cuda::ptr_rebind

Defined in the header <cuda/memory>.

namespace cuda {

template <typename U, typename T>
[[nodiscard]] __host__ __device__
U* ptr_rebind(T* ptr) noexcept;

template <typename U, typename T>
[[nodiscard]] __host__ __device__
const U* ptr_rebind(const T* ptr) noexcept;

template <typename U, typename T>
[[nodiscard]] __host__ __device__
volatile U* ptr_rebind(volatile T* ptr) noexcept;

template <typename U, typename T>
[[nodiscard]] __host__ __device__
const volatile U* ptr_rebind(const volatile T* ptr) noexcept;

} // namespace cuda

The functions return the pointer ptr cast to type U* or const U*. They are shorter and safer alternative to reinterpret_cast.

Parameters

• ptr: The pointer.

Return value

• The pointer cast to type U* or const U*.

Constraints

• ptr must be aligned to alignof(U) and alignof(T).

Performance considerations

• The returned pointer is decorated with __builtin_assume_aligned with the alignof(U) value to help the compiler generate better code.

• The returned pointer maintains the same memory space, for example shared memory, as the input pointer.

Example

#include <cuda/memory>
#include <cuda/std/cstdint>

__global__ void kernel(const int* ptr, volatile int* ptr2) {
    auto ptr_res1 = cuda::ptr_rebind<uint64_t>(ptr);  // ptr_res1: const uint64_t*
    auto ptr_res2 = cuda::ptr_rebind<uint64_t>(ptr2); // ptr_res2: volatile uint64_t*
}

int main() {
    int* ptr;
    cudaMalloc(&ptr, 100 * sizeof(int));
    kernel<<<1, 1>>>(ptr);
    cudaDeviceSynchronize();
    return 0;
}

See it on Godbolt 🔗