From 9ed721de0e9ce564b7c852e38359398b019a5c2f Mon Sep 17 00:00:00 2001
From: Joshua Bakita <bakitajoshua@gmail.com>
Date: Thu, 16 Mar 2023 02:18:35 -0400
Subject: Introduce pysmctrl: A python interface to libsmctrl

Initially supports the GPU information functions via:
- pysmctrl.get_gpc_info(dev_id)
- pysmctrl.get_tpc_info(dev_id)
- pysmctrl.get_tpc_info_cuda(cuda_dev_id)
All functions are extensively documented. See pysmctrl/__init__.py
for details.

Device partitioning functions have yet to be mapped into Python, as
these will require more testing.

As part of this:
- libsmctrl_get_*_info() functions have been modified to consistently
  return positive error codes.
- libsmctrl_get_tpc_info() now uses nvdebug-style device numbering and
  uses libsmctrl_get_gpc_info() under the covers. This should be more
  reliable.
- libsmctrl_get_tpc_info_cuda() has been introduced as an improved
  version of the old libsmctrl_get_tpc_info() function. This continues
  to use CUDA-style device numbering, but is now resiliant to CUDA
  failures.
- Various minor style improvements in libsmctrl.c
---
 pysmctrl/__init__.py | 82 ++++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 82 insertions(+)
 create mode 100644 pysmctrl/__init__.py

(limited to 'pysmctrl')

diff --git a/pysmctrl/__init__.py b/pysmctrl/__init__.py
new file mode 100644
index 0000000..5dc6175
--- /dev/null
+++ b/pysmctrl/__init__.py
@@ -0,0 +1,82 @@
+import ctypes, ctypes.util
+import os
+
+# If this is failing, make sure that the directory containing libsmctrl.so is
+# in your LD_LIBRARY_PATH environment variable. You likely need something like:
+# LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/playpen/jbakita/gpu_subdiv/libsmctrl/
+libsmctrl_path = ctypes.util.find_library("libsmctrl")
+if not libsmctrl_path:
+    libsmctrl_path = __path__[0] + "/../libsmctrl.so"
+libsmctrl = ctypes.CDLL(libsmctrl_path)
+
+def get_gpc_info(device_num):
+    """
+    Obtain list of thread processing clusters (TPCs) enabled for each general
+    processing cluster (GPC) in the specified GPU.
+
+    Parameters
+    ----------
+    device_num : int
+        Which device to obtain information for (starts as 0, order is defined
+        by nvdebug module). May not match CUDA device numbering.
+
+    Returns
+    -------
+    list of int64
+        A list as long as the number of GPCs enabled, where each list entry is
+        a bitmask. A bit set at index `i` indicates that TPC `i` is part of the
+        GPC at that list index. Obtained via GPU register reads in `nvdebug`.
+    """
+    num_gpcs = ctypes.c_uint()
+    tpc_masks = ctypes.pointer(ctypes.c_ulonglong())
+    res = libsmctrl.libsmctrl_get_gpc_info(ctypes.byref(num_gpcs), ctypes.byref(tpc_masks), device_num)
+    if res != 0:
+        print("pysmctrl: Unable to call libsmctrl_get_gpc_info(). Raising error %d..."%res)
+        raise OSError(res, os.strerror(res))
+    return [tpc_masks[i] for i in range(num_gpcs.value)]
+
+def get_tpc_info(device_num):
+    """
+    Obtain a count of the total number of thread processing clusters (TPCs)
+    enabled on the specified GPU.
+
+    Parameters
+    ----------
+    device_num : int
+        Which device to obtain TPC count for (starts as 0, order is defined by
+        `nvdebug` module). May not match CUDA device numbering.
+
+    Returns
+    -------
+    int
+        Count of enabled TPCs. Obtained via GPU register reads in `nvdebug`.
+    """
+    num_tpcs = ctypes.c_uint()
+    res = libsmctrl.libsmctrl_get_tpc_info(ctypes.byref(num_tpcs), device_num)
+    if res != 0:
+        print("pysmctrl: Unable to call libsmctrl_get_tpc_info(). Raising error %d..."%res)
+        raise OSError(res, os.strerror(res))
+    return num_tpcs.value
+
+def get_tpc_info_cuda(device_num):
+    """
+    Obtain a count of the total number of thread processing clusters (TPCs)
+    enabled on the specified GPU.
+
+    Parameters
+    ----------
+    device_num : int
+        Which device to obtain TPC count for, as a CUDA device ID.
+
+    Returns
+    -------
+    int
+        Count of enabled TPCs. Obtained via calculations on data from CUDA.
+    """
+    num_tpcs = ctypes.c_uint()
+    res = libsmctrl.libsmctrl_get_tpc_info_cuda(ctypes.byref(num_tpcs), device_num)
+    if res != 0:
+        print("pysmctrl: Unable to call libsmctrl_get_tpc_info_cuda(). Raising error %d..."%res)
+        raise OSError(res, os.strerror(res))
+    return num_tpcs.value
+
-- 
cgit v1.2.2