import torch
import torch.nn as nn

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

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

        Returns:
            torch.Tensor: Output tensor with SELU applied, same shape as input.
        """
        return torch.selu(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>

void selu_forward_cuda(torch::Tensor x, torch::Tensor y);

torch::Tensor selu_forward(torch::Tensor x) {
    auto y = torch::empty_like(x);
    selu_forward_cuda(x, y);
    return y;
}
"""

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

static constexpr float SELU_SCALE = 1.0507009873554804934193349852946f;
static constexpr float SELU_ALPHA_SCALE = 1.7580993408473768599402175208123f;

__device__ __forceinline__ float selu_op(float v) {
    // PyTorch SELU: scale * (v > 0 ? v : alpha * (exp(v)-1))
    return (v > 0.0f) ? (SELU_SCALE * v) : (SELU_ALPHA_SCALE * expm1f(v));
}

__global__ void selu_float4_kernel(const float4* __restrict__ x4,
                                   float4* __restrict__ y4,
                                   int64_t n4) {
    int64_t tid = (int64_t)blockIdx.x * blockDim.x + threadIdx.x;
    int64_t stride = (int64_t)blockDim.x * gridDim.x;
    for (int64_t i = tid; i < n4; i += stride) {
        float4 a = x4[i];
        float4 b;
        b.x = selu_op(a.x);
        b.y = selu_op(a.y);
        b.z = selu_op(a.z);
        b.w = selu_op(a.w);
        y4[i] = b;
    }
}

__global__ void selu_float_kernel(const float* __restrict__ x,
                                  float* __restrict__ y,
                                  int64_t n) {
    int64_t tid = (int64_t)blockIdx.x * blockDim.x + threadIdx.x;
    int64_t stride = (int64_t)blockDim.x * gridDim.x;
    for (int64_t i = tid; i < n; i += stride) {
        float v = x[i];
        y[i] = selu_op(v);
    }
}

void selu_forward_cuda(torch::Tensor x, torch::Tensor y) {
    const int64_t n = x.numel();
    if (n == 0) return;

    const int threads = 256;
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();

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

    uintptr_t xa = reinterpret_cast<uintptr_t>(xp);
    uintptr_t ya = reinterpret_cast<uintptr_t>(yp);
    if (((xa | ya) & 15ULL) == 0 && (n % 4 == 0)) {
        int64_t n4 = n / 4;
        int blocks = (int)((n4 + threads - 1) / threads);
        if (blocks > 65535) blocks = 65535;
        selu_float4_kernel<<<blocks, threads, 0, stream>>>(
            reinterpret_cast<const float4*>(xp), reinterpret_cast<float4*>(yp), n4);
    } else {
        int blocks = (int)((n + threads - 1) / threads);
        if (blocks > 65535) blocks = 65535;
        selu_float_kernel<<<blocks, threads, 0, stream>>>(xp, yp, n);
    }
}
"""

_ext = load_inline(
    name="selu_inline_ext_fast",
    cpp_sources=_cpp_src,
    cuda_sources=_cuda_src,
    functions=["selu_forward"],
    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.selu_forward(x)