import torch
import torch.nn as nn

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

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

        Returns:
            torch.Tensor: Output tensor with ReLU applied, same shape as input.
        """
        return torch.relu(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

_cuda_src = r'''
#include <torch/extension.h>
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAException.h>
#include <stdint.h>

__device__ __forceinline__ float relu_f(float x) {
    return x > 0.0f ? x : 0.0f;
}

__global__ void relu_vec4_kernel(const float* __restrict__ x,
                                 float* __restrict__ y,
                                 int64_t n4,
                                 int tail) {
    int64_t tid = (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 = tid; i < n4; i += stride) {
        float4 v = x4[i];
        v.x = relu_f(v.x);
        v.y = relu_f(v.y);
        v.z = relu_f(v.z);
        v.w = relu_f(v.w);
        y4[i] = v;
    }

    if (tail && blockIdx.x == 0) {
        int t = threadIdx.x;
        if (t < tail) {
            int64_t idx = n4 * 4 + t;
            y[idx] = relu_f(x[idx]);
        }
    }
}

__global__ void relu_scalar_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) {
        y[i] = relu_f(x[i]);
    }
}

torch::Tensor relu_forward(torch::Tensor x) {
    TORCH_CHECK(x.is_cuda(), "x must be a CUDA tensor");
    TORCH_CHECK(x.scalar_type() == at::kFloat, "only float32 is supported");
    TORCH_CHECK(x.is_contiguous(), "x must be contiguous");

    auto y = torch::empty_like(x);
    const int64_t n = x.numel();
    if (n == 0) return y;

    const int threads = 256;
    const float* xp = x.data_ptr<float>();
    float* yp = y.data_ptr<float>();

    uintptr_t ax = reinterpret_cast<uintptr_t>(xp);
    uintptr_t ay = reinterpret_cast<uintptr_t>(yp);
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();

    if (((ax | ay) & 15) == 0) {
        int64_t n4 = n >> 2;
        int tail = (int)(n & 3);
        int64_t work = n4 > 0 ? n4 : 1;
        int blocks = (int)((work + threads - 1) / threads);
        if (blocks > 4096) blocks = 4096;
        relu_vec4_kernel<<<blocks, threads, 0, stream>>>(xp, yp, n4, tail);
    } else {
        int64_t work = n;
        int blocks = (int)((work + threads - 1) / threads);
        if (blocks > 4096) blocks = 4096;
        relu_scalar_kernel<<<blocks, threads, 0, stream>>>(xp, yp, n);
    }
    C10_CUDA_KERNEL_LAUNCH_CHECK();
    return y;
}
'''

_relu_ext = load_inline(
    name='relu_forward_ext_opt',
    cpp_sources='',
    cuda_sources=_cuda_src,
    functions=['relu_forward'],
    extra_cuda_cflags=['-O3'],
    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 _relu_ext.relu_forward(x)