Raistmer

SKA and Profiler disagree on registers number

Discussion created by Raistmer on May 7, 2010
Latest reply on May 8, 2010 by Raistmer
SKA shows no scratchshould be used but profiler uses scratch registers.

For same kernel SKA shows:
Name,Scratch Reg,GPR,Min,Max,Avg,Est Cycles,Est Cycles(Bi),ALU:Fetch(Bi),BottleNeck(Bi),%s\Clock(Bi),Throughput(Bi)
Radeon HD 4870,0,119,244.90,36705.76,6919.27,6919.27,6919.27,1.19,ALU Ops,0.00,2 M Threads\Sec

0 scratch registers and 119 GPR.
And indeed, HD4870 assembly doesn't contain scratch read/write instructions.

But profiler shows 121 GPR and scratch registers usage!
And its assembly shows scratch instructions.
MaxScratchRegsNeeded = 98
SQ_PGM_RESOURCES:NUM_GPRS = 121


Why disagreement in so important thing for performance as scratch registers usage?

Outcomes