153 #define _tile_dpbssd(dst, src0, src1)                                          \  argument
154   __builtin_ia32_tdpbssd((dst), (src0), (src1))
172 #define _tile_dpbsud(dst, src0, src1)                                          \  argument
173   __builtin_ia32_tdpbsud((dst), (src0), (src1))
191 #define _tile_dpbusd(dst, src0, src1)                                          \  argument
192   __builtin_ia32_tdpbusd((dst), (src0), (src1))
210 #define _tile_dpbuud(dst, src0, src1)                                          \  argument
211   __builtin_ia32_tdpbuud((dst), (src0), (src1))
228 #define _tile_dpbf16ps(dst, src0, src1)                                        \  argument
229   __builtin_ia32_tdpbf16ps((dst), (src0), (src1))
369 static __inline__ void __tile_dpbssd(__tile1024i *dst, __tile1024i src0,  in __tile_dpbssd()  argument
371   dst->tile = _tile_dpbssd_internal(src0.row, src1.col, src0.col, dst->tile,  in __tile_dpbssd()
372                                     src0.tile, src1.tile);  in __tile_dpbssd()
392 static __inline__ void __tile_dpbsud(__tile1024i *dst, __tile1024i src0,  in __tile_dpbsud()  argument
394   dst->tile = _tile_dpbsud_internal(src0.row, src1.col, src0.col, dst->tile,  in __tile_dpbsud()
395                                     src0.tile, src1.tile);  in __tile_dpbsud()
415 static __inline__ void __tile_dpbusd(__tile1024i *dst, __tile1024i src0,  in __tile_dpbusd()  argument
417   dst->tile = _tile_dpbusd_internal(src0.row, src1.col, src0.col, dst->tile,  in __tile_dpbusd()
418                                     src0.tile, src1.tile);  in __tile_dpbusd()
438 static __inline__ void __tile_dpbuud(__tile1024i *dst, __tile1024i src0,  in __tile_dpbuud()  argument
440   dst->tile = _tile_dpbuud_internal(src0.row, src1.col, src0.col, dst->tile,  in __tile_dpbuud()
441                                     src0.tile, src1.tile);  in __tile_dpbuud()
490 static __inline__ void __tile_dpbf16ps(__tile1024i *dst, __tile1024i src0,  in __tile_dpbf16ps()  argument
492   dst->tile = _tile_dpbf16ps_internal(src0.row, src1.col, src0.col, dst->tile,  in __tile_dpbf16ps()
493                                       src0.tile, src1.tile);  in __tile_dpbf16ps()
512 static __inline__ void __tile_dpfp16ps(__tile1024i *dst, __tile1024i src0,  in __tile_dpfp16ps()  argument
514   dst->tile = _tile_dpfp16ps_internal(src0.row, src1.col, src0.col, dst->tile,  in __tile_dpfp16ps()
515                                       src0.tile, src1.tile);  in __tile_dpfp16ps()