Thanks for the answer, it is a pity that you it is not possible to use this
functions, especially since it also seems not possible to use the cublas
functions in the source modules. In order to be able to use the gpu array
functions in a large loop one has to prevent the copy to the main memory. To
take from the simple speed test example:
##################
# GPUArray SECTION
# The result is copied back to main memory on each iteration, this is a
bottleneck
....
....
for i in range(n_iter):
a_gpu = pycuda.cumath.sin(a_gpu)
end.record() # end timing
.....
....
would it be possible to not copy the result to the main memory (I also do not
see why the result needs to be copied back to the main memory, it looks more
logic to me only to copy when you ask for it).
kind regards
Fil
17.02.2015, 17:45, "Andreas Kloeckner" <[email protected]>:
> Fil Peters <[email protected]> writes:
>> Hello,
>>
>> I am just new to pycuda and started testing it. I was wondering if it is
>> possible to use the gpuarray functions in a sourcemodule.
>>
>> For example, I was trying to covert the following code into a pycuda
>> sourcemodule:
>>
>> numpy code:
>>
>> fac1=np.float32(0.5)
>> fac2=np.float32(1.001)
>> for i in range(niter):
>> c = a
>> d = b
>> a = (c + d)*fac1
>> b = c*fac2
>> f += a*b + a*a
>> k = np.dot(a,b)
>>
>> Until the line "f += a*b + a*a" it works well:
>>
>> mod = SourceModule("""
>>
>> __global__ void vecmul(float *dest, float *in1, float *in2, float *in3,
>> float *in4, int niter)
>> {
>>
>> const int i = blockDim.x*blockIdx.x + threadIdx.x;
>>
>> for(int n = 0; n < niter; n++) {
>> in3[i] = in1[i];
>> in4[i] = in2[i];
>> in1[i] = (in3[i]+in4[i])*0.5 ;
>> in2[i] = in3[i]*1.001;
>> dest[i] += in1[i] * in2[i] + in1[i]*in1[i];
>> }
>> }
>> """)
>>
>> 'of course I realize that the dot product is not very useful in this loop,
>> but in my final program I will need to reuse this value in the loop).
>> So for this specific case the question is how to incorporate the
>> function gpuarray.dot() in the code, or if that is not possible how to
>> include a reduction kernel in the sourcemodule.
>
> No, you can't, sorry.
>
> Andreas
_______________________________________________
PyCUDA mailing list
[email protected]
http://lists.tiker.net/listinfo/pycuda
