sync 91d990483a17

Browse files

Files changed (7) hide show

README.md +11 -7
build/webgpu/bench.json +0 -1
build/webgpu/manifest.json +31 -56
build/webgpu/metadata.json +11 -8
build/webgpu/test.json +0 -1
build/webgpu/unary-scalar.wgsl.jinja +20 -10
build/webgpu/unary-vec4.wgsl.jinja +27 -8

README.md CHANGED Viewed

@@ -18,15 +18,15 @@ See the [ONNX `Not` spec](https://onnx.ai/onnx/operators/onnx__Not.html) for the
 ## Inputs
-| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
 | --- | --- | --- | --- | --- | --- | --- |
-| `X` | `x` | `B` | — | — | Input boolean tensor. | required |
 ## Outputs
-| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
 | --- | --- | --- | --- | --- | --- | --- |
-| `Y` | `y` | `B` | same as `X` | same as `X` | Output boolean tensor with each element logically negated. | required |
 ## Type constraints
@@ -36,7 +36,7 @@ See the [ONNX `Not` spec](https://onnx.ai/onnx/operators/onnx__Not.html) for the
 ## Files
-- [`metadata.json`](build/webgpu/metadata.json) — kernel metadata (id, digests, 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
@@ -45,10 +45,14 @@ See the [ONNX `Not` spec](https://onnx.ai/onnx/operators/onnx__Not.html) for the
 ## Use with `@huggingface/kernels`
-The loader derives every required output's shape and logical dtype from the manifest contract and this call.
-It then allocates the result tensors automatically.
 The `version: 1` option selects the published kernel contract; it is independent of any operator opset, contrib `since_version`, or model version.
 Replace each `*Data` placeholder with a typed array containing the corresponding input data.

 ## Inputs
+| Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence |
 | --- | --- | --- | --- | --- | --- | --- |
+| `x` | `X` | `B` | — | — | Input boolean tensor. | required |
 ## Outputs
+| Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence |
 | --- | --- | --- | --- | --- | --- | --- |
+| `y` | `Y` | `B` | same as `x` | same as `x` | Output boolean tensor with each element logically negated. | required |
 ## Type constraints
 ## Files
+- [`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
 ## 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.
 The `version: 1` option selects the published kernel contract; it is independent of any operator opset, contrib `since_version`, or model version.
+It follows the `v1` branch as fixes land. To pin exact artifact bytes, pass a 40-character commit `revision` instead of `version`.
 Replace each `*Data` placeholder with a typed array containing the corresponding input data.

build/webgpu/bench.json CHANGED Viewed

@@ -1,5 +1,4 @@
 {
-  "op": "ai.onnx.Not",
   "cases": [
     {
       "name": "1m_bool",

 {
   "cases": [
     {
       "name": "1m_bool",

build/webgpu/manifest.json CHANGED Viewed

@@ -2,85 +2,60 @@
   "domain": "ai.onnx",
   "name": "Not",
   "sinceVersion": 1,
-  "description": "Computes the elementwise logical NOT of a boolean tensor. Each output element is `true` where the corresponding input element is `false`, and vice versa.",
-  "inputs": [{ "role": "X", "dtype": "B", "description": "Input boolean tensor." }],
-  "outputs": [
-    {
-      "role": "Y",
-      "dtype": "B",
-      "rank": "ranks.X",
-      "description": "Output boolean tensor with each element logically negated.",
-      "shape": "shapes.X"
-    }
-  ],
   "typeConstraints": { "B": ["bool"] },
-  "args": {
-    "x": { "kind": "tensor", "semantic": "X", "role": "input" },
-    "y": { "kind": "tensor", "semantic": "Y", "role": "output" }
-  },
-  "tunables": { "WORKGROUP_SIZE": 256 },
   "derive": {
     "wideVec4StorageOk": "device.features.has(\"subgroups\") or not (has(device.adapterInfo, \"subgroupMinSize\") and has(device.adapterInfo, \"subgroupMaxSize\") and device.adapterInfo.subgroupMinSize == 8 and device.adapterInfo.subgroupMaxSize <= 32)"
   },
   "variants": [
     {
       "id": "same_layout_vec4",
-      "when": ["numel(shapes.X) > 0", "numel(shapes.X) % 4 == 0", "numel(shapes.X) == numel(shapes.Y)", "wideVec4StorageOk"],
       "passes": [
         {
           "id": "main",
           "name": "Not.vec4",
-          "source": { "shader": "unary-vec4.wgsl.jinja", "inputs": { "op": "\"not\"" } },
           "bindings": [
-            {
-              "name": "x",
-              "arg": "x",
-              "semantic": "X",
-              "buffer": { "type": "read-only-storage" },
-              "elementType": "vec4<u32>"
-            },
-            { "name": "y", "arg": "y", "semantic": "Y", "buffer": { "type": "storage" }, "elementType": "vec4<u32>" },
-            {
-              "name": "params",
-              "semantic": "kernel.params",
-              "buffer": { "type": "uniform" },
-              "struct": {
-                "name": "Params",
-                "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.Y) / 4" }]
-              }
-            }
           ],
-          "dispatch": { "threads": "numel(shapes.Y) / 4", "workgroupSize": "tunables.WORKGROUP_SIZE" }
         }
-      ],
-      "priority": 20,
-      "constants": { "scalar": "dtypes.B", "usesF16": false, "vectorScalar": "\"vec4<\" ~ dtypes.B ~ \">\"" }
     },
     {
       "id": "elementwise",
-      "when": "numel(shapes.X) == numel(shapes.Y)",
       "passes": [
         {
           "id": "main",
           "name": "Not",
-          "source": { "shader": "unary-scalar.wgsl.jinja", "inputs": { "op": "\"not\"" } },
           "bindings": [
-            {
-              "name": "x",
-              "arg": "x",
-              "semantic": "X",
-              "buffer": { "type": "read-only-storage" },
-              "elementType": "u32"
-            },
-            { "name": "y", "arg": "y", "semantic": "Y", "buffer": { "type": "storage" }, "elementType": "u32" },
-            {
-              "name": "params",
-              "semantic": "kernel.params",
-              "buffer": { "type": "uniform" },
-              "struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.Y)" }] }
-            }
           ],
-          "dispatch": { "threads": "numel(shapes.Y)", "workgroupSize": "tunables.WORKGROUP_SIZE" }
         }
       ]
     }

   "domain": "ai.onnx",
   "name": "Not",
   "sinceVersion": 1,
+  "inputs": { "x": { "onnx": "X", "dtype": "B" } },
+  "outputs": { "y": { "onnx": "Y", "dtype": "B", "rank": "ranks.x", "shape": "shapes.x" } },
   "typeConstraints": { "B": ["bool"] },
+  "tunables": { "WORKGROUP_SIZE": { "default": 256 } },
   "derive": {
     "wideVec4StorageOk": "device.features.has(\"subgroups\") or not (has(device.adapterInfo, \"subgroupMinSize\") and has(device.adapterInfo, \"subgroupMaxSize\") and device.adapterInfo.subgroupMinSize == 8 and device.adapterInfo.subgroupMaxSize <= 32)"
   },
+  "when": ["numel(shapes.x) == numel(shapes.y)"],
   "variants": [
     {
       "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>" },
+            { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.y) / 4" }] }
           ],
+          "dispatch": {
+            "x": "min(ceilDiv((ceilDiv(numel(shapes.y) / 4, 4 if numel(shapes.y) * dtypeBytes(tensorDtypes.y) <= 16777216 else 1)), (tunables.WORKGROUP_SIZE)), 65535)",
+            "y": "ceilDiv(ceilDiv((ceilDiv(numel(shapes.y) / 4, 4 if numel(shapes.y) * dtypeBytes(tensorDtypes.y) <= 16777216 else 1)), (tunables.WORKGROUP_SIZE)), 65535)",
+            "z": 1
+          }
         }
+      ]
     },
     {
       "id": "elementwise",
       "passes": [
         {
           "id": "main",
           "name": "Not",
+          "shader": "unary-scalar.wgsl.jinja",
+          "derive": { "op": "\"not\"" },
           "bindings": [
+            { "arg": "x", "elementType": "u32" },
+            { "arg": "y", "elementType": "u32" },
+            { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.y)" }] }
           ],
+          "dispatch": {
+            "x": "min(ceilDiv((ceilDiv(numel(shapes.y), 4)), (tunables.WORKGROUP_SIZE)), 65535)",
+            "y": "ceilDiv(ceilDiv((ceilDiv(numel(shapes.y), 4)), (tunables.WORKGROUP_SIZE)), 65535)",
+            "z": 1
+          }
         }
       ]
     }

build/webgpu/metadata.json CHANGED Viewed

@@ -1,19 +1,22 @@
 {
   "name": "ai.onnx.Not",
-  "id": "_ai_onnx_not_webgpu_57bebe0",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
   "digest": {
     "algorithm": "sha256",
     "files": {
-      "bench.json": "WHSdkLkXazF/f8HKIJ1KSyIV89eguvAi7ac99EgCUW4=",
-      "manifest.json": "OtseGXk5ZdhyPpPY4Kzp6vaC6TJkQtX4HnbcOMZz5o8=",
-      "test.json": "0pg588+zzH/qMBLQH3y4kuOCNObQPNAswB/BJASd2gA=",
-      "unary-scalar.wgsl.jinja": "zhTCLhU7z3PVyvPl4rMfDCv55JFjr8wNm4OvsfKqUGQ=",
-      "unary-vec4.wgsl.jinja": "xgxipqkd61CfxU63Rjg+Cn6ISilONZmzX8Stwl1Os7M="
     }
   },
-  "provenance": { "kernel": { "sha": "2e7068faf55e7f43df740015f6d1ee49391a41c5", "dirty": false } },
-  "webgpu": { "manifestSpec": "1.0", "specialized": true, "opPath": "ops/ai.onnx.Not" }
 }

 {
   "name": "ai.onnx.Not",
+  "id": "_ai_onnx_not_webgpu_36db185",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
   "digest": {
     "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"] }
+  }
 }

build/webgpu/test.json CHANGED Viewed

@@ -1,5 +1,4 @@
 {
-  "op": "ai.onnx.Not",
   "cases": [
     {
       "name": "bool_vector_basic",

 {
   "cases": [
     {
       "name": "bool_vector_basic",

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

@@ -1,15 +1,25 @@
-// Scalar unary fallback. Each branch retains the operation's numeric hardening,
-// including Payne-Hanek trigonometric range reduction and NaN/overflow guards.
-{{ env.wgsl.resourceDeclarations }}
 @compute @workgroup_size({{ tunables.WORKGROUP_SIZE }})
-fn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups) nwg: vec3<u32>) {
   // 2D-folded flat index: gid.y carries the high bits when the element count exceeds the
-  // maxComputeWorkgroupsPerDimension limit (the dispatch caps x and spills the rest into y).
-  let i = gid.x + gid.y * nwg.x * {{ tunables.WORKGROUP_SIZE }}u;
-  if (i >= params.count) {
-    return;
   }
   y[i] = select(0u, 1u, x[i] == 0u);
 }

+{% macro flat_tail_open() %}
 @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
+  // dispatch's per-axis workgroup fold width (the dispatch caps x and spills the rest into y).
+  let invocation = gid.x + gid.y * {{ DISPATCH_FOLD_WIDTH }}u * {{ tunables.WORKGROUP_SIZE }}u;
+  // Tail-safe scalar x4 keeps vector-like dispatch density without requiring
+  // the logical tensor length (or its storage binding) to be vec4 aligned.
+{% set itemsPerInvocation = itemsPerInvocation if itemsPerInvocation is defined else 4 %}
+  let begin = invocation * {{ itemsPerInvocation }}u;
+  let end = min(begin + {{ itemsPerInvocation }}u, params.count);
+  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() -}}
 }

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

@@ -1,19 +1,38 @@
-// Loads and stores vec4<T> (128 bits) while retaining scalar per-component
-// arithmetic, including per-component helper calls for guard-heavy operations.
-{% if usesF16 %}
-enable f16;
-{% endif %}
 {{ env.wgsl.resourceDeclarations }}
 @compute @workgroup_size({{ tunables.WORKGROUP_SIZE }})
-fn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups) nwg: vec3<u32>) {
   // 2D-folded flat index: gid.y carries the high bits when the element count exceeds the
-  // maxComputeWorkgroupsPerDimension limit (the dispatch caps x and spills the rest into y).
-  let i = gid.x + gid.y * nwg.x * {{ tunables.WORKGROUP_SIZE }}u;
   if (i >= params.count) {
     return;
   }
   let xv = x[i];
   y[i] = select(vec4<u32>(0u), vec4<u32>(1u), xv == vec4<u32>(0u));
 }

+// Loads and stores vec4<T> while evaluating the selected unary operation per
+// 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 %}
+  }
+{% endif %}
 }