* added matmul_v2 bf16/fp32 FWD kernel

added matmul_v2 bf16/fp32 FWD kernel

* added formatting

* removed some tests due to timeout in CI

* refactored tests

* merged tests classes into one file

* minor change

* removed test guard for CUDA

* remove skipIf

* changes after review

* formated one file

* minor change

* added skipping UT in CUDA place

* [ROCM] fix RNN miopen as weight need to permuted, test=develop

* [ROCM] fix data share when is_test, test=develop

* update, test=develop

* - fix to #33282

* - Increased threshold for elementwise_mul_bf16 grad

* -disabled faulty UT

* - fix to approval

As the title

* base changes for split op

* 90% of split functionality added

* full fp32 functionality

* added bf16 test

* added submemory caching

* added bf test to static mode whitelist

* minor change

* enabled split op for inference

* minor fix

* minor fix

* optimize attr default value, test=develop

* refine, test=develop

* refine, test=develop

* refine, test=develop

* fix bug in AttrReader, test=develop

* fix bug, test=develop

* fix double_grad, test=develop

* refine, test=develop

* refine, test=develop

* fix checker null, test=develop

* for test, test=develop

* refine, test=develop

* refine, test=develop

* refine, test=develop

* refine, test=develop

* refine, test=develop

* refine, test=develop

* new api diagonal, test=develop

* add new api diagonal, test=develop

* new api diagonal, test=develop

* add new api paddle.diagonal, test=develop

* use framework::stride replace ComputeDimStride

* replace cudaMalloc/cudaMemcpy by TensorFormVector in cudaKernel and cudaGradKernel

* perfect funciton: when attr(offset) is exceed attr(axis1) or attr(axis2), set the diagonal dim is 0

* fix RP-Mac-CI bug: replace framework::stride() by ComputDimStride.

* perfect code-block

* perfect code of python API diagonal

* api supports dtype of float16 and bool

* api supports dtype of float16 and bool

* modify unittest code

* modify unittest code

* perfect dtype describe

* perfect code-block

* Add oneDNN AXPY handler.

* Add fallback for small tensors.

* Fix ifdefs

* Remove unnecessary namespace prefixes and add missing headers.

* Guard handler_axpy with proper ifdefs.

* Compilation of this function is possible only when Paddle is not build
with CUDA nor HIP.

* Move AXPY handler code to separate files.

* Use oneDNN AXPY handler in SGD op.

* Use axpy handler only when Paddle is built with oneDNN.

* Add test for SUM BF16 with big rows.

* Fix SFINAE rules for elementwise_add_to.

* Add test case for SGD with big rows.

* update

* update
Co-authored-by: NAdam Osewski <adam.osewski@intel.com>

* use BatchMatMul

* replace TensorCopy with ShareDataWith

* remove check fp16 grad

* fix format

* add grad_check

* fix grad check

* Combine amp and qat
* add unit test

* enable npu alignment

* support flatten_params/grads

* support clip by global norm

* remove memset in coalesce_tensor_op

* fix npu kernel of sum op when input is one tensor

* add ut for flatten_param_grads+regularizer

* fix ut

* fix typo

* batch_norm_act_fuse_pass_init

* repair the unittest of batch_norm_act

* Add bf16 support for save and load ops

* Add bf16 test condition

* Add matmul and chagne fluid.io to paddle.static

* Reduce the test duration

* add atan2_op

* fix

* Add lookup_table_v2 BF16

* Reuse lookup table UT

* Change op_type to op_version

* Remove check_dygraph

* Remove skip_check_grad_ci

* - Draft of implementation of refactoring

- compilation fix

* - Fixes after review

* - Removed unnecessary comment