m53297601
/
graphengine
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

!1443 synchronize latest ascend softare suite 02 Apr 2021

Browse Source 
From: @nicholas_yhr
Reviewed-by: @lilongfei15,@ljl0711
Signed-off-by: @lilongfei15
pull/1443/MERGE
mindspore-ci-bot 4 years ago committed by Gitee
parent 0679af1d75 2fbf01c53f
commit 24d3b54ab8
53 changed files with 681 additions and 365 deletions
Show Stats Download Patch File Download Diff File

9
ge/analyzer/analyzer.cc
Unescape View File

@ -221,7 +221,10 @@ ge::Status Analyzer::SaveAnalyzerDataToFile(uint64_t session_id, uint64_t graph_
try {
json_file_ << jsn.dump(kJsonDumpLevel) << std::endl;
} catch (nlohmann::detail::type_error &e) {
GELOGE(FAILED, "[Json.dump][GraphInfo]json.dump to analyze file [%s] failed because [%s], session_id:%lu, graph_id:%lu", json_file_name_.c_str(), e.what(), session_id, graph_id);
GELOGE(FAILED,
"[Json.dump][GraphInfo]json.dump to analyze file [%s] failed because [%s],"
"session_id:%lu, graph_id:%lu",
json_file_name_.c_str(), e.what(), session_id, graph_id);
ret_failed = true;
}
json_file_.close();
@ -241,7 +244,9 @@ ge::Status Analyzer::DoAnalyze(DataInfo &data_info) {
GE_CHECK_NOTNULL(graph_info);
auto status = SaveOpInfo(desc, data_info, graph_info);
if (status != SUCCESS) {
GELOGE(status, "[Check][SaveOpInfo]save op info: desc_name [%s] desc_type [%s] failed!", desc->GetName().c_str(), desc->GetType().c_str());
GELOGE(status,
"[Check][SaveOpInfo]save op info: desc_name [%s] desc_type [%s] failed!",
desc->GetName().c_str(), desc->GetType().c_str());
return FAILED;
}
// create json file

3
ge/common/formats/format_transfers/datatype_transfer.cc
Unescape View File

@ -154,7 +154,8 @@ Status DataTypeTransfer::TransDataType(const CastArgs &args, TransResult &result
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to alloc the memory for dst buf %zu, data size %zu", total_size, args.src_data_size);
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to alloc the memory for dst buf %zu, data size %zu", total_size, args.src_data_size);
return ACL_ERROR_GE_MEMORY_ALLOCATION;
}

3
ge/common/formats/format_transfers/format_transfer_c1hwncoc0_hwcn.cc
Unescape View File

@ -73,7 +73,8 @@ Status CheckArgsForC1hwncoc0ToHwcn(const TransArgs &args) {
Status GetDstDataAfterTrans(const TransArgs &args, TransResult &result, int size, int64_t total_size) {
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), total_size, ShapeToString(args.dst_shape).c_str());
return ACL_ERROR_GE_MEMORY_ALLOCATION;

6
ge/common/formats/format_transfers/format_transfer_dhwcn_fracz3D.cc
Unescape View File

@ -94,7 +94,8 @@ Status TransFormatDhwckToFz3D(const TransArgs &args, TransResult &result) {
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size], std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
return ACL_ERROR_GE_MEMORY_ALLOCATION;
@ -122,7 +123,8 @@ Status TransFormatDhwckToFz3D(const TransArgs &args, TransResult &result) {
args.data + src_idx * data_size, static_cast<size_t>(data_size));
}
if (ret != EOK) {
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d, pad mode %d",
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
"Failed to operate the dst memory at offset %ld, error-code %d, pad mode %d",
dst_offset, ret, pad_zero);
return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
}

6
ge/common/formats/format_transfers/format_transfer_dhwnc_fracz3D_transpose.cc
Unescape View File

@ -95,7 +95,8 @@ Status TransFormatDhwncToFz3DTranspose(const TransArgs &args, TransResult &resul
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size], std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
return ACL_ERROR_GE_MEMORY_ALLOCATION;
@ -123,7 +124,8 @@ Status TransFormatDhwncToFz3DTranspose(const TransArgs &args, TransResult &resul
args.data + src_idx * data_size, static_cast<size_t>(data_size));
}
if (ret != EOK) {
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d, pad mode %d",
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
"Failed to operate the dst memory at offset %ld, error-code %d, pad mode %d",
dst_offset, ret, pad_zero);
return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
}

30
ge/common/formats/format_transfers/format_transfer_fractal_nz.cc
Unescape View File

@ -139,7 +139,8 @@ Status TransFormatFromNdToFracNz(const TransArgs &args, TransResult &result, con
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size](), std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
return ACL_ERROR_GE_MEMORY_ALLOCATION;
@ -175,7 +176,8 @@ Status TransFormatFromNdToFracNz(const TransArgs &args, TransResult &result, con
auto ret = memcpy_s(dst.get() + dst_offset, static_cast<size_t>(protected_size), args.data + src_offset,
static_cast<size_t>(size * w0));
if (ret != EOK) {
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
"Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
}
}
@ -189,7 +191,8 @@ Status TransFormatFromNdToFracNz(const TransArgs &args, TransResult &result, con
auto ret = memcpy_s(dst.get() + dst_offset, static_cast<size_t>(protected_size), args.data + src_offset,
static_cast<size_t>(size));
if (ret != EOK) {
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
"Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
}
}
@ -210,7 +213,8 @@ Status TransFormatFromFracNzToNd(const TransArgs &args, TransResult &result, con
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size], std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
return ACL_ERROR_GE_MEMORY_ALLOCATION;
@ -246,7 +250,8 @@ Status TransFormatFromFracNzToNd(const TransArgs &args, TransResult &result, con
ret = memcpy_s(dst.get() + dst_offset, static_cast<size_t>(protected_size), args.data + src_offset,
static_cast<size_t>(size * w0));
if (ret != EOK) {
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
"Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
}
}
@ -260,7 +265,8 @@ Status TransFormatFromFracNzToNd(const TransArgs &args, TransResult &result, con
ret = memcpy_s(dst.get() + dst_offset, static_cast<size_t>(protected_size), args.data + src_offset,
static_cast<size_t>(size));
if (ret != EOK) {
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
"Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
}
}
@ -274,14 +280,16 @@ Status TransFormatFromFracNzToNd(const TransArgs &args, TransResult &result, con
Status FormatTransferFractalNz::TransFormat(const TransArgs &args, TransResult &result) {
if (!IsDataTypeSupport(args.src_data_type)) {
GELOGE(ACL_ERROR_GE_DATATYPE_INVALID, "Trans format from %s to %s, src shape %s, dst shape %s, data type %s is not supported",
GELOGE(ACL_ERROR_GE_DATATYPE_INVALID,
"Trans format from %s to %s, src shape %s, dst shape %s, data type %s is not supported",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), ShapeToString(args.src_shape).c_str(),
ShapeToString(args.dst_shape).c_str(), TypeUtils::DataTypeToSerialString(args.src_data_type).c_str());
return ACL_ERROR_GE_DATATYPE_INVALID;
}
if (!CheckShape(args.src_format, args.src_shape) || !IsShapeValid(args.dst_shape)) {
GELOGE(ACL_ERROR_GE_SHAPE_INVALID, "Trans format from %s to %s, src shape %s, dst shape %s, data type %s is not supported",
GELOGE(ACL_ERROR_GE_SHAPE_INVALID,
"Trans format from %s to %s, src shape %s, dst shape %s, data type %s is not supported",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), ShapeToString(args.src_shape).c_str(),
ShapeToString(args.dst_shape).c_str(), TypeUtils::DataTypeToSerialString(args.src_data_type).c_str());
@ -325,7 +333,8 @@ Status FormatTransferFractalNz::TransShape(Format src_format, const ShapeVector
Status FormatTransferFractalNzND::TransFormat(const TransArgs &args, TransResult &result) {
if (!IsDataTypeSupport(args.src_data_type)) {
GELOGE(ACL_ERROR_GE_DATATYPE_INVALID, "Trans format from %s to %s, src shape %s, dst shape %s, data type %s is not supported",
GELOGE(ACL_ERROR_GE_DATATYPE_INVALID,
"Trans format from %s to %s, src shape %s, dst shape %s, data type %s is not supported",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), ShapeToString(args.src_shape).c_str(),
ShapeToString(args.dst_shape).c_str(), TypeUtils::DataTypeToSerialString(args.src_data_type).c_str());
@ -333,7 +342,8 @@ Status FormatTransferFractalNzND::TransFormat(const TransArgs &args, TransResult
}
if (!IsShapeValid(args.src_shape) || !CheckShape(args.dst_format, args.dst_shape)) {
GELOGE(ACL_ERROR_GE_SHAPE_INVALID, "Trans format from %s to %s, src shape %s, dst shape %s, data type %s is not supported",
GELOGE(ACL_ERROR_GE_SHAPE_INVALID,
"Trans format from %s to %s, src shape %s, dst shape %s, data type %s is not supported",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), ShapeToString(args.src_shape).c_str(),
ShapeToString(args.dst_shape).c_str(), TypeUtils::DataTypeToSerialString(args.src_data_type).c_str());

26
ge/common/formats/format_transfers/format_transfer_fractal_z.cc
Unescape View File

@ -127,7 +127,8 @@ Status TransFormatFromNchwToFz(const TransArgs &args, TransResult &result) {
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size], std::default_delete<uint8_t[]>());
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
dst == nullptr,
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
return ACL_ERROR_GE_MEMORY_ALLOCATION;);
@ -173,8 +174,9 @@ Status TransFormatFromNchwToFz(const TransArgs &args, TransResult &result) {
}
}
if (ret != EOK) {
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d pad mode %d", offset,
ret, need_pad_zero);
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
"Failed to operate the dst memory at offset %ld, error-code %d pad mode %d",
offset, ret, need_pad_zero);
return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
}
}
@ -213,7 +215,8 @@ Status TransFormatHwcnToFz(const TransArgs &args, TransResult &result) {
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size], std::default_delete<uint8_t[]>());
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
dst == nullptr,
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
return ACL_ERROR_GE_MEMORY_ALLOCATION;);
@ -235,7 +238,8 @@ Status TransFormatHwcnToFz(const TransArgs &args, TransResult &result) {
static_cast<size_t>(data_size));
} else {
if (protected_size < data_size) {
GELOGE(ACL_ERROR_GE_PARAM_INVALID, "Failed to operate the dst memory, protected_size is %ld and size is %ld",
GELOGE(ACL_ERROR_GE_PARAM_INVALID,
"Failed to operate the dst memory, protected_size is %ld and size is %ld",
protected_size, data_size);
return ACL_ERROR_GE_PARAM_INVALID;
}
@ -247,7 +251,8 @@ Status TransFormatHwcnToFz(const TransArgs &args, TransResult &result) {
}
}
if (ret != EOK) {
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d, pad mode %d",
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
"Failed to operate the dst memory at offset %ld, error-code %d, pad mode %d",
dst_offset, ret, pad_zero);
return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
}
@ -288,7 +293,8 @@ Status TransFormatNhwcToFz(const TransArgs &args, TransResult &result) {
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size], std::default_delete<uint8_t[]>());
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
dst == nullptr,
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
return ACL_ERROR_GE_MEMORY_ALLOCATION;);
@ -310,7 +316,8 @@ Status TransFormatNhwcToFz(const TransArgs &args, TransResult &result) {
static_cast<size_t>(data_size));
} else {
if (protected_size < data_size) {
GELOGE(ACL_ERROR_GE_PARAM_INVALID, "Failed to operate the dst memory, protected_size is %ld and size is %ld",
GELOGE(ACL_ERROR_GE_PARAM_INVALID,
"Failed to operate the dst memory, protected_size is %ld and size is %ld",
protected_size, data_size);
return ACL_ERROR_GE_PARAM_INVALID;
}
@ -322,7 +329,8 @@ Status TransFormatNhwcToFz(const TransArgs &args, TransResult &result) {
}
}
if (ret != EOK) {
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d, pad mode %d",
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
"Failed to operate the dst memory at offset %ld, error-code %d, pad mode %d",
dst_offset, ret, pad_zero);
return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
}

30
ge/common/formats/format_transfers/format_transfer_fractal_zz.cc
Unescape View File

@ -140,7 +140,8 @@ Status TransFormatFromNdToFracZz(const TransArgs &args, TransResult &result, con
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size](), std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
return ACL_ERROR_GE_MEMORY_ALLOCATION;
@ -179,7 +180,8 @@ Status TransFormatFromNdToFracZz(const TransArgs &args, TransResult &result, con
auto ret = memcpy_s(dst.get() + dst_offset, static_cast<size_t>(protected_size), args.data + src_offset,
static_cast<size_t>(size * w0));
if (ret != EOK) {
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
"Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
}
}
@ -195,7 +197,8 @@ Status TransFormatFromNdToFracZz(const TransArgs &args, TransResult &result, con
auto ret = memcpy_s(dst.get() + dst_offset, static_cast<size_t>(protected_size), args.data + src_offset,
static_cast<size_t>(size));
if (ret != EOK) {
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
"Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
}
}
@ -217,7 +220,8 @@ Status TransFormatFromFracZzToNd(const TransArgs &args, TransResult &result, con
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size](), std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
return ACL_ERROR_GE_MEMORY_ALLOCATION;
@ -257,7 +261,8 @@ Status TransFormatFromFracZzToNd(const TransArgs &args, TransResult &result, con
auto ret = memcpy_s(dst.get() + dst_offset, static_cast<size_t>(protected_size), args.data + src_offset,
static_cast<size_t>(size * w0));
if (ret != EOK) {
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
"Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
}
}
@ -273,7 +278,8 @@ Status TransFormatFromFracZzToNd(const TransArgs &args, TransResult &result, con
auto ret = memcpy_s(dst.get() + dst_offset, static_cast<size_t>(protected_size), args.data + src_offset,
static_cast<size_t>(size));
if (ret != EOK) {
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED,
"Failed to operate the dst memory at offset %ld, error-code %d", dst_offset, ret);
return ACL_ERROR_GE_MEMORY_OPERATE_FAILED;
}
}
@ -288,14 +294,16 @@ Status TransFormatFromFracZzToNd(const TransArgs &args, TransResult &result, con
Status FormatTransferFractalZz::TransFormat(const TransArgs &args, TransResult &result) {
if (!IsDataTypeSupport(args.src_data_type)) {
GELOGE(ACL_ERROR_GE_DATATYPE_INVALID, "Not support trans format from %s to %s, src shape %s, dst shape %s, data type %s",
GELOGE(ACL_ERROR_GE_DATATYPE_INVALID,
"Not support trans format from %s to %s, src shape %s, dst shape %s, data type %s",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), ShapeToString(args.src_shape).c_str(),
ShapeToString(args.dst_shape).c_str(), TypeUtils::DataTypeToSerialString(args.src_data_type).c_str());
return ACL_ERROR_GE_DATATYPE_INVALID;
}
if (!CheckShape(args.src_format, args.src_shape) || !IsShapeValid(args.dst_shape)) {
GELOGE(ACL_ERROR_GE_SHAPE_INVALID, "Not support trans format from %s to %s, src shape %s, dst shape %s, data type %s",
GELOGE(ACL_ERROR_GE_SHAPE_INVALID,
"Not support trans format from %s to %s, src shape %s, dst shape %s, data type %s",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), ShapeToString(args.src_shape).c_str(),
ShapeToString(args.dst_shape).c_str(), TypeUtils::DataTypeToSerialString(args.src_data_type).c_str());
@ -339,7 +347,8 @@ Status FormatTransferFractalZz::TransShape(Format src_format, const ShapeVector
Status FormatTransferFractalZzND::TransFormat(const TransArgs &args, TransResult &result) {
if (!IsDataTypeSupport(args.src_data_type)) {
GELOGE(ACL_ERROR_GE_DATATYPE_INVALID, "Not support trans format from %s to %s, src shape %s, dst shape %s, data type %s",
GELOGE(ACL_ERROR_GE_DATATYPE_INVALID,
"Not support trans format from %s to %s, src shape %s, dst shape %s, data type %s",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), ShapeToString(args.src_shape).c_str(),
ShapeToString(args.dst_shape).c_str(), TypeUtils::DataTypeToSerialString(args.src_data_type).c_str());
@ -347,7 +356,8 @@ Status FormatTransferFractalZzND::TransFormat(const TransArgs &args, TransResult
}
if (!IsShapeValid(args.src_shape) || !CheckShape(args.dst_format, args.dst_shape)) {
GELOGE(ACL_ERROR_GE_SHAPE_INVALID, "Not support trans format from %s to %s, src shape %s, dst shape %s, data type %s",
GELOGE(ACL_ERROR_GE_SHAPE_INVALID,
"Not support trans format from %s to %s, src shape %s, dst shape %s, data type %s",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), ShapeToString(args.src_shape).c_str(),
ShapeToString(args.dst_shape).c_str(), TypeUtils::DataTypeToSerialString(args.src_data_type).c_str());

3
ge/common/formats/format_transfers/format_transfer_fracz_hwcn.cc
Unescape View File

@ -74,7 +74,8 @@ Status CheckArgsForFracZToHwcn(const TransArgs &args) {
Status GetDstDataAfterTrans(const TransArgs &args, TransResult &result, const int size, const int64_t total_size) {
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), total_size, ShapeToString(args.dst_shape).c_str());
return ACL_ERROR_GE_MEMORY_ALLOCATION;

10
ge/common/formats/format_transfers/format_transfer_fracz_nchw.cc
Unescape View File

@ -59,9 +59,10 @@ Status CheckArgsForFracZToNchw(const TransArgs &args) {
}
int64_t c1 = Ceil(dst_shape.at(kNchwC), c0);
int64_t n0 = Ceil(dst_shape.at(kNchwN), static_cast<int64_t>(kNiSize));
if (src_shape.at(kFracZHWC1) != dst_shape.at(kNchwH) * dst_shape.at(kNchwW) * c1 || src_shape.at(kFracZC0) != c0 ||
src_shape.at(kFracZNi) != kNiSize || src_shape.at(kFracZN0) != n0) {
GELOGE(ACL_ERROR_GE_SHAPE_INVALID, "Failed to check relationship between src and dst shape, src shape %s, dst shape %s",
if (src_shape.at(kFracZHWC1) != dst_shape.at(kNchwH) * dst_shape.at(kNchwW) * c1 ||
src_shape.at(kFracZC0) != c0 || src_shape.at(kFracZNi) != kNiSize || src_shape.at(kFracZN0) != n0) {
GELOGE(ACL_ERROR_GE_SHAPE_INVALID,
"Failed to check relationship between src and dst shape, src shape %s, dst shape %s",
ShapeToString(src_shape).c_str(), ShapeToString(dst_shape).c_str());
return ACL_ERROR_GE_SHAPE_INVALID;
}
@ -72,7 +73,8 @@ Status CheckArgsForFracZToNchw(const TransArgs &args) {
Status GetDstDataAfterTrans(const TransArgs &args, TransResult &result, const int size, const int64_t total_size) {
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), total_size, ShapeToString(args.dst_shape).c_str());
return ACL_ERROR_GE_MEMORY_ALLOCATION;

10
ge/common/formats/format_transfers/format_transfer_fracz_nhwc.cc
Unescape View File

@ -59,9 +59,10 @@ Status CheckArgsForFracZToNhwc(const TransArgs &args) {
}
int64_t c1 = Ceil(dst_shape.at(kNhwcC), c0);
int64_t n0 = Ceil(dst_shape.at(kNhwcN), static_cast<int64_t>(kNiSize));
if (src_shape.at(kFracZHWC1) != dst_shape.at(kNhwcH) * dst_shape.at(kNhwcW) * c1 || src_shape.at(kFracZC0) != c0 ||
src_shape.at(kFracZNi) != kNiSize || src_shape.at(kFracZN0) != n0) {
GELOGE(PARAM_INVALID, "Failed to check relationship between src and dst shape, src shape %s, dst shape %s",
if (src_shape.at(kFracZHWC1) != dst_shape.at(kNhwcH) * dst_shape.at(kNhwcW) * c1 ||
src_shape.at(kFracZC0) != c0 || src_shape.at(kFracZNi) != kNiSize || src_shape.at(kFracZN0) != n0) {
GELOGE(PARAM_INVALID,
"Failed to check relationship between src and dst shape, src shape %s, dst shape %s",
ShapeToString(src_shape).c_str(), ShapeToString(dst_shape).c_str());
return PARAM_INVALID;
}
@ -72,7 +73,8 @@ Status CheckArgsForFracZToNhwc(const TransArgs &args) {
Status GetDstDataAfterTrans(const TransArgs &args, TransResult &result, int size, int64_t total_size) {
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(OUT_OF_MEMORY, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
GELOGE(OUT_OF_MEMORY,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), total_size, ShapeToString(args.dst_shape).c_str());
return OUT_OF_MEMORY;

3
ge/common/formats/format_transfers/format_transfer_hwcn_c1hwncoc0.cc
Unescape View File

@ -91,7 +91,8 @@ Status CheckArgsForHwcnToC1hwncoc0(const TransArgs &args) {
Status GetDstDataAfterTrans(const TransArgs &args, TransResult &result, const int size, const int64_t total_size) {
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), total_size, ShapeToString(args.dst_shape).c_str());
return ACL_ERROR_GE_MEMORY_ALLOCATION;

3
ge/common/formats/format_transfers/format_transfer_nc1hwc0_nchw.cc
Unescape View File

@ -72,7 +72,8 @@ Status CheckArgsForNc1hwc0ToNchw(const TransArgs &args) {
Status GetDstDataAfterTrans(const TransArgs &args, TransResult &result, const int size, const int64_t total_size) {
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(OUT_OF_MEMORY, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
GELOGE(OUT_OF_MEMORY,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), total_size, ShapeToString(args.dst_shape).c_str());
return OUT_OF_MEMORY;

6
ge/common/formats/format_transfers/format_transfer_nc1hwc0_nhwc.cc
Unescape View File

@ -61,7 +61,8 @@ Status CheckArgsForNc1hwc0ToNhwc(const TransArgs &args) {
if (src_shape.at(kNc1hwc0H) != dst_shape.at(kNhwcH) || src_shape.at(kNc1hwc0W) != dst_shape.at(kNhwcW) ||
src_shape.at(kNc1hwc0N) != dst_shape.at(kNhwcN) || src_shape.at(kNc1hwc0C0) != c0 ||
src_shape.at(kNc1hwc0C1) != (Ceil(dst_shape.at(kNhwcC), c0))) {
GELOGE(ACL_ERROR_GE_SHAPE_INVALID, "Failed to check relationship between src and dst shape, src shape %s, dst shape %s",
GELOGE(ACL_ERROR_GE_SHAPE_INVALID,
"Failed to check relationship between src and dst shape, src shape %s, dst shape %s",
ShapeToString(src_shape).c_str(), ShapeToString(dst_shape).c_str());
return ACL_ERROR_GE_SHAPE_INVALID;
}
@ -72,7 +73,8 @@ Status CheckArgsForNc1hwc0ToNhwc(const TransArgs &args) {
Status GetDstDataAfterTrans(const TransArgs &args, TransResult &result, const int size, const int64_t total_size) {
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), total_size, ShapeToString(args.dst_shape).c_str());
return ACL_ERROR_GE_MEMORY_ALLOCATION;

15
ge/common/formats/format_transfers/format_transfer_nchw_fz_c04.cc
Unescape View File

@ -125,7 +125,8 @@ Status TransFormatFromNchwToFzC04(const TransArgs &args, TransResult &result) {
return ACL_ERROR_GE_INTERNAL_ERROR);
auto t1 = h_o * w_o;
auto t2 = n_o * c_o;
GE_IF_BOOL_EXEC(!CheckInt64MulOverflow(t1, t2), GELOGE(INTERNAL_ERROR, "int64 mul overflow.A[%ld], B[%ld]", t1, t2);
GE_IF_BOOL_EXEC(!CheckInt64MulOverflow(t1, t2),
GELOGE(INTERNAL_ERROR, "int64 mul overflow.A[%ld], B[%ld]", t1, t2);
return ACL_ERROR_GE_INTERNAL_ERROR);
int64_t total_ele_cnt = n_o * c_o * h_o * w_o;
@ -140,7 +141,8 @@ Status TransFormatFromNchwToFzC04(const TransArgs &args, TransResult &result) {
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[dst_size], std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
return ACL_ERROR_GE_MEMORY_ALLOCATION;
@ -212,7 +214,8 @@ Status PaddingNC(const TransArgs &args, TransArgs &args_tmp, std::shared_ptr<uin
return ACL_ERROR_GE_INTERNAL_ERROR);
auto t1 = h_o * w_o;
auto t2 = n_o * c_o;
GE_IF_BOOL_EXEC(!CheckInt64MulOverflow(t1, t2), GELOGE(ACL_ERROR_GE_INTERNAL_ERROR, "int64 mul overflow.A[%ld], B[%ld]", t1, t2);
GE_IF_BOOL_EXEC(!CheckInt64MulOverflow(t1, t2),
GELOGE(ACL_ERROR_GE_INTERNAL_ERROR,"int64 mul overflow.A[%ld], B[%ld]", t1, t2);
return ACL_ERROR_GE_INTERNAL_ERROR);
int64_t total_ele_cnt = n_o * c_o * h_o * w_o;
@ -228,7 +231,8 @@ Status PaddingNC(const TransArgs &args, TransArgs &args_tmp, std::shared_ptr<uin
dst.reset(new (std::nothrow) uint8_t[dst_size], std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), dst_size);
return ACL_ERROR_GE_MEMORY_ALLOCATION;
@ -275,7 +279,8 @@ Status FormatTransferNchwToFZC04::TransFormat(const TransArgs &args, TransResult
}
std::vector<int64_t> expect_shape;
ret = TransShape(args_tmp.src_format, args_tmp.src_shape, args_tmp.src_data_type, args_tmp.dst_format, expect_shape);
ret = TransShape(args_tmp.src_format, args_tmp.src_shape, args_tmp.src_data_type,
args_tmp.dst_format, expect_shape);
if (ret != SUCCESS) {
return ret;
}

3
ge/common/formats/format_transfers/format_transfer_nhwc_nc1hwc0.cc
Unescape View File

@ -92,7 +92,8 @@ Status CheckArgsForNhwcToNc1hwc0(const TransArgs &args) {
Status GetDstDataAfterTrans(const TransArgs &args, TransResult &result, const int size, const int64_t total_size) {
std::shared_ptr<uint8_t> dst(new (std::nothrow) uint8_t[total_size], std::default_delete<uint8_t[]>());
if (dst == nullptr) {
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION, "Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
GELOGE(ACL_ERROR_GE_MEMORY_ALLOCATION,
"Failed to trans format from %s to %s, can not alloc the memory for dst buf %ld, shape %s",
TypeUtils::FormatToSerialString(args.src_format).c_str(),
TypeUtils::FormatToSerialString(args.dst_format).c_str(), total_size, ShapeToString(args.dst_shape).c_str());
return ACL_ERROR_GE_MEMORY_ALLOCATION;

4
ge/generator/ge_generator.cc
Unescape View File

@ -592,8 +592,8 @@ Status GeGenerator::SetModelNameForDump(const GeRootModelPtr &ge_root_model) {
ErrorManager::GetInstance().ATCReportErrMessage("E10000", {"parameter"}, {"output"});
GELOGE(FAILED, "[Check][GetModelNameStep]Get model_name failed. Param --output is invalid, root graph name: %s",
ge_root_model->GetRootGraph()->GetName().c_str());
REPORT_CALL_ERROR("E19999", "Get model_name failed. Param --output is invalid,",
"root graph name: %s", ge_root_model->GetRootGraph()->GetName().c_str());
REPORT_CALL_ERROR("E19999", "Get model_name failed. Param --output is invalid, root graph name: %s",
ge_root_model->GetRootGraph()->GetName().c_str());
return PARAM_INVALID;
}
map<string, GeModelPtr> name_to_ge_model = ge_root_model->GetSubgraphInstanceNameToModel();

66
ge/graph/build/memory/block_mem_assigner.cc
Unescape View File

@ -597,9 +597,11 @@ void BlockMemAssigner::GetOutAndWorkSpaceMem(vector<int64_t> &all_memory_size) {
int64_t size = 0;
GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(output_desc, size) != SUCCESS, GELOGI("Get size failed"));
GE_IF_BOOL_EXEC(size < 0,
GELOGE(FAILED, "[Check][TensorSize]tensor_size:%ld is invalid, maybe it is unknown shape node, Node_name:%s",
GELOGE(FAILED, "[Check][TensorSize]tensor_size:%ld is invalid, "
"maybe it is unknown shape node, Node_name:%s",
size, node_op_desc->GetName().c_str());
REPORT_INNER_ERROR("E19999", "tensor_size:%ld is invalid, maybe it is unknown shape node, Node_name:%s",
REPORT_INNER_ERROR("E19999", "tensor_size:%ld is invalid, "
"maybe it is unknown shape node, Node_name:%s",
size, node_op_desc->GetName().c_str());
return;);
batch_all_memory_size[batch_label].emplace_back(size);
@ -1105,7 +1107,8 @@ MemoryBlock *BlockMemAssigner::ApplyMemory(size_t block_size, size_t real_size,
OpMemoryType mem_type, const NodePtr &n, uint32_t out_index,
const vector<bool> &workspace_reuse_flag, const bool is_op_reuse_mem,
const bool continuous, int64_t memory_type) {
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(n == nullptr,
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
n == nullptr,
REPORT_INNER_ERROR("E19999", "Input parameter n(type:node_ptr) is null, apply memory failed");
return nullptr, "[Check][Param]Input parameter n(type:node_ptr) is null.");
auto node_op_desc = n->GetOpDesc();
@ -1159,10 +1162,12 @@ MemoryBlock *BlockMemAssigner::ApplyMemory(size_t block_size, size_t real_size,
}
auto block = new (std::nothrow) MemoryBlock(block_size, node_op_desc->GetStreamId(), is_reuse_memory, memory_type);
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(block == nullptr,
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
block == nullptr,
REPORT_INNER_ERROR("E19999", "new a memoryblock object failed. node_name:%s out_index:%u",
n->GetName().c_str(), out_index);
return nullptr, "[New][Object]new MemoryBlock failed, node_name:%s out_index:%u", n->GetName().c_str(), out_index);
return nullptr,
"[New][Object]new MemoryBlock failed, node_name:%s out_index:%u", n->GetName().c_str(), out_index);
// Data and netoutput need zero copy block
block->is_zero_copy_ = IsZeroCopyBlock(n, continuous);
@ -1221,11 +1226,13 @@ void BlockMemAssigner::ContinuousOutRefCheck(bool &isAllOutputRef, bool &isOutpu
Status BlockMemAssigner::ApplyContinuousMemory(const NodePtr &n, const vector<int64_t> &ranges,
const bool is_op_reuse_mem) {
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(n == nullptr,
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
n == nullptr,
REPORT_INNER_ERROR("E19999", "Input parameter n(type:node_ptr) is null");
return INTERNAL_ERROR, "[check][param]Input parameter n(type:NodePtr) is null.");
auto node_op_desc = n->GetOpDesc();
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(node_op_desc == nullptr,
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
node_op_desc == nullptr,
REPORT_INNER_ERROR("E19999", "Input parameter n(type:OpDescPtr) is null");
return INTERNAL_ERROR, "[Check][Param]Input parameter n(type:OpDescPtr) is null");
@ -1319,26 +1326,33 @@ Status BlockMemAssigner::ApplyContinuousMemory(const NodePtr &n, const vector<in
MemoryBlock *BlockMemAssigner::ApplyOutMemory(const NodePtr &n, uint32_t index, const vector<int64_t> &ranges,
const bool is_op_reuse_mem, const bool continuous) {
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(n == nullptr,
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
n == nullptr,
REPORT_INNER_ERROR("E19999", "Input parameter n(type:NodePtr) is null");
return nullptr, "[Check][Param]Input parameter n(type:NodePtr) is null");
auto node_op_desc = n->GetOpDesc();
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(node_op_desc == nullptr,
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
node_op_desc == nullptr,
REPORT_INNER_ERROR("E19999", "Input parameter n(type:OpDescPtr) is null");
return nullptr, "[Check][Param]Input parameter n(type:OpDescPtr) is null");
MemoryBlock *block = nullptr;
NodeIndexIO node_index_io(n, index, kOut);
int64_t size = 0;
auto output_op_desc = node_op_desc->GetOutputDescPtr(index);
GE_IF_BOOL_EXEC(output_op_desc == nullptr,
REPORT_INNER_ERROR("E19999", "get output_desc failed, node_name:%s, output_index:%u", n->GetName().c_str(), index);
GE_IF_BOOL_EXEC(
output_op_desc == nullptr,
REPORT_INNER_ERROR("E19999", "get output_desc failed, node_name:%s, output_index:%u",
n->GetName().c_str(), index);
GELOGE(FAILED, "[Get][OutputDesc]node_name:%s, output_index:%u", n->GetName().c_str(), index);
return nullptr);
GE_IF_BOOL_EXEC(ge::TensorUtils::GetSize(*output_op_desc, size) != SUCCESS, GELOGI("Get size failed"));
size_t no_align_size = 0;
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(GetNoAlignSize(*node_op_desc, index, no_align_size) != SUCCESS,
REPORT_CALL_ERROR("E19999", "Get no align size failed, node_name:%s, output_index:%u", n->GetName().c_str(), index);
return nullptr, "[Get][TensorSize]Get no align size, node_name:%s, output_index:%u", n->GetName().c_str(), index);
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
GetNoAlignSize(*node_op_desc, index, no_align_size) != SUCCESS,
REPORT_CALL_ERROR("E19999", "Get no align size failed, node_name:%s, output_index:%u",
n->GetName().c_str(), index);
return nullptr,
"[Get][TensorSize]Get no align size, node_name:%s, output_index:%u", n->GetName().c_str(), index);
std::string symbol;
bool reuse_input = false;
@ -1392,21 +1406,25 @@ MemoryBlock *BlockMemAssigner::ApplyOutMemory(const NodePtr &n, uint32_t index,
vector<bool> workspace_reuse_flag;
block = ApplyMemory(block_size, size, no_align_size, kOutput, n, index,
workspace_reuse_flag, is_op_reuse_mem, continuous, memory_type);
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(block == nullptr,
GE_CHK_BOOL_TRUE_EXEC_WITH_LOG(
block == nullptr,
REPORT_CALL_ERROR("E19999", "apply out Memory failed, node_name:%s, block_size:%ld, out_index:%u",
n->GetName().c_str(), block_size, index);
return nullptr, "[Apply][Memory]node_name:%s, block_size:%ld, out_index:%u",
return nullptr,
"[Apply][Memory]node_name:%s, block_size:%ld, out_index:%u",
n->GetName().c_str(), block_size, index);
}
int out_count = 0;
GE_IF_BOOL_EXEC(index >= n->GetAllOutDataAnchors().size(),
GE_IF_BOOL_EXEC(
index >= n->GetAllOutDataAnchors().size(),
REPORT_INNER_ERROR("E19999", "out index:%u exceed out_size:%lu, node_name:%s",
index, n->GetAllOutDataAnchors().size(), n->GetName().c_str());
GELOGE(FAILED, "[Check][OutIndex]index:%u exceed out_size:%lu, node_name:%s",
index, n->GetAllOutDataAnchors().size(), n->GetName().c_str());
return nullptr);
auto out_data_anchor = n->GetOutDataAnchor(index);
GE_IF_BOOL_EXEC(out_data_anchor == nullptr,
GE_IF_BOOL_EXEC(
out_data_anchor == nullptr,
REPORT_INNER_ERROR("E19999", "out anchor is null, index:%u, node_name:%s", index, n->GetName().c_str());
GELOGE(FAILED, "[Check][OutAnchor]is null, index:%u, node_name:%s", index, n->GetName().c_str());
return nullptr);
@ -1618,7 +1636,8 @@ Status BlockMemAssigner::AssignOutputMemoryWithReuse(const NodePtr &node, vector
REPORT_INNER_ERROR("E19999", "Attr[%s] size:%zu not equal to node output size:%zu, node_name:%s",
ATTR_NAME_OUTPUT_MEM_TYPE_LIST.c_str(), memorys_type.size(),
op_desc->GetOutputsSize(), op_desc->GetName().c_str());
GELOGE(INTERNAL_ERROR,
GELOGE(
INTERNAL_ERROR,
"[Check][MemTypeAttr]Attr %s size:%zu not equal to node output size:%zu, node_name:%s",
ATTR_NAME_OUTPUT_MEM_TYPE_LIST.c_str(), memorys_type.size(),
op_desc->GetOutputsSize(), op_desc->GetName().c_str());
@ -1748,9 +1767,11 @@ void BlockMemAssigner::AssignMemoryWithReuse(vector<int64_t> &ranges) {
if (has_tvm_workspace_mem_type_attr && (temp.size() != tvm_workspace_memory_type.size())) {
REPORT_INNER_ERROR("E19999", "Attr[%s]size:%zu is not equal to workspace size:%zu, node_name:%s",
TVM_ATTR_NAME_WORKSPACE_TYPE.c_str(), tvm_workspace_memory_type.size(), temp.size(), n->GetName().c_str());
TVM_ATTR_NAME_WORKSPACE_TYPE.c_str(), tvm_workspace_memory_type.size(),
temp.size(), n->GetName().c_str());
GELOGE(INTERNAL_ERROR, "[Check][Attr]Attr %s size:%zu is not equal to workspace size:%zu, node_name:%s",
TVM_ATTR_NAME_WORKSPACE_TYPE.c_str(), tvm_workspace_memory_type.size(), temp.size(), n->GetName().c_str());
TVM_ATTR_NAME_WORKSPACE_TYPE.c_str(), tvm_workspace_memory_type.size(),
temp.size(), n->GetName().c_str());
return;
}
for (size_t i = 0; i < temp.size(); i++) {
@ -2161,7 +2182,8 @@ bool BlockMemAssigner::GetWorkSpaceMemoryType(const NodePtr &node, size_t index,
if (has_workspace_mem_type_attr && (workspace_memory_type.size() <= index)) {
REPORT_INNER_ERROR("E19999", "get workspace mem_type failed, "
"index %zu invalid, bigger than attr %s size:%zu, node_name:%s",
index, TVM_ATTR_NAME_WORKSPACE_TYPE.c_str(), workspace_memory_type.size(), node->GetName().c_str());
index, TVM_ATTR_NAME_WORKSPACE_TYPE.c_str(),
workspace_memory_type.size(), node->GetName().c_str());
GELOGE(INTERNAL_ERROR, "[Get][WorkspaceMemType]index %zu invalid, bigger than attr %s size:%zu, node_name:%s",
index, TVM_ATTR_NAME_WORKSPACE_TYPE.c_str(), workspace_memory_type.size(), node->GetName().c_str());
return false;

31
ge/graph/load/model_manager/davinci_model.cc
Unescape View File

@ -2875,9 +2875,7 @@ Status DavinciModel::UpdateKnownNodeArgs(const vector<void *> &inputs, const vec
GELOGI("DavinciModel::UpdateKnownNodeArgs in");
GE_CHK_STATUS_RET(CreateKnownZeroCopyMap(inputs, outputs),
"DavinciModel::UpdateKnownNodeArgs create map for input/output zero copy.");
if (!base_addr_not_changed_) {
total_io_addrs_.clear();
orig_total_io_addrs_.clear();
for (size_t task_index = 0; task_index < task_list_.size(); ++task_index) {
auto &task = task_list_[task_index];
if (task != nullptr) {
@ -2888,11 +2886,6 @@ Status DavinciModel::UpdateKnownNodeArgs(const vector<void *> &inputs, const vec
}
}
}
// cache latest iterator io addr
orig_total_io_addrs_ = total_io_addrs_;
} else {
total_io_addrs_ = orig_total_io_addrs_;
}
GE_CHK_STATUS_RET(UpdateKnownZeroCopyAddr(total_io_addrs_, false), "DavinciModel::UpdateKnownZeroCopyAddr failed.");
if (total_args_size_ == 0) {
@ -2931,6 +2924,14 @@ Status DavinciModel::InitTaskInfo(domi::ModelTaskDef &model_task_def) {
return SUCCESS;
}
Status DavinciModel::CheckCapability(rtFeatureType_t featureType, int32_t featureInfo, bool &is_support) const {
int64_t value = RT_CAPABILITY_SUPPORT;
auto rt_ret = rtGetRtCapability(featureType, featureInfo, &value);
GE_CHK_BOOL_RET_STATUS(rt_ret == RT_ERROR_NONE, FAILED, "call rtGetRtCapability failed!");
is_support = (value == RT_CAPABILITY_SUPPORT) ? true : false;
return SUCCESS;
}
Status DavinciModel::MallocKnownArgs() {
GELOGI("DavinciModel::MallocKnownArgs in");
const auto &model_task_def = ge_model_->GetModelTaskDefPtr();
@ -2949,20 +2950,22 @@ Status DavinciModel::MallocKnownArgs() {
return ret;
}
}
rtError_t rt_ret;
// malloc args memory
if (total_args_size_ == 0) {
GELOGW("DavinciModel::MallocKnownArgs total_args_size_ equals to zero.");
return SUCCESS;
}
bool is_support = false;
GE_CHK_STATUS_RET_NOLOG(CheckCapability(FEATURE_TYPE_MEMORY, MEMORY_INFO_TS_4G_LIMITED, is_support));
auto mem_type = is_support ? RT_MEMORY_TS_4G : RT_MEMORY_HBM;
rtError_t rt_ret = rtMalloc(&args_, total_args_size_, RT_MEMORY_HBM);
if (total_args_size_ != 0) {
rt_ret = rtMalloc(&args_, total_args_size_, mem_type);
if (rt_ret != RT_ERROR_NONE) {
GELOGE(RT_FAILED, "Call rtMalloc failed, ret: 0x%X", rt_ret);
return RT_ERROR_TO_GE_STATUS(rt_ret);
}
}
// malloc dynamic and static hybrid memory
if (total_hybrid_args_size_ != 0) {
rt_ret = rtMalloc(&hybrid_addrs_, total_hybrid_args_size_, RT_MEMORY_HBM);
rt_ret = rtMalloc(&hybrid_addrs_, total_hybrid_args_size_, mem_type);
if (rt_ret != RT_ERROR_NONE) {
GELOGE(RT_FAILED, "Call rtMalloc failed, ret: 0x%X", rt_ret);
return RT_ERROR_TO_GE_STATUS(rt_ret);
@ -2971,7 +2974,7 @@ Status DavinciModel::MallocKnownArgs() {
// malloc fixed addr memory, eg: rts op
if (total_fixed_addr_size_ != 0) {
GELOGI("Begin to allocate fixed addr.");
rt_ret = rtMalloc(&fixed_addrs_, total_fixed_addr_size_, RT_MEMORY_HBM);
rt_ret = rtMalloc(&fixed_addrs_, total_fixed_addr_size_, mem_type);
if (rt_ret != RT_ERROR_NONE) {
GELOGE(RT_FAILED, "Call rtMalloc failed, ret: 0x%X", rt_ret);
return RT_ERROR_TO_GE_STATUS(rt_ret);

4
ge/graph/load/model_manager/davinci_model.h
Unescape View File

@ -530,11 +530,11 @@ class DavinciModel {
}
void SetKnownNode(bool known_node) { known_node_ = known_node; }
bool IsKnownNode() { return known_node_; }
Status CheckCapability(rtFeatureType_t featureType, int32_t featureInfo, bool &is_support) const;
Status MallocKnownArgs();
Status UpdateKnownNodeArgs(const vector<void *> &inputs, const vector<void *> &outputs);
Status CreateKnownZeroCopyMap(const vector<void *> &inputs, const vector<void *> &outputs);
Status UpdateKnownZeroCopyAddr(vector<void *> &total_io_addrs, bool update_args = true);
void SetKnownNodeAddrNotChanged(bool base_addr_not_changed) { base_addr_not_changed_ = base_addr_not_changed; }
Status GetOrigInputInfo(uint32_t index, OriginInputInfo &orig_input_info) const;
Status GetAllAippInputOutputDims(uint32_t index, vector<InputOutputDims> &input_dims,
@ -1007,8 +1007,6 @@ class DavinciModel {
map<const void *, void *> known_input_data_info_;
map<const void *, void *> known_output_data_info_;
vector<void *> total_io_addrs_;
vector<void *> orig_total_io_addrs_;
bool base_addr_not_changed_ = false;
vector<vector<int64_t>> batch_info_;
vector<vector<int64_t>> combined_batch_info_;

3
ge/graph/load/model_manager/model_utils.cc
Unescape View File

@ -384,7 +384,8 @@ Status ModelUtils::GetVarAddr(const RuntimeParam &model_param, const ConstOpDesc
switch (mem_type) {
case RT_MEMORY_RDMA_HBM:
if (offset < 0) {
GELOGE(PARAM_INVALID, "rdma var addr is invalid, addr=%p", reinterpret_cast<uint8_t *>(offset));
GELOGE(PARAM_INVALID, "rdma var addr is invalid, addr=%p",
reinterpret_cast<uint8_t *>(static_cast<uintptr_t>(offset)));
return PARAM_INVALID;
}
var_addr = reinterpret_cast<uint8_t *>(static_cast<uintptr_t>(offset));

101
ge/graph/load/model_manager/task_info/kernel_task_info.cc
Unescape View File

@ -124,7 +124,8 @@ Status KernelTaskInfo::Init(const domi::TaskDef &task_def, DavinciModel *davinci
return FAILED;
}
ret = InitTVMTask(args_offset_tmp[0], kernel_def);
io_addr_offset_ = args_offset_tmp[0];
ret = InitTVMTask(io_addr_offset_, kernel_def);
} else if (kernel_type_ == ccKernelType::CUSTOMIZED) {
ret = InitAICPUCustomTask(context.op_index(), kernel_def);
} else if (kernel_type_ == ccKernelType::AI_CPU || kernel_type_ == ccKernelType::CUST_AI_CPU) {
@ -380,7 +381,8 @@ Status KernelTaskInfo::Distribute() {
GELOGD("KernelTaskInfo Distribute Start.");
if (davinci_model_->IsKnownNode()) {
if (kernel_type_ == ccKernelType::TE) {
args_ = davinci_model_->GetCurrentArgsAddr(args_offset_);
args_ = l2_buffer_on_ ? davinci_model_->GetCurrentHybridArgsAddr(hybrid_args_offset_)
: davinci_model_->GetCurrentArgsAddr(args_offset_);
} else if (kernel_type_ == ccKernelType::AI_CPU || kernel_type_ == ccKernelType::CUST_AI_CPU) {
args_ = davinci_model_->GetCurrentHybridArgsAddr(hybrid_args_offset_);
}
@ -449,29 +451,41 @@ void KernelTaskInfo::SetIoAddrs(const OpDescPtr &op_desc) {
}
}
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) {
Status KernelTaskInfo::CopyNoncontinuousArgs(uint16_t offset) {
GE_CHECK_NOTNULL(davinci_model_);
// copy new io addrs
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();
errno_t sec_ret = memcpy_s(reinterpret_cast<void *>(io_addr), addrs_size, io_addrs.data(), addrs_size);
auto addr_size = kAddrLen * io_addrs.size();
// copy io addr
errno_t sec_ret = memcpy_s(args_addr.get() + offset, addr_size, io_addrs.data(), addr_size);
if (sec_ret != EOK) {
GELOGE(FAILED, "memcpy failed, ret: %d", sec_ret);
return FAILED;
}
// copy args to device
rtError_t rt_ret = rtMemcpy(args_, args_size_, args_addr.get(), args_size_, RT_MEMCPY_HOST_TO_DEVICE);
if (rt_ret != RT_ERROR_NONE) {
GELOGE(RT_FAILED, "Call rt api(rtMemcpy) failed, ret: 0x%X", rt_ret);
return RT_ERROR_TO_GE_STATUS(rt_ret);
}
}
GELOGD("Copy noncontinuous args success, kernel type %d.", kernel_type_);
return SUCCESS;
}
GELOGI("KernelTaskInfo::UpdateArgs success.");
Status KernelTaskInfo::UpdateArgs() {
GELOGI("KernelTaskInfo::UpdateArgs in.");
GE_CHECK_NOTNULL(davinci_model_);
if (kernel_type_ == ccKernelType::TE) {
if (l2_buffer_on_) {
return CopyNoncontinuousArgs(io_addr_offset_);
}
davinci_model_->SetTotalIOAddrs(io_addrs_);
} else if (kernel_type_ == ccKernelType::AI_CPU || kernel_type_ == ccKernelType::CUST_AI_CPU) {
return CopyNoncontinuousArgs(sizeof(aicpu::AicpuParamHead));
}
return SUCCESS;
}
@ -516,8 +530,8 @@ Status KernelTaskInfo::UpdateL2Data(const domi::KernelDef &kernel_def) {
return SUCCESS;
}
char *sm_contrl = const_cast<char *>(sm_desc.data());
rtL2Ctrl_t *l2_ctrl_info = reinterpret_cast<rtL2Ctrl_t *>(sm_contrl);
char *sm_control = const_cast<char *>(sm_desc.data());
rtL2Ctrl_t *l2_ctrl_info = reinterpret_cast<rtL2Ctrl_t *>(sm_control);
uint64_t gen_base_addr = davinci_model_->GetRtBaseAddr();
// There is no weight for te op now. Update L2_mirror_addr by data memory base.
@ -545,19 +559,31 @@ Status KernelTaskInfo::UpdateL2Data(const domi::KernelDef &kernel_def) {
return SUCCESS;
}
void KernelTaskInfo::SetContinuousArgs(uint32_t args_size, DavinciModel *davinci_model) {
args_offset_ = davinci_model->GetTotalArgsSize();
davinci_model->SetTotalArgsSize(args_size);
}
void KernelTaskInfo::SetNoncontinuousArgs(uint32_t args_size, DavinciModel *davinci_model) {
hybrid_args_offset_ = davinci_model->GetHybridArgsSize();
davinci_model->SetHybridArgsSize(args_size);
}
Status KernelTaskInfo::CalculateArgs(const domi::TaskDef &task_def, DavinciModel *davinci_model) {
GE_CHECK_NOTNULL(davinci_model);
const domi::KernelDef &kernel_def = task_def.kernel();
const domi::KernelContext &context = kernel_def.context();
kernel_type_ = static_cast<ccKernelType>(context.kernel_type());
if (kernel_type_ == ccKernelType::TE) {
uint32_t args_size = kernel_def.args_size();
args_offset_ = davinci_model->GetTotalArgsSize();
davinci_model->SetTotalArgsSize(args_size);
GELOGI("kernel task name , args_size %u, args_offset %u", args_size, args_offset_);
if (kernel_type_ == ccKernelType::TE) {
if (kernel_def.sm_desc().empty()) {
SetContinuousArgs(args_size, davinci_model);
return SUCCESS;
}
l2_buffer_on_ = true;
SetNoncontinuousArgs(args_size, davinci_model);
} else if (kernel_type_ == ccKernelType::AI_CPU || kernel_type_ == ccKernelType::CUST_AI_CPU) {
hybrid_args_offset_ = davinci_model->GetHybridArgsSize();
davinci_model->SetHybridArgsSize(kernel_def.args_size());
GELOGI("aicpu kernel task name , args_size %u, args_offset %u", kernel_def.args_size(), hybrid_args_offset_);
SetNoncontinuousArgs(args_size, davinci_model);
}
return SUCCESS;
}
@ -568,8 +594,23 @@ Status KernelTaskInfo::InitTVMTask(uint16_t offset, const domi::KernelDef &kerne
// get tvm op desc
OpDescPtr op_desc = davinci_model_->GetOpByIndex(ctx_.opIndex);
GE_CHECK_NOTNULL(op_desc);
args_addr = std::unique_ptr<uint8_t[]>(new (std::nothrow) uint8_t[args_size_]);
errno_t sec_ret = memcpy_s(args_addr.get(), args_size_, kernel_def.args().data(), args_size_);
if (sec_ret != EOK) {
GELOGE(FAILED, "memcpy failed, ret: %d", sec_ret);
return FAILED;
}
Status ge_ret = UpdateL2Data(kernel_def);
// update origin l2 data
if (ge_ret != SUCCESS) {
return ge_ret;
}
if (davinci_model_->IsKnownNode()) {
args_ = davinci_model_->GetCurrentArgsAddr(args_offset_);
args_ = l2_buffer_on_ ? davinci_model_->GetCurrentHybridArgsAddr(hybrid_args_offset_)
: davinci_model_->GetCurrentArgsAddr(args_offset_);
InitDumpTask(offset);
return SUCCESS;
}
@ -609,12 +650,6 @@ Status KernelTaskInfo::InitTVMTask(uint16_t offset, const domi::KernelDef &kerne
GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
return RT_ERROR_TO_GE_STATUS(rt_ret);
}
vector<uint8_t> args_info(args_size_);
errno_t sec_ret = memcpy_s(args_info.data(), args_size_, kernel_def.args().data(), args_size_);
if (sec_ret != EOK) {
GELOGE(FAILED, "memcpy failed, ret: %d", sec_ret);
return FAILED;
}
if ((args_size_ <= offset) || (args_size_ - offset < kAddrLen * tensor_device_addrs.size())) {
GELOGE(FAILED, "offset >= kernelInfo.argsSize or copy content beyond applied memory.");
@ -628,7 +663,7 @@ Status KernelTaskInfo::InitTVMTask(uint16_t offset, const domi::KernelDef &kerne
GELOGE(RT_FAILED, "Call rt api failed, ret: 0x%X", rt_ret);
return RT_ERROR_TO_GE_STATUS(rt_ret);
}
sec_ret = memcpy_s(args_info.data() + offset, args_size_ - offset, tensor_device_addrs.data(),
sec_ret = memcpy_s(args_addr.get() + offset, args_size_ - offset, tensor_device_addrs.data(),
kAddrLen * tensor_device_addrs.size());
if (sec_ret != EOK) {
GELOGE(FAILED, "memcpy failed, ret: %d", sec_ret);
@ -640,19 +675,13 @@ Status KernelTaskInfo::InitTVMTask(uint16_t offset, const domi::KernelDef &kerne
GE_CHK_BOOL_TRUE_EXEC_INFO(davinci_model_->GetOpDugReg(), dump_args_ = static_cast<char *>(args_) + offset,
"Op debug is open in TVM task info");
Status ge_ret = UpdateL2Data(kernel_def);
// update origin l2 data
if (ge_ret != SUCCESS) {
return ge_ret;
}
vector<void *> virtual_io_addrs; // use virtual address for zero copy key.
virtual_io_addrs.insert(virtual_io_addrs.end(), input_data_addrs.begin(), input_data_addrs.end());
virtual_io_addrs.insert(virtual_io_addrs.end(), output_data_addrs.begin(), output_data_addrs.end());
if (op_desc->GetType() == ATOMICADDRCLEAN) {
virtual_io_addrs.insert(virtual_io_addrs.end(), workspace_data_addrs.begin(), workspace_data_addrs.end());
}
davinci_model_->SetZeroCopyAddr(op_desc, virtual_io_addrs, args_info.data(), args_, args_size_, offset);
davinci_model_->SetZeroCopyAddr(op_desc, virtual_io_addrs, args_addr.get(), args_, args_size_, offset);
GELOGD("Do InitTVMTask end");
return SUCCESS;

5
ge/graph/load/model_manager/task_info/kernel_task_info.h
Unescape View File

@ -129,6 +129,9 @@ class KernelTaskInfo : public TaskInfo {
bool IsL1FusionOp(const OpDescPtr &op_desc);
void SetIoAddrs(const OpDescPtr &op_desc);
void InitDumpTask(uint32_t offset);
void SetContinuousArgs(uint32_t args_size, DavinciModel *davinci_model);
void SetNoncontinuousArgs(uint32_t args_size, DavinciModel *davinci_model);
Status CopyNoncontinuousArgs(uint16_t offset);
// For super kernel
Status SaveSKTDumpInfo();
@ -163,6 +166,8 @@ class KernelTaskInfo : public TaskInfo {
uint32_t hybrid_args_offset_ = 0;
int64_t fixed_addr_offset_ = 0;
std::unique_ptr<uint8_t[]> args_addr = nullptr;
uint16_t io_addr_offset_ = 0;
bool l2_buffer_on_ = false;
bool call_save_dump_ = false;
// aicpu ext_info device mem

103
ge/graph/passes/base_pass.cc
Unescape View File

@ -30,8 +30,15 @@ constexpr int kMaxRePassTimes = 10000;
constexpr size_t kMaxOneInNodes = 1000;
// Each iteration, we take about 0.3k memory on the stack, we should change the recursion to loop later
constexpr int kMaxRecursiveDepth = 20;
struct DuringPassNodeSets {
std::unordered_set<Node *> nodes_seen;
std::unordered_set<NodePtr> nodes_deleted;
std::unordered_set<NodePtr> nodes_re_pass;
std::unordered_set<NodePtr> nodes_re_pass_immediately;
std::unordered_set<NodePtr> nodes_last;
};
void GetAllNodesNoInputEdge(const ComputeGraphPtr &graph, std::queue<NodePtr> &input_edge_nodes,
void GetAllNodesNoInputEdge(const ComputeGraphPtr &graph, std::deque<NodePtr> &input_edge_nodes,
std::unordered_set<Node *> &nodes_seen, std::unordered_set<NodePtr> &nodes_last) {
nodes_last.clear();
for (auto &node : graph->GetDirectNode()) {
@ -40,7 +47,7 @@ void GetAllNodesNoInputEdge(const ComputeGraphPtr &graph, std::queue<NodePtr> &i
}
size_t in_nums = node->GetInNodes().size();
if (in_nums == 0) {
input_edge_nodes.push(node);
input_edge_nodes.push_back(node);
nodes_seen.insert(node.get());
} else if (in_nums > kMaxOneInNodes) {
nodes_last.insert(node);
@ -48,7 +55,7 @@ void GetAllNodesNoInputEdge(const ComputeGraphPtr &graph, std::queue<NodePtr> &i
}
}
void AddNextIterNodes(const Node::Vistor<NodePtr> &nodes, std::queue<NodePtr> &nodes_to_pass,
void AddNextIterNodes(const Node::Vistor<NodePtr> &nodes, std::deque<NodePtr> &nodes_to_pass,
std::unordered_set<Node *> &nodes_seen, std::unordered_set<NodePtr> &nodes_last) {
for (auto &node : nodes) {
if (node == nullptr) {
@ -60,13 +67,30 @@ void AddNextIterNodes(const Node::Vistor<NodePtr> &nodes, std::queue<NodePtr> &n
bool all_in_nodes_seen = node->IsAllInNodesSeen(nodes_seen);
if (all_in_nodes_seen && nodes_seen.insert(node.get()).second) {
nodes_to_pass.push(node);
nodes_to_pass.push_back(node);
}
}
}
void PushToRePassIfSeen(NodePtr &node, const std::pair<std::string, BaseNodePass *> &name_to_pass,
std::unordered_set<Node *> &nodes_seen, std::unordered_set<NodePtr> &nodes_to_re_pass,
std::unordered_set<NodePtr> &nodes_re_pass) {
for (const auto &node_to_re_pass : nodes_to_re_pass) {
if (node_to_re_pass == nullptr) {
GELOGW("Found null re-pass node when executing %s on node %s type %s", name_to_pass.first.c_str(),
node->GetName().c_str(), node->GetType().c_str());
continue;
}
if (nodes_seen.count(node_to_re_pass.get()) > 0 || node_to_re_pass->IsAllInNodesSeen(nodes_seen)) {
GELOGD("The node %s will be re-pass.", node_to_re_pass->GetName().c_str());
nodes_re_pass.insert(node_to_re_pass);
} else {
GELOGD("The node %s are not all seen, don't set repass this time", node_to_re_pass->GetName().c_str());
}
}
}
Status RunPasses(NodePtr &node, const NamesToPass &names_to_passes, std::unordered_set<NodePtr> &nodes_re_pass,
std::unordered_set<NodePtr> &nodes_deleted, std::unordered_set<Node *> &nodes_seen) {
Status RunPasses(NodePtr &node, const NamesToPass &names_to_passes, DuringPassNodeSets &during_pass_node_set) {
if (node == nullptr) {
GELOGE(FAILED, "parameter is null.");
return FAILED;
@ -90,22 +114,15 @@ Status RunPasses(NodePtr &node, const NamesToPass &names_to_passes, std::unorder
}
auto nodes_to_re_pass = name_to_pass.second->GetNodesNeedRePass();
for (const auto &node_to_re_pass : nodes_to_re_pass) {
if (node_to_re_pass == nullptr) {
GELOGW("Found null re-pass node when executing %s on node %s type %s", name_to_pass.first.c_str(),
node->GetName().c_str(), node->GetType().c_str());
continue;
}
if (nodes_seen.count(node_to_re_pass.get()) > 0 || node_to_re_pass->IsAllInNodesSeen(nodes_seen)) {
GELOGD("The node %s will be re-pass later", node_to_re_pass->GetName().c_str());
nodes_re_pass.insert(node_to_re_pass);
} else {
GELOGD("The node %s are not all seen, don't set repass this time", node_to_re_pass->GetName().c_str());
}
}
PushToRePassIfSeen(node, name_to_pass, during_pass_node_set.nodes_seen, nodes_to_re_pass,
during_pass_node_set.nodes_re_pass);
auto nodes_to_re_pass_immediately = name_to_pass.second->GetNodesNeedRePassImmediately();
PushToRePassIfSeen(node, name_to_pass, during_pass_node_set.nodes_seen, nodes_to_re_pass_immediately,
during_pass_node_set.nodes_re_pass_immediately);
auto nodes_deleted_by_pass = name_to_pass.second->GetNodesDeleted();
nodes_deleted.insert(nodes_deleted_by_pass.begin(), nodes_deleted_by_pass.end());
during_pass_node_set.nodes_deleted.insert(nodes_deleted_by_pass.begin(), nodes_deleted_by_pass.end());
if (nodes_deleted_by_pass.count(node) > 0) {
GELOGD("The node %s was deleted by pass %s, stop the remain passes", node->GetName().c_str(),
name_to_pass.first.c_str());
@ -181,36 +198,33 @@ Status GEPass::Run(const NamesToPass &names_to_passes) {
Status GEPass::RunPassesOneGraph(const NamesToPass &names_to_passes) {
GELOGD("Begin to run pass on graph, passes count %zu", names_to_passes.size());
std::queue<NodePtr> nodes;
std::unordered_set<Node *> nodes_seen;
std::unordered_set<NodePtr> nodes_deleted;
std::unordered_set<NodePtr> nodes_re_pass;
std::unordered_set<NodePtr> nodes_last;
GetAllNodesNoInputEdge(graph_, nodes, nodes_seen, nodes_last);
std::deque<NodePtr> nodes;
DuringPassNodeSets during_pass_node_set;
GetAllNodesNoInputEdge(graph_, nodes, during_pass_node_set.nodes_seen, during_pass_node_set.nodes_last);
GELOGD("Start points count %zu", nodes.size());
int re_pass_times = 0;
do {
for (auto &node : nodes_re_pass) {
nodes.push(node);
nodes_seen.insert(node.get());
for (auto &node : during_pass_node_set.nodes_re_pass) {
nodes.push_back(node);
during_pass_node_set.nodes_seen.insert(node.get());
}
nodes_re_pass.clear();
during_pass_node_set.nodes_re_pass.clear();
while (!nodes.empty()) {
NodePtr node = nodes.front();
nodes.pop();
nodes.pop_front();
(void)nodes_re_pass.erase(node);
(void)during_pass_node_set.nodes_re_pass.erase(node);
GE_IF_BOOL_EXEC(node == nullptr, GELOGW("node is null"); continue);
if (nodes_deleted.count(node) > 0) {
if (during_pass_node_set.nodes_deleted.count(node) > 0) {
GELOGD("The node %s was deleted before, skip it.", node->GetName().c_str());
continue;
}
AddNextIterNodes(node->GetOutNodes(), nodes, nodes_seen, nodes_last);
AddNextIterNodes(node->GetOutNodes(), nodes, during_pass_node_set.nodes_seen, during_pass_node_set.nodes_last);
auto ret = RunPasses(node, names_to_passes, nodes_re_pass, nodes_deleted, nodes_seen);
auto ret = RunPasses(node, names_to_passes, during_pass_node_set);
if (ret != SUCCESS) {
GELOGE(ret, "Failed to process passes on node %s type %s, error code: %u",
node->GetName().c_str(), node->GetType().c_str(), ret);
@ -227,7 +241,7 @@ Status GEPass::RunPassesOneGraph(const NamesToPass &names_to_passes) {
if (has_sub_graph) {
GELOGD("There are subgraphs on node %s, run passes for for the second time", node->GetName().c_str());
SetFlagOption(kOptimizeAfterSubGraph, names_to_passes);
ret = RunPasses(node, names_to_passes, nodes_re_pass, nodes_deleted, nodes_seen);
ret = RunPasses(node, names_to_passes, during_pass_node_set);
if (ret != SUCCESS) {
GELOGE(ret, "Failed to process passes on node %s type %s, error code: %u",
node->GetName().c_str(), node->GetType().c_str(), ret);
@ -239,16 +253,21 @@ Status GEPass::RunPassesOneGraph(const NamesToPass &names_to_passes) {
// should be called each time at the begin of the iteration
ClearOption(names_to_passes);
}
for (const auto &node : during_pass_node_set.nodes_re_pass_immediately) {
GELOGD("The node %s will be re-pass immediately.", node->GetName().c_str());
nodes.push_front(node);
}
during_pass_node_set.nodes_re_pass_immediately.clear();
}
for (auto &node : nodes_last) {
bool all_in_nodes_seen = node->IsAllInNodesSeen(nodes_seen);
if (all_in_nodes_seen && nodes_seen.insert(node.get()).second) {
nodes.push(node);
for (auto &node : during_pass_node_set.nodes_last) {
bool all_in_nodes_seen = node->IsAllInNodesSeen(during_pass_node_set.nodes_seen);
if (all_in_nodes_seen && during_pass_node_set.nodes_seen.insert(node.get()).second) {
nodes.push_back(node);
}
}
nodes_last.clear();
} while ((!nodes_re_pass.empty() || !nodes.empty()) && ++re_pass_times < kMaxRePassTimes);
during_pass_node_set.nodes_last.clear();
} while ((!during_pass_node_set.nodes_re_pass.empty() || !nodes.empty()) && ++re_pass_times < kMaxRePassTimes);
if (re_pass_times == kMaxRePassTimes) {
GELOGW("re_pass_times should not come to %d", kMaxRePassTimes);

12
ge/graph/passes/base_pass.h
Unescape View File

@ -53,6 +53,8 @@ class BaseNodePass {
std::unordered_set<NodePtr> GetNodesNeedRePass() { return nodes_need_re_pass_; }
std::unordered_set<NodePtr> GetNodesNeedRePassImmediately() { return nodes_need_re_pass_immediately_; }
std::unordered_set<NodePtr> GetNodesDeleted() { return nodes_deleted_; }
void SetOption(NodePassOption option, const std::string &value) { options_[option] = value; }
@ -62,6 +64,7 @@ class BaseNodePass {
void init() {
nodes_need_re_pass_.clear();
nodes_deleted_.clear();
nodes_need_re_pass_immediately_.clear();
}
protected:
@ -79,6 +82,14 @@ class BaseNodePass {
///
void AddRePassNode(NodePtr &node) { nodes_need_re_pass_.insert(node); }
///
/// Add a node to be optimized immediately again. If you add a new node to the graph, or
/// change a node connections, and you want to make sure the node will be
/// optimized by other passes, call this function.
/// @param node
///
void AddImmediateRePassNode(NodePtr &node) { nodes_need_re_pass_immediately_.insert(node); }
///
/// Add a node and it's input/output data nodes to be optimized again.
/// @param node
@ -109,6 +120,7 @@ class BaseNodePass {
private:
std::unordered_set<NodePtr> nodes_need_re_pass_;
std::unordered_set<NodePtr> nodes_need_re_pass_immediately_;
std::unordered_set<NodePtr> nodes_deleted_;
std::map<NodePassOption, std::string> options_;
};

Some files were not shown because too many files have changed in this diff Show More

Write Preview
Loading…
Cancel
Save