summaryrefslogtreecommitdiff
AgeCommit message (Collapse)AuthorFilesLines
2009-01-15Added copyright boilerplate.HEADmasterM Joonas Pihlaja7-8/+181
2009-01-14Added bench.luaM Joonas Pihlaja1-0/+110
2009-01-14Added READMEM Joonas Pihlaja1-0/+109
2009-01-14Comment in the branch to the blocked SSE2 version.M Joonas Pihlaja1-1/+1
2009-01-14Added some usage text to the driver.M Joonas Pihlaja1-3/+56
2009-01-14Rename inv32 -> inv.M Joonas Pihlaja3-10/+10
2009-01-14Separate special cased unpremultipliers from just the basic method.M Joonas Pihlaja8-131/+322
2009-01-13Separate unpremultiplier methods, clean them up, and kill the bad ones.M Joonas Pihlaja5-342/+297
2009-01-13Added prototype to header comment.M Joonas Pihlaja1-1/+5
2009-01-13Removed crufty SSE2 versions.M Joonas Pihlaja4-495/+264
2009-01-13Fix some typos and misc. cleanup.M Joonas Pihlaja1-10/+11
2009-01-13[sse2-test] Identify runs of zero in the input in the single pixel cases.M Joonas Pihlaja1-3/+6
2009-01-13[sse2-test] sync it.M Joonas Pihlaja1-45/+247
2009-01-13[tester] sync it.M Joonas Pihlaja1-14/+63
2009-01-13Move numeric arguments to the environment.M Joonas Pihlaja1-3/+14
2009-01-10.gitignoreM Joonas Pihlaja1-0/+2
2009-01-10Add new tighter SSE2 unpremultiply.M Joonas Pihlaja2-2/+107
The same issues as with memcpy are now affecting SSE2 enabled unpremultiplication: the choice of movntd vs. movdqa vs. movdqu for writing to the destination buffer is crucial, but depends on how the result is used and what size it is. If it fits in L2 and is going to be used quickly, then it makes sense to use movdqa/dqu to write to the destination. If it's not going to be used or doesn't fit then it's far better to use movntdq.
2009-01-03Remove the old unused SSE2 version.M Joonas Pihlaja1-165/+0
2009-01-03Import sources from oneshots/M Joonas Pihlaja4-0/+929