Hi,
Here is a small diff to allow int8 to work correctly.
In file pycuda/tools.py in the function dtype_to_ctype, it use "signed
char" for numpy.int8, but the function parse_c_arg don't understand
it.
We can change dtype_to_ctype to return just "char", but that would be
inconsistent with what is done for othe dtype.
Here is a patch that make parse_c_arg understand "signed char"
diff --git a/pycuda/tools.py b/pycuda/tools.py
index f16cc60..3ce4648 100644
--- a/pycuda/tools.py
+++ b/pycuda/tools.py
@@ -473,7 +473,7 @@ def parse_c_arg(c_arg):
dtype = np.uint32
elif tp in ["short", "short int"]: dtype = np.int16
elif tp in ["unsigned short", "unsigned short int"]: dtype = np.uint16
- elif tp in ["char"]: dtype = np.int8
+ elif tp in ["char", "signed char"]: dtype = np.int8
elif tp in ["unsigned char"]: dtype = np.uint8
elif tp in ["bool"]: dtype = np.bool
else:
Here is code that trigger the problem:
import numpy
import pycuda.autoinit
import pycuda.gpuarray
data = numpy.asarray(numpy.random.rand(5,5)*50, dtype='int8')
data_gpu = pycuda.gpuarray.to_gpu(data)
data_gpu*2
thanks
Frédéric Bastien
_______________________________________________
PyCUDA mailing list
[email protected]
http://lists.tiker.net/listinfo/pycuda
