import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs L2 normalization.
    """
    def __init__(self):
        """
        Initializes the L2Norm layer.

        Args:
            dim (int): Dimension along which to normalize.
        """
        super(Model, self).__init__()

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Applies L2 normalization to the input tensor.

        Args:
            x (torch.Tensor): Input tensor of shape (*, dim, *).

        Returns:
            torch.Tensor: Output tensor with L2 normalization applied, same shape as input.
        """
        return x / torch.norm(x, p=2, dim=1, keepdim=True)

batch_size = 32768
# choose dim so total <2^31
dim = 65535

def get_inputs():
    x = torch.rand(batch_size, dim)
    return [x]

def get_init_inputs():
    return []

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

_cpp_src = r"""
#include <torch/extension.h>
torch::Tensor l2norm_forward(torch::Tensor x);
"""

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

__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;
}

template<int BLOCK>
__global__ void l2norm_rows_vec4_inplace_kernel(float* __restrict__ x, int rows, int cols) {
    int row = blockIdx.x;
    float* rowx = x + (long long)row * cols;

    int start = (32 - (int)(((long long)row * (long long)cols) & 31LL)) & 31;
    if (start >= cols) start = 0;
    int n_aligned = cols - start;
    int vecN = n_aligned >> 2;
    int tail0 = start + (vecN << 2);

    float sum = 0.0f;
    if (threadIdx.x < start) {
        float v = rowx[threadIdx.x];
        sum = fmaf(v, v, sum);
    }
    const float4* __restrict__ row4 = reinterpret_cast<const float4*>(rowx + start);
    for (int i = threadIdx.x; i < vecN; i += BLOCK) {
        float4 a = row4[i];
        sum = fmaf(a.x, a.x, sum);
        sum = fmaf(a.y, a.y, sum);
        sum = fmaf(a.z, a.z, sum);
        sum = fmaf(a.w, a.w, sum);
    }
    if (threadIdx.x < (cols - tail0)) {
        float v = rowx[tail0 + threadIdx.x];
        sum = fmaf(v, v, sum);
    }

    sum = warp_reduce_sum(sum);
    __shared__ float warp_sums[32];
    int lane = threadIdx.x & 31;
    int warp = threadIdx.x >> 5;
    if (lane == 0) warp_sums[warp] = sum;
    __syncthreads();

    float block_sum = 0.0f;
    if (threadIdx.x < (BLOCK >> 5)) block_sum = warp_sums[lane];
    if (warp == 0) block_sum = warp_reduce_sum(block_sum);

    __shared__ float inv_shared;
    if (threadIdx.x == 0) inv_shared = rsqrtf(block_sum);
    __syncthreads();
    float inv = inv_shared;

    if (threadIdx.x < start) rowx[threadIdx.x] = rowx[threadIdx.x] * inv;
    float4* __restrict__ out4 = reinterpret_cast<float4*>(rowx + start);
    for (int i = threadIdx.x; i < vecN; i += BLOCK) {
        float4 a = out4[i];
        a.x *= inv;
        a.y *= inv;
        a.z *= inv;
        a.w *= inv;
        out4[i] = a;
    }
    if (threadIdx.x < (cols - tail0)) rowx[tail0 + threadIdx.x] = rowx[tail0 + threadIdx.x] * inv;
}

torch::Tensor l2norm_forward(torch::Tensor x) {
    TORCH_CHECK(x.is_cuda(), "x must be CUDA");
    TORCH_CHECK(x.scalar_type() == torch::kFloat32, "x must be float32");
    TORCH_CHECK(x.dim() == 2, "x must be 2D");
    auto xc = x.contiguous();
    int rows = (int)xc.size(0);
    int cols = (int)xc.size(1);
    constexpr int BLOCK = 1024;
    l2norm_rows_vec4_inplace_kernel<BLOCK><<<rows, BLOCK, 0, at::cuda::getCurrentCUDAStream()>>>(xc.data_ptr<float>(), rows, cols);
    return xc;
}
"""

_ext = load_inline(
    name="l2norm_ext_final",
    cpp_sources=_cpp_src,
    cuda_sources=_cuda_src,
    functions=["l2norm_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.l2norm_forward(x)