Request for Explanation of values set to some variables in jit_avx512_common_conv_kernel.cpp

[kasturedeeksha]

Request for Explanation of values set to some variables in jit_avx512_common_conv_kernel.cpp

[shu1chen]

@kasturedeeksha From my understanding, the number of blocks jcp.nb_ic_blocking and jcp.nb_oc_blocking are initialized with value 1, meaning the data is initially not blocked. The values can be updated in some conditions, e.g. jcp.nb_ic_blocking can be updated with value i in line 2104 and jcp.nb_oc_blocking can be updated with value best_nb_oc_blocking in line 1197.

@msotoflo I see in the commit history that you updated the heuristic for nb_oc_blocking 4 years ago. Do you possibly have any suggestions for manually adjusting the values of jcp.nb_ic_blocking and jcp.nb_oc_blocking?

[msotoflo]

Hello @kasturedeeksha

There might be multiple reasons why these params were set to 1. Although primarily, depending on which direction the convolution is executed (e.g. forward vs backward propagation) is what kind of impact nb_ic_blocking or nb_oc_blocking will have. For forward direction, nb_oc_blocking will allow loop unrolling on the output elements. The best value for this param is chosen in src/cpu/x64/jit_avx512_common_conv_kernel.cpp#L1197 while taking into account other factors. jcp.nb_ic_blocking will do the same for backward propagation.

Additionally, I would like to point out that these are somewhat legacy JIT kernels. We now mostly depend on BRGeMM JIT Implementations to support the convolution primitive and give better performance.

Might I ask what is your interest in this specific implementation and what your approach is to running these primitives? We expect users to execute OneDNN primitives through the API and let OneDNN choose the best primitive implementation (as well as its internal parameters).