[webgpu] Support convTranspose vec4 #6603
Conversation
axinging
force-pushed
the
conv2d_backprop_vec4
branch
4 times, most recently
from
7660b60 to
1a7c882
Compare
|
@qjia7 @xhcao @gyagp @haoyunfeix PTAL |
| return 'return W[getIndexFromCoords4D(coord, uniforms.wShape)];'; | ||
| case 4: | ||
| return ` | ||
| let coord1 = vec4<i32>(coordX, coordY, col + 1, rowInner); |
There was a problem hiding this comment.
Do you need to check whether col + 1|col + 2|col + 3 out of boundary?
There was a problem hiding this comment.
I am not sure if this necessary, but in the updated commit, I do add this check, PTAL
There was a problem hiding this comment.
Ok, it seem that it's not needed since you have make sure that convInfo.inChannels % 4 === 0
|
|
||
| constructor(convInfo: backend_util.Conv2DInfo) { | ||
| this.outputShape = convInfo.inShape; | ||
|
|
||
| util.assert( | ||
| convInfo.dataFormat === 'channelsLast', | ||
| () => 'TODO: NCHW is unimplemented'); | ||
| this.isChannelsLast = convInfo.dataFormat === 'channelsLast'; |
There was a problem hiding this comment.
nit: Currently NCHW is unimplemented. So this.isChannelsLast can be ignored in this PR.
| this.isChannelsLast = convInfo.dataFormat === 'channelsLast'; | ||
| this.isVec4 = (((convInfo.inChannels % 4 === 0) && this.isChannelsLast) || | ||
| (convInfo.outWidth % 4 === 0 && !this.isChannelsLast)) && | ||
| convInfo.outChannels % 4 === 0; |
There was a problem hiding this comment.
const dimAOuter = convInfo.inShape[1] * convInfo.inShape[2];
const dimBOuter = convInfo.inShape[3];
const dimInner =
convInfo.filterHeight * convInfo.filterWidth * convInfo.outChannels;
So I suppose the vec4 condition should be convInfo.inChannels % 4 === 0 && convInfo.outChannels % 4 === 0
| } | ||
| } | ||
|
|
||
| ${conv2dTransposeCommonSnippet(elementsSize[2])} |
There was a problem hiding this comment.
${conv2dTransposeCommonSnippet(this.isVec4? 4 : 1)} ?
axinging
force-pushed
the
conv2d_backprop_vec4
branch
from
1a7c882 to
8e399a0
Compare
| return 'return W[getIndexFromCoords4D(coord, uniforms.wShape)];'; | ||
| case 4: | ||
| return ` | ||
| let coord1 = vec4<i32>(coordX, coordY, col + 1, rowInner); |
There was a problem hiding this comment.
Ok, it seem that it's not needed since you have make sure that convInfo.inChannels % 4 === 0
Done |
PERF
hand_detector model on intel CFL:
hand_detector model on intel TGL:
To see the logs from the Cloud Build CI, please join either our discussion or announcement mailing list.
This change is