[PIR] OneDNN Pir onednn instruction

[wanghuancoder]

PR types

New features

PR changes

Others

Description

PIR支持OneDNN，机制建设——OneDNN Dialect添加。
目前，本地测试Program中有use_mkldnn的单测，经过ProgramTranslator已经能够成功运行到Instruction。
并支持了OneDNNPhiKernelInstruction，并跑通了Conv2d。

IR转换如下：

IR before lowering = {
 (%0) = "pd_op.feed" () {col:(Int32)1,is_persisable:[false],name:"Filter",stop_gradient:[false]} : () -> pd_op.tensor<6x3x3x3xf32>
 (%1) = "pd_op.feed" () {col:(Int32)0,is_persisable:[false],name:"Input",stop_gradient:[false]} : () -> pd_op.tensor<2x3x5x5xf32>
 (%2) = "pd_onednn_op.conv2d" (%1, %0) {data_format:"NCHW",dilations:[(Int32)1,(Int32)1],groups:(Int32)1,is_persisable:[false],is_test:false,padding_algorithm:"EXPLICIT",paddings:[(Int32)0,(Int32)0],stop_gradient:[false],strides:[(Int32)1,(Int32)1]} : (pd_op.tensor<2x3x5x5xf32>, pd_op.tensor<6x3x3x3xf32>) -> pd_op.tensor<2x6x3x3xf32>
 (%3) = "pd_op.fetch" (%2) {col:(Int32)0,name:"Output"} : (pd_op.tensor<2x6x3x3xf32>) -> pd_op.tensor<2x6x3x3xf32>
}

IR after lowering = {
 (%0) = "(phi_kernel)" () {col:(Int32)1,is_persisable:[false],kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"",name:"Filter",op_name:"pd_op.feed",stop_gradient:[false]} : () -> cpu_tensor<6x3x3x3xf32>
 (%1) = "(phi_kernel)" () {col:(Int32)0,is_persisable:[false],kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"",name:"Input",op_name:"pd_op.feed",stop_gradient:[false]} : () -> cpu_tensor<2x3x5x5xf32>
 (%2) = "pd_onednn_kernel.phi_kernel" (%1, %0) {data_format:"NCHW",dilations:[(Int32)1,(Int32)1],dynamic_fallback:false,extra_args:["is_test"],groups:(Int32)1,is_onednn_only:false,is_persisable:[false],is_test:false,kernel_key:<backend:ONEDNN|layout:ONEDNN|dtype:float32>,kernel_name:"conv2d",layout_transform_arg:"data_format",layout_transform_inputs:["input"],op_name:"pd_onednn_op.conv2d",padding_algorithm:"EXPLICIT",paddings:[(Int32)0,(Int32)0],stop_gradient:[false],strides:[(Int32)1,(Int32)1]} : (cpu_tensor<2x3x5x5xf32>, cpu_tensor<6x3x3x3xf32>) -> cpu_tensor<2x6x3x3xf32>
 (%3) = "onednn_to_paddle_layout(phi_kernel)" (%2) {dst_layout:(Int32)0,is_persisable:[false],kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"onednn_to_paddle_layout",op_name:"pd_op.onednn_to_paddle_layout"} : (cpu_tensor<2x6x3x3xf32>) -> cpu_tensor<2x6x3x3xf32>
 (%4) = "fetch(phi_kernel)" (%3) {col:(Int32)0,kernel_key:<backend:CPU|layout:Undefined(AnyLayout)|dtype:float32>,kernel_name:"fetch",name:"Output",op_name:"pd_op.fetch"} : (cpu_tensor<2x6x3x3xf32>) -> cpu_tensor<2x6x3x3xf32>
}

Pcard-67164

[winter-wang]

LGTM

[kangguangli]

LGTM for program translator part.

[Aurelius84]

LGTM

[Aurelius84]

REMOVE_ITEM支持移除多个，可以合并：https://cmake.org/cmake/help/v3.0/command/list.html

[wanghuancoder]

done, thx!

[yuanlehome]

LGTM

[XiaoguangHu01]

LGTM