sync 2e7068faf55e

Browse files

Files changed (8) hide show

README.md +60 -0
build/webgpu/bench.json +29 -0
build/webgpu/compare-broadcast-vec4.wgsl.jinja +116 -0
build/webgpu/compare-broadcast.wgsl.jinja +103 -0
build/webgpu/compare-vec4.wgsl.jinja +17 -0
build/webgpu/manifest.json +215 -0
build/webgpu/metadata.json +20 -0
build/webgpu/test.json +536 -0

README.md CHANGED Viewed

@@ -1,3 +1,63 @@
 ---
 license: apache-2.0
 ---

 ---
+library_name: kernels
 license: apache-2.0
+tags:
+- kernel
+- webgpu
+- wgsl
 ---
+# ai.onnx.Less
+`ai.onnx`  ·  standard ONNX operator  ·  ONNX opset ≥ 13
+## Description
+Performs an elementwise `less-than` comparison between tensors `A` and `B` with NumPy-style multidirectional broadcasting, producing a boolean result tensor `C`. Each output element is `true` where the corresponding element of `A` is strictly less than that of `B`.
+See the [ONNX `Less` spec](https://onnx.ai/onnx/operators/onnx__Less.html) for the reference semantics.
+## Inputs
+| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- | --- |
+| `A` | `a` | `T` | — | — | First input operand for the less-than comparison. | required |
+| `B` | `b` | `T` | — | — | Second input operand for the less-than comparison. | required |
+## Outputs
+| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- | --- |
+| `C` | `c` | `B` | derived | broadcast result of `A` and `B` | Boolean result tensor; true where A < B. | required |
+## Type constraints
+| Variable | Allowed dtypes |
+| --- | --- |
+| `T` | `float32`, `float16`, `int32`, `int16`, `uint32`, `int8`, `uint8` |
+| `B` | `bool` |
+## 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
+- [`compare-broadcast-vec4.wgsl.jinja`](build/webgpu/compare-broadcast-vec4.wgsl.jinja)
+- [`compare-broadcast.wgsl.jinja`](build/webgpu/compare-broadcast.wgsl.jinja)
+- [`compare-vec4.wgsl.jinja`](build/webgpu/compare-vec4.wgsl.jinja)
+## Use with `@huggingface/kernels`
+The loader derives every required output's shape and logical dtype from the manifest contract and this call.
+It then allocates the result tensors automatically.
+The `version: 1` option selects the published kernel contract; it is independent of any operator opset, contrib `since_version`, or model version.
+Replace each `*Data` placeholder with a typed array containing the corresponding input data.
+```js
+import { getKernel } from "@huggingface/kernels";
+const kernel = await getKernel("webgpu-kernels/ai.onnx.Less", { version: 1 });
+const { c } = await kernel({ a: { data: aData, shape: [] }, b: { data: bData, shape: [] } });
+```

build/webgpu/bench.json ADDED Viewed

	@@ -0,0 +1,29 @@

+{
+  "op": "ai.onnx.Less",
+  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
+  "cases": [
+    {
+      "name": "1m_f32",
+      "preset": "smoke",
+      "inputs": { "a": { "dtype": "float32", "shape": [1048576] }, "b": { "dtype": "float32", "shape": [1048576] } },
+      "outputs": { "c": { "dtype": "bool", "shape": [1048576] } },
+      "bench": {
+        "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.a) + numel(shapes.b) + numel(shapes.c)) * 4" }]
+      }
+    },
+    {
+      "name": "1m_f32_broadcast_scalar_odd_last_dim",
+      "preset": "edge",
+      "inputs": {
+        "a": { "dtype": "float32", "shape": [1024, 1025], "dist": "normal", "seed": 7, "scale": 2 },
+        "b": { "dtype": "float32", "shape": [1025], "dist": "normal", "seed": 8, "scale": 2 }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [1024, 1025], "dist": "empty" } },
+      "bench": {
+        "metrics": [
+          { "type": "bandwidth", "value": "(numel(shapes.a) + numel(shapes.b) + numel(shapes.c)) * 4", "name": "GB/s" }
+        ]
+      }
+    }
+  ]
+}

build/webgpu/compare-broadcast-vec4.wgsl.jinja ADDED Viewed

	@@ -0,0 +1,116 @@

+{% macro offset_fn(fn_name, opShape, opRank, op_same, op_numel, outShape, outRank, out_numel) %}
+fn {{ fn_name }}({% if out_numel != 0 and op_numel != 1 %}out_index: u32{% endif %}) -> u32 {
+{% if out_numel == 0 %}
+  return 0u;
+{% elif op_numel == 1 %}
+  return 0u;
+{% elif op_same %}
+  return out_index;
+{% else %}
+  var offset = 0u;
+{% for axis in range(outRank) %}
+{% set op_axis = axis - (outRank - opRank) %}
+{% if op_axis >= 0 and opShape[op_axis] != 1 %}
+{% set c_stride = namespace(value=1) %}
+{% for j in range(axis + 1, outRank) %}
+{% set c_stride.value = c_stride.value * outShape[j] %}
+{% endfor %}
+{% set op_stride = namespace(value=1) %}
+{% for j in range(op_axis + 1, opRank) %}
+{% set op_stride.value = op_stride.value * opShape[j] %}
+{% endfor %}
+{% if c_stride.value == 1 %}
+  let coord{{ axis }} = out_index % {{ outShape[axis] }}u;
+{% else %}
+  let coord{{ axis }} = (out_index / {{ c_stride.value }}u) % {{ outShape[axis] }}u;
+{% endif %}
+{% if op_stride.value == 1 %}
+  offset = offset + coord{{ axis }};
+{% else %}
+  offset = offset + coord{{ axis }} * {{ op_stride.value }}u;
+{% endif %}
+{% endif %}
+{% endfor %}
+  return offset;
+{% endif %}
+}
+{%- endmacro %}{% macro broadcast_offset_call(fn_name, opShape, outShape, out_index) %}
+{% set op_numel = namespace(value=1) %}
+{% for d in opShape %}{% set op_numel.value = op_numel.value * d %}{% endfor %}
+{% set out_numel = namespace(value=1) %}
+{% for d in outShape %}{% set out_numel.value = out_numel.value * d %}{% endfor %}
+{{ fn_name }}({% if out_numel.value != 0 and op_numel.value != 1 %}{{ out_index }}{% endif %})
+{%- endmacro %}
+{% if usesF16 %}
+enable f16;
+{% endif %}
+{{ env.wgsl.resourceDeclarations }}
+// Vec4 broadcast comparison op (Equal/Greater/GreaterOrEqual/Less/LessOrEqual),
+// vectorized over the innermost output axis. Each thread writes
+// a vec4<u32> of {0,1} results. The guard requires the output's innermost
+// (stride-1) axis to be a multiple of 4, so a vec4 group of 4 consecutive
+// outputs never crosses that axis. Per the broadcast rules, for the innermost
+// axis each operand is one of: scalar (whole operand is one element), splat
+// (innermost axis 1), or contiguous (innermost axis matches output). Operands
+// stay scalar-typed bindings (4 coalesced scalar loads == one vec4 of bandwidth)
+// and compare in their native type (no f32 widening — exact for int operands).
+// Grid-strides a clamped dispatch so outputs beyond one dispatch dimension are covered.
+{% set a_numel = namespace(value=1) %}
+{% for d in source.aShape %}{% set a_numel.value = a_numel.value * d %}{% endfor %}
+{% set b_numel = namespace(value=1) %}
+{% for d in source.bShape %}{% set b_numel.value = b_numel.value * d %}{% endfor %}
+{% set c_numel = namespace(value=1) %}
+{% for d in source.cShape %}{% set c_numel.value = c_numel.value * d %}{% endfor %}
+{% set a_same = namespace(value=(source.aRank == source.cRank)) %}
+{% if a_same.value %}{% for axis in range(source.cRank) %}{% if source.aShape[axis] != source.cShape[axis] %}{% set a_same.value = false %}{% endif %}{% endfor %}{% endif %}
+{% set b_same = namespace(value=(source.bRank == source.cRank)) %}
+{% if b_same.value %}{% for axis in range(source.cRank) %}{% if source.bShape[axis] != source.cShape[axis] %}{% set b_same.value = false %}{% endif %}{% endfor %}{% endif %}
+{% set a_inner = source.aShape[source.aRank - 1] if source.aRank >= 1 else 1 %}
+{% set b_inner = source.bShape[source.bRank - 1] if source.bRank >= 1 else 1 %}
+{% if a_numel.value == 1 %}{% set a_mode = "scalar" %}
+{% elif a_inner == 1 %}{% set a_mode = "splat" %}
+{% else %}{% set a_mode = "contig" %}{% endif %}
+{% if b_numel.value == 1 %}{% set b_mode = "scalar" %}
+{% elif b_inner == 1 %}{% set b_mode = "splat" %}
+{% else %}{% set b_mode = "contig" %}{% endif %}
+{% if a_mode != "scalar" %}
+{{ offset_fn("a_offset", source.aShape, source.aRank, a_same.value, a_numel.value, source.cShape, source.cRank, c_numel.value) }}
+{% endif %}
+{% if b_mode != "scalar" %}
+{{ offset_fn("b_offset", source.bShape, source.bRank, b_same.value, b_numel.value, source.cShape, source.cRank, c_numel.value) }}
+{% endif %}
+{% set OP = {"equal": "==", "greater": ">", "greaterOrEqual": ">=", "less": "<", "lessOrEqual": "<="}[source.op] %}
+const WG: u32 = {{ tunables.WORKGROUP_SIZE }}u;
+@compute @workgroup_size({{ tunables.WORKGROUP_SIZE }})
+fn main(
+  @builtin(global_invocation_id) gid: vec3<u32>,
+  @builtin(num_workgroups) nwg: vec3<u32>
+) {
+  let stride = nwg.x * WG;
+  for (var i4 = gid.x; i4 < params.count; i4 += stride) {
+    let base = i4 * 4u;
+{% if a_mode == "scalar" %}
+    let av = vec4<{{ scalar }}>(a[0]);
+{% elif a_mode == "splat" %}
+    let av = vec4<{{ scalar }}>(a[{{ broadcast_offset_call("a_offset", source.aShape, source.cShape, "base") }}]);
+{% else %}
+    let ao = {{ broadcast_offset_call("a_offset", source.aShape, source.cShape, "base") }};
+    let av = vec4<{{ scalar }}>(a[ao], a[ao + 1u], a[ao + 2u], a[ao + 3u]);
+{% endif %}
+{% if b_mode == "scalar" %}
+    let bv = vec4<{{ scalar }}>(b[0]);
+{% elif b_mode == "splat" %}
+    let bv = vec4<{{ scalar }}>(b[{{ broadcast_offset_call("b_offset", source.bShape, source.cShape, "base") }}]);
+{% else %}
+    let bo = {{ broadcast_offset_call("b_offset", source.bShape, source.cShape, "base") }};
+    let bv = vec4<{{ scalar }}>(b[bo], b[bo + 1u], b[bo + 2u], b[bo + 3u]);
+{% endif %}
+    c[i4] = select(vec4<u32>(0u), vec4<u32>(1u), av {{ OP }} bv);
+  }
+}

build/webgpu/compare-broadcast.wgsl.jinja ADDED Viewed

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

+{% macro flat_tail_open() %}
+@compute @workgroup_size({{ tunables.WORKGROUP_SIZE }})
+fn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups) nwg: vec3<u32>) {
+  // 2D-folded flat index: gid.y carries the high bits when the element count exceeds the
+  // maxComputeWorkgroupsPerDimension limit (the dispatch caps x and spills the rest into y).
+  let invocation = gid.x + gid.y * nwg.x * {{ tunables.WORKGROUP_SIZE }}u;
+{% if source.itemsPerInvocation is defined %}
+  // Tail-safe scalar x4 keeps vector-like dispatch density without requiring
+  // the logical tensor length (or its storage binding) to be vec4 aligned.
+  let begin = invocation * {{ source.itemsPerInvocation }}u;
+  let end = min(begin + {{ source.itemsPerInvocation }}u, params.count);
+  for (var i = begin; i < end; i = i + 1u) {
+{%- else %}
+  let i = invocation;
+  if (i >= params.count) {
+    return;
+  }
+{%- endif %}
+{% endmacro %}
+{% macro flat_tail_close() %}
+{% if source.itemsPerInvocation is defined %}
+  }
+{% endif %}
+{% endmacro %}
+{% macro offset_fn(fn_name, opShape, opRank, op_same, op_numel, outShape, outRank, out_numel) %}
+fn {{ fn_name }}({% if out_numel != 0 and op_numel != 1 %}out_index: u32{% endif %}) -> u32 {
+{% if out_numel == 0 %}
+  return 0u;
+{% elif op_numel == 1 %}
+  return 0u;
+{% elif op_same %}
+  return out_index;
+{% else %}
+  var offset = 0u;
+{% for axis in range(outRank) %}
+{% set op_axis = axis - (outRank - opRank) %}
+{% if op_axis >= 0 and opShape[op_axis] != 1 %}
+{% set c_stride = namespace(value=1) %}
+{% for j in range(axis + 1, outRank) %}
+{% set c_stride.value = c_stride.value * outShape[j] %}
+{% endfor %}
+{% set op_stride = namespace(value=1) %}
+{% for j in range(op_axis + 1, opRank) %}
+{% set op_stride.value = op_stride.value * opShape[j] %}
+{% endfor %}
+{% if c_stride.value == 1 %}
+  let coord{{ axis }} = out_index % {{ outShape[axis] }}u;
+{% else %}
+  let coord{{ axis }} = (out_index / {{ c_stride.value }}u) % {{ outShape[axis] }}u;
+{% endif %}
+{% if op_stride.value == 1 %}
+  offset = offset + coord{{ axis }};
+{% else %}
+  offset = offset + coord{{ axis }} * {{ op_stride.value }}u;
+{% endif %}
+{% endif %}
+{% endfor %}
+  return offset;
+{% endif %}
+}
+{%- endmacro %}{% macro broadcast_offset_call(fn_name, opShape, outShape, out_index) %}
+{% set op_numel = namespace(value=1) %}
+{% for d in opShape %}{% set op_numel.value = op_numel.value * d %}{% endfor %}
+{% set out_numel = namespace(value=1) %}
+{% for d in outShape %}{% set out_numel.value = out_numel.value * d %}{% endfor %}
+{{ fn_name }}({% if out_numel.value != 0 and op_numel.value != 1 %}{{ out_index }}{% endif %})
+{%- endmacro %}{% macro broadcast_offset_fn(fn_name, opShape, opRank, outShape, outRank) %}
+{% set op_numel = namespace(value=1) %}
+{% for d in opShape %}
+{% set op_numel.value = op_numel.value * d %}
+{% endfor %}
+{% set out_numel = namespace(value=1) %}
+{% for d in outShape %}
+{% set out_numel.value = out_numel.value * d %}
+{% endfor %}
+{% set op_same = namespace(value=(opRank == outRank)) %}
+{% if op_same.value %}
+{% for axis in range(outRank) %}
+{% if opShape[axis] != outShape[axis] %}
+{% set op_same.value = false %}
+{% endif %}
+{% endfor %}
+{% endif %}
+{{ offset_fn(fn_name, opShape, opRank, op_same.value, op_numel.value, outShape, outRank, out_numel.value) }}
+{%- endmacro %}{% macro binary_broadcast_offsets() %}
+{{ broadcast_offset_fn("a_offset", source.aShape, source.aRank, source.cShape, source.cRank) }}
+{{ broadcast_offset_fn("b_offset", source.bShape, source.bRank, source.cShape, source.cRank) }}
+{%- endmacro %}
+{% if usesF16 %}
+enable f16;
+{% endif %}
+{{ env.wgsl.resourceDeclarations }}
+{{ binary_broadcast_offsets() }}
+{{ flat_tail_open() }}
+  c[i] = select(0u, 1u, a[{{ broadcast_offset_call("a_offset", source.aShape, source.cShape, "i") }}] < b[{{ broadcast_offset_call("b_offset", source.bShape, source.cShape, "i") }}]);
+{{ flat_tail_close() -}}
+}

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

	@@ -0,0 +1,17 @@

+{% if usesF16 %}
+enable f16;
+{% endif %}
+{{ env.wgsl.resourceDeclarations }}
+@compute @workgroup_size({{ tunables.WORKGROUP_SIZE }})
+fn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups) nwg: vec3<u32>) {
+  // 2D-folded flat index: gid.y carries the high bits when the element count exceeds the
+  // maxComputeWorkgroupsPerDimension limit (the dispatch caps x and spills the rest into y).
+  let i = gid.x + gid.y * nwg.x * {{ tunables.WORKGROUP_SIZE }}u;
+  if (i >= params.count) {
+    return;
+  }
+  let av = a[i];
+  let bv = b[i];
+  c[i] = select(vec4<u32>(0u), vec4<u32>(1u), av < bv);
+}

build/webgpu/manifest.json ADDED Viewed

	@@ -0,0 +1,215 @@

+{
+  "domain": "ai.onnx",
+  "name": "Less",
+  "sinceVersion": 13,
+  "description": "Performs an elementwise `less-than` comparison between tensors `A` and `B` with NumPy-style multidirectional broadcasting, producing a boolean result tensor `C`. Each output element is `true` where the corresponding element of `A` is strictly less than that of `B`.",
+  "inputs": [
+    { "role": "A", "dtype": "T", "description": "First input operand for the less-than comparison." },
+    { "role": "B", "dtype": "T", "description": "Second input operand for the less-than comparison." }
+  ],
+  "outputs": [
+    {
+      "role": "C",
+      "dtype": "B",
+      "rank": "max(ranks.A, ranks.B)",
+      "description": "Boolean result tensor; true where A < B.",
+      "shape": "broadcastShape(shapes.A, shapes.B)"
+    }
+  ],
+  "typeConstraints": { "T": ["float32", "float16", "int32", "int16", "uint32", "int8", "uint8"], "B": ["bool"] },
+  "args": {
+    "a": { "kind": "tensor", "semantic": "A", "role": "input" },
+    "b": { "kind": "tensor", "semantic": "B", "role": "input" },
+    "c": { "kind": "tensor", "semantic": "C", "role": "output" }
+  },
+  "tunables": { "WORKGROUP_SIZE": 256 },
+  "variants": [
+    {
+      "id": "same_shape_vec4",
+      "priority": 20,
+      "when": ["sameShape(shapes.A, shapes.C)", "sameShape(shapes.B, shapes.C)", "numel(shapes.C) > 0", "numel(shapes.C) % 4 == 0", "f16Ok(dtypes.T)"],
+      "constants": { "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"", "usesF16": "dtypes.T == \"f16\"" },
+      "passes": [
+        {
+          "id": "main",
+          "name": "Less.vec4",
+          "source": { "shader": "compare-vec4.wgsl.jinja", "inputs": { "op": "\"less\"" } },
+          "bindings": [
+            {
+              "name": "a",
+              "arg": "a",
+              "semantic": "A",
+              "buffer": { "type": "read-only-storage" },
+              "elementType": "$vectorScalar"
+            },
+            {
+              "name": "b",
+              "arg": "b",
+              "semantic": "B",
+              "buffer": { "type": "read-only-storage" },
+              "elementType": "$vectorScalar"
+            },
+            { "name": "c", "arg": "c", "semantic": "C", "buffer": { "type": "storage" }, "elementType": "vec4<u32>" },
+            {
+              "name": "params",
+              "semantic": "kernel.params",
+              "buffer": { "type": "uniform" },
+              "struct": {
+                "name": "Params",
+                "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.C) / 4" }]
+              }
+            }
+          ],
+          "dispatch": { "threads": "numel(shapes.C) / 4", "workgroupSize": "tunables.WORKGROUP_SIZE" }
+        }
+      ]
+    },
+    {
+      "id": "broadcast_vec4",
+      "when": ["ranks.A <= ranks.C", "ranks.B <= ranks.C", "ranks.C >= 1", "dim(shapes.C, ranks.C - 1) % 4 == 0", "numel(shapes.C) % 4 == 0", "numel(shapes.C) >= 4", "f16Ok(dtypes.T)"],
+      "constants": { "scalar": "dtypes.T", "usesF16": "dtypes.T == \"f16\"" },
+      "passes": [
+        {
+          "id": "main",
+          "name": "Less",
+          "source": {
+            "shader": "compare-broadcast-vec4.wgsl.jinja",
+            "inputs": {
+              "aShape": "shapes.A",
+              "bShape": "shapes.B",
+              "cShape": "shapes.C",
+              "aRank": "ranks.A",
+              "bRank": "ranks.B",
+              "cRank": "ranks.C",
+              "op": "\"less\""
+            }
+          },
+          "bindings": [
+            {
+              "name": "a",
+              "arg": "a",
+              "semantic": "A",
+              "buffer": { "type": "read-only-storage" },
+              "elementType": "$scalar"
+            },
+            {
+              "name": "b",
+              "arg": "b",
+              "semantic": "B",
+              "buffer": { "type": "read-only-storage" },
+              "elementType": "$scalar"
+            },
+            { "name": "c", "arg": "c", "semantic": "C", "buffer": { "type": "storage" }, "elementType": "vec4<u32>" },
+            {
+              "name": "params",
+              "semantic": "kernel.params",
+              "buffer": { "type": "uniform" },
+              "struct": {
+                "name": "Params",
+                "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.C) / 4" }]
+              }
+            }
+          ],
+          "dispatch": { "gridStride": "numel(shapes.C) / 4", "workgroupSize": "tunables.WORKGROUP_SIZE" }
+        }
+      ],
+      "priority": 10
+    },
+    {
+      "id": "broadcast",
+      "when": ["ranks.A <= ranks.C", "ranks.B <= ranks.C", "f16Ok(dtypes.T)"],
+      "constants": { "scalar": "dtypes.T", "usesF16": "dtypes.T == \"f16\"" },
+      "passes": [
+        {
+          "id": "main",
+          "name": "Less",
+          "source": {
+            "shader": "compare-broadcast.wgsl.jinja",
+            "inputs": {
+              "aShape": "shapes.A",
+              "bShape": "shapes.B",
+              "cShape": "shapes.C",
+              "aRank": "ranks.A",
+              "bRank": "ranks.B",
+              "cRank": "ranks.C",
+              "op": "\"less\""
+            }
+          },
+          "bindings": [
+            {
+              "name": "a",
+              "arg": "a",
+              "semantic": "A",
+              "buffer": { "type": "read-only-storage" },
+              "elementType": "$scalar"
+            },
+            {
+              "name": "b",
+              "arg": "b",
+              "semantic": "B",
+              "buffer": { "type": "read-only-storage" },
+              "elementType": "$scalar"
+            },
+            { "name": "c", "arg": "c", "semantic": "C", "buffer": { "type": "storage" }, "elementType": "u32" },
+            {
+              "name": "params",
+              "semantic": "kernel.params",
+              "buffer": { "type": "uniform" },
+              "struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.C)" }] }
+            }
+          ],
+          "dispatch": { "threads": "numel(shapes.C)", "workgroupSize": "tunables.WORKGROUP_SIZE" }
+        }
+      ]
+    },
+    {
+      "id": "same_shape_scalar_x4",
+      "priority": 15,
+      "when": ["sameShape(shapes.A, shapes.C)", "sameShape(shapes.B, shapes.C)", "numel(shapes.C) > 0", "numel(shapes.C) % 4 != 0", "ranks.A <= ranks.C", "ranks.B <= ranks.C", "f16Ok(dtypes.T)"],
+      "constants": { "scalar": "dtypes.T", "usesF16": "dtypes.T == \"f16\"" },
+      "passes": [
+        {
+          "id": "main",
+          "name": "Less",
+          "source": {
+            "shader": "compare-broadcast.wgsl.jinja",
+            "inputs": {
+              "aShape": "shapes.A",
+              "bShape": "shapes.B",
+              "cShape": "shapes.C",
+              "aRank": "ranks.A",
+              "bRank": "ranks.B",
+              "cRank": "ranks.C",
+              "op": "\"less\"",
+              "itemsPerInvocation": 4
+            }
+          },
+          "bindings": [
+            {
+              "name": "a",
+              "arg": "a",
+              "semantic": "A",
+              "buffer": { "type": "read-only-storage" },
+              "elementType": "$scalar"
+            },
+            {
+              "name": "b",
+              "arg": "b",
+              "semantic": "B",
+              "buffer": { "type": "read-only-storage" },
+              "elementType": "$scalar"
+            },
+            { "name": "c", "arg": "c", "semantic": "C", "buffer": { "type": "storage" }, "elementType": "u32" },
+            {
+              "name": "params",
+              "semantic": "kernel.params",
+              "buffer": { "type": "uniform" },
+              "struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.C)" }] }
+            }
+          ],
+          "dispatch": { "threads": "ceilDiv(numel(shapes.C), 4)", "workgroupSize": "tunables.WORKGROUP_SIZE" }
+        }
+      ]
+    }
+  ]
+}

build/webgpu/metadata.json ADDED Viewed

	@@ -0,0 +1,20 @@

+{
+  "name": "ai.onnx.Less",
+  "id": "_ai_onnx_less_webgpu_b3a83b0",
+  "version": 1,
+  "license": "Apache-2.0",
+  "backend": { "type": "webgpu" },
+  "digest": {
+    "algorithm": "sha256",
+    "files": {
+      "bench.json": "bi459h1j97iJXf9cdTyyrHoisesrs12uXVRMJmdV3jk=",
+      "compare-broadcast-vec4.wgsl.jinja": "tSx+Yubq+i8IwWTdoOppEsdRyIkK2mENofLE+InFA0w=",
+      "compare-broadcast.wgsl.jinja": "HVI9RZdy0/w305zyaf8ucuqiO7eZVyDwD4NIyvr5ouA=",
+      "compare-vec4.wgsl.jinja": "cXSOC920mmdSmQNgxaMiUZDh4jmFywSvCjPxMr8SAjI=",
+      "manifest.json": "ohxKA8+i8wSkSf6usZjszwgTM6UC/vDXuvvuSeZjOVQ=",
+      "test.json": "1ZrEwckIRspwkQp9fLSrgGzAcBEUXKJKNXIU81H+DYE="
+    }
+  },
+  "provenance": { "kernel": { "sha": "2e7068faf55e7f43df740015f6d1ee49391a41c5", "dirty": false } },
+  "webgpu": { "manifestSpec": "1.0", "specialized": true, "opPath": "ops/ai.onnx.Less" }
+}

build/webgpu/test.json ADDED Viewed

	@@ -0,0 +1,536 @@

+{
+  "op": "ai.onnx.Less",
+  "fixtureArrays": {
+    "onnx_backend_input_a": [1.764052391052246, 0.40015721321105957, 0.978738009929657, 2.2408931255340576, 1.8675580024719238, -0.9772778749465942, 0.9500884413719177, -0.15135720372200012, -0.10321885347366333, 0.4105985164642334, 0.14404356479644775, 1.4542734622955322, 0.7610377073287964, 0.12167501449584961, 0.44386324286460876, 0.3336743414402008, 1.4940791130065918, -0.2051582634449005, 0.3130677044391632, -0.8540957570075989, -2.5529897212982178, 0.653618574142456, 0.8644362092018127, -0.7421650290489197, 2.269754648208618, -1.4543657302856445, 0.04575851559638977, -0.18718385696411133, 1.5327792167663574, 1.4693588018417358, 0.154947429895401, 0.37816253304481506, -0.8877857327461243, -1.980796456336975, -0.34791216254234314, 0.15634897351264954, 1.2302906513214111, 1.202379822731018, -0.38732680678367615, -0.302302747964859, -1.0485529899597168, -1.420017957687378, -1.7062702178955078, 1.950775384902954, -0.5096521973609924, -0.4380742907524109, -1.2527953386306763, 0.7774903774261475, -1.6138978004455566, -0.21274028718471527, -0.8954665660858154, 0.38690251111984253, -0.5108051300048828, -1.18063223361969, -0.02818222902715206, 0.4283318817615509, 0.06651721894741058, 0.30247190594673157, -0.6343221068382263, -0.3627411723136902],
+    "broadcast_vec4_lhs_splat_lower_rank_input_b": [9, 10, 11, 10, 19, 20, 21, 20, 29, 30, 31, 30, 30, 20, 10, 0, 10, 20, 30, 40, 31, 29, 30, 10]
+  },
+  "cases": [
+    {
+      "name": "f32_negative_subnormal_less_than_zero_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: Metal flushes float32 subnormals to zero, so a subnormal compares equal to 0 on GPU; cannot reproduce the strict-inequality result."
+      },
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Less",
+        "notes": "A negative subnormal is strictly less than zero; flushing it to zero flips the predicate."
+      },
+      "inputs": {
+        "a": { "dtype": "float32", "shape": [4], "data": { "kind": "values", "values": [-1e-39, -1e-40, 0.0, 1e-40] } },
+        "b": { "dtype": "float32", "shape": [4], "data": { "kind": "constant", "value": 0.0 } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "f32_negative_subnormal_less_than_zero_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: Metal flushes float32 subnormals to zero, so a subnormal compares equal to 0 on GPU; cannot reproduce the strict-inequality result."
+      },
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Less",
+        "notes": "Scalar-path companion: negative subnormals are strictly less than zero and must keep mask lanes true."
+      },
+      "inputs": {
+        "a": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [-1e-40, 0.0, 1e-40] } },
+        "b": { "dtype": "float32", "shape": [3], "data": { "kind": "constant", "value": 0.0 } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [3], "tolerance": 0 } }
+    },
+    {
+      "name": "broadcast",
+      "inputs": {
+        "a": {
+          "dtype": "float32",
+          "shape": [2, 3],
+          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, 4.0, 5.0, 6.0] }
+        },
+        "b": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [2.0, 2.0, 5.0] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [2, 3] } }
+    },
+    {
+      "name": "ort_int16_scalar_rhs",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Less_int16_Scalar1"
+      },
+      "inputs": {
+        "a": { "dtype": "int16", "shape": [4], "data": { "kind": "values", "values": [1, 0, 2, -1] } },
+        "b": { "dtype": "int16", "shape": [1], "data": { "kind": "values", "values": [1] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "rank0_scalar_scalar_output",
+      "inputs": {
+        "a": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [-1.0] } },
+        "b": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [0.0] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [], "tolerance": 0 } }
+    },
+    {
+      "name": "int32_exact_above_float24",
+      "inputs": {
+        "a": {
+          "dtype": "int32",
+          "shape": [4],
+          "data": { "kind": "values", "values": [16777216, 16777217, -16777217, -16777216] }
+        },
+        "b": {
+          "dtype": "int32",
+          "shape": [4],
+          "data": { "kind": "values", "values": [16777217, 16777216, -16777216, -16777217] }
+        }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_int8_scalar_rhs",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Less_int8_Scalar1"
+      },
+      "inputs": {
+        "a": { "dtype": "int8", "shape": [4], "data": { "kind": "values", "values": [1, 0, 2, -1] } },
+        "b": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [1] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_float_vector",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Less"
+      },
+      "inputs": {
+        "a": { "dtype": "float32", "shape": [4], "data": { "kind": "values", "values": [1.0, 0.0, -1.0, -1.0] } },
+        "b": { "dtype": "float32", "shape": [4], "data": { "kind": "values", "values": [1.0, 1.0, 2.0, -1.0] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_float_scalar_lhs",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Less_Scalar0"
+      },
+      "inputs": {
+        "a": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [1.0] } },
+        "b": { "dtype": "float32", "shape": [4], "data": { "kind": "values", "values": [1.0, 1.5, 2.0, -1.0] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_float_scalar_rhs",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Less_Scalar1"
+      },
+      "inputs": {
+        "a": { "dtype": "float32", "shape": [4], "data": { "kind": "values", "values": [1.0, 0.5, 2.0, -1.0] } },
+        "b": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [1.0] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_uint8_scalar_rhs",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Less_uint8_Scalar1"
+      },
+      "inputs": {
+        "a": { "dtype": "uint8", "shape": [4], "data": { "kind": "values", "values": [1, 0, 2, 3] } },
+        "b": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [2] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_uint32_scalar_rhs",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Less_uint32_Scalar1"
+      },
+      "inputs": {
+        "a": { "dtype": "uint32", "shape": [4], "data": { "kind": "values", "values": [1, 0, 2, 3] } },
+        "b": { "dtype": "uint32", "shape": [1], "data": { "kind": "values", "values": [2] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "nan_comparisons_are_false",
+      "inputs": {
+        "a": {
+          "dtype": "float32",
+          "shape": [4],
+          "data": { "kind": "values", "values": ["NaN", 1.0, "Infinity", "-Infinity"] }
+        },
+        "b": {
+          "dtype": "float32",
+          "shape": [4],
+          "data": { "kind": "values", "values": [0.0, "NaN", "Infinity", "-Infinity"] }
+        }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_int32_broadcast_ab",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Less_broadcastAB"
+      },
+      "inputs": {
+        "a": {
+          "dtype": "int32",
+          "shape": [4, 2],
+          "data": { "kind": "values", "values": [10, 11, 12, 13, 14, 15, 16, 17] }
+        },
+        "b": { "dtype": "int32", "shape": [2], "data": { "kind": "values", "values": [15, 7] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4, 2], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_int32_broadcast_ba",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Less_broadcastBA"
+      },
+      "inputs": {
+        "a": { "dtype": "int32", "shape": [2], "data": { "kind": "values", "values": [15, 7] } },
+        "b": {
+          "dtype": "int32",
+          "shape": [4, 2],
+          "data": { "kind": "values", "values": [10, 11, 12, 13, 14, 15, 16, 17] }
+        }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4, 2], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_int32_multidirectional_broadcast_ab",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Less_multidirectional_broadcastAB"
+      },
+      "inputs": {
+        "a": { "dtype": "int32", "shape": [4, 1], "data": { "kind": "values", "values": [10, 11, 12, 13] } },
+        "b": { "dtype": "int32", "shape": [2], "data": { "kind": "values", "values": [15, 7] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4, 2], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_int32_multidirectional_broadcast_ba",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Less_multidirectional_broadcastBA"
+      },
+      "inputs": {
+        "a": { "dtype": "int32", "shape": [2], "data": { "kind": "values", "values": [15, 7] } },
+        "b": { "dtype": "int32", "shape": [4, 1], "data": { "kind": "values", "values": [10, 11, 12, 13] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4, 2], "tolerance": 0 } }
+    },
+    {
+      "name": "onnx_backend_bcast_float32_rank3",
+      "provenance": {
+        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_less_bcast",
+        "test": "test_less_bcast"
+      },
+      "inputs": {
+        "a": {
+          "dtype": "float32",
+          "shape": [3, 4, 5],
+          "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_input_a" } }
+        },
+        "b": {
+          "dtype": "float32",
+          "shape": [5],
+          "data": {
+            "kind": "values",
+            "values": [-0.6724604368209839, -0.35955315828323364, -0.8131462931632996, -1.7262825965881348, 0.17742614448070526]
+          }
+        }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [3, 4, 5], "tolerance": 0 } }
+    },
+    {
+      "name": "onnx_backend_less",
+      "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_less", "test": "test_less" },
+      "inputs": {
+        "a": {
+          "dtype": "float32",
+          "shape": [3, 4, 5],
+          "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/onnx_backend_input_a" } }
+        },
+        "b": {
+          "dtype": "float32",
+          "shape": [3, 4, 5],
+          "data": {
+            "kind": "values",
+            "values": [-0.6724604368209839, -0.35955315828323364, -0.8131462931632996, -1.7262825965881348, 0.17742614448070526, -0.4017809331417084, -1.630198359489441, 0.46278226375579834, -0.9072983860969543, 0.05194539576768875, 0.7290905714035034, 0.12898291647434235, 1.1394007205963135, -1.234825849533081, 0.4023416340351105, -0.6848101019859314, -0.8707971572875977, -0.5788496732711792, -0.3115525245666504, 0.056165341287851334, -1.1651498079299927, 0.9008265137672424, 0.4656624495983124, -1.5362436771392822, 1.4882521629333496, 1.895889163017273, 1.1787796020507812, -0.1799248307943344, -1.0707526206970215, 1.0544517040252686, -0.4031769335269928, 1.222445011138916, 0.2082749754190445, 0.9766390323638916, 0.3563663959503174, 0.7065731883049011, 0.01050002034753561, 1.7858705520629883, 0.12691208720207214, 0.4019893705844879, 1.8831506967544556, -1.3477590084075928, -1.2704850435256958, 0.969396710395813, -1.1731233596801758, 1.9436211585998535, -0.4136189818382263, -0.747454822063446, 1.922942042350769, 1.4805147647857666, 1.8675589561462402, 0.9060446619987488, -0.8612256646156311, 1.910064935684204, -0.26800337433815, 0.8024563789367676, 0.9472519755363464, -0.15501008927822113, 0.6140793561935425, 0.922206699848175]
+          }
+        }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [3, 4, 5], "tolerance": 0 } }
+    },
+    {
+      "name": "onnx_backend_less_int8",
+      "provenance": {
+        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_less_int8",
+        "test": "test_less_int8"
+      },
+      "inputs": {
+        "a": {
+          "dtype": "int8",
+          "shape": [3, 4, 5],
+          "data": {
+            "kind": "values",
+            "values": [0, -1, 0, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, -1, 0, 0, 0, 2, 0, 0, 1, -1, 0, 0, 1, 0, 0, 0, 0, 1, -1, -1, 0, 0, 1, 0, 0, -1, 0, -1, -1, 1, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0]
+          }
+        },
+        "b": {
+          "dtype": "int8",
+          "shape": [3, 4, 5],
+          "data": {
+            "kind": "values",
+            "values": [0, -1, 0, -2, 0, -1, -1, 0, 0, 1, -1, 0, 0, -1, 0, 0, 0, 0, 2, 1, 0, 0, 1, 0, 0, -1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, -1, 0, 0, 0, 0, 0, -1, 0, 0, 0, -1, 0, 1, -2, 0, 0]
+          }
+        }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [3, 4, 5], "tolerance": 0 } }
+    },
+    {
+      "name": "onnx_backend_less_uint8",
+      "provenance": {
+        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_less_uint8",
+        "test": "test_less_uint8"
+      },
+      "inputs": {
+        "a": {
+          "dtype": "uint8",
+          "shape": [3, 4, 5],
+          "data": {
+            "kind": "values",
+            "values": [13, 0, 18, 7, 12, 14, 19, 8, 13, 17, 17, 10, 22, 19, 7, 12, 13, 2, 11, 12, 17, 13, 6, 14, 15, 7, 5, 21, 17, 16, 3, 18, 9, 10, 0, 18, 0, 20, 13, 6, 10, 23, 19, 1, 13, 6, 23, 23, 12, 23, 20, 14, 1, 1, 16, 12, 17, 10, 20, 21]
+          }
+        },
+        "b": {
+          "dtype": "uint8",
+          "shape": [3, 4, 5],
+          "data": {
+            "kind": "values",
+            "values": [18, 21, 10, 19, 0, 5, 21, 5, 4, 18, 23, 11, 20, 12, 20, 10, 16, 2, 12, 2, 23, 0, 7, 6, 6, 16, 18, 9, 3, 0, 10, 2, 18, 18, 20, 13, 1, 12, 0, 9, 22, 7, 19, 21, 1, 21, 3, 19, 21, 4, 2, 5, 16, 11, 18, 4, 4, 1, 23, 9]
+          }
+        }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [3, 4, 5], "tolerance": 0 } }
+    },
+    {
+      "name": "onnx_backend_less_uint32",
+      "provenance": {
+        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_less_uint32",
+        "test": "test_less_uint32"
+      },
+      "inputs": {
+        "a": {
+          "dtype": "uint32",
+          "shape": [3, 4, 5],
+          "data": {
+            "kind": "values",
+            "values": [19, 2, 12, 9, 19, 8, 5, 20, 21, 11, 1, 11, 15, 12, 15, 12, 21, 18, 13, 12, 9, 21, 6, 4, 16, 8, 5, 15, 5, 5, 17, 0, 16, 16, 0, 10, 12, 0, 7, 22, 19, 5, 4, 19, 4, 5, 12, 17, 19, 12, 6, 7, 20, 16, 0, 23, 15, 10, 3, 3]
+          }
+        },
+        "b": {
+          "dtype": "uint32",
+          "shape": [3, 4, 5],
+          "data": {
+            "kind": "values",
+            "values": [16, 16, 9, 20, 20, 5, 16, 17, 21, 9, 12, 15, 6, 14, 1, 4, 16, 4, 7, 23, 14, 7, 22, 2, 22, 16, 15, 15, 18, 11, 23, 14, 7, 2, 22, 17, 3, 7, 13, 16, 5, 4, 21, 13, 5, 19, 11, 8, 0, 17, 5, 4, 1, 5, 16, 6, 8, 22, 19, 4]
+          }
+        }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [3, 4, 5], "tolerance": 0 } }
+    },
+    {
+      "name": "empty_input_zero_dim",
+      "inputs": {
+        "a": { "dtype": "float32", "shape": [2, 0], "data": { "kind": "values", "values": [] } },
+        "b": { "dtype": "float32", "shape": [0], "data": { "kind": "values", "values": [] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [2, 0], "tolerance": 0 } }
+    },
+    {
+      "name": "broadcast_vec4_inner4_gpu",
+      "provenance": {
+        "source": "synthetic",
+        "test": "broadcast_vec4 inner-dim %4==0 coverage",
+        "notes": "Broadcasts A=[2,1,4] and B to C with innermost extent four. Exercises vec4 broadcast indexing and verifies that the grid-stride count is expressed in vec4 groups (numel(C)/4), including bounds-safe iteration across six groups."
+      },
+      "inputs": {
+        "a": {
+          "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] }
+        },
+        "b": {
+          "dtype": "float32",
+          "shape": [2, 3, 4],
+          "data": {
+            "kind": "values",
+            "values": [0.0, 2.0, 4.0, 4.0, 1.0, 1.0, 1.0, 1.0, 2.0, 3.0, 4.0, 5.0, 5.0, 5.0, 5.0, 5.0, 4.0, 7.0, 6.0, 9.0, 9.0, 9.0, 9.0, 9.0]
+          }
+        }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [2, 3, 4], "tolerance": 0 } }
+    },
+    {
+      "name": "f16_same_shape",
+      "inputs": {
+        "a": {
+          "dtype": "float16",
+          "shape": [8],
+          "data": { "kind": "values", "values": [1.0, 2.0, 3.0, "NaN", "Infinity", "-Infinity", -2.5, 65504.0] }
+        },
+        "b": {
+          "dtype": "float16",
+          "shape": [8],
+          "data": { "kind": "values", "values": [2.0, 2.0, 1.0, 1.0, 1.0, 1.0, -2.5, 65504.0] }
+        }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [8] } }
+    },
+    {
+      "name": "f16_scalar_rhs",
+      "inputs": {
+        "a": {
+          "dtype": "float16",
+          "shape": [8],
+          "data": { "kind": "values", "values": [-2.0, -1.0, 0.0, 1.0, 2.0, 3.0, "NaN", "Infinity"] }
+        },
+        "b": { "dtype": "float16", "shape": [1], "data": { "kind": "values", "values": [1.0] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [8] } }
+    },
+    {
+      "name": "int8_extremes_less",
+      "inputs": {
+        "a": { "dtype": "int8", "shape": [6], "data": { "kind": "values", "values": [-128, 127, -128, 127, 0, -1] } },
+        "b": { "dtype": "int8", "shape": [6], "data": { "kind": "values", "values": [127, -128, -128, 127, -128, 0] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [6], "tolerance": 0 } }
+    },
+    {
+      "name": "broadcast_scalar_path_odd_inner_dim",
+      "inputs": {
+        "a": {
+          "dtype": "int32",
+          "shape": [4, 5],
+          "data": { "kind": "values", "values": [1, 3, 5, 7, 9, 2, 4, 6, 8, 10, 10, 8, 6, 4, 2, 9, 7, 5, 3, 1] }
+        },
+        "b": { "dtype": "int32", "shape": [5], "data": { "kind": "values", "values": [5, 5, 5, 5, 5] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4, 5], "tolerance": 0 } }
+    },
+    {
+      "name": "f32_broadcast_odd_inner_dim_4x17",
+      "provenance": {
+        "notes": "Compact f32 sibling for the odd-inner broadcast benchmark; preserves the [rows, odd] x [odd] shape family without benchmark-scale rows."
+      },
+      "inputs": {
+        "a": {
+          "dtype": "float32",
+          "shape": [4, 17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.037, "cosStep": 0.019, "scale": 0.5 }
+        },
+        "b": {
+          "dtype": "float32",
+          "shape": [17],
+          "data": { "kind": "fillFloat32", "sinStep": 0.011, "cosStep": 0.043, "scale": 0.5 }
+        }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4, 17], "tolerance": 0 } }
+    },
+    {
+      "name": "uint8_extremes_less",
+      "inputs": {
+        "a": { "dtype": "uint8", "shape": [4], "data": { "kind": "values", "values": [0, 255, 128, 127] } },
+        "b": { "dtype": "uint8", "shape": [4], "data": { "kind": "values", "values": [255, 0, 127, 128] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "rank7_broadcast_scalar_tail",
+      "inputs": {
+        "a": {
+          "dtype": "float32",
+          "shape": [2, 1, 2, 1, 2, 1, 3],
+          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23 }
+        },
+        "b": {
+          "dtype": "float32",
+          "shape": [1, 2, 1, 2, 1, 2, 1],
+          "data": { "kind": "fillFloat32", "sinStep": 0.31, "cosStep": 0.07 }
+        }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [2, 2, 2, 2, 2, 2, 3], "tolerance": 0 } }
+    },
+    {
+      "name": "broadcast_vec4_lhs_splat_lower_rank",
+      "provenance": {
+        "source": "authored for render coverage",
+        "test": "compare-broadcast-vec4 lhs splat arm",
+        "notes": "A is rank 2 against a rank 3 output and its innermost axis is 1, so compare-broadcast-vec4 renders `a_same` false straight from the rank test and takes the a_mode == \"splat\" branch, which reads one A element per vec4 group instead of four contiguous ones. Every other vec4-eligible case has A either the same rank as C (contiguous) or a single element (scalar), so neither the lower-rank A arm nor the A-splat arm had ever rendered. Each row of A holds a different value and B differs at every position, so a wrong a_offset changes the result."
+      },
+      "inputs": {
+        "a": { "dtype": "float32", "shape": [3, 1], "data": { "kind": "values", "values": [10.0, 20.0, 30.0] } },
+        "b": {
+          "dtype": "float32",
+          "shape": [2, 3, 4],
+          "data": {
+            "kind": "values",
+            "values": { "$ref": "#/fixtureArrays/broadcast_vec4_lhs_splat_lower_rank_input_b" }
+          }
+        }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [2, 3, 4], "tolerance": 0 } }
+    },
+    {
+      "name": "broadcast_vec4_rhs_splat_lower_rank",
+      "provenance": {
+        "source": "authored for render coverage",
+        "test": "compare-broadcast-vec4 rhs splat arm",
+        "notes": "B is rank 2 with innermost axis 1 against a rank-3 output, so each vec4 output group broadcasts one B value while A remains contiguous. This mirrors the corresponding left-hand splat case."
+      },
+      "inputs": {
+        "a": {
+          "dtype": "float32",
+          "shape": [2, 3, 4],
+          "data": {
+            "kind": "values",
+            "values": { "$ref": "#/fixtureArrays/broadcast_vec4_lhs_splat_lower_rank_input_b" }
+          }
+        },
+        "b": { "dtype": "float32", "shape": [3, 1], "data": { "kind": "values", "values": [10.0, 20.0, 30.0] } }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [2, 3, 4], "tolerance": 0 } }
+    },
+    {
+      "name": "broadcast_vec4_true_scalar_lhs",
+      "provenance": {
+        "source": "authored for render coverage",
+        "test": "compare-broadcast-vec4 rank-0 lhs arm",
+        "notes": "A is a true rank-0 scalar, so `source.aRank >= 1` is false and compare-broadcast-vec4 takes the left-hand scalar arm. C's innermost dimension is 4 and numel is 8, selecting broadcast_vec4 while distinguishing A-scalar from B-scalar indexing."
+      },
+      "inputs": {
+        "a": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [20.0] } },
+        "b": {
+          "dtype": "float32",
+          "shape": [2, 4],
+          "data": { "kind": "values", "values": [10.0, 20.0, 30.0, 20.0, -5.0, 20.0, 25.0, 0.0] }
+        }
+      },
+      "outputs": { "c": { "dtype": "bool", "shape": [2, 4], "tolerance": 0 } }
+    }
+  ]
+}