sync 2e7068faf55e

Browse files

Files changed (7) hide show

README.md +57 -0
build/webgpu/bench.json +49 -0
build/webgpu/manifest.json +80 -0
build/webgpu/metadata.json +19 -0
build/webgpu/test.json +178 -0
build/webgpu/unary-scalar.wgsl.jinja +26 -0
build/webgpu/unary-vec4.wgsl.jinja +19 -0

README.md CHANGED Viewed

@@ -1,3 +1,60 @@
 ---
 license: apache-2.0
 ---

 ---
+library_name: kernels
 license: apache-2.0
+tags:
+- kernel
+- webgpu
+- wgsl
 ---
+# ai.onnx.Sqrt
+`ai.onnx`  ·  standard ONNX operator  ·  ONNX opset ≥ 13
+## Description
+Computes the elementwise square root of the input tensor: `Y = X^0.5`. Returns `NaN` for negative input values.
+See the [ONNX `Sqrt` spec](https://onnx.ai/onnx/operators/onnx__Sqrt.html) for the reference semantics.
+## Inputs
+| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- | --- |
+| `X` | `x` | `T` | — | — | Values whose principal square roots are computed elementwise. | required |
+## Outputs
+| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- | --- |
+| `Y` | `y` | `T` | same as `X` | same as `X` | Output tensor containing the elementwise square roots of `X`. | required |
+## Type constraints
+| Variable | Allowed dtypes |
+| --- | --- |
+| `T` | `float32`, `float16` |
+## 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
+- [`unary-scalar.wgsl.jinja`](build/webgpu/unary-scalar.wgsl.jinja)
+- [`unary-vec4.wgsl.jinja`](build/webgpu/unary-vec4.wgsl.jinja)
+## 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.
+```js
+import { getKernel } from "@huggingface/kernels";
+const kernel = await getKernel("webgpu-kernels/ai.onnx.Sqrt", { version: 1 });
+const { y } = await kernel({ x: { data: xData, shape: [] } });
+```

build/webgpu/bench.json ADDED Viewed

	@@ -0,0 +1,49 @@

+{
+  "op": "ai.onnx.Sqrt",
+  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
+  "cases": [
+    {
+      "name": "sqrt-f32-1m",
+      "preset": "smoke",
+      "vars": { "dtype": "float32", "count": 1048576 },
+      "inputs": {
+        "x": { "shape": [1048576], "dtype": "float32", "dist": "uniform", "seed": 610, "scale": 1, "offset": 1 }
+      },
+      "outputs": { "y": { "shape": [1048576], "dtype": "float32" } },
+      "bench": {
+        "primary": true,
+        "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }]
+      }
+    },
+    {
+      "name": "sqrt-f32-8m",
+      "preset": "smoke",
+      "vars": { "dtype": "float32", "count": 8388608 },
+      "inputs": {
+        "x": { "shape": [8388608], "dtype": "float32", "seed": 7013, "dist": "uniform", "scale": 1, "offset": 1 }
+      },
+      "outputs": { "y": { "shape": [8388608], "dtype": "float32" } },
+      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
+    },
+    {
+      "name": "sqrt-f16-8m",
+      "preset": "smoke",
+      "vars": { "dtype": "float16", "count": 8388608 },
+      "inputs": {
+        "x": { "shape": [8388608], "dtype": "float16", "seed": 7014, "dist": "uniform", "scale": 1, "offset": 1 }
+      },
+      "outputs": { "y": { "shape": [8388608], "dtype": "float16" } },
+      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
+    },
+    {
+      "name": "sqrt-f32-8m-scalar-fallback",
+      "preset": "edge",
+      "vars": { "dtype": "float32", "count": 8388609 },
+      "inputs": {
+        "x": { "shape": [8388609], "dtype": "float32", "dist": "uniform", "seed": 7211, "scale": 1, "offset": 1 }
+      },
+      "outputs": { "y": { "shape": [8388609], "dtype": "float32", "dist": "empty" } },
+      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
+    }
+  ]
+}

build/webgpu/manifest.json ADDED Viewed

	@@ -0,0 +1,80 @@

+{
+  "domain": "ai.onnx",
+  "name": "Sqrt",
+  "sinceVersion": 13,
+  "description": "Computes the elementwise square root of the input tensor: `Y = X^0.5`. Returns `NaN` for negative input values.",
+  "inputs": [
+    { "role": "X", "dtype": "T", "description": "Values whose principal square roots are computed elementwise." }
+  ],
+  "outputs": [
+    {
+      "role": "Y",
+      "dtype": "T",
+      "rank": "ranks.X",
+      "description": "Output tensor containing the elementwise square roots of `X`.",
+      "shape": "shapes.X"
+    }
+  ],
+  "typeConstraints": { "T": ["float32", "float16"] },
+  "args": {
+    "x": { "kind": "tensor", "semantic": "X", "role": "input" },
+    "y": { "kind": "tensor", "semantic": "Y", "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 ~ \">\""
+      },
+      "passes": [
+        {
+          "id": "main",
+          "name": "Sqrt.vec4",
+          "source": { "shader": "unary-vec4.wgsl.jinja", "inputs": { "op": "\"sqrt\"" } },
+          "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": "scalar",
+      "when": ["numel(shapes.x) == numel(shapes.y)", "f16Ok(dtypes.T)"],
+      "constants": { "scalar": "dtypes.T", "usesF16": "dtypes.T == \"f16\"" },
+      "passes": [
+        {
+          "id": "main",
+          "name": "Sqrt",
+          "source": { "shader": "unary-scalar.wgsl.jinja", "inputs": { "op": "\"sqrt\"", "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" }
+        }
+      ]
+    }
+  ]
+}

build/webgpu/metadata.json ADDED Viewed

	@@ -0,0 +1,19 @@

+{
+  "name": "ai.onnx.Sqrt",
+  "id": "_ai_onnx_sqrt_webgpu_a12c6e5",
+  "version": 1,
+  "license": "Apache-2.0",
+  "backend": { "type": "webgpu" },
+  "digest": {
+    "algorithm": "sha256",
+    "files": {
+      "bench.json": "zZsOr8xlZCQSGFaKHtZrqo5rA+Cm+Wm57it9ocFzj+o=",
+      "manifest.json": "n37vXFoJbr2LZzaamwpaAPZqtoeGiEYLn52ql30g+xo=",
+      "test.json": "RSo58XZfuOywGddYHTpyTvrpBXY9gOZL6LWTwjwpgtg=",
+      "unary-scalar.wgsl.jinja": "xsLYVBQLpl1Xq5FyPATjZem+6Vqb1o3ef/RsvBuKwAw=",
+      "unary-vec4.wgsl.jinja": "8lb1ey0+QO0x7qrQ5uqlrgEeJJtWToXkKBbKSo9EyHc="
+    }
+  },
+  "provenance": { "kernel": { "sha": "2e7068faf55e7f43df740015f6d1ee49391a41c5", "dirty": false } },
+  "webgpu": { "manifestSpec": "1.0", "specialized": true, "opPath": "ops/ai.onnx.Sqrt" }
+}

build/webgpu/test.json ADDED Viewed

	@@ -0,0 +1,178 @@

+{
+  "op": "ai.onnx.Sqrt",
+  "cases": [
+    {
+      "name": "vector_positive",
+      "inputs": { "x": { "dtype": "float32", "shape": [32], "data": { "kind": "constant", "value": 4.0 } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [32], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "rank0_scalar_positive",
+      "inputs": { "x": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [9.0] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "special_values_neg_inf_nan",
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [6],
+          "data": { "kind": "values", "values": [-1.0, -4.0, 0.0, 4.0, "Infinity", "NaN"] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [6], "tolerance": 0.000001, "allowNaN": true } }
+    },
+    {
+      "name": "positive_subnormal_values_gpu_gap",
+      "skipGpu": {
+        "category": "permanent",
+        "reason": "Portable WGSL floating-point semantics do not guarantee preservation of the subnormal values required by this fixture. Backend evidence: WebGPU/Metal flushes subnormals to zero (f32 and f16); the kernel cannot preserve denormal inputs/outputs bit-exactly."
+      },
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Sqrt_Float",
+        "notes": "Positive subnormal inputs should produce tiny nonzero square roots; zero-flushing loses the signal."
+      },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [4], "data": { "kind": "values", "values": [1e-45, 1e-40, 1e-39, 1e-38] } }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [4], "tolerance": 1e-27 } }
+    },
+    {
+      "name": "positive_subnormal_values_scalar_gpu_gap",
+      "skipGpu": {
+        "category": "permanent",
+        "reason": "Portable WGSL floating-point semantics do not guarantee preservation of the subnormal values required by this fixture. Backend evidence: WebGPU/Metal flushes subnormals to zero (f32 and f16); the kernel cannot preserve denormal inputs/outputs bit-exactly."
+      },
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Sqrt_Float",
+        "notes": "Scalar-path companion: positive subnormal inputs should produce tiny nonzero square roots."
+      },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [1e-45, 1e-40, 1e-39] } }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 1e-27 } }
+    },
+    {
+      "name": "ort_float_2x2",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Sqrt_Float"
+      },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [2, 2], "data": { "kind": "values", "values": [1.0, 4.0, 0.0, 9.0] } }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 2], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "ort_float16_2x2",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Sqrt_Float",
+        "notes": "ORT TestUnaryFloat16 coverage paired with the float Sqrt fixture."
+      },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [2, 2], "data": { "kind": "values", "values": [1.0, 4.0, 0.0, 9.0] } }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 2], "tolerance": 0.00098 } }
+    },
+    {
+      "name": "negative_domain_nonfinite_and_signed_zero",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Sqrt_Float",
+        "notes": "Additional node-level domain edge: negative finite inputs produce NaN while zeros and infinities stay on the boundary."
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [6],
+          "data": { "kind": "values", "values": [-1.0, 0.0, 0.0, 4.0, "Infinity", "NaN"] }
+        }
+      },
+      "outputs": {
+        "y": {
+          "dtype": "float32",
+          "shape": [6],
+          "tolerance": 0,
+          "allowNaN": true,
+          "data": { "kind": "values", "values": ["NaN", 0.0, 0.0, 2.0, "Infinity", "NaN"] }
+        }
+      }
+    },
+    {
+      "name": "large_exact_square_values",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Sqrt_Float",
+        "notes": "Additional node-level edge with exactly representable large square inputs and exact integer roots."
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [5],
+          "data": { "kind": "values", "values": [65536.0, 1048576.0, 16777216.0, 67108864.0, 268435456.0] }
+        }
+      },
+      "outputs": {
+        "y": {
+          "dtype": "float32",
+          "shape": [5],
+          "tolerance": 0,
+          "data": { "kind": "values", "values": [256.0, 1024.0, 4096.0, 8192.0, 16384.0] }
+        }
+      }
+    },
+    {
+      "name": "onnx_backend_example",
+      "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_sqrt_example" },
+      "inputs": { "x": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [1.0, 4.0, 9.0] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "onnx_backend_sqrt",
+      "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_sqrt" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 4, 5],
+          "data": {
+            "kind": "values",
+            "values": [1.764052391052246, 0.40015721321105957, 0.978738009929657, 2.2408931255340576, 1.8675580024719238, 0.9772778749465942, 0.9500884413719177, 0.15135720372200012, 0.10321885347366333, 0.4105985164642334, 0.14404356479644775, 1.4542734622955322, 0.7610377073287964, 0.12167501449584961, 0.44386324286460876, 0.3336743414402008, 1.4940791130065918, 0.2051582634449005, 0.3130677044391632, 0.8540957570075989, 2.5529897212982178, 0.653618574142456, 0.8644362092018127, 0.7421650290489197, 2.269754648208618, 1.4543657302856445, 0.04575851559638977, 0.18718385696411133, 1.5327792167663574, 1.4693588018417358, 0.154947429895401, 0.37816253304481506, 0.8877857327461243, 1.980796456336975, 0.34791216254234314, 0.15634897351264954, 1.2302906513214111, 1.202379822731018, 0.38732680678367615, 0.302302747964859, 1.0485529899597168, 1.420017957687378, 1.7062702178955078, 1.950775384902954, 0.5096521973609924, 0.4380742907524109, 1.2527953386306763, 0.7774903774261475, 1.6138978004455566, 0.21274028718471527, 0.8954665660858154, 0.38690251111984253, 0.5108051300048828, 1.18063223361969, 0.02818222902715206, 0.4283318817615509, 0.06651721894741058, 0.30247190594673157, 0.6343221068382263, 0.3627411723136902]
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3, 4, 5], "tolerance": 0.00001, "allowNaN": true } }
+    },
+    {
+      "name": "onnx_backend_sqrt_example",
+      "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_sqrt_example" },
+      "inputs": { "x": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [1.0, 4.0, 9.0] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 0.00001 } }
+    },
+    {
+      "name": "empty_input_zero_dim",
+      "inputs": { "x": { "dtype": "float32", "shape": [0], "data": { "kind": "values", "values": [] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [0], "tolerance": 0 } }
+    },
+    {
+      "name": "f16_scalar_fallback_negative_and_nan",
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [5], "data": { "kind": "values", "values": [-4.0, -1.0, 0.0, 4.0, "NaN"] } }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [5], "tolerance": 0.001, "allowNaN": true } }
+    },
+    {
+      "name": "f16_scalar_fallback_large_and_infinity",
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [7],
+          "data": { "kind": "values", "values": [65504.0, 100.0, 1.0, 0.0, 0.0, "Infinity", "NaN"] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [7], "tolerance": 1, "allowNaN": true } }
+    }
+  ]
+}

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

	@@ -0,0 +1,26 @@

+{% 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 %}
+{{ env.wgsl.resourceDeclarations }}
+{{ flat_tail_open() }}
+  y[i] = {{ scalar }}(sqrt(f32(x[i])));
+{{ flat_tail_close() -}}
+}

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

	@@ -0,0 +1,19 @@

+// 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] = {{ vectorScalar }}(sqrt(vec4<f32>(xv)));
+}