import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Simple model that performs L1 normalization.
    """
    def __init__(self):
        """
        Initializes the L1 normalization layer.
        """
        super(Model, self).__init__()

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Applies L1 normalization to the input tensor.

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

        Returns:
            torch.Tensor: Output tensor with L1 normalization applied, same shape as input.
        """
        return x / torch.mean(torch.abs(x), 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 l1norm_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) {
    #pragma unroll
    for (int offset = 16; offset > 0; offset >>= 1) val += __shfl_down_sync(0xffffffff, val, offset);
    return val;
}

__inline__ __device__ float block_reduce_sum(float val) {
    static __shared__ float shared[32];
    int lane = threadIdx.x & 31;
    int wid = threadIdx.x >> 5;
    val = warp_reduce_sum(val);
    if (lane == 0) shared[wid] = val;
    __syncthreads();
    val = (threadIdx.x < (blockDim.x >> 5)) ? shared[lane] : 0.0f;
    if (wid == 0) val = warp_reduce_sum(val);
    return val;
}

__global__ __launch_bounds__(256, 3)
void l1norm_inplace_kernel(float* __restrict__ x, int cols) {
    int row = blockIdx.x;
    long long base = (long long)row * (long long)cols;
    float sum = 0.0f;
    const int BD = 256;
    int col = threadIdx.x;
    for (; col + 7*BD < cols; col += 8*BD) {
        sum += fabsf(x[base + col]);
        sum += fabsf(x[base + col + BD]);
        sum += fabsf(x[base + col + 2*BD]);
        sum += fabsf(x[base + col + 3*BD]);
        sum += fabsf(x[base + col + 4*BD]);
        sum += fabsf(x[base + col + 5*BD]);
        sum += fabsf(x[base + col + 6*BD]);
        sum += fabsf(x[base + col + 7*BD]);
    }
    for (; col < cols; col += BD) sum += fabsf(x[base + col]);
    sum = block_reduce_sum(sum);
    __shared__ float scale_s;
    if (threadIdx.x == 0) scale_s = ((float)cols) / sum;
    __syncthreads();
    float scale = scale_s;
    col = threadIdx.x;
    for (; col + 7*BD < cols; col += 8*BD) {
        x[base + col] = x[base + col] * scale;
        x[base + col + BD] = x[base + col + BD] * scale;
        x[base + col + 2*BD] = x[base + col + 2*BD] * scale;
        x[base + col + 3*BD] = x[base + col + 3*BD] * scale;
        x[base + col + 4*BD] = x[base + col + 4*BD] * scale;
        x[base + col + 5*BD] = x[base + col + 5*BD] * scale;
        x[base + col + 6*BD] = x[base + col + 6*BD] * scale;
        x[base + col + 7*BD] = x[base + col + 7*BD] * scale;
    }
    for (; col < cols; col += BD) x[base + col] = x[base + col] * scale;
}

torch::Tensor l1norm_forward(torch::Tensor x) {
    int rows = (int)x.size(0);
    int cols = (int)x.size(1);
    cudaStream_t stream = at::cuda::getCurrentCUDAStream();
    l1norm_inplace_kernel<<<rows, 256, 0, stream>>>(x.data_ptr<float>(), cols);
    return x;
}
'''

_ext = load_inline(name='l1norm_ext_final', cpp_sources=_cpp_src, cuda_sources=_cuda_src, functions=['l1norm_forward'], extra_cuda_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.l1norm_forward(x)