Gpu estimation command (#48)

* gpu usage command integration

* changed name of memory output field

* changed comment

Author: johncalesp

deepview_profile/__main__.py

@@ -6,6 +6,7 @@
 import deepview_profile.commands.interactive
 import deepview_profile.commands.memory
 import deepview_profile.commands.time
+import deepview_profile.commands.gpu_estimation
 
 
 def main():
@@ -23,6 +24,7 @@ def main():
     deepview_profile.commands.interactive.register_command(subparsers)
     deepview_profile.commands.memory.register_command(subparsers)
     deepview_profile.commands.time.register_command(subparsers)
+    deepview_profile.commands.gpu_estimation.register_command(subparsers)
     args = parser.parse_args()
 
     if args.version:

deepview_profile/commands/gpu_estimation.py

@@ -0,0 +1,119 @@
+import subprocess
+import sys
+import os
+import sqlite3
+from collections import defaultdict
+
+NS_TIME = 1e9
+
+def register_command(subparsers):
+    parser = subparsers.add_parser(
+        "gpu-usage-estimation",
+        help="Estimation of gpu operation time. We recommend running your training cycle for 100 iterations"
+    )
+
+    parser.add_argument(
+        "path_to_file",
+        help="path of the file you want to analyze"
+    )
+    parser.set_defaults(func=main)
+
+def joinIntervals(arr):
+    # arr = tuple(type,start,end,streamid)
+    eventDict = defaultdict(int)
+    filteredArr = []
+    prevRecord = list(arr[0])
+    for i in range(1,len(arr)):
+        newRecord = list(arr[i])
+        if prevRecord[1] <= newRecord[1] <= prevRecord[2]:
+            prevRecord[1] = min(prevRecord[1], newRecord[1])
+            prevRecord[2] = max(prevRecord[2], newRecord[2])
+        else:
+            filteredArr.append(prevRecord)
+            prevRecord = newRecord
+    filteredArr.append(prevRecord) # append the last record
+    for item in filteredArr:
+        eventDict[item[0]] += (item[2]-item[1])
+    return eventDict
+
+def sql_command_execution(db_path):
+    connection = sqlite3.connect(db_path, detect_types=sqlite3.PARSE_DECLTYPES|sqlite3.PARSE_COLNAMES)
+    cursor = connection.cursor()
+
+    try:
+        timeline_data = cursor.execute("""
+            SELECT "memOps" as name, start,end, streamId
+            FROM CUPTI_ACTIVITY_KIND_MEMCPY
+            UNION ALL
+            SELECT "kernelOps" as name, start, end, streamId
+            FROM CUPTI_ACTIVITY_KIND_KERNEL
+            UNION ALL
+            SELECT "memOps" as name, start, end, streamId
+            FROM CUPTI_ACTIVITY_KIND_MEMSET
+            ORDER by start ASC;
+            """).fetchall()
+        profiling_duration = cursor.execute("""
+            SELECT duration FROM ANALYSIS_DETAILS;
+            """).fetchone()[0]
+        cupti_api_duration = cursor.execute("""
+            SELECT max(end)-min(start) from CUPTI_ACTIVITY_KIND_RUNTIME;
+            """).fetchone()[0]
+        cursor.close()
+    except sqlite3.Error as er:
+        print("There was an error reading the information from the sqlite database")
+        print('SQLite error: %s' % (' '.join(er.args)))
+        cursor.close()
+        sys.exit(1)
+
+    if not timeline_data:
+        print("There are no traces of gpu activity")
+        sys.exit()
+    gpu_activity_time = joinIntervals(timeline_data)
+    percgpu_activity = ((gpu_activity_time["kernelOps"]+gpu_activity_time["memOps"])/cupti_api_duration)*100
+    data = [round(profiling_duration/NS_TIME,3),
+             round(cupti_api_duration/NS_TIME,3),
+             round(gpu_activity_time["kernelOps"]/NS_TIME,3),
+             round(gpu_activity_time["memOps"]/NS_TIME,3),
+             round(percgpu_activity,3)]
+    
+    return data
+
+def remove_files(curr_dir):
+    nsysfile = os.path.join(curr_dir,"gpu_estimation.nsys-rep")
+    sqlitefile = os.path.join(curr_dir,"gpu_estimation.sqlite")
+    subprocess.run(["rm",nsysfile], capture_output=True, text=True)
+    subprocess.run(["rm",sqlitefile], capture_output=True, text=True)
+
+
+def actual_main(args):
+    result = subprocess.run(["which","nsys"], capture_output=True, text=True)
+    if not result.stdout:
+        print("Please make sure the command nsys is included in your path")
+        print("You can try: export PATH=[path/to/bin]:$PATH")
+        print("You can verify using:","\nwhich nsys","\nnsys --version")
+        sys.exit(1)
+
+    curr_dir = subprocess.run(["pwd"], capture_output=True, text=True).stdout.strip()
+    nsys_output = subprocess.run(["nsys","profile","--trace=cuda,osrt","--cpuctxsw=none","--sample=none","--force-overwrite=true","--stats=true","--output=gpu_estimation","python", args.path_to_file], 
+                                 stdout=subprocess.PIPE, 
+                                 stderr=subprocess.PIPE, 
+                                 text=True)
+    if nsys_output.stderr:
+        print("An error ocurred during the analysis")
+        print("Please make sure that your training is executing on GPU")
+        print("Error:",nsys_output.stderr)
+        # remove generated files
+        remove_files(curr_dir)
+        sys.exit(1)
+
+    db_path = os.path.join(curr_dir,"gpu_estimation.sqlite")
+    summary = sql_command_execution(db_path)
+    headers = ["Estimate Profiling time","CUDA API Time","Kernel Ops Time","Memory Ops time","GPU Perc"]
+    format_row = "{:^25}" * len((headers))
+    print(format_row.format(*headers))
+    print(format_row.format(*summary))
+    # remove generated files
+    remove_files(curr_dir)
+
+def main(args):
+    actual_main(args)