An operator can have different kernel implementations, and each operator will have a map to store the related kernels. Fluid uses `OpKernelType` as a key to identify a unique Kernel. Before an operator runs, an certain kernel must be chosen by a key of `OpKernelType`. Currently, `OpKernelType` is defined as follows:
An operator can have different kernel implementations, and each operator will have a map to store the related kernels. Fluid uses `OpKernelType` as a key to identify a unique kernel. Before an operator runs, a certain type of kernel must be chosen via a key of `OpKernelType`. Currently, `OpKernelType` is defined as follows:
```cpp
struct OpKernelType {
...
...
@@ -10,13 +10,13 @@ struct OpKernelType {
```
For more details, please refer to [codes](https://github.com/PaddlePaddle/Paddle/blob/2d5ec16bc8a09fb8e0f62c89b116b0cd1d333907/paddle/framework/operator.h#L348-L374) in github.
It contains two keys, `Place` and `DataType`. And these two keys will be hashed to a unique key to represent a certain type of kernel. However, these two keys are not enough. We need a more complete representation of `OpKernelType`.
It contains two keys, `Place` and `DataType`. And these two keys will be hashed to a unique key to represent a certain type of kernel. However, these two keys do not provide enough information. We need a more complete representation of `OpKernelType`.
We often implement a kernel of an operator with some computing library in certain device(place). Please remind that computing library and device are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices.
We often implement a kernel of an operator with some computing library on certain device(place). Please note that computing library and device do not have a one-to-one correspondence. A device can have a lot of computing libraries and a computing library can also support different devices.
For example, Eigen library can support Nvidia GPU/AMD GPU/CPU. And MKLDNN library can support Intel CPU/Intel FPGA. Both `Place` and `Library` should be a key of `OpKernelType`.
For example, Eigen library supports Nvidia GPU/AMD GPU/CPU and MKLDNN library supports Intel CPU/Intel FPGA. Both `Place` and `Library` should be a key of `OpKernelType`.
It's obvious that different DataTypes, like fp64/fp32/int8 will have different kernels. But the data layout of a Tensor will also lead to different implementation. Please refer to the batch norm operator [kernels](https://github.com/PaddlePaddle/Paddle/blob/a948fac4d0ad7e0412d373b8aabeb711c2899563/paddle/operators/batch_norm_op.cc#L180-L209). Data Layout should also be taken into consideration.
Different DataTypes, such as fp64/fp32/int8, will obviously have different kernels. But different data layout of a Tensor will also lead to different implementations. Please refer to the batch norm operator [kernels](https://github.com/PaddlePaddle/Paddle/blob/a948fac4d0ad7e0412d373b8aabeb711c2899563/paddle/operators/batch_norm_op.cc#L180-L209) as an example. Data layout should also be taken into consideration.
`Place` is to represent the device memory where data is locating.
`Place` represents the device memory where data is located.
### Library
...
...
@@ -52,10 +52,10 @@ One operator kernel is usually implemented based on one library. `Library` is de
enum Library { Plain, MKLDNN, CUDNN };
```
We use `Plain` enumerator to represent default library. Since most operators in Fluid are implemented based on `Eigen` library, we take `Eigen` library as the `Plain` enumerator.
A library usually has a corresponding `DeviceContext` which contains some handles needed by computation. Fluid now have two default DeviceContexts in CPU and CUDA, `CPUDeviceContext` and `CUDADeviceContext`. `CPUDeviceContext` contains a Eigen library handle and `CDUADeviceContext` contains a Eigen library handle and cuBLAS handle.
We use `Plain` enumerator to represent default library. Since most operators in Fluid are implemented based on the `Eigen` library, we take `Eigen` library as the `Plain` enumerator.
A library usually has a corresponding `DeviceContext` which contains some handles needed for computation. Fluid now has two default DeviceContexts for CPU and CUDA, namely, `CPUDeviceContext` and `CUDADeviceContext`. `CPUDeviceContext` contains an Eigen library handle and `CDUADeviceContext` contains an Eigen library handle and a cuBLAS handle.
If we want to support new Library, a new enumerator need to be added to `Library` and a new corresponding `LibraryDeviceContext` will be created.
If we want to support new library, a new enumerator need to be added to `Library` and a corresponding new `LibraryDeviceContext` need to be created.
### DataType
...
...
@@ -67,15 +67,15 @@ If we want to support new Library, a new enumerator need to be added to `Library
Actually, a Tensor is a view of a block of memory. Besides a pointer to the memory, we also have to get some other descriptions of this block of memory, such as shape(ddim), stride, and layout.
Different layout leads to different implementation of operator kernel. There are mainly 4 principles we have to follow to support layout in our fluid framework.
Different layout leads to different implementation of the operator kernel. There are mainly 4 principles we have to follow to support layout in our Fluid framework.
- We take layout as a data member of Tensor. Layout is actually a enum variable. If fluid is built with MKLDNN, then, the memory format in MKLDNN will be added into this enum variable too.
- We take layout as a data member of Tensor. Layout is actually a enum variable. If Fluid is built with MKLDNN, then the memory format in MKLDNN will also be added into this enum variable.
- Users have to set layout for input data. And some operators like fill_constant/random, also have to set layout of generating data. Of course, we can have some default layout, like NCHW.
- Users have to set layout for input data. And some operators like fill_constant/random, also have to set layout for generating data. Of course, we can have some default layout, like NCHW.
- The inference of Layout is at run-time, not compile-time.
- The inference of Layout is at run-time, not at compile-time.
- Every operator have to implement different kernels for different layouts. Let's take MKLDNN as an example, if we want to implement a MKLDNN convolution operator, we have to realize all the kernels for different layout, list at [here](http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html). And we will have a special macro to do registering kernels for MKLDNN operators.
- Every operator has to implement different kernels for different layouts. Let's take MKLDNN as an example. If we want to implement an MKLDNN convolution operator, we have to implement all the kernels for different layouts, which are listed [here](http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html). And we will have a special macro to register kernels for MKLDNN operators.
<spanid="design-doc-the-keys-of-operator-kernel-type"></span><h1>Design Doc: The Keys of Operator Kernel Type<aclass="headerlink"href="#design-doc-the-keys-of-operator-kernel-type"title="Permalink to this headline">¶</a></h1>
<divclass="section"id="problem">
<spanid="problem"></span><h2>Problem<aclass="headerlink"href="#problem"title="Permalink to this headline">¶</a></h2>
<p>An operator can have different kernel implementations, and each operator will have a map to store the related kernels. Fluid uses <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code> as a key to identify a unique Kernel. Before an operator runs, an certain kernel must be chosen by a key of <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code>. Currently, <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code> is defined as follows:</p>
<p>An operator can have different kernel implementations, and each operator will have a map to store the related kernels. Fluid uses <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code> as a key to identify a unique kernel. Before an operator runs, a certain type of kernel must be chosen via a key of <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code>. Currently, <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code> is defined as follows:</p>
<p>For more details, please refer to <aclass="reference external"href="https://github.com/PaddlePaddle/Paddle/blob/2d5ec16bc8a09fb8e0f62c89b116b0cd1d333907/paddle/framework/operator.h#L348-L374">codes</a> in github.</p>
<p>It contains two keys, <codeclass="docutils literal"><spanclass="pre">Place</span></code> and <codeclass="docutils literal"><spanclass="pre">DataType</span></code>. And these two keys will be hashed to a unique key to represent a certain type of kernel. However, these two keys are not enough. We need a more complete representation of <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code>.</p>
<p>We often implement a kernel of an operator with some computing library in certain device(place). Please remind that computing library and device are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices.</p>
<p>For example, Eigen library can support Nvidia GPU/AMD GPU/CPU. And MKLDNN library can support Intel CPU/Intel FPGA. Both <codeclass="docutils literal"><spanclass="pre">Place</span></code> and <codeclass="docutils literal"><spanclass="pre">Library</span></code> should be a key of <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code>.</p>
<p>It’s obvious that different DataTypes, like fp64/fp32/int8 will have different kernels. But the data layout of a Tensor will also lead to different implementation. Please refer to the batch norm operator <aclass="reference external"href="https://github.com/PaddlePaddle/Paddle/blob/a948fac4d0ad7e0412d373b8aabeb711c2899563/paddle/operators/batch_norm_op.cc#L180-L209">kernels</a>. Data Layout should also be taken into consideration.</p>
<p>It contains two keys, <codeclass="docutils literal"><spanclass="pre">Place</span></code> and <codeclass="docutils literal"><spanclass="pre">DataType</span></code>. And these two keys will be hashed to a unique key to represent a certain type of kernel. However, these two keys do not provide enough information. We need a more complete representation of <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code>.</p>
<p>We often implement a kernel of an operator with some computing library on certain device(place). Please note that computing library and device do not have a one-to-one correspondence. A device can have a lot of computing libraries and a computing library can also support different devices.</p>
<p>For example, Eigen library supports Nvidia GPU/AMD GPU/CPU and MKLDNN library supports Intel CPU/Intel FPGA. Both <codeclass="docutils literal"><spanclass="pre">Place</span></code> and <codeclass="docutils literal"><spanclass="pre">Library</span></code> should be a key of <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code>.</p>
<p>Different DataTypes, such as fp64/fp32/int8, will obviously have different kernels. But different data layout of a Tensor will also lead to different implementations. Please refer to the batch norm operator <aclass="reference external"href="https://github.com/PaddlePaddle/Paddle/blob/a948fac4d0ad7e0412d373b8aabeb711c2899563/paddle/operators/batch_norm_op.cc#L180-L209">kernels</a> as an example. Data layout should also be taken into consideration.</p>
</div>
<divclass="section"id="solution">
<spanid="solution"></span><h2>Solution<aclass="headerlink"href="#solution"title="Permalink to this headline">¶</a></h2>
...
...
@@ -236,14 +236,14 @@
<spanclass="p">};</span>
</pre></div>
</div>
<p>Following is the details:</p>
<p>The details are as follows:</p>
<divclass="section"id="place">
<spanid="place"></span><h3>Place<aclass="headerlink"href="#place"title="Permalink to this headline">¶</a></h3>
<p><codeclass="docutils literal"><spanclass="pre">Place</span></code> is defined as follows:</p>
<p><codeclass="docutils literal"><spanclass="pre">Place</span></code> is defined as:</p>
<p>We use <codeclass="docutils literal"><spanclass="pre">Plain</span></code> enumerator to represent default library. Since most operators in Fluid are implemented based on <codeclass="docutils literal"><spanclass="pre">Eigen</span></code> library, we take <codeclass="docutils literal"><spanclass="pre">Eigen</span></code> library as the <codeclass="docutils literal"><spanclass="pre">Plain</span></code> enumerator.
A library usually has a corresponding <codeclass="docutils literal"><spanclass="pre">DeviceContext</span></code> which contains some handles needed by computation. Fluid now have two default DeviceContexts in CPU and CUDA, <codeclass="docutils literal"><spanclass="pre">CPUDeviceContext</span></code> and <codeclass="docutils literal"><spanclass="pre">CUDADeviceContext</span></code>. <codeclass="docutils literal"><spanclass="pre">CPUDeviceContext</span></code> contains a Eigen library handle and <codeclass="docutils literal"><spanclass="pre">CDUADeviceContext</span></code> contains a Eigen library handle and cuBLAS handle.</p>
<p>If we want to support new Library, a new enumerator need to be added to <codeclass="docutils literal"><spanclass="pre">Library</span></code> and a new corresponding <codeclass="docutils literal"><spanclass="pre">LibraryDeviceContext</span></code> will be created.</p>
<p>We use <codeclass="docutils literal"><spanclass="pre">Plain</span></code> enumerator to represent default library. Since most operators in Fluid are implemented based on the <codeclass="docutils literal"><spanclass="pre">Eigen</span></code> library, we take <codeclass="docutils literal"><spanclass="pre">Eigen</span></code> library as the <codeclass="docutils literal"><spanclass="pre">Plain</span></code> enumerator.
A library usually has a corresponding <codeclass="docutils literal"><spanclass="pre">DeviceContext</span></code> which contains some handles needed for computation. Fluid now has two default DeviceContexts for CPU and CUDA, namely, <codeclass="docutils literal"><spanclass="pre">CPUDeviceContext</span></code> and <codeclass="docutils literal"><spanclass="pre">CUDADeviceContext</span></code>. <codeclass="docutils literal"><spanclass="pre">CPUDeviceContext</span></code> contains an Eigen library handle and <codeclass="docutils literal"><spanclass="pre">CDUADeviceContext</span></code> contains an Eigen library handle and a cuBLAS handle.</p>
<p>If we want to support new library, a new enumerator need to be added to <codeclass="docutils literal"><spanclass="pre">Library</span></code> and a corresponding new <codeclass="docutils literal"><spanclass="pre">LibraryDeviceContext</span></code> need to be created.</p>
</div>
<divclass="section"id="datatype">
<spanid="datatype"></span><h3>DataType<aclass="headerlink"href="#datatype"title="Permalink to this headline">¶</a></h3>
...
...
@@ -262,12 +262,12 @@ A library usually has a corresponding <code class="docutils literal"><span class
<divclass="section"id="layout">
<spanid="layout"></span><h3>Layout<aclass="headerlink"href="#layout"title="Permalink to this headline">¶</a></h3>
<p>Actually, a Tensor is a view of a block of memory. Besides a pointer to the memory, we also have to get some other descriptions of this block of memory, such as shape(ddim), stride, and layout.</p>
<p>Different layout leads to different implementation of operator kernel. There are mainly 4 principles we have to follow to support layout in our fluid framework.</p>
<p>Different layout leads to different implementation of the operator kernel. There are mainly 4 principles we have to follow to support layout in our Fluid framework.</p>
<ulclass="simple">
<li>We take layout as a data member of Tensor. Layout is actually a enum variable. If fluid is built with MKLDNN, then, the memory format in MKLDNN will be added into this enum variable too.</li>
<li>Users have to set layout for input data. And some operators like fill_constant/random, also have to set layout of generating data. Of course, we can have some default layout, like NCHW.</li>
<li>The inference of Layout is at run-time, not compile-time.</li>
<li>Every operator have to implement different kernels for different layouts. Let’s take MKLDNN as an example, if we want to implement a MKLDNN convolution operator, we have to realize all the kernels for different layout, list at <aclass="reference external"href="http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html">here</a>. And we will have a special macro to do registering kernels for MKLDNN operators.</li>
<li>We take layout as a data member of Tensor. Layout is actually a enum variable. If Fluid is built with MKLDNN, then the memory format in MKLDNN will also be added into this enum variable.</li>
<li>Users have to set layout for input data. And some operators like fill_constant/random, also have to set layout for generating data. Of course, we can have some default layout, like NCHW.</li>
<li>The inference of Layout is at run-time, not at compile-time.</li>
<li>Every operator has to implement different kernels for different layouts. Let’s take MKLDNN as an example. If we want to implement an MKLDNN convolution operator, we have to implement all the kernels for different layouts, which are listed <aclass="reference external"href="http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html">here</a>. And we will have a special macro to register kernels for MKLDNN operators.</li>
</ul>
<p><codeclass="docutils literal"><spanclass="pre">Layout</span></code> is also defined as a enum variable:</p>
An operator can have different kernel implementations, and each operator will have a map to store the related kernels. Fluid uses `OpKernelType` as a key to identify a unique Kernel. Before an operator runs, an certain kernel must be chosen by a key of `OpKernelType`. Currently, `OpKernelType` is defined as follows:
An operator can have different kernel implementations, and each operator will have a map to store the related kernels. Fluid uses `OpKernelType` as a key to identify a unique kernel. Before an operator runs, a certain type of kernel must be chosen via a key of `OpKernelType`. Currently, `OpKernelType` is defined as follows:
```cpp
struct OpKernelType {
...
...
@@ -10,13 +10,13 @@ struct OpKernelType {
```
For more details, please refer to [codes](https://github.com/PaddlePaddle/Paddle/blob/2d5ec16bc8a09fb8e0f62c89b116b0cd1d333907/paddle/framework/operator.h#L348-L374) in github.
It contains two keys, `Place` and `DataType`. And these two keys will be hashed to a unique key to represent a certain type of kernel. However, these two keys are not enough. We need a more complete representation of `OpKernelType`.
It contains two keys, `Place` and `DataType`. And these two keys will be hashed to a unique key to represent a certain type of kernel. However, these two keys do not provide enough information. We need a more complete representation of `OpKernelType`.
We often implement a kernel of an operator with some computing library in certain device(place). Please remind that computing library and device are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices.
We often implement a kernel of an operator with some computing library on certain device(place). Please note that computing library and device do not have a one-to-one correspondence. A device can have a lot of computing libraries and a computing library can also support different devices.
For example, Eigen library can support Nvidia GPU/AMD GPU/CPU. And MKLDNN library can support Intel CPU/Intel FPGA. Both `Place` and `Library` should be a key of `OpKernelType`.
For example, Eigen library supports Nvidia GPU/AMD GPU/CPU and MKLDNN library supports Intel CPU/Intel FPGA. Both `Place` and `Library` should be a key of `OpKernelType`.
It's obvious that different DataTypes, like fp64/fp32/int8 will have different kernels. But the data layout of a Tensor will also lead to different implementation. Please refer to the batch norm operator [kernels](https://github.com/PaddlePaddle/Paddle/blob/a948fac4d0ad7e0412d373b8aabeb711c2899563/paddle/operators/batch_norm_op.cc#L180-L209). Data Layout should also be taken into consideration.
Different DataTypes, such as fp64/fp32/int8, will obviously have different kernels. But different data layout of a Tensor will also lead to different implementations. Please refer to the batch norm operator [kernels](https://github.com/PaddlePaddle/Paddle/blob/a948fac4d0ad7e0412d373b8aabeb711c2899563/paddle/operators/batch_norm_op.cc#L180-L209) as an example. Data layout should also be taken into consideration.
`Place` is to represent the device memory where data is locating.
`Place` represents the device memory where data is located.
### Library
...
...
@@ -52,10 +52,10 @@ One operator kernel is usually implemented based on one library. `Library` is de
enum Library { Plain, MKLDNN, CUDNN };
```
We use `Plain` enumerator to represent default library. Since most operators in Fluid are implemented based on `Eigen` library, we take `Eigen` library as the `Plain` enumerator.
A library usually has a corresponding `DeviceContext` which contains some handles needed by computation. Fluid now have two default DeviceContexts in CPU and CUDA, `CPUDeviceContext` and `CUDADeviceContext`. `CPUDeviceContext` contains a Eigen library handle and `CDUADeviceContext` contains a Eigen library handle and cuBLAS handle.
We use `Plain` enumerator to represent default library. Since most operators in Fluid are implemented based on the `Eigen` library, we take `Eigen` library as the `Plain` enumerator.
A library usually has a corresponding `DeviceContext` which contains some handles needed for computation. Fluid now has two default DeviceContexts for CPU and CUDA, namely, `CPUDeviceContext` and `CUDADeviceContext`. `CPUDeviceContext` contains an Eigen library handle and `CDUADeviceContext` contains an Eigen library handle and a cuBLAS handle.
If we want to support new Library, a new enumerator need to be added to `Library` and a new corresponding `LibraryDeviceContext` will be created.
If we want to support new library, a new enumerator need to be added to `Library` and a corresponding new `LibraryDeviceContext` need to be created.
### DataType
...
...
@@ -67,15 +67,15 @@ If we want to support new Library, a new enumerator need to be added to `Library
Actually, a Tensor is a view of a block of memory. Besides a pointer to the memory, we also have to get some other descriptions of this block of memory, such as shape(ddim), stride, and layout.
Different layout leads to different implementation of operator kernel. There are mainly 4 principles we have to follow to support layout in our fluid framework.
Different layout leads to different implementation of the operator kernel. There are mainly 4 principles we have to follow to support layout in our Fluid framework.
- We take layout as a data member of Tensor. Layout is actually a enum variable. If fluid is built with MKLDNN, then, the memory format in MKLDNN will be added into this enum variable too.
- We take layout as a data member of Tensor. Layout is actually a enum variable. If Fluid is built with MKLDNN, then the memory format in MKLDNN will also be added into this enum variable.
- Users have to set layout for input data. And some operators like fill_constant/random, also have to set layout of generating data. Of course, we can have some default layout, like NCHW.
- Users have to set layout for input data. And some operators like fill_constant/random, also have to set layout for generating data. Of course, we can have some default layout, like NCHW.
- The inference of Layout is at run-time, not compile-time.
- The inference of Layout is at run-time, not at compile-time.
- Every operator have to implement different kernels for different layouts. Let's take MKLDNN as an example, if we want to implement a MKLDNN convolution operator, we have to realize all the kernels for different layout, list at [here](http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html). And we will have a special macro to do registering kernels for MKLDNN operators.
- Every operator has to implement different kernels for different layouts. Let's take MKLDNN as an example. If we want to implement an MKLDNN convolution operator, we have to implement all the kernels for different layouts, which are listed [here](http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html). And we will have a special macro to register kernels for MKLDNN operators.
<spanid="design-doc-the-keys-of-operator-kernel-type"></span><h1>Design Doc: The Keys of Operator Kernel Type<aclass="headerlink"href="#design-doc-the-keys-of-operator-kernel-type"title="永久链接至标题">¶</a></h1>
<p>An operator can have different kernel implementations, and each operator will have a map to store the related kernels. Fluid uses <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code> as a key to identify a unique Kernel. Before an operator runs, an certain kernel must be chosen by a key of <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code>. Currently, <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code> is defined as follows:</p>
<p>An operator can have different kernel implementations, and each operator will have a map to store the related kernels. Fluid uses <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code> as a key to identify a unique kernel. Before an operator runs, a certain type of kernel must be chosen via a key of <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code>. Currently, <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code> is defined as follows:</p>
<p>For more details, please refer to <aclass="reference external"href="https://github.com/PaddlePaddle/Paddle/blob/2d5ec16bc8a09fb8e0f62c89b116b0cd1d333907/paddle/framework/operator.h#L348-L374">codes</a> in github.</p>
<p>It contains two keys, <codeclass="docutils literal"><spanclass="pre">Place</span></code> and <codeclass="docutils literal"><spanclass="pre">DataType</span></code>. And these two keys will be hashed to a unique key to represent a certain type of kernel. However, these two keys are not enough. We need a more complete representation of <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code>.</p>
<p>We often implement a kernel of an operator with some computing library in certain device(place). Please remind that computing library and device are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices.</p>
<p>For example, Eigen library can support Nvidia GPU/AMD GPU/CPU. And MKLDNN library can support Intel CPU/Intel FPGA. Both <codeclass="docutils literal"><spanclass="pre">Place</span></code> and <codeclass="docutils literal"><spanclass="pre">Library</span></code> should be a key of <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code>.</p>
<p>It’s obvious that different DataTypes, like fp64/fp32/int8 will have different kernels. But the data layout of a Tensor will also lead to different implementation. Please refer to the batch norm operator <aclass="reference external"href="https://github.com/PaddlePaddle/Paddle/blob/a948fac4d0ad7e0412d373b8aabeb711c2899563/paddle/operators/batch_norm_op.cc#L180-L209">kernels</a>. Data Layout should also be taken into consideration.</p>
<p>It contains two keys, <codeclass="docutils literal"><spanclass="pre">Place</span></code> and <codeclass="docutils literal"><spanclass="pre">DataType</span></code>. And these two keys will be hashed to a unique key to represent a certain type of kernel. However, these two keys do not provide enough information. We need a more complete representation of <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code>.</p>
<p>We often implement a kernel of an operator with some computing library on certain device(place). Please note that computing library and device do not have a one-to-one correspondence. A device can have a lot of computing libraries and a computing library can also support different devices.</p>
<p>For example, Eigen library supports Nvidia GPU/AMD GPU/CPU and MKLDNN library supports Intel CPU/Intel FPGA. Both <codeclass="docutils literal"><spanclass="pre">Place</span></code> and <codeclass="docutils literal"><spanclass="pre">Library</span></code> should be a key of <codeclass="docutils literal"><spanclass="pre">OpKernelType</span></code>.</p>
<p>Different DataTypes, such as fp64/fp32/int8, will obviously have different kernels. But different data layout of a Tensor will also lead to different implementations. Please refer to the batch norm operator <aclass="reference external"href="https://github.com/PaddlePaddle/Paddle/blob/a948fac4d0ad7e0412d373b8aabeb711c2899563/paddle/operators/batch_norm_op.cc#L180-L209">kernels</a> as an example. Data layout should also be taken into consideration.</p>
<p>We use <codeclass="docutils literal"><spanclass="pre">Plain</span></code> enumerator to represent default library. Since most operators in Fluid are implemented based on <codeclass="docutils literal"><spanclass="pre">Eigen</span></code> library, we take <codeclass="docutils literal"><spanclass="pre">Eigen</span></code> library as the <codeclass="docutils literal"><spanclass="pre">Plain</span></code> enumerator.
A library usually has a corresponding <codeclass="docutils literal"><spanclass="pre">DeviceContext</span></code> which contains some handles needed by computation. Fluid now have two default DeviceContexts in CPU and CUDA, <codeclass="docutils literal"><spanclass="pre">CPUDeviceContext</span></code> and <codeclass="docutils literal"><spanclass="pre">CUDADeviceContext</span></code>. <codeclass="docutils literal"><spanclass="pre">CPUDeviceContext</span></code> contains a Eigen library handle and <codeclass="docutils literal"><spanclass="pre">CDUADeviceContext</span></code> contains a Eigen library handle and cuBLAS handle.</p>
<p>If we want to support new Library, a new enumerator need to be added to <codeclass="docutils literal"><spanclass="pre">Library</span></code> and a new corresponding <codeclass="docutils literal"><spanclass="pre">LibraryDeviceContext</span></code> will be created.</p>
<p>We use <codeclass="docutils literal"><spanclass="pre">Plain</span></code> enumerator to represent default library. Since most operators in Fluid are implemented based on the <codeclass="docutils literal"><spanclass="pre">Eigen</span></code> library, we take <codeclass="docutils literal"><spanclass="pre">Eigen</span></code> library as the <codeclass="docutils literal"><spanclass="pre">Plain</span></code> enumerator.
A library usually has a corresponding <codeclass="docutils literal"><spanclass="pre">DeviceContext</span></code> which contains some handles needed for computation. Fluid now has two default DeviceContexts for CPU and CUDA, namely, <codeclass="docutils literal"><spanclass="pre">CPUDeviceContext</span></code> and <codeclass="docutils literal"><spanclass="pre">CUDADeviceContext</span></code>. <codeclass="docutils literal"><spanclass="pre">CPUDeviceContext</span></code> contains an Eigen library handle and <codeclass="docutils literal"><spanclass="pre">CDUADeviceContext</span></code> contains an Eigen library handle and a cuBLAS handle.</p>
<p>If we want to support new library, a new enumerator need to be added to <codeclass="docutils literal"><spanclass="pre">Library</span></code> and a corresponding new <codeclass="docutils literal"><spanclass="pre">LibraryDeviceContext</span></code> need to be created.</p>
<p>Actually, a Tensor is a view of a block of memory. Besides a pointer to the memory, we also have to get some other descriptions of this block of memory, such as shape(ddim), stride, and layout.</p>
<p>Different layout leads to different implementation of operator kernel. There are mainly 4 principles we have to follow to support layout in our fluid framework.</p>
<p>Different layout leads to different implementation of the operator kernel. There are mainly 4 principles we have to follow to support layout in our Fluid framework.</p>
<ulclass="simple">
<li>We take layout as a data member of Tensor. Layout is actually a enum variable. If fluid is built with MKLDNN, then, the memory format in MKLDNN will be added into this enum variable too.</li>
<li>Users have to set layout for input data. And some operators like fill_constant/random, also have to set layout of generating data. Of course, we can have some default layout, like NCHW.</li>
<li>The inference of Layout is at run-time, not compile-time.</li>
<li>Every operator have to implement different kernels for different layouts. Let’s take MKLDNN as an example, if we want to implement a MKLDNN convolution operator, we have to realize all the kernels for different layout, list at <aclass="reference external"href="http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html">here</a>. And we will have a special macro to do registering kernels for MKLDNN operators.</li>
<li>We take layout as a data member of Tensor. Layout is actually a enum variable. If Fluid is built with MKLDNN, then the memory format in MKLDNN will also be added into this enum variable.</li>
<li>Users have to set layout for input data. And some operators like fill_constant/random, also have to set layout for generating data. Of course, we can have some default layout, like NCHW.</li>
<li>The inference of Layout is at run-time, not at compile-time.</li>
<li>Every operator has to implement different kernels for different layouts. Let’s take MKLDNN as an example. If we want to implement an MKLDNN convolution operator, we have to implement all the kernels for different layouts, which are listed <aclass="reference external"href="http://01org.github.io/mkl-dnn/structmkldnn_1_1memory.html">here</a>. And we will have a special macro to register kernels for MKLDNN operators.</li>
</ul>
<p><codeclass="docutils literal"><spanclass="pre">Layout</span></code> is also defined as a enum variable:</p>
