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>

// CUDA launcher declared in cuda source
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 <cuda.h>
#include <cuda_runtime.h>
#include <stdint.h>

__device__ __forceinline__ float gelu_fast_scalar(float x) {
    // Exact GELU is 0.5*x*(1+erf(x/sqrt(2))).  This is a high-accuracy
    // erf approximation (max abs error ~1.5e-7), easily within fp32 tolerance
    // while avoiding the slower libdevice erff path.
    const float inv_sqrt2 = 0.70710678118654752440f;
    float z = x * inv_sqrt2;
    float s = 1.0f;
    if (z < 0.0f) { s = -1.0f; z = -z; }
    const float p  = 0.3275911f;
    const float a1 = 0.254829592f;
    const float a2 = -0.284496736f;
    const float a3 = 1.421413741f;
    const float a4 = -1.453152027f;
    const float a5 = 1.061405429f;
    float t = 1.0f / (1.0f + p * z);
    float poly = (((((a5 * t + a4) * t) + a3) * t + a2) * t + a1) * t;
    float erfv = s * (1.0f - poly * __expf(-z * z));
    return 0.5f * x * (1.0f + erfv);
}

__global__ __launch_bounds__(256, 4) void gelu_float4_kernel(const float* __restrict__ x,
                                                            float* __restrict__ y,
                                                            int64_t n4) {
    int64_t idx = (int64_t)blockIdx.x * blockDim.x + threadIdx.x;
    int64_t stride = (int64_t)blockDim.x * gridDim.x;
    const float4* __restrict__ x4 = reinterpret_cast<const float4*>(x);
    float4* __restrict__ y4 = reinterpret_cast<float4*>(y);
    for (int64_t i = idx; i < n4; i += stride) {
        float4 v = x4[i];
        v.x = gelu_fast_scalar(v.x);
        v.y = gelu_fast_scalar(v.y);
        v.z = gelu_fast_scalar(v.z);
        v.w = gelu_fast_scalar(v.w);
        y4[i] = v;
    }
}

__global__ __launch_bounds__(256, 4) void gelu_scalar_kernel(const float* __restrict__ x,
                                                            float* __restrict__ y,
                                                            int64_t n) {
    int64_t idx = (int64_t)blockIdx.x * blockDim.x + threadIdx.x;
    int64_t stride = (int64_t)blockDim.x * gridDim.x;
    for (int64_t i = idx; i < n; i += stride) {
        y[i] = gelu_fast_scalar(x[i]);
    }
}

__global__ void gelu_tail_kernel(const float* __restrict__ x, float* __restrict__ y,
                                 int64_t start, int64_t n) {
    int i = blockIdx.x * blockDim.x + threadIdx.x;
    int64_t idx = start + i;
    if (idx < n) y[idx] = gelu_fast_scalar(x[idx]);
}

void gelu_forward_cuda(torch::Tensor x, torch::Tensor y) {
    const int threads = 256;
    int64_t n = x.numel();
    if (n == 0) return;
    const float* xp = x.data_ptr<float>();
    float* yp = y.data_ptr<float>();
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();

    // The benchmark input is contiguous and CUDA allocations are sufficiently aligned.
    // Use vectorized loads/stores whenever the data pointers and length permit it.
    uintptr_t xa = reinterpret_cast<uintptr_t>(xp);
    uintptr_t ya = reinterpret_cast<uintptr_t>(yp);
    if (((xa | ya) & 15) == 0 && n >= 4) {
        int64_t n4 = n / 4;
        int blocks = (int)((n4 + threads - 1) / threads);
        if (blocks > 65535) blocks = 65535;
        gelu_float4_kernel<<<blocks, threads, 0, stream>>>(xp, yp, n4);
        int64_t done = n4 * 4;
        if (done < n) {
            gelu_tail_kernel<<<1, 32, 0, stream>>>(xp, yp, done, n);
        }
    } else {
        int blocks = (int)((n + threads - 1) / threads);
        if (blocks > 65535) blocks = 65535;
        gelu_scalar_kernel<<<blocks, threads, 0, stream>>>(xp, yp, n);
    }
}
"""

_ext = load_inline(
    name="gelu_fast_ext_v1",
    cpp_sources=[_cpp_src],
    cuda_sources=[_cuda_src],
    functions=["gelu_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.gelu_forward(x)