ó
    “~iJ  ã                   óx  • S SK Jr  \R                  SS j5       r\R                  SS j5       r\R
                  SS j5       r\R
                  SS j5       r\R
                  SS j5       r\R
                  SS j5       r	\R
                  SS	 j5       r
\R
                  SS
 j5       r\R
                  SS j5       rg)é    )ÚcoreNc           
      óR   • [         R                  " SS/ [         R                  SSU S9$ )Nzmov.u64 $0, %globaltimer;z=lFé   ©ÚdtypeÚis_pureÚpackÚ	_semantic)r   Úinline_asm_elementwiseÚint64©r
   s    Ú]/home/mande/repo/quber/.venv/lib/python3.13/site-packages/triton/language/extra/cuda/utils.pyÚglobaltimerr      s.   € ä×&Ò&Ð'BÀDÈ"ÔTX×T^ÑT^ÐhmÐtuØ1:ñ<ð <ó    c           
      óR   • [         R                  " SS/ [         R                  SSU S9$ )Nzmov.u32 $0, %smid;z=rTr   r   )r   r   Úint32r   s    r   Úsmidr   
   s+   € ä×&Ò&Ð';¸TÀ2ÌTÏZÉZÐaeÐlmØ1:ñ<ð <r   c                 óp   • [         R                  " U R                  R                  R                  S-  5      $ )Né    ©r   Ú	constexprÚbuilderÚoptionsÚ	num_warpsr   s    r   Únum_threadsr      s(   € ä�>Š>˜)×+Ñ+×3Ñ3×=Ñ=ÀÑBÓCÐCr   c                 ój   • [         R                  " U R                  R                  R                  5      $ ©Nr   r   s    r   r   r      s#   € ä�>Š>˜)×+Ñ+×3Ñ3×=Ñ=Ó>Ð>r   c           
      óT   • [         R                  " SSU /[         R                  SSUS9$ )Na¸  {                                      
.reg .b32 a<2>, b<2>;                  
prmt.b32 a0, 0, $2, 0x5746;            
and.b32 b0, a0, 0x7f007f00;            
and.b32 b1, a0, 0x00ff00ff;            
and.b32 a1, a0, 0x00800080;            
shr.b32  b0, b0, 1;                    
add.u32 b1, b1, a1;                    
lop3.b32 $0, b0, 0x80008000, a0, 0xf8; 
shl.b32 $1, b1, 7;                     
}                                      
z=r,=r,rTé   r   )r   r   Úfloat16)Úargr
   s     r   Úconvert_fp8e4b15_to_float16r"   !   s2   € ä×&Ò&ð
	4ð 6?ÀÀÌTÏ\É\ÐcgÐnoØñð r   c           
      ó†   • SnU(       a  US-  nOUS-  nUS-  n[         R                  " USU /[         R                  SSUS9$ )	NaN  {
            .reg .pred p<4>;
            .reg .b32 a<2>, b<2>;
            .reg .b16 c<4>;
            .reg .b16 max_val_f16;
            .reg .b32 max_val_f16x2;
            mov.b16 max_val_f16,   0x3F00;
            mov.b32 max_val_f16x2, 0x3F003F00;
            and.b32 a0, $1, 0x7fff7fff;
            and.b32 a1, $2, 0x7fff7fff;zSmin.f16x2 a0, a0, max_val_f16x2;
                  min.f16x2 a1, a1, max_val_f16x2;a×  setp.lt.f16x2  p0|p1, a0, max_val_f16x2;
                  setp.lt.f16x2  p2|p3, a1, max_val_f16x2;
                  mov.b32 {c0, c1}, a0;
                  mov.b32 {c2, c3}, a1;
                  selp.b16  c0, c0, max_val_f16, p0;
                  selp.b16  c1, c1, max_val_f16, p1;
                  selp.b16  c2, c2, max_val_f16, p2;
                  selp.b16  c3, c3, max_val_f16, p3;
                  mov.b32 a0, {c0, c1};
                  mov.b32 a1, {c2, c3};zömad.lo.u32 a0, a0, 2, 0x00800080;
              mad.lo.u32 a1, a1, 2, 0x00800080;
              lop3.b32 b0, $1, 0x80008000, a0, 0xea;
              lop3.b32 b1, $2, 0x80008000, a1, 0xea;
              prmt.b32 $0, b0, b1, 0x7531;
              }z=r,r,rTr   r   )r   r   Úfloat8e4b15)r!   Ú	has_minx2r
   Úasms       r   Úconvert_float16_to_fp8e4b15r'   2   sm   € ð	+€Cö Øð 6ñ 	6‰ð 	ð 	+ñ 		+ˆð ð ñ €Cô ×&Ò& s¨H°s°eÄ4×CSÑCSÐ]aÐhiØ1:ñ<ð <r   c                 ó8  • U R                   R                  R                  5       (       aH  [        XS9nUR                  R	                  5       (       a  UR                  [        R                  US9nU$ U R                   R                  R                  5       (       d+  U R                   R                  R	                  5       (       d   eU nU R                   R                  R	                  5       (       a  UR                  [        R                  SUS9n[        XcUS9nU$ )Nr   Úrtz)Úfp_downcast_roundingr
   ©r%   r
   )ÚtypeÚscalarÚis_fp8e4b15r"   Úis_fp32Útor   Úfloat32Úis_fp16r    r'   )r!   Údst_tyr*   r%   r
   Ú
upcast_valÚdowncast_vals          r   Úconvert_custom_float8r6   V   sÍ   € à
‡x�x‡�×"Ñ"×$Ñ$Ü0°ÑJˆ
Ø�=‰=× Ñ ×"Ñ"Ø#Ÿ™¤t§|¡|¸y˜ÐIˆJØÐà�8‰8�?‰?×"Ñ"×$Ñ$¨¯©¯©×(?Ñ(?×(AÑ(AÐAÐAØ€LØ
‡x�x‡�×Ñ× Ñ Ø#—‘¤t§|¡|È%Ð[d�ÐeˆÜ.¨|Ð\eÑf€LØÐr   c                 ó   • [        XUSUS9$ )NTr+   ©r6   ©r!   r3   r*   r
   s       r   Úconvert_custom_float8_sm80r:   f   s   € ä  Ð.BÈdÐ^gÑhÐhr   c                 ó   • [        XUSUS9$ )NFr+   r8   r9   s       r   Úconvert_custom_float8_sm70r<   k   s   € ä  Ð.BÈeÐ_hÑiÐir   r   )NN)Útriton.languager   Úexternr   r   Úbuiltinr   r   r"   r'   r6   r:   r<   © r   r   Ú<module>rA      sç   ðÝ  ð ‡�ó<ó ð<ð
 ‡�ó<ó ð<ð
 ‡�óDó ðDð ‡�ó?ó ð?ð ‡�óó ðð  ‡�ó <ó ð <ðF ‡�óó ðð ‡�óió ðið ‡�ójó ñjr   