import torch
import torch.nn as nn

# Default ModelNew is PyTorch-only. run_and_check: use check_kernel=False with backend=cuda
# until the candidate file contains a real CUDA kernel (see kernel_static_checker).


class Model(nn.Module):
    """
    Fused **2D bidirectional cross-scan** block for robotics **SSM-hybrid** backbones
    (e.g. Vision Mamba / VMamba-style encoders on overhead maps, costmaps, or
    multi-sensor tiles).

    One forward fuses four fixed scan permutations of ``(H, W)`` into 1D sequences,
    applies a **shared depthwise Conv1d** (lightweight stand-in for a selective SSM
    along the scan), maps each result back to the grid, **sums** the four
    directional contributions, and adds a **residual**.

    Scan order (same spirit as 2D cross-scan / S2D):
      LR — row-major left→right, rows top→bottom  
      RL — row-major right→left per row  
      TB — column-wise top→bottom, columns left→right  
      BT — flip vertical then same as TB  
    """

    def __init__(self, channels: int, kernel_size: int = 3):
        super().__init__()
        assert channels >= 1
        assert kernel_size % 2 == 1, "use odd kernel_size for same-length conv1d"
        self.channels = channels
        pad = kernel_size // 2
        self.dw_conv = nn.Conv1d(
            channels,
            channels,
            kernel_size,
            padding=pad,
            groups=channels,
            bias=True,
        )

    def _apply_scan_conv_unscan(self, x: torch.Tensor) -> torch.Tensor:
        """Shared depthwise conv along flattened 2D, then map back to (B,C,H,W)."""
        B, C, H, W = x.shape
        seq = x.flatten(2)
        y = self.dw_conv(seq)
        return y.view(B, C, H, W)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Args:
            x: (B, C, H, W) spatial features (map / image / fused tile).
        Returns:
            (B, C, H, W) after fused four-way cross-scan + residual.
        """
        B, C, H, W = x.shape
        assert C == self.channels

        acc = torch.zeros_like(x)

        # LR: row-major
        acc = acc + self._apply_scan_conv_unscan(x)

        # RL: flip W, scan, unflip
        xr = x.flip(dims=(-1,))
        acc = acc + self._apply_scan_conv_unscan(xr).flip(dims=(-1,))

        # TB: columns as sequences (permute to W,H, flatten)
        xt = x.permute(0, 1, 3, 2).contiguous()
        Bt, Ct, Wt, Ht = xt.shape
        seq_t = xt.flatten(2)
        yt = self.dw_conv(seq_t).view(Bt, Ct, Wt, Ht).permute(0, 1, 3, 2).contiguous()
        acc = acc + yt

        # BT: flip H then same TB path
        xb = x.flip(dims=(-2,))
        xbt = xb.permute(0, 1, 3, 2).contiguous()
        Bb, Cb, Wb, Hb = xbt.shape
        seq_b = xbt.flatten(2)
        yb = self.dw_conv(seq_b).view(Bb, Cb, Wb, Hb).permute(0, 1, 3, 2).contiguous().flip(
            dims=(-2,)
        )
        acc = acc + yb

        return x + acc


class ModelNew(nn.Module):
    """KernelBench candidate entry point; default delegates to ``Model``."""

    def __init__(self, channels: int, kernel_size: int = 3):
        super().__init__()
        self._impl = Model(channels, kernel_size=kernel_size)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        return self._impl(x)


batch_size = 4
channels = 64
height = 32
width = 32
kernel_size = 3


def get_inputs():
    p = popcorn_pri
    x = torch.randn(p.jitter_int(batch_size), channels, p.jitter_int(height), p.jitter_int(width))
    return [x]


def get_init_inputs():
    return [channels, kernel_size]

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

cuda_source = '''
#include <torch/extension.h>
#include <cuda.h>
#include <cuda_runtime.h>

__global__ void cross_scan_conv_kernel(
    const float* __restrict__ x,
    const float* __restrict__ weight,
    const float* __restrict__ bias,
    float* __restrict__ out,
    int B, int C, int H, int W, int K
) {
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    int total = B * C * H * W;
    if (idx >= total) return;

    int w = idx % W;
    int h = (idx / W) % H;
    int c = (idx / (W * H)) % C;
    int b = idx / (W * H * C);

    float sum = 0.0f;

    // LR: row-major flatten
    int pos_lr = h * W + w;
    int L = H * W;
    for (int k = 0; k < K; ++k) {
        int p = pos_lr + (k - K/2);
        if (p >= 0 && p < L) {
            int hh = p / W;
            int ww = p % W;
            sum += x[((b * C + c) * H + hh) * W + ww] * weight[c * K + k];
        }
    }
    sum += bias[c];

    // RL: flip W
    int pos_rl = h * W + (W - 1 - w);
    for (int k = 0; k < K; ++k) {
        int p = pos_rl + (k - K/2);
        if (p >= 0 && p < L) {
            int hh = p / W;
            int ww = W - 1 - (p % W);
            sum += x[((b * C + c) * H + hh) * W + ww] * weight[c * K + k];
        }
    }
    sum += bias[c];

    // TB: permute to (B,C,W,H), flatten -> vertical scans
    int pos_tb = w * H + h;
    int Lt = W * H;
    for (int k = 0; k < K; ++k) {
        int p = pos_tb + (k - K/2);
        if (p >= 0 && p < Lt) {
            int ww = p / H;
            int hh = p % H;
            sum += x[((b * C + c) * H + hh) * W + ww] * weight[c * K + k];
        }
    }
    sum += bias[c];

    // BT: flip H then TB
    int pos_bt = w * H + (H - 1 - h);
    for (int k = 0; k < K; ++k) {
        int p = pos_bt + (k - K/2);
        if (p >= 0 && p < Lt) {
            int ww = p / H;
            int hh = H - 1 - (p % H);
            sum += x[((b * C + c) * H + hh) * W + ww] * weight[c * K + k];
        }
    }
    sum += bias[c];

    // residual
    float xval = x[idx];
    out[idx] = xval + sum;
}

torch::Tensor cross_scan_conv_cuda(
    torch::Tensor x,
    torch::Tensor weight,
    torch::Tensor bias,
    int kernel_size
) {
    int B = x.size(0);
    int C = x.size(1);
    int H = x.size(2);
    int W = x.size(3);
    auto out = torch::zeros_like(x);

    int total = B * C * H * W;
    int threads = 256;
    int blocks = (total + threads - 1) / threads;

    cross_scan_conv_kernel<<<blocks, threads>>>(
        x.data_ptr<float>(),
        weight.data_ptr<float>(),
        bias.data_ptr<float>(),
        out.data_ptr<float>(),
        B, C, H, W, kernel_size
    );

    return out;
}
'''

cpp_source = '''
torch::Tensor cross_scan_conv_cuda(torch::Tensor x, torch::Tensor weight, torch::Tensor bias, int kernel_size);
'''

module = load_inline(
    name='cross_scan_conv',
    cpp_sources=cpp_source,
    cuda_sources=cuda_source,
    functions=['cross_scan_conv_cuda'],
    verbose=True,
    extra_cuda_cflags=['-O3']
)

class ModelNew(nn.Module):
    def __init__(self, channels: int, kernel_size: int = 3):
        super().__init__()
        self.channels = channels
        self.kernel_size = kernel_size
        pad = kernel_size // 2
        self.dw_conv = nn.Conv1d(
            channels,
            channels,
            kernel_size,
            padding=pad,
            groups=channels,
            bias=True,
        )
        self.module = module

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        weight = self.dw_conv.weight
        bias = self.dw_conv.bias
        return self.module.cross_scan_conv_cuda(x, weight, bias, self.kernel_size)