1 /*===--------------- amxintrin.h - AMX intrinsics -*- C/C++ -*---------------===
3 * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 * See https://llvm.org/LICENSE.txt for license information.
5 * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 *===------------------------------------------------------------------------===
11 #error "Never use <amxintrin.h> directly; include <immintrin.h> instead."
12 #endif /* __IMMINTRIN_H */
18 /* Define the default attributes for the functions in this file. */
19 #define __DEFAULT_FN_ATTRS_TILE \
20 __attribute__((__always_inline__, __nodebug__, __target__("amx-tile")))
21 #define __DEFAULT_FN_ATTRS_INT8 \
22 __attribute__((__always_inline__, __nodebug__, __target__("amx-int8")))
23 #define __DEFAULT_FN_ATTRS_BF16 \
24 __attribute__((__always_inline__, __nodebug__, __target__("amx-bf16")))
25 #define __DEFAULT_FN_ATTRS_FP16 \
26 __attribute__((__always_inline__, __nodebug__, __target__("amx-fp16")))
28 /// Load tile configuration from a 64-byte memory location specified by
29 /// "mem_addr". The tile configuration includes the tile type palette, the
30 /// number of bytes per row, and the number of rows. If the specified
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
33 /// configurations will result in #GP fault.
35 /// \headerfile <immintrin.h>
37 /// This intrinsic corresponds to the <c> LDTILECFG </c> instruction.
40 /// A pointer to 512-bits configuration
41 static __inline__
void __DEFAULT_FN_ATTRS_TILE
42 _tile_loadconfig(const void *__config
) {
43 __builtin_ia32_tile_loadconfig(__config
);
46 /// Stores the current tile configuration to a 64-byte memory location
47 /// specified by "mem_addr". The tile configuration includes the tile type
48 /// palette, the number of bytes per row, and the number of rows. If tiles
49 /// are not configured, all zeroes will be stored to memory.
51 /// \headerfile <immintrin.h>
53 /// This intrinsic corresponds to the <c> STTILECFG </c> instruction.
56 /// A pointer to 512-bits configuration
57 static __inline__
void __DEFAULT_FN_ATTRS_TILE
58 _tile_storeconfig(void *__config
) {
59 __builtin_ia32_tile_storeconfig(__config
);
62 /// Release the tile configuration to return to the init state, which
63 /// releases all storage it currently holds.
65 /// \headerfile <immintrin.h>
67 /// This intrinsic corresponds to the <c> TILERELEASE </c> instruction.
68 static __inline__
void __DEFAULT_FN_ATTRS_TILE
_tile_release(void) {
69 __builtin_ia32_tilerelease();
72 /// Load tile rows from memory specifieid by "base" address and "stride" into
73 /// destination tile "dst" using the tile configuration previously configured
74 /// via "_tile_loadconfig".
76 /// \headerfile <immintrin.h>
78 /// This intrinsic corresponds to the <c> TILELOADD </c> instruction.
81 /// A destination tile. Max size is 1024 Bytes.
83 /// A pointer to base address.
85 /// The stride between the rows' data to be loaded in memory.
86 #define _tile_loadd(dst, base, stride) \
87 __builtin_ia32_tileloadd64((dst), ((const void *)(base)), \
88 (__SIZE_TYPE__)(stride))
90 /// Load tile rows from memory specifieid by "base" address and "stride" into
91 /// destination tile "dst" using the tile configuration previously configured
92 /// via "_tile_loadconfig". This intrinsic provides a hint to the implementation
93 /// that the data will likely not be reused in the near future and the data
94 /// caching can be optimized accordingly.
96 /// \headerfile <immintrin.h>
98 /// This intrinsic corresponds to the <c> TILELOADDT1 </c> instruction.
101 /// A destination tile. Max size is 1024 Bytes.
103 /// A pointer to base address.
105 /// The stride between the rows' data to be loaded in memory.
106 #define _tile_stream_loadd(dst, base, stride) \
107 __builtin_ia32_tileloaddt164((dst), ((const void *)(base)), \
108 (__SIZE_TYPE__)(stride))
110 /// Store the tile specified by "src" to memory specifieid by "base" address and
111 /// "stride" using the tile configuration previously configured via
112 /// "_tile_loadconfig".
114 /// \headerfile <immintrin.h>
116 /// This intrinsic corresponds to the <c> TILESTORED </c> instruction.
119 /// A destination tile. Max size is 1024 Bytes.
121 /// A pointer to base address.
123 /// The stride between the rows' data to be stored in memory.
124 #define _tile_stored(dst, base, stride) \
125 __builtin_ia32_tilestored64((dst), ((void *)(base)), (__SIZE_TYPE__)(stride))
127 /// Zero the tile specified by "tdest".
129 /// \headerfile <immintrin.h>
131 /// This intrinsic corresponds to the <c> TILEZERO </c> instruction.
134 /// The destination tile to be zero. Max size is 1024 Bytes.
135 #define _tile_zero(tile) __builtin_ia32_tilezero((tile))
137 /// Compute dot-product of bytes in tiles with a source/destination accumulator.
138 /// Multiply groups of 4 adjacent pairs of signed 8-bit integers in src0 with
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".
143 /// \headerfile <immintrin.h>
145 /// This intrinsic corresponds to the <c> TDPBSSD </c> instruction.
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.
153 #define _tile_dpbssd(dst, src0, src1) \
154 __builtin_ia32_tdpbssd((dst), (src0), (src1))
156 /// Compute dot-product of bytes in tiles with a source/destination accumulator.
157 /// Multiply groups of 4 adjacent pairs of signed 8-bit integers in src0 with
158 /// corresponding unsigned 8-bit integers in src1, producing 4 intermediate
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".
162 /// \headerfile <immintrin.h>
164 /// This intrinsic corresponds to the <c> TDPBSUD </c> instruction.
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.
172 #define _tile_dpbsud(dst, src0, src1) \
173 __builtin_ia32_tdpbsud((dst), (src0), (src1))
175 /// Compute dot-product of bytes in tiles with a source/destination accumulator.
176 /// Multiply groups of 4 adjacent pairs of unsigned 8-bit integers in src0 with
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".
181 /// \headerfile <immintrin.h>
183 /// This intrinsic corresponds to the <c> TDPBUSD </c> instruction.
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.
191 #define _tile_dpbusd(dst, src0, src1) \
192 __builtin_ia32_tdpbusd((dst), (src0), (src1))
194 /// Compute dot-product of bytes in tiles with a source/destination accumulator.
195 /// Multiply groups of 4 adjacent pairs of unsigned 8-bit integers in src0 with
196 /// corresponding unsigned 8-bit integers in src1, producing 4 intermediate
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".
200 /// \headerfile <immintrin.h>
202 /// This intrinsic corresponds to the <c> TDPBUUD </c> instruction.
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.
210 #define _tile_dpbuud(dst, src0, src1) \
211 __builtin_ia32_tdpbuud((dst), (src0), (src1))
213 /// Compute dot-product of BF16 (16-bit) floating-point pairs in tiles src0 and
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
218 /// \headerfile <immintrin.h>
220 /// This intrinsic corresponds to the <c> TDPBF16PS </c> instruction.
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.
228 #define _tile_dpbf16ps(dst, src0, src1) \
229 __builtin_ia32_tdpbf16ps((dst), (src0), (src1))
231 /// AMX tile register size can be configured, the maximum size is 16x64=1024
232 /// bytes. Since there is no 2D type in llvm IR, we use vector type to
233 /// represent 2D tile and the fixed size is maximum amx tile register size.
234 typedef int _tile1024i
__attribute__((__vector_size__(1024), __aligned__(64)));
236 /// This is internal intrinsic. C/C++ user should avoid calling it directly.
237 static __inline__ _tile1024i __DEFAULT_FN_ATTRS_INT8
238 _tile_loadd_internal(unsigned short m
, unsigned short n
, const void *base
,
239 __SIZE_TYPE__ stride
) {
240 return __builtin_ia32_tileloadd64_internal(m
, n
, base
,
241 (__SIZE_TYPE__
)(stride
));
244 /// This is internal intrinsic. C/C++ user should avoid calling it directly.
245 static __inline__ _tile1024i __DEFAULT_FN_ATTRS_INT8
246 _tile_loaddt1_internal(unsigned short m
, unsigned short n
, const void *base
,
247 __SIZE_TYPE__ stride
) {
248 return __builtin_ia32_tileloaddt164_internal(m
, n
, base
,
249 (__SIZE_TYPE__
)(stride
));
252 /// This is internal intrinsic. C/C++ user should avoid calling it directly.
253 static __inline__ _tile1024i __DEFAULT_FN_ATTRS_INT8
254 _tile_dpbssd_internal(unsigned short m
, unsigned short n
, unsigned short k
,
255 _tile1024i dst
, _tile1024i src1
, _tile1024i src2
) {
256 return __builtin_ia32_tdpbssd_internal(m
, n
, k
, dst
, src1
, src2
);
259 /// This is internal intrinsic. C/C++ user should avoid calling it directly.
260 static __inline__ _tile1024i __DEFAULT_FN_ATTRS_INT8
261 _tile_dpbsud_internal(unsigned short m
, unsigned short n
, unsigned short k
,
262 _tile1024i dst
, _tile1024i src1
, _tile1024i src2
) {
263 return __builtin_ia32_tdpbsud_internal(m
, n
, k
, dst
, src1
, src2
);
266 /// This is internal intrinsic. C/C++ user should avoid calling it directly.
267 static __inline__ _tile1024i __DEFAULT_FN_ATTRS_INT8
268 _tile_dpbusd_internal(unsigned short m
, unsigned short n
, unsigned short k
,
269 _tile1024i dst
, _tile1024i src1
, _tile1024i src2
) {
270 return __builtin_ia32_tdpbusd_internal(m
, n
, k
, dst
, src1
, src2
);
273 /// This is internal intrinsic. C/C++ user should avoid calling it directly.
274 static __inline__ _tile1024i __DEFAULT_FN_ATTRS_INT8
275 _tile_dpbuud_internal(unsigned short m
, unsigned short n
, unsigned short k
,
276 _tile1024i dst
, _tile1024i src1
, _tile1024i src2
) {
277 return __builtin_ia32_tdpbuud_internal(m
, n
, k
, dst
, src1
, src2
);
280 /// This is internal intrinsic. C/C++ user should avoid calling it directly.
281 static __inline__
void __DEFAULT_FN_ATTRS_INT8
282 _tile_stored_internal(unsigned short m
, unsigned short n
, void *base
,
283 __SIZE_TYPE__ stride
, _tile1024i tile
) {
284 return __builtin_ia32_tilestored64_internal(m
, n
, base
,
285 (__SIZE_TYPE__
)(stride
), tile
);
288 /// This is internal intrinsic. C/C++ user should avoid calling it directly.
289 static __inline__ _tile1024i __DEFAULT_FN_ATTRS_BF16
290 _tile_dpbf16ps_internal(unsigned short m
, unsigned short n
, unsigned short k
,
291 _tile1024i dst
, _tile1024i src1
, _tile1024i src2
) {
292 return __builtin_ia32_tdpbf16ps_internal(m
, n
, k
, dst
, src1
, src2
);
295 /// This is internal intrinsic. C/C++ user should avoid calling it directly.
296 static __inline__ _tile1024i __DEFAULT_FN_ATTRS_FP16
297 _tile_dpfp16ps_internal(unsigned short m
, unsigned short n
, unsigned short k
,
298 _tile1024i dst
, _tile1024i src1
, _tile1024i src2
) {
299 return __builtin_ia32_tdpfp16ps_internal(m
, n
, k
, dst
, src1
, src2
);
302 /// This struct pack the shape and tile data together for user. We suggest
303 /// initializing the struct as early as possible, because compiler depends
304 /// on the shape information to do configure. The constant value is preferred
305 /// for optimization by compiler.
306 typedef struct __tile1024i_str
{
307 const unsigned short row
;
308 const unsigned short col
;
312 /// Load tile rows from memory specifieid by "base" address and "stride" into
313 /// destination tile "dst".
315 /// \headerfile <immintrin.h>
317 /// This intrinsic corresponds to the <c> TILELOADD </c> instruction.
320 /// A destination tile. Max size is 1024 Bytes.
322 /// A pointer to base address.
324 /// The stride between the rows' data to be loaded in memory.
325 __DEFAULT_FN_ATTRS_TILE
326 static __inline__
void __tile_loadd(__tile1024i
*dst
, const void *base
,
327 __SIZE_TYPE__ stride
) {
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
333 /// that the data will likely not be reused in the near future and the data
334 /// caching can be optimized accordingly.
336 /// \headerfile <immintrin.h>
338 /// This intrinsic corresponds to the <c> TILELOADDT1 </c> instruction.
341 /// A destination tile. Max size is 1024 Bytes.
343 /// A pointer to base address.
345 /// The stride between the rows' data to be loaded in memory.
346 __DEFAULT_FN_ATTRS_TILE
347 static __inline__
void __tile_stream_loadd(__tile1024i
*dst
, const void *base
,
348 __SIZE_TYPE__ stride
) {
349 dst
->tile
= _tile_loaddt1_internal(dst
->row
, dst
->col
, base
, stride
);
352 /// Compute dot-product of bytes in tiles with a source/destination accumulator.
353 /// Multiply groups of 4 adjacent pairs of signed 8-bit integers in src0 with
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".
358 /// \headerfile <immintrin.h>
360 /// This intrinsic corresponds to the <c> TDPBSSD </c> instruction.
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.
368 __DEFAULT_FN_ATTRS_INT8
369 static __inline__
void __tile_dpbssd(__tile1024i
*dst
, __tile1024i src0
,
371 dst
->tile
= _tile_dpbssd_internal(src0
.row
, src1
.col
, src0
.col
, dst
->tile
,
372 src0
.tile
, src1
.tile
);
375 /// Compute dot-product of bytes in tiles with a source/destination accumulator.
376 /// Multiply groups of 4 adjacent pairs of signed 8-bit integers in src0 with
377 /// corresponding unsigned 8-bit integers in src1, producing 4 intermediate
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".
381 /// \headerfile <immintrin.h>
383 /// This intrinsic corresponds to the <c> TDPBSUD </c> instruction.
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.
391 __DEFAULT_FN_ATTRS_INT8
392 static __inline__
void __tile_dpbsud(__tile1024i
*dst
, __tile1024i src0
,
394 dst
->tile
= _tile_dpbsud_internal(src0
.row
, src1
.col
, src0
.col
, dst
->tile
,
395 src0
.tile
, src1
.tile
);
398 /// Compute dot-product of bytes in tiles with a source/destination accumulator.
399 /// Multiply groups of 4 adjacent pairs of unsigned 8-bit integers in src0 with
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".
404 /// \headerfile <immintrin.h>
406 /// This intrinsic corresponds to the <c> TDPBUSD </c> instruction.
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.
414 __DEFAULT_FN_ATTRS_INT8
415 static __inline__
void __tile_dpbusd(__tile1024i
*dst
, __tile1024i src0
,
417 dst
->tile
= _tile_dpbusd_internal(src0
.row
, src1
.col
, src0
.col
, dst
->tile
,
418 src0
.tile
, src1
.tile
);
421 /// Compute dot-product of bytes in tiles with a source/destination accumulator.
422 /// Multiply groups of 4 adjacent pairs of unsigned 8-bit integers in src0 with
423 /// corresponding unsigned 8-bit integers in src1, producing 4 intermediate
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".
427 /// \headerfile <immintrin.h>
429 /// This intrinsic corresponds to the <c> TDPBUUD </c> instruction.
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.
437 __DEFAULT_FN_ATTRS_INT8
438 static __inline__
void __tile_dpbuud(__tile1024i
*dst
, __tile1024i src0
,
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
447 /// \headerfile <immintrin.h>
449 /// This intrinsic corresponds to the <c> TILESTORED </c> instruction.
452 /// A pointer to base address.
454 /// The stride between the rows' data to be stored in memory.
455 __DEFAULT_FN_ATTRS_TILE
456 static __inline__
void __tile_stored(void *base
, __SIZE_TYPE__ stride
,
458 _tile_stored_internal(src
.row
, src
.col
, base
, stride
, src
.tile
);
461 /// Zero the tile specified by "dst".
463 /// \headerfile <immintrin.h>
465 /// This intrinsic corresponds to the <c> TILEZERO </c> instruction.
468 /// The destination tile to be zero. Max size is 1024 Bytes.
469 __DEFAULT_FN_ATTRS_TILE
470 static __inline__
void __tile_zero(__tile1024i
*dst
) {
471 dst
->tile
= __builtin_ia32_tilezero_internal(dst
->row
, dst
->col
);
474 /// Compute dot-product of BF16 (16-bit) floating-point pairs in tiles src0 and
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
479 /// \headerfile <immintrin.h>
481 /// This intrinsic corresponds to the <c> TDPBF16PS </c> instruction.
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.
489 __DEFAULT_FN_ATTRS_BF16
490 static __inline__
void __tile_dpbf16ps(__tile1024i
*dst
, __tile1024i src0
,
492 dst
->tile
= _tile_dpbf16ps_internal(src0
.row
, src1
.col
, src0
.col
, dst
->tile
,
493 src0
.tile
, src1
.tile
);
496 /// Compute dot-product of FP16 (16-bit) floating-point pairs in tiles src0 and
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
501 /// \headerfile <immintrin.h>
503 /// This intrinsic corresponds to the <c> TDPFP16PS </c> instruction.
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.
511 __DEFAULT_FN_ATTRS_FP16
512 static __inline__
void __tile_dpfp16ps(__tile1024i
*dst
, __tile1024i src0
,
514 dst
->tile
= _tile_dpfp16ps_internal(src0
.row
, src1
.col
, src0
.col
, dst
->tile
,
515 src0
.tile
, src1
.tile
);
518 #undef __DEFAULT_FN_ATTRS_TILE
519 #undef __DEFAULT_FN_ATTRS_INT8
520 #undef __DEFAULT_FN_ATTRS_BF16
521 #undef __DEFAULT_FN_ATTRS_FP16
523 #endif /* __x86_64__ */
524 #endif /* __AMXINTRIN_H */