On Mon, 20 Mar 2023 01:07:25 GMT, Sergey Bylokhov <s...@openjdk.org> wrote:
> Support of premultiplied alpha is added to the "accelerated" code path of the > CMM. It is implemented on top of the new feature added to the littlecms > library in 2.13 and 2.15. > > The next formats are now "supported": TYPE_INT_ARGB_PRE, TYPE_4BYTE_ABGR_PRE, > and any custom images which use the ComponentColorModel+8-bit-precision like > RGBApre or ApreBGR. > > After this patch, we will fully support all our standard types(8-bit > precision), and mostly any combinations of blits between them, having two > exceptions: > * lcms does not convert pre-alpha for transparent src if dst is opaque > * lcms does not set correct alpha(=1.0) for transparent dst if src is opaque > > Both of them are "features", so we probably need to implement a workaround > someday, but for now we will use our generic/slow code. > > Performance results for the ColorConvertOp.filter() for images with > premultiplied alpha. > > | Test | Base(avgt in us/op) | Fix(avgt in us/op) | Improvement % | > | ------------- | ------------- | ------------- |------------- | > | 32 Threads, from=sRGB:size=10:to=CIEXYZ | 97| 12| 708% | > | 32 Threads, from=sRGB:size=100:to=CIEXYZ | 8 013 | 617 | 1199% | > | 32 Threads, from=sRGB:size=1000:to=CIEXYZ | 651 838 | 104 969| 521% | > | 1 Thread, from=sRGB:size=10:to=CIEXYZ | 21| 5| 320% | > | 1 Thread, from=sRGB:size=100:to=CIEXYZ | 677 | 300 | 126% | > | 1 Thread, from=sRGB:size=1000:to=CIEXYZ | 60 095 | 30 536 | 97% | Sad, does it catches the difference between slow/fast implementations or some issues related to the alpha_pre support? ------------- PR Comment: https://git.openjdk.org/jdk/pull/13095#issuecomment-1502489899
