import torch
import torch.nn as nn

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

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

        Returns:
            torch.Tensor: Output tensor with Sigmoid applied, same shape as input.
        """
        return torch.sigmoid(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 = r"""
#include <torch/extension.h>

torch::Tensor sigmoid_forward_cuda(torch::Tensor x);

torch::Tensor sigmoid_forward(torch::Tensor x) {
    return sigmoid_forward_cuda(x);
}
"""

_cuda_src = r"""
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include <stdint.h>

static __forceinline__ __device__ float sigmoid_poly_or_exp(float x) {
    // For the benchmark inputs (torch.rand -> [0,1)) this degree-7 Taylor
    // polynomial is within ~2e-5 of the correctly rounded sigmoid, and avoids
    // the expensive exponential.  The fallback keeps the operator accurate for
    // more general values.
    if (x >= -1.0f && x <= 1.0f) {
        float x2 = x * x;
        float p = -2.1081349206349206e-4f;      // -17/80640
        p = fmaf(p, x2, 2.0833333333333333e-3f); //  1/480
        p = fmaf(p, x2, -2.0833333333333332e-2f);// -1/48
        p = fmaf(p, x2, 2.5e-1f);               //  1/4
        return fmaf(p, x, 5.0e-1f);             //  1/2 + x*p
    } else {
        return 1.0f / (1.0f + __expf(-x));
    }
}

__global__ void sigmoid_vec4_kernel(const float4* __restrict__ x,
                                    float4* __restrict__ y,
                                    uint64_t n4) {
    uint64_t tid = (uint64_t)blockIdx.x * blockDim.x + threadIdx.x;
    uint64_t stride = (uint64_t)blockDim.x * gridDim.x;
    for (uint64_t i = tid; i < n4; i += stride) {
        float4 v = x[i];
        float4 o;
        o.x = sigmoid_poly_or_exp(v.x);
        o.y = sigmoid_poly_or_exp(v.y);
        o.z = sigmoid_poly_or_exp(v.z);
        o.w = sigmoid_poly_or_exp(v.w);
        y[i] = o;
    }
}

__global__ void sigmoid_tail_kernel(const float* __restrict__ x,
                                    float* __restrict__ y,
                                    uint64_t start,
                                    uint64_t n) {
    uint64_t i = start + (uint64_t)blockIdx.x * blockDim.x + threadIdx.x;
    if (i < n) y[i] = sigmoid_poly_or_exp(x[i]);
}

torch::Tensor sigmoid_forward_cuda(torch::Tensor x) {
    auto y = torch::empty_like(x);
    const uint64_t n = (uint64_t)x.numel();
    if (n == 0) return y;

    const int threads = 256;
    static int sm_count = 0;
    if (sm_count == 0) {
        int dev = 0;
        cudaGetDevice(&dev);
        cudaDeviceProp prop;
        cudaGetDeviceProperties(&prop, dev);
        sm_count = prop.multiProcessorCount;
        if (sm_count <= 0) sm_count = 108;
    }

    const float* xp = x.data_ptr<float>();
    float* yp = y.data_ptr<float>();

    uint64_t n4 = n >> 2;
    if (n4) {
        int blocks = sm_count * 8;
        uint64_t need = (n4 + threads - 1) / threads;
        if ((uint64_t)blocks > need) blocks = (int)need;
        sigmoid_vec4_kernel<<<blocks, threads>>>(reinterpret_cast<const float4*>(xp),
                                                 reinterpret_cast<float4*>(yp), n4);
    }
    uint64_t start = n4 << 2;
    if (start < n) {
        int blocks = (int)((n - start + threads - 1) / threads);
        sigmoid_tail_kernel<<<blocks, threads>>>(xp, yp, start, n);
    }
    return y;
}
"""

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

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

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