.last_completed;
+        for &(value, sync) in self.pending.iter() {
+            let status = unsafe { gl.get_sync_status(sync) };
+            if status == glow::SIGNALED {
+                max_value = value;
+            }
+        }
+        max_value
+    }
+
+    fn maintain(&mut self, gl: &glow::Context) {
+        let latest = self.get_latest(gl);
+        for &(value, sync) in self.pending.iter() {
+            if value <= latest {
+                unsafe {
+                    gl.delete_sync(sync);
+                }
+            }
+        }
+        self.pending.retain(|&(value, _)| value > latest);
+        self.last_completed = latest;
+    }
+}
+
 #[derive(Debug)]
 struct TextureCopyInfo {
     external_format: u32,
@@ -419,6 +457,14 @@ enum Command {
         copy: crate::BufferTextureCopy,
     },
     SetIndexBuffer(glow::Buffer),
+    BeginQuery(glow::Query, BindTarget),
+    EndQuery(BindTarget),
+    CopyQueryResults {
+        query_range: Range<u32>,
+        dst: glow::Buffer,
+        dst_target: BindTarget,
+        dst_offset: wgt::BufferAddress,
+    },
     InsertDebugMarker(Range<u32>),
     PushDebugGroup(Range<u32>),
     PopDebugGroup,
@@ -428,7 +474,8 @@ enum Command {
 pub struct CommandBuffer {
     label: Option<String>,
     commands: Vec<Command>,
-    data: Vec<u8>,
+    data_bytes: Vec<u8>,
+    data_words: Vec<u32>,
 }
 
 #[derive(Default)]

wgpu-hal/src/gles/queue.rs

@@ -1,6 +1,6 @@
 use super::Command as C;
 use glow::HasContext;
-use std::ops::Range;
+use std::{mem, ops::Range, slice};
 
 const DEBUG_ID: u32 = 0;
 
@@ -16,7 +16,7 @@ fn is_3d_target(target: super::BindTarget) -> bool {
 }
 
 impl super::Queue {
-    unsafe fn process(&mut self, command: &C, data: &[u8]) {
+    unsafe fn process(&mut self, command: &C, data_bytes: &[u8], data_words: &[u32]) {
         let gl = &self.shared.context;
         match *command {
             C::Draw {
@@ -265,8 +265,34 @@ impl super::Queue {
             C::SetIndexBuffer(buffer) => {
                 gl.bind_buffer(glow::ELEMENT_ARRAY_BUFFER, Some(buffer));
             }
+            C::BeginQuery(query, target) => {
+                gl.begin_query(target, query);
+            }
+            C::EndQuery(target) => {
+                gl.end_query(target);
+            }
+            C::CopyQueryResults {
+                ref query_range,
+                dst,
+                dst_target,
+                dst_offset,
+            } => {
+                self.temp_query_results.clear();
+                for &query in
+                    data_words[query_range.start as usize..query_range.end as usize].iter()
+                {
+                    let result = gl.get_query_parameter_u32(query, glow::QUERY_RESULT);
+                    self.temp_query_results.push(result as u64);
+                }
+                let query_data = slice::from_raw_parts(
+                    self.temp_query_results.as_ptr() as *const u8,
+                    self.temp_query_results.len() * mem::size_of::<u64>(),
+                );
+                gl.bind_buffer(dst_target, Some(dst));
+                gl.buffer_sub_data_u8_slice(dst_target, dst_offset as i32, query_data);
+            }
             C::InsertDebugMarker(ref range) => {
-                let marker = extract_marker(data, range);
+                let marker = extract_marker(data_bytes, range);
                 gl.debug_message_insert(
                     glow::DEBUG_SOURCE_APPLICATION,
                     glow::DEBUG_TYPE_MARKER,
@@ -276,7 +302,7 @@ impl super::Queue {
                 );
             }
             C::PushDebugGroup(ref range) => {
-                let marker = extract_marker(data, range);
+                let marker = extract_marker(data_bytes, range);
                 gl.push_debug_group(glow::DEBUG_SOURCE_APPLICATION, DEBUG_ID, marker);
             }
             C::PopDebugGroup => {
@@ -290,13 +316,23 @@ impl crate::Queue<super::Api> for super::Queue {
     unsafe fn submit(
         &mut self,
         command_buffers: &[&super::CommandBuffer],
-        signal_fence: Option<(&mut super::Resource, crate::FenceValue)>,
+        signal_fence: Option<(&mut super::Fence, crate::FenceValue)>,
     ) -> Result<(), crate::DeviceError> {
         for cmd_buf in command_buffers.iter() {
             for command in cmd_buf.commands.iter() {
-                self.process(command, &cmd_buf.data);
+                self.process(command, &cmd_buf.data_bytes, &cmd_buf.data_words);
             }
         }
+
+        if let Some((fence, value)) = signal_fence {
+            let gl = &self.shared.context;
+            fence.maintain(gl);
+            let sync = gl
+                .fence_sync(glow::SYNC_GPU_COMMANDS_COMPLETE, 0)
+                .map_err(|_| crate::DeviceError::OutOfMemory)?;
+            fence.pending.push((value, sync));
+        }
+
         Ok(())
     }