Megvii Engine Team
a1e6720756
feat(dnn): enable bool comparison
GitOrigin-RevId: 735693b81e
4 years ago
Megvii Engine Team
4a178a8dba
feat(windows/cuda/cmake): support cmake cuda build on windows
GitOrigin-RevId: 4d9832e559
4 years ago
Megvii Engine Team
1915593b6c
fix(version_depend): add a fake version on dev
make dev useable on all megskull version
GitOrigin-RevId: 8c4739d470
4 years ago
Megvii Engine Team
56381f808b
fix(dnn/arm): use vcvtq_f32_s32 for all arm code
GitOrigin-RevId: 27effe7d24
4 years ago
Megvii Engine Team
1173205726
fix(gopt): nchw_nchwxx useable and opt pass use nchw_nchwxx_valid
GitOrigin-RevId: 60942aca5b
4 years ago
Megvii Engine Team
eb18eba87d
fix(gopt): fix nchw44 nchw44_dot gopt test
GitOrigin-RevId: 06b38dcd30
4 years ago
Megvii Engine Team
eab7ab0530
fix(gopt): gen nchw_nchw44 when kernel is optimized
GitOrigin-RevId: 89083be200
4 years ago
Megvii Engine Team
777f3ea970
refactor(gopt): format code
GitOrigin-RevId: 9d5c87000f
4 years ago
Megvii Engine Team
30ce3c60bd
Revert "fix(mgb/opr): change EQ opr's backward_graph to nullptr instead of InvalidGrad"
This reverts commit 6355337641
.
GitOrigin-RevId: 6897f6d45d
4 years ago
Megvii Engine Team
14e71b551b
feat(imperative): add helper for dnn opr caller
GitOrigin-RevId: 828cd69803
4 years ago
Megvii Engine Team
230ab45a1e
fix(mgb/naive): fix naive convolution no dispatch kernel in handle
GitOrigin-RevId: 4038fe23a4
4 years ago
Megvii Engine Team
1bce857cb8
fix(mgb/opr-mm): use comp_node of config as default in CollectiveComm
GitOrigin-RevId: 6b43c9fc93
4 years ago
Megvii Engine Team
27205461ae
feat(mgb/opr-mm): add register info cache for multi-machine oprs
GitOrigin-RevId: d5ae3c5a7c
4 years ago
Megvii Engine Team
96ec586d28
fix(dnn): fix bool cvt
GitOrigin-RevId: 2f883dcbe0
4 years ago
Megvii Engine Team
f829f836b9
test(mgb/index): add empty index desc tests
GitOrigin-RevId: 1a71ad3ede
4 years ago
Megvii Engine Team
e73f2799d0
fix(mgb/index): enable index desc empty
GitOrigin-RevId: 4f0ab7c6e7
4 years ago
Megvii Engine Team
ff60fdb82d
feat(dnn): add bool type cvt on gpu
GitOrigin-RevId: ab0fecf368
4 years ago
Megvii Engine Team
e8571cca51
fix(mgb/cuda): fix cuda host alloc set device
GitOrigin-RevId: f4756e8981
4 years ago
Megvii Engine Team
f7b5eced23
refactor(mgb/opr-mm): set False as default value of local_grad
GitOrigin-RevId: 2f9603b087
4 years ago
Megvii Engine Team
c7b6ef35c1
feat(dnn/cuda): add warp perspective backward mat idx
GitOrigin-RevId: b4b494bb69
5 years ago
Megvii Engine Team
09b5f3d434
fix(mgb/core): fix multi thread pool deactive and multi thread conflict
GitOrigin-RevId: 36787a08a5
4 years ago
Megvii Engine Team
ef239f835f
feat(windows/python_whl): make windows HAPPY for build megbrain python package
GitOrigin-RevId: 92b2c07bf9
4 years ago
Megvii Engine Team
e258812f12
feat(dnn): add bool dtype
GitOrigin-RevId: 98c8a092b4
4 years ago
Megvii Engine Team
734c498d27
perf(mgb/core): improve DevMemAlloc when it has single stream
GitOrigin-RevId: 61874faa6d
4 years ago
Megvii Engine Team
39bd66fc63
fix(mgb): fix TensorRT missing cudaSetDevice
GitOrigin-RevId: 40eb119e48
4 years ago
Megvii Engine Team
ab9dfbcefc
test(mgb): fix tensorrt tests missing cudaSetDevice
GitOrigin-RevId: faeb6ae070
4 years ago
Megvii Engine Team
b43fb1a97c
perf(mgb): add CUDA host memory allocator
test(mgb): add SimpleCachingAlloc test
GitOrigin-RevId: 17f381e4ac
4 years ago
Megvii Engine Team
2afceb4187
fix(mgb/atlas): use dyn output alloc if enable dynamic batchsize
GitOrigin-RevId: 45a6c6ad51
4 years ago
Megvii Engine Team
6bcc6faec8
feat(mge/imperative/opr): modify batch_norm to support frozen BN
fix(mge/imperative): cmake uses MGE_BUILD_IMPERATIVE_RT flag
GitOrigin-RevId: 8ea21af9da
4 years ago
Megvii Engine Team
54d18115b6
fix(imperative): fix grad of BatchNorm
GitOrigin-RevId: 1e8d8afaf2
4 years ago
Megvii Engine Team
80c4705317
perf(mgb): use midout in megbrain to reduce binary size
GitOrigin-RevId: ddc8af79af
4 years ago
Megvii Engine Team
4348960c40
fix(mge/gopt): fix fp16 compute mode
GitOrigin-RevId: 350625d1aa
4 years ago
Megvii Engine Team
9f4060b050
fix(mgb/gopt): add ShuffleShuffleRemovePass assert
GitOrigin-RevId: 69e07c3197
4 years ago
Megvii Engine Team
9b0978597c
chore(core): remove index_sequence polyfill for pre-C++14
GitOrigin-RevId: 8010000bc2
4 years ago
Megvii Engine Team
074645d82e
feat(core): add array_skip
GitOrigin-RevId: ae0d12fbb8
4 years ago
Megvii Engine Team
2faab90bd0
feat(core): add a naive if_constexpr for C++14
GitOrigin-RevId: bd5811ebde
4 years ago
Megvii Engine Team
8f42fe0d92
feat(core): import a std::apply polyfill
GitOrigin-RevId: da95c85b57
4 years ago
Megvii Engine Team
be5dde9da8
fix(mgb/opr-mm): fix group manager sync error
GitOrigin-RevId: 344aae534c
4 years ago
Megvii Engine Team
2dc9190055
fix(fastrun/persistent_cache): fix fastrun crash
GitOrigin-RevId: b3f7bdf7dc
4 years ago
Megvii Engine Team
c1fb3c8c1a
fix(mgb/gopt): fix ConvertF32ToF16Pass endpoints
GitOrigin-RevId: 850eaa0906
4 years ago
Megvii Engine Team
618faf6497
feat(mgb/profiler): dump static infer info
GitOrigin-RevId: bb9150eb83
4 years ago
Megvii Engine Team
5c224c71f0
refactor(mge/imperative): drop ComputingGraphImpl in proxy graph
GitOrigin-RevId: a6e6f2e76b
4 years ago
Megvii Engine Team
5e7d2a91c2
refactor(mgb): add TensorND::proxy_to_default_cpu
GitOrigin-RevId: 3ab8525f1c
4 years ago
Megvii Engine Team
9e9e8ca0fe
fix(mgb): fix some compile errors
GitOrigin-RevId: ec38f8e15a
4 years ago
Megvii Engine Team
d782edf80f
refactor(mgb): decouple node insertion from ComputingGraphImpl
GitOrigin-RevId: 59b45fcb17
5 years ago
Megvii Engine Team
d42cf4cd65
refactor(mgb): replace static_cast<ComputingGraphImpl*> with a checked version
GitOrigin-RevId: d05b114668
5 years ago
Megvii Engine Team
55844d3e24
fix(core): fix CPU compnode get_ram_status_byte free_memory error
GitOrigin-RevId: a6d4351840
5 years ago
Megvii Engine Team
672d4ad0e0
fix(jit): more testcases on the grad of JITExecutor
GitOrigin-RevId: c3bb405979
5 years ago
Megvii Engine Team
bc95e873ef
fix(jit): fix jit grad
a) fix shape mismatch when take grad of JITExecutor including Dimshuffle
b) avoid redundant computation in the grad of JITExecutor
c) not pass unused vars as inputs to the grad of JITExecutor to save device memory
d) traverse internal graph only once in JITExecutor ctor instead of traverse
whole graph in each call of setup_args()
e) expand the gradient graph into the origin graph if all inputs are const
GitOrigin-RevId: ba6a2b29e9
5 years ago
Megvii Engine Team
4cf193ab97
fix(mgb/opr-mm): fix device id in TestOprIORemote.IdentityWithGopt
GitOrigin-RevId: 0184036b8b
4 years ago