metrics.py

#Base classes for different metrics

print_error = lambda msg: False

class MetricBase(object):
    # Derived classes can override these
    level = 0
    name = ""
    domain = ""
    area = ""
    server = False
    htoff = False
    sample = []
    sibling = None
    metricgroup = []
    desc = "Missing description"

    def __init__(self, **kwargs):
        self.errcount = 0
        for k, v in kwargs.items():
            setattr(self, k, v)

    def compute(self, EV):
        try:
            self.val = self._compute(EV)
            self.thresh = self.val > 0
        except ZeroDivisionError:
            print_error("{0} zero division".format(self.__class__.__name__))
            self.errcount += 1
        self.val = 0
        self.thresh = False
        return self.val

    def _compute(self, EV):
        raise NotImplementedError()

class FrontendBound(MetricBase):
    level = 1
    name = "Frontend_Bound"
    domain = "Slots"
    area = "FE"
    desc = ("\n"
            "This category reflects slots where the Frontend of the\n"
            "processor undersupplies its Backend.")
    metricgroup = ['TopDownL1']

class FrontendLatency(MetricBase):
    level = 2
    name = "Frontend_Latency"
    domain = "Slots"
    area = "FE"
    desc = ("\n"
            "This metric represents slots fraction the CPU was stalled\n"
            "due to Frontend latency issues.  For example; instruction-\n"
            "cache misses; iTLB misses or fetch stalls after a branch\n"
            "misprediction are categorized under Frontend Latency. In\n"
            "such cases; the Frontend eventually delivers no uops for\n"
            "some period.")
    htoff = False
    sample = []
    errcount = 0
    sibling = None
    server = False
    metricgroup = ['Frontend_Bound', 'TopDownL2']

class BadSpeculation(MetricBase):
    level = 1
    name = "Bad_Speculation"
    domain = "Slots"
    area = "BAD"
    desc = ("\n"
            "This category represents slots fraction wasted due to\n"
            "incorrect speculations. This includes slots used to issue\n"
            "uops that do not eventually get retired and slots for which\n"
            "the issue-pipeline was blocked due to recovery from earlier\n"
            "incorrect speculation. For example; wasted work due to miss-\n"
            "predicted branches are categorized under Bad Speculation\n"
            "category. Incorrect data speculation followed by Memory\n"
            "Ordering Nukes is another example.")
    metricgroup = ['Bad_Speculation', 'TopDownL1']

class Retiring(MetricBase):
    level = 1
    name = "Retiring"
    domain = "Slots"
    area = "RET"
    desc = ("\n"
            "This category represents slots fraction utilized by useful\n"
            "work i.e. issued uops that eventually get retired. Ideally;"
            "all pipeline slots would be attributed to the Retiring\n"
            "category.  Retiring of 100% would indicate the maximum 4\n"
            "uops retired per cycle has been achieved.  Maximizing\n"
            "Retiring typically increases the Instruction-Per-Cycle\n"
            "metric. Note that a high Retiring value does not necessary\n"
            "mean there is no room for more performance.  For example;\n"
            "Microcode assists are categorized under Retiring. They hurt\n"
            "performance and can often be avoided. . A high Retiring\n"
            "value for non-vectorized code may be a good hint for\n"
            "programmer to consider vectorizing his code.  Doing so\n"
            "essentially lets more computations be done without\n"
            "significantly increasing number of instructions thus\n"
            "improving the performance.")
    metricgroup = ['TopDownL1']

class BackendBound(MetricBase):
    level = 1
    name = "Backend_Bound"
    domain = "Slots"
    area = "BE"
    desc = ("\n"
            "This category represents slots fraction where no uops are\n"
            "being delivered due to a lack of required resources for\n"
            "accepting new uops in the Backend. Backend is the portion of\n"
            "the processor core where the out-of-order scheduler\n"
            "dispatches ready uops into their respective execution units;\n"
            "and once completed these uops get retired according to\n"
            "program order. For example; stalls due to data-cache misses\n"
            "or stalls due to the divider unit being overloaded are both\n"
            "categorized under Backend Bound. Backend Bound is further\n"
            "divided into two main categories: Memory Bound and Core\n"
            "Bound.")
    metricgroup = ['TopDownL1']

class ICacheMisses(MetricBase):
    level = 3
    name = "ICache Misses"
    domain = "Clocks"
    area = "FE"
    desc = ("\n"
            "This metric represents cycles fraction the CPU was stalled\n"
            "due to instruction cache misses.. Using compiler's Profile-\n"
            "Guided Optimization (PGO) can reduce i-cache misses through\n"
            "improved hot code layout.")
    metricgroup = ['Frontend_Latency']

class ITLBMisses(MetricBase):
    level = 3
    name = "ITLB_Misses"
    domain = "Clocks"
    area = "FE"
    desc = ("\n"
            "This metric represents cycles fraction the CPU was stalled\n"
            "due to instruction TLB misses.. Consider large 2M pages for\n"
            "code (selectively prefer hot large-size function, due to\n"
            "limited 2M entries). Linux options: standard binaries use\n"
            "libhugetlbfs; Hfsort.. https://github.com/libhugetlbfs/libhu\n"
            "getlbfs;https://research.fb.com/publications/optimizing-\n"
            "function-placement-for-large-scale-data-center-\n"
            "applications-2/\n")
    metricgroup = ['Frontend_Latency', 'TLB']

class BranchResteers(MetricBase):
    level = 3
    name = "Branch_Resteers"
    domain = "Clocks"
    area = "FE"
    desc = ("\nThis metric represents cycles fraction the CPU was stalled\n"
            "due to Branch Resteers. Branch Resteers estimates the\n"
            "Frontend delay in fetching operations from corrected path;\n"
            "following all sorts of miss-predicted branches. For example;\n"
            "branchy code with lots of miss-predictions might get\n"
            "categorized under Branch Resteers. Note the value of this\n"
            "node may overlap with its siblings.")
    metricgroup = ['Bad_Speculation', 'Frontend_Latency']

class MSSwitches(MetricBase):
    level = 3
    name = "MS_Switches"
    domain = "Clocks"
    area = "FE"
    desc = ("\n"
            "This metric estimates the fraction of cycles when the CPU\n"
            "was stalled due to switches of uop delivery to the Microcode\n"
            "Sequencer (MS). Commonly used instructions are optimized for\n"
            "delivery by the DSB (decoded i-cache) or MITE (legacy\n"
            "instruction decode) pipelines. Certain operations cannot be\n"
            "handled natively by the execution pipeline; and must be\n"
            "performed by microcode (small programs injected into the\n"
            "execution stream). Switching to the MS too often can\n"
            "negatively impact performance. The MS is designated to\n"
            "deliver long uop flows required by CISC instructions like\n"
            "CPUID; or uncommon conditions like Floating Point Assists\n"
            "when dealing with Denormals.")
    metricgroup = ['Frontend_Latency', 'Microcode_Sequencer']

class IFetchLine(MetricBase):
    name = "IFetchLine"
    domain = "Metric"
    maxval = 1
    desc = ("\n"
            "This metric represents cycles fraction the fetch stalls\n"
            "due to an instruction cache miss.")
    metricgroup = []