sync 2e7068faf55e

Browse files

Files changed (14) hide show

README.md +86 -0
build/webgpu/bench.json +155 -0
build/webgpu/manifest.json +1085 -0
build/webgpu/metadata.json +26 -0
build/webgpu/reduce-axis-split-reduce.wgsl.jinja +49 -0
build/webgpu/reduce-axis0-splitk-combine.wgsl.jinja +32 -0
build/webgpu/reduce-axis0-splitk-reduce.wgsl.jinja +44 -0
build/webgpu/reduce-axis0-tilecols.wgsl.jinja +103 -0
build/webgpu/reduce-flat-partial.wgsl.jinja +60 -0
build/webgpu/reduce-noop-empty-axes.wgsl.jinja +13 -0
build/webgpu/reduce-row-subgroup.wgsl.jinja +69 -0
build/webgpu/reduce-row-tree.wgsl.jinja +89 -0
build/webgpu/reduce-serial-axis.wgsl.jinja +154 -0
build/webgpu/test.json +1346 -0

README.md CHANGED Viewed

@@ -1,3 +1,89 @@
 ---
 license: apache-2.0
 ---

 ---
+library_name: kernels
 license: apache-2.0
+tags:
+- kernel
+- webgpu
+- wgsl
 ---
+# ai.onnx.ReduceMean
+`ai.onnx`  ·  standard ONNX operator  ·  ONNX opset ≥ 18
+## Description
+Computes the mean of the input tensor's elements along the specified axes. The output rank matches the input if `keepdims` is 1; otherwise the reduced dimensions are pruned. When `axes` is empty, behavior is controlled by `noop_with_empty_axes`: reduce over all axes (default) or act as a no-op.
+See the [ONNX `ReduceMean` spec](https://onnx.ai/onnx/operators/onnx__ReduceMean.html) for the reference semantics.
+## Inputs
+| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- | --- |
+| `data` | `x` | `T` | — | — | Input tensor to reduce. | required |
+## Outputs
+| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- | --- |
+| `reduced` | `y` | `T` | derived | — | Reduced output tensor with mean values along the specified axes. | required |
+## Attributes
+Default values (overridable per request):
+| Attribute | Default | Description |
+| --- | --- | --- |
+| `keepdims` | `1` | If 1, retains the reduced dimensions with size 1 in the output shape; if 0, the reduced dimensions are removed. |
+| `noop_with_empty_axes` | `0` | When `axes` is empty, `0` (the default) reduces over all axes and `1` skips reduction as a no-op. |
+| `axes` | `[]` | Values of the optional ONNX `axes` tensor input, supplied through this request attribute; an empty list follows `noop_with_empty_axes`. |
+## Type constraints
+| Variable | Allowed dtypes |
+| --- | --- |
+| `T` | `float32`, `float16`, `int32` |
+## Device requirements
+Some implementation variants require `subgroups`. These are route-specific capabilities, not package-wide requirements; availability also depends on the request shape and dtype.
+## 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
+- [`reduce-axis-split-reduce.wgsl.jinja`](build/webgpu/reduce-axis-split-reduce.wgsl.jinja)
+- [`reduce-axis0-splitk-combine.wgsl.jinja`](build/webgpu/reduce-axis0-splitk-combine.wgsl.jinja)
+- [`reduce-axis0-splitk-reduce.wgsl.jinja`](build/webgpu/reduce-axis0-splitk-reduce.wgsl.jinja)
+- [`reduce-axis0-tilecols.wgsl.jinja`](build/webgpu/reduce-axis0-tilecols.wgsl.jinja)
+- [`reduce-flat-partial.wgsl.jinja`](build/webgpu/reduce-flat-partial.wgsl.jinja)
+- [`reduce-noop-empty-axes.wgsl.jinja`](build/webgpu/reduce-noop-empty-axes.wgsl.jinja)
+- [`reduce-row-subgroup.wgsl.jinja`](build/webgpu/reduce-row-subgroup.wgsl.jinja)
+- [`reduce-row-tree.wgsl.jinja`](build/webgpu/reduce-row-tree.wgsl.jinja)
+- [`reduce-serial-axis.wgsl.jinja`](build/webgpu/reduce-serial-axis.wgsl.jinja)
+## Use with `@huggingface/kernels`
+The loader automatically allocates outputs whose metadata it can derive from the manifest contract and this call.
+The explicit `outputs` entries provide shape and logical dtype metadata for the results listed below:
+- `y`
+Each entry either requests an optional result or supplies metadata that cannot be inferred from the inputs.
+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.ReduceMean", { version: 1 });
+// Explicit destinations request optional results or supply metadata that cannot be inferred.
+const { y } = await kernel({ x: { data: xData, shape: [] } }, {
+  outputs: { y: { shape: [], dtype: "float32" } },
+});
+```

build/webgpu/bench.json ADDED Viewed

	@@ -0,0 +1,155 @@

+{
+  "op": "ai.onnx.ReduceMean",
+  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256], "MIDDLE_AXIS_TILE_COLS": [16, 32, 64] },
+  "cases": [
+    {
+      "name": "reduce-mean-axis0-1024x512",
+      "preset": "smoke",
+      "vars": { "dtype": "float32", "rows": 1024, "cols": 512 },
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": { "x": { "shape": [1024, 512], "dtype": "float32", "dist": "normal", "seed": 881, "scale": 1 } },
+      "outputs": { "y": { "shape": [512], "dtype": "float32" } },
+      "bench": {
+        "primary": true,
+        "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype)" }]
+      }
+    },
+    {
+      "name": "reduce-mean-axis1-1024x4096",
+      "preset": "smoke",
+      "vars": { "dtype": "float32", "rows": 1024, "cols": 4096 },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": { "x": { "shape": [1024, 4096], "dtype": "float32", "dist": "normal", "seed": 880, "scale": 1 } },
+      "outputs": { "y": { "shape": [1024], "dtype": "float32" } },
+      "bench": {
+        "primary": true,
+        "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype)" }]
+      }
+    },
+    {
+      "name": "reducemean-lastaxis-f32-4096x4096",
+      "preset": "smoke",
+      "vars": { "rows": 4096, "cols": 4096 },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": { "x": { "shape": [4096, 4096], "dtype": "float32", "dist": "normal", "seed": 117, "scale": 0.2 } },
+      "outputs": { "y": { "shape": [4096], "dtype": "float32" } },
+      "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4" }] }
+    },
+    {
+      "name": "reducemean-rank3-lastaxis-f32-64x64x4096",
+      "preset": "smoke",
+      "vars": { "rows": 4096, "cols": 4096 },
+      "attrs": { "axes": [2], "keepdims": 0 },
+      "inputs": { "x": { "shape": [64, 64, 4096], "dtype": "float32", "dist": "normal", "seed": 218, "scale": 0.2 } },
+      "outputs": { "y": { "shape": [64, 64], "dtype": "float32" } },
+      "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4" }] }
+    },
+    {
+      "name": "reducemean-rank3-lastaxis-smallcols-f32-1024x32x16",
+      "preset": "smoke",
+      "vars": { "rows": 32768, "cols": 16 },
+      "attrs": { "axes": [2], "keepdims": 0 },
+      "inputs": { "x": { "shape": [1024, 32, 16], "dtype": "float32", "dist": "normal", "seed": 412, "scale": 0.2 } },
+      "outputs": { "y": { "shape": [1024, 32], "dtype": "float32" } },
+      "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4" }] }
+    },
+    {
+      "name": "axis0-large-65536x512",
+      "preset": "smoke",
+      "vars": { "rows": 65536, "cols": 512 },
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": { "x": { "shape": [65536, 512], "dtype": "float32", "dist": "normal", "scale": 0.2, "seed": 311 } },
+      "outputs": { "y": { "shape": [512], "dtype": "float32" } },
+      "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4" }] }
+    },
+    {
+      "name": "axis0-tall-16384x1024",
+      "preset": "smoke",
+      "vars": { "rows": 16384, "cols": 1024 },
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": { "x": { "shape": [16384, 1024], "dtype": "float32", "dist": "normal", "scale": 0.2, "seed": 312 } },
+      "outputs": { "y": { "shape": [1024], "dtype": "float32" } },
+      "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4" }] }
+    },
+    {
+      "name": "reducemean-rank4-axis1-serial-32x256x28x28",
+      "preset": "smoke",
+      "vars": { "dtype": "float32" },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": { "shape": [32, 256, 28, 28], "dtype": "float32", "dist": "normal", "scale": 0.2, "seed": 5511 }
+      },
+      "outputs": { "y": { "shape": [32, 28, 28], "dtype": "float32" } },
+      "bench": { "metrics": [{ "type": "bandwidth", "value": "32 * 256 * 28 * 28 * dtypeBytes(args.dtype)" }] }
+    },
+    {
+      "name": "reducemean-rank3-axis1-serial-8x1024x768",
+      "preset": "smoke",
+      "vars": { "dtype": "float32" },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": { "x": { "shape": [8, 1024, 768], "dtype": "float32", "dist": "normal", "scale": 0.2, "seed": 5512 } },
+      "outputs": { "y": { "shape": [8, 768], "dtype": "float32" } },
+      "bench": { "metrics": [{ "type": "bandwidth", "value": "8 * 1024 * 768 * dtypeBytes(args.dtype)" }] }
+    },
+    {
+      "name": "reducemean-lastaxis-scalar-subgroup-cols-unaligned-4096x4095",
+      "preset": "smoke",
+      "vars": { "dtype": "float32", "rows": 4096, "cols": 4095 },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": { "x": { "shape": [4096, 4095], "dtype": "float32", "dist": "normal", "scale": 0.2, "seed": 5513 } },
+      "outputs": { "y": { "shape": [4096], "dtype": "float32" } },
+      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * dtypeBytes(args.dtype)" }] }
+    },
+    {
+      "name": "reducemean-fullreduce-serial-numel-not-mul4-r3",
+      "preset": "stress",
+      "attrs": { "keepdims": 0 },
+      "inputs": { "x": { "shape": [1001, 1001, 1], "dtype": "float32", "dist": "normal", "seed": 601, "scale": 1 } },
+      "outputs": { "y": { "shape": [], "dtype": "float32" } },
+      "bench": { "metrics": [{ "type": "bandwidth", "value": "1001 * 1001 * 1 * 4" }] }
+    },
+    {
+      "name": "reducemean-int32-lastaxis-serial-4096x1024",
+      "preset": "stress",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": { "shape": [4096, 1024], "dtype": "int32", "dist": "randint", "seed": 602, "min": -1000, "max": 1000 }
+      },
+      "outputs": { "y": { "shape": [4096], "dtype": "int32" } },
+      "bench": { "metrics": [{ "type": "bandwidth", "value": "4096 * 1024 * 4" }] }
+    },
+    {
+      "name": "reducemean-axis0-f32-131072x15-narrow-splitk-boundary",
+      "preset": "stress",
+      "vars": { "rows": 131072, "cols": 15 },
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": { "x": { "shape": [131072, 15], "dtype": "float32", "dist": "normal", "seed": 603, "scale": 0.2 } },
+      "outputs": { "y": { "shape": [15], "dtype": "float32", "dist": "empty" } },
+      "bench": { "primary": true, "metrics": [{ "type": "bandwidth", "value": "args.rows * args.cols * 4" }] }
+    },
+    {
+      "name": "reducemean-spatial-axes23-f32-2x64x256x256-globalpool-pattern",
+      "preset": "stress",
+      "provenance": {
+        "source": "authored for variant coverage",
+        "notes": "Rank-4 contiguous-suffix reduction that verifies the vec4 subgroup reducer and portable workgroup-tree fallback."
+      },
+      "vars": { "batch": 2, "channels": 64, "height": 256, "width": 256 },
+      "attrs": { "axes": [2, 3], "keepdims": 1 },
+      "inputs": { "x": { "shape": [2, 64, 256, 256], "dtype": "float32", "dist": "normal", "seed": 909, "scale": 0.2 } },
+      "outputs": { "y": { "shape": [2, 64, 1, 1], "dtype": "float32", "dist": "empty" } },
+      "bench": {
+        "metrics": [{ "type": "bandwidth", "value": "args.batch * args.channels * args.height * args.width * 4" }]
+      }
+    },
+    {
+      "name": "reducemean-rank4-axis1-largeaxis-2x131072x1x8",
+      "preset": "smoke",
+      "vars": { "numel": 2097152 },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": { "x": { "shape": [2, 131072, 1, 8], "dtype": "float32", "dist": "normal", "seed": 120, "scale": 0.2 } },
+      "outputs": { "y": { "shape": [2, 1, 8], "dtype": "float32" } },
+      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.numel * 4" }] }
+    }
+  ]
+}

build/webgpu/manifest.json ADDED Viewed

	@@ -0,0 +1,1085 @@

+{
+  "domain": "ai.onnx",
+  "name": "ReduceMean",
+  "sinceVersion": 18,
+  "description": "Computes the mean of the input tensor's elements along the specified axes. The output rank matches the input if `keepdims` is 1; otherwise the reduced dimensions are pruned. When `axes` is empty, behavior is controlled by `noop_with_empty_axes`: reduce over all axes (default) or act as a no-op.",
+  "inputs": [{ "role": "data", "dtype": "T", "description": "Input tensor to reduce." }],
+  "outputs": [
+    {
+      "role": "reduced",
+      "dtype": "T",
+      "rank": "ranks.data if attrs.keepdims == 1 or ((attrs.axes | length) == 0 and attrs.noop_with_empty_axes == 1) else (ranks.data - (attrs.axes | length) if (attrs.axes | length) > 0 else 0)",
+      "description": "Reduced output tensor with mean values along the specified axes."
+    }
+  ],
+  "attributes": { "keepdims": 1, "noop_with_empty_axes": 0, "axes": [] },
+  "attributeDescriptions": {
+    "keepdims": "If 1, retains the reduced dimensions with size 1 in the output shape; if 0, the reduced dimensions are removed.",
+    "noop_with_empty_axes": "When `axes` is empty, `0` (the default) reduces over all axes and `1` skips reduction as a no-op.",
+    "axes": "Values of the optional ONNX `axes` tensor input, supplied through this request attribute; an empty list follows `noop_with_empty_axes`."
+  },
+  "attributeConstraints": { "keepdims": { "values": [0, 1] }, "noop_with_empty_axes": { "values": [0, 1] } },
+  "typeConstraints": { "T": ["float32", "float16", "int32"] },
+  "args": {
+    "x": { "kind": "tensor", "semantic": "data", "role": "input" },
+    "y": { "kind": "tensor", "semantic": "reduced", "role": "output" }
+  },
+  "derive": {
+    "deviceWorkgroupCap": "min(device.limits.maxComputeInvocationsPerWorkgroup, device.limits.maxComputeWorkgroupSizeX)",
+    "wave32Adapter": "has(device.adapterInfo, \"subgroupMinSize\") and has(device.adapterInfo, \"subgroupMaxSize\") and device.adapterInfo.subgroupMinSize == 32 and device.adapterInfo.subgroupMaxSize == 32",
+    "reportedNonWave32Adapter": "not wave32Adapter and (has(device.adapterInfo, \"subgroupMinSize\") or has(device.adapterInfo, \"subgroupMaxSize\"))",
+    "reduceWorkgroupSize": "min(tunables.WORKGROUP_SIZE, deviceWorkgroupCap)",
+    "treeWorkgroupOk": "reduceWorkgroupSize > 0 and pow2ceil(reduceWorkgroupSize) == reduceWorkgroupSize and reduceWorkgroupSize * dtypeBytes(\"float32\") <= device.limits.maxComputeWorkgroupStorageSize",
+    "subgroupWorkgroupFloor": "min(reduceWorkgroupSize, max(1, device.adapterInfo.subgroupMaxSize))",
+    "lastAxisRows": "rows(shapes.data, ranks.data - 1) if ranks.data > 0 else 1",
+    "lastAxisCols": "dim(shapes.data, ranks.data - 1) if ranks.data > 0 else 1",
+    "rowSerialPreferred": "lastAxisRows >= tunables.ROW_SERIAL_MIN_ROWS and lastAxisCols <= tunables.ROW_SERIAL_MAX_COLS",
+    "axis0Rows": "dim(shapes.data, 0) if ranks.data >= 2 else 0",
+    "axis0Cols": "dim(shapes.data, 1) if ranks.data >= 2 else 0",
+    "axis0SplitCount": "min(tunables.AXIS0_MAX_SPLITS, pow2ceil(ceilDiv(axis0Rows, tunables.AXIS0_SPLIT_TARGET_ROWS)))",
+    "axis0SplitScratchBytes": "axis0SplitCount * axis0Cols * dtypeBytes(\"float32\")",
+    "axis0SplitPathFits": "axis0SplitCount <= device.limits.maxComputeWorkgroupsPerDimension and ceilDiv(ceilDiv(axis0Cols, reduceWorkgroupSize), device.limits.maxComputeWorkgroupsPerDimension) <= device.limits.maxComputeWorkgroupsPerDimension and axis0SplitScratchBytes <= device.limits.maxStorageBufferBindingSize and axis0SplitScratchBytes <= device.limits.maxBufferSize",
+    "reduceAxis": "(attrs.axes[0] + ranks.data if attrs.axes[0] < 0 else attrs.axes[0]) if ((attrs.axes | length) == 1 and isUniqueIntList(attrs.axes, 0 - ranks.data, ranks.data, 1)) else ranks.data",
+    "axisSplitDim": "dim(shapes.data, reduceAxis) if ranks.data >= 2 and reduceAxis < ranks.data else 0",
+    "axisSplitInner": "inner(shapes.data, reduceAxis) if ranks.data >= 2 and reduceAxis < ranks.data else 1",
+    "axisSplitOutputs": "numel(shapes.reduced)",
+    "axisSplitCount": "min(tunables.AXIS0_MAX_SPLITS, pow2ceil(ceilDiv(axisSplitDim, tunables.AXIS0_SPLIT_TARGET_ROWS)))",
+    "axisSplitScratchBytes": "axisSplitCount * axisSplitOutputs * 4",
+    "axisSplitPathFits": "axisSplitCount <= device.limits.maxComputeWorkgroupsPerDimension and ceilDiv(ceilDiv(axisSplitOutputs, reduceWorkgroupSize), device.limits.maxComputeWorkgroupsPerDimension) <= device.limits.maxComputeWorkgroupsPerDimension and axisSplitScratchBytes <= device.limits.maxStorageBufferBindingSize and axisSplitScratchBytes <= device.limits.maxBufferSize",
+    "axis0TilePathFits": "treeWorkgroupOk and tunables.AXIS0_TILE_COLS > 0 and tunables.AXIS0_TILE_COLS <= reduceWorkgroupSize and reduceWorkgroupSize % tunables.AXIS0_TILE_COLS == 0",
+    "flatItems": "floor(numel(shapes.data) / tunables.VECTOR_WIDTH)",
+    "flatSplitCount": "max(1, min(tunables.FULL_REDUCE_MAX_SPLITS, ceilDiv(flatItems, reduceWorkgroupSize)))",
+    "flatScratchBytes": "flatSplitCount * dtypeBytes(\"float32\")",
+    "flatPathFits": "treeWorkgroupOk and flatSplitCount <= device.limits.maxComputeWorkgroupsPerDimension and flatScratchBytes <= device.limits.maxStorageBufferBindingSize and flatScratchBytes <= device.limits.maxBufferSize",
+    "flatParallelCovered": "(dtypes.T == \"f32\" or dtypes.T == \"f16\") and f16Ok(dtypes.T) and numel(shapes.reduced) == 1 and numel(shapes.data) >= tunables.FULL_REDUCE_MIN_ELEMENTS and flatPathFits",
+    "contiguousSuffixParallelCovered": "(dtypes.T == \"f32\" or dtypes.T == \"f16\") and f16Ok(dtypes.T) and numel(shapes.reduced) > 0 and numel(shapes.data) % numel(shapes.reduced) == 0 and numel(shapes.data) / numel(shapes.reduced) >= tunables.CONTIGUOUS_SUFFIX_MIN_COLS and ((ranks.data == 3 and hasAxis(attrs.axes, 0, 3) == false and hasAxis(attrs.axes, 1, 3) and hasAxis(attrs.axes, 2, 3) and numel(shapes.reduced) == dim(shapes.data, 0)) or (ranks.data == 4 and hasAxis(attrs.axes, 0, 4) == false and hasAxis(attrs.axes, 1, 4) == false and hasAxis(attrs.axes, 2, 4) and hasAxis(attrs.axes, 3, 4) and numel(shapes.reduced) == dim(shapes.data, 0) * dim(shapes.data, 1)) or (ranks.data == 4 and hasAxis(attrs.axes, 0, 4) == false and hasAxis(attrs.axes, 1, 4) and hasAxis(attrs.axes, 2, 4) and hasAxis(attrs.axes, 3, 4) and numel(shapes.reduced) == dim(shapes.data, 0)))"
+  },
+  "tunables": {
+    "WORKGROUP_SIZE": 256,
+    "VECTOR_WIDTH": 4,
+    "ROW_PARALLEL_MIN_COLS": 64,
+    "SUBGROUP_MIN_COLS": 256,
+    "SUBGROUP_SMALL_ROW_LIMIT": 32768,
+    "AXIS0_SPLIT_MIN_ROWS": 8192,
+    "AXIS0_SPLIT_TARGET_ROWS": 256,
+    "AXIS0_MAX_SPLITS": 128,
+    "AXIS0_TILE_MIN_ROWS": 64,
+    "AXIS0_TILE_MIN_COLS": 16,
+    "AXIS0_TILE_COLS": 16,
+    "AXIS_SPLIT_TILE_COLS": 8,
+    "MIDDLE_AXIS_TILE_COLS": 16,
+    "FULL_REDUCE_MIN_ELEMENTS": 8192,
+    "FULL_REDUCE_MAX_SPLITS": 256,
+    "CONTIGUOUS_SUFFIX_MIN_COLS": 256,
+    "ROW_SERIAL_MIN_ROWS": 8192,
+    "ROW_SERIAL_MAX_COLS": 1024
+  },
+  "bindingSets": {
+    "elementwise": [
+      { "name": "x", "arg": "x", "semantic": "data", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.reduced)" }] }
+      }
+    ],
+    "lastAxisVec4": [
+      {
+        "name": "x",
+        "arg": "x",
+        "semantic": "data",
+        "buffer": { "type": "read-only-storage" },
+        "elementType": "$vectorScalar"
+      },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "rows", "type": "u32", "value": "rows(shapes.data, ranks.data - 1)" },
+            { "name": "cols", "type": "u32", "value": "dim(shapes.data, ranks.data - 1)" },
+            { "name": "chunkCount", "type": "u32", "value": "dim(shapes.data, ranks.data - 1) / tunables.VECTOR_WIDTH" }
+          ]
+        }
+      }
+    ],
+    "suffixVec4": [
+      {
+        "name": "x",
+        "arg": "x",
+        "semantic": "data",
+        "buffer": { "type": "read-only-storage" },
+        "elementType": "$vectorScalar"
+      },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "rows", "type": "u32", "value": "numel(shapes.reduced)" },
+            { "name": "cols", "type": "u32", "value": "numel(shapes.data) / numel(shapes.reduced)" },
+            {
+              "name": "chunkCount",
+              "type": "u32",
+              "value": "numel(shapes.data) / numel(shapes.reduced) / tunables.VECTOR_WIDTH"
+            }
+          ]
+        }
+      }
+    ],
+    "suffixScalar": [
+      { "name": "x", "arg": "x", "semantic": "data", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "rows", "type": "u32", "value": "numel(shapes.reduced)" },
+            { "name": "cols", "type": "u32", "value": "numel(shapes.data) / numel(shapes.reduced)" }
+          ]
+        }
+      }
+    ],
+    "lastAxisScalar": [
+      { "name": "x", "arg": "x", "semantic": "data", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "rows", "type": "u32", "value": "rows(shapes.data, ranks.data - 1)" },
+            { "name": "cols", "type": "u32", "value": "dim(shapes.data, ranks.data - 1)" }
+          ]
+        }
+      }
+    ],
+    "lastAxisScalarSubgroup": [
+      { "name": "x", "arg": "x", "semantic": "data", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "rows", "type": "u32", "value": "rows(shapes.data, ranks.data - 1)" },
+            { "name": "cols", "type": "u32", "value": "dim(shapes.data, ranks.data - 1)" },
+            { "name": "chunkCount", "type": "u32", "value": "dim(shapes.data, ranks.data - 1)" }
+          ]
+        }
+      }
+    ],
+    "scalar": [
+      { "name": "x", "arg": "x", "semantic": "data", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "rows", "type": "u32", "value": "1" },
+            { "name": "cols", "type": "u32", "value": "1" },
+            { "name": "outCount", "type": "u32", "value": "1" }
+          ]
+        }
+      }
+    ],
+    "rank1Axis0": [
+      { "name": "x", "arg": "x", "semantic": "data", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "rows", "type": "u32", "value": "dim(shapes.data, 0)" },
+            { "name": "cols", "type": "u32", "value": "1" },
+            { "name": "outCount", "type": "u32", "value": "numel(shapes.reduced)" }
+          ]
+        }
+      }
+    ],
+    "rank2Serial": [
+      { "name": "x", "arg": "x", "semantic": "data", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "rows", "type": "u32", "value": "dim(shapes.data, 0)" },
+            { "name": "cols", "type": "u32", "value": "dim(shapes.data, 1)" },
+            { "name": "outCount", "type": "u32", "value": "numel(shapes.reduced)" }
+          ]
+        }
+      }
+    ],
+    "rank2SerialAxis1": [
+      { "name": "x", "arg": "x", "semantic": "data", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "cols", "type": "u32", "value": "dim(shapes.data, 1)" },
+            { "name": "outCount", "type": "u32", "value": "numel(shapes.reduced)" }
+          ]
+        }
+      }
+    ],
+    "axis0Parallel": [
+      { "name": "x", "arg": "x", "semantic": "data", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "rows", "type": "u32", "value": "dim(shapes.data, 0)" },
+            { "name": "cols", "type": "u32", "value": "dim(shapes.data, 1)" }
+          ]
+        }
+      }
+    ],
+    "fullReduceSerial": [
+      { "name": "x", "arg": "x", "semantic": "data", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "rows", "type": "u32", "value": "numel(shapes.data)" },
+            { "name": "cols", "type": "u32", "value": "1" },
+            { "name": "outCount", "type": "u32", "value": "numel(shapes.reduced)" }
+          ]
+        }
+      }
+    ],
+    "axisSplitReduce": [
+      { "name": "x", "arg": "x", "semantic": "data", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
+      { "name": "partials", "semantic": "partials", "buffer": { "type": "storage" }, "elementType": "$partialElement" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "axisDim", "type": "u32", "value": "axisSplitDim" },
+            { "name": "inner", "type": "u32", "value": "axisSplitInner" },
+            { "name": "outputs", "type": "u32", "value": "axisSplitOutputs" }
+          ]
+        }
+      }
+    ],
+    "axisSplitCombine": [
+      {
+        "name": "partials",
+        "semantic": "partials",
+        "buffer": { "type": "read-only-storage" },
+        "elementType": "$partialElement"
+      },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "rows", "type": "u32", "value": "axisSplitDim" },
+            { "name": "cols", "type": "u32", "value": "axisSplitOutputs" }
+          ]
+        }
+      }
+    ],
+    "axis0SplitReduce": [
+      { "name": "x", "arg": "x", "semantic": "data", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
+      { "name": "partials", "semantic": "partials", "buffer": { "type": "storage" }, "elementType": "$partialElement" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "rows", "type": "u32", "value": "dim(shapes.data, 0)" },
+            { "name": "cols", "type": "u32", "value": "dim(shapes.data, 1)" }
+          ]
+        }
+      }
+    ],
+    "axis0SplitCombine": [
+      {
+        "name": "partials",
+        "semantic": "partials",
+        "buffer": { "type": "read-only-storage" },
+        "elementType": "$partialElement"
+      },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "rows", "type": "u32", "value": "dim(shapes.data, 0)" },
+            { "name": "cols", "type": "u32", "value": "dim(shapes.data, 1)" }
+          ]
+        }
+      }
+    ],
+    "rankNAxis": [
+      { "name": "x", "arg": "x", "semantic": "data", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "axisDim", "type": "u32", "value": "axisSplitDim" },
+            { "name": "outCount", "type": "u32", "value": "numel(shapes.reduced)" }
+          ]
+        }
+      }
+    ],
+    "rank3Axis1Tiled": [
+      { "name": "x", "arg": "x", "semantic": "data", "buffer": { "type": "read-only-storage" }, "elementType": "f32" },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "f32" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "rows", "type": "u32", "value": "dim(shapes.data, 1)" },
+            { "name": "cols", "type": "u32", "value": "dim(shapes.data, 2)" },
+            {
+              "name": "colTiles",
+              "type": "u32",
+              "value": "ceilDiv(dim(shapes.data, 2), tunables.MIDDLE_AXIS_TILE_COLS)"
+            }
+          ]
+        }
+      }
+    ],
+    "flatPartialF32": [
+      { "name": "x", "arg": "x", "semantic": "data", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
+      { "name": "partials", "semantic": "partials", "buffer": { "type": "storage" }, "elementType": "f32" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "count4", "type": "u32", "value": "floor(numel(shapes.data) / tunables.VECTOR_WIDTH)" },
+            { "name": "numel", "type": "u32", "value": "numel(shapes.data)" }
+          ]
+        }
+      }
+    ],
+    "flatCombineF32": [
+      { "name": "partials", "semantic": "partials", "buffer": { "type": "read-only-storage" }, "elementType": "f32" },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [
+            { "name": "rows", "type": "u32", "value": "numel(shapes.data)" },
+            { "name": "cols", "type": "u32", "value": "1" }
+          ]
+        }
+      }
+    ],
+    "multiAxis": [
+      { "name": "x", "arg": "x", "semantic": "data", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
+      { "name": "y", "arg": "y", "semantic": "reduced", "buffer": { "type": "storage" }, "elementType": "$T" },
+      {
+        "name": "params",
+        "semantic": "kernel.params",
+        "buffer": { "type": "uniform" },
+        "struct": {
+          "name": "Params",
+          "fields": [{ "name": "outCount", "type": "u32", "value": "numel(shapes.reduced)" }]
+        }
+      }
+    ]
+  },
+  "variants": [
+    {
+      "id": "contiguous_suffix_subgroup_vec4",
+      "priority": 30,
+      "requires": { "features": ["subgroups"] },
+      "when": ["device.wgslLanguageFeatures.has(\"subgroup_id\")", "not flatParallelCovered", "contiguousSuffixParallelCovered", "(numel(shapes.data) / numel(shapes.reduced)) % tunables.VECTOR_WIDTH == 0"],
+      "constants": {
+        "scalar": "dtypes.T",
+        "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"",
+        "workgroupSize": "min(reduceWorkgroupSize, max(subgroupWorkgroupFloor, pow2ceil(ceilDiv(numel(shapes.data) / numel(shapes.reduced), tunables.VECTOR_WIDTH))))"
+      },
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.ContiguousSuffixSubgroupVec4",
+          "source": {
+            "shader": "reduce-row-subgroup.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "vec4": true,
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\""
+            }
+          },
+          "subgroupCollectivesWidth": "portable",
+          "bindings": "suffixVec4",
+          "dispatch": { "workgroups": "numel(shapes.reduced)" }
+        }
+      ]
+    },
+    {
+      "id": "contiguous_suffix_tree_vec4",
+      "priority": 22,
+      "when": ["not flatParallelCovered", "contiguousSuffixParallelCovered", "(numel(shapes.data) / numel(shapes.reduced)) % tunables.VECTOR_WIDTH == 0", "treeWorkgroupOk"],
+      "constants": {
+        "scalar": "dtypes.T",
+        "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"",
+        "workgroupSize": "min(reduceWorkgroupSize, pow2ceil(ceilDiv(numel(shapes.data) / numel(shapes.reduced), tunables.VECTOR_WIDTH)))"
+      },
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.ContiguousSuffixTreeVec4",
+          "source": {
+            "shader": "reduce-row-tree.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "vec4": true,
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\""
+            }
+          },
+          "bindings": "suffixVec4",
+          "dispatch": { "workgroups": "numel(shapes.reduced)" }
+        }
+      ]
+    },
+    {
+      "id": "contiguous_suffix_tree",
+      "priority": 21,
+      "when": ["not flatParallelCovered", "contiguousSuffixParallelCovered", "treeWorkgroupOk"],
+      "constants": {
+        "workgroupSize": "min(reduceWorkgroupSize, pow2ceil(numel(shapes.data) / numel(shapes.reduced)))",
+        "scalar": "dtypes.T"
+      },
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.ContiguousSuffixTree",
+          "source": {
+            "shader": "reduce-row-tree.wgsl.jinja",
+            "inputs": { "op": "\"mean\"", "castF32": "dtypes.T == \"f16\"", "usesF16": "dtypes.T == \"f16\"" }
+          },
+          "bindings": "suffixScalar",
+          "dispatch": { "workgroups": "numel(shapes.reduced)" }
+        }
+      ]
+    },
+    {
+      "id": "noop_empty_axes",
+      "priority": 40,
+      "when": ["dtypes.T == \"f32\"", "attrs.noop_with_empty_axes == 1", "(attrs.axes | length) == 0", "sameShape(shapes.data, shapes.reduced)"],
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.NoopEmptyAxes",
+          "source": { "shader": "reduce-noop-empty-axes.wgsl.jinja" },
+          "bindings": "elementwise",
+          "dispatch": { "threads": "numel(shapes.reduced)", "workgroupSize": "reduceWorkgroupSize" }
+        }
+      ]
+    },
+    {
+      "id": "multi_axis_rank3",
+      "priority": 8,
+      "when": ["not flatParallelCovered", "not contiguousSuffixParallelCovered", "f16Ok(dtypes.T)", "ranks.data == 3", "(attrs.keepdims == 1 and ranks.reduced == 3 and (dim(shapes.reduced, 0) == 1 if hasAxis(attrs.axes, 0, 3) else dim(shapes.reduced, 0) == dim(shapes.data, 0)) and (dim(shapes.reduced, 1) == 1 if hasAxis(attrs.axes, 1, 3) else dim(shapes.reduced, 1) == dim(shapes.data, 1)) and (dim(shapes.reduced, 2) == 1 if hasAxis(attrs.axes, 2, 3) else dim(shapes.reduced, 2) == dim(shapes.data, 2))) or (attrs.keepdims == 0 and ranks.reduced == 1)"],
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.MultiAxisRank3",
+          "source": {
+            "shader": "reduce-serial-axis.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "indexing": "\"multiaxis\"",
+              "rank": 3,
+              "reduce": ["hasAxis(attrs.axes, 0, 3)", "hasAxis(attrs.axes, 1, 3)", "hasAxis(attrs.axes, 2, 3)"],
+              "dataShape": "shapes.data",
+              "outputShape": "shapes.reduced",
+              "outputRank": "ranks.reduced",
+              "keepDims": "attrs.keepdims != 0",
+              "intMode": "dtypes.T == \"i32\"",
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\""
+            }
+          },
+          "bindings": "multiAxis",
+          "dispatch": { "threads": "numel(shapes.reduced)", "workgroupSize": "reduceWorkgroupSize" }
+        }
+      ],
+      "constants": { "scalar": "dtypes.T" }
+    },
+    {
+      "id": "multi_axis_rank4",
+      "priority": 8,
+      "when": ["not flatParallelCovered", "not contiguousSuffixParallelCovered", "f16Ok(dtypes.T)", "ranks.data == 4", "attrs.noop_with_empty_axes == 0", "numel(shapes.reduced) == (1 if hasAxis(attrs.axes, 0, 4) else dim(shapes.data, 0)) * (1 if hasAxis(attrs.axes, 1, 4) else dim(shapes.data, 1)) * (1 if hasAxis(attrs.axes, 2, 4) else dim(shapes.data, 2)) * (1 if hasAxis(attrs.axes, 3, 4) else dim(shapes.data, 3))", "((attrs.keepdims == 1 and ranks.reduced == 4) or (attrs.keepdims == 0 and ranks.reduced < 4))"],
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.MultiAxisRank4",
+          "source": {
+            "shader": "reduce-serial-axis.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "indexing": "\"multiaxis\"",
+              "rank": 4,
+              "reduce": ["hasAxis(attrs.axes, 0, 4)", "hasAxis(attrs.axes, 1, 4)", "hasAxis(attrs.axes, 2, 4)", "hasAxis(attrs.axes, 3, 4)"],
+              "dataShape": "shapes.data",
+              "outputShape": "shapes.reduced",
+              "outputRank": "ranks.reduced",
+              "keepDims": "attrs.keepdims != 0",
+              "intMode": "dtypes.T == \"i32\"",
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\""
+            }
+          },
+          "bindings": "multiAxis",
+          "dispatch": { "threads": "numel(shapes.reduced)", "workgroupSize": "reduceWorkgroupSize" }
+        }
+      ],
+      "constants": { "scalar": "dtypes.T" }
+    },
+    {
+      "id": "tree_last_axis_vec4",
+      "priority": 23,
+      "demoteWhen": ["rowSerialPreferred"],
+      "when": ["not flatParallelCovered", "(dtypes.T == \"f32\" or dtypes.T == \"f16\")", "f16Ok(dtypes.T)", "ranks.data >= 1", "reduceAxis == ranks.data - 1", "numel(shapes.reduced) == rows(shapes.data, ranks.data - 1)", "attrs.noop_with_empty_axes == 0", "lastAxisCols >= tunables.ROW_PARALLEL_MIN_COLS", "lastAxisCols % tunables.VECTOR_WIDTH == 0", "treeWorkgroupOk"],
+      "constants": {
+        "scalar": "dtypes.T",
+        "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"",
+        "workgroupSize": "min(reduceWorkgroupSize, pow2ceil(ceilDiv(lastAxisCols, tunables.VECTOR_WIDTH)))"
+      },
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.TreeRowVec4",
+          "source": {
+            "shader": "reduce-row-tree.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "vec4": true,
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\""
+            }
+          },
+          "bindings": "lastAxisVec4",
+          "dispatch": { "workgroups": "lastAxisRows" }
+        }
+      ]
+    },
+    {
+      "id": "rank0_scalar",
+      "priority": 40,
+      "constants": { "axis": 0, "scalar": "dtypes.T" },
+      "when": ["f16Ok(dtypes.T)", "ranks.data == 0", "ranks.reduced == 0"],
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.Rank0Scalar",
+          "source": {
+            "shader": "reduce-serial-axis.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "indexing": "\"axis2d\"",
+              "intMode": "dtypes.T == \"i32\"",
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\"",
+              "logicalBool": "tensorDtypes.data == \"bool\""
+            }
+          },
+          "bindings": "scalar",
+          "dispatch": { "x": 1 }
+        }
+      ]
+    },
+    {
+      "id": "rank1_axis0",
+      "constants": { "axis": 0, "scalar": "dtypes.T" },
+      "when": ["not flatParallelCovered", "f16Ok(dtypes.T)", "ranks.data == 1", "reduceAxis == 0", "((attrs.keepdims == 0 and ranks.reduced == 0) or (attrs.keepdims == 1 and ranks.reduced == 1 and dim(shapes.reduced, 0) == 1))"],
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.Rank1Axis0",
+          "source": {
+            "shader": "reduce-serial-axis.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "indexing": "\"axis2d\"",
+              "intMode": "dtypes.T == \"i32\"",
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\"",
+              "logicalBool": "tensorDtypes.data == \"bool\""
+            }
+          },
+          "bindings": "rank1Axis0",
+          "dispatch": { "threads": "numel(shapes.reduced)", "workgroupSize": "reduceWorkgroupSize" }
+        }
+      ]
+    },
+    {
+      "id": "axis1_parallel",
+      "priority": 20,
+      "demoteWhen": ["rowSerialPreferred"],
+      "when": ["not flatParallelCovered", "(dtypes.T == \"f32\" or dtypes.T == \"f16\")", "f16Ok(dtypes.T)", "ranks.data >= 2", "reduceAxis == ranks.data - 1", "numel(shapes.reduced) == rows(shapes.data, ranks.data - 1)", "lastAxisCols >= tunables.ROW_PARALLEL_MIN_COLS", "treeWorkgroupOk"],
+      "constants": { "workgroupSize": "min(reduceWorkgroupSize, pow2ceil(dim(shapes.data, ranks.data - 1)))" },
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.Axis1Parallel",
+          "source": {
+            "shader": "reduce-row-tree.wgsl.jinja",
+            "inputs": { "op": "\"mean\"", "castF32": "dtypes.T == \"f16\"", "usesF16": "dtypes.T == \"f16\"" }
+          },
+          "bindings": "lastAxisScalar",
+          "dispatch": { "workgroups": "rows(shapes.data, ranks.data - 1)" }
+        }
+      ]
+    },
+    {
+      "id": "axis_split",
+      "priority": 24,
+      "when": ["not flatParallelCovered", "(dtypes.T == \"f32\" or dtypes.T == \"f16\")", "f16Ok(dtypes.T)", "attrs.noop_with_empty_axes == 0", "ranks.data >= 2", "reduceAxis < ranks.data - 1", "not (ranks.data == 2 and reduceAxis == 0)", "axisSplitDim >= tunables.AXIS0_SPLIT_MIN_ROWS", "axisSplitOutputs >= 1", "axisSplitOutputs <= 4096", "axisSplitOutputs == rows(shapes.data, reduceAxis)", "axisSplitPathFits"],
+      "derive": { "splitCount": "axisSplitCount" },
+      "constants": { "partialElement": "\"f32\"", "workgroupSize": "reduceWorkgroupSize", "split": "splitCount" },
+      "intermediates": [{ "id": "partials", "dtype": "float32", "shape": "[splitCount * axisSplitOutputs]" }],
+      "passes": [
+        {
+          "id": "split_reduce",
+          "name": "ReduceMean.AxisSplitReduce",
+          "source": {
+            "shader": "reduce-axis-split-reduce.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "workgroupSize": "reduceWorkgroupSize",
+              "split": "splitCount",
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\""
+            }
+          },
+          "bindings": "axisSplitReduce",
+          "dispatch": { "threads": "axisSplitOutputs", "workgroupSize": "reduceWorkgroupSize", "y": "splitCount" }
+        },
+        {
+          "id": "combine",
+          "name": "ReduceMean.AxisSplitCombine",
+          "source": {
+            "shader": "reduce-axis0-splitk-combine.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "workgroupSize": "reduceWorkgroupSize",
+              "split": "splitCount",
+              "outputF16": "dtypes.T == \"f16\""
+            }
+          },
+          "bindings": "axisSplitCombine",
+          "dispatch": { "threads": "axisSplitOutputs", "workgroupSize": "reduceWorkgroupSize" }
+        }
+      ]
+    },
+    {
+      "id": "axis_split_tiled_narrow",
+      "priority": 25,
+      "when": ["not flatParallelCovered", "(dtypes.T == \"f32\" or dtypes.T == \"f16\")", "f16Ok(dtypes.T)", "attrs.noop_with_empty_axes == 0", "ranks.data >= 2", "reduceAxis < ranks.data - 1", "axisSplitDim >= tunables.AXIS0_SPLIT_MIN_ROWS", "axisSplitOutputs >= 1", "axisSplitOutputs <= 2 * tunables.AXIS_SPLIT_TILE_COLS", "reduceWorkgroupSize % tunables.AXIS_SPLIT_TILE_COLS == 0", "axisSplitOutputs == rows(shapes.data, reduceAxis)", "axisSplitPathFits"],
+      "derive": { "splitCount": "axisSplitCount" },
+      "constants": {
+        "partialElement": "\"f32\"",
+        "scalar": "dtypes.T",
+        "workgroupSize": "reduceWorkgroupSize",
+        "split": "splitCount",
+        "tileCols": "tunables.AXIS_SPLIT_TILE_COLS"
+      },
+      "intermediates": [{ "id": "partials", "dtype": "float32", "shape": "[splitCount * axisSplitOutputs]" }],
+      "passes": [
+        {
+          "id": "split_reduce",
+          "name": "ReduceMean.AxisSplitTiledReduce",
+          "source": {
+            "shader": "reduce-axis0-tilecols.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "workgroupSize": "reduceWorkgroupSize",
+              "split": "splitCount",
+              "tileCols": "tunables.AXIS_SPLIT_TILE_COLS",
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\""
+            }
+          },
+          "bindings": "axisSplitReduce",
+          "dispatch": { "workgroups": "ceilDiv((axisSplitOutputs), (constants.tileCols))", "y": "splitCount" }
+        },
+        {
+          "id": "combine",
+          "name": "ReduceMean.AxisSplitCombine",
+          "source": {
+            "shader": "reduce-axis0-splitk-combine.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "workgroupSize": "reduceWorkgroupSize",
+              "split": "splitCount",
+              "outputF16": "dtypes.T == \"f16\""
+            }
+          },
+          "bindings": "axisSplitCombine",
+          "dispatch": { "threads": "axisSplitOutputs", "workgroupSize": "reduceWorkgroupSize" }
+        }
+      ]
+    },
+    {
+      "id": "axis0_splitk",
+      "priority": 22,
+      "when": ["not flatParallelCovered", "(dtypes.T == \"f32\" or dtypes.T == \"f16\")", "f16Ok(dtypes.T)", "ranks.data == 2", "reduceAxis == 0", "axis0Rows >= tunables.AXIS0_SPLIT_MIN_ROWS", "dim(shapes.data, 1) > 0", "((attrs.keepdims == 0 and ranks.reduced == 1 and dim(shapes.reduced, 0) == dim(shapes.data, 1)) or (attrs.keepdims == 1 and ranks.reduced == 2 and dim(shapes.reduced, 0) == 1 and dim(shapes.reduced, 1) == dim(shapes.data, 1)))", "axis0SplitPathFits"],
+      "derive": { "splitCount": "axis0SplitCount" },
+      "constants": { "partialElement": "\"f32\"", "workgroupSize": "reduceWorkgroupSize", "split": "splitCount" },
+      "intermediates": [{ "id": "partials", "dtype": "float32", "shape": "[splitCount * dim(shapes.data, 1)]" }],
+      "passes": [
+        {
+          "id": "split_reduce",
+          "name": "ReduceMean.Axis0SplitKReduce",
+          "source": {
+            "shader": "reduce-axis0-splitk-reduce.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "workgroupSize": "reduceWorkgroupSize",
+              "split": "splitCount",
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\""
+            }
+          },
+          "bindings": "axis0SplitReduce",
+          "dispatch": { "threads": "dim(shapes.data, 1)", "workgroupSize": "reduceWorkgroupSize", "y": "splitCount" }
+        },
+        {
+          "id": "combine",
+          "name": "ReduceMean.Axis0SplitKCombine",
+          "source": {
+            "shader": "reduce-axis0-splitk-combine.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "workgroupSize": "reduceWorkgroupSize",
+              "split": "splitCount",
+              "outputF16": "dtypes.T == \"f16\""
+            }
+          },
+          "bindings": "axis0SplitCombine",
+          "dispatch": { "threads": "dim(shapes.data, 1)", "workgroupSize": "reduceWorkgroupSize" }
+        }
+      ]
+    },
+    {
+      "id": "axis0_tilecols",
+      "priority": 20,
+      "when": ["not flatParallelCovered", "f16Ok(dtypes.T)", "ranks.data == 2", "reduceAxis == 0", "axis0Rows >= tunables.AXIS0_TILE_MIN_ROWS", "axis0Cols >= tunables.AXIS0_TILE_MIN_COLS", "((attrs.keepdims == 0 and ranks.reduced == 1 and dim(shapes.reduced, 0) == dim(shapes.data, 1)) or (attrs.keepdims == 1 and ranks.reduced == 2 and dim(shapes.reduced, 0) == 1 and dim(shapes.reduced, 1) == dim(shapes.data, 1)))", "axis0TilePathFits"],
+      "constants": {
+        "scalar": "dtypes.T",
+        "workgroupSize": "reduceWorkgroupSize",
+        "tileCols": "tunables.AXIS0_TILE_COLS"
+      },
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.Axis0TileCols",
+          "source": {
+            "shader": "reduce-axis0-tilecols.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "intMode": "dtypes.T == \"i32\"",
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\""
+            }
+          },
+          "bindings": "axis0Parallel",
+          "dispatch": { "workgroups": "ceilDiv((dim(shapes.data, 1)), (constants.tileCols))" }
+        }
+      ]
+    },
+    {
+      "id": "all_axes_flat",
+      "priority": 31,
+      "constants": { "scalar": "dtypes.T", "workgroupSize": "reduceWorkgroupSize", "split": "flatSplitCount" },
+      "when": ["flatParallelCovered"],
+      "intermediates": [{ "id": "partials", "dtype": "float32", "shape": "[flatSplitCount]" }],
+      "passes": [
+        {
+          "id": "flat_partial",
+          "name": "ReduceMean.AllAxesFlatPartial",
+          "source": {
+            "shader": "reduce-flat-partial.wgsl.jinja",
+            "inputs": { "op": "\"mean\"", "castF32": "dtypes.T == \"f16\"", "usesF16": "dtypes.T == \"f16\"" }
+          },
+          "bindings": "flatPartialF32",
+          "dispatch": { "x": "flatSplitCount" }
+        },
+        {
+          "id": "combine",
+          "name": "ReduceMean.AllAxesFlatCombine",
+          "source": {
+            "shader": "reduce-axis0-splitk-combine.wgsl.jinja",
+            "inputs": { "op": "\"mean\"", "outputF16": "dtypes.T == \"f16\"" }
+          },
+          "bindings": "flatCombineF32",
+          "dispatch": { "x": 1 }
+        }
+      ]
+    },
+    {
+      "id": "rank3_axis1_tiled",
+      "priority": 23,
+      "demoteWhen": ["reportedNonWave32Adapter"],
+      "when": ["not flatParallelCovered", "dtypes.T == \"f32\"", "ranks.data == 3", "reduceAxis == 1", "dim(shapes.data, 1) >= tunables.AXIS0_TILE_MIN_ROWS", "dim(shapes.data, 2) >= tunables.MIDDLE_AXIS_TILE_COLS", "tunables.MIDDLE_AXIS_TILE_COLS > 0", "tunables.MIDDLE_AXIS_TILE_COLS <= reduceWorkgroupSize", "reduceWorkgroupSize % tunables.MIDDLE_AXIS_TILE_COLS == 0", "treeWorkgroupOk", "((attrs.keepdims == 0 and ranks.reduced == 2 and dim(shapes.reduced, 0) == dim(shapes.data, 0) and dim(shapes.reduced, 1) == dim(shapes.data, 2)) or (attrs.keepdims == 1 and ranks.reduced == 3 and dim(shapes.reduced, 0) == dim(shapes.data, 0) and dim(shapes.reduced, 1) == 1 and dim(shapes.reduced, 2) == dim(shapes.data, 2)))"],
+      "constants": { "workgroupSize": "reduceWorkgroupSize", "tileCols": "tunables.MIDDLE_AXIS_TILE_COLS" },
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.Rank3Axis1Tiled",
+          "source": { "shader": "reduce-axis0-tilecols.wgsl.jinja", "inputs": { "op": "\"mean\"", "batched": true } },
+          "bindings": "rank3Axis1Tiled",
+          "dispatch": {
+            "workgroups": "dim(shapes.data, 0) * ceilDiv(dim(shapes.data, 2), tunables.MIDDLE_AXIS_TILE_COLS)"
+          }
+        }
+      ]
+    },
+    {
+      "id": "rankn_single_axis_generic",
+      "priority": 12,
+      "supersededBy": ["axis_split_tiled_narrow", "axis_split", "subgroup_last_axis_vec4", "subgroup_last_axis", "tree_last_axis_vec4", "rank3_axis1_tiled"],
+      "when": ["not flatParallelCovered", "f16Ok(dtypes.T)", "ranks.data >= 3", "attrs.noop_with_empty_axes == 0", "reduceAxis < ranks.data", "numel(shapes.reduced) == rows(shapes.data, reduceAxis)", "((attrs.keepdims == 0 and ranks.reduced == ranks.data - 1) or (attrs.keepdims == 1 and ranks.reduced == ranks.data and dim(shapes.reduced, reduceAxis) == 1))"],
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.RankNSingleAxisGeneric",
+          "source": {
+            "shader": "reduce-serial-axis.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "indexing": "\"rankn\"",
+              "rank": "ranks.data",
+              "axis": "reduceAxis",
+              "dataShape": "shapes.data",
+              "outputShape": "shapes.reduced",
+              "outputRank": "ranks.reduced",
+              "keepDims": "attrs.keepdims != 0",
+              "intMode": "dtypes.T == \"i32\"",
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\""
+            }
+          },
+          "bindings": "rankNAxis",
+          "dispatch": { "threads": "numel(shapes.reduced)", "workgroupSize": "reduceWorkgroupSize" }
+        }
+      ],
+      "constants": { "scalar": "dtypes.T" }
+    },
+    {
+      "id": "subgroup_last_axis_vec4",
+      "priority": 25,
+      "requires": { "features": ["subgroups"] },
+      "when": ["device.wgslLanguageFeatures.has(\"subgroup_id\")", "not flatParallelCovered", "(dtypes.T == \"f32\" or dtypes.T == \"f16\")", "f16Ok(dtypes.T)", "ranks.data >= 1", "reduceAxis == ranks.data - 1", "numel(shapes.reduced) == rows(shapes.data, ranks.data - 1)", "dim(shapes.data, ranks.data - 1) >= 4", "dim(shapes.data, ranks.data - 1) % tunables.VECTOR_WIDTH == 0", "(lastAxisCols >= tunables.SUBGROUP_MIN_COLS or lastAxisRows < tunables.SUBGROUP_SMALL_ROW_LIMIT)", "not rowSerialPreferred"],
+      "constants": {
+        "scalar": "dtypes.T",
+        "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"",
+        "workgroupSize": "min(reduceWorkgroupSize, max(subgroupWorkgroupFloor, pow2ceil(ceilDiv(lastAxisCols, tunables.VECTOR_WIDTH))))"
+      },
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.SubgroupRowVec4",
+          "source": {
+            "shader": "reduce-row-subgroup.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "vec4": true,
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\""
+            }
+          },
+          "subgroupCollectivesWidth": "portable",
+          "bindings": "lastAxisVec4",
+          "dispatch": { "workgroups": "rows(shapes.data, ranks.data - 1)" }
+        }
+      ]
+    },
+    {
+      "id": "subgroup_last_axis",
+      "priority": 24,
+      "requires": { "features": ["subgroups"] },
+      "when": ["device.wgslLanguageFeatures.has(\"subgroup_id\")", "not flatParallelCovered", "(dtypes.T == \"f32\" or dtypes.T == \"f16\")", "f16Ok(dtypes.T)", "ranks.data >= 1", "reduceAxis == ranks.data - 1", "numel(shapes.reduced) == rows(shapes.data, ranks.data - 1)", "dim(shapes.data, ranks.data - 1) > 0", "dim(shapes.data, ranks.data - 1) % tunables.VECTOR_WIDTH != 0", "(lastAxisCols >= tunables.SUBGROUP_MIN_COLS or lastAxisRows < tunables.SUBGROUP_SMALL_ROW_LIMIT)", "not rowSerialPreferred"],
+      "constants": {
+        "scalar": "dtypes.T",
+        "workgroupSize": "min(reduceWorkgroupSize, max(subgroupWorkgroupFloor, pow2ceil(lastAxisCols)))"
+      },
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.SubgroupRow",
+          "source": {
+            "shader": "reduce-row-subgroup.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "vec4": false,
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\""
+            }
+          },
+          "subgroupCollectivesWidth": "portable",
+          "bindings": "lastAxisScalarSubgroup",
+          "dispatch": { "workgroups": "rows(shapes.data, ranks.data - 1)" }
+        }
+      ]
+    },
+    {
+      "id": "axis0",
+      "priority": 0,
+      "supersededBy": ["axis_split_tiled_narrow", "axis0_splitk", "axis0_tilecols"],
+      "constants": { "axis": 0, "scalar": "dtypes.T" },
+      "when": ["not flatParallelCovered", "f16Ok(dtypes.T)", "ranks.data == 2", "reduceAxis == 0", "((attrs.keepdims == 0 and ranks.reduced == 1 and dim(shapes.reduced, 0) == dim(shapes.data, 1)) or (attrs.keepdims == 1 and ranks.reduced == 2 and dim(shapes.reduced, 0) == 1 and dim(shapes.reduced, 1) == dim(shapes.data, 1)))"],
+      "passes": [
+        {
+          "id": "main",
+          "name": "axis0",
+          "source": {
+            "shader": "reduce-serial-axis.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "indexing": "\"axis2d\"",
+              "intMode": "dtypes.T == \"i32\"",
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\""
+            }
+          },
+          "bindings": "rank2Serial",
+          "constants": { "axis": 0 },
+          "dispatch": { "threads": "numel(shapes.reduced)", "workgroupSize": "reduceWorkgroupSize" }
+        }
+      ]
+    },
+    {
+      "id": "axis1",
+      "priority": 0,
+      "constants": { "axis": 1, "scalar": "dtypes.T" },
+      "when": ["not flatParallelCovered", "f16Ok(dtypes.T)", "ranks.data == 2", "reduceAxis == 1", "((attrs.keepdims == 0 and ranks.reduced == 1 and dim(shapes.reduced, 0) == dim(shapes.data, 0)) or (attrs.keepdims == 1 and ranks.reduced == 2 and dim(shapes.reduced, 0) == dim(shapes.data, 0) and dim(shapes.reduced, 1) == 1))"],
+      "passes": [
+        {
+          "id": "main",
+          "name": "axis1",
+          "source": {
+            "shader": "reduce-serial-axis.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "indexing": "\"axis2d\"",
+              "intMode": "dtypes.T == \"i32\"",
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\""
+            }
+          },
+          "bindings": "rank2SerialAxis1",
+          "constants": { "axis": 1 },
+          "dispatch": { "threads": "numel(shapes.reduced)", "workgroupSize": "reduceWorkgroupSize" }
+        }
+      ]
+    },
+    {
+      "id": "all_axes_no_keepdims",
+      "priority": 30,
+      "constants": { "axis": 0, "scalar": "dtypes.T" },
+      "when": ["not flatParallelCovered", "f16Ok(dtypes.T)", "ranks.data >= 3", "attrs.keepdims == 0 and ranks.reduced == 0"],
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.Rank3AllAxesNoKeepdims",
+          "source": {
+            "shader": "reduce-serial-axis.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "indexing": "\"axis2d\"",
+              "intMode": "dtypes.T == \"i32\"",
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\""
+            }
+          },
+          "bindings": "fullReduceSerial",
+          "dispatch": { "threads": "numel(shapes.reduced)", "workgroupSize": "reduceWorkgroupSize" }
+        }
+      ]
+    },
+    {
+      "id": "all_axes_keepdims",
+      "priority": 30,
+      "constants": { "axis": 0, "scalar": "dtypes.T" },
+      "when": ["not flatParallelCovered", "f16Ok(dtypes.T)", "ranks.data >= 3", "attrs.keepdims == 1 and ranks.reduced == ranks.data and numel(shapes.reduced) == 1"],
+      "passes": [
+        {
+          "id": "main",
+          "name": "ReduceMean.Rank3AllAxesKeepdims",
+          "source": {
+            "shader": "reduce-serial-axis.wgsl.jinja",
+            "inputs": {
+              "op": "\"mean\"",
+              "indexing": "\"axis2d\"",
+              "intMode": "dtypes.T == \"i32\"",
+              "castF32": "dtypes.T == \"f16\"",
+              "usesF16": "dtypes.T == \"f16\""
+            }
+          },
+          "bindings": "fullReduceSerial",
+          "dispatch": { "threads": "numel(shapes.reduced)", "workgroupSize": "reduceWorkgroupSize" }
+        }
+      ]
+    }
+  ]
+}

build/webgpu/metadata.json ADDED Viewed

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

+{
+  "name": "ai.onnx.ReduceMean",
+  "id": "_ai_onnx_reducemean_webgpu_01e9079",
+  "version": 1,
+  "license": "Apache-2.0",
+  "backend": { "type": "webgpu" },
+  "digest": {
+    "algorithm": "sha256",
+    "files": {
+      "bench.json": "/K+s8cJOdBn/VX7RsBXRL096sIV729idGwO0eC0jDFY=",
+      "manifest.json": "bJgZfjv4JjiLNg4BxmMQFWswJgqkmUgNaFdwrJGd07A=",
+      "reduce-axis-split-reduce.wgsl.jinja": "Q50AhcbpYvCADMPrbGtPXKbmfT8f6LmQY+6+/CEX224=",
+      "reduce-axis0-splitk-combine.wgsl.jinja": "ybuwLy7He8M/Lkwk+LD6l9m4QI1b2d5pHzDUp/hmDPY=",
+      "reduce-axis0-splitk-reduce.wgsl.jinja": "jc7OBcuREabFLa+kwdZs2oEI3W/BWfHh8nyeUQnIVyY=",
+      "reduce-axis0-tilecols.wgsl.jinja": "mgYY0RTgZxPka4Q7qze7veEiPTC3jL1K7kYZWk/b2f4=",
+      "reduce-flat-partial.wgsl.jinja": "Jz5ZMpVLRJ702gjunfgrUEwqOase9mEBr5SRHlwA0zU=",
+      "reduce-noop-empty-axes.wgsl.jinja": "IO2MEzyBGJr3Z8AVP7ELJXqfhwcG/vtBfa0wCdkbuwo=",
+      "reduce-row-subgroup.wgsl.jinja": "fMRk3pR/DtSfjeRScsPXSY3hG9l13La8Xqyhjj46MyA=",
+      "reduce-row-tree.wgsl.jinja": "/ofIrGSbKtQbrDSfclyWqXfO2lRS2PJaE5Ws8IJqBJ0=",
+      "reduce-serial-axis.wgsl.jinja": "aUda1ztZ4NwVPwicDZz5P1sIhLT/LT9HrQnjpkdDmp4=",
+      "test.json": "RgkWmqzv4eLtCd+VbotifqQ1COTLqkSZ2gJ65sPQQaQ="
+    }
+  },
+  "provenance": { "kernel": { "sha": "2e7068faf55e7f43df740015f6d1ee49391a41c5", "dirty": false } },
+  "webgpu": { "manifestSpec": "1.0", "specialized": true, "opPath": "ops/ai.onnx.ReduceMean" }
+}

build/webgpu/reduce-axis-split-reduce.wgsl.jinja ADDED Viewed

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

+// Pass 1 of a split-K reduction for an (outer, axis, inner) flattening. Each
+// (output, segment) invocation reduces one axis slice at stride inner and writes
+// partials[segment * outputs + output]. Adjacent output threads read adjacent
+// inner-axis elements. The combine pass folds the segments and finalizes the
+// selected reduction.
+//
+// logsumexp writes three partial planes per output and segment: the segment
+// maximum, the sum of exp(x - maximum), and a packed NaN marker.
+{% set castF32 = source.castF32 is defined and source.castF32 %}
+{% set xa = "f32(" if castF32 else "" %}
+{% set ax = ")" if castF32 else "" %}
+{% if source.usesF16 is defined and source.usesF16 %}
+enable f16;
+{% endif %}
+{{ env.wgsl.resourceDeclarations }}
+/* Exact max/min reduction identities. WGSL rejects infinity during constant
+ * evaluation, so f32 identities are constructed at runtime from their
+ * IEEE-754 bit patterns. */
+const WG: u32 = {{ workgroupSize }}u;
+const SPLIT: u32 = {{ split }}u;
+@compute @workgroup_size(WG, 1, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>,
+        @builtin(workgroup_id) wg: vec3<u32>,
+        @builtin(num_workgroups) nwg: vec3<u32>) {
+  // 2D-folded output index: wg.z carries the high bits past the
+  // per-dimension dispatch limit on the x dimension.
+  let output_index = (wg.x + wg.z * nwg.x) * WG + (gid.x % WG);
+  let seg = wg.y;
+  if (output_index >= params.outputs) { return; }
+  let outer_index = output_index / params.inner;
+  let inner_index = output_index % params.inner;
+  let input_base = outer_index * params.axisDim * params.inner + inner_index;
+  // Even split of [0, axisDim) across SPLIT segments (last may be shorter).
+  let chunk = (params.axisDim + SPLIT - 1u) / SPLIT;
+  let a0 = seg * chunk;
+  var a1 = a0 + chunk;
+  if (a1 > params.axisDim) { a1 = params.axisDim; }
+  var acc = 0.0;
+  for (var axis_index = a0; axis_index < a1; axis_index = axis_index + 1u) {
+    acc = acc + {{ xa }}x[input_base + axis_index * params.inner]{{ ax }};
+  }
+  partials[seg * params.outputs + output_index] = acc;
+}

build/webgpu/reduce-axis0-splitk-combine.wgsl.jinja ADDED Viewed

	@@ -0,0 +1,32 @@

+// Pass 2 of the split-K column-wise reduction. One invocation per output column
+// folds the segment partials and applies the selected reduction's final step.
+// Segments are folded in ascending order for deterministic results. This order
+// differs from the single-pass reduction but remains within the f32 tolerance.
+{% set yv = "f16(" if source.outputF16 else "" %}
+{% set vy = ")" if source.outputF16 else "" %}
+{% if source.outputF16 %}
+enable f16;
+{% endif %}
+{{ env.wgsl.resourceDeclarations }}
+/* Exact max/min reduction identities. WGSL rejects infinity during constant
+ * evaluation, so f32 identities are constructed at runtime from their
+ * IEEE-754 bit patterns. */
+const WG: u32 = {{ workgroupSize }}u;
+const SPLIT: u32 = {{ split }}u;
+@compute @workgroup_size(WG, 1, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>,
+        @builtin(num_workgroups) nwg: vec3<u32>) {
+  let stride = nwg.x * WG;
+  let start = (gid.y * nwg.x * WG) + gid.x;
+  for (var col = start; col < params.cols; col = col + stride) {
+    var total = 0.0;
+    for (var seg = 0u; seg < SPLIT; seg = seg + 1u) {
+      let p = partials[seg * params.cols + col];
+      total = total + p;
+    }
+    y[col] = {{ yv }}total / f32(params.rows){{ vy }};
+  }
+}

build/webgpu/reduce-axis0-splitk-reduce.wgsl.jinja ADDED Viewed

	@@ -0,0 +1,44 @@

+// Pass 1 of a split-K column-wise reduction. Splitting rows across workgroup
+// segments increases residency for tall matrices. Each (column, segment)
+// invocation reduces one row slice and writes partials[segment * columns +
+// column]. Adjacent column threads keep row reads coalesced.
+//
+// logsumexp writes the segment maximum and sum of exp(x - maximum) as separate
+// partial planes; the combine pass merges them stably and handles NaN and +Inf.
+{% set castF32 = source.castF32 is defined and source.castF32 %}
+{% set xa = "f32(" if castF32 else "" %}
+{% set ax = ")" if castF32 else "" %}
+{% if source.usesF16 is defined and source.usesF16 %}
+enable f16;
+{% endif %}
+{{ env.wgsl.resourceDeclarations }}
+/* Exact max/min reduction identities. WGSL rejects infinity during constant
+ * evaluation, so f32 identities are constructed at runtime from their
+ * IEEE-754 bit patterns. */
+const WG: u32 = {{ workgroupSize }}u;
+const SPLIT: u32 = {{ split }}u;
+@compute @workgroup_size(WG, 1, 1)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>,
+        @builtin(workgroup_id) wg: vec3<u32>,
+        @builtin(num_workgroups) nwg: vec3<u32>) {
+  // 2D-folded column index: wg.z carries the high bits past the maxComputeWorkgroupsPerDimension
+  // workgroup-per-dimension dispatch limit on the x dimension.
+  let col = (wg.x + wg.z * nwg.x) * WG + (gid.x % WG);
+  let seg = wg.y;
+  if (col >= params.cols) { return; }
+  // Even split of [0, rows) across SPLIT segments (last segment may be shorter).
+  let chunk = (params.rows + SPLIT - 1u) / SPLIT;
+  let r0 = seg * chunk;
+  var r1 = r0 + chunk;
+  if (r1 > params.rows) { r1 = params.rows; }
+  var acc = 0.0;
+  for (var row = r0; row < r1; row = row + 1u) {
+    acc = acc + {{ xa }}x[row * params.cols + col]{{ ax }};
+  }
+  partials[seg * params.cols + col] = acc;
+}

build/webgpu/reduce-axis0-tilecols.wgsl.jinja ADDED Viewed

	@@ -0,0 +1,103 @@

+// Tiled column-wise reduction. Each workgroup owns TILE_COLS columns;
+// ROW_LANES threads stride the rows of one column, then lane 0 folds their
+// partials. In split mode, workgroup y selects an axis segment and finalization
+// is deferred to the combine pass. Cooperative row lanes preserve occupancy
+// when the flattened output has only a few elements.
+{% set splitMode = source.split is defined %}
+{% if splitMode %}
+{% set rowBegin = "row_begin + row_lane" %}
+{% set rowEnd = "row_end" %}
+{% set elem = "x[input_base + row * params.inner]" %}
+{% else %}
+{% set rowBegin = "row_lane" %}
+{% set rowEnd = "params.rows" %}
+{% set elem = "x[inputBase + row * params.cols + col]" %}
+{% endif %}
+{% set castF32 = source.castF32 is defined and source.castF32 %}
+{% set intMode = source.intMode is defined and source.intMode %}
+{% set scalar = "f32" if castF32 else scalar %}
+{% if castF32 %}
+{% set elem = "f32(" ~ elem ~ ")" %}
+{% endif %}
+{% set yv = "f16(" if castF32 else "" %}
+{% set vy = ")" if castF32 else "" %}
+{% if source.usesF16 is defined and source.usesF16 %}
+enable f16;
+{% endif %}
+{{ env.wgsl.resourceDeclarations }}
+/* Exact max/min reduction identities. WGSL rejects infinity during constant
+ * evaluation, so f32 identities are constructed at runtime from their
+ * IEEE-754 bit patterns. */
+const WG: u32 = {{ workgroupSize }}u;
+const TILE_COLS: u32 = {{ tileCols }}u;
+const ROW_LANES: u32 = WG / TILE_COLS;
+{% if splitMode %}
+const SPLIT: u32 = {{ source.split }}u;
+{% endif %}
+var<workgroup> partial: array<{{ scalar if (source.op == "max" or source.op == "min" or intMode) else "f32" }}, WG>;
+@compute @workgroup_size(WG, 1, 1)
+fn main(@builtin(workgroup_id) wg: vec3<u32>, @builtin(local_invocation_id) lid: vec3<u32>{% if not splitMode %},
+        @builtin(num_workgroups) nwg: vec3<u32>{% endif %}) {
+  let tid = lid.x;
+  let col_lane = tid % TILE_COLS;
+  let row_lane = tid / TILE_COLS;
+{% if splitMode %}
+  // Narrow outputs: wg.x covers every column tile, wg.y is the axis segment.
+  let col = wg.x * TILE_COLS + col_lane;
+  let outputIndex = col;
+  let in_bounds = col < params.outputs;
+  let seg = wg.y;
+  let outer_index = col / params.inner;
+  let inner_index = col % params.inner;
+  let input_base = outer_index * params.axisDim * params.inner + inner_index;
+  // Even split of [0, axisDim) across SPLIT segments (last may be shorter).
+  let chunk = (params.axisDim + SPLIT - 1u) / SPLIT;
+  let row_begin = seg * chunk;
+  let row_end = min(row_begin + chunk, params.axisDim);
+{% else %}
+  // 2D-folded tile index: wg.y carries the high bits past the dispatch limit.
+  // The batched form reuses this same coalesced axis-0 reduction for a middle
+  // axis by assigning consecutive tiles to each outer slice.
+  let tile = wg.x + wg.y * nwg.x;
+{% if source.batched is defined and source.batched %}
+  let outer = tile / params.colTiles;
+  let col = (tile % params.colTiles) * TILE_COLS + col_lane;
+  let inputBase = outer * params.rows * params.cols;
+  let outputIndex = outer * params.cols + col;
+{% else %}
+  let col = tile * TILE_COLS + col_lane;
+  let inputBase = 0u;
+  let outputIndex = col;
+{% endif %}
+  let in_bounds = col < params.cols;
+{% endif %}
+  var acc = {% if intMode %}{{ scalar }}(0){% else %}0.0{% endif %};
+  if (in_bounds) {
+    for (var row = {{ rowBegin }}; row < {{ rowEnd }}; row = row + ROW_LANES) {
+      acc = acc + {{ elem }};
+    }
+  }
+  partial[tid] = acc;
+  workgroupBarrier();
+  if (row_lane == 0u && in_bounds) {
+    var total = partial[col_lane];
+    for (var lane = 1u; lane < ROW_LANES; lane = lane + 1u) {
+      total = total + partial[lane * TILE_COLS + col_lane];
+    }
+{% if splitMode %}
+    partials[seg * params.outputs + outputIndex] = total;
+{% else %}
+{% if source.op == "mean" and intMode %}
+    // Integer division truncates toward zero.
+    y[outputIndex] = total / {{ scalar }}(params.rows);
+{% else %}
+    y[outputIndex] = {{ yv }}total / f32(params.rows){{ vy }};
+{% endif %}
+{% endif %}
+  }
+}

build/webgpu/reduce-flat-partial.wgsl.jinja ADDED Viewed

	@@ -0,0 +1,60 @@

+// Pass 1 of a flat reduction to one scalar. Instead of assigning the whole
+// tensor to one invocation, SPLIT workgroups traverse the flat input, accumulate
+// in registers, reduce within each workgroup, and write one partial each. A
+// following combine pass folds the partials and applies the operation finalizer.
+//
+// Scalar f32 bindings keep arbitrary element counts legal. The grid-stride loop
+// manually assembles full vec4 groups from contiguous scalars, and one global
+// thread folds the final zero-to-three scalar elements exactly once.
+{% set castF32 = source.castF32 is defined and source.castF32 %}
+{% set xa = "f32(" if castF32 else "" %}
+{% set ax = ")" if castF32 else "" %}
+{% if source.usesF16 is defined and source.usesF16 %}
+enable f16;
+{% endif %}
+{{ env.wgsl.resourceDeclarations }}
+/* Exact max/min reduction identities. WGSL rejects infinity during constant
+ * evaluation, so f32 identities are constructed at runtime from their
+ * IEEE-754 bit patterns. */
+const WG: u32 = {{ workgroupSize }}u;
+var<workgroup> red: array<{{ "i32" if source.intMode else "f32" }}, WG>;
+@compute @workgroup_size(WG)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>,
+        @builtin(local_invocation_id) lid: vec3<u32>,
+        @builtin(workgroup_id) wg: vec3<u32>,
+        @builtin(num_workgroups) nwg: vec3<u32>) {
+  let tid = lid.x;
+  let gstride = nwg.x * WG;
+  var acc = 0.0;
+  // Grid-stride over the flat vec4 groups (params.count4 = numel / 4, floored).
+  for (var i = gid.x; i < params.count4; i = i + gstride) {
+    let b = 4u * i;
+    let v = vec4<{{ "i32" if source.intMode else "f32" }}>({{ xa }}x[b]{{ ax }}, {{ xa }}x[b + 1u]{{ ax }}, {{ xa }}x[b + 2u]{{ ax }}, {{ xa }}x[b + 3u]{{ ax }});
+    acc = acc + v.x + v.y + v.z + v.w;
+  }
+  // Scalar tail (the 0..3 elements past the last full vec4). One global thread
+  // folds it so it is counted exactly once; the count is tiny so serializing it
+  // is negligible. numel % 4 == 0 makes this loop empty.
+  if (gid.x == 0u) {
+    for (var i = 4u * params.count4; i < params.numel; i = i + 1u) {
+      let s = {{ xa }}x[i]{{ ax }};
+      acc = acc + s;
+    }
+  }
+  red[tid] = acc;
+  workgroupBarrier();
+  var stride: u32 = WG / 2u;
+  loop {
+    if (stride == 0u) { break; }
+    if (tid < stride) {
+      red[tid] = red[tid] + red[tid + stride];
+    }
+    stride = stride / 2u;
+    workgroupBarrier();
+  }
+  if (tid == 0u) {
+    partials[wg.x] = red[0];
+  }
+}

build/webgpu/reduce-noop-empty-axes.wgsl.jinja ADDED Viewed

	@@ -0,0 +1,13 @@

+{{ env.wgsl.resourceDeclarations }}
+@compute @workgroup_size({{ reduceWorkgroupSize }})
+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 * {{ reduceWorkgroupSize }}u;
+  if (i >= params.count) {
+    return;
+  }
+  let v = x[i];
+  y[i] = v;
+}

build/webgpu/reduce-row-subgroup.wgsl.jinja ADDED Viewed

	@@ -0,0 +1,69 @@

+// Subgroup row reduction for a contiguous last axis. One workgroup owns each
+// output row. Threads reduce strided chunks, subgroup leaders deposit their
+// partials in workgroup memory, and thread 0 folds those slots and finalizes.
+//
+// Accumulators are f32 except for integer max/min. Mean, l2, and logsum apply
+// division, square root, or logarithm only at finalization. logsumexp uses
+// max-subtraction, bit-exact NaN detection, NaN propagation, and a +Inf
+// short-circuit. Runtime bitcasts provide max/min infinity identities because
+// WGSL constant evaluation rejects infinite constants. f16 storage is widened
+// before accumulation and narrowed only for the final store.
+{% set castF32 = source.castF32 is defined and source.castF32 %}
+{% set scalar = "f32" if castF32 else scalar %}
+{% set xv = ("vec4<f32>(" if source.vec4 else "f32(") if castF32 else "" %}
+{% set vx = ")" if castF32 else "" %}
+{% set yv = "f16(" if castF32 else "" %}
+{% set vy = ")" if castF32 else "" %}
+enable subgroups;
+{% if source.usesF16 is defined and source.usesF16 %}
+enable f16;
+{% endif %}
+{{ env.wgsl.resourceDeclarations }}
+/* Exact max/min reduction identities. WGSL rejects infinity during constant
+ * evaluation, so f32 identities are constructed at runtime from their
+ * IEEE-754 bit patterns. */
+const WG: u32 = {{ workgroupSize }}u;
+var<workgroup> wgPartial: array<{{ scalar }}, WG>;
+{% macro emit_reduce(name, collective, combine) %}
+fn {{ name }}(value: {{ scalar }}, sgLid: u32, sgId: u32, numSg: u32) -> {{ scalar }} {
+  let sgVal = {{ collective }}(value);
+  if (numSg == 1u) {
+    return sgVal;
+  }
+  if (sgLid == 0u) {
+    wgPartial[sgId] = sgVal;
+  }
+  workgroupBarrier();
+  var total = wgPartial[0];
+  for (var i = 1u; i < numSg; i = i + 1u) {
+    {{ combine }}
+  }
+  workgroupBarrier();
+  return total;
+}
+{%- endmacro %}{{ emit_reduce("reduce_row", "subgroupAdd", "total = total + wgPartial[i];") }}
+@compute @workgroup_size(WG, 1, 1)
+fn main(@builtin(workgroup_id) wg: vec3<u32>,
+        @builtin(num_workgroups) nwg: vec3<u32>,
+        @builtin(local_invocation_id) lid: vec3<u32>,
+        @builtin(subgroup_invocation_id) sgLid: u32,
+        @builtin(subgroup_id) sgId: u32,
+        @builtin(num_subgroups) numSg: u32) {
+  let row = wg.x + wg.y * nwg.x;
+  if (row >= params.rows) {
+    return;
+  }
+  let tid = lid.x;
+  let base = row * params.chunkCount;  let INIT: f32 = 0.0;{% if source.vec4 %}
+  var acc4 = vec4<{{ scalar }}>(INIT);
+  for (var c = tid; c < params.chunkCount; c = c + WG) {
+    let v = {{ xv }}x[base + c]{{ vx }};    acc4 = acc4 + v;  }  let acc = (acc4.x + acc4.y) + (acc4.z + acc4.w);{% else %}
+  var acc = INIT;
+  for (var c = tid; c < params.chunkCount; c = c + WG) {
+    let v = {{ xv }}x[base + c]{{ vx }};    acc = acc + v;  }
+{%- endif %}
+  let total = reduce_row(acc, sgLid, sgId, numSg);
+  if (tid == 0u) {    y[row] = {{ yv }}total / f32(params.cols){{ vy }};  }}

build/webgpu/reduce-row-tree.wgsl.jinja ADDED Viewed

	@@ -0,0 +1,89 @@

+// Portable one-workgroup-per-row reduction for the Reduce value family.
+// Threads stride a contiguous row, accumulate locally, and fold their values
+// through a shared-memory tree without relying on subgroups.
+//
+// Max, min, and product use the native f32/i32/u32 value type. Additive
+// transforms accumulate float inputs in f32 but retain integer accumulation
+// for sum, L1, and sum-of-squares. Mean, L2, log-sum, and log-sum-exp are
+// f32-only because their finalizers divide, take a square root, or take a log.
+// f16 storage widens through f32 for both accumulation and the shared tree,
+// then narrows only at the final store.
+{% set isVec4 = source.vec4 is defined and source.vec4 %}
+{% set rowIsEmpty = "params.chunkCount == 0u" if isVec4 else "params.cols == 0u" %}
+{% set castF32 = source.castF32 is defined and source.castF32 %}
+{% set scalar = "f32" if castF32 else scalar %}
+{% set xv = ("vec4<f32>(" if source.vec4 else "f32(") if castF32 else "" %}
+{% set vx = ")" if castF32 else "" %}
+{% set yv = "f16(" if castF32 else "" %}
+{% set vy = ")" if castF32 else "" %}
+{% if source.usesF16 is defined and source.usesF16 %}
+enable f16;
+{% endif %}
+{{ env.wgsl.resourceDeclarations }}
+/* Exact max/min reduction identities. WGSL rejects infinity during constant
+ * evaluation, so f32 identities are constructed at runtime from their
+ * IEEE-754 bit patterns. */
+const WG: u32 = {{ workgroupSize }}u;
+{% set is_int = scalar == "i32" or scalar == "u32" %}
+{% set intAdditive = is_int and (source.op == "sum" or source.op == "l1" or source.op == "sumsquare") %}
+{% set accType = scalar if source.op == "max" or source.op == "min" or source.op == "prod" or intAdditive else "f32" %}
+fn identity() -> {{ accType }} { return {{ accType }}(0); }
+fn combine(a: {{ accType }}, b: {{ accType }}) -> {{ accType }} {
+  return a + b;
+}
+var<workgroup> partial: array<{{ accType }}, WG>;
+@compute @workgroup_size(WG, 1, 1)
+fn main(@builtin(workgroup_id) wg: vec3<u32>,
+        @builtin(num_workgroups) nwg: vec3<u32>,
+        @builtin(local_invocation_id) lid: vec3<u32>) {
+  let row = wg.x + wg.y * nwg.x;
+  if (row >= params.rows) {
+    return;
+  }
+  let tid = lid.x;
+{% if source.vec4 %}
+  let base = row * params.chunkCount;
+{% else %}
+  let base = row * params.cols;
+{% endif %}
+  if ({{ rowIsEmpty }}) {
+    if (tid == 0u) { y[row] = {{ yv }}0.0{{ vy }}; }
+    return;
+  }
+{% if source.vec4 %}
+  var acc4 = vec4<{{ accType }}>(identity());
+  for (var col = tid; col < params.chunkCount; col = col + WG) {
+    let value = {{ xv }}x[base + col]{{ vx }};
+    acc4 = acc4 + value;
+  }
+  let acc = (acc4.x + acc4.y) + (acc4.z + acc4.w);
+{% else %}
+  var acc = identity();
+  for (var col = tid; col < params.cols; col = col + WG) {
+    let value = {{ xv }}x[base + col]{{ vx }};
+    acc = combine(acc, value);
+  }
+{% endif %}
+  partial[tid] = acc;
+  workgroupBarrier();
+  for (var step = WG >> 1u; step > 0u; step = step >> 1u) {
+    if (tid < step) {
+      partial[tid] = combine(partial[tid], partial[tid + step]);
+    }
+    workgroupBarrier();
+  }
+  if (tid == 0u) {
+    let total = partial[0];
+    y[row] = {{ yv }}total / f32(params.cols){{ vy }};
+  }
+}

build/webgpu/reduce-serial-axis.wgsl.jinja ADDED Viewed

	@@ -0,0 +1,154 @@

+// Serial one-thread-per-output reduction for the no-feature tier. f16 storage
+// is widened before every accumulation and narrowed only for the final store.
+{% set castF32 = source.castF32 is defined and source.castF32 %}
+{% set intMode = source.intMode is defined and source.intMode %}
+{% set yv = "f16(" if castF32 else "" %}
+{% set vy = ")" if castF32 else "" %}
+{% if source.usesF16 is defined and source.usesF16 %}
+enable f16;
+{% endif %}
+{{ env.wgsl.resourceDeclarations }}
+/* Exact max/min reduction identities. WGSL rejects infinity during constant
+ * evaluation, so f32 identities are constructed at runtime from their
+ * IEEE-754 bit patterns. */
+{% if source.indexing == "rankn" %}
+fn input_offset(out_index: u32, reduce_index: u32) -> u32 {
+  var rem = out_index;
+{% for axis in range(source.outputRank) %}
+{% set out_stride = namespace(value=1) %}
+{% for j in range(axis + 1, source.outputRank) %}
+{% set out_stride.value = out_stride.value * source.outputShape[j] %}
+{% endfor %}
+{% set safe_out_stride = 1 if out_stride.value == 0 else out_stride.value %}
+{% if not source.keepDims or axis != source.axis %}
+  let out_coord{{ axis }} = rem / {{ safe_out_stride }}u;
+{% endif %}
+  rem = rem % {{ safe_out_stride }}u;
+{% endfor %}
+{% for axis in range(source.rank) %}
+{% if axis == source.axis %}
+  let coord{{ axis }} = reduce_index;
+{% elif source.keepDims %}
+  let coord{{ axis }} = out_coord{{ axis }};
+{% elif axis < source.axis %}
+  let coord{{ axis }} = out_coord{{ axis }};
+{% else %}
+  let coord{{ axis }} = out_coord{{ axis - 1 }};
+{% endif %}
+{% endfor %}
+{% set src = namespace(value="coord0") %}
+{% for axis in range(1, source.rank) %}
+{% set src.value = "(" ~ src.value ~ " * " ~ source.dataShape[axis] ~ "u + coord" ~ axis ~ ")" %}
+{% endfor %}
+  return {{ src.value }};
+}
+{% endif %}
+{% if source.indexing == "multiaxis" %}
+{% set hasReducedAxis = namespace(value=false) %}
+{% for a in range(source.rank) %}{% if source.reduce[a] %}{% set hasReducedAxis.value = true %}{% endif %}{% endfor %}
+// One thread per output element walks the Cartesian product of the reduced axes,
+// linearized as reduce_linear. Specialized shapes make every input offset a sum
+// of coordinate-times-constant terms.
+fn input_offset(out_index: u32{% if hasReducedAxis.value %}, reduce_linear: u32{% endif %}) -> u32 {
+  var rem = out_index;
+{% for oaxis in range(source.outputRank) %}
+{% set ostride = namespace(value=1) %}
+{% for j in range(oaxis + 1, source.outputRank) %}
+{% set ostride.value = ostride.value * source.outputShape[j] %}
+{% endfor %}
+{% set osafe = 1 if ostride.value == 0 else ostride.value %}
+{% if not source.keepDims or not source.reduce[oaxis] %}
+  let out_coord{{ oaxis }} = rem / {{ osafe }}u;
+{% endif %}
+  rem = rem % {{ osafe }}u;
+{% endfor %}
+{% if hasReducedAxis.value %}
+  var rrem = reduce_linear;
+{% endif %}
+{% for a in range(source.rank) if source.reduce[a] %}
+{% set rstride = namespace(value=1) %}
+{% for b in range(a + 1, source.rank) if source.reduce[b] %}
+{% set rstride.value = rstride.value * source.dataShape[b] %}
+{% endfor %}
+{% set rsafe = 1 if rstride.value == 0 else rstride.value %}
+  let red_coord{{ a }} = rrem / {{ rsafe }}u;
+  rrem = rrem % {{ rsafe }}u;
+{% endfor %}
+{% set oc = namespace(i=0) %}
+{% for a in range(source.rank) %}
+{% if source.reduce[a] %}
+  let coord{{ a }} = red_coord{{ a }};
+{% elif source.keepDims %}
+  let coord{{ a }} = out_coord{{ a }};
+{% else %}
+  let coord{{ a }} = out_coord{{ oc.i }};
+{% set oc.i = oc.i + 1 %}
+{% endif %}
+{% endfor %}
+{% set src = namespace(value="coord0") %}
+{% for a in range(1, source.rank) %}
+{% set src.value = "(" ~ src.value ~ " * " ~ source.dataShape[a] ~ "u + coord" ~ a ~ ")" %}
+{% endfor %}
+  return {{ src.value }};
+}
+{% endif %}
+{% if source.indexing == "multiaxis" %}
+{% set mcount = namespace(value=1) %}
+{% for a in range(source.rank) if source.reduce[a] %}
+{% set mcount.value = mcount.value * source.dataShape[a] %}
+{% endfor %}
+{% set count = mcount.value ~ "u" %}
+{% if hasReducedAxis.value %}
+{% set at = "x[input_offset(i, r)]" %}
+{% else %}
+{% set at = "x[input_offset(i)]" %}
+{% endif %}
+{% elif source.indexing == "rankn" %}
+{% set count = "params.axisDim" %}
+{% set at = "x[input_offset(i, r)]" %}
+{% elif axis == 0 %}
+{% set count = "params.rows" %}
+{% set at = "x[r * params.cols + i]" %}
+{% else %}
+{% set count = "params.cols" %}
+{% set at = "x[i * params.cols + r]" %}
+{% endif %}
+{% if castF32 %}
+{% set at = "f32(" ~ at ~ ")" %}
+{% endif %}
+@compute @workgroup_size({{ reduceWorkgroupSize }})
+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 * {{ reduceWorkgroupSize }}u;
+  if (i >= params.outCount) {
+    return;
+  }
+  if ({{ count }} == 0u) {
+{% if intMode %}
+    y[i] = {{ scalar }}(0);
+{% else %}
+    y[i] = {{ yv }}0.0{{ vy }};
+{% endif %}
+    return;
+  }
+{% if intMode %}
+  // Integer reduction accumulates in the output type. sum/prod/l1/sumsquare stay
+  // in integer arithmetic; l2/logsum widen for sqrt/log and truncate afterward.
+  var acc = {{ scalar }}({{ "1" if source.op == "prod" else "0" }});
+{% else %}
+  var acc = {{ "1.0" if source.op == "prod" else "0.0" }};
+{% endif %}
+  for (var r = 0u; r < {{ count }}; r = r + 1u) {
+    acc = acc + {{ at }};
+  }
+{% if intMode %}
+  y[i] = acc / {{ scalar }}({{ count }});
+{% else %}
+  y[i] = {{ yv }}acc / f32({{ count }}){{ vy }};
+{% endif %}
+}

build/webgpu/test.json ADDED Viewed

	@@ -0,0 +1,1346 @@

+{
+  "op": "ai.onnx.ReduceMean",
+  "fixtureArrays": {
+    "rank3_axis2_last_keepdims_input_x": [1, 2, 3, 4, -1, -2, -3, -4, 0.5, 1.5, 2.5, 3.5, 10, 20, 30, 40, -10, -20, -30, -40, 2, 4, 6, 8]
+  },
+  "cases": [
+    {
+      "name": "contiguous_suffix_axes23_parallel",
+      "attrs": { "axes": [2, 3], "keepdims": 1 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 3, 16, 16],
+          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.07, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 1, 1], "tolerance": 0.00001 } }
+    },
+    {
+      "name": "all_axes_flat_rank1_boundary_8192",
+      "provenance": {
+        "notes": "The parallel full-reduction threshold must supersede the rank1 serial and row-reduction fallbacks."
+      },
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": { "x": { "dtype": "float32", "shape": [8192], "data": { "kind": "constant", "value": 1.0 } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [], "tolerance": 0 } }
+    },
+    {
+      "name": "all_axes_flat_fullreduce_32x32x32_keepdims",
+      "attrs": { "keepdims": 1 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [32, 32, 32],
+          "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.5 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1], "tolerance": 0.01, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "all_axes_flat_fullreduce_32x32x32_keepdims_offset_mean_scale_lock",
+      "provenance": {
+        "notes": "The all_axes_flat route reduces 32768 elements through an f32 partial pass and a single-workgroup combine; the zero-mean sibling produces 4.5e-4 against a 0.01 absolute tolerance (min detectable uniform scale error 22), so the combine's divide by numel is unobservable. Offsetting x about 1.0 makes the full-reduce mean O(1)."
+      },
+      "attrs": { "keepdims": 1 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [32, 32, 32],
+          "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.027, "scale": 0.5, "offset": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1], "tolerance": 0.0001, "relTolerance": 0.00001 } }
+    },
+    {
+      "name": "dispatch_cliff_axis1_rank2",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [16776961, 1], "data": { "kind": "linspace", "start": -1.0, "end": 1.0 } }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [16776961], "tolerance": 0.0001 } }
+    },
+    {
+      "name": "axis0",
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 4],
+          "data": { "kind": "fillFloat32", "sinStep": 0.23, "cosStep": 0.11 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [4], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "axis0_tiled_64x32",
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [64, 32],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [32], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "f32_subnormal_axis0_tilecols_mean_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 in f32; bit-exact subnormal preservation is unattainable on GPU."
+      },
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean",
+        "notes": "Tiled axis-0 companion: the column mean of equal finite subnormal values remains subnormal."
+      },
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": { "x": { "dtype": "float32", "shape": [64, 16], "data": { "kind": "constant", "value": 1e-40 } } },
+      "outputs": {
+        "y": { "dtype": "float32", "shape": [16], "tolerance": 2e-45, "data": { "kind": "constant", "value": 1e-40 } }
+      }
+    },
+    {
+      "name": "axis1",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 4],
+          "data": { "kind": "fillFloat32", "sinStep": 0.23, "cosStep": 0.11 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "f32_axis1_parallel_cancellation_order_gpu_gap",
+      "skipGpu": {
+        "category": "todo",
+        "reason": "The current parallel reduction changes the fixture's required sequential evaluation order, so f32 rounding is not bit-exact. An order-preserving reduction route can implement this behavior."
+      },
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean",
+        "notes": "Mean inherits the same cancellation-order trap as ReduceSum: serial float32 summation yields 0, while the parallel row tree can preserve the small lane terms before division."
+      },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 1024],
+          "data": { "kind": "cycle", "values": [100000000000000000000.0, 1.0, -100000000000000000000.0, 0.0] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1], "tolerance": 0 } }
+    },
+    {
+      "name": "f32_subnormal_axis1_mean_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 in f32; bit-exact subnormal preservation is unattainable on GPU."
+      },
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean",
+        "notes": "A row mean over equal finite subnormal values remains subnormal; reduction kernels must not flush the input or final quotient to zero."
+      },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 3],
+          "data": { "kind": "values", "values": [1e-40, 1e-40, 1e-40, -1e-40, -1e-40, -1e-40] }
+        }
+      },
+      "outputs": {
+        "y": {
+          "dtype": "float32",
+          "shape": [2],
+          "tolerance": 2e-45,
+          "data": { "kind": "values", "values": [1e-40, -1e-40] }
+        }
+      }
+    },
+    {
+      "name": "f32_subnormal_axis0_mean_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 in f32; bit-exact subnormal preservation is unattainable on GPU."
+      },
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean",
+        "notes": "Axis-0 companion: the mean of equal finite subnormal column values remains subnormal."
+      },
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 2],
+          "data": { "kind": "values", "values": [1e-40, -1e-40, 1e-40, -1e-40, 1e-40, -1e-40] }
+        }
+      },
+      "outputs": {
+        "y": {
+          "dtype": "float32",
+          "shape": [2],
+          "tolerance": 2e-45,
+          "data": { "kind": "values", "values": [1e-40, -1e-40] }
+        }
+      }
+    },
+    {
+      "name": "f32_subnormal_last_axis_vec4_mean_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 in f32; bit-exact subnormal preservation is unattainable on GPU."
+      },
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean",
+        "notes": "Vec4 last-axis companion: the mean of equal finite subnormal values should remain subnormal."
+      },
+      "attrs": { "axes": [-1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 4],
+          "data": { "kind": "values", "values": [1e-40, 1e-40, 1e-40, 1e-40, -1e-40, -1e-40, -1e-40, -1e-40] }
+        }
+      },
+      "outputs": {
+        "y": {
+          "dtype": "float32",
+          "shape": [2],
+          "tolerance": 2e-45,
+          "data": { "kind": "values", "values": [1e-40, -1e-40] }
+        }
+      }
+    },
+    {
+      "name": "f32_subnormal_last_axis_odd_mean_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 in f32; bit-exact subnormal preservation is unattainable on GPU."
+      },
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean",
+        "notes": "Odd-width last-axis companion: the mean of equal finite subnormal values should remain subnormal in the non-vec4 subgroup reducer."
+      },
+      "attrs": { "axes": [-1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 3],
+          "data": { "kind": "values", "values": [1e-40, 1e-40, 1e-40, -1e-40, -1e-40, -1e-40] }
+        }
+      },
+      "outputs": {
+        "y": {
+          "dtype": "float32",
+          "shape": [2],
+          "tolerance": 2e-45,
+          "data": { "kind": "values", "values": [1e-40, -1e-40] }
+        }
+      }
+    },
+    {
+      "name": "f32_subnormal_rank3_axis1_mean_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 in f32; bit-exact subnormal preservation is unattainable on GPU."
+      },
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean",
+        "notes": "Rank-3 axis-1 companion: the mean of equal finite subnormal values remains subnormal through middle-axis indexing."
+      },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 3, 2],
+          "data": {
+            "kind": "values",
+            "values": [1e-40, -1e-40, 1e-40, -1e-40, 1e-40, -1e-40, -1e-40, 1e-40, -1e-40, 1e-40, -1e-40, 1e-40]
+          }
+        }
+      },
+      "outputs": {
+        "y": {
+          "dtype": "float32",
+          "shape": [2, 2],
+          "tolerance": 2e-45,
+          "data": { "kind": "values", "values": [1e-40, -1e-40, -1e-40, 1e-40] }
+        }
+      }
+    },
+    {
+      "name": "f32_subnormal_rank3_all_axes_mean_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 in f32; bit-exact subnormal preservation is unattainable on GPU."
+      },
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean_default_axes_do_not_keep_dims",
+        "notes": "Rank-3 default-axes companion: the mean of equal finite subnormal values should remain subnormal in scalar output form."
+      },
+      "attrs": { "keepdims": 0 },
+      "inputs": { "x": { "dtype": "float32", "shape": [2, 3, 2], "data": { "kind": "constant", "value": 1e-40 } } },
+      "outputs": {
+        "y": { "dtype": "float32", "shape": [], "tolerance": 2e-45, "data": { "kind": "values", "values": [1e-40] } }
+      }
+    },
+    {
+      "name": "f32_subnormal_rank3_all_axes_keepdims_mean_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 in f32; bit-exact subnormal preservation is unattainable on GPU."
+      },
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean_default_axes_keepdims",
+        "notes": "Rank-3 default-axes keepdims companion: the mean of equal finite subnormals should remain subnormal in shape [1,1,1]."
+      },
+      "attrs": { "keepdims": 1 },
+      "inputs": { "x": { "dtype": "float32", "shape": [2, 3, 2], "data": { "kind": "constant", "value": 1e-40 } } },
+      "outputs": {
+        "y": {
+          "dtype": "float32",
+          "shape": [1, 1, 1],
+          "tolerance": 2e-45,
+          "data": { "kind": "values", "values": [1e-40] }
+        }
+      }
+    },
+    {
+      "name": "axis1_empty_cols_identity_zero",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": { "x": { "dtype": "float32", "shape": [2, 0], "data": { "kind": "values", "values": [] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [2], "tolerance": 0 } }
+    },
+    {
+      "name": "axis0_empty_rows_identity_zero",
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": { "x": { "dtype": "float32", "shape": [0, 3], "data": { "kind": "values", "values": [] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 0 } }
+    },
+    {
+      "name": "axis1_zero_rows_noop",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": { "x": { "dtype": "float32", "shape": [0, 3], "data": { "kind": "values", "values": [] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [0], "tolerance": 0 } }
+    },
+    {
+      "name": "axis_minus_one",
+      "attrs": { "axes": [-1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 4],
+          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, -1.0, -2.0, -3.0, -4.0, 0.5, 1.5, 2.5, 3.5] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "rank3_axis2_last_keepdims",
+      "attrs": { "axes": [2], "keepdims": 1 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 3, 4],
+          "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/rank3_axis2_last_keepdims_input_x" } }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 1], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "rank4_axis1_channel_no_keepdims",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 3, 2, 2],
+          "data": {
+            "kind": "values",
+            "values": [1.0, -2.0, 3.0, -4.0, 10.0, 20.0, -30.0, -40.0, 0.25, -0.5, 0.75, -1.0, -5.0, 6.0, -7.0, 8.0, 0.0, 0.0, 1.5, -1.5, 100.0, -200.0, 300.0, -400.0]
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 2, 2], "tolerance": 0.00001 } }
+    },
+    {
+      "name": "rank1_axis0_scalar_output",
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [6],
+          "data": { "kind": "values", "values": [1.0, -2.0, 3.5, 4.5, -1.0, 0.0] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "rank3_axis2_empty_axis_identity_zero",
+      "attrs": { "axes": [2], "keepdims": 1 },
+      "inputs": { "x": { "dtype": "float32", "shape": [2, 3, 0], "data": { "kind": "values", "values": [] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 1], "tolerance": 0 } }
+    },
+    {
+      "name": "rank4_axis1_empty_axis_identity_zero",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": { "x": { "dtype": "float32", "shape": [2, 0, 2, 2], "data": { "kind": "values", "values": [] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 2, 2], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_axis1_rank3_no_keepdims",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean_do_not_keepdims"
+      },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 2, 2],
+          "data": { "kind": "values", "values": [5.0, 1.0, 20.0, 2.0, 30.0, 1.0, 40.0, 2.0, 55.0, 1.0, 60.0, 2.0] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3, 2], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "ort_axis1_rank3_keepdims",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean_keepdims"
+      },
+      "attrs": { "axes": [1], "keepdims": 1 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 2, 2],
+          "data": { "kind": "values", "values": [5.0, 1.0, 20.0, 2.0, 30.0, 1.0, 40.0, 2.0, 55.0, 1.0, 60.0, 2.0] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3, 1, 2], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "ort_axis0_rank1_scalar",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean_do_not_keepdims_2"
+      },
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": { "x": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [1.0, 2.0, 3.0] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "ort_rank0_scalar",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean0DTensor"
+      },
+      "inputs": { "x": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [2.0] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_axis0_singleton_keepdims_noop",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean_keepdims_results_in_noop"
+      },
+      "attrs": { "axes": [0], "keepdims": 1 },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 3], "data": { "kind": "values", "values": [1.0, 2.0, 3.0] } }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 3], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "ort_axis0_singleton_no_keepdims_shape_change",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean_keepdims_results_in_shape_change"
+      },
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 3], "data": { "kind": "values", "values": [1.0, 2.0, 3.0] } }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "ort_default_axes_rank3_no_keepdims_scalar",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean_default_axes_do_not_keep_dims",
+        "notes": "Default axes reduce all input dimensions to a rank-0 scalar when keepdims=0."
+      },
+      "attrs": { "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 2, 2],
+          "data": { "kind": "values", "values": [5.0, 1.0, 20.0, 2.0, 30.0, 1.0, 40.0, 2.0, 55.0, 1.0, 60.0, 2.0] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "onnx_backend_reduce_mean_do_not_keepdims_example",
+      "attrs": { "keepdims": 0, "axes": [1] },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 2, 2],
+          "data": { "kind": "values", "values": [5.0, 1.0, 20.0, 2.0, 30.0, 1.0, 40.0, 2.0, 55.0, 1.0, 60.0, 2.0] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3, 2] } },
+      "provenance": {
+        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_mean_do_not_keepdims_example",
+        "notes": "The ONNX int64 axes input is materialized as this compile-time axes list."
+      }
+    },
+    {
+      "name": "onnx_backend_reduce_mean_do_not_keepdims_random",
+      "attrs": { "keepdims": 0, "axes": [1] },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 2, 2],
+          "data": {
+            "kind": "values",
+            "values": [0.9762700796127319, 4.3037872314453125, 2.055267572402954, 0.8976636528968811, -1.5269039869308472, 2.917882204055786, -1.248255729675293, 7.835460186004639, 9.273255348205566, -2.331169605255127, 5.834500789642334, 0.577898383140564]
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3, 2] } },
+      "provenance": {
+        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_mean_do_not_keepdims_random",
+        "notes": "The ONNX int64 axes input is materialized as this compile-time axes list."
+      }
+    },
+    {
+      "name": "onnx_backend_reduce_mean_keepdims_example",
+      "attrs": { "keepdims": 1, "axes": [1] },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 2, 2],
+          "data": { "kind": "values", "values": [5.0, 1.0, 20.0, 2.0, 30.0, 1.0, 40.0, 2.0, 55.0, 1.0, 60.0, 2.0] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3, 1, 2] } },
+      "provenance": {
+        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_mean_keepdims_example",
+        "notes": "The ONNX int64 axes input is materialized as this compile-time axes list."
+      }
+    },
+    {
+      "name": "onnx_backend_reduce_mean_keepdims_random",
+      "attrs": { "keepdims": 1, "axes": [1] },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 2, 2],
+          "data": {
+            "kind": "values",
+            "values": [0.9762700796127319, 4.3037872314453125, 2.055267572402954, 0.8976636528968811, -1.5269039869308472, 2.917882204055786, -1.248255729675293, 7.835460186004639, 9.273255348205566, -2.331169605255127, 5.834500789642334, 0.577898383140564]
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3, 1, 2] } },
+      "provenance": {
+        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_mean_keepdims_random",
+        "notes": "The ONNX int64 axes input is materialized as this compile-time axes list."
+      }
+    },
+    {
+      "name": "onnx_backend_reduce_mean_negative_axes_keepdims_example",
+      "attrs": { "keepdims": 1, "axes": [-2] },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 2, 2],
+          "data": { "kind": "values", "values": [5.0, 1.0, 20.0, 2.0, 30.0, 1.0, 40.0, 2.0, 55.0, 1.0, 60.0, 2.0] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3, 1, 2] } },
+      "provenance": {
+        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_mean_negative_axes_keepdims_example",
+        "notes": "The ONNX int64 axes input is materialized as this compile-time axes list."
+      }
+    },
+    {
+      "name": "onnx_backend_reduce_mean_negative_axes_keepdims_random",
+      "attrs": { "keepdims": 1, "axes": [-2] },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 2, 2],
+          "data": {
+            "kind": "values",
+            "values": [0.9762700796127319, 4.3037872314453125, 2.055267572402954, 0.8976636528968811, -1.5269039869308472, 2.917882204055786, -1.248255729675293, 7.835460186004639, 9.273255348205566, -2.331169605255127, 5.834500789642334, 0.577898383140564]
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3, 1, 2] } },
+      "provenance": {
+        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_mean_negative_axes_keepdims_random",
+        "notes": "The ONNX int64 axes input is materialized as this compile-time axes list."
+      }
+    },
+    {
+      "name": "onnx_backend_reduce_mean_default_axes_keepdims_example",
+      "provenance": {
+        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_mean_default_axes_keepdims_example"
+      },
+      "attrs": { "keepdims": 1 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 2, 2],
+          "data": { "kind": "values", "values": [5.0, 1.0, 20.0, 2.0, 30.0, 1.0, 40.0, 2.0, 55.0, 1.0, 60.0, 2.0] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1] } }
+    },
+    {
+      "name": "ort_default_axes_keepdims_all_rank3",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean_default_axes_keepdims"
+      },
+      "attrs": { "keepdims": 1 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 2, 2],
+          "data": { "kind": "values", "values": [5.0, 1.0, 20.0, 2.0, 30.0, 1.0, 40.0, 2.0, 55.0, 1.0, 60.0, 2.0] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "onnx_backend_reduce_mean_default_axes_keepdims_random",
+      "provenance": {
+        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reduce_mean_default_axes_keepdims_random"
+      },
+      "attrs": { "keepdims": 1 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 2, 2],
+          "data": {
+            "kind": "values",
+            "values": [0.9762700796127319, 4.3037872314453125, 2.055267572402954, 0.8976636528968811, -1.5269039869308472, 2.917882204055786, -1.248255729675293, 7.835460186004639, 9.273255348205566, -2.331169605255127, 5.834500789642334, 0.577898383140564]
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 1, 1] } }
+    },
+    {
+      "name": "subgroup_vec4_last_axis_2x256",
+      "attrs": { "axes": [-1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 256],
+          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2], "tolerance": 0.0002, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "subgroup_scalar_last_axis_2x65",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 65],
+          "data": { "kind": "fillFloat32", "sinStep": 0.23, "cosStep": 0.11 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2], "tolerance": 0.0002, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "ort_int32_large_values_no_overflow_gpu_gap",
+      "skipGpu": {
+        "category": "todo",
+        "reason": "The current integer reduction route uses an i32 accumulator, so the fixture's 6e9 intermediate sum overflows before division. A portable multiword accumulator can implement this behavior."
+      },
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean_int32_LargeValues_NoOverflow"
+      },
+      "attrs": { "axes": [0], "keepdims": 1 },
+      "inputs": {
+        "x": {
+          "dtype": "int32",
+          "shape": [3],
+          "data": { "kind": "values", "values": [2000000000, 2000000000, 2000000000] }
+        }
+      },
+      "outputs": {
+        "y": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2000000000] }, "tolerance": 0 }
+      }
+    },
+    {
+      "name": "ort_noop_empty_axes_identity",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean_noop_axes_input_initializer_opset_18",
+        "notes": "The omitted axes input exercises empty-axes behavior."
+      },
+      "attrs": { "keepdims": 0, "noop_with_empty_axes": 1 },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [1, 2, 2], "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0] } }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 2], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_int32_multi_axis_keepdims",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean_int32"
+      },
+      "attrs": { "axes": [0, 2], "keepdims": 1 },
+      "inputs": {
+        "x": {
+          "dtype": "int32",
+          "shape": [3, 2, 2],
+          "data": { "kind": "values", "values": [10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120] }
+        }
+      },
+      "outputs": {
+        "y": { "dtype": "int32", "shape": [1, 2, 1], "data": { "kind": "values", "values": [55, 75] }, "tolerance": 0 }
+      }
+    },
+    {
+      "name": "ort_float_multi_axis_keepdims",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean"
+      },
+      "attrs": { "axes": [0, 2], "keepdims": 1 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 2, 2],
+          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 2, 1], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "rank3_lastaxis_cols1024_tree_nosubgroup",
+      "attrs": { "axes": [2], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 2, 1024],
+          "data": {
+            "kind": "cycle",
+            "values": [1.0, -2.0, 0.5, 3.25, -1.5, 2.0, -0.75, 4.0, -3.5, 1.25, 0.0, -2.25, 5.0, -4.0, 2.75, -1.0]
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 2], "tolerance": 0.00001 } }
+    },
+    {
+      "name": "axis0_splitk_8192x32",
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [8192, 32],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [32], "tolerance": 0.0001 } }
+    },
+    {
+      "name": "axis0_splitk_8192x32_offset_mean_scale_lock",
+      "provenance": {
+        "notes": "axis0_splitk_8192x32 draws x from a zero-mean sinusoid, so the mean over 8192 rows cancels to 7e-5 while the absolute tolerance is 1e-4: the whole output fits inside the tolerance and any multiplicative error passes (min detectable uniform scale error 1.43). Offsetting x to oscillate about 1.0 makes each column mean O(1), so the axis0_splitk divisor is under test - a per-split partial that divides by the split height instead of the axis length, a missing final divide, or a combine that double-counts a partial all fail here."
+      },
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [8192, 32],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2, "offset": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [32], "tolerance": 0.0001, "relTolerance": 0.00001 } }
+    },
+    {
+      "name": "axis0_splitk_8192x48_keepdims",
+      "attrs": { "axes": [0], "keepdims": 1 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [8192, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 48], "tolerance": 0.0001 } }
+    },
+    {
+      "name": "axis0_splitk_8192x48_keepdims_offset_mean_scale_lock",
+      "provenance": {
+        "notes": "The keepdims sibling axis0_splitk_8192x48_keepdims averages zero-mean data to 1e-5 against a 1e-4 absolute tolerance (min detectable uniform scale error 10.0), so the keepdims axis0_splitk combine is only checked for shape, never for magnitude. Offsetting x about 1.0 makes the mean O(1) so the divisor on the keepdims path is checked too."
+      },
+      "attrs": { "axes": [0], "keepdims": 1 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [8192, 48],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2, "offset": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [1, 48], "tolerance": 0.0001, "relTolerance": 0.00001 } }
+    },
+    {
+      "name": "f32_rank4_axis2_no_keepdims",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean",
+        "notes": "Rank-4 single-axis reduce over a middle (non-last, non-axis1) dimension."
+      },
+      "attrs": { "axes": [2], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 3, 4, 2],
+          "data": { "kind": "fillFloat32", "sinStep": 0.37, "cosStep": 0.13, "scale": 0.5 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 2], "tolerance": 0.00001 } }
+    },
+    {
+      "name": "f32_rank4_multi_axis_23_keepdims",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean",
+        "notes": "Rank-4 multi-axis reduce over the trailing spatial axes [2,3]."
+      },
+      "attrs": { "axes": [2, 3], "keepdims": 1 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 3, 4, 2],
+          "data": { "kind": "fillFloat32", "sinStep": 0.29, "cosStep": 0.17, "scale": 0.5 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 3, 1, 1], "tolerance": 0.00001 } }
+    },
+    {
+      "name": "f32_rank4_default_all_axes_scalar",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean_default_axes_do_not_keep_dims",
+        "notes": "Reduces every axis of a rank-4 float32 tensor to a rank-0 scalar."
+      },
+      "attrs": { "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 3, 4, 2],
+          "data": { "kind": "fillFloat32", "sinStep": 0.41, "cosStep": 0.19, "scale": 0.5 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [], "tolerance": 0.00001 } }
+    },
+    {
+      "name": "f32_last_axis_inf_nan_propagation",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean",
+        "notes": "inf/NaN propagation through the subgroup vec4 last-axis reducer (cols%4==0): row means of +inf -> inf, mixed +inf/-inf -> NaN, any NaN -> NaN. Not subnormal, so Metal FTZ does not apply."
+      },
+      "attrs": { "axes": [-1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [4, 4],
+          "data": {
+            "kind": "values",
+            "values": [1.0, 2.0, 3.0, 4.0, "Infinity", 1.0, 2.0, 3.0, "Infinity", "-Infinity", 1.0, 1.0, "NaN", 1.0, 2.0, 3.0]
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [4], "tolerance": 0.000001, "allowNaN": true } }
+    },
+    {
+      "name": "rank3_multi_axes_12_keepdims",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/reduction/reduction_ops_test.cc",
+        "test": "ReductionOpTest.ReduceMean",
+        "notes": "Multi-axis rank3 reduce over axes=[1,2] with keepdims=1, exercising the multi_axis_rank3 mask branch beyond the existing [0,2] case."
+      },
+      "attrs": { "axes": [1, 2], "keepdims": 1 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 3, 4],
+          "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/rank3_axis2_last_keepdims_input_x" } }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 1, 1], "tolerance": 0.00001 } }
+    },
+    {
+      "name": "int32_mean_truncation_toward_zero",
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": {
+        "x": { "dtype": "int32", "shape": [2, 4], "data": { "kind": "values", "values": [-3, 5, -7, 9, -4, 2, -8, 2] } }
+      },
+      "outputs": {
+        "y": { "dtype": "int32", "shape": [4], "data": { "kind": "values", "values": [-3, 3, -7, 5] }, "tolerance": 0 }
+      }
+    },
+    {
+      "name": "reduce_size1_axis_returns_input_value",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 1, 4],
+          "data": { "kind": "values", "values": [1.0, -2.0, 3.0, -4.0, 5.0, -6.0, 7.0, -8.0] }
+        }
+      },
+      "outputs": {
+        "y": {
+          "dtype": "float32",
+          "shape": [2, 4],
+          "data": { "kind": "values", "values": [1.0, -2.0, 3.0, -4.0, 5.0, -6.0, 7.0, -8.0] },
+          "tolerance": 0
+        }
+      }
+    },
+    {
+      "name": "all_axes_flat_numel_boundary_8192",
+      "attrs": { "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 8192],
+          "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02, "scale": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [], "tolerance": 0.001, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "all_axes_flat_numel_boundary_8192_offset_mean_scale_lock",
+      "provenance": {
+        "notes": "At exactly FULL_REDUCE_MIN_ELEMENTS the split count is at its low boundary, which is where a combine that divides by the partial count instead of numel would first show; the zero-mean sibling averages to 2.2e-3 against a 1e-3 tolerance (min detectable uniform scale error 0.45) and cannot show it. Offsetting x about 1.0 makes the mean O(1) at the same boundary."
+      },
+      "attrs": { "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [1, 8192],
+          "data": { "kind": "fillFloat32", "sinStep": 0.01, "cosStep": 0.02, "scale": 1.0, "offset": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [], "tolerance": 0.0001, "relTolerance": 0.00001 } }
+    },
+    {
+      "name": "axis0_narrow_f32_8192x3_splitk_guard_lock",
+      "provenance": {
+        "notes": "Compact lock below the historical 16-column split-K guard. Constant ones validate that mean finalization remains exactly one on the split-K route."
+      },
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": { "x": { "dtype": "float32", "shape": [8192, 3], "data": { "kind": "constant", "value": 1.0 } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 0 } }
+    },
+    {
+      "name": "contiguous_suffix_axes12_parallel",
+      "provenance": {
+        "notes": "Contiguous axes {1,2} exercise the shared cooperative suffix reduction instead of one serial lane per output."
+      },
+      "attrs": { "axes": [1, 2], "keepdims": 1 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 16, 16],
+          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.07, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3, 1, 1], "tolerance": 0.00001 } }
+    },
+    {
+      "name": "rank3_axis1_tiled_middle_reduction",
+      "provenance": {
+        "notes": "Compact route fixture for the coalesced multi-lane middle-axis reduction used by the 8x1024x768 case."
+      },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 64, 64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.061, "scale": 0.7 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 64], "tolerance": 0.00002 } }
+    },
+    {
+      "name": "axis_split_rank3_axis1_2x8192x4",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 8192, 4],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 4], "tolerance": 0.0001 } }
+    },
+    {
+      "name": "axis_split_rank3_axis1_2x8192x4_offset_mean_scale_lock",
+      "provenance": {
+        "notes": "axis_split_rank3_axis1_2x8192x4 selects axis_split_tiled_narrow, whose reduce pass writes per-split partials and whose combine performs the divide; its zero-mean data averages to 9e-5 against a 1e-4 tolerance (min detectable uniform scale error 1.08), so neither the partial nor the combine divisor is observable. Offsetting x about 1.0 puts the mean at O(1)."
+      },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 8192, 4],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2, "offset": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 4], "tolerance": 0.0001, "relTolerance": 0.00001 } }
+    },
+    {
+      "name": "f16_axis_split_tiled_narrow_2x8192x4",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 8192, 4],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 4], "tolerance": 0.05, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_axis_split_tiled_narrow_2x8192x4_offset_mean_scale_lock",
+      "provenance": {
+        "notes": "Offsetting the input around 1.0 keeps the mean at O(1), so the tightened tolerance detects errors in the float16 partial representation and the axis-split combine divisor at roughly two float16 ulps."
+      },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 8192, 4],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2, "offset": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 4], "tolerance": 0.002, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "axis_split_rank3_axis1_wide_2x8192x32",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 8192, 32],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 32], "tolerance": 0.0001 } }
+    },
+    {
+      "name": "axis_split_rank3_axis1_wide_2x8192x32_offset_mean_scale_lock",
+      "provenance": {
+        "notes": "The wide axis_split route (inner 32, not the tiled-narrow path) averages zero-mean data to 7e-5 against a 1e-4 tolerance (min detectable uniform scale error 1.43). Offsetting x about 1.0 makes the mean O(1) so the wide route's split count, partial stride, and final divide are all under test."
+      },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [2, 8192, 32],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2, "offset": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 32], "tolerance": 0.0001, "relTolerance": 0.00001 } }
+    },
+    {
+      "name": "f16_axis_split_wide_2x8192x32",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 8192, 32],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 32], "tolerance": 0.05, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_axis_split_wide_2x8192x32_offset_mean_scale_lock",
+      "provenance": {
+        "notes": "f16 wide axis_split: 0.05 absolute tolerance against a 7e-5 mean is a min detectable uniform scale error of 710, so nothing multiplicative is observable on this route. Offsetting x about 1.0 makes the mean O(1) with an f16-resolution tolerance."
+      },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 8192, 32],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2, "offset": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2, 32], "tolerance": 0.002, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_rank3_axis1_serial",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [3, 2, 2],
+          "data": { "kind": "values", "values": [5.0, 1.0, 20.0, 2.0, 30.0, 1.0, 40.0, 2.0, 55.0, 1.0, 60.0, 2.0] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [3, 2], "tolerance": 0.02 } }
+    },
+    {
+      "name": "f16_last_axis_serial_fallback",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 65],
+          "data": { "kind": "fillFloat32", "sinStep": 0.23, "cosStep": 0.11 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2], "tolerance": 0.02, "relTolerance": 0.0001 } }
+    },
+    {
+      "name": "f16_last_axis_serial_fallback_offset_mean_scale_lock",
+      "provenance": {
+        "notes": "The 65-column f16 row uses a different workgroup geometry (pow2ceil(65)=128) from the 1023-column scalar case, and its zero-mean sibling averages to 0.054 against a 0.02 absolute tolerance (min detectable uniform scale error 0.37) - still above the 0.1 blindness threshold. Offsetting x about 1.0 makes the mean O(1) at this geometry."
+      },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [2, 65],
+          "data": { "kind": "fillFloat32", "sinStep": 0.23, "cosStep": 0.11, "offset": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [2], "tolerance": 0.002, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_all_axes",
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": { "x": { "dtype": "float16", "shape": [8192], "data": { "kind": "constant", "value": 1.0 } } },
+      "outputs": { "y": { "dtype": "float16", "shape": [], "tolerance": 0.02 } }
+    },
+    {
+      "name": "f16_axis0_splitk_8192x8",
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [8192, 8],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [8], "tolerance": 0.05, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_axis0_splitk_8192x8_offset_mean_scale_lock",
+      "provenance": {
+        "notes": "The most tolerance-blind fixture in this op: f16_axis0_splitk_8192x8 has a min detectable uniform scale error of 1483 (0.05 absolute tolerance, 3.4e-5 max output), so halving, doubling, or dividing by the split height instead of 8192 all pass. Same rank-2 axis-0 f16 shape with x offset to oscillate about 1.0, so the mean is O(1) and the divisor is checked."
+      },
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [8192, 8],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2, "offset": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [8], "tolerance": 0.002, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_last_axis_vec4_8x1024",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [8, 1024],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [8], "tolerance": 0.05, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_last_axis_vec4_8x1024_offset_mean_scale_lock",
+      "provenance": {
+        "notes": "Last-axis vec4 (tree_last_axis_vec4 on the baseline tier, subgroup_last_axis_vec4 with subgroups): the zero-mean sibling averages 1024 columns to 4.8e-3 against a 0.05 absolute tolerance (min detectable uniform scale error 10.5), so a divide by the vec4 count rather than the column count is invisible. Offsetting x about 1.0 makes each row mean O(1)."
+      },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [8, 1024],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2, "offset": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [8], "tolerance": 0.002, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_last_axis_scalar_8x1023",
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [8, 1023],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [8], "tolerance": 0.05, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_last_axis_scalar_8x1023_offset_mean_scale_lock",
+      "provenance": {
+        "notes": "The scalar last-axis route at 1023 columns (workgroup floor 256, so each lane walks a strided span with a ragged tail); the zero-mean sibling leaves 4.7e-3 against a 0.05 absolute tolerance (min detectable uniform scale error 10.6). Offsetting x about 1.0 makes the mean O(1) so a divisor of 1024 instead of 1023, or a lane span miscount, fails."
+      },
+      "attrs": { "axes": [1], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [8, 1023],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2, "offset": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [8], "tolerance": 0.002, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_all_axes_flat_65543",
+      "attrs": { "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [65543],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [], "tolerance": 0.05, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_all_axes_flat_65543_offset_mean_scale_lock",
+      "provenance": {
+        "notes": "65543 is prime to the workgroup span, so the all_axes_flat partial pass runs a ragged tail; the zero-mean sibling reduces it to 5.4e-5 against a 0.05 absolute tolerance (min detectable uniform scale error 924), so a tail element counted twice or a divide by the padded length passes. Offsetting x about 1.0 makes the mean O(1) and any miscount of the element total shows up directly."
+      },
+      "attrs": { "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [65543],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2, "offset": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [], "tolerance": 0.002, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_suffix_vec4_4x8x128",
+      "attrs": { "axes": [1, 2], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [4, 8, 128],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [4], "tolerance": 0.05, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_suffix_vec4_4x8x128_offset_mean_scale_lock",
+      "provenance": {
+        "notes": "The contiguous-suffix vec4 route (tree on the baseline tier, subgroup on the subgroup tiers) reduces 1024 suffix elements per row; the zero-mean sibling leaves 3.3e-3 against a 0.05 absolute tolerance (min detectable uniform scale error 15.3), so the suffix element count used as the divisor is unchecked. Offsetting x about 1.0 makes the suffix mean O(1) on both tiers."
+      },
+      "attrs": { "axes": [1, 2], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [4, 8, 128],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2, "offset": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [4], "tolerance": 0.002, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_suffix_scalar_4x7x37",
+      "attrs": { "axes": [1, 2], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [4, 7, 37],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [4], "tolerance": 0.05, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_suffix_scalar_4x7x37_offset_mean_scale_lock",
+      "provenance": {
+        "notes": "The scalar contiguous-suffix route with a non-vectorizable suffix of 259: the zero-mean sibling leaves 5.1e-3 against a 0.05 absolute tolerance (min detectable uniform scale error 9.8). Offsetting x about 1.0 makes the mean O(1), so a divisor taken from a padded or rounded suffix length rather than 7*37 fails."
+      },
+      "attrs": { "axes": [1, 2], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [4, 7, 37],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2, "offset": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [4], "tolerance": 0.002, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_axis0_tilecols_4096x64",
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [4096, 64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [64], "tolerance": 0.05, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "f16_axis0_tilecols_4096x64_offset_mean_scale_lock",
+      "provenance": {
+        "notes": "axis0_tilecols reduces 4096 rows per column tile in one workgroup; the zero-mean sibling averages to 1e-4 against a 0.05 absolute tolerance (min detectable uniform scale error 509), so a divide by the tile column count rather than the row count is invisible. Offsetting x about 1.0 makes the column means O(1)."
+      },
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [4096, 64],
+          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.07, "scale": 0.2, "offset": 1.0 }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [64], "tolerance": 0.002, "relTolerance": 0.002 } }
+    },
+    {
+      "name": "int32_axis0_tiled_64x32",
+      "attrs": { "axes": [0], "keepdims": 0 },
+      "inputs": {
+        "x": {
+          "dtype": "int32",
+          "shape": [64, 32],
+          "data": { "kind": "cycle", "values": [16777217, 3, -5, 16777219, 7, -11, 2] }
+        }
+      },
+      "outputs": { "y": { "dtype": "int32", "shape": [32], "tolerance": 0 } }
+    }
+  ]
+}