amd zen2 predefined events need to be updated

[Vince Weaver]

as reported by Steve Kaufmann, the zen2 events are not the same as zen1 and papi_events.csv needs to be updated.

unfortunately the events between zen1 and zen2 seem to be very different so it’s not a quick fix

[Sebastian Mobo]

I took a look through some documentation for Zen 2 performance events, and added presets based on that in pull request #127. That said, I don’t have hardware that I can actually run tests on, so I’m not sure if they’re fully correct yet.

[David Álvarez]

Hi!

How’s the status of this issue? I have tested pull request #127 and found that FP_INS, FP_OPS and L1_DCM do not pass testing on a Zen2 System I have access to, but the rest seem to work fine. Specifically, L1_DCM is counting CHANGE_TO_X accesses to the L2 cache, and if I remove that, then I pass the L1 cache tests. In master there is a FP_INS counter for Zen2 that works fine, but FP_OPS is also wrong apparently.

Can I help in some way to move this forward? I can provide testing on a Zen2 system

‌

[Giuseppe Congiu]

@Anthony Danalis This is a pretty old issue. Do you know if this has been resolved?

[Giuseppe Congiu]

# New FLOP event on zen2
# PPR (under section 2.1.15.3. --
# https://www.amd.com/system/files/TechDocs/54945_3.03_ppr_ZP_B2_pub.zip)
# explains that FLOP events require MergeEvent support, which was included•
# in the 5.6 kernel.•
# Hence, a kernel version 5.6 or greater is required.
# NOTE: without the MergeEvent support in the kernel, there is no guarantee
# that this SSE/AVX FLOP event produces any useful data whatsoever.••
PRESET,PAPI_FP_OPS,NOT_DERIVED,RETIRED_SSE_AVX_FLOPS:ANY
# Since FP_OPS counts both single- and double-prec operations
# correctly, we don't need to confuse the user with additional
# DP_OPS and SP_OPS events. So, I'm taking them out.••
#PRESET,PAPI_DP_OPS,NOT_DERIVED,RETIRED_SSE_AVX_FLOPS:ANY
#PRESET,PAPI_SP_OPS,NOT_DERIVED,RETIRED_SSE_AVX_FLOPS:ANY
#
# Floating-point instructions (including non-numeric floating-point instructions,
# e.g. Move or Merge Scalar Double-Precision Floating-Point values)
PRESET,PAPI_FP_INS,NOT_DERIVED,RETIRED_MMX_FP_INSTRUCTIONS:SSE_INSTR:MMX_INSTR:X87_INSTR
# Since FP_INS counts both single- and double-prec instuctions
# correctly, we don't need to confuse the user with additional
# VEC_DP and VEC_SP events. So, I'm taking them out.••

From `papi_events.csv`. Please, consider the comments above.