信号与同步设计 | CANN/pto-isa

信号与同步设计【免费下载链接】pto-isaParallel Tile Operation (PTO) is a virtual instruction set architecture designed by Ascend CANN, focusing on tile-level operations. This repository offers high-performance, cross-platform tile operations across Ascend platforms.项目地址: https://gitcode.com/cann/pto-isa信号矩阵布局通信算子通常需要一个信号矩阵来协调多 rank 同步。典型布局信号矩阵每个 rank 一份存放在通信窗口/对称内存中 ┌─────────────────────────────────────────────────────┐ │ [0..MAX_RANKS-1] 跨 rank 阶段计数器RS/AG done│ │ [MAX_RANKS] 本地广播标志block 0 → all │ │ [MAX_RANKS1] intra-rank block 到达计数器 │ │ ...按需扩展更多阶段 │ └─────────────────────────────────────────────────────┘设计原则每个 phase 使用独立的信号区域避免信号复用导致竞争使用NotifyOp::AtomicAdd实现计数型同步多方通知一方使用NotifyOp::Set实现标志型同步一方通知多方跨 Rank Barrier 模式AICORE inline void DeviceBarrier(__gm__ DeviceContext *ctx, __gm__ int32_t *signal_base, int phase, int my_rank, int nranks, int block_idx, int num_comm_blocks) { // 步骤 1Intra-rank barrier所有 block 必须到达 __gm__ int32_t *intra_counter signal_base INTRA_RANK_OFFSET phase; if (block_idx ! 0) { comm::Signal arrSig(intra_counter); comm::TNOTIFY(arrSig, 1, comm::NotifyOp::AtomicAdd); } else { if (num_comm_blocks 1) { comm::Signal arrSig(intra_counter); comm::TWAIT(arrSig, num_comm_blocks - 1, comm::WaitCmp::GE); } } pipe_barrier(PIPE_ALL); // 步骤 2Cross-rank barrier仅 block 0 执行 if (block_idx 0) { for (int r 0; r nranks; r) { if (r my_rank) continue; __gm__ int32_t *remote_sig GetRemotePtr(ctx, signal_base my_rank, r); comm::Signal sig(remote_sig); comm::TNOTIFY(sig, 1, comm::NotifyOp::AtomicAdd); } for (int r 0; r nranks; r) { if (r my_rank) continue; comm::Signal sig(signal_base r); comm::TWAIT(sig, 1, comm::WaitCmp::GE); } __gm__ int32_t *local_flag signal_base LOCAL_FLAG_OFFSET phase; comm::Signal localSig(local_flag); comm::TNOTIFY(localSig, 1, comm::NotifyOp::Set); } else { __gm__ int32_t *local_flag signal_base LOCAL_FLAG_OFFSET phase; comm::Signal localSig(local_flag); comm::TWAIT(localSig, 1, comm::WaitCmp::GE); } pipe_barrier(PIPE_ALL); }Intra-kernel 同步模式通信 kernel 内部不同阶段之间使用pipe_barrier(PIPE_ALL)分隔// 阶段 1ReduceScatter ReduceScatterPhase(...); pipe_barrier(PIPE_ALL); // 阶段 2跨 rank barrier DeviceBarrier(...); // 阶段 3AllGather AllGatherPhase(...); pipe_barrier(PIPE_ALL);手动流水线同步当使用底层TLOAD/TSTORE_IMPL而非高层TPUT/TGET时需要手动管理流水线同步TLOAD(tile, srcG); set_flag(PIPE_MTE2, PIPE_MTE3, EVENT_ID0); wait_flag(PIPE_MTE2, PIPE_MTE3, EVENT_ID0); TSTORE_IMPLTileData, Global, AtomicType::AtomicNone(dstG, tile); set_flag(PIPE_MTE3, PIPE_MTE2, EVENT_ID0); wait_flag(PIPE_MTE3, PIPE_MTE2, EVENT_ID0);注意TPUT/TGET 内部已处理流水线同步直接使用时无需手动 set_flag/wait_flag。仅在使用底层TLOAD/TSTORE_IMPL构建自定义流水线时才需要。【免费下载链接】pto-isaParallel Tile Operation (PTO) is a virtual instruction set architecture designed by Ascend CANN, focusing on tile-level operations. This repository offers high-performance, cross-platform tile operations across Ascend platforms.项目地址: https://gitcode.com/cann/pto-isa创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考