Files

Luo Cheng bb3868d8cd [CPU] OneDNN 2.6 migration (#11627 )

* Migrate on OneDNN 2.7

* [CPU] Enabled brconv implementation

* Post ops optimizations

* [CPU] Enabled I8 precision on activations for Convolution node

* [CPU][WA] Disabled Deconvolution + post ops fusing optimization

* Fixed FQ post op optimization

* [CPU] Optimize post ops processing

* [WA] Add node name if tensor names are empty

* [WA] remove  layout compatibility chheck that leads to the fase-positive exceptions

* [CPU] Optimize processing for FQ + Sum + FQ post ops pattern

* [CPU][WA] Enabled ReduceSum -> AvgPool transformation due to perf issues

* fix compiler error

* rebase onednn master

* cherry pick from 2.7 to 2.6

* [WA] make cpu case to run completed

* fix xmm zero check

* reopen 'FuseDeconvolutionAndSimpleOperation' Transform  to fix CPU 'ConvolutionBackpropDataLayerTest' fail issue

* [WR] Removed failed the ReduceMean tests caused by 21f3555.

* group deconv may crash on memory out of bound

* [WA] Remove the moc fail case by #af4731a1

* testcase conv maxpool will check brgconv instead of jit

* test subgraph added nhwc format check

* fix gemm bf16 win crash

* fix avx2 groupconv accuracy problem

* [WA] remove invalid FQ tests

* WR to disable the LPT multiplyToGroupConv test because the  transformation was disabled in d5e16f

* add gemm int8 binary postops to fix GroupConvolutionQDqTransformation fail

* add gemm int8 binary postops to fix GroupConvolutionQDqTransformation fail

* fix gemm bf16 fail

* Fix ConcatConvSumInPlaceTest

* Add cpuDebugFuncTests target

* [WA] bf16 crash due to MemoryInput/Output

* OVClassBasicTest case typo

* testcase subgraph sets default ENFORCE_BF16 to NO

* fix clang check

* Fix primType check issue

* Fix cpplint error

* MemoryInput/Output support bf16; Enforce bf16 'NO' should enable snipptes

* disable BF16 fusing fakequant testcase

* testcase init support amx check

* testcase for conv brgconv avx512/amx

* testcase for conv brgconv avx512/amx

* WR enforce reorder bug and add NSPC into deconv supported list.

* Compiling issue fix.

* [WA] skip fakequantize fusing in bf16

* mix legacy/new binary postops

* make nightly case run. tested on amx/avx512/avx2.

* [CPU] Add BF16 AMX test for Matmul

* Add CPU dump check tool

* Add verbose log

* Generate exec graph in cpu dump check tool

* fix binary prelu post Ops

* fix cpplint

* Update ONEDNN version to fix AVX2 bug.

* cpu dump check supports compare dump files

* Add a new CPU_DEBUG_CAPS: OV_CPU_SUMMARY_PERF

* change VERBOSE_LOG to DEBUG_LOG

* fix oneDNN register_jit_code log

* fix cpplint

* Add OV_CPU_DEBUG_LOG controls debug logs to show

* Revert reorder WR.

* Enhanced CPU debug logs and breakpoint support

* Enhanced cpu_dump_check with --ports

* Fix DEBUG_LOG compile issue

* GroupDeconvolutionLayerCPUTest extend to add amx test cases

* Add Node into DBUEG_LOG

* cpu_dump_check: Dump results even no port is specified

* FIx MergeTransposeAndReorder for blocked input

* Fix cpu_dump_check result names

* Enhance DEBUG_LOG on edges

* Cpu dump check support shape mismatch

* Fix bi-directionl inplace

* Cpu dump check support inference_precion_hing f32.

* fix windows dump fail.

* fix depthwise nwc conv

* add rtol arg

* win debugbreak

* fix pooling accuracy

* GroupDeconvolutionLayerCPUTest remove invalid test param for nspc

* recover ov onednn fork

* revert af4731a1f1 '[WA] remove layout compatibility chheck'

* [WA] disable avx2 conv3d fusing case

* [WA] disable avx2 conv3d fusing case

* [WA] Disabled weights md transpose in FC to prevent perf degradations

Co-authored-by: dmitrygo <dmitry.gorokhov@intel.com>
Co-authored-by: Vladislav Golubev <vladislav.golubev@intel.com>
Co-authored-by: Zhang Yi3 <yi3.zhang@intel.com>
Co-authored-by: liubo-intel <bo4.liu@intel.com>
Co-authored-by: Luwei Zhou <luwei.zhou@intel.com>
Co-authored-by: Li, Tingqian <tingqian.li@intel.com>
Co-authored-by: xuchen-intel <chen.xu@intel.com>
Co-authored-by: ceciliapeng2011 <cecilia.peng@intel.com>

2022-06-06 18:30:32 +08:00

cpu_dump_check.py

[CPU] OneDNN 2.6 migration (#11627 )

2022-06-06 18:30:32 +08:00

README.md

[CPU] OneDNN 2.6 migration (#11627 )

2022-06-06 18:30:32 +08:00

requirements.txt

[CPU] OneDNN 2.6 migration (#11627 )

2022-06-06 18:30:32 +08:00

README.md

CPU Dump Check Tool

Compile CPU plugin with -DENABLE_DEBUG_CAPS=ON, then this tool allows:

dump each output tensors from CPU plugin:

python3 cpu_dump_check.py -m=/path/to/model dump1

comparing two dumps and analyze differences:

python3 cpu_dump_check.py -m=/path/to/model dump1 dump2

visualize first error map:

python3 cpu_dump_check.py -m=/path/to/model dump1 dump2 -v