sync 91d990483a17

Browse files

Files changed (7) hide show

README.md +13 -9
build/webgpu/bench.json +0 -1
build/webgpu/manifest.json +31 -71
build/webgpu/metadata.json +11 -8
build/webgpu/test.json +3 -6
build/webgpu/unary-scalar.wgsl.jinja +8 -7
build/webgpu/unary-vec4.wgsl.jinja +27 -8

README.md CHANGED Viewed

@@ -18,15 +18,15 @@ See the [ONNX `Sign` spec](https://onnx.ai/onnx/operators/onnx__Sign.html) for t
 ## Inputs
-| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
-| --- | --- | --- | --- | --- | --- | --- |
-| `input` | `input` | `T` | — | — | Input tensor of any numeric type. | required |
 ## Outputs
-| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
-| --- | --- | --- | --- | --- | --- | --- |
-| `output` | `output` | `T` | same as `input` | same as `input` | Elementwise sign of the input; same shape and type as the input. | required |
 ## Type constraints
@@ -36,7 +36,7 @@ See the [ONNX `Sign` spec](https://onnx.ai/onnx/operators/onnx__Sign.html) for t
 ## 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 `Sign` spec](https://onnx.ai/onnx/operators/onnx__Sign.html) for t
 ## 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 |
+| --- | --- | --- | --- | --- | --- |
+| `input` | `T` | — | — | Input tensor of any numeric type. | required |
 ## Outputs
+| Name | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- |
+| `output` | `T` | same as `input` | same as `input` | Elementwise sign of the input; same shape and type 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": "ai.onnx.Sign",
   "cases": [
     {
       "name": "1m_f32",

 {
   "cases": [
     {
       "name": "1m_f32",

build/webgpu/manifest.json CHANGED Viewed

@@ -2,98 +2,58 @@
   "domain": "ai.onnx",
   "name": "Sign",
   "sinceVersion": 13,
-  "description": "Computes the elementwise sign of the input tensor: outputs `1` for positive values, `-1` for negative values, and `0` where the input is zero. The output has the same shape and type as the input.",
-  "inputs": [{ "role": "input", "dtype": "T", "description": "Input tensor of any numeric type." }],
-  "outputs": [
-    {
-      "role": "output",
-      "dtype": "T",
-      "rank": "ranks.input",
-      "description": "Elementwise sign of the input; same shape and type as the input.",
-      "shape": "shapes.input"
-    }
-  ],
   "typeConstraints": { "T": ["float32", "float16", "int32", "uint32", "int16", "int8", "uint8"] },
-  "args": {
-    "input": { "kind": "tensor", "semantic": "input", "role": "input" },
-    "output": { "kind": "tensor", "semantic": "output", "role": "output" }
-  },
-  "tunables": { "WORKGROUP_SIZE": 256 },
-  "constants": { "scalar": "dtypes.T", "usesF16": "dtypes.T == \"f16\"" },
   "variants": [
     {
       "id": "same_layout_vec4",
       "when": ["numel(shapes.input) > 0", "numel(shapes.input) % 4 == 0", "numel(shapes.input) == numel(shapes.output)", "f16Ok(dtypes.T)"],
-      "constants": { "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"" },
       "passes": [
         {
           "id": "main",
           "name": "Sign.vec4",
-          "source": { "shader": "unary-vec4.wgsl.jinja", "inputs": { "op": "\"sign\"" } },
           "bindings": [
-            {
-              "name": "x",
-              "arg": "input",
-              "semantic": "input",
-              "buffer": { "type": "read-only-storage" },
-              "elementType": "$vectorScalar"
-            },
-            {
-              "name": "y",
-              "arg": "output",
-              "semantic": "output",
-              "buffer": { "type": "storage" },
-              "elementType": "$vectorScalar"
-            },
-            {
-              "name": "params",
-              "semantic": "kernel.params",
-              "buffer": { "type": "uniform" },
-              "struct": {
-                "name": "Params",
-                "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.output) / 4" }]
-              }
-            }
           ],
-          "dispatch": { "threads": "numel(shapes.output) / 4", "workgroupSize": "tunables.WORKGROUP_SIZE" }
         }
-      ],
-      "priority": 20
     },
     {
       "id": "elementwise",
-      "when": "numel(shapes.input) == numel(shapes.output) and (f16Ok(dtypes.T))",
       "passes": [
         {
           "id": "main",
           "name": "Sign",
-          "source": { "shader": "unary-scalar.wgsl.jinja", "inputs": { "op": "\"sign\"", "itemsPerInvocation": 4 } },
           "bindings": [
-            {
-              "name": "input",
-              "arg": "input",
-              "semantic": "input",
-              "buffer": { "type": "read-only-storage" },
-              "elementType": "$scalar"
-            },
-            {
-              "name": "output",
-              "arg": "output",
-              "semantic": "output",
-              "buffer": { "type": "storage" },
-              "elementType": "$scalar"
-            },
-            {
-              "name": "params",
-              "semantic": "kernel.params",
-              "buffer": { "type": "uniform" },
-              "struct": {
-                "name": "Params",
-                "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.output)" }]
-              }
-            }
           ],
-          "dispatch": { "threads": "ceilDiv(numel(shapes.output), 4)", "workgroupSize": "tunables.WORKGROUP_SIZE" }
         }
       ]
     }

   "domain": "ai.onnx",
   "name": "Sign",
   "sinceVersion": 13,
+  "inputs": { "input": { "dtype": "T" } },
+  "outputs": { "output": { "dtype": "T", "rank": "ranks.input", "shape": "shapes.input" } },
   "typeConstraints": { "T": ["float32", "float16", "int32", "uint32", "int16", "int8", "uint8"] },
+  "tunables": { "WORKGROUP_SIZE": { "default": 256 } },
+  "derive": { "scalar": "dtypes.T", "usesF16": "dtypes.T == \"f16\"" },
   "variants": [
     {
       "id": "same_layout_vec4",
+      "priority": 20,
       "when": ["numel(shapes.input) > 0", "numel(shapes.input) % 4 == 0", "numel(shapes.input) == numel(shapes.output)", "f16Ok(dtypes.T)"],
+      "derive": { "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"" },
       "passes": [
         {
           "id": "main",
           "name": "Sign.vec4",
+          "shader": "unary-vec4.wgsl.jinja",
+          "derive": {
+            "op": "\"sign\"",
+            "vec4PerThread": "4 if numel(shapes.output) * dtypeBytes(tensorDtypes.output) <= 16777216 else 1"
+          },
           "bindings": [
+            { "arg": "input", "name": "x", "elementType": "$vectorScalar" },
+            { "arg": "output", "name": "y", "elementType": "$vectorScalar" },
+            { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.output) / 4" }] }
           ],
+          "dispatch": {
+            "x": "min(ceilDiv((ceilDiv(numel(shapes.output) / 4, 4 if numel(shapes.output) * dtypeBytes(tensorDtypes.output) <= 16777216 else 1)), (tunables.WORKGROUP_SIZE)), 65535)",
+            "y": "ceilDiv(ceilDiv((ceilDiv(numel(shapes.output) / 4, 4 if numel(shapes.output) * dtypeBytes(tensorDtypes.output) <= 16777216 else 1)), (tunables.WORKGROUP_SIZE)), 65535)",
+            "z": 1
+          }
         }
+      ]
     },
     {
       "id": "elementwise",
+      "when": ["numel(shapes.input) == numel(shapes.output) and (f16Ok(dtypes.T))"],
       "passes": [
         {
           "id": "main",
           "name": "Sign",
+          "shader": "unary-scalar.wgsl.jinja",
+          "derive": { "op": "\"sign\"", "itemsPerInvocation": 4 },
           "bindings": [
+            "input",
+            "output",
+            { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.output)" }] }
           ],
+          "dispatch": {
+            "x": "min(ceilDiv((ceilDiv(numel(shapes.output), 4)), (tunables.WORKGROUP_SIZE)), 65535)",
+            "y": "ceilDiv(ceilDiv((ceilDiv(numel(shapes.output), 4)), (tunables.WORKGROUP_SIZE)), 65535)",
+            "z": 1
+          }
         }
       ]
     }

build/webgpu/metadata.json CHANGED Viewed

@@ -1,19 +1,22 @@
 {
   "name": "ai.onnx.Sign",
-  "id": "_ai_onnx_sign_webgpu_6d0a2b1",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
   "digest": {
     "algorithm": "sha256",
     "files": {
-      "bench.json": "4q8mZQrcY+ViFixjxj3zYSAeXdqwAnfqK6JL0eXfAeU=",
-      "manifest.json": "4+AgHdYHHwfbRxAG9bwppknCrSxXYIjvelehzFR1osw=",
-      "test.json": "xB83KjUJDVMn8MAFK2+ZMNngy6SCZSwPM0ALPFALosA=",
-      "unary-scalar.wgsl.jinja": "dCTx9hbmVW/fySnWmzBzZT9z4Ij9/iG6RFliizgwwxs=",
-      "unary-vec4.wgsl.jinja": "h0Y4ksIDPy57eIshgKVikk0FJSpUp/mXgHTtmxmVRNw="
     }
   },
-  "provenance": { "kernel": { "sha": "2e7068faf55e7f43df740015f6d1ee49391a41c5", "dirty": false } },
-  "webgpu": { "manifestSpec": "1.0", "specialized": true, "opPath": "ops/ai.onnx.Sign" }
 }

 {
   "name": "ai.onnx.Sign",
+  "id": "_ai_onnx_sign_webgpu_7e75e78",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
   "digest": {
     "algorithm": "sha256",
     "files": {
+      "bench.json": "BeK/pm3J1N63lmR/lECkTp1EWNnUeMFlM0F3OdgdhPk=",
+      "manifest.json": "EvDo7Z3CuI8sTL6W9FMajRDcM3UUbuipgGymQxGltlM=",
+      "test.json": "wqsQX1gxCRJBO5+/hR/fIFpxvnVtWZirFG1q+C+tF5I=",
+      "unary-scalar.wgsl.jinja": "ZmU+8SDLS8BCho41AuG6LIxEFTVx+xq8TBb3kyuTC+s=",
+      "unary-vec4.wgsl.jinja": "5eC2XytxRMc9QBwFeqlwy9ubxkWVr9aClnppDxHbvWQ="
     }
   },
+  "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.Sign",
   "cases": [
     {
       "name": "int16_scalar_boundaries",
@@ -44,7 +43,7 @@
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/math/sign_test.cc",
         "test": "MathOpTest.Sign_float",
-        "notes": "Vec4 companion for signed subnormal Sign behavior; optimized unary paths must preserve nonzero signs."
       },
       "inputs": {
         "input": {
@@ -236,7 +235,7 @@
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/math/sign_test.cc",
         "test": "MathOpTest.Sign_float",
-        "notes": "Scalar-path companion for ORT CPU-verified NaN preservation."
       },
       "inputs": {
         "input": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": ["NaN", -2.0, 2.0] } }
@@ -301,9 +300,7 @@
     },
     {
       "name": "f32_scalar_tail_4095",
-      "provenance": {
-        "notes": "Compact sibling for the large f32 Sign scalar-fallback benchmark; odd numel forces the elementwise path while preserving exact sign semantics."
-      },
       "inputs": {
         "input": {
           "dtype": "float32",

 {
   "cases": [
     {
       "name": "int16_scalar_boundaries",
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/math/sign_test.cc",
         "test": "MathOpTest.Sign_float",
+        "notes": "Vectorized Sign must preserve the signs of nonzero subnormal inputs."
       },
       "inputs": {
         "input": {
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/math/sign_test.cc",
         "test": "MathOpTest.Sign_float",
+        "notes": "Scalar Sign must preserve NaN inputs."
       },
       "inputs": {
         "input": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": ["NaN", -2.0, 2.0] } }
     },
     {
       "name": "f32_scalar_tail_4095",
+      "provenance": { "notes": "An odd float32 element count exercises scalar Sign with exact sign semantics." },
       "inputs": {
         "input": {
           "dtype": "float32",

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

@@ -1,21 +1,22 @@
 {% macro flat_tail_open() %}
 @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 invocation = gid.x + gid.y * nwg.x * {{ 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.
-  let begin = invocation * {{ source.itemsPerInvocation }}u;
-  let end = min(begin + {{ source.itemsPerInvocation }}u, params.count);
   for (var i = begin; i < end; i = i + 1u) {
 {%- endmacro %}
 {% macro flat_tail_close() %}
   }
 {% endmacro %}
-// Scalar unary fallback. Each branch retains the operation's numeric hardening,
-// including Payne-Hanek trigonometric range reduction and NaN/overflow guards.
 {% if usesF16 %}
 enable f16;
 {% endif %}

 {% 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.
 {% if usesF16 %}
 enable f16;
 {% endif %}

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

@@ -1,19 +1,35 @@
-// 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];
 {% if scalar == "i32" %}
   y[i] = select(select(vec4<i32>(0i), vec4<i32>(1i), xv > vec4<i32>(0i)), vec4<i32>(-1i), xv < vec4<i32>(0i));
@@ -28,4 +44,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups)
   let signed = select(select(vec4<f32>(0.0), vec4<f32>(1.0), fv > vec4<f32>(0.0)), vec4<f32>(-1.0), fv < vec4<f32>(0.0));
   y[i] = {{ vectorScalar }}(select(signed, fv, fv_is_nan));
 {% endif %}
 }

+// 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];
 {% if scalar == "i32" %}
   y[i] = select(select(vec4<i32>(0i), vec4<i32>(1i), xv > vec4<i32>(0i)), vec4<i32>(-1i), xv < vec4<i32>(0i));
   let signed = select(select(vec4<f32>(0.0), vec4<f32>(1.0), fv > vec4<f32>(0.0)), vec4<f32>(-1.0), fv < vec4<f32>(0.0));
   y[i] = {{ vectorScalar }}(select(signed, fv, fv_is_nan));
 {% endif %}
+{% if vec4PerThread > 1 %}
+  }
+{% endif %}
 }