submission 66694

#!POPCORN leaderboard vectorsum_v2

import torch
import triton
import triton.language as tl
from task import input_t, output_t


@triton.jit
def sum_kernel(x_ptr, partial_sums_ptr, n_elements, BLOCK_SIZE: tl.constexpr):
    pid = tl.program_id(0)
    block_start = pid * BLOCK_SIZE
    offsets = block_start + tl.arange(0, BLOCK_SIZE)
    mask = offsets < n_elements

    x = tl.load(x_ptr + offsets, mask=mask, other=0.0)
    block_sum = tl.sum(x, axis=0)
    tl.store(partial_sums_ptr + pid, block_sum)


def custom_kernel(data: input_t) -> output_t:
    input, output = data
    n_elements = input.numel()

    BLOCK_SIZE = 4096

    n_blocks = triton.cdiv(n_elements, BLOCK_SIZE)
    partial_sums = torch.empty(n_blocks, device=input.device, dtype=input.dtype)
    sum_kernel[(n_blocks,)](input, partial_sums, n_elements, BLOCK_SIZE=BLOCK_SIZE)

    # Keep reducing on GPU while there are many elements left
    # Stop when small enough that CPU reduction is faster
    while partial_sums.numel() > BLOCK_SIZE:
        n_elements = partial_sums.numel()
        n_blocks = triton.cdiv(n_elements, BLOCK_SIZE)
        next_level = torch.empty(n_blocks, device=input.device, dtype=input.dtype)
        sum_kernel[(n_blocks,)](partial_sums, next_level, n_elements, BLOCK_SIZE=BLOCK_SIZE)
        partial_sums = next_level

    # Final small sum on CPU is faster than launching another tiny kernel
    return partial_sums.sum()