webgpu: Support 2x4 output for Conv2DBackpropInput #7371
Conversation
|
With the second commit a734dd8, the time of Conv2DBackpropInput is further reduced (26.32ms -> 14.17ms -> 10.27ms on CFL) |
xhcao
left a comment
xhcao
left a comment
There was a problem hiding this comment.
LGTM, please take care of WGSL code typeset.
Linchenn
left a comment
Linchenn
left a comment
There was a problem hiding this comment.
LGTM, great finding!
Reviewable status:
complete! 2 of 1 approvals obtained (waiting on @gyagp and @qjia7)
tfjs-backend-webgpu/src/conv_backprop_webgpu.ts line 142 at r2 (raw file):
dotProd[0] = dotProd[0] + tmpval; } } else if (bDyCVal2) {
We may not need if (bDyCVal2) because we have checked (bDyCVal == false && bDyCVal2 == false) before. Or we could remove if (bDyCVal == false && bDyCVal2 == false) above.
qjia7
left a comment
qjia7
left a comment
There was a problem hiding this comment.
Reviewable status:
tfjs-backend-webgpu/src/conv_backprop_webgpu.ts line 142 at r2 (raw file):
Previously, Linchenn wrote…
We may not need
if (bDyCVal2)because we have checked(bDyCVal == false && bDyCVal2 == false)before. Or we could removeif (bDyCVal == false && bDyCVal2 == false)above.
Done.
3 participants
Conv2DBackpropInput in ArPortraitDepth
26.32ms -> 14.17ms on CFL
6.91ms -> 4.70ms on TGL
To see the logs from the Cloud Build CI, please join either our discussion or announcement mailing list.
This change is