139 /// corresponding signed 8-bit integers in src1, producing 4 intermediate 32-bit
140 /// results. Sum these 4 results with the corresponding 32-bit integer in "dst",
141 /// and store the 32-bit result back to tile "dst".
159 /// 32-bit results. Sum these 4 results with the corresponding 32-bit integer
160 /// in "dst", and store the 32-bit result back to tile "dst".
177 /// corresponding signed 8-bit integers in src1, producing 4 intermediate 32-bit
178 /// results. Sum these 4 results with the corresponding 32-bit integer in "dst",
179 /// and store the 32-bit result back to tile "dst".
197 /// 32-bit results. Sum these 4 results with the corresponding 32-bit integer in
198 /// "dst", and store the 32-bit result back to tile "dst".
214 /// src1, accumulating the intermediate single-precision (32-bit) floating-point
215 /// elements with elements in "dst", and store the 32-bit result back to tile
238 _tile_loadd_internal(unsigned short m, unsigned short n, const void *base,  in _tile_loadd_internal()  argument
240   return __builtin_ia32_tileloadd64_internal(m, n, base,  in _tile_loadd_internal()
246 _tile_loaddt1_internal(unsigned short m, unsigned short n, const void *base,  in _tile_loaddt1_internal()  argument
248   return __builtin_ia32_tileloaddt164_internal(m, n, base,  in _tile_loaddt1_internal()
254 _tile_dpbssd_internal(unsigned short m, unsigned short n, unsigned short k,  in _tile_dpbssd_internal()  argument
256   return __builtin_ia32_tdpbssd_internal(m, n, k, dst, src1, src2);  in _tile_dpbssd_internal()
261 _tile_dpbsud_internal(unsigned short m, unsigned short n, unsigned short k,  in _tile_dpbsud_internal()  argument
263   return __builtin_ia32_tdpbsud_internal(m, n, k, dst, src1, src2);  in _tile_dpbsud_internal()
268 _tile_dpbusd_internal(unsigned short m, unsigned short n, unsigned short k,  in _tile_dpbusd_internal()  argument
270   return __builtin_ia32_tdpbusd_internal(m, n, k, dst, src1, src2);  in _tile_dpbusd_internal()
275 _tile_dpbuud_internal(unsigned short m, unsigned short n, unsigned short k,  in _tile_dpbuud_internal()  argument
277   return __builtin_ia32_tdpbuud_internal(m, n, k, dst, src1, src2);  in _tile_dpbuud_internal()
282 _tile_stored_internal(unsigned short m, unsigned short n, void *base,  in _tile_stored_internal()  argument
284   return __builtin_ia32_tilestored64_internal(m, n, base,  in _tile_stored_internal()
290 _tile_dpbf16ps_internal(unsigned short m, unsigned short n, unsigned short k,  in _tile_dpbf16ps_internal()  argument
292   return __builtin_ia32_tdpbf16ps_internal(m, n, k, dst, src1, src2);  in _tile_dpbf16ps_internal()
297 _tile_dpfp16ps_internal(unsigned short m, unsigned short n, unsigned short k,  in _tile_dpfp16ps_internal()  argument
299   return __builtin_ia32_tdpfp16ps_internal(m, n, k, dst, src1, src2);  in _tile_dpfp16ps_internal()
354 /// corresponding signed 8-bit integers in src1, producing 4 intermediate 32-bit
355 /// results. Sum these 4 results with the corresponding 32-bit integer in "dst",
356 /// and store the 32-bit result back to tile "dst".
378 /// 32-bit results. Sum these 4 results with the corresponding 32-bit integer
379 /// in "dst", and store the 32-bit result back to tile "dst".
400 /// corresponding signed 8-bit integers in src1, producing 4 intermediate 32-bit
401 /// results. Sum these 4 results with the corresponding 32-bit integer in "dst",
402 /// and store the 32-bit result back to tile "dst".
424 /// 32-bit results. Sum these 4 results with the corresponding 32-bit integer in
425 /// "dst", and store the 32-bit result back to tile "dst".
475 /// src1, accumulating the intermediate single-precision (32-bit) floating-point
476 /// elements with elements in "dst", and store the 32-bit result back to tile
497 /// src1, accumulating the intermediate single-precision (32-bit) floating-point
498 /// elements with elements in "dst", and store the 32-bit result back to tile