+
    åûjÁ1  ã                  ó�   € R t ^ RIHt ^ RIHtHtHt ^ RIt^ RIH	u H
t ^ RIHt RR R lltR R ltR	 R
 ltR R ltR R ltR# )a*  Pure-PyTorch implementation of deform_conv2d matching torchvision.ops.deform_conv2d.

Implements both DCNv1 (mask=None) and DCNv2 (mask provided) as described in:
  - Deformable Convolutional Networks (https://arxiv.org/abs/1703.06211)
  - Deformable ConvNets v2 (https://arxiv.org/abs/1811.11168)
)Úannotations)ÚOptionalÚTupleÚUnionN)ÚTensorc               ó<   € V ^8„  d   QhRRRRRRRRRRR	RR
RRRRR/	# )é   Úinputr   ÚoffsetÚweightÚbiasúOptional[Tensor]ÚstridezUnion[int, Tuple[int, int]]ÚpaddingÚdilationÚmaskÚreturn© )Úformats   "Ú_C:\Ai_Coding\apps\studio\ComfyUI\.venv\Lib\site-packages\kornia\feature\aliked\deform_conv2d.pyÚ__annotate__r   "   sm   € ÷ Pñ PØðPàðPð ðPð ð	Pð
 (ðPð )ðPð *ðPð ðPð ñPó    c                ó²  € V P                   w  r‰r«VP                   w  rÍrï\        V\        4      '       d   WU3p\        V\        4      '       d   Wf3p\        V\        4      '       d   WD3pVw  ppVw  ppVw  ppW�,          pWï,          pVP                   ^,          ^V,          ,          pV	V,          pV
^V,          ,           VV^,
          ,          ,
          ^,
          V,          ^,           pV^V,          ,           VV^,
          ,          ,
          ^,
          V,          ^,           p\        P                  ! 4       '       d   \        WW~V3WEVVV3VV4
      pM\        WW~V3WEVVV3VV4
      pV^8X  da   VP                  W‰V,          VV,          4      p\        P                  ! VP                  ^WÉV,          4      P                  VRR4      V4      pMQVV,          pVP                  VVVVVV,          4      pVP                  VVVV4      p \        P                  ! RVV 4      pVP                  WŒVV4      pVe   VVP                  ^R^^4      ,           pV# )aô  Deformable 2D convolution (DCNv1/v2) without torchvision dependency.

Matches the signature and behaviour of ``torchvision.ops.deform_conv2d``.

Args:
    input: input feature map ``(B, C_in, H_in, W_in)``.
    offset: offset tensor ``(B, 2 * n_offset_groups * kH * kW, H_out, W_out)``.
        Layout per kernel point: ``(dy, dx)`` interleaved.
    weight: convolution kernel ``(C_out, C_in / groups, kH, kW)``.
    bias: optional bias ``(C_out,)``.
    stride: convolution stride ``(sH, sW)``.
    padding: zero-padding ``(pH, pW)``.
    dilation: kernel dilation ``(dH, dW)``.
    mask: optional DCNv2 modulation mask
        ``(B, n_offset_groups * kH * kW, H_out, W_out)`` in ``[0, 1]``.

Returns:
    Output feature map ``(B, C_out, H_out, W_out)``.
zbgckn,gock->bgonéÿÿÿÿ)ÚshapeÚ
isinstanceÚintÚtorchÚis_grad_enabledÚ_sample_with_autogradÚ_grid_sample_offsetsÚreshapeÚbmmÚexpandÚeinsum)!r	   r
   r   r   r   r   r   r   ÚBÚC_inÚH_inÚW_inÚC_outÚC_in_per_groupÚkHÚkWÚsHÚsWÚpHÚpWÚdHÚdWÚgroupsÚKÚ
n_off_grpsÚc_per_off_grpÚH_outÚW_outÚsampledÚcolsÚoutÚC_out_per_groupÚweight_flats!   &&&&&&&&                         r   Údeform_conv2dr>   "   s%  € ð:  Ÿ+™+Ñ€AˆTØ$*§L¡LÑ!€E˜2ä�'œ3×ÒØÐ$ˆÜ�(œC× Ò ØÐ'ˆÜ�&œ#×ÒØÐ!ˆà�F€BˆØ�F€BˆØ�F€BˆàÕ#€FØ
�€AØ—‘˜a• Q¨¥UÕ+€JØ˜JÕ&€Mà�A˜•F�]˜R 2¨¥6�]Õ*¨QÕ.°2Õ5¸Õ9€EØ�A˜•F�]˜R 2¨¥6�]Õ*¨QÕ.°2Õ5¸Õ9€Eô ×Ò×ÒÜ'Ø˜4 b ¨6¸HÀuÈeÀnÐV`Ðboó
‰ô 'Ø˜4 b ¨6¸HÀuÈeÀnÐV`Ðboó
ˆð
 �„{à�‰˜q¨¥(¨E°E­MÓ:ˆÜ�iŠi˜Ÿ™ q¨%¸µÓ:×AÑAÀ!ÀRÈÓLÈdÓS‰à 6�/ˆØ—/‘/ ! V¨^¸QÀÈÅÓNˆØ—n‘n V¨_¸nÈaÓPˆÜ�lŠlÐ-¨w¸ÓDˆØ
�+‰+�a  uÓ
-€Cð ÒØ�D—L‘L  B¨¨1Ó-Õ-ˆà€Jr   c               óD   € V ^8„  d   QhRRRRRRRRRRR	RR
RRRRRRRRR/# ©r   r	   r   r
   r   r   Úkernel_sizezTuple[int, int]r   r   r   Úout_sizer5   r   r6   r   r   )r   s   "r   r   r   u   sv   € ÷ -ñ -Øð-àð-ð ð-ð !ð	-ð
 ð-ð ð-ð ð-ð ð-ð ð-ð ð-ð ñ-r   c
           	     óþ  € V P                   ^ ,          p
W4WVV3w  w  r¼w  rÞw  ppw  ppw  ppW¼,          pVP                  W¨^V,          VV4      pVR,          pVR,          p\        P                  ! VV P                  V P
                  R7      p\        P                  ! VV P                  V P
                  R7      pVV,          V,
          P                  ^^^V^4      pVV,          V,
          P                  ^^^^V4      p\        P                  ! W°P                  V P
                  R7      V,          p\        P                  ! WÀP                  V P
                  R7      V,          pVP                  V4      P                  ^^V^^4      pVP                  V4      P                  ^^V^^4      pVV,           V,           p VV,           V,           p!\        V V V!W‰4      p"Ve   V"VP                  W¨^VVV4      ,          p"V"# )z¶Sample the deformable-convolution taps with :func:`_bilinear_sample` and modulate them by ``mask``.

Returns:
    Sampled tensor ``(B, n_off_grps, c_per_off_grp, K, H_out, W_out)``.
©ÚdeviceÚdtype)ºNNNrG   :é    Nr   rG   rG   )rG   rG   :é   Nr   rG   rG   )	r   r!   r   ÚarangerE   rF   Úrepeat_interleaveÚrepeatÚ_bilinear_sample)#r	   r
   r   rA   r   r   r   rB   r5   r6   r%   r+   r,   r-   r.   r/   r0   r1   r2   r7   r8   r4   Úoff_yÚoff_xÚgrid_ohÚgrid_owÚbase_hÚbase_wÚkh_offsÚkw_offsÚkern_hÚkern_wÚsample_hÚsample_wr9   s#   &&&&&&&&&&                         r   r   r   u   s»  € ð" 	�‰�A�€AØ=HÐRYÐemÐ=mÑ:�H€R‰hˆr™˜˜R¡( 2 r©N¨U°EØ
�€Að �^‰^˜A¨1¨q­5°%¸Ó?€FØÐ#Õ$€EØÐ#Õ$€Eô �lŠl˜5¨¯©¸U¿[¹[ÔI€GÜ�lŠl˜5¨¯©¸U¿[¹[ÔI€Gà˜�l˜RÕ×(Ñ(¨¨A¨q°%¸Ó;€FØ˜�l˜RÕ×(Ñ(¨¨A¨q°!°UÓ;€Fä�lŠl˜2§l¡l¸%¿+¹+ÔFÈÕK€GÜ�lŠl˜2§l¡l¸%¿+¹+ÔFÈÕK€GØ×&Ñ& rÓ*×2Ñ2°1°a¸¸A¸qÓA€FØ�^‰^˜BÓ×'Ñ'¨¨1¨a°°AÓ6€Fð ˜� Õ&€HØ˜� Õ&€Hä˜u h°¸*ÓT€GØÒØ˜DŸL™L¨¸¸1¸eÀUÓKÕKˆØ€Nr   c               ó<   € V ^8„  d   QhRRRRRRRRRRRRRR	R
RRR/	# )r   Ún_outr   Úkr   ÚpadÚdilÚsizerE   ztorch.devicerF   ztorch.dtyper   r   r   )r   s   "r   r   r   ¥   sT   € ÷ )ñ )Øð)Øð)Ø #ð)Ø*-ð)Ø47ð)Ø?Bð)ØLXð)Øalð)àñ)r   c                ó&  € ^^V,          ,
          V,
          p\         P                  ! Wˆ^V,          V ,          ,           ^V,          WgR7      p	\         P                  ! ^ ^V,          V,          ^V,          WgR7      p
V
R,          V	R,          ,           # )aª  Return ``2 p + 1 - size`` for the pixel each zero-offset tap reads along one axis, shape ``(k, n_out)``.

Tap ``j`` of output ``o`` reads pixel ``p = o * stride - pad + j * dil``, which ``grid_sample`` with
``align_corners=False`` places at ``(2 p + 1 - size) / size``. The values are integers, exact in float32, and two
integer ``arange`` calls keep the sizes symbolic under ``torch.compile``, without a host-to-device copy.
rD   )rG   N)NrG   )r   rJ   )r[   r\   r   r]   r^   r_   rE   rF   ÚstartÚoutsÚkerns   &&&&&&&&   r   Ú_doubled_baserd   ¥   sm   € ð ��C•�K˜$Õ€EÜ�<Š<˜ q¨6¥z°EÕ'9Õ9¸1¸v½:ÈfÔb€DÜ�<Š<˜˜1˜s�7 Q�;¨¨C­¸ÔL€DØ��=˜4 �=Õ(Ð(r   c               óD   € V ^8„  d   QhRRRRRRRRRRR	RR
RRRRRRRRR/# r@   r   )r   s   "r   r   r   ´   sƒ   € ÷ 1Sñ 1SØð1Sàð1Sð ð1Sð !ð	1Sð
 ð1Sð ð1Sð ð1Sð ð1Sð ð1Sð ð1Sð ñ1Sr   c
           
     óê  € V P                   w  r«rÍW4WVV3w  w  rïw  ppw  ppw  ppw  ppWï,          p\        P                  ! V P                  \        P                  4      pVP                  W¨Wï^VV4      P                  V4      p\        VVVVVWÀP                  V4      p\        VVVVVWÐP                  V4      p\        P                  ! VR,          VR	,          ^R7      P                  V4      P                  R
^4      p\        P                  ! VR,          VR,          ^R7      P                  V4      P                  R
^4      p\        P                  ! VV3RR7      P                  W¨,          VV,          V^4      pV P                  V4      P                  W¨,          WœV4      p \        P                  ! V VRRRR7      p!V!P                  W¨V	VVV4      p!\        P                  ! V4      P!                  ^R7      P#                  4       P                  W¨^VVV4      p"Vf5   V!P%                  V"\'        R4      4      P                  V P                  4      # \        P                  ! VVP                  4      p#VP                  W¨^VVV4      P                  V#4      P%                  V"\'        R4      4      p$V!V$,          P                  \        P                  ! V P                  VP                  4      4      # )aÍ  Sample the deformable-convolution taps with one ``grid_sample`` and modulate them by ``mask``.

Gives the values of :func:`_sample_with_autograd`. Offset groups are folded into the batch of ``grid_sample``
and kernel points into its output height. The normalized coordinates are built straight from the offsets, one
fused multiply-add and one division per axis on a zero-offset base grid, which keeps the kernel launches that
bound small-image inference on CUDA few. Half-precision inputs are sampled in float32: rounding the normalized
coordinates to half precision would add error that the absolute-coordinate sampler does not have.

Returns:
    Sampled tensor ``(B, n_off_grps, c_per_off_grp, K, H_out, W_out)``.
)Úalpha)ÚdimÚbilinearÚzerosF)ÚmodeÚpadding_modeÚalign_cornersÚnan)rG   NrG   N)rG   rG   rG   rG   rH   éýÿÿÿ)NrG   NrG   )rG   rG   rG   rG   rI   r   )r   r   Úpromote_typesrF   Úfloat32r!   Útord   rE   ÚaddÚdiv_Úclamp_ÚstackÚFÚgrid_sampleÚisfiniteÚallÚlogical_not_Úmasked_fillÚfloat)%r	   r
   r   rA   r   r   r   rB   r5   r6   r%   Ú_C_inr'   r(   r+   r,   r-   r.   r/   r0   r1   r2   r7   r8   r4   rF   ÚoffÚbase_yÚbase_xÚgrid_yÚgrid_xÚgridÚgroupedr9   Ú	nonfiniteÚmodulation_dtypeÚ
modulations%   &&&&&&&&&&                           r   r    r    ´   s…  € ð. !Ÿ;™;Ñ€AˆdØ=HÐRYÐemÐ=mÑ:�H€R‰hˆr�2™˜˜R¡( 2 r©N¨U°EØ
�€AÜ×Ò §¡¬U¯]©]Ó;€Eà
�.‰.˜¨°°5¸%Ó
@×
CÑ
CÀEÓ
J€CÜ˜5 " b¨"¨b°$¿¹ÀeÓL€FÜ˜5 " b¨"¨b°$¿¹ÀeÓL€Fô �YŠY�vÐ.Õ/°°]Õ1CÈ1ÔM×RÑRÐSWÓX×_Ñ_Ð`bÐdeÓf€FÜ�YŠY�vÐ.Õ/°°]Õ1CÈ1ÔM×RÑRÐSWÓX×_Ñ_Ð`bÐdeÓf€FÜ�;Š;˜ Ð'¨RÔ0×8Ñ8¸½ÈÈUÍÐTYÐ[\Ó]€DØ�h‰h�u‹o×%Ñ% a¥n°mÈ4ÓP€GÜ�mŠm˜G T°
ÈÐ`eÔf€GØ�o‰o˜a¨]¸A¸uÀeÓL€Gô —’˜sÓ#×'Ñ'¨AÐ'Ó.×;Ñ;Ó=×EÑEÀaÐUVÐXYÐ[`ÐbgÓh€IØ‚|Ø×"Ñ" 9¬e°E«lÓ;×>Ñ>¸u¿{¹{ÓKÐKä×*Ò*¨5°$·*±*Ó=Ðà�‰�Q A q¨%°Ó7×:Ñ:Ð;KÓL×XÑXÐYbÔdiÐjoÓdpÓqð ð �jÕ ×$Ñ$¤U×%8Ò%8¸¿¹ÀdÇjÁjÓ%QÓRÐRr   c               ó0   € V ^8„  d   QhRRRRRRRRRRRR/# )	r   r	   r   rX   rY   r5   r   r6   r   r   )r   s   "r   r   r   è   sT   € ÷ Pyñ PyØðPyàðPyð ðPyð ð	Pyð
 ðPyð ñPyr   c                ój  a aaaaa a!a"€ S P                   w  opoo VP                  4       pVP                  4       pW,
          pW',
          p	VP                  4       pVP                  4       pV^,           p
V^,           pR VV V 3R llpV! Wg4      pV! Wk4      pV! W§4      pV! W«4      pVP                  ^ \	        S^,
          ^ 4      4      pV
P                  ^ \	        S^,
          ^ 4      4      pVP                  ^ \	        S ^,
          ^ 4      4      pVP                  ^ \	        S ^,
          ^ 4      4      pS P                  SSSSS ,          4      o"VP                   ^,          VP                   ^,          ,          VP                   ^,          ,          o!R VV VV!V"V3R llpV! VV4      pV! VV4      pV! VV4      pV! VV4      pVP                   ^,          pVP                   ^,          pVP                   ^,          pSSSVVV3pVP                  V4      pVP                  V4      pVP                  V4      pVP                  V4      pVP                  ^4      pV	P                  ^4      p	VP                  ^4      pVP                  ^4      pVP                  ^4      pVP                  ^4      pVV,          ^V,
          ,          ^V	,
          ,          VV,          ^V,
          ,          V	,          ,           VV,          V,          ^V	,
          ,          ,           VV,          V,          V	,          ,           # )añ  Sample input at fractional (sample_h, sample_w) positions with zero-padding.

Uses the same corner-validity bilinear interpolation as torchvision's
deformable-conv C++ kernel.

Args:
    input: ``(B, C_in, H_in, W_in)``
    sample_h: ``(B, n_off_grps, K, H_out, W_out)``
    sample_w: ``(B, n_off_grps, K, H_out, W_out)``
    n_off_grps: number of offset groups.
    c_per_off_grp: input channels per offset group.

Returns:
    Sampled tensor ``(B, n_off_grps, c_per_off_grp, K, H_out, W_out)``.
c               ó$   € V ^8„  d   QhRRRRRR/# )r   Úhr   Úwr   r   )r   s   "r   r   Ú&_bilinear_sample.<locals>.__annotate__  s&   € ÷ Oñ O�&ð O˜Vð O¨ñ Or   c                ó|   <€ V ^ 8¬  V S8  ,          V^ 8¬  ,          VS8  ,          P                  SP                  4      # )rH   )rr   rF   )rŒ   r�   r'   r(   r	   s   &&€€€r   Ú_validÚ _bilinear_sample.<locals>._valid  s7   ø€ Ø�a‘˜A ™HÕ%¨¨a©Õ0°A¸±HÕ=×AÑAÀ%Ç+Á+ÓNÐNr   c               ó$   € V ^8„  d   QhRRRRRR/# )r   Úhhr   Úwwr   r   )r   s   "r   r   rŽ     s!   € ÷ 0ñ 0�Fð 0 ð 0¨6ñ 0r   c                ó¼   <€ V S,          V,           P                  SSS4      pVP                  ^4      P                  RRSR4      p\        P                  ! S^V4      # )r   r   )r!   Ú	unsqueezer#   r   Úgather)	r“   r”   Úidxr%   r(   r6   Ú	flat_sizeÚ
input_flatr5   s	   && €€€€€€r   Ú_gatherÚ!_bilinear_sample.<locals>._gather  sQ   ø€ Ø�D�y˜2�~×&Ñ& q¨*°iÓ@ˆØ�m‰m˜AÓ×%Ñ% b¨"¨m¸RÓ@ˆÜ�|Š|˜J¨¨3Ó/Ð/r   )r   ÚfloorÚlongÚclampÚmaxr!   r–   )#r	   rX   rY   r5   r6   r~   Úh0Úw0ÚlhÚlwÚh1Úw1r�   Úm00Úm01Úm10Úm11Úh0cÚh1cÚw0cÚw1cr›   Úv00Úv01Úv10Úv11r4   r7   r8   r   r%   r'   r(   r™   rš   s#   f&&ff                         @@@@@r   rM   rM   è   s®  ÿ€ ð, !Ÿ;™;Ñ€A€uˆd�Dà	�‰Ó	€BØ	�‰Ó	€Bà	�€BØ	�€Bà	�‰‹€BØ	�‰‹€BØ	ˆa�€BØ	ˆa�€B÷Oò Oñ �‹.€CÙ
�‹.€CÙ
�‹.€CÙ
�‹.€Cà
�(‰(�1”c˜$ �( AÓ&Ó
'€CØ
�(‰(�1”c˜$ �( AÓ&Ó
'€CØ
�(‰(�1”c˜$ �( AÓ&Ó
'€CØ
�(‰(�1”c˜$ �( AÓ&Ó
'€Cð —‘˜q *¨m¸TÀD½[ÓI€Jà—‘˜qÕ! H§N¡N°1Õ$5Õ5¸¿¹ÀqÕ8IÕI€I÷0õ 0ñ
 �#�sÓ
€CÙ
�#�sÓ
€CÙ
�#�sÓ
€CÙ
�#�sÓ
€Cà�‰�qÕ€AØ�N‰N˜1Õ€EØ�N‰N˜1Õ€EØ�
˜M¨1¨e°UÐ;€Eà
�+‰+�eÓ
€CØ
�+‰+�eÓ
€CØ
�+‰+�eÓ
€CØ
�+‰+�eÓ
€Cà	�‰�a‹€BØ	�‰�a‹€BØ
�-‰-˜Ó
€CØ
�-‰-˜Ó
€CØ
�-‰-˜Ó
€CØ
�-‰-˜Ó
€Cà��9˜˜B�Õ 1 r¥6Õ*¨S°3­Y¸!¸b½&Õ-AÀBÕ-FÕFÈÈsÍÐUWÍÐ[\Ð_aÕ[aÕIbÕbÐehÐknÕenÐqsÕesÐvxÕexÕxÐxr   )N©rI   rI   )rH   rH   r³   N)Ú__doc__Ú
__future__r   Útypingr   r   r   r   Útorch.nn.functionalÚnnÚ
functionalrw   r   r>   r   rd   r    rM   r   r   r   Ú<module>rº      sA   ðñ$õ #ç )Ñ )ã ß Ð Ý ÷Põf-õ`)õ1S÷hPyr   