From: @zhangxiaokun9 Reviewed-by: @wangxiaotian22,@xchu42,@wangxiaotian22,@xchu42 Signed-off-by: @ji_chentags/v1.2.0
@@ -108,6 +108,7 @@ std::mutex DavinciModel::tvm_bin_mutex_; | |||
DavinciModel::DavinciModel(int32_t priority, const std::shared_ptr<ModelListener> &listener) | |||
: weights_mem_base_(nullptr), | |||
var_mem_base_(nullptr), | |||
fixed_mem_base_(0), | |||
mem_base_(nullptr), | |||
is_inner_mem_base_(false), | |||
is_inner_weight_base_(false), | |||
@@ -139,6 +140,7 @@ DavinciModel::DavinciModel(int32_t priority, const std::shared_ptr<ModelListener | |||
is_l1_fusion_enable_(false), | |||
is_first_execute_(true) { | |||
op_list_.clear(); | |||
skt_info_ = {0, 0, 0, 0, nullptr, nullptr, {}, {}, {}, {}, {}, RT_KERNEL_DEFAULT, -1, 0, nullptr}; | |||
} | |||
DavinciModel::~DavinciModel() { | |||
@@ -261,6 +263,7 @@ Status DavinciModel::Assign(const GeModelPtr &ge_model) { | |||
/// @return: void | |||
/// | |||
void DavinciModel::Shrink() { | |||
skt_info_ = {0, 0, 0, 0, nullptr, nullptr, {}, {}, {}, {}, {}, RT_KERNEL_DEFAULT, -1, 0, nullptr}; | |||
ge_model_.reset(); // delete object. | |||
} | |||
@@ -668,6 +671,7 @@ Status DavinciModel::Init(void *dev_ptr, size_t mem_size, void *weight_ptr, size | |||
data_inputer_ = new (std::nothrow) DataInputer(); | |||
GE_CHK_BOOL_RET_STATUS(data_inputer_ != nullptr, MEMALLOC_FAILED, "data_inputer_ is nullptr."); | |||
} | |||
fixed_mem_base_ = reinterpret_cast<uintptr_t>(mem_base_); | |||
GE_TIMESTAMP_END(InitModelMem, "GraphLoader::InitModelMem"); | |||
for (const ge::NodePtr &node : compute_graph->GetDirectNode()) { | |||
@@ -2826,7 +2830,32 @@ Status DavinciModel::CreateKnownZeroCopyMap(const vector<void *> &inputs, const | |||
return SUCCESS; | |||
} | |||
void DavinciModel::SetTotalIOAddrs(const vector<void *> &io_addrs) { | |||
if (fixed_mem_base_ == reinterpret_cast<uintptr_t>(mem_base_)) { | |||
total_io_addrs_.insert(total_io_addrs_.end(), io_addrs.begin(), io_addrs.end()); | |||
return; | |||
} | |||
for (size_t i = 0; i < io_addrs.size(); ++i) { | |||
uintptr_t addr = reinterpret_cast<uintptr_t>(io_addrs[i]); | |||
if ((fixed_mem_base_ <= addr) && (addr < fixed_mem_base_ + runtime_param_.mem_size)) { | |||
total_io_addrs_.emplace_back(mem_base_ + (addr - fixed_mem_base_)); | |||
} else { | |||
total_io_addrs_.emplace_back(io_addrs[i]); | |||
} | |||
} | |||
} | |||
Status DavinciModel::UpdateKnownZeroCopyAddr(vector<void *> &total_io_addrs) { | |||
if (fixed_mem_base_ != reinterpret_cast<uintptr_t>(mem_base_)) { | |||
for (size_t i = 0; i < total_io_addrs.size(); ++i) { | |||
uintptr_t addr = reinterpret_cast<uintptr_t>(total_io_addrs[i]); | |||
if ((fixed_mem_base_ <= addr) && (addr < fixed_mem_base_ + runtime_param_.mem_size)) { | |||
total_io_addrs[i] = mem_base_ + (addr - fixed_mem_base_); | |||
} | |||
} | |||
} | |||
for (size_t i = 0; i < total_io_addrs.size(); ++i) { | |||
auto it_in = knonw_input_data_info_.find(total_io_addrs[i]); | |||
if (it_in != knonw_input_data_info_.end()) { | |||
@@ -76,6 +76,25 @@ struct timeInfo { | |||
int64_t dumpEndTime; | |||
}; | |||
// For super kernel | |||
struct SuperKernelTaskInfo { | |||
uint32_t last_block_dim; | |||
uint32_t last_args_size; | |||
uint32_t last_task_id; | |||
uint32_t last_stream_id; | |||
void *last_stream; | |||
void *last_sm_desc; | |||
std::vector<void *> kernel_list; | |||
std::vector<void *> arg_list; | |||
std::vector<uint32_t> dump_flag_list; | |||
std::vector<OpDescPtr> op_desc_list; | |||
std::vector<uintptr_t> dump_args_list; | |||
uint32_t last_dump_flag; | |||
int64_t last_group_key; | |||
uintptr_t last_dump_args; | |||
OpDescPtr last_op; | |||
}; | |||
struct TaskMemInfo { | |||
int64_t input_size{0}; | |||
int64_t output_size{0}; | |||
@@ -261,6 +280,9 @@ class DavinciModel { | |||
// get updated task info list | |||
std::vector<TaskInfoPtr> GetTaskList() { return task_list_; } | |||
// Modified from KernelTaskInfo. | |||
SuperKernelTaskInfo &GetSuperKernelTaskInfo() { return skt_info_; } | |||
/// | |||
/// @ingroup ge | |||
/// @brief get model input and output format | |||
@@ -481,9 +503,7 @@ class DavinciModel { | |||
void *cur_args = static_cast<char *>(args_) + offset; | |||
return cur_args; | |||
} | |||
void SetTotalIOAddrs(vector<void *> &io_addrs) { | |||
total_io_addrs_.insert(total_io_addrs_.end(), io_addrs.begin(), io_addrs.end()); | |||
} | |||
void SetTotalIOAddrs(const vector<void *> &io_addrs); | |||
void SetHybridArgsSize(uint32_t args_size) { total_hybrid_args_size_ += args_size; } | |||
uint32_t GetHybridArgsSize() { | |||
return total_hybrid_args_size_; | |||
@@ -533,6 +553,7 @@ class DavinciModel { | |||
uint8_t *weights_mem_base_; | |||
uint8_t *var_mem_base_; | |||
// memory address of model | |||
uintptr_t fixed_mem_base_; // Initial of mem_base_, keep forever. | |||
uint8_t *mem_base_; | |||
uint8_t *p2p_mem_base_; | |||
bool is_inner_mem_base_; | |||
@@ -996,6 +1017,9 @@ class DavinciModel { | |||
std::multimap<uint32_t, uint32_t> op_id_map_; | |||
std::vector<ProfileInfo> profile_list_; | |||
// For super kernel. | |||
SuperKernelTaskInfo skt_info_; | |||
}; | |||
} // namespace ge | |||
#endif // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_DAVINCI_MODEL_H_ |
@@ -59,40 +59,40 @@ Status HcclTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_m | |||
GELOGI("HcclTaskInfo Init, op_index is: %u", op_index); | |||
// Get HCCL op | |||
op_desc_ = davinci_model->GetOpByIndex(op_index); | |||
GE_CHECK_NOTNULL(op_desc_); | |||
const auto op_desc = davinci_model_->GetOpByIndex(op_index); | |||
GE_CHECK_NOTNULL(op_desc); | |||
// Create the kernel hccl infos | |||
CreateKernelHcclInfo(op_desc_); | |||
CreateKernelHcclInfo(op_desc); | |||
// Initialize the hccl_type of all kernel hccl info | |||
HcomOmeUtil::GetHcclType(task_def, kernel_hccl_infos_); | |||
// Only in Horovod scenario should get the inputName and GeShape | |||
ret = HcomOmeUtil::GetHorovodInputs(op_desc_, kernel_hccl_infos_); | |||
ret = HcomOmeUtil::GetHorovodInputs(op_desc, kernel_hccl_infos_); | |||
if (ret != SUCCESS) { | |||
GELOGE(ret, "davinci_model: GetHorovodInputs fail! domi error: %u", ret); | |||
return ret; | |||
} | |||
Status dmrt = HcomOmeUtil::GetHcclDataType(op_desc_, kernel_hccl_infos_); | |||
Status dmrt = HcomOmeUtil::GetHcclDataType(op_desc, kernel_hccl_infos_); | |||
if (dmrt != SUCCESS) { | |||
GELOGE(dmrt, "davinci_model: GetHcomDataType fail! domi error: %u", dmrt); | |||
return dmrt; | |||
} | |||
dmrt = HcomOmeUtil::GetHcclCount(op_desc_, kernel_hccl_infos_); | |||
dmrt = HcomOmeUtil::GetHcclCount(op_desc, kernel_hccl_infos_); | |||
if (dmrt != SUCCESS) { | |||
GELOGE(dmrt, "davinci_model: GetHcomCount fail! domi error: %u", dmrt); | |||
return dmrt; | |||
} | |||
// Only HCOMBROADCAST and HVDCALLBACKBROADCAST need to get the rootId | |||
dmrt = HcomOmeUtil::GetAllRootId(op_desc_, kernel_hccl_infos_); | |||
dmrt = HcomOmeUtil::GetAllRootId(op_desc, kernel_hccl_infos_); | |||
if (dmrt != SUCCESS) { | |||
GELOGE(dmrt, "davinci_model: Get rootId fail! domi error: %u", dmrt); | |||
return dmrt; | |||
} | |||
// GE's new process: hccl declares the number of streams required, creates a stream by GE, and sends it to hccl | |||
ret = SetFollowStream(op_desc_, davinci_model); | |||
ret = SetFollowStream(op_desc, davinci_model); | |||
if (ret != SUCCESS) { | |||
GELOGE(ret, "SetStream Fail."); | |||
return ret; | |||
@@ -100,21 +100,22 @@ Status HcclTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_m | |||
if (davinci_model_->IsKnownNode()) { | |||
args_ = davinci_model_->GetCurrentArgsAddr(args_offset_); | |||
GELOGI("Known node %s args addr %p, offset %u.", op_desc_->GetName().c_str(), args_, args_offset_); | |||
GELOGI("Known node %s args addr %p, offset %u.", op_desc->GetName().c_str(), args_, args_offset_); | |||
} | |||
ret = SetAddrs(op_desc_, kernel_hccl_infos_); | |||
ret = SetAddrs(op_desc, kernel_hccl_infos_); | |||
if (ret != SUCCESS) { | |||
GELOGE(ret, "Setaddrs Fail."); | |||
return ret; | |||
} | |||
// GE's new process: hccl declares the need for Workspace size, and GE allocates Workspace | |||
ret = SetWorkspace(op_desc_, kernel_hccl_infos_); | |||
ret = SetWorkspace(op_desc, kernel_hccl_infos_); | |||
if (ret != SUCCESS) { | |||
GELOGE(ret, "SetWorkspace Fail."); | |||
return ret; | |||
} | |||
SetIoAddrs(op_desc); | |||
GELOGI("HcclTaskInfo Init Success"); | |||
return SUCCESS; | |||
} | |||
@@ -229,20 +230,19 @@ Status HcclTaskInfo::CalculateArgs(const domi::TaskDef &task_def, DavinciModel * | |||
return SUCCESS; | |||
} | |||
Status HcclTaskInfo::UpdateArgs() { | |||
GELOGI("HcclTaskInfo::UpdateArgs in."); | |||
void HcclTaskInfo::SetIoAddrs(const OpDescPtr &op_desc) { | |||
const RuntimeParam &rts_param = davinci_model_->GetRuntimeParam(); | |||
input_data_addrs_ = ModelUtils::GetInputDataAddrs(rts_param, op_desc_); | |||
output_data_addrs_ = ModelUtils::GetOutputDataAddrs(rts_param, op_desc_); | |||
workspace_data_addrs_ = ModelUtils::GetWorkspaceDataAddrs(rts_param, op_desc_); | |||
vector<void *> io_addrs; | |||
io_addrs.insert(io_addrs.end(), input_data_addrs_.begin(), input_data_addrs_.end()); | |||
io_addrs.insert(io_addrs.end(), output_data_addrs_.begin(), output_data_addrs_.end()); | |||
io_addrs.insert(io_addrs.end(), workspace_data_addrs_.begin(), workspace_data_addrs_.end()); | |||
davinci_model_->SetTotalIOAddrs(io_addrs); | |||
const auto input_data_addrs = ModelUtils::GetInputDataAddrs(rts_param, op_desc); | |||
const auto output_data_addrs = ModelUtils::GetOutputDataAddrs(rts_param, op_desc); | |||
const auto workspace_data_addrs = ModelUtils::GetWorkspaceDataAddrs(rts_param, op_desc); | |||
io_addrs_.insert(io_addrs_.end(), input_data_addrs.begin(), input_data_addrs.end()); | |||
io_addrs_.insert(io_addrs_.end(), output_data_addrs.begin(), output_data_addrs.end()); | |||
io_addrs_.insert(io_addrs_.end(), workspace_data_addrs.begin(), workspace_data_addrs.end()); | |||
} | |||
Status HcclTaskInfo::UpdateArgs() { | |||
GELOGI("HcclTaskInfo::UpdateArgs in."); | |||
davinci_model_->SetTotalIOAddrs(io_addrs_); | |||
GELOGI("HcclTaskInfo::UpdateArgs success."); | |||
return SUCCESS; | |||
} | |||
@@ -261,9 +261,11 @@ Status HcclTaskInfo::SetAddrs(const std::shared_ptr<OpDesc> &op_desc, | |||
HcclReduceOp op_type = HCCL_REDUCE_SUM; | |||
GE_CHECK_NOTNULL(davinci_model_); | |||
GELOGI("Calc opType[%s] input address before. Node name[%s]", op_desc->GetType().c_str(), op_desc->GetName().c_str()); | |||
vector<void *> input_data_addrs; | |||
vector<void *> output_data_addrs; | |||
if (!davinci_model_->IsKnownNode()) { | |||
input_data_addrs_ = ModelUtils::GetInputDataAddrs(davinci_model_->GetRuntimeParam(), op_desc); | |||
output_data_addrs_ = ModelUtils::GetOutputDataAddrs(davinci_model_->GetRuntimeParam(), op_desc); | |||
input_data_addrs = ModelUtils::GetInputDataAddrs(davinci_model_->GetRuntimeParam(), op_desc); | |||
output_data_addrs = ModelUtils::GetOutputDataAddrs(davinci_model_->GetRuntimeParam(), op_desc); | |||
} | |||
void *input_data_addr = nullptr; | |||
void *output_data_addr = nullptr; | |||
@@ -275,8 +277,8 @@ Status HcclTaskInfo::SetAddrs(const std::shared_ptr<OpDesc> &op_desc, | |||
output_data_addr = reinterpret_cast<void *>(reinterpret_cast<uint64_t *>(args_) + op_desc->GetInputsSize() + i); | |||
GELOGI("Hccl task info known input addr %p, output addr %p.", input_data_addr, output_data_addr); | |||
} else { | |||
input_data_addr = input_data_addrs_.empty() ? nullptr : input_data_addrs_[i]; | |||
output_data_addr = output_data_addrs_.empty() ? nullptr : output_data_addrs_[i]; | |||
input_data_addr = input_data_addrs.empty() ? nullptr : input_data_addrs[i]; | |||
output_data_addr = output_data_addrs.empty() ? nullptr : output_data_addrs[i]; | |||
} | |||
kernel_hccl_infos[i].inputDataAddr = input_data_addr; | |||
if (hccl_type == HCOMALLGATHER || hccl_type == HCOMRECEIVE || hccl_type == HVDCALLBACKALLGATHER) { | |||
@@ -366,8 +368,8 @@ Status HcclTaskInfo::SetWorkspace(const std::shared_ptr<OpDesc> &op_desc, | |||
workspace_addr = reinterpret_cast<void *>(reinterpret_cast<uint64_t *>(args_) + op_desc->GetInputsSize() + | |||
op_desc->GetOutputsSize()); | |||
} else { | |||
workspace_data_addrs_ = ModelUtils::GetWorkspaceDataAddrs(davinci_model_->GetRuntimeParam(), op_desc); | |||
workspace_addr = workspace_data_addrs_.empty() ? nullptr : workspace_data_addrs_[0]; | |||
const auto workspace_data_addrs = ModelUtils::GetWorkspaceDataAddrs(davinci_model_->GetRuntimeParam(), op_desc); | |||
workspace_addr = workspace_data_addrs.empty() ? nullptr : workspace_data_addrs[0]; | |||
} | |||
} | |||
} | |||
@@ -35,7 +35,6 @@ class HcclTaskInfo : public TaskInfo { | |||
ops_kernel_store_(nullptr), | |||
private_def_(nullptr), | |||
private_def_len_(0), | |||
op_desc_(nullptr), | |||
args_(nullptr), | |||
args_offset_(0) {} | |||
@@ -52,7 +51,7 @@ class HcclTaskInfo : public TaskInfo { | |||
Status UpdateArgs() override; | |||
private: | |||
ge::Status SetAddrs(const std::string &hccl_type, const std::shared_ptr<OpDesc> &op); | |||
void SetIoAddrs(const OpDescPtr &op_desc); | |||
Status SetAddrs(const std::shared_ptr<OpDesc> &op_desc, std::vector<GETaskKernelHcclInfo> &kernel_hccl_infos); | |||
@@ -76,10 +75,7 @@ class HcclTaskInfo : public TaskInfo { | |||
uint32_t private_def_len_; | |||
static std::mutex hccl_follow_stream_mutex_; | |||
vector<GETaskKernelHcclInfo> kernel_hccl_infos_; | |||
vector<void *> input_data_addrs_; | |||
vector<void *> output_data_addrs_; | |||
vector<void *> workspace_data_addrs_; | |||
OpDescPtr op_desc_; | |||
vector<void *> io_addrs_; | |||
void *args_; | |||
uint32_t args_offset_; | |||
}; | |||
@@ -30,11 +30,7 @@ | |||
namespace ge { | |||
Status KernelExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_model) { | |||
GELOGI("KernelExTaskInfo Init Start."); | |||
if (davinci_model == nullptr) { | |||
GELOGE(PARAM_INVALID, "davinci_model is null!"); | |||
return PARAM_INVALID; | |||
} | |||
GE_CHECK_NOTNULL(davinci_model); | |||
davinci_model_ = davinci_model; | |||
Status ret = SetStream(task_def.stream_id(), davinci_model_->GetStreamList()); | |||
if (ret != SUCCESS) { | |||
@@ -51,7 +47,6 @@ Status KernelExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davin | |||
GELOGE(INTERNAL_ERROR, "Init aicpu task info error, index is out of range!"); | |||
return INTERNAL_ERROR; | |||
} | |||
op_desc_ = op_desc; | |||
// 2. Reconstruct kernelExDef.args to STR_FWK_OP_KERNEL | |||
STR_FWK_OP_KERNEL fwk_op_kernel = {0}; | |||
@@ -79,8 +74,8 @@ Status KernelExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davin | |||
return RT_ERROR_TO_GE_STATUS(rt_ret);) | |||
} | |||
GELOGI("Node[%s] type[%s] kernel_ext_info size=%zu, ext_info_addr_=%p", op_desc_->GetName().c_str(), | |||
op_desc_->GetType().c_str(), ext_info.size(), ext_info_addr_); | |||
GELOGI("Node[%s] type[%s] kernel_ext_info size=%zu, ext_info_addr_=%p", op_desc->GetName().c_str(), | |||
op_desc->GetType().c_str(), ext_info.size(), ext_info_addr_); | |||
// 2.1 get loop cond variable for tensor array write | |||
uint64_t step_id_addr = 0; | |||
@@ -132,6 +127,7 @@ Status KernelExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davin | |||
GE_IF_BOOL_EXEC(rt_ret != RT_ERROR_NONE, GELOGE(RT_FAILED, "rtMemcpy error, ret: Ox%X", rt_ret); | |||
return RT_ERROR_TO_GE_STATUS(rt_ret);) | |||
SetIoAddrs(op_desc); | |||
GELOGI("KernelExTaskInfo knonw node Init Success."); | |||
return SUCCESS; | |||
} | |||
@@ -195,7 +191,7 @@ Status KernelExTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davin | |||
return RT_ERROR_TO_GE_STATUS(rt_ret);) | |||
davinci_model_->SetZeroCopyAddr(op_desc, io_addrs, io_addrs.data(), input_output_addr_, addrs_size, 0); | |||
SetIoAddrs(op_desc); | |||
GELOGI("KernelExTaskInfo Init Success. session id: %lu", session_id); | |||
return SUCCESS; | |||
} | |||
@@ -236,36 +232,38 @@ Status KernelExTaskInfo::CalculateArgs(const domi::TaskDef &task_def, DavinciMod | |||
return SUCCESS; | |||
} | |||
Status KernelExTaskInfo::UpdateArgs() { | |||
GELOGI("KernelExTaskInfo::UpdateArgs in."); | |||
void KernelExTaskInfo::SetIoAddrs(const OpDescPtr &op_desc) { | |||
const RuntimeParam &rts_param = davinci_model_->GetRuntimeParam(); | |||
vector<void *> input_data_addrs = ModelUtils::GetInputDataAddrs(rts_param, op_desc_); | |||
vector<void *> output_data_addrs = ModelUtils::GetOutputDataAddrs(rts_param, op_desc_); | |||
vector<void *> io_addrs; | |||
if (!op_desc_->HasAttr(ATTR_DYNAMIC_SHAPE_FIXED_ADDR)) { | |||
io_addrs.insert(io_addrs.end(), input_data_addrs.begin(), input_data_addrs.end()); | |||
io_addrs.insert(io_addrs.end(), output_data_addrs.begin(), output_data_addrs.end()); | |||
vector<void *> input_data_addrs = ModelUtils::GetInputDataAddrs(rts_param, op_desc); | |||
vector<void *> output_data_addrs = ModelUtils::GetOutputDataAddrs(rts_param, op_desc); | |||
if (!op_desc->HasAttr(ATTR_DYNAMIC_SHAPE_FIXED_ADDR)) { | |||
io_addrs_.insert(io_addrs_.end(), input_data_addrs.begin(), input_data_addrs.end()); | |||
io_addrs_.insert(io_addrs_.end(), output_data_addrs.begin(), output_data_addrs.end()); | |||
} else { | |||
string peer_input_name; | |||
if (AttrUtils::GetStr(op_desc_, ATTR_DYNAMIC_SHAPE_FIXED_ADDR, peer_input_name)) { | |||
if (AttrUtils::GetStr(op_desc, ATTR_DYNAMIC_SHAPE_FIXED_ADDR, peer_input_name)) { | |||
uint32_t output_index = davinci_model_->GetFixedAddrOutputIndex(peer_input_name); | |||
if (output_index > output_data_addrs.size()) { | |||
GELOGE(FAILED, "The output data addr size[%zu] and output index[%u] are inconsistent.", | |||
output_data_addrs.size(), output_index); | |||
return FAILED; | |||
return; | |||
} | |||
io_addrs.insert(io_addrs.end(), input_data_addrs.begin(), input_data_addrs.end()); | |||
io_addrs_.insert(io_addrs_.end(), input_data_addrs.begin(), input_data_addrs.end()); | |||
for (size_t i = 0; i < output_data_addrs.size(); ++i) { | |||
if (i == output_index) { | |||
void *fixed_addr = davinci_model_->GetCurrentFixedAddr(fixed_addr_offset_); | |||
io_addrs.emplace_back(fixed_addr); | |||
io_addrs_.emplace_back(fixed_addr); | |||
continue; | |||
} | |||
io_addrs.emplace_back(output_data_addrs[i]); | |||
io_addrs_.emplace_back(output_data_addrs[i]); | |||
} | |||
} | |||
} | |||
davinci_model_->SetTotalIOAddrs(io_addrs); | |||
} | |||
Status KernelExTaskInfo::UpdateArgs() { | |||
GELOGI("KernelExTaskInfo::UpdateArgs in."); | |||
davinci_model_->SetTotalIOAddrs(io_addrs_); | |||
GELOGI("KernelExTaskInfo::UpdateArgs success."); | |||
return SUCCESS; | |||
} | |||
@@ -59,6 +59,7 @@ class KernelExTaskInfo : public TaskInfo { | |||
}; | |||
private: | |||
Status CopyTaskInfo(const domi::KernelExDef &kernel_def, const RuntimeParam &rts_param, const OpDescPtr &op_desc); | |||
void SetIoAddrs(const OpDescPtr &op_desc); | |||
uint32_t task_id_; | |||
uint32_t stream_id_; | |||
@@ -69,7 +70,7 @@ class KernelExTaskInfo : public TaskInfo { | |||
void *input_output_addr_; | |||
void *ext_info_addr_; | |||
void *dump_args_; | |||
OpDescPtr op_desc_ = nullptr; | |||
vector<void *> io_addrs_; | |||
uint32_t args_offset_ = 0; | |||
int64_t fixed_addr_offset_ = 0; | |||
}; | |||
@@ -53,9 +53,6 @@ const int kArgsAttrHandle = 4; | |||
} // namespace | |||
namespace ge { | |||
KernelTaskInfo::SuperKernelTaskInfo KernelTaskInfo::skt_info_ = { | |||
0, 0, 0, 0, nullptr, nullptr, {}, {}, {}, {}, {}, RT_KERNEL_DEFAULT, kInvalidGroupKey, 0, nullptr}; | |||
Status KernelTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci_model) { | |||
GE_CHECK_NOTNULL(davinci_model); | |||
davinci_model_ = davinci_model; | |||
@@ -137,6 +134,7 @@ Status KernelTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci | |||
ret = InitCceTask(kernel_def); | |||
} | |||
SetIoAddrs(op_desc_); | |||
GELOGD("KernelTaskInfo init finish, result=%u.", ret); | |||
return ret; | |||
} | |||
@@ -148,9 +146,10 @@ Status KernelTaskInfo::SaveSKTDumpInfo() { | |||
return SUCCESS; | |||
} | |||
// all op in super kernel share one taskid and streamid | |||
for (size_t i = 0; i < skt_info_.op_desc_list.size(); i++) { | |||
davinci_model_->SaveDumpTask(skt_info_.last_task_id, skt_info_.last_stream_id, skt_info_.op_desc_list[i], | |||
skt_info_.dump_args_list[i]); | |||
const SuperKernelTaskInfo &skt_info = davinci_model_->GetSuperKernelTaskInfo(); | |||
for (size_t i = 0; i < skt_info.op_desc_list.size(); i++) { | |||
davinci_model_->SaveDumpTask(skt_info.last_task_id, skt_info.last_stream_id, skt_info.op_desc_list[i], | |||
skt_info.dump_args_list[i]); | |||
} | |||
return SUCCESS; | |||
} | |||
@@ -164,9 +163,10 @@ void KernelTaskInfo::UpdateSKTTaskId() { | |||
GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret); | |||
return; | |||
} | |||
skt_info_.last_task_id = task_id; | |||
skt_info_.last_stream_id = stream_id; | |||
skt_id_ = skt_info_.last_task_id; | |||
SuperKernelTaskInfo &skt_info = davinci_model_->GetSuperKernelTaskInfo(); | |||
skt_info.last_task_id = task_id; | |||
skt_info.last_stream_id = stream_id; | |||
skt_id_ = skt_info.last_task_id; | |||
GELOGI("UpdateTaskId:UpdateSKTTaskId [%u],stream id [%u]", task_id, stream_id); | |||
} | |||
@@ -191,23 +191,25 @@ Status KernelTaskInfo::SKTFinalize() { | |||
UpdateSKTTaskId(); | |||
GE_CHK_STATUS_RET(SaveSKTDumpInfo(), "skt save dump info failed"); | |||
GELOGI("SuperKernel Distribute [skt_id:%u]", skt_id_); | |||
skt_info_.kernel_list.clear(); | |||
skt_info_.arg_list.clear(); | |||
skt_info_.dump_flag_list.clear(); | |||
skt_info_.op_desc_list.clear(); | |||
skt_info_.dump_args_list.clear(); | |||
skt_info_.last_stream = nullptr; | |||
skt_info_.last_block_dim = 0; | |||
skt_info_.last_sm_desc = sm_desc_; | |||
skt_info_.last_group_key = kInvalidGroupKey; | |||
skt_info_.last_dump_flag = RT_KERNEL_DEFAULT; | |||
skt_info_.last_dump_args = 0; | |||
skt_info_.last_op = nullptr; | |||
SuperKernelTaskInfo &skt_info = davinci_model_->GetSuperKernelTaskInfo(); | |||
skt_info.kernel_list.clear(); | |||
skt_info.arg_list.clear(); | |||
skt_info.dump_flag_list.clear(); | |||
skt_info.op_desc_list.clear(); | |||
skt_info.dump_args_list.clear(); | |||
skt_info.last_stream = nullptr; | |||
skt_info.last_block_dim = 0; | |||
skt_info.last_sm_desc = sm_desc_; | |||
skt_info.last_group_key = kInvalidGroupKey; | |||
skt_info.last_dump_flag = RT_KERNEL_DEFAULT; | |||
skt_info.last_dump_args = 0; | |||
skt_info.last_op = nullptr; | |||
return SUCCESS; | |||
} | |||
uint32_t KernelTaskInfo::GetDumpFlag() { | |||
for (auto flag : skt_info_.dump_flag_list) { | |||
const SuperKernelTaskInfo &skt_info = davinci_model_->GetSuperKernelTaskInfo(); | |||
for (auto flag : skt_info.dump_flag_list) { | |||
if (flag == RT_KERNEL_DUMPFLAG) { | |||
return RT_KERNEL_DUMPFLAG; | |||
} | |||
@@ -216,19 +218,20 @@ uint32_t KernelTaskInfo::GetDumpFlag() { | |||
} | |||
Status KernelTaskInfo::SuperKernelLaunch() { | |||
if (skt_info_.kernel_list.empty()) { | |||
const SuperKernelTaskInfo &skt_info = davinci_model_->GetSuperKernelTaskInfo(); | |||
if (skt_info.kernel_list.empty()) { | |||
GELOGI("SuperKernelLaunch: Skt_kernel_list has no task, just return"); | |||
return SUCCESS; | |||
} | |||
rtError_t rt_ret; | |||
auto &skt_kernel_list = skt_info_.kernel_list; | |||
auto &skt_arg_list = skt_info_.arg_list; | |||
auto &skt_kernel_list = skt_info.kernel_list; | |||
auto &skt_arg_list = skt_info.arg_list; | |||
GELOGI("SuperKernelLaunch: Skt_kernel_list size[%zu] skt_arg_list[%zu]", skt_kernel_list.size(), skt_arg_list.size()); | |||
if (skt_kernel_list.size() == kSKTSingleSize && skt_arg_list.size() == kSKTSingleSize) { | |||
rt_ret = rtKernelLaunchWithFlag(skt_info_.kernel_list[0], static_cast<uint32_t>(skt_info_.last_block_dim), | |||
skt_info_.arg_list[0], skt_info_.last_args_size, | |||
static_cast<rtSmDesc_t *>(skt_info_.last_sm_desc), skt_info_.last_stream, | |||
skt_info_.last_dump_flag); | |||
rt_ret = rtKernelLaunchWithFlag(skt_info.kernel_list[0], static_cast<uint32_t>(skt_info.last_block_dim), | |||
skt_info.arg_list[0], skt_info.last_args_size, | |||
static_cast<rtSmDesc_t *>(skt_info.last_sm_desc), skt_info.last_stream, | |||
skt_info.last_dump_flag); | |||
if (rt_ret != RT_ERROR_NONE) { | |||
GELOGE(RT_FAILED, "SuperKernelLaunch: Call rt api failed, ret: 0x%X", rt_ret); | |||
return RT_ERROR_TO_GE_STATUS(rt_ret); | |||
@@ -247,14 +250,14 @@ Status KernelTaskInfo::SuperKernelLaunch() { | |||
} | |||
// Call the fuse API | |||
std::unique_ptr<skt::SuperKernel> superKernel = nullptr; | |||
ge_ret = factory->FuseKernels(skt_kernel_list, skt_arg_list, skt_info_.last_block_dim, superKernel); | |||
ge_ret = factory->FuseKernels(skt_kernel_list, skt_arg_list, skt_info.last_block_dim, superKernel); | |||
if (ge_ret != SUCCESS) { | |||
GELOGE(ge_ret, "SuperKernelLaunch: fuse call failed"); | |||
return ge_ret; | |||
} | |||
// Launch a super kernel | |||
skt_dump_flag_ = GetDumpFlag(); | |||
ge_ret = superKernel->Launch(skt_info_.last_stream, skt_dump_flag_); | |||
ge_ret = superKernel->Launch(skt_info.last_stream, skt_dump_flag_); | |||
if (ge_ret != SUCCESS) { | |||
GELOGE(ge_ret, "SuperKernelLaunch: launch failed"); | |||
return ge_ret; | |||
@@ -269,23 +272,26 @@ Status KernelTaskInfo::SuperKernelLaunch() { | |||
} | |||
Status KernelTaskInfo::SaveSuperKernelInfo() { | |||
skt_info_.kernel_list.push_back(stub_func_); | |||
skt_info_.arg_list.push_back(args_); | |||
skt_info_.last_stream = stream_; | |||
skt_info_.last_block_dim = block_dim_; | |||
skt_info_.last_args_size = args_size_; | |||
skt_info_.last_sm_desc = sm_desc_; | |||
skt_info_.last_dump_flag = dump_flag_; | |||
skt_info_.dump_flag_list.push_back(dump_flag_); | |||
skt_info_.op_desc_list.push_back(op_desc_); | |||
skt_info_.dump_args_list.push_back(reinterpret_cast<uintptr_t>(skt_dump_args_)); | |||
skt_info_.last_group_key = group_key_; | |||
skt_info_.last_dump_args = reinterpret_cast<uintptr_t>(skt_dump_args_); | |||
skt_info_.last_op = op_desc_; | |||
SuperKernelTaskInfo &skt_info = davinci_model_->GetSuperKernelTaskInfo(); | |||
skt_info.kernel_list.push_back(stub_func_); | |||
skt_info.arg_list.push_back(args_); | |||
skt_info.last_stream = stream_; | |||
skt_info.last_block_dim = block_dim_; | |||
skt_info.last_args_size = args_size_; | |||
skt_info.last_sm_desc = sm_desc_; | |||
skt_info.last_dump_flag = dump_flag_; | |||
skt_info.dump_flag_list.push_back(dump_flag_); | |||
skt_info.op_desc_list.push_back(op_desc_); | |||
skt_info.dump_args_list.push_back(reinterpret_cast<uintptr_t>(skt_dump_args_)); | |||
skt_info.last_group_key = group_key_; | |||
skt_info.last_dump_args = reinterpret_cast<uintptr_t>(skt_dump_args_); | |||
skt_info.last_op = op_desc_; | |||
// last node in a stream, just launch | |||
if (IsMarkedLastNode()) { | |||
return SuperKernelLaunch(); | |||
} | |||
GELOGI("Save Current task [block_dim:%u, size:%zu].", block_dim_, skt_info.kernel_list.size()); | |||
return SUCCESS; | |||
} | |||
@@ -322,8 +328,9 @@ bool KernelTaskInfo::IsMarkedFirstNode() { | |||
// then may be saved to skt task list; else | |||
// call skt launch those saved tasks before | |||
bool KernelTaskInfo::FirstCallSKTLaunchCheck() { | |||
return ((block_dim_ != skt_info_.last_block_dim) || (stream_ != skt_info_.last_stream) || | |||
(has_group_key_ && (group_key_ != skt_info_.last_group_key))); | |||
const SuperKernelTaskInfo &skt_info = davinci_model_->GetSuperKernelTaskInfo(); | |||
return ((block_dim_ != skt_info.last_block_dim) || (stream_ != skt_info.last_stream) || | |||
(has_group_key_ && (group_key_ != skt_info.last_group_key))); | |||
} | |||
// current task has group_id or has n ATTR_N_BATCH_SPLIT then save it to skt task list; else | |||
@@ -362,7 +369,6 @@ Status KernelTaskInfo::SuperKernelDistribute() { | |||
GELOGE(ret, "Call SuperKernelLaunch failed!"); | |||
return ret; | |||
} | |||
GELOGI("Save Current task [block_dim:%u, size:%zu].", block_dim_, skt_info_.kernel_list.size()); | |||
} | |||
return SUCCESS; | |||
} | |||
@@ -391,10 +397,11 @@ Status KernelTaskInfo::Distribute() { | |||
call_save_dump_ = true; | |||
} else { | |||
/* default: not skt launch */ | |||
const SuperKernelTaskInfo &skt_info = davinci_model_->GetSuperKernelTaskInfo(); | |||
GELOGD( | |||
"KernelTaskInfo Distribute Start, sktenable:%d taskid:%u sktid:%u last_sktid:%u stubfunc_name:%s " | |||
"stubfunc:%p blockdim:%u stream:%p", | |||
call_skt, task_id_, skt_id_, skt_info_.last_task_id, stub_func_name_.c_str(), stub_func_, block_dim_, stream_); | |||
call_skt, task_id_, skt_id_, skt_info.last_task_id, stub_func_name_.c_str(), stub_func_, block_dim_, stream_); | |||
// l1 fusion enable and env flag open (kCloseSkt for skt debug) | |||
bool open_dump = false; | |||
auto all_dump_model = davinci_model_->GetDumpProperties().GetAllDumpModel(); | |||
@@ -422,23 +429,29 @@ Status KernelTaskInfo::Distribute() { | |||
"KernelTaskInfo Distribute Success. sktenable:%d taskid:%d sktid:%d stubfunc_name:%s stubfunc:%p " | |||
"blockdim:%d stream:%p", | |||
call_skt, task_id_, skt_id_, stub_func_name_.c_str(), stub_func_, block_dim_, stream_); | |||
op_desc_.reset(); // Not hold OpDesc after distribute. | |||
return SUCCESS; | |||
} | |||
Status KernelTaskInfo::UpdateArgs() { | |||
GELOGI("KernelTaskInfo::UpdateArgs in."); | |||
void KernelTaskInfo::SetIoAddrs(const OpDescPtr &op_desc) { | |||
const RuntimeParam &rts_param = davinci_model_->GetRuntimeParam(); | |||
vector<void *> input_data_addrs = ModelUtils::GetInputDataAddrs(rts_param, op_desc_); | |||
vector<void *> output_data_addrs = ModelUtils::GetOutputDataAddrs(rts_param, op_desc_); | |||
vector<void *> input_data_addrs = ModelUtils::GetInputDataAddrs(rts_param, op_desc); | |||
vector<void *> output_data_addrs = ModelUtils::GetOutputDataAddrs(rts_param, op_desc); | |||
vector<void *> io_addrs; | |||
io_addrs.insert(io_addrs.end(), input_data_addrs.begin(), input_data_addrs.end()); | |||
io_addrs.insert(io_addrs.end(), output_data_addrs.begin(), output_data_addrs.end()); | |||
io_addrs_.insert(io_addrs_.end(), input_data_addrs.begin(), input_data_addrs.end()); | |||
io_addrs_.insert(io_addrs_.end(), output_data_addrs.begin(), output_data_addrs.end()); | |||
if (kernel_type_ == ccKernelType::TE) { | |||
vector<void *> workspace_data_addrs = ModelUtils::GetWorkspaceDataAddrs(rts_param, op_desc_); | |||
io_addrs.insert(io_addrs.end(), workspace_data_addrs.begin(), workspace_data_addrs.end()); | |||
davinci_model_->SetTotalIOAddrs(io_addrs); | |||
vector<void *> workspace_data_addrs = ModelUtils::GetWorkspaceDataAddrs(rts_param, op_desc); | |||
io_addrs_.insert(io_addrs_.end(), workspace_data_addrs.begin(), workspace_data_addrs.end()); | |||
} | |||
} | |||
Status KernelTaskInfo::UpdateArgs() { | |||
GELOGI("KernelTaskInfo::UpdateArgs in."); | |||
if (kernel_type_ == ccKernelType::TE) { | |||
davinci_model_->SetTotalIOAddrs(io_addrs_); | |||
} else if (kernel_type_ == ccKernelType::AI_CPU || kernel_type_ == ccKernelType::CUST_AI_CPU) { | |||
vector<void *> io_addrs = io_addrs_; | |||
davinci_model_->UpdateKnownZeroCopyAddr(io_addrs); | |||
uintptr_t io_addr = reinterpret_cast<uintptr_t>(args_addr.get()) + sizeof(aicpu::AicpuParamHead); | |||
auto addrs_size = sizeof(uint64_t) * io_addrs.size(); | |||
@@ -789,7 +802,6 @@ Status KernelTaskInfo::InitCceTask(const domi::KernelDef &kernel_def) { | |||
GELOGE(FAILED, "flowtable is null."); | |||
return FAILED; | |||
} | |||
flowtable_size_ = flowtable.size(); | |||
} | |||
// get smDesc stored in model | |||
@@ -854,14 +866,14 @@ Status KernelTaskInfo::InitAicpuTask(uint32_t op_index, const domi::KernelDef &k | |||
GELOGI("Do InitAicpuTask"); | |||
so_name_ = kernel_def.so_name(); | |||
kernel_name_ = kernel_def.kernel_name(); | |||
GELOGI("node[%s] test so name %s, kernel name %s", op_desc_->GetName().c_str(), so_name_.c_str(), | |||
kernel_name_.c_str()); | |||
OpDescPtr op_desc = davinci_model_->GetOpByIndex(op_index); | |||
if (op_desc == nullptr) { | |||
GELOGE(INTERNAL_ERROR, "index is out of range, index: %u", op_index); | |||
return INTERNAL_ERROR; | |||
} | |||
GELOGI("node[%s] test so name %s, kernel name %s", op_desc->GetName().c_str(), so_name_.c_str(), | |||
kernel_name_.c_str()); | |||
if (kernel_type_ == ccKernelType::CUST_AI_CPU) { | |||
bool loaded = false; | |||
@@ -885,8 +897,8 @@ Status KernelTaskInfo::InitAicpuTask(uint32_t op_index, const domi::KernelDef &k | |||
GELOGE(init_ret, "Init aicpu task ext info failed, ext_info size=%zu", ext_info.size()); | |||
return init_ret; | |||
} | |||
GELOGI("Node[%s] type[%s] kernel_ext_info size=%zu, aicpu_ext_info_addr_=%p", op_desc_->GetName().c_str(), | |||
op_desc_->GetType().c_str(), ext_info.size(), aicpu_ext_info_addr_); | |||
GELOGI("Node[%s] type[%s] kernel_ext_info size=%zu, aicpu_ext_info_addr_=%p", op_desc->GetName().c_str(), | |||
op_desc->GetType().c_str(), ext_info.size(), aicpu_ext_info_addr_); | |||
aicpu_param_head->extInfoAddr = reinterpret_cast<uintptr_t>(aicpu_ext_info_addr_); | |||
aicpu_param_head->extInfoLength = static_cast<uintptr_t>(ext_info.size()); | |||
@@ -38,7 +38,6 @@ class KernelTaskInfo : public TaskInfo { | |||
flowtable_(nullptr), | |||
block_dim_(0), | |||
args_size_(0), | |||
flowtable_size_(0), | |||
task_id_(0), | |||
stream_id_(0), | |||
so_name_(""), | |||
@@ -128,6 +127,7 @@ class KernelTaskInfo : public TaskInfo { | |||
Status SuperKernelDistribute(); | |||
bool IsL1FusionOp(const OpDescPtr &op_desc); | |||
void SetIoAddrs(const OpDescPtr &op_desc); | |||
// For super kernel | |||
Status SaveSKTDumpInfo(); | |||
@@ -148,7 +148,6 @@ class KernelTaskInfo : public TaskInfo { | |||
void *flowtable_; | |||
uint32_t block_dim_; | |||
uint32_t args_size_; | |||
uint32_t flowtable_size_; | |||
uint32_t task_id_; | |||
uint32_t stream_id_; | |||
std::string so_name_; | |||
@@ -156,7 +155,8 @@ class KernelTaskInfo : public TaskInfo { | |||
ccKernelType kernel_type_; | |||
uint32_t dump_flag_; | |||
void *dump_args_; | |||
OpDescPtr op_desc_; | |||
OpDescPtr op_desc_; // Clear after distribute. | |||
vector<void *> io_addrs_; | |||
DavinciModel *davinci_model_; | |||
uint32_t args_offset_ = 0; | |||
uint32_t hybrid_args_offset_ = 0; | |||
@@ -186,25 +186,6 @@ class KernelTaskInfo : public TaskInfo { | |||
void *output_addrs = nullptr; | |||
void *attr_handle = nullptr; | |||
} custom_info_; | |||
// For super kernel | |||
static struct SuperKernelTaskInfo { | |||
uint32_t last_block_dim; | |||
uint32_t last_args_size; | |||
uint32_t last_task_id; | |||
uint32_t last_stream_id; | |||
void *last_stream; | |||
void *last_sm_desc; | |||
std::vector<void *> kernel_list; | |||
std::vector<void *> arg_list; | |||
std::vector<uint32_t> dump_flag_list; | |||
std::vector<OpDescPtr> op_desc_list; | |||
std::vector<uintptr_t> dump_args_list; | |||
uint32_t last_dump_flag; | |||
int64_t last_group_key; | |||
uintptr_t last_dump_args; | |||
OpDescPtr last_op; | |||
} skt_info_; | |||
}; | |||
} // namespace ge | |||
#endif // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_KERNEL_TASK_INFO_H_ |
@@ -30,14 +30,13 @@ Status MemcpyAsyncTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *da | |||
return ret; | |||
} | |||
memcpy_async_ = task_def.memcpy_async(); | |||
count_ = memcpy_async_.count(); | |||
kind_ = memcpy_async_.kind(); | |||
dst_max_ = memcpy_async_.dst_max(); | |||
OpDescPtr op_desc = davinci_model_->GetOpByIndex(memcpy_async_.op_index()); | |||
op_desc_ = op_desc; | |||
const domi::MemcpyAsyncDef &memcpy_async = task_def.memcpy_async(); | |||
count_ = memcpy_async.count(); | |||
kind_ = memcpy_async.kind(); | |||
dst_max_ = memcpy_async.dst_max(); | |||
OpDescPtr op_desc = davinci_model_->GetOpByIndex(memcpy_async.op_index()); | |||
if (op_desc == nullptr) { | |||
GELOGE(INTERNAL_ERROR, "Task op index:%u out of range", memcpy_async_.op_index()); | |||
GELOGE(INTERNAL_ERROR, "Task op index:%u out of range", memcpy_async.op_index()); | |||
return INTERNAL_ERROR; | |||
} | |||
@@ -46,13 +45,14 @@ Status MemcpyAsyncTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *da | |||
dst_ = reinterpret_cast<uint8_t *>(reinterpret_cast<uintptr_t>(src_) + sizeof(void *)); | |||
// for zero copy | |||
kind_ = RT_MEMCPY_ADDR_DEVICE_TO_DEVICE; | |||
GE_CHK_STATUS_RET(SetIoAddrs(op_desc, memcpy_async), "Set addrs failed"); | |||
GELOGI("MemcpyAsyncTaskInfo op name %s, src_ %p, dst_ %p, args_offset %u.", | |||
op_desc->GetName().c_str(), src_, dst_, args_offset_); | |||
return SUCCESS; | |||
} | |||
const RuntimeParam &rts_param = davinci_model_->GetRuntimeParam(); | |||
ret = ModelUtils::GetRtAddress(rts_param, memcpy_async_.src(), src_); | |||
ret = ModelUtils::GetRtAddress(rts_param, memcpy_async.src(), src_); | |||
if (ret != SUCCESS) { | |||
return ret; | |||
} | |||
@@ -61,23 +61,23 @@ Status MemcpyAsyncTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *da | |||
vector<int64_t> memory_type_list; | |||
(void)AttrUtils::GetListInt(op_desc, ATTR_NAME_OUTPUT_MEM_TYPE_LIST, memory_type_list); | |||
if (!memory_type_list.empty() && memory_type_list[0] == RT_MEMORY_TS_4G) { // TS Feature, Just one. | |||
uint64_t mem_offset = memcpy_async_.dst() - rts_param.logic_mem_base; | |||
dst_ = static_cast<uint8_t *>(rts_param.ts_mem_mall->Acquire(mem_offset, memcpy_async_.dst_max())); | |||
uint64_t mem_offset = memcpy_async.dst() - rts_param.logic_mem_base; | |||
dst_ = static_cast<uint8_t *>(rts_param.ts_mem_mall->Acquire(mem_offset, memcpy_async.dst_max())); | |||
if (dst_ == nullptr) { | |||
return FAILED; | |||
} | |||
} else { | |||
ret = ModelUtils::GetRtAddress(rts_param, memcpy_async_.dst(), dst_); | |||
ret = ModelUtils::GetRtAddress(rts_param, memcpy_async.dst(), dst_); | |||
if (ret != SUCCESS) { | |||
return ret; | |||
} | |||
} | |||
GELOGI("MemcpyAsyncTaskInfo Init Success, logic[0x%lx, 0x%lx], src:%p, dst:%p, max:%lu, count:%lu", | |||
memcpy_async_.src(), memcpy_async_.dst(), src_, dst_, dst_max_, count_); | |||
davinci_model_->DisableZeroCopy(src_); | |||
davinci_model_->DisableZeroCopy(dst_); | |||
GE_CHK_STATUS_RET(SetIoAddrs(op_desc, memcpy_async), "Set addrs failed"); | |||
GELOGI("MemcpyAsyncTaskInfo Init Success, logic[0x%lx, 0x%lx], src:%p, dst:%p, max:%lu, count:%lu", | |||
memcpy_async.src(), memcpy_async.dst(), src_, dst_, dst_max_, count_); | |||
return SUCCESS; | |||
} | |||
@@ -115,29 +115,33 @@ Status MemcpyAsyncTaskInfo::CalculateArgs(const domi::TaskDef &task_def, Davinci | |||
return SUCCESS; | |||
} | |||
Status MemcpyAsyncTaskInfo::UpdateArgs() { | |||
GELOGI("MemcpyAsyncTaskInfo::UpdateArgs in."); | |||
GE_CHECK_NOTNULL(davinci_model_); | |||
Status ret = ModelUtils::GetRtAddress(davinci_model_->GetRuntimeParam(), memcpy_async_.src(), src_); | |||
if (ret != SUCCESS) { | |||
return ret; | |||
} | |||
ret = ModelUtils::GetRtAddress(davinci_model_->GetRuntimeParam(), memcpy_async_.dst(), dst_); | |||
Status MemcpyAsyncTaskInfo::SetIoAddrs(const OpDescPtr &op_desc, const domi::MemcpyAsyncDef &memcpy_async) { | |||
uint8_t *src = nullptr; | |||
Status ret = ModelUtils::GetRtAddress(davinci_model_->GetRuntimeParam(), memcpy_async.src(), src); | |||
if (ret != SUCCESS) { | |||
return ret; | |||
} | |||
io_addrs_.emplace_back(reinterpret_cast<void *>(src)); | |||
vector<void *> io_addrs; | |||
io_addrs.emplace_back(reinterpret_cast<void *>(src_)); | |||
if (op_desc_->HasAttr(ATTR_DYNAMIC_SHAPE_FIXED_ADDR)) { | |||
if (op_desc->HasAttr(ATTR_DYNAMIC_SHAPE_FIXED_ADDR)) { | |||
void *fixed_addr = davinci_model_->GetCurrentFixedAddr(fixed_addr_offset_); | |||
io_addrs.emplace_back(fixed_addr); | |||
io_addrs_.emplace_back(fixed_addr); | |||
} else { | |||
io_addrs.emplace_back(reinterpret_cast<void *>(dst_)); | |||
uint8_t *dst = nullptr; | |||
ret = ModelUtils::GetRtAddress(davinci_model_->GetRuntimeParam(), memcpy_async.dst(), dst); | |||
if (ret != SUCCESS) { | |||
return ret; | |||
} | |||
io_addrs_.emplace_back(reinterpret_cast<void *>(dst)); | |||
} | |||
davinci_model_->SetTotalIOAddrs(io_addrs); | |||
return SUCCESS; | |||
} | |||
Status MemcpyAsyncTaskInfo::UpdateArgs() { | |||
GELOGI("MemcpyAsyncTaskInfo::UpdateArgs in."); | |||
GE_CHECK_NOTNULL(davinci_model_); | |||
davinci_model_->SetTotalIOAddrs(io_addrs_); | |||
GELOGI("MemcpyAsyncTaskInfo::UpdateArgs success."); | |||
return SUCCESS; | |||
} | |||
@@ -39,16 +39,17 @@ class MemcpyAsyncTaskInfo : public TaskInfo { | |||
Status CalculateArgs(const domi::TaskDef &task_def, DavinciModel *davinci_model) override; | |||
private: | |||
Status SetIoAddrs(const OpDescPtr &op_desc, const domi::MemcpyAsyncDef &memcpy_async); | |||
uint8_t *dst_; | |||
uint64_t dst_max_; | |||
uint8_t *src_; | |||
uint64_t count_; | |||
uint32_t kind_; | |||
OpDescPtr op_desc_; | |||
vector<void *> io_addrs_; | |||
int64_t fixed_addr_offset_; | |||
DavinciModel *davinci_model_ = nullptr; | |||
uint32_t args_offset_ = 0; | |||
domi::MemcpyAsyncDef memcpy_async_; | |||
}; | |||
} // namespace ge | |||
#endif // GE_GRAPH_LOAD_NEW_MODEL_MANAGER_TASK_INFO_MEMCPY_ASYNC_TASK_INFO_H_ |
@@ -329,7 +329,7 @@ set(COMMON_FORMAT_SRC_FILES | |||
"${GE_CODE_DIR}/ge/common/formats/format_transfers/format_transfer_fracz_nchw.cc" | |||
"${GE_CODE_DIR}/ge/common/formats/format_transfers/format_transfer_fracz_nhwc.cc" | |||
"${GE_CODE_DIR}/ge/common/formats/format_transfers/format_transfer_fracz_hwcn.cc" | |||
"${GE_CODE_DIR}/ge/common/formats/utils/formats_trans_utils.cc" | |||
"${GE_CODE_DIR}/ge/common/formats/utils/formats_trans_utils.cc" | |||
) | |||
set(GRAPH_OPTIMIZE_COMMON_SRC_FILES | |||
@@ -388,6 +388,7 @@ set(DISTINCT_GRAPH_LOAD_SRC_FILES | |||
"${GE_CODE_DIR}/ge/graph/load/new_model_manager/task_info/kernel_ex_task_info.cc" | |||
"${GE_CODE_DIR}/ge/graph/load/new_model_manager/task_info/kernel_task_info.cc" | |||
"${GE_CODE_DIR}/ge/graph/load/new_model_manager/task_info/label_set_task_info.cc" | |||
"${GE_CODE_DIR}/ge/graph/load/new_model_manager/task_info/memcpy_addr_async_task_info.cc" | |||
"${GE_CODE_DIR}/ge/graph/load/new_model_manager/task_info/memcpy_async_task_info.cc" | |||
"${GE_CODE_DIR}/ge/graph/load/new_model_manager/task_info/profiler_trace_task_info.cc" | |||
"${GE_CODE_DIR}/ge/graph/load/new_model_manager/task_info/stream_active_task_info.cc" | |||
@@ -565,6 +566,11 @@ set(DISTINCT_GRAPH_LOAD_TEST_FILES | |||
"graph/load/new_model_manager_event_manager_unittest.cc" | |||
#"graph/load/output_net_output_unittest.cc" | |||
"graph/load/tbe_handle_store_unittest.cc" | |||
"graph/load/hccl_task_info_unittest.cc" | |||
"graph/load/kernel_ex_task_info_unittest.cc" | |||
"graph/load/kernel_task_info_unittest.cc" | |||
"graph/load/memcpy_addr_async_task_info_unittest.cc" | |||
"graph/load/memcpy_async_task_info_unittest.cc" | |||
#"graph/graph_load_unittest.cc" | |||
"graph/ge_executor_unittest.cc" | |||
) | |||
@@ -918,7 +924,7 @@ target_compile_definitions(ut_libge_distinct_load_utest PRIVATE | |||
google=ascend_private | |||
) | |||
target_link_libraries(ut_libge_distinct_load_utest | |||
target_link_libraries(ut_libge_distinct_load_utest | |||
${COMMON_SHARED_LIBRARIES} | |||
$<BUILD_INTERFACE:intf_pub> | |||
ge_execute_common ge_ut_common_format ge_load_common | |||
@@ -0,0 +1,134 @@ | |||
/** | |||
* Copyright 2019-2020 Huawei Technologies Co., Ltd | |||
* | |||
* Licensed under the Apache License, Version 2.0 (the "License"); | |||
* you may not use this file except in compliance with the License. | |||
* You may obtain a copy of the License at | |||
* | |||
* http://www.apache.org/licenses/LICENSE-2.0 | |||
* | |||
* Unless required by applicable law or agreed to in writing, software | |||
* distributed under the License is distributed on an "AS IS" BASIS, | |||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | |||
* See the License for the specific language governing permissions and | |||
* limitations under the License. | |||
*/ | |||
#include <gtest/gtest.h> | |||
#define private public | |||
#define protected public | |||
#include "graph/load/new_model_manager/davinci_model.h" | |||
#include "graph/load/new_model_manager/task_info/hccl_task_info.h" | |||
namespace ge { | |||
class UtestHcclTaskInfo : public testing::Test { | |||
protected: | |||
void SetUp() {} | |||
void TearDown() {} | |||
}; | |||
// test success GetTaskID | |||
TEST_F(UtestHcclTaskInfo, success_get_task_id) { | |||
domi::ModelTaskDef model_task_def; | |||
domi::TaskDef *task = model_task_def.add_task(); | |||
task->set_type(RT_MODEL_TASK_KERNEL); | |||
TaskInfoPtr task_info = TaskInfoFactory::Instance().Create(static_cast<rtModelTaskType_t>(task->type())); | |||
EXPECT_EQ(task_info->GetTaskID(), 0); | |||
HcclTaskInfo hccl_task_info; | |||
EXPECT_EQ(hccl_task_info.GetTaskID(), 0); | |||
} | |||
// test init EventRecordTaskInfo | |||
TEST_F(UtestHcclTaskInfo, success_create_stream) { | |||
DavinciModel model(0, nullptr); | |||
HcclTaskInfo hccl_task_info; | |||
EXPECT_EQ(hccl_task_info.CreateStream(3, &model, 0), SUCCESS); | |||
} | |||
// test hccl_Distribute | |||
TEST_F(UtestHcclTaskInfo, success_distribute7) { | |||
DavinciModel model(0, nullptr); | |||
domi::ModelTaskDef model_task_def; | |||
domi::TaskDef *task7 = model_task_def.add_task(); | |||
task7->set_type(RT_MODEL_TASK_HCCL); | |||
TaskInfoPtr task_info7 = TaskInfoFactory::Instance().Create(static_cast<rtModelTaskType_t>(task7->type())); | |||
Status ret = task_info7->Init(task7[0], &model); | |||
EXPECT_EQ(FAILED, ret); | |||
std::vector<TaskInfoPtr> task_list; | |||
task_list.push_back(task_info7); | |||
model.task_list_ = task_list; | |||
EXPECT_EQ(task_info7->Release(), SUCCESS); | |||
} | |||
// test hccl_Distribute | |||
TEST_F(UtestHcclTaskInfo, success_distribute7_with_hccl_type) { | |||
DavinciModel model(0, nullptr); | |||
rtStream_t stream = nullptr; | |||
rtStreamCreate(&stream, 0); | |||
model.stream_list_ = { stream }; | |||
domi::TaskDef task_def; | |||
HcclTaskInfo hccl_task_info; | |||
EXPECT_EQ(hccl_task_info.Init(task_def, nullptr), PARAM_INVALID); | |||
domi::KernelHcclDef *kernel_hccl_def = task_def.mutable_kernel_hccl(); | |||
kernel_hccl_def->set_op_index(0); | |||
kernel_hccl_def->set_hccl_type("HcomBroadcast"); | |||
model.op_list_[0] = std::make_shared<OpDesc>("FrameworkOp", "FrameworkOp"); | |||
EXPECT_EQ(hccl_task_info.Init(task_def, &model), SUCCESS); | |||
task_def.clear_kernel_hccl(); | |||
} | |||
// test hccl_GetPrivateDefByTaskDef | |||
TEST_F(UtestHcclTaskInfo, success_hccl_get_private_def_by_task_def) { | |||
DavinciModel model(0, nullptr); | |||
domi::ModelTaskDef model_task_def; | |||
domi::TaskDef *task7 = model_task_def.add_task(); | |||
task7->set_type(RT_MODEL_TASK_HCCL); | |||
// for SetStream | |||
rtStream_t stream = nullptr; | |||
rtStreamCreate(&stream, 0); | |||
model.stream_list_.push_back(stream); | |||
// for GetPrivateDefByTaskDef | |||
task7->set_ops_kernel_store_ptr(10); | |||
std::string value = "hccl_task"; | |||
task7->set_private_def(value); | |||
TaskInfoPtr task_info7 = TaskInfoFactory::Instance().Create(static_cast<rtModelTaskType_t>(task7->type())); | |||
// for Distribute | |||
EXPECT_EQ(task_info7->Init(task7[0], &model), PARAM_INVALID); | |||
EXPECT_EQ(task_info7->Release(), SUCCESS); | |||
} | |||
// test hccl_task_TransToGETaskInfo | |||
TEST_F(UtestHcclTaskInfo, success_hccl_trans_to_ge_task_info) { | |||
DavinciModel model(0, nullptr); | |||
domi::ModelTaskDef model_task_def; | |||
domi::TaskDef *task7 = model_task_def.add_task(); | |||
// for type | |||
task7->set_type(RT_MODEL_TASK_HCCL); | |||
TaskInfoPtr task_info7 = TaskInfoFactory::Instance().Create(static_cast<rtModelTaskType_t>(task7->type())); | |||
GETaskInfo ge_task; | |||
HcclTaskInfo hccl_task_info; | |||
hccl_task_info.TransToGETaskInfo(ge_task); | |||
EXPECT_EQ(task_info7->Release(), SUCCESS); | |||
} | |||
} // namespace ge |
@@ -0,0 +1,144 @@ | |||
/** | |||
* Copyright 2019-2020 Huawei Technologies Co., Ltd | |||
* | |||
* Licensed under the Apache License, Version 2.0 (the "License"); | |||
* you may not use this file except in compliance with the License. | |||
* You may obtain a copy of the License at | |||
* | |||
* http://www.apache.org/licenses/LICENSE-2.0 | |||
* | |||
* Unless required by applicable law or agreed to in writing, software | |||
* distributed under the License is distributed on an "AS IS" BASIS, | |||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | |||
* See the License for the specific language governing permissions and | |||
* limitations under the License. | |||
*/ | |||
#include <gtest/gtest.h> | |||
#define private public | |||
#define protected public | |||
#include "graph/load/new_model_manager/davinci_model.h" | |||
#include "graph/load/new_model_manager/task_info/kernel_ex_task_info.h" | |||
#include "cce/aicpu_engine_struct.h" | |||
namespace ge { | |||
extern OpDescPtr CreateOpDesc(string name, string type); | |||
class UtestKernelExTaskInfo : public testing::Test { | |||
protected: | |||
void SetUp() {} | |||
void TearDown() {} | |||
}; | |||
// test kernel_ex_task_Release | |||
TEST_F(UtestKernelExTaskInfo, success_kernel_ex_task_init) { | |||
domi::TaskDef task_def; | |||
KernelExTaskInfo kernel_ex_task_info; | |||
EXPECT_EQ(kernel_ex_task_info.Init(task_def, nullptr), PARAM_INVALID); | |||
DavinciModel model(0, nullptr); | |||
EXPECT_EQ(kernel_ex_task_info.Init(task_def, &model), FAILED); | |||
rtStream_t stream = nullptr; | |||
rtStreamCreate(&stream, 0); | |||
model.stream_list_.push_back(stream); | |||
domi::KernelExDef *kernel_ex_def = task_def.mutable_kernel_ex(); | |||
kernel_ex_def->set_op_index(1); | |||
model.op_list_[0] = CreateOpDesc("FrameworkOp", "FrameworkOp"); | |||
EXPECT_EQ(kernel_ex_task_info.Init(task_def, &model), INTERNAL_ERROR); | |||
kernel_ex_def->clear_op_index(); | |||
kernel_ex_def->set_op_index(0); | |||
EXPECT_EQ(kernel_ex_task_info.Init(task_def, &model), FAILED); | |||
kernel_ex_def->set_task_info("KernelEx"); | |||
kernel_ex_def->set_task_info_size(1); | |||
EXPECT_EQ(kernel_ex_task_info.Init(task_def, &model), FAILED); | |||
constexpr uint32_t arg_size = sizeof(STR_FWK_OP_KERNEL); | |||
string value1(arg_size, 'a'); | |||
kernel_ex_def->set_args_size(arg_size); | |||
kernel_ex_def->set_args(value1); | |||
OpDescPtr v_op_desc = CreateOpDesc("ge_global_step", "Variable"); | |||
model.variable_op_list_.push_back(v_op_desc); | |||
model.op_list_[0]->SetWorkspace({100331008}); // offset | |||
model.op_list_[0]->SetWorkspaceBytes({150}); // length | |||
EXPECT_EQ(kernel_ex_task_info.Init(task_def, &model), FAILED); | |||
task_def.clear_kernel_ex(); | |||
} | |||
// test kernel_ex_task_Release | |||
TEST_F(UtestKernelExTaskInfo, success_kernel_ex_task_release) { | |||
KernelExTaskInfo kernel_ex_task_info; | |||
EXPECT_EQ(kernel_ex_task_info.Release(), SUCCESS); | |||
kernel_ex_task_info.kernel_buf_ = nullptr; | |||
rtMalloc(&kernel_ex_task_info.input_output_addr_, 64, RT_MEMORY_HBM); | |||
EXPECT_EQ(kernel_ex_task_info.Release(), SUCCESS); | |||
kernel_ex_task_info.input_output_addr_ = nullptr; | |||
rtMalloc(&kernel_ex_task_info.kernel_buf_, 64, RT_MEMORY_HBM); | |||
EXPECT_EQ(kernel_ex_task_info.Release(), SUCCESS); | |||
rtMalloc(&kernel_ex_task_info.kernel_buf_, 64, RT_MEMORY_HBM); | |||
rtMalloc(&kernel_ex_task_info.input_output_addr_, 64, RT_MEMORY_HBM); | |||
EXPECT_EQ(kernel_ex_task_info.Release(), SUCCESS); | |||
} | |||
// test kernel_ex_task_Release | |||
TEST_F(UtestKernelExTaskInfo, success_kernel_ex_task_info_copy) { | |||
DavinciModel model(0, nullptr); | |||
model.runtime_param_.mem_base = (uint8_t *)0x12345; | |||
model.runtime_param_.mem_size = 100332000; | |||
rtStream_t stream = nullptr; | |||
rtStreamCreate(&stream, 0); | |||
model.stream_list_.push_back(stream); | |||
domi::TaskDef task_def; | |||
KernelExTaskInfo kernel_ex_task_info; | |||
domi::KernelExDef *kernel_ex_def = task_def.mutable_kernel_ex(); | |||
kernel_ex_def->set_task_info_size(150); | |||
kernel_ex_def->set_op_index(0); | |||
model.op_list_[0] = CreateOpDesc("FrameworkOp", "FrameworkOp"); | |||
EXPECT_EQ(kernel_ex_task_info.Init(task_def, &model), FAILED); // workspace empty. | |||
model.op_list_[0]->SetWorkspace({100331008}); // offset | |||
model.op_list_[0]->SetWorkspaceBytes({0}); // length | |||
EXPECT_EQ(kernel_ex_task_info.Init(task_def, &model), FAILED); // workspace addr is null. | |||
model.op_list_[0]->SetWorkspace({100331008}); // offset | |||
model.op_list_[0]->SetWorkspaceBytes({10}); // length | |||
EXPECT_EQ(kernel_ex_task_info.Init(task_def, &model), FAILED); // workspace addr is small. | |||
model.op_list_[0]->SetWorkspace({100331008}); // offset | |||
model.op_list_[0]->SetWorkspaceBytes({150}); // length | |||
EXPECT_EQ(kernel_ex_task_info.Init(task_def, &model), SUCCESS); | |||
task_def.clear_kernel_ex(); | |||
model.runtime_param_.mem_base = nullptr; | |||
} | |||
TEST_F(UtestKernelExTaskInfo, kernel_ex_task_info_calculate_args) { | |||
DavinciModel model(0, nullptr); | |||
domi::TaskDef task_def; | |||
domi::KernelExDef *kernel_ex_def = task_def.mutable_kernel_ex(); | |||
kernel_ex_def->set_op_index(0); | |||
model.op_list_[0] = CreateOpDesc("FrameworkOp", "FrameworkOp"); | |||
AttrUtils::SetStr(model.op_list_[0], ATTR_DYNAMIC_SHAPE_FIXED_ADDR, "Hello Mr Tree"); | |||
KernelExTaskInfo kernel_ex_task_info; | |||
EXPECT_EQ(kernel_ex_task_info.CalculateArgs(task_def, &model), FAILED); | |||
} | |||
} // namespace ge |
@@ -0,0 +1,138 @@ | |||
/** | |||
* Copyright 2019-2020 Huawei Technologies Co., Ltd | |||
* | |||
* Licensed under the Apache License, Version 2.0 (the "License"); | |||
* you may not use this file except in compliance with the License. | |||
* You may obtain a copy of the License at | |||
* | |||
* http://www.apache.org/licenses/LICENSE-2.0 | |||
* | |||
* Unless required by applicable law or agreed to in writing, software | |||
* distributed under the License is distributed on an "AS IS" BASIS, | |||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | |||
* See the License for the specific language governing permissions and | |||
* limitations under the License. | |||
*/ | |||
#include <gtest/gtest.h> | |||
#define private public | |||
#define protected public | |||
#include "graph/load/new_model_manager/davinci_model.h" | |||
#include "graph/load/new_model_manager/task_info/memcpy_addr_async_task_info.h" | |||
namespace ge { | |||
class UtestMemcpyAddrAsyncTaskInfo : public testing::Test { | |||
protected: | |||
void SetUp() {} | |||
void TearDown() {} | |||
}; | |||
extern OpDescPtr CreateOpDesc(string name, string type); | |||
TEST_F(UtestMemcpyAddrAsyncTaskInfo, success_memcpy_addr_async_task_init) { | |||
DavinciModel model(0, nullptr); | |||
domi::TaskDef task_def; | |||
task_def.set_stream_id(0); | |||
domi::MemcpyAsyncDef *memcpy_async = task_def.mutable_memcpy_async(); | |||
memcpy_async->set_dst(10); | |||
memcpy_async->set_dst_max(512); | |||
memcpy_async->set_src(10); | |||
memcpy_async->set_count(1); | |||
memcpy_async->set_kind(RT_MEMCPY_ADDR_DEVICE_TO_DEVICE); | |||
memcpy_async->set_op_index(6); | |||
model.runtime_param_.logic_mem_base = 0x8003000; | |||
model.runtime_param_.logic_weight_base = 0x8008000; | |||
model.runtime_param_.logic_var_base = 0x800e000; | |||
model.runtime_param_.mem_size = 0x5000; | |||
model.runtime_param_.weight_size = 0x6000; | |||
model.runtime_param_.var_size = 0x1000; | |||
// DavinciModel is null | |||
MemcpyAddrAsyncTaskInfo memcpy_addr_async_task_info; | |||
EXPECT_EQ(memcpy_addr_async_task_info.Init(task_def, nullptr), PARAM_INVALID); | |||
// SetStream failed. | |||
EXPECT_EQ(memcpy_addr_async_task_info.Init(task_def, &model), FAILED); | |||
// GetOpByIndex src failed | |||
rtStream_t stream = nullptr; | |||
rtStreamCreate(&stream, 0); | |||
model.stream_list_.push_back(stream); | |||
EXPECT_EQ(memcpy_addr_async_task_info.Init(task_def, &model), INTERNAL_ERROR); | |||
// GetRuntimeAddress src failed. | |||
model.op_list_[6] = CreateOpDesc("memcpyaddrasync", MEMCPYADDRASYNC); | |||
EXPECT_EQ(memcpy_addr_async_task_info.Init(task_def, &model), PARAM_INVALID); | |||
// GetRuntimeAddress dst failed. | |||
memcpy_async->set_src(0x08003000); | |||
EXPECT_EQ(memcpy_addr_async_task_info.Init(task_def, &model), PARAM_INVALID); | |||
memcpy_async->set_dst(0x08008000); | |||
EXPECT_EQ(memcpy_addr_async_task_info.Init(task_def, &model), SUCCESS); | |||
task_def.clear_memcpy_async(); | |||
} | |||
TEST_F(UtestMemcpyAddrAsyncTaskInfo, success_memcpy_async_task_init_failed) { | |||
DavinciModel model(0, nullptr); | |||
domi::TaskDef task_def; | |||
task_def.set_stream_id(0); | |||
rtStream_t stream = nullptr; | |||
rtStreamCreate(&stream, 0); | |||
model.stream_list_.push_back(stream); | |||
domi::MemcpyAsyncDef *memcpy_async = task_def.mutable_memcpy_async(); | |||
memcpy_async->set_dst(10); | |||
memcpy_async->set_dst_max(512); | |||
memcpy_async->set_src(10); | |||
memcpy_async->set_count(1); | |||
memcpy_async->set_kind(RT_MEMCPY_ADDR_DEVICE_TO_DEVICE); | |||
memcpy_async->set_op_index(6); | |||
model.runtime_param_.logic_mem_base = 0x8003000; | |||
model.runtime_param_.logic_weight_base = 0x8008000; | |||
model.runtime_param_.logic_var_base = 0x800e000; | |||
model.runtime_param_.mem_size = 0x5000; | |||
model.runtime_param_.weight_size = 0x6000; | |||
model.runtime_param_.var_size = 0x1000; | |||
GeTensorDesc tensor(GeShape(), FORMAT_NCHW, DT_FLOAT); | |||
model.op_list_[6] = CreateOpDesc("memcpyasync", MEMCPYADDRASYNC); | |||
model.op_list_[6]->AddInputDesc(tensor); | |||
model.op_list_[6]->AddOutputDesc(tensor); | |||
model.op_list_[6]->SetInputOffset({1024}); | |||
model.op_list_[6]->SetOutputOffset({5120}); | |||
// DavinciModel is null | |||
MemcpyAddrAsyncTaskInfo memcpy_addr_async_task_info; | |||
EXPECT_EQ(memcpy_addr_async_task_info.Init(task_def, &model), PARAM_INVALID); | |||
task_def.clear_memcpy_async(); | |||
} | |||
TEST_F(UtestMemcpyAddrAsyncTaskInfo, success_memcpy_async_calculate_args) { | |||
DavinciModel model(0, nullptr); | |||
domi::TaskDef task_def; | |||
domi::MemcpyAsyncDef *memcpy_async = task_def.mutable_memcpy_async(); | |||
memcpy_async->set_dst(0x08003000); | |||
memcpy_async->set_dst_max(512); | |||
memcpy_async->set_src(0x08008000); | |||
memcpy_async->set_count(1); | |||
memcpy_async->set_kind(RT_MEMCPY_DEVICE_TO_DEVICE); | |||
memcpy_async->set_op_index(0); | |||
// DavinciModel is null | |||
MemcpyAddrAsyncTaskInfo memcpy_addr_async_task_info; | |||
EXPECT_EQ(memcpy_addr_async_task_info.CalculateArgs(task_def, &model), SUCCESS); | |||
} | |||
} // namespace ge |
@@ -0,0 +1,273 @@ | |||
/** | |||
* Copyright 2019-2020 Huawei Technologies Co., Ltd | |||
* | |||
* Licensed under the Apache License, Version 2.0 (the "License"); | |||
* you may not use this file except in compliance with the License. | |||
* You may obtain a copy of the License at | |||
* | |||
* http://www.apache.org/licenses/LICENSE-2.0 | |||
* | |||
* Unless required by applicable law or agreed to in writing, software | |||
* distributed under the License is distributed on an "AS IS" BASIS, | |||
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | |||
* See the License for the specific language governing permissions and | |||
* limitations under the License. | |||
*/ | |||
#include <gtest/gtest.h> | |||
#define private public | |||
#define protected public | |||
#include "graph/load/new_model_manager/davinci_model.h" | |||
#include "graph/load/new_model_manager/task_info/memcpy_async_task_info.h" | |||
namespace ge { | |||
class UtestMemcpyAsyncTaskInfo : public testing::Test { | |||
protected: | |||
void SetUp() {} | |||
void TearDown() {} | |||
}; | |||
OpDescPtr CreateOpDesc(string name = "", string type = "") { | |||
auto op_desc = std::make_shared<OpDesc>(name, type); | |||
op_desc->SetStreamId(0); | |||
op_desc->SetId(0); | |||
AttrUtils::SetFloat(op_desc, ATTR_NAME_ALPHA, 0); | |||
AttrUtils::SetFloat(op_desc, ATTR_NAME_BETA, 0); | |||
op_desc->SetWorkspace({}); | |||
op_desc->SetWorkspaceBytes({}); | |||
op_desc->SetInputOffset({}); | |||
op_desc->SetOutputOffset({}); | |||
AttrUtils::SetListStr(op_desc, ATTR_NAME_WEIGHT_NAME, {}); | |||
AttrUtils::SetInt(op_desc, POOLING_ATTR_MODE, 0); | |||
AttrUtils::SetInt(op_desc, POOLING_ATTR_PAD_MODE, 0); | |||
AttrUtils::SetInt(op_desc, POOLING_ATTR_DATA_MODE, 0); | |||
AttrUtils::SetInt(op_desc, POOLING_ATTR_CEIL_MODE, 0); | |||
AttrUtils::SetInt(op_desc, POOLING_ATTR_NAN_OPT, 0); | |||
AttrUtils::SetListInt(op_desc, POOLING_ATTR_WINDOW, {}); | |||
AttrUtils::SetListInt(op_desc, POOLING_ATTR_PAD, {}); | |||
AttrUtils::SetListInt(op_desc, POOLING_ATTR_STRIDE, {}); | |||
AttrUtils::SetListInt(op_desc, ATTR_NAME_ACTIVE_STREAM_LIST, {1, 1}); | |||
AttrUtils::SetInt(op_desc, ATTR_NAME_STREAM_SWITCH_COND, 0); | |||
return op_desc; | |||
} | |||
TEST_F(UtestMemcpyAsyncTaskInfo, success_memcpy_async_task_init) { | |||
DavinciModel model(0, nullptr); | |||
domi::TaskDef task_def; | |||
task_def.set_stream_id(0); | |||
domi::MemcpyAsyncDef *memcpy_async = task_def.mutable_memcpy_async(); | |||
memcpy_async->set_dst(10); | |||
memcpy_async->set_dst_max(512); | |||
memcpy_async->set_src(10); | |||
memcpy_async->set_count(1); | |||
memcpy_async->set_kind(RT_MEMCPY_DEVICE_TO_DEVICE); | |||
memcpy_async->set_op_index(6); | |||
model.runtime_param_.logic_mem_base = 0x8003000; | |||
model.runtime_param_.logic_weight_base = 0x8008000; | |||
model.runtime_param_.logic_var_base = 0x800e000; | |||
model.runtime_param_.mem_size = 0x5000; | |||
model.runtime_param_.weight_size = 0x6000; | |||
model.runtime_param_.var_size = 0x1000; | |||
MemcpyAsyncTaskInfo memcpy_async_task_info; | |||
// GetOpByIndex src failed | |||
rtStream_t stream = nullptr; | |||
rtStreamCreate(&stream, 0); | |||
model.stream_list_.push_back(stream); | |||
EXPECT_EQ(memcpy_async_task_info.Init(task_def, &model), INTERNAL_ERROR); | |||
model.op_list_[6] = CreateOpDesc("memcpyasync", MEMCPYASYNC); | |||
memcpy_async->set_src(0x08008000); | |||
EXPECT_EQ(memcpy_async_task_info.Init(task_def, &model), PARAM_INVALID); | |||
// set OpDesc attr | |||
std::vector<int64_t> memory_type = { RT_MEMORY_TS_4G }; | |||
AttrUtils::SetListInt(model.op_list_[6], ATTR_NAME_OUTPUT_MEM_TYPE_LIST, memory_type); | |||
GeTensorDesc tensor(GeShape(), FORMAT_NCHW, DT_FLOAT); | |||
model.op_list_[6]->AddInputDesc(tensor); | |||
model.op_list_[6]->AddOutputDesc(tensor); | |||
memcpy_async->set_dst_max(0); | |||
EXPECT_EQ(memcpy_async_task_info.Init(task_def, &model), FAILED); | |||
memcpy_async->set_dst_max(0); | |||
model.op_list_[6]->SetInputOffset({1024}); | |||
model.op_list_[6]->SetOutputOffset({5120}); | |||
EXPECT_EQ(memcpy_async_task_info.Init(task_def, &model), FAILED); | |||
task_def.clear_memcpy_async(); | |||
} | |||
TEST_F(UtestMemcpyAsyncTaskInfo, success_memcpy_async_task_init_failed) { | |||
DavinciModel model(0, nullptr); | |||
domi::TaskDef task_def; | |||
task_def.set_stream_id(0); | |||
domi::MemcpyAsyncDef *memcpy_async = task_def.mutable_memcpy_async(); | |||
memcpy_async->set_dst(10); | |||
memcpy_async->set_dst_max(512); | |||
memcpy_async->set_src(10); | |||
memcpy_async->set_count(1); | |||
memcpy_async->set_kind(RT_MEMCPY_DEVICE_TO_DEVICE); | |||
memcpy_async->set_op_index(6); | |||
model.runtime_param_.logic_mem_base = 0x8003000; | |||
model.runtime_param_.logic_weight_base = 0x8008000; | |||
model.runtime_param_.logic_var_base = 0x800e000; | |||
model.runtime_param_.mem_size = 0x5000; | |||
model.runtime_param_.weight_size = 0x6000; | |||
model.runtime_param_.var_size = 0x1000; | |||
// DavinciModel is null | |||
MemcpyAsyncTaskInfo memcpy_async_task_info; | |||
EXPECT_EQ(memcpy_async_task_info.Init(task_def, nullptr), PARAM_INVALID); | |||
// SetStream failed | |||
EXPECT_EQ(memcpy_async_task_info.Init(task_def, nullptr), PARAM_INVALID); | |||
// GetOpByIndex failed | |||
rtStream_t stream = nullptr; | |||
rtStreamCreate(&stream, 0); | |||
model.stream_list_.push_back(stream); | |||
EXPECT_EQ(memcpy_async_task_info.Init(task_def, &model), INTERNAL_ERROR); | |||
model.op_list_[6] = CreateOpDesc("memcpyasync", MEMCPYASYNC); | |||
EXPECT_EQ(memcpy_async_task_info.Init(task_def, &model), PARAM_INVALID); | |||
memcpy_async->set_src(0x08008000); | |||
EXPECT_EQ(memcpy_async_task_info.Init(task_def, &model), PARAM_INVALID); | |||
memcpy_async->set_dst(0x08003000); | |||
// set OpDesc attr | |||
std::vector<int64_t> memory_type = { RT_MEMORY_TS_4G }; | |||
AttrUtils::SetListInt(model.op_list_[6], ATTR_NAME_OUTPUT_MEM_TYPE_LIST, memory_type); | |||
memcpy_async->set_dst_max(0); | |||
EXPECT_EQ(memcpy_async_task_info.Init(task_def, nullptr), PARAM_INVALID); | |||
memcpy_async->set_dst_max(512); | |||
GeTensorDesc tensor(GeShape(), FORMAT_NCHW, DT_FLOAT); | |||
model.op_list_[6]->AddInputDesc(tensor); | |||
model.op_list_[6]->AddOutputDesc(tensor); | |||
model.op_list_[6]->SetInputOffset({1024}); | |||
model.op_list_[6]->SetOutputOffset({5120}); | |||
EXPECT_EQ(memcpy_async_task_info.Init(task_def, &model), SUCCESS); | |||
memcpy_async->set_dst(0x08009000); | |||
EXPECT_EQ(memcpy_async_task_info.Init(task_def, &model), SUCCESS); | |||
task_def.clear_memcpy_async(); | |||
} | |||
TEST_F(UtestMemcpyAsyncTaskInfo, success_memcpy_async_task_distribute) { | |||
DavinciModel model(0, nullptr); | |||
model.SetKnownNode(true); | |||
domi::TaskDef task_def; | |||
task_def.set_stream_id(0); | |||
domi::MemcpyAsyncDef *memcpy_async = task_def.mutable_memcpy_async(); | |||
memcpy_async->set_dst(10); | |||
memcpy_async->set_dst_max(512); | |||
memcpy_async->set_src(10); | |||
memcpy_async->set_count(1); | |||
memcpy_async->set_kind(RT_MEMCPY_DEVICE_TO_DEVICE); | |||
memcpy_async->set_op_index(6); | |||
model.runtime_param_.logic_mem_base = 0x8003000; | |||
model.runtime_param_.logic_weight_base = 0x8008000; | |||
model.runtime_param_.logic_var_base = 0x800e000; | |||
model.runtime_param_.mem_size = 0x5000; | |||
model.runtime_param_.weight_size = 0x6000; | |||
model.runtime_param_.var_size = 0x1000; | |||
MemcpyAsyncTaskInfo memcpy_async_task_info; | |||
// GetOpByIndex src failed | |||
rtStream_t stream = nullptr; | |||
rtStreamCreate(&stream, 0); | |||
model.stream_list_.push_back(stream); | |||
EXPECT_EQ(memcpy_async_task_info.Init(task_def, &model), INTERNAL_ERROR); | |||
model.op_list_[6] = CreateOpDesc("memcpyasync", MEMCPYASYNC); | |||
memcpy_async->set_src(0x08008000); | |||
EXPECT_EQ(memcpy_async_task_info.Init(task_def, &model), PARAM_INVALID); | |||
// set OpDesc attr | |||
AttrUtils::SetStr(model.op_list_[6], ATTR_DYNAMIC_SHAPE_FIXED_ADDR, "Hello Mr Tree"); | |||
GeTensorDesc tensor(GeShape(), FORMAT_NCHW, DT_FLOAT); | |||
model.op_list_[6]->AddInputDesc(tensor); | |||
model.op_list_[6]->AddOutputDesc(tensor); | |||
memcpy_async->set_dst_max(0); | |||
EXPECT_EQ(memcpy_async_task_info.Init(task_def, &model), SUCCESS); | |||
memcpy_async->set_dst_max(0); | |||
model.op_list_[6]->SetInputOffset({1024}); | |||
model.op_list_[6]->SetOutputOffset({5120}); | |||
EXPECT_EQ(memcpy_async_task_info.Init(task_def, &model), SUCCESS); | |||
task_def.clear_memcpy_async(); | |||
} | |||
TEST_F(UtestMemcpyAsyncTaskInfo, success_distribute) { | |||
DavinciModel model(0, nullptr); | |||
model.ge_model_ = MakeShared<GeModel>(); | |||
auto model_task_def = MakeShared<domi::ModelTaskDef>(); | |||
domi::TaskDef *task_def = model_task_def->add_task(); | |||
task_def->set_type(RT_MODEL_TASK_MEMCPY_ASYNC); | |||
domi::KernelDef *kernel_def = task_def->mutable_kernel(); | |||
domi::KernelContext *ctx = kernel_def->mutable_context(); | |||
ctx->set_op_index(0); | |||
model.op_list_[0] = CreateOpDesc("memcpyasync", MEMCPYASYNC); | |||
TaskInfoPtr task_info = TaskInfoFactory::Instance().Create(static_cast<rtModelTaskType_t>(task_def->type())); | |||
model.task_list_ = { task_info }; | |||
model.ge_model_->SetModelTaskDef(model_task_def); | |||
EXPECT_EQ(model.DistributeTask(), SUCCESS); | |||
EXPECT_EQ(task_info->Distribute(), SUCCESS); | |||
task_info->Release(); | |||
} | |||
TEST_F(UtestMemcpyAsyncTaskInfo, success_memcpy_async_calculate_args) { | |||
DavinciModel model(0, nullptr); | |||
domi::TaskDef task_def; | |||
domi::MemcpyAsyncDef *memcpy_async = task_def.mutable_memcpy_async(); | |||
memcpy_async->set_dst(0x08003000); | |||
memcpy_async->set_dst_max(512); | |||
memcpy_async->set_src(0x08008000); | |||
memcpy_async->set_count(1); | |||
memcpy_async->set_kind(RT_MEMCPY_DEVICE_TO_DEVICE); | |||
memcpy_async->set_op_index(0); | |||
model.op_list_[0] = CreateOpDesc("memcpyasync", MEMCPYASYNC); | |||
AttrUtils::SetStr(model.op_list_[0], ATTR_DYNAMIC_SHAPE_FIXED_ADDR, "Hello Mr Tree"); | |||
// DavinciModel is null | |||
MemcpyAsyncTaskInfo memcpy_async_task_info; | |||
EXPECT_EQ(memcpy_async_task_info.CalculateArgs(task_def, &model), SUCCESS); | |||
} | |||
TEST_F(UtestMemcpyAsyncTaskInfo, memcpy_async_update_args) { | |||
DavinciModel model(0, nullptr); | |||
MemcpyAsyncTaskInfo memcpy_async_task_info; | |||
memcpy_async_task_info.davinci_model_ = &model; | |||
EXPECT_EQ(memcpy_async_task_info.UpdateArgs(), SUCCESS); | |||
} | |||
} // namespace ge |