sync 91d990483a17

Browse files

Files changed (6) hide show

README.md +13 -9
build/webgpu/bench.json +0 -1
build/webgpu/bitcast.wgsl.jinja +6 -6
build/webgpu/manifest.json +33 -85
build/webgpu/metadata.json +10 -7
build/webgpu/test.json +2 -3

README.md CHANGED Viewed

@@ -18,15 +18,15 @@ See the [ONNX `BitCast` spec](https://onnx.ai/onnx/operators/onnx__BitCast.html)
 ## Inputs
-| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
-| --- | --- | --- | --- | --- | --- | --- |
-| `input` | `input` | `T` | — | — | Input tensor to be bitwise reinterpreted. | required |
 ## Outputs
-| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
-| --- | --- | --- | --- | --- | --- | --- |
-| `output` | `output` | `U` | same as `input` | same as `input` | Output tensor with the same shape as the input, reinterpreted as the target type. | required |
 ## Attributes
@@ -45,7 +45,7 @@ Attributes and default values (overridable per request):
 ## 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
@@ -53,10 +53,14 @@ Attributes and default values (overridable per request):
 ## 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 | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- |
+| `input` | `T` | — | — | Input tensor to be bitwise reinterpreted. | required |
 ## Outputs
+| Name | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- |
+| `output` | `U` | same as `input` | same as `input` | Output tensor with the same shape as the input, reinterpreted as the target type. | required |
 ## Attributes
 ## 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.BitCast",
   "cases": [
     {
       "name": "bitcast-f32-to-u32-33m",

 {
   "cases": [
     {
       "name": "bitcast-f32-to-u32-33m",

build/webgpu/bitcast.wgsl.jinja CHANGED Viewed

@@ -1,14 +1,14 @@
 {{ env.wgsl.resourceDeclarations }}
-{% set vectorized = source.vectorized if source.vectorized is defined else false %}
 // Each invocation reinterprets one 32-bit slot, or four contiguous slots on
-// the vectorized route. Framework-widened int8/uint8 values retain their
-// explicit low-byte masking and sign extension.
-@compute @workgroup_size({{ source.workgroupSize }})
-fn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups) nwg: vec3<u32>) {
   // The flat dispatch is folded across x/y at the device limit, so gid.y
   // carries the high portion of the slot index.
-  let i = gid.x + gid.y * nwg.x * {{ source.workgroupSize }}u;
   if (i >= params.count) {
     return;
   }

 {{ env.wgsl.resourceDeclarations }}
+{% set vectorized = vectorizedSpec if vectorizedSpec is defined else false %}
 // Each invocation reinterprets one 32-bit slot, or four contiguous slots on
+// the vectorized route. Logical int8/uint8 values stored in 32-bit slots retain
+// their explicit low-byte masking and sign extension.
+@compute @workgroup_size({{ workgroupSizeSpec }})
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
   // The flat dispatch is folded across x/y at the device limit, so gid.y
   // carries the high portion of the slot index.
+  let i = gid.x + gid.y * {{ DISPATCH_FOLD_WIDTH }}u * {{ workgroupSizeSpec }}u;
   if (i >= params.count) {
     return;
   }

build/webgpu/manifest.json CHANGED Viewed

@@ -2,40 +2,22 @@
   "domain": "ai.onnx",
   "name": "BitCast",
   "sinceVersion": 26,
-  "description": "Reinterprets the raw bit pattern of a tensor as a different data type without any value conversion. The target type must have the same bit-width as the input type, and the output tensor has the same shape as the input.",
-  "inputs": [{ "role": "input", "dtype": "T", "description": "Input tensor to be bitwise reinterpreted." }],
-  "outputs": [
-    {
-      "role": "output",
-      "dtype": "U",
-      "rank": "ranks.input",
-      "description": "Output tensor with the same shape as the input, reinterpreted as the target type.",
-      "shape": "shapes.input"
-    }
-  ],
-  "attributes": {},
-  "attributeDescriptions": {
-    "to": "Required TensorProto DataType enum integer naming the output dtype; the target type must have the same bit-width as the input type."
-  },
   "attributeConstraints": { "to": { "required": true } },
   "typeConstraints": {
     "T": ["float32", "int8", "int32", "uint8", "uint32"],
     "U": ["float32", "int8", "int32", "uint8", "uint32"]
   },
-  "args": {
-    "input": { "kind": "tensor", "semantic": "input", "role": "input" },
-    "output": { "kind": "tensor", "semantic": "output", "role": "output" }
-  },
-  "tunables": { "WORKGROUP_SIZE": 256 },
   "derive": {
     "deviceWorkgroupCap": "min(device.limits.maxComputeInvocationsPerWorkgroup, device.limits.maxComputeWorkgroupSizeX)",
     "bitcastShapeOk": "ranks.input >= 0 and ranks.output == ranks.input and numel(shapes.input) == numel(shapes.output)",
     "bitcastTypeOk": "attrs.to == onnxDtypeCode(logicalDtypes.U) and (((tensorDtypes.input == \"int8\" or tensorDtypes.input == \"uint8\") and (tensorDtypes.output == \"int8\" or tensorDtypes.output == \"uint8\")) or (tensorDtypes.input != \"int8\" and tensorDtypes.input != \"uint8\" and tensorDtypes.output != \"int8\" and tensorDtypes.output != \"uint8\"))",
     "bitcastWorkgroupSize": "min(tunables.WORKGROUP_SIZE, deviceWorkgroupCap)",
-    "bitcastDispatchFits": "ceilDiv(ceilDiv(numel(shapes.output), bitcastWorkgroupSize), device.limits.maxComputeWorkgroupsPerDimension) <= device.limits.maxComputeWorkgroupsPerDimension",
-    "bitcastBaseOk": "bitcastShapeOk and bitcastTypeOk and bitcastDispatchFits"
-  },
-  "constants": {
     "inScalar": "dtypes.T",
     "outScalar": "dtypes.U",
     "inputIsInt8": "tensorDtypes.input == \"int8\"",
@@ -43,84 +25,50 @@
     "outputIsInt8": "tensorDtypes.output == \"int8\"",
     "outputIsUint8": "tensorDtypes.output == \"uint8\""
   },
-  "bindingSets": {
-    "vec4Slots": [
-      {
-        "name": "input",
-        "arg": "input",
-        "semantic": "input",
-        "buffer": { "type": "read-only-storage" },
-        "elementType": "$inVec4"
-      },
-      {
-        "name": "output",
-        "arg": "output",
-        "semantic": "output",
-        "buffer": { "type": "storage" },
-        "elementType": "$outVec4"
-      },
-      {
-        "name": "params",
-        "semantic": "kernel.params",
-        "buffer": { "type": "uniform" },
-        "struct": {
-          "name": "Params",
-          "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.output) / 4" }]
-        }
-      }
-    ],
-    "scalarSlots": [
-      {
-        "name": "input",
-        "arg": "input",
-        "semantic": "input",
-        "buffer": { "type": "read-only-storage" },
-        "elementType": "$inScalar"
-      },
-      {
-        "name": "output",
-        "arg": "output",
-        "semantic": "output",
-        "buffer": { "type": "storage" },
-        "elementType": "$outScalar"
-      },
-      {
-        "name": "params",
-        "semantic": "kernel.params",
-        "buffer": { "type": "uniform" },
-        "struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.output)" }] }
-      }
-    ]
-  },
   "variants": [
     {
       "id": "slot32_vec4",
       "priority": 20,
-      "when": ["numel(shapes.input) % 4 == 0", "bitcastBaseOk"],
-      "constants": { "inVec4": "\"vec4<\" ~ dtypes.T ~ \">\"", "outVec4": "\"vec4<\" ~ dtypes.U ~ \">\"" },
       "passes": [
         {
           "id": "main",
           "name": "BitCast.vec4",
-          "source": {
-            "shader": "bitcast.wgsl.jinja",
-            "inputs": { "vectorized": true, "workgroupSize": "bitcastWorkgroupSize" }
-          },
-          "bindings": "vec4Slots",
-          "dispatch": { "threads": "numel(shapes.output) / 4", "workgroupSize": "bitcastWorkgroupSize" }
         }
       ]
     },
     {
       "id": "slot32",
-      "when": ["bitcastBaseOk"],
       "passes": [
         {
           "id": "main",
           "name": "BitCast",
-          "source": { "shader": "bitcast.wgsl.jinja", "inputs": { "workgroupSize": "bitcastWorkgroupSize" } },
-          "bindings": "scalarSlots",
-          "dispatch": { "threads": "numel(shapes.output)", "workgroupSize": "bitcastWorkgroupSize" }
         }
       ]
     }

   "domain": "ai.onnx",
   "name": "BitCast",
   "sinceVersion": 26,
+  "inputs": { "input": { "dtype": "T" } },
+  "outputs": { "output": { "dtype": "U", "rank": "ranks.input", "shape": "shapes.input" } },
+  "attributes": { "to": {} },
   "attributeConstraints": { "to": { "required": true } },
   "typeConstraints": {
     "T": ["float32", "int8", "int32", "uint8", "uint32"],
     "U": ["float32", "int8", "int32", "uint8", "uint32"]
   },
+  "tunables": { "WORKGROUP_SIZE": { "default": 256 } },
   "derive": {
     "deviceWorkgroupCap": "min(device.limits.maxComputeInvocationsPerWorkgroup, device.limits.maxComputeWorkgroupSizeX)",
     "bitcastShapeOk": "ranks.input >= 0 and ranks.output == ranks.input and numel(shapes.input) == numel(shapes.output)",
     "bitcastTypeOk": "attrs.to == onnxDtypeCode(logicalDtypes.U) and (((tensorDtypes.input == \"int8\" or tensorDtypes.input == \"uint8\") and (tensorDtypes.output == \"int8\" or tensorDtypes.output == \"uint8\")) or (tensorDtypes.input != \"int8\" and tensorDtypes.input != \"uint8\" and tensorDtypes.output != \"int8\" and tensorDtypes.output != \"uint8\"))",
     "bitcastWorkgroupSize": "min(tunables.WORKGROUP_SIZE, deviceWorkgroupCap)",
+    "bitcastDispatchFits": "ceilDiv(ceilDiv(numel(shapes.output), bitcastWorkgroupSize), min(device.limits.maxComputeWorkgroupsPerDimension, 65535)) <= min(device.limits.maxComputeWorkgroupsPerDimension, 65535)",
+    "bitcastBaseOk": "bitcastShapeOk and bitcastTypeOk and bitcastDispatchFits",
     "inScalar": "dtypes.T",
     "outScalar": "dtypes.U",
     "inputIsInt8": "tensorDtypes.input == \"int8\"",
     "outputIsInt8": "tensorDtypes.output == \"int8\"",
     "outputIsUint8": "tensorDtypes.output == \"uint8\""
   },
+  "when": ["bitcastBaseOk"],
   "variants": [
     {
       "id": "slot32_vec4",
       "priority": 20,
+      "when": ["numel(shapes.input) % 4 == 0"],
+      "derive": { "inVec4": "\"vec4<\" ~ dtypes.T ~ \">\"", "outVec4": "\"vec4<\" ~ dtypes.U ~ \">\"" },
       "passes": [
         {
           "id": "main",
           "name": "BitCast.vec4",
+          "shader": "bitcast.wgsl.jinja",
+          "derive": { "vectorizedSpec": true, "workgroupSizeSpec": "bitcastWorkgroupSize" },
+          "bindings": [
+            { "arg": "input", "elementType": "$inVec4" },
+            { "arg": "output", "elementType": "$outVec4" },
+            { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.output) / 4" }] }
+          ],
+          "dispatch": {
+            "x": "min(ceilDiv((numel(shapes.output) / 4), (bitcastWorkgroupSize)), 65535)",
+            "y": "ceilDiv(ceilDiv((numel(shapes.output) / 4), (bitcastWorkgroupSize)), 65535)",
+            "z": 1
+          }
         }
       ]
     },
     {
       "id": "slot32",
       "passes": [
         {
           "id": "main",
           "name": "BitCast",
+          "shader": "bitcast.wgsl.jinja",
+          "derive": { "workgroupSizeSpec": "bitcastWorkgroupSize" },
+          "bindings": [
+            "input",
+            "output",
+            { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.output)" }] }
+          ],
+          "dispatch": {
+            "x": "min(ceilDiv((numel(shapes.output)), (bitcastWorkgroupSize)), 65535)",
+            "y": "ceilDiv(ceilDiv((numel(shapes.output)), (bitcastWorkgroupSize)), 65535)",
+            "z": 1
+          }
         }
       ]
     }

build/webgpu/metadata.json CHANGED Viewed

@@ -1,18 +1,21 @@
 {
   "name": "ai.onnx.BitCast",
-  "id": "_ai_onnx_bitcast_webgpu_953c2be",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
   "digest": {
     "algorithm": "sha256",
     "files": {
-      "bench.json": "lVPh5n2BwzVpRwCSVTTJRFxMXRTPmVeRM1Iy1ZVmjag=",
-      "bitcast.wgsl.jinja": "lXNQpNlvG/xrJll5JpnrFRGYtGajmNiP3uv9e846ObQ=",
-      "manifest.json": "gj3f/7xbrjxRy/dTZy9BN+nmJNOqanpBjiwjeIoFuFk=",
-      "test.json": "bz92JkCXkvPo9upgVfsjbWlUBWrz1pIqKHUyZS/K2uA="
     }
   },
-  "provenance": { "kernel": { "sha": "2e7068faf55e7f43df740015f6d1ee49391a41c5", "dirty": false } },
-  "webgpu": { "manifestSpec": "1.0", "specialized": true, "opPath": "ops/ai.onnx.BitCast" }
 }

 {
   "name": "ai.onnx.BitCast",
+  "id": "_ai_onnx_bitcast_webgpu_ab905c6",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
   "digest": {
     "algorithm": "sha256",
     "files": {
+      "bench.json": "7aNcJ8LB0eyjvFh7jXkwAw1p2vgZejuWw6H9iwqbgqY=",
+      "bitcast.wgsl.jinja": "1vfOUaYUxSx36bWOnKDPyXUPGprDxhPVEv10804xfVI=",
+      "manifest.json": "UplDa7horbJakWMkW0Gkocun6e1kcewHPZAvi6weNHs=",
+      "test.json": "FXueeEbCTZUCVYOfr0NjJrQAoVhhvtwqADXKZI6829c="
     }
   },
+  "provenance": { "kernel": { "sha": "91d990483a174128daf7673f3f37a7c890493ae1", "dirty": false } },
+  "webgpu": {
+    "manifestSpec": "2.0",
+    "variants": { "slot32_vec4": ["bitcast.wgsl.jinja"], "slot32": ["bitcast.wgsl.jinja"] }
+  }
 }

build/webgpu/test.json CHANGED Viewed

@@ -1,5 +1,4 @@
 {
-  "op": "ai.onnx.BitCast",
   "cases": [
     {
       "name": "float32_to_int32",
@@ -227,7 +226,7 @@
     {
       "name": "int8_identity_pairwise_route",
       "provenance": {
-        "notes": "Identity reinterpretation bridges the independently selectable signed and unsigned 8-bit input/output routes; the ORT-derived int8-to-uint8 sibling retains cross-signedness coverage."
       },
       "attrs": { "to": 3 },
       "inputs": {
@@ -247,7 +246,7 @@
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/tensor/bitcast_op_test.cc",
         "test": "BitCastTest.Int8ToUInt8",
-        "notes": "A four-element sibling pins the vectorized signed-to-unsigned byte reinterpretation as well as the scalar ORT fixture below."
       },
       "attrs": { "to": 2 },
       "inputs": {

 {
   "cases": [
     {
       "name": "float32_to_int32",
     {
       "name": "int8_identity_pairwise_route",
       "provenance": {
+        "notes": "Identity reinterpretation exercises signed 8-bit input and output selection independently while preserving every payload bit."
       },
       "attrs": { "to": 3 },
       "inputs": {
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/tensor/bitcast_op_test.cc",
         "test": "BitCastTest.Int8ToUInt8",
+        "notes": "Four elements exercise vectorized signed-to-unsigned byte reinterpretation while preserving every payload bit."
       },
       "attrs": { "to": 2 },
       "inputs": {