On Thu, 3 Apr 2025 09:26:03 GMT, Nikita Gubarkov <ngubar...@openjdk.org> wrote:
>>>I just accepted any pixel-interleaved surface with 4-byte stride and assumed >>>that I can blit it directly. >> >> So you implemented logic similar to [this >> one](https://github.com/openjdk/jdk/blob/bd74922157230c866802b4c5269da81e872525aa/src/java.desktop/share/native/libawt/java2d/loops/LoopMacros.h#L847). >> >> I haven’t found yet whether we have logic to fall back to a "generic Java >> loop" if the gap is not stored in the array and that memcpy fails. If we’re >> missing this, we should either: >> >> - Add the fallback, >> - Implement another version of the loop that handles the tail, or >> - Update the current ISOSCALE_BLIT. >> >> A similar approach could be implemented for Vulkan as well. > > I get the idea. Actually, my Vulkan blit routines know the component order > (needed for swizzling anyway), so instead of copying `scanlineStride` or > `width * pixelStride`, I can copy `(width - 1) * pixelStride + maxBandOffset > + 1` (I don't like breaking the alignment at the end, but anyway). > However, supporting the gap at the end of the pixel seems like a lot of > burden to me - not all current code seems to consider this possibility and > for the future code it's too easy to get wrong. > > Some side thoughts: I think we rely on predefined formats and loops too much. > As I already mentioned, `SurfaceDataRasInfo` doesn't have info about its > bands and doesn't even have a total size. We rely on specific loops > registered for specific surface types, which only works well when we created > those surfaces by ourselves. Example: > I profiled Intellij IDEA to see which blit loops are often used there and > noticed that icons are loaded as 4-byte RGBA images. See the issue? That's a > `TYPE_CUSTOM` image with some generic `SurfaceType` - it doesn't have native > raster ops initialized, so it can only go through a generic software loop! > That's how I came to https://github.com/openjdk/jdk/pull/24378. So what I was > trying to do with that Vulkan work is to make it more generic - register > loops for more generic surface types and dynamically inspect their properties > to see whether we can actually do an optimal blit. And given that 99% of the > time those are 3/4-bytes per pixel RGBA/ARGB/BGR/etc., it's very easy to > generalize, but we don't seem to have enough flexibility for that. There is no reason to change our current implementation of ComponentSampleModel, since a similar raster can be created manually and should be properly handled by accelerated pipelines. DataBuffer manualBuffer = new DataBufferByte( scanlineStride * (SIZE - 1) + pixelStride * SIZE ); WritableRaster manualRaster = Raster.createWritableRaster(sampleModel, manualBuffer, null); BufferedImage manualImage = new BufferedImage(colorModel, manualRaster, isAlphaPremultiplied, null); ------------- PR Review Comment: https://git.openjdk.org/jdk/pull/24111#discussion_r2047838110
