Files
calculet-npu-research-archive/history/topics/02-performance-optimization.tsv
T

55 lines
9.3 KiB
Plaintext

ff9f1440163846534a91e5c6f6e1072e76a98456 904993432e59dbe1ba6e2060cac875117e9e18b9 2026-05-25T14:13:59+08:00 patrickyang update infer launcher to avoid 0 value for csr. Update job engine to get actual task time for debug usage. improve compability to supoort arm
045ef4e602bc4284763491772ffdaa563c7b5e81 3b4a6b5da9706e6a40b964051bd816d5fe0f6418 2026-05-22T14:50:09+08:00 patrickyang reduce risk of false sharing
e759441adb552be27579871112abf452ede026ba 50891ffaf41e1b1ac54a8af52af57383b2426e01 2026-05-21T13:38:37+08:00 patrickyang update log lock to time out while loop
e407b669edbfc892b69851edbb7bf27b9f7addf2 44b2100ffbc5cf5ac428988d62d334face045f06 2026-05-18T15:52:20+08:00 patrickyang update opdump model runner to print infer time
2d6f1ddc7fa08d848b0466cd10943d35a713d1b3 da56db4214491f46be58a11bccf0b62b788107b6 2026-04-20T15:52:56+08:00 patrickyang support to get copy data time consumption, infer time consumption
afbc3ab041906876b2042f63a5bd2fc8bf5234e5 ad34aeee82a8f511a43f2b571c1e5c9d9537542a 2026-04-14T14:09:19+08:00 patrickyang fix multi-threaed env doesn't support time statistics
cd6af5c537398b47b7b01bc11289c387f8d7f454 18323edfcc5e5e3988f31c3deda890d32733be21 2026-04-14T10:58:54+08:00 wantaobo upt infer time for cnn_hparam
d8ae5feb40becd16fa19222cec2815540097c979 616337c570262b3ae1c90ed5c473b426e86484b4 2026-04-13T11:28:31+08:00 patrickyang slightly improve job engine performance at high-freq job execution
64a25123a39635aa8205ab5dff3def93f0b6cf27 d6ea019bf5a3058063947f8d9da55135a48585b9 2026-04-08T10:53:05+08:00 patrickyang upt perf
e2a805571b1a859f12a8f0bc400b2e4bf1f7ae84 7d76e34e5388905fa35b20c810b4fdd4f8b343ab 2026-04-07T16:49:51+08:00 patrickyang analyze wait job done time consume
f605e590665ccf878745346491e0abeef8300638 a67c45863600284fd5f5d5e7ff75bbf27000af6d 2026-04-02T18:31:11+08:00 patrickyang print out each step time consume during infer
f85400d4c88067269da8c40c676e1420eb9bb909 74fe4ab382d517dadf71308cf0b6dd9ad6266078 2026-04-01T11:23:19+08:00 patrickyang support recording data transfer time
29629fd97c915b8f24424633ca18553d981c73d0 b14a1cd6a6ef4f439394a3f3e0f7886a2b8576bf 2026-03-25T15:27:13+08:00 patrickyang rm independent optimization option
b14a1cd6a6ef4f439394a3f3e0f7886a2b8576bf 27b2c8ca75784a376628621e39e543af2a733907 2026-03-25T10:38:13+08:00 patrickyang add more trace log. may slightly impact rt performance
df9caaad5332dba3b10b366a4b946fb7aeb53cef 4ef93c4310288f716d210f487e4e689c58c00229 2026-03-13T11:45:53+08:00 wantaobo upt device time
4ef93c4310288f716d210f487e4e689c58c00229 ff15fff2eebdf4e64ed5192a1eea648fd6286a27 2026-03-13T11:45:02+08:00 wantaobo upt cost time
f967dbe0fff392e22a2a0baa77a34ce2f946886a edb26bec507c7c16be4b658314d497171f9b39d5 2026-03-13T11:20:28+08:00 patrickyang support get infer time consume via job end queue
e31e719a66354bb00eebaba21fe56836059253ba 08f3a6159c3807465db5f5c74918e8ce85269a3b 2026-02-05T16:52:58+08:00 patrickyang update deploy function. now automatically refresh device memory each time when deploy calbin
a03c388d225f0229c4b5fb5ed8b1cab1a623c3bc 0eb11558ce0c54c42cbda06ea35cb9dcb5e200a1 2026-02-02T16:56:01+08:00 patrickyang support ping-pong sync to provent transfer speed is slower than infer
3178a5444a3d7e6d66aac727957d6e4dc7cd54f2 749b0def192ba19783fd007a4ee21e3d3b9da0e2 2026-01-04T14:41:58+08:00 patrickyang improve performance
a298898a064a9f854d021597827ec4d84d09d3e4 78e5765d213aa9a095604efac528a63c9f4eb807 2026-01-04T14:01:17+08:00 patrickyang fix bug causing segmentation fault due to mis-manage buffer life time
a7569f32c574ebbcdc726728493230ad9d75ff6d 3db4f2f1090aa4d8fc6f404fbc611068ef9a1df3 2025-12-30T18:11:23+08:00 patrickyang replace map with vector for better performance
5c676deb55589e7f336647b60a4f2fed5fd15735 d0a569f5653dedc5af3797d9d26d57ce26f2bf61 2025-12-29T17:17:23+08:00 patrickyang Large update: add export macro for ABI stable and C/C++ compatible with extern C. Improve job engine performance
d0a569f5653dedc5af3797d9d26d57ce26f2bf61 8c35e5ca78d5e7db9ab3545e9d173a40150609f0 2025-12-29T17:09:22+08:00 patrickyang disable assert function. improve performance by reducing atomic ops
8c35e5ca78d5e7db9ab3545e9d173a40150609f0 a1a0173983a37b8adeddb0f9c2b2d1a84d88178b 2025-12-29T17:02:55+08:00 patrickyang implement unbalanced thread safe btree. improve thread-safe map try-get function performance
a1a0173983a37b8adeddb0f9c2b2d1a84d88178b 382786360d6aea8f8daf54a558faaaeb778eed5b 2025-12-25T14:52:18+08:00 patrickyang reduce some register rd/wr for continue run the same model
ed1eac8f2ee674e5bf80989271500056860e6a48 b73d76da9b8a9355bbce22f51d2dbf7b35baefcf 2025-12-25T10:23:08+08:00 patrickyang improve pcie driver performance for multi-threads env: Less mutex require
7cb8946b22dc82d1442ee346baeffd6f5fe78dcc 7c183be2c9b7ae4045faedcacbef6fe58ef28450 2025-12-22T13:37:13+08:00 patrickyang temp disable bandwidth update
de9677369c5c56f4b34e89e0c62976e1ba11ba3b a9b9c982674d3ac3fb4b5bef331b1f47f5f37f79 2025-12-16T18:08:06+08:00 patrickyang refactor bandwidth update logic
fbdb86fae305b559c24bfc0744695e98f0725a7c 4cc24d02898e43f70d981d2bb1098b796b75063b 2025-12-16T14:24:06+08:00 patrickyang improve search performance
4cc24d02898e43f70d981d2bb1098b796b75063b 306b4913ebe46c986a9c43cc91cc62b8bdf096ef 2025-12-16T11:01:58+08:00 patrickyang improve performance for search table
5e48b29b8e950c5a261f6aa4c9ed49cd374eaa57 adfc15ea9f96a6e2185131f4353ada7ee89cc66d 2025-12-16T09:50:04+08:00 patrickyang improve job engine performance
1e1d375d56951b52843cabbd68b24480d554a2f6 e919b71433dbfbab0b83c8ad4233d1519fb59188 2025-12-09T17:23:36+08:00 patrickyang update calbin print error for less char size. improve configure performance for ccu with multi-chip
6fbfdafdbc27a3af89554e5555f822244d80e4e9 9fec930a18007bb18ee892bc1ac5f14d34a88bbd 2025-12-04T17:48:56+08:00 patrickyang improve deploy performance. fix protocol internal for repair dead process
894674b583d964b002d22c1045f7e8b84da2ab2f 56cfe55b1b24d037da17e6a28dd0d6a3e2089235 2025-12-03T11:49:49+08:00 patrickyang improve cli and driver performance
be87db88f7d075d5448bc6a19b722301ea9a6e5d 554671db59fa65872fdb15459c44dd85e6722d37 2025-11-26T16:32:35+08:00 patrickyang update device to support update bandwidth
7938fc0caa5fe3f67e7ca2e0b72c348224de023d 64f389018c9710ffc49bc6336cf81426ba5ef079 2025-11-12T15:15:22+08:00 patrickyang improve deploy performance. Discard unordered_multimap. implement with unordered_map
bbde495fd949804273e2939b68849793f750adc7 08132b58e84f66ac56d5325db769926556159460 2025-11-07T17:36:02+08:00 patrickyang fix bugs cause bus error. sync class status to shared mem lifetime
45ccc877f411385a3906e5bc80b830f6bb0d8b89 4efc33f7101958ac44dd691e68628b945a0b0496 2025-09-15T16:09:37+08:00 patrickyang update calbin performance at string handling
18c9cb15118f87968f1fb47e1cbd9738289005e7 d28d57fa1553adccf0bbff3f47f05fddf8ee996b 2025-09-09T15:40:10+08:00 patrickyang update calrt to support backdoor data and comparing data with golden. optimize deploy performance
e6a5e63bfa5574ef044552cf377e3eaf8791c48a 67a827171a7b639dbc8477ce47f13993ed08b489 2025-08-07T18:32:09+08:00 patrickyang improve interval map perf
8c1e65b8d6fbda3a65a0fb0c77cd7a24f26402c0 c4c747b7980329807b5636b84f6aff09de043537 2025-07-30T14:30:16+08:00 patrickyang update runtime doc
ba56be32d2d93090917607833fd068b735093911 2fb50f3ebe0d35ef6c107d3d1ed0f02665068381 2025-07-29T11:41:32+08:00 patrickyang improve log perf. fix mis-order print issues when error occurs, triggering by abort
fae87906d397c98ec5505db70a5334bd93272fbb b2a485fe8061d571184425331e36b8c58e2ad47a 2025-07-22T19:10:02+08:00 patrickyang update runtimeAPI
a1e801eb6ca74c7ffdab54a4c02442a80648c504 adef041a6f571d75095ee11d50f842388aeac93a 2025-07-22T14:43:40+08:00 patrickyang add runtime api doc
54e0b1bc760c1c56ed124aac43450ccf0bad5de6 bd3f9aacf35c00b8d1b7671a4e1cd66ff4fc7e49 2025-07-02T11:23:45+08:00 patrickyang add time counter for calrt main
a8ed61e0fa39fc710b52b5e35d131f0423289cc2 99991320774979d141235b80e630d4607e59e0f2 2025-06-16T17:47:01+08:00 patrickyang add time monitor for hostrt->device rt performance
fafbcf40d2c0cb5e28dd032ef0ccd710adef2aac 52ad8dc70666813ab0a74bc9b81e17d4c5f9fcd7 2025-05-09T16:27:56+08:00 patrickyang compile with O2 optimization
dbfdb3da8d269f1927c2ffe7c5ed694ffd1bd424 a0f9e78090831821676669bffe968123399ede20 2025-04-16T16:00:58+08:00 patrickyang support shard thread safe map for high volumn concurrent or multi-thread performance
0afc7d49c882e707581a625aee105f98e3ac9de9 37ee65c5586a73dd2f2fa55e524b9b882e9a683f 2025-04-10T18:39:56+08:00 patrickyang temporary work for multiple times infer and wait result
37ee65c5586a73dd2f2fa55e524b9b882e9a683f 702635cde6e4af1e6ad8bb63127d2e3a1b861f8c 2025-04-10T18:39:43+08:00 patrickyang temporary work for multiple times infer and wait result
98810ac4a7f9dc3e14fd6cfcf8afc6e90f843a45 952acc3cfd75739858654beee313901283a14465 2025-04-01T11:05:52+08:00 patrickyang optimize runtime workflow. Refine object ownership: remove part of massive shared_ptr and unnecssary copy operation
f2ecf793ec3326ef23eb85dd349768c2635f25fa 4c7d655462d72230a99a7c5a28d0acc5ddb0c6d7 2025-01-09T18:28:16+08:00 patrickyang optimize switch case. Add debug macro in cmakefiles
4188323d5be76113b57a656b14ef66332472c916 f7afeb8e7c757d1f4f768fdd024f95c3a8f9c59f 2025-01-09T18:17:29+08:00 patrickyang optimize gen cmd logic. point to point strategy currently
5244c9bb9b24b4c527cc8b018f83b327b252d77d 7ab99fc8756e91ea2ce271462a0577723b590a7e 2024-12-25T16:19:13+08:00 patrickyang add execution time counter