import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs a HardSigmoid activation.
    """
    def __init__(self):
        super(Model, self).__init__()
    
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Applies HardSigmoid activation to the input tensor.

        Args:
            x (torch.Tensor): Input tensor of any shape.

        Returns:
            torch.Tensor: Output tensor with HardSigmoid applied, same shape as input.
        """
        return torch.nn.functional.hardsigmoid(x)

batch_size = 4096
dim = 393216

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

def get_init_inputs():
    return []  # No special initialization inputs needed

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

_cpp_src = '#include <torch/extension.h>\ntorch::Tensor hardsigmoid_cuda(torch::Tensor x);\n'

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

__device__ __forceinline__ float hs_f32(float v) {
    return __saturatef(fmaf(v, 0.16666666666666666f, 0.5f));
}

__device__ __forceinline__ void do4(const float4* __restrict__ x4,
                                    float4* __restrict__ y4,
                                    int i) {
    float4 v = x4[i];
    v.x = hs_f32(v.x);
    v.y = hs_f32(v.y);
    v.z = hs_f32(v.z);
    v.w = hs_f32(v.w);
    y4[i] = v;
}

__global__ void hs4(const float* __restrict__ x, float* __restrict__ y, int n4) {
    int tid = blockIdx.x * blockDim.x + threadIdx.x;
    int stride = blockDim.x * gridDim.x;
    const float4* __restrict__ x4 = (const float4*)x;
    float4* __restrict__ y4 = (float4*)y;
    int step = stride * 4;
    for (int i = tid; i < n4; i += step) {
        do4(x4, y4, i);
        int j = i + stride;
        if (j < n4) do4(x4, y4, j);
        j += stride;
        if (j < n4) do4(x4, y4, j);
        j += stride;
        if (j < n4) do4(x4, y4, j);
    }
}

__global__ void hs1(const float* __restrict__ x, float* __restrict__ y, long long n) {
    long long tid = (long long)blockIdx.x * blockDim.x + threadIdx.x;
    long long stride = (long long)blockDim.x * gridDim.x;
    for (long long i = tid; i < n; i += stride) {
        y[i] = hs_f32(x[i]);
    }
}

torch::Tensor hardsigmoid_cuda(torch::Tensor x) {
    auto y = torch::empty_like(x);
    long long n = x.numel();
    if (n == 0) return y;

    const int threads = 256;
    cudaStream_t s = at::cuda::getCurrentCUDAStream();
    const float* xp = x.data_ptr<float>();
    float* yp = y.data_ptr<float>();

    if ((((uintptr_t)xp | (uintptr_t)yp) & 15) == 0 && (n & 3) == 0 && n / 4 <= 2147483647LL) {
        int n4 = (int)(n / 4);
        int blocks = (n4 + threads - 1) / threads;
        if (blocks > 4096) blocks = 4096;
        hs4<<<blocks, threads, 0, s>>>(xp, yp, n4);
    } else {
        long long nb = (n + threads - 1) / threads;
        int blocks = (int)(nb > 4096 ? 4096 : nb);
        hs1<<<blocks, threads, 0, s>>>(xp, yp, n);
    }
    return y;
}
'''

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

class ModelNew(nn.Module):
    def __init__(self):
        super(ModelNew, self).__init__()

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return _ext.hardsigmoid_cuda(x)