sync 91d990483a17

Browse files

Files changed (7) hide show

README.md +11 -7
build/webgpu/bench.json +0 -1
build/webgpu/manifest.json +43 -50
build/webgpu/metadata.json +11 -8
build/webgpu/test.json +3 -4
build/webgpu/unary-scalar.wgsl.jinja +12 -12
build/webgpu/unary-vec4.wgsl.jinja +31 -13

README.md CHANGED Viewed

@@ -18,15 +18,15 @@ See the [ONNX `Cos` spec](https://onnx.ai/onnx/operators/onnx__Cos.html) for the
 ## Inputs
-| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
 | --- | --- | --- | --- | --- | --- | --- |
-| `input` | `x` | `T` | — | — | Angles in radians whose cosine is computed elementwise. | required |
 ## Outputs
-| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
 | --- | --- | --- | --- | --- | --- | --- |
-| `output` | `y` | `T` | same as `input` | same as `input` | Elementwise cosine of the input tensor. | required |
 ## Type constraints
@@ -36,7 +36,7 @@ See the [ONNX `Cos` spec](https://onnx.ai/onnx/operators/onnx__Cos.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 `Cos` spec](https://onnx.ai/onnx/operators/onnx__Cos.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` | `input` | `T` | — | — | Angles in radians whose cosine is computed elementwise. | required |
 ## Outputs
+| Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence |
 | --- | --- | --- | --- | --- | --- | --- |
+| `y` | `output` | `T` | same as `x` | same as `x` | Elementwise cosine of the input tensor. | 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.Cos",
   "cases": [
     {
       "name": "1m_f32",

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

build/webgpu/manifest.json CHANGED Viewed

@@ -2,30 +2,16 @@
   "domain": "ai.onnx",
   "name": "Cos",
   "sinceVersion": 7,
-  "description": "Computes the cosine of each element in the input tensor. The output has the same shape and type as the input.",
-  "inputs": [
-    { "role": "input", "dtype": "T", "description": "Angles in radians whose cosine is computed elementwise." }
-  ],
-  "outputs": [
-    {
-      "role": "output",
-      "dtype": "T",
-      "rank": "ranks.input",
-      "description": "Elementwise cosine of the input tensor.",
-      "shape": "shapes.input"
-    }
-  ],
   "typeConstraints": { "T": ["float32", "float16"] },
-  "args": {
-    "x": { "kind": "tensor", "semantic": "input", "role": "input" },
-    "y": { "kind": "tensor", "semantic": "output", "role": "output" }
-  },
-  "tunables": { "WORKGROUP_SIZE": 256 },
   "variants": [
     {
       "id": "same_layout_vec4",
       "when": ["numel(shapes.x) > 0", "numel(shapes.x) % 4 == 0", "numel(shapes.x) == numel(shapes.y)", "f16Ok(dtypes.T)"],
-      "constants": {
         "scalar": "dtypes.T",
         "usesF16": "dtypes.T == \"f16\"",
         "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\""
@@ -34,47 +20,54 @@
         {
           "id": "main",
           "name": "Cos.vec4",
-          "source": { "shader": "unary-vec4.wgsl.jinja", "inputs": { "op": "\"cos\"" } },
-          "bindings": [
-            { "name": "x", "arg": "x", "buffer": { "type": "read-only-storage" }, "elementType": "$vectorScalar" },
-            { "name": "y", "arg": "y", "buffer": { "type": "storage" }, "elementType": "$vectorScalar" },
-            {
-              "name": "params",
-              "semantic": "kernel.params",
-              "buffer": { "type": "uniform" },
-              "struct": {
-                "name": "Params",
-                "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.y) / 4" }]
-              }
-            }
-          ],
-          "dispatch": { "threads": "numel(shapes.y) / 4", "workgroupSize": "tunables.WORKGROUP_SIZE" }
         }
-      ],
-      "priority": 20
     },
     {
       "id": "elementwise",
       "when": ["numel(shapes.x) == numel(shapes.y)", "f16Ok(dtypes.T)"],
-      "constants": { "scalar": "dtypes.T", "usesF16": "dtypes.T == \"f16\"" },
       "passes": [
         {
           "id": "main",
           "name": "Cos",
-          "source": { "shader": "unary-scalar.wgsl.jinja", "inputs": { "op": "\"cos\"", "itemsPerInvocation": 4 } },
-          "bindings": [
-            { "name": "x", "arg": "x", "buffer": { "type": "read-only-storage" }, "elementType": "$scalar" },
-            { "name": "y", "arg": "y", "buffer": { "type": "storage" }, "elementType": "$scalar" },
-            {
-              "name": "params",
-              "semantic": "kernel.params",
-              "buffer": { "type": "uniform" },
-              "struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.y)" }] }
-            }
-          ],
-          "dispatch": { "threads": "ceilDiv(numel(shapes.y), 4)", "workgroupSize": "tunables.WORKGROUP_SIZE" }
         }
       ]
     }
-  ]
 }

   "domain": "ai.onnx",
   "name": "Cos",
   "sinceVersion": 7,
+  "inputs": { "x": { "onnx": "input", "dtype": "T" } },
+  "outputs": { "y": { "onnx": "output", "dtype": "T", "rank": "ranks.x", "shape": "shapes.x" } },
   "typeConstraints": { "T": ["float32", "float16"] },
+  "tunables": { "WORKGROUP_SIZE": { "default": 256 } },
   "variants": [
     {
       "id": "same_layout_vec4",
+      "priority": 20,
       "when": ["numel(shapes.x) > 0", "numel(shapes.x) % 4 == 0", "numel(shapes.x) == numel(shapes.y)", "f16Ok(dtypes.T)"],
+      "derive": {
         "scalar": "dtypes.T",
         "usesF16": "dtypes.T == \"f16\"",
         "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\""
         {
           "id": "main",
           "name": "Cos.vec4",
+          "shader": "unary-vec4.wgsl.jinja",
+          "derive": {
+            "op": "\"cos\"",
+            "vec4PerThread": "4 if numel(shapes.y) * dtypeBytes(tensorDtypes.y) <= 16777216 else 1"
+          },
+          "bindings": ["x", "y", "params_unary"],
+          "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",
       "when": ["numel(shapes.x) == numel(shapes.y)", "f16Ok(dtypes.T)"],
+      "derive": { "scalar": "dtypes.T", "usesF16": "dtypes.T == \"f16\"" },
       "passes": [
         {
           "id": "main",
           "name": "Cos",
+          "shader": "unary-scalar.wgsl.jinja",
+          "derive": { "op": "\"cos\"", "itemsPerInvocation": 4 },
+          "bindings": ["x_2", "y_2", "params_2_unary"],
+          "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
+          }
         }
       ]
     }
+  ],
+  "bindings": {
+    "x": { "buffer": "read-only-storage", "elementType": "$vectorScalar" },
+    "y": { "buffer": "storage", "elementType": "$vectorScalar" },
+    "params_unary": {
+      "buffer": "uniform",
+      "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.y) / 4" }],
+      "name": "params"
+    },
+    "x_2": { "buffer": "read-only-storage", "name": "x", "elementType": "$scalar" },
+    "y_2": { "buffer": "storage", "name": "y", "elementType": "$scalar" },
+    "params_2_unary": {
+      "buffer": "uniform",
+      "name": "params",
+      "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.y)" }]
+    }
+  }
 }

build/webgpu/metadata.json CHANGED Viewed

@@ -1,19 +1,22 @@
 {
   "name": "ai.onnx.Cos",
-  "id": "_ai_onnx_cos_webgpu_cac3d94",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
   "digest": {
     "algorithm": "sha256",
     "files": {
-      "bench.json": "aMFVQ+6rhvrMngby96efMmsldwq0VRJ9P0ICGr91lUY=",
-      "manifest.json": "jHDCQwVeQCIpu3FV1PiYQD9ymkzPseA2K6yGVJf3wBk=",
-      "test.json": "4f0wcy6ce8pie62+H13dy87SkZOfWv4TkyeqalgROmw=",
-      "unary-scalar.wgsl.jinja": "VPyTY5cV6Zjidm4VpLmKEhg9VJDzG6N8+WZ/QJc+mlo=",
-      "unary-vec4.wgsl.jinja": "1g3ktdx4TPnyT48hkpcIvkPRV8APmLzL3NESPtJPryA="
     }
   },
-  "provenance": { "kernel": { "sha": "2e7068faf55e7f43df740015f6d1ee49391a41c5", "dirty": false } },
-  "webgpu": { "manifestSpec": "1.0", "specialized": true, "opPath": "ops/ai.onnx.Cos" }
 }

 {
   "name": "ai.onnx.Cos",
+  "id": "_ai_onnx_cos_webgpu_b8cb940",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
   "digest": {
     "algorithm": "sha256",
     "files": {
+      "bench.json": "hZB5P81WgcOr8dQM++HnJ3SeV7se1ce5xBKtyIc7wAU=",
+      "manifest.json": "zfXDJZ9MmehTi2+hmO6LmOc6d0G/3MoiI/BBrdGNbec=",
+      "test.json": "nCOlkbGujYFc4xvpMBa8GIhuE9gsEvUwrN7hbkN056A=",
+      "unary-scalar.wgsl.jinja": "u8Oh5cH6SLO8tlNHu5CQY6yrH7ZuJnUDAV2p++iIGTc=",
+      "unary-vec4.wgsl.jinja": "h+TvunU/ZWWcEFtLA63R/pj1TbBBTYdsbcKWfpAWVbY="
     }
   },
+  "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.Cos",
   "cases": [
     {
       "name": "f32_values",
@@ -30,7 +29,7 @@
       "name": "f32_large_argument_range_reduction_accuracy_gpu_gap",
       "skipGpu": {
         "category": "todo",
-        "reason": "The current f32 large-argument range reduction differs from the CPU reference by about one ULP. A more accurate range reduction or software-extended precision could close this implementable gap."
       },
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
@@ -53,12 +52,12 @@
       "name": "f32_large_argument_range_reduction_accuracy_scalar_gpu_gap",
       "skipGpu": {
         "category": "todo",
-        "reason": "The current scalar f32 large-argument range reduction differs from the CPU reference by about one ULP. A more accurate range reduction or software-extended precision could close this implementable gap."
       },
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
         "test": "MathOpTest.CosFloat",
-        "notes": "Scalar-path companion for strict large-argument Cos range-reduction accuracy."
       },
       "inputs": {
         "x": {

 {
   "cases": [
     {
       "name": "f32_values",
       "name": "f32_large_argument_range_reduction_accuracy_gpu_gap",
       "skipGpu": {
         "category": "todo",
+        "reason": "The package's f32 large-argument range reduction differs from the correctly rounded expected value by about one ULP. More accurate range reduction or software-extended precision could close the gap."
       },
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
       "name": "f32_large_argument_range_reduction_accuracy_scalar_gpu_gap",
       "skipGpu": {
         "category": "todo",
+        "reason": "The package's scalar f32 large-argument range reduction differs from the correctly rounded expected value by about one ULP. More accurate range reduction or software-extended precision could close the gap."
       },
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
         "test": "MathOpTest.CosFloat",
+        "notes": "Unaligned scalar-path inputs exercise strict large-argument Cos range-reduction accuracy."
       },
       "inputs": {
         "x": {

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 %}
@@ -102,11 +103,10 @@ fn reduce_pio2_fast(ax: f32) -> Pio2 {
 {% set needPreciseTrigTwoPi = false %}
 // Backend-stable f32 sine/cosine core.
 //
-// Shader transcendental accuracy is implementation-defined, and some portable
-// backends are only accurate to roughly 1e-4. Each path retains the most accurate
-// available phase representation, reduces it to [-pi, pi], then uses these
-// polynomials. The half-pi core is degree 13 for sine and degree 12 for cosine;
-// truncation error is well below one f32 ULP over its documented interval.
 {% if needPreciseTrigCentered %}
 const PRECISE_TRIG_PI: f32 = 3.141592653589793;
 {% 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 %}
 {% set needPreciseTrigTwoPi = false %}
 // Backend-stable f32 sine/cosine core.
 //
+// Shader transcendental accuracy is implementation-defined. Each path retains
+// its available phase representation, reduces it to [-pi, pi], then evaluates
+// explicit polynomials in a fixed order. The half-pi core is degree 13 for sine
+// and degree 12 for cosine.
 {% if needPreciseTrigCentered %}
 const PRECISE_TRIG_PI: f32 = 3.141592653589793;
 {% endif %}

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

@@ -1,8 +1,5 @@
-// 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 }}
 {% macro emit_reduce_pio2() %}
@@ -87,11 +84,10 @@ fn reduce_pio2_fast(ax: f32) -> Pio2 {
 {% set needPreciseTrigTwoPi = false %}
 // Backend-stable f32 sine/cosine core.
 //
-// Shader transcendental accuracy is implementation-defined, and some portable
-// backends are only accurate to roughly 1e-4. Each path retains the most accurate
-// available phase representation, reduces it to [-pi, pi], then uses these
-// polynomials. The half-pi core is degree 13 for sine and degree 12 for cosine;
-// truncation error is well below one f32 ULP over its documented interval.
 {% if needPreciseTrigCentered %}
 const PRECISE_TRIG_PI: f32 = 3.141592653589793;
 {% endif %}
@@ -184,15 +180,37 @@ fn cos_accurate(x: f32) -> f32 {
 {{ emit_trig_reduction_support() }}
 {{ emit_cos_accurate() }}
 @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];
   let fv = vec4<f32>(xv);
   y[i] = {{ vectorScalar }}(vec4<f32>(cos_accurate(fv.x), cos_accurate(fv.y), cos_accurate(fv.z), cos_accurate(fv.w)));
 }

+// Loads and stores vec4<T> while evaluating the selected unary operation per
+// component.
 {{ env.wgsl.resourceDeclarations }}
 {% macro emit_reduce_pio2() %}
 {% set needPreciseTrigTwoPi = false %}
 // Backend-stable f32 sine/cosine core.
 //
+// Shader transcendental accuracy is implementation-defined. Each path retains
+// its available phase representation, reduces it to [-pi, pi], then evaluates
+// explicit polynomials in a fixed order. The half-pi core is degree 13 for sine
+// and degree 12 for cosine.
 {% if needPreciseTrigCentered %}
 const PRECISE_TRIG_PI: f32 = 3.141592653589793;
 {% endif %}
 {{ emit_trig_reduction_support() }}
 {{ emit_cos_accurate() }}
+{% 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];
   let fv = vec4<f32>(xv);
   y[i] = {{ vectorScalar }}(vec4<f32>(cos_accurate(fv.x), cos_accurate(fv.y), cos_accurate(fv.z), cos_accurate(fv.w)));
+{% if vec4PerThread > 1 %}
+  }
+{% endif %}
 }