p12346870
/
mindspore
forked from mindspore-Ecosystem/mindspore
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

!22891 [Debugger] Fix buffer size for reading npy header 

Merge pull request !22891 from TinaMengtingZhang/fix_read_npy_header_len
This commit is contained in:
i-robot 2021-09-04 02:29:06 +00:00 committed by Gitee
parent c928278b62 b039cc3682
commit c9b8d6e91a
2 changed files with 22 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

32
mindspore/ccsrc/debug/debug_services.cc
View File

@ -411,7 +411,7 @@ void DebugServices::CheckWatchpoints(std::vector<std::string> *const name, std::
MS_LOG(INFO) << "tensor list size: " << tensor_list_size;
if (tensor_list_size == 0) return;
// default value for number of threads
const int default_thread_num = 32;
const int default_thread_num = 16;
int max_thread_num = default_thread_num;
if (max_thread_num > tensor_list_size) {
max_thread_num = tensor_list_size;
@ -528,20 +528,28 @@ void DebugServices::ReadTensorFromNpy(const std::string &tensor_name, const std:
<< " ErrInfo:" << strerror(errno);
return;
}
uint64_t file_size = infile.tellg();
infile.seekg(0, std::ios::beg);
auto buffer = std::make_unique<std::vector<char>>(file_size);
if (!infile.read(buffer->data(), file_size)) {
MS_LOG(ERROR) << "Failed to read file (In ReadTensorFromNpy) " << file_path;
return;
}
const int substr_len = 2;
const int header_len_offset = 8;
const int header_offset = 9;
const int type_offset = 10;
uint16_t header_len = *reinterpret_cast<uint16_t *>(buffer->data() + header_len_offset);
std::string header(buffer->data() + header_offset, header_len);
buffer.reset();
// get header length
infile.seekg(0, std::ios::beg);
auto header_len_buffer = std::make_unique<std::vector<char>>(header_len_offset + 2);
if (!infile.read(header_len_buffer->data(), header_len_offset + 2)) {
MS_LOG(ERROR) << "Failed to parse header length from " << file_path;
return;
}
uint16_t header_len = *reinterpret_cast<uint16_t *>(header_len_buffer->data() + header_len_offset);
header_len_buffer.reset();
// read in header
infile.seekg(0, std::ios::beg);
auto header_buffer = std::make_unique<std::vector<char>>(header_len_offset + header_len);
if (!infile.read(header_buffer->data(), header_len_offset + header_len)) {
MS_LOG(ERROR) << "Failed to read header from " << file_path;
return;
}
std::string header(header_buffer->data() + header_offset, header_len);
header_buffer.reset();
std::size_t type_i = header.find("descr") + type_offset;
if (header.length() < type_i + substr_len) {
MS_LOG(ERROR) << "Cannot get tensor_type, header length is " << header.length();
@ -566,7 +574,7 @@ void DebugServices::ReadTensorFromNpy(const std::string &tensor_name, const std:
has_enough_memory = tensor_loader_->CheckMemoryAvailable(tensor_name, data_size);
}
if (!has_enough_memory) {
MS_LOG(ERROR) << "No enough space available for loading " << tensor_name << " into host memory.";
MS_LOG(ERROR) << "No enough memory available for loading " << tensor_name << " into host memory.";
*no_mem_to_read = true;
} else {
infile.seekg(header_len + type_offset);

3
mindspore/ccsrc/debug/debugger/offline_debug/dbg_services.cc
View File

@ -66,8 +66,9 @@ int32_t DbgServices::Initialize(std::string net_name, std::string dump_folder_pa
debug_services_->SetDumpDir(dump_folder_path);
debug_services_->SetSyncMode(is_sync_mode);
// Set the memory ratio used by tensor cache. Leave 50% for other debugger backend usage.
const int32_t kMegabytesToBytes = pow(2, 20); // max_mem_usage will be bytes in unit in debugger backend.
auto cache_mem_ratio = 0.5;
debug_services_->SetMemLimit(max_mem_usage * pow(2, 20) * cache_mem_ratio);
debug_services_->SetMemLimit(max_mem_usage * kMegabytesToBytes * cache_mem_ratio);
return 0;
}