
    Vh$!                     .   d dl Z d dlZd dlmZ d dlmZ ddlmZmZ	 ddl
mZmZmZmZ ddlmZmZmZmZmZ ddlmZ d	d
lmZmZmZmZmZmZ  e j<                  e      Z ejB                  jD                  Z"ed        Z#d Z$d Z% ede#d      Z& eejN                  d      Z( eejR                  de"jR                  jT                        Z+ e	jX                  e"jN                        ddd       Z- e	jX                  e"jR                        d	d	ddd       Z.y)    N)counters)CKGemmTemplate   )irlowering)autotune_select_algorithmExternKernelChoiceSymbolicGridFnTritonTemplate)use_aten_gemm_kernelsuse_ck_gemm_templateuse_cpp_bmm_templateuse_cutlass_templateuse_triton_template)V   )_is_static_problemaddmm_epiloguemm_args
mm_configs
mm_optionsshould_fallback_to_atenc                :     |||d          |||d         z  | dfS )NBLOCK_MBLOCK_Nr    )bmnmetacdivs        J/home/dcms/DCMS/lib/python3.12/site-packages/torch/_inductor/kernel/bmm.pybmm_gridr#   %   s*    DO$tAtI'??AFF    c                 2    | dkD  s
|dkD  s|dkD  ry| |z  dkD  S )N   Ti   r   )r   r   ks      r"   _is_large_block_for_cpur(   *   s&    3w!c'QWq55=r$   c                N    |dk(  rt        | ||dt              S t        | ||      S )Ncpug      ?)scaleexclude)r   r(   )r   r   r'   device_types       r"   bmm_configsr.   1   s-    e!Q6MNNaAr$   bmma  
{{def_kernel("A", "B")}}
    M = {{size("A", -2)}}
    N = {{size("B", -1)}}
    K = {{size("A", -1)}}

    stride_aq = {{stride("A", 0)}}
    stride_am = {{stride("A", 1)}}
    stride_ak = {{stride("A", 2)}}

    stride_bq = {{stride("B", 0)}}
    stride_bk = {{stride("B", 1)}}
    stride_bn = {{stride("B", 2)}}

    # based on triton.ops.matmul
    pid = tl.program_id(0)
    grid_m = (M + BLOCK_M - 1) // BLOCK_M
    grid_n = (N + BLOCK_N - 1) // BLOCK_N

    # re-order program ID for better L2 performance
    width = GROUP_M * grid_n
    group_id = pid // width
    group_size = min(grid_m - group_id * GROUP_M, GROUP_M)
    pid_m = group_id * GROUP_M + (pid % group_size)
    pid_n = (pid % width) // (group_size)

    rm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
    rn = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
    if (stride_am == 1 and stride_ak == M) or (stride_am == K and stride_ak == 1):
        ram = tl.max_contiguous(tl.multiple_of(rm % M, BLOCK_M), BLOCK_M)
    else:
        ram = rm % M
    if (stride_bk == 1 and stride_bn == K) or (stride_bk == N and stride_bn == 1):
        rbn = tl.max_contiguous(tl.multiple_of(rn % N, BLOCK_N), BLOCK_N)
    else:
        rbn = rn % N

    rk = tl.arange(0, BLOCK_K)

    idx_q = tl.program_id(1)  # batch dimension for BMM
    A = A + (ram[:, None] * stride_am + rk[None, :] * stride_ak + idx_q*stride_aq)
    B = B + (rk[:, None] * stride_bk + rbn[None, :] * stride_bn + idx_q*stride_bq)

    acc = tl.zeros((BLOCK_M, BLOCK_N), dtype=ACC_TYPE)
    for k in range(K, 0, -BLOCK_K):
        if EVEN_K:
            a = tl.load(A)
            b = tl.load(B)
        else:
            a = tl.load(A, mask=rk[None, :] < k, other=0.)
            b = tl.load(B, mask=rk[:, None] < k, other=0.)
        acc += tl.dot(a, b, allow_tf32=ALLOW_TF32)
        A += BLOCK_K * stride_ak
        B += BLOCK_K * stride_bk

    # rematerialize rm and rn to save registers
    rm = pid_m * BLOCK_M + tl.arange(0, BLOCK_M)
    rn = pid_n * BLOCK_N + tl.arange(0, BLOCK_N)
    idx_q = tl.program_id(1)  # batch dimension for BMM
    idx_m = rm[:, None]
    idx_n = rn[None, :]
    mask = (idx_m < M) & (idx_n < N)

    # inductor generates a suffix
    {{store_output(("idx_q", "idx_m", "idx_n"), "acc", "mask")}}
)namegridsourcezat::bmm_outzat::baddbmm_out)op_overloadlayoutc                   t        d | |fD              r| j                         d   dk(  s|j                         d   dk(  rWt        j                  | d      } t        j                  |d      }t        j                  t        j
                  | |      d      S d }d fd} ||       r0t        j                  j                  j                  d	   } || |      }  ||      r0t        j                  j                  j                  d   } |||      }t        | ||
      \  }}}	}} }t        d   d| d| d|	 xx   dz  cc<   t        j                  d|||	| j                         |j                         |       t               rt         j#                  | |f|      gng }
t%        |      rOt'        |||	t)        j*                  |             D ]*  }t-        j.                  |
f| |f|dt1        ||||	|       , t3        |      \  }}|r+|r)t5        ||||	      rddlm} |j;                  |
|| |g       t=        || |      rddlm } |jC                  |
|| |g       tE        ||||	      rtG        jH                  |
|| |g       tK        |
      r'|
jM                  t         j#                  | |f|             tO        d|
| |g|      S )Nc              3   V   K   | ]!  }|j                         j                  d k(   # yw)r*   N)
get_devicetype).0xs     r"   	<genexpr>ztuned_bmm.<locals>.<genexpr>   s!     
>A1<<>%'
>s   ')r   r   )axisc                     t        j                  |       syt        j                  | d      \  }}t        |t         j                        S )NTF)freeze)r   is_storage_and_layoutas_storage_and_layout
isinstanceFlexibleLayout)t_r5   s      r"   is_valid_to_require_contiguousz1tuned_bmm.<locals>.is_valid_to_require_contiguous   s<    ++A.005AIAvfb&7&788r$   c                     |d   dk(  xr | d   dk(  xs |d   | d   k\  xs |d   dk(  xr | d   dk(  xs |d   | d   k\  S )Nr=   r   r   )sizesstridess     r"    is_preferred_layout_as_bmm_inputz3tuned_bmm.<locals>.is_preferred_layout_as_bmm_input   sf     q QeBi1n&PuRy8PU"+"Sb	Q(R'"+r:RUr$   c                     |j                   d   j                         }|j                   d   j                         } ||      st        j                  j                  |       } | S )Nval)r    sizestrider   ExternKernelrequire_contiguous)rE   meta_trJ   rK   rL   s       r"   may_require_contiguousz)tuned_bmm.<locals>.may_require_contiguous   sT    KK&++-Ekk%(//1G3E7COO66q9Hr$   r   r4   aten_mm_infoz	aten.bmm_rF   zPTuned aten.bmm: m=%s, n=%s, k=%s, mat1_dtype=%s, mat2_dtype=%s, output_layout=%sr-   input_nodesr5   )CUTLASS3xGemmTemplate)CppBmmTemplater/   )(allget_sizeL	unsqueezesum_mulr   graphcurrent_nodeargsr   r   loginfo	get_dtyper   aten_bmmbindr   r.   r   get_device_typebmm_templatemaybe_append_choicer   r   r   codegen.cuda.gemm_templaterY   add_cutlass_gemm_choicesr   codegen.cpp_bmm_templaterZ   add_choicesr   r   add_ck_gemm_choicesr   appendr   )mat1mat2r5   rG   rT   	meta_mat1	meta_mat2r   r   r'   choicesconfigstatic_shape
is_nonzerorY   rZ   rL   s                   @r"   	tuned_bmmrz      s   

>$
>>==?1"dmmoa&8A&=;;tR(D;;tQ'D66!%%d+!44	9	U	 *$/,,11!4I)$	:D)$/,,11!4I)$	:D")$V"DAq!VT4 ^y1QCq45:5HHZ			 8M7Nx}}dD\623TVG6"!!Qr7I7I$7OP 	F,,!4L VQ1f5		  2&9L*
';FAq!'LF66wtUFD$/=""4L	
 FAq!,**7FT4LIw'x}}dD\6:;$UGdD\6JJr$   )alphabetar5   c                T   t        ||| |      \  }}}}}}} t        d   d| d| d| xx   dz  cc<   t        j                  d||||j	                         |j	                         | j	                         |       t               rt        j                  | ||f|||      gng }	t        |      rjt        |||t        j                  |            D ]E  }
t        j                  |	f| ||f|d	t        |
||||      dt        |j                   ||      d
 G t#        d|	| ||g|      S )Nr4   rU   zaten.baddbmm_rF   r   z\Tuned aten.baddbmm: m=%s, n=%s, k=%s, mat1_dtype=%s, mat2_dtype=%s, inp=%s, output_layout=%s)r{   r|   rV   rW   )prefix_argsepilogue_fnbaddbmm)r   r   rd   re   rf   r   aten_baddbmmrh   r   r.   r   ri   rj   rk   r   r   dtyper   )inprr   rs   r{   r|   r5   r   r   r'   rv   rw   s              r"   tuned_baddbmmr      sS   '.tT3v'N$Aq!VT4 ^}QCq1QC89Q>9HHf				 !" 
		Ct,fE		MN 
 6"!!Qr7I7I$7OP 	F,, $- VQ1f5	
 *6<<E	 %Y#tT9JFSSr$   )/loggingtorchtorch._dynamo.utilsr   7torch._inductor.codegen.rocm.ck_universal_gemm_templater    r   r   r]   select_algorithmr   r	   r
   r   utilsr   r   r   r   r   virtualizedr   	mm_commonr   r   r   r   r   r   	getLogger__name__rd   opsatenr#   r(   r.   rj   r/   rg   r   outr   register_loweringrz   r   r   r$   r"   <module>r      s     ( R       g!yy~~ G G 		AEN eii7!	MM$$,,2B2B
 TXX$( SK SKl T\\",-Ad !T #!Tr$   