clang 20.0.0git
amxfp16intrin.h
Go to the documentation of this file.
1/*===------------- amxfp16intrin.h - AMX_FP16 intrinsics -*- C++ -*---------===
2 *
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
6 *
7 *===------------------------------------------------------------------------===
8 */
9
10#ifndef __IMMINTRIN_H
11#error "Never use <amxfp16intrin.h> directly; use <immintrin.h> instead."
12#endif /* __IMMINTRIN_H */
13
14#ifndef __AMX_FP16INTRIN_H
15#define __AMX_FP16INTRIN_H
16#ifdef __x86_64__
17
18/// Compute dot-product of FP16 (16-bit) floating-point pairs in tiles \a a
19/// and \a b, accumulating the intermediate single-precision (32-bit)
20/// floating-point elements with elements in \a dst, and store the 32-bit
21/// result back to tile \a dst.
22///
23/// \headerfile <immintrin.h>
24///
25/// \code
26/// void _tile_dpfp16ps (__tile dst, __tile a, __tile b)
27/// \endcode
28///
29/// \code{.operation}
30/// FOR m := 0 TO dst.rows - 1
31/// tmp := dst.row[m]
32/// FOR k := 0 TO (a.colsb / 4) - 1
33/// FOR n := 0 TO (dst.colsb / 4) - 1
34/// tmp.fp32[n] += FP32(a.row[m].fp16[2*k+0]) *
35/// FP32(b.row[k].fp16[2*n+0])
36/// tmp.fp32[n] += FP32(a.row[m].fp16[2*k+1]) *
37/// FP32(b.row[k].fp16[2*n+1])
38/// ENDFOR
39/// ENDFOR
40/// write_row_and_zero(dst, m, tmp, dst.colsb)
41/// ENDFOR
42/// zero_upper_rows(dst, dst.rows)
43/// zero_tileconfig_start()
44/// \endcode
45///
46/// This intrinsic corresponds to the \c TDPFP16PS instruction.
47///
48/// \param dst
49/// The destination tile. Max size is 1024 Bytes.
50/// \param a
51/// The 1st source tile. Max size is 1024 Bytes.
52/// \param b
53/// The 2nd source tile. Max size is 1024 Bytes.
54#define _tile_dpfp16ps(dst, a, b) \
55 __builtin_ia32_tdpfp16ps(dst, a, b)
56
57#endif /* __x86_64__ */
58#endif /* __AMX_FP16INTRIN_H */