source: libcfa/src/concurrency/kernel/cluster.hfa@ 77125cc

//
// Cforall Version 1.0.0 Copyright (C) 2022 University of Waterloo
//
// The contents of this file are covered under the licence agreement in the
// file "LICENCE" distributed with Cforall.
//
// cluster.hfa -- file that includes helpers for subsystem that need cluster wide support
//
// Author           : Thierry Delisle
// Created On       : Tue Mar 15 16:40:12 2022
// Last Modified By :
// Last Modified On :
// Update Count     :
//

#pragma once

#include "device/cpu.hfa"
#include "kernel/private.hfa"

#include <limits.h>

//-----------------------------------------------------------------------
// Calc moving average based on existing average, before and current time.
static inline unsigned long long moving_average(unsigned long long currtsc, unsigned long long instsc, unsigned long long old_avg, bool strict) {
        (void)strict; // disable the warning around the fact this is unused in release.

        const unsigned long long new_val = currtsc > instsc ? currtsc - instsc : 0;
        const unsigned long long total_weight = 16;
        const unsigned long long new_weight   = 4;
        const unsigned long long old_weight = total_weight - new_weight;
        const unsigned long long ret = ((new_weight * new_val) + (old_weight * old_avg)) / total_weight;
        return ret;
}

static inline void touch_tsc(__timestamp_t * tscs, size_t idx, unsigned long long ts_prev, unsigned long long ts_next, bool strict) {
        if (ts_next == ULLONG_MAX) return;
        unsigned long long now = rdtscl();
        unsigned long long pma = __atomic_load_n(&tscs[ idx ].t.ma, __ATOMIC_RELAXED);
        __atomic_store_n(&tscs[ idx ].t.tv, ts_next, __ATOMIC_RELAXED);
        __atomic_store_n(&tscs[ idx ].t.ma, moving_average(now, ts_prev, pma, strict), __ATOMIC_RELAXED);
}

//-----------------------------------------------------------------------
// Calc age a timestamp should be before needing help.
forall(Data_t * | { unsigned long long ts(Data_t & this); })
static inline unsigned long long calc_cutoff(
        const unsigned long long ctsc,
        unsigned procid,
        size_t count,
        Data_t * data,
        __timestamp_t * tscs,
        const unsigned shard_factor,
        bool strict
) {
        unsigned start = procid;
        unsigned long long max = 0;
        for(i; shard_factor) {
                unsigned long long ptsc = ts(data[start + i]);
                if(ptsc != ULLONG_MAX) {
                        /* paranoid */ verify( start + i < count );
                        unsigned long long tsc = moving_average(ctsc, ptsc, tscs[start + i].t.ma, strict);
                        if(tsc > max) max = tsc;
                }
        }
        return 8 * max;
}

static inline unsigned cache_id(struct cluster * cltr, unsigned idx) with (cltr->sched) {
        // Figure out the current cpu and make sure it is valid
        const int cpu = __kernel_getcpu();
        /* paranoid */ verify(cpu >= 0);
        /* paranoid */ verify(cpu < cpu_info.hthrd_count);
        unsigned this_cache = cpu_info.llc_map[cpu].cache;

        // Super important: don't write the same value over and over again
        // We want to maximise our chances that his particular values stays in cache
        if(caches[idx].id != this_cache)
                __atomic_store_n(&caches[idx].id, this_cache, __ATOMIC_RELAXED);

        return this_cache;
}

static struct {
        const unsigned readyq;
        const unsigned io;
} __shard_factor = { 2, 1 };

// Local Variables: //
// mode: c //
// tab-width: 4 //
// End: //