1 v2.30 Intel Intelligent Storage Acceleration Library Release Notes
2 ==================================================================
7 3. CHANGE LOG & FEATURES ADDED
12 * Perf tests do not run in Windows environment.
14 * 32-bit lib is not supported in Windows.
21 * Windows nasm support fix.
25 * Fix documentation on gf_vect_mad(). Min length listed as 32 instead of
26 required min 64 bytes.
30 * Fix lack of install for pkg-config files
34 * Fixes for sanitizer warnings.
38 * Fix for nasm on Mac OS X/darwin.
42 * Fix for crc32_iscsi(). Potential read-over for small buffer. For an input
43 buffer length of less than 8 bytes and aligned to an 8 byte boundary, function
44 could read past length. Previously had the possibility to cause a seg fault
45 only for length 0 and invalid buffer passed. Calculated CRC is unchanged.
47 * Fix for compression/decompression of > 4GB files. For streaming compression
48 of extremely large files, the total_out parameter would wrap and could
49 potentially flag an otherwise valid lookback distance as being invalid.
50 Total_out is still 32bit for zlib compatibility. No inconsistent compressed
51 buffers were generated by the issue.
55 * Fix for histogram generation base function.
56 * Fix library build warnings on macOS.
57 * Fix igzip to use bsf instruction when tzcnt is not available.
61 * Fix ISA-L builds for other architectures. Base function and examples
62 sanitized for non-IA builds.
64 * Fix fuzz test script to work with llvm 6.0 builtin libFuzz.
68 * Inflate total_out behavior corrected for in-progress decompression.
69 Previously total_out represented the total bytes decompressed into the output
70 buffer or temp internal buffer. This is changed to be only the bytes put into
73 * Fixed issue with isal_create_hufftables_subset. Affects semi-dynamic
74 compression use case when explicitly creating hufftables from histogram. The
75 _hufftables_subset function could fail to generate length symbols for any
76 length that were never seen.
80 * Fix erasure code test that violates rs matrix bounds.
82 * Fix 0 length file and looping errors in igzip_inflate_test.
86 * Mac OS X/darwin systems no longer require the --target=darwin config option.
87 The autoconf canonical build should detect.
91 * Fix igzip using 32K window and a shared object
93 * Fix igzip undefined instruction error on Nehalem.
95 * Fixed issue in crc performance tests where OS optimizations turned cold cache
96 tests into warm tests.
100 * Fix for windows register save in gf_6vect_mad_avx2.asm. Only affects windows
101 versions of ec_encode_data_update() running with AVX2. A GP register was not
102 properly restored resulting in corruption on return.
106 * Building in unit directories is no longer supported removing the issue of
107 leftover object files causing the top-level make build to fail.
111 * Fix for windows register save overlap in gf_{3-6}vect_dot_prod_sse.asm. Only
112 affects windows versions of erasure code. GP register saves/restore were
113 pushed to same stack area as XMM.
115 3. CHANGE LOG & FEATURES ADDED
116 ------------------------------
119 * Igzip compression enhancements.
120 - New functions for dictionary acceleration. Split dictionary processing and
121 resetting can greatly accelerate the performance of compressing many small
122 files with a dictionary.
123 - New static level 0 header decode tables. Accelerates decompressing small
124 files that are level 0 compressed by skipping the known header parsing.
125 - New feature for igzip cli tool: support for concatenated .gz files. On
126 decompression, igzip will process a series of independent, concatenated .gz
127 files into one output stream.
130 - New vclmul version of crc32_iscsi().
131 - Updates for aarch64.
136 - New AVX512 vclmul versions of crc16_t10dif(), crc32_ieee(), crc32_gzip_refl.
138 * Erasure code improvements
139 - Added AVX512 ec functions with 5 and 6 outputs. Can improve performance for
140 codes with 5 or more parity by running in batches of up to 6 at a time.
144 * New next-arch versions of 64-bit CRC. All norm and reflected 64-bit
145 polynomials are expanded to utilize vpclmulqdq.
149 * New multi-threaded compression option for igzip cli tool
153 * Adler32 added to external API.
154 * Multi-arch improvements.
155 * Performance test improvements.
159 * Igzip performance improvements and features.
160 - Performance improvements for uncompressable files. Random or uncompressable
161 files can be up to 3x faster in level 1 or 2 compression.
162 - Additional small file performance improvments.
163 - New options in igzip cli: use name from header or not, test compressed file.
165 * Multi-arch autoconf script.
166 - Autoconf should detect architecture and run base functions at minimum.
170 * Igzip small file performance improvements and new features.
171 - Better performance on small files.
172 - New gzip/zlib header and trailer handling.
173 - New gzip/zlib header parsing helper functions.
174 - New user-space compression/decompression tool igzip.
176 * New mem unit added with first function isal_zero_detect().
180 * Igzip inflate (decompression) performance improvements.
181 - Implemented multi-byte decode for inflate. Decode can pack up to three
182 symbols into the decode table making some compressed streams decompress much
183 faster depending on the prevalence of short codes.
187 * Igzip: AVX2 version of level 3 compression added.
189 * Erasure code examples
190 - New examples for standard EC encode and decode.
191 - Example of piggyback EC encode and decode.
196 - New compression levels added. ISA-L fast deflate now has more levels to
197 balance speed vs. target compression level. Level 0, 1 are as in previous
198 generations. New levels 2 & 3 target higher compression roughly comparable
199 to zlib levels 2-3. Level 3 is currently only optimized for processors with
202 * New T10dif & copy function - crc16_t10dif_copy()
203 - CRC and copy was added to emulate T10dif operations such as DIF insert and
204 strip. This function stitches together CRC and memcpy operations
205 eliminating an extra data read.
207 * CRC32 iscsi performance improvements
208 - Fixes issue under some distributions where warm cache performance was
214 - Optimized deflate_hash in compression functions.
215 Improves performance of using preset dictionary.
216 - Removed alignment restrictions on input structure.
222 - Add optimized Adler-32 checksum.
224 - Implement zlib compression format.
226 - Add stateful dictionary support.
228 - Add struct reset functions for both deflate and inflate.
230 * Reflected IEEE format CRC32 is released out. Function interface is named
233 * Exact work condition of Erasure Code Reed-Solomon Matrix is determined by new
234 added program gen_rs_matrix_limits.
238 * New 2-pass fully-dynamic deflate compression (level -1). ISA-L fast deflate
239 now has two levels. Level 0 (default) is the same as previous generations.
240 Setting to level 1 will switch to the fully-dynamic compression that will
241 typically reach higher compression ratios.
243 * RAID AVX512 functions.
247 * New fast decompression (inflate)
249 * Compression improvements (deflate)
250 - Speed and compression ratio improvements.
251 - Fast custom Huffman code generation.
253 * Run-time option of gzip crc calculation and headers/trailer.
254 * Choice of static header (BTYPE 01) blocks.
255 * LARGE_WINDOW, 32K history, now default.
256 * Stateless full flush mode.
259 - Six new 64-bit polynomials supported. Normal and reflected versions of ECMA,
260 ISO and Jones polynomials.
264 * Units added: crc, raid, igzip (deflate compression).
268 * Erasure code updates. New AVX512 versions.
270 * Nasm support. ISA-L ported to build with nasm or yasm assembler.
272 * Windows DLL support. Windows builds DLL by default.
276 * Autoconf and autotools build allows easier porting to additional systems.
277 Previous make system still available to embedded users with Makefile.unx.
279 * Includes update for building on Mac OS X/darwin systems. Add --target=darwin
284 * Erasure code improvments
285 - 32-bit port of optimized gf_vect_dot_prod() functions. This makes
286 ec_encode_data() functions much faster on 32-bit processors.
287 - Avoton performance improvements. Performance on Avoton for
288 gf_vect_dot_prod() and ec_encode_data() can improve by as much as 20%.
292 * Incremental erasure code. New functions added to erasure code to handle
293 single source update of code blocks. The function ec_encode_data_update()
294 works with parameters similar to ec_encode_data() but are called incrementally
295 with each source block. These versions are useful when source blocks are not
296 all available at once.
300 * Erasure code updates
301 - New AVX and AVX2 support functions.
302 - Changes min len requirement on gf_vect_dot_prod() to 32 from 16.
303 - Tests include both source and parity recovery with ec_encode_data().
304 - New encoding examples with Vandermonde or Cauchy matrix.
308 * First open release of erasure code unit that is part of ISA-L.