import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs Max Pooling 2D.
    """
    def __init__(self, kernel_size: int, stride: int, padding: int, dilation: int):
        """
        Initializes the Max Pooling 2D layer.

        Args:
            kernel_size (int): Size of the pooling window.
            stride (int): Stride of the pooling window.
            padding (int): Padding to be applied before pooling.
            dilation (int): Spacing between kernel elements.
        """
        super(Model, self).__init__()
        self.maxpool = nn.MaxPool2d(kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation)

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

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

        Returns:
            torch.Tensor: Output tensor after Max Pooling 2D, shape (batch_size, channels, pooled_height, pooled_width).
        """
        return self.maxpool(x)

batch_size = 32
channels = 64
height = 512
width = 512
kernel_size = 4
stride = 1
padding = 1
dilation = 1

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

def get_init_inputs():
    return [kernel_size, stride, padding, dilation]

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

_cpp_src = r"""
#include <torch/extension.h>
torch::Tensor maxpool2d_forward(torch::Tensor x);
"""

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

#define PITCH 35

__global__ __launch_bounds__(1024, 1)
void maxpool4x4s1p1_sep_kernel(const float* __restrict__ x, float* __restrict__ y) {
    __shared__ float smem[35 * 35];
    __shared__ float hmax[35 * 32];
    const int bx = blockIdx.x;
    const int by = blockIdx.y;
    const int ox0 = bx << 5;
    const int oy0 = by << 5;
    const int plane = blockIdx.z;
    const int tx = threadIdx.x;
    const int ty = threadIdx.y;
    const int tid = (ty << 5) + tx;
    const int in_base_y = oy0 - 1;
    const int in_base_x = ox0 - 1;
    const int plane_in_off = plane << 18;

    if (bx != 0 && bx != 15 && by != 0 && by != 15) {
        for (int idx = tid; idx < 1225; idx += 1024) {
            int sy = idx / PITCH;
            int sx = idx - sy * PITCH;
            smem[idx] = x[plane_in_off + ((in_base_y + sy) << 9) + (in_base_x + sx)];
        }
    } else {
        for (int idx = tid; idx < 1225; idx += 1024) {
            int sy = idx / PITCH;
            int sx = idx - sy * PITCH;
            int iy = in_base_y + sy;
            int ix = in_base_x + sx;
            float v = -FLT_MAX;
            if ((unsigned)iy < 512u && (unsigned)ix < 512u) {
                v = x[plane_in_off + (iy << 9) + ix];
            }
            smem[idx] = v;
        }
    }
    __syncthreads();

    for (int idx = tid; idx < 1120; idx += 1024) {
        int sy = idx >> 5;
        int sx = idx & 31;
        int s = sy * PITCH + sx;
        float m = smem[s];
        float v = smem[s + 1]; m = v > m ? v : m;
        v = smem[s + 2]; m = v > m ? v : m;
        v = smem[s + 3]; m = v > m ? v : m;
        hmax[idx] = m;
    }
    __syncthreads();

    const int ox = ox0 + tx;
    const int oy = oy0 + ty;
    if (ox < 511 && oy < 511) {
        int h = (ty << 5) + tx;
        float m = hmax[h];
        float v = hmax[h + 32]; m = v > m ? v : m;
        v = hmax[h + 64]; m = v > m ? v : m;
        v = hmax[h + 96]; m = v > m ? v : m;
        y[((plane * 511 + oy) * 511) + ox] = m;
    }
}

torch::Tensor maxpool2d_forward(torch::Tensor x) {
    auto y = torch::empty({x.size(0), x.size(1), 511, 511}, x.options());
    int planes = (int)(x.size(0) * x.size(1));
    dim3 block(32, 32);
    dim3 grid(16, 16, planes);
    maxpool4x4s1p1_sep_kernel<<<grid, block, 0, at::cuda::getCurrentCUDAStream()>>>(x.data_ptr<float>(), y.data_ptr<float>());
    return y;
}
"""

_ext = load_inline(name="maxpool2d_4x4_s1_p1_final",
                   cpp_sources=_cpp_src,
                   cuda_sources=_cuda_src,
                   functions=["maxpool2d_forward"],
                   extra_cuda_cflags=["-O3"],
                   verbose=False)

class ModelNew(nn.Module):
    def __init__(self, kernel_size: int, stride: int, padding: int, dilation: int):
        super(ModelNew, self).__init__()
        self.kernel_size = kernel_size
        self.stride = stride
        self.padding = padding
        self.dilation = dilation
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return _ext.maxpool2d_forward(x)