Block 层扩展开发指南

【免费下载链接】cannbot-skillsCANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体，本仓库为其提供可复用的 Skills 模块。项目地址: https://gitcode.com/cann/cannbot-skills

适用路径：blaze_custom（路径 A）

§1 BlockMmad 接口规范

模板签名（8 参数 + SFINAE，block_mmad.h）

template <class DispatchPolicy_, class AType_, class LayoutA_, class BType_, class LayoutB_, class CType_, class LayoutC_, class Enable = void> class BlockMmad;

blaze 库为 10 参数（多BiasType_+LayoutBias_），见references/fundamentals/blaze-framework-overview.md§4.4。

必需内部类型

using L0CType = conditional_t<is_same_v<AType, int8_t>, int32_t, float>; static constexpr bool transA = TagToTrans<LayoutA>::value; struct Params { GM_ADDR aGmAddr; GM_ADDR bGmAddr; GM_ADDR cGmAddr; }; struct L1Params { uint64_t kL1; };

接口合约

§2 K-L1/K-L0 双层循环骨架

for (uint64_t iter0 = 0; iter0 < kL1Iter_; ++iter0) { uint64_t l1BufId = abL1LoopCnt_ & L1_BUFFER_MASK; auto curKL1 = (iter0 + 1 == kL1Iter_) ? (k_ - iter0 * kL1_) : kL1_; WaitFlag<HardEvent::MTE1_MTE2>(l1BufId); Copy(copyGM2L1, tensorAL1, gmA.Slice(...)); Copy(copyGM2L1, tensorBL1, gmB.Slice(...)); SetFlag<HardEvent::MTE2_MTE1>(l1BufId); WaitFlag<HardEvent::MTE2_MTE1>(l1BufId); for (uint16_t iter1 = 0; iter1 < CeilDiv(curKL1, baseK_); ++iter1) { uint64_t l0BufId = l0PingPong_ & 0x1; WaitFlag<HardEvent::M_MTE1>(l0BufId); Copy(copyL12L0A, tensorAL0, tensorBlockAL1); Copy(copyL12L0B, tensorBL0, tensorBlockBL1); SetFlag<HardEvent::MTE1_M>(l0BufId); WaitFlag<HardEvent::MTE1_M>(l0BufId); bool isLast = (iter0 + 1 == kL1Iter_) && (iter1 + 1 == kL0Iter); bool isFirst = (iter0 == 0 && iter1 == 0); MmadParams params{(uint16_t)curM, (uint16_t)curN, (uint16_t)curKL0, isLast ? FINAL_ACCUMULATION : NON_FINAL_ACCUMULATION, isFirst}; Mmad(mmadAtom, tensorL0C, tensorAL0, tensorBL0); SetFlag<HardEvent::M_MTE1>(l0BufId); l0PingPong_++; } SetFlag<HardEvent::MTE1_MTE2>(l1BufId); abL1LoopCnt_++; }

§3 事件管理（Init-Drain 模式）

构造预发、析构排空，确保事件平衡：

BlockMmad() { for (uint8_t i = 0; i < L1_BUFFER_NUM; ++i) SetFlag<HardEvent::MTE1_MTE2>(i); SetFlag<HardEvent::M_MTE1>(ZERO_FLAG); SetFlag<HardEvent::M_MTE1>(FIRST_FLAG); SetMMLayoutTransform(true); } ~BlockMmad() { for (uint8_t i = 0; i < L1_BUFFER_NUM; ++i) WaitFlag<HardEvent::MTE1_MTE2>(i); WaitFlag<HardEvent::M_MTE1>(ZERO_FLAG); WaitFlag<HardEvent::M_MTE1>(FIRST_FLAG); SetMMLayoutTransform(false); }

构造SetFlag与析构WaitFlag必须完全对称。遗漏析构 WaitFlag → 第二次 kernel 调用 hang。

完整配对规则 →references/fundamentals/blaze-sync-patterns.md§3

§4 tensor_api 必要知识

MakeCopy / Copy（dst 在前）

auto copyGM2L1 = MakeCopy(CopyGM2L1{}); Copy(copyGM2L1, dstL1Tensor, srcGMTensor);

13 种 CopyOperation：CopyGM2L1、CopyL12L0A、CopyL12L0B、CopyL0C2GM、CopyL0C2UB等。Routing 按(dstLoc, srcLoc, dstPattern, srcPattern)编译期派发。

MakeMmad / Mmad

MmadParams params{m, n, k, unitFlag, cmatrixInitVal}; Mmad(MmadAtom<MmadTraits<MmadOperation>>{}.with(params), dstL0C, srcL0A, srcL0B);

unitFlag：FINAL_ACCUMULATION（K 末条）/NON_FINAL_ACCUMULATION。cmatrixInitVal：true= 清零再累加（K 首条）。

完整 API →references/fundamentals/tensor-api-reference.md§2.4-§2.5

§5 Layout 与数据通路

L1 双缓冲布局

L1: [A0|B0] | [A1|B1] （half0 / half1 ping-pong）

偏移：l1BufferAOffset_[bufId] = (bufId & 1) * l1HalfSize + (bufId >> 1) * aL1OneBuffer_

GM→L1 数据通路

L0 固定：L0A =NZLayoutPtn，L0B =ZNLayoutPtn，C0 = 16。

Layout 详解 →references/fundamentals/blaze-matmul-layout.md

【免费下载链接】cannbot-skillsCANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体，本仓库为其提供可复用的 Skills 模块。项目地址: https://gitcode.com/cann/cannbot-skills