提交 · 10af44abbaf037cc6c7c06586aed110772e53125 · MegEngine 天元 / MegEngine

15 10月, 2021 1 次提交

fix(dnn/cuda): fix cudnn conv impl for nchw4_nchw hybrid layout · 10af44ab

由 Megvii Engine Team 提交于 10月 14, 2021

the conv_bias algo *_IMPLICIT_GEMM in cudnn less than 8.0.0 is disabled due to the incorrect result for int8x4->f32 configs

GitOrigin-RevId: 7cc52d0a85c5ba345af52c61534e7e82f42cc088

10af44ab

12 10月, 2021 1 次提交
- M
  style(all): reformat c++ code · 369c2ccc
  由 Megvii Engine Team 提交于 10月 12, 2021
```
GitOrigin-RevId: 3ffd1b211f140e8ca05661a3c71801083ae4a951
```
  369c2ccc
26 9月, 2021 2 次提交
- M
  feat(gopt): add nhwc fuse conv typecvt optpass · fca19535
  由 Megvii Engine Team 提交于 9月 14, 2021
```
GitOrigin-RevId: adc230120323835d27786e8908e0f427f3e2f2df
```
  fca19535
- M
  Revert "perf(cuda/conv): cache serval cudnn api" · de363c04
  由 Megvii Engine Team 提交于 9月 15, 2021
```
This reverts commit 188c62cdd65ba27793b0af89164c6eee0122b21f.

GitOrigin-RevId: 92a82b8cd9f1bed053c7be89fa61ffb05acd1279
```
  de363c04
17 9月, 2021 1 次提交
- M
  feat(dnn): add an get_all_algorithms_safe interface · d69b5903
  由 Megvii Engine Team 提交于 9月 06, 2021
```
GitOrigin-RevId: e3734e4531c2f91205814ec6b490f5d6392f3fa3
```
  d69b5903
16 9月, 2021 1 次提交
- M
  feat(dnn): add heuristic cache · 694aa1bd
  由 Megvii Engine Team 提交于 8月 20, 2021
```
GitOrigin-RevId: 35e942b5e39c60d9d4e0ffe41497103e8c0a8822
```
  694aa1bd
01 9月, 2021 1 次提交
- M
  feat(dnn/cuda): add nhwc int8 imma conv and conv fuse typecvt · 11f022ff
  由 Megvii Engine Team 提交于 8月 11, 2021
```
GitOrigin-RevId: 229e1eb4be0fe29932117f226523b446c9ca665e
```
  11f022ff
18 8月, 2021 1 次提交
- M
  refactor(mgb): refactor has-usable-algo function for global optimizer · d195fdec
  由 Megvii Engine Team 提交于 8月 13, 2021
```
GitOrigin-RevId: 66105166505ba81d2ef984bbc8085af5333a1c9b
```
  d195fdec
09 8月, 2021 2 次提交
- M
  feat(dnn/cuda): integrate cutlass operation table and replace all cutlass wrappers · 9b4b910d
  由 Megvii Engine Team 提交于 7月 12, 2021
```
GitOrigin-RevId: 2a70335441e8a844dcf3c2d00bbf6db381ad9623
```
  9b4b910d
- M
  feat(dnn/cuda): use cutlass remove shared load imma conv kernel · b18feaab
  由 Megvii Engine Team 提交于 7月 06, 2021
```
GitOrigin-RevId: 0b5574f52669ba88237967d3f00fceca0857b80f
```
  b18feaab
07 8月, 2021 1 次提交
- M
  feat(imperative): add support for quantized conv transpose2d · 3eb0505f
  由 Megvii Engine Team 提交于 7月 23, 2021
```
GitOrigin-RevId: ffd6431299b2ae008fbdd1eed6458437e6b6a45f
```
  3eb0505f
30 7月, 2021 3 次提交
- M
  feat(mgb): cuda conv support nhwc format and fp16 dtype · 3b452d8c
  由 Megvii Engine Team 提交于 7月 28, 2021
```
GitOrigin-RevId: b8ddcd108a4370a0b093c51bd90ebde0e007cb24
```
  3b452d8c
- M
  refactor(mgb/dnn): fix group conv is_available · 2aba0378
  由 Megvii Engine Team 提交于 7月 08, 2021
```
GitOrigin-RevId: b2799091689336cfc626315885f3d296fd13e70f
```
  2aba0378
- M
  refactor(mgb): refactor group conv · 6ce212d2
  由 Megvii Engine Team 提交于 6月 30, 2021
```
GitOrigin-RevId: 7afd3126900938661927c047aac87cb4aeaa9f13
```
  6ce212d2
19 7月, 2021 12 次提交
- M
  feat(dnn/cuda): generate cutlass kimpls using cmake and bazel · 4eda3388
  由 Megvii Engine Team 提交于 6月 21, 2021
```
GitOrigin-RevId: da3bcfb85af3f148116b8b6a51f214c6cef8684e
```
  4eda3388
- M
  fix(dnn/cuda): fix testcase for fallback nchw qs8 conv · 8d248a6a
  由 Megvii Engine Team 提交于 6月 23, 2021
```
GitOrigin-RevId: 646440db59f0157a3fdbd8061167f9ac04dbd422
```
  8d248a6a
- M
  feat(dnn/cuda): add nhwc int4 conv support · f41a8086
  由 Megvii Engine Team 提交于 6月 11, 2021
```
GitOrigin-RevId: 5236b235d0310feecac0fbd1dc76ff0755cf9426
```
  f41a8086
- M
  refactor(dnn/cuda): refactor cutlass kernel generator for gemm and gemv · 5a14a892
  由 Megvii Engine Team 提交于 5月 31, 2021
```
GitOrigin-RevId: 11d78ab2270f0720d7d79e186124a1254c467980
```
  5a14a892
- M
  refactor(dnn/cuda): refactor kernel generator for cutlass convolution kernels · 4abf7bd3
  由 Megvii Engine Team 提交于 5月 27, 2021
```
GitOrigin-RevId: 7882f9c68c0e0f35159f12a07ccdf027f4540ff1
```
  4abf7bd3
- M
  feat(dnn/cuda): add convolution with i8 input and u4 output · b4687ce8
  由 Megvii Engine Team 提交于 5月 24, 2021
```
GitOrigin-RevId: 8be439abf1f448a6be33ea0e57c48e674c5c94c4
```
  b4687ce8
- M
  fix(dnn/cuda): fix recursive algo search for fallback_nchw_qs8 · 00083d13
  由 Megvii Engine Team 提交于 5月 20, 2021
```
GitOrigin-RevId: 6be2991224bced3a38a17b6b888fd4f324d03f9f
```
  00083d13
- M
  feat(dnn/cuda): add convolution with i8 input and i4 output · 66f70578
  由 Megvii Engine Team 提交于 5月 18, 2021
```
GitOrigin-RevId: 10512645d5d5ac3d985720788760bf8a3855c1f1
```
  66f70578
- M
  feat(gopt/inference): allow Float32 output dtype in EnableNCHW4Pass · 7d3df995
  由 Megvii Engine Team 提交于 6月 11, 2021
```
GitOrigin-RevId: 81100dbaf7d9e8f5aade6c630ead79ea5ba45e52
```
  7d3df995
- M
  fix(dnn/cuda): fix AlgoFallbackNCHWQS8 to support Float32 dst · 633016a9
  由 Megvii Engine Team 提交于 6月 10, 2021
```
GitOrigin-RevId: 06f90f5cf384bc4ddb2f97860e4f530ee9a85705
```
  633016a9
- M
  refactor(mgb): make conv-backward-data handle noncontiguous tensors · 44376f70
  由 Megvii Engine Team 提交于 6月 11, 2021
```
GitOrigin-RevId: 0a8f66f9d378b6466bc383a94c57ec80bcc5cb74
```
  44376f70
- M
  refactor(mgb): make conv handle noncontiguous tensors · 7b2a76d1
  由 Megvii Engine Team 提交于 6月 09, 2021
```
GitOrigin-RevId: 86282709b3013ae44e13f74639a1c45a0dad97b0
```
  7b2a76d1
21 6月, 2021 14 次提交
- M
  feat(dnn/cuda): add nchw conv u4xs4 support · 24a38781
  由 Megvii Engine Team 提交于 4月 28, 2021
```
GitOrigin-RevId: 5edba47bd96371b0afa99d85645e41af5d2f3180
```
  24a38781
- M
  feat(dnn/cuda): add base class for cutlass uint4 and int4 algos · e661ae90
  由 Megvii Engine Team 提交于 5月 21, 2021
```
GitOrigin-RevId: a4d42f032c7e53f2966016092ba52c091575be77
```
  e661ae90
- M
  feat(dnn/cuda): add cutlass impls for uint4 x int4 conv bias · 319436dd
  由 Megvii Engine Team 提交于 5月 17, 2021
```
GitOrigin-RevId: cf4536855ac3faf5a929b1077dac91092b2f008f
```
  319436dd
- M
  feat(dnn/cuda): add cutlass impls for int4 conv bias · d28eba4e
  由 Megvii Engine Team 提交于 5月 07, 2021
```
GitOrigin-RevId: 878bb8c95511de578cb5c880a412a3e1c9f0dea9
```
  d28eba4e
- M
  feat(dnn/cuda): add reduce_filter_and_update_bias · 14b65e4d
  由 Megvii Engine Team 提交于 5月 10, 2021
```
GitOrigin-RevId: 31b6e6b0abe2790029e63c9f91c64290a1801958
```
  14b65e4d
- M
  fix(dnn): fixes for int4 · 86b69cac
  由 Megvii Engine Team 提交于 4月 26, 2021
```
GitOrigin-RevId: 845e164fd3143d2092627374167e74f37b952aed
```
  86b69cac
- M
  feat(dnn/cuda): add conv u4xs4 sass kernel · 4a802d21
  由 Megvii Engine Team 提交于 4月 16, 2021
```
GitOrigin-RevId: 4defcf5f1f33f91c5d92df0282f337e9080dc322
```
  4a802d21
- M
  feat(dnn/cuda): fallback conv qs4 support channel not aligend to 64 · c218d4b0
  由 Megvii Engine Team 提交于 4月 12, 2021
```
GitOrigin-RevId: f0d080f35cc3194f6d45dba5fc58986b4f6c06ab
```
  c218d4b0
- M
  perf(dnn/cuda): optimize int4 sass conv main loop and epilogue without fuse_z · cff61a53
  由 Megvii Engine Team 提交于 4月 12, 2021
```
GitOrigin-RevId: 4274e58d64b8de532c03a137cc86eb6274977a2b
```
  cff61a53
- M
  feat(dnn/cuda): support conv_bias for nchw64 and qint4 · e250afb0
  由 Megvii Engine Team 提交于 3月 23, 2021
```
GitOrigin-RevId: 1c65ba87d7a2276171addcb8883464ada95306a4
```
  e250afb0
- M
  refactor(dnn): refactor lowbit tensor format · 3b9b8780
  由 Megvii Engine Team 提交于 3月 24, 2021
```
GitOrigin-RevId: b646dc085b171589d5dec24650ccf7148f8a03af
```
  3b9b8780
- M
  test(dnn/cuda): add testcase for transforming tensor layout between nchw and nchw64 · 19a554d6
  由 Megvii Engine Team 提交于 3月 18, 2021
```
GitOrigin-RevId: 75d579635ad177d9391b8da6ca45fab1086d3f6a
```
  19a554d6
- M
  feat(dnn/cuda): add relayout format to support layout transform between NCHW and NCHW64 · 71c2f612
  由 Megvii Engine Team 提交于 3月 16, 2021
```
GitOrigin-RevId: 1445ecfabe106eee57494b6bee4df14b9b81556b
```
  71c2f612
- M
  feat(dnn/cuda): add cuda conv bias impls for NCHW format tensors with qint4 data type · df009e89
  由 Megvii Engine Team 提交于 3月 12, 2021
```
GitOrigin-RevId: a0a08cf42c01974d376c9852cd40bb9e3ccb09f4
```
  df009e89

MegEngine 天元 / MegEngine 大约 1 年 前同步成功

MegEngine 天元 / MegEngine
大约 1 年前同步成功