Optional use of GPU (plasma-umass#830)

* Only init GPU when requested / if it's default on system.

* Suppress a mypy complaint.

* Force init when no GPU is available.

* Bumped version.

* Removed.

scalene/find_browser.py

@@ -21,7 +21,7 @@ def find_browser(browserClass: Optional[str] = None) -> Optional[str]:
         # we need to refer to it as such to prevent this error:
         #   'MacOSXOSAScript' object has no attribute 'name'
         browser = webbrowser.get(browserClass)
-        return browser._name if browser._name not in text_browsers else None
+        return browser._name if browser._name not in text_browsers else None # type: ignore[attr-defined]
     except webbrowser.Error:
         # Return None if there is an error in getting the browser
         return None

scalene/scalene_config.py

@@ -1,7 +1,7 @@
 """Current version of Scalene; reported by --version."""
 
-scalene_version = "1.5.42.2"
-scalene_date = "2024.06.18"
+scalene_version = "1.5.42.3"
+scalene_date = "2024.07.05"
 
 # Port to use for Scalene UI
 SCALENE_PORT = 11235

scalene/scalene_profiler.py

@@ -101,11 +101,6 @@
 if sys.platform != "win32":
     import resource
 
-if platform.system() == "Darwin":
-    from scalene.scalene_apple_gpu import ScaleneAppleGPU as ScaleneGPU
-else:
-    from scalene.scalene_gpu import ScaleneGPU  # type: ignore
-
 from scalene.scalene_parseargs import ScaleneParseArgs, StopJupyterExecution
 from scalene.scalene_sigqueue import ScaleneSigQueue
 
@@ -214,10 +209,7 @@ def last_profiled_tuple() -> Tuple[Filename, LineNumber, ByteCodeIndex]:
     __stats = ScaleneStatistics()
     __output = ScaleneOutput()
     __json = ScaleneJSON()
-    __gpu = ScaleneGPU()
-
-    __output.gpu = __gpu.has_gpu()
-    __json.gpu = __gpu.has_gpu()
+    __gpu = None # initialized after parsing arguments in `main`
     __invalidate_queue: List[Tuple[Filename, LineNumber]] = []
     __invalidate_mutex: threading.Lock
     __profiler_base: str
@@ -762,7 +754,10 @@ def cpu_signal_handler(
                     )
                 return
 
-            (gpu_load, gpu_mem_used) = Scalene.__gpu.get_stats()
+            if Scalene.__gpu:
+                (gpu_load, gpu_mem_used) = Scalene.__gpu.get_stats()
+            else:
+                (gpu_load, gpu_mem_used) = (0.0, 0.0)
 
             # Process this CPU sample.
             Scalene.process_cpu_sample(
@@ -1442,7 +1437,7 @@ def after_fork_in_child() -> None:
         Scalene.__is_child = True
 
         Scalene.clear_metrics()
-        if Scalene.__gpu.has_gpu():
+        if Scalene.__gpu and Scalene.__gpu.has_gpu():
             Scalene.__gpu.nvml_reinit()
         # Note: __parent_pid of the topmost process is its own pid.
         Scalene.__pid = Scalene.__parent_pid
@@ -1887,6 +1882,19 @@ def main() -> None:
             args,
             left,
         ) = ScaleneParseArgs.parse_args()
+        # Try to profile a GPU if one is found and `--gpu` is selected / it's the default (see ScaleneArguments).
+        if args.gpu:
+            if platform.system() == "Darwin":
+                from scalene.scalene_apple_gpu import ScaleneAppleGPU as ScaleneGPU
+            else:
+                from scalene.scalene_gpu import ScaleneGPU  # type: ignore
+            Scalene.__gpu = ScaleneGPU()
+            Scalene.__output.gpu = Scalene.__gpu.has_gpu()
+            Scalene.__json.gpu = Scalene.__gpu.has_gpu()
+        else:
+            Scalene.__gpu = None
+            Scalene.__output.gpu = False
+            Scalene.__json.gpu = False
         Scalene.set_initialized()
         Scalene.run_profiler(args, left)