On Thu, Jun 12, 2025 at 09:26:25AM +0200, Roger Pau Monné wrote: > On Wed, Jun 11, 2025 at 08:33:55PM +0100, Andrew Cooper wrote: > > On 11/06/2025 6:16 pm, Roger Pau Monne wrote: > > > We can discuss whether we want both the fast and the slow variants. The > > > slow (brute force) was added as a result of me playing with weird region > > > layouts where the fast one didn't manage to compress, or the resulting > > > coefficients had a poor compression ratio. However at this point the > > > slow variant has only proven helpful in synthetic cases, I haven't (yet?) > > > seen a real host memory layout that would benefit from it. > > > > I'm going to hold off on opinions until I've read the rest of the series. > > > > One question through. Can we round offset up to the next power of two, > > so we can replace the divide with a shift? > > I've tried to round up both offset and size, but that resulted in no > compression in some cases. I can try to maybe round just one of > those? > > Note that the divide is done once with offset and once with size, > depending on the direction of the translation. I can explore this a > bit.
So for the 4s GNR example, rounding the offset to the nearest (lower) power of two will result in: PFN fast compression result diverge, expected: offset 0000063e80000 size 0000008300000 (91%) got: offset 0000040000000 size 0000033e80000 (18%) Ranges: 0000000000000-0000008080000 0000063e80000-000006be80000 00000c7e80000-00000cfe80000 000012be80000-0000133e80000 And that's just rounding the offset, if we then also round the size there's no compression possible, as size would become 0x40000000 and thus offset == size. Thanks, Roger.
