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 +207 -0
build/webgpu/unary-scalar.wgsl.jinja +53 -0
build/webgpu/unary-vec4.wgsl.jinja +48 -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.Log
+`ai.onnx`  ·  standard ONNX operator  ·  ONNX opset ≥ 13
+## Description
+Computes the natural logarithm of each element of the input tensor, producing an output of the same shape.
+See the [ONNX `Log` spec](https://onnx.ai/onnx/operators/onnx__Log.html) for the reference semantics.
+## Inputs
+| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- | --- |
+| `input` | `x` | `T` | — | — | Values whose natural logarithms are computed elementwise. | required |
+## Outputs
+| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- | --- |
+| `output` | `y` | `T` | same as `input` | same as `input` | The natural log of the input tensor, computed elementwise. | 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.Log", { version: 1 });
+const { y } = await kernel({ x: { data: xData, shape: [] } });
+```

build/webgpu/bench.json ADDED Viewed

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

+{
+  "op": "ai.onnx.Log",
+  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
+  "cases": [
+    {
+      "name": "log-f32-1m",
+      "preset": "smoke",
+      "vars": { "dtype": "float32", "count": 1048576 },
+      "inputs": {
+        "x": { "shape": [1048576], "dtype": "float32", "dist": "uniform", "seed": 910, "scale": 1, "offset": 1 }
+      },
+      "outputs": { "y": { "shape": [1048576], "dtype": "float32" } },
+      "bench": {
+        "primary": true,
+        "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }]
+      }
+    },
+    {
+      "name": "log-f32-8m",
+      "preset": "smoke",
+      "vars": { "dtype": "float32", "count": 8388608 },
+      "inputs": {
+        "x": { "shape": [8388608], "dtype": "float32", "seed": 7011, "dist": "uniform", "scale": 1, "offset": 1 }
+      },
+      "outputs": { "y": { "shape": [8388608], "dtype": "float32" } },
+      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
+    },
+    {
+      "name": "log-f16-8m",
+      "preset": "smoke",
+      "vars": { "dtype": "float16", "count": 8388608 },
+      "inputs": {
+        "x": { "shape": [8388608], "dtype": "float16", "seed": 7012, "dist": "uniform", "scale": 1, "offset": 1 }
+      },
+      "outputs": { "y": { "shape": [8388608], "dtype": "float16" } },
+      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
+    },
+    {
+      "name": "log-f32-odd-scalar-fallback",
+      "preset": "edge",
+      "vars": { "dtype": "float32", "count": 1048577 },
+      "inputs": {
+        "x": { "shape": [1048577], "dtype": "float32", "dist": "uniform", "seed": 7031, "scale": 1, "offset": 1 }
+      },
+      "outputs": { "y": { "shape": [1048577], "dtype": "float32" } },
+      "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": "Log",
+  "sinceVersion": 13,
+  "description": "Computes the natural logarithm of each element of the input tensor, producing an output of the same shape.",
+  "inputs": [
+    { "role": "input", "dtype": "T", "description": "Values whose natural logarithms are computed elementwise." }
+  ],
+  "outputs": [
+    {
+      "role": "output",
+      "dtype": "T",
+      "rank": "ranks.input",
+      "description": "The natural log of the input tensor, computed elementwise.",
+      "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 ~ \">\""
+      },
+      "passes": [
+        {
+          "id": "main",
+          "name": "Log.vec4",
+          "source": { "shader": "unary-vec4.wgsl.jinja", "inputs": { "op": "\"log\"" } },
+          "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": "Log",
+          "source": { "shader": "unary-scalar.wgsl.jinja", "inputs": { "op": "\"log\"", "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.Log",
+  "id": "_ai_onnx_log_webgpu_2cf8591",
+  "version": 1,
+  "license": "Apache-2.0",
+  "backend": { "type": "webgpu" },
+  "digest": {
+    "algorithm": "sha256",
+    "files": {
+      "bench.json": "r7qXaSJXGGynT+ARGoJs6iYyKU+pTj7T7TbCcS1LRLs=",
+      "manifest.json": "8qaXwo8hQlTvqDgxd6nCxHGAf8k11eDjoUNfKRvG164=",
+      "test.json": "B1RS6+XfT6gcxEtTg2ZGEnEkyMawcYaeCBTIuntIGtc=",
+      "unary-scalar.wgsl.jinja": "mfdeF8kKBQgPX6wT3gRXd/fnkMK6WBCOHOO7Y+rNFVY=",
+      "unary-vec4.wgsl.jinja": "1Kt6VZ/JRGUyA4Im6S/4ND8TGN3YtreG3gTtpiyTKEg="
+    }
+  },
+  "provenance": { "kernel": { "sha": "2e7068faf55e7f43df740015f6d1ee49391a41c5", "dirty": false } },
+  "webgpu": { "manifestSpec": "1.0", "specialized": true, "opPath": "ops/ai.onnx.Log" }
+}

build/webgpu/test.json ADDED Viewed

	@@ -0,0 +1,207 @@

+{
+  "op": "ai.onnx.Log",
+  "cases": [
+    {
+      "name": "vector_positive",
+      "inputs": { "x": { "dtype": "float32", "shape": [32], "data": { "kind": "constant", "value": 2.5 } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [32], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "rank0_scalar_positive",
+      "inputs": { "x": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [4.0] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "domain_edges_zero_negative_inf",
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [6],
+          "data": { "kind": "values", "values": [0.0, -1.0, -5.0, "Infinity", "NaN", 1.0] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [6], "tolerance": 0.000001, "allowNaN": true } }
+    },
+    {
+      "name": "f32_positive_subnormal_inputs_finite_logs_vec4_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.Log",
+        "notes": "Positive subnormal inputs are inside Log's domain and should produce finite large negative values; flushing them to zero produces -Infinity."
+      },
+      "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": 0.000001 } }
+    },
+    {
+      "name": "f32_positive_subnormal_inputs_finite_logs_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.Log",
+        "notes": "Scalar-path companion for positive subnormal Log inputs."
+      },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [1e-45, 1e-40, 1e-38] } }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "f16_positive_values",
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [5], "data": { "kind": "values", "values": [0.25, 0.5, 1.0, 2.0, 8.0] } }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [5], "tolerance": 0.002 } }
+    },
+    {
+      "name": "subnormal_positive_f32",
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [4],
+          "data": { "kind": "values", "values": [1.17549435e-38, 1e-37, 1e-20, 1e-10] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [4], "tolerance": 0.00001 } }
+    },
+    {
+      "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.Log",
+        "notes": "Positive subnormal inputs are valid for Log; flushing them to zero turns finite logs into -Infinity."
+      },
+      "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": 0.00001 } }
+    },
+    {
+      "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.Log",
+        "notes": "Scalar-path companion: positive subnormal inputs are valid and should produce finite logs."
+      },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [1e-45, 1e-40, 1e-39] } }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 0.00001 } }
+    },
+    {
+      "name": "ort_float_2x2",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Log"
+      },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [2, 2], "data": { "kind": "values", "values": [1.0, 2.0, 5.0, 10.0] } }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 2], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "ort_float16_2x2_projection",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Log",
+        "notes": "Float16 projection of ORT's float Log node values."
+      },
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [2, 2], "data": { "kind": "values", "values": [1.0, 2.0, 5.0, 10.0] } }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 2], "tolerance": 0.002 } }
+    },
+    {
+      "name": "signed_zero_negative_infinity_domain",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Log",
+        "notes": "Additional node-level domain boundary: log(+/-0) is -Infinity, negative finite and -Infinity inputs produce NaN, and log(+Infinity) is +Infinity."
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [7],
+          "data": { "kind": "values", "values": [0.0, 0.0, -1.0, "-Infinity", 1.0, "Infinity", "NaN"] }
+        }
+      },
+      "outputs": {
+        "y": {
+          "dtype": "float32",
+          "shape": [7],
+          "tolerance": 0,
+          "allowNaN": true,
+          "data": { "kind": "values", "values": ["-Infinity", "-Infinity", "NaN", "NaN", 0.0, "Infinity", "NaN"] }
+        }
+      }
+    },
+    {
+      "name": "onnx_backend_example",
+      "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_log_example" },
+      "inputs": { "x": { "dtype": "float32", "shape": [2], "data": { "kind": "values", "values": [1.0, 10.0] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [2], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "onnx_backend_log",
+      "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_log" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 4, 5],
+          "data": {
+            "kind": "values",
+            "values": [5.8360395431518555, 1.4920592308044434, 2.6610958576202393, 9.401724815368652, 6.472471237182617, 0.37633413076400757, 2.5859384536743164, 0.859540581703186, 0.901929497718811, 1.507719874382019, 1.1549344062805176, 4.281371593475342, 2.1404964923858643, 1.1293870210647583, 1.5587172508239746, 1.3960884809494019, 4.4552321434021, 0.8145183324813843, 1.3676141500473022, 0.4256679117679596, 0.07784857600927353, 1.9224849939346313, 2.3736672401428223, 0.47608205676078796, 9.677026748657227, 0.23354846239089966, 1.0468215942382812, 0.8292912244796753, 4.6310296058654785, 4.346446990966797, 1.1675965785980225, 1.4596000909805298, 0.4115660488605499, 0.1379593163728714, 0.706160843372345, 1.1692341566085815, 3.4222238063812256, 3.3280277252197266, 0.6788691878318787, 0.7391142249107361, 0.3504444658756256, 0.24170967936515808, 0.18154163658618927, 7.034140110015869, 0.6007044911384583, 0.6452777981758118, 0.285705029964447, 2.176004409790039, 0.1991100162267685, 0.8083661198616028, 0.4084170162677765, 1.4724129438400269, 0.6000123023986816, 0.3070845305919647, 0.9722112417221069, 1.5346952676773071, 1.0687793493270874, 1.3531996011734009, 0.5302948355674744, 0.6957665085792542]
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3, 4, 5], "tolerance": 0.00001, "allowNaN": true } }
+    },
+    {
+      "name": "onnx_backend_log_example",
+      "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_log_example" },
+      "inputs": { "x": { "dtype": "float32", "shape": [2], "data": { "kind": "values", "values": [1.0, 10.0] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [2], "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_domain_edges_scalar_path",
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [5],
+          "data": { "kind": "values", "values": [0.0, -1.0, "-Infinity", "Infinity", 1.0] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [5], "allowNaN": true } },
+      "tolerance": 0.001
+    },
+    {
+      "name": "f16_domain_edges_vec4_path",
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [4], "data": { "kind": "values", "values": [0.0, -2.0, "Infinity", 1.0] } }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [4], "allowNaN": true } },
+      "tolerance": 0.001
+    }
+  ]
+}

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

	@@ -0,0 +1,53 @@

+{% 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 }}
+{% set unaryDomainGuard = device.adapterInfo.architecture == "" or device.adapterInfo.architecture == "apple" %}
+{% macro emit_log_safe(guarded) %}
+fn log_safe(x: f32) -> f32 {
+{%- if guarded %}
+  // A no-NaN compilation path can return a finite value for out-of-domain
+  // operands. Preserve IEEE domain behavior from the loaded bits: both
+  // signed zeros map to -Inf; negative finite values, -Inf, and input NaNs map
+  // to a quiet NaN.
+  let bits = bitcast<u32>(x);
+  let magnitude = bits & 0x7fffffffu;
+  if (magnitude == 0u) {
+    return negative_infinity();
+  }
+  let exponent_all = (magnitude & 0x7f800000u) == 0x7f800000u;
+  let input_nan = exponent_all && (magnitude & 0x007fffffu) != 0u;
+  if ((bits & 0x80000000u) != 0u || input_nan) {
+    return bitcast<f32>(bits | 0x7fc00000u);
+  }
+{%- endif %}
+  return log(x);
+}{% endmacro %}
+{% if unaryDomainGuard %}fn negative_infinity() -> f32 {
+  var bits = 0xff800000u;
+  return bitcast<f32>(bits);
+}
+{% endif %}
+{{ emit_log_safe(unaryDomainGuard) }}
+{{ flat_tail_open() }}
+  y[i] = {{ scalar }}(log_safe(f32(x[i])));
+{{ flat_tail_close() -}}
+}

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

	@@ -0,0 +1,48 @@

+// 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 }}
+{% set unaryDomainGuard = device.adapterInfo.architecture == "" or device.adapterInfo.architecture == "apple" %}
+{% macro emit_log_safe(guarded) %}
+fn log_safe(x: f32) -> f32 {
+{%- if guarded %}
+  // A no-NaN compilation path can return a finite value for out-of-domain
+  // operands. Preserve IEEE domain behavior from the loaded bits: both
+  // signed zeros map to -Inf; negative finite values, -Inf, and input NaNs map
+  // to a quiet NaN.
+  let bits = bitcast<u32>(x);
+  let magnitude = bits & 0x7fffffffu;
+  if (magnitude == 0u) {
+    return negative_infinity();
+  }
+  let exponent_all = (magnitude & 0x7f800000u) == 0x7f800000u;
+  let input_nan = exponent_all && (magnitude & 0x007fffffu) != 0u;
+  if ((bits & 0x80000000u) != 0u || input_nan) {
+    return bitcast<f32>(bits | 0x7fc00000u);
+  }
+{%- endif %}
+  return log(x);
+}{% endmacro %}
+{% if unaryDomainGuard %}fn negative_infinity() -> f32 {
+  var bits = 0xff800000u;
+  return bitcast<f32>(bits);
+}
+{% endif %}
+{{ emit_log_safe(unaryDomainGuard) }}
+@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>(
+    log_safe(fv.x), log_safe(fv.y), log_safe(fv.z), log_safe(fv.w)));
+}