Qianshui-Jiang opened a new pull request, #11571: URL: https://github.com/apache/tvm/pull/11571
This PR mainly about mapping oneDNN OP implementation in X86 Relay Op Strategy. we've observed that nn.dense kernel that could be dispatched to DNNL by assigning "-libs=mkldnn" and there is also conv2d kernel implemented in runtime/contrib/dnnl. so we mapping it in X86 Relay Op Strategy and optimized the kernel implementation to let DNNL choose blocked format according to different input shape, as [performance-profiling example](https://oneapi-src.github.io/oneDNN/page_performance_profiling_cpp.html#doxid-performance-profiling-cpp) discribed in oneDNN doc. Here is the details: - Adjust DNNL Conv2D implementation to let it support NHWC format and automate reorder for input/weights/outputs to abtain best performace. - Add 'target.libs=mkldnn' branch in Relay X86 OP strategy for both NCHW and NHWC Conv2D kernel. - Add 2 test funtions for case mentioned above. We are trying to enable more DNNL kernels including different format and datatypes this way. -- 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]
