sync 91d990483a17

Browse files

Files changed (6) hide show

README.md +13 -9
build/webgpu/bench.json +0 -1
build/webgpu/elementwise-bias-gelu.wgsl.jinja +11 -58
build/webgpu/manifest.json +24 -68
build/webgpu/metadata.json +10 -7
build/webgpu/test.json +1 -2

README.md CHANGED Viewed

@@ -18,15 +18,15 @@ See the [ONNX Runtime `Gelu` contrib-operator spec](https://github.com/microsoft
 ## Inputs
-| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
-| --- | --- | --- | --- | --- | --- | --- |
-| `X` | `X` | `T` | — | — | Values transformed elementwise by the exact GELU activation. | required |
 ## Outputs
-| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
-| --- | --- | --- | --- | --- | --- | --- |
-| `Y` | `Y` | `T` | same as `X` | same as `X` | Output tensor after applying GELU; same shape as the input. | required |
 ## Type constraints
@@ -36,7 +36,7 @@ See the [ONNX Runtime `Gelu` contrib-operator spec](https://github.com/microsoft
 ## 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
@@ -44,10 +44,14 @@ See the [ONNX Runtime `Gelu` contrib-operator spec](https://github.com/microsoft
 ## 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 | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- |
+| `X` | `T` | — | — | Values transformed elementwise by the exact GELU activation. | required |
 ## Outputs
+| Name | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- |
+| `Y` | `T` | same as `X` | same as `X` | Output tensor after applying GELU; same shape as the input. | 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": "com.microsoft.Gelu",
   "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
   "cases": [
     {

 {
   "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
   "cases": [
     {

build/webgpu/elementwise-bias-gelu.wgsl.jinja CHANGED Viewed

@@ -1,26 +1,17 @@
-{% if usesF16 %}
-enable f16;
-{% endif %}
 {{ env.wgsl.resourceDeclarations }}
 {% set wg = workgroupSize if workgroupSize is defined else tunables.WORKGROUP_SIZE %}
-// Bias plus GELU, with a specialization-selected tanh or erf approximation.
-// The optional bias is a rank-1 vector broadcast over the innermost (hidden)
-// axis: bias index = element_index % HIDDEN. The `vec4` path requires
-// HIDDEN % 4 == 0 and numel % 4 == 0 so a vec4 group never crosses the hidden
-// axis (the bias slice is then contiguous). `vec4Tail` keeps scalar bindings but
-// evaluates four guarded lanes per invocation, so odd hidden sizes retain the
-// same parallel efficiency without crossing row/bias boundaries. Gelu math and overflow guards match
-// the vectorized unary implementation: tanh saturates to +/-1 by |x|~9, and the
-// erf path uses the same rational approximation as Gelu.
-{% if approximate == "erf" %}
 fn erf_approx(x: f32) -> f32 {
   let ax = abs(x);
   // The polynomial has a small nonzero floor near zero. Use erf(x) ~=
   // 2/sqrt(pi)*x below 2^-20 to preserve erf(0) == 0, odd symmetry, and the
-  // correctly rounded f32 result. The exactly representable threshold keeps
-  // scalar and vector branching identical. NaN falls through to the polynomial
-  // and propagates.
   if (ax < 9.5367431640625e-7) {
     return 1.1283791670955126 * x;
   }
@@ -29,62 +20,24 @@ fn erf_approx(x: f32) -> f32 {
   let y = 1.0 - (((((1.061405429 * t - 1.453152027) * t) + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * exp(-(ax * ax));
   return sign * y;
 }
-{% else %}
-fn tanh_safe(x: f32) -> f32 {
-  if (x > 10.0) { return 1.0; }
-  if (x < -10.0) { return -1.0; }
-  return tanh(x);
-}
-{% endif %}
 fn gelu_value(v: f32) -> f32 {
-{% if approximate == "erf" %}
   return 0.5 * v * (1.0 + erf_approx(v * 0.7071067811865476));
-{% else %}
-  return 0.5 * v * (1.0 + tanh_safe(0.7978845608028654 * (v + 0.044715 * v * v * v)));
-{% endif %}
 }
-{% if hasBias %}
-const HIDDEN: u32 = {{ hidden }}u;
-{% endif %}
 @compute @workgroup_size({{ wg }})
-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 past the
-  // maxComputeWorkgroupsPerDimension dispatch limit (outputs > 16.7M elements).
-  let i = gid.x + gid.y * nwg.x * {{ wg }}u;
   if (i >= params.count) {
     return;
   }
-{% if vec4Tail %}
-  let base = i * 4u;
-{% for lane in range(4) %}
-  if (base + {{ lane }}u < params.count) {
-    let xv{{ lane }} = f32(x[base + {{ lane }}u]);
-{% if hasBias %}
-    let v{{ lane }} = xv{{ lane }} + f32(bias[(base + {{ lane }}u) % HIDDEN]);
-{% else %}
-    let v{{ lane }} = xv{{ lane }};
-{% endif %}
-    y[base + {{ lane }}u] = {{ scalar }}(gelu_value(v{{ lane }}));
-  }
-{% endfor %}
-{% elif vec4 %}
   let xv = vec4<f32>(x[i]);
-{% if hasBias %}
-  let bcol = (i * 4u) % HIDDEN;
-  let v = xv + vec4<f32>(f32(bias[bcol]), f32(bias[bcol + 1u]), f32(bias[bcol + 2u]), f32(bias[bcol + 3u]));
-{% else %}
   let v = xv;
-{% endif %}
   y[i] = vec4<{{ scalar }}>(vec4<f32>(gelu_value(v.x), gelu_value(v.y), gelu_value(v.z), gelu_value(v.w)));
 {% else %}
   let xv = f32(x[i]);
-{% if hasBias %}
-  let v = xv + f32(bias[i % HIDDEN]);
-{% else %}
   let v = xv;
-{% endif %}
   y[i] = {{ scalar }}(gelu_value(v));
 {% endif %}
 }

 {{ env.wgsl.resourceDeclarations }}
 {% set wg = workgroupSize if workgroupSize is defined else tunables.WORKGROUP_SIZE %}
+// Computes GELU, optionally after adding a rank-1 bias broadcast over the
+// innermost axis. A present bias uses element_index % hidden_extent. The `vec4`
+// path requires the innermost extent and element count to be divisible by four;
+// `vec4Tail` instead uses scalar bindings with four guarded lanes.
+// GELU uses the rational erf approximation defined below.
 fn erf_approx(x: f32) -> f32 {
   let ax = abs(x);
   // The polynomial has a small nonzero floor near zero. Use erf(x) ~=
   // 2/sqrt(pi)*x below 2^-20 to preserve erf(0) == 0, odd symmetry, and the
+  // small-input linear behavior. The threshold is exactly representable in
+  // f32. NaN falls through to the polynomial and propagates.
   if (ax < 9.5367431640625e-7) {
     return 1.1283791670955126 * x;
   }
   let y = 1.0 - (((((1.061405429 * t - 1.453152027) * t) + 1.421413741) * t - 0.284496736) * t + 0.254829592) * t * exp(-(ax * ax));
   return sign * y;
 }
 fn gelu_value(v: f32) -> f32 {
   return 0.5 * v * (1.0 + erf_approx(v * 0.7071067811865476));
 }
 @compute @workgroup_size({{ wg }})
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
   // 2D-folded flat index: gid.y carries the high bits past the
+  // per-axis dispatch fold width (outputs > 16.7M elements).
+  let i = gid.x + gid.y * {{ DISPATCH_FOLD_WIDTH }}u * {{ wg }}u;
   if (i >= params.count) {
     return;
   }
+{% if vec4 %}
   let xv = vec4<f32>(x[i]);
   let v = xv;
   y[i] = vec4<{{ scalar }}>(vec4<f32>(gelu_value(v.x), gelu_value(v.y), gelu_value(v.z), gelu_value(v.w)));
 {% else %}
   let xv = f32(x[i]);
   let v = xv;
   y[i] = {{ scalar }}(gelu_value(v));
 {% endif %}
 }

build/webgpu/manifest.json CHANGED Viewed

@@ -2,99 +2,55 @@
   "domain": "com.microsoft",
   "name": "Gelu",
   "sinceVersion": 1,
-  "description": "Applies the Gaussian Error Linear Unit (GELU) activation elementwise: `Y = 0.5 * X * (1 + erf(X / sqrt(2)))`. The output has the same shape as the input. Float16 and float32 are supported; the schema's double and bfloat16 types are not.",
-  "inputs": [
-    { "role": "X", "dtype": "T", "description": "Values transformed elementwise by the exact GELU activation." }
-  ],
-  "outputs": [
-    {
-      "role": "Y",
-      "dtype": "T",
-      "rank": "ranks.X",
-      "shape": "shapes.X",
-      "description": "Output tensor after applying GELU; same shape as the input."
-    }
-  ],
   "typeConstraints": { "T": ["float32", "float16"] },
-  "args": {
-    "X": { "kind": "tensor", "semantic": "X", "role": "input" },
-    "Y": { "kind": "tensor", "semantic": "Y", "role": "output" }
-  },
-  "tunables": { "WORKGROUP_SIZE": 256 },
-  "constants": {
-    "scalar": "dtypes.T",
-    "usesF16": "dtypes.T == \"f16\"",
-    "approximate": "\"erf\"",
-    "vec4Tail": false,
-    "hasBias": false
-  },
   "variants": [
     {
       "id": "vec4",
       "priority": 20,
-      "when": ["numel(shapes.X) == numel(shapes.Y)", "f16Ok(dtypes.T)", "numel(shapes.X) > 0", "numel(shapes.X) % 4 == 0"],
-      "constants": { "vec4": true, "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"" },
       "passes": [
         {
           "id": "main",
           "name": "Gelu.vec4",
           "shader": "elementwise-bias-gelu.wgsl.jinja",
           "bindings": [
-            {
-              "name": "x",
-              "arg": "X",
-              "semantic": "X",
-              "buffer": { "type": "read-only-storage" },
-              "elementType": "$vectorScalar"
-            },
-            {
-              "name": "y",
-              "arg": "Y",
-              "semantic": "Y",
-              "buffer": { "type": "storage" },
-              "elementType": "$vectorScalar"
-            },
-            {
-              "name": "params",
-              "semantic": "kernel.params",
-              "buffer": { "type": "uniform" },
-              "struct": {
-                "name": "Params",
-                "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.X) / 4" }]
-              }
-            }
           ],
-          "dispatch": { "threads": "numel(shapes.X) / 4", "workgroupSize": "tunables.WORKGROUP_SIZE" }
         }
       ]
     },
     {
       "id": "scalar",
       "priority": 0,
-      "when": ["numel(shapes.X) == numel(shapes.Y)", "f16Ok(dtypes.T)"],
-      "constants": { "vec4": false },
       "passes": [
         {
           "id": "main",
           "name": "Gelu.scalar",
           "shader": "elementwise-bias-gelu.wgsl.jinja",
           "bindings": [
-            {
-              "name": "x",
-              "arg": "X",
-              "semantic": "X",
-              "buffer": { "type": "read-only-storage" },
-              "elementType": "$scalar"
-            },
-            { "name": "y", "arg": "Y", "semantic": "Y", "buffer": { "type": "storage" }, "elementType": "$scalar" },
-            {
-              "name": "params",
-              "semantic": "kernel.params",
-              "buffer": { "type": "uniform" },
-              "struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.X)" }] }
-            }
           ],
-          "dispatch": { "threads": "numel(shapes.X)", "workgroupSize": "tunables.WORKGROUP_SIZE" }
         }
       ]
     }

   "domain": "com.microsoft",
   "name": "Gelu",
   "sinceVersion": 1,
+  "inputs": { "X": { "dtype": "T" } },
+  "outputs": { "Y": { "dtype": "T", "rank": "ranks.X", "shape": "shapes.X" } },
   "typeConstraints": { "T": ["float32", "float16"] },
+  "tunables": { "WORKGROUP_SIZE": { "default": 256 } },
+  "derive": { "scalar": "dtypes.T", "approximate": "\"erf\"", "vec4Tail": false, "hasBias": false },
+  "when": ["numel(shapes.X) == numel(shapes.Y)", "f16Ok(dtypes.T)"],
   "variants": [
     {
       "id": "vec4",
       "priority": 20,
+      "when": ["numel(shapes.X) > 0", "numel(shapes.X) % 4 == 0"],
+      "derive": { "vec4": true, "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"" },
       "passes": [
         {
           "id": "main",
           "name": "Gelu.vec4",
           "shader": "elementwise-bias-gelu.wgsl.jinja",
           "bindings": [
+            { "arg": "X", "name": "x", "elementType": "$vectorScalar" },
+            { "arg": "Y", "name": "y", "elementType": "$vectorScalar" },
+            { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.X) / 4" }] }
           ],
+          "dispatch": {
+            "x": "min(ceilDiv((numel(shapes.X) / 4), (tunables.WORKGROUP_SIZE)), 65535)",
+            "y": "ceilDiv(ceilDiv((numel(shapes.X) / 4), (tunables.WORKGROUP_SIZE)), 65535)",
+            "z": 1
+          }
         }
       ]
     },
     {
       "id": "scalar",
       "priority": 0,
+      "derive": { "vec4": false },
       "passes": [
         {
           "id": "main",
           "name": "Gelu.scalar",
           "shader": "elementwise-bias-gelu.wgsl.jinja",
           "bindings": [
+            { "arg": "X", "name": "x", "elementType": "$scalar" },
+            { "arg": "Y", "name": "y", "elementType": "$scalar" },
+            { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.X)" }] }
           ],
+          "dispatch": {
+            "x": "min(ceilDiv((numel(shapes.X)), (tunables.WORKGROUP_SIZE)), 65535)",
+            "y": "ceilDiv(ceilDiv((numel(shapes.X)), (tunables.WORKGROUP_SIZE)), 65535)",
+            "z": 1
+          }
         }
       ]
     }

build/webgpu/metadata.json CHANGED Viewed

@@ -1,18 +1,21 @@
 {
   "name": "com.microsoft.Gelu",
-  "id": "_com_microsoft_gelu_webgpu_6511522",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
   "digest": {
     "algorithm": "sha256",
     "files": {
-      "bench.json": "cFEvXUyEhaDZGPv9ppB2H/jtirbnrkEOd8ngzRycX+k=",
-      "elementwise-bias-gelu.wgsl.jinja": "Eg8N2jJCMce+IsYNcCzuxvs8CSv7yTQXdorvv+c0L58=",
-      "manifest.json": "89wktvMtLewo412huOWKAmpUsL1Q9k9RYkeWTxrhepE=",
-      "test.json": "CmfJcfaIcytJIPtgyCKKLh72PzJvUlrpycr/FB0EZr8="
     }
   },
-  "provenance": { "kernel": { "sha": "2e7068faf55e7f43df740015f6d1ee49391a41c5", "dirty": false } },
-  "webgpu": { "manifestSpec": "1.0", "specialized": true, "opPath": "ops/com.microsoft.Gelu" }
 }

 {
   "name": "com.microsoft.Gelu",
+  "id": "_com_microsoft_gelu_webgpu_9075789",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
   "digest": {
     "algorithm": "sha256",
     "files": {
+      "bench.json": "zb+BUPCMklZQithfY94RzWttjR/SZK0PIWgZjexuaxs=",
+      "elementwise-bias-gelu.wgsl.jinja": "K4LR9m7TibnO/vuJpCLk9+CTk1slr+E/XCGlmgH7hww=",
+      "manifest.json": "R8nYWq7dFuIme6gF1PUJWYKRv8u6u67dxzGzsZFLs38=",
+      "test.json": "K4VvkDT8J9wTZh+uYL3GSIAzvOw0PuV04RAIyhxmgxM="
     }
   },
+  "provenance": { "kernel": { "sha": "91d990483a174128daf7673f3f37a7c890493ae1", "dirty": false } },
+  "webgpu": {
+    "manifestSpec": "2.0",
+    "variants": { "vec4": ["elementwise-bias-gelu.wgsl.jinja"], "scalar": ["elementwise-bias-gelu.wgsl.jinja"] }
+  }
 }

build/webgpu/test.json CHANGED Viewed

@@ -1,5 +1,4 @@
 {
-  "op": "com.microsoft.Gelu",
   "cases": [
     {
       "name": "dispatch_cliff_scalar_f32",
@@ -105,7 +104,7 @@
       "provenance": {
         "source": "onnxruntime/test/contrib_ops/activation_op_test.cc",
         "test": "ActivationOpTest.Gelu",
-        "notes": "Scalar-path companion for com.microsoft.Gelu subnormal linear-region behavior."
       },
       "inputs": {
         "X": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [1e-40, -1e-40, 1e-38] } }

 {
   "cases": [
     {
       "name": "dispatch_cliff_scalar_f32",
       "provenance": {
         "source": "onnxruntime/test/contrib_ops/activation_op_test.cc",
         "test": "ActivationOpTest.Gelu",
+        "notes": "Subnormal inputs exercise the exact GELU linear region on the scalar path."
       },
       "inputs": {
         "X": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [1e-40, -1e-40, 1e-38] } }