[Feature]: hip_vector_types accessor internal array

[NB4444]

Suggestion Description

In 3cd3b3f it seems that for one of the two implementations .data was removed from hip_vector_types and instead make use of get_native_vector(). Would it be possible to make it directly accessible with the [] operator, instead of either using .data or get_native_vector() depending on the implementation?

For example like:

HIP_vector_base<unsigned int, 2U> vector;

int index; # either 0 or 1

unsigned int val = vector[index];

Operating System

No response

GPU

No response

ROCm Component

No response

[ppanchad-amd]

Hi @NB4444. Internal ticket has been created to address this request. Thanks!

[AidanBeltonS]

Hi @NB4444,_ there has been a recent commit to amd-staging 0f46bd5 that adds support for the [] operator for HIP_vector_type.

Let me know if this resolves your issue

[Snektron]

@AidanBeltonS Looks like the compiler is having trouble compiling operator[]: its spilling the vector to shared memory. When using get_native_vector(x)[idx], the compiler is able to optimize the function into a serial check for the right index:

	v_cmp_eq_u32_e32 vcc, 1, v10
	v_cndmask_b32_e32 v11, v14, v15, vcc
	v_cmp_eq_u32_e32 vcc, 2, v10
	v_cndmask_b32_e32 v11, v11, v16, vcc
	v_cmp_eq_u32_e32 vcc, 3, v10
	v_cndmask_b32_e32 v10, v11, v17, vcc

but when using operator[], the function lowers to a spill instead:

	v_mov_b32_e32 v11, v36
	v_mov_b32_e32 v12, v15
	ds_write_b128 v66, v[10:13]
	v_lshl_add_u32 v11, v61, 2, v66
	ds_read_b32 v11, v11

The actual performance varies a bit because of that, but I think that in general spilling can be considered universally worse. At present I don't know if this can be fixed by changing the implementation of operator[] to simply call get_native_vector(), and I'm also not sure why the compiler cannot just optimize this properly...