Use of undeclared identifier, when summing two axes

[Baekalfen]

The code below crashes Bohrium. Both sums seem to be required for the crash to occur.

dim = 16
c = np.arange(dim**3, dtype=np.float32).reshape((dim,dim,dim))
k3 = np.sum(c, axis=2)
k4 = np.sum(c, axis=1)

The kernel is simply missing the array in global memory, it tries to write to.

#pragma OPENCL EXTENSION cl_khr_fp64 : enable
#include <kernel_dependencies/complex_opencl.h>
#include <kernel_dependencies/integer_operations.h>

__kernel void execute_9775374476049381469(ulong vo0, ulong vs0_0, ulong vs0_1, ulong vo1, ulong vs1_0, ulong vs1_1, ulong vs1_2, const long c1, const long c2, const float c3, const float c4) {
    // The IDs of the threaded blocks:
    const uint g0 = get_global_id(0); if (g0 >= 16) { return; } // Prevent overflow

    {const ulong i0 = g0;
        for (ulong i1 = 0; i1 < 16; ++i1) {
            const ulong idx0 = (vo0 +i0*vs0_0 +i1*vs0_1);
            a0[idx0] = c4;
        }
        for (ulong i1 = 0; i1 < 16; ++i1) {
            float t1;
            t1 = c3;
            for (ulong i2 = 0; i2 < 16; ++i2) {
                uint t2;
                float t3;
                float t0;
                t2 = (vo1 +i0*vs1_0 +i1*vs1_1 +i2*vs1_2);
                t3 = t2;
                t1 += t3;
                t0 += t3;
            }
        }
    }
}

[Baekalfen]

Here is a simpler example. The dimensions and axis seems to play a role.

dim = 50
a = np.arange(dim**4, dtype=np.float32).reshape((dim,dim,dim,dim))
k = np.sum(a, axis=1)

[omegahm]

These work fine on my MacBook with Bohrium compiled from master. As a side effect, I print k or k3 and k4. Is this still an issue?

[Baekalfen]

I don't have access to my test-machine again before monday. But If I remember correct, the problem disappears when you the print the variables, as this will allocate the arrays in global memory.

[omegahm]

Then indeed mine also crash when executing

import bohrium as np
dim = 50
a = np.arange(dim**4, dtype=np.float32).reshape((dim,dim,dim,dim))
k = np.sum(a, axis=1)