151 #define _tile_dpbssd(dst, src0, src1)                                          \  argument
152   __builtin_ia32_tdpbssd((dst), (src0), (src1))
170 #define _tile_dpbsud(dst, src0, src1)                                          \  argument
171   __builtin_ia32_tdpbsud((dst), (src0), (src1))
189 #define _tile_dpbusd(dst, src0, src1)                                          \  argument
190   __builtin_ia32_tdpbusd((dst), (src0), (src1))
208 #define _tile_dpbuud(dst, src0, src1)                                          \  argument
209   __builtin_ia32_tdpbuud((dst), (src0), (src1))
226 #define _tile_dpbf16ps(dst, src0, src1)                                        \  argument
227   __builtin_ia32_tdpbf16ps((dst), (src0), (src1))
245                       _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpbssd_internal()  argument
246   return __builtin_ia32_tdpbssd_internal(m, n, k, dst, src1, src2);  in _tile_dpbssd_internal()
252                       _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpbsud_internal()  argument
253   return __builtin_ia32_tdpbsud_internal(m, n, k, dst, src1, src2);  in _tile_dpbsud_internal()
259                       _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpbusd_internal()  argument
260   return __builtin_ia32_tdpbusd_internal(m, n, k, dst, src1, src2);  in _tile_dpbusd_internal()
266                       _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpbuud_internal()  argument
267   return __builtin_ia32_tdpbuud_internal(m, n, k, dst, src1, src2);  in _tile_dpbuud_internal()
281                         _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpbf16ps_internal()  argument
282   return __builtin_ia32_tdpbf16ps_internal(m, n, k, dst, src1, src2);  in _tile_dpbf16ps_internal()
332                           __tile1024i src1) {  in __tile_dpbssd()  argument
333   dst->tile = _tile_dpbssd_internal(src0.row, src1.col, src0.col, dst->tile,  in __tile_dpbssd()
334                                     src0.tile, src1.tile);  in __tile_dpbssd()
355                           __tile1024i src1) {  in __tile_dpbsud()  argument
356   dst->tile = _tile_dpbsud_internal(src0.row, src1.col, src0.col, dst->tile,  in __tile_dpbsud()
357                                     src0.tile, src1.tile);  in __tile_dpbsud()
378                           __tile1024i src1) {  in __tile_dpbusd()  argument
379   dst->tile = _tile_dpbusd_internal(src0.row, src1.col, src0.col, dst->tile,  in __tile_dpbusd()
380                                     src0.tile, src1.tile);  in __tile_dpbusd()
401                           __tile1024i src1) {  in __tile_dpbuud()  argument
402   dst->tile = _tile_dpbuud_internal(src0.row, src1.col, src0.col, dst->tile,  in __tile_dpbuud()
403                                     src0.tile, src1.tile);  in __tile_dpbuud()
454                             __tile1024i src1) {  in __tile_dpbf16ps()  argument
455   dst->tile = _tile_dpbf16ps_internal(src0.row, src1.col, src0.col, dst->tile,  in __tile_dpbf16ps()
456                                       src0.tile, src1.tile);  in __tile_dpbf16ps()