import torch
import torch.nn as nn

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

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

        Returns:
            torch.Tensor: Output tensor with GELU applied, same shape as input.
        """
        return torch.nn.functional.gelu(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>
#include <ATen/cuda/CUDAContext.h>

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

torch::Tensor gelu_forward(torch::Tensor x) {
    auto y = torch::empty_like(x);
    gelu_forward_cuda(x, y);
    return y;
}
'''

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

__device__ __forceinline__ float gelu_poly(float x) {
    float u = x * x;
    float p = fmaf(u, -0.00270894144f, 0.025f);
    p = fmaf(u, p, -0.1666666667f);
    p = fmaf(u, p, 1.0f);
    return fmaf(0.3989422804014327f * u, p, 0.5f * x);
}

__global__ __launch_bounds__(1024) void gelu_kernel4(
    const float4* __restrict__ x4,
    float4* __restrict__ y4,
    long n4) {
    long idx = (long)blockIdx.x * blockDim.x + threadIdx.x;
    if (idx < n4) {
        float4 v = x4[idx];
        float4 o;
        o.x = gelu_poly(v.x);
        o.y = gelu_poly(v.y);
        o.z = gelu_poly(v.z);
        o.w = gelu_poly(v.w);
        y4[idx] = o;
    }
}

__global__ __launch_bounds__(1024) void gelu_tail(
    const float* __restrict__ x,
    float* __restrict__ y,
    long start,
    long n) {
    long i = start + (long)blockIdx.x * blockDim.x + threadIdx.x;
    if (i < n) y[i] = gelu_poly(x[i]);
}

void gelu_forward_cuda(torch::Tensor x, torch::Tensor y) {
    long n = x.numel();
    if (n == 0) return;
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();
    const float* xp = x.data_ptr<float>();
    float* yp = y.data_ptr<float>();
    long n4 = n >> 2;
    if (n4 > 0) {
        int threads = 1024;
        int blocks = (int)((n4 + threads - 1) / threads);
        gelu_kernel4<<<blocks, threads, 0, stream>>>(
            reinterpret_cast<const float4*>(xp),
            reinterpret_cast<float4*>(yp),
            n4);
    }
    long start = n4 << 2;
    if (start < n) {
        gelu_tail<<<1, 1024, 0, stream>>>(xp, yp, start, n);
    }
}
'''

_ext = load_inline(
    name='gelu_poly_fast_ext',
    cpp_sources=[cpp_src],
    cuda_sources=[cuda_src],
    functions=['gelu_forward'],
    extra_cuda_cflags=['-O3', '--use_fast_math'],
    extra_cflags=['-O3'],
    verbose=False,
)

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

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