sync c928d21e6cc1
Browse files
build/webgpu/binary-broadcast.wgsl.jinja
CHANGED
|
@@ -4,23 +4,14 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>, @builtin(num_workgroups)
|
|
| 4 |
// 2D-folded flat index: gid.y carries the high bits when the element count exceeds the
|
| 5 |
// maxComputeWorkgroupsPerDimension limit (the dispatch caps x and spills the rest into y).
|
| 6 |
let invocation = gid.x + gid.y * nwg.x * {{ tunables.WORKGROUP_SIZE }}u;
|
| 7 |
-
{% if source.itemsPerInvocation is defined %}
|
| 8 |
// Tail-safe scalar x4 keeps vector-like dispatch density without requiring
|
| 9 |
// the logical tensor length (or its storage binding) to be vec4 aligned.
|
| 10 |
let begin = invocation * {{ source.itemsPerInvocation }}u;
|
| 11 |
let end = min(begin + {{ source.itemsPerInvocation }}u, params.count);
|
| 12 |
for (var i = begin; i < end; i = i + 1u) {
|
| 13 |
-
{%-
|
| 14 |
-
let i = invocation;
|
| 15 |
-
if (i >= params.count) {
|
| 16 |
-
return;
|
| 17 |
-
}
|
| 18 |
-
{%- endif %}
|
| 19 |
-
{% endmacro %}
|
| 20 |
{% macro flat_tail_close() %}
|
| 21 |
-
{% if source.itemsPerInvocation is defined %}
|
| 22 |
}
|
| 23 |
-
{% endif %}
|
| 24 |
{% endmacro %}
|
| 25 |
|
| 26 |
{% macro offset_fn(fn_name, opShape, opRank, op_same, op_numel, outShape, outRank, out_numel) %}
|
|
|
|
| 4 |
// 2D-folded flat index: gid.y carries the high bits when the element count exceeds the
|
| 5 |
// maxComputeWorkgroupsPerDimension limit (the dispatch caps x and spills the rest into y).
|
| 6 |
let invocation = gid.x + gid.y * nwg.x * {{ tunables.WORKGROUP_SIZE }}u;
|
|
|
|
| 7 |
// Tail-safe scalar x4 keeps vector-like dispatch density without requiring
|
| 8 |
// the logical tensor length (or its storage binding) to be vec4 aligned.
|
| 9 |
let begin = invocation * {{ source.itemsPerInvocation }}u;
|
| 10 |
let end = min(begin + {{ source.itemsPerInvocation }}u, params.count);
|
| 11 |
for (var i = begin; i < end; i = i + 1u) {
|
| 12 |
+
{%- endmacro %}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
{% macro flat_tail_close() %}
|
|
|
|
| 14 |
}
|
|
|
|
| 15 |
{% endmacro %}
|
| 16 |
|
| 17 |
{% macro offset_fn(fn_name, opShape, opRank, op_same, op_numel, outShape, outRank, out_numel) %}
|
build/webgpu/manifest.json
CHANGED
|
@@ -200,7 +200,8 @@
|
|
| 200 |
"bRank": "ranks.B",
|
| 201 |
"cRank": "ranks.C",
|
| 202 |
"op": "\"sub\"",
|
| 203 |
-
"cDtype": "tensorDtypes.C"
|
|
|
|
| 204 |
}
|
| 205 |
},
|
| 206 |
"bindings": [
|
|
@@ -226,7 +227,7 @@
|
|
| 226 |
"struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.C)" }] }
|
| 227 |
}
|
| 228 |
],
|
| 229 |
-
"dispatch": { "threads": "numel(shapes.C)", "workgroupSize": "tunables.WORKGROUP_SIZE" }
|
| 230 |
}
|
| 231 |
]
|
| 232 |
}
|
|
|
|
| 200 |
"bRank": "ranks.B",
|
| 201 |
"cRank": "ranks.C",
|
| 202 |
"op": "\"sub\"",
|
| 203 |
+
"cDtype": "tensorDtypes.C",
|
| 204 |
+
"itemsPerInvocation": 4
|
| 205 |
}
|
| 206 |
},
|
| 207 |
"bindings": [
|
|
|
|
| 227 |
"struct": { "name": "Params", "fields": [{ "name": "count", "type": "u32", "value": "numel(shapes.C)" }] }
|
| 228 |
}
|
| 229 |
],
|
| 230 |
+
"dispatch": { "threads": "ceilDiv(numel(shapes.C), 4)", "workgroupSize": "tunables.WORKGROUP_SIZE" }
|
| 231 |
}
|
| 232 |
]
|
| 233 |
}
|
build/webgpu/metadata.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
{
|
| 2 |
"name": "ai.onnx.Sub",
|
| 3 |
-
"id": "
|
| 4 |
"version": 1,
|
| 5 |
"license": "Apache-2.0",
|
| 6 |
"backend": { "type": "webgpu" },
|
|
@@ -9,12 +9,12 @@
|
|
| 9 |
"files": {
|
| 10 |
"bench.json": "lWLk58oMaDoHDeONhX1uGhYlFv2T6DHU9iZLWHsLcEs=",
|
| 11 |
"binary-broadcast-vec4.wgsl.jinja": "RwqJ+9q1E9Hmx4WPB8XnYYgYVhaO4X6vHghv0iHkXOE=",
|
| 12 |
-
"binary-broadcast.wgsl.jinja": "
|
| 13 |
"binary-vec4.wgsl.jinja": "VtSzBDt9S766ph3QPveK+YJgNlh/FBOKvznRIQWfKIg=",
|
| 14 |
-
"manifest.json": "
|
| 15 |
"test.json": "/HGe/vteCb46Xnc+SHBVlJ6Xcpv+Jx+rXV+GXJARgD8="
|
| 16 |
}
|
| 17 |
},
|
| 18 |
-
"provenance": { "kernel": { "sha": "
|
| 19 |
"webgpu": { "manifestSpec": "1.0", "specialized": true, "opPath": "ops/ai.onnx.Sub" }
|
| 20 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"name": "ai.onnx.Sub",
|
| 3 |
+
"id": "_ai_onnx_sub_webgpu_0449631",
|
| 4 |
"version": 1,
|
| 5 |
"license": "Apache-2.0",
|
| 6 |
"backend": { "type": "webgpu" },
|
|
|
|
| 9 |
"files": {
|
| 10 |
"bench.json": "lWLk58oMaDoHDeONhX1uGhYlFv2T6DHU9iZLWHsLcEs=",
|
| 11 |
"binary-broadcast-vec4.wgsl.jinja": "RwqJ+9q1E9Hmx4WPB8XnYYgYVhaO4X6vHghv0iHkXOE=",
|
| 12 |
+
"binary-broadcast.wgsl.jinja": "QSH3z+74piXaHqqqH98mYE2rc80oT1WDMldkaEwrZWM=",
|
| 13 |
"binary-vec4.wgsl.jinja": "VtSzBDt9S766ph3QPveK+YJgNlh/FBOKvznRIQWfKIg=",
|
| 14 |
+
"manifest.json": "Vwbwn6nbn9LV1r0M6kaVuu76MEZk3AtAqs3Z+q+li1I=",
|
| 15 |
"test.json": "/HGe/vteCb46Xnc+SHBVlJ6Xcpv+Jx+rXV+GXJARgD8="
|
| 16 |
}
|
| 17 |
},
|
| 18 |
+
"provenance": { "kernel": { "sha": "c928d21e6cc1310861cba3bafb75f5f679ecf5f3", "dirty": false } },
|
| 19 |
"webgpu": { "manifestSpec": "1.0", "specialized": true, "opPath": "ops/ai.onnx.Sub" }
|
| 20 |
}
|