A while ago, a developer showed the performance results using Intel Graphics. I am new to TVM. Do you mind explaining in detail how a user can reproduce the results ?
Thanks!
"Optimize convolutions specifically for Intel Graphics which compared with other opencl based backends, adds specific intel_sub_group operations. And we support consistent data packing (thanks to @yzhliu’s implementation ) for end-to-end networks. Note that currently conv2d is most optimized for workloads for resnet. We compared with state-of-the-art deep learning libraries for intel GPUs. Following is some end-to-end benchmark for Intel® Processor Graphics Gen9 using opt_level 3. "
|Networks|TVM|clDNN (Intel’s deep learning library)|
|ResNet-18|74ms|70ms| |ResNet-50|262 ms|208.4 ms|