diff --git a/docs/demo_guides/amlogic_npu.md b/docs/demo_guides/amlogic_npu.md
index 19396472882..8941b001cc0 100644
--- a/docs/demo_guides/amlogic_npu.md
+++ b/docs/demo_guides/amlogic_npu.md
@@ -223,12 +223,13 @@ Paddle Lite 已支持晶晨 NPU 的预测部署。
3）`build.sh` 根据入参生成针对不同操作系统、体系结构的二进制程序，需查阅注释信息配置正确的参数值。
4）`run_with_adb.sh` 入参包括模型名称、操作系统、体系结构、目标设备、设备序列号等，需查阅注释信息配置正确的参数值。
5）`run_with_ssh.sh` 入参包括模型名称、操作系统、体系结构、目标设备、ip地址、用户名、用户密码等，需查阅注释信息配置正确的参数值。
+ 6）下述命令行示例中涉及的具体IP、SSH账号密码、设备序列号等均为示例环境，请用户根据自身实际设备环境修改。
在 ARM CPU 上运行 mobilenet_v1_int8_224_per_layer 全量化模型
$ cd PaddleLite-generic-demo/image_classification_demo/shell
For C308X
- $ ./run_with_adb.sh mobilenet_v1_int8_224_per_layer linux arm64 cpu
+ $ ./run_with_ssh.sh mobilenet_v1_int8_224_per_layer linux arm64 cpu 192.168.100.244 22 root 123456
(C308X)
warmup: 1 repeat: 5, average: 167.6916 ms, max: 207.458000 ms, min: 159.823239 ms
results: 3
@@ -240,7 +241,7 @@ Paddle Lite 已支持晶晨 NPU 的预测部署。
Postprocess time: 0.542000 ms
For A311D
- $ ./run_with_adb.sh mobilenet_v1_int8_224_per_layer linux arm64 cpu
+ $ ./run_with_adb.sh mobilenet_v1_int8_224_per_layer linux arm64 cpu 0123456789ABCDEF
(A311D)
warmup: 1 repeat: 15, average: 81.678067 ms, max: 81.945999 ms, min: 81.591003 ms
results: 3
@@ -252,7 +253,7 @@ Paddle Lite 已支持晶晨 NPU 的预测部署。
Postprocess time: 0.407000 ms
For S905D3(Android版)
- $ ./run_with_ssh.sh mobilenet_v1_int8_224_per_layer android armeabi-v7a cpu
+ $ ./run_with_adb.sh mobilenet_v1_int8_224_per_layer android armeabi-v7a cpu c8631471d5cd
(S905D3(Android版))
warmup: 1 repeat: 5, average: 280.465997 ms, max: 358.815002 ms, min: 268.549812 ms
results: 3
@@ -269,7 +270,7 @@ Paddle Lite 已支持晶晨 NPU 的预测部署。
$ cd PaddleLite-generic-demo/image_classification_demo/shell
For C308X
- $ ./run_with_adb.sh mobilenet_v1_int8_224_per_layer linux arm64 amlogic_npu
+ $ ./run_with_ssh.sh mobilenet_v1_int8_224_per_layer linux arm64 amlogic_npu 192.168.100.244 22 root 123456
(C308X)
warmup: 1 repeat: 5, average: 6.982800 ms, max: 7.045000 ms, min: 6.951000 ms
results: 3
@@ -281,7 +282,7 @@ Paddle Lite 已支持晶晨 NPU 的预测部署。
Postprocess time: 0.509000 ms
For A311D
- $ ./run_with_adb.sh mobilenet_v1_int8_224_per_layer linux arm64 amlogic_npu
+ $ ./run_with_adb.sh mobilenet_v1_int8_224_per_layer linux arm64 amlogic_npu 0123456789ABCDEF
( A311D)
warmup: 1 repeat: 15, average: 5.567867 ms, max: 5.723000 ms, min: 5.461000 ms
results: 3
@@ -293,7 +294,7 @@ Paddle Lite 已支持晶晨 NPU 的预测部署。
Postprocess time: 0.411000 ms
For S905D3(Android版)
- $ ./run_with_adb.sh mobilenet_v1_int8_224_per_layer android armeabi-v7a amlogic_npu
+ $ ./run_with_adb.sh mobilenet_v1_int8_224_per_layer android armeabi-v7a amlogic_npu c8631471d5cd
(S905D3(Android版))
warmup: 1 repeat: 5, average: 13.4116 ms, max: 15.751210 ms, min: 12.433400 ms
results: 3
diff --git a/docs/demo_guides/rockchip_npu.md b/docs/demo_guides/rockchip_npu.md
index 4906745c637..ae92f94442e 100644
--- a/docs/demo_guides/rockchip_npu.md
+++ b/docs/demo_guides/rockchip_npu.md
@@ -208,6 +208,7 @@ Paddle Lite 已支持 Rockchip NPU 的预测部署。
3）`build.sh` 根据入参生成针对不同操作系统、体系结构的二进制程序，需查阅注释信息配置正确的参数值。
4）`run_with_adb.sh` 入参包括模型名称、操作系统、体系结构、目标设备、设备序列号等，需查阅注释信息配置正确的参数值。
5）`run_with_ssh.sh` 入参包括模型名称、操作系统、体系结构、目标设备、ip 地址、用户名、用户密码等，需查阅注释信息配置正确的参数值。
+ 6）下述命令行示例中涉及的具体IP、SSH账号密码、设备序列号等均为示例环境，请用户根据自身实际设备环境修改。
在 ARM CPU 上运行 mobilenet_v1_int8_224_per_layer 全量化模型
$ cd PaddleLite-generic-demo/image_classification_demo/shell
diff --git a/docs/demo_guides/verisilicon_timvx.md b/docs/demo_guides/verisilicon_timvx.md
new file mode 100644
index 00000000000..011fd95f7af
--- /dev/null
+++ b/docs/demo_guides/verisilicon_timvx.md
@@ -0,0 +1,434 @@
+# 芯原 TIM-VX 部署示例
+
+Paddle Lite 已支持通过 TIM-VX 的方式调用芯原 NPU 算力的预测部署。
+其接入原理是与其他接入 Paddle Lite 的新硬件类似，即加载并分析 Paddle 模型，首先将 Paddle 算子转成 NNAdapter 标准算子，其次再通过 TIM-VX 的组网 API 进行网络构建，在线编译模型并执行模型。
+
+需要注意的是，芯原（verisilicon）作为 IP 设计厂商，本身并不提供实体SoC产品，而是授权其 IP 给芯片厂商，如：晶晨（Amlogic），瑞芯微（Rockchip）等。因此本文是适用于被芯原授权了 NPU IP 的芯片产品。只要芯片产品没有大副修改芯原的底层库，则该芯片就可以使用本文档作为 Paddle Lite 推理部署的参考和教程。在本文中，晶晨 SoC 中的 NPU 和 瑞芯微 SoC 中的 NPU 统称为芯原 NPU。
+
+本文档与[ 晶晨 NPU 部署示例 ](./amlogic_npu)和[ 瑞芯微 NPU 部署示例 ](./rockchip_npu)中所描述的部署示例相比，虽然涉及的部分芯片产品相同，但前者是通过 IP 厂商芯原的 TIM-VX 框架接入 Paddle Lite，后二者是通过各自芯片 DDK 接入 Paddle Lite。接入方式不同，支持的算子和模型范围也有所区别。
+
+## 支持现状
+
+### 已支持的芯片
+
+- Amlogic A311D
+
+- Amlogic S905D3
+
+ 注意：理论上支持所有经过芯原授权了 NPU IP 的 SoC（须有匹配版本的 NPU 驱动，下文描述），上述为经过测试的部分。
+
+### 已支持的 Paddle 模型
+
+#### 模型
+
+- [mobilenet_v1_int8_224_per_layer](https://paddlelite-demo.bj.bcebos.com/models/mobilenet_v1_int8_224_per_layer.tar.gz)
+- [resnet50_int8_224_per_layer](https://paddlelite-demo.bj.bcebos.com/models/resnet50_int8_224_per_layer.tar.gz)
+- [ssd_mobilenet_v1_relu_voc_int8_300_per_layer](https://paddlelite-demo.bj.bcebos.com/models/ssd_mobilenet_v1_relu_voc_int8_300_per_layer.tar.gz)
+
+#### 性能
+
+- 测试环境
+ - 编译环境
+ - Ubuntu 16.04，GCC 5.4 for ARMLinux armhf and aarch64
+
+ - 硬件环境
+ - Amlogic A311D
+ - CPU：4 x ARM Cortex-A73 \+ 2 x ARM Cortex-A53
+ - NPU：5 TOPs for INT8
+ - Amlogic S905D3(Android 版本)
+ - CPU：2 x ARM Cortex-55
+ - NPU：1.2 TOPs for INT8
+
+- 测试方法
+ - warmup=1, repeats=5，统计平均时间，单位是 ms
+ - 线程数为1，`paddle::lite_api::PowerMode CPU_POWER_MODE`设置为` paddle::lite_api::PowerMode::LITE_POWER_HIGH `
+ - 分类模型的输入图像维度是{1, 3, 224, 224}
+
+- 测试结果
+
+ |模型 |A311D||S905D3(Android 版本)||
+ |---|---|---|---|---|
+ | |CPU(ms) | NPU(ms) |CPU(ms) | NPU(ms) |
+ |mobilenet_v1_int8_224_per_layer| 81.632133 | 5.112500 | 280.465997 | 12.808100 |
+ |resnet50_int8_224_per_layer| 390.498300| 17.583200 | 787.532340 | 41.313999 |
+ |ssd_mobilenet_v1_relu_voc_int8_300_per_layer| 134.991560| 15.216700 | 295.48919| 40.108970 |
+
+### 已支持（或部分支持）NNAdapter 的 Paddle 算子
+
+您可以查阅[ NNAdapter 算子支持列表](https://github.com/PaddlePaddle/Paddle-Lite/blob/develop/lite/kernels/nnadapter/converter/all.h)获得各算子在不同新硬件上的最新支持信息。
+
+## 参考示例演示
+
+### 测试设备
+
+- Khadas VIM3 开发板（SoC 为 Amlogic A311D）
+
+ 
+
+
+
+- Khadas VIM3L 开发板（SoC 为 Amlogic S905D3)
+
+ 
+
+### 准备设备环境
+
+- A311D
+
+ - 需要驱动版本为 6.4.4.3（下载驱动请联系开发板厂商）。
+
+ - 注意是 64 位系统。
+
+ - 提供了网络连接 SSH 登录的方式，部分系统提供了adb连接的方式。
+
+ - 可通过 `dmesg | grep Galcore` 查询系统版本：
+
+ ```shell
+ $ dmesg | grep Galcore
+ [ 24.140820] Galcore version 6.4.4.3.310723AAA
+ ```
+
+- S905D3(Android 版本)
+
+ - 需要驱动版本为 6.4.4.3（下载驱动请联系开发板厂商）。
+ - 注意是 32 位系统。
+ - `adb root + adb remount` 以获得修改系统库的权限。
+
+ ```shell
+ $ dmesg | grep Galcore
+ [ 9.020168] <6>[ 9.020168@0] Galcore version 6.4.4.3.310723a
+ ```
+
+ - 示例程序和 Paddle Lite 库的编译需要采用交叉编译方式，通过 `adb`或`ssh` 进行设备的交互和示例程序的运行。
+
+
+### 准备交叉编译环境
+
+- 为了保证编译环境一致，建议参考[ Docker 环境准备](../source_compile/docker_environment)中的 Docker 开发环境进行配置；
+- 由于有些设备只提供网络访问方式（根据开发版的实际情况），需要通过 `scp` 和 `ssh` 命令将交叉编译生成的Paddle Lite 库和示例程序传输到设备上执行，因此，在进入 Docker 容器后还需要安装如下软件：
+
+ ```
+ # apt-get install openssh-client sshpass
+ ```
+
+### 运行图像分类示例程序
+
+- 下载 Paddle Lite 通用示例程序[PaddleLite-generic-demo.tar.gz](https://paddlelite-demo.bj.bcebos.com/devices/generic/PaddleLite-generic-demo.tar.gz)，解压后目录主体结构如下：
+
+ ```shell
+ - PaddleLite-generic-demo
+ - image_classification_demo
+ - assets
+ - images
+ - tabby_cat.jpg # 测试图片
+ - tabby_cat.raw # 经过 convert_to_raw_image.py 处理后的 RGB Raw 图像
+ - labels
+ - synset_words.txt # 1000 分类 label 文件
+ - models
+ - mobilenet_v1_int8_224_per_layer
+ - __model__ # Paddle fluid 模型组网文件，可使用 netron 查看网络结构
+ — conv1_weights # Paddle fluid 模型参数文件
+ - batch_norm_0.tmp_2.quant_dequant.scale # Paddle fluid 模型量化参数文件
+ — subgraph_partition_config_file.txt # 自定义子图分割配置文件
+ ...
+ - shell
+ - CMakeLists.txt # 示例程序 CMake 脚本
+ - build.linux.arm64 # arm64 编译工作目录
+ - image_classification_demo # 已编译好的，适用于 arm64 的示例程序
+ - build.linux.armhf # armhf编译工作目录
+ - image_classification_demo # 已编译好的，适用于 armhf 的示例程序
+ - build.android.armeabi-v7a # Android armv7编译工作目录
+ - image_classification_demo # 已编译好的，适用于 Android armv7 的示例程序
+ ...
+ - image_classification_demo.cc # 示例程序源码
+ - build.sh # 示例程序编译脚本
+ - run.sh # 示例程序本地运行脚本
+ - run_with_ssh.sh # 示例程序ssh运行脚本
+ - run_with_adb.sh # 示例程序adb运行脚本
+ - libs
+ - PaddleLite
+ - linux
+ - arm64 # Linux 64 位系统
+ - include # Paddle Lite 头文件
+ - lib # Paddle Lite 库文件
+ - verisilicon_timvx # 芯原 TIM-VX DDK、NNAdapter 运行时库、device HAL 库
+ - libnnadapter.so # NNAdapter 运行时库
+ - libGAL.so # 芯原 DDK
+ - libVSC.so # 芯原 DDK
+ - libOpenVX.so # 芯原 DDK
+ - libarchmodelSw.so # 芯原 DDK
+ - libNNArchPerf.so # 芯原 DDK
+ - libOvx12VXCBinary.so # 芯原 DDK
+ - libNNVXCBinary.so # 芯原 DDK
+ - libOpenVXU.so # 芯原 DDK
+ - libNNGPUBinary.so # 芯原 DDK
+ - libovxlib.so # 芯原 DDK
+ - libOpenCL.so # OpenCL
+ - libverisilicon_timvx.so # # NNAdapter device HAL 库
+ - libtim-vx.so # 芯原 TIM-VX
+ - libgomp.so.1 # gnuomp 库
+ - libpaddle_full_api_shared.so # 预编译 PaddleLite full api 库
+ - libpaddle_light_api_shared.so # 预编译 PaddleLite light api 库
+ ...
+ - android
+ - armeabi-v7a # Android 32 位系统
+ - include # Paddle Lite 头文件
+ - lib # Paddle Lite 库文件
+ - verisilicon_timvx # 芯原 TIM-VX DDK、NNAdapter 运行时库、device HAL 库
+ - libnnadapter.so # NNAdapter 运行时库
+ - libGAL.so # 芯原 DDK
+ - libVSC.so # 芯原 DDK
+ - libOpenVX.so # 芯原 DDK
+ - libarchmodelSw.so # 芯原 DDK
+ - libNNArchPerf.so # 芯原 DDK
+ - libOvx12VXCBinary.so # 芯原 DDK
+ - libNNVXCBinary.so # 芯原 DDK
+ - libOpenVXU.so # 芯原 DDK
+ - libNNGPUBinary.so # 芯原 DDK
+ - libovxlib.so # 芯原 DDK
+ - libOpenCL.so # OpenCL
+ - libverisilicon_timvx.so # # NNAdapter device HAL 库
+ - libtim-vx.so # 芯原 TIM-VX
+ - libgomp.so.1 # gnuomp 库
+ - libc++_shared.so
+ - libpaddle_full_api_shared.so # 预编译 Paddle Lite full api 库
+ - libpaddle_light_api_shared.so # 预编译 Paddle Lite light api 库
+ - OpenCV # OpenCV 预编译库
+ - ssd_detection_demo # 基于 ssd 的目标检测示例程序
+ ```
+
+- 按照以下命令分别运行转换后的ARM CPU模型和 芯原 TIM-VX 模型，比较它们的性能和结果；
+
+ ```shell
+ 注意：
+ 1）`run_with_adb.sh` 不能在 Docker 环境执行，否则可能无法找到设备，也不能在设备上运行。
+ 2）`run_with_ssh.sh` 不能在设备上运行，且执行前需要配置目标设备的 IP 地址、SSH 账号和密码。
+ 3）`build.sh` 根据入参生成针对不同操作系统、体系结构的二进制程序，需查阅注释信息配置正确的参数值。
+ 4）`run_with_adb.sh` 入参包括模型名称、操作系统、体系结构、目标设备、设备序列号等，需查阅注释信息配置正确的参数值。
+ 5）`run_with_ssh.sh` 入参包括模型名称、操作系统、体系结构、目标设备、ip地址、用户名、用户密码等，需查阅注释信息配置正确的参数值。
+ 6）下述命令行示例中涉及的具体IP、SSH账号密码、设备序列号等均为示例环境，请用户根据自身实际设备环境修改。
+
+ 在 ARM CPU 上运行 mobilenet_v1_int8_224_per_layer 全量化模型
+ $ cd PaddleLite-generic-demo/image_classification_demo/shell
+
+ For A311D
+ $ ./run_with_ssh.sh mobilenet_v1_int8_224_per_layer linux arm64 cpu 192.168.100.30 22 khadas khadas
+ (A311D)
+ warmup: 1 repeat: 15, average: 81.678067 ms, max: 81.945999 ms, min: 81.591003 ms
+ results: 3
+ Top0 Egyptian cat - 0.512545
+ Top1 tabby, tabby cat - 0.402567
+ Top2 tiger cat - 0.067904
+ Preprocess time: 1.352000 ms
+ Prediction time: 81.678067 ms
+ Postprocess time: 0.407000 ms
+
+ For S905D3(Android版)
+ $ ./run_with_adb.sh mobilenet_v1_int8_224_per_layer android armeabi-v7a cpu c8631471d5cd
+ (S905D3(Android版))
+ warmup: 1 repeat: 5, average: 280.465997 ms, max: 358.815002 ms, min: 268.549812 ms
+ results: 3
+ Top0 Egyptian cat - 0.512545
+ Top1 tabby, tabby cat - 0.402567
+ Top2 tiger cat - 0.067904
+ Preprocess time: 3.199000 ms
+ Prediction time: 280.465997 ms
+ Postprocess time: 0.596000 ms
+
+ ------------------------------
+
+ 在 芯原 NPU 上运行 mobilenet_v1_int8_224_per_layer 全量化模型
+ $ cd PaddleLite-generic-demo/image_classification_demo/shell
+
+ For A311D
+ $ ./run_with_ssh.sh mobilenet_v1_int8_224_per_layer linux arm64 verisilicon_timvx 192.168.100.30 22 khadas khadas
+ (A311D)
+ warmup: 1 repeat: 15, average: 5.112500 ms, max: 5.223000 ms, min: 5.009130 ms
+ results: 3
+ Top0 Egyptian cat - 0.508929
+ Top1 tabby, tabby cat - 0.415333
+ Top2 tiger cat - 0.064347
+ Preprocess time: 1.356000 ms
+ Prediction time: 5.112500 ms
+ Postprocess time: 0.411000 ms
+
+ For S905D3(Android版)
+ $ ./run_with_adb.sh mobilenet_v1_int8_224_per_layer android armeabi-v7a verisilicon_timvx c8631471d5cd
+ (S905D3(Android版))
+ warmup: 1 repeat: 5, average: 13.4116 ms, max: 14.7615 ms, min: 12.80810 ms
+ results: 3
+ Top0 Egyptian cat - 0.508929
+ Top1 tabby, tabby cat - 0.415333
+ Top2 tiger cat - 0.064347
+ Preprocess time: 3.170000 ms
+ Prediction time: 13.4116 ms
+ Postprocess time: 0.634000 ms
+ ```
+
+- 如果需要更改测试图片，可将图片拷贝到 `PaddleLite-generic-demo/image_classification_demo/assets/images` 目录下，然后调用 `convert_to_raw_image.py` 生成相应的 RGB Raw 图像，最后修改 `run_with_adb.sh`、`run_with_ssh.sh` 的 IMAGE_NAME 变量即可；
+- 重新编译示例程序：
+ ```shell
+ 注意：
+ 1）请根据 `buid.sh` 配置正确的参数值。
+ 2）需在 Docker 环境中编译。
+
+ # 对于 A311D
+ ./build.sh linux arm64
+
+ # 对于 S905D3(Android版)
+ ./build.sh android armeabi-v7a
+ ```
+
+### 更新模型
+- 通过 Paddle 训练或 X2Paddle 转换得到 MobileNetv1 foat32 模型[ mobilenet_v1_fp32_224 ](https://paddlelite-demo.bj.bcebos.com/models/mobilenet_v1_fp32_224_fluid.tar.gz)
+- 通过 Paddle+PaddleSlim 后量化方式，生成[ mobilenet_v1_int8_224_per_layer 量化模型](https://paddlelite-demo.bj.bcebos.com/devices/rockchip/mobilenet_v1_int8_224_fluid.tar.gz)
+- 下载[ PaddleSlim-quant-demo.tar.gz ](https://paddlelite-demo.bj.bcebos.com/tools/PaddleSlim-quant-demo.tar.gz)，解压后清单如下：
+ ```shell
+ - PaddleSlim-quant-demo
+ - image_classification_demo
+ - quant_post # 后量化
+ - quant_post_rockchip_npu.sh # 一键量化脚本，Amlogic 和瑞芯微底层都使用芯原的 NPU，所以通用
+ - README.md # 环境配置说明，涉及 PaddlePaddle、PaddleSlim 的版本选择、编译和安装步骤
+ - datasets # 量化所需要的校准数据集合
+ - ILSVRC2012_val_100 # 从 ImageNet2012 验证集挑选的 100 张图片
+ - inputs # 待量化的 fp32 模型
+ - mobilenet_v1
+ - resnet50
+ - outputs # 产出的全量化模型
+ - scripts # 后量化内置脚本
+ ```
+- 查看 `README.md` 完成 PaddlePaddle 和 PaddleSlim 的安装
+- 直接执行 `./quant_post_rockchip_npu.sh` 即可在 `outputs` 目录下生成mobilenet_v1_int8_224_per_layer 量化模型
+ ```shell
+ ----------- Configuration Arguments -----------
+ activation_bits: 8
+ activation_quantize_type: moving_average_abs_max
+ algo: KL
+ batch_nums: 10
+ batch_size: 10
+ data_dir: ../dataset/ILSVRC2012_val_100
+ is_full_quantize: 1
+ is_use_cache_file: 0
+ model_path: ../models/mobilenet_v1
+ optimize_model: 1
+ output_path: ../outputs/mobilenet_v1
+ quantizable_op_type: conv2d,depthwise_conv2d,mul
+ use_gpu: 0
+ use_slim: 1
+ weight_bits: 8
+ weight_quantize_type: abs_max
+ ------------------------------------------------
+ quantizable_op_type:['conv2d', 'depthwise_conv2d', 'mul']
+ 2021-08-30 05:52:10,048-INFO: Load model and set data loader ...
+ 2021-08-30 05:52:10,129-INFO: Optimize FP32 model ...
+ I0830 05:52:10.139564 14447 graph_pattern_detector.cc:91] --- detected 14 subgraphs
+ I0830 05:52:10.148236 14447 graph_pattern_detector.cc:91] --- detected 13 subgraphs
+ 2021-08-30 05:52:10,167-INFO: Collect quantized variable names ...
+ 2021-08-30 05:52:10,168-WARNING: feed is not supported for quantization.
+ 2021-08-30 05:52:10,169-WARNING: fetch is not supported for quantization.
+ 2021-08-30 05:52:10,170-INFO: Preparation stage ...
+ 2021-08-30 05:52:11,853-INFO: Run batch: 0
+ 2021-08-30 05:52:16,963-INFO: Run batch: 5
+ 2021-08-30 05:52:21,037-INFO: Finish preparation stage, all batch:10
+ 2021-08-30 05:52:21,048-INFO: Sampling stage ...
+ 2021-08-30 05:52:31,800-INFO: Run batch: 0
+ 2021-08-30 05:53:23,443-INFO: Run batch: 5
+ 2021-08-30 05:54:03,773-INFO: Finish sampling stage, all batch: 10
+ 2021-08-30 05:54:03,774-INFO: Calculate KL threshold ...
+ 2021-08-30 05:54:28,580-INFO: Update the program ...
+ 2021-08-30 05:54:29,194-INFO: The quantized model is saved in ../outputs/mobilenet_v1
+ post training quantization finish, and it takes 139.42292165756226.
+
+ ----------- Configuration Arguments -----------
+ batch_size: 20
+ class_dim: 1000
+ data_dir: ../dataset/ILSVRC2012_val_100
+ image_shape: 3,224,224
+ inference_model: ../outputs/mobilenet_v1
+ input_img_save_path: ./img_txt
+ save_input_img: False
+ test_samples: -1
+ use_gpu: 0
+ ------------------------------------------------
+ Testbatch 0, acc1 0.8, acc5 1.0, time 1.63 sec
+ End test: test_acc1 0.76, test_acc5 0.92
+ --------finish eval int8 model: mobilenet_v1-------------
+ ```
+ - 参考[模型转化方法](../user_guides/model_optimize_tool)，利用 opt 工具转换生成 TIM-VX 模型，仅需要将 `valid_targets` 设置为 `verisilicon_timvx`, `arm` 即可。
+ ```shell
+ $ ./opt --model_dir=mobilenet_v1_int8_224_per_layer \
+ --optimize_out_type=naive_buffer \
+ --optimize_out=opt_model \
+ --valid_targets=verisilicon_timvx,arm
+ ```
+### 更新支持 TIM-VX 的 Paddle Lite 库
+
+- 下载 Paddle Lite 源码
+
+ ```shell
+ $ git clone https://github.com/PaddlePaddle/Paddle-Lite.git
+ $ cd Paddle-Lite
+ $ git checkout
+ # 注意：编译中依赖的 verisilicon_timvx 相关代码和依赖项会在后续编译脚本中自动下载，无需用户手动下载。
+ ```
+
+- 编译并生成 `Paddle Lite+Verisilicon_TIMVX` 的部署库
+
+ - For A311D
+ - tiny_publish 编译方式
+ ```shell
+ $ ./lite/tools/build_linux.sh --with_extra=ON --with_log=ON --with_nnadapter=ON --nnadapter_with_verisilicon_timvx=ON --nnadapter_verisilicon_timvx_src_git_tag=main --nnadapter_verisilicon_timvx_viv_sdk_url=http://paddlelite-demo.bj.bcebos.com/devices/verisilicon/sdk/viv_sdk_linux_arm64_6_4_4_3_generic.tgz
+
+ ```
+ - full_publish 编译方式
+ ```shell
+ $ ./lite/tools/build_linux.sh --with_extra=ON --with_log=ON --with_nnadapter=ON --nnadapter_with_verisilicon_timvx=ON --nnadapter_verisilicon_timvx_src_git_tag=main --nnadapter_verisilicon_timvx_viv_sdk_url=http://paddlelite-demo.bj.bcebos.com/devices/verisilicon/sdk/viv_sdk_linux_arm64_6_4_4_3_generic.tgz full_publish
+
+ ```
+ - 替换头文件和库
+ ```shell
+ # 替换 include 目录
+ $ cp -rf build.lite.linux.armv8.gcc/inference_lite_lib.armlinux.armv8.nnadapter/cxx/include/ PaddleLite-generic-demo/libs/PaddleLite/linux/arm64/include/
+ # 替换 NNAdapter 运行时库
+ $ cp -rf build.lite.linux.armv8.gcc/inference_lite_lib.armlinux.armv8.nnadapter/cxx/lib/libnnadapter.so PaddleLite-generic-demo/libs/PaddleLite/linux/arm64/lib/verisilicon_timvx/
+ # 替换 NNAdapter device HAL 库
+ $ cp -rf build.lite.linux.armv8.gcc/inference_lite_lib.armlinux.armv8.nnadapter/cxx/lib/libverisilicon_timvx.so PaddleLite-generic-demo/libs/PaddleLite/linux/arm64/lib/verisilicon_timvx/
+ # 替换 芯原 TIM-VX 库
+ $ cp -rf build.lite.linux.armv8.gcc/inference_lite_lib.armlinux.armv8.nnadapter/cxx/lib/libtim-vx.so PaddleLite-generic-demo/libs/PaddleLite/linux/arm64/lib/verisilicon_timvx/
+ # 替换 libpaddle_light_api_shared.so
+ $ cp -rf build.lite.linux.armv8.gcc/inference_lite_lib.armlinux.armv8.nnadapter/cxx/lib/libpaddle_light_api_shared.so PaddleLite-generic-demo/libs/PaddleLite/linux/arm64/lib/
+ # 替换 libpaddle_full_api_shared.so (仅在 full_publish 编译方式下)
+ $ cp -rf build.lite.linux.armv8.gcc/inference_lite_lib.armlinux.armv8.nnadapter/cxx/lib/libpaddle_full_api_shared.so PaddleLite-generic-demo/libs/PaddleLite/linux/arm64/lib/
+ ```
+
+ - S905D3(Android 版)
+ - tiny_publish 编译方式
+ ```shell
+ $ ./lite/tools/build_android.sh --arch=armv7 --toolchain=clang --android_stl=c++_shared --with_extra=ON --with_exception=ON --with_cv=ON --with_log=ON --with_nnadapter=ON --nnadapter_with_verisilicon_timvx=ON --nnadapter_verisilicon_timvx_src_git_tag=main --nnadapter_verisilicon_timvx_viv_sdk_url=http://paddlelite-demo.bj.bcebos.com/devices/verisilicon/sdk/viv_sdk_android_9_armeabi_v7a_6_4_4_3_generic.tgz
+ ```
+
+ - full_publish 编译方式
+ ```shell
+ $ ./lite/tools/build_android.sh --arch=armv7 --toolchain=clang --android_stl=c++_shared --with_extra=ON --with_exception=ON --with_cv=ON --with_log=ON --with_nnadapter=ON --nnadapter_with_verisilicon_timvx=ON --nnadapter_verisilicon_timvx_src_git_tag=main --nnadapter_verisilicon_timvx_viv_sdk_url=http://paddlelite-demo.bj.bcebos.com/devices/verisilicon/sdk/viv_sdk_android_9_armeabi_v7a_6_4_4_3_generic.tgz full_publish
+ ```
+ - 替换头文件和库
+ ```shell
+ # 替换 include 目录
+ $ cp -rf build.lite.android.armv7.clang/inference_lite_lib.android.armv7.nnadapter/cxx/include/ PaddleLite-generic-demo/libs/PaddleLite/linux/armhf/include/
+ # 替换 NNAdapter 运行时库
+ $ cp -rf build.lite.android.armv7.clang/inference_lite_lib.android.armv7.nnadapter/cxx/lib/libnnadapter.so PaddleLite-generic-demo/libs/PaddleLite/android/armeabi-v7a/lib/verisilicon_timvx/
+ # 替换 NNAdapter device HAL 库
+ $ cp -rf build.lite.android.armv7.clang/inference_lite_lib.android.armv7.nnadapter/cxx/lib/libverisilicon_timvx.so PaddleLite-generic-demo/libs/PaddleLite/android/armeabi-v7a/lib/verisilicon_timvx/
+ # 替换 芯原 TIM-VX 库
+ $ cp -rf build.lite.android.armv7.clang/inference_lite_lib.android.armv7.nnadapter/cxx/lib/libtim-vx.so PaddleLite-generic-demo/libs/PaddleLite/android/armeabi-v7a/lib/verisilicon_timvx/
+ # 替换 libpaddle_light_api_shared.so
+ $ cp -rf build.lite.android.armv7.clang/inference_lite_lib.android.armv7.nnadapter/cxx/lib/libpaddle_light_api_shared.so PaddleLite-generic-demo/libs/PaddleLite/android/armeabi-v7a/lib/
+ # 替换 libpaddle_full_api_shared.so(仅在 full_publish 编译方式下)
+ $ cp -rf build.lite.android.armv7.clang/inference_lite_lib.android.armv7.nnadapter/cxx/lib/libpaddle_full_api_shared.so PaddleLite-generic-demo/libs/PaddleLite/android/armeabi-v7a/lib/
+ ```
+
+- 替换头文件后需要重新编译示例程序
+
+## 其它说明
+
+- Paddle Lite 研发团队正在持续扩展基于TIM-VX的算子和模型。
diff --git a/docs/index.rst b/docs/index.rst
index 37a923d3a1d..7bc950b092a 100644
--- a/docs/index.rst
+++ b/docs/index.rst
@@ -78,6 +78,7 @@ Welcome to Paddle-Lite's documentation!
demo_guides/baidu_xpu
demo_guides/rockchip_npu
demo_guides/amlogic_npu
+ demo_guides/verisilicon_timvx
demo_guides/mediatek_apu
demo_guides/imagination_nna
demo_guides/bitmain
diff --git a/docs/performance/benchmark.md b/docs/performance/benchmark.md
index 60f13c0c2e2..db19374b1f3 100644
--- a/docs/performance/benchmark.md
+++ b/docs/performance/benchmark.md
@@ -137,7 +137,10 @@
请参考 [Paddle Lite 使用瑞芯微 NPU 预测部署](../demo_guides/rockchip_npu)的最新性能数据
## 晶晨 NPU 的性能数据
-请参考 [Paddle Lite 使用晶晨NPU 预测部署](../demo_guides/amlogic_npu)的最新性能数据
+请参考 [Paddle Lite 使用晶晨 NPU 预测部署](../demo_guides/amlogic_npu)的最新性能数据
+
+## 芯原 TIM-VX 的性能数据
+请参考 [Paddle Lite 使用芯原 TIM-VX 预测部署](../demo_guides/verisilicon_timvx)的最新性能数据
## 联发科 APU 的性能数据
请参考 [Paddle Lite 使用联发科 APU 预测部署](../demo_guides/mediatek_apu)的最新性能数据
diff --git a/docs/quick_start/support_model_list.md b/docs/quick_start/support_model_list.md
index fd24a5b5d78..5c65c3176ff 100644
--- a/docs/quick_start/support_model_list.md
+++ b/docs/quick_start/support_model_list.md
@@ -6,7 +6,7 @@
| 类别 | 类别细分 | 模型 | 支持平台 |
|-|-|:-|:-|
-| CV | 分类 | [MobileNetV1](https://paddlelite-demo.bj.bcebos.com/models/mobilenet_v1_fp32_224_fluid.tar.gz) | ARM, X86, GPU(OPENCL,METAL), HuaweiKirinNPU, RockchipNPU, MediatekAPU, KunlunxinXPU, HuaweiAscendNPU |
+| CV | 分类 | [MobileNetV1](https://paddlelite-demo.bj.bcebos.com/models/mobilenet_v1_fp32_224_fluid.tar.gz) | ARM, X86, GPU(OPENCL,METAL), HuaweiKirinNPU, RockchipNPU, MediatekAPU, KunlunxinXPU, HuaweiAscendNPU, VerisiliconTIMVX |
| CV | 分类 | [MobileNetV2](https://paddlelite-demo.bj.bcebos.com/models/mobilenet_v2_fp32_224_fluid.tar.gz) | ARM, X86, GPU(OPENCL,METAL), HuaweiKirinNPU, KunlunxinXPU, HuaweiAscendNPU |
| CV | 分类 | [MobileNetV3_large](https://paddle-inference-dist.bj.bcebos.com/AI-Rank/mobile/MobileNetV3_large_x1_0.tar.gz) | ARM, X86, GPU(OPENCL,METAL), HuaweiAscendNPU |
| CV | 分类 | [MobileNetV3_small](https://paddle-inference-dist.bj.bcebos.com/AI-Rank/mobile/MobileNetV3_small_x1_0.tar.gz) | ARM, X86, GPU(OPENCL,METAL), HuaweiAscendNPU |
@@ -19,8 +19,8 @@
| CV | 分类 | HRNet_W18_C | ARM, X86 |
| CV | 分类 | RegNetX_4GF | ARM, X86 |
| CV | 分类 | Xception41 | ARM, X86 |
-| CV | 分类 | [ResNet18](https://paddlelite-demo.bj.bcebos.com/models/resnet18_fp32_224_fluid.tar.gz) | ARM, X86, GPU(OPENCL,METAL), HuaweiKirinNPU, RockchipNPU, KunlunxinXPU, HuaweiAscendNPU |
-| CV | 分类 | [ResNet50](https://paddlelite-demo.bj.bcebos.com/models/resnet50_fp32_224_fluid.tar.gz) | ARM, X86, GPU(OPENCL,METAL), HuaweiKirinNPU, RockchipNPU, KunlunxinXPU, HuaweiAscendNPU |
+| CV | 分类 | [ResNet18](https://paddlelite-demo.bj.bcebos.com/models/resnet18_fp32_224_fluid.tar.gz) | ARM, X86, GPU(OPENCL,METAL), HuaweiKirinNPU, RockchipNPU, KunlunxinXPU, HuaweiAscendNPU, VerisiliconTIMVX |
+| CV | 分类 | [ResNet50](https://paddlelite-demo.bj.bcebos.com/models/resnet50_fp32_224_fluid.tar.gz) | ARM, X86, GPU(OPENCL,METAL), HuaweiKirinNPU, RockchipNPU, KunlunxinXPU, HuaweiAscendNPU, VerisiliconTIMVX |
| CV | 分类 | [ResNet101](https://paddlelite-demo.bj.bcebos.com/NNAdapter/models/PaddleClas/ResNet101.tgz) | ARM, X86, HuaweiKirinNPU, RockchipNPU, KunlunxinXPU, HuaweiAscendNPU |
| CV | 分类 | [ResNeXt50](https://paddlelite-demo.bj.bcebos.com/NNAdapter/models/PaddleClas/ResNeXt50_32x4d.tgz) | ARM, X86, HuaweiAscendNPU |
| CV | 分类 | [MnasNet](https://paddlelite-demo.bj.bcebos.com/models/mnasnet_fp32_224_fluid.tar.gz)| ARM, HuaweiKirinNPU, HuaweiAscendNPU |
@@ -32,7 +32,7 @@
| CV | 分类 | VGG16 | ARM, X86, GPU(OPENCL), KunlunxinXPU, HuaweiAscendNPU |
| CV | 分类 | VGG19 | ARM, X86, GPU(OPENCL,METAL), KunlunxinXPU, HuaweiAscendNPU|
| CV | 分类 | GoogleNet | ARM, X86, KunlunxinXPU, HuaweiAscendNPU |
-| CV | 检测 | [SSD-MobileNetV1](https://paddlelite-demo.bj.bcebos.com/models/ssd_mobilenet_v1_pascalvoc_fp32_300_fluid.tar.gz) | ARM, HuaweiKirinNPU*, HuaweiAscendNPU* |
+| CV | 检测 | [SSD-MobileNetV1](https://paddlelite-demo.bj.bcebos.com/models/ssd_mobilenet_v1_pascalvoc_fp32_300_fluid.tar.gz) | ARM, HuaweiKirinNPU*, HuaweiAscendNPU*, VerisiliconTIMVX* |
| CV | 检测 | [SSD-MobileNetV3-large](https://paddle-inference-dist.bj.bcebos.com/AI-Rank/mobile/ssdlite_mobilenet_v3_large.tar.gz) | ARM, X86, GPU(OPENCL,METAL) |
| CV | 检测 | [SSD-VGG16](https://paddlelite-demo.bj.bcebos.com/NNAdapter/models/PaddleDetection/ssd_vgg16_300_240e_voc.tgz) | ARM, X86, HuaweiAscendNPU* |
| CV | 检测 | [YOLOv3-DarkNet53](https://paddlelite-demo.bj.bcebos.com/NNAdapter/models/PaddleDetection/yolov3_darknet53_270e_coco.tgz) | ARM, X86, HuaweiAscendNPU* |
diff --git a/docs/quick_start/support_operation_list.md b/docs/quick_start/support_operation_list.md
index a85d9b708ff..2c98239f2a6 100644
--- a/docs/quick_start/support_operation_list.md
+++ b/docs/quick_start/support_operation_list.md
@@ -10,99 +10,99 @@ Host 端 Kernel 是算子在任意 CPU 上纯 C/C++ 的具体实现，具有可
以 ARM CPU 为例，如果模型中某个算子没有 ARM 端 Kernel，但是有 Host 端 Kernel，那么模型优化阶段该算子会选择 Host 端 Kernel，该模型还是可以顺利部署。
-| OP_name| ARM | OpenCL | Metal | 昆仑芯XPU | Host | X86 | 比特大陆 | 英特尔FPGA | 寒武纪mlu | 华为昇腾NPU | 联发科APU | 瑞芯微NPU | 华为麒麟NPU | 颖脉NNA | 晶晨NPU |
-|-:|-| -| -| -| -| -| -| -| -| -| -| -| -| -| -|
-| affine_channel|Y| | | | | | | | | | | | | | |
-| affine_grid|Y| | | | | | | | | | | | | | |
-| arg_max|Y|Y| |Y|Y| | | |Y|Y| | | | | |
-| assign_value| | | |Y|Y| |Y| | |Y| | | | | |
-| batch_norm|Y|Y|Y|Y| |Y|Y| |Y|Y| | | | | |
-| bilinear_interp|Y|Y|Y|Y| |Y|Y| | |Y| | | | | |
-| bilinear_interp_v2|Y|Y|Y|Y| |Y| | | |Y| | | | | |
-| box_coder|Y|Y|Y|Y|Y|Y|Y| | | | | | | | |
-| calib|Y| | |Y| | | | |Y| | | | | | |
-| cast| | | |Y|Y| |Y| |Y|Y| | | | | |
-| concat|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|
-| conv2d|Y|Y|Y|Y| |Y|Y|Y|Y|Y|Y|Y|Y|Y|Y|
-| conv2d_transpose|Y|Y|Y|Y| |Y|Y| | |Y| | | | |Y|
-| density_prior_box| | | |Y|Y|Y|Y| | | | | | | | |
-| depthwise_conv2d|Y|Y|Y|Y| |Y|Y|Y|Y|Y|Y|Y|Y|Y|Y|
-| depthwise_conv2d_transpose| |Y| | | | |Y| | | | | | | | |
-| dropout|Y|Y|Y|Y| |Y|Y| |Y|Y| | | | | |
-| elementwise_add|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|
-| elementwise_div|Y| | |Y| |Y|Y| | |Y|Y|Y|Y| |Y|
-| elementwise_floordiv|Y| | | | |Y| | | | | | | | | |
-| elementwise_max|Y| | |Y| |Y| | | |Y| | | | | |
-| elementwise_min|Y| | | | |Y| | | |Y| | | | | |
-| elementwise_mod|Y| | | | |Y| | | | | | | | | |
-| elementwise_mul|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|
-| elementwise_pow|Y| | | | |Y| | | |Y| | | | | |
-| elementwise_sub|Y|Y|Y|Y| |Y|Y| | |Y|Y|Y|Y| |Y|
-| elu|Y| | | |Y| | | | | | | | | | |
-| erf|Y| | | | | | | | | | | | | | |
-| expand| |Y| | |Y| | | | | | | | | | |
-| expand_as| | | | |Y| | | | | | | | | | |
-| fc|Y|Y|Y| | |Y| | |Y|Y|Y|Y|Y|Y|Y|
-| feed| | |Y| |Y| | | | | | | | | | |
-| fetch| | |Y| |Y| | | | | | | | | | |
-| fill_constant| | | |Y|Y| |Y| | |Y| | | | | |
-| fill_constant_batch_size_like| | | |Y|Y| | | | | | | | | | |
-| flatten| |Y|Y|Y|Y| |Y| |Y|Y|Y|Y|Y| |Y|
-| flatten2| |Y|Y|Y|Y| |Y| |Y|Y|Y|Y|Y| |Y|
-| flatten_contiguous_range| | | |Y|Y| | | | |Y|Y|Y|Y| |Y|
-|fusion_elementwise_add_activation|Y|Y|Y| | |Y| | | |Y|Y|Y|Y| |Y|
-|fusion_elementwise_div_activation|Y| | | | |Y| | | |Y|Y|Y|Y| |Y|
-|fusion_elementwise_max_activation|Y| | | | |Y| | | |Y| | | | | |
-|fusion_elementwise_min_activation|Y| | | | |Y| | | |Y| | | | | |
-|fusion_elementwise_mul_activation|Y| | | | |Y| | | |Y|Y|Y|Y| |Y|
-|fusion_elementwise_pow_activation| | | | | | | | | |Y| | | | | |
-|fusion_elementwise_sub_activation|Y|Y| | | |Y| | | |Y|Y|Y|Y| |Y|
-| grid_sampler|Y|Y| |Y| |Y| | | | | | | | | |
-| instance_norm|Y|Y| |Y| |Y| | | |Y| | | | | |
-| io_copy| |Y|Y|Y| | | | |Y| | | | | | |
-| io_copy_once| |Y|Y|Y| | | | | | | | | | | |
-| layout|Y|Y| | | |Y| | |Y| | | | | | |
-| layout_once|Y|Y| | | | | | | | | | | | | |
-| leaky_relu|Y|Y|Y|Y|Y|Y|Y| |Y|Y| | | | | |
-| lod_array_length| | | | |Y| | | | | | | | | | |
-| matmul|Y|Y|Y|Y| |Y|Y| | |Y| | | | | |
-| mul|Y| | |Y| |Y|Y| | | | | | | | |
-| multiclass_nms|Y| | | |Y| |Y| | | | | | | | |
-| multiclass_nms2|Y| | | |Y| |Y| | | | | | | | |
-| multiclass_nms3|Y| | | |Y| | | | | | | | | | |
-| nearest_interp|Y|Y|Y|Y| |Y|Y| |Y|Y| | | | | |
-| nearest_interp_v2|Y|Y|Y|Y| |Y| | | |Y| | | | | |
-| pad2d|Y|Y|Y|Y|Y| | | | |Y| | | | | |
-| pool2d|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y|Y|Y|
-| prelu|Y|Y| |Y|Y| | | | |Y| | | | | |
-| prior_box|Y| | |Y|Y| |Y| | | | | | | | |
-| range| | | | |Y| | | | |Y| | | | | |
-| reduce_mean|Y|Y| |Y| |Y|Y| | |Y| | | | | |
-| relu|Y|Y|Y|Y|Y|Y|Y| |Y|Y|Y|Y|Y|Y|Y|
-| relu6|Y|Y|Y|Y|Y|Y| | |Y|Y|Y|Y|Y|Y|Y|
-| reshape| |Y| |Y|Y| |Y| |Y|Y|Y|Y|Y| |Y|
-| reshape2| |Y|Y|Y|Y| |Y| |Y|Y|Y|Y|Y| |Y|
-| scale|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|
-| search_fc| | | |Y| | | | | | | | | | | |
-| sequence_topk_avg_pooling| | | |Y| |Y| | | | | | | | | |
-| shuffle_channel|Y|Y|Y| |Y| | | | | | | | | | |
-| sigmoid|Y|Y|Y|Y|Y|Y|Y| |Y|Y|Y|Y|Y| |Y|
-| slice|Y|Y|Y|Y| |Y|Y| |Y|Y| | | | | |
-| softmax|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y|Y|Y|
-| softplus|Y| | | | | | | | | | | | | | |
-| squeeze| |Y| |Y|Y| |Y| |Y|Y| | | | | |
-| squeeze2| |Y| |Y|Y| |Y| |Y|Y| | | | | |
-| stack| | | |Y|Y|Y| | | |Y| | | | | |
-| subgraph| | | | | | | | |Y| | | | | | |
-| sync_batch_norm|Y|Y| | | |Y| | | | | | | | | |
-| tanh|Y|Y| |Y|Y|Y| | | |Y|Y|Y|Y| |Y|
-| thresholded_relu|Y| | | |Y| | | | | | | | | | |
-| transpose|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|
-| transpose2|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|
-| unsqueeze| |Y| |Y|Y| | | | |Y| | | | | |
-| unsqueeze2| |Y| |Y|Y| | | | |Y| | | | | |
-| write_back| | | | |Y| | | | | | | | | | |
-| yolo_box|Y|Y|Y|Y|Y| |Y| | | | | | | | |
+| OP_name| ARM | OpenCL | Metal | 昆仑芯XPU | Host | X86 | 比特大陆 | 英特尔FPGA | 寒武纪mlu | 华为昇腾NPU | 联发科APU | 瑞芯微NPU | 华为麒麟NPU | 颖脉NNA | 晶晨NPU | 芯原TIM-VX |
+|-:|-| -| -| -| -| -| -| -| -| -| -| -| -| -| -| -|
+| affine_channel|Y| | | | | | | | | | | | | | | |
+| affine_grid|Y| | | | | | | | | | | | | | | |
+| arg_max|Y|Y| |Y|Y| | | |Y|Y| | | | | | |
+| assign_value| | | |Y|Y| |Y| | |Y| | | | | | |
+| batch_norm|Y|Y|Y|Y| |Y|Y| |Y|Y| | | | | |Y|
+| bilinear_interp|Y|Y|Y|Y| |Y|Y| | |Y| | | | | |Y|
+| bilinear_interp_v2|Y|Y|Y|Y| |Y| | | |Y| | | | | |Y|
+| box_coder|Y|Y|Y|Y|Y|Y|Y| | | | | | | | | |
+| calib|Y| | |Y| | | | |Y| | | | | | | |
+| cast| | | |Y|Y| |Y| |Y|Y| | | | | | |
+| concat|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|Y|
+| conv2d|Y|Y|Y|Y| |Y|Y|Y|Y|Y|Y|Y|Y|Y|Y|Y|
+| conv2d_transpose|Y|Y|Y|Y| |Y|Y| | |Y| | | | |Y|Y|
+| density_prior_box| | | |Y|Y|Y|Y| | | | | | | | | |
+| depthwise_conv2d|Y|Y|Y|Y| |Y|Y|Y|Y|Y|Y|Y|Y|Y|Y|Y|
+| depthwise_conv2d_transpose| |Y| | | | |Y| | | | | | | | | |
+| dropout|Y|Y|Y|Y| |Y|Y| |Y|Y| | | | | |Y|
+| elementwise_add|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|Y|
+| elementwise_div|Y| | |Y| |Y|Y| | |Y|Y|Y|Y| |Y|Y|
+| elementwise_floordiv|Y| | | | |Y| | | | | | | | | | |
+| elementwise_max|Y| | |Y| |Y| | | |Y| | | | | |Y|
+| elementwise_min|Y| | | | |Y| | | |Y| | | | | |Y|
+| elementwise_mod|Y| | | | |Y| | | | | | | | | | |
+| elementwise_mul|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|Y|
+| elementwise_pow|Y| | | | |Y| | | |Y| | | | | |Y|
+| elementwise_sub|Y|Y|Y|Y| |Y|Y| | |Y|Y|Y|Y| |Y|Y|
+| elu|Y| | | |Y| | | | | | | | | | | |
+| erf|Y| | | | | | | | | | | | | | | |
+| expand| |Y| | |Y| | | | | | | | | | | |
+| expand_as| | | | |Y| | | | | | | | | | | |
+| fc|Y|Y|Y| | |Y| | |Y|Y|Y|Y|Y|Y|Y|Y|
+| feed| | |Y| |Y| | | | | | | | | | | |
+| fetch| | |Y| |Y| | | | | | | | | | | |
+| fill_constant| | | |Y|Y| |Y| | |Y| | | | | | |
+| fill_constant_batch_size_like| | | |Y|Y| | | | | | | | | | |Y|
+| flatten| |Y|Y|Y|Y| |Y| |Y|Y|Y|Y|Y| |Y|Y|
+| flatten2| |Y|Y|Y|Y| |Y| |Y|Y|Y|Y|Y| |Y|Y|
+| flatten_contiguous_range| | | |Y|Y| | | | |Y|Y|Y|Y| |Y|Y|
+|fusion_elementwise_add_activation|Y|Y|Y| | |Y| | | |Y|Y|Y|Y| |Y| |
+|fusion_elementwise_div_activation|Y| | | | |Y| | | |Y|Y|Y|Y| |Y| |
+|fusion_elementwise_max_activation|Y| | | | |Y| | | |Y| | | | | | |
+|fusion_elementwise_min_activation|Y| | | | |Y| | | |Y| | | | | | |
+|fusion_elementwise_mul_activation|Y| | | | |Y| | | |Y|Y|Y|Y| |Y| |
+|fusion_elementwise_pow_activation| | | | | | | | | |Y| | | | | | |
+|fusion_elementwise_sub_activation|Y|Y| | | |Y| | | |Y|Y|Y|Y| |Y| |
+| grid_sampler|Y|Y| |Y| |Y| | | | | | | | | | |
+| instance_norm|Y|Y| |Y| |Y| | | |Y| | | | | | |
+| io_copy| |Y|Y|Y| | | | |Y| | | | | | | |
+| io_copy_once| |Y|Y|Y| | | | | | | | | | | | |
+| layout|Y|Y| | | |Y| | |Y| | | | | | | |
+| layout_once|Y|Y| | | | | | | | | | | | | | |
+| leaky_relu|Y|Y|Y|Y|Y|Y|Y| |Y|Y| | | | | |Y|
+| lod_array_length| | | | |Y| | | | | | | | | | | |
+| matmul|Y|Y|Y|Y| |Y|Y| | |Y| | | | | |Y|
+| mul|Y| | |Y| |Y|Y| | | | | | | | | |
+| multiclass_nms|Y| | | |Y| |Y| | | | | | | | | |
+| multiclass_nms2|Y| | | |Y| |Y| | | | | | | | | |
+| multiclass_nms3|Y| | | |Y| | | | | | | | | | | |
+| nearest_interp|Y|Y|Y|Y| |Y|Y| |Y|Y| | | | | |Y|
+| nearest_interp_v2|Y|Y|Y|Y| |Y| | | |Y| | | | | |Y|
+| pad2d|Y|Y|Y|Y|Y| | | | |Y| | | | | | |
+| pool2d|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y|Y|Y|Y|
+| prelu|Y|Y| |Y|Y| | | | |Y| | | | | | |
+| prior_box|Y| | |Y|Y| |Y| | | | | | | | | |
+| range| | | | |Y| | | | |Y| | | | | | |
+| reduce_mean|Y|Y| |Y| |Y|Y| | |Y| | | | | | |
+| relu|Y|Y|Y|Y|Y|Y|Y| |Y|Y|Y|Y|Y|Y|Y|Y|
+| relu6|Y|Y|Y|Y|Y|Y| | |Y|Y|Y|Y|Y|Y|Y|Y|
+| reshape| |Y| |Y|Y| |Y| |Y|Y|Y|Y|Y| |Y|Y|
+| reshape2| |Y|Y|Y|Y| |Y| |Y|Y|Y|Y|Y| |Y|Y|
+| scale|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|Y|
+| search_fc| | | |Y| | | | | | | | | | | | |
+| sequence_topk_avg_pooling| | | |Y| |Y| | | | | | | | | | |
+| shuffle_channel|Y|Y|Y| |Y| | | | | | | | | | |Y|
+| sigmoid|Y|Y|Y|Y|Y|Y|Y| |Y|Y|Y|Y|Y| |Y|Y|
+| slice|Y|Y|Y|Y| |Y|Y| |Y|Y| | | | | |Y|
+| softmax|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y|Y|Y|Y|
+| softplus|Y| | | | | | | | | | | | | | | |
+| squeeze| |Y| |Y|Y| |Y| |Y|Y| | | | | |Y|
+| squeeze2| |Y| |Y|Y| |Y| |Y|Y| | | | | |Y|
+| stack| | | |Y|Y|Y| | | |Y| | | | | | |
+| subgraph| | | | | | | | |Y| | | | | | | |
+| sync_batch_norm|Y|Y| | | |Y| | | | | | | | | | |
+| tanh|Y|Y| |Y|Y|Y| | | |Y|Y|Y|Y| |Y| |
+| thresholded_relu|Y| | | |Y| | | | | | | | | | | |
+| transpose|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|Y|
+| transpose2|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|Y|
+| unsqueeze| |Y| |Y|Y| | | | |Y| | | | | |Y|
+| unsqueeze2| |Y| |Y|Y| | | | |Y| | | | | |Y|
+| write_back| | | | |Y| | | | | | | | | | | |
+| yolo_box|Y|Y|Y|Y|Y| |Y| | | | | | | | | |
### 附加算子
@@ -110,273 +110,273 @@ Host 端 Kernel 是算子在任意 CPU 上纯 C/C++ 的具体实现，具有可
加上附加算子共计 269 个，需要在编译时打开 `--with_extra=ON` 开关才会编译，具体请参考[参数详情](../source_compile/compile_options)。
-| OP_name| ARM | OpenCL | Metal | 昆仑芯XPU | Host | X86 | 比特大陆 | 英特尔FPGA | 寒武纪mlu | 华为昇腾NPU | 联发科APU | 瑞芯微NPU | 华为麒麟NPU | 颖脉NNA | 晶晨NPU |
-|-:|-| -| -| -| -| -| -| -| -| -| -| -| -| -| -|
-| abs|Y|Y| |Y|Y| | | | |Y| | | | | |
-| affine_channel|Y| | | | | | | | | | | | | | |
-| affine_grid|Y| | | | | | | | | | | | | | |
-| anchor_generator| | | |Y|Y| | | | | | | | | | |
-| arg_max|Y|Y| |Y|Y| | | |Y|Y| | | | | |
-| arg_min| | | | | | | | | |Y| | | | | |
-| argsort| | | | |Y| | | | | | | | | | |
-| assign| | | |Y|Y| | | | |Y| | | | | |
-| assign_value| | | |Y|Y| |Y| | |Y| | | | | |
-| attention_padding_mask| | | | | | | | | | | | | | | |
-| axpy|Y| | | | | | | | | | | | | | |
-| batch_norm|Y|Y|Y|Y| |Y|Y| |Y|Y| | | | | |
-| beam_search| | | | |Y| | | | | | | | | | |
-| beam_search_decode| | | | |Y| | | | | | | | | | |
-| bilinear_interp|Y|Y|Y|Y| |Y|Y| | |Y| | | | | |
-| bilinear_interp_v2|Y|Y|Y|Y| |Y| | | |Y| | | | | |
-| bmm| | | |Y| | | | | | | | | | | |
-| box_clip| | | |Y|Y| | | | | | | | | | |
-| box_coder|Y|Y|Y|Y|Y|Y|Y| | | | | | | | |
-| calib|Y| | |Y| | | | |Y| | | | | | |
-| calib_once|Y| | |Y| | | | |Y| | | | | | |
-| cast| | | |Y|Y| |Y| |Y|Y| | | | | |
-| clip|Y|Y| |Y| |Y| | | |Y| | | | | |
-| collect_fpn_proposals| | | | |Y| | | | | | | | | | |
-| concat|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|
-| conditional_block| | | | |Y| | | | | | | | | | |
-| conv2d|Y|Y|Y|Y| |Y|Y|Y|Y|Y|Y|Y|Y|Y|Y|
-| conv2d_transpose|Y|Y|Y|Y| |Y|Y| | |Y| | | | |Y|
-| correlation| | | |Y|Y| | | | | | | | | | |
-| cos| |Y| | |Y| | | | | | | | | | |
-| cos_sim| | | | |Y| | | | | | | | | | |
-| crf_decoding| | | | |Y| | | | | | | | | | |
-| crop| | | | |Y| | | | | | | | | | |
-| crop_tensor| | | | |Y| | | | | | | | | | |
-| ctc_align| | | | |Y| | | | | | | | | | |
-| cumsum| | | | |Y| | | | |Y| | | | | |
-| decode_bboxes|Y| | | | | | | | | | | | | | |
-| deformable_conv|Y| | | |Y| | | | |Y| | | | | |
-| density_prior_box| | | |Y|Y|Y|Y| | | | | | | | |
-| depthwise_conv2d|Y|Y|Y|Y| |Y|Y|Y|Y|Y|Y|Y|Y|Y|Y|
-| depthwise_conv2d_transpose| |Y| | | | |Y| | | | | | | | |
-| distribute_fpn_proposals| | | | |Y| | | | | | | | | | |
-| dropout|Y|Y|Y|Y| |Y|Y| |Y|Y| | | | | |
-| elementwise_add|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|
-| elementwise_div|Y| | |Y| |Y|Y| | |Y|Y|Y|Y| |Y|
-| elementwise_floordiv|Y| | | | |Y| | | | | | | | | |
-| elementwise_max|Y| | |Y| |Y| | | |Y| | | | | |
-| elementwise_min|Y| | | | |Y| | | |Y| | | | | |
-| elementwise_mod|Y| | | | |Y| | | | | | | | | |
-| elementwise_mul|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|
-| elementwise_pow|Y| | | | |Y| | | |Y| | | | | |
-| elementwise_sub|Y|Y|Y|Y| |Y|Y| | |Y|Y|Y|Y| |Y|
-| elu|Y| | | |Y| | | | | | | | | | |
-| equal| | | | |Y| | | | |Y| | | | | |
-| erf|Y| | | | | | | | | | | | | | |
-| exp|Y|Y|Y|Y|Y| | | | |Y| | | | | |
-| expand| |Y| | |Y| | | | | | | | | | |
-| expand_as| | | | |Y| | | | | | | | | | |
-| expand_v2| | | |Y|Y| | | | |Y| | | | | |
-| fake_channel_wise_dequantize_max_abs| | | | | | | | | | | | | | | |
-| fake_channel_wise_quantize_dequantize_abs_max| | | | | | | | | | | | | | | |
-| fake_dequantize_max_abs| | | | | | | | | | | | | | | |
-| fake_quantize_abs_max| | | | | | | | | | | | | | | |
-| fake_quantize_dequantize_abs_max| | | | | | | | | | | | | | | |
-|fake_quantize_dequantize_moving_average_abs_max| | | | | | | | | | | | | | | |
-| fake_quantize_moving_average_abs_max| | | | | | | | | | | | | | | |
-| fake_quantize_range_abs_max| | | | | | | | | | | | | | | |
-| fc|Y|Y|Y| | |Y| | |Y|Y|Y|Y|Y|Y|Y|
-| feed| | |Y| |Y| | | | | | | | | | |
-| fetch| | |Y| |Y| | | | | | | | | | |
-| fill_any_like| | | |Y|Y| | | | |Y| | | | | |
-| fill_constant| | | |Y|Y| |Y| | |Y| | | | | |
-| fill_constant_batch_size_like| | | |Y|Y| | | | | | | | | | |
-| fill_zeros_like| | | |Y|Y| | | | | | | | | | |
-| flatten| |Y|Y|Y|Y| |Y| |Y|Y|Y|Y|Y| |Y|
-| flatten2| |Y|Y|Y|Y| |Y| |Y|Y|Y|Y|Y| |Y|
-| flatten_contiguous_range| | | |Y|Y| | | | |Y|Y|Y|Y| |Y|
-| flip| | | | |Y| | | | | | | | | | |
-| floor|Y| | | |Y| | | | | | | | | | |
-| fusion_elementwise_add_activation|Y|Y|Y| | |Y| | | |Y|Y|Y|Y| |Y|
-| fusion_elementwise_div_activation|Y| | | | |Y| | | |Y|Y|Y|Y| |Y|
-| fusion_elementwise_max_activation|Y| | | | |Y| | | |Y| | | | | |
-| fusion_elementwise_min_activation|Y| | | | |Y| | | |Y| | | | | |
-| fusion_elementwise_mul_activation|Y| | | | |Y| | | |Y|Y|Y|Y| |Y|
-| fusion_elementwise_pow_activation| | | | | | | | | |Y| | | | | |
-| fusion_elementwise_sub_activation|Y|Y| | | |Y| | | |Y|Y|Y|Y| |Y|
-| gather|Y| | |Y|Y|Y| | |Y| | | | | | |
-| gather_nd| | | | |Y| | | | | | | | | | |
-| gather_tree| | | | |Y| | | | | | | | | | |
-| gelu|Y| | |Y| |Y| | | |Y| | | | | |
-| generate_proposals| | | |Y|Y| | | | | | | | | | |
-| generate_proposals_v2|Y| | | | | | | | | | | | | | |
-| greater_equal| | | | |Y| | | | |Y| | | | | |
-| greater_than| |Y| | |Y| | | | |Y| | | | | |
-| grid_sampler|Y|Y| |Y| |Y| | | | | | | | | |
-| group_norm|Y| | | | |Y| | | | | | | | | |
-| gru|Y| | |Y| |Y| | | | | | | | | |
-| gru_unit|Y| | |Y| |Y| | | | | | | | | |
-| hard_sigmoid|Y|Y|Y|Y|Y| |Y| | |Y| | | | | |
-| hard_swish|Y|Y|Y|Y|Y|Y|Y| | |Y| | | | | |
-| im2sequence|Y| | |Y| | |Y| | | | | | | | |
-| increment| | | |Y|Y| | | | | | | | | | |
-| index_select| | | | |Y| | | | | | | | | | |
-| instance_norm|Y|Y| |Y| |Y| | | |Y| | | | | |
-| inverse| | | | |Y| | | | | | | | | | |
-| io_copy| |Y|Y|Y| | | | |Y| | | | | | |
-| io_copy_once| |Y|Y|Y| | | | | | | | | | | |
-| is_empty| | | |Y|Y| | | | | | | | | | |
-| layer_norm|Y| | |Y| |Y| | | |Y| | | | | |
-| layout|Y|Y| | | |Y| | |Y| | | | | | |
-| layout_once|Y|Y| | | | | | | | | | | | | |
-| leaky_relu|Y|Y|Y|Y|Y|Y|Y| |Y|Y| | | | | |
-| less_equal| | | | |Y| | | | |Y| | | | | |
-| less_than| | | |Y|Y| | | | |Y| | | | | |
-| linspace| | | | |Y| | | | | | | | | | |
-| lod_array_length| | | | |Y| | | | | | | | | | |
-| lod_reset| | | | |Y| | | | | | | | | | |
-| log|Y| | |Y|Y| | | | |Y| | | | | |
-| logical_and| | | |Y|Y| | | | | | | | | | |
-| logical_not| | | |Y|Y| | | | | | | | | | |
-| logical_or| | | | |Y| | | | | | | | | | |
-| logical_xor| | | | |Y| | | | | | | | | | |
-| lookup_table|Y| | |Y| |Y| | | | | | | | | |
-| lookup_table_dequant|Y| | | | | | | | | | | | | | |
-| lookup_table_v2|Y| | |Y| |Y| | | |Y| | | | | |
-| lrn|Y|Y| |Y| | | | |Y| | | | | | |
-| lstm|Y| | | | | | | | | | | | | | |
-| match_matrix_tensor| | | |Y| |Y| | | | | | | | | |
-| matmul|Y|Y|Y|Y| |Y|Y| | |Y| | | | | |
-| matmul_v2|Y|Y| |Y| | | | | |Y| | | | | |
-| matrix_nms| | | | |Y| | | | | | | | | | |
-| max_pool2d_with_index| | | |Y| | |Y| | | | | | | | |
-| mean|Y| | | | | | | | | | | | | | |
-| merge_lod_tensor|Y| | | | | | | | | | | | | | |
-| meshgrid| | | | |Y| | | | | | | | | | |
-| mish|Y| | | | |Y| | | | | | | | | |
-| mul|Y| | |Y| |Y|Y| | | | | | | | |
-| multiclass_nms|Y| | | |Y| |Y| | | | | | | | |
-| multiclass_nms2|Y| | | |Y| |Y| | | | | | | | |
-| multiclass_nms3|Y| | | |Y| | | | | | | | | | |
-| nearest_interp|Y|Y|Y|Y| |Y|Y| |Y|Y| | | | | |
-| nearest_interp_v2|Y|Y|Y|Y| |Y| | | |Y| | | | | |
-| negative|Y| | | | | | | | | | | | | | |
-| norm|Y| | |Y|Y| |Y| |Y|Y| | | | | |
-| not_equal| | | | |Y| | | | |Y| | | | | |
-| one_hot| | | | |Y| | | | | | | | | | |
-| one_hot_v2| | | | |Y| | | | | | | | | | |
-| p_norm|Y| | | |Y| | | | |Y| | | | | |
-| pad2d|Y|Y|Y|Y|Y| | | | |Y| | | | | |
-| pad3d| | | | |Y| | | | |Y| | | | | |
-| pixel_shuffle|Y|Y| |Y|Y| | | | | | | | | | |
-| polygon_box_transform| | | | |Y| | | | | | | | | | |
-| pool2d|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y|Y|Y|
-| pow|Y| | |Y| |Y| | | |Y| | | | | |
-| prelu|Y|Y| |Y|Y| | | | |Y| | | | | |
-| print| | | | |Y| | | | | | | | | | |
-| prior_box|Y| | |Y|Y| |Y| | | | | | | | |
-| range| | | | |Y| | | | |Y| | | | | |
-| read_from_array| | | |Y|Y| | | | | | | | | | |
-| reciprocal|Y| | |Y|Y| | | | | | | | | | |
-| reduce_all| | | |Y|Y| | | | | | | | | | |
-| reduce_any| | | |Y|Y| | | | | | | | | | |
-| reduce_max|Y|Y| |Y| |Y|Y| | | | | | | | |
-| reduce_mean|Y|Y| |Y| |Y|Y| | |Y| | | | | |
-| reduce_min|Y| | |Y| |Y| | | | | | | | | |
-| reduce_prod|Y| | |Y| |Y| | | | | | | | | |
-| reduce_sum|Y| | |Y| |Y|Y| | | | | | | | |
-| relu|Y|Y|Y|Y|Y|Y|Y| |Y|Y|Y|Y|Y|Y|Y|
-| relu6|Y|Y|Y|Y|Y|Y| | |Y|Y|Y|Y|Y|Y|Y|
-| relu_clipped|Y| | | |Y| | | | | | | | | | |
-| reshape| |Y| |Y|Y| |Y| |Y|Y|Y|Y|Y| |Y|
-| reshape2| |Y|Y|Y|Y| |Y| |Y|Y|Y|Y|Y| |Y|
-| retinanet_detection_output| | | | |Y| | | | | | | | | | |
-| reverse| | | | |Y| | | | | | | | | | |
-| rnn|Y| | |Y| |Y| | | | | | | | | |
-| roi_align| | | |Y|Y| | | | | | | | | | |
-| roi_perspective_transform| | | | |Y| | | | | | | | | | |
-| rsqrt|Y|Y| |Y|Y|Y| | | | | | | | | |
-| scale|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|
-| scatter|Y| | | | | | | | | | | | | | |
-| scatter_nd_add| | | | |Y| | | | | | | | | | |
-| search_aligned_mat_mul| | | | | |Y| | | | | | | | | |
-| search_attention_padding_mask| | | | | |Y| | | | | | | | | |
-| search_fc| | | |Y| | | | | | | | | | | |
-| search_grnn| | | |Y| |Y| | | | | | | | | |
-| search_group_padding| | | | | |Y| | | | | | | | | |
-| search_seq_arithmetic| | | |Y| |Y| | | | | | | | | |
-| search_seq_depadding| | | | | |Y| | | | | | | | | |
-| search_seq_fc| | | | | |Y| | | | | | | | | |
-| search_seq_softmax| | | | | |Y| | | | | | | | | |
-| select_input| | | | |Y| | | | | | | | | | |
-| sequence_arithmetic| | | |Y| |Y| | | | | | | | | |
-| sequence_concat| | | |Y| |Y| | | | | | | | | |
-| sequence_conv|Y| | | | |Y| | | | | | | | | |
-| sequence_expand| | | | |Y| | | | | | | | | | |
-| sequence_expand_as|Y| | | | |Y| | | | | | | | | |
-| sequence_mask| | | |Y|Y| | | | | | | | | | |
-| sequence_pad| | | |Y|Y| | | | | | | | | | |
-| sequence_pool|Y| | |Y| |Y| | | | | | | | | |
-| sequence_reshape| | | | | |Y| | | | | | | | | |
-| sequence_reverse| | | |Y| |Y| | | | | | | | | |
-| sequence_softmax| | | | |Y| | | | | | | | | | |
-| sequence_topk_avg_pooling| | | |Y| |Y| | | | | | | | | |
-| sequence_unpad| | | |Y|Y| | | | | | | | | | |
-| shape| |Y| |Y|Y| |Y| | |Y| | | | | |
-| shuffle_channel|Y|Y|Y| |Y| | | | | | | | | | |
-| sigmoid|Y|Y|Y|Y|Y|Y|Y| |Y|Y|Y|Y|Y| |Y|
-| sign|Y| | |Y| | | | | | | | | | | |
-| sin| |Y| | |Y| | | | | | | | | | |
-| slice|Y|Y|Y|Y| |Y|Y| |Y|Y| | | | | |
-| softmax|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y|Y|Y|
-| softplus|Y| | | | | | | | | | | | | | |
-| softsign| | | |Y| |Y| | | | | | | | | |
-| sparse_conv2d|Y| | | | | | | | | | | | | | |
-| split| |Y|Y|Y|Y| |Y| |Y|Y| | |Y| | |
-| split_lod_tensor|Y| | | | | | | | | | | | | | |
-| sqrt|Y|Y| |Y| |Y|Y| | | | | | | | |
-| square|Y|Y| |Y|Y|Y|Y| | | | | | | | |
-| squeeze| |Y| |Y|Y| |Y| |Y|Y| | | | | |
-| squeeze2| |Y| |Y|Y| |Y| |Y|Y| | | | | |
-| stack| | | |Y|Y|Y| | | |Y| | | | | |
-| strided_slice| | | | |Y| | | | | | | | | | |
-| subgraph| | | | | | | | |Y| | | | | | |
-| sum|Y| | |Y| | | | | | | | | | | |
-| swish|Y|Y|Y|Y|Y| |Y| | |Y| | | | | |
-| sync_batch_norm|Y|Y| | | |Y| | | | | | | | | |
-| tanh|Y|Y| |Y|Y|Y| | | |Y|Y|Y|Y| |Y|
-| tensor_array_to_tensor| | | | |Y| | | | | | | | | | |
-| thresholded_relu|Y| | | |Y| | | | | | | | | | |
-| tile| | | | |Y| | | | | | | | | | |
-| top_k| | | |Y|Y| | | | |Y| | | | | |
-| top_k_v2| | | | |Y| | | | |Y| | | | | |
-| transpose|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|
-| transpose2|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|
-| tril_triu| | | | |Y| | | | | | | | | | |
-| uniform_random| | | | |Y| | | | | | | | | | |
-| unique_with_counts| | | | |Y| | | | | | | | | | |
-| unsqueeze| |Y| |Y|Y| | | | |Y| | | | | |
-| unsqueeze2| |Y| |Y|Y| | | | |Y| | | | | |
-| unstack| | | |Y|Y| | | | | | | | | | |
-| var_conv_2d| | | |Y| |Y| | | | | | | | | |
-| where| | | | |Y| | | | | | | | | | |
-| where_index| | | | |Y| | | | | | | | | | |
-| while| | | | |Y| | | | | | | | | | |
-| write_back| | | | |Y| | | | | | | | | | |
-| write_to_array| | | |Y|Y| | | | | | | | | | |
-| yolo_box|Y|Y|Y|Y|Y| |Y| | | | | | | | |
-| __xpu__bigru| | | |Y| | | | | | | | | | | |
-| __xpu__conv2d| | | |Y| | | | | | | | | | | |
-| __xpu__dynamic_lstm_fuse_op| | | |Y| | | | | | | | | | | |
-| __xpu__embedding_with_eltwise_add| | | |Y| | | | | | | | | | | |
-| __xpu__fc| | | |Y| | | | | | | | | | | |
-| __xpu__generate_sequence| | | |Y| | | | | | | | | | | |
-| __xpu__logit| | | |Y| | | | | | | | | | | |
-| __xpu__mmdnn_bid_emb_att| | | |Y| | | | | | | | | | | |
-| __xpu__mmdnn_bid_emb_grnn_att| | | |Y| | | | | | | | | | | |
-| __xpu__mmdnn_bid_emb_grnn_att2| | | |Y| | | | | | | | | | | |
-| __xpu__mmdnn_match_conv_topk| | | |Y| | | | | | | | | | | |
-| __xpu__mmdnn_merge_all| | | |Y| | | | | | | | | | | |
-| __xpu__mmdnn_search_attention| | | |Y| | | | | | | | | | | |
-| __xpu__mmdnn_search_attention2| | | |Y| | | | | | | | | | | |
-| __xpu__multi_encoder| | | |Y| | | | | | | | | | | |
-| __xpu__multi_softmax| | | |Y| | | | | | | | | | | |
-| __xpu__resnet50| | | |Y| | | | | | | | | | | |
-| __xpu__resnet_cbam| | | |Y| | | | | | | | | | | |
-| __xpu__sfa_head| | | |Y| | | | | | | | | | | |
-| __xpu__softmax_topk| | | |Y| | | | | | | | | | | |
-| __xpu__squeeze_excitation_block| | | |Y| | | | | | | | | | | |
+| OP_name| ARM | OpenCL | Metal | 昆仑芯XPU | Host | X86 | 比特大陆 | 英特尔FPGA | 寒武纪mlu | 华为昇腾NPU | 联发科APU | 瑞芯微NPU | 华为麒麟NPU | 颖脉NNA | 晶晨NPU | 芯原TIM-VX |
+|-:|-| -| -| -| -| -| -| -| -| -| -| -| -| -| -| -|
+| abs|Y|Y| |Y|Y| | | | |Y| | | | | | |
+| affine_channel|Y| | | | | | | | | | | | | | | |
+| affine_grid|Y| | | | | | | | | | | | | | | |
+| anchor_generator| | | |Y|Y| | | | | | | | | | | |
+| arg_max|Y|Y| |Y|Y| | | |Y|Y| | | | | | |
+| arg_min| | | | | | | | | |Y| | | | | | |
+| argsort| | | | |Y| | | | | | | | | | | |
+| assign| | | |Y|Y| | | | |Y| | | | | | |
+| assign_value| | | |Y|Y| |Y| | |Y| | | | | | |
+| attention_padding_mask| | | | | | | | | | | | | | | | |
+| axpy|Y| | | | | | | | | | | | | | | |
+| batch_norm|Y|Y|Y|Y| |Y|Y| |Y|Y| | | | | |Y|
+| beam_search| | | | |Y| | | | | | | | | | | |
+| beam_search_decode| | | | |Y| | | | | | | | | | | |
+| bilinear_interp|Y|Y|Y|Y| |Y|Y| | |Y| | | | | |Y|
+| bilinear_interp_v2|Y|Y|Y|Y| |Y| | | |Y| | | | | |Y|
+| bmm| | | |Y| | | | | | | | | | | | |
+| box_clip| | | |Y|Y| | | | | | | | | | | |
+| box_coder|Y|Y|Y|Y|Y|Y|Y| | | | | | | | | |
+| calib|Y| | |Y| | | | |Y| | | | | | | |
+| calib_once|Y| | |Y| | | | |Y| | | | | | | |
+| cast| | | |Y|Y| |Y| |Y|Y| | | | | | |
+| clip|Y|Y| |Y| |Y| | | |Y| | | | | | |
+| collect_fpn_proposals| | | | |Y| | | | | | | | | | | |
+| concat|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|Y|
+| conditional_block| | | | |Y| | | | | | | | | | | |
+| conv2d|Y|Y|Y|Y| |Y|Y|Y|Y|Y|Y|Y|Y|Y|Y|Y|
+| conv2d_transpose|Y|Y|Y|Y| |Y|Y| | |Y| | | | |Y|Y|
+| correlation| | | |Y|Y| | | | | | | | | | | |
+| cos| |Y| | |Y| | | | | | | | | | | |
+| cos_sim| | | | |Y| | | | | | | | | | | |
+| crf_decoding| | | | |Y| | | | | | | | | | | |
+| crop| | | | |Y| | | | | | | | | | | |
+| crop_tensor| | | | |Y| | | | | | | | | | | |
+| ctc_align| | | | |Y| | | | | | | | | | | |
+| cumsum| | | | |Y| | | | |Y| | | | | | |
+| decode_bboxes|Y| | | | | | | | | | | | | | | |
+| deformable_conv|Y| | | |Y| | | | |Y| | | | | | |
+| density_prior_box| | | |Y|Y|Y|Y| | | | | | | | | |
+| depthwise_conv2d|Y|Y|Y|Y| |Y|Y|Y|Y|Y|Y|Y|Y|Y|Y|Y|
+| depthwise_conv2d_transpose| |Y| | | | |Y| | | | | | | | | |
+| distribute_fpn_proposals| | | | |Y| | | | | | | | | | | |
+| dropout|Y|Y|Y|Y| |Y|Y| |Y|Y| | | | | |Y|
+| elementwise_add|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|Y|
+| elementwise_div|Y| | |Y| |Y|Y| | |Y|Y|Y|Y| |Y|Y|
+| elementwise_floordiv|Y| | | | |Y| | | | | | | | | | |
+| elementwise_max|Y| | |Y| |Y| | | |Y| | | | | |Y|
+| elementwise_min|Y| | | | |Y| | | |Y| | | | | |Y|
+| elementwise_mod|Y| | | | |Y| | | | | | | | | | |
+| elementwise_mul|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|Y|
+| elementwise_pow|Y| | | | |Y| | | |Y| | | | | |Y|
+| elementwise_sub|Y|Y|Y|Y| |Y|Y| | |Y|Y|Y|Y| |Y|Y|
+| elu|Y| | | |Y| | | | | | | | | | | |
+| equal| | | | |Y| | | | |Y| | | | | | |
+| erf|Y| | | | | | | | | | | | | | | |
+| exp|Y|Y|Y|Y|Y| | | | |Y| | | | | | |
+| expand| |Y| | |Y| | | | | | | | | | | |
+| expand_as| | | | |Y| | | | | | | | | | | |
+| expand_v2| | | |Y|Y| | | | |Y| | | | | | |
+| fake_channel_wise_dequantize_max_abs| | | | | | | | | | | | | | | | |
+| fake_channel_wise_quantize_dequantize_abs_max| | | | | | | | | | | | | | | | |
+| fake_dequantize_max_abs| | | | | | | | | | | | | | | | |
+| fake_quantize_abs_max| | | | | | | | | | | | | | | | |
+| fake_quantize_dequantize_abs_max| | | | | | | | | | | | | | | | |
+|fake_quantize_dequantize_moving_average_abs_max| | | | | | | | | | | | | | | | |
+| fake_quantize_moving_average_abs_max| | | | | | | | | | | | | | | | |
+| fake_quantize_range_abs_max| | | | | | | | | | | | | | | | |
+| fc|Y|Y|Y| | |Y| | |Y|Y|Y|Y|Y|Y|Y|Y|
+| feed| | |Y| |Y| | | | | | | | | | | |
+| fetch| | |Y| |Y| | | | | | | | | | | |
+| fill_any_like| | | |Y|Y| | | | |Y| | | | | |Y|
+| fill_constant| | | |Y|Y| |Y| | |Y| | | | | | |
+| fill_constant_batch_size_like| | | |Y|Y| | | | | | | | | | |Y|
+| fill_zeros_like| | | |Y|Y| | | | | | | | | | | |
+| flatten| |Y|Y|Y|Y| |Y| |Y|Y|Y|Y|Y| |Y|Y|
+| flatten2| |Y|Y|Y|Y| |Y| |Y|Y|Y|Y|Y| |Y|Y|
+| flatten_contiguous_range| | | |Y|Y| | | | |Y|Y|Y|Y| |Y|Y|
+| flip| | | | |Y| | | | | | | | | | | |
+| floor|Y| | | |Y| | | | | | | | | | | |
+| fusion_elementwise_add_activation|Y|Y|Y| | |Y| | | |Y|Y|Y|Y| |Y| |
+| fusion_elementwise_div_activation|Y| | | | |Y| | | |Y|Y|Y|Y| |Y| |
+| fusion_elementwise_max_activation|Y| | | | |Y| | | |Y| | | | | | |
+| fusion_elementwise_min_activation|Y| | | | |Y| | | |Y| | | | | | |
+| fusion_elementwise_mul_activation|Y| | | | |Y| | | |Y|Y|Y|Y| |Y| |
+| fusion_elementwise_pow_activation| | | | | | | | | |Y| | | | | | |
+| fusion_elementwise_sub_activation|Y|Y| | | |Y| | | |Y|Y|Y|Y| |Y| |
+| gather|Y| | |Y|Y|Y| | |Y| | | | | | | |
+| gather_nd| | | | |Y| | | | | | | | | | | |
+| gather_tree| | | | |Y| | | | | | | | | | | |
+| gelu|Y| | |Y| |Y| | | |Y| | | | | | |
+| generate_proposals| | | |Y|Y| | | | | | | | | | | |
+| generate_proposals_v2|Y| | | | | | | | | | | | | | | |
+| greater_equal| | | | |Y| | | | |Y| | | | | | |
+| greater_than| |Y| | |Y| | | | |Y| | | | | | |
+| grid_sampler|Y|Y| |Y| |Y| | | | | | | | | | |
+| group_norm|Y| | | | |Y| | | | | | | | | | |
+| gru|Y| | |Y| |Y| | | | | | | | | | |
+| gru_unit|Y| | |Y| |Y| | | | | | | | | | |
+| hard_sigmoid|Y|Y|Y|Y|Y| |Y| | |Y| | | | | |Y|
+| hard_swish|Y|Y|Y|Y|Y|Y|Y| | |Y| | | | | |Y|
+| im2sequence|Y| | |Y| | |Y| | | | | | | | | |
+| increment| | | |Y|Y| | | | | | | | | | | |
+| index_select| | | | |Y| | | | | | | | | | | |
+| instance_norm|Y|Y| |Y| |Y| | | |Y| | | | | | |
+| inverse| | | | |Y| | | | | | | | | | | |
+| io_copy| |Y|Y|Y| | | | |Y| | | | | | | |
+| io_copy_once| |Y|Y|Y| | | | | | | | | | | | |
+| is_empty| | | |Y|Y| | | | | | | | | | | |
+| layer_norm|Y| | |Y| |Y| | | |Y| | | | | | |
+| layout|Y|Y| | | |Y| | |Y| | | | | | | |
+| layout_once|Y|Y| | | | | | | | | | | | | | |
+| leaky_relu|Y|Y|Y|Y|Y|Y|Y| |Y|Y| | | | | |Y|
+| less_equal| | | | |Y| | | | |Y| | | | | | |
+| less_than| | | |Y|Y| | | | |Y| | | | | | |
+| linspace| | | | |Y| | | | | | | | | | | |
+| lod_array_length| | | | |Y| | | | | | | | | | | |
+| lod_reset| | | | |Y| | | | | | | | | | | |
+| log|Y| | |Y|Y| | | | |Y| | | | | | |
+| logical_and| | | |Y|Y| | | | | | | | | | | |
+| logical_not| | | |Y|Y| | | | | | | | | | | |
+| logical_or| | | | |Y| | | | | | | | | | | |
+| logical_xor| | | | |Y| | | | | | | | | | | |
+| lookup_table|Y| | |Y| |Y| | | | | | | | | | |
+| lookup_table_dequant|Y| | | | | | | | | | | | | | | |
+| lookup_table_v2|Y| | |Y| |Y| | | |Y| | | | | | |
+| lrn|Y|Y| |Y| | | | |Y| | | | | | | |
+| lstm|Y| | | | | | | | | | | | | | | |
+| match_matrix_tensor| | | |Y| |Y| | | | | | | | | | |
+| matmul|Y|Y|Y|Y| |Y|Y| | |Y| | | | | |Y|
+| matmul_v2|Y|Y| |Y| | | | | |Y| | | | | |Y|
+| matrix_nms| | | | |Y| | | | | | | | | | | |
+| max_pool2d_with_index| | | |Y| | |Y| | | | | | | | | |
+| mean|Y| | | | | | | | | | | | | | | |
+| merge_lod_tensor|Y| | | | | | | | | | | | | | | |
+| meshgrid| | | | |Y| | | | | | | | | | | |
+| mish|Y| | | | |Y| | | | | | | | | | |
+| mul|Y| | |Y| |Y|Y| | | | | | | | | |
+| multiclass_nms|Y| | | |Y| |Y| | | | | | | | | |
+| multiclass_nms2|Y| | | |Y| |Y| | | | | | | | | |
+| multiclass_nms3|Y| | | |Y| | | | | | | | | | | |
+| nearest_interp|Y|Y|Y|Y| |Y|Y| |Y|Y| | | | | |Y|
+| nearest_interp_v2|Y|Y|Y|Y| |Y| | | |Y| | | | | |Y|
+| negative|Y| | | | | | | | | | | | | | | |
+| norm|Y| | |Y|Y| |Y| |Y|Y| | | | | | |
+| not_equal| | | | |Y| | | | |Y| | | | | | |
+| one_hot| | | | |Y| | | | | | | | | | | |
+| one_hot_v2| | | | |Y| | | | | | | | | | | |
+| p_norm|Y| | | |Y| | | | |Y| | | | | | |
+| pad2d|Y|Y|Y|Y|Y| | | | |Y| | | | | | |
+| pad3d| | | | |Y| | | | |Y| | | | | | |
+| pixel_shuffle|Y|Y| |Y|Y| | | | | | | | | | | |
+| polygon_box_transform| | | | |Y| | | | | | | | | | | |
+| pool2d|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y|Y|Y|Y|
+| pow|Y| | |Y| |Y| | | |Y| | | | | | |
+| prelu|Y|Y| |Y|Y| | | | |Y| | | | | | |
+| print| | | | |Y| | | | | | | | | | | |
+| prior_box|Y| | |Y|Y| |Y| | | | | | | | | |
+| range| | | | |Y| | | | |Y| | | | | | |
+| read_from_array| | | |Y|Y| | | | | | | | | | | |
+| reciprocal|Y| | |Y|Y| | | | | | | | | | | |
+| reduce_all| | | |Y|Y| | | | | | | | | | | |
+| reduce_any| | | |Y|Y| | | | | | | | | | | |
+| reduce_max|Y|Y| |Y| |Y|Y| | | | | | | | | |
+| reduce_mean|Y|Y| |Y| |Y|Y| | |Y| | | | | | |
+| reduce_min|Y| | |Y| |Y| | | | | | | | | | |
+| reduce_prod|Y| | |Y| |Y| | | | | | | | | | |
+| reduce_sum|Y| | |Y| |Y|Y| | | | | | | | | |
+| relu|Y|Y|Y|Y|Y|Y|Y| |Y|Y|Y|Y|Y|Y|Y|Y|
+| relu6|Y|Y|Y|Y|Y|Y| | |Y|Y|Y|Y|Y|Y|Y|Y|
+| relu_clipped|Y| | | |Y| | | | | | | | | | | |
+| reshape| |Y| |Y|Y| |Y| |Y|Y|Y|Y|Y| |Y|Y|
+| reshape2| |Y|Y|Y|Y| |Y| |Y|Y|Y|Y|Y| |Y|Y|
+| retinanet_detection_output| | | | |Y| | | | | | | | | | | |
+| reverse| | | | |Y| | | | | | | | | | | |
+| rnn|Y| | |Y| |Y| | | | | | | | | | |
+| roi_align| | | |Y|Y| | | | | | | | | | | |
+| roi_perspective_transform| | | | |Y| | | | | | | | | | | |
+| rsqrt|Y|Y| |Y|Y|Y| | | | | | | | | | |
+| scale|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|Y|
+| scatter|Y| | | | | | | | | | | | | | | |
+| scatter_nd_add| | | | |Y| | | | | | | | | | | |
+| search_aligned_mat_mul| | | | | |Y| | | | | | | | | | |
+| search_attention_padding_mask| | | | | |Y| | | | | | | | | | |
+| search_fc| | | |Y| | | | | | | | | | | | |
+| search_grnn| | | |Y| |Y| | | | | | | | | | |
+| search_group_padding| | | | | |Y| | | | | | | | | | |
+| search_seq_arithmetic| | | |Y| |Y| | | | | | | | | | |
+| search_seq_depadding| | | | | |Y| | | | | | | | | | |
+| search_seq_fc| | | | | |Y| | | | | | | | | | |
+| search_seq_softmax| | | | | |Y| | | | | | | | | | |
+| select_input| | | | |Y| | | | | | | | | | | |
+| sequence_arithmetic| | | |Y| |Y| | | | | | | | | | |
+| sequence_concat| | | |Y| |Y| | | | | | | | | | |
+| sequence_conv|Y| | | | |Y| | | | | | | | | | |
+| sequence_expand| | | | |Y| | | | | | | | | | | |
+| sequence_expand_as|Y| | | | |Y| | | | | | | | | | |
+| sequence_mask| | | |Y|Y| | | | | | | | | | | |
+| sequence_pad| | | |Y|Y| | | | | | | | | | | |
+| sequence_pool|Y| | |Y| |Y| | | | | | | | | | |
+| sequence_reshape| | | | | |Y| | | | | | | | | | |
+| sequence_reverse| | | |Y| |Y| | | | | | | | | | |
+| sequence_softmax| | | | |Y| | | | | | | | | | | |
+| sequence_topk_avg_pooling| | | |Y| |Y| | | | | | | | | | |
+| sequence_unpad| | | |Y|Y| | | | | | | | | | | |
+| shape| |Y| |Y|Y| |Y| | |Y| | | | | | |
+| shuffle_channel|Y|Y|Y| |Y| | | | | | | | | | |Y|
+| sigmoid|Y|Y|Y|Y|Y|Y|Y| |Y|Y|Y|Y|Y| |Y|Y|
+| sign|Y| | |Y| | | | | | | | | | | | |
+| sin| |Y| | |Y| | | | | | | | | | | |
+| slice|Y|Y|Y|Y| |Y|Y| |Y|Y| | | | | |Y|
+| softmax|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y|Y|Y|Y|
+| softplus|Y| | | | | | | | | | | | | | | |
+| softsign| | | |Y| |Y| | | | | | | | | | |
+| sparse_conv2d|Y| | | | | | | | | | | | | | | |
+| split| |Y|Y|Y|Y| |Y| |Y|Y| | |Y| | |Y|
+| split_lod_tensor|Y| | | | | | | | | | | | | | | |
+| sqrt|Y|Y| |Y| |Y|Y| | | | | | | | | |
+| square|Y|Y| |Y|Y|Y|Y| | | | | | | | | |
+| squeeze| |Y| |Y|Y| |Y| |Y|Y| | | | | |Y|
+| squeeze2| |Y| |Y|Y| |Y| |Y|Y| | | | | |Y|
+| stack| | | |Y|Y|Y| | | |Y| | | | | | |
+| strided_slice| | | | |Y| | | | | | | | | | | |
+| subgraph| | | | | | | | |Y| | | | | | | |
+| sum|Y| | |Y| | | | | | | | | | | | |
+| swish|Y|Y|Y|Y|Y| |Y| | |Y| | | | | | |
+| sync_batch_norm|Y|Y| | | |Y| | | | | | | | | | |
+| tanh|Y|Y| |Y|Y|Y| | | |Y|Y|Y|Y| |Y| |
+| tensor_array_to_tensor| | | | |Y| | | | | | | | | | | |
+| thresholded_relu|Y| | | |Y| | | | | | | | | | | |
+| tile| | | | |Y| | | | | | | | | | | |
+| top_k| | | |Y|Y| | | | |Y| | | | | | |
+| top_k_v2| | | | |Y| | | | |Y| | | | | | |
+| transpose|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|Y|
+| transpose2|Y|Y|Y|Y| |Y|Y| |Y|Y|Y|Y|Y| |Y|Y|
+| tril_triu| | | | |Y| | | | | | | | | | | |
+| uniform_random| | | | |Y| | | | | | | | | | | |
+| unique_with_counts| | | | |Y| | | | | | | | | | | |
+| unsqueeze| |Y| |Y|Y| | | | |Y| | | | | |Y|
+| unsqueeze2| |Y| |Y|Y| | | | |Y| | | | | |Y|
+| unstack| | | |Y|Y| | | | | | | | | | | |
+| var_conv_2d| | | |Y| |Y| | | | | | | | | | |
+| where| | | | |Y| | | | | | | | | | | |
+| where_index| | | | |Y| | | | | | | | | | | |
+| while| | | | |Y| | | | | | | | | | | |
+| write_back| | | | |Y| | | | | | | | | | | |
+| write_to_array| | | |Y|Y| | | | | | | | | | | |
+| yolo_box|Y|Y|Y|Y|Y| |Y| | | | | | | | | |
+| __xpu__bigru| | | |Y| | | | | | | | | | | | |
+| __xpu__conv2d| | | |Y| | | | | | | | | | | | |
+| __xpu__dynamic_lstm_fuse_op| | | |Y| | | | | | | | | | | | |
+| __xpu__embedding_with_eltwise_add| | | |Y| | | | | | | | | | | | |
+| __xpu__fc| | | |Y| | | | | | | | | | | | |
+| __xpu__generate_sequence| | | |Y| | | | | | | | | | | | |
+| __xpu__logit| | | |Y| | | | | | | | | | | | |
+| __xpu__mmdnn_bid_emb_att| | | |Y| | | | | | | | | | | | |
+| __xpu__mmdnn_bid_emb_grnn_att| | | |Y| | | | | | | | | | | | |
+| __xpu__mmdnn_bid_emb_grnn_att2| | | |Y| | | | | | | | | | | | |
+| __xpu__mmdnn_match_conv_topk| | | |Y| | | | | | | | | | | | |
+| __xpu__mmdnn_merge_all| | | |Y| | | | | | | | | | | | |
+| __xpu__mmdnn_search_attention| | | |Y| | | | | | | | | | | | |
+| __xpu__mmdnn_search_attention2| | | |Y| | | | | | | | | | | | |
+| __xpu__multi_encoder| | | |Y| | | | | | | | | | | | |
+| __xpu__multi_softmax| | | |Y| | | | | | | | | | | | |
+| __xpu__resnet50| | | |Y| | | | | | | | | | | | |
+| __xpu__resnet_cbam| | | |Y| | | | | | | | | | | | |
+| __xpu__sfa_head| | | |Y| | | | | | | | | | | | |
+| __xpu__softmax_topk| | | |Y| | | | | | | | | | | | |
+| __xpu__squeeze_excitation_block| | | |Y| | | | | | | | | | | | |
diff --git a/docs/source_compile/include/multi_device_support/nnadapter_support_verisilicon_timvx.rst b/docs/source_compile/include/multi_device_support/nnadapter_support_verisilicon_timvx.rst
new file mode 100644
index 00000000000..fcb3c51d3bb
--- /dev/null
+++ b/docs/source_compile/include/multi_device_support/nnadapter_support_verisilicon_timvx.rst
@@ -0,0 +1,27 @@
+NNAdapter 支持芯原 TIM-VX
+^^^^^^^^^^^^^^^^^^^^^^^^
+
+.. list-table::
+
+ * - 参数
+ - 说明
+ - 可选范围
+ - 默认值
+ * - nnadapter_with_verisilicon_timvx
+ - 是否编译芯原 TIM-VX 的 NNAdapter HAL 库
+ - OFF / ON
+ - OFF
+ * - nnadapter_verisilicon_timvx_src_git_tag
+ - 设置芯原 TIM-VX 的代码分支
+ - TIM-VX repo 分支名
+ - main
+ * - nnadapter_verisilicon_timvx_viv_sdk_url
+ - 设置芯原 TIM-VX SDK 的下载链接
+ - 用户自定义
+ - Android系统：http://paddlelite-demo.bj.bcebos.com/devices/verisilicon/sdk/viv_sdk_android_9_armeabi_v7a_6_4_4_3_generic.tgz
+ Linux系统：http://paddlelite-demo.bj.bcebos.com/devices/verisilicon/sdk/viv_sdk_linux_arm64_6_4_4_3_generic.tgz
+ * - nnadapter_verisilicon_timvx_viv_sdk_root
+ - 设置芯原 TIM-VX 的本地路径
+ - 用户自定义
+ - 空值
+详细请参考 `芯原 TIM-VX 部署示例 `_
diff --git a/docs/source_compile/linux_x86_compile_android.rst b/docs/source_compile/linux_x86_compile_android.rst
index e32c4ff45d1..c69e0e7e358 100644
--- a/docs/source_compile/linux_x86_compile_android.rst
+++ b/docs/source_compile/linux_x86_compile_android.rst
@@ -242,3 +242,5 @@ Paddle Lite 仓库中\ ``/lite/tools/build_android.sh``\ 脚本文件用于构
.. include:: include/multi_device_support/nnadapter_support_mediatek_apu.rst
.. include:: include/multi_device_support/nnadapter_support_amlogic_npu.rst
+
+.. include:: include/multi_device_support/nnadapter_support_verisilicon_timvx.rst
\ No newline at end of file
diff --git a/docs/source_compile/linux_x86_compile_arm_linux.rst b/docs/source_compile/linux_x86_compile_arm_linux.rst
index 136309891ad..4b0635827b1 100644
--- a/docs/source_compile/linux_x86_compile_arm_linux.rst
+++ b/docs/source_compile/linux_x86_compile_arm_linux.rst
@@ -186,3 +186,5 @@ Paddle Lite 仓库中 \ ``./lite/tools/build_linux.sh``\ 脚本文件用于构
.. include:: include/multi_device_support/nnadapter_support_rockchip_npu.rst
.. include:: include/multi_device_support/nnadapter_support_amlogic_npu.rst
+
+.. include:: include/multi_device_support/nnadapter_support_verisilicon_timvx.rst
\ No newline at end of file