[SYCL] alloca in a kernel

CUDA C++ supports dynamic memory allocation using either the built-in function malloc or using the operator new. However, allocations by malloc and new contribute to significant runtime performance overhead due to dynamic allocation on the heap.

In CUDA 11.3, CUDA C++ introduces support for using the memory allocator alloca in device code as a preview feature. Unlike malloc and new, the built-in function alloca allocates memory on the current thread’s stack, offering a faster and more convenient way to allocate small chunks of memory dynamically. This is especially useful when the size of an allocation is not known in advance at compile time.

When memory is allocated using alloca, the stack pointer of the thread’s stack is moved based on the requested memory allocation size to reserve or otherwise allocate the memory. The memory allocated is aligned at a 16-byte boundary, making possible accesses using all basic types, including vector types, without alignment constraints.

https://developer.nvidia.com/blog/programming-efficiently-with-the-cuda-11-3-compiler-toolchain/