@driftengine/core 4.9.0 → 4.11.0

package/dist/render/backend/webgpu/materialStore.js

@@ -0,0 +1,265 @@
+import { DYNAMIC_ALIGNMENT, MAX_RING_BYTES } from './uniformRing.js';
+/**
+ * The lit stage's material blocks, each kept in its slot for as long as later draws ask for the
+ * same numbers, so a material that has not changed since the last frame costs no upload at all.
+ *
+ * **Keyed by content, because nothing else names a material.** A material here is whatever the
+ * setters left in the block when a draw took it — a consumer never hands over an object that could
+ * carry an identity — so the block's own words are the key: hashed to find a bucket, compared word
+ * for word to accept a match, and bound by the slot's offset. What it gives up is a hash and a
+ * compare of a few hundred bytes per material change, against copying and uploading the whole
+ * block every frame, which a stage of four hundred materials measured at megabytes a frame.
+ *
+ * **A slot the frame has asked for is never given away within it**, for the reason `UniformRing`
+ * exists: a queue write does not interleave with recorded draws, so rewriting a slot a recorded draw
+ * reads would hand that draw the new numbers. A frame asking for more distinct blocks than there are
+ * slots is refused, counted and skipped, as the ring was. Across frames rewriting is safe, because a
+ * write issued after a submit is ordered after the commands in it.
+ *
+ * **What a new block displaces is a slot nobody asked for this frame or the last**, found by a hand
+ * that sweeps the slots in turn; only when every slot was asked for recently does it take one the
+ * last frame used. What would make that wrong is a scene whose distinct materials a frame approach
+ * the store's size: it then rewrites some of its steady materials every frame, which costs their
+ * upload and nothing else.
+ */
+export class MaterialStore {
+    device;
+    usage;
+    label;
+    budget;
+    /** Replaced by `growTo`, so every bind group over it is rebuilt when that says so. */
+    buffer;
+    /** Bytes per slot, rounded up to the alignment a dynamic offset requires. */
+    slotSize;
+    words;
+    stride;
+    capacity;
+    staging;
+    ints;
+    /** Per slot: its block's hash, the next slot in its bucket or -1, and whether it holds one. */
+    hashes;
+    next;
+    held;
+    /** Per slot: the frame that last asked for it, and whether it was written since the last flush. */
+    askedIn;
+    pending;
+    /** Per slot: how many holders pinned it. A pinned slot is never displaced. */
+    pins;
+    /**
+     * How many times the store has forgotten everything it held, which growth does: a holder of an
+     * offset compares this, because an offset from before is a slot of a buffer that is gone.
+     */
+    epoch = 0;
+    /** Per bucket: its first slot, or -1. A power of two at least twice the slots. */
+    heads;
+    mask;
+    lowPending;
+    highPending = -1;
+    hand = 0;
+    /** Starts past zero, so a slot never asked for reads as long unasked. */
+    frame = 2;
+    constructor(device, bytesPerBlock, slots, usage, label = 'materials.store', budget = null) {
+        this.device = device;
+        this.usage = usage;
+        this.label = label;
+        this.budget = budget;
+        this.words = Math.ceil(bytesPerBlock / 4);
+        this.slotSize = Math.ceil(bytesPerBlock / DYNAMIC_ALIGNMENT) * DYNAMIC_ALIGNMENT;
+        this.stride = this.slotSize / 4;
+        this.capacity = slots;
+        this.lowPending = slots;
+        this.staging = new ArrayBuffer(this.slotSize * slots);
+        this.ints = new Int32Array(this.staging);
+        this.hashes = new Int32Array(slots);
+        this.next = new Int32Array(slots);
+        this.held = new Uint8Array(slots);
+        this.askedIn = new Int32Array(slots);
+        this.pending = new Uint8Array(slots);
+        this.pins = new Int32Array(slots);
+        this.heads = new Int32Array(bucketsFor(slots)).fill(-1);
+        this.mask = this.heads.length - 1;
+        this.buffer = device.createBuffer({ label, size: this.staging.byteLength, usage });
+    }
+    /** How many distinct blocks a frame may hold. */
+    get slots() {
+        return this.capacity;
+    }
+    /** Start a frame: every slot the last one asked for may be displaced from here on, last. */
+    beginFrame() {
+        this.frame += 1;
+    }
+    /**
+     * The byte offset of a slot holding exactly `block`, written there if no slot did, or null when
+     * every slot is one this frame has asked for. `block` is the whole block, word for word.
+     */
+    place(block) {
+        this.budget?.ask();
+        const hash = hashWords(block, this.words);
+        const bucket = hash & this.mask;
+        for (let slot = this.heads[bucket]; slot >= 0; slot = this.next[slot]) {
+            if (this.hashes[slot] === hash && this.holds(slot, block)) {
+                this.askedIn[slot] = this.frame;
+                return slot * this.slotSize;
+            }
+        }
+        const slot = this.displace();
+        if (slot < 0) {
+            this.budget?.drop();
+            return null;
+        }
+        if (this.held[slot] === 1)
+            this.unlink(slot);
+        this.ints.set(block, slot * this.stride);
+        this.hashes[slot] = hash;
+        this.next[slot] = this.heads[bucket];
+        this.heads[bucket] = slot;
+        this.held[slot] = 1;
+        this.askedIn[slot] = this.frame;
+        this.pending[slot] = 1;
+        if (slot < this.lowPending)
+            this.lowPending = slot;
+        if (slot > this.highPending)
+            this.highPending = slot;
+        return slot * this.slotSize;
+    }
+    /** Upload every slot written since the last flush, one write per run of neighbouring slots. */
+    flush() {
+        let slot = this.lowPending;
+        while (slot <= this.highPending) {
+            if (this.pending[slot] !== 1) {
+                slot += 1;
+                continue;
+            }
+            const from = slot;
+            while (slot <= this.highPending && this.pending[slot] === 1) {
+                this.pending[slot] = 0;
+                slot += 1;
+            }
+            const at = from * this.slotSize;
+            this.device.queue.writeBuffer(this.buffer, at, this.staging, at, (slot - from) * this.slotSize);
+        }
+        this.lowPending = this.capacity;
+        this.highPending = -1;
+    }
+    /**
+     * Make room for at least `slots`, forgetting every block held, and say whether the buffer was
+     * replaced. Called where `UniformRing.growTo` is, at a frame's start, for the reasons it gives,
+     * and refused past the same ceiling.
+     */
+    growTo(slots) {
+        if (slots <= this.capacity)
+            return false;
+        if (this.slotSize * slots > MAX_RING_BYTES)
+            return false;
+        this.buffer.destroy();
+        this.capacity = slots;
+        this.staging = new ArrayBuffer(this.slotSize * slots);
+        this.ints = new Int32Array(this.staging);
+        this.hashes = new Int32Array(slots);
+        this.next = new Int32Array(slots);
+        this.held = new Uint8Array(slots);
+        this.askedIn = new Int32Array(slots);
+        this.pending = new Uint8Array(slots);
+        this.pins = new Int32Array(slots);
+        this.heads = new Int32Array(bucketsFor(slots)).fill(-1);
+        this.mask = this.heads.length - 1;
+        this.lowPending = slots;
+        this.epoch += 1;
+        this.highPending = -1;
+        this.hand = 0;
+        this.buffer = this.device.createBuffer({
+            label: this.label,
+            size: this.staging.byteLength,
+            usage: this.usage,
+        });
+        return true;
+    }
+    /**
+     * Keep the block at `offset` where it is until `unpin`: a holder binding it by offset frame after
+     * frame — a list recorded once — whether or not anything asks for it. Counted, so two holders of
+     * one block each let go of their own pin.
+     */
+    pin(offset) {
+        const slot = offset / this.slotSize;
+        this.pins[slot] = this.pins[slot] + 1;
+    }
+    unpin(offset) {
+        const slot = offset / this.slotSize;
+        if (this.pins[slot] > 0)
+            this.pins[slot] = this.pins[slot] - 1;
+    }
+    dispose() {
+        this.buffer.destroy();
+    }
+    holds(slot, block) {
+        const base = slot * this.stride;
+        for (let word = 0; word < this.words; word++) {
+            if (this.ints[base + word] !== block[word])
+                return false;
+        }
+        return true;
+    }
+    /**
+     * The slot a new block goes in: the first from the hand nobody asked for this frame or the last,
+     * else the first this frame has not asked for, else none. One sweep at most.
+     */
+    displace() {
+        let fallback = -1;
+        for (let step = 0; step < this.capacity; step++) {
+            const slot = this.hand;
+            this.hand = slot + 1 === this.capacity ? 0 : slot + 1;
+            if (this.pins[slot] > 0)
+                continue;
+            const asked = this.askedIn[slot];
+            if (asked < this.frame - 1)
+                return slot;
+            if (fallback < 0 && asked !== this.frame)
+                fallback = slot;
+        }
+        return fallback;
+    }
+    unlink(slot) {
+        const bucket = this.hashes[slot] & this.mask;
+        let at = this.heads[bucket];
+        if (at === slot) {
+            this.heads[bucket] = this.next[slot];
+            return;
+        }
+        while (at >= 0) {
+            const after = this.next[at];
+            if (after === slot) {
+                this.next[at] = this.next[slot];
+                return;
+            }
+            at = after;
+        }
+    }
+}
+function bucketsFor(slots) {
+    let buckets = 16;
+    while (buckets < slots * 2)
+        buckets *= 2;
+    return buckets;
+}
+/**
+ * MurmurHash3's 32-bit mix over the block's words. Not FNV, whose multiply carries a difference only
+ * upward: the bucket is the hash's low bits, and a material's words are mostly floats, whose low
+ * mantissa bits are zero for every value a person types — 0.5, 1, 2 — so blocks differing only in
+ * those would all have shared one bucket. The compare that follows a match is what decides.
+ */
+export function hashWords(block, words) {
+    let hash = 0x9747b28c;
+    for (let word = 0; word < words; word++) {
+        let k = Math.imul(block[word], 0xcc9e2d51);
+        k = (k << 15) | (k >>> 17);
+        hash ^= Math.imul(k, 0x1b873593);
+        hash = (hash << 13) | (hash >>> 19);
+        hash = (Math.imul(hash, 5) + 0xe6546b64) | 0;
+    }
+    hash ^= words * 4;
+    hash ^= hash >>> 16;
+    hash = Math.imul(hash, 0x85ebca6b);
+    hash ^= hash >>> 13;
+    hash = Math.imul(hash, 0xc2b2ae35);
+    return hash ^ (hash >>> 16);
+}

package/dist/render/backend/webgpu/motionPass.js

@@ -197,12 +197,17 @@ export class MotionPass {
         const eight = offsets.joints2 >= 0 && offsets.weights2 >= 0;
         /*
          * Five byte counts, each under 256 because a vertex is at most 172 bytes, as five base-256
-         * digits: exact below 2^40, so two layouts can never share a key.
+         * digits, and two bits for the weights' formats: exact below 2^42, so two layouts can never
+         * share a key.
          */
-        const key = (((stride * 256 + offsets.joints) * 256 + offsets.weights) * 256 + offsets.joints2 + 1) *
+        const key = ((((stride * 256 + offsets.joints) * 256 + offsets.weights) * 256 + offsets.joints2 + 1) *
             256 +
             offsets.weights2 +
-            1;
+            1) *
+            4 +
+            /* And whether each four weights are floats or fixed point, two bits: `vertexPacking.ts`. */
+            (offsets.weightsFormat === 'float32x4' ? 0 : 2) +
+            (offsets.weights2Format === 'float32x4' ? 0 : 1);
         const pipelines = cloth ? this.clothPipelines : this.skinnedPipelines;
         const held = pipelines.get(key);
         if (held !== undefined)
@@ -224,11 +229,15 @@ export class MotionPass {
                         attributes: [
                             { shaderLocation: 0, offset: 0, format: 'float32x3' },
                             { shaderLocation: 11, offset: offsets.joints, format: 'float32x4' },
-                            { shaderLocation: 12, offset: offsets.weights, format: 'float32x4' },
+                            { shaderLocation: 12, offset: offsets.weights, format: offsets.weightsFormat },
                             ...(eight
                                 ? [
                                     { shaderLocation: 14, offset: offsets.joints2, format: 'float32x4' },
-                                    { shaderLocation: 15, offset: offsets.weights2, format: 'float32x4' },
+                                    {
+                                        shaderLocation: 15,
+                                        offset: offsets.weights2,
+                                        format: offsets.weights2Format,
+                                    },
                                 ]
                                 : []),
                         ],
@@ -570,7 +579,7 @@ export class MotionPass {
                 }
                 pass.setBindGroup(0, this.group, [i * MOTION_DRAW_STRIDE]);
                 pass.setVertexBuffer(0, mesh.vertexBuffers[0]);
-                pass.setIndexBuffer(mesh.indexBuffer, 'uint32');
+                pass.setIndexBuffer(mesh.indexBuffer, mesh.indexFormat);
                 pass.drawIndexed(mesh.indexCount, instances);
             }
         }

package/dist/render/backend/webgpu/pipelineCache.d.ts

@@ -52,6 +52,12 @@ export declare class PipelineCache {
     readonly glassShadows: boolean;
     /** How to describe each lit pipeline again, so `enable` can rebuild every one. */
     private readonly litDescribers;
+    /**
+     * Bumped when a switch's rebuilt set lands, which replaces pipelines already handed out: a holder
+     * of one — a recorded render bundle — compares this and records again rather than drawing on
+     * with the pipelines from before the switch.
+     */
+    version: number;
     /** Bumped by every switch, so a lit build started before one is not kept after it. */
     private litGeneration;
     /** The last rebuild, which the next one waits for so they land in the order they were asked. */
@@ -73,6 +79,15 @@ export declare class PipelineCache {
      * wait for readiness rather than discovering it mid-frame.
      */
     private readonly compiling;
+    /** Keys whose compile off the main thread failed, which `lit` then builds where asked for. */
+    private readonly failed;
+    /**
+     * Whether a lit pipeline first asked for inside a frame is compiled off the main thread while
+     * its draw is left out — `RenderQuality.pipelineCompile` `'skip'` — rather than built where it is
+     * asked for, which holds the GPU process for the compile. The renderer turns it off while it
+     * records static draws, which must hold every draw they were given.
+     */
+    skipsCompiling: boolean;
     constructor(device: GPUDevice, format: GPUTextureFormat, sampleCount?: number, glassShadows?: boolean, 
     /** The profile's `clusteredLights`, which compiles the clustered arm in or leaves it out. */
     clusteredLights?: boolean);
@@ -90,7 +105,8 @@ export declare class PipelineCache {
     /** Whether this is a lit pipeline, which `enable` rebuilds. See `litSwitches`. */
     lit?: boolean): Promise<GPURenderPipeline>;
     /**
-     * Resolve once every pipeline asked for so far has finished compiling.
+     * Resolve once every pipeline asked for so far has finished compiling, a switch's rebuild
+     * among them.
      *
      * A renderer awaits this before its first frame, which is what moves the compile out of the
      * queue and in front of it. A failure is not rethrown: a pipeline that cannot be built will
@@ -112,6 +128,17 @@ export declare class PipelineCache {
     get(key: string, describe: () => GPURenderPipelineDescriptor, 
     /** Whether this is a lit pipeline, which `enable` rebuilds. See `litSwitches`. */
     lit?: boolean): GPURenderPipeline;
+    /**
+     * A lit pipeline for a draw inside a frame: the one built, or one built here — or, where draws
+     * skip (`skipsCompiling`), none, with its compile started off the main thread for a later frame to
+     * find. A draw handed none is left out.
+     *
+     * **What skipping gives up** is the draw for the frames its compile takes, a few to a few dozen on
+     * a phone, against a frame held for the compile where it is asked for: 130 to 400 ms, reported on
+     * a phone for a material first drawn mid-scene. A compile that fails is built here the next time,
+     * so the device reports it rather than the draw vanishing for good.
+     */
+    lit(key: string, describe: () => GPURenderPipelineDescriptor): GPURenderPipeline | null;
     /**
      * Turn one of the lit stage's switches on, for every lit pipeline built from now on and every one
      * already built. See `litSwitches`.

package/dist/render/backend/webgpu/pipelineCache.js

@@ -51,6 +51,12 @@ export class PipelineCache {
     glassShadows;
     /** How to describe each lit pipeline again, so `enable` can rebuild every one. */
     litDescribers = new Map();
+    /**
+     * Bumped when a switch's rebuilt set lands, which replaces pipelines already handed out: a holder
+     * of one — a recorded render bundle — compares this and records again rather than drawing on
+     * with the pipelines from before the switch.
+     */
+    version = 0;
     /** Bumped by every switch, so a lit build started before one is not kept after it. */
     litGeneration = 0;
     /** The last rebuild, which the next one waits for so they land in the order they were asked. */
@@ -72,6 +78,15 @@ export class PipelineCache {
      * wait for readiness rather than discovering it mid-frame.
      */
     compiling = new Map();
+    /** Keys whose compile off the main thread failed, which `lit` then builds where asked for. */
+    failed = new Set();
+    /**
+     * Whether a lit pipeline first asked for inside a frame is compiled off the main thread while
+     * its draw is left out — `RenderQuality.pipelineCompile` `'skip'` — rather than built where it is
+     * asked for, which holds the GPU process for the compile. The renderer turns it off while it
+     * records static draws, which must hold every draw they were given.
+     */
+    skipsCompiling = false;
     constructor(device, format, sampleCount = 1, glassShadows = true, 
     /** The profile's `clusteredLights`, which compiles the clustered arm in or leaves it out. */
     clusteredLights = false) {
@@ -121,12 +136,19 @@ export class PipelineCache {
                 this.pipelines.set(key, built);
             this.compiling.delete(key);
             return built;
+        }, 
+        /* Out of the list, or `ready` waits on it for ever; `lit` builds it where asked next. */
+        (error) => {
+            this.compiling.delete(key);
+            this.failed.add(key);
+            throw error;
         });
         this.compiling.set(key, promise);
         return promise;
     }
     /**
-     * Resolve once every pipeline asked for so far has finished compiling.
+     * Resolve once every pipeline asked for so far has finished compiling, a switch's rebuild
+     * among them.
      *
      * A renderer awaits this before its first frame, which is what moves the compile out of the
      * queue and in front of it. A failure is not rethrown: a pipeline that cannot be built will
@@ -137,6 +159,8 @@ export class PipelineCache {
         while (this.compiling.size > 0) {
             await Promise.allSettled([...this.compiling.values()]);
         }
+        /* And a switch's rebuild, whose set replaces the pipelines in one swap: see `enable`. */
+        await this.litRebuild;
     }
     /** The pipeline for this key if it is already built, without building one. */
     peek(key) {
@@ -165,6 +189,26 @@ export class PipelineCache {
         this.pipelines.set(key, built);
         return built;
     }
+    /**
+     * A lit pipeline for a draw inside a frame: the one built, or one built here — or, where draws
+     * skip (`skipsCompiling`), none, with its compile started off the main thread for a later frame to
+     * find. A draw handed none is left out.
+     *
+     * **What skipping gives up** is the draw for the frames its compile takes, a few to a few dozen on
+     * a phone, against a frame held for the compile where it is asked for: 130 to 400 ms, reported on
+     * a phone for a material first drawn mid-scene. A compile that fails is built here the next time,
+     * so the device reports it rather than the draw vanishing for good.
+     */
+    lit(key, describe) {
+        const existing = this.pipelines.get(key);
+        if (existing !== undefined)
+            return existing;
+        if (!this.skipsCompiling || this.failed.has(key))
+            return this.get(key, describe, true);
+        if (!this.compiling.has(key))
+            void this.getAsync(key, describe, true).catch(() => { });
+        return null;
+    }
     /**
      * Turn one of the lit stage's switches on, for every lit pipeline built from now on and every one
      * already built. See `litSwitches`.
@@ -201,6 +245,7 @@ export class PipelineCache {
                 if (result.status === 'fulfilled')
                     this.pipelines.set(result.value[0], result.value[1]);
             }
+            this.version += 1;
         });
         return this.litRebuild;
     }

package/dist/render/backend/webgpu/postPass.d.ts

@@ -75,7 +75,7 @@ export declare const AO_FRAG_SIZE: 96;
 export declare const AO_BLUR_FIELDS: UniformFields;
 export declare const AO_BLUR_SIZE: 48;
 export declare const MEDIUM_FRAG_FIELDS: UniformFields;
-export declare const MEDIUM_FRAG_SIZE: 240;
+export declare const MEDIUM_FRAG_SIZE: 304;
 export declare const MEDIUM_UPSAMPLE_FIELDS: UniformFields;
 export declare const MEDIUM_UPSAMPLE_SIZE: 32;
 export declare const BLOOM_PREFILTER_FIELDS: UniformFields;