[Tuning] rel-4.5 fdb update

[cderb]

No description provided.

[junliume]

@cderb and @JehandadKhan: do we attempt to merge into release/rocm-rel-4.5 directly or through release/rocm-rel-4.5-staging?

[junliume]

@cderb and @JehandadKhan:
Failure on this issue, restarted to check if it is run to run:

FAILED: filesystem::recursive_directory_iterator increment error: No such file or directory

[atamazov]

FAILED: filesystem::recursive_directory_iterator increment error: No such file or directory

Happened on ixt-sjc2-22 which is Navi21/36 during smoke testing with OCL backend. We do not supply fdb for it.

[junliume]

@cderb @JehandadKhan : would this "fix" this issue: https://ontrack-internal.amd.com/browse/SWDEV-305355?

[JehandadKhan]

@cderb @JehandadKhan : would this "fix" this issue: https://ontrack-internal.amd.com/browse/SWDEV-305355?

Yes

[atamazov]

Unfortunately, not good to go because #1196 has issues.

issue fixed: fdb entries no longer contain duplicate algorithms.

[atamazov]

Now in testing.

[atamazov]

gfx906 perf testing results (ROCm 4.0)

🟢 Looks good. release/rocm-rel-4.5-staging, prior vs after merging cderb/fdb_tuning_4.5, tested with gfx906/60 & ROCm 4.0. Performance gain, based on sum of all times/directions:

		GPU time	Wall time	Aux Wall time
FP32/HIP	Default find mode	1.14	1.08	1.45
	Immediate mode	1.13	1.11	1.50
FP16/HIP	Default find mode	1.00	0.90	0.74
	Immediate mode	1.00	0.99	1.19
BF16/HIP	Default find mode	1.00	0.95	1.09
	Immediate mode	1.00	1.00	1.13
FP32/OpenCL	Default find mode	0.99	1.02	0.99
	Immediate mode	1.00	1.00	0.98