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))
255                       _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpbssd_internal()  argument
256   return __builtin_ia32_tdpbssd_internal(m, n, k, dst, src1, src2);  in _tile_dpbssd_internal()
262                       _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpbsud_internal()  argument
263   return __builtin_ia32_tdpbsud_internal(m, n, k, dst, src1, src2);  in _tile_dpbsud_internal()
269                       _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpbusd_internal()  argument
270   return __builtin_ia32_tdpbusd_internal(m, n, k, dst, src1, src2);  in _tile_dpbusd_internal()
276                       _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpbuud_internal()  argument
277   return __builtin_ia32_tdpbuud_internal(m, n, k, dst, src1, src2);  in _tile_dpbuud_internal()
291                         _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpbf16ps_internal()  argument
292   return __builtin_ia32_tdpbf16ps_internal(m, n, k, dst, src1, src2);  in _tile_dpbf16ps_internal()
298                         _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpfp16ps_internal()  argument
299   return __builtin_ia32_tdpfp16ps_internal(m, n, k, dst, src1, src2);  in _tile_dpfp16ps_internal()
370                                      __tile1024i src1) {  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()
393                                      __tile1024i src1) {  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()
416                                      __tile1024i src1) {  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()
439                                      __tile1024i src1) {  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()
491                                        __tile1024i src1) {  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()
513                                        __tile1024i src1) {  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()