sync 91d990483a17

Browse files

Files changed (7) hide show

README.md +11 -7
build/webgpu/bench.json +0 -1
build/webgpu/manifest.json +36 -48
build/webgpu/metadata.json +11 -8
build/webgpu/test.json +1 -2
build/webgpu/unary-scalar.wgsl.jinja +12 -11
build/webgpu/unary-vec4.wgsl.jinja +29 -10

README.md CHANGED Viewed

@@ -18,15 +18,15 @@ See the [ONNX `Relu` spec](https://onnx.ai/onnx/operators/onnx__Relu.html) for t
 ## Inputs
-| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
 | --- | --- | --- | --- | --- | --- | --- |
-| `X` | `x` | `T` | — | — | Values clamped elementwise to a minimum of zero. | required |
 ## Outputs
-| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
 | --- | --- | --- | --- | --- | --- | --- |
-| `Y` | `y` | `T` | same as `X` | same as `X` | Output tensor; same shape as the input. | required |
 ## Type constraints
@@ -36,7 +36,7 @@ See the [ONNX `Relu` spec](https://onnx.ai/onnx/operators/onnx__Relu.html) for t
 ## 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
@@ -45,10 +45,14 @@ See the [ONNX `Relu` spec](https://onnx.ai/onnx/operators/onnx__Relu.html) for t
 ## 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.

 ## Inputs
+| Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence |
 | --- | --- | --- | --- | --- | --- | --- |
+| `x` | `X` | `T` | — | — | Values clamped elementwise to a minimum of zero. | required |
 ## Outputs
+| Name | Upstream name | Logical dtype | Rank | Shape | Description | Presence |
 | --- | --- | --- | --- | --- | --- | --- |
+| `y` | `Y` | `T` | same as `x` | same as `x` | Output tensor; same shape as the input. | required |
 ## Type constraints
 ## Files
+- [`metadata.json`](build/webgpu/metadata.json) — kernel metadata (id, digests, per-variant templates, 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
 ## Use with `@huggingface/kernels`
+```sh
+npm install --save-exact @huggingface/kernels@0.0.1-preview.2
+```
+Required output shapes and logical data types are inferred from the supplied inputs and attributes; result tensors are allocated automatically.
 The `version: 1` option selects the published kernel contract; it is independent of any operator opset, contrib `since_version`, or model version.
+It follows the `v1` branch as fixes land. To pin exact artifact bytes, pass a 40-character commit `revision` instead of `version`.
 Replace each `*Data` placeholder with a typed array containing the corresponding input data.

build/webgpu/bench.json CHANGED Viewed

@@ -1,5 +1,4 @@
 {
-  "op": "ai.onnx.Relu",
   "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
   "cases": [
     {

 {
   "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
   "cases": [
     {

build/webgpu/manifest.json CHANGED Viewed

@@ -2,77 +2,65 @@
   "domain": "ai.onnx",
   "name": "Relu",
   "sinceVersion": 14,
-  "description": "Applies the rectified linear unit function elementwise: `y = max(0, x)`. The output has the same shape and type as the input.",
-  "inputs": [{ "role": "X", "dtype": "T", "description": "Values clamped elementwise to a minimum of zero." }],
-  "outputs": [
-    {
-      "role": "Y",
-      "dtype": "T",
-      "rank": "ranks.x",
-      "shape": "shapes.x",
-      "description": "Output tensor; same shape as the input."
-    }
-  ],
   "typeConstraints": { "T": ["float32", "float16", "int32", "int16", "int8"] },
-  "args": {
-    "x": { "kind": "tensor", "semantic": "X", "role": "input" },
-    "y": { "kind": "tensor", "semantic": "Y", "role": "output" }
-  },
-  "tunables": { "WORKGROUP_SIZE": 256 },
   "derive": {
     "wideVec4StorageOk": "dtypes.T != \"f32\" or device.features.has(\"subgroups\") or not (has(device.adapterInfo, \"subgroupMinSize\") and has(device.adapterInfo, \"subgroupMaxSize\") and device.adapterInfo.subgroupMinSize == 8 and device.adapterInfo.subgroupMaxSize <= 32)"
   },
   "variants": [
     {
       "id": "same_layout_vec4",
-      "when": ["numel(shapes.X) > 0", "numel(shapes.X) % 4 == 0", "numel(shapes.X) == numel(shapes.Y)", "f16Ok(dtypes.T)", "wideVec4StorageOk"],
       "passes": [
         {
           "id": "main",
           "name": "Relu.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": "\"relu\"" } }
         }
-      ],
-      "priority": 20,
-      "constants": {
-        "scalar": "dtypes.T",
-        "usesF16": "dtypes.T == \"f16\"",
-        "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\""
-      }
     },
     {
       "id": "scalar",
-      "when": "f16Ok(dtypes.T)",
       "passes": [
         {
           "id": "main",
           "name": "scalar",
-          "source": { "shader": "unary-scalar.wgsl.jinja", "inputs": { "op": "\"relu\"", "itemsPerInvocation": 4 } },
           "bindings": [
-            { "name": "x", "arg": "x", "semantic": "X", "buffer": { "type": "read-only-storage" }, "elementType": "$T" },
-            { "name": "y", "arg": "y", "semantic": "Y", "buffer": { "type": "storage" }, "elementType": "$T" },
-            {
-              "name": "params",
-              "semantic": "kernel.params",
-              "buffer": { "type": "uniform" },
-              "struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.X)" }] }
-            }
           ],
-          "dispatch": { "threads": "ceilDiv(numel(shapes.X), 4)", "workgroupSize": "tunables.WORKGROUP_SIZE" }
         }
       ]
     }

   "domain": "ai.onnx",
   "name": "Relu",
   "sinceVersion": 14,
+  "inputs": { "x": { "onnx": "X", "dtype": "T" } },
+  "outputs": { "y": { "onnx": "Y", "dtype": "T", "rank": "ranks.x", "shape": "shapes.x" } },
   "typeConstraints": { "T": ["float32", "float16", "int32", "int16", "int8"] },
+  "tunables": { "WORKGROUP_SIZE": { "default": 256 } },
   "derive": {
     "wideVec4StorageOk": "dtypes.T != \"f32\" or device.features.has(\"subgroups\") or not (has(device.adapterInfo, \"subgroupMinSize\") and has(device.adapterInfo, \"subgroupMaxSize\") and device.adapterInfo.subgroupMinSize == 8 and device.adapterInfo.subgroupMaxSize <= 32)"
   },
+  "when": ["f16Ok(dtypes.T)"],
   "variants": [
     {
       "id": "same_layout_vec4",
+      "priority": 20,
+      "when": ["numel(shapes.x) > 0", "numel(shapes.x) % 4 == 0", "numel(shapes.x) == numel(shapes.y)", "wideVec4StorageOk"],
+      "derive": {
+        "scalar": "dtypes.T",
+        "usesF16": "dtypes.T == \"f16\"",
+        "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\""
+      },
       "passes": [
         {
           "id": "main",
           "name": "Relu.vec4",
+          "shader": "unary-vec4.wgsl.jinja",
+          "derive": {
+            "op": "\"relu\"",
+            "vec4PerThread": "4 if numel(shapes.y) * dtypeBytes(tensorDtypes.y) <= 16777216 else 1"
+          },
           "bindings": [
+            { "arg": "x", "elementType": "$vectorScalar" },
+            { "arg": "y", "elementType": "$vectorScalar" },
+            { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.y) / 4" }] }
           ],
+          "dispatch": {
+            "x": "min(ceilDiv((ceilDiv(numel(shapes.y) / 4, 4 if numel(shapes.y) * dtypeBytes(tensorDtypes.y) <= 16777216 else 1)), (tunables.WORKGROUP_SIZE)), 65535)",
+            "y": "ceilDiv(ceilDiv((ceilDiv(numel(shapes.y) / 4, 4 if numel(shapes.y) * dtypeBytes(tensorDtypes.y) <= 16777216 else 1)), (tunables.WORKGROUP_SIZE)), 65535)",
+            "z": 1
+          }
         }
+      ]
     },
     {
       "id": "scalar",
+      "derive": { "scalar": "dtypes.T", "usesF16": "dtypes.T == \"f16\"" },
       "passes": [
         {
           "id": "main",
           "name": "scalar",
+          "shader": "unary-scalar.wgsl.jinja",
+          "derive": { "op": "\"relu\"", "itemsPerInvocation": 4 },
           "bindings": [
+            "x",
+            "y",
+            { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.x)" }] }
           ],
+          "dispatch": {
+            "x": "min(ceilDiv((ceilDiv(numel(shapes.x), 4)), (tunables.WORKGROUP_SIZE)), 65535)",
+            "y": "ceilDiv(ceilDiv((ceilDiv(numel(shapes.x), 4)), (tunables.WORKGROUP_SIZE)), 65535)",
+            "z": 1
+          }
         }
       ]
     }

build/webgpu/metadata.json CHANGED Viewed

@@ -1,19 +1,22 @@
 {
   "name": "ai.onnx.Relu",
-  "id": "_ai_onnx_relu_webgpu_38d226d",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
   "digest": {
     "algorithm": "sha256",
     "files": {
-      "bench.json": "pCIgxzDmyCADelXHsyZ3NqiYlep/O2kEemkYPRpHi4U=",
-      "manifest.json": "J17/rYpIwl28VO9DLnSBUimzXY15xRye4otbVNfgPOg=",
-      "test.json": "h0j4IDFTdi+LEdq6CjTJnSNQ1PTE9tpICFf8PVVFEKs=",
-      "unary-scalar.wgsl.jinja": "Vgm24ufHWZEC2cUeI0dA2snfsng2zIRpM1wC8URlqKA=",
-      "unary-vec4.wgsl.jinja": "xSW7Bah2pFe8MxojMV3phM8GEeicxBAk++vGxLURc0Q="
     }
   },
-  "provenance": { "kernel": { "sha": "2e7068faf55e7f43df740015f6d1ee49391a41c5", "dirty": false } },
-  "webgpu": { "manifestSpec": "1.0", "specialized": true, "opPath": "ops/ai.onnx.Relu" }
 }

 {
   "name": "ai.onnx.Relu",
+  "id": "_ai_onnx_relu_webgpu_251441b",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
   "digest": {
     "algorithm": "sha256",
     "files": {
+      "bench.json": "ebSV1lr1/T3ypGQCKBHUbO1bygE2a5cJF9HMlB3unpI=",
+      "manifest.json": "Cd+eoyUA3ITLJr1g3yz5nfXW3t5ESuzBdfAnc0e4LP4=",
+      "test.json": "DDwaG+obBvpm6dyBb1kNPQ2G5uLX77yTTruGxnKJUlM=",
+      "unary-scalar.wgsl.jinja": "cmT3ZpyQMwDNhN4i9ZSCIPZJVcin47pLs0Ce9C0JpYk=",
+      "unary-vec4.wgsl.jinja": "v/Orpo0E19349fM5UjI6yJmPHAGtpWEeOIlGA4Lo6yg="
     }
   },
+  "provenance": { "kernel": { "sha": "91d990483a174128daf7673f3f37a7c890493ae1", "dirty": false } },
+  "webgpu": {
+    "manifestSpec": "2.0",
+    "variants": { "same_layout_vec4": ["unary-vec4.wgsl.jinja"], "scalar": ["unary-scalar.wgsl.jinja"] }
+  }
 }

build/webgpu/test.json CHANGED Viewed

@@ -1,5 +1,4 @@
 {
-  "op": "ai.onnx.Relu",
   "fixtureArrays": {
     "onnx_backend_input_x": [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]
   },
@@ -62,7 +61,7 @@
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/activation/activation_op_test.cc",
         "test": "ActivationOpTest.Relu",
-        "notes": "Scalar-path companion: positive subnormal activations must pass through Relu unchanged."
       },
       "inputs": {
         "x": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [-1e-40, 0.0, 1e-40] } }

 {
   "fixtureArrays": {
     "onnx_backend_input_x": [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]
   },
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/activation/activation_op_test.cc",
         "test": "ActivationOpTest.Relu",
+        "notes": "Scalar Relu must preserve positive subnormal activations."
       },
       "inputs": {
         "x": { "dtype": "float32", "shape": [3], "data": { "kind": "values", "values": [-1e-40, 0.0, 1e-40] } }

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

@@ -1,27 +1,28 @@
 {% 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 T == "f16" %}
 enable f16;
 {% endif %}
 {{ env.wgsl.resourceDeclarations }}
-{% if T != "i32" %}
 fn is_nan_f32(value: f32) -> bool {
   let bits = bitcast<u32>(value);
   return (bits & 0x7f800000u) == 0x7f800000u && (bits & 0x007fffffu) != 0u;
@@ -29,7 +30,7 @@ fn is_nan_f32(value: f32) -> bool {
 {% endif %}
 {{ flat_tail_open() }}
-{% if T == "i32" %}
   let value = x[i];
   y[i] = select(value, 0i, value < 0i);
 {% else %}
@@ -38,7 +39,7 @@ fn is_nan_f32(value: f32) -> bool {
   if (is_nan_f32(value)) {
     out = value;
   }
-  y[i] = {{ T }}(out);
 {% endif %}
 {{ flat_tail_close() -}}
 }

 {% macro flat_tail_open() %}
 @compute @workgroup_size({{ tunables.WORKGROUP_SIZE }})
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
   // 2D-folded flat index: gid.y carries the high bits when the element count exceeds the
+  // dispatch's per-axis workgroup fold width (the dispatch caps x and spills the rest into y).
+  let invocation = gid.x + gid.y * {{ DISPATCH_FOLD_WIDTH }}u * {{ 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.
+{% set itemsPerInvocation = itemsPerInvocation if itemsPerInvocation is defined else 4 %}
+  let begin = invocation * {{ itemsPerInvocation }}u;
+  let end = min(begin + {{ itemsPerInvocation }}u, params.count);
   for (var i = begin; i < end; i = i + 1u) {
 {%- endmacro %}
 {% macro flat_tail_close() %}
   }
 {% endmacro %}
+// Scalar unary elementwise implementation. Specialization emits only the
+// selected operation and any numerical helper it requires.
+{% if usesF16 %}
 enable f16;
 {% endif %}
 {{ env.wgsl.resourceDeclarations }}
+{% if scalar != "i32" %}
 fn is_nan_f32(value: f32) -> bool {
   let bits = bitcast<u32>(value);
   return (bits & 0x7f800000u) == 0x7f800000u && (bits & 0x007fffffu) != 0u;
 {% endif %}
 {{ flat_tail_open() }}
+{% if scalar == "i32" %}
   let value = x[i];
   y[i] = select(value, 0i, value < 0i);
 {% else %}
   if (is_nan_f32(value)) {
     out = value;
   }
+  y[i] = {{ scalar }}(out);
 {% endif %}
 {{ flat_tail_close() -}}
 }

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

@@ -1,8 +1,5 @@
-// 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 }}
 {% macro emit_is_nan_f32() %}
@@ -10,10 +7,10 @@ fn is_nan_f32(value: f32) -> bool {
   let bits = bitcast<u32>(value);
   return (bits & 0x7f800000u) == 0x7f800000u && (bits & 0x007fffffu) != 0u;
 }{% endmacro %}
-{% if (source.op == "clip" and scalar != "i32" and scalar != "u32") or (source.op == "relu" and scalar != "i32") or source.op == "sin" or source.op == "tan" %}
 {{ emit_is_nan_f32() }}
 {% endif %}
-{% if source.op == "relu" and scalar != "i32" %}
 fn relu_value(value: f32) -> f32 {
   // Some shader compilers apply no-NaN fast-math to vector select/compare
   // expressions. Inspecting the IEEE payload explicitly keeps Relu's required
@@ -25,14 +22,33 @@ fn relu_value(value: f32) -> f32 {
 }
 {% endif %}
 @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" %}
   y[i] = select(xv, vec4<i32>(0i), xv < vec4<i32>(0i));
@@ -43,4 +59,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups)
   y[i] = {{ vectorScalar }}(vec4<f32>(
     relu_value(fv.x), relu_value(fv.y), relu_value(fv.z), relu_value(fv.w)));
 {% endif %}
 }

+// Loads and stores vec4<T> while evaluating the selected unary operation per
+// component.
 {{ env.wgsl.resourceDeclarations }}
 {% macro emit_is_nan_f32() %}
   let bits = bitcast<u32>(value);
   return (bits & 0x7f800000u) == 0x7f800000u && (bits & 0x007fffffu) != 0u;
 }{% endmacro %}
+{% if (op == "clip" and scalar != "i32" and scalar != "u32") or (op == "relu" and scalar != "i32") or op == "sin" or op == "tan" %}
 {{ emit_is_nan_f32() }}
 {% endif %}
+{% if op == "relu" and scalar != "i32" %}
 fn relu_value(value: f32) -> f32 {
   // Some shader compilers apply no-NaN fast-math to vector select/compare
   // expressions. Inspecting the IEEE payload explicitly keeps Relu's required
 }
 {% endif %}
+{% set vec4PerThread = vec4PerThread %}
+{% if vec4PerThread > 1 %}
+const ITEMS: u32 = {{ vec4PerThread }}u;
+{% endif %}
 @compute @workgroup_size({{ tunables.WORKGROUP_SIZE }})
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
   // 2D-folded flat index: gid.y carries the high bits when the element count exceeds the
+  // per-axis dispatch fold width (the dispatch caps x and spills the rest into y).
+{% if vec4PerThread > 1 %}
+  // Each invocation walks ITEMS vec4 groups a span apart. Consecutive lanes
+  // access consecutive words on every step, while each lane can keep several
+  // independent loads in flight.
+  let tid = gid.x + gid.y * {{ DISPATCH_FOLD_WIDTH }}u * {{ tunables.WORKGROUP_SIZE }}u;
+  let span = (params.count + ITEMS - 1u) / ITEMS;
+  for (var j = 0u; j < ITEMS; j = j + 1u) {
+    let i = tid + j * span;
+    if (i >= params.count) {
+      break;
+    }
+{% else %}
+  let i = gid.x + gid.y * {{ DISPATCH_FOLD_WIDTH }}u * {{ tunables.WORKGROUP_SIZE }}u;
   if (i >= params.count) {
     return;
   }
+{% endif %}
   let xv = x[i];
 {% if scalar == "i32" %}
   y[i] = select(xv, vec4<i32>(0i), xv < vec4<i32>(0i));
   y[i] = {{ vectorScalar }}(vec4<f32>(
     relu_value(fv.x), relu_value(fv.y), relu_value(fv.z), relu_value(fv.w)));
 {% endif %}
+{% if vec4PerThread > 1 %}
+  }
+{% endif %}
 }