94 lines · c
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-exception6 *7 *===------------------------------------------------------------------------===8 */9 10#ifndef __IMMINTRIN_H11#error "Never use <amxfp16intrin.h> directly; use <immintrin.h> instead."12#endif /* __IMMINTRIN_H */13 14#ifndef __AMX_FP16INTRIN_H15#define __AMX_FP16INTRIN_H16#ifdef __x86_64__17 18/* Define the default attributes for the functions in this file. */19#define __DEFAULT_FN_ATTRS \20 __attribute__((__always_inline__, __nodebug__, __target__("amx-fp16")))21 22/// Compute dot-product of FP16 (16-bit) floating-point pairs in tiles \a a23/// and \a b, accumulating the intermediate single-precision (32-bit)24/// floating-point elements with elements in \a dst, and store the 32-bit25/// result back to tile \a dst.26///27/// \headerfile <immintrin.h>28///29/// \code30/// void _tile_dpfp16ps (__tile dst, __tile a, __tile b)31/// \endcode32///33/// \code{.operation}34/// FOR m := 0 TO dst.rows - 135/// tmp := dst.row[m]36/// FOR k := 0 TO (a.colsb / 4) - 137/// FOR n := 0 TO (dst.colsb / 4) - 138/// tmp.fp32[n] += FP32(a.row[m].fp16[2*k+0]) *39/// FP32(b.row[k].fp16[2*n+0])40/// tmp.fp32[n] += FP32(a.row[m].fp16[2*k+1]) *41/// FP32(b.row[k].fp16[2*n+1])42/// ENDFOR43/// ENDFOR44/// write_row_and_zero(dst, m, tmp, dst.colsb)45/// ENDFOR46/// zero_upper_rows(dst, dst.rows)47/// zero_tileconfig_start()48/// \endcode49///50/// This intrinsic corresponds to the \c TDPFP16PS instruction.51///52/// \param dst53/// The destination tile. Max size is 1024 Bytes.54/// \param a55/// The 1st source tile. Max size is 1024 Bytes.56/// \param b57/// The 2nd source tile. Max size is 1024 Bytes.58#define _tile_dpfp16ps(dst, a, b) \59 __builtin_ia32_tdpfp16ps(dst, a, b)60 61/// This is internal intrinsic. C/C++ user should avoid calling it directly.62static __inline__ _tile1024i __DEFAULT_FN_ATTRS63_tile_dpfp16ps_internal(unsigned short m, unsigned short n, unsigned short k,64 _tile1024i dst, _tile1024i src1, _tile1024i src2) {65 return __builtin_ia32_tdpfp16ps_internal(m, n, k, dst, src1, src2);66}67 68/// Compute dot-product of FP16 (16-bit) floating-point pairs in tiles src0 and69/// src1, accumulating the intermediate single-precision (32-bit) floating-point70/// elements with elements in "dst", and store the 32-bit result back to tile71/// "dst".72///73/// \headerfile <immintrin.h>74///75/// This intrinsic corresponds to the <c> TDPFP16PS </c> instruction.76///77/// \param dst78/// The destination tile. Max size is 1024 Bytes.79/// \param src080/// The 1st source tile. Max size is 1024 Bytes.81/// \param src182/// The 2nd source tile. Max size is 1024 Bytes.83__DEFAULT_FN_ATTRS84static __inline__ void __tile_dpfp16ps(__tile1024i *dst, __tile1024i src0,85 __tile1024i src1) {86 dst->tile = _tile_dpfp16ps_internal(src0.row, src1.col, src0.col, dst->tile,87 src0.tile, src1.tile);88}89 90#undef __DEFAULT_FN_ATTRS91 92#endif /* __x86_64__ */93#endif /* __AMX_FP16INTRIN_H */94