diff options
Diffstat (limited to 'release_docs/RELEASE.txt')
| -rw-r--r-- | release_docs/RELEASE.txt | 119 |
1 files changed, 119 insertions, 0 deletions
diff --git a/release_docs/RELEASE.txt b/release_docs/RELEASE.txt index 2d8eb32..bb0448e 100644 --- a/release_docs/RELEASE.txt +++ b/release_docs/RELEASE.txt @@ -368,6 +368,78 @@ New Features Library: -------- + - Added support for _Float16 16-bit half-precision floating-point datatype + + Support for the _Float16 C datatype has been added on platforms where: + + - The _Float16 datatype and its associated macros (FLT16_MIN, FLT16_MAX, + FLT16_EPSILON, etc.) are available + - A simple test program that converts between the _Float16 datatype and + other datatypes with casts can be successfully compiled and run at + configure time. Some compilers appear to be buggy or feature-incomplete + in this regard and will generate calls to compiler-internal functions + for converting between the _Float16 datatype and other datatypes, but + will not link these functions into the build, resulting in build + failures. + + The following new macros have been added: + + H5_HAVE__FLOAT16 - This macro is defined in H5pubconf.h and will have + the value 1 if support for the _Float16 datatype is + available. It will not be defined otherwise. + + H5_SIZEOF__FLOAT16 - This macro is defined in H5pubconf.h and will have + a value corresponding to the size of the _Float16 + datatype, as computed by sizeof(). It will have the + value 0 if support for the _Float16 datatype is not + available. + + H5_HAVE_FABSF16 - This macro is defined in H5pubconf.h and will have the + value 1 if the fabsf16 function is available for use. + + H5_LDOUBLE_TO_FLOAT16_CORRECT - This macro is defined in H5pubconf.h and + will have the value 1 if the platform can + correctly convert long double values to + _Float16. Some compilers have issues with + this. + + H5T_NATIVE_FLOAT16 - This macro maps to the ID of an HDF5 datatype representing + the native C _Float16 datatype for the platform. If + support for the _Float16 datatype is not available, the + macro will map to H5I_INVALID_HID and should not be used. + + H5T_IEEE_F16BE - This macro maps to the ID of an HDF5 datatype representing + a big-endian IEEE 754 16-bit floating-point datatype. This + datatype is available regardless of whether _Float16 support + is available or not. + + H5T_IEEE_F16LE - This macro maps to the ID of an HDF5 datatype representing + a little-endian IEEE 754 16-bit floating-point datatype. + This datatype is available regardless of whether _Float16 + support is available or not. + + The following new hard datatype conversion paths have been added, but + will only be used when _Float16 support is available: + + H5T_NATIVE_SCHAR <-> H5T_NATIVE_FLOAT16 | H5T_NATIVE_UCHAR <-> H5T_NATIVE_FLOAT16 + H5T_NATIVE_SHORT <-> H5T_NATIVE_FLOAT16 | H5T_NATIVE_USHORT <-> H5T_NATIVE_FLOAT16 + H5T_NATIVE_INT <-> H5T_NATIVE_FLOAT16 | H5T_NATIVE_UINT <-> H5T_NATIVE_FLOAT16 + H5T_NATIVE_LONG <-> H5T_NATIVE_FLOAT16 | H5T_NATIVE_ULONG <-> H5T_NATIVE_FLOAT16 + H5T_NATIVE_LLONG <-> H5T_NATIVE_FLOAT16 | H5T_NATIVE_ULLONG <-> H5T_NATIVE_FLOAT16 + H5T_NATIVE_FLOAT <-> H5T_NATIVE_FLOAT16 | H5T_NATIVE_DOUBLE <-> H5T_NATIVE_FLOAT16 + H5T_NATIVE_LDOUBLE <-> H5T_NATIVE_FLOAT16 + + The H5T_NATIVE_LDOUBLE -> H5T_NATIVE_FLOAT16 hard conversion path will only + be available and used if H5_LDOUBLE_TO_FLOAT16_CORRECT has a value of 1. Otherwise, + the conversion will be emulated in software by the library. + + Note that in the absence of any compiler flags for architecture-specific + tuning, the generated code for datatype conversions with the _Float16 type + may perform conversions by first promoting the type to float. Use of + architecture-specific tuning compiler flags may instead allow for the + generation of specialized instructions, such as AVX512-FP16 instructions, + if available. + - Implemented optimized support for vector I/O in the Subfiling VFD Previously, the Subfiling VFD would handle vector I/O requests by @@ -591,6 +663,42 @@ Bug Fixes since HDF5-1.14.0 release Library ------- + - Fixed a bug that causes the library to incorrectly identify + the endian-ness of 16-bit and smaller C floating-point datatypes + + When detecting the endian-ness of an in-memory C floating-point + datatype, the library previously always assumed that the type + was at least 32 bits in size. This resulted in invalid memory + accesses and would usually cause the library to identify the + datatype as having an endian-ness of H5T_ORDER_VAX. This has + now been fixed. + + - Fixed a bug that causes an invalid memory access issue when + converting 16-bit floating-point values to integers with the + library's software conversion function + + The H5T__conv_f_i function previously always assumed that + floating-point values were at least 32 bits in size and would + access invalid memory when attempting to convert 16-bit + floating-point values to integers. To fix this, parts of the + H5T__conv_f_i function had to be rewritten, which also resulted + in a significant speedup when converting floating-point values + to integers where the library does not have a hard conversion + path. This is the case for any floating-point values with a + datatype not represented by H5T_NATIVE_FLOAT16 (if _Float16 is + supported), H5T_NATIVE_FLOAT, H5T_NATIVE_DOUBLE or + H5T_NATIVE_LDOUBLE. + + - Fixed a bug that can cause incorrect data when overflows occur + while converting integer values to floating-point values with + the library's software conversion function + + The H5T__conv_i_f function had a bug which previously caused it + to return incorrect data when an overflow occurs and an application's + conversion exception callback function decides not to handle the + overflow. Rather than return positive infinity, the library would + return truncated data. This has now been fixed. + - Fixed error when overwriting certain nested variable length types Previously, when using a datatype that included a variable length type @@ -1330,6 +1438,17 @@ Bug Fixes since HDF5-1.14.0 release Testing ------- + - Fixed a bug in the dt_arith test when H5_WANT_DCONV_EXCEPTION is not + defined + + The dt_arith test program's test_particular_fp_integer sub-test tries + to ensure that the library correctly raises a datatype conversion + exception when converting a floating-point value to an integer overflows. + However, this test would run even when H5_WANT_DCONV_EXCEPTION isn't + defined, causing the test to fail due to the library not raising + datatype conversion exceptions. This has now been fixed by not running + the test when H5_WANT_DCONV_EXCEPTION is not defined. + - Disabled running of MPI Atomicity tests for OpenMPI major versions < 5 Support for MPI atomicity operations is not implemented for major |