Thank you for the DirectML update

[BadWXfcst]

Whatever magic @lshqqytiger implemented in DirectML drastically increased the speed and VRAM availability of my old 8GB AMD GPU. Thank you.

[vladmandic]

@lshqqytiger FYI :)

[lshqqytiger]

There was no expectation about speed or vram on recent update.. interesting.

[TeutonJon78]

I'm finding the opposite on my RX580. At the same settings yesterday I could generated 512x912, but now I'm getting OOM.

Commit cdc803b - 512x912 works fine
Commit 0a52c44 -- RuntimeError: Could not allocate tensor with 1703411712 bytes. There is not enough GPU video memory available!

And same at current head (a7f5e75), OOM

@lshqqytiger, any thoughts?

Edit: VRAM utilization is sitting at 6.3 GB instead of the normal 7.4 GB.