86 #define _tile_loadd(dst, base, stride)                                         \  argument
106 #define _tile_stream_loadd(dst, base, stride)                                  \  argument
124 #define _tile_stored(dst, base, stride)                                        \  argument
153 #define _tile_dpbssd(dst, src0, src1)                                          \  argument
172 #define _tile_dpbsud(dst, src0, src1)                                          \  argument
191 #define _tile_dpbusd(dst, src0, src1)                                          \  argument
210 #define _tile_dpbuud(dst, src0, src1)                                          \  argument
228 #define _tile_dpbf16ps(dst, src0, src1)                                        \  argument
255                       _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpbssd_internal()
262                       _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpbsud_internal()
269                       _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpbusd_internal()
276                       _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpbuud_internal()
291                         _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpbf16ps_internal()
298                         _tile1024i dst, _tile1024i src1, _tile1024i src2) {  in _tile_dpfp16ps_internal()
326 static __inline__ void __tile_loadd(__tile1024i *dst, const void *base,  in __tile_loadd()
347 static __inline__ void __tile_stream_loadd(__tile1024i *dst, const void *base,  in __tile_stream_loadd()
369 static __inline__ void __tile_dpbssd(__tile1024i *dst, __tile1024i src0,  in __tile_dpbssd()
392 static __inline__ void __tile_dpbsud(__tile1024i *dst, __tile1024i src0,  in __tile_dpbsud()
415 static __inline__ void __tile_dpbusd(__tile1024i *dst, __tile1024i src0,  in __tile_dpbusd()
438 static __inline__ void __tile_dpbuud(__tile1024i *dst, __tile1024i src0,  in __tile_dpbuud()
470 static __inline__ void __tile_zero(__tile1024i *dst) {  in __tile_zero()
490 static __inline__ void __tile_dpbf16ps(__tile1024i *dst, __tile1024i src0,  in __tile_dpbf16ps()
512 static __inline__ void __tile_dpfp16ps(__tile1024i *dst, __tile1024i src0,  in __tile_dpfp16ps()