mindspore-Ecosystem
/
mindspore
21
0
Fork 27
Code Issues Pull Requests 6 Projects Releases Wiki Activity

!49948 NPU V2 overflow check 

Merge pull request !49948 from archer2049/master_npu3_merge
This commit is contained in:
i-robot 2023-03-08 06:47:14 +00:00 committed by Gitee
parent 9236c3b4e9 2261cf7bca
commit 522d61880a
No known key found for this signature in database
GPG Key ID: 173E9B9CA92EEF8F
37 changed files with 1280 additions and 240 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
docs/api/api_python/amp/mindspore.amp.all_finite.rst
View File

@ -12,7 +12,6 @@ mindspore.amp.all_finite
参数：
- **inputs** (Union(tuple(Tensor), list(Tensor))) - 可迭代的Tensor。
- **status** (Tensor) - 溢出检测时所需要的初始状态，仅在Ascend需要。默认值：None。
返回：
Tensor，布尔类型的标量Tensor。

12
docs/api/api_python/amp/mindspore.amp.init_status.rst
View File

@ -1,12 +0,0 @@
mindspore.amp.init_status
===========================
.. py:function:: mindspore.amp.init_status()
初始化溢出状态检测变量。
.. note::
该接口仅在Ascend后端有效，在GPU、CPU上调用的返回值没有作用。
返回：
Tensor，shape为 (8,) 。

1
docs/api/api_python/mindspore.amp.rst
View File

@ -35,5 +35,4 @@ mindspore.amp
:nosignatures:
:template: classtemplate.rst
mindspore.amp.init_status
mindspore.amp.all_finite

1
docs/api/api_python_en/mindspore.amp.rst
View File

@ -35,5 +35,4 @@ Overflow Detection
:nosignatures:
:template: classtemplate.rst
mindspore.amp.init_status
mindspore.amp.all_finite

2
mindspore/ccsrc/include/common/utils/utils.h
View File

@ -572,6 +572,8 @@ constexpr auto kNonZeroOpName = "NonZero";
constexpr auto kNPUAllocFloatStatusOpName = "NPUAllocFloatStatus";
constexpr auto kNPUClearFloatStatusOpName = "NPUClearFloatStatus";
constexpr auto kNPUGetFloatStatusOpName = "NPUGetFloatStatus";
constexpr auto kNPUClearFloatStatusV2OpName = "NPUClearFloatStatusV2";
constexpr auto kNPUGetFloatStatusV2OpName = "NPUGetFloatStatusV2";
constexpr auto kNthElementOpName = "NthElement";
constexpr auto kOneHotOpName = "OneHot";
constexpr auto kOneHotDOpName = "OneHotD";

94
mindspore/ccsrc/kernel/kash/kernel_pack.cc
View File

@ -24,14 +24,31 @@
namespace mindspore {
namespace kernel {
constexpr size_t kJsonSuffixLength = 5;
constexpr char kMagic[] = "magic";
constexpr char kBlockDim[] = "blockDim";
constexpr char kKernelName[] = "kernelName";
constexpr char kBinFileName[] = "binFileName";
constexpr char kBinFileSuffix[] = "binFileSuffix";
constexpr char kCoreType[] = "core_type";
constexpr char kTaskRation[] = "taskRation";
constexpr char kWorkspace[] = "workspace";
constexpr char kParameters[] = "parameters";
constexpr char kOpParaSize[] = "opParaSize";
constexpr char kSHA256[] = "sha256";
constexpr char kKBHit[] = "KBHit";
constexpr char kKernelList[] = "kernelList";
constexpr char kModeInArgsFirstField[] = "modeInArgsFirstField";
constexpr char kBatchBindOnly[] = "batchBindOnly";
constexpr char kArgsRemap[] = "args_remap";
constexpr char kSize[] = "size";
constexpr char kGlobalWorkspaceSpecWorkspace[] = "globalworkspace_spec_workspace";
namespace {
bool CheckHash(const std::string &json_file, const std::string &bin_file, const nlohmann::json &js) {
if (js.find("sha256") == js.end()) {
MS_LOG(ERROR) << "No sha256 found in " << json_file;
if (js.find(kSHA256) == js.end()) {
return false;
}
std::string sha256_cal = system::sha256::GetHashFromFile(bin_file);
std::string sha256_str = js["sha256"];
std::string sha256_str = js[kSHA256];
if (sha256_cal.empty() || sha256_cal != sha256_str) {
MS_LOG(WARNING) << "Check sha256 for [" << bin_file << "] failed, it will try to rebuild the op.";
return false;
@ -154,9 +171,9 @@ bool KernelPack::ReadFromJsonFile(const std::string &json_f, const std::string &
}
// cuda json file may have workspace information
if (js.find("workspace") != js.end()) {
auto workspace = js.at("workspace");
std::vector<size_t> sizes = workspace.at("size");
if (js.find(kWorkspace) != js.end()) {
auto workspace = js.at(kWorkspace);
std::vector<size_t> sizes = workspace.at(kSize);
for (auto size : sizes) {
kernel_json_info_.workspaces.push_back(size);
}
@ -165,7 +182,7 @@ bool KernelPack::ReadFromJsonFile(const std::string &json_f, const std::string &
return true;
}
std::string binfile_suffix = js["binFileSuffix"];
std::string binfile_suffix = js[kBinFileSuffix];
std::string bin_f = json_f.substr(0, json_f.length() - kJsonSuffixLength) + binfile_suffix;
if (binfile_suffix == ".so") {
// change "xx/xx.so" -> "xx/libxx.so"
@ -282,18 +299,18 @@ void KernelPack::ParseWorkSpace(const std::string &key, const nlohmann::json &js
}
try {
auto workspace = js.at(key);
if (workspace.find("num") == workspace.end() || workspace.find("size") == workspace.end()) {
if (workspace.find("num") == workspace.end() || workspace.find(kSize) == workspace.end()) {
MS_LOG(WARNING) << "'num' and 'size' ars necessary in workspace, but not found. " << js.dump(indent);
return;
}
size_t num = workspace.at("num");
std::vector<size_t> sizes = workspace.at("size");
std::vector<size_t> sizes = workspace.at(kSize);
if (num != sizes.size()) {
MS_LOG(WARNING) << "'num' and length of 'size' must be same. " << js.dump(indent);
return;
}
if (workspace.find("type") != workspace.end()) {
std::vector<size_t> type = workspace.at("type");
if (workspace.find(kType) != workspace.end()) {
std::vector<size_t> type = workspace.at(kType);
if (num != type.size()) {
MS_LOG(WARNING) << "'num' and length of 'type' must be same. " << js.dump(indent);
return;
@ -383,24 +400,47 @@ void KernelPack::ParseArgsRemap(const std::string &key, const nlohmann::json &js
}
}
void KernelPack::ParseGlogbleWorkSpace(const std::string &key, const nlohmann::json &js,
KernelJsonInfo *kernel_json_info) {
MS_EXCEPTION_IF_NULL(kernel_json_info);
if (js.find(key) == js.end()) {
return;
}
try {
auto globalWorkspace = js.at(key);
if (globalWorkspace.find(kSize) != globalWorkspace.end()) {
kernel_json_info->global_workspace.size = globalWorkspace.at(kSize);
kernel_json_info->global_workspace.is_overflow = true;
}
if (globalWorkspace.find(kType) != globalWorkspace.end()) {
kernel_json_info->global_workspace.type = globalWorkspace.at(kType);
kernel_json_info->global_workspace.is_overflow = true;
}
} catch (std::exception &e) {
MS_LOG(ERROR) << "Parse json value failed, jsong is:" + js.dump() + ", error info: " << e.what();
}
}
void KernelPack::ParseKernelJson(const nlohmann::json &js) {
using KernelJsonParser = std::function<void(const std::string &, const nlohmann::json &, KernelJsonInfo *)>;
const std::map<std::string, KernelJsonParser> kernel_json_map = {{"magic", ParseMagic},
{"blockDim", ParseBlockDim},
{"kernelName", ParseKernelName},
{"binFileName", ParseBinFileName},
{"binFileSuffix", ParseBinFileSuffix},
{"core_type", ParseCoreType},
{"taskRation", ParseTaskRatio},
{"workspace", ParseWorkSpace},
{"parameters", ParseParameters},
{"opParaSize", ParseOpParaSize},
{"sha256", ParseSHA256},
{"KBHit", ParseKBHit},
{"kernelList", ParseKernelList},
{"modeInArgsFirstField", ParseModeInArgsFirstField},
{"batchBindOnly", ParseBatchBindOnly},
{"args_remap", ParseArgsRemap}};
const std::map<std::string, KernelJsonParser> kernel_json_map = {
{kMagic, ParseMagic},
{kBlockDim, ParseBlockDim},
{kKernelName, ParseKernelName},
{kBinFileName, ParseBinFileName},
{kBinFileSuffix, ParseBinFileSuffix},
{kCoreType, ParseCoreType},
{kTaskRation, ParseTaskRatio},
{kWorkspace, ParseWorkSpace},
{kParameters, ParseParameters},
{kOpParaSize, ParseOpParaSize},
{kSHA256, ParseSHA256},
{kKBHit, ParseKBHit},
{kKernelList, ParseKernelList},
{kModeInArgsFirstField, ParseModeInArgsFirstField},
{kBatchBindOnly, ParseBatchBindOnly},
{kArgsRemap, ParseArgsRemap},
{kGlobalWorkspaceSpecWorkspace, ParseGlogbleWorkSpace}};
auto iter = kernel_json_map.begin();
while (iter != kernel_json_map.end()) {
iter->second(iter->first, js, &kernel_json_info_);

8
mindspore/ccsrc/kernel/kernel.h
View File

@ -123,6 +123,12 @@ struct FlexArray {
char contents[];
};
struct GlobalWorkspace {
size_t size;
size_t type;
bool is_overflow = false;
};
struct KernelJsonInfo {
std::string bin_file_name;
std::string bin_file_suffix;
@ -133,6 +139,7 @@ struct KernelJsonInfo {
std::string sha256;
std::vector<size_t> workspaces_type;
std::vector<size_t> workspaces;
GlobalWorkspace global_workspace;
bool has_kernel_list = false;
uint32_t op_para_size;
int32_t KBHit;
@ -185,6 +192,7 @@ class BACKEND_EXPORT KernelPack {
static void ParseModeInArgsFirstField(const std::string &key, const nlohmann::json &js,
KernelJsonInfo *kernel_json_info);
static void ParseArgsRemap(const std::string &key, const nlohmann::json &js, KernelJsonInfo *kernel_json_info);
static void ParseGlogbleWorkSpace(const std::string &key, const nlohmann::json &js, KernelJsonInfo *kernel_json_info);
KernelJsonInfo kernel_json_info_;
FlexArray *json_;

16
mindspore/ccsrc/plugin/device/ascend/hal/device/ascend_memory_adapter.cc
View File

@ -19,6 +19,7 @@
#include <algorithm>
#include "ir/func_graph.h"
#include "runtime/mem.h"
#include "acl/acl_rt.h"
#include "utils/ms_context.h"
#include "utils/convert_utils_base.h"
#include "graphengine/inc/external/runtime/rt_error_codes.h"
@ -35,6 +36,7 @@ constexpr double kHalfRatio = 0.5;
// The Ascend max available device memory is 32GB.
constexpr float kAscendMaxDeviceMemory = 32;
constexpr uint64_t kOverflowAddrSize = 512;
constexpr char kGlobalOverflowWorkspace[] = "GLOBAL_OVERFLOW_WORKSPACE";
size_t AscendMemAdapter::GetRoundDownAlignSize(size_t input_size) {
return (input_size / kAscendMemAlignSize) * kAscendMemAlignSize;
@ -180,16 +182,16 @@ uint8_t *AscendMemAdapter::MallocDynamicDevMem(size_t size, const std::string &t
return memory_block_ptr;
}
uint8_t *AscendMemAdapter::MallocOverflowMem(const CNodePtr &kernel) {
uint8_t *AscendMemAdapter::MallocOverflowMem() {
std::lock_guard<std::mutex> locker(overflow_mutex_);
auto funcGraph = kernel->func_graph();
MS_EXCEPTION_IF_NULL(funcGraph);
if (overflow_memory_info_map_.find(funcGraph->ToString()) != overflow_memory_info_map_.cend()) {
return overflow_memory_info_map_.find(funcGraph->ToString())->second;
if (overflow_memory_info_map_.find(kGlobalOverflowWorkspace) != overflow_memory_info_map_.cend()) {
auto addr = overflow_memory_info_map_.find(kGlobalOverflowWorkspace);
return addr->second;
} else {
auto overflow_memory_ptr = MallocStaticDevMem(kOverflowAddrSize, "overflow memory ptr");
auto overflow_memory_ptr = MallocStaticDevMem(kOverflowAddrSize, "global overflow memory ptr");
MS_EXCEPTION_IF_NULL(overflow_memory_ptr);
(void)overflow_memory_info_map_.emplace(funcGraph->ToString(), overflow_memory_ptr);
(void)aclrtMemset(overflow_memory_ptr, kOverflowAddrSize, 0, kOverflowAddrSize);
(void)overflow_memory_info_map_.emplace(kGlobalOverflowWorkspace, overflow_memory_ptr);
return overflow_memory_ptr;
}
}

2
mindspore/ccsrc/plugin/device/ascend/hal/device/ascend_memory_adapter.h
View File

@ -39,7 +39,7 @@ class AscendMemAdapter {
uint8_t *MallocStaticDevMem(size_t size, const std::string &tag = "");
uint8_t *MallocDynamicDevMem(size_t size, const std::string &tag = "");
uint8_t *MallocOverflowMem(const CNodePtr &kernel);
uint8_t *MallocOverflowMem();
bool FreeStaticDevMem(void *) const { return true; }
void ResetDynamicMemory();

156
mindspore/ccsrc/plugin/device/ascend/hal/device/kernel_adjust.cc
View File

@ -735,74 +735,56 @@ void KernelAdjust::InsertProfilingKernel(const ProfilingTraceInfo &profiling_tra
}
#endif
CNodePtr KernelAdjust::CreateNPUGetFloatStatus(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const CNodePtr &npu_alloc_cnode) const {
CNodePtr KernelAdjust::CreateNPUGetFloatStatusV2(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const AnfNodePtr &status_value_node) const {
MS_EXCEPTION_IF_NULL(kernel_graph_ptr);
MS_EXCEPTION_IF_NULL(npu_alloc_cnode);
auto npu_get_primitive = std::make_shared<Primitive>(kNPUGetFloatStatusOpName);
std::vector<AnfNodePtr> npu_get_inputs = {NewValueNode(npu_get_primitive), npu_alloc_cnode};
MS_EXCEPTION_IF_NULL(status_value_node);
auto npu_get_primitive = std::make_shared<Primitive>(kNPUGetFloatStatusV2OpName);
std::vector<AnfNodePtr> npu_get_inputs = {NewValueNode(npu_get_primitive), status_value_node};
auto npu_get_cnode = kernel_graph_ptr->NewCNode(npu_get_inputs);
MS_EXCEPTION_IF_NULL(npu_get_cnode);
npu_alloc_cnode->set_scope(kDefaultScope);
npu_get_cnode->set_abstract(npu_alloc_cnode->abstract());
status_value_node->set_scope(kDefaultScope);
ShapeVector npu_output_shape = {kNPUShape};
common::AnfAlgo::SetOutputInferTypeAndShape({kNumberTypeInt32}, {npu_output_shape}, npu_get_cnode.get());
kernel::KernelBuildInfo::KernelBuildInfoBuilder selected_kernel_builder;
selected_kernel_builder.SetInputsFormat({kOpFormat_DEFAULT});
selected_kernel_builder.SetInputsDeviceType({kNumberTypeFloat32});
selected_kernel_builder.SetInputsDeviceType({kNumberTypeInt32});
selected_kernel_builder.SetFusionType(kernel::kPatternOpaque);
selected_kernel_builder.SetProcessor(kernel::Processor::AICORE);
selected_kernel_builder.SetKernelType(KernelType::TBE_KERNEL);
selected_kernel_builder.SetOutputsFormat({kOpFormat_DEFAULT});
selected_kernel_builder.SetOutputsDeviceType({kNumberTypeFloat32});
selected_kernel_builder.SetOutputsDeviceType({kNumberTypeInt32});
AnfAlgo::SetSelectKernelBuildInfo(selected_kernel_builder.Build(), npu_get_cnode.get());
return npu_get_cnode;
}
CNodePtr KernelAdjust::CreateNPUClearStatus(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const CNodePtr &npu_alloc_cnode) const {
CNodePtr KernelAdjust::CreateNPUClearStatusV2(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const AnfNodePtr &status_value_node) const {
MS_EXCEPTION_IF_NULL(kernel_graph_ptr);
MS_EXCEPTION_IF_NULL(npu_alloc_cnode);
auto npu_clear_primitive = std::make_shared<Primitive>(kNPUClearFloatStatusOpName);
std::vector<AnfNodePtr> npu_clear_inputs = {NewValueNode(npu_clear_primitive), npu_alloc_cnode};
MS_EXCEPTION_IF_NULL(status_value_node);
auto npu_clear_primitive = std::make_shared<Primitive>(kNPUClearFloatStatusV2OpName);
std::vector<AnfNodePtr> npu_clear_inputs = {NewValueNode(npu_clear_primitive), status_value_node};
auto npu_clear_cnode = kernel_graph_ptr->NewCNode(npu_clear_inputs);
MS_EXCEPTION_IF_NULL(npu_clear_cnode);
npu_alloc_cnode->set_scope(kDefaultScope);
npu_clear_cnode->set_abstract(npu_alloc_cnode->abstract());
status_value_node->set_scope(kDefaultScope);
npu_clear_cnode->set_abstract(status_value_node->abstract());
ShapeVector npu_output_shape = {kNPUShape};
common::AnfAlgo::SetOutputInferTypeAndShape({kNumberTypeInt32}, {npu_output_shape}, npu_clear_cnode.get());
kernel::KernelBuildInfo::KernelBuildInfoBuilder selected_kernel_builder;
selected_kernel_builder.SetInputsFormat({kOpFormat_DEFAULT});
selected_kernel_builder.SetInputsDeviceType({kNumberTypeFloat32});
selected_kernel_builder.SetInputsDeviceType({kNumberTypeInt32});
selected_kernel_builder.SetFusionType(kernel::kPatternOpaque);
selected_kernel_builder.SetProcessor(kernel::Processor::AICORE);
selected_kernel_builder.SetKernelType(KernelType::TBE_KERNEL);
selected_kernel_builder.SetOutputsFormat({kOpFormat_DEFAULT});
selected_kernel_builder.SetOutputsDeviceType({kNumberTypeFloat32});
selected_kernel_builder.SetOutputsDeviceType({kNumberTypeInt32});
AnfAlgo::SetSelectKernelBuildInfo(selected_kernel_builder.Build(), npu_clear_cnode.get());
return npu_clear_cnode;
}
CNodePtr KernelAdjust::CreateNPUAllocStatus(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr) const {
MS_EXCEPTION_IF_NULL(kernel_graph_ptr);
// create npu_alloc_cnode
auto npu_alloc_primitive = std::make_shared<Primitive>(kNPUAllocFloatStatusOpName);
std::vector<AnfNodePtr> npu_alloc_inputs = {NewValueNode(npu_alloc_primitive)};
auto npu_alloc_cnode = kernel_graph_ptr->NewCNode(npu_alloc_inputs);
MS_EXCEPTION_IF_NULL(npu_alloc_cnode);
npu_alloc_cnode->set_scope(kDefaultScope);
ShapeVector npu_output_shape = {kNPUShape};
common::AnfAlgo::SetOutputInferTypeAndShape({kNumberTypeFloat32}, {npu_output_shape}, npu_alloc_cnode.get());
kernel::KernelBuildInfo::KernelBuildInfoBuilder selected_kernel_builder;
selected_kernel_builder.SetFusionType(kernel::kPatternOpaque);
selected_kernel_builder.SetProcessor(kernel::Processor::AICORE);
selected_kernel_builder.SetKernelType(KernelType::TBE_KERNEL);
selected_kernel_builder.SetOutputsFormat({kOpFormat_DEFAULT});
selected_kernel_builder.SetOutputsDeviceType({kNumberTypeFloat32});
AnfAlgo::SetSelectKernelBuildInfo(selected_kernel_builder.Build(), npu_alloc_cnode.get());
return npu_alloc_cnode;
}
CNodePtr KernelAdjust::CreateAssignAdd(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const CNodePtr &npu_alloc_cnode, const AnfNodePtr &specify_para) const {
MS_EXCEPTION_IF_NULL(kernel_graph_ptr);
@ -836,39 +818,41 @@ CNodePtr KernelAdjust::CreateAssignAdd(const std::shared_ptr<session::KernelGrap
return assign_add_cnode;
}
CNodePtr KernelAdjust::CreateAssign(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const AnfNodePtr &specify_para) const {
MS_EXCEPTION_IF_NULL(kernel_graph_ptr);
MS_EXCEPTION_IF_NULL(specify_para);
std::vector<float> reset(kNPUShape, 0.0);
ShapeVector reset_shape({kNPUShape});
auto shp_buf_size = sizeof(float) * reset.size();
auto reset_tensor = std::make_shared<tensor::Tensor>(kNumberTypeFloat32, reset_shape, reset.data(), shp_buf_size);
auto reset_value_node = std::make_shared<ValueNode>(reset_tensor);
MS_EXCEPTION_IF_NULL(reset_value_node);
reset_value_node->set_abstract(specify_para->abstract());
kernel_graph_ptr->AddValueNodeToGraph(reset_value_node);
AnfNodePtr KernelAdjust::CreateZerosValueNode(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr) const {
std::vector<int32_t> zeros(kNPUShape, 0);
ShapeVector zeros_shape({kNPUShape});
auto shp_buf_size = sizeof(float) * zeros.size();
auto zeros_tensor = std::make_shared<tensor::Tensor>(kNumberTypeInt32, zeros_shape, zeros.data(), shp_buf_size);
auto zeros_value_node = std::make_shared<ValueNode>(zeros_tensor);
MS_EXCEPTION_IF_NULL(zeros_value_node);
kernel_graph_ptr->AddValueNodeToGraph(zeros_value_node);
auto kernel_info = std::make_shared<device::KernelInfo>();
MS_EXCEPTION_IF_NULL(kernel_info);
reset_value_node->set_kernel_info(kernel_info);
zeros_value_node->set_kernel_info(kernel_info);
kernel::KernelBuildInfo::KernelBuildInfoBuilder builder1;
builder1.SetOutputsFormat({kOpFormat_DEFAULT});
builder1.SetOutputsDeviceType({kNumberTypeFloat32});
AnfAlgo::SetSelectKernelBuildInfo(builder1.Build(), reset_value_node.get());
builder1.SetOutputsDeviceType({kNumberTypeInt32});
AnfAlgo::SetSelectKernelBuildInfo(builder1.Build(), zeros_value_node.get());
return zeros_value_node;
}
CNodePtr KernelAdjust::CreateAssign(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const AnfNodePtr &specify_para, const AnfNodePtr &data) const {
MS_EXCEPTION_IF_NULL(kernel_graph_ptr);
MS_EXCEPTION_IF_NULL(specify_para);
auto assign_primitive = std::make_shared<Primitive>(kAssignOpName);
std::vector<AnfNodePtr> assign_inputs = {NewValueNode(assign_primitive), specify_para, reset_value_node};
std::vector<AnfNodePtr> assign_inputs = {NewValueNode(assign_primitive), specify_para, data};
auto assign_cnode = kernel_graph_ptr->NewCNode(assign_inputs);
MS_EXCEPTION_IF_NULL(assign_cnode);
assign_cnode->set_scope(kDefaultScope);
assign_cnode->set_abstract(specify_para->abstract());
kernel::KernelBuildInfo::KernelBuildInfoBuilder selected_kernel_builder = CreateMngKernelBuilder(
{kOpFormat_DEFAULT, kOpFormat_DEFAULT}, {TypeId::kNumberTypeFloat32, TypeId::kNumberTypeFloat32});
{kOpFormat_DEFAULT, kOpFormat_DEFAULT}, {TypeId::kNumberTypeInt32, TypeId::kNumberTypeInt32});
selected_kernel_builder.SetOutputsFormat({kOpFormat_DEFAULT});
selected_kernel_builder.SetOutputsDeviceType({kNumberTypeFloat32});
selected_kernel_builder.SetOutputsDeviceType({kNumberTypeInt32});
AnfAlgo::SetSelectKernelBuildInfo(selected_kernel_builder.Build(), assign_cnode.get());
std::vector<std::string> input_names = {"ref", "value"};
@ -944,7 +928,8 @@ void KernelAdjust::InsertGradientOverflowCheckOperations(
MS_EXCEPTION_IF_NULL(kernel_graph_ptr);
bool first_grad_op = true;
CNodePtr npu_alloc_cnode;
auto status_value_node = CreateZerosValueNode(kernel_graph_ptr);
auto reset_value_node = CreateZerosValueNode(kernel_graph_ptr);
std::vector<CNodePtr> new_execution_order;
auto execution_order = kernel_graph_ptr->execution_order();
for (size_t i = 0; i < execution_order.size() - 1; i++) {
@ -956,39 +941,37 @@ void KernelAdjust::InsertGradientOverflowCheckOperations(
if (cur_full_name.find(kGradients) == std::string::npos && next_full_name.find(kGradients) != std::string::npos) {
if (first_grad_op) {
npu_alloc_cnode = CreateNPUAllocStatus(kernel_graph_ptr);
auto npu_clear_cnode = CreateNPUClearStatus(kernel_graph_ptr, npu_alloc_cnode);
auto assign_cnode = CreateAssign(kernel_graph_ptr, specify_para);
AnfAlgo::SetStreamId(next_stream_id, npu_alloc_cnode.get());
auto npu_clear_cnode = CreateNPUClearStatusV2(kernel_graph_ptr, status_value_node);
auto assign_cnode = CreateAssign(kernel_graph_ptr, specify_para, reset_value_node);
AnfAlgo::SetStreamId(next_stream_id, status_value_node.get());
AnfAlgo::SetStreamId(next_stream_id, npu_clear_cnode.get());
AnfAlgo::SetStreamId(next_stream_id, assign_cnode.get());
new_execution_order.push_back(npu_alloc_cnode);
new_execution_order.push_back(npu_clear_cnode);
new_execution_order.push_back(assign_cnode);
first_grad_op = false;
} else {
auto npu_clear_cnode = CreateNPUClearStatus(kernel_graph_ptr, npu_alloc_cnode);
auto npu_clear_cnode = CreateNPUClearStatusV2(kernel_graph_ptr, status_value_node);
AnfAlgo::SetStreamId(next_stream_id, npu_clear_cnode.get());
new_execution_order.push_back(npu_clear_cnode);
}
}
if (cur_full_name.find(kGradients) != std::string::npos && next_full_name.find(kGradients) == std::string::npos) {
auto npu_get_cnode = CreateNPUGetFloatStatus(kernel_graph_ptr, npu_alloc_cnode);
auto assign_add_cnode = CreateAssignAdd(kernel_graph_ptr, npu_alloc_cnode, specify_para);
auto npu_get_cnode = CreateNPUGetFloatStatusV2(kernel_graph_ptr, status_value_node);
auto assign_status_node = CreateAssign(kernel_graph_ptr, specify_para, npu_get_cnode);
AnfAlgo::SetStreamId(cur_stream_id, npu_get_cnode.get());
AnfAlgo::SetStreamId(cur_stream_id, npu_get_cnode.get());
AnfAlgo::SetStreamId(cur_stream_id, assign_add_cnode.get());
new_execution_order.push_back(npu_get_cnode);
new_execution_order.push_back(assign_add_cnode);
new_execution_order.push_back(assign_status_node);
}
if (i == execution_order.size() - kLastHandleDiff) {
new_execution_order.push_back(execution_order[i + 1]);
if (next_full_name.find(kGradients) != std::string::npos) {
auto npu_get_cnode = CreateNPUGetFloatStatus(kernel_graph_ptr, npu_alloc_cnode);
auto assign_add_cnode = CreateAssignAdd(kernel_graph_ptr, npu_alloc_cnode, specify_para);
auto npu_get_cnode = CreateNPUGetFloatStatusV2(kernel_graph_ptr, status_value_node);
auto assign_status_node = CreateAssign(kernel_graph_ptr, specify_para, npu_get_cnode);
AnfAlgo::SetStreamId(cur_stream_id, npu_get_cnode.get());
AnfAlgo::SetStreamId(cur_stream_id, assign_add_cnode.get());
AnfAlgo::SetStreamId(cur_stream_id, assign_status_node.get());
new_execution_order.push_back(npu_get_cnode);
new_execution_order.push_back(assign_add_cnode);
new_execution_order.push_back(assign_status_node);
}
}
}
@ -1030,18 +1013,16 @@ void KernelAdjust::InsertDynamicLossScaleCheckOperations(const std::shared_ptr<s
bool first_layer_op = true;
std::vector<CNodePtr> new_execution_order;
int64_t cur_param = static_cast<int64_t>(dynamic_loss_scale_param_list->size()) - 1;
CNodePtr npu_alloc_cnode;
auto status_value_node = CreateZerosValueNode(kernel_graph_ptr);
auto reset_value_node = CreateZerosValueNode(kernel_graph_ptr);
std::set<int64_t> viewed_id;
for (size_t i = 0; i < execution_order.size(); ++i) {
auto cur_node = execution_order[i];
auto cur_stream_id = AnfAlgo::GetStreamId(cur_node);
if (common::AnfAlgo::HasNodeAttr(kSplitOverFlow, cur_node) || (i == end_gradient_index)) {
if (first_layer_op) {
npu_alloc_cnode = CreateNPUAllocStatus(kernel_graph_ptr);
AnfAlgo::SetStreamId(cur_stream_id, npu_alloc_cnode.get());
(void)new_execution_order.emplace_back(npu_alloc_cnode);
for (const auto &param : *dynamic_loss_scale_param_list) {
auto assign_cnode = CreateAssign(kernel_graph_ptr, param);
auto assign_cnode = CreateAssign(kernel_graph_ptr, param, reset_value_node);
AnfAlgo::SetStreamId(cur_stream_id, assign_cnode.get());
(void)new_execution_order.emplace_back(assign_cnode);
}
@ -1055,22 +1036,19 @@ void KernelAdjust::InsertDynamicLossScaleCheckOperations(const std::shared_ptr<s
(void)new_execution_order.emplace_back(cur_node);
continue;
}
if (viewed_id.count(cur_param) != 0) {
auto assign_cnode = CreateAssign(kernel_graph_ptr, dynamic_loss_scale_param_list->at(cur_param));
AnfAlgo::SetStreamId(cur_stream_id, assign_cnode.get());
(void)new_execution_order.emplace_back(assign_cnode);
}
auto npu_get_cnode = CreateNPUGetFloatStatus(kernel_graph_ptr, npu_alloc_cnode);
auto npu_get_cnode = CreateNPUGetFloatStatusV2(kernel_graph_ptr, status_value_node);
AnfAlgo::SetStreamId(cur_stream_id, npu_get_cnode.get());
(void)new_execution_order.emplace_back(npu_get_cnode);
auto assign_add_cnode =
CreateAssignAdd(kernel_graph_ptr, npu_alloc_cnode, dynamic_loss_scale_param_list->at(cur_param));
AnfAlgo::SetStreamId(cur_stream_id, assign_add_cnode.get());
(void)new_execution_order.emplace_back(assign_add_cnode);
auto assign_status_node =
CreateAssign(kernel_graph_ptr, dynamic_loss_scale_param_list->at(cur_param), npu_get_cnode);
AnfAlgo::SetStreamId(cur_stream_id, assign_status_node.get());
(void)new_execution_order.emplace_back(assign_status_node);
(void)viewed_id.insert(cur_param);
cur_param--;
}
auto npu_clear_cnode = CreateNPUClearStatus(kernel_graph_ptr, npu_alloc_cnode);
auto npu_clear_cnode = CreateNPUClearStatusV2(kernel_graph_ptr, status_value_node);
AnfAlgo::SetStreamId(cur_stream_id, npu_clear_cnode.get());
(void)new_execution_order.emplace_back(npu_clear_cnode);
}

14
mindspore/ccsrc/plugin/device/ascend/hal/device/kernel_adjust.h
View File

@ -80,15 +80,15 @@ class KernelAdjust {
KernelAdjust() = default;
~KernelAdjust() = default;
CNodePtr CreateNPUGetFloatStatus(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const CNodePtr &npu_alloc_cnode) const;
CNodePtr CreateNPUClearStatus(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const CNodePtr &npu_alloc_cnode) const;
CNodePtr CreateNPUAllocStatus(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr) const;
AnfNodePtr CreateZerosValueNode(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr) const;
CNodePtr CreateNPUGetFloatStatusV2(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const AnfNodePtr &status_value_node) const;
CNodePtr CreateNPUClearStatusV2(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const AnfNodePtr &status_value_node) const;
CNodePtr CreateAssignAdd(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const CNodePtr &npu_alloc_cnode, const AnfNodePtr &specify_para) const;
CNodePtr CreateAssign(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const AnfNodePtr &specify_para) const;
CNodePtr CreateAssign(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr, const AnfNodePtr &specify_para,
const AnfNodePtr &data) const;
void ReorderGetNext(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr) const;
CNodePtr CreateStreamSwitchOp(const std::shared_ptr<session::KernelGraph> &kernel_graph_ptr,
const std::map<std::string, mindspore::ParameterPtr> &switch_loop_input,

5
mindspore/ccsrc/plugin/device/ascend/kernel/hccl/hccl_kernel.cc
View File

@ -305,10 +305,11 @@ std::vector<TaskInfoPtr> HcclKernel::GenTask(const std::vector<AddressPtr> &inpu
}
std::vector<void *> global_workspace_addr;
auto overflow_memory_ptr =
device::ascend::AscendMemAdapter::GetInstance().MallocOverflowMem(anf_node_.lock()->cast<CNodePtr>());
auto overflow_memory_ptr = device::ascend::AscendMemAdapter::GetInstance().MallocOverflowMem();
MS_EXCEPTION_IF_NULL(overflow_memory_ptr);
global_workspace_addr.push_back(reinterpret_cast<void *>(overflow_memory_ptr));
MS_LOG(DEBUG) << "Assign overflow memory for node " << anf_node->fullname_with_scope() << ", addr is "
<< reinterpret_cast<void *>(overflow_memory_ptr);
HcclTaskInfoPtr hcclTaskInfo =
std::make_shared<HcclTaskInfo>(unique_name_, stream_id, hccl::HcclAdapter::GetHcclType(anf_node), input_data_addr,

7
mindspore/ccsrc/plugin/device/ascend/kernel/tbe/dynamic_tbe_kernel_mod.cc
View File

@ -289,6 +289,13 @@ bool DynamicTbeKernelMod::Launch(const std::vector<AddressPtr> &inputs, const st
runtimeargs.push_back(tiling_data_ptr_);
}
AddressPtr overflow_address_ptr = GetOverflowAddress();
if (overflow_address_ptr != nullptr) {
runtimeargs.emplace_back(overflow_address_ptr->addr);
MS_LOG(DEBUG) << "Assign overflow memory for node " << node->fullname_with_scope() << ", addr is "
<< overflow_address_ptr->addr;
}
rtL2Ctrl_t *l2ctrl = nullptr;
auto args_size = static_cast<uint32_t>(UlongToUint(sizeof(void *)) * runtimeargs.size());
auto node_info = cnode->fullname_with_scope();

38
mindspore/ccsrc/plugin/device/ascend/kernel/tbe/tbe_json/single_tbe_json_creator.cc
View File

@ -58,6 +58,43 @@ bool SingleTbeJsonCreator::GenJson(const AnfNodePtr &anf_node, nlohmann::json *k
return true;
}
void NpuClearV2PostProcessing(const AnfNodePtr &anf_node, std::vector<nlohmann::json> *op_list_json) {
if (op_list_json->size() != 2) {
MS_LOG(ERROR) << "Op list json's size is not equal to 2, abort post processing.";
}
auto compute_json = (*op_list_json)[1];
std::vector<nlohmann::json> empty_vector_json;
compute_json[kJInputDesc] = empty_vector_json;
compute_json[kJOutputDataDesc] = empty_vector_json;
compute_json[kJOutputDesc] = empty_vector_json;
op_list_json->clear();
(*op_list_json).emplace_back(compute_json);
MS_LOG(DEBUG) << "Op list json after post processing:" << compute_json.dump();
}
void NpuGetV2PostProcessing(const AnfNodePtr &anf_node, std::vector<nlohmann::json> *op_list_json) {
if (op_list_json->size() != 2) {
MS_LOG(ERROR) << "Op list json's size is not equal to 2, abort post processing.";
}
auto compute_json = (*op_list_json)[1];
std::vector<nlohmann::json> empty_vector_json;
compute_json[kJInputDesc] = empty_vector_json;
op_list_json->clear();
(*op_list_json).emplace_back(compute_json);
MS_LOG(DEBUG) << "Op list json after post processing:" << compute_json.dump();
}
void SingleTbeJsonCreator::OpListPostProcessing(const AnfNodePtr &anf_node, std::vector<nlohmann::json> *op_list_json) {
auto kernel_name = common::AnfAlgo::GetCNodeName(anf_node);
if (kernel_name == kNPUClearFloatStatusV2OpName) {
NpuClearV2PostProcessing(anf_node, op_list_json);
} else if (kernel_name == kNPUGetFloatStatusV2OpName) {
NpuGetV2PostProcessing(anf_node, op_list_json);
}
}
bool SingleTbeJsonCreator::GenOpListJson(const AnfNodePtr &anf_node, std::vector<nlohmann::json> *op_list_json) {
MS_EXCEPTION_IF_NULL(anf_node);
MS_EXCEPTION_IF_NULL(op_list_json);
@ -69,6 +106,7 @@ bool SingleTbeJsonCreator::GenOpListJson(const AnfNodePtr &anf_node, std::vector
}
GenDataJson(anf_node, compute_json, op_list_json);
(*op_list_json).push_back(compute_json);
OpListPostProcessing(anf_node, op_list_json);
MS_LOG(DEBUG) << "End.";
return true;
}

1
mindspore/ccsrc/plugin/device/ascend/kernel/tbe/tbe_json/single_tbe_json_creator.h
View File

@ -29,6 +29,7 @@ class SingleTbeJsonCreator : public TbeJsonCreator {
protected:
bool GenOpListJson(const AnfNodePtr &anf_node, std::vector<nlohmann::json> *op_list_json);
void OpListPostProcessing(const AnfNodePtr &anf_node, std::vector<nlohmann::json> *op_list_json);
void GenDataJson(const AnfNodePtr &anf_node, const nlohmann::json &compute_json,
std::vector<nlohmann::json> *op_list_json) const;
virtual void GenInputDescJson(const AnfNodePtr &anf_node, size_t real_input_index, nlohmann::json *input_desc);

20
mindspore/ccsrc/plugin/device/ascend/kernel/tbe/tbe_kernel_compile.cc
View File

@ -571,15 +571,17 @@ void TbeKernelCompileManager::Query(const std::string &type) {
std::pair<std::vector<CNodePtr>, std::vector<CNodePtr>> TbeKernelCompileManager::GenKernelMod(
const std::vector<CNodePtr> &node_list) {
MS_LOG(INFO) << "Gen kernel mod start!";
std::vector<CNodePtr> success_node;
std::vector<CNodePtr> failed_node;
std::vector<CNodePtr> success_nodes;
std::vector<CNodePtr> failed_nodes;
for (auto &node : node_list) {
MS_EXCEPTION_IF_NULL(node);
if (AnfAlgo::GetKernelMod(node) != nullptr) {
(void)success_node.emplace_back(node);
(void)success_nodes.emplace_back(node);
continue; // kernel mod already exist, continue;
}
auto op_name = common::AnfAlgo::GetCNodeName(node);
auto full_name = node->fullname_with_scope();
if (common::AnfAlgo::HasNodeAttr(kAttrOriFusionName, node)) {
full_name = common::AnfAlgo::GetNodeAttr<std::string>(node, kAttrOriFusionName);
@ -592,7 +594,7 @@ std::pair<std::vector<CNodePtr>, std::vector<CNodePtr>> TbeKernelCompileManager:
kernel_pack = bin_map->SearchInFile(json_name);
if (kernel_pack == nullptr) {
MS_LOG(INFO) << "Can not find .json file or the .o file for op:" << json_name << trace::DumpSourceLines(node);
(void)failed_node.emplace_back(node);
(void)failed_nodes.emplace_back(node);
continue;
}
}
@ -612,11 +614,17 @@ std::pair<std::vector<CNodePtr>, std::vector<CNodePtr>> TbeKernelCompileManager:
kernel_mod_ptr->SetInputSizeList(iter->second.input_size_list);
kernel_mod_ptr->SetOutputSizeList(iter->second.output_size_list);
kernel_mod_ptr->SetWorkspaceSizeList(kernel_info_json.workspaces);
if (op_name == kNPUClearFloatStatusV2OpName || op_name == kNPUGetFloatStatusV2OpName) {
constexpr size_t io_byte_size = 32;
const std::vector<size_t> size_list = {io_byte_size};
kernel_mod_ptr->SetInputSizeList(size_list);
kernel_mod_ptr->SetOutputSizeList(size_list);
}
AnfAlgo::SetKernelMod(kernel_mod_ptr, node.get());
(void)success_node.emplace_back(node);
(void)success_nodes.emplace_back(node);
}
MS_LOG(INFO) << "Gen kernel mod end!";
return std::make_pair(success_node, failed_node);
return std::make_pair(success_nodes, failed_nodes);
}
void TbeKernelCompileManager::UpdateFusionTypeAndOutputDataDesc(const std::vector<CNodePtr> &nodes) {

78
mindspore/ccsrc/plugin/device/ascend/kernel/tbe/tbe_kernel_mod.cc
View File

@ -21,12 +21,12 @@
#include "utils/ms_context.h"
#include "plugin/device/ascend/hal/device/ge_runtime/task_info.h"
#include "runtime/device/kernel_runtime.h"
#include "plugin/device/ascend/hal/device/ascend_memory_adapter.h"
namespace mindspore {
namespace kernel {
using TbeTaskInfoPtr = std::shared_ptr<mindspore::ge::model_runner::TbeTaskInfo>;
using tbe::KernelManager;
using AddressPtrList = std::vector<mindspore::kernel::AddressPtr>;
bool TbeKernelMod::Launch(const std::vector<mindspore::kernel::AddressPtr> &inputs,
const std::vector<mindspore::kernel::AddressPtr> &workspace,
const std::vector<mindspore::kernel::AddressPtr> &outputs, void *stream_ptr) {
@ -60,14 +60,23 @@ bool TbeKernelMod::Launch(const std::vector<mindspore::kernel::AddressPtr> &inpu
return false;
}
auto node = anf_node_.lock();
MS_EXCEPTION_IF_NULL(node);
auto cnode = node->cast<CNodePtr>();
MS_EXCEPTION_IF_NULL(cnode);
std::vector<mindspore::kernel::AddressPtr> real_inputs;
std::vector<mindspore::kernel::AddressPtr> real_outputs;
GetRealIOAddress(cnode, inputs, outputs, &real_inputs, &real_outputs);
// pack all addresses into a vector.
std::vector<void *> runtimeargs;
(void)std::transform(std::begin(inputs), std::end(inputs), std::back_inserter(runtimeargs),
(void)std::transform(std::begin(real_inputs), std::end(real_inputs), std::back_inserter(runtimeargs),
[](const AddressPtr &input) -> void * {
MS_EXCEPTION_IF_NULL(input);
return input->addr;
});
(void)std::transform(std::begin(outputs), std::end(outputs), std::back_inserter(runtimeargs),
(void)std::transform(std::begin(real_outputs), std::end(real_outputs), std::back_inserter(runtimeargs),
[](const AddressPtr &output) -> void * {
MS_EXCEPTION_IF_NULL(output);
return output->addr;
@ -79,6 +88,14 @@ bool TbeKernelMod::Launch(const std::vector<mindspore::kernel::AddressPtr> &inpu
return addr->addr;
});
}
AddressPtr overflow_address_ptr = GetOverflowAddress();
if (overflow_address_ptr != nullptr) {
runtimeargs.emplace_back(overflow_address_ptr->addr);
MS_LOG(DEBUG) << "Assign overflow memory for node " << cnode->fullname_with_scope() << ", addr is "
<< overflow_address_ptr->addr;
}
rtL2Ctrl_t *l2ctrl = nullptr;
const void *stubFunc = reinterpret_cast<void *>(func_stub);
auto argsSize = static_cast<uint32_t>(UlongToUint(sizeof(void *)) * runtimeargs.size());
@ -106,13 +123,22 @@ std::vector<TaskInfoPtr> TbeKernelMod::GenTask(const std::vector<AddressPtr> &in
std::vector<void *> output_data_addrs;
std::vector<void *> workspace_addrs;
auto node = anf_node_.lock();
MS_EXCEPTION_IF_NULL(node);
auto cnode = node->cast<CNodePtr>();
MS_EXCEPTION_IF_NULL(cnode);
std::vector<mindspore::kernel::AddressPtr> real_inputs;
std::vector<mindspore::kernel::AddressPtr> real_outputs;
GetRealIOAddress(cnode, inputs, outputs, &real_inputs, &real_outputs);
// pack all addresses into a vector.
(void)std::transform(std::begin(inputs), std::end(inputs), std::back_inserter(input_data_addrs),
(void)std::transform(std::begin(real_inputs), std::end(real_inputs), std::back_inserter(input_data_addrs),
[](const AddressPtr &input) -> void * {
MS_EXCEPTION_IF_NULL(input);
return input->addr;
});
(void)std::transform(std::begin(outputs), std::end(outputs), std::back_inserter(output_data_addrs),
(void)std::transform(std::begin(real_outputs), std::end(real_outputs), std::back_inserter(output_data_addrs),
[](const AddressPtr &output) -> void * {
MS_EXCEPTION_IF_NULL(output);
return output->addr;
@ -125,6 +151,13 @@ std::vector<TaskInfoPtr> TbeKernelMod::GenTask(const std::vector<AddressPtr> &in
});
}
AddressPtr overflow_address_ptr = GetOverflowAddress();
if (overflow_address_ptr != nullptr) {
workspace_addrs.emplace_back(overflow_address_ptr->addr);
MS_LOG(DEBUG) << "Assign overflow memory for node " << cnode->fullname_with_scope() << ", addr is "
<< overflow_address_ptr->addr;
}
stream_id_ = stream_id;
auto funcstub = KernelManager::GenFuncStub(*kernel_pack_, false, &block_dim_, nullptr);
if (funcstub == 0) {
@ -146,5 +179,40 @@ vector<size_t> TbeKernelMod::GenParameters() {
auto kernel_json_info = kernel_pack_->kernel_json_info();
return kernel_json_info.parameters;
}
AddressPtr TbeKernelMod::GetOverflowAddress() {
AddressPtr overflow_address_ptr = nullptr;
auto is_overflow = kernel_pack_.get()->kernel_json_info().global_workspace.is_overflow;
if (is_overflow) {
constexpr size_t size = 32;
auto overflow_memory_ptr = device::ascend::AscendMemAdapter::GetInstance().MallocOverflowMem();
MS_EXCEPTION_IF_NULL(overflow_memory_ptr);
overflow_address_ptr = std::make_shared<kernel::Address>();
overflow_address_ptr->addr = reinterpret_cast<void *>(overflow_memory_ptr);
overflow_address_ptr->size = size;
}
return overflow_address_ptr;
}
void TbeKernelMod::GetRealIOAddress(const AnfNodePtr &cnode, const vector<AddressPtr> &inputs,
const vector<AddressPtr> &outputs,
vector<mindspore::kernel::AddressPtr> *real_inputs,
vector<mindspore::kernel::AddressPtr> *real_outputs) const {
auto op_name = common::AnfAlgo::GetCNodeName(cnode);
MS_EXCEPTION_IF_NULL(real_inputs);
MS_EXCEPTION_IF_NULL(real_outputs);
*real_inputs = inputs;
*real_outputs = outputs;
if (op_name == kNPUClearFloatStatusV2OpName) {
// NPUClearFloatStatusV2 has no input output.
real_inputs->clear();
real_outputs->clear();
MS_LOG(INFO) << "Clear Node " << cnode->fullname_with_scope() << "'s inputs and outputs";
} else if (op_name == kNPUGetFloatStatusV2OpName) {
// NPUGetFloatStatusV2 has no input
real_inputs->clear();
MS_LOG(INFO) << "Clear Node " << cnode->fullname_with_scope() << "'s inputs";
}
}
} // namespace kernel
} // namespace mindspore

4
mindspore/ccsrc/plugin/device/ascend/kernel/tbe/tbe_kernel_mod.h
View File

@ -42,6 +42,10 @@ class TbeKernelMod : public AscendKernelMod {
std::vector<TaskInfoPtr> GenTask(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &workspaces,
const std::vector<AddressPtr> &outputs, uint32_t stream_id) override;
std::vector<size_t> GenParameters() override;
AddressPtr GetOverflowAddress();
void GetRealIOAddress(const AnfNodePtr &cnode, const std::vector<AddressPtr> &inputs,
const std::vector<AddressPtr> &outputs, std::vector<AddressPtr> *real_inputs,
std::vector<AddressPtr> *real_outputs) const;
protected:
KernelPackPtr kernel_pack_;

2
mindspore/ccsrc/plugin/device/ascend/kernel/tbe/tbe_utils.cc
View File

@ -208,6 +208,8 @@ nlohmann::json TbeUtils::GenSocInfo() {
soc_info_json["op_debug_config"] = GetOpDebugConfig();
soc_info_json["autoTilingMode"] = context_ptr->get_param<std::string>(MS_CTX_TUNE_MODE);
soc_info_json["deviceId"] = std::to_string(context_ptr->get_param<uint32_t>(MS_CTX_DEVICE_ID));
soc_info_json["status_check"] = "true";
std::string config_path;
if (!Common::CommonFuncForConfigPath("", common::GetEnv("OP_BANK_PATH"), &config_path)) {
MS_LOG(EXCEPTION) << "Invalid environment variable 'OP_BANK_PATH', the path is " << common::GetEnv("OP_BANK_PATH")

2
mindspore/core/ops/core_ops.h
View File

@ -1608,6 +1608,8 @@ GVAR_DEF(PrimitivePtr, kPrimPush, std::make_shared<Primitive>("Push"));
GVAR_DEF(PrimitivePtr, kPrimNPUGetFloatStatus, std::make_shared<Primitive>("NPUGetFloatStatus"));
GVAR_DEF(PrimitivePtr, kPrimNPUAllocFloatStatus, std::make_shared<Primitive>("NPUAllocFloatStatus"));
GVAR_DEF(PrimitivePtr, kPrimNPUClearFloatStatus, std::make_shared<Primitive>("NPUClearFloatStatus"));
GVAR_DEF(PrimitivePtr, kPrimNPUGetFloatStatusV2, std::make_shared<Primitive>("NPUGetFloatStatusV2"));
GVAR_DEF(PrimitivePtr, kPrimNPUClearFloatStatusV2, std::make_shared<Primitive>("NPUClearFloatStatusV2"));
GVAR_DEF(PrimitivePtr, kPrimPyFunc, std::make_shared<Primitive>("PyFunc"));
GVAR_DEF(PrimitivePtr, kPrimDynamicLossScale, std::make_shared<Primitive>("_DynamicLossScale"));
GVAR_DEF(PrimitivePtr, kPrimScaleGrad, std::make_shared<Primitive>("ScaleGrad"));

99
mindspore/core/ops/npu_clear_float_status_v2.cc Normal file
View File

@ -0,0 +1,99 @@
/**
* Copyright 2023 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 <map>
#include <set>
#include <string>
#include "ops/npu_clear_float_status_v2.h"
#include "ops/op_utils.h"
#include "abstract/param_validator.h"
#include "utils/check_convert_utils.h"
#include "abstract/ops/primitive_infer_map.h"
#include "mindapi/src/helper.h"
namespace mindspore {
namespace ops {
namespace {
abstract::ShapePtr NPUClearFloatStatusV2InferShape(const PrimitivePtr &,
const std::vector<AbstractBasePtr> &input_args) {
auto input_shape = CheckAndConvertUtils::ConvertShapePtrToShapeMap(input_args[0]->BuildShape())[kShape];
// dynamic rank
if (IsDynamicRank(input_shape)) {
return std::make_shared<abstract::Shape>(ShapeVector{abstract::Shape::kShapeRankAny});
}
// dynamic shape
if (IsDynamic(input_shape)) {
ShapeVector out_shape_dyn;
for (size_t i = 0; i < input_shape.size(); ++i) {
out_shape_dyn.push_back(abstract::Shape::kShapeDimAny);
}
return std::make_shared<abstract::Shape>(out_shape_dyn);
}
const int64_t normal_shape_size = 1;
const int64_t normal_shape_len = 8;
if (input_shape.size() != normal_shape_size) {
MS_EXCEPTION(ValueError) << "Input_x must be a 1-dimensional tensor, but got " << std::to_string(input_shape.size())
<< "-dimensional tensor.";
}
if (input_shape[0] != normal_shape_len) {
MS_EXCEPTION(ValueError) << "The first dimension of input_x must be 8, but got " << std::to_string(input_shape[0]);
}
std::vector<int64_t> output_shape = {normal_shape_len};
return std::make_shared<abstract::Shape>(output_shape);
}
TypePtr NPUClearFloatStatusV2InferType(const PrimitivePtr &primitive, const std::vector<AbstractBasePtr> &input_args) {
std::map<std::string, TypePtr> types;
std::set<TypePtr> valid_types = {kInt32};
TypePtr input_x_type = input_args[0]->BuildType();
(void)types.emplace("input_x", input_x_type);
(void)CheckAndConvertUtils::CheckTensorTypeSame(types, valid_types, primitive->name());
return kInt32;
}
} // namespace
MIND_API_OPERATOR_IMPL(NPUClearFloatStatusV2, BaseOperator);
AbstractBasePtr NPUClearFloatStatusV2Infer(const abstract::AnalysisEnginePtr &, const PrimitivePtr &primitive,
const std::vector<AbstractBasePtr> &input_args) {
MS_EXCEPTION_IF_NULL(primitive);
const int64_t input_num = 1;
CheckAndConvertUtils::CheckInputArgs(input_args, kEqual, input_num, primitive->name());
auto infer_type = NPUClearFloatStatusV2InferType(primitive, input_args);
auto infer_shape = NPUClearFloatStatusV2InferShape(primitive, input_args);
return abstract::MakeAbstract(infer_shape, infer_type);
}
// AG means auto generated
class MIND_API AGNPUClearFloatStatusV2Infer : public abstract::OpInferBase {
public:
BaseShapePtr InferShape(const PrimitivePtr &primitive,
const std::vector<AbstractBasePtr> &input_args) const override {
return NPUClearFloatStatusV2InferShape(primitive, input_args);
}
TypePtr InferType(const PrimitivePtr &primitive, const std::vector<AbstractBasePtr> &input_args) const override {
return NPUClearFloatStatusV2InferType(primitive, input_args);
}
AbstractBasePtr InferShapeAndType(const abstract::AnalysisEnginePtr &engine, const PrimitivePtr &primitive,
const std::vector<AbstractBasePtr> &input_args) const override {
return NPUClearFloatStatusV2Infer(engine, primitive, input_args);
}
};
REGISTER_PRIMITIVE_OP_INFER_IMPL(NPUClearFloatStatusV2, prim::kPrimNPUClearFloatStatusV2, AGNPUClearFloatStatusV2Infer,
false);
} // namespace ops
} // namespace mindspore

39
mindspore/core/ops/npu_clear_float_status_v2.h Normal file
View File

@ -0,0 +1,39 @@
/**
* Copyright 2023 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.
*/
#ifndef MINDSPORE_CORE_OPS_NPU_CLEAR_FLOAT_STATUS_V2_H_
#define MINDSPORE_CORE_OPS_NPU_CLEAR_FLOAT_STATUS_V2_H_
#include <vector>
#include <memory>
#include "ops/base_operator.h"
#include "mindapi/base/types.h"
namespace mindspore {
namespace ops {
constexpr auto kNameNPUClearFloatStatusV2 = "NPUClearFloatStatusV2";
class MIND_API NPUClearFloatStatusV2 : public BaseOperator {
public:
MIND_API_BASE_MEMBER(NPUClearFloatStatusV2);
NPUClearFloatStatusV2() : BaseOperator(kNameNPUClearFloatStatusV2) { InitIOName({"addr"}, {"data"}); }
void Init() const {}
};
MIND_API abstract::AbstractBasePtr NPUClearFloatStatusV2Infer(const abstract::AnalysisEnginePtr &,
const PrimitivePtr &primitive,
const std::vector<abstract::AbstractBasePtr> &input_args);
} // namespace ops
} // namespace mindspore
#endif // MINDSPORE_CORE_OPS_NPU_CLEAR_FLOAT_STATUS_V2_H_

99
mindspore/core/ops/npu_get_float_status_v2.cc Normal file
View File

@ -0,0 +1,99 @@
/**
* Copyright 2023 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 <map>
#include <set>
#include <string>
#include "ops/npu_get_float_status_v2.h"
#include "ops/op_utils.h"
#include "abstract/param_validator.h"
#include "utils/check_convert_utils.h"
#include "abstract/ops/primitive_infer_map.h"
#include "mindapi/src/helper.h"
namespace mindspore {
namespace ops {
namespace {
abstract::ShapePtr NPUGetFloatStatusV2InferShape(const PrimitivePtr &, const std::vector<AbstractBasePtr> &input_args) {
auto input_shape = CheckAndConvertUtils::ConvertShapePtrToShapeMap(input_args[0]->BuildShape())[kShape];
// dynamic rank
if (IsDynamicRank(input_shape)) {
return std::make_shared<abstract::Shape>(ShapeVector{abstract::Shape::kShapeRankAny});
}
// dynamic shape
if (IsDynamic(input_shape)) {
ShapeVector out_shape_dyn;
for (size_t i = 0; i < input_shape.size(); ++i) {
out_shape_dyn.push_back(abstract::Shape::kShapeDimAny);
}
return std::make_shared<abstract::Shape>(out_shape_dyn);
}
const int64_t normal_shape_size = 1;
const int64_t normal_shape_len = 8;
if (input_shape.size() != normal_shape_size) {
MS_EXCEPTION(ValueError) << "Input_x must be a 1-dimensional tensor, but got " << std::to_string(input_shape.size())
<< "-dimensional tensor.";
}
if (input_shape[0] != normal_shape_len) {
MS_EXCEPTION(ValueError) << "The first dimension of input_x must be 8, but got " << std::to_string(input_shape[0]);
}
std::vector<int64_t> output_shape = {normal_shape_len};
return std::make_shared<abstract::Shape>(output_shape);
}
TypePtr NPUGetFloatStatusV2InferType(const PrimitivePtr &primitive, const std::vector<AbstractBasePtr> &input_args) {
std::map<std::string, TypePtr> types;
std::set<TypePtr> valid_types = {kInt32};
TypePtr input_x_type = input_args[0]->BuildType();
(void)types.emplace("input_x", input_x_type);
(void)CheckAndConvertUtils::CheckTensorTypeSame(types, valid_types, primitive->name());
return kInt32;
}
} // namespace
MIND_API_OPERATOR_IMPL(NPUGetFloatStatusV2, BaseOperator);
AbstractBasePtr NPUGetFloatStatusV2Infer(const abstract::AnalysisEnginePtr &, const PrimitivePtr &primitive,
const std::vector<AbstractBasePtr> &input_args) {
MS_EXCEPTION_IF_NULL(primitive);
const int64_t input_num = 1;
CheckAndConvertUtils::CheckInputArgs(input_args, kEqual, input_num, primitive->name());
auto infer_type = NPUGetFloatStatusV2InferType(primitive, input_args);
auto infer_shape = NPUGetFloatStatusV2InferShape(primitive, input_args);
return abstract::MakeAbstract(infer_shape, infer_type);
}
// AG means auto generated
class MIND_API AGNPUGetFloatStatusV2Infer : public abstract::OpInferBase {
public:
BaseShapePtr InferShape(const PrimitivePtr &primitive,
const std::vector<AbstractBasePtr> &input_args) const override {
return NPUGetFloatStatusV2InferShape(primitive, input_args);
}
TypePtr InferType(const PrimitivePtr &primitive, const std::vector<AbstractBasePtr> &input_args) const override {
return NPUGetFloatStatusV2InferType(primitive, input_args);
}
AbstractBasePtr InferShapeAndType(const abstract::AnalysisEnginePtr &engine, const PrimitivePtr &primitive,
const std::vector<AbstractBasePtr> &input_args) const override {
return NPUGetFloatStatusV2Infer(engine, primitive, input_args);
}
};
REGISTER_PRIMITIVE_OP_INFER_IMPL(NPUGetFloatStatusV2, prim::kPrimNPUGetFloatStatusV2, AGNPUGetFloatStatusV2Infer,
false);
} // namespace ops
} // namespace mindspore

39
mindspore/core/ops/npu_get_float_status_v2.h Normal file
View File

@ -0,0 +1,39 @@
/**
* Copyright 2023 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.
*/
#ifndef MINDSPORE_CORE_OPS_NPU_GET_FLOAT_STATUS_V2_H_
#define MINDSPORE_CORE_OPS_NPU_GET_FLOAT_STATUS_V2_H_
#include <vector>
#include <memory>
#include "ops/base_operator.h"
#include "mindapi/base/types.h"
namespace mindspore {
namespace ops {
constexpr auto kNameNPUGetFloatStatusV2 = "NPUGetFloatStatusV2";
class MIND_API NPUGetFloatStatusV2 : public BaseOperator {
public:
MIND_API_BASE_MEMBER(NPUGetFloatStatusV2);
NPUGetFloatStatusV2() : BaseOperator(kNameNPUGetFloatStatusV2) { InitIOName({"addr"}, {"data"}); }
void Init() const {}
};
MIND_API abstract::AbstractBasePtr NPUGetFloatStatusV2Infer(const abstract::AnalysisEnginePtr &,
const PrimitivePtr &primitive,
const std::vector<abstract::AbstractBasePtr> &input_args);
} // namespace ops
} // namespace mindspore
#endif // MINDSPORE_CORE_OPS_NPU_GET_FLOAT_STATUS_V2_H_

2
mindspore/python/mindspore/_extends/parallel_compile/tbe_compiler/tbe_helper.py
View File

@ -298,7 +298,7 @@ def get_options_info(job_content):
options["op_impl_mode_list"] = job_content["SocInfo"]["op_impl_mode_list"]
options["kernel_meta_temp_dir"] = job_content["SocInfo"]["kernel_meta_temp_dir"]
options["deterministic"] = job_content["SocInfo"]["deterministic"]
options["status_check"] = "false"
options["status_check"] = job_content["SocInfo"]["status_check"]
return options

48
mindspore/python/mindspore/amp.py
View File

@ -1,4 +1,4 @@
# Copyright 2020 Huawei Technologies Co., Ltd
# Copyright 2023 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.
@ -16,7 +16,8 @@
from __future__ import absolute_import
from abc import ABC, abstractmethod
from mindspore.ops._primitive_cache import _get_cache_prim
from mindspore.ops.operations.math_ops import NPUGetFloatStatusV2, NPUClearFloatStatusV2
from ._checkparam import Validator as validator
from .common import dtype as mstype
from . import context
@ -58,34 +59,7 @@ def _overflow(inputs):
return 1 - status.all()
def init_status():
r"""
Returns a Tensor indicating initialized status for overflow detection.
Note:
Only Ascend need status to capture overflow status, you can also call
this function on GPU or CPU, but the return value is useless.
Returns:
Tensor, has the shape of `(8,)`.
Supported Platforms:
``Ascend`` ``GPU`` ``CPU``
Examples:
>>> status = amp.init_status()
"""
if _ascend_target():
status = ops.NPUAllocFloatStatus()()
clear_status = ops.NPUClearFloatStatus()(status)
status = ops.depend(status, clear_status)
else:
status = Tensor([0, 0, 0, 0, 0, 0, 0, 0], mstype.float32)
return status
def all_finite(inputs, status=None):
def all_finite(inputs):
r"""
Returns a scalar Tensor indicating whether the inputs are finite.
@ -98,8 +72,6 @@ def all_finite(inputs, status=None):
Args:
inputs (Union(tuple(Tensor), list(Tensor))): a iterable Tensor.
status (Tensor): the status Tensor for overflow detection, only required on
Ascend. Default: None.
Returns:
Tensor, a scalar Tensor and the dtype is bool.
@ -112,13 +84,13 @@ def all_finite(inputs, status=None):
>>> output = amp.all_finite(x)
"""
if _ascend_target():
if status is None:
raise ValueError("The status must be initialized on Ascend, but get 'None'.")
status = Tensor([0] * 8, mstype.int32)
status = ops.depend(status, inputs)
get_status = ops.NPUGetFloatStatus()(status)
get_status = _get_cache_prim(NPUGetFloatStatusV2)()(status)
status = ops.depend(status, get_status)
status_finite = status.sum() == 0
_ = ops.NPUClearFloatStatus()(status)
clear_status = _get_cache_prim(NPUClearFloatStatusV2)()(status)
get_status = ops.depend(get_status, clear_status)
status_finite = get_status.equal(Tensor(0, mstype.int32)).all()
return status_finite
outputs = _hypermap(_partial(_overflow), inputs)
flag_sum = ops.addn(outputs).reshape(())
@ -329,5 +301,5 @@ class DynamicLossScaler(LossScaler):
__all__ = [
"DynamicLossScaleManager", "LossScaleManager", "FixedLossScaleManager",
"build_train_network", "DynamicLossScaler", "StaticLossScaler", "LossScaler",
"auto_mixed_precision", "init_status", "all_finite"
"auto_mixed_precision", "all_finite"
]

53
mindspore/python/mindspore/boost/boost_cell_wrapper.py
View File

@ -27,6 +27,7 @@ from mindspore.common import Tensor
from mindspore.common.sparse_tensor import RowTensorInner
from mindspore.common.parameter import Parameter, ParameterTuple
from mindspore.nn.wrap.grad_reducer import DistributedGradReducer
from mindspore.ops.operations.math_ops import NPUGetFloatStatusV2, NPUClearFloatStatusV2
from mindspore.ops import functional as F
from mindspore.ops import composite as C
from mindspore.ops import operations as P
@ -460,6 +461,9 @@ class BoostTrainOneStepWithLossScaleCell(BoostTrainOneStepCell):
self.is_distributed = (self.parallel_mode != ParallelMode.STAND_ALONE)
self.gpu_target = (context.get_context("device_target") == "GPU")
self.loss_scaling_manager = None
self.base0 = Tensor(0, mstype.int32)
self.reduce_all = P.ReduceAll(keep_dims=False)
self.equal = P.Equal()
if self.auto_boost.boost_config.get("loss_scale_group", False):
self.enable_enhanced_amp = True
@ -535,12 +539,13 @@ class BoostTrainOneStepWithLossScaleCell(BoostTrainOneStepCell):
bool, overflow value.
float, update ratio.
"""
flag_sum = self.reduce_sum(param, (0,))
flag_sum = self.equal(self.base0, param)
if self.reducer_flag:
flag_reduce = self.allreduce(flag_sum)
overflow = self.less_equal(self.base, flag_reduce)
overflow = not self.reduce_all(flag_reduce)
else:
overflow = self.less_equal(self.base, flag_sum)
overflow = not self.reduce_all(flag_sum)
if overflow:
update_ratio = self.reduce_ratio
else:
@ -609,13 +614,11 @@ class BoostTrainOneStepWithLossScaleCell(BoostTrainOneStepCell):
The second value is the same as the input of `compute_input`, but contains some information about the
execution order.
"""
status = False
status = Tensor([0] * 8, mstype.int32)
if not self.gpu_target:
# init overflow buffer
status = P.NPUAllocFloatStatus()()
status = F.depend(status, pre_cond)
# clear overflow buffer
clear_status = P.NPUClearFloatStatus()(status)
clear_status = NPUClearFloatStatusV2()(status)
compute_input = F.depend(compute_input, clear_status)
return status, compute_input
@ -636,22 +639,36 @@ class BoostTrainOneStepWithLossScaleCell(BoostTrainOneStepCell):
"""
if not self.gpu_target:
status = F.depend(status, compute_output)
get_status = P.NPUGetFloatStatus()(status)
status = F.depend(status, get_status)
# sum overflow buffer elements, 0:not overflow , >0:overflow
flag_sum = self.reduce_sum(status, (0,))
get_status = NPUGetFloatStatusV2()(status)
if self.is_distributed:
# sum overflow flag over devices
flag_reduce = self.allreduce(get_status)
# get_status not equal to [0]*8 means overflow
flag = self.equal(self.base0, flag_reduce)
status = F.depend(status, flag)
clear_status = NPUClearFloatStatusV2()(status)
flag = F.depend(flag, clear_status)
overall_finite = self.reduce_all(flag)
else:
status = F.depend(status, get_status)
clear_status = NPUClearFloatStatusV2()(status)
get_status = F.depend(get_status, clear_status)
flag = self.equal(self.base0, get_status)
overall_finite = self.reduce_all(flag)
overflow = not overall_finite
else:
flag_sum = self.hyper_map(F.partial(_grad_overflow), compute_output)
flag_sum = P.AddN()(flag_sum)
# convert flag_sum to scalar
flag_sum = P.Reshape()(flag_sum, (()))
if self.is_distributed:
# sum overflow flag over devices
flag_reduce = self.allreduce(flag_sum)
overflow = self.less_equal(self.base, flag_reduce)
else:
overflow = self.less_equal(self.base, flag_sum)
if self.is_distributed:
# sum overflow flag over devices
flag_reduce = self.allreduce(flag_sum)
overflow = self.less_equal(self.base, flag_reduce)
else:
overflow = self.less_equal(self.base, flag_sum)
return overflow
def _process_loss_scale(self, overflow):
@ -688,7 +705,7 @@ class BoostTrainOneStepWithLossScaleCell(BoostTrainOneStepCell):
self.optimizer_loss_scale = [self.parent.count(x) for x in parent_set]
self.reduce_ratio = Tensor(1.0 / (2 ** 0.5), mstype.float32)
self.growth_ratio = Tensor(2 ** (1.0 / 1000.0), mstype.float32)
self.overflow_status_list = ParameterTuple(Parameter(Tensor(np.zeros(shape=[8]), mstype.float32),
self.overflow_status_list = ParameterTuple(Parameter(Tensor(np.zeros(shape=[8]), mstype.int32),
name='mix_layer_status_{}'.format(x), requires_grad=False)
for x in range(loss_scale_number))
self.loss_scaling_manager.set_loss_scale_status(loss_scale_number, self.loss_scaling_manager.get_loss_scale())

44
mindspore/python/mindspore/nn/wrap/loss_scale.py
View File

@ -23,6 +23,7 @@ from mindspore.nn.cell import Cell
from mindspore.common import Tensor
from mindspore.common.sparse_tensor import RowTensorInner
from mindspore.common.parameter import Parameter
from mindspore.ops.operations.math_ops import NPUGetFloatStatusV2, NPUClearFloatStatusV2
from mindspore.ops import functional as F
from mindspore.ops import composite as C
from mindspore.ops import operations as P
@ -309,8 +310,11 @@ class TrainOneStepWithLossScaleCell(TrainOneStepCell):
super(TrainOneStepWithLossScaleCell, self).__init__(network, optimizer, sens=None)
self.hyper_map = C.HyperMap()
self.base = Tensor(1, mstype.float32)
self.base0 = Tensor(0, mstype.int32)
self.reduce_sum = P.ReduceSum(keep_dims=False)
self.reduce_all = P.ReduceAll(keep_dims=False)
self.less_equal = P.LessEqual()
self.equal = P.Equal()
self.allreduce = P.AllReduce()
self.is_distributed = (self.parallel_mode != ParallelMode.STAND_ALONE)
self.gpu_target = (context.get_context("device_target") == "GPU")
@ -390,13 +394,11 @@ class TrainOneStepWithLossScaleCell(TrainOneStepCell):
The second value is the same as the input of `compute_input`, but contains some information about the
execution order.
"""
status = False
status = Tensor([0] * 8, mstype.int32)
if not self.gpu_target:
# init overflow buffer
status = P.NPUAllocFloatStatus()()
status = F.depend(status, pre_cond)
# clear overflow buffer
clear_status = P.NPUClearFloatStatus()(status)
clear_status = NPUClearFloatStatusV2()(status)
compute_input = F.depend(compute_input, clear_status)
return status, compute_input
@ -419,22 +421,36 @@ class TrainOneStepWithLossScaleCell(TrainOneStepCell):
"""
if not self.gpu_target:
status = F.depend(status, compute_output)
get_status = P.NPUGetFloatStatus()(status)
status = F.depend(status, get_status)
# sum overflow buffer elements, 0:not overflow , >0:overflow
flag_sum = self.reduce_sum(status, (0,))
get_status = NPUGetFloatStatusV2()(status)
if self.is_distributed:
# sum overflow flag over devices
flag_reduce = self.allreduce(get_status)
# get_status not equal to [0]*8 means overflow
flag = self.equal(self.base0, flag_reduce)
status = F.depend(status, flag)
clear_status = NPUClearFloatStatusV2()(status)
flag = F.depend(flag, clear_status)
overall_finite = self.reduce_all(flag)
else:
status = F.depend(status, get_status)
clear_status = NPUClearFloatStatusV2()(status)
get_status = F.depend(get_status, clear_status)
flag = self.equal(self.base0, get_status)
overall_finite = self.reduce_all(flag)
overflow = not overall_finite
else:
flag_sum = self.hyper_map(F.partial(_grad_overflow), compute_output)
flag_sum = P.AddN()(flag_sum)
# convert flag_sum to scalar
flag_sum = P.Reshape()(flag_sum, (()))
if self.is_distributed:
# sum overflow flag over devices
flag_reduce = self.allreduce(flag_sum)
overflow = self.less_equal(self.base, flag_reduce)
else:
overflow = self.less_equal(self.base, flag_sum)
if self.is_distributed:
# sum overflow flag over devices
flag_reduce = self.allreduce(flag_sum)
overflow = self.less_equal(self.base, flag_reduce)
else:
overflow = self.less_equal(self.base, flag_sum)
return overflow
def process_loss_scale(self, overflow):

2
mindspore/python/mindspore/ops/_op_impl/tbe/__init__.py
View File

@ -37,3 +37,5 @@ from .scatter_nd_d import _scatter_nd_d_tbe # in python no check supported
from .assign_add_ds import _assign_add_ds_tbe # "Frac_nz in pangu not support"
from .atomic_addr_clean import _atomic_addr_clean_tbe # need to clean addr larger than 2G, int32 is not enough
from .assign import _assign_tbe # Different formats of assign inputs cause memory to increase
from .npu_clear_float_status_v2 import _npu_clear_float_status_v2_tbe # io mismatch
from .npu_get_float_status_v2 import _npu_get_float_status_v2_tbe # io mismatch

35
mindspore/python/mindspore/ops/_op_impl/tbe/npu_clear_float_status_v2.py Normal file
View File

@ -0,0 +1,35 @@
# Copyright 2023 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.
# ============================================================================
"""NPUClearFloatStatusV2 op"""
from mindspore.ops.op_info_register import op_info_register, TBERegOp, DataType
npu_clear_float_status_v2_op_info = TBERegOp("NPUClearFloatStatusV2") \
.fusion_type("OPAQUE") \
.async_flag(False) \
.binfile_name("n_p_u_clear_float_status_v2.so") \
.compute_cost(10) \
.kernel_name("n_p_u_clear_float_status_v2") \
.partial_flag(True) \
.input(0, "addr", False, "required", "all") \
.output(0, "data", False, "required", "all") \
.dtype_format(DataType.I32_Default, DataType.I32_Default) \
.get_op_info()
@op_info_register(npu_clear_float_status_v2_op_info)
def _npu_clear_float_status_v2_tbe():
"""NPUClearFloatStatusV2 TBE register"""
return

35
mindspore/python/mindspore/ops/_op_impl/tbe/npu_get_float_status_v2.py Normal file
View File

@ -0,0 +1,35 @@
# Copyright 2023 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.
# ============================================================================
"""NPUGetFloatStatusV2 op"""
from mindspore.ops.op_info_register import op_info_register, TBERegOp, DataType
npu_get_float_status_v2_op_info = TBERegOp("NPUGetFloatStatusV2") \
.fusion_type("ELEMWISE") \
.async_flag(False) \
.binfile_name("n_p_u_get_float_status_v2.so") \
.compute_cost(10) \
.kernel_name("n_p_u_get_float_status_v2") \
.partial_flag(True) \
.input(0, "addr", False, "required", "all") \
.output(0, "data", False, "required", "all") \
.dtype_format(DataType.I32_Default, DataType.I32_Default) \
.get_op_info()
@op_info_register(npu_get_float_status_v2_op_info)
def _npu_get_float_status_v2_tbe():
"""NPUGetFloatStatusV2 TBE register"""
return

170
mindspore/python/mindspore/ops/operations/math_ops.py
View File

@ -20,6 +20,7 @@ from __future__ import division
import numpy as np
from mindspore import context
from mindspore import log as logger
from mindspore.ops import signature as sig
from mindspore._checkparam import Validator as validator
from mindspore._checkparam import Rel
@ -4339,6 +4340,7 @@ class NPUAllocFloatStatus(Primitive):
@prim_attr_register
def __init__(self):
"""Initialize NPUAllocFloatStatus"""
logger.warning("The 'NPUAllocFloatStatus' operator will be deprecated in the future. Please don't use it.")
class NPUGetFloatStatus(Primitive):
@ -4408,6 +4410,7 @@ class NPUGetFloatStatus(Primitive):
@prim_attr_register
def __init__(self):
"""Initialize NPUGetFloatStatus"""
logger.warning("The 'NPUGetFloatStatus' operator will be deprecated in the future. Please don't use it.")
class NPUClearFloatStatus(Primitive):
@ -4471,6 +4474,173 @@ class NPUClearFloatStatus(Primitive):
@prim_attr_register
def __init__(self):
"""Initialize NPUClearFloatStatus"""
logger.warning("The 'NPUClearFloatStatus' operator will be deprecated in the future. Please don't use it.")
class NPUGetFloatStatusV2(Primitive):
"""
Get the flag for storage overflow status. This flag is located in a register at a
fixed address on the `Ascend` device, and overflow information is automatically
written to this register.
The flag is a one-dimensional Tensor with shape :math:`(8,)` and data type `mindspore.dtype.int32`.
If the value of flag is zero, no overflow has occurred, otherwise, overflow.
When performing overflow detection on the network, you should first call `NPUClearFloatStatusV2` to
reset the register before the detection, and then call `NPUGetFloatStatusV2` to get the register
status after the network execution is completed.
Note:
- In order to avoid mis-optimization by the compiler, additional input is added to
this operator. The input is defined as a shape of: math:`(8,)` and data type of
`mindspore.dtype.int32` Tensor, meaningless.
- Since this op lacks contextual dependencies with parameters in the network,
:class:`mindspore.ops.Depend` needs to be used to ensure order of execution.
Inputs:
Tensor, an additional input created to avoid compiler optimization, is specified as shape :math:`(8,)`,
data type is `mindspore.dtype.int32`, and has no actual meaning.
Usually use the output of `NPUClearFloatStatusV2`.
Outputs:
Tensor, shape and data type are the same as input. If all are zero, it means no overflow, otherwise, overflow.
Raises:
TypeError: If `x` is not a Tensor.
TypeError: If dtype of `x` is not int32.
ValueError: If shape of `x` is not equal to :math:`(8,)`.
Supported Platforms:
``Ascend``
Examples:
>>> import mindspore as ms
>>> import numpy as np
>>> from mindspore import ops, nn, Tensor
>>> from mindspore.ops.operations.math_ops import NPUGetFloatStatusV2, NPUClearFloatStatusV2
>>> class Net(nn.Cell):
... def __init__(self):
... super().__init__()
... self.clear_status = NPUClearFloatStatusV2()
... self.get_status = NPUGetFloatStatusV2()
... self.sub = ops.Sub()
... self.neg = ops.Neg()
... self.equal = ops.Equal()
... self.reduce_all = ops.ReduceAll(keep_dims=False)
... self.base = Tensor([0], dtype=ms.int32)
...
... def construct(self, x):
... init = Tensor([0]*8, dtype=ms.int32)
... clear_status = self.clear_status(init)
... x = ops.depend(x, clear_status)
... res = self.sub(x, self.neg(x))
... init = ops.depend(init, res)
... get_status = self.get_status(init)
... flag = self.equal(self.base, get_status)
... overall_finite = self.reduce_all(flag)
... overflow = not overall_finite
... return overflow
...
>>> value = 65504
>>> data = np.full((2, 3), value, dtype=np.float16)
>>> x = Tensor(data, dtype=ms.float16)
>>> net = Net()
>>> res = net(x)
>>> print(res)
True
>>> value = 10
>>> data = np.full((2, 3), value, dtype=np.float16)
>>> x = Tensor(data, dtype=ms.float16)
>>> net = Net()
>>> res = net(x)
>>> print(res)
False
"""
@prim_attr_register
def __init__(self):
"""Initialize NPUGetFloatStatusV2"""
class NPUClearFloatStatusV2(Primitive):
"""
Clear the flag for storage overflow status. This flag is located in a register at a
fixed address on the `Ascend` device, and overflow information is automatically
written to this register.
The flag is a one-dimensional Tensor with shape :math:`(8,)` and data type `mindspore.dtype.int32`.
If the value of flag is zero, no overflow has occurred, otherwise, overflow.
When performing overflow detection on the network, you should first call `NPUClearFloatStatusV2` to
reset the register before the detection, and then call `NPUGetFloatStatusV2` to get the register
status after the network execution is completed.
Note:
- In order to avoid mis-optimization by the compiler, additional input and output are added to
this operator. The input and output are defined as a shape of: math:`(8,)` and data type of
`mindspore.dtype.int32` Tensor, meaningless.
- Since this op lacks contextual dependencies with parameters in the network,
:class:`mindspore.ops.Depend` needs to be used to ensure order of execution.
Inputs:
Tensor, an additional input created to avoid compiler optimization, is specified as shape :math:`(8,)`,
data type is `mindspore.dtype.int32`, and has no actual meaning.
Outputs:
Tensor, shape and data type are the same as input, meaningless.
Raises:
TypeError: If `x` is not a Tensor.
TypeError: If dtype of `x` is not int32.
ValueError: If shape of `x` is not equal to :math:`(8,)`.
Supported Platforms:
``Ascend``
Examples:
>>> import mindspore as ms
>>> import numpy as np
>>> from mindspore import ops, nn, Tensor
>>> from mindspore.ops.operations.math_ops import NPUGetFloatStatusV2, NPUClearFloatStatusV2
>>> class Net(nn.Cell):
... def __init__(self):
... super().__init__()
... self.clear_status = NPUClearFloatStatusV2()
... self.get_status = NPUGetFloatStatusV2()
... self.sub = ops.Sub()
... self.neg = ops.Neg()
... self.equal = ops.Equal()
... self.reduce_all = ops.ReduceAll(keep_dims=False)
... self.base = Tensor([0], dtype=ms.int32)
...
... def construct(self, x):
... init = Tensor([0]*8, dtype=ms.int32)
... clear_status = self.clear_status(init)
... x = ops.depend(x, clear_status)
... res = self.sub(x, self.neg(x))
... init = ops.depend(init, res)
... get_status = self.get_status(init)
... flag = self.equal(self.base, get_status)
... overall_finite = self.reduce_all(flag)
... overflow = not overall_finite
... return overflow
...
>>> value = 65504
>>> data = np.full((2, 3), value, dtype=np.float16)
>>> x = Tensor(data, dtype=ms.float16)
>>> net = Net()
>>> res = net(x)
>>> print(res)
True
>>> value = 10
>>> data = np.full((2, 3), value, dtype=np.float16)
>>> x = Tensor(data, dtype=ms.float16)
>>> net = Net()
>>> res = net(x)
>>> print(res)
False
"""
@prim_attr_register
def __init__(self):
"""Initialize NPUClearFloatStatusV2"""
class Cos(Primitive):

7
tests/st/mix_precision/test_mix_precision_func.py
View File

@ -15,6 +15,7 @@
import numpy as np
import pytest
import mindspore
from mindspore import Tensor, Parameter
from mindspore.common import dtype as mstype
@ -60,22 +61,20 @@ def test_dynamic_loss_scaler(mode):
Expectation: the `scale_value` can be adjusted correctly.
"""
context.set_context(mode=mode)
status = amp.init_status()
loss_scaler = amp.DynamicLossScaler(scale_value=2**10, scale_factor=2, scale_window=50)
grads = (Tensor(np.array([0.5, 1.0]), mindspore.float16),
Tensor(np.array([0.2]), mindspore.float16))
unscaled_grads = loss_scaler.unscale(grads)
grads_finite = amp.all_finite(unscaled_grads, status)
grads_finite = amp.all_finite(unscaled_grads)
loss_scaler.counter = Parameter(Tensor(49, dtype=mstype.int32))
loss_scaler.adjust(grads_finite)
assert loss_scaler.scale_value.asnumpy() == np.array(2048.)
status = amp.init_status()
grads = (Tensor(np.array([2., 1.0]), mindspore.float16),
Tensor(np.array([0.2]), mindspore.float16))
unscaled_grads = loss_scaler.unscale(grads)
grads_finite = amp.all_finite(unscaled_grads, status)
grads_finite = amp.all_finite(unscaled_grads)
loss_scaler.scale_value = Parameter(Tensor(2**10, dtype=mstype.float32))
loss_scaler.adjust(grads_finite)
assert loss_scaler.scale_value.asnumpy() == np.array(1024.)

175
tests/st/ops/ascend/test_npu_overflow_v2.py Normal file
View File

@ -0,0 +1,175 @@
# Copyright 2023 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.
# ============================================================================
import pytest
import numpy as np
import mindspore as ms
from mindspore import Tensor, nn, ops
from mindspore import dtype as mstype
from mindspore.ops._primitive_cache import _get_cache_prim
from mindspore.ops.operations.math_ops import NPUGetFloatStatusV2, NPUClearFloatStatusV2
class OverflowCheckNet(nn.Cell):
def __init__(self):
super(OverflowCheckNet, self).__init__()
self.base1 = Tensor(1, mstype.float32)
self.base2 = Tensor(0, mstype.int32)
self.reduce_sum = ops.ReduceSum(keep_dims=False)
self.less_equal = ops.LessEqual()
self.reduce_all = ops.ReduceAll(keep_dims=False)
self.equal = ops.Equal()
def start_overflow_check_v1(self, pre_cond, compute_input):
status = False
# init overflow buffer
status = ops.NPUAllocFloatStatus()()
status = ops.depend(status, pre_cond)
# clear overflow buffer
clear_status = ops.NPUClearFloatStatus()(status)
compute_input = ops.depend(compute_input, clear_status)
return status, compute_input
def get_overflow_status_v1(self, status, compute_output):
status = ops.depend(status, compute_output)
get_status = ops.NPUGetFloatStatus()(status)
status = ops.depend(status, get_status)
# sum overflow buffer elements, 0:not overflow , >0:overflow
flag_sum = self.reduce_sum(status, (0,))
overflow = self.less_equal(self.base1, flag_sum)
return overflow
def start_overflow_check_v2(self, pre_cond, compute_input):
status = Tensor([0] * 8, mstype.int32)
status = ops.depend(status, pre_cond)
# clear overflow buffer
clear_status = _get_cache_prim(NPUClearFloatStatusV2)()(status)
compute_input = ops.depend(compute_input, clear_status)
return status, compute_input
def get_overflow_status_v2(self, status, compute_output):
status = ops.depend(status, compute_output)
get_status = _get_cache_prim(NPUGetFloatStatusV2)()(status)
status = ops.depend(status, get_status)
clear_status = _get_cache_prim(NPUClearFloatStatusV2)()(status)
get_status = ops.depend(get_status, clear_status)
flag = self.equal(self.base2, get_status)
overall_finite = self.reduce_all(flag)
return not overall_finite
class OverFlowNetV2GetStatusAfterClear(OverflowCheckNet):
def __init__(self):
super(OverFlowNetV2GetStatusAfterClear, self).__init__()
self.mul = ops.Mul()
self.sub = ops.Sub()
def construct(self, x1, x2):
y1 = self.mul(x1, x1)
status, compute_input = self.start_overflow_check_v2(y1, x2)
y2 = self.sub(y1, compute_input)
cond = self.get_overflow_status_v2(status, y2)
return cond
class OverFlowNetV2GetStatus(OverflowCheckNet):
def __init__(self):
super(OverFlowNetV2GetStatus, self).__init__()
self.add = ops.Add()
self.mul = ops.Mul()
def construct(self, x1, x2):
y1 = self.add(x1, x1)
status, compute_input = self.start_overflow_check_v2(y1, x2)
y2 = self.mul(y1, compute_input)
cond = self.get_overflow_status_v2(status, y2)
return cond
class OverflowCheckV1vsV2(OverflowCheckNet):
def __init__(self):
super(OverflowCheckV1vsV2, self).__init__()
self.add = ops.Add()
self.atan2 = ops.Atan2()
def construct(self, x1, x2, version):
y1 = self.add(x1, x1)
if version == 1:
status, compute_input = self.start_overflow_check_v1(y1, x2)
y2 = self.atan2(y1, compute_input)
cond = self.get_overflow_status_v1(status, y2)
else:
status, compute_input = self.start_overflow_check_v2(y1, x2)
y2 = self.atan2(y1, compute_input)
cond = self.get_overflow_status_v2(status, y2)
return cond
@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
def test_v2_overflow_get_after_clear(mode):
"""
Feature: overflow check v2
Description: Verify the result of get_status after clear
Expectation: success
"""
ms.set_context(mode=mode)
net = OverFlowNetV2GetStatusAfterClear()
output = net(Tensor(65504, mstype.float16), Tensor(1, mstype.float16))
assert not output
@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
def test_v2_clear_overflow_get(mode):
"""
Feature: overflow check v2
Description: Verify the result of get_status when overflow
Expectation: success
"""
ms.set_context(mode=mode)
net = OverFlowNetV2GetStatus()
output = net(Tensor(1, mstype.float16), Tensor(65504, mstype.float16))
assert output
@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.env_onecard
@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
def test_v1_vs_v2_overflow_check(mode):
"""
Feature: overflow check v1 vs v2
Description: Verify the result of atan2 when inputs include 0
Expectation: success
"""
ms.set_context(mode=mode)
input1 = np.random.random((2, 4)).astype(np.float32)
input2 = np.random.random((2, 4)).astype(np.float32)
input1[0] = 0
input2[1] = 0
net = OverflowCheckV1vsV2()
overflow_v1 = net(Tensor(input1), Tensor(input2), 1)
overflow_v2 = net(Tensor(input1), Tensor(input2), 2)
assert overflow_v1
assert not overflow_v2

83
tests/st/train/test_amp_overflow.py Normal file
View File

@ -0,0 +1,83 @@
# Copyright 2023 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.
# ============================================================================
'''test overflow'''
import pytest
import numpy as np
from mindspore import Tensor, Parameter, nn, ops
import mindspore.amp as amp
import mindspore as ms
class Net(nn.Cell):
def __init__(self, in_features, out_features):
super(Net, self).__init__()
self.weight = Parameter(Tensor(np.full([in_features, out_features], 2, np.float16)),
name='weight')
self.matmul = ops.MatMul()
def construct(self, x):
output = self.matmul(x, self.weight)
return output
@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.platform_x86_gpu_training
@pytest.mark.platform_x86_cpu_training
@pytest.mark.env_onecard
@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
def test_functional_amp_overflow(mode):
"""
Feature: mindspore.amp.overflow
Description: test amp overflow
Expectation: Success.
"""
ms.set_context(mode=mode)
size, in_features, out_features = 1, 2, 2
net = Net(in_features, out_features)
loss_fn = nn.MSELoss()
def forward_fn(data, label):
logits = net(data)
loss = loss_fn(logits, label)
return loss, logits
grad_fn = ops.value_and_grad(forward_fn, grad_position=None, weights=net.trainable_params())
@ms.jit
def train_step(data, label):
(loss, _), grads = grad_fn(data, label)
is_finite = amp.all_finite(grads)
return loss, is_finite
shape = (size, in_features)
inputs = [
Tensor(np.full(shape, -np.inf, np.float16)),
Tensor(np.full(shape, 0, np.float16)),
Tensor(np.full(shape, 40000, np.float16)),
Tensor(np.full(shape, 10, np.float16)),
Tensor(np.full(shape, np.inf, np.float16)),
]
label = Tensor(np.full([out_features,], 0, np.float16))
datasets = list(zip(inputs, [label for _ in range(len(inputs))]))
expect_results = [False, True, False, True, False]
outputs = []
for data, label in datasets:
_, is_finite = train_step(data, label)
outputs.append(is_finite.asnumpy().tolist())
assert outputs == expect_results

114
tests/st/train/test_loss_scale_overflow.py Normal file
View File

@ -0,0 +1,114 @@
# Copyright 2023 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.
# ============================================================================
'''test overflow'''
import pytest
import numpy as np
import mindspore as ms
from mindspore import Tensor, Parameter, nn, ops, boost
from mindspore import dtype as mstype
class Net(nn.Cell):
def __init__(self, in_features, out_features):
super(Net, self).__init__()
self.weight = Parameter(Tensor(np.full([in_features, out_features], 2, np.float16)),
name='weight')
self.matmul = ops.MatMul()
def construct(self, x):
output = self.matmul(x, self.weight)
return output
@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
@pytest.mark.parametrize('mode', [ms.GRAPH_MODE, ms.PYNATIVE_MODE])
def test_train_one_step_with_loss_scale_cell_overflow(mode):
"""
Feature: mindspore.TrainOneStepWithLossScaleCell.overflow
Description: test TrainOneStepWithLossScaleCell overflow
Expectation: Success.
"""
ms.set_context(mode=mode)
size, in_features, out_features = 1, 2, 2
net = Net(in_features, out_features)
loss = nn.MSELoss()
optimizer = nn.Momentum(net.trainable_params(),
learning_rate=0.1, momentum=0.9)
net_with_loss = nn.WithLossCell(net, loss)
shape = (size, in_features)
inputs = [
Tensor(np.full(shape, -np.inf, np.float16)),
Tensor(np.full(shape, 0, np.float16)),
Tensor(np.full(shape, 40000, np.float16)),
Tensor(np.full(shape, 10, np.float16)),
Tensor(np.full(shape, np.inf, np.float16)),
]
label = Tensor(np.full([out_features,], 0, np.float16))
datasets = list(zip(inputs, [label for _ in range(len(inputs))]))
scaling_sens = Tensor([8], dtype=mstype.float16)
train_network = nn.TrainOneStepWithLossScaleCell(
net_with_loss, optimizer, scale_sense=scaling_sens)
expect_results = [True, False, True, False, True]
outputs = []
for x, label in datasets:
_, overflow, _ = train_network(x, label)
outputs.append(overflow.asnumpy().tolist())
assert outputs == expect_results
@pytest.mark.level0
@pytest.mark.platform_arm_ascend_training
@pytest.mark.platform_x86_ascend_training
@pytest.mark.platform_x86_gpu_training
@pytest.mark.env_onecard
@pytest.mark.parametrize('mode', [ms.PYNATIVE_MODE])
def test_boost_train_one_step_with_loss_scale_cell_overflow(mode):
"""
Feature: mindspore.BoostTrainOneStepWithLossScaleCell.overflow
Description: test BoostTrainOneStepWithLossScaleCell overflow
Expectation: Success.
"""
ms.set_context(mode=mode)
size, in_features, out_features = 1, 2, 2
net = Net(in_features, out_features)
loss = nn.MSELoss()
optimizer = nn.Momentum(net.trainable_params(),
learning_rate=0.1, momentum=0.9)
net_with_loss = nn.WithLossCell(net, loss)
shape = (size, in_features)
inputs = [
Tensor(np.full(shape, -np.inf, np.float16)),
Tensor(np.full(shape, 0, np.float16)),
Tensor(np.full(shape, 40000, np.float16)),
Tensor(np.full(shape, 10, np.float16)),
Tensor(np.full(shape, np.inf, np.float16)),
]
label = Tensor(np.full([out_features,], 0, np.float16))
datasets = list(zip(inputs, [label for _ in range(len(inputs))]))
scaling_sens = Tensor([8], dtype=mstype.float16)
train_network = boost.BoostTrainOneStepWithLossScaleCell(
net_with_loss, optimizer, scale_sense=scaling_sens)
expect_results = [True, False, True, False, True]
outputs = []
for x, label in datasets:
_, overflow, _ = train_network(x, label)
outputs.append(overflow)
assert outputs == expect_results

12
tests/ut/cpp/tbe/tbe_json_creator_test.cc
View File

@ -76,7 +76,7 @@ TEST_F(TestHWTBEJsonCreator, DISABLED_test_tbe_single_common) {
auto tbe_json_creator_build = std::make_shared<BuildTbeJsonCreator>();
nlohmann::json kernel_json;
EXPECT_TRUE(tbe_json_creator_select->GenJson(relu1, &kernel_json));
EXPECT_EQ(tbe_json_creator_select->GetJsonHash(), 10654173078034037040U)
EXPECT_EQ(tbe_json_creator_select->GetJsonHash(), 12207851473833394607U)
<< "Error json is:" << kernel_json << ", for expected json, see file: tbe_single_common_select.json";
EXPECT_TRUE(tbe_json_creator_build->GenJson(relu1, &kernel_json));
EXPECT_EQ(tbe_json_creator_build->GetJsonHash(), 2389029245513168162U)
@ -118,7 +118,7 @@ TEST_F(TestHWTBEJsonCreator, DISABLED_test_tbe_single_conv2d_backprop_filter) {
auto tbe_json_creator_build = std::make_shared<BuildTbeJsonCreator>();
nlohmann::json kernel_json;
EXPECT_TRUE(tbe_json_creator_select->GenJson(conv2d_backprop_filter, &kernel_json));
EXPECT_EQ(tbe_json_creator_select->GetJsonHash(), 16416634683849134630U)
EXPECT_EQ(tbe_json_creator_select->GetJsonHash(), 14683931476519216146U)
<< "Error json is:" << kernel_json
<< ", for expected json, see file: tbe_single_conv2d_backprop_filter_select.json";
EXPECT_TRUE(tbe_json_creator_build->GenJson(conv2d_backprop_filter, &kernel_json));
@ -177,7 +177,7 @@ TEST_F(TestHWTBEJsonCreator, DISABLED_test_tbe_single_dynamic_rnn) {
auto tbe_json_creator_build = std::make_shared<BuildTbeJsonCreator>();
nlohmann::json kernel_json;
EXPECT_TRUE(tbe_json_creator_select->GenJson(dynamic_rnn, &kernel_json));
EXPECT_EQ(tbe_json_creator_select->GetJsonHash(), 3107761065269367419U)
EXPECT_EQ(tbe_json_creator_select->GetJsonHash(), 16143536111232395651U)
<< "Error json is:" << kernel_json << ", for expected json, see file: tbe_single_dynamic_rnn_select.json";
EXPECT_TRUE(tbe_json_creator_build->GenJson(dynamic_rnn, &kernel_json));
EXPECT_EQ(tbe_json_creator_build->GetJsonHash(), 14916511955212123861U)
@ -230,7 +230,7 @@ TEST_F(TestHWTBEJsonCreator, DISABLED_test_tbe_single_layer_norm) {
auto tbe_json_creator_build = std::make_shared<BuildTbeJsonCreator>();
nlohmann::json kernel_json;
EXPECT_TRUE(tbe_json_creator_select->GenJson(layer_norm, &kernel_json));
EXPECT_EQ(tbe_json_creator_select->GetJsonHash(), 6592146268336877821U)
EXPECT_EQ(tbe_json_creator_select->GetJsonHash(), 1161191001728520611U)
<< "Error json is:" << kernel_json << ", for expected json, see file: tbe_single_layer_norm_select.json";
EXPECT_TRUE(tbe_json_creator_build->GenJson(layer_norm, &kernel_json));
EXPECT_EQ(tbe_json_creator_build->GetJsonHash(), 2848618249728529296U)
@ -306,7 +306,7 @@ TEST_F(TestHWTBEJsonCreator, test_tbe_fusion_common) {
nlohmann::json fusion_json;
auto tbe_json_creator = std::make_shared<FusionBuildTbeJsonCreator>();
EXPECT_TRUE(tbe_json_creator->GenJson(fusion_scope_info, &fusion_json));
EXPECT_EQ(tbe_json_creator->GetJsonHash(), 9482071119130243510U)
EXPECT_EQ(tbe_json_creator->GetJsonHash(), 18379117451241093022U)
<< "Error json is:" << fusion_json << ", for expected json, see file: tbe_fusion_common.json";
}
@ -367,7 +367,7 @@ TEST_F(TestHWTBEJsonCreator, test_fusion_add_conv2d) {
nlohmann::json fusion_json;
auto tbe_json_creator = std::make_shared<FusionBuildTbeJsonCreator>();
EXPECT_TRUE(tbe_json_creator->GenJson(fusion_scope_info, &fusion_json));
EXPECT_EQ(tbe_json_creator->GetJsonHash(), 1515571995667332418U)
EXPECT_EQ(tbe_json_creator->GetJsonHash(), 16132617067967162574U)
<< "Error json is:" << fusion_json << ", for expected json, see file: test_fusion_add_conv2d.json";
}