Performance degradation for higher batch size on PPC64LE

[pratiklp00]

Hi,
While running the large batch size model we have observed the performance degradation. We were able to trace it back to #4920

[martin-frbg]

That's bad but my impression was that it is a question of right vs faster. It is possible of course that it is a specific corner case that needed fixing rather than the algorithm in general, or that #4920 papers over an actual issue elsewhere.

[pratiklp00]

Hi @martin-frbg Thank you for your response. We haven't observed the zgemm/cgemm test failure on PPC64LE. Would it be possible to implement a conditional check for PPC so that we can continue using the previous code? If not, could you guide me on the best way to proceed with this?

[ChipKerchner]

PPC64LE is seeing as much as a 2X slowdown with the above patch. I suspect that there might be an issue with C/ZGEMM for x86 (stride issue, overlapping data, ?) and though this fix prevents that, it had an adverse effect of slowing some cases significantly. We are not seeing it most likely because we specialize these cases.

[pratiklp00]

@martin-frbg Hi Martin we would like to make below change. If you have any other suggestions please let me know.

#if defined(POWER)
if (width < switch_ratio ) {
#else
if (width < switch_ratio && width > 1) {
#endif