Add python script to detect nvidia CC

Author: zhi-wang

cmake/nvidiacc.py

@@ -0,0 +1,90 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+'''
+This script detects the compute capabilities (X.Y) of the the GPU cards and
+returns them in a comma-separated XY string, if multiple GPUs are detected.
+
+The auto-detected compute capabilities (XY) are used by the CMakeLists.txt only
+if they are not explicitly set by the user. For the following situations:
+    - the compiler does not support the XY value, e.g., XY=52;
+    - the compiler is older than the device;
+    - you don't have python3 in your system;
+    - etc.
+you must set the compute capabilities explicitly when cmake is configured.
+
+Reference:
+https://gist.github.com/f0k/63a664160d016a491b2cbea15913d549
+by Jan Schlüter
+'''
+
+
+import ctypes
+import sys
+
+
+def CommaSeparatedCCString():
+    libnames = ('libcuda.so', 'libcuda.dylib', 'nvcuda.dll', 'cuda.dll')
+    for libname in libnames:
+        try:
+            cuda = ctypes.CDLL(libname)
+        except OSError:
+            continue
+        else:
+            break
+    else:
+        raise OSError('could not load any of: ' + ' '.join(libnames))
+
+    ############################################################
+
+    def checkCall(f, *args):
+        CUDA_SUCCESS = 0
+        result = ctypes.c_int()
+
+        result = f(*args)
+        if result != CUDA_SUCCESS:
+            error_str = ctypes.c_char_p()
+            cuda.cuGetErrorString(result, ctypes.byref(error_str))
+            print('{} failed with error code {}: {}'.format(f.__name__, result, error_str.value.decode()))
+            sys.exit(1)
+
+    ############################################################
+
+    cc_set = set()
+
+    # from cuda.h
+    CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR = 75
+    CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR = 76
+
+    checkCall(cuda.cuInit, 0)
+
+    nGpus = ctypes.c_int()
+    checkCall(cuda.cuDeviceGetCount, ctypes.byref(nGpus))
+
+    for i in range(nGpus.value):
+        device = ctypes.c_int()
+        checkCall(cuda.cuDeviceGet, ctypes.byref(device), i)
+
+        cc_major = ctypes.c_int()
+        cc_minor = ctypes.c_int()
+        checkCall(cuda.cuDeviceGetAttribute, ctypes.byref(cc_major), CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, device)
+        checkCall(cuda.cuDeviceGetAttribute, ctypes.byref(cc_minor), CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, device)
+        cc = cc_major.value * 10 + cc_minor.value
+        cc_set.add(cc)
+
+    cc_list = []
+    for c in sorted(cc_set):
+        cc_list.append(c)
+
+    cc_len = len(cc_list)
+    if cc_len:
+        cc_str = '{}'.format(cc_list[0])
+        for i in range(2, cc_len):
+            cc_str = cc_str + ',{}'.format(cc_list[i])
+        return cc_str
+    else:
+        return ''
+
+
+if __name__ == '__main__':
+    print(CommaSeparatedCCString())

src/cudart/gpucard.cpp

@@ -114,8 +114,7 @@ static void getDeviceAttribute(DeviceAttribute& a, int device = 0)
    else
       found_cc = false;
 
-   // Number of CUDA cores per multiprocessor, not tabulated in
-   // cuda-c-programming-guide;
+   // Number of CUDA cores per multiprocessor, not tabulated in cuda-c-programming-guide;
    // documented in "Compute Capability - architecture"
    // 8.6: 128
    // 7.0 7.2 7.5 8.0: 64