Hugging Face's logo

Kernels:
webgpu-kernels
/
ai.onnx.Pad

Kernel card Files
xet
 Community
Xenova HF Staff commited on
Commit
ad8a994
·
verified ·
1 Parent(s): b35cc85

sync 6fdf6301e2bb

Browse files
Files changed (7) hide show
  1. README.md +3 -3
  2. build/webgpu/bench.json +3 -3
  3. build/webgpu/manifest.json +12 -29
  4. build/webgpu/metadata.json +8 -8
  5. build/webgpu/pad-quad-vec4.wgsl.jinja +0 -2
  6. build/webgpu/pad.wgsl.jinja +3 -6
  7. build/webgpu/test.json +103 -0
README.md CHANGED
@@ -42,21 +42,21 @@ Attributes and default values (overridable per request):
42
 
43
  | Variable | Allowed dtypes |
44
  | --- | --- |
45
- | `T` | `float32`, `float16`, `uint32`, `int32`, `uint8`, `int8`, `bool` |
46
 
47
  ## Files
48
 
49
  - [`metadata.json`](build/webgpu/metadata.json) — kernel metadata (id, digests, per-variant templates, provenance)
50
  - [`manifest.json`](build/webgpu/manifest.json) — the op contract (source of truth)
51
  - [`test.json`](build/webgpu/test.json) — correctness cases
52
- - [`bench.json`](build/webgpu/bench.json) — benchmark + tuning cases
53
  - [`pad-quad-vec4.wgsl.jinja`](build/webgpu/pad-quad-vec4.wgsl.jinja)
54
  - [`pad.wgsl.jinja`](build/webgpu/pad.wgsl.jinja)
55
 
56
  ## Use with `@huggingface/kernels`
57
 
58
  ```sh
59
- npm install --save-exact @huggingface/kernels@0.0.1-preview.2
60
  ```
61
 
62
  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.
 
42
 
43
  | Variable | Allowed dtypes |
44
  | --- | --- |
45
+ | `T` | `float32`, `float16`, `uint32`, `int32`, `uint8`, `int8`, `bool`, `int64` |
46
 
47
  ## Files
48
 
49
  - [`metadata.json`](build/webgpu/metadata.json) — kernel metadata (id, digests, per-variant templates, provenance)
50
  - [`manifest.json`](build/webgpu/manifest.json) — the op contract (source of truth)
51
  - [`test.json`](build/webgpu/test.json) — correctness cases
52
+ - [`bench.json`](build/webgpu/bench.json) — benchmark cases
53
  - [`pad-quad-vec4.wgsl.jinja`](build/webgpu/pad-quad-vec4.wgsl.jinja)
54
  - [`pad.wgsl.jinja`](build/webgpu/pad.wgsl.jinja)
55
 
56
  ## Use with `@huggingface/kernels`
57
 
58
  ```sh
59
+ npm install --save-exact @huggingface/kernels@0.0.1-preview.3
60
  ```
61
 
62
  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.
build/webgpu/bench.json CHANGED
@@ -77,7 +77,7 @@
77
  "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.data) + numel(shapes.output)) * 2" }] }
78
  },
79
  {
80
- "name": "rank3_value_input_forces_scalar_path_large",
81
  "preset": "stress",
82
  "attrs": { "mode": "constant", "pads": [0, 0, 0, 0, 64, 0] },
83
  "inputs": {
@@ -91,7 +91,7 @@
91
  }
92
  },
93
  {
94
- "name": "rank4_edge_inner_unaligned_forces_scalar_path",
95
  "preset": "stress",
96
  "attrs": { "mode": "edge", "pads": [0, 0, 1, 1, 0, 0, 1, 1] },
97
  "inputs": { "data": { "dtype": "float32", "shape": [1, 64, 127, 127], "dist": "normal", "seed": 12, "scale": 2 } },
@@ -99,7 +99,7 @@
99
  "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.data) + numel(shapes.output)) * 4" }] }
100
  },
101
  {
102
- "name": "rank2_constant_inner_unaligned_forces_scalar_path",
103
  "preset": "stress",
104
  "attrs": { "mode": "constant", "pads": [1, 1, 1, 0] },
105
  "inputs": {
 
77
  "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.data) + numel(shapes.output)) * 2" }] }
78
  },
79
  {
80
+ "name": "rank3_value_input_scalar_path_large",
81
  "preset": "stress",
82
  "attrs": { "mode": "constant", "pads": [0, 0, 0, 0, 64, 0] },
83
  "inputs": {
 
91
  }
92
  },
93
  {
94
+ "name": "rank4_edge_inner_unaligned_scalar_path",
95
  "preset": "stress",
96
  "attrs": { "mode": "edge", "pads": [0, 0, 1, 1, 0, 0, 1, 1] },
97
  "inputs": { "data": { "dtype": "float32", "shape": [1, 64, 127, 127], "dist": "normal", "seed": 12, "scale": 2 } },
 
99
  "bench": { "metrics": [{ "type": "bandwidth", "value": "(numel(shapes.data) + numel(shapes.output)) * 4" }] }
100
  },
101
  {
102
+ "name": "rank2_constant_inner_unaligned_scalar_path",
103
  "preset": "stress",
104
  "attrs": { "mode": "constant", "pads": [1, 1, 1, 0] },
105
  "inputs": {
build/webgpu/manifest.json CHANGED
@@ -9,18 +9,21 @@
9
  "pads": { "required": true },
10
  "mode": { "values": ["constant", "reflect", "edge", "wrap"] }
11
  },
12
- "typeConstraints": { "T": ["float32", "float16", "uint32", "int32", "uint8", "int8", "bool"] },
13
  "tunables": { "WORKGROUP_SIZE": { "default": 256 } },
14
  "derive": {
15
  "shapeContract": "ranks.data == ranks.output and ranks.output >= 1",
16
  "defaultValueContract": "not present.constant_value and shapeContract and f16Ok(dtypes.T)",
17
  "vec4Contract": "defaultValueContract and numel(shapes.output) > 0 and dim(shapes.output, ranks.output - 1) % 4 == 0",
18
  "scalar": "dtypes.T",
19
- "mode": "attrs.mode"
20
- },
21
- "bindings": {
22
- "params": { "buffer": "uniform", "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.output)" }] }
 
 
23
  },
 
24
  "variants": [
25
  {
26
  "id": "constant_value_input",
@@ -32,13 +35,6 @@
32
  "id": "main",
33
  "name": "Pad",
34
  "shader": "pad.wgsl.jinja",
35
- "derive": {
36
- "dataShape": "shapes.data",
37
- "outputShape": "shapes.output",
38
- "rank": "ranks.output",
39
- "pads": "attrs.pads",
40
- "itemsPerInvocation": 4
41
- },
42
  "bindings": [
43
  "data",
44
  { "arg": "constant_value", "name": "padValue", "elementType": "$scalar", "length": 1 },
@@ -56,20 +52,14 @@
56
  {
57
  "id": "quad_vec4",
58
  "priority": 11,
59
- "when": ["vec4Contract"],
60
  "derive": { "outputElement": "\"vec4<\" ~ dtypes.T ~ \">\"", "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"" },
61
  "passes": [
62
  {
63
  "id": "main",
64
- "name": "Pad.quadVec4",
65
  "shader": "pad-quad-vec4.wgsl.jinja",
66
- "derive": {
67
- "dataShape": "shapes.data",
68
- "outputShape": "shapes.output",
69
- "rank": "ranks.output",
70
- "pads": "attrs.pads",
71
- "value": 0
72
- },
73
  "bindings": ["data", { "arg": "output", "elementType": "$outputElement" }],
74
  "dispatch": {
75
  "x": "min(ceilDiv((numel(shapes.output) / 4), (tunables.WORKGROUP_SIZE)), min(device.limits.maxComputeWorkgroupsPerDimension, 65535))",
@@ -88,14 +78,7 @@
88
  "id": "main",
89
  "name": "Pad",
90
  "shader": "pad.wgsl.jinja",
91
- "derive": {
92
- "dataShape": "shapes.data",
93
- "outputShape": "shapes.output",
94
- "rank": "ranks.output",
95
- "pads": "attrs.pads",
96
- "value": 0,
97
- "itemsPerInvocation": 4
98
- },
99
  "bindings": ["data", "output", "params"],
100
  "dispatch": {
101
  "x": "min(ceilDiv((ceilDiv(numel(shapes.output), 4)), (tunables.WORKGROUP_SIZE)), 65535)",
 
9
  "pads": { "required": true },
10
  "mode": { "values": ["constant", "reflect", "edge", "wrap"] }
11
  },
12
+ "typeConstraints": { "T": ["float32", "float16", "uint32", "int32", "uint8", "int8", "bool", "int64"] },
13
  "tunables": { "WORKGROUP_SIZE": { "default": 256 } },
14
  "derive": {
15
  "shapeContract": "ranks.data == ranks.output and ranks.output >= 1",
16
  "defaultValueContract": "not present.constant_value and shapeContract and f16Ok(dtypes.T)",
17
  "vec4Contract": "defaultValueContract and numel(shapes.output) > 0 and dim(shapes.output, ranks.output - 1) % 4 == 0",
18
  "scalar": "dtypes.T",
19
+ "mode": "attrs.mode",
20
+ "dataShape": "shapes.data",
21
+ "outputShape": "shapes.output",
22
+ "rank": "ranks.output",
23
+ "pads": "attrs.pads",
24
+ "itemsPerInvocation": 4
25
  },
26
+ "bindings": { "params": { "struct": [{ "name": "count", "type": "u32", "value": "numel(shapes.output)" }] } },
27
  "variants": [
28
  {
29
  "id": "constant_value_input",
 
35
  "id": "main",
36
  "name": "Pad",
37
  "shader": "pad.wgsl.jinja",
 
 
 
 
 
 
 
38
  "bindings": [
39
  "data",
40
  { "arg": "constant_value", "name": "padValue", "elementType": "$scalar", "length": 1 },
 
52
  {
53
  "id": "quad_vec4",
54
  "priority": 11,
55
+ "when": ["dtypes.T != \"vec2<u32>\"", "vec4Contract"],
56
  "derive": { "outputElement": "\"vec4<\" ~ dtypes.T ~ \">\"", "vectorScalar": "\"vec4<\" ~ dtypes.T ~ \">\"" },
57
  "passes": [
58
  {
59
  "id": "main",
60
+ "name": "Pad.QuadVec4",
61
  "shader": "pad-quad-vec4.wgsl.jinja",
62
+ "derive": { "value": 0 },
 
 
 
 
 
 
63
  "bindings": ["data", { "arg": "output", "elementType": "$outputElement" }],
64
  "dispatch": {
65
  "x": "min(ceilDiv((numel(shapes.output) / 4), (tunables.WORKGROUP_SIZE)), min(device.limits.maxComputeWorkgroupsPerDimension, 65535))",
 
78
  "id": "main",
79
  "name": "Pad",
80
  "shader": "pad.wgsl.jinja",
81
+ "derive": { "value": 0 },
 
 
 
 
 
 
 
82
  "bindings": ["data", "output", "params"],
83
  "dispatch": {
84
  "x": "min(ceilDiv((ceilDiv(numel(shapes.output), 4)), (tunables.WORKGROUP_SIZE)), 65535)",
build/webgpu/metadata.json CHANGED
@@ -1,22 +1,22 @@
1
  {
2
  "name": "ai.onnx.Pad",
3
- "id": "_ai_onnx_pad_webgpu_30935c6",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "backend": { "type": "webgpu" },
7
  "digest": {
8
  "algorithm": "sha256",
9
  "files": {
10
- "bench.json": "8tPycaU74JVcyHJOuvQfZrz8BEwiW5yt868REE1nvEI=",
11
- "manifest.json": "izmrfOCxDl4LceDqRx5gtmM5Q1dfUM4oJcnITWPO8mI=",
12
- "pad-quad-vec4.wgsl.jinja": "nKXmrJKUYEPgQR1ugQHSwvGjzW1gaerShR3b8fHLnoA=",
13
- "pad.wgsl.jinja": "uAqBSKweEQls2Ml+WxW1nkrsPhVMvLOYPNTtGcTwSgM=",
14
- "test.json": "pw+IPrkqhtk9kFg2+5puugkg5AE10N7CwLH0HWkNtks="
15
  }
16
  },
17
- "provenance": { "kernel": { "sha": "91d990483a174128daf7673f3f37a7c890493ae1", "dirty": false } },
18
  "webgpu": {
19
- "manifestSpec": "2.0",
20
  "variants": {
21
  "constant_value_input": ["pad.wgsl.jinja"],
22
  "quad_vec4": ["pad-quad-vec4.wgsl.jinja"],
 
1
  {
2
  "name": "ai.onnx.Pad",
3
+ "id": "_ai_onnx_pad_webgpu_f6b74c2",
4
  "version": 1,
5
  "license": "Apache-2.0",
6
  "backend": { "type": "webgpu" },
7
  "digest": {
8
  "algorithm": "sha256",
9
  "files": {
10
+ "bench.json": "UTHQCS/ec5z6S0xAS0MTjeXNVOdoMNrZl0ucq2NtMB8=",
11
+ "manifest.json": "bLjZA68bmGDgk9+7gG8Pgax+tfEag6hXtZuz84wxwh8=",
12
+ "pad-quad-vec4.wgsl.jinja": "p5a1bip066TyzgzmqEPY2T3HhvWviGjm/xvpNODTII4=",
13
+ "pad.wgsl.jinja": "Jhj/0J43kh7fQPKnCIdRD3aybTq0koRB8GGGmRZ8plE=",
14
+ "test.json": "9PTxHWGE1cnwGEiI0w5XTIsJlRK1lHpxQL1arNL4qxo="
15
  }
16
  },
17
+ "provenance": { "kernel": { "sha": "6fdf6301e2bbcc2f03bf1eaf493b7ad55ef33afc", "dirty": false } },
18
  "webgpu": {
19
+ "manifestSpec": "2.1",
20
  "variants": {
21
  "constant_value_input": ["pad.wgsl.jinja"],
22
  "quad_vec4": ["pad-quad-vec4.wgsl.jinja"],
build/webgpu/pad-quad-vec4.wgsl.jinja CHANGED
@@ -8,7 +8,6 @@
8
  // coordinate mapping.
9
  {{ env.wgsl.resourceDeclarations }}
10
 
11
- {% set rank = rank %}
12
  fn map_coord(coord: i32, dim: i32) -> i32 {
13
  if (coord >= 0i && coord < dim) {
14
  return coord;
@@ -37,7 +36,6 @@ fn map_coord(coord: i32, dim: i32) -> i32 {
37
  return period - wrapped;
38
  {% endif %}
39
  }
40
-
41
  {% set countVecs = namespace(value=1) %}
42
  {% for j in range(rank) %}
43
  {% set countVecs.value = countVecs.value * outputShape[j] %}
 
8
  // coordinate mapping.
9
  {{ env.wgsl.resourceDeclarations }}
10
 
 
11
  fn map_coord(coord: i32, dim: i32) -> i32 {
12
  if (coord >= 0i && coord < dim) {
13
  return coord;
 
36
  return period - wrapped;
37
  {% endif %}
38
  }
 
39
  {% set countVecs = namespace(value=1) %}
40
  {% for j in range(rank) %}
41
  {% set countVecs.value = countVecs.value * outputShape[j] %}
build/webgpu/pad.wgsl.jinja CHANGED
@@ -12,9 +12,7 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
12
  for (var i = begin; i < end; i = i + 1u) {
13
  {%- endmacro %}
14
  {% macro flat_tail_close() %}
15
- }
16
- {% endmacro %}
17
-
18
  {{ env.wgsl.resourceDeclarations }}
19
 
20
  fn map_coord(coord: i32, dim: i32) -> i32 {
@@ -45,7 +43,6 @@ fn map_coord(coord: i32, dim: i32) -> i32 {
45
  return period - wrapped;
46
  {% endif %}
47
  }
48
-
49
  fn data_offset_or_invalid(out_index: u32) -> i32 {
50
  var rem = out_index;
51
  var offset = 0i;
@@ -99,10 +96,10 @@ fn data_offset_or_invalid(out_index: u32) -> i32 {
99
  {% if valueFromInput %}
100
  output[i] = padValue[0];
101
  {% else %}
102
- output[i] = {{ scalar }}({{ value }});
103
  {% endif %}
104
  } else {
105
  output[i] = data[u32(src)];
106
  }
107
- {{ flat_tail_close() -}}
108
  }
 
12
  for (var i = begin; i < end; i = i + 1u) {
13
  {%- endmacro %}
14
  {% macro flat_tail_close() %}
15
+ }{% endmacro %}
 
 
16
  {{ env.wgsl.resourceDeclarations }}
17
 
18
  fn map_coord(coord: i32, dim: i32) -> i32 {
 
43
  return period - wrapped;
44
  {% endif %}
45
  }
 
46
  fn data_offset_or_invalid(out_index: u32) -> i32 {
47
  var rem = out_index;
48
  var offset = 0i;
 
96
  {% if valueFromInput %}
97
  output[i] = padValue[0];
98
  {% else %}
99
+ output[i] = {{ scalar }}({% if scalar == "vec2<u32>" %}0u{% else %}{{ value }}{% endif %});
100
  {% endif %}
101
  } else {
102
  output[i] = data[u32(src)];
103
  }
104
+ {{ flat_tail_close() }}
105
  }
build/webgpu/test.json CHANGED
@@ -1371,6 +1371,109 @@
1371
  "constant_value": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [-1.0] } }
1372
  },
1373
  "outputs": { "output": { "dtype": "float32", "shape": [1, 2, 1, 2, 1, 2, 2, 4], "tolerance": 0 } }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1374
  }
1375
  ]
1376
  }
 
1371
  "constant_value": { "dtype": "float32", "shape": [], "data": { "kind": "values", "values": [-1.0] } }
1372
  },
1373
  "outputs": { "output": { "dtype": "float32", "shape": [1, 2, 1, 2, 1, 2, 2, 4], "tolerance": 0 } }
1374
+ },
1375
+ {
1376
+ "name": "int64_full_range_rank2_constant",
1377
+ "attrs": { "pads": [1, 2, 1, 0] },
1378
+ "inputs": {
1379
+ "data": {
1380
+ "dtype": "int64",
1381
+ "shape": [2, 3],
1382
+ "data": {
1383
+ "kind": "cycle",
1384
+ "values": ["0", "4294967296", "8589934593", "-4294967297", "9223372036854775807", "-9223372036854775808", "9223372036854775806", "-1"]
1385
+ }
1386
+ },
1387
+ "constant_value": {
1388
+ "dtype": "int64",
1389
+ "shape": [],
1390
+ "data": {
1391
+ "kind": "cycle",
1392
+ "values": ["0", "4294967296", "8589934593", "-4294967297", "9223372036854775807", "-9223372036854775808", "9223372036854775806", "-1"]
1393
+ }
1394
+ }
1395
+ },
1396
+ "outputs": { "output": { "dtype": "int64", "shape": [4, 5], "tolerance": 0, "relTolerance": 0 } },
1397
+ "tolerance": 0,
1398
+ "relTolerance": 0
1399
+ },
1400
+ {
1401
+ "name": "int64_full_range_ort_value_input_constant_1d",
1402
+ "provenance": {
1403
+ "source": "onnxruntime/test/providers/cpu/tensor/pad_test.cc",
1404
+ "test": "PadOpTest.Pad_Constant_1D",
1405
+ "notes": "Uses Pad's optional `constant_value` input. Pads remain supplied by the `pads` attribute, while the fill value comes directly from the input tensor."
1406
+ },
1407
+ "attrs": { "mode": "constant", "pads": [1, 2] },
1408
+ "inputs": {
1409
+ "data": {
1410
+ "dtype": "int64",
1411
+ "shape": [2],
1412
+ "data": {
1413
+ "kind": "cycle",
1414
+ "values": ["4294967296", "8589934593", "-4294967297", "9223372036854775807", "-9223372036854775808", "9223372036854775806", "-1", "0"]
1415
+ }
1416
+ },
1417
+ "constant_value": {
1418
+ "dtype": "int64",
1419
+ "shape": [],
1420
+ "data": {
1421
+ "kind": "cycle",
1422
+ "values": ["4294967296", "8589934593", "-4294967297", "9223372036854775807", "-9223372036854775808", "9223372036854775806", "-1", "0"]
1423
+ }
1424
+ }
1425
+ },
1426
+ "outputs": { "output": { "dtype": "int64", "shape": [5], "tolerance": 0, "relTolerance": 0 } },
1427
+ "tolerance": 0,
1428
+ "relTolerance": 0
1429
+ },
1430
+ {
1431
+ "name": "int64_default_zero",
1432
+ "inputs": {
1433
+ "data": {
1434
+ "dtype": "int64",
1435
+ "shape": [2],
1436
+ "data": { "kind": "values", "values": ["-9223372036854775808", "9223372036854775807"] }
1437
+ }
1438
+ },
1439
+ "outputs": { "output": { "dtype": "int64", "shape": [5], "tolerance": 0, "relTolerance": 0 } },
1440
+ "attrs": { "pads": [1, 2] }
1441
+ },
1442
+ {
1443
+ "name": "int64_padding_edge",
1444
+ "inputs": {
1445
+ "data": {
1446
+ "dtype": "int64",
1447
+ "shape": [3],
1448
+ "data": { "kind": "values", "values": ["-9223372036854775808", "9223372036854775807", "4294967296"] }
1449
+ }
1450
+ },
1451
+ "outputs": { "output": { "dtype": "int64", "shape": [6], "tolerance": 0, "relTolerance": 0 } },
1452
+ "attrs": { "pads": [1, 2], "mode": "edge" }
1453
+ },
1454
+ {
1455
+ "name": "int64_padding_reflect",
1456
+ "inputs": {
1457
+ "data": {
1458
+ "dtype": "int64",
1459
+ "shape": [3],
1460
+ "data": { "kind": "values", "values": ["-9223372036854775808", "9223372036854775807", "4294967296"] }
1461
+ }
1462
+ },
1463
+ "outputs": { "output": { "dtype": "int64", "shape": [6], "tolerance": 0, "relTolerance": 0 } },
1464
+ "attrs": { "pads": [1, 2], "mode": "reflect" }
1465
+ },
1466
+ {
1467
+ "name": "int64_padding_wrap",
1468
+ "inputs": {
1469
+ "data": {
1470
+ "dtype": "int64",
1471
+ "shape": [3],
1472
+ "data": { "kind": "values", "values": ["-9223372036854775808", "9223372036854775807", "4294967296"] }
1473
+ }
1474
+ },
1475
+ "outputs": { "output": { "dtype": "int64", "shape": [6], "tolerance": 0, "relTolerance": 0 } },
1476
+ "attrs": { "pads": [1, 2], "mode": "wrap" }
1477
  }
1478
  ]
1479
  }