sync 91d990483a17

Browse files

Files changed (7) hide show

README.md +17 -14
build/webgpu/bench.json +0 -1
build/webgpu/datamove-elementwise-copy.wgsl.jinja +0 -3
build/webgpu/datamove-flat-copy.wgsl.jinja +1 -4
build/webgpu/manifest.json +26 -71
build/webgpu/metadata.json +11 -8
build/webgpu/test.json +13 -14

README.md CHANGED Viewed

@@ -12,22 +12,22 @@ tags:
 ## Description
-Reshapes the input tensor to a new shape while preserving the total number of elements. In this preallocated kernel ABI, the caller must resolve and validate the target shape—including `-1`, zero dimensions, and `allowzero`—before allocating `reshaped`; the kernel copies the elements and requires the input and output element counts to match.
 See the [ONNX `Reshape` spec](https://onnx.ai/onnx/operators/onnx__Reshape.html) for the reference semantics.
 ## Inputs
-| Name | Bind key | Logical dtype | WebGPU storage | Rank | Shape | Description | Presence |
-| --- | --- | --- | --- | --- | --- | --- | --- |
-| `data` | `data` | `T` | runtime-selected; narrow integers and bool use 32-bit slots | — | — | The input tensor to reshape. | required |
-| `shape` | `shape` | `S` | `int32` | `1` | — | Logical int64 1-D ONNX target-shape tensor, stored as int32 by WebGPU so `-1` remains representable. This standalone kernel does not inspect its values; the caller must validate them and allocate `reshaped` with the resolved shape. | required |
 ## Outputs
-| Name | Bind key | Logical dtype | Rank | Shape | Description | Presence |
-| --- | --- | --- | --- | --- | --- | --- |
-| `reshaped` | `reshaped` | `T` | derived | — | The reshaped output tensor with the caller-resolved target shape and the same elements and data type as the input. | required |
 ## Attributes
@@ -35,7 +35,7 @@ Default values (overridable per request):
 | Attribute | Default | Description |
 | --- | --- | --- |
-| `allowzero` | `0` | Controls caller-side target-shape resolution: `0` copies the corresponding input dimension for a zero entry, while `1` preserves a literal zero dimension. The standalone kernel does not interpret this attribute or the `shape` tensor. |
 ## Type constraints
@@ -46,7 +46,7 @@ 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
@@ -55,15 +55,18 @@ Default values (overridable per request):
 ## Use with `@huggingface/kernels`
-The loader automatically allocates outputs whose metadata it can derive from the manifest contract and this call.
-The explicit `outputs` entries provide shape and logical dtype metadata for the results listed below:
-- `reshaped`
-Each entry either requests an optional result or supplies metadata that cannot be inferred from the inputs.
 The `version: 1` option selects the published kernel contract; it is independent of any operator opset, contrib `since_version`, or model version.
 Replace each `*Data` placeholder with a typed array containing the corresponding input data.

 ## Description
+Reshapes the input tensor while preserving its elements. The target shape—including `-1`, zero dimensions, and `allowzero`—must be resolved and validated before `reshaped` is allocated; input and output element counts must match.
 See the [ONNX `Reshape` spec](https://onnx.ai/onnx/operators/onnx__Reshape.html) for the reference semantics.
 ## Inputs
+| Name | Logical dtype | WebGPU storage | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- | --- |
+| `data` | `T` | runtime-selected; narrow integers and bool use 32-bit slots | — | — | The input tensor to reshape. | required |
+| `shape` | `S` | `int32` | `1` | — | Logical int64 1-D ONNX target-shape tensor, stored as int32 by WebGPU so `-1` remains representable. The values must be validated and `reshaped` allocated with the resolved shape before dispatch. | required |
 ## Outputs
+| Name | Logical dtype | Rank | Shape | Description | Presence |
+| --- | --- | --- | --- | --- | --- |
+| `reshaped` | `T` | derived | — | The output tensor with the resolved target shape and the same elements and data type as the input. | required |
 ## Attributes
 | Attribute | Default | Description |
 | --- | --- | --- |
+| `allowzero` | `0` | Controls target-shape resolution: `0` copies the corresponding input dimension for a zero entry, while `1` preserves a literal zero dimension. The standalone kernel does not interpret this attribute or the `shape` tensor. |
 ## 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
+```
+Outputs with inferable metadata are allocated automatically. Explicit `outputs` entries request optional results or provide metadata that cannot be inferred from the supplied inputs and attributes.
+This example supplies explicit metadata for:
+- `reshaped`
 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.Reshape",
   "tunableSpace": { "WORKGROUP_SIZE": [64, 128, 256] },
   "cases": [
     {

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

build/webgpu/datamove-elementwise-copy.wgsl.jinja CHANGED Viewed

@@ -1,6 +1,3 @@
-{% if usesF16 %}
-enable f16;
-{% endif %}
 {{ env.wgsl.resourceDeclarations }}
 const WG: u32 = {{ tunables.WORKGROUP_SIZE }}u;





1	{{ env.wgsl.resourceDeclarations }}
2
3	const WG: u32 = {{ tunables.WORKGROUP_SIZE }}u;

build/webgpu/datamove-flat-copy.wgsl.jinja CHANGED Viewed

@@ -1,9 +1,6 @@
-{% if usesF16 %}
-enable f16;
-{% endif %}
 {{ env.wgsl.resourceDeclarations }}
-const COUNT: u32 = {{ source.count }}u;
 const WG: u32 = {{ tunables.WORKGROUP_SIZE }}u;
 @compute @workgroup_size({{ tunables.WORKGROUP_SIZE }})

 {{ env.wgsl.resourceDeclarations }}
+const COUNT: u32 = {{ count }}u;
 const WG: u32 = {{ tunables.WORKGROUP_SIZE }}u;
 @compute @workgroup_size({{ tunables.WORKGROUP_SIZE }})

build/webgpu/manifest.json CHANGED Viewed

@@ -2,108 +2,63 @@
   "domain": "ai.onnx",
   "name": "Reshape",
   "sinceVersion": 14,
-  "description": "Reshapes the input tensor to a new shape while preserving the total number of elements. In this preallocated kernel ABI, the caller must resolve and validate the target shape—including `-1`, zero dimensions, and `allowzero`—before allocating `reshaped`; the kernel copies the elements and requires the input and output element counts to match.",
-  "inputs": [
-    { "role": "data", "dtype": "T", "description": "The input tensor to reshape." },
-    {
-      "role": "shape",
-      "dtype": "S",
-      "description": "Logical int64 1-D ONNX target-shape tensor, stored as int32 by WebGPU so `-1` remains representable. This standalone kernel does not inspect its values; the caller must validate them and allocate `reshaped` with the resolved shape.",
-      "rank": 1
-    }
-  ],
-  "outputs": [
-    {
-      "role": "reshaped",
-      "dtype": "T",
-      "description": "The reshaped output tensor with the caller-resolved target shape and the same elements and data type as the input.",
-      "rank": "dim(shapes.shape, 0)"
-    }
-  ],
-  "attributes": { "allowzero": 0 },
-  "attributeDescriptions": {
-    "allowzero": "Controls caller-side target-shape resolution: `0` copies the corresponding input dimension for a zero entry, while `1` preserves a literal zero dimension. The standalone kernel does not interpret this attribute or the `shape` tensor."
   },
   "attributeConstraints": { "allowzero": { "values": [0, 1] } },
   "typeConstraints": {
     "T": ["float32", "float16", "int32", "int16", "uint32", "uint8", "int8", "bool"],
     "S": ["int64"]
   },
-  "args": {
-    "data": { "kind": "tensor", "semantic": "data", "role": "input" },
-    "shape": { "kind": "tensor", "semantic": "shape", "role": "shape", "dtype": "int32", "narrowing": "checked" },
-    "reshaped": { "kind": "tensor", "semantic": "reshaped", "role": "output" }
-  },
-  "tunables": { "WORKGROUP_SIZE": 256 },
   "derive": { "reshapeModeOk": "attrs.allowzero == 0 or attrs.allowzero == 1" },
-  "constants": { "usesF16": "dtypes.T == \"f16\"" },
   "variants": [
     {
       "id": "copy_vec4",
       "priority": 20,
       "when": ["reshapeModeOk", "numel(shapes.data) > 0", "numel(shapes.data) % 4 == 0", "numel(shapes.data) == numel(shapes.reshaped)", "f16Ok(dtypes.T)"],
-      "constants": { "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"" },
       "passes": [
         {
           "id": "main",
           "name": "Reshape.copyVec4",
-          "source": { "shader": "datamove-flat-copy.wgsl.jinja", "inputs": { "count": "numel(shapes.reshaped) / 4" } },
           "bindings": [
-            {
-              "name": "src",
-              "arg": "data",
-              "semantic": "data",
-              "buffer": { "type": "read-only-storage" },
-              "elementType": "$vectorScalar"
-            },
-            {
-              "name": "dst",
-              "arg": "reshaped",
-              "semantic": "reshaped",
-              "buffer": { "type": "storage" },
-              "elementType": "$vectorScalar"
-            }
           ],
-          "dispatch": { "gridStride": "numel(shapes.reshaped) / 4", "workgroupSize": "tunables.WORKGROUP_SIZE" },
           "viewAlias": [{ "input": "src", "output": "dst" }]
         }
       ]
     },
     {
       "id": "copy",
-      "when": "reshapeModeOk and numel(shapes.data) == numel(shapes.reshaped) and f16Ok(dtypes.T)",
-      "constants": { "scalar": "dtypes.T" },
       "passes": [
         {
           "id": "main",
           "name": "Reshape",
           "shader": "datamove-elementwise-copy.wgsl.jinja",
           "bindings": [
-            {
-              "name": "x",
-              "arg": "data",
-              "semantic": "data",
-              "buffer": { "type": "read-only-storage" },
-              "elementType": "$scalar"
-            },
-            {
-              "name": "y",
-              "arg": "reshaped",
-              "semantic": "reshaped",
-              "buffer": { "type": "storage" },
-              "elementType": "$scalar"
-            },
-            {
-              "name": "params",
-              "semantic": "kernel.params",
-              "buffer": { "type": "uniform" },
-              "struct": {
-                "name": "Params",
-                "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.reshaped)" }]
-              }
-            }
           ],
-          "dispatch": { "gridStride": "numel(shapes.reshaped)", "workgroupSize": "tunables.WORKGROUP_SIZE" },
           "viewAlias": [{ "input": "x", "output": "y" }]
         }
       ]

   "domain": "ai.onnx",
   "name": "Reshape",
   "sinceVersion": 14,
+  "inputs": {
+    "data": { "dtype": "T" },
+    "shape": { "dtype": "S", "rank": 1, "storage": "int32", "narrowing": "checked" }
   },
+  "outputs": { "reshaped": { "dtype": "T", "rank": "dim(shapes.shape, 0)" } },
+  "attributes": { "allowzero": { "default": 0 } },
   "attributeConstraints": { "allowzero": { "values": [0, 1] } },
   "typeConstraints": {
     "T": ["float32", "float16", "int32", "int16", "uint32", "uint8", "int8", "bool"],
     "S": ["int64"]
   },
+  "tunables": { "WORKGROUP_SIZE": { "default": 256 } },
   "derive": { "reshapeModeOk": "attrs.allowzero == 0 or attrs.allowzero == 1" },
   "variants": [
     {
       "id": "copy_vec4",
       "priority": 20,
       "when": ["reshapeModeOk", "numel(shapes.data) > 0", "numel(shapes.data) % 4 == 0", "numel(shapes.data) == numel(shapes.reshaped)", "f16Ok(dtypes.T)"],
+      "derive": { "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"" },
       "passes": [
         {
           "id": "main",
           "name": "Reshape.copyVec4",
+          "shader": "datamove-flat-copy.wgsl.jinja",
+          "derive": { "count": "numel(shapes.reshaped) / 4" },
           "bindings": [
+            { "arg": "data", "name": "src", "elementType": "$vectorScalar" },
+            { "arg": "reshaped", "name": "dst", "elementType": "$vectorScalar" }
           ],
+          "dispatch": {
+            "x": "min(ceilDiv((numel(shapes.reshaped) / 4), (tunables.WORKGROUP_SIZE)), min(device.limits.maxComputeWorkgroupsPerDimension, 65535))",
+            "y": 1,
+            "z": 1
+          },
           "viewAlias": [{ "input": "src", "output": "dst" }]
         }
       ]
     },
     {
       "id": "copy",
+      "when": ["reshapeModeOk and numel(shapes.data) == numel(shapes.reshaped) and f16Ok(dtypes.T)"],
+      "derive": { "scalar": "dtypes.T" },
       "passes": [
         {
           "id": "main",
           "name": "Reshape",
           "shader": "datamove-elementwise-copy.wgsl.jinja",
           "bindings": [
+            { "arg": "data", "name": "x", "elementType": "$scalar" },
+            { "arg": "reshaped", "name": "y", "elementType": "$scalar" },
+            { "name": "params", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.reshaped)" }] }
           ],
+          "dispatch": {
+            "x": "min(ceilDiv((numel(shapes.reshaped)), (tunables.WORKGROUP_SIZE)), min(device.limits.maxComputeWorkgroupsPerDimension, 65535))",
+            "y": 1,
+            "z": 1
+          },
           "viewAlias": [{ "input": "x", "output": "y" }]
         }
       ]

build/webgpu/metadata.json CHANGED Viewed

@@ -1,19 +1,22 @@
 {
   "name": "ai.onnx.Reshape",
-  "id": "_ai_onnx_reshape_webgpu_65ee152",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
   "digest": {
     "algorithm": "sha256",
     "files": {
-      "bench.json": "69A1NMlREheggpSNj8htmManeB/z0HH/QNeacSzURGI=",
-      "datamove-elementwise-copy.wgsl.jinja": "J5yC2bAddPiP+odXLgVGS3TJ9jeNsfRTvedKrj/fhZg=",
-      "datamove-flat-copy.wgsl.jinja": "rEpeYMvIZQuQnB2f3lt51jrrklKICsqUTzCJPRTknNk=",
-      "manifest.json": "61gkdCMoUp39wr9rBxQ0oncJwE5qd9Az+fRQDDEOGB0=",
-      "test.json": "+cht0hPHFoAOvZ6pfI97ReMTuoOgC0MZRnnXDu4549c="
     }
   },
-  "provenance": { "kernel": { "sha": "2e7068faf55e7f43df740015f6d1ee49391a41c5", "dirty": false } },
-  "webgpu": { "manifestSpec": "1.0", "specialized": true, "opPath": "ops/ai.onnx.Reshape" }
 }

 {
   "name": "ai.onnx.Reshape",
+  "id": "_ai_onnx_reshape_webgpu_c55e31e",
   "version": 1,
   "license": "Apache-2.0",
   "backend": { "type": "webgpu" },
   "digest": {
     "algorithm": "sha256",
     "files": {
+      "bench.json": "AAAfqBx6yJ8rtDAPF9mbztPnteMhWDLHiIiJJf1Mc9U=",
+      "datamove-elementwise-copy.wgsl.jinja": "Q1WuZCqcDf9b6cbT4rLilpeWj6327t/bSFmKr6IVBoQ=",
+      "datamove-flat-copy.wgsl.jinja": "g9d62mer2bmHfbScqX5CIT5Zq/Pdinac0W2dH0CD+5s=",
+      "manifest.json": "WrN4wPoSc9+xsh3es62hMt0yDgsN3Pl+4f//L/X0X4s=",
+      "test.json": "1XC5Gsp15I4tUcAsnBJkDezeVpvO/Pm+10uzzDJtedY="
     }
   },
+  "provenance": { "kernel": { "sha": "91d990483a174128daf7673f3f37a7c890493ae1", "dirty": false } },
+  "webgpu": {
+    "manifestSpec": "2.0",
+    "variants": { "copy_vec4": ["datamove-flat-copy.wgsl.jinja"], "copy": ["datamove-elementwise-copy.wgsl.jinja"] }
+  }
 }

build/webgpu/test.json CHANGED Viewed

@@ -1,5 +1,4 @@
 {
-  "op": "ai.onnx.Reshape",
   "fixtureArrays": {
     "onnx_backend_input_data": [0.54881352186203, 0.7151893377304077, 0.6027633547782898, 0.5448831915855408, 0.42365479469299316, 0.6458941102027893, 0.4375872015953064, 0.891772985458374, 0.9636627435684204, 0.3834415078163147, 0.7917250394821167, 0.5288949012756348, 0.5680445432662964, 0.9255966544151306, 0.07103605568408966, 0.08712930232286453, 0.020218396559357643, 0.832619845867157, 0.7781567573547363, 0.8700121641159058, 0.978618323802948, 0.7991585731506348, 0.4614793658256531, 0.7805292010307312]
   },
@@ -202,7 +201,7 @@
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/tensor/tensor_op_test.cc",
         "test": "TensorOpTest.Reshape",
-        "notes": "Same shape transform as ORT's basic Reshape test, using supported int32 data storage."
       },
       "inputs": {
         "data": { "dtype": "int32", "shape": [2, 3], "data": { "kind": "values", "values": [1, 2, 3, 4, 5, 6] } },
@@ -215,7 +214,7 @@
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/tensor/tensor_op_test.cc",
         "test": "TensorOpTest.Reshape_UnknownDimWithoutAllowZero",
-        "notes": "ONNX int64 shape tensor represented with framework int32 slots."
       },
       "inputs": {
         "data": {
@@ -232,7 +231,7 @@
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/tensor/tensor_op_test.cc",
         "test": "TensorOpTest.ReshapeWithEmptyInputAndDynamicShape",
-        "notes": "The ORT success subcase reshapes {1,0} using {1,1,-1} to {1,1,0}; the sibling {1,0,-1} case is an expected failure and is not encoded here."
       },
       "inputs": {
         "data": { "dtype": "float32", "shape": [1, 0], "data": { "kind": "values", "values": [] } },
@@ -249,7 +248,7 @@
       "outputs": { "reshaped": { "dtype": "float32", "shape": [3, 4, 0] } },
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_allowzero_reordered",
-        "notes": "ONNX int64 metadata/index tensors use framework int32/uint32 slots where representable. ONNX Runtime validates `allowzero=1` and resolves `[3, 4, 0]`; the standalone WebGPU copy consumes the caller-preallocated `[3, 4, 0]` output and does not inspect the shape tensor."
       },
       "attrs": { "allowzero": 1 }
     },
@@ -266,7 +265,7 @@
       "outputs": { "reshaped": { "dtype": "float32", "shape": [2, 3, 2, 2] } },
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_extended_dims",
-        "notes": "ONNX int64 metadata/index tensors use framework int32/uint32 slots where representable."
       }
     },
     {
@@ -282,7 +281,7 @@
       "outputs": { "reshaped": { "dtype": "float32", "shape": [24] } },
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_one_dim",
-        "notes": "ONNX int64 metadata/index tensors use framework int32/uint32 slots where representable."
       }
     },
     {
@@ -298,7 +297,7 @@
       "outputs": { "reshaped": { "dtype": "float32", "shape": [2, 12] } },
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_reduced_dims",
-        "notes": "ONNX int64 metadata/index tensors use framework int32/uint32 slots where representable."
       }
     },
     {
@@ -314,7 +313,7 @@
       "outputs": { "reshaped": { "dtype": "float32", "shape": [4, 2, 3] } },
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_reordered_all_dims",
-        "notes": "ONNX int64 metadata/index tensors use framework int32/uint32 slots where representable."
       }
     },
     {
@@ -330,7 +329,7 @@
       "outputs": { "reshaped": { "dtype": "float32", "shape": [2, 4, 3] } },
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_reordered_last_dims",
-        "notes": "ONNX int64 metadata/index tensors use framework int32/uint32 slots where representable."
       }
     },
     {
@@ -346,14 +345,14 @@
       "outputs": { "reshaped": { "dtype": "float32", "shape": [2, 3, 4, 1] } },
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_zero_dim",
-        "notes": "ONNX int64 metadata/index tensors use framework int32/uint32 slots where representable."
       }
     },
     {
       "name": "onnx_backend_negative_dim_infer_middle",
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_negative_dim",
-        "notes": "Official shape tensor is int64; this framework stores signed Reshape metadata as int32."
       },
       "inputs": {
         "data": {
@@ -369,7 +368,7 @@
       "name": "onnx_backend_negative_dim_infer_leading",
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_negative_extended_dims",
-        "notes": "Official shape tensor is int64; this framework stores signed Reshape metadata as int32."
       },
       "inputs": {
         "data": {
@@ -385,7 +384,7 @@
       "name": "onnx_backend_zero_copy_and_negative_dim",
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_zero_and_negative_dim",
-        "notes": "Official shape tensor is int64; this framework stores signed Reshape metadata as int32."
       },
       "inputs": {
         "data": {

 {
   "fixtureArrays": {
     "onnx_backend_input_data": [0.54881352186203, 0.7151893377304077, 0.6027633547782898, 0.5448831915855408, 0.42365479469299316, 0.6458941102027893, 0.4375872015953064, 0.891772985458374, 0.9636627435684204, 0.3834415078163147, 0.7917250394821167, 0.5288949012756348, 0.5680445432662964, 0.9255966544151306, 0.07103605568408966, 0.08712930232286453, 0.020218396559357643, 0.832619845867157, 0.7781567573547363, 0.8700121641159058, 0.978618323802948, 0.7991585731506348, 0.4614793658256531, 0.7805292010307312]
   },
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/tensor/tensor_op_test.cc",
         "test": "TensorOpTest.Reshape",
+        "notes": "A basic Reshape transform uses supported int32 payload storage."
       },
       "inputs": {
         "data": { "dtype": "int32", "shape": [2, 3], "data": { "kind": "values", "values": [1, 2, 3, 4, 5, 6] } },
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/tensor/tensor_op_test.cc",
         "test": "TensorOpTest.Reshape_UnknownDimWithoutAllowZero",
+        "notes": "The int32 shape tensor exactly represents the source ONNX int64 values."
       },
       "inputs": {
         "data": {
       "provenance": {
         "source": "onnxruntime/test/providers/cpu/tensor/tensor_op_test.cc",
         "test": "TensorOpTest.ReshapeWithEmptyInputAndDynamicShape",
+        "notes": "Input shape {1,0} with requested shape {1,1,-1} infers the zero-length dimension and produces {1,1,0}."
       },
       "inputs": {
         "data": { "dtype": "float32", "shape": [1, 0], "data": { "kind": "values", "values": [] } },
       "outputs": { "reshaped": { "dtype": "float32", "shape": [3, 4, 0] } },
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_allowzero_reordered",
+        "notes": "The integer shape tensor exactly represents the source ONNX int64 values. With `allowzero=1`, the declared output shape is [3,4,0], so the empty copy performs no element reads."
       },
       "attrs": { "allowzero": 1 }
     },
       "outputs": { "reshaped": { "dtype": "float32", "shape": [2, 3, 2, 2] } },
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_extended_dims",
+        "notes": "The integer shape tensor exactly represents the source ONNX int64 values."
       }
     },
     {
       "outputs": { "reshaped": { "dtype": "float32", "shape": [24] } },
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_one_dim",
+        "notes": "The integer shape tensor exactly represents the source ONNX int64 values."
       }
     },
     {
       "outputs": { "reshaped": { "dtype": "float32", "shape": [2, 12] } },
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_reduced_dims",
+        "notes": "The integer shape tensor exactly represents the source ONNX int64 values."
       }
     },
     {
       "outputs": { "reshaped": { "dtype": "float32", "shape": [4, 2, 3] } },
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_reordered_all_dims",
+        "notes": "The integer shape tensor exactly represents the source ONNX int64 values."
       }
     },
     {
       "outputs": { "reshaped": { "dtype": "float32", "shape": [2, 4, 3] } },
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_reordered_last_dims",
+        "notes": "The integer shape tensor exactly represents the source ONNX int64 values."
       }
     },
     {
       "outputs": { "reshaped": { "dtype": "float32", "shape": [2, 3, 4, 1] } },
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_zero_dim",
+        "notes": "The integer shape tensor exactly represents the source ONNX int64 values."
       }
     },
     {
       "name": "onnx_backend_negative_dim_infer_middle",
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_negative_dim",
+        "notes": "Projects signed int64 shape metadata to int32 storage."
       },
       "inputs": {
         "data": {
       "name": "onnx_backend_negative_dim_infer_leading",
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_negative_extended_dims",
+        "notes": "Projects signed int64 shape metadata to int32 storage."
       },
       "inputs": {
         "data": {
       "name": "onnx_backend_zero_copy_and_negative_dim",
       "provenance": {
         "source": "cmake/external/onnx/onnx/backend/test/data/node/test_reshape_zero_and_negative_dim",
+        "notes": "Projects signed int64 shape metadata to int32 storage."
       },
       "inputs": {
         "data": {