core: manual buffer sync with sync-rows, reuse unchanged buffers and pipelines across activations

Also lets the screenshot suite point at another reference directory and settle longer.

Amp-Thread-ID: https://ampcode.com/threads/T-01a0b2dd-db32-712d-99ce-12ce4a86d979
Co-authored-by: Heaust Azure <heaust.azure@gmail.com>

core/render_graph/gpu_resource.rs

@@ -9,6 +9,7 @@ use serde_json::Value;
 use crate::gpu::Wgpu;
 use crate::graph::{Binding, Execution, Extent, Pass, RenderGraph, RenderPipeline, Texture};
 use crate::render_data::RenderData;
+use crate::store::Loadout;
 
 pub struct GpuResources {
     pub buffers: HashMap<String, GpuBuffer>,
@@ -38,6 +39,7 @@ pub struct ProfileMap {
 pub struct GpuBuffer {
     pub buffer: wgpu::Buffer,
     pub source: String,
+    pub usage: wgpu::BufferUsages,
     pub sync_each_frame: bool,
 }
 
@@ -65,28 +67,48 @@ pub enum GpuPass {
 }
 
 impl GpuResources {
+    /// Builds the GPU side of `graph`. Buffers whose source array, size, and usage are unchanged
+    /// and pipelines whose declaration is unchanged are taken over from `previous` instead of being
+    /// recreated and re-uploaded; bind groups and bundles are always rebuilt.
     pub fn activate(
         graph: &RenderGraph,
         gpu: &Wgpu,
         data: &RenderData,
-        previous: Option<&Self>,
+        previous: Option<&Loadout>,
     ) -> Result<Self, String> {
         let mut buffers = HashMap::new();
         for source in &graph.resources.buffers {
             let rows = data.rows(&source.array).ok_or("GRAPH_ARRAY_UNKNOWN")?;
-            let buffer = gpu.device.create_buffer(&wgpu::BufferDescriptor {
-                label: Some(&source.id),
-                size: u64::from(rows.bytes.max(4)),
-                usage: buffer_usage(&source.usage)?,
-                mapped_at_creation: false,
-            });
-            gpu.queue
-                .write_buffer(&buffer, 0, data.bytes(&source.array).unwrap());
+            let size = u64::from(rows.bytes.max(4));
+            let usage = buffer_usage(&source.usage)?;
+            let reused = previous
+                .and_then(|previous| previous.resources.buffers.get(&source.id))
+                .filter(|buffer| {
+                    buffer.source == source.array
+                        && buffer.usage == usage
+                        && buffer.buffer.size() == size
+                })
+                .map(|buffer| buffer.buffer.clone());
+            let buffer = match reused {
+                Some(buffer) => buffer,
+                None => {
+                    let buffer = gpu.device.create_buffer(&wgpu::BufferDescriptor {
+                        label: Some(&source.id),
+                        size,
+                        usage,
+                        mapped_at_creation: false,
+                    });
+                    gpu.queue
+                        .write_buffer(&buffer, 0, data.bytes(&source.array).unwrap());
+                    buffer
+                }
+            };
             buffers.insert(
                 source.id.clone(),
                 GpuBuffer {
                     buffer,
                     source: source.array.clone(),
+                    usage,
                     sync_each_frame: source.sync == "frame",
                 },
             );
@@ -102,11 +124,12 @@ impl GpuResources {
                     let key = source.key()?;
                     if !source.transient {
                         if let Some(texture) = previous
-                            .and_then(|resources| {
-                                resources
+                            .and_then(|previous| {
+                                previous
+                                    .resources
                                     .texture_slots
                                     .get(&source.id)
-                                    .map(|slot| &resources.textures[*slot])
+                                    .map(|slot| &previous.resources.textures[*slot])
                             })
                             .filter(|texture| texture.key == key)
                         {
@@ -147,7 +170,23 @@ impl GpuResources {
             .render
             .iter()
             .map(|source| {
-                create_render_pipeline(source, gpu).map(|pipeline| (source.id.clone(), pipeline))
+                let reused = previous.and_then(|previous| {
+                    previous
+                        .graph
+                        .pipelines
+                        .render
+                        .iter()
+                        .find(|candidate| candidate.same_as(source))
+                        .and_then(|candidate| {
+                            previous.resources.render_pipelines.get(&candidate.id)
+                        })
+                        .cloned()
+                });
+                match reused {
+                    Some(pipeline) => Ok(pipeline),
+                    None => create_render_pipeline(source, gpu),
+                }
+                .map(|pipeline| (source.id.clone(), pipeline))
             })
             .collect::<Result<HashMap<_, _>, _>>()?;
         let compute_pipelines = graph
@@ -155,6 +194,21 @@ impl GpuResources {
             .compute
             .iter()
             .map(|source| {
+                let reused = previous.and_then(|previous| {
+                    previous
+                        .graph
+                        .pipelines
+                        .compute
+                        .iter()
+                        .find(|candidate| candidate.same_as(source))
+                        .and_then(|candidate| {
+                            previous.resources.compute_pipelines.get(&candidate.id)
+                        })
+                        .cloned()
+                });
+                if let Some(pipeline) = reused {
+                    return Ok::<_, String>((source.id.clone(), pipeline));
+                }
                 let module = gpu
                     .device
                     .create_shader_module(wgpu::ShaderModuleDescriptor {
@@ -171,7 +225,7 @@ impl GpuResources {
                             compilation_options: Default::default(),
                             cache: None,
                         });
-                Ok::<_, String>((source.id.clone(), pipeline))
+                Ok((source.id.clone(), pipeline))
             })
             .collect::<Result<HashMap<_, _>, _>>()?;
 
@@ -251,6 +305,28 @@ impl GpuResources {
         Ok(resources)
     }
 
+    /// Copies rows `first .. first + count` of a source array into every GPU buffer bound to it.
+    pub fn sync_rows(
+        &self,
+        name: &str,
+        first: u32,
+        count: u32,
+        gpu: &Wgpu,
+        data: &RenderData,
+    ) -> Result<(), String> {
+        let rows = data.rows(name).ok_or("ROWS_UNKNOWN")?;
+        let end = first
+            .checked_add(count)
+            .filter(|end| *end <= rows.rows)
+            .ok_or("ROW_RANGE")?;
+        let start = (first * rows.stride) as usize;
+        let bytes = &data.bytes(name).unwrap()[start..(end * rows.stride) as usize];
+        for buffer in self.buffers.values().filter(|buffer| buffer.source == name) {
+            gpu.queue.write_buffer(&buffer.buffer, start as u64, bytes);
+        }
+        Ok(())
+    }
+
     pub fn texture_view(&self, id: &str) -> Option<&wgpu::TextureView> {
         self.texture_slots
             .get(id)

core/render_graph/render_graph.rs

@@ -85,7 +85,7 @@ pub struct PipelineDeclarations {
     pub compute: Vec<ComputePipeline>,
 }
 
-#[derive(Clone, Deserialize)]
+#[derive(Clone, Deserialize, PartialEq)]
 #[serde(rename_all = "camelCase")]
 pub struct RenderPipeline {
     pub id: String,
@@ -102,7 +102,7 @@ pub struct RenderPipeline {
     pub multisample: Value,
 }
 
-#[derive(Clone, Deserialize)]
+#[derive(Clone, Deserialize, PartialEq, Eq)]
 pub struct ComputePipeline {
     pub id: String,
     pub code: String,
@@ -110,7 +110,7 @@ pub struct ComputePipeline {
     pub entry: String,
 }
 
-#[derive(Clone, Deserialize)]
+#[derive(Clone, Deserialize, PartialEq, Eq)]
 pub struct VertexStage {
     #[serde(default = "vertex_entry")]
     pub entry: String,
@@ -127,7 +127,7 @@ impl Default for VertexStage {
     }
 }
 
-#[derive(Clone, Default, Deserialize)]
+#[derive(Clone, Default, Deserialize, PartialEq, Eq)]
 #[serde(rename_all = "camelCase")]
 pub struct VertexBuffer {
     pub array_stride: u64,
@@ -137,7 +137,7 @@ pub struct VertexBuffer {
     pub attributes: Vec<VertexAttribute>,
 }
 
-#[derive(Clone, Deserialize)]
+#[derive(Clone, Deserialize, PartialEq, Eq)]
 #[serde(rename_all = "camelCase")]
 pub struct VertexAttribute {
     pub format: String,
@@ -145,7 +145,7 @@ pub struct VertexAttribute {
     pub shader_location: u32,
 }
 
-#[derive(Clone, Deserialize)]
+#[derive(Clone, Deserialize, PartialEq)]
 pub struct FragmentStage {
     #[serde(default = "fragment_entry")]
     pub entry: String,
@@ -162,7 +162,7 @@ impl Default for FragmentStage {
     }
 }
 
-#[derive(Clone, Deserialize)]
+#[derive(Clone, Deserialize, PartialEq)]
 #[serde(rename_all = "camelCase")]
 pub struct FragmentTarget {
     pub format: String,
@@ -519,7 +519,7 @@ impl RenderGraph {
             .resources
             .buffers
             .iter()
-            .any(|buffer| !matches!(buffer.sync.as_str(), "frame" | "loadout"))
+            .any(|buffer| !matches!(buffer.sync.as_str(), "frame" | "manual"))
         {
             return Err("GRAPH_BUFFER_SYNC");
         }
@@ -561,6 +561,24 @@ impl Pass {
     }
 }
 
+impl RenderPipeline {
+    /// Same GPU pipeline regardless of declaration ID.
+    pub(crate) fn same_as(&self, other: &Self) -> bool {
+        self.code == other.code
+            && self.vertex == other.vertex
+            && self.fragment == other.fragment
+            && self.primitive == other.primitive
+            && self.depth_stencil == other.depth_stencil
+            && self.multisample == other.multisample
+    }
+}
+
+impl ComputePipeline {
+    pub(crate) fn same_as(&self, other: &Self) -> bool {
+        self.code == other.code && self.entry == other.entry
+    }
+}
+
 impl Texture {
     pub(crate) fn key(&self) -> Result<String, &'static str> {
         let mut usage = self.usage.clone();

core/render_graph/store.rs

@@ -26,12 +26,7 @@ impl Store {
 
     pub fn switch(&mut self, id: &str, gpu: &Wgpu, data: &RenderData) -> Result<(), String> {
         let graph = self.graphs.get(id).ok_or("GRAPH_UNKNOWN")?.clone();
-        let mut resources = GpuResources::activate(
-            &graph,
-            gpu,
-            data,
-            self.active.as_ref().map(|active| &active.resources),
-        )?;
+        let mut resources = GpuResources::activate(&graph, gpu, data, self.active.as_ref())?;
         for (name, levels) in &self.texture_sources {
             for (mip_level, image) in levels {
                 if resources.needs_upload(name, *mip_level) {
@@ -72,6 +67,20 @@ impl Store {
         }
     }
 
+    pub fn sync_rows(
+        &self,
+        name: &str,
+        first: u32,
+        count: u32,
+        gpu: &Wgpu,
+        data: &RenderData,
+    ) -> Result<(), String> {
+        match &self.active {
+            Some(active) => active.resources.sync_rows(name, first, count, gpu, data),
+            None => Ok(()),
+        }
+    }
+
     pub fn active_mut(&mut self) -> Option<&mut Loadout> {
         self.active.as_mut()
     }
@@ -104,12 +113,7 @@ impl Store {
             return Ok(());
         };
         let graph = active.graph.clone();
-        let mut resources = GpuResources::activate(
-            &graph,
-            gpu,
-            data,
-            self.active.as_ref().map(|active| &active.resources),
-        )?;
+        let mut resources = GpuResources::activate(&graph, gpu, data, self.active.as_ref())?;
         for (name, levels) in &self.texture_sources {
             for (mip_level, image) in levels {
                 if resources.needs_upload(name, *mip_level) {