import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs Instance Normalization.
    """
    def __init__(self, num_features: int):
        """
        Initializes the InstanceNorm layer.

        Args:
            num_features (int): Number of features in the input tensor.
        """
        super(Model, self).__init__()
        self.inorm = nn.InstanceNorm2d(num_features=num_features)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Applies Instance Normalization to the input tensor.

        Args:
            x (torch.Tensor): Input tensor of shape (batch_size, num_features, height, width).

        Returns:
            torch.Tensor: Output tensor with Instance Normalization applied, same shape as input.
        """
        return self.inorm(x)

batch_size = 112  # heavier workload
features = 64
dim1 = 512
dim2 = 512

def get_inputs():
    x = torch.rand(batch_size, features, dim1, dim2)
    return [x]

def get_init_inputs():
    return [features]

import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline

_cpp_src = r'''
#include <torch/extension.h>
torch::Tensor instance_norm_forward(torch::Tensor x);
'''

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

static __inline__ __device__ float warp_reduce_sum(float val) {
    unsigned mask = 0xffffffffu;
    val += __shfl_down_sync(mask, val, 16);
    val += __shfl_down_sync(mask, val, 8);
    val += __shfl_down_sync(mask, val, 4);
    val += __shfl_down_sync(mask, val, 2);
    val += __shfl_down_sync(mask, val, 1);
    return val;
}

__device__ __forceinline__ void acc4(float4 v, float &s, float &q) {
    s += v.x + v.y + v.z + v.w;
    q += v.x*v.x + v.y*v.y + v.z*v.z + v.w*v.w;
}

__device__ __forceinline__ float4 norm4(float4 v, float m, float r) {
    v.x = (v.x - m) * r;
    v.y = (v.y - m) * r;
    v.z = (v.z - m) * r;
    v.w = (v.w - m) * r;
    return v;
}

template<int BLOCK>
__global__ void inorm2d_kernel(const float* __restrict__ x, float* __restrict__ y,
                               int rows, int inner4) {
    int row = blockIdx.x;
    if (row >= rows) return;
    int tid = threadIdx.x;
    size_t base4 = (size_t)row * (size_t)inner4;
    const float4* __restrict__ x4 = reinterpret_cast<const float4*>(x) + base4;

    float sum = 0.0f;
    float sumsq = 0.0f;

    for (int i = tid; i < inner4; i += BLOCK) {
        float4 v = x4[i];
        acc4(v, sum, sumsq);
    }

    sum = warp_reduce_sum(sum);
    sumsq = warp_reduce_sum(sumsq);

    __shared__ float s_sum[32];
    __shared__ float s_sumsq[32];
    __shared__ float s_mean;
    __shared__ float s_invstd;

    int lane = tid & 31;
    int warp = tid >> 5;
    if (lane == 0) {
        s_sum[warp] = sum;
        s_sumsq[warp] = sumsq;
    }
    __syncthreads();

    if (warp == 0) {
        float tsum = s_sum[lane];
        float tsumsq = s_sumsq[lane];
        tsum = warp_reduce_sum(tsum);
        tsumsq = warp_reduce_sum(tsumsq);
        if (lane == 0) {
            float denom = 1.0f / (float)(inner4 * 4);
            float mean = tsum * denom;
            s_mean = mean;
            s_invstd = rsqrtf(tsumsq * denom - mean * mean + 1.0e-5f);
        }
    }
    __syncthreads();

    float mean = s_mean;
    float invstd = s_invstd;
    float4* __restrict__ y4 = reinterpret_cast<float4*>(y) + base4;
    for (int i = tid; i < inner4; i += BLOCK) {
        y4[i] = norm4(x4[i], mean, invstd);
    }
}

torch::Tensor instance_norm_forward(torch::Tensor x) {
    auto y = torch::empty_like(x);
    int rows = (int)(x.size(0) * x.size(1));
    int inner4 = (int)((x.size(2) * x.size(3)) >> 2);
    const int BLOCK = 1024;
    inorm2d_kernel<BLOCK><<<rows, BLOCK, 0, at::cuda::getCurrentCUDAStream()>>>(
        x.data_ptr<float>(), y.data_ptr<float>(), rows, inner4);
    C10_CUDA_KERNEL_LAUNCH_CHECK();
    return y;
}
'''

_ext = load_inline(
    name='inorm2d_ext_final',
    cpp_sources=[_cpp_src],
    cuda_sources=[_cuda_src],
    functions=['instance_norm_forward'],
    extra_cflags=['-O3'],
    extra_cuda_cflags=['-O3', '--use_fast_math'],
    verbose=False,
)

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

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