mirror of
https://github.com/HDFGroup/hdf5.git
synced 2025-04-18 17:40:55 +08:00
8baa675250
Changes since 19990616 ---------------------- ./src/H5T.c ./src/H5Tconv.c All conversion functions take an extra argument called `stride' which is the number of bytes to advance the source and destination pointers after each element is converted. If the value is zero then the old behavior is preserved (source and destination values are packed). This feature was necessary to implement the compound datatype conversion optimizations and it causes a minor change to the API (application-defined type conversion functions take an extra size_t stride argument). ./src/H5Tconv.c An additional compound data type conversion function was added which is applied unless the destination type is larger than the source type. I'm measuring significant performance increases for certain operations: Test Name Struct-Conv Noop-Conv New(Old) MB/s New(Old) MB/s ---------- ------------- ------------- Reordering 2.062(0.3936) 54087(0.9047) Subsetting 2.901(0.6581) 40192(1.1100) Shrinking 1.976(0.3925) 33628(1.1500) ---------- ------------- ------------- ./test/dtypes.c Added various compound datatype conversion tests. Fixed return values from functions. ./src/H5T.c Conversion timers are updated only if H5T debugging is turned on a runtime (in addition to compile time). This allows the data type layer to be compiled with debugging support without having to pay a big runtime penalty if the debugging isn't actually used. ./src/H5A.c ./src/H5D.c ./src/H5Ofill.c ./src/H5P.c ./src/H5T.c ./src/H5Tconv.c ./src/H5Tpkg.h ./src/H5Tprivate.h ./src/H5Tpublic.h Added the stride argument to the H5T_convert() calls. The stride is always zero, which means that the source and destination data values are packed. ./configure.in ./configure [REGENERATED] If API tracing is turned off then libhdf5.settings will say `no' instead of nothing. ./test/flush1.c ./test/flush2.c Added better error messages in a couple places.