sync.h

Go to the documentation of this file.

// Copyright 2023 ETH Zurich and University of Bologna.
// Licensed under the Apache License, Version 2.0, see LICENSE for details.
// SPDX-License-Identifier: Apache-2.0
//
// Luca Colagrande <colluca@iis.ee.ethz.ch>
// Viviane Potocnik <vivianep@iis.ee.ethz.ch>
 
#pragma once
 
#include "../../deps/riscv-opcodes/encoding.h"
 
#include <math.h>
 
#define SNRT_BROADCAST_MASK ((SNRT_CLUSTER_NUM - 1) * SNRT_CLUSTER_OFFSET)
 
//================================================================================
// Mutex functions
//================================================================================
 
inline volatile uint32_t *snrt_mutex() { return &_snrt_mutex; }
 
inline void snrt_mutex_acquire(volatile uint32_t *pmtx) {
    asm volatile(
        "li            t0,1          # t0 = 1\n"
        "1:\n"
        "  amoswap.w.aq  t0,t0,(%0)   # t0 = oldlock & lock = 1\n"
        "  bnez          t0,1b      # Retry if previously set)\n"
        : "+r"(pmtx)
        :
        : "t0");
}
 
inline void snrt_mutex_ttas_acquire(volatile uint32_t *pmtx) {
    asm volatile(
        "1:\n"
        "  lw t0, 0(%0)\n"
        "  bnez t0, 1b\n"
        "  li t0,1          # t0 = 1\n"
        "2:\n"
        "  amoswap.w.aq  t0,t0,(%0)   # t0 = oldlock & lock = 1\n"
        "  bnez          t0,2b      # Retry if previously set)\n"
        : "+r"(pmtx)
        :
        : "t0");
}
 
inline void snrt_mutex_release(volatile uint32_t *pmtx) {
    asm volatile("amoswap.w.rl  x0,x0,(%0)   # Release lock by storing 0\n"
                 : "+r"(pmtx));
}
 
//================================================================================
// Barrier functions
//================================================================================
 
inline void snrt_wake_all(uint32_t core_mask) {
#ifdef SNRT_SUPPORTS_MULTICAST
    // Multicast cluster interrupt to every other cluster's core
    // Note: we need to address another cluster's address space
    //       because the cluster XBAR has not been extended to support
    //       multicast yet. We address the second cluster, if we are the
    //       first cluster, and the first cluster otherwise.
    if (snrt_cluster_num() > 0) {
        volatile snitch_cluster_t *cluster;
        if (snrt_cluster_idx() == 0)
            cluster = snrt_cluster(1);
        else
            cluster = snrt_cluster(0);
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Waddress-of-packed-member"
        uint32_t *addr = (uint32_t *)&(cluster->peripheral_reg.cl_clint_set.w);
#pragma clang diagnostic pop
        snrt_enable_multicast(SNRT_BROADCAST_MASK);
        *addr = core_mask;
        snrt_disable_multicast();
    }
#else
    for (int i = 0; i < snrt_cluster_num(); i++) {
        if (snrt_cluster_idx() != i) {
            snrt_cluster(i)->peripheral_reg.cl_clint_set.f.cl_clint_set =
                core_mask;
        }
    }
#endif
}
 
inline void snrt_cluster_hw_barrier() {
    asm volatile("csrr x0, 0x7C2" ::: "memory");
}
 
static inline void snrt_inter_cluster_barrier() {
    // Everyone increments a shared counter
    uint32_t cnt =
        __atomic_add_fetch(&(_snrt_barrier.cnt), 1, __ATOMIC_RELAXED);
 
    // All but the last cluster enter WFI, while the last cluster resets the
    // counter for the next barrier and multicasts an interrupt to wake up the
    // other clusters.
    if (cnt == snrt_cluster_num()) {
        _snrt_barrier.cnt = 0;
        // Wake all clusters
        snrt_wake_all(1 << snrt_cluster_core_idx());
    } else {
        snrt_wfi();
        // Clear interrupt for next barrier
        snrt_int_clr_mcip();
    }
}
 
inline void snrt_global_barrier() {
    snrt_cluster_hw_barrier();
 
    // Synchronize all DM cores in software
    if (snrt_is_dm_core()) {
        snrt_inter_cluster_barrier();
    }
    // Synchronize cores in a cluster with the HW barrier
    snrt_cluster_hw_barrier();
}
 
inline void snrt_partial_barrier(snrt_barrier_t *barr, uint32_t n) {
    // Remember previous iteration
    uint32_t prev_it = barr->iteration;
    uint32_t cnt = __atomic_add_fetch(&barr->cnt, 1, __ATOMIC_RELAXED);
 
    // Increment the barrier counter
    if (cnt == n) {
        barr->cnt = 0;
        __atomic_add_fetch(&barr->iteration, 1, __ATOMIC_RELAXED);
    } else {
        // Some threads have not reached the barrier --> Let's wait
        while (prev_it == barr->iteration)
            ;
    }
}
 
//================================================================================
// Reduction functions
//================================================================================
 
inline uint32_t snrt_global_all_to_all_reduction(uint32_t value) {
    // Reduce cores within cluster in TCDM
    uint32_t *cluster_result = &(cls()->reduction);
    uint32_t tmp = __atomic_fetch_add(cluster_result, value, __ATOMIC_RELAXED);
 
    // Wait for writeback to ensure AMO is seen by all cores after barrier
    snrt_wait_writeback(tmp);
    snrt_cluster_hw_barrier();
 
    // Reduce DM cores across clusters in global memory
    if (snrt_is_dm_core()) {
        __atomic_add_fetch(&_reduction_result, *cluster_result,
                           __ATOMIC_RELAXED);
        snrt_inter_cluster_barrier();
        *cluster_result = _reduction_result;
    }
    snrt_cluster_hw_barrier();
    return *cluster_result;
}
 
inline void snrt_global_reduction_dma(double *dst_buffer, double *src_buffer,
                                      size_t len) {
    // If we have a single cluster, no reduction has to be done
    if (snrt_cluster_num() > 1) {
        // Iterate levels in the binary reduction tree
        int num_levels = ceil(log2(snrt_cluster_num()));
        for (unsigned int level = 0; level < num_levels; level++) {
            // Determine whether the current cluster is an active cluster.
            // An active cluster is a cluster that participates in the current
            // level of the reduction tree. Every second cluster among the
            // active ones is a sender.
            uint32_t is_active = (snrt_cluster_idx() % (1 << level)) == 0;
            uint32_t is_sender = (snrt_cluster_idx() % (1 << (level + 1))) != 0;
 
            // If the cluster is a sender, it sends the data in its source
            // buffer to the respective receiver's destination buffer
            if (is_active && is_sender) {
                if (!snrt_is_compute_core()) {
                    uint64_t dst = (uint64_t)dst_buffer -
                                   (1 << level) * SNRT_CLUSTER_OFFSET;
                    snrt_dma_start_1d(dst, (uint64_t)src_buffer,
                                      len * sizeof(double));
                    snrt_dma_wait_all();
                }
            }
 
            // Synchronize senders and receivers
            snrt_global_barrier();
 
            // Every cluster which is not a sender performs the reduction
            if (is_active && !is_sender) {
                // Computation is parallelized over the compute cores
                if (snrt_is_compute_core()) {
                    uint32_t items_per_core =
                        len / snrt_cluster_compute_core_num();
                    uint32_t core_offset =
                        snrt_cluster_core_idx() * items_per_core;
                    for (uint32_t i = 0; i < items_per_core; i++) {
                        uint32_t abs_i = core_offset + i;
                        src_buffer[abs_i] += dst_buffer[abs_i];
                    }
                }
            }
 
            // Synchronize compute and DM cores for next tree level
            snrt_cluster_hw_barrier();
        }
    }
}
 
//================================================================================
// Memory consistency
//================================================================================
 
inline void snrt_wait_writeback(uint32_t val) {
    asm volatile("mv %0, %0" : "+r"(val)::);
}
 
//================================================================================
// Multicast functions
//================================================================================
 
inline void snrt_enable_multicast(uint32_t mask) { write_csr(0x7c4, mask); }
 
inline void snrt_disable_multicast() { write_csr(0x7c4, 0); }