import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs 1D Average Pooling.
    """
    def __init__(self, kernel_size: int, stride: int = 1, padding: int = 0):
        """
        Initializes the 1D Average Pooling layer.

        Args:
            kernel_size (int): Size of the pooling window.
            stride (int, optional): Stride of the pooling operation. Defaults to 1.
            padding (int, optional): Padding applied to the input tensor. Defaults to 0.
        """
        super(Model, self).__init__()
        self.avg_pool = nn.AvgPool1d(kernel_size=kernel_size, stride=stride, padding=padding)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Applies 1D Average Pooling to the input tensor.

        Args:
            x (torch.Tensor): Input tensor of shape (batch_size, in_channels, input_length).

        Returns:
            torch.Tensor: Output tensor with 1D Average Pooling applied, shape (batch_size, in_channels, output_length).
        """
        return self.avg_pool(x)

batch_size = 64
in_channels = 128
input_length = 65536
kernel_size = 8
stride = 1
padding = 4

def get_inputs():
    x = torch.rand(batch_size, in_channels, input_length)
    return [x]

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

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

_cpp_src = r"""
#include <torch/extension.h>
torch::Tensor avgpool1d_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>

static constexpr int BLOCK = 256;
static constexpr int TILE = 8192;
static constexpr int K = 8;
static constexpr int P = 4;
static constexpr int Q = TILE / BLOCK;

__global__ __launch_bounds__(BLOCK, 4)
void avgpool1d_k8s1p4_kernel(const float* __restrict__ x, float* __restrict__ y,
                              int rows, int L, int Lout) {
    int tile = blockIdx.x;
    int row = blockIdx.y;
    int tid = threadIdx.x;
    int out0 = tile * TILE;
    int tile_count = Lout - out0;
    if (tile_count > TILE) tile_count = TILE;
    if (tile_count <= 0) return;

    extern __shared__ float sh[];
    int load_count = tile_count + K - 1;
    int start = out0 - P;
    const float* rowx = x + ((long long)row) * L;
    if (start >= 0 && start + load_count <= L) {
        for (int i = tid; i < load_count; i += BLOCK) {
            sh[i] = rowx[start + i];
        }
    } else {
        for (int i = tid; i < load_count; i += BLOCK) {
            int xi = start + i;
            float v = 0.0f;
            if ((unsigned)xi < (unsigned)L) v = rowx[xi];
            sh[i] = v;
        }
    }
    __syncthreads();

    float* rowy = y + ((long long)row) * Lout;
    for (int q = 0; q < Q; ++q) {
        int j = q * BLOCK + tid;
        if (j < tile_count) {
            float s = sh[j] + sh[j+1] + sh[j+2] + sh[j+3]
                    + sh[j+4] + sh[j+5] + sh[j+6] + sh[j+7];
            rowy[out0 + j] = s * 0.125f;
        }
    }
}

__global__ void avgpool1d_generic_kernel(const float* __restrict__ x, float* __restrict__ y,
                                         int rows, int L, int Lout, int kernel_size,
                                         int stride, int padding) {
    long long total = (long long)rows * Lout;
    long long idx = (long long)blockIdx.x * blockDim.x + threadIdx.x;
    long long step = (long long)blockDim.x * gridDim.x;
    for (; idx < total; idx += step) {
        int o = (int)(idx % Lout);
        int row = (int)(idx / Lout);
        int start = o * stride - padding;
        float sum = 0.0f;
        const float* rowx = x + ((long long)row) * L;
        for (int k = 0; k < kernel_size; ++k) {
            int xi = start + k;
            if ((unsigned)xi < (unsigned)L) sum += rowx[xi];
        }
        y[idx] = sum / (float)kernel_size;
    }
}

torch::Tensor avgpool1d_forward(torch::Tensor x, int64_t kernel_size, int64_t stride, int64_t padding) {
    const int64_t N = x.size(0);
    const int64_t C = x.size(1);
    const int64_t L64 = x.size(2);
    const int64_t Lout64 = (L64 + 2 * padding - kernel_size) / stride + 1;
    auto y = torch::empty({N, C, Lout64}, x.options());
    int rows = (int)(N * C);
    int L = (int)L64;
    int Lout = (int)Lout64;
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();
    if (x.scalar_type() == at::kFloat && kernel_size == 8 && stride == 1 && padding == 4) {
        dim3 block(BLOCK);
        dim3 grid((Lout + TILE - 1) / TILE, rows);
        size_t smem = (TILE + K - 1) * sizeof(float);
        avgpool1d_k8s1p4_kernel<<<grid, block, smem, stream>>>(
            x.data_ptr<float>(), y.data_ptr<float>(), rows, L, Lout);
    } else {
        int threads = 256;
        long long total = (long long)rows * Lout;
        int blocks = (int)((total + threads - 1) / threads);
        if (blocks > 65535) blocks = 65535;
        avgpool1d_generic_kernel<<<blocks, threads, 0, stream>>>(
            x.data_ptr<float>(), y.data_ptr<float>(), rows, L, Lout, (int)kernel_size, (int)stride, (int)padding);
    }
    return y;
}
"""

_ext = load_inline(
    name="avgpool1d_ext_final",
    cpp_sources=_cpp_src,
    cuda_sources=_cuda_src,
    functions=["avgpool1d_forward"],
    extra_cflags=["-O3"],
    extra_cuda_cflags=["-O3", "--use_fast_math"],
    verbose=False,
)

class ModelNew(nn.Module):
    def __init__(self, kernel_size: int, stride: int = 1, padding: int = 0):
        super(ModelNew, self).__init__()
        self.kernel_size = int(kernel_size)
        self.stride = int(stride)
        self.padding = int(padding)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return _ext.avgpool1d_forward(x, self.kernel_size, self.stride, self.padding)