Menu
This question and its answer, which was recently tagged as an Epic Answer, has prompted me to wonder; Can I measure the performance of a running application in Windows in terms of its CPU branch prediction failures? I know that some static analysis tools exist, that might help with optimizing code for good performance in branch-prediction situations, and that manual techniques could help by simply making changes and re-testing, but I'm looking for some automatic mechanism that can report a total number of branch prediction failures, over a period of time, as a Windows application runs, and I'm hoping that some Profiler tool for Visual C++ could help me.
For the sake of this question, the application in question is either built with a native-compiler such as Visual C++ for Windows, or using some other native compiler, such as GCC, FreePascal, Delphi, or TurboAssembler. The executable may not have any debug information at all. I want to know if I can detect, and count branch prediction failures, perhaps by reading internal CPU information through some Windows service like WMI, or perhaps by running entirely inside a virtualized environment running Windows, such as using VirtualBox, and then running a completely virtualized windows environment with my test application, inside VirtualBox, and doing runtime analysis of the virtual CPU. Or some other technique that I don't know of, thus this question.
Yes, I googled. The only thing that looks promising is this PDF from AMD. Page 18 mentions something very close to what I'd like to do, but seems written for those working without any operating system, on raw evaluation hardware platforms:
5.1. Branches. Applicability. Conditional branch mispredictions may be a significant issue in code with a lot of decision-making logic.
Conditional branches may be mispredicted when the likelihood of choosing the true or false path is random or near a 50-50 split. The branch prediction hardware cannot "learn" a pattern and branches are not predicted correctly. Collection. Collect the events in this table to measure branch prediction performance:
Branches Compute the rate at which branches are taken and the ratio of the number of instructions per branch using these formulas: Branch taken rate = Taken_branches / Ret_instructions Branch taken ratio = Taken_branches / Branches
Instructions per branch = Ret_instructions / Branches
Update: I guess I could say I'm looking for a way to read the Intel Core i7 PMU module, or equivalent functions of other CPUs. It looks like Intel VTUNE (from the comments by Adrian) is very close to what I asked for.
842
Branch prediction is an approach to computer architecture that attempts to mitigate the costs of branching. Branch predication speeds up the processing of branch instructions with CPUs using pipelining. The technique involves only executing certain instructions if certain predicates are true.
It predicts where to fetch from next cycle, and then based on that prediction it predicts where to fetch from the cycle after that, and so on. It doesn't use any information about the decoded instructions, but only past behavior of the branches.
Branch prediction is very important to the performance of a deeply pipelined processor. Branch prediction enables the processor to begin executing instructions long before the branch outcome is certain. Branch delay is the penalty that is incurred in the absence of a correct prediction.
VTune Performance Analyzer can do it! Btw if you are studying these topics, take a look at "Optimization Cookbook" from Intel Press.
Note: Comments state the same answer but with some uncertainty, I used VTune and I measured the branch prediction rate for an Intel CPU. So I'm 100% sure.
here is the link for VTune
here is the link for the book
88
If you love us? You can donate to us via Paypal or buy me a coffee so we can maintain and grow! Thank you!
Donate Us With
