dsbarinov1 commented on code in PR #13393: URL: https://github.com/apache/tvm/pull/13393#discussion_r1028135144
########## docs/how_to/deploy/adreno.rst: ########## @@ -0,0 +1,314 @@ +Deploy to Adreno GPU +======================================= + +**Authors**: Daniil Barinov, Egor Churaev, Andrey Malyshev + +Introduction +------------ + +Adreno is a series of graphics processing unit (GPU) semiconductor +intellectual property cores developed by Qualcomm and used in many of +their SoCs. + +The Adreno GPU accelerates the rendering of complex geometries to +deliver high-performance graphics and a rich user experience with low +power consumption. + +This guide will demonstrate :ref:`the benefits of using textures with Adreno<Advantages of the Textures>`, +how to :ref:`build TVM with OpenCL-SDK<Building TVM for Adreno>` (needed by Adreno devices) and TVM RPC +enabled. It will also provide :ref:`example code<Build and deploy model for Adreno>` to better understand the differences in compiling and deploying models +for Adreno devices. + +Advantages of the Textures +-------------------------- + +One of the Adreno's advantages is the clever handling of textures. At +the moment, TVM is able to benefit from this by having texture support +for Adreno. The graph below shows the Adreno A5x architecture. + +|High-level overview of the Adreno A5x architecture for OpenCL| + +*Fig. 1 High-level overview of the Adreno A5x architecture for OpenCL* + +*source:* `OpenCL Optimization and Best Practices for Qualcomm Adreno GPUs <https://dl.acm.org/doi/10.1145/3204919.3204935>`_ + +Reasons of using textures: + +- Texture processor (TP) has a dedicated L1 cache, which is read-only cache and stores data + fetched from level-2 (L2) cache for texture operations (primary + reason) + +- The handling of image boundaries is built-in. + +- Supports numerous image format and data type combinations with + support for automatic format conversions + +Overall, with textures, it is possible to achieve a significant performance boost +compared to OpenCL buffer based solutions. + +Building TVM for Adreno +----------------------- + +This section gives instructions on how to build the Android part of TVM +with OpenCL-SDK and TVM RPC Server in order to deploy models on Adreno. + +Since the process of building TVM for Adreno is exactly the same as the +process of building TVM for Android, please refer to these instructions: +`TVM RPC +Server <https://github.com/apache/tvm/tree/main/apps/cpp_rpc>`_. + +Alternatively, to build a TVM via docker using OpenCL-Headers and set-up +with Android TVM RPC, refer to this guide: `Deploy the Pretrained Model on Android <https://tvm.apache.org/docs/how_to/deploy_models/deploy_model_on_android.html>`_. + +**Prerequisites**: Android NDK, Android Debug Bridge and OpenCL-SDK must +be installed and Android part of TVM must be built: + +- Read documentation about *Android NDK installation* here: https://developer.android.com/ndk +- To get access to adb tools you can see *Android Debug Bridge installation* here: https://developer.android.com/studio/command-line/adb +- For *OpenCL-SDK installation* please refer to official github repository: https://github.com/KhronosGroup/OpenCL-SDK.git + +You can also build the android part of TVM locally. From the root +folder of TVM: + +:: + + mkdir build_android + cd build_android + cmake .. -DUSE_OPENCL=path/to/OpenCL -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_HOME}/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_NATIVE_API_LEVEL=android-28 -DCMAKE_FIND_ROOT_PATH_MODE_PACKAGE=ON -DANDROID_STL=c++_static -DUSE_CPP_RPC=ON + make -jN tvm_runtime tvm_rpc + +where **N** is the number of cores available on your *CPU*. + +At this stage you have built TVM for Adreno. + +Build and deploy model for Adreno +--------------------------------- + +In this section we will focus on target, needed to compile and deploy models for Adreno, demonstrate +the differences in generated kernels with and without textures and, in addition, the +possibility of choosing a different precision for model compilation will +be considered. + +For the complete step-py-step process of compiling and deploying models on +Adreno, including selection of precision, running the inference of the +model, getting the predictions, and measuring the performance please refer to this tutorial: `How To Deploy model on Adreno <https://tvm.apache.org/docs/how_to/deploy_models/deploy_model_on_adreno.html>`_ Review Comment: I could not find references for creating relative links in other _apache_ documents. @driazati what do you think about that? Should we somehow make relative links to the other docs? -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
