Inductor annotations (pytorch#130429)

Add NVTX annotations around training phases and buffer computations

RFC/discussion: https://dev-discuss.pytorch.org/t/rfc-performance-profiling-at-scale-with-details-nvtx-annotations/2224

<img width="2160" alt="Screenshot 2024-07-10 at 11 48 04" src="https://github.com/pytorch/pytorch/assets/1175576/9ade139c-d393-473f-9b68-6c25da367dc4">

Pull Request resolved: pytorch#130429
Approved by: https://github.com/aorenste, https://github.com/eellison, https://github.com/albanD

Co-authored-by: Cedric GESTES <cedric.gestes@flex.ai>

Author: 2 people

test/inductor/test_inductor_annotations.py

@@ -0,0 +1,41 @@
+# Owner(s): ["module: inductor"]
+import torch
+import torch._inductor.config as inductor_config
+from torch._inductor.test_case import run_tests, TestCase
+from torch._inductor.utils import run_and_get_code
+from torch.testing._internal.triton_utils import requires_cuda
+
+
+class InductorAnnotationTestCase(TestCase):
+    def get_code(self):
+        def f(a, b):
+            return a + b, a * b
+
+        a = torch.randn(5, device="cuda")
+        b = torch.randn(5, device="cuda")
+        f_comp = torch.compile(f)
+
+        _, code = run_and_get_code(f_comp, a, b)
+        return code[0]
+
+    @requires_cuda
+    def test_no_annotations(self):
+        code = self.get_code()
+
+        self.assertTrue("from torch.cuda import nvtx" not in code)
+        self.assertTrue("training_annotation" not in code)
+
+    @inductor_config.patch(annotate_training=True)
+    @requires_cuda
+    def test_training_annotation(self):
+        code = self.get_code()
+
+        self.assertTrue("from torch.cuda import nvtx" in code)
+        self.assertEqual(
+            code.count("training_annotation = nvtx._device_range_start('inference')"), 1
+        )
+        self.assertEqual(code.count("nvtx._device_range_end(training_annotation)"), 1)
+
+
+if __name__ == "__main__":
+    run_tests()

torch/_C/_nvtx.pyi

@@ -5,3 +5,5 @@ def rangePop() -> int: ...
 def rangeStartA(message: str) -> int: ...
 def rangeEnd(int) -> None: ...
 def markA(message: str) -> None: ...
+def deviceRangeStart(message: str, stream: int) -> object: ...
+def deviceRangeEnd(range_handle: object, stream: int) -> None: ...

torch/_inductor/codegen/wrapper.py

@@ -772,6 +772,8 @@ def write_header(self) -> None:
             )
         except (AttributeError, ImportError):
             pass
+        if config.annotate_training:
+            self.header.writeline("from torch.cuda import nvtx")
 
     def include_extra_header(self, header: str):
         pass
@@ -889,6 +891,11 @@ def {self.launcher_fn_name}(args):
         with self.prefix.indent():
             if config.triton.debug_sync_graph:
                 self.prefix.writeline(V.graph.device_ops.synchronize())
+            phase = V.graph.get_training_phase()
+            if config.annotate_training:
+                self.prefix.writeline(
+                    f"training_annotation = nvtx._device_range_start('{phase}')"
+                )
             if V.graph.graph_inputs:
                 lhs = ", ".join(V.graph.graph_input_names)
                 if len(V.graph.graph_input_names) == 1:
@@ -1175,6 +1182,10 @@ def _generate(self, is_inference):
             if config.triton.autotune_at_compile_time:
                 self.generate_and_run_autotune_block()
 
+            if config.annotate_training:
+                self.wrapper_call.writeline(
+                    "nvtx._device_range_end(training_annotation)"
+                )
             self.generate_return(output_refs)
 
         self.finalize_prefix()

torch/_inductor/config.py

@@ -773,6 +773,10 @@ def decide_compile_threads() -> int:
 )
 
 
+# Adds NVTX annotations aroung training phases
+annotate_training: bool = os.environ.get("TORCHINDUCTOR_ANNOTATE_TRAINING", "0") == "1"
+
+
 # config specific to codegen/cpp.py
 class cpp:
     # set to torch.get_num_threads()

torch/_inductor/graph.py

@@ -519,6 +519,13 @@ def set_current_device(self, device: torch.device) -> Iterator[None]:
         finally:
             self.current_device = prior
 
+    def get_training_phase(self) -> str:
+        if self.is_inference:
+            return "inference"
+        if self.is_backward:
+            return "backward"
+        return "forward"
+
     @staticmethod
     def decide_layout_opt(gm: GraphModule, *, is_inference: bool) -> bool:
         """

torch/csrc/cuda/shared/nvtx.cpp

@@ -6,10 +6,41 @@
 #else
 #include <nvToolsExt.h>
 #endif
+#include <cuda_runtime.h>
 #include <torch/csrc/utils/pybind.h>
 
 namespace torch::cuda::shared {
 
+struct RangeHandle {
+  nvtxRangeId_t id;
+  const char* msg;
+};
+
+static void device_callback_range_end(void* userData) {
+  RangeHandle* handle = ((RangeHandle*)userData);
+  nvtxRangeEnd(handle->id);
+  free((void*)handle->msg);
+  free((void*)handle);
+}
+
+static void device_nvtxRangeEnd(void* handle, std::intptr_t stream) {
+  cudaLaunchHostFunc((cudaStream_t)stream, device_callback_range_end, handle);
+}
+
+static void device_callback_range_start(void* userData) {
+  RangeHandle* handle = ((RangeHandle*)userData);
+  handle->id = nvtxRangeStartA(handle->msg);
+}
+
+static void* device_nvtxRangeStart(const char* msg, std::intptr_t stream) {
+  RangeHandle* handle = (RangeHandle*)calloc(sizeof(RangeHandle), 1);
+  handle->msg = strdup(msg);
+  handle->id = 0;
+  cudaLaunchHostFunc(
+      (cudaStream_t)stream, device_callback_range_start, (void*)handle);
+  return handle;
+}
+
 void initNvtxBindings(PyObject* module) {
   auto m = py::handle(module).cast<py::module>();
 
@@ -23,6 +54,8 @@ void initNvtxBindings(PyObject* module) {
   nvtx.def("rangeStartA", nvtxRangeStartA);
   nvtx.def("rangeEnd", nvtxRangeEnd);
   nvtx.def("markA", nvtxMarkA);
+  nvtx.def("deviceRangeStart", device_nvtxRangeStart);
+  nvtx.def("deviceRangeEnd", device_nvtxRangeEnd);
 }
 
 } // namespace torch::cuda::shared

torch/cuda/nvtx.py

@@ -66,6 +66,38 @@ def range_end(range_id) -> None:
     _nvtx.rangeEnd(range_id)
 
 
+def _device_range_start(msg: str, stream: int = 0) -> object:
+    """
+    Marks the start of a range with string message.
+    It returns an opaque heap-allocated handle for this range
+    to pass to the corresponding call to device_range_end().
+
+    A key difference between this and range_start is that the
+    range_start marks the range right away, while _device_range_start
+    marks the start of the range as soon as all the tasks on the
+    CUDA stream are completed.
+
+    Returns: An opaque heap-allocated handle that should be passed to _device_range_end().
+
+    Args:
+        msg (str): ASCII message to associate with the range.
+        stream (int): CUDA stream id.
+    """
+    return _nvtx.deviceRangeStart(msg, stream)
+
+
+def _device_range_end(range_handle: object, stream: int = 0) -> None:
+    """
+    Mark the end of a range for a given range_handle as soon as all the tasks
+    on the CUDA stream are completed.
+
+    Args:
+        range_handle: an unique handle for the start range.
+        stream (int): CUDA stream id.
+    """
+    _nvtx.deviceRangeEnd(range_handle, stream)
+
+
 def mark(msg):
     """
     Describe an instantaneous event that occurred at some point.

torch/utils/hipify/cuda_to_hip_mappings.py

@@ -5035,6 +5035,7 @@
             "cudaLaunchCooperativeKernel",
             ("hipLaunchCooperativeKernel", CONV_EXEC, API_RUNTIME),
         ),
+        ("cudaLaunchHostFunc", ("hipLaunchHostFunc", CONV_EXEC, API_RUNTIME, HIP_UNSUPPORTED)),
         (
             "cudaSetupArgument",
             ("hipSetupArgument", CONV_EXEC, API_RUNTIME, HIP_UNSUPPORTED),
@@ -7965,6 +7966,7 @@
         ("nvtxRangePop", ("roctxRangePop", CONV_OTHER, API_ROCTX)),
         ("nvtxRangeStartA", ("roctxRangeStartA", CONV_OTHER, API_ROCTX)),
         ("nvtxRangeEnd", ("roctxRangeStop", CONV_OTHER, API_ROCTX)),
+        ("nvtxRangeId_t", ("int", CONV_OTHER, API_ROCTX)),
         ("nvmlReturn_t", ("rsmi_status_t", CONV_OTHER, API_ROCMSMI)),
         ("NVML_SUCCESS", ("RSMI_STATUS_SUCCESS", CONV_OTHER, API_ROCMSMI)),
         ("NVML_P2P_CAPS_INDEX_READ", ("RSMI_STATUS_SUCCESS", CONV_OTHER, API_ROCMSMI)),