lib/Headers/amxintrin.h

20   __attribute__((__always_inline__, __nodebug__, __target__("amx-tile")))
28 /// Load tile configuration from a 64-byte memory location specified by
29 /// "mem_addr". The tile configuration includes the tile type palette, the
31 /// palette_id is zero, that signifies the init state for both the tile
32 /// config and the tile data, and the tiles are zeroed. Any invalid
46 /// Stores the current tile configuration to a 64-byte memory location
47 /// specified by "mem_addr". The tile configuration includes the tile type
62 /// Release the tile configuration to return to the init state, which
72 /// Load tile rows from memory specifieid by "base" address and "stride" into
73 /// destination tile "dst" using the tile configuration previously configured
81 ///    A destination tile. Max size is 1024 Bytes.
90 /// Load tile rows from memory specifieid by "base" address and "stride" into
91 /// destination tile "dst" using the tile configuration previously configured
101 ///    A destination tile. Max size is 1024 Bytes.
110 /// Store the tile specified by "src" to memory specifieid by "base" address and
111 /// "stride" using the tile configuration previously configured via
119 ///    A destination tile. Max size is 1024 Bytes.
127 /// Zero the tile specified by "tdest".
133 /// \param tile
134 ///    The destination tile to be zero. Max size is 1024 Bytes.
135 #define _tile_zero(tile) __builtin_ia32_tilezero((tile))
141 /// and store the 32-bit result back to tile "dst".
148 ///    The destination tile. Max size is 1024 Bytes.
150 ///    The 1st source tile. Max size is 1024 Bytes.
152 ///    The 2nd source tile. Max size is 1024 Bytes.
160 /// in "dst", and store the 32-bit result back to tile "dst".
167 ///    The destination tile. Max size is 1024 Bytes.
169 ///    The 1st source tile. Max size is 1024 Bytes.
171 ///    The 2nd source tile. Max size is 1024 Bytes.
179 /// and store the 32-bit result back to tile "dst".
186 ///    The destination tile. Max size is 1024 Bytes.
188 ///    The 1st source tile. Max size is 1024 Bytes.
190 ///    The 2nd source tile. Max size is 1024 Bytes.
198 /// "dst", and store the 32-bit result back to tile "dst".
205 ///    The destination tile. Max size is 1024 Bytes.
207 ///    The 1st source tile. Max size is 1024 Bytes.
209 ///    The 2nd source tile. Max size is 1024 Bytes.
215 /// elements with elements in "dst", and store the 32-bit result back to tile
223 ///    The destination tile. Max size is 1024 Bytes.
225 ///    The 1st source tile. Max size is 1024 Bytes.
227 ///    The 2nd source tile. Max size is 1024 Bytes.
231 /// AMX tile register size can be configured, the maximum size is 16x64=1024
233 /// represent 2D tile and the fixed size is maximum amx tile register size.
283                       __SIZE_TYPE__ stride, _tile1024i tile) {
285                                               (__SIZE_TYPE__)(stride), tile);
302 /// This struct pack the shape and tile data together for user. We suggest
309   _tile1024i tile;
312 /// Load tile rows from memory specifieid by "base" address and "stride" into
313 /// destination tile "dst".
320 ///    A destination tile. Max size is 1024 Bytes.
328   dst->tile = _tile_loadd_internal(dst->row, dst->col, base, stride);
331 /// Load tile rows from memory specifieid by "base" address and "stride" into
332 /// destination tile "dst". This intrinsic provides a hint to the implementation
341 ///    A destination tile. Max size is 1024 Bytes.
349   dst->tile = _tile_loaddt1_internal(dst->row, dst->col, base, stride);
356 /// and store the 32-bit result back to tile "dst".
363 ///    The destination tile. Max size is 1024 Bytes.
365 ///    The 1st source tile. Max size is 1024 Bytes.
367 ///    The 2nd source tile. Max size is 1024 Bytes.
371   dst->tile = _tile_dpbssd_internal(src0.row, src1.col, src0.col, dst->tile,
372                                     src0.tile, src1.tile);
379 /// in "dst", and store the 32-bit result back to tile "dst".
386 ///    The destination tile. Max size is 1024 Bytes.
388 ///    The 1st source tile. Max size is 1024 Bytes.
390 ///    The 2nd source tile. Max size is 1024 Bytes.
394   dst->tile = _tile_dpbsud_internal(src0.row, src1.col, src0.col, dst->tile,
395                                     src0.tile, src1.tile);
402 /// and store the 32-bit result back to tile "dst".
409 ///    The destination tile. Max size is 1024 Bytes.
411 ///    The 1st source tile. Max size is 1024 Bytes.
413 ///    The 2nd source tile. Max size is 1024 Bytes.
417   dst->tile = _tile_dpbusd_internal(src0.row, src1.col, src0.col, dst->tile,
418                                     src0.tile, src1.tile);
425 /// "dst", and store the 32-bit result back to tile "dst".
432 ///    The destination tile. Max size is 1024 Bytes.
434 ///    The 1st source tile. Max size is 1024 Bytes.
436 ///    The 2nd source tile. Max size is 1024 Bytes.
440   dst->tile = _tile_dpbuud_internal(src0.row, src1.col, src0.col, dst->tile,
441                                     src0.tile, src1.tile);
444 /// Store the tile specified by "src" to memory specifieid by "base" address and
458   _tile_stored_internal(src.row, src.col, base, stride, src.tile);
461 /// Zero the tile specified by "dst".
468 ///    The destination tile to be zero. Max size is 1024 Bytes.
471   dst->tile = __builtin_ia32_tilezero_internal(dst->row, dst->col);
476 /// elements with elements in "dst", and store the 32-bit result back to tile
484 ///    The destination tile. Max size is 1024 Bytes.
486 ///    The 1st source tile. Max size is 1024 Bytes.
488 ///    The 2nd source tile. Max size is 1024 Bytes.
492   dst->tile = _tile_dpbf16ps_internal(src0.row, src1.col, src0.col, dst->tile,
493                                       src0.tile, src1.tile);
498 /// elements with elements in "dst", and store the 32-bit result back to tile
506 ///    The destination tile. Max size is 1024 Bytes.
508 ///    The 1st source tile. Max size is 1024 Bytes.
510 ///    The 2nd source tile. Max size is 1024 Bytes.
514   dst->tile = _tile_dpfp16ps_internal(src0.row, src1.col, src0.col, dst->tile,
515                                       src0.tile, src1.tile);