sync 6fdf6301e2bb

Browse files

Files changed (5) hide show

README.md +5 -3
build/webgpu/manifest.json +3 -7
build/webgpu/metadata.json +6 -6
build/webgpu/unary-scalar.wgsl.jinja +4 -6
build/webgpu/unary-vec4.wgsl.jinja +19 -11

README.md CHANGED Viewed

@@ -39,14 +39,14 @@ See the [ONNX `Not` spec](https://onnx.ai/onnx/operators/onnx__Not.html) for the
 - [`metadata.json`](build/webgpu/metadata.json) — kernel metadata (id, digests, per-variant templates, provenance)
 - [`manifest.json`](build/webgpu/manifest.json) — the op contract (source of truth)
 - [`test.json`](build/webgpu/test.json) — correctness cases
-- [`bench.json`](build/webgpu/bench.json) — benchmark + tuning cases
 - [`unary-scalar.wgsl.jinja`](build/webgpu/unary-scalar.wgsl.jinja)
 - [`unary-vec4.wgsl.jinja`](build/webgpu/unary-vec4.wgsl.jinja)
 ## Use with `@huggingface/kernels`
 ```sh
-npm install --save-exact @huggingface/kernels@0.0.1-preview.2
 ```
 Required output shapes and logical data types are inferred from the supplied inputs and attributes; result tensors are allocated automatically.
@@ -56,9 +56,11 @@ It follows the `v1` branch as fixes land. To pin exact artifact bytes, pass a 40
 Replace each `*Data` placeholder with a typed array containing the corresponding input data.
 ```js
 import { getKernel } from "@huggingface/kernels";
 const kernel = await getKernel("webgpu-kernels/ai.onnx.Not", { version: 1 });
-const { y } = await kernel({ x: { data: xData, shape: [] } });
 ```

 - [`metadata.json`](build/webgpu/metadata.json) — kernel metadata (id, digests, per-variant templates, provenance)
 - [`manifest.json`](build/webgpu/manifest.json) — the op contract (source of truth)
 - [`test.json`](build/webgpu/test.json) — correctness cases
+- [`bench.json`](build/webgpu/bench.json) — benchmark cases
 - [`unary-scalar.wgsl.jinja`](build/webgpu/unary-scalar.wgsl.jinja)
 - [`unary-vec4.wgsl.jinja`](build/webgpu/unary-vec4.wgsl.jinja)
 ## Use with `@huggingface/kernels`
 ```sh
+npm install --save-exact @huggingface/kernels@0.0.1-preview.3
 ```
 Required output shapes and logical data types are inferred from the supplied inputs and attributes; result tensors are allocated automatically.
 Replace each `*Data` placeholder with a typed array containing the corresponding input data.
+A `bool` input takes 0/1 values in a `Uint8Array` together with `dtype: "bool"`, since no typed-array class implies it.
 ```js
 import { getKernel } from "@huggingface/kernels";
 const kernel = await getKernel("webgpu-kernels/ai.onnx.Not", { version: 1 });
+const { y } = await kernel({ x: { data: xData, shape: [2], dtype: "bool" } });
 ```

build/webgpu/manifest.json CHANGED Viewed

@@ -15,16 +15,13 @@
       "id": "same_layout_vec4",
       "priority": 20,
       "when": ["numel(shapes.x) > 0", "numel(shapes.x) % 4 == 0", "wideVec4StorageOk"],
-      "derive": { "scalar": "dtypes.B", "usesF16": false, "vectorScalar": "\"vec4<\" ~ dtypes.B ~ \">\"" },
       "passes": [
         {
           "id": "main",
-          "name": "Not.vec4",
           "shader": "unary-vec4.wgsl.jinja",
-          "derive": {
-            "op": "\"not\"",
-            "vec4PerThread": "4 if numel(shapes.y) * dtypeBytes(tensorDtypes.y) <= 16777216 else 1"
-          },
           "bindings": [
             { "arg": "x", "elementType": "vec4<u32>" },
             { "arg": "y", "elementType": "vec4<u32>" },
@@ -45,7 +42,6 @@
           "id": "main",
           "name": "Not",
           "shader": "unary-scalar.wgsl.jinja",
-          "derive": { "op": "\"not\"" },
           "bindings": [
             { "arg": "x", "elementType": "u32" },
             { "arg": "y", "elementType": "u32" },

       "id": "same_layout_vec4",
       "priority": 20,
       "when": ["numel(shapes.x) > 0", "numel(shapes.x) % 4 == 0", "wideVec4StorageOk"],
+      "derive": { "vectorScalar": "\"vec4<\" ~ dtypes.B ~ \">\"", "scalar": "dtypes.B" },
       "passes": [
         {
           "id": "main",
+          "name": "Not.Vec4",
           "shader": "unary-vec4.wgsl.jinja",
+          "derive": { "vec4PerThread": "4 if numel(shapes.y) * dtypeBytes(tensorDtypes.y) <= 16777216 else 1" },
           "bindings": [
             { "arg": "x", "elementType": "vec4<u32>" },
             { "arg": "y", "elementType": "vec4<u32>" },
           "id": "main",
           "name": "Not",
           "shader": "unary-scalar.wgsl.jinja",
           "bindings": [
             { "arg": "x", "elementType": "u32" },
             { "arg": "y", "elementType": "u32" },

build/webgpu/metadata.json CHANGED Viewed

@@ -1,6 +1,6 @@
 {
   "name": "ai.onnx.Not",
-  "id": "_ai_onnx_not_webgpu_36db185",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
@@ -8,15 +8,15 @@
     "algorithm": "sha256",
     "files": {
       "bench.json": "L8DnqPVQmtoUbVzOJf4ZOBc4B5szJWInfBVo8dLAsjI=",
-      "manifest.json": "XcMLy6Mv7EKh+Xx5cHodb3BrUbRuFMx5wwBDmoC34qQ=",
       "test.json": "C0OWAWpvBPaJ3TPJV+Q7uEeKYDziZXz17iNTn7Bl8+s=",
-      "unary-scalar.wgsl.jinja": "aF5jl3TxM+kRr5JdAJgNXYpsijeufs+DRUN95lLAl50=",
-      "unary-vec4.wgsl.jinja": "UQ/vPY94Jhd4/nX5WCB8jGObg2b9oNcqq4hpdqDXUVM="
     }
   },
-  "provenance": { "kernel": { "sha": "91d990483a174128daf7673f3f37a7c890493ae1", "dirty": false } },
   "webgpu": {
-    "manifestSpec": "2.0",
     "variants": { "same_layout_vec4": ["unary-vec4.wgsl.jinja"], "elementwise": ["unary-scalar.wgsl.jinja"] }
   }
 }

 {
   "name": "ai.onnx.Not",
+  "id": "_ai_onnx_not_webgpu_6e19d7e",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
     "algorithm": "sha256",
     "files": {
       "bench.json": "L8DnqPVQmtoUbVzOJf4ZOBc4B5szJWInfBVo8dLAsjI=",
+      "manifest.json": "BCYOzxneNalRr+4OjUZTMTByUEqUz/iK1HylpIyPjmE=",
       "test.json": "C0OWAWpvBPaJ3TPJV+Q7uEeKYDziZXz17iNTn7Bl8+s=",
+      "unary-scalar.wgsl.jinja": "ZdY8qjNQvzHtQc/mLDR5TsKg/J35QwBfwnkZoOLJDwc=",
+      "unary-vec4.wgsl.jinja": "ajrPgCRHsVACLo9RyHJHftcz65FhdgUch0tjkcQwA9c="
     }
   },
+  "provenance": { "kernel": { "sha": "6fdf6301e2bbcc2f03bf1eaf493b7ad55ef33afc", "dirty": false } },
   "webgpu": {
+    "manifestSpec": "2.1",
     "variants": { "same_layout_vec4": ["unary-vec4.wgsl.jinja"], "elementwise": ["unary-scalar.wgsl.jinja"] }
   }
 }

build/webgpu/unary-scalar.wgsl.jinja CHANGED Viewed

@@ -12,14 +12,12 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
   for (var i = begin; i < end; i = i + 1u) {
 {%- endmacro %}
 {% macro flat_tail_close() %}
-  }
-{% endmacro %}
-// Scalar unary elementwise implementation. Specialization emits only the
-// selected operation and any numerical helper it requires.
 {{ env.wgsl.resourceDeclarations }}
 {{ flat_tail_open() }}
   y[i] = select(0u, 1u, x[i] == 0u);
-{{ flat_tail_close() -}}
 }

   for (var i = begin; i < end; i = i + 1u) {
 {%- endmacro %}
 {% macro flat_tail_close() %}
+  }{% endmacro %}
+// Scalar unary elementwise kernel: applies the selected operation, with the
+// numerical helpers it uses, to each element.
 {{ env.wgsl.resourceDeclarations }}
 {{ flat_tail_open() }}
   y[i] = select(0u, 1u, x[i] == 0u);
+{{ flat_tail_close() }}
 }

build/webgpu/unary-vec4.wgsl.jinja CHANGED Viewed

@@ -2,34 +2,42 @@
 // component.
 {{ env.wgsl.resourceDeclarations }}
-{% set vec4PerThread = vec4PerThread %}
 {% if vec4PerThread > 1 %}
 const ITEMS: u32 = {{ vec4PerThread }}u;
-{% endif %}
 @compute @workgroup_size({{ tunables.WORKGROUP_SIZE }})
 fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-  // 2D-folded flat index: gid.y carries the high bits when the element count exceeds the
-  // per-axis dispatch fold width (the dispatch caps x and spills the rest into y).
 {% if vec4PerThread > 1 %}
   // Each invocation walks ITEMS vec4 groups a span apart. Consecutive lanes
   // access consecutive words on every step, while each lane can keep several
   // independent loads in flight.
-  let tid = gid.x + gid.y * {{ DISPATCH_FOLD_WIDTH }}u * {{ tunables.WORKGROUP_SIZE }}u;
   let span = (params.count + ITEMS - 1u) / ITEMS;
   for (var j = 0u; j < ITEMS; j = j + 1u) {
     let i = tid + j * span;
     if (i >= params.count) {
       break;
     }
 {% else %}
-  let i = gid.x + gid.y * {{ DISPATCH_FOLD_WIDTH }}u * {{ tunables.WORKGROUP_SIZE }}u;
-  if (i >= params.count) {
-    return;
-  }
 {% endif %}
   let xv = x[i];
   y[i] = select(vec4<u32>(0u), vec4<u32>(1u), xv == vec4<u32>(0u));
 {% if vec4PerThread > 1 %}

 // component.
 {{ env.wgsl.resourceDeclarations }}
+{% macro flat_index_2d(workgroupSize, name="i", bound="params.count", guardInline=false) %}
+{% set wgTerm = workgroupSize ~ "u" if workgroupSize is number else workgroupSize %}
+  // 2D-folded flat index: gid.y carries the high bits past the dispatch's
+  // per-axis workgroup fold width.
+{% if bound == "" %}
+  let {{ name }} = gid.x + gid.y * {{ DISPATCH_FOLD_WIDTH }}u * {{ wgTerm }};{% else %}
+  let {{ name }} = gid.x + gid.y * {{ DISPATCH_FOLD_WIDTH }}u * {{ wgTerm }};
+  if ({{ name }} >= {{ bound }}) {
+    return;
+  }{% endif %}{% endmacro %}
 {% if vec4PerThread > 1 %}
 const ITEMS: u32 = {{ vec4PerThread }}u;
+{% endif %}
 @compute @workgroup_size({{ tunables.WORKGROUP_SIZE }})
 fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
 {% if vec4PerThread > 1 %}
   // Each invocation walks ITEMS vec4 groups a span apart. Consecutive lanes
   // access consecutive words on every step, while each lane can keep several
   // independent loads in flight.
+{{ flat_index_2d(tunables.WORKGROUP_SIZE, "tid", "") }}
   let span = (params.count + ITEMS - 1u) / ITEMS;
+  // Lanes at or past span exist only because the dispatch rounds up to whole
+  // workgroups. Lane span + k would start on lane k's second group and rewrite
+  // up to ITEMS - 1 groups another lane already stored.
+  if (tid >= span) {
+    return;
+  }
   for (var j = 0u; j < ITEMS; j = j + 1u) {
     let i = tid + j * span;
     if (i >= params.count) {
       break;
     }
 {% else %}
+{{ flat_index_2d(tunables.WORKGROUP_SIZE) }}
 {% endif %}
   let xv = x[i];
   y[i] = select(vec4<u32>(0u), vec4<u32>(1u), xv == vec4<u32>(0u));
 {% if vec4PerThread > 1 %}