mindspore 2.2.0__cp38-cp38-manylinux1_x86_64.whl → 2.2.11__cp38-cp38-manylinux1_x86_64.whl

mindspore/profiler/parser/ascend_hccl_generator.py

@@ -82,6 +82,238 @@ def count_average(data):
 class AscendHCCLGenerator:
     """Generate ascend hccl data from files."""
 
+    def __init__(self, source_path, steptrace):
+        self.root_path = source_path
+        self.steptrace = steptrace
+        self.hccl_raw = []
+        self.hccl_data_df = np.dtype(
+            [('model_id', int), ('iteration_id', int), ('name', object), ('pid', int), ('tid', int), ('ts', float),
+             ('te', float), ('dur', float), ('ph', object),
+             ('task_type', object), ('link_info', object), ('transport_type', object), ('size', int), ('tag', object)])
+
+    @staticmethod
+    def _cost_analyse(iteration):
+        """analyse communication cost and wait cost"""
+        communication_cost = np.sum(iteration[iteration['name'] != 'Notify_Wait']['dur'])
+        wait_cost = np.sum(iteration[iteration['name'] == 'Notify_Wait']['dur'])
+        return communication_cost, wait_cost
+
+    @staticmethod
+    def _rdma_analyse(groupby_transport):
+        """rdma analyse"""
+        thread_groups = np.unique(groupby_transport['tid'])
+        thread_information = []
+        for thread_index in thread_groups:
+            groupby_thread = groupby_transport[groupby_transport['tid'] == thread_index]
+            rdma_communication_time = 0
+            rdma_communication_size = 0
+            rdma_communication_wait_time = 0
+            start_index = 0
+            end_index = groupby_thread.size - 2
+            while start_index < end_index:
+                first_task_type = groupby_thread[start_index]['task_type']
+                if first_task_type == 'RDMASend':
+                    second_index = start_index + 1
+                    third_index = start_index + 2
+                    second_task_type = groupby_thread[second_index]['task_type']
+                    third_task_type = groupby_thread[third_index]['task_type']
+                    if second_task_type == 'RDMASend' and third_task_type == 'Notify Wait':
+                        rdma_send_cost = groupby_thread[start_index]['dur']
+                        notify_record_cost = groupby_thread[second_index]['dur']
+                        notify_wait_cost = groupby_thread[third_index]['dur']
+                        rdma_communication_time += rdma_send_cost + notify_record_cost + notify_wait_cost
+                        rdma_communication_wait_time += notify_wait_cost
+                        rdma_communication_size += groupby_thread[start_index]['size'] + groupby_thread[second_index][
+                            'size']
+                        start_index += 2
+                start_index += 1
+            rdma_communication_wait_time = rdma_communication_wait_time / 1e3
+            rdma_communication_size = rdma_communication_size / 1e3
+            rdma_communication_time = rdma_communication_time / 1e3
+            rdma_bandwidth = rdma_communication_size / (rdma_communication_time / 1e3) \
+                if rdma_communication_size else 0
+            thread_information.append(
+                [rdma_communication_time, rdma_communication_size, rdma_bandwidth, rdma_communication_wait_time])
+        if len(thread_information) > 1:
+            thread_information = np.sum(thread_information, axis=0).tolist()
+
+        return thread_information
+
+    def parse(self):
+        """Analyse the original hccl data generator hccl data."""
+        hccl_data = []
+        for hccl_file in find_files(self.root_path, "hccl_*.json"):
+            with open(hccl_file) as fr:
+                hccl_data.append(self._original_data_analyse(json.load(fr)))
+        hccl_data = np.concatenate(hccl_data)
+
+        for model_id in np.unique(hccl_data['model_id']):
+            hccl_data_model = hccl_data[hccl_data['model_id'] == model_id]
+            for iteration_id in np.unique(hccl_data_model['iteration_id']):
+                hccl_data_model_iteration = hccl_data_model[hccl_data_model['iteration_id'] == iteration_id]
+
+                hccl_abstract_data = hccl_data_model_iteration[hccl_data_model_iteration['task_type'] == '']
+                hccl_detail_data = hccl_data_model_iteration[hccl_data_model_iteration['task_type'] != '']
+                hccl_abstract_data = np.sort(hccl_abstract_data, order='ts')
+                hccl_detail_data = np.sort(hccl_detail_data, order='ts')
+
+                tag = np.searchsorted(hccl_abstract_data['ts'], hccl_detail_data['ts'], side='right') - 1
+
+                hccl_detail_data['tag'] = [x[-1] for x in
+                                           np.char.split(hccl_abstract_data[tag]['name'].astype(str), sep='/')]
+
+                self.hccl_raw.append(self._iteration_analyse(hccl_detail_data, iteration_id))
+
+        self.hccl_raw = sorted(self.hccl_raw, key=lambda x: x[0])
+        self.hccl_raw.append(copy.deepcopy(self.hccl_raw[-1]))
+        self.hccl_raw[-1][0] = '-'
+        for _, value in self.hccl_raw[-1][4].items():
+            value[0] = '-'
+
+    def write(self, hccl_raw_path):
+        """
+        Write the flops.csv and flops_summary.json
+
+        Args:
+            hccl_raw_path(str): hccl_raw.csv path.
+        """
+        try:
+            with os.fdopen(os.open(hccl_raw_path,
+                                   os.O_WRONLY | os.O_CREAT | os.O_TRUNC, stat.S_IWUSR | stat.S_IRUSR), 'w',
+                           newline='') as hccl_row:
+                writer = csv.writer(hccl_row)
+                writer.writerow(
+                    ['step_num', 'communication_cost', 'wait_cost', 'link_info', 'communication_operator_cost'])
+                for row in self.hccl_raw:
+                    row[3] = json.dumps(row[3])
+                    row[4] = json.dumps(row[4])
+                writer.writerows(self.hccl_raw)
+        except (IOError, OSError) as err:
+            logging.critical('Errot occurred when write aicore detail file: %s', err)
+            raise ProfilerIOException() from err
+        if os.path.exists(hccl_raw_path):
+            os.chmod(hccl_raw_path, stat.S_IREAD | stat.S_IWRITE)
+
+    def _original_data_analyse(self, original_data):
+        """analyse original data"""
+
+        groups_steptrace = {model_id: np.sort(self.steptrace[self.steptrace['Model ID'] == model_id],
+                                              order='Iteration ID')
+                            for model_id in np.unique(self.steptrace['Model ID'])}
+
+        target_data = []
+        for row in original_data:
+            model_id = row.get('args', {}).get('model id')
+            if row.get('ph') == 'X' and model_id is not None:
+                name = row.get('name')
+                pid = row.get('pid')
+                tid = row.get('tid')
+                ts = row.get('ts')
+                dur = row.get('dur')
+                te = ts + dur
+                ph = row.get('ph')
+                task_type = row.get('args', {}).get('task type', '')
+                src_rank = row.get('args', {}).get('src rank', 0)
+                dst_rank = row.get('args', {}).get('dst rank', 0)
+                if src_rank == int('0xffffffff', 16):
+                    src_rank = dst_rank
+                if dst_rank == int('0xffffffff', 16):
+                    dst_rank = src_rank
+                transport_type = row.get('args', {}).get('transport type', '')
+                if transport_type == 'LOCAL':
+                    src_rank, dst_rank = dst_rank, src_rank
+                link_info = str(src_rank) + '-' + str(dst_rank)
+                size = row.get('args', {}).get('size(Byte)', 0)
+                size = size if isinstance(size, int) else int(size, 16)
+                steptrace = groups_steptrace.get(model_id, None)
+                if steptrace is None:
+                    logging.warning('Could not find model: %s in hccl json, skip.', model_id)
+                    continue
+                tag = np.searchsorted(steptrace['Iteration End'], te * 1e-3, side='left')
+                iteration_id = steptrace[tag]['Iteration ID']
+                target_data.append(
+                    tuple([model_id, iteration_id, name, pid, tid,
+                           ts, te, dur, ph, task_type,
+                           link_info, transport_type, size, -1]))
+
+        hccl_data = np.array(target_data, dtype=self.hccl_data_df)
+
+        return hccl_data
+
+    def _iteration_analyse(self, hccl_detail_data, iteration):
+        """analyse data by iteration """
+        communication_cost, wait_cost = self._cost_analyse(hccl_detail_data)
+        link_info = self._link_info_analyse(hccl_detail_data)
+        communication_operator_cost = self._communication_operator_cost_analyse(hccl_detail_data, iteration)
+        return [iteration, communication_cost, wait_cost, link_info, communication_operator_cost]
+
+    def _link_info_analyse(self, hccl_detail_data):
+        """analyse link info data"""
+        groupby_iteration = hccl_detail_data[hccl_detail_data['task_type'] != 'Notify Record']
+        link_info_groups = np.unique(groupby_iteration['link_info'])
+        link_info_information = dict()
+        for link_info_index in link_info_groups:
+            groupby_link_info = groupby_iteration[groupby_iteration['link_info'] == link_info_index]
+            transport_groups = np.unique(groupby_iteration['transport_type'])
+            transport_information = dict()
+            for transport_index in transport_groups:
+                groupby_transport = groupby_link_info[groupby_link_info['transport_type'] == transport_index]
+                if transport_index == 'SDMA' and groupby_transport.size > 0:
+                    groupby_sdma = \
+                        groupby_transport[np.isin(groupby_transport['task_type'], ['Memcpy', 'Reduce Inline'])][
+                            ['dur', 'size']]
+                    sdma_communication_time = np.sum(groupby_sdma['dur']) * 1e-3
+                    sdma_communication_size = np.sum(groupby_sdma['size']) * 1e-3
+                    sdma_bandwidth = sdma_communication_size / sdma_communication_time * 1e-3 \
+                        if sdma_communication_time != 0 else 0
+                    transport_information['SDMA'] = [sdma_communication_time, sdma_communication_size, sdma_bandwidth]
+                elif transport_index == 'RDMA' and groupby_transport.size > 0:
+                    transport_information['RDMA'] = self._rdma_analyse(groupby_transport)
+            link_info_information[link_info_index] = transport_information
+        return link_info_information
+
+    def _communication_operator_cost_analyse(self, hccl_detail_data, iteration_index):
+        """analyse communication operator cost"""
+        groupby_iteration = hccl_detail_data[hccl_detail_data['task_type'] != 'Notify Record']
+        tag_groups = np.unique(groupby_iteration['tag'])
+        tag_information = dict()
+        for tag_index in tag_groups:
+            groupby_tag = groupby_iteration[groupby_iteration['tag'] == tag_index]
+            link_groups = np.unique(groupby_iteration['link_info'])
+            link_info_information = dict()
+            for link_info_index in link_groups:
+                groupby_link_info = groupby_tag[groupby_tag['link_info'] == link_info_index]
+                transport_groups = np.unique(groupby_link_info['transport_type'])
+                transport_information = dict()
+                for transport_index in transport_groups:
+                    groupby_transport = groupby_link_info[groupby_link_info['transport_type'] == transport_index]
+                    if transport_index == 'SDMA':
+                        groupby_sdma = \
+                            groupby_transport[np.isin(groupby_transport['task_type'], ['Memcpy', 'Reduce Inline'])][
+                                ['dur', 'size']]
+                        sdma_communication_time = np.sum(groupby_sdma['dur']) * 1e-3
+                        sdma_communication_size = np.sum(groupby_sdma['size']) * 1e-3
+                        sdma_bandwidth = sdma_communication_size / sdma_communication_time * 1e-3 \
+                            if sdma_communication_time != 0 else 0
+                        transport_information['SDMA'] = [
+                            sdma_communication_time, sdma_communication_size,
+                            sdma_bandwidth
+                        ]
+                    elif transport_index == 'RDMA':
+                        transport_information['RDMA'] = self._rdma_analyse(groupby_transport)
+                    link_info_information[link_info_index] = transport_information
+                communication_cost = np.sum(groupby_tag[groupby_tag['name'] != 'Notify_Wait']['dur'])
+                wait_cost = np.sum(groupby_tag[groupby_tag['name'] == 'Notify_Wait']['dur'])
+                tag_information[tag_index] = [
+                    str(iteration_index), communication_cost, wait_cost,
+                    link_info_information
+                ]
+        return tag_information
+
+
+class AscendHCCLGeneratorOld:
+    """Generate ascend hccl data from files."""
+
     def __init__(self, source_path):
         self.root_path = source_path
         self.hccl_raw = []

mindspore/profiler/parser/ascend_msprof_exporter.py

@@ -15,6 +15,8 @@
 """msprof PROF data export api file"""
 import os
 import shutil
+import json
+from json import JSONDecodeError
 from collections import defaultdict
 from subprocess import CalledProcessError, TimeoutExpired
 from subprocess import Popen, PIPE
@@ -39,7 +41,7 @@ class AscendMsprofExporter:
         >> ms_exporter = AscendMsprofExporter("path/to/profiler/data")
         >> ms_exporter.export(start_time)
     """
-
+    DRV_VERSION = 467473
     _hiai_msprof_tail = "Ascend/latest/tools/profiler/bin"
     _msprof_cmd = "msprof"
     _ascend_mark = "Ascend"
@@ -49,46 +51,79 @@ class AscendMsprofExporter:
     _op_summary_mark = "op_summary"
     _op_statistic_mark = "op_statistic"
 
-    def __init__(self, source_path, time_out=3000):
+    def __init__(self, source_path, time_out=3600):
         self._time_out = time_out
         self.source_path = source_path
-        self.prof_root_dir = os.path.abspath(os.path.join(self.source_path, os.path.pardir))
+        self.prof_root_dir = os.path.abspath(os.path.join(self.source_path, os.path.pardir))  # PROF*/
 
         self._check_msprof_env()
 
+    def get_drv_version(self):
+        """Get the drv_version for choosing the export mode."""
+        host_dir = os.path.join(self.prof_root_dir, 'host')
+        cmd = ['python',
+               '/usr/local/Ascend/latest/tools/profiler/profiler_tool/analysis/interface/get_msprof_info.py',
+               '-dir', host_dir]
+        try:
+            outs, _ = self._run_cmd(cmd)
+            if not outs:
+                logger.warning('Check the drvVersion can`t find the result, use single export mode instead.')
+                return False
+            result = json.loads(outs)
+            logger.info('get drv_version result is : %s', result)
+            status = result.get('status', 1)
+            if status == 1:
+                return False
+            drv_version = result.get('data', {}).get('version_info', {}).get('drv_version', 0)
+            if drv_version >= self.DRV_VERSION:
+                return True
+            return False
+        except (RuntimeError, JSONDecodeError, AttributeError) as err:
+            logger.warning('Get the drvVersion error, use single-export mode instead. detail : %s', err)
+            return False
+
     def export(self, model_iteration_dict=None):
         """start_time is the time to collect PROF data"""
 
-        if not model_iteration_dict:
-            model_iteration_dict = self._generate_step_trace(self.prof_root_dir, self.source_path)
+        flag = self.get_drv_version()
+        if not flag or model_iteration_dict:
+            flag = False
+            if not model_iteration_dict:
+                model_iteration_dict = self._generate_step_trace(self.prof_root_dir, self.source_path)
 
-        if model_iteration_dict:
-            for model_id, value in model_iteration_dict.items():
-                for iteration_id in value:
-                    msprof_export_cmd = self._msprof_command_generator(self.prof_root_dir, model_id, iteration_id)
-                    self._run_cmd(msprof_export_cmd)
-            self._check_export_files(self.source_path, model_iteration_dict)
+            if model_iteration_dict:
+                for model_id, value in model_iteration_dict.items():
+                    for iteration_id in value:
+                        msprof_export_cmd = self._msprof_command_generator_old(self.prof_root_dir, model_id,
+                                                                               iteration_id)
+                        self._run_cmd(msprof_export_cmd)
 
-    def _run_cmd(self, cmd, raise_error=True):
+                self._check_export_files_old(self.source_path, model_iteration_dict)
+        else:
+            msprof_export_cmd = self._msprof_command_generator(self.prof_root_dir)
+            self._run_cmd(msprof_export_cmd)
+            self._check_export_files(self.source_path)
+
+        return flag
+
+    def _run_cmd(self, cmd):
         """run shell command"""
         try:
             proc = Popen(cmd, stdout=PIPE, stderr=PIPE, text=True)
         except (FileNotFoundError, PermissionError, CalledProcessError) as exc:
-            raise RuntimeError(exc)
+            raise RuntimeError(exc) from exc
         try:
             outs, errs = proc.communicate(timeout=self._time_out)
-        except TimeoutExpired:
+        except TimeoutExpired as err:
             proc.kill()
             msg = "The possible cause is that too much data is collected " \
                   "and the export time is too long."
             logger.error(msg)
-            raise TimeoutError(msg)
+            raise TimeoutError(msg) from err
         logger.info(outs)
-        if raise_error and errs != "":
-            raise RuntimeError(errs)
-        return outs
+        return outs, errs
 
-    def _msprof_command_generator(self, output, model_id=None, iter_id=None):
+    def _msprof_command_generator_old(self, output, model_id=None, iter_id=None):
         """msprof export helper"""
         export_cmd = [self._msprof_cmd, "--export=on", "--output={}".format(output)]
         if isinstance(model_id, int) and model_id >= 0:
@@ -97,6 +132,10 @@ class AscendMsprofExporter:
             export_cmd.append("--iteration-id={}".format(iter_id))
         return export_cmd
 
+    def _msprof_command_generator(self, output):
+        """msprof export helper"""
+        return [self._msprof_cmd, "--export=on", "--output={}".format(output)]
+
     def _check_msprof_env(self):
         """Check the existence of msprof binary tool"""
 
@@ -109,7 +148,7 @@ class AscendMsprofExporter:
             return False
 
         msprof_cmd = ["which", self._msprof_cmd]
-        outs = self._run_cmd(msprof_cmd, raise_error=False)
+        outs, _ = self._run_cmd(msprof_cmd)
         if outs != "":
             return
         logger.warning("[Profiler]The msprof command was not found. Searching from environment variables...")
@@ -143,8 +182,7 @@ class AscendMsprofExporter:
         summary_path = os.path.join(device_path, self._summary_dir)
         timeline_path = os.path.join(device_path, self._timeline_dir)
 
-        msprof_export_cmd = self._msprof_command_generator(prof_path)
-        self._run_cmd(msprof_export_cmd)
+        self._run_cmd(self._msprof_command_generator_old(prof_path))
 
         if not os.path.isdir(summary_path):
             msg = "Path {} is not a existing directory. Make sure there is " \
@@ -158,16 +196,15 @@ class AscendMsprofExporter:
             return None
 
         step_trace = defaultdict(list)
-        with open(step_trace_file, newline='', mode='r') as csvfile:
+        with os.fdopen(os.open(step_trace_file, os.O_RDONLY, 0o600), newline='', mode='r') as csvfile:
             reader = csv.reader(csvfile, delimiter=',', quotechar='"')
-            header = next(reader)
-            for index, value in enumerate(header):
+            for index, value in enumerate(next(reader)):
                 if value == 'Model ID':
-                    Model_ID = index
+                    model_id = index
                 if value == 'Iteration ID':
-                    Iteration_ID = index
+                    iteration_id = index
             for row in reader:
-                step_trace[int(row[Model_ID])].append(int(row[Iteration_ID]))
+                step_trace[int(row[model_id])].append(int(row[iteration_id]))
 
         if os.path.isdir(summary_path):
             shutil.rmtree(summary_path)
@@ -176,16 +213,15 @@ class AscendMsprofExporter:
 
         return step_trace
 
-    def _check_export_files(self, source_path, step_trace):
+    def _check_export_files_old(self, source_path, step_trace):
         """Check the existence of op_summary & op_statistic files."""
         summary_path = os.path.join(source_path, self._summary_dir)
         if not os.path.isdir(summary_path):
             raise RuntimeError("Path {} is not a existing directory.".format(summary_path))
-        summary_file_list = os.listdir(summary_path)
         op_summary = set()
         op_statistic = set()
 
-        for summary_file in summary_file_list:
+        for summary_file in os.listdir(summary_path):
             if summary_file.startswith(self._op_summary_mark):
                 op_summary.add(summary_file)
             elif summary_file.startswith(self._op_statistic_mark):
@@ -196,18 +232,25 @@ class AscendMsprofExporter:
         if not op_statistic:
             raise RuntimeError("The op_statistics file was not found, perhaps the original data was not collected.")
 
-        device_id = source_path.split('_')[-1].replace("/", "")
+        logger.info("Finish checking files.")
 
-        for model_id, value in step_trace.items():
-            for iteration_id in value:
-                tag = f"_{device_id}_{model_id}_{iteration_id}.csv"
-                op_summary_file_name = self._op_summary_mark + tag
-                op_statistic_file = self._op_statistic_mark + tag
-                if op_summary_file_name not in op_summary:
-                    logger.warning("[Profiler]The file {} was not found, " \
-                                   "perhaps the original data was not collected.".format(op_summary_file_name))
-                if op_statistic_file not in op_statistic:
-                    logger.warning("[Profiler]The file {} was not found, " \
-                                   "perhaps the original data was not collected.".format(op_statistic_file))
+    def _check_export_files(self, source_path):
+        """Check the existence of op_summary & op_statistic files."""
+        summary_path = os.path.join(source_path, self._summary_dir)
+        if not os.path.isdir(summary_path):
+            raise RuntimeError("Path {} is not a existing directory.".format(summary_path))
+        summary_file_list = os.listdir(summary_path)
+        op_summary = set()
+        op_statistic = set()
+
+        for summary_file in summary_file_list:
+            if summary_file.startswith(self._op_summary_mark):
+                op_summary.add(summary_file)
+            elif summary_file.startswith(self._op_statistic_mark):
+                op_statistic.add(summary_file)
 
+        if not op_summary:
+            raise RuntimeError("The op_summary file was not found, perhaps the original data was not collected.")
+        if not op_statistic:
+            raise RuntimeError("The op_statistics file was not found, perhaps the original data was not collected.")
         logger.info("Finish checking files.")