submission 170701

#!POPCORN leaderboard grayscale_v2

# This is a submission template for popcorn leaderboard 'grayscale_v2'.
# Your task is as follows:
# > Implement an RGB to grayscale conversion kernel that matches the reference implementation.
# > The kernel should convert square RGB images with even sizes to grayscale using the standard coefficients:
# > Y = 0.2989 R + 0.5870 G + 0.1140 B
# > 
# > Input: RGB tensor of shape (H, W, 3) with values in [0, 1]
# > Output: Grayscale tensor of shape (H, W) with values in [0, 1]
# The deadline for this leaderboard is 2025-12-30 00:00:00+00:00

# You can automatically route this file to specific GPUs by adding a line
# `#!POPCORN gpus <GPUs>` to the header of this file.
# Happy hacking!


import subprocess

subprocess.run(["pip", "install", "cuda-cccl[cu12]==0.4.3"])

subprocess.run(["pip", "install", "cupy-cuda12x"])

from task import input_t, output_t

import cuda.compute
from cuda.compute import gpu_struct
import cupy as cp
import numpy as np
import torch

@gpu_struct
class Pixel:
    r: np.float32
    g: np.float32
    b: np.float32

def as_grayscale(p: Pixel) -> np.float32:
    return (
        np.float32(0.2989) * p.r +
        np.float32(0.587) * p.g +
        np.float32(0.114) * p.b
    )

build_in = cp.empty(1, dtype=Pixel)
build_out = torch.empty(1, dtype=torch.float32, device="cuda")

transformer = cuda.compute.make_unary_transform(build_in, build_out, as_grayscale)

def custom_kernel(data: input_t) -> output_t:
    d_in, d_out = data
    size = len(d_in)
    transformer(d_in, d_out, size * size)

    return d_out