Add option to use NVIDIA Nsight Compute CLI profiler (#21)

* Use NVIDIA Nsight Compute CLI profiler

* Add profile and profiler-args options to argument parser.

* Add missing comma to profiler-args option.

* Use profile args in version 1 of the plugin

* Change profiler-args option to take all remaining arguments

* Change profiler_args type from string to list of strings

* Add profile option to version 2 of the plugin

* Add profiler usage instructions

README.md

@@ -23,3 +23,12 @@ V2 brings support of multiple source and header files.
 %%cuda_run
 # This line just to bypass an exeption and can contain any text
 ```
+
+- To profile your CUDA kernels using NVIDIA Nsight Compute CLI profiler you need to run
+```
+%%cu --profile
+```
+- You can add options to the profiler. Keep in mind that any argument after "--profiler-args" will be considered as a profiler argument. For example, to select which sections to collect metrics for you need to run
+```
+%%cu --profile --profiler-args --section SpeedOfLight --section MemoryWorkloadAnalysis --section Occupancy
+```

common/helper.py

@@ -3,8 +3,27 @@ import argparse
 
 def get_argparser():
     parser = argparse.ArgumentParser(description='NVCCPlugin params')
-    parser.add_argument("-t", "--timeit", action='store_true',
-                        help='flag to return timeit result instead of stdout')
+    parser.add_argument(
+        '-t', 
+        '--timeit',
+        action='store_true',
+        help='If set, returns the output of the "timeit" built-in ipython magic instead of stdout.',
+    )
+    parser.add_argument(
+        '-p', 
+        '--profile', 
+        action='store_true',
+        help='If set, runs the nvidia nsight compute profiler. Has no effect if used with --timeit.',
+    )
+    parser.add_argument(
+        '-a',
+        '--profiler-args',
+        type=str,
+        nargs=argparse.REMAINDER,
+        default=[],
+        help='Extra options that can be passed to the nvidia nsight compute profiler. '
+             'Must be the last option given to the argument parser so you can pass arguments with dashes.',
+    )
     return parser
 
 

v1/v1.py

@@ -7,6 +7,7 @@ from IPython.core.magic import Magics, cell_magic, magics_class
 from common import helper
 
 compiler = '/usr/local/cuda/bin/nvcc'
+profiler = '/usr/local/cuda/bin/ncu'
 ext = '.cu'
 
 
@@ -23,15 +24,18 @@ class NVCCPlugin(Magics):
         subprocess.check_output(
             [compiler, file_path + ext, "-o", file_path + ".out", '-Wno-deprecated-gpu-targets'], stderr=subprocess.STDOUT)
 
-    def run(self, file_path, timeit=False):
+    def run(self, file_path, timeit=False, profile=False, profiler_args=[]):
         if timeit:
             stmt = f"subprocess.check_output(['{file_path}.out'], stderr=subprocess.STDOUT)"
             output = self.shell.run_cell_magic(
                 magic_name="timeit", line="-q -o import subprocess", cell=stmt)
             output = str(output) # convert TimeitResult object to human readable string
         else:
-            output = subprocess.check_output(
-                [file_path + ".out"], stderr=subprocess.STDOUT)
+            run_args = []
+            if profile:
+                run_args.extend([profiler] + profiler_args)
+            run_args.append(file_path + ".out")
+            output = subprocess.check_output(run_args, stderr=subprocess.STDOUT)
             output = output.decode('utf8')
             
         helper.print_out(output)
@@ -51,7 +55,7 @@ class NVCCPlugin(Magics):
                 f.write(cell)
             try:
                 self.compile(file_path)
-                output = self.run(file_path, timeit=args.timeit)
+                output = self.run(file_path, timeit=args.timeit, profile=args.profile, profiler_args=args.profiler_args)
             except subprocess.CalledProcessError as e:
                 helper.print_out(e.output.decode("utf8"))
                 output = None

v2/v2.py

@@ -6,6 +6,7 @@ from IPython.core.magic_arguments import argument, magic_arguments, parse_argstr
 from common import helper
 
 compiler = '/usr/local/cuda/bin/nvcc'
+profiler = '/usr/local/cuda/bin/ncu'
 
 
 @magics_class
@@ -32,15 +33,18 @@ class NVCCPluginV2(Magics):
         res = res.decode()
         helper.print_out(res)
 
-    def run(self, timeit=False):
+    def run(self, timeit=False, profile=False, profiler_args=[]):
         if timeit:
             stmt = f"subprocess.check_output(['{self.out}'], stderr=subprocess.STDOUT)"
             output = self.shell.run_cell_magic(
                 magic_name="timeit", line="-q -o import subprocess", cell=stmt)
             output = str(output) # convert TimeitResult object to human readable string
         else:
-            output = subprocess.check_output(
-                [self.out], stderr=subprocess.STDOUT)
+            run_args = []
+            if profile:
+                run_args.extend([profiler] + profiler_args)
+            run_args.append(self.out)
+            output = subprocess.check_output(run_args, stderr=subprocess.STDOUT)
             output = output.decode('utf8')
 
         helper.print_out(output)
@@ -72,7 +76,7 @@ class NVCCPluginV2(Magics):
         if args.compile:
             try:
                 self.compile(self.output_dir, file_path, self.out)
-                output = self.run(timeit=args.timeit)
+                output = self.run(timeit=args.timeit, profile=args.profile, profiler_args=args.profiler_args)
             except subprocess.CalledProcessError as e:
                 helper.print_out(e.output.decode("utf8"))
                 output = None
@@ -95,7 +99,7 @@ class NVCCPluginV2(Magics):
                         for x in cuda_src if x[-3:] == '.cu']
             print(f'found sources: {cuda_src}')
             self.compile(self.output_dir, ' '.join(cuda_src), self.out)
-            output = self.run(timeit=args.timeit)
+            output = self.run(timeit=args.timeit, profile=args.profile, profiler_args=args.profiler_args)
         except subprocess.CalledProcessError as e:
             helper.print_out(e.output.decode("utf8"))
             output = None