import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Model that performs a 3D transposed convolution, applies Softmax and Sigmoid.
    """
    def __init__(self, in_channels, out_channels, kernel_size, stride, padding, output_padding, bias=True):
        super(Model, self).__init__()
        self.conv_transpose = nn.ConvTranspose3d(in_channels, out_channels, kernel_size, stride=stride, padding=padding, output_padding=output_padding, bias=bias)
        self.softmax = nn.Softmax(dim=1)
        self.sigmoid = nn.Sigmoid()

    def forward(self, x):
        """
        Args:
            x (torch.Tensor): Input tensor of shape (batch_size, in_channels, D, H, W).

        Returns:
            torch.Tensor: Output tensor of shape (batch_size, out_channels, D, H, W).
        """
        x = self.conv_transpose(x)
        x = self.softmax(x)
        x = self.sigmoid(x)
        return x

batch_size = 16
in_channels = 32
out_channels = 64
D, H, W = 16, 32, 32
kernel_size = 3
stride = 2
padding = 1
output_padding = 1

def get_inputs():
    return [torch.rand(batch_size, in_channels, D, H, W)]

def get_init_inputs():
    return [in_channels, out_channels, kernel_size, stride, padding, output_padding]

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

cpp_src = 'torch::Tensor convt_softmax_sigmoid(torch::Tensor x, torch::Tensor w, c10::optional<torch::Tensor> b);'

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

__device__ __forceinline__ float warp_reduce_max(float v){
    #pragma unroll
    for(int offset=16; offset>0; offset/=2) v = fmaxf(v, __shfl_down_sync(0xffffffff, v, offset));
    return v;
}
__device__ __forceinline__ float warp_reduce_sum(float v){
    #pragma unroll
    for(int offset=16; offset>0; offset/=2) v += __shfl_down_sync(0xffffffff, v, offset);
    return v;
}
__device__ __forceinline__ int dim_indices(int o, int dim, int* ids, int* ks){
    if((o & 1) == 0){ ids[0] = o >> 1; ks[0] = 1; return 1; }
    int c = 0;
    int i0 = (o + 1) >> 1;
    if(i0 < dim){ ids[c] = i0; ks[c] = 0; c++; }
    int i1 = (o - 1) >> 1;
    if(i1 >= 0){ ids[c] = i1; ks[c] = 2; c++; }
    return c;
}

template<int WARPS>
__global__ void convt_softmax_sigmoid_kernel(const float* __restrict__ x, const float* __restrict__ w, const float* __restrict__ b, float* __restrict__ y,
                                            int N, int Di, int Hi, int Wi, int Do, int Ho, int Wo, int has_bias){
    int tid = threadIdx.x;
    int lane = tid & 31;
    int warp = tid >> 5;
    int spatial_o = Do * Ho * Wo;
    int col = blockIdx.x * WARPS + warp;
    if(col >= N * spatial_o) return;
    int n = col / spatial_o;
    int s = col - n * spatial_o;
    int ow = s % Wo;
    int tmp = s / Wo;
    int oh = tmp % Ho;
    int od = tmp / Ho;
    int dids[2], hids[2], wids[2], dks[2], hks[2], wks[2];
    int nd = dim_indices(od, Di, dids, dks);
    int nh = dim_indices(oh, Hi, hids, hks);
    int nw = dim_indices(ow, Wi, wids, wks);
    int co0 = lane, co1 = lane + 32;
    float acc0 = has_bias ? __ldg(b + co0) : 0.0f;
    float acc1 = has_bias ? __ldg(b + co1) : 0.0f;
    int spatial_i = Di * Hi * Wi;
    long long nbase = (long long)n * 32 * spatial_i;
    #pragma unroll
    for(int a=0; a<2; ++a){
      if(a >= nd) continue;
      int id = dids[a], kd = dks[a];
      #pragma unroll
      for(int bb=0; bb<2; ++bb){
        if(bb >= nh) continue;
        int ih = hids[bb], kh = hks[bb];
        #pragma unroll
        for(int c=0; c<2; ++c){
          if(c >= nw) continue;
          int iw = wids[c], kw = wks[c];
          long long in_base = nbase + ((long long)id * Hi + ih) * Wi + iw;
          int kidx = (kd * 3 + kh) * 3 + kw;
          #pragma unroll
          for(int ci=0; ci<32; ++ci){
            float xv = __ldg(x + in_base + (long long)ci * spatial_i);
            long long wb = (((long long)ci * 64 + co0) * 27 + kidx);
            acc0 = fmaf(xv, __ldg(w + wb), acc0);
            acc1 = fmaf(xv, __ldg(w + wb + (long long)32 * 27), acc1);
          }
        }
      }
    }
    float m = warp_reduce_max(fmaxf(acc0, acc1));
    m = __shfl_sync(0xffffffff, m, 0);
    float e0 = expf(acc0 - m);
    float e1 = expf(acc1 - m);
    float sum = warp_reduce_sum(e0 + e1);
    sum = __shfl_sync(0xffffffff, sum, 0);
    float p0 = e0 / sum;
    float p1 = e1 / sum;
    y[((long long)n * 64 + co0) * spatial_o + s] = 1.0f / (1.0f + expf(-p0));
    y[((long long)n * 64 + co1) * spatial_o + s] = 1.0f / (1.0f + expf(-p1));
}

torch::Tensor convt_softmax_sigmoid(torch::Tensor x, torch::Tensor w, c10::optional<torch::Tensor> bopt){
    int N = x.size(0), Di = x.size(2), Hi = x.size(3), Wi = x.size(4);
    int Do = Di * 2, Ho = Hi * 2, Wo = Wi * 2;
    auto y = torch::empty({N, 64, Do, Ho, Wo}, x.options());
    const float* bp = nullptr; int has_bias = 0;
    if(bopt.has_value() && bopt.value().defined()) { bp = bopt.value().data_ptr<float>(); has_bias = 1; }
    int cols = N * Do * Ho * Wo;
    constexpr int WARPS = 8;
    convt_softmax_sigmoid_kernel<WARPS><<<(cols + WARPS - 1)/WARPS, 32*WARPS, 0, at::cuda::getCurrentCUDAStream()>>>(x.data_ptr<float>(), w.data_ptr<float>(), bp, y.data_ptr<float>(), N, Di, Hi, Wi, Do, Ho, Wo, has_bias);
    return y;
}
'''

_ext = load_inline(name='convt_fused_ext_final_submit', cpp_sources=cpp_src, cuda_sources=cuda_src, functions=['convt_softmax_sigmoid'], verbose=False, extra_cuda_cflags=['-O3', '--use_fast_math'])

class ModelNew(nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size, stride, padding, output_padding, bias=True):
        super(ModelNew, self).__init__()
        self.conv_transpose = nn.ConvTranspose3d(in_channels, out_channels, kernel_size, stride=stride, padding=padding, output_padding=output_padding, bias=bias)
    def forward(self, x):
        return _ext.convt_softmax_sigmoid(x, self.conv_transpose.weight, self.conv_transpose.bias)