triton-windows 3.3.0.post19__cp312-cp312-win_amd64.whl → 3.4.0.post20__cp312-cp312-win_amd64.whl

triton/language/extra/cuda/__init__.py

@@ -1,13 +1,16 @@
 from . import libdevice
 
 from .utils import (globaltimer, num_threads, num_warps, smid, convert_custom_float8_sm70, convert_custom_float8_sm80)
-
-from ._experimental_tma import *  # noqa: F403
-from ._experimental_tma import __all__ as _tma_all
+from .gdc import (gdc_launch_dependents, gdc_wait)
 
 __all__ = [
-    "libdevice", "globaltimer", "num_threads", "num_warps", "smid", "convert_custom_float8_sm70",
-    "convert_custom_float8_sm80", *_tma_all
+    "libdevice",
+    "globaltimer",
+    "num_threads",
+    "num_warps",
+    "smid",
+    "convert_custom_float8_sm70",
+    "convert_custom_float8_sm80",
+    "gdc_launch_dependents",
+    "gdc_wait",
 ]
-
-del _tma_all

triton/language/extra/cuda/gdc.py

@@ -0,0 +1,42 @@
+"""
+Grid Dependency Control (GDC) is a mechanism used when enabling programmatic dependent launch to launch and
+synchronize grids. These APIs expose GDC to the programmer.
+
+Programmatic dependent launch is supported on SM90 (Hopper) and beyond.
+For PTX reference on grid dependency control see https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#parallel-synchronization-and-communication-instructions-griddepcontrol.
+"""
+
+from triton.language import core
+
+
+@core.extern
+def gdc_wait(_builder=None):
+    """
+    GDC wait is a blocking instruction that waits for all instructions in a prior kernel to complete before continuing.
+    This ensures all memory operations happening before the wait is visible to instructions after it,
+    e.g. if the prior kernel writes to address "x" the new values will be visible in this kernel after the wait.
+
+    This instruction is also safe to execute when programatic dependent launch is disabled.
+
+    See https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#parallel-synchronization-and-communication-instructions-griddepcontrol for more details.
+    """
+    core.inline_asm_elementwise("griddepcontrol.wait; // dummy $0", "=r", [], dtype=core.int32, is_pure=False, pack=1,
+                                _builder=_builder)
+
+
+@core.extern
+def gdc_launch_dependents(_builder=None):
+    """
+    This operation when launched with programmatic dependent launch signals that
+    the next program may launch once all programs in the current kernel
+    call this function or complete.
+
+    Repeated calls to this function have no effect past the first call, and the first call should be
+    treated by the programmer as a hint to the runtime system to launch the next kernel.
+
+    This instruction is also safe to execute when programatic dependent launch is disabled.
+
+    See https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#parallel-synchronization-and-communication-instructions-griddepcontrol for more details.
+    """
+    core.inline_asm_elementwise("griddepcontrol.launch_dependents; // dummy $0", "=r", [], dtype=core.int32,
+                                is_pure=False, pack=1, _builder=_builder)