From 06b82af5aab6c176666024f93f18e4f2242f3246 Mon Sep 17 00:00:00 2001
From: shadowcodex <1348053+shadowcodex@users.noreply.github.com>
Date: Tue, 11 Aug 2026 19:42:02 -0500
Subject: [PATCH] webgpu: correct two defects in per-frame attribute buffer
 handling

---
 packages/brometal/src/runtime/webgpu.ts      | 106 +++++++++++++++----
 scripts/gpu/entry.ts                         |  38 +++++++
 scripts/gpu/fixtures/gpu-batch.shader.gen.ts |  33 ++++++
 scripts/gpu/fixtures/gpu-batch.shader.ts     |  20 ++++
 4 files changed, 177 insertions(+), 20 deletions(-)
 create mode 100644 scripts/gpu/fixtures/gpu-batch.shader.gen.ts
 create mode 100644 scripts/gpu/fixtures/gpu-batch.shader.ts

diff --git a/packages/brometal/src/runtime/webgpu.ts b/packages/brometal/src/runtime/webgpu.ts
index fffed19..c121c27 100644
--- a/packages/brometal/src/runtime/webgpu.ts
+++ b/packages/brometal/src/runtime/webgpu.ts
@@ -337,6 +337,14 @@ interface GpuAttributeState {
   buffer: GPUBuffer;
   capacity: number;
   elementCount: number;
+  /**
+   * Byte offset the most recent upload was written at, and the frame it
+   * happened in. A second upload within one frame appends rather than
+   * overwriting — see `uploadAttribute`.
+   */
+  offset: number;
+  writtenThisFrame: number;
+  frame: number;
 }
 
 interface GpuTextureBinding {
@@ -594,6 +602,20 @@ export function createWebgpuProgram<A extends GpuRecord, I extends GpuRecord, U
   const instanceAttributes = {} as { [K in keyof I]: AttributeHandle };
   let isInstanced = false;
 
+  /**
+   * Buffers that a larger allocation replaced during a frame.
+   *
+   * Do not destroy these immediately. A draw command already recorded into the
+   * open render pass still refers to them, and destroying one fails the whole
+   * submit with "used in submit while destroyed" — every draw in the frame
+   * fails, not only the draw that grew.
+   *
+   * The uniform ring below has the same problem and uses the same method:
+   * destroy at the next frame boundary, when the submit that could refer to
+   * them is complete.
+   */
+  const retired: GPUBuffer[] = [];
+
   const uploadAttribute = (entry: AttributeLayoutEntry, data: Float32Array): void => {
     if (data.length % entry.size !== 0) {
       throw new Error(
@@ -602,20 +624,50 @@ export function createWebgpuProgram<A extends GpuRecord, I extends GpuRecord, U
     }
     const states = entry.divisor === 1 ? instanceStates : vertexStates;
     let state = states.get(entry.name);
-    if (state === undefined || state.capacity < data.byteLength) {
-      state?.buffer.destroy();
-      state = {
-        buffer: device.createBuffer({
-          size: data.byteLength,
-          usage: GPUBufferUsage.VERTEX | GPUBufferUsage.COPY_DST,
-        }),
-        capacity: data.byteLength,
-        elementCount: 0,
-      };
-      states.set(entry.name, state);
+
+    // A frame is one command encoder, submitted once at the end. queue.writeBuffer
+    // is ordered against that submit, not against the draw commands inside it, so
+    // every draw in the frame reads whatever was written LAST. Two draws that each
+    // write at offset 0 therefore both read the second batch.
+    //
+    // The uniform ring below has the same problem and uses the same method: a
+    // second upload in the same frame writes at a new offset, and the draw binds
+    // the vertex buffer there. This is what lets one program draw several batches
+    // in one frame.
+    const repeat = state !== undefined && state.frame === internals.frame;
+    const offset = repeat ? state!.writtenThisFrame : 0;
+    const needed = offset + data.byteLength;
+
+    if (state === undefined || state.capacity < needed) {
+      const grown = Math.max(needed, (state?.capacity ?? 0) * 2);
+      const replacement = device.createBuffer({
+        size: grown,
+        usage: GPUBufferUsage.VERTEX | GPUBufferUsage.COPY_DST,
+      });
+      if (state !== undefined) {
+        // Draw commands from earlier in this frame still point into the old
+        // buffer, at their own offsets. It stays alive until the frame boundary.
+        retired.push(state.buffer);
+        state.buffer = replacement;
+        state.capacity = grown;
+      } else {
+        state = {
+          buffer: replacement,
+          capacity: grown,
+          elementCount: 0,
+          offset: 0,
+          writtenThisFrame: 0,
+          frame: -1,
+        };
+        states.set(entry.name, state);
+      }
     }
+
     state.elementCount = data.length / entry.size;
-    device.queue.writeBuffer(state.buffer, 0, data as unknown as BufferSource);
+    state.offset = offset;
+    state.writtenThisFrame = offset + data.byteLength;
+    state.frame = internals.frame;
+    device.queue.writeBuffer(state.buffer, offset, data as unknown as BufferSource);
   };
 
   for (const entry of compiled.layout.attributes) {
@@ -742,6 +794,20 @@ export function createWebgpuProgram<A extends GpuRecord, I extends GpuRecord, U
       if (pass === null) {
         throw new Error('BroMetal: draw() must be called inside renderer.loop()');
       }
+      // Per-frame bookkeeping before any early exit. A frame in which every draw
+      // is skipped must still release retired buffers and restart the slot ring,
+      // or the buffers live until dispose() and the ring can overwrite an offset
+      // a recorded draw still uses.
+      if (internals.frame !== lastFrame) {
+        // The GPU has the previous frame, so buffers it retired are safe now.
+        lastFrame = internals.frame;
+        for (const buffer of retired) {
+          buffer.destroy();
+        }
+        retired.length = 0;
+        slot = -1;
+        uniformsDirty = true;
+      }
       const vertexCount = resolveCount(vertexStates, 'vertex');
       const instanceCount = isInstanced ? resolveCount(instanceStates, 'instance') : 1;
       for (const entry of compiled.layout.attributes) {
@@ -750,13 +816,7 @@ export function createWebgpuProgram<A extends GpuRecord, I extends GpuRecord, U
           throw new Error(`BroMetal: attribute '${entry.name}' has no data — call set(...) before draw()`);
         }
       }
-      if (internals.frame !== lastFrame) {
-        // New frame: restart the slot ring. Forcing a write keeps this frame's
-        // ascending slots from ever overwriting an offset already referenced.
-        lastFrame = internals.frame;
-        slot = -1;
-        uniformsDirty = true;
-      }
+
       flushUniforms();
       if (bindGroup === null) {
         bindGroup = buildBindGroup();
@@ -765,7 +825,9 @@ export function createWebgpuProgram<A extends GpuRecord, I extends GpuRecord, U
       pass.setBindGroup(0, bindGroup, uniformBuffer === null ? [] : [currentOffset]);
       compiled.layout.attributes.forEach((entry, slot) => {
         const states = entry.divisor === 1 ? instanceStates : vertexStates;
-        pass.setVertexBuffer(slot, states.get(entry.name)!.buffer);
+        const state = states.get(entry.name)!;
+        // Bind at the offset holding this draw's data, not at 0.
+        pass.setVertexBuffer(slot, state.buffer, state.offset);
       });
       if (indexBuffer !== null) {
         pass.setIndexBuffer(indexBuffer, indexFormat);
@@ -785,6 +847,10 @@ export function createWebgpuProgram<A extends GpuRecord, I extends GpuRecord, U
       instanceStates.clear();
       indexBuffer?.destroy();
       indexBuffer = null;
+      for (const buffer of retired) {
+        buffer.destroy();
+      }
+      retired.length = 0;
       uniformBuffer?.destroy();
       placeholderTexture.destroy();
     },
diff --git a/scripts/gpu/entry.ts b/scripts/gpu/entry.ts
index b684e25..3457617 100644
--- a/scripts/gpu/entry.ts
+++ b/scripts/gpu/entry.ts
@@ -19,6 +19,7 @@ import computeShader from './fixtures/gpu-compute.shader.gen';
 import readbackShader from './fixtures/gpu-readback.shader.gen';
 import targetWriteShader from './fixtures/gpu-target-write.shader.gen';
 import targetReadShader from './fixtures/gpu-target-read.shader.gen';
+import batchShader from './fixtures/gpu-batch.shader.gen';
 
 interface Check {
   name: string;
@@ -246,6 +247,43 @@ async function run(): Promise<void> {
     });
   }
 
+  // Two batches through one program in one frame. queue.writeBuffer is ordered
+  // against the frame's single submit, not against the draws inside it, so
+  // without per-draw offsets both draws read whatever was written last and the
+  // left half comes out the colour of the right. Growing the buffer between the
+  // two uploads also retires the first one mid-frame, which is the second
+  // defect: destroying it immediately fails the whole submit.
+  {
+    const batch = createProgram(renderer, batchShader);
+    await new Promise<void>((resolve) => {
+      const stop = renderer.loop(() => {
+        // Left half, red. Two vertices' worth of tint.
+        batch.attributes.aPosition.set(new Float32Array([-1, -1, 0, 0, -1, 0, -1, 1, 0]));
+        batch.attributes.aTint.set(new Float32Array([1, 0, 0, 1, 0, 0, 1, 0, 0]));
+        batch.draw();
+        // Right half, blue. A second upload in the same frame, to the same
+        // attributes, and larger so the buffer grows and the first is retired.
+        batch.attributes.aPosition.set(
+          new Float32Array([0, -1, 0, 1, -1, 0, 0, 1, 0, 1, 1, 0, 0, 1, 0, 1, -1, 0]),
+        );
+        batch.attributes.aTint.set(
+          new Float32Array([0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 1]),
+        );
+        batch.draw();
+        stop();
+        resolve();
+      });
+    });
+
+    const [leftRed, , leftBlue] = samplePixel(canvas, 40, 32);
+    const [rightRed, , rightBlue] = samplePixel(canvas, 216, 32);
+    checks.push({
+      name: 'two batches in one frame keep their own attribute data',
+      passed: leftRed! > 150 && leftBlue! < 100 && rightBlue! > 150 && rightRed! < 100,
+      detail: `left rgb(${leftRed},_,${leftBlue}) expected red, right rgb(${rightRed},_,${rightBlue}) expected blue`,
+    });
+  }
+
   window.__GPU_RESULTS__ = { backend: renderer.backend, mode: 'webgpu', checks };
 }
 
diff --git a/scripts/gpu/fixtures/gpu-batch.shader.gen.ts b/scripts/gpu/fixtures/gpu-batch.shader.gen.ts
new file mode 100644
index 0000000..8c63c7e
--- /dev/null
+++ b/scripts/gpu/fixtures/gpu-batch.shader.gen.ts
@@ -0,0 +1,33 @@
+/* Generated by BroMetal. Do not edit — recompile with `npx brometal dev`. */
+import type { CompiledShader } from 'brometal';
+
+const GpuBatch: CompiledShader<{ aPosition: 'vec3'; aTint: 'vec3' }, Record<string, never>, Record<string, never>> = {
+  wgslSrc: `struct BmVSIn {
+  @location(0) aPosition : vec3f,
+  @location(1) aTint : vec3f,
+}
+struct BmVSOut {
+  @builtin(position) bm_position : vec4f,
+  @location(0) vTint : vec3f,
+}
+@vertex
+fn vs_main(bm_in : BmVSIn) -> BmVSOut {
+  var bm_out : BmVSOut;
+  bm_out.vTint = bm_in.aTint;
+  bm_out.bm_position = vec4f(bm_in.aPosition.x, bm_in.aPosition.y, 0.0, 1.0);
+  bm_out.bm_position.z = (bm_out.bm_position.z + bm_out.bm_position.w) * 0.5;
+  return bm_out;
+}
+@fragment
+fn fs_main(bm_in : BmVSOut) -> @location(0) vec4f {
+  return vec4f(bm_in.vTint, 1.0);
+}
+`,
+  attributes: { aPosition: 'vec3', aTint: 'vec3' },
+  instanceAttributes: {},
+  uniforms: {},
+  layout: {"attributes":[{"name":"aPosition","type":"vec3","location":0,"size":3,"divisor":0},{"name":"aTint","type":"vec3","location":1,"size":3,"divisor":0}],"uniforms":[],"uniformBlockSize":0},
+
+};
+
+export default GpuBatch;
diff --git a/scripts/gpu/fixtures/gpu-batch.shader.ts b/scripts/gpu/fixtures/gpu-batch.shader.ts
new file mode 100644
index 0000000..1e5111d
--- /dev/null
+++ b/scripts/gpu/fixtures/gpu-batch.shader.ts
@@ -0,0 +1,20 @@
+import { shader, vec4 } from 'brometal';
+
+/**
+ * Fixture: draws a solid colour taken from a per-vertex attribute, so two
+ * uploads to the same attribute in one frame produce two visibly different
+ * draws — unless the second upload overwrote the first.
+ */
+export const GpuBatch = shader({
+  attributes: { aPosition: 'vec3', aTint: 'vec3' },
+  varyings: { vTint: 'vec3' },
+
+  vertex({ aPosition, aTint }, _uniforms, v) {
+    v.vTint = aTint;
+    return vec4(aPosition.x, aPosition.y, 0, 1);
+  },
+
+  fragment(_uniforms, { vTint }) {
+    return vec4(vTint, 1);
+  },
+});