import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Model that performs a batch matrix multiplication, instance normalization, summation, residual addition, and multiplication.
    """
    def __init__(self, in_features, out_features, eps=1e-5, momentum=0.1):
        super(Model, self).__init__()
        self.bmm = nn.Linear(in_features, out_features)
        self.instance_norm = nn.InstanceNorm2d(out_features, eps=eps, momentum=momentum)

    def forward(self, x, y):
        """
        Args:
            x (torch.Tensor): Input tensor of shape (batch_size, in_features).
            y (torch.Tensor): Input tensor of shape (batch_size, out_features).

        Returns:
            torch.Tensor: Output tensor of shape (batch_size, out_features).
        """
        x = self.bmm(x)
        x = self.instance_norm(x.unsqueeze(1).unsqueeze(1)).squeeze(1).squeeze(1)
        x = x + y
        x = x * y
        return x

batch_size = 1024  # Increased batch size
in_features = 8192  # Increased input features
out_features = 8192  # Increased output features

def get_inputs():
    return [torch.rand(batch_size, in_features), torch.rand(batch_size, out_features)]

def get_init_inputs():
    return [in_features, out_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 fused_forward(torch::Tensor x, torch::Tensor y, torch::Tensor weight, torch::Tensor bias, double eps);
'''

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

__inline__ __device__ void warp_reduce_pair(float &a, float &b) {
    #pragma unroll
    for (int offset = 16; offset > 0; offset >>= 1) {
        a += __shfl_down_sync(0xffffffff, a, offset);
        b += __shfl_down_sync(0xffffffff, b, offset);
    }
}

__inline__ __device__ void block_reduce_pair(float &a, float &b) {
    static __shared__ float shared_a[32];
    static __shared__ float shared_b[32];
    int lane = threadIdx.x & 31;
    int wid = threadIdx.x >> 5;
    warp_reduce_pair(a, b);
    if (lane == 0) {
        shared_a[wid] = a;
        shared_b[wid] = b;
    }
    __syncthreads();
    a = (threadIdx.x < (blockDim.x >> 5)) ? shared_a[lane] : 0.0f;
    b = (threadIdx.x < (blockDim.x >> 5)) ? shared_b[lane] : 0.0f;
    if (wid == 0) warp_reduce_pair(a, b);
}

__global__ void norm_residual_kernel(const float* __restrict__ z,
                                     const float* __restrict__ y,
                                     float* __restrict__ out,
                                     int B, int O, float eps) {
    int row = blockIdx.x;
    const float* zr = z + (long long)row * O;
    const float* yr = y + (long long)row * O;
    float* orow = out + (long long)row * O;

    float s = 0.0f;
    float ss = 0.0f;
    for (int i = threadIdx.x; i < O; i += blockDim.x) {
        float v = zr[i];
        s += v;
        ss += v * v;
    }
    block_reduce_pair(s, ss);
    __shared__ float sh_mean;
    __shared__ float sh_inv;
    if (threadIdx.x == 0) {
        float mean = s / (float)O;
        float var = ss / (float)O - mean * mean;
        if (var < 0.0f) var = 0.0f;
        sh_mean = mean;
        sh_inv = rsqrtf(var + eps);
    }
    __syncthreads();
    float mean = sh_mean;
    float inv = sh_inv;
    for (int i = threadIdx.x; i < O; i += blockDim.x) {
        float yy = yr[i];
        float n = (zr[i] - mean) * inv;
        orow[i] = (n + yy) * yy;
    }
}

torch::Tensor fused_forward(torch::Tensor x, torch::Tensor y, torch::Tensor weight, torch::Tensor bias, double eps) {
    torch::Tensor z = at::linear(x, weight, bias);
    auto out = torch::empty_like(y);
    int B = (int)y.size(0);
    int O = (int)y.size(1);
    norm_residual_kernel<<<B, 256, 0, at::cuda::getCurrentCUDAStream()>>>(
        z.data_ptr<float>(), y.data_ptr<float>(), out.data_ptr<float>(), B, O, (float)eps);
    return out;
}
'''

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

class ModelNew(nn.Module):
    def __init__(self, in_features, out_features, eps=1e-5, momentum=0.1):
        super(ModelNew, self).__init__()
        self.bmm = nn.Linear(in_features, out_features)
        self.instance_norm = nn.InstanceNorm2d(out_features, eps=eps, momentum=momentum)
        self.eps = eps

    def forward(self, x, y):
        return _ext.fused_forward(x, y, self.bmm.weight, self.bmm.bias, float(self.eps))