TVM性能评估分析（三）

 

 

 Figure 1. TVM’s WebGPU backend close to native GPU performance when deploying models to the web.

 

 

 Figure 2.  WebGPU is to write shaders for primitive operators in deep neural networks

 

 

 Figure 3.  Build a WebGPU runtime inside TVM’s JS runtime

 

 

 Figure 4. Comparing the execution of a full computational graph via TVM’s WebGPU backend and native targets

 

 

 Figure 5. 2D convolution with data layout in NCHW4c and weight layout in OIHW4o4i. Left: The input tensor in NCHW4c layout. One moving filter of the kernel is colored in blue. One element of the input and kernel is colored in grey. Mid: The packed input and kernel in the grey block. Right: The output in NCHW4c layout. Inside the one element depicted, there are four packed elements in channel sub-dimension.

 

 

 Figure 6. Workflow of running quantized models

 

 

 Figure 7.  A full deep learning compiler stack to support machine learning workloads for diverse hardware backends.

 

 

 Figure 8. Golang Interface over TVM Runtime

 

 

 Figure 9.  Import, Compile, Integrate and Deploy