Commit Graph
651 Commits
Author SHA1 Message Date
Youmu Liu 2741b34583 left and right use diff pack kernel 2026-06-12 18:25:04 +08:00
Youmu Liu 06a119cc6f make a trial on packing GEMM kernel 2026-06-11 20:03:02 +08:00
Youmu Liu 6b449a341e all mat kernel func use new task pack convension 2026-06-11 17:27:44 +08:00
Youmu Liu 09c1cace48 remove mat panel, mat_task goes on span like 2026-06-10 18:10:52 +08:00
Youmu Liu 0356f5aff8 mat task adapts L4 2026-06-09 18:13:36 +08:00
Youmu Liu 413e41db45 other gemm kernel use half-vector-machine style 2026-06-08 18:09:15 +08:00
Youmu Liu 9d5c0fd8b4 other gemm kernel use half-vector-machine style 2026-06-08 16:18:03 +08:00
Nitload 1c515ea2ae FMA-avx512-f64 kernel use half-vector-machine design 2026-06-08 01:02:34 +08:00
Youmu Liu 4336b0ca5a Move some marcros from x86 ro base head 2026-06-05 13:32:59 +08:00
Youmu Liu 94b2d759ae Gemm/Fma kernel on avx2 use new solution, with less permute and shuff instruction 2026-06-04 21:05:38 +08:00
Youmu Liu 2690cf9341 Gemm/Fma kernel on avx512 use new solution, with less permute and shuff instruction 2026-06-04 16:43:17 +08:00
Youmu Liu 6c38bcfade Find solution to fma, change some type define macros 2026-06-03 18:06:37 +08:00
Youmu Liu 9f3fc3b345 Compress L4 metadata into 32 bytes 2026-06-02 18:27:58 +08:00
Nitload 07f8178431 Remove id from L4 head, and other some useless types. 2026-06-02 00:35:08 +08:00
Nitload 9a09d32960 TLSF in l4 use a shorter head, but may be useless
Signed-off-by: Nitload <YoumuLiu_NSI@whv0t.onmicrosoft.com>
2026-06-01 00:36:46 +08:00
Youmu Liu ddb59c38ce Little change in C code, remove isa specific code in std header 2026-05-28 18:44:53 +08:00
Youmu Liu 15f90aed29 L4 is now a burden 2026-05-21 19:40:00 +08:00
Youmu Liu 0681237c3b L4 is now a burden 2026-05-20 18:07:54 +08:00
Youmu Liu 4694fd9433 L4 is now a burden 2026-05-20 18:07:21 +08:00
Youmu Liu 53ed4de530 L4 is now a burden 2026-05-19 18:09:24 +08:00
Youmu Liu f3cfa37b1c working on L4's tlsf 2026-05-15 18:53:58 +08:00
Youmu Liu b87943495d working on L4's tlsf 2026-05-14 20:15:07 +08:00
Youmu Liu c9fb11b47b working on L4's tlsf 2026-05-13 20:13:37 +08:00
Youmu Liu ae57de4040 tlsf is ok, and fast 2026-05-13 15:08:30 +08:00
Youmu Liu bf9879f59f tlsf briefly ok, debugging 2026-05-12 18:25:48 +08:00
Youmu Liu a99b19fa3f add libs' cmakelists file 2026-05-11 18:11:28 +08:00
Youmu Liu b59d414e20 Remove hlsf, TLSF.alloc is ok 2026-05-11 18:08:35 +08:00
Youmu Liu 5706c89563 Remove ignored files 2026-05-11 09:30:04 +08:00
Youmu Liu 6baae849f8 Improve .gitignore 2026-05-11 09:29:18 +08:00
Nitload 8105af6ca3 Update on 20260511-001810 by Nitload Youmu. Additional Message : Abort HLSF design, use a non-standard TLSF as instead
Signed-off-by: Nitload <YoumuLiu_NSI@whv0t.onmicrosoft.com>
2026-05-11 00:18:10 +08:00
Nitload 5e73469c2e after merge 2026-05-10 06:39:52 +08:00
Youmu Liu 9234dbe99b Abort hlsf, goes back to tlsf, may use 'Arena-ed tlsf' 2026-05-09 18:27:34 +08:00
Youmu Liu 942933dcac Abort hlsf, goes back to tlsf, may use 'Arena-ed tlsf' 2026-05-09 18:18:57 +08:00
Youmu Liu d1aca8f745 hlsf goes bad 2026-05-08 18:09:12 +08:00
Youmu Liu 8aa319a120 going on reconstruct of HLSF 2026-05-07 18:03:33 +08:00
Youmu Liu c343e763ed before rewrite of HLSF 2026-05-07 11:16:41 +08:00
Youmu Liu e636410ba6 mat queue and mat predictor confuses. i need time to design and organize 2026-05-06 18:07:31 +08:00
Nitload 156f6afe85 after merge 2026-05-04 17:25:01 +08:00
Nitload e1167bb071 before merge 2026-05-04 17:22:29 +08:00
Youmu Liu 1338bf2801 add design of .ngra file 2026-04-30 20:38:44 +08:00
Youmu Liu 37a96ba4fc matrix parallel start to use two different policy 2026-04-29 20:01:07 +08:00
Youmu Liu ae664cb541 matrix parallel start to use two different policy 2026-04-27 20:08:10 +08:00
Youmu Liu ecb1f04770 matrix thread pool use ParaWaveFront and heavy sync 2026-04-26 17:38:10 +08:00
Youmu Liu 251762791e L4 will not main stream in matrix schedualing 2026-04-24 17:31:49 +08:00
Youmu Liu 41350b0487 add a pre-kP evaluation function in c 2026-04-24 15:46:58 +08:00
Youmu Liu 2f6617447c L4 is now on new algo, using tricky way to init entryies 2026-04-23 18:07:14 +08:00
Youmu Liu 1a6ce24a27 L4 is now on new algo, using new eviction arrays, using union of hlsf and cache line head. 2026-04-23 17:39:38 +08:00
Youmu Liu 5252be9436 L4 removes RRPV array, will not support cross-predictor reuse, will use out-line predictcount and map count array, will not use inline predict count. 2026-04-22 18:11:00 +08:00
Youmu Liu 4f3e206dcb HLSf use nuint rather than long, L4 now uses hlsf rather than array to alloc cache line. 2026-04-21 20:32:28 +08:00
Youmu Liu 2e2d64a2ce HLSf use nuint rather than long, L4 now uses hlsf rather than array to alloc cache line. 2026-04-20 18:10:14 +08:00