* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
#include "ccu_kernel_broadcast_mesh1d.h"
namespace ops_hccl {
constexpr int OUTPUT_XN_ID = 1;
constexpr int TOKEN_XN_ID = 2;
constexpr int CKE_IDX_0 = 0;
constexpr int POST_SYNC_ID = 3;
static CcuResult ParseKernelArg(BroadcastMesh1DContext &ctx, CcuKernelArgBroadcastMesh1D *kernelArg)
{
ctx.arg = kernelArg;
return CCU_SUCCESS;
}
static CcuResult InitResource(BroadcastMesh1DContext &ctx)
{
const auto *arg = ctx.arg;
uint32_t channelIdx = 0;
if (arg->channelCount == 0) {
HCCL_ERROR("[CcuKernelBroadcastMesh1D] channels is empty!");
return CcuResult::CCU_E_INTERNAL;
}
HCCL_INFO("[CcuKernelBroadcastMesh1D] channels.size: [%u]", arg->channelCount);
ctx.output.resize(arg->rankSize);
ctx.token.resize(arg->rankSize);
for (uint64_t peerId = 0; peerId < arg->rankSize; peerId++) {
if (peerId != arg->rankId) {
ctx.output[peerId] = ccu::GetResByChannel<ccu::Variable>(arg->channels[channelIdx], OUTPUT_XN_ID);
ctx.token[peerId] = ccu::GetResByChannel<ccu::Variable>(arg->channels[channelIdx], TOKEN_XN_ID);
channelIdx++;
}
}
ctx.resourceAllocated = false;
return CCU_SUCCESS;
}
static CcuResult LoadArgs(BroadcastMesh1DContext &ctx)
{
const auto *arg = ctx.arg;
uint32_t argId = 0;
CCU_CHK_RET(ccu::LoadArg(ctx.input, argId++));
CCU_CHK_RET(ccu::LoadArg(ctx.output[arg->rankId], argId++));
CCU_CHK_RET(ccu::LoadArg(ctx.token[arg->rankId], argId++));
CCU_CHK_RET(ccu::LoadArg(ctx.offset, argId++));
CCU_CHK_RET(ccu::LoadArg(ctx.goSize.addrOffset, argId++));
CCU_CHK_RET(ccu::LoadArg(ctx.goSize.loopParam, argId++));
CCU_CHK_RET(ccu::LoadArg(ctx.goSize.parallelParam, argId++));
CCU_CHK_RET(ccu::LoadArg(ctx.goSize.residual, argId++));
return CCU_SUCCESS;
}
static CcuResult PreSync(BroadcastMesh1DContext &ctx)
{
const auto *arg = ctx.arg;
for (uint32_t i = 0; i < arg->channelCount; i++) {
CCU_CHK_RET(ccu::WriteVariableWithNotify(arg->channels[i], ctx.output[arg->rankId],
OUTPUT_XN_ID, CKE_IDX_0, 1 << OUTPUT_XN_ID));
CCU_CHK_RET(ccu::WriteVariableWithNotify(arg->channels[i], ctx.token[arg->rankId],
TOKEN_XN_ID, CKE_IDX_0, 1 << TOKEN_XN_ID));
}
uint32_t allBit = (1 << OUTPUT_XN_ID) | (1 << TOKEN_XN_ID);
for (uint32_t i = 0; i < arg->channelCount; i++) {
CCU_CHK_RET(ccu::NotifyWait(arg->channels[i], CKE_IDX_0, allBit));
}
return CCU_SUCCESS;
}
static CcuResult PostSync(BroadcastMesh1DContext &ctx)
{
const auto *arg = ctx.arg;
for (uint32_t i = 0; i < arg->channelCount; i++) {
ccu::NotifyRecord(arg->channels[i], CKE_IDX_0, 1 << POST_SYNC_ID);
}
for (uint32_t i = 0; i < arg->channelCount; i++) {
ccu::NotifyWait(arg->channels[i], CKE_IDX_0, 1 << POST_SYNC_ID);
}
return CCU_SUCCESS;
}
static CcuResult DoBroadcast(BroadcastMesh1DContext &ctx)
{
const auto *arg = ctx.arg;
ccu::LocalAddr src;
ccu::LocalAddr localdst;
std::vector<ccu::RemoteAddr> dst;
dst.resize(arg->rankSize - 1);
src.addr = ctx.input;
src.token = ctx.token[arg->rankId];
uint32_t dstId = 0;
uint32_t curId = 0;
for (uint64_t rankIdx = 0; rankIdx < arg->rankSize; rankIdx++) {
if (rankIdx != arg->rootId) {
curId = dstId;
dst[curId].addr = ctx.output[rankIdx];
dst[curId].addr += ctx.offset;
dst[curId].token = ctx.token[rankIdx];
dstId++;
} else {
localdst.addr = ctx.output[rankIdx];
localdst.addr += ctx.offset;
localdst.token = ctx.token[rankIdx];
}
}
CCU_CHK_RET(GroupBroadcast(ctx, arg->channels, arg->channelCount,
localdst, dst, src, ctx.goSize));
return CCU_SUCCESS;
}
CcuResult CcuBroadcastMesh1DKernel(CcuKernelArg arg)
{
auto *kernelArg = static_cast<CcuKernelArgBroadcastMesh1D *>(arg);
BroadcastMesh1DContext ctx;
ctx.resourceAllocated = false;
ctx.moConfig.msInterleave = 0;
ctx.moConfig.loopCount = 0;
ctx.moConfig.memSlice = 0;
ctx.moRes.eventCount = 0;
ctx.moRes.bufCount = 0;
ctx.enginePool = 0;
HCCL_INFO("[CcuKernelBroadcastMesh1D] BroadcastMesh1D run");
CCU_CHK_RET(ParseKernelArg(ctx, kernelArg));
CCU_CHK_RET(InitResource(ctx));
CCU_CHK_RET(LoadArgs(ctx));
CCU_CHK_RET(PreSync(ctx));
if (ctx.arg->rankId == ctx.arg->rootId) {
CCU_CHK_RET(DoBroadcast(ctx));
}
CCU_CHK_RET(PostSync(ctx));
HCCL_INFO("[CcuKernelBroadcastMesh1D] BroadcastMesh1D end");
return CCU_SUCCESS;
}
}