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 k, 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>

// Specialized fast path for the benchmark case: k = stride = 11, padding = 0.
// One CTA computes one output row for one (N,C).  This keeps all threads in a
// block reading the same 11 input rows and avoids the index arithmetic of a
// fully flattened mapping.
__global__ void avgpool2d_k11s11p0_kernel(const float* __restrict__ x,
                                           float* __restrict__ y,
                                           int NC, int H, int W,
                                           int outH, int outW) {
    int row_id = blockIdx.x;                 // 0 .. NC*outH-1
    int ow = threadIdx.x;
    if (ow >= outW) return;

    int oh = row_id % outH;
    int nc = row_id / outH;
    int ih0 = oh * 11;
    int iw0 = ow * 11;

    const float* base = x + ((long long)nc * H + ih0) * W + iw0;
    float sum = 0.0f;

    #pragma unroll
    for (int kh = 0; kh < 11; ++kh) {
        const float* r = base + (long long)kh * W;
        // 11 scalar loads; unrolled to minimize loop overhead.  The pooling
        // windows are non-overlapping, so every value is used exactly once.
        sum += r[0];
        sum += r[1];
        sum += r[2];
        sum += r[3];
        sum += r[4];
        sum += r[5];
        sum += r[6];
        sum += r[7];
        sum += r[8];
        sum += r[9];
        sum += r[10];
    }
    y[(long long)row_id * outW + ow] = sum * (1.0f / 121.0f);
}

// Generic fp32 AvgPool2d, matching PyTorch defaults: ceil_mode=False,
// count_include_pad=True, divisor_override=None.
__global__ void avgpool2d_generic_kernel(const float* __restrict__ x,
                                         float* __restrict__ y,
                                         long long total,
                                         int C, int H, int W,
                                         int outH, int outW,
                                         int k, int stride, int padding) {
    long long idx = (long long)blockIdx.x * blockDim.x + threadIdx.x;
    if (idx >= total) return;

    int ow = (int)(idx % outW);
    long long t = idx / outW;
    int oh = (int)(t % outH);
    t /= outH;
    int c = (int)(t % C);
    int n = (int)(t / C);

    int ih_start = oh * stride - padding;
    int iw_start = ow * stride - padding;
    int ih_end = ih_start + k;
    int iw_end = iw_start + k;

    int vh0 = ih_start < 0 ? 0 : ih_start;
    int vw0 = iw_start < 0 ? 0 : iw_start;
    int vh1 = ih_end > H ? H : ih_end;
    int vw1 = iw_end > W ? W : iw_end;

    float sum = 0.0f;
    const float* xnc = x + ((long long)n * C + c) * H * W;
    for (int ih = vh0; ih < vh1; ++ih) {
        const float* row = xnc + (long long)ih * W;
        for (int iw = vw0; iw < vw1; ++iw) {
            sum += row[iw];
        }
    }
    y[idx] = sum / (float)(k * k);
}

torch::Tensor avgpool2d_forward(torch::Tensor x, int64_t k64, int64_t stride64, int64_t padding64) {
    const int64_t N64 = x.size(0);
    const int64_t C64 = x.size(1);
    const int64_t H64 = x.size(2);
    const int64_t W64 = x.size(3);
    const int k = (int)k64;
    const int stride = (int)stride64;
    const int padding = (int)padding64;

    const int64_t outH64 = (H64 + 2 * (int64_t)padding - k) / stride + 1;
    const int64_t outW64 = (W64 + 2 * (int64_t)padding - k) / stride + 1;
    auto y = torch::empty({N64, C64, outH64, outW64}, x.options());

    const float* xp = x.data_ptr<float>();
    float* yp = y.data_ptr<float>();
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();

    if (k == 11 && stride == 11 && padding == 0 && x.scalar_type() == at::kFloat && x.is_contiguous()) {
        int NC = (int)(N64 * C64);
        int H = (int)H64, W = (int)W64, outH = (int)outH64, outW = (int)outW64;
        dim3 block(256);
        dim3 grid((unsigned int)(NC * outH));
        avgpool2d_k11s11p0_kernel<<<grid, block, 0, stream>>>(xp, yp, NC, H, W, outH, outW);
    } else {
        long long total = N64 * C64 * outH64 * outW64;
        dim3 block(256);
        dim3 grid((unsigned int)((total + block.x - 1) / block.x));
        avgpool2d_generic_kernel<<<grid, block, 0, stream>>>(xp, yp, total, (int)C64, (int)H64, (int)W64,
                                                             (int)outH64, (int)outW64, k, stride, padding);
    }
    return y;
}
"""

_ext = load_inline(
    name="avgpool2d_k11_ext",
    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)