Migrate oneDNN operators to under PHI

[yaomichael]

需求描述 Feature Description

design doc: https://github.com/PaddlePaddle/docs/blob/develop/docs/design/phi/design_en.md

with Op name (op_type) as the unit, MKLDNN currently has a total of 100 OpKernels, which we divide into three categories:

1. Migrate first: a total of 53 (42 to be migrated, 11 have been migrated by YuanRisheng). The parameters required by this type of Kernel are consistent with the CPU&GPU Kernel. Currently, the migration is non-blocking. In order to verify the mechanism, we have migrated 11 of them;
2. Migrate later: 23 in total. This kind of Kernel has the problem of Extra parameters in OpMaker that we mentioned before. It requires more parameters than CPU&GPU Kernel, so this part of Kernel is recommended to be migrated after we complete the extra parameter cleaning. The specific time will be tell you later (expectedly in mid to late September);
3. No need to migrate: 24 in total. This type of Kernel mainly includes two types of operators, one is the historical operators that have been abandoned, such as reshape, flatten, etc, these abandoned operators will be gradually removed from the framework, so they will no need to be migrated; the other is framework function related operators. For example, related to fusion or quantize, such operators may not be migrated to PHI in the future, and may continue to be maintained in the main framework.

替代实现 Alternatives

No response

[paddle-bot]

您好，我们已经收到了您的问题，会安排技术人员尽快解答您的问题，请耐心等待。请您再次检查是否提供了清晰的问题描述、复现代码、环境&版本、报错信息等。同时，您也可以通过查看官网API文档、常见问题、历史Issue、AI社区来寻求解答。祝您生活愉快～

Hi! We've received your issue and please be patient to get responded. We will arrange technicians to answer your questions as soon as possible. Please make sure that you have posted enough message to demo your request. You may also check out the API，FAQ，Github Issue and AI community to get the answer.Have a nice day!

[yaomichael]

@chenwhql @jiangjiajun @jczaja @wozna

[Silv3S]

Merged

• abs, abs_grad, relu6 [PHI] Migrate relu6 and abs kernels #45397
• gaussian_random [PHI] Migrate gaussian_random kernel #45481
• Moving fluid/mkldnn_reuse.h to phi/onednn_reuse.h [PHI] Move oneDNN helper classes to new location #45626
• scale [PHI] Migrate scale kernel #45537
• reduce_sum, reduce_sum_grad, reduce_mean, reduce_mean_grad, reduce_min, reduce_max [PHI] Migrate reduce sum+grad, mean+grad, min and max oneDNN kernels #45536
• cast, clip, clip_grad [PHI] Migrate cast, clip+grad and pool+grad oneDNN kernels #45775
• shape [PHI] Shape op migration #46051
• softmax_grad [PHI] migrate softmax_grad kernel #46257
• slice, slice_grad, split, pad2d, pad3d [PHI] Migrate slice, slice_grad, split, pad and pad3d oneDNN kernels #46101
• concat, concat_grad, expand_v2, expand_v2_grad, fill_constant, nearest_interp_v2, bilinear_interp_v2 [PHI] Migrate concat+grad, expand+grad, fill_constant, nearest_interp and bilinear_interp oneDNN kernels #45863
• revert pool2d, pool2d_grad [PHI] Revert pool and pool_grad oneDNN kernel conversion #45989
• gelu, gelu_grad [PHI] Migrate gelu kernels #45596
• sgd, stack [PHI] Migrate sgd and stack oneDNN kernels #46374
• sum [PHI] Sum op migration #46239
• relu6_grad [PHI] relu6_grad kernel #46501
• hard_swish_grad unit test Enable hard_swish_grad unit test #46621
• fix accuracy drop in gaussian_random [PHI] gaussian_random kernel fix accuracy drop #46747
• transpose2_grad [PHI] transpose2_grad op migration #46139

---

Cherry-picks to release/2.4 branch

• shape [cherry-pick] [PHI] Shape op migration (#46051) #46724
• softmax_grad [cherry-pick] [PHI] migrate softmax_grad kernel (#46257) #46725
• slice, slice_grad, split, pad2d, pad3d [cherry-pick] [PHI] Migrate slice, slice_grad, split, pad and pad3d oneDNN kernels (#46101) #46726
• concat, concat_grad, expand_v2, expand_v2_grad, fill_constant, nearest_interp_v2, bilinear_interp_v2 [cherry-pick] [PHI] Migrate concat+grad, expand+grad, fill_constant … oneDNN kernels (#45863) #46727
• revert pool2d, pool2d_grad [cherry-pick] [PHI] Revert pool+grad oneDNN kernel conversion (#45989) #46860
• gelu, gelu_grad [cherry-pick] [PHI] relu6_grad kernel (#46501) #46862
• sgd, stack [cherry-pick] [PHI] Migrate sgd and stack oneDNN kernels (#46374) #46729
• sum [cherry-pick] [PHI] Sum op migration (#46239) #46730
• relu6_grad [cherry-pick] [PHI] relu6_grad kernel (#46501) #46862
• hard_swish_grad unit test [cherry-pick] [PHI] Enable hard_swish_grad unit test (#46621) #46857
• fix accuracy drop in gaussian_random [cherry-pick] [PHI] gaussian_random kernel fix accuracy drop (#46747) #46858
• transpose2_grad [cherry-pick] [PHI] transpose2_grad op migration (#46139) #46873

[chenwhql]

@Silv3S @piotrekobi Our kernel migration work will be released in Paddle v2.4, and the release/2.4 branch bas been created at 9.9, so the PRs merged into develop after 9.9 need to be cherry-pick to release/2.4, remind attention it, thanks.

https://github.com/PaddlePaddle/Paddle/tree/release/2.4

[Silv3S]

Part II

Merged

• conv2d, conv2d_grad Adapting device-specific Extra Attributes for the PHI kernel #46342
• softmax [PHI] Migrate softmax kernel #47339
• softplus [PHI] Migrate softplus kernel #47406
• pool2d, pool2d_grad [PHI] Migrate pool2d and pool2d_grad kernels #47423
• depthwise_conv2d, conv3d [PHI] Migrate depthwise_conv2d and conv3d kernels #47658
• depthwise_conv2d_grad, conv3d_grad [PHI] Migrate depthwise_conv2d_grad and conv3d_grad kernels #47686
• batch_norm [PHI] Migrate batch_norm #47652
• prelu, prelu_grad [phi] migrate prelu #47422
• Remove redundant FLUID files mkldnn directory cleanup #47779
• conv2d_transpose [PHI] Migrate conv_transpose kernel #48119
• matmul_v2_grad [PHI] Migrate matmul_grad kernel #48023
• mul_grad [PHI] Migrate mul_grad kernel #48061
• elementwise_add_grad, elementwise_sub_grad, elementwise_mul_grad, elementwise_div_grad, elementwise_div [PHI] Migrate elementwise_div + all elementwise grad kernels #48210
• batch_norm_grad [PHI] Migrate batch_norm_grad kernel #48288
• Remove redundant FLUID files [PHI decoupling] Move MKLDNN code #48352
• matmul_v2 [PHI] Migrate matmul kernel #48162
• transpose2 [PHI] traspose2 kernel migration #47748
• elementwise_sub [PHI] Migrate elementwise_sub kernel #48611
• mul Migrate mul_mkldnn_op to phi matmul_kernel #48299
• elementwise_add, elementwise_mul [PHI] Migrate elementwise_(add/mul) kernels #48625
• squeeze2, squeeze2_grad [PHI] Migrate squeeze and squeeze_grad kernels #48634
• reshape2 [PHI] Migrate reshape kernel #48749
• Copy method [PHI] OneDNN version of Copy #48539