import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Performs a 3D transposed convolution operation with asymmetric input and square kernel.
    The input is padded before the convolution.

    Args:
        in_channels (int): Number of channels in the input tensor.
        out_channels (int): Number of channels produced by the convolution.
        kernel_size (int): Size of the square convolution kernel.
        stride (int, optional): Stride of the convolution. Defaults to 1.
        padding (int, optional): Padding applied to the input. Defaults to 0.
        groups (int, optional): Number of blocked connections from input channels to output channels. Defaults to 1.
        bias (bool, optional): If `True`, adds a learnable bias to the output. Defaults to `False`.
    """
    def __init__(self, in_channels: int, out_channels: int, kernel_size: int, stride: int = 1, padding: int = 0, output_padding: int = 0, groups: int = 1, bias: bool = False):
        super(Model, self).__init__()
        self.conv_transpose3d = nn.ConvTranspose3d(in_channels, out_channels, kernel_size=(kernel_size, kernel_size, kernel_size), stride=stride, padding=padding, groups=groups, bias=bias)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Performs the 3D transposed convolution.

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

        Returns:
            torch.Tensor: Output tensor of shape (batch_size, out_channels, depth_out, height_out, width_out).
        """
        return self.conv_transpose3d(x)

# Test code
batch_size = 4
in_channels = 32
out_channels = 32
kernel_size = 3
depth = 32
height = 64
width = 128
stride = 2
padding = 1
groups = 4

def get_inputs():
    x = torch.rand(batch_size, in_channels, depth, height, width)
    return [x]

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

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

_cpp_src = r'''
#include <torch/extension.h>

torch::Tensor convt3d_forward(torch::Tensor x, torch::Tensor weight, torch::Tensor bias,
                              int64_t stride, int64_t padding, int64_t groups);
'''

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

__global__ void convt3d_s2p1k3_kernel(const float* __restrict__ x,
                                      const float* __restrict__ w,
                                      const float* __restrict__ b,
                                      float* __restrict__ y,
                                      int N, int Cin, int D, int H, int W,
                                      int Cout, int outD, int outH, int outW,
                                      int groups, int has_bias) {
    int ow0 = threadIdx.x;
    int row = blockIdx.x;        // od*outH + oh
    int noc = blockIdx.y;        // n*Cout + oc
    int oh = row % outH;
    int od = row / outH;
    int oc = noc % Cout;
    int n  = noc / Cout;

    int outCpg = Cout / groups;
    int inCpg  = Cin / groups;
    int g = oc / outCpg;
    int ocg = oc - g * outCpg;
    int ic0 = g * inCpg;

    // For stride=2, padding=1, kernel=3, each output coordinate receives
    // one tap on even coordinates and two taps on odd coordinates.
    int nd = (od & 1) ? 2 : 1;
    int d_id0 = (od & 1) ? ((od + 1) >> 1) : (od >> 1);
    int d_k0  = (od & 1) ? 0 : 1;
    int d_id1 = (od - 1) >> 1;

    int nh = (oh & 1) ? 2 : 1;
    int h_id0 = (oh & 1) ? ((oh + 1) >> 1) : (oh >> 1);
    int h_k0  = (oh & 1) ? 0 : 1;
    int h_id1 = (oh - 1) >> 1;

    int DHW = D * H * W;
    int HW = H * W;
    int in_n_base = n * Cin * DHW;
    int out_base = ((n * Cout + oc) * outD + od) * outH * outW + oh * outW;
    int w_chan_stride = outCpg * 27;
    int w_oc_base = ocg * 27;

    for (int ow = ow0; ow < outW; ow += blockDim.x) {
        int nw = (ow & 1) ? 2 : 1;
        int w_id0 = (ow & 1) ? ((ow + 1) >> 1) : (ow >> 1);
        int w_k0  = (ow & 1) ? 0 : 1;
        int w_id1 = (ow - 1) >> 1;

        float acc = has_bias ? b[oc] : 0.0f;

        #pragma unroll
        for (int dd = 0; dd < 2; ++dd) {
            if (dd >= nd) continue;
            int id = (dd == 0) ? d_id0 : d_id1;
            int kd = (dd == 0) ? d_k0  : 2;
            #pragma unroll
            for (int hh = 0; hh < 2; ++hh) {
                if (hh >= nh) continue;
                int ih = (hh == 0) ? h_id0 : h_id1;
                int kh = (hh == 0) ? h_k0  : 2;
                #pragma unroll
                for (int ww = 0; ww < 2; ++ww) {
                    if (ww >= nw) continue;
                    int iw = (ww == 0) ? w_id0 : w_id1;
                    int kw = (ww == 0) ? w_k0  : 2;
                    int kidx = (kd * 3 + kh) * 3 + kw;
                    int x_sp_base = in_n_base + id * HW + ih * W + iw;
                    int w_base = w_oc_base + kidx;

                    for (int ci = 0; ci < inCpg; ++ci) {
                        int ic = ic0 + ci;
                        acc = fmaf(x[x_sp_base + ic * DHW], w[ic * w_chan_stride + w_base], acc);
                    }
                }
            }
        }
        y[out_base + ow] = acc;
    }
}

torch::Tensor convt3d_forward(torch::Tensor x, torch::Tensor weight, torch::Tensor bias,
                              int64_t stride, int64_t padding, int64_t groups) {
    const c10::cuda::CUDAGuard device_guard(x.device());
    int N = (int)x.size(0);
    int Cin = (int)x.size(1);
    int D = (int)x.size(2);
    int H = (int)x.size(3);
    int W = (int)x.size(4);
    int Cout = (int)(weight.size(1) * groups);
    int K = (int)weight.size(2);

    int outD = (D - 1) * (int)stride - 2 * (int)padding + K;
    int outH = (H - 1) * (int)stride - 2 * (int)padding + K;
    int outW = (W - 1) * (int)stride - 2 * (int)padding + K;

    auto y = torch::empty({N, Cout, outD, outH, outW}, x.options());

    // The benchmark case is float32, K=3, stride=2, padding=1.
    dim3 block(256);
    dim3 grid((unsigned)(outD * outH), (unsigned)(N * Cout));
    int has_bias = (bias.numel() != 0);
    convt3d_s2p1k3_kernel<<<grid, block, 0, at::cuda::getCurrentCUDAStream()>>>(
        x.data_ptr<float>(), weight.data_ptr<float>(), has_bias ? bias.data_ptr<float>() : nullptr,
        y.data_ptr<float>(), N, Cin, D, H, W, Cout, outD, outH, outW, (int)groups, has_bias);
    return y;
}
'''

_ext = load_inline(
    name='convt3d_s2p1k3_ext_v1',
    cpp_sources=_cpp_src,
    cuda_sources=_cuda_src,
    functions=['convt3d_forward'],
    extra_cflags=['-O3'],
    extra_cuda_cflags=['-O3', '--use_fast_math'],
    with_cuda=True,
    verbose=False,
)

class ModelNew(nn.Module):
    def __init__(self, in_channels: int, out_channels: int, kernel_size: int, stride: int = 1, padding: int = 0, output_padding: int = 0, groups: int = 1, bias: bool = False):
        super(ModelNew, self).__init__()
        # Match the reference exactly: output_padding is accepted but is not
        # forwarded to nn.ConvTranspose3d in the provided Model.
        self.conv_transpose3d = nn.ConvTranspose3d(
            in_channels, out_channels,
            kernel_size=(kernel_size, kernel_size, kernel_size),
            stride=stride, padding=padding, groups=groups, bias=bias
        )
        self._stride = stride
        self._padding = padding
        self._groups = groups
        self.register_buffer('_empty_bias', torch.empty(0), persistent=False)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        b = self.conv_transpose3d.bias
        if b is None:
            b = self._empty_bias
        return _ext.convt3d_forward(x, self.conv_transpose3d.weight, b, self._stride, self._padding, self.conv_transpose3d.groups)