lvchangquan
0b09fdf94c
fix an allreduce bug with two streams sync problem
4 years ago
mindspore-ci-bot
19158780b5
!15978 Add GPU BCEWithLogitsLoss
From: @TFbunny
Reviewed-by: @tom__chen,@robingrosman
Signed-off-by: @robingrosman
4 years ago
TFBunny
9eae68efaa
add gpu BCEWithLogitsLoss kernel
4 years ago
Parastoo Ashtari
7b9a73fb1b
Fixing multi graph suspend for debugger in GPU
removed the suspend from preExecute function to avoid double stopping in multigraph models
replaced else if with else in postExecute
add else if to check the smoke ascend test
improve the format
4 years ago
He Wei
121a6a28d9
[auto-monad] Enforce order of exection for Loads user nodes in frontend
4 years ago
mindspore-ci-bot
3ed60633c8
!15710 gpu inference syupport multi-outputs
From: @wilfchen
Reviewed-by: @limingqi107,@cristoval
Signed-off-by: @cristoval
4 years ago
wilfChen
662bce82ae
inference support multi outputs
4 years ago
limingqi107
179c677fef
fix graph output address set in the one time memory application scenarios
4 years ago
mindspore-ci-bot
4189a0c06f
!15563 [GraphKernel] fix precision error when open graph kernel
From: @zengzitao
Reviewed-by: @limingqi107,@anyrenwei
Signed-off-by: @anyrenwei
4 years ago
zengzitao
1fd87c6d83
add address state to fix cache problem when open graph kernel
4 years ago
limingqi107
9fe2aa1647
fix the device address set of same tensor
4 years ago
limingqi107
888bb91290
fix the input is the output of graph
4 years ago
limingqi107
22972a89a7
support the output address of graph reapply
5 years ago
kswang
2a48b2ecb8
reconstruct session code
5 years ago
tanghuikang
c8a14ba016
Clean code
5 years ago
mindspore-ci-bot
33edd67261
!14667 add dump data function for CPU
From: @zhangbuxue
Reviewed-by:
Signed-off-by:
5 years ago
buxue
ab59f62147
add dump data function for CPU
5 years ago
kswang
97a97e02db
extract load input
5 years ago
louei5
f23ce6c7d9
optimize record gpu memory information
5 years ago
lizhenyu
3f9d9c5b2e
add error log when set device id failed
5 years ago
dayschan
11ee3b1624
add context graph_kernel_flags
used the flag "opt_level" to control GraphKernel,
0 means disabled while non-zero value means enabled.
the default value is controlled by context "enable_graph_kernel",
but if it's also set in "graph_kernel_flags", then the flag will prevail.
supported the whitelist and blacklist operators for GraphKernelExpander.
"enable_expand_ops", "enable_expand_ops_only", "disable_expand_ops".
5 years ago
mindspore-ci-bot
0da0f72c6c
!13329 [MS][RDR] optimize saving FuncGraph to using same params with save_graphs
From: @louie5
Reviewed-by: @zh_qh
Signed-off-by:
5 years ago
louei5
fd51088da8
fix rdr saving FuncGraph and KernelGraph different with save_graphs
5 years ago
mindspore-ci-bot
5fd3d140b6
!13344 add DeviceContext module
From: @zyli2020
Reviewed-by:
Signed-off-by:
5 years ago
lizhenyu
95565aa7b8
add hardware abstract layer
5 years ago
luopengting
c8ba7694c5
refactor RDR to support single name
1. support single name
2. add hash method for pair
3. move constructor and destructor of MemAddressInfo as public
4. remove graph_id
5. modify interval for somas info
5 years ago
TFBunny
4d35303265
support string in GPU print
5 years ago
mindspore-ci-bot
6f6d14d944
!13102 Add unique id for .dat and .dot file to avoid covering
From: @irmo
Reviewed-by:
Signed-off-by:
5 years ago
huanghui
a2ba47e18a
1. Add unique id for .dat and .dot file to avoid covering
2. Dump the end graph in gpu session and cu session
5 years ago
Islam Amin
cbbffbedef
fix gpu dump naming
5 years ago
mindspore-ci-bot
a21c8e13b5
!13010 Add device id log
From: @zpac
Reviewed-by: @cristoval,@wilfchen
Signed-off-by: @cristoval
5 years ago
tanghuikang
6102202abd
Not save InitDatasetQueue and GetNext op in PyNative Mode
5 years ago
ZPaC
f2edee750a
Add device id log
5 years ago
mindspore-ci-bot
7104e42304
!12808 Add graph_ to execution order filename
From: @islam_amin
Reviewed-by: @john_tzanakakis,@tom__chen
Signed-off-by:
5 years ago
caifubi
171b468bb3
PyNative AllReduce Bucket
5 years ago
Islam Amin
ed2f8876b9
adding graph_ to exec order filename
5 years ago
mindspore-ci-bot
7296659f14
!12764 [Ascend][GPU] Add execution order dumping of final execution graphs
From: @islam_amin
Reviewed-by: @john_tzanakakis,@yelihua
Signed-off-by: @yelihua
5 years ago
mindspore-ci-bot
00f25c8409
!12728 fix precision error after cache modification
From: @simson_wu
Reviewed-by: @chujinjin,@zhoufeng54
Signed-off-by: @chujinjin
5 years ago
Islam Amin
187222d461
Adding dump of order of execution for final exec graphs on ascend and gpu
5 years ago
dayschan
c165ab5bb1
Combine the GraphKernelOptimization of Gpu and Ascend
removed one cse from GPU passes,
some common passes was enabled for Ascend.
5 years ago
simson
c29d8f66d8
fix precision error after cache modification
5 years ago
mindspore-ci-bot
5524280075
!12550 [MS][RDR] recording func_graph in pipeline and task debug info
From: @louie5
Reviewed-by:
Signed-off-by:
5 years ago
mindspore-ci-bot
4dedab3775
!12593 Not AllocateMemory when CompileGraph in PyNative mode
From: @HulkTang
Reviewed-by: @zhoufeng54
Signed-off-by:
5 years ago
louei5
9a48405a41
recording func_graph in pipeline and task debug information
5 years ago
Islam Amin
722eb2ec5a
ascend graph dump trigger at data dump
5 years ago
tanghuikang
c346a96529
Not AllocateMemory when CompileGraph in PyNative mode
5 years ago
He Wei
7d9a783993
[auto-monad] Support side-effects by auto-monad
The basic idea is: exploits data dependency to control the execution order
of side-effect operations, and keep the semantics of ANF unchanged.
The ControlDepend primitive is removed and there are two primitives added:
1. UpdateState:
```
a = Assign(para, value)
```
became:
```
a = Assign(para, value, u)
u = UpdateState(u, a)
```
2. Load:
```
x = Add(para, value)
```
became:
```
p = Load(para, u)
x = Add(p, value)
u = UpdateState(u, p)
```
5 years ago
mindspore-ci-bot
0ff27ef3b4
!11930 【GraphKernel】Replace Assign with InplaceAssign
From: @dayschan
Reviewed-by: @gaoxiong1,@dylangeng
Signed-off-by: @gaoxiong1
5 years ago
mindspore-ci-bot
a24ff36d9c
!11777 stitch fusion
From: @r1chardf1d0
Reviewed-by:
Signed-off-by:
5 years ago
dayschan
08345c54ea
[GraphKernel] Replace Assign with InplaceAssign
1. added a pass to replace Assign with InplaceAssign.
2. bugfix in eliminate_redundant_output. the side-effect node should not be eliminated.
3. bugfix in graph_kernel/splitter.py, the kernel includes InplaceAssign should be a composite node.
4. added two tool functions GetAllInputDeviceTypes and GetAllOutputDeviceTypes into AnfAlgo.
5. do not fuse a single Assign in pass BasicOpsFusion.
5 years ago