
    rhBS                        d dl mZ d dlZd dlmZmZmZ d dlZd dlm	Z	 ddl
mZmZ ddlmZmZmZmZ ddlmZmZmZmZ dd	lmZmZmZmZmZmZ dd
lm Z  ddl!m"Z" erd dl#m$Z$ ddlm%Z%  ejL                  e'      Z(ejR                  jT                  Z*ed        Z+ed        Z,d Z-dZ.	  ede+de.z   dz   e.z   dz         Z/dZ0 ede,de0z   dz   e0z   dz         Z1 eejd                  dde*jd                  jf                        Z4d  Z5 ee5d      Z6 G d! d"e      Z7	 	 	 	 	 	 	 	 	 	 	 	 	 	 	 	 	 	 	 	 d)d#Z8d$ Z9d% Z: ee*jd                        	 	 	 	 	 	 	 	 	 	 	 	 	 	 	 	 	 	 d*d&       Z2 ee*jv                        d'        Z;d( Z< ee*jd                  e<       y)+    )annotationsN)OptionalTYPE_CHECKING	TypedDict)CKGroupedConvFwdTemplate   )configir)add_layout_constraintconstrain_to_fx_strides	loweringsregister_lowering)autotune_select_algorithmExternKernelChoiceSymbolicGridFnTritonTemplate)is_onesis_zerospad_listlikesympy_productuse_ck_conv_templateuse_triton_template)V   )mm_config_kwargs)Sequence)	TensorBoxc               F     || |z  |z  |d          |||d         |d   fS NBLOCK_MBLOCK_NGROUPS )nchwmetacdivs         n/var/www/html/ai-insurance-compliance-backend/venv/lib/python3.12/site-packages/torch/_inductor/kernel/conv.pyconv2d_gridr+   .   s9     	QUQYY(QY X     c               L     || |z  |z  |z  |d          |||d         |d   fS r   r#   )r$   r%   dr&   r'   r(   r)   s          r*   conv3d_gridr/   7   s=     	QUQY]DO,QY X r,   c                8    | dkD  s
|dkD  s|dkD  ry| |z  |z  dkD  S )N   Ti   r#   )mr$   ks      r*   _is_large_block_for_cpur4   @   s+    3w!c'QWq519ur,   a  
        idx_x_h = i - PADDING_H + idx_y_h * STRIDE_H
        idx_x_w = j - PADDING_W + idx_y_w * STRIDE_W
        idx_x_c = tl.arange(0, BLOCK_K) + k

        x_ptrs = x_base + (
            (idx_x_h * stride_xh)[:, None]
            + (idx_x_w * stride_xw)[:, None]
            + (idx_x_c * stride_xc)[None, :]
        )
        mask_x = (
            (idx_n < BATCH)[:, None]
            & (idx_x_h >= 0)[:, None]
            & (idx_x_h < IN_H)[:, None]
            & (idx_x_w >= 0)[:, None]
            & (idx_x_w < IN_W)[:, None]
            & (idx_x_c < GROUP_IN_C)[None, :]
        )
        matrix_x = tl.load(x_ptrs, mask=mask_x, other=0.0)

        w_ptrs = w_base + (
            (idx_x_c * stride_wc_in)[:, None] + (i * stride_wh) + (j * stride_ww)
        )
        mask_w = (idx_x_c[:, None] < GROUP_IN_C) & (idx_y_c[None, :] < GROUP_OUT_C)
        matrix_w = tl.load(w_ptrs, mask=mask_w, other=0.0)
        acc += tl.dot(matrix_x, matrix_w, allow_tf32=ALLOW_TF32)
convolution2dag  
{{def_kernel("X", "W")}}
    # Tensor dimensions
    BATCH = {{size("X", 0)}}
    IN_C = {{size("X", 1)}}
    IN_H = {{size("X", 2)}}
    IN_W = {{size("X", 3)}}
    OUT_C = {{size(None, 1)}}
    OUT_H = {{size(None, 2)}}
    OUT_W = {{size(None, 3)}}

    # Strides:
    stride_xn = {{stride("X", 0)}}
    stride_xc = {{stride("X", 1)}}
    stride_xh = {{stride("X", 2)}}
    stride_xw = {{stride("X", 3)}}
    stride_wc_out = {{stride("W", 0)}}
    stride_wc_in = {{stride("W", 1)}}
    stride_wh = {{stride("W", 2)}}
    stride_ww = {{stride("W", 3)}}

    nhw = tl.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M)
    idx_y_w = nhw % OUT_W
    nh = nhw // OUT_W
    idx_y_h = nh % OUT_H
    idx_n = nh // OUT_H
    idx_y_c = tl.program_id(1) * BLOCK_N + tl.arange(0, BLOCK_N)

{% if GROUPS == 1 %}
    group = 0
    GROUP_IN_C = IN_C
    GROUP_OUT_C = OUT_C
{% else %}
    group = tl.program_id(2)
    GROUP_IN_C = IN_C // GROUPS
    GROUP_OUT_C = OUT_C // GROUPS
{% endif %}

    x_base = X + (group * stride_xc * GROUP_IN_C + idx_n * stride_xn)[:, None]
    w_base = (
        W + (group * stride_wc_out * GROUP_OUT_C + idx_y_c * stride_wc_out)[None, :]
    )

    acc = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)

{% if UNROLL %}
{% for i in range(KERNEL_H) %}
{% for j in range(KERNEL_W) %}
    i = {{i}}
    j = {{j}}
    for k in range(0, GROUP_IN_C, BLOCK_K):
        a  
{% endfor %}
{% endfor %}
{% else %}
    # Could be simplified, but slightly slower:
    # for i in range(KERNEL_H):
    #     for j in range(KERNEL_W):
    #         for k in range(0, GROUP_IN_C, BLOCK_K):
    BLOCK_K_COUNT = (GROUP_IN_C + BLOCK_K - 1) // BLOCK_K
    for ijk in range(KERNEL_H * KERNEL_W * BLOCK_K_COUNT):
        k = (ijk % BLOCK_K_COUNT) * BLOCK_K
        ij = ijk // BLOCK_K_COUNT
        i = ij // KERNEL_W
        j = ij % KERNEL_W
        a  
{% endif %}

    mask = (
        (idx_n < BATCH)[:, None]
        & (idx_y_h < OUT_H)[:, None]
        & (idx_y_w < OUT_W)[:, None]
        & (idx_y_c < GROUP_OUT_C)[None, :]
    )
    idx_n = idx_n[:, None]
    idx_c = idx_y_c[None, :] + group * GROUP_OUT_C
    idx_h = idx_y_h[:, None]
    idx_w = idx_y_w[:, None]

    # inductor generates a suffix
    {{store_output(("idx_n", "idx_c", "idx_h", "idx_w"), "acc", "mask")}}
)namegridsourcea  
        idx_x_d = d - PADDING_D + idx_y_d * STRIDE_D
        idx_x_h = i - PADDING_H + idx_y_h * STRIDE_H
        idx_x_w = j - PADDING_W + idx_y_w * STRIDE_W
        idx_x_c = tl.arange(0, BLOCK_K) + k

        x_ptrs = x_base + (
            (idx_x_d * stride_xd)[:, None]
            + (idx_x_h * stride_xh)[:, None]
            + (idx_x_w * stride_xw)[:, None]
            + (idx_x_c * stride_xc)[None, :]
        )
        mask_x = (
            (idx_n < BATCH)[:, None]
            & (idx_x_d >= 0)[:, None]
            & (idx_x_d < IN_D)[:, None]
            & (idx_x_h >= 0)[:, None]
            & (idx_x_h < IN_H)[:, None]
            & (idx_x_w >= 0)[:, None]
            & (idx_x_w < IN_W)[:, None]
            & (idx_x_c < GROUP_IN_C)[None, :]
        )
        matrix_x = tl.load(x_ptrs, mask=mask_x, other=0.0)

        w_ptrs = w_base + (
            (idx_x_c * stride_wc_in)[:, None] +
            (d * stride_wd) + (i * stride_wh) + (j * stride_ww)
        )
        mask_w = (idx_x_c[:, None] < GROUP_IN_C) & (idx_y_c[None, :] < GROUP_OUT_C)
        matrix_w = tl.load(w_ptrs, mask=mask_w, other=0.0)
        acc += tl.dot(matrix_x, matrix_w, allow_tf32=ALLOW_TF32)
convolution3daH  
{{def_kernel("X", "W")}}
    # Tensor dimensions
    BATCH = {{size("X", 0)}}
    IN_C = {{size("X", 1)}}
    IN_D = {{size("X", 2)}}
    IN_H = {{size("X", 3)}}
    IN_W = {{size("X", 4)}}
    OUT_C = {{size(None, 1)}}
    OUT_D = {{size(None, 2)}}
    OUT_H = {{size(None, 3)}}
    OUT_W = {{size(None, 4)}}

    # Strides:
    stride_xn = {{stride("X", 0)}}
    stride_xc = {{stride("X", 1)}}
    stride_xd = {{stride("X", 2)}}
    stride_xh = {{stride("X", 3)}}
    stride_xw = {{stride("X", 4)}}
    stride_wc_out = {{stride("W", 0)}}
    stride_wc_in = {{stride("W", 1)}}
    stride_wd = {{stride("W", 2)}}
    stride_wh = {{stride("W", 3)}}
    stride_ww = {{stride("W", 4)}}

    ndhw = tl.program_id(0) * BLOCK_M + tl.arange(0, BLOCK_M)
    idx_y_w = ndhw % OUT_W
    ndh = ndhw // OUT_W
    idx_y_h = ndh % OUT_H
    nd = ndh // OUT_H
    idx_y_d = nd % OUT_D
    idx_n = nd // OUT_D
    idx_y_c = tl.program_id(1) * BLOCK_N + tl.arange(0, BLOCK_N)

{% if GROUPS == 1 %}
    group = 0
    GROUP_IN_C = IN_C
    GROUP_OUT_C = OUT_C
{% else %}
    group = tl.program_id(2)
    GROUP_IN_C = IN_C // GROUPS
    GROUP_OUT_C = OUT_C // GROUPS
{% endif %}

    x_base = X + (group * stride_xc * GROUP_IN_C + idx_n * stride_xn)[:, None]
    w_base = (
        W + (group * stride_wc_out * GROUP_OUT_C + idx_y_c * stride_wc_out)[None, :]
    )

    acc = tl.zeros((BLOCK_M, BLOCK_N), dtype=tl.float32)

{% if UNROLL %}
{% for d in range(KERNEL_D) %}
{% for i in range(KERNEL_H) %}
{% for j in range(KERNEL_W) %}
    d = {{d}}
    i = {{i}}
    j = {{j}}
    for k in range(0, GROUP_IN_C, BLOCK_K):
        aF  
{% endfor %}
{% endfor %}
{% endfor %}
{% else %}
    # Could be simplified, but slightly slower:
    # for d in range(KERNEL_D):
    #   for i in range(KERNEL_H):
    #     for j in range(KERNEL_W):
    #         for k in range(0, GROUP_IN_C, BLOCK_K):
    BLOCK_K_COUNT = (GROUP_IN_C + BLOCK_K - 1) // BLOCK_K
    for dijk in range(KERNEL_D * KERNEL_H * KERNEL_W * BLOCK_K_COUNT):
        k = (dijk % BLOCK_K_COUNT) * BLOCK_K
        dij = dijk // BLOCK_K_COUNT
        j = dij % KERNEL_W
        di = dij // KERNEL_W
        i = di % KERNEL_H
        d = di // KERNEL_H
        a  
{% endif %}

    mask = (
        (idx_n < BATCH)[:, None]
        & (idx_y_d < OUT_D)[:, None]
        & (idx_y_h < OUT_H)[:, None]
        & (idx_y_w < OUT_W)[:, None]
        & (idx_y_c < GROUP_OUT_C)[None, :]
    )
    idx_n = idx_n[:, None]
    idx_c = idx_y_c[None, :] + group * GROUP_OUT_C
    idx_d = idx_y_d[:, None]
    idx_h = idx_y_h[:, None]
    idx_w = idx_y_w[:, None]

    # inductor generates a suffix
    {{store_output(("idx_n", "idx_c", "idx_d", "idx_h", "idx_w"), "acc", "mask")}}
zat::convolutionF)has_out_variantop_overloadc          
         t        j                  t        j                  |d      d      }t        j                  | j                  dddd      |j                  dd      |j                  dddd            S )Nr   r      r   )out)torchsqueezematmulpermute)xr'   r?   s      r*   conv1x1_via_mmrE   U  s]    emmAr*B/A<<			!Q1qyyACKK1a4K r,   c                  J    e Zd ZU ded<   ded<   ded<   ded<   ded<   ded	<   y
)ConvLayoutParamstuple[int, ...]stridepaddingdilationbool
transposedoutput_paddingintgroupsN)__name__
__module____qualname____annotations__r#   r,   r*   rG   rG   _  s%    ##Kr,   rG   c	                l   t         j                  j                  5  t        j                  j
                  j                  t        j                  | d      t        j                  |d      t        j                  |d      t         j                  j                  j                  |      t         j                  j                  j                  |      t         j                  j                  j                  |      |t         j                  j                  j                  |      |	      }	t        j                  |	j                               }
t        j                  |	j                               }ddd       t        j                  | j                         | j!                         
|      S # 1 sw Y   =xY w)z)Determine output layout for a convolutionT)guard_shapeN)r   graph	fake_moder@   opsatenconvolutionr
   ir_node_to_tensorsizevars
size_hintsconvert_shape_to_inductorsizerI   FixedLayoutget_device_or_error	get_dtype)rD   weightbiasrI   rJ   rK   rM   rN   rP   outputsizess              r*   conv_layoutrh   h  s0    
		 ?++  5  T:  48GG''/GG''0GG''1GG''7

 ,,V[[];--fmmo>? >>			 ? ?s   EF**F3c                    t        t        t        |                   }|j                  d|j	                  d             |S )Nr   r=   )listreversedrangeinsertpop)rankorders     r*   channels_last_orderrq     s0    %+&'E	LLEIIbM"Lr,   c                   t        |j                               }t        |dz
        D ]   }t        t        j
                     |d      }" t        t        j                     |ddg      }t        j                  j                  | t        |            } t        t        |            }|j                  |j                  d             t        t        j                     | |      } | j                         ^ }}t        t        j                     | t        |      |g      } |t        t        j                      | |      }nt        t        j"                     || |      }t        t        j                     |g |d      }t        t        |            }	|	j%                  d|	j                  d             t        t        j                     ||	      S )Nr   r=   dimr   r   )lenget_sizerl   LrZ   rA   rC   r
   ExternKernelrequire_stride_orderrq   rj   appendrn   reshaper   mmaddmmrm   )
rD   rd   re   ro   _	x_permuterg   in_chanresultresult_permutes
             r*   convert_1x1_conv_to_mmr     sa   v !D4!8_ 14<<R01t||_VaV,F
,,Q0CD0IJAU4[!IY]]1%&	$,,9%AjjlOUG	$,,M%0':;A|477Av&4::tQ/t||_V\u\b\2F%+&N!^//34T\\?6>22r,   c	                l    t        |      }t        |      }t        |      }t        |      }t        |t              s)t        j                  j
                  j                  |      }t        |t              sJ t        t        j                  j
                  j                  |            }t        t        j                  j
                  j                  |            }||||||dt        j                         }	t         j                               t        j                               dz
  k(  rVt        t        j                     t        t        t        j                       dg j                               |fi d      S t        j                  j
                  j                  j                               ^}
}}t         j                               dk(  rt        |      dk(  r|	dk(  rj#                  d|z   d|z   d|z   d|z   d	       t        t        j$                      d
       t        t        j$                     d
      t        t        j                     t         |fi d
      S t        |      t'        |      }t'        |      }t'        |      }t'        |      } fd}t(        j*                  xs t(        j,                  }t(        j.                  s	|r |       rt1        |      rt1        |      rvt3        |      rkt1        |      r`|s^t3        |      rS|dk(  rNt        j                  j
                  j5                  t7         j                               d      rt9         |      S |d|	dk7  r_t         d fi }t        t        j:                     |t        t        j<                     ||j                         d   gdgz  z               S  j?                          j?                          t        j                  j@                  rud
k(  rpt        j                  xjB                  dz  c_!        t        jD                  jG                          t        jD                  jG                        tI         d fi }ntI         d fi }t        jJ                  t        j                  j
                  jM                  |jN                              }t        jD                  jQ                   |       t        jD                  jQ                  |      g d}| g}d d<   |jS                  dd       n\ |g}|j?                          |jU                          t        j                  j
                  j                  |j                                g }tV        jX                  jZ                  j]                  d      rt_        j`                  |||fi g}tV        jX                  jZ                  j]                  d      rTtc        |      rHt1        |      r<|s9t3        |      r-t        j                  j
                  je                  | j                         d         rt1        |      r@t1        |      r5t3        |      r*|dk(  r%|jg                  th        ja                  ||             t        jj                  jm                  |	      } |t7         j                         d   g j                         d
d        |
|fi to        |	tp              D ]<  }d
k(  rts        jt                  |f f||d   |d   |d   |d   |d   |d   |t1        |      tV        jv                  jx                  jz                  |j|                  |j~                  d|j                   dk(  st        jt                  |fi d fd|d|d   d|d   d|d
   d|d   d|d   d|d
   d|d   d|d   d|d
   d|dt1        |      dtV        jv                  jx                  jz                  d|j|                  d |j~                  |j                   ? t        |      r/t        j                  || f||fn	t               z   ||||!       t        d"|||      S )#N)rI   rJ   rK   rM   rN   rP   r   r   rs   r>   xpu)r   )r   )rI   rJ   rK   rN   r   c                    t         j                  j                  rdk(  ryt        d fi } t	        j
                  t         j                  j                  j                  | j                              }|t        j                  k(  S )Nr   T)
r   rW   
layout_optrh   r
   get_stride_orderr]   r^   rI   NHWC_STRIDE_ORDER)layoutreq_stride_orderkwargsndimrd   rD   s     r*   channels_last_convz'convolution.<locals>.channels_last_conv  sl    77$!)Q77..GG''6
  2#7#777r,   cpure   ATENTRITON)input_nodesr   KERNEL_HKERNEL_WSTRIDE_HSTRIDE_W	PADDING_H	PADDING_Wr"   UNROLL
ALLOW_TF32
num_stages	num_warpsr   r   KERNEL_Dr   r   STRIDE_Dr   r   	PADDING_Dr   r   r"   r   r   r   r   )r   rI   rJ   rK   rP   n_spatial_dimensionsr[   )Ftuple
isinstancerO   r   rW   r]   evaluate_static_shapeevaluate_static_shapesr
   get_device_typeru   rv   rw   rZ   rA   r[   expandupdate	unsqueezer   r	   max_autotunemax_autotune_gemmconv_1x1_as_mmr   r   statically_known_gtr   r   addviewrealizer   num_channels_last_convrx   require_channels_lastrh   r   r^   rI   ry   rm   freeze_layoutr@   	_inductorutils_use_conv_autotune_backendaten_convolutionbindr   statically_known_equalsrz   aten_conv1x1_via_mmchoicesget_conv_configsr   r4   conv2d_templatemaybe_append_choicebackendscudnn
allow_tf32r   r   r   conv3d_templater   r   add_ck_conv_choicesr   )rD   rd   re   rI   rJ   rK   rM   rN   rP   device_typeout_chanr   kernel_shaper   autotuning_gemmr   r   r   ordered_kwargs_for_cpp_kernelargsr   conv_configscfgr   r   s   ``                     @@r*   r[   r[     s    6]FGnGXH>*Nfc"!!77?fc""" 177##::6BCFAGG$$;;GDEG  ( F $$Q'K
1::<C 12Q66$++q1*<qzz|*<=vtVvV
 	

 ()ww'7'7'N'N($Hg 1::<A#l"3q"8[E=Q-'> 8O"&"7		
 dnnaQ'4>>"6q164262
 	

 |D&$'F7D)GHd+H!.$7N8 ))EV-E-EO 
		?7I7KL!FOWH^$aKGG00qzz|1LaP%a66K50Q77{AdiiL(9!(<'=s
'JK
 	
 IIK
NN
 	wwdai	&&!+&OO11!4 66v>Q77Q77..GG''6
 OO004DE55f>NO%! |6{v%,,Q764 	//@G77?!!- 	
 	88B'H^$GG44Wajjl1oN L!!!NN.33D&AByy11+>1::<?>QZZ\!"-=>?
 {,CD	
 0	C qy33!"F!)!_)!_#AY#AY%aj%aj! #<0$~~33>>"~~!mm!" jj#& 33!"F " *!_	
 *!_ *!_ $AY $AY $AY &aj &aj &aj "  #<0!"  %~~33>>#$  #~~%& "mmjj)70	b F# 44F$2BwP!%		
 %]GT6JJr,   c                (    t        | ||||||||	      S N)r[   )rD   rd   re   rI   rJ   rK   rM   rN   rP   	benchmarkdeterministiccudnn_enabledr   s                r*   _convolutionr     s%      	64(JPV r,   c                    | j                   t        j                  j                  j                  j
                  k(  sJ t        j                  j                  r||fS t        | g|i |S r   )
targetr@   rY   rZ   r[   defaultr   rW   r   r   )fx_noder   r   s      r*   constrain_conv_to_fx_stridesr     sT    >>UYY^^77?????wwV|&w@@@@r,   )rD   r   rd   r   re   Optional[TensorBox]rI   Sequence[int]rJ   rH   rK   rH   rM   rL   rN   rH   rP   rO   returnz	ir.Layout)rD   r   rd   r   re   r   rI   r   rJ   r   rK   r   rM   rL   rN   r   rP   rO   )=
__future__r   loggingtypingr   r   r   r@   -torch._inductor.codegen.rocm.ck_conv_templater    r	   r
   loweringr   r   r   rw   r   select_algorithmr   r   r   r   r   r   r   r   r   r   r   virtualizedr   	mm_commonr   collections.abcr   r   	getLoggerrQ   logrY   rZ   r+   r/   r4   LOOP_BODY_2Dr   LOOP_BODY_3Dr   r[   r   r   rE   r   rG   rh   rq   r   r   r   r#   r,   r*   <module>r      s   "  5 5  R      ' (g! yy~~    8
 !		3h i4jkCH IDJKUYvB !		;x y<z{O` aPbccgR &	  ((	  )> y       	 
       $     F3. 4##$oKoKoK oK 	oK
 oK oK oK "oK oK %oKd 4$$% &(A d&&(D Er,   