webgpu: support MaxPoolGrad kernel #7261
Conversation
| const maxPoolPositions = backend.runWebGPUProgram( | ||
| maxPoolPositionsProgram, [x], 'int32', uniformData); | ||
|
|
||
| const avgPoolBackpropProgram = new MaxPool2DBackpropProgram(convInfo); |
There was a problem hiding this comment.
nit: avgPoolXXX -> maxPoolXXX
| @@ -0,0 +1,93 @@ | |||
| /** | |||
There was a problem hiding this comment.
Will MaxPool3DBackpropProgram share the same file? If yes, maybe we should remove 2d from file name?
There was a problem hiding this comment.
Currently, I am not sure the file will be shared by 3d, so we keep the name with 2d here. If we put the 3d in this file, then we could rename the file.
| dispatch: [number, number, number]; | ||
| variableNames = ['dy', 'maxPos']; | ||
| uniforms = | ||
| `stride : vec2<i32>, pad : vec2<i32>, dilation : vec2<i32>, filterDims : vec2<i32>, |
There was a problem hiding this comment.
Can we unify to use strides, pads, dilations from this PR? And you can do the refactor for other places in a separate PR.
|
@gyagp Please take a look, thanks |
| this.dispatch = computeDispatch( | ||
| this.dispatchLayout, this.outputShape, this.workgroupSize); | ||
|
|
||
| this.shaderKey = `max_pool2d_backprop`; |
There was a problem hiding this comment.
| this.shaderKey = `max_pool2d_backprop`; | |
| this.shaderKey = `maxPool2DBackprop`; |
What's the convention for shaderKey? Maybe className.replace('Program', '') and variants split by '_'? We need to somehow unify this.
3 participants
To see the logs from the Cloud Build CI, please join either our discussion or announcement mailing list.
This change is