How to invoke a kernel which has a double pointer as argument? - cupy

Question

I'm trying to invoke a custom kernel with a double pointer as argument using CuPy.
Given a dataset of size n of d-dimensional points, I allocate an array of n floats for each dimension.
Then I put all the pointers to those arrays into another array. That's why I'm dealing with a double pointer.
My custom kernel has a signature like:
__global__ void myKernel(float **dataset, int n, int d, int* output, ...)
In native cuda code (compiled through nvcc) my code works as expected.
My python attempt to invoke the kernel looks like:
import cupy as cp
from sklearn.datasets import make_blobs
X, y = make_blobs(n_samples=1_000, n_features=2, cluster_std=.5, center_box=(- 10.0, 10.0), shuffle=True, random_state=None)
with open('path', 'r') as f:
code = f.read()
output = cp.empty(len(X), dtype=cp.int32)
kern = cp.RawKernel(code, 'myKernel')
d1 = cp.array(X[:, 0])
d2 = cp.array(X[:, 1])
blocks = ...
grid = ...
args = (cp.array([d1.data.ptr, d2.data.ptr]), 2, len(X), output, ...)
shared_mem = ...
kern((grid, 1, 1), (blocks, 1, 1), args=args, shared_mem=shared_mem)
cp.cuda.Stream.null.synchronize()
As you can probably spot, my issue is in the args = ... row.
The code does not raise errors, but the data in the output vector is definitely wrong.
Am I passing the data properly? Is there a better way?

Answer 1

Could you try creating a minimum reproducer? I see no problem passing a double-pointer argument.
import cupy as cp
import sys
code = '''
extern "C" __global__ void myKernel(float **dataset, int n, int* output) {
for (int i = 0; i < n; ++i) {
printf("[");
for (int j = 0; j < 10; ++j) {
printf("%f, ", dataset[i][j]);
}
printf("]\\n");
}
}
'''
kern = cp.RawKernel(code, 'myKernel')
d0 = cp.arange(10, dtype=cp.float32)
d1 = cp.arange(10, dtype=cp.float32) * 2
dataset = cp.array([d0.data.ptr, d1.data.ptr])
output = cp.empty(10, dtype=cp.int32)
def run():
print('--expected--')
print(d0)
print(d1)
print('--actual--')
sys.stdout.flush()
kern((1, 1, 1), (1, 1, 1), args=(dataset, len(dataset), output))
cp.cuda.Stream.null.synchronize()
print('--end--')
print()
run()
Output:
--expected--
[0. 1. 2. 3. 4. 5. 6. 7. 8. 9.]
[ 0. 2. 4. 6. 8. 10. 12. 14. 16. 18.]
--actual--
[0.000000, 1.000000, 2.000000, 3.000000, 4.000000, 5.000000, 6.000000, 7.000000, 8.000000, 9.000000, ]
[0.000000, 2.000000, 4.000000, 6.000000, 8.000000, 10.000000, 12.000000, 14.000000, 16.000000, 18.000000, ]
--end--