import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs 2D Average Pooling.
    """
    def __init__(self, kernel_size: int, stride: int = None, padding: int = 0):
        """
        Initializes the Average Pooling layer.

        Args:
            kernel_size (int): Size of the pooling window.
            stride (int, optional): Stride of the pooling operation. Defaults to None (same as kernel_size).
            padding (int, optional): Padding applied to the input tensor. Defaults to 0.
        """
        super(Model, self).__init__()
        self.avg_pool = nn.AvgPool2d(kernel_size=kernel_size, stride=stride, padding=padding)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Applies 2D Average Pooling to the input tensor.

        Args:
            x (torch.Tensor): Input tensor of shape (batch_size, channels, height, width).

        Returns:
            torch.Tensor: Output tensor with Average Pooling applied.
        """
        return self.avg_pool(x)

batch_size = 16
channels = 64
height = 2048
width = 2048
kernel_size = 11

def get_inputs():
    x = torch.rand(batch_size, channels, height, width)
    return [x]

def get_init_inputs():
    return [kernel_size]

import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline

_cpp_src = r'''
#include <torch/extension.h>
torch::Tensor avgpool2d_forward(torch::Tensor x, int64_t kernel_size, int64_t stride, int64_t padding);
'''

_cuda_src = r'''
#include <torch/extension.h>
#include <ATen/cuda/CUDAContext.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <cstdint>

#define CHECK_CUDA(x) TORCH_CHECK(x.is_cuda(), #x " must be a CUDA tensor")
#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
#define CHECK_FLOAT(x) TORCH_CHECK(x.scalar_type() == at::kFloat, #x " must be float32")

__global__ void avgpool11_warp4_grid_kernel(const float* __restrict__ x, float* __restrict__ y,
                                            int64_t C, int64_t H, int64_t W,
                                            int64_t outH, int64_t outW, int64_t groupsW) {
    extern __shared__ float smem[];
    int lane = threadIdx.x & 31;
    int warp_in_block = threadIdx.x >> 5;
    int warps_per_block = blockDim.x >> 5;
    int64_t gw = (int64_t)blockIdx.x * warps_per_block + warp_in_block;
    if (gw >= groupsW) return;
    int64_t oh = blockIdx.y;
    int64_t nc = blockIdx.z;
    int64_t n = nc / C;
    int64_t c = nc - n * C;

    int64_t ow_base = gw * 4;
    int64_t in_col0 = ow_base * 11;
    int valid_cols = (int)((outW - ow_base) * 11);
    if (valid_cols > 44) valid_cols = 44;

    float* ws = smem + warp_in_block * 44;
    int64_t base = ((n * C + c) * H + oh * 11) * W + in_col0;

    int col = lane;
    float sum = 0.0f;
    if (col < valid_cols) {
#pragma unroll
        for (int ky = 0; ky < 11; ++ky) sum += x[base + (int64_t)ky * W + col];
    }
    ws[col] = sum;

    if (lane < 12) {
        col = 32 + lane;
        sum = 0.0f;
        if (col < valid_cols) {
#pragma unroll
            for (int ky = 0; ky < 11; ++ky) sum += x[base + (int64_t)ky * W + col];
        }
        ws[col] = sum;
    }

    __syncwarp();

    if (lane < 4) {
        int64_t ow = ow_base + lane;
        if (ow < outW) {
            float s = 0.0f;
            int off = lane * 11;
#pragma unroll
            for (int i = 0; i < 11; ++i) s += ws[off + i];
            y[((nc) * outH + oh) * outW + ow] = s * (1.0f / 121.0f);
        }
    }
}

__global__ void avgpool_generic_kernel(const float* __restrict__ x, float* __restrict__ y,
                                       int64_t total, int64_t C, int64_t H, int64_t W,
                                       int64_t outH, int64_t outW,
                                       int k, int stride, int pad) {
    int64_t idx = (int64_t)blockIdx.x * blockDim.x + threadIdx.x;
    if (idx >= total) return;
    int64_t ow = idx % outW;
    int64_t t = idx / outW;
    int64_t oh = t % outH;
    t /= outH;
    int64_t c = t % C;
    int64_t n = t / C;
    int64_t hstart = oh * (int64_t)stride - pad;
    int64_t wstart = ow * (int64_t)stride - pad;
    float s = 0.0f;
    for (int ky = 0; ky < k; ++ky) {
        int64_t ih = hstart + ky;
        if ((uint64_t)ih < (uint64_t)H) {
            int64_t row = ((n * C + c) * H + ih) * W;
            for (int kx = 0; kx < k; ++kx) {
                int64_t iw = wstart + kx;
                if ((uint64_t)iw < (uint64_t)W) s += x[row + iw];
            }
        }
    }
    y[idx] = s / (float)(k * k);
}

torch::Tensor avgpool2d_forward(torch::Tensor x, int64_t kernel_size, int64_t stride, int64_t padding) {
    CHECK_CUDA(x); CHECK_CONTIGUOUS(x); CHECK_FLOAT(x);
    TORCH_CHECK(x.dim() == 4, "input must be NCHW");
    int64_t N = x.size(0), C = x.size(1), H = x.size(2), W = x.size(3);
    int64_t outH = (H + 2 * padding - kernel_size) / stride + 1;
    int64_t outW = (W + 2 * padding - kernel_size) / stride + 1;
    auto y = torch::empty({N, C, outH, outW}, x.options());
    if (N == 0 || C == 0 || outH <= 0 || outW <= 0) return y;
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();
    const float* xp = x.data_ptr<float>();
    float* yp = y.data_ptr<float>();
    if (kernel_size == 11 && stride == 11 && padding == 0) {
        const int threads = 512;
        const int warps = threads / 32;
        int64_t groupsW = (outW + 3) / 4;
        dim3 grid((unsigned int)((groupsW + warps - 1) / warps), (unsigned int)outH, (unsigned int)(N * C));
        avgpool11_warp4_grid_kernel<<<grid, threads, warps * 44 * sizeof(float), stream>>>(xp, yp, C, H, W, outH, outW, groupsW);
    } else {
        int64_t total = N * C * outH * outW;
        const int threads = 256;
        int blocks = (int)((total + threads - 1) / threads);
        avgpool_generic_kernel<<<blocks, threads, 0, stream>>>(xp, yp, total, C, H, W, outH, outW, (int)kernel_size, (int)stride, (int)padding);
    }
    return y;
}
'''

_ext = load_inline(
    name='avgpool2d_ext_final',
    cpp_sources=_cpp_src,
    cuda_sources=_cuda_src,
    functions=['avgpool2d_forward'],
    extra_cuda_cflags=['-O3', '--use_fast_math'],
    extra_cflags=['-O3'],
    with_cuda=True,
    verbose=False,
)

class ModelNew(nn.Module):
    def __init__(self, kernel_size: int, stride: int = None, padding: int = 0):
        super(ModelNew, self).__init__()
        self.kernel_size = int(kernel_size)
        self.stride = int(kernel_size if stride is None else stride)
        self.padding = int(padding)
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return _ext.avgpool2d_forward(x, self.kernel_size, self.stride, self.padding)