CUDA thrust device pointer with transform copy crash - cuda

Question

In CUDA 9.2 I have something like this:
#ifdef __CUDA_ARCH__
struct Context { float n[4]; } context;
#else
typedef __m128 Context;
#endif
struct A { float k[2]; };
struct B { float q[4]; };
struct FTransform : thrust::unary_function<A, B>
{
const Context context;
FTransform(Context context) : context(context){}
__device__ __host__ B operator()(const A& a) const
{
B b{{a.k[0], a.k[1], a.k[0]*context.n[0], a.k[1]*context.n[1]}};
return b;
}
};
void DoThrust(B* _bs, const Context& context, A* _as, uint32_t count)
{
thrust::device_ptr<B> bs = thrust::device_pointer_cast(_bs);
thrust::device_ptr<A> as = thrust::device_pointer_cast(_as);
FTransform fTransform(context);
auto first = thrust::make_transform_iterator(as, fTransform);
auto last = thrust::make_transform_iterator(as + count, fTransform);
thrust::copy(first, last, bs);
}
int main(int c, char **argv)
{
const uint32_t Count = 4;
Context context;
A* as;
B* bs;
cudaMalloc(&as, Count*sizeof(A));
cudaMalloc(&bs, Count*sizeof(B));
A hostAs[Count];
cudaMemcpy(as, hostAs, Count * sizeof(A), cudaMemcpyHostToDevice);
DoThrust(bs, context, as, Count);
B hostBs[Count];
cudaMemcpy(hostBs, bs, Count * sizeof(B), cudaMemcpyDeviceToHost);//crash
return 0;
}
Then when I call a standard cudaMemcpy() call later on the results I get the exception "an illegal memory access was encountered".
If I replace the thrust code with a non-thrust equivalent there is no error and everything works fine. Various combinations of trying to copy to device_vectors etc I get different crashes that seem to be thrust trying to release the device_ptr's for some reason - so maybe it is here for some reason?
== UPDATE ==
Ok that was confusing it appears it's due to the functor FTransform context member variable in my actual more complicated case. This specifically:
struct FTransform : thrust::unary_function<A, B>
{
#ifdef __CUDA_ARCH__
struct Context { float v[4]; } context;
#else
__m128 context;
#endif
...
};
So I guess it's an alignment problem somehow => in fact it is, as this works:
#ifdef __CUDA_ARCH__
struct __align__(16) Context { float v[4]; } context;
#else
__m128 context;
#endif

Answer 1

The solution is to ensure that if you use aligned types in thrust functor members (such as __m128 SSE types) that are copied to the GPU, that they are defined as aligned both during NVCC's CPU and GPU code build passes - and not accidentally assume even if a type may seem to naturally align to it's equivalent in the other pass that it will be ok, as otherwise bad hard to understand things may happen.
So for example the _ align _(16) is necessary in code like this:
struct FTransform : thrust::unary_function<A, B>
{
#ifdef __CUDA_ARCH__
struct __align__(16) Context { float v[4]; } context;
#else
__m128 context;
#endif
FTransform(Context context) : context(context){}
__device__ __host__ B operator()(const A& a) const; // function makes use of context
};