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>
#define CHECK_CUDA(x) TORCH_CHECK(x.is_cuda(), #x " must be a CUDA tensor")

template<int WARPS_PER_BLOCK>
__global__ __launch_bounds__(WARPS_PER_BLOCK*32, 4)
void avgpool_k11_pairwarp_kernel(const float* __restrict__ x, float* __restrict__ y,
                                  int total_pairs, int OH, int OW, int H, int W) {
    int warp_in_block = threadIdx.x >> 5;
    int lane = threadIdx.x & 31;
    int pair_idx = (int)blockIdx.x * WARPS_PER_BLOCK + warp_in_block;
    if (pair_idx >= total_pairs) return;
    int pairs_per_row = (OW + 1) >> 1;
    int pair_ow = pair_idx % pairs_per_row;
    int t = pair_idx / pairs_per_row;
    int oh = t % OH;
    int nc = t / OH;
    int ow0 = pair_ow << 1;
    const float* base = x + ((long long)nc * H + oh * 11) * W + ow0 * 11;
    float s = 0.0f;
    if (lane < 22 && (ow0 + (lane >= 11)) < OW) {
#pragma unroll
        for (int r = 0; r < 11; ++r) {
            s += base[r * W + lane];
        }
    }
    unsigned mask0 = 0x000007ffu;
    if (lane < 11) {
        float v;
        v = __shfl_down_sync(mask0, s, 8); if (lane + 8 < 11) s += v;
        v = __shfl_down_sync(mask0, s, 4); if (lane + 4 < 11) s += v;
        v = __shfl_down_sync(mask0, s, 2); if (lane + 2 < 11) s += v;
        v = __shfl_down_sync(mask0, s, 1); if (lane + 1 < 11) s += v;
        if (lane == 0) y[((long long)nc * OH + oh) * OW + ow0] = s * 0.0082644628099173553719f;
    } else if (lane >= 11 && lane < 22 && ow0 + 1 < OW) {
        unsigned mask1 = 0x003ff800u;
        float v;
        v = __shfl_down_sync(mask1, s, 8); if (lane + 8 < 22) s += v;
        v = __shfl_down_sync(mask1, s, 4); if (lane + 4 < 22) s += v;
        v = __shfl_down_sync(mask1, s, 2); if (lane + 2 < 22) s += v;
        v = __shfl_down_sync(mask1, s, 1); if (lane + 1 < 22) s += v;
        if (lane == 11) y[((long long)nc * OH + oh) * OW + ow0 + 1] = s * 0.0082644628099173553719f;
    }
}

__global__ void avgpool_generic_kernel(const float* __restrict__ x, float* __restrict__ y,
                                       int total_out, int C, int H, int W, int OH, int OW,
                                       int K, int S, int P) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    if (idx >= total_out) return;
    int ow = idx % OW;
    int tmp = idx / OW;
    int oh = tmp % OH;
    tmp /= OH;
    int c = tmp % C;
    int n = tmp / C;
    int ih0 = oh * S - P;
    int iw0 = ow * S - P;
    float sum = 0.0f;
    for (int kh = 0; kh < K; ++kh) {
        int ih = ih0 + kh;
        if ((unsigned)ih < (unsigned)H) {
            const float* row = x + (((long long)n * C + c) * H + ih) * W;
            for (int kw = 0; kw < K; ++kw) {
                int iw = iw0 + kw;
                if ((unsigned)iw < (unsigned)W) sum += row[iw];
            }
        }
    }
    y[idx] = sum / (float)(K * K);
}

torch::Tensor avgpool2d_forward(torch::Tensor x, int64_t kernel_size, int64_t stride, int64_t padding) {
    CHECK_CUDA(x);
    TORCH_CHECK(x.scalar_type() == at::kFloat, "only float32 supported");
    TORCH_CHECK(x.dim() == 4, "input must be NCHW");
    auto xc = x.contiguous();
    int N = (int)xc.size(0), C = (int)xc.size(1), H = (int)xc.size(2), W = (int)xc.size(3);
    int K = (int)kernel_size, S = (int)stride, P = (int)padding;
    int OH = (H + 2 * P - K) / S + 1;
    int OW = (W + 2 * P - K) / S + 1;
    auto y = torch::empty({N, C, OH, OW}, xc.options());
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();
    int total = N*C*OH*OW;
    if (K == 11 && S == 11 && P == 0) {
        constexpr int WARPS = 8;
        int total_pairs = N * C * OH * ((OW + 1) >> 1);
        int blocks = (total_pairs + WARPS - 1) / WARPS;
        avgpool_k11_pairwarp_kernel<WARPS><<<blocks, WARPS*32, 0, stream>>>(xc.data_ptr<float>(), y.data_ptr<float>(), total_pairs, OH, OW, H, W);
    } else {
        int threads=256, blocks=(total+threads-1)/threads;
        avgpool_generic_kernel<<<blocks, threads, 0, stream>>>(xc.data_ptr<float>(), y.data_ptr<float>(), total, C, H, W, OH, OW, K, S, P);
    }
    return y;
}
'''
_ext = load_inline(name='avgpool2d_ext_final_pair', cpp_sources=_cpp_src, cuda_sources=_cuda_src, functions=['avgpool2d_forward'], extra_cuda_cflags=['-O3','--use_fast_math'], extra_cflags=['-O3'], 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)