sync 2e7068faf55e

Browse files

Files changed (7) hide show

README.md +57 -0
build/webgpu/bench.json +41 -0
build/webgpu/manifest.json +74 -0
build/webgpu/metadata.json +19 -0
build/webgpu/test.json +210 -0
build/webgpu/unary-scalar.wgsl.jinja +36 -0
build/webgpu/unary-vec4.wgsl.jinja +29 -0

README.md CHANGED Viewed

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

 ---
+library_name: kernels
 license: apache-2.0
+tags:
+- kernel
+- webgpu
+- wgsl
 ---
+# ai.onnx.Neg
+`ai.onnx`  ·  standard ONNX operator  ·  ONNX opset ≥ 13
+## Description
+Applies elementwise arithmetic negation to a tensor: `y = -x`. The output has the same shape and type as the input.
+See the [ONNX `Neg` spec](https://onnx.ai/onnx/operators/onnx__Neg.html) for the reference semantics.
+## Inputs
+| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- | --- |
+| `X` | `x` | `T` | — | — | Input tensor whose elements are to be negated. | required |
+## Outputs
+| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- | --- |
+| `Y` | `y` | `T` | same as `X` | same as `X` | Output tensor with each element flipped in sign. | required |
+## Type constraints
+| Variable | Allowed dtypes |
+| --- | --- |
+| `T` | `float32`, `float16`, `int32`, `int8` |
+## Files
+- [`metadata.json`](build/webgpu/metadata.json) — kernel metadata (id, digests, provenance)
+- [`manifest.json`](build/webgpu/manifest.json) — the op contract (source of truth)
+- [`test.json`](build/webgpu/test.json) — correctness cases
+- [`bench.json`](build/webgpu/bench.json) — benchmark + tuning cases
+- [`unary-scalar.wgsl.jinja`](build/webgpu/unary-scalar.wgsl.jinja)
+- [`unary-vec4.wgsl.jinja`](build/webgpu/unary-vec4.wgsl.jinja)
+## Use with `@huggingface/kernels`
+The loader derives every required output's shape and logical dtype from the manifest contract and this call.
+It then allocates the result tensors automatically.
+The `version: 1` option selects the published kernel contract; it is independent of any operator opset, contrib `since_version`, or model version.
+Replace each `*Data` placeholder with a typed array containing the corresponding input data.
+```js
+import { getKernel } from "@huggingface/kernels";
+const kernel = await getKernel("webgpu-kernels/ai.onnx.Neg", { version: 1 });
+const { y } = await kernel({ x: { data: xData, shape: [] } });
+```

build/webgpu/bench.json ADDED Viewed

	@@ -0,0 +1,41 @@

+{
+  "op": "ai.onnx.Neg",
+  "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
+  "cases": [
+    {
+      "name": "neg-f32-1m",
+      "preset": "smoke",
+      "vars": { "dtype": "float32", "count": 1048576 },
+      "inputs": { "x": { "shape": [1048576], "dtype": "float32", "dist": "normal", "seed": 920, "scale": 1 } },
+      "outputs": { "y": { "shape": [1048576], "dtype": "float32" } },
+      "bench": {
+        "primary": true,
+        "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }]
+      }
+    },
+    {
+      "name": "neg-f32-8m",
+      "preset": "smoke",
+      "vars": { "dtype": "float32", "count": 8388608 },
+      "inputs": { "x": { "shape": [8388608], "dtype": "float32", "seed": 7017, "dist": "normal", "scale": 1 } },
+      "outputs": { "y": { "shape": [8388608], "dtype": "float32" } },
+      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
+    },
+    {
+      "name": "neg-f16-8m",
+      "preset": "smoke",
+      "vars": { "dtype": "float16", "count": 8388608 },
+      "inputs": { "x": { "shape": [8388608], "dtype": "float16", "seed": 7018, "dist": "normal", "scale": 1 } },
+      "outputs": { "y": { "shape": [8388608], "dtype": "float16" } },
+      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
+    },
+    {
+      "name": "neg-f32-scalar-fallback-8m",
+      "preset": "edge",
+      "vars": { "dtype": "float32", "count": 8388607 },
+      "inputs": { "x": { "shape": [8388607], "dtype": "float32", "dist": "normal", "seed": 7023, "scale": 1 } },
+      "outputs": { "y": { "shape": [8388607], "dtype": "float32", "dist": "empty" } },
+      "bench": { "metrics": [{ "type": "bandwidth", "value": "args.count * dtypeBytes(args.dtype) * 2" }] }
+    }
+  ]
+}

build/webgpu/manifest.json ADDED Viewed

	@@ -0,0 +1,74 @@

+{
+  "domain": "ai.onnx",
+  "name": "Neg",
+  "sinceVersion": 13,
+  "description": "Applies elementwise arithmetic negation to a tensor: `y = -x`. The output has the same shape and type as the input.",
+  "inputs": [{ "role": "X", "dtype": "T", "description": "Input tensor whose elements are to be negated." }],
+  "outputs": [
+    {
+      "role": "Y",
+      "dtype": "T",
+      "rank": "ranks.X",
+      "description": "Output tensor with each element flipped in sign.",
+      "shape": "shapes.X"
+    }
+  ],
+  "typeConstraints": { "T": ["float32", "float16", "int32", "int8"] },
+  "args": {
+    "x": { "kind": "tensor", "semantic": "X", "role": "input" },
+    "y": { "kind": "tensor", "semantic": "Y", "role": "output" }
+  },
+  "tunables": { "WORKGROUP_SIZE": 256 },
+  "constants": { "scalar": "dtypes.T", "usesF16": "dtypes.T == \"f16\"", "isInt8": "tensorDtypes.X == \"int8\"" },
+  "variants": [
+    {
+      "id": "same_layout_vec4",
+      "when": ["numel(shapes.X) > 0", "numel(shapes.X) % 4 == 0", "numel(shapes.X) == numel(shapes.Y)", "f16Ok(dtypes.T)"],
+      "constants": { "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"" },
+      "passes": [
+        {
+          "id": "main",
+          "name": "Neg.vec4",
+          "bindings": [
+            { "name": "x", "arg": "x", "buffer": { "type": "read-only-storage" }, "elementType": "$vectorScalar" },
+            { "name": "y", "arg": "y", "buffer": { "type": "storage" }, "elementType": "$vectorScalar" },
+            {
+              "name": "params",
+              "semantic": "kernel.params",
+              "buffer": { "type": "uniform" },
+              "struct": {
+                "name": "Params",
+                "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.y) / 4" }]
+              }
+            }
+          ],
+          "dispatch": { "threads": "numel(shapes.Y) / 4", "workgroupSize": "tunables.WORKGROUP_SIZE" },
+          "source": { "shader": "unary-vec4.wgsl.jinja", "inputs": { "op": "\"neg\"" } }
+        }
+      ],
+      "priority": 20
+    },
+    {
+      "id": "scalar",
+      "when": ["numel(shapes.X) == numel(shapes.Y)", "f16Ok(dtypes.T)"],
+      "passes": [
+        {
+          "id": "main",
+          "name": "Neg",
+          "source": { "shader": "unary-scalar.wgsl.jinja", "inputs": { "op": "\"neg\"", "itemsPerInvocation": 4 } },
+          "bindings": [
+            { "name": "x", "arg": "x", "buffer": { "type": "read-only-storage" }, "elementType": "$scalar" },
+            { "name": "y", "arg": "y", "buffer": { "type": "storage" }, "elementType": "$scalar" },
+            {
+              "name": "params",
+              "semantic": "kernel.params",
+              "buffer": { "type": "uniform" },
+              "struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.y)" }] }
+            }
+          ],
+          "dispatch": { "threads": "ceilDiv(numel(shapes.X), 4)", "workgroupSize": "tunables.WORKGROUP_SIZE" }
+        }
+      ]
+    }
+  ]
+}

build/webgpu/metadata.json ADDED Viewed

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

+{
+  "name": "ai.onnx.Neg",
+  "id": "_ai_onnx_neg_webgpu_8236ad6",
+  "version": 1,
+  "license": "Apache-2.0",
+  "backend": { "type": "webgpu" },
+  "digest": {
+    "algorithm": "sha256",
+    "files": {
+      "bench.json": "n5005lamF4m4p0BxJ1kSTJvbtpJuoIEeQ1FPQ0HMuKA=",
+      "manifest.json": "PQmzvtWQy5xB+2hy5pYhEUyxUTmHQcfwmJ1WfnkIS5Q=",
+      "test.json": "x27LNvCdyCMV4Lg4DSMz7vUiHWGXz7+o9vbFgSRs368=",
+      "unary-scalar.wgsl.jinja": "B6v4kydeS0fCwDbfgbbFxI9czECCbxFO++n1qKzvul8=",
+      "unary-vec4.wgsl.jinja": "7MxDZBYskcJujwaEid6Vs5Y8f+Wb9aGLE4wjXvJroso="
+    }
+  },
+  "provenance": { "kernel": { "sha": "2e7068faf55e7f43df740015f6d1ee49391a41c5", "dirty": false } },
+  "webgpu": { "manifestSpec": "1.0", "specialized": true, "opPath": "ops/ai.onnx.Neg" }
+}

build/webgpu/test.json ADDED Viewed

	@@ -0,0 +1,210 @@

+{
+  "op": "ai.onnx.Neg",
+  "cases": [
+    {
+      "name": "vector",
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [32], "data": { "kind": "fillFloat32", "sinStep": 0.1, "cosStep": 0.2 } }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [32], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "rank0_scalar",
+      "inputs": { "x": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [-3.5] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [], "tolerance": 0.000001 } }
+    },
+    {
+      "name": "int32_exact_above_float24",
+      "inputs": {
+        "x": {
+          "dtype": "int32",
+          "shape": [4],
+          "data": { "kind": "values", "values": [16777217, -16777217, 123456789, -123456789] }
+        }
+      },
+      "outputs": { "y": { "dtype": "int32", "shape": [4] } }
+    },
+    {
+      "name": "float_special_values",
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [5],
+          "data": { "kind": "values", "values": ["-Infinity", 0.0, 0.0, "Infinity", "NaN"] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [5], "tolerance": 0, "allowNaN": true } }
+    },
+    {
+      "name": "f16_values",
+      "inputs": {
+        "x": { "dtype": "float16", "shape": [5], "data": { "kind": "values", "values": [-10.0, -1.5, 0.0, 2.0, 8.0] } }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [5], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_float_2x2",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Neg_float"
+      },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [2, 2], "data": { "kind": "values", "values": [1.0, -2.0, 0.0, -10.0] } }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [2, 2], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_int8_values",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Neg_int8"
+      },
+      "inputs": { "x": { "dtype": "int8", "shape": [4], "data": { "kind": "values", "values": [1, -2, 0, -10] } } },
+      "outputs": { "y": { "dtype": "int8", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_int8_min_value_overflow_edge",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Neg_int8",
+        "notes": "Extends ORT's int8 Neg coverage with INT8_MIN, whose mathematical negation is not representable in int8 storage."
+      },
+      "inputs": { "x": { "dtype": "int8", "shape": [4], "data": { "kind": "values", "values": [-128, -127, -1, 0] } } },
+      "outputs": { "y": { "dtype": "int8", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_int32_values",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Neg_int32"
+      },
+      "inputs": { "x": { "dtype": "int32", "shape": [4], "data": { "kind": "values", "values": [1, -2, 0, -10] } } },
+      "outputs": { "y": { "dtype": "int32", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_int16_values_gpu_gap",
+      "skipGpu": {
+        "category": "todo",
+        "reason": "The widened-i32 Neg route is not yet declared and needs an explicit signed 16-bit narrowing step before it can cover the full int16 domain."
+      },
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Neg_int16",
+        "notes": "ORT has signed-integer Neg coverage; this fixture uses the actual ONNX int16 dtype, which is not currently in the WebGPU Neg manifest."
+      },
+      "inputs": { "x": { "dtype": "int16", "shape": [4], "data": { "kind": "values", "values": [1, -2, 0, -10] } } },
+      "outputs": { "y": { "dtype": "int16", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "ort_int16_min_value_overflow_edge_gpu_gap",
+      "skipGpu": {
+        "category": "todo",
+        "reason": "The current widened-i32 Neg kernel produces 32768 for -(-32768) instead of restoring the required wrapped int16 result; add explicit narrowing before enabling int16."
+      },
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Neg_int16",
+        "notes": "Extends ORT's signed int16 Neg coverage with INT16_MIN, whose mathematical negation is not representable in int16 storage."
+      },
+      "inputs": {
+        "x": { "dtype": "int16", "shape": [4], "data": { "kind": "values", "values": [-32768, -32767, -1, 0] } }
+      },
+      "outputs": { "y": { "dtype": "int16", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "f32_subnormal_sign_flip_vec4",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Neg_float",
+        "notes": "Finite signed subnormal float32 inputs are valid; Neg should flip their sign without flushing their magnitude to zero."
+      },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [4],
+          "data": { "kind": "values", "values": [-1e-40, 1e-40, -2e-40, 2e-40] }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [4], "tolerance": 0 } }
+    },
+    {
+      "name": "f32_subnormal_sign_flip_scalar",
+      "provenance": {
+        "source": "onnxruntime/test/providers/cpu/math/element_wise_ops_test.cc",
+        "test": "MathOpTest.Neg_float",
+        "notes": "Scalar-path companion for signed subnormal Neg sign flipping."
+      },
+      "inputs": {
+        "x": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [-1e-40, 0.0, 1e-40] } }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3], "tolerance": 0 } }
+    },
+    {
+      "name": "onnx_backend_example",
+      "provenance": {
+        "source": "cmake/external/onnx/onnx/backend/test/data/node/test_neg_example",
+        "test": "test_neg_example"
+      },
+      "inputs": { "x": { "dtype": "float32", "shape": [2], "data": { "kind": "values", "values": [-4.0, 2.0] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [2], "tolerance": 0 } }
+    },
+    {
+      "name": "onnx_backend_neg",
+      "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_neg" },
+      "inputs": {
+        "x": {
+          "dtype": "float32",
+          "shape": [3, 4, 5],
+          "data": {
+            "kind": "values",
+            "values": [1.764052391052246, 0.40015721321105957, 0.978738009929657, 2.2408931255340576, 1.8675580024719238, -0.9772778749465942, 0.9500884413719177, -0.15135720372200012, -0.10321885347366333, 0.4105985164642334, 0.14404356479644775, 1.4542734622955322, 0.7610377073287964, 0.12167501449584961, 0.44386324286460876, 0.3336743414402008, 1.4940791130065918, -0.2051582634449005, 0.3130677044391632, -0.8540957570075989, -2.5529897212982178, 0.653618574142456, 0.8644362092018127, -0.7421650290489197, 2.269754648208618, -1.4543657302856445, 0.04575851559638977, -0.18718385696411133, 1.5327792167663574, 1.4693588018417358, 0.154947429895401, 0.37816253304481506, -0.8877857327461243, -1.980796456336975, -0.34791216254234314, 0.15634897351264954, 1.2302906513214111, 1.202379822731018, -0.38732680678367615, -0.302302747964859, -1.0485529899597168, -1.420017957687378, -1.7062702178955078, 1.950775384902954, -0.5096521973609924, -0.4380742907524109, -1.2527953386306763, 0.7774903774261475, -1.6138978004455566, -0.21274028718471527, -0.8954665660858154, 0.38690251111984253, -0.5108051300048828, -1.18063223361969, -0.02818222902715206, 0.4283318817615509, 0.06651721894741058, 0.30247190594673157, -0.6343221068382263, -0.3627411723136902]
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float32", "shape": [3, 4, 5], "tolerance": 0 } }
+    },
+    {
+      "name": "onnx_backend_neg_example",
+      "provenance": { "source": "cmake/external/onnx/onnx/backend/test/data/node/test_neg_example" },
+      "inputs": { "x": { "dtype": "float32", "shape": [2], "data": { "kind": "values", "values": [-4.0, 2.0] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [2], "tolerance": 0.00001 } }
+    },
+    {
+      "name": "vec4_f16_lanes",
+      "inputs": {
+        "x": {
+          "dtype": "float16",
+          "shape": [16],
+          "data": {
+            "kind": "values",
+            "values": [-6.0, -4.0, -3.0, -2.0, -1.5, -1.0, -0.5, -0.25, 0.0, 0.25, 0.5, 1.0, 1.5, 2.0, 4.0, 6.0]
+          }
+        }
+      },
+      "outputs": { "y": { "dtype": "float16", "shape": [16], "tolerance": 0 } }
+    },
+    {
+      "name": "empty_input_zero_dim",
+      "inputs": { "x": { "dtype": "float32", "shape": [0], "data": { "kind": "values", "values": [] } } },
+      "outputs": { "y": { "dtype": "float32", "shape": [0], "tolerance": 0 } }
+    },
+    {
+      "name": "int32_min_overflow_wrap",
+      "inputs": {
+        "x": {
+          "dtype": "int32",
+          "shape": [4],
+          "data": { "kind": "values", "values": [-2147483648, -2147483647, -1, 0] }
+        }
+      },
+      "outputs": {
+        "y": {
+          "dtype": "int32",
+          "shape": [4],
+          "tolerance": 0,
+          "data": { "kind": "values", "values": [-2147483648, 2147483647, 1, 0] }
+        }
+      }
+    }
+  ]
+}

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

	@@ -0,0 +1,36 @@

+{% macro flat_tail_open() %}
+@compute @workgroup_size({{ tunables.WORKGROUP_SIZE }})
+fn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups) nwg: vec3<u32>) {
+  // 2D-folded flat index: gid.y carries the high bits when the element count exceeds the
+  // maxComputeWorkgroupsPerDimension limit (the dispatch caps x and spills the rest into y).
+  let invocation = gid.x + gid.y * nwg.x * {{ tunables.WORKGROUP_SIZE }}u;
+  // Tail-safe scalar x4 keeps vector-like dispatch density without requiring
+  // the logical tensor length (or its storage binding) to be vec4 aligned.
+  let begin = invocation * {{ source.itemsPerInvocation }}u;
+  let end = min(begin + {{ source.itemsPerInvocation }}u, params.count);
+  for (var i = begin; i < end; i = i + 1u) {
+{%- endmacro %}
+{% macro flat_tail_close() %}
+  }
+{% endmacro %}
+// Scalar unary fallback. Each branch retains the operation's numeric hardening,
+// including Payne-Hanek trigonometric range reduction and NaN/overflow guards.
+{% if usesF16 %}
+enable f16;
+{% endif %}
+{{ env.wgsl.resourceDeclarations }}
+{{ flat_tail_open() }}
+{% if scalar == "i32" %}
+{% if isInt8 %}
+  // int8 is stored in i32 lanes; -(INT8_MIN) wraps to INT8_MIN in the logical
+  // dtype. Sign-extend the low byte to reproduce that wraparound.
+  y[i] = ((-x[i]) << 24u) >> 24u;
+{% else %}
+  y[i] = -x[i];
+{% endif %}
+{% else %}
+  y[i] = {{ scalar }}(-f32(x[i]));
+{% endif %}
+{{ flat_tail_close() -}}
+}

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

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

+// Loads and stores vec4<T> (128 bits) while retaining scalar per-component
+// arithmetic, including per-component helper calls for guard-heavy operations.
+{% if usesF16 %}
+enable f16;
+{% endif %}
+{{ env.wgsl.resourceDeclarations }}
+@compute @workgroup_size({{ tunables.WORKGROUP_SIZE }})
+fn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups) nwg: vec3<u32>) {
+  // 2D-folded flat index: gid.y carries the high bits when the element count exceeds the
+  // maxComputeWorkgroupsPerDimension limit (the dispatch caps x and spills the rest into y).
+  let i = gid.x + gid.y * nwg.x * {{ tunables.WORKGROUP_SIZE }}u;
+  if (i >= params.count) {
+    return;
+  }
+  let xv = x[i];
+{% if scalar == "i32" %}
+{% if isInt8 %}
+  // int8 is stored in i32 lanes; -(INT8_MIN) wraps to INT8_MIN in the logical
+  // dtype. Sign-extend the low byte componentwise to reproduce that wraparound.
+  y[i] = ((-xv) << vec4<u32>(24u)) >> vec4<u32>(24u);
+{% else %}
+  y[i] = -xv;
+{% endif %}
+{% else %}
+  y[i] = {{ vectorScalar }}(-vec4<f32>(xv));
+{% endif %}
+}