[WebGPU] support pool3d kernels #7405
Conversation
| const convInfo = backend_util.computePool3DInfo( | ||
| x.shape as [number, number, number, number, number], filterSize, strides, | ||
| dilations, pad, dimRoundingMode, dataFormat); | ||
| const maxPoolProgram = new Pool3DProgram(convInfo, 'avg'); |
There was a problem hiding this comment.
maxPoolProgram -> avgPoolProgram or program
| { | ||
| type: 'int32', | ||
| data: [ | ||
| convInfo.dilationDepth, convInfo.dilationHeight, convInfo.dilationWidth |
There was a problem hiding this comment.
In L30, you always use dilations = [1, 1, 1]. Do we still need to pass it to uniform? And I see the API doc also doesn't have this parameter.
There was a problem hiding this comment.
Removed. Thanks for finding there is no dilations for 3d pool kernels.
| }, | ||
| { | ||
| type: 'int32', | ||
| data: [convInfo.inDepth, convInfo.inHeight, convInfo.inWidth] |
There was a problem hiding this comment.
Although it doesn't support channelsFirst now, the convInfo.inDepth, convInfo.inHeight and convInfo.inWidth had handled the dataFormat attribute, if we use uniforms.xShape, we will also need handle the dataFormat attribute in the shader again.
If we decide that we must reuse the uniforms.xShape, we could modify all related code in the future, not only one here.
| } | ||
| let idyR = i32(dyR); | ||
| for (var wC = 0; wC < uniforms.filterDims[2]; wC++) { |
There was a problem hiding this comment.
For wD and wR, you both add the dilation, if dilations are needed, please add it here to keep consistent. Otherwise, remove all dilations.
This patch supports AvgPool3D, MaxPool3D and MaxPool3DGrad kernels.
| dispatch: [number, number, number]; | ||
| variableNames = ['dy', 'maxPos']; | ||
| uniforms = | ||
| `strides : vec3<i32>, pads : vec3<i32>, dilations : vec3<i32>, filterDims : vec3<i32>, |
There was a problem hiding this comment.
nit: dilations should be removed?
There was a problem hiding this comment.
Done. Had forgotten to remove dilations information for MaxPool3DGrad kernel. Thanks.
3 participants
This patch supports AvgPool3D, MaxPool3D and
MaxPool3DGrad kernels.
To see the logs from the Cloud Build CI, please join either our discussion or announcement mailing list.
This change is