import torch
import torch.nn as nn

class Model(nn.Module):
    """
    Performs a depthwise 2D convolution with asymmetric input and square kernel.

    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.
        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, bias: bool = False):
        super(Model, self).__init__()
        self.conv2d = nn.Conv2d(in_channels, out_channels, kernel_size=(kernel_size, kernel_size), stride=stride, padding=padding, groups=in_channels, bias=bias)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """
        Performs the depthwise 2D convolution.

        Args:
            x (torch.Tensor): Input tensor of shape (batch_size, in_channels, height_in, width_in).

        Returns:
            torch.Tensor: Output tensor of shape (batch_size, out_channels, height_out, width_out).
        """
        return self.conv2d(x)

# Test code
batch_size = 64
in_channels = 128
out_channels = 128
kernel_size = 3
width_in = 512
height_in = 256
stride = 1
padding = 0

def get_inputs():
    x = torch.rand(batch_size, in_channels, height_in, width_in)
    return [x]

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

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

cpp_src = r'''
#include <torch/extension.h>
torch::Tensor dwconv_forward(torch::Tensor x, torch::Tensor weight, torch::Tensor bias, int64_t stride, int64_t padding);
'''

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

#define ROWS_PER_BLOCK 8

__global__ void dw3x3_s1p0_kernel(const float* __restrict__ x,
                                  const float* __restrict__ w,
                                  const float* __restrict__ b,
                                  float* __restrict__ y,
                                  int C, int H, int W, int OC, int OH, int OW, int mult) {
    int oc = blockIdx.x;
    int n = blockIdx.z;
    int c = oc / mult;
    int y0 = blockIdx.y * ROWS_PER_BLOCK;
    int tid = threadIdx.x;
    __shared__ float sw[9];
    __shared__ float sb;
    if (tid < 9) sw[tid] = w[oc * 9 + tid];
    if (tid == 9) sb = b[oc];
    __syncthreads();
    const float w00=sw[0], w01=sw[1], w02=sw[2], w10=sw[3], w11=sw[4], w12=sw[5], w20=sw[6], w21=sw[7], w22=sw[8];
    const float bv=sb;
    const int in_plane = H * W, out_plane = OH * OW;
    const float* base = x + ((n * C + c) * in_plane);
    float* out = y + ((n * OC + oc) * out_plane);
    #pragma unroll
    for (int dy=0; dy<ROWS_PER_BLOCK; ++dy) {
        int oy = y0 + dy;
        if (oy < OH) {
            const float* r0 = base + oy * W; const float* r1 = r0 + W; const float* r2 = r1 + W;
            float* o = out + oy * OW;
            int ox = tid;
            if (ox < OW) {
                float v = bv;
                v = fmaf(r0[ox], w00, v); v = fmaf(r0[ox+1], w01, v); v = fmaf(r0[ox+2], w02, v);
                v = fmaf(r1[ox], w10, v); v = fmaf(r1[ox+1], w11, v); v = fmaf(r1[ox+2], w12, v);
                v = fmaf(r2[ox], w20, v); v = fmaf(r2[ox+1], w21, v); v = fmaf(r2[ox+2], w22, v);
                o[ox] = v;
            }
            ox = tid + 256;
            if (ox < OW) {
                float v = bv;
                v = fmaf(r0[ox], w00, v); v = fmaf(r0[ox+1], w01, v); v = fmaf(r0[ox+2], w02, v);
                v = fmaf(r1[ox], w10, v); v = fmaf(r1[ox+1], w11, v); v = fmaf(r1[ox+2], w12, v);
                v = fmaf(r2[ox], w20, v); v = fmaf(r2[ox+1], w21, v); v = fmaf(r2[ox+2], w22, v);
                o[ox] = v;
            }
        }
    }
}

__global__ void dw3x3_s1p0_nobias_kernel(const float* __restrict__ x,
                                  const float* __restrict__ w,
                                  float* __restrict__ y,
                                  int C, int H, int W, int OC, int OH, int OW, int mult) {
    int oc = blockIdx.x; int n = blockIdx.z; int c = oc / mult;
    int y0 = blockIdx.y * ROWS_PER_BLOCK; int tid = threadIdx.x;
    __shared__ float sw[9];
    if (tid < 9) sw[tid] = w[oc * 9 + tid];
    __syncthreads();
    const float w00=sw[0], w01=sw[1], w02=sw[2], w10=sw[3], w11=sw[4], w12=sw[5], w20=sw[6], w21=sw[7], w22=sw[8];
    const int in_plane = H * W, out_plane = OH * OW;
    const float* base = x + ((n * C + c) * in_plane); float* out = y + ((n * OC + oc) * out_plane);
    #pragma unroll
    for (int dy=0; dy<ROWS_PER_BLOCK; ++dy) {
        int oy = y0 + dy;
        if (oy < OH) {
            const float* r0 = base + oy * W; const float* r1 = r0 + W; const float* r2 = r1 + W; float* o = out + oy * OW;
            int ox = tid;
            if (ox < OW) {
                float v = 0.0f;
                v = fmaf(r0[ox], w00, v); v = fmaf(r0[ox+1], w01, v); v = fmaf(r0[ox+2], w02, v);
                v = fmaf(r1[ox], w10, v); v = fmaf(r1[ox+1], w11, v); v = fmaf(r1[ox+2], w12, v);
                v = fmaf(r2[ox], w20, v); v = fmaf(r2[ox+1], w21, v); v = fmaf(r2[ox+2], w22, v);
                o[ox] = v;
            }
            ox = tid + 256;
            if (ox < OW) {
                float v = 0.0f;
                v = fmaf(r0[ox], w00, v); v = fmaf(r0[ox+1], w01, v); v = fmaf(r0[ox+2], w02, v);
                v = fmaf(r1[ox], w10, v); v = fmaf(r1[ox+1], w11, v); v = fmaf(r1[ox+2], w12, v);
                v = fmaf(r2[ox], w20, v); v = fmaf(r2[ox+1], w21, v); v = fmaf(r2[ox+2], w22, v);
                o[ox] = v;
            }
        }
    }
}

torch::Tensor dwconv_forward(torch::Tensor x, torch::Tensor weight, torch::Tensor bias, int64_t stride, int64_t padding) {
    int N = x.size(0), C = x.size(1), H = x.size(2), W = x.size(3); int OC = weight.size(0); int K = weight.size(2);
    int OH = (H + 2*(int)padding - K) / (int)stride + 1; int OW = (W + 2*(int)padding - K) / (int)stride + 1;
    auto y = torch::empty({N, OC, OH, OW}, x.options()); int mult = OC / C;
    dim3 block(256); dim3 grid(OC, (OH + ROWS_PER_BLOCK - 1) / ROWS_PER_BLOCK, N);
    if (bias.numel() == 0) dw3x3_s1p0_nobias_kernel<<<grid, block>>>(x.data_ptr<float>(), weight.data_ptr<float>(), y.data_ptr<float>(), C, H, W, OC, OH, OW, mult);
    else dw3x3_s1p0_kernel<<<grid, block>>>(x.data_ptr<float>(), weight.data_ptr<float>(), bias.data_ptr<float>(), y.data_ptr<float>(), C, H, W, OC, OH, OW, mult);
    return y;
}
'''
_ext = load_inline(name='dwconv_ext_final', cpp_sources=cpp_src, cuda_sources=cuda_src, functions=['dwconv_forward'], extra_cuda_cflags=['-O3', '--use_fast_math'], verbose=False)

class ModelNew(nn.Module):
    def __init__(self, in_channels: int, out_channels: int, kernel_size: int, stride: int = 1, padding: int = 0, bias: bool = False):
        super(ModelNew, self).__init__()
        self.conv2d = nn.Conv2d(in_channels, out_channels, kernel_size=(kernel_size, kernel_size), stride=stride, padding=padding, groups=in_channels, bias=bias)
        self.stride = stride
        self.padding = padding
        if not bias:
            self.register_buffer('_empty_bias', torch.empty(0))
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        b = self.conv2d.bias if self.conv2d.bias is not None else self._empty_bias
        return _ext.dwconv_forward(x, self.conv2d.weight, b, self.stride, self.padding)