* Add mobilenetv3_large performance test

* Disable the BF16 test if the device does not support BF16 computations

* Change test timeout

* remove

* pass

* more pass

* add DistributedFusedLamb op

* polish code

* fix compile error

* compatible with pten changement

* fix rocm compile error

* improve converage

* update upstream/develop

* fix cast_with_ptr.h

* add FLAGS_distributed_lamb_divide_nranks_when_allreduce=1

* fix clip before allreduce

* add use_master_param_norm

* code polish

* fix bug

* fix ROCM ci

* mish unit tests

* code format

* remove unused imports

* code format

* remove hard-coded shape values

* remove timeouts

* remove timeouts v2

* restore timeouts

* rebuild matmul pass: trt and gpu_cpu

* rebuild matmul pass: trt and gpu_cpu

* rebuild matmul pass: trt and gpu_cpu

* rebuild matmul pass: trt and gpu_cpu

* amp list updated

* tests updated

* gray list updated

* amp list updated

* test updated

* Upadate pass in quant2_int8_mkldnn_pass

* Back to the previous scale_matmul order

* Change place of cpu_quantize_placement_pass

* base changes for mul reimplementation

* empty commit

* tmp save

* full implementation of mul bf16/fp32 fwd bwd

* CI fix

* CI rerun

* changed unity build cmake to avoid gpu issues

* removed mul mkldnn from unity build

* added skipping tests if not cpu_bf16

* CI fix

* CI fix

* CI fix

* fix conv act int8 scale

* add unit test for conv+hard_swish

* make post training quant API support dataloader

* Quantize nearest_interp and nearest_interp_v2

* Check if avx_core supported

* Add depthwise_conv2d to supported quantization list

* Fix scatter_op fp16 perf problem.

* Add scatter into black list.

* Add scatter into black list for dygraph.

* support FP16 for more ops

* add amp list tests

* refine reduce_mean_grad

* fix OP benchmark ci

* fix fp16 reduce_mean

* updat ut, but still have some problems

* remove mean/reduce_mean fp16 kernel

* support multi precision update for LAMB

* hide some api

* fix ci uts

* fix lamb output of dygraph

* remove some changes to some PR

* try to fix Py3 CI compile error

* fix test_imperative_optimizer, add lars ut, add layer_norm ut

* fix ut, fix format

* fix ut

* fix windows ci

* add map_matmul passes to quant2_int8_mkldnn_pass

* fix fc+act fuse (activation scale)

* ci fix, c++17 structured bindings not available

* fix ci static check

* reshape+transpose+matmul_v2

* in_name->input_name

* fix pr-ci-static-check

* fix single card 8 unittests in new executor

* fix

* fix

* Support sub graph quant-post

* quantize slice op

* correct test

* fix code formatting

* dequantize matmul and matmul_v2 Y weights in qat2_int8

* review fix

* split conv and mul tests, add matmul test

* fixup

* fix ci build

* remove unused variables

* formatting fix

* remove extra newline at end of file

* test

* test

* rm test

* update

* update

* update

* add unittest

* update

* update save

* fix a quantization bug

本PR是fused_transformer的layer层代码，包含FusedFeedForward的layer层代码和FusedTransformerEncoderLayer的代码。

* add pow2_warmup op

* remove contrib __all__

* add AttrT

* rename

* follow comments

* fix duplicate PADDLE_RESTRICT

* quant support matmul_v2

* fix format

* add sparse_embedding doc

* delete wrong space

* fix error for sample code

* fix error for doc compile

* delete __all__

* modify sample code

Sparse tensor core for convolution requires the input channel dimension is 2:4 structed sparse.
So we have to mask the input channel dimension for using sparse tensor core

* Create stateful OneDNNAXPYHandler object.

This makes it possible to call it multiple times without recreating the
oneDNN primitives every time.

* Prepare SGDOpKernel to reuse its implementation from OneDNN kernel.

* OneDNN SGD kernel.

* Update call to use new OneDNNAXPYHandler object api.

* Setup seed in proper place.

* Enable OneDNN kernel only for single case.

* For dense param and sparse grad.

* Small refactor.

* Enable oneDNN by op attr or by cmd line flag.

* Use int64_t type for number of elements.

* Support dense param and grad from OneDNN kernel.

* Enable SGD OneDNN kernel when use MP BF16 optimizer.

* Force non-copyable/movable OneDNNAXPYHandler.

* Reuse OneDNNAXPYHandler for spare tensors in SUM op.

* Fix SFINAE rules.

* Remove recording event inside AXPY.

* Get rid of internal primitive caching.

* Stop use PP cache mechanims to store mem and primitive obj.
* Handler obj store and reuse needed desc & prim

* Do not derive from MKLDNNHandlerT