[ptweezy]

encode_deflate_icf_06, the AVX-512 encoder for levels 1 to 3, gathers the lit/len and dist codes of each 16 ICF entries with two vpgatherdd, which are slow on AMD. This PR adds encode_deflate_icf_10, encode_df_06.asm built with USE_BYTE_PLANES, which looks up the codes of 64 entries at a time with VBMI byte permutes when a call has at least 288 entries, and gathers only the last entries.

Which CPUs run it

AMD processors with AVX-512 VBMI (Zen 4, Zen 5), at levels 1 to 3: an optional 7th function and 8th argument of mbin_dispatch_init6 select it only on AuthenticAMD processors with VBMI and the G2 features. Selected on every processor with VBMI, it made level 1 3-4% slower on Emerald Rapids and Granite Rapids; on Ice Lake, it made whole files 0.5% slower and 16 and 64 KB slices 1.4% and 2.2% faster (run, both builds on top of my other open changes).

Output does not change, and encode_deflate_icf_06 assembles to the same code as on master.

Related PRs

• igzip: save xmm6-xmm15 at 16-byte steps in the win64 ICF encoders #450 fixes the win64 FUNC_SAVE of encode_df_06.asm, which encode_df_10.asm includes and which corrupts the upper halves of xmm6-xmm14.
• This PR conflicts with igzip: build deflate Huffman codes faster #444 in Makefile.nmake, cmake/igzip.cmake and igzip/Makefile.am.

Results

MB/s, master → this PR, in-process A/B on GitHub runners (Zen 4, Zen 5, harness): median over all runs (two per runner) of the best of 11 rounds; in parentheses, the median over runs of the change in median-round throughput; Silesia files and 256 slices of the Silesia tar.

Runner CPU	L1	L2	L3	L1, 16 KB slices	L1, 64 KB slices
AMD EPYC 9V74 (Zen 4), 3 runners	537 → 552 (+2.8%)	511 → 523 (+2.5%)	302 → 307 (+1.5%)	395 → 405 (+2.5%)	496 → 511 (+3.0%)
AMD EPYC 9V45 (Zen 5), 6 runners	691 → 721 (+3.5%)	659 → 683 (+3.3%)	429 → 438 (+1.9%)	492 → 507 (+2.4%)	630 → 655 (+3.1%)

Testing

• make check passes on Linux and natively on Windows 11 (mingw, Ryzen 7 9800X3D), alone and with igzip: save xmm6-xmm15 at 16-byte steps in the win64 ICF encoders #450. On Windows with igzip: save xmm6-xmm15 at 16-byte steps in the win64 ICF encoders #450, every call to the new encoder and its dispatcher keeps xmm6-xmm15, the callee-saved GPRs, rsp and DF.
• encode_deflate_icf_10 and _06 return the same output and bit state for 45,000 random token streams ending at PROT_NONE pages; level 1-3 output is the same with either in the dispatch slot.
• With faked CPUID and XGETBV (AMD, Intel, Hygon and malformed vendor strings, all combinations of the checked feature bits and XCR0 states), the dispatcher picks the new encoder only on AMD with AVX-512, VBMI and the G2 features.