brintos

brintos / llvm-project-archived public Read only

0
0
Text · 6.8 KiB · 87ee8f3 Raw
168 lines · c
1/*===--------- amxcomplexintrin.h - AMXCOMPLEX 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 <amxcomplexintrin.h> directly; include <immintrin.h> instead."12#endif // __IMMINTRIN_H13 14#ifndef __AMX_COMPLEXINTRIN_H15#define __AMX_COMPLEXINTRIN_H16#ifdef __x86_64__17 18#define __DEFAULT_FN_ATTRS_COMPLEX                                             \19  __attribute__((__always_inline__, __nodebug__, __target__("amx-complex")))20 21/// Perform matrix multiplication of two tiles containing complex elements and22///    accumulate the results into a packed single precision tile. Each dword23///    element in input tiles \a a and \a b is interpreted as a complex number24///    with FP16 real part and FP16 imaginary part.25/// Calculates the imaginary part of the result. For each possible combination26///    of (row of \a a, column of \a b), it performs a set of multiplication27///    and accumulations on all corresponding complex numbers (one from \a a28///    and one from \a b). The imaginary part of the \a a element is multiplied29///    with the real part of the corresponding \a b element, and the real part30///    of the \a a element is multiplied with the imaginary part of the31///    corresponding \a b elements. The two accumulated results are added, and32///    then accumulated into the corresponding row and column of \a dst.33///34/// \headerfile <x86intrin.h>35///36/// \code37/// void _tile_cmmimfp16ps(__tile dst, __tile a, __tile b);38/// \endcode39///40/// \code{.operation}41/// FOR m := 0 TO dst.rows - 142///	tmp := dst.row[m]43///	FOR k := 0 TO (a.colsb / 4) - 144///		FOR n := 0 TO (dst.colsb / 4) - 145///			tmp.fp32[n] += FP32(a.row[m].fp16[2*k+0]) * FP32(b.row[k].fp16[2*n+1])46///			tmp.fp32[n] += FP32(a.row[m].fp16[2*k+1]) * FP32(b.row[k].fp16[2*n+0])47///		ENDFOR48///	ENDFOR49///	write_row_and_zero(dst, m, tmp, dst.colsb)50/// ENDFOR51/// zero_upper_rows(dst, dst.rows)52/// zero_tileconfig_start()53/// \endcode54///55/// This intrinsic corresponds to the \c TCMMIMFP16PS instruction.56///57/// \param dst58///    The destination tile. Max size is 1024 Bytes.59/// \param a60///    The 1st source tile. Max size is 1024 Bytes.61/// \param b62///    The 2nd source tile. Max size is 1024 Bytes.63#define _tile_cmmimfp16ps(dst, a, b) __builtin_ia32_tcmmimfp16ps(dst, a, b)64 65/// Perform matrix multiplication of two tiles containing complex elements and66///    accumulate the results into a packed single precision tile. Each dword67///    element in input tiles \a a and \a b is interpreted as a complex number68///    with FP16 real part and FP16 imaginary part.69/// Calculates the real part of the result. For each possible combination70///    of (row of \a a, column of \a b), it performs a set of multiplication71///    and accumulations on all corresponding complex numbers (one from \a a72///    and one from \a b). The real part of the \a a element is multiplied73///    with the real part of the corresponding \a b element, and the negated74///    imaginary part of the \a a element is multiplied with the imaginary75///    part of the corresponding \a b elements. The two accumulated results76///    are added, and then accumulated into the corresponding row and column77///    of \a dst.78///79/// \headerfile <x86intrin.h>80///81/// \code82/// void _tile_cmmrlfp16ps(__tile dst, __tile a, __tile b);83/// \endcode84///85/// \code{.operation}86/// FOR m := 0 TO dst.rows - 187///	tmp := dst.row[m]88///	FOR k := 0 TO (a.colsb / 4) - 189///		FOR n := 0 TO (dst.colsb / 4) - 190///			tmp.fp32[n] += FP32(a.row[m].fp16[2*k+0]) * FP32(b.row[k].fp16[2*n+0])91///			tmp.fp32[n] += FP32(-a.row[m].fp16[2*k+1]) * FP32(b.row[k].fp16[2*n+1])92///		ENDFOR93///	ENDFOR94///	write_row_and_zero(dst, m, tmp, dst.colsb)95/// ENDFOR96/// zero_upper_rows(dst, dst.rows)97/// zero_tileconfig_start()98/// \endcode99///100/// This intrinsic corresponds to the \c TCMMIMFP16PS instruction.101///102/// \param dst103///    The destination tile. Max size is 1024 Bytes.104/// \param a105///    The 1st source tile. Max size is 1024 Bytes.106/// \param b107///    The 2nd source tile. Max size is 1024 Bytes.108#define _tile_cmmrlfp16ps(dst, a, b) __builtin_ia32_tcmmrlfp16ps(dst, a, b)109 110static __inline__ _tile1024i __DEFAULT_FN_ATTRS_COMPLEX111_tile_cmmimfp16ps_internal(unsigned short m, unsigned short n, unsigned short k,112                           _tile1024i dst, _tile1024i src1, _tile1024i src2) {113  return __builtin_ia32_tcmmimfp16ps_internal(m, n, k, dst, src1, src2);114}115 116static __inline__ _tile1024i __DEFAULT_FN_ATTRS_COMPLEX117_tile_cmmrlfp16ps_internal(unsigned short m, unsigned short n, unsigned short k,118                           _tile1024i dst, _tile1024i src1, _tile1024i src2) {119  return __builtin_ia32_tcmmrlfp16ps_internal(m, n, k, dst, src1, src2);120}121 122/// Perform matrix multiplication of two tiles containing complex elements and123/// accumulate the results into a packed single precision tile. Each dword124/// element in input tiles src0 and src1 is interpreted as a complex number with125/// FP16 real part and FP16 imaginary part.126/// This function calculates the imaginary part of the result.127///128/// \headerfile <immintrin.h>129///130/// This intrinsic corresponds to the <c> TCMMIMFP16PS </c> instruction.131///132/// \param dst133///    The destination tile. Max size is 1024 Bytes.134/// \param src0135///    The 1st source tile. Max size is 1024 Bytes.136/// \param src1137///    The 2nd source tile. Max size is 1024 Bytes.138static __inline__ void __DEFAULT_FN_ATTRS_COMPLEX139__tile_cmmimfp16ps(__tile1024i *dst, __tile1024i src0, __tile1024i src1) {140  dst->tile = _tile_cmmimfp16ps_internal(src0.row, src1.col, src0.col,141                                         dst->tile, src0.tile, src1.tile);142}143 144/// Perform matrix multiplication of two tiles containing complex elements and145/// accumulate the results into a packed single precision tile. Each dword146/// element in input tiles src0 and src1 is interpreted as a complex number with147/// FP16 real part and FP16 imaginary part.148/// This function calculates the real part of the result.149///150/// \headerfile <immintrin.h>151///152/// This intrinsic corresponds to the <c> TCMMRLFP16PS </c> instruction.153///154/// \param dst155///    The destination tile. Max size is 1024 Bytes.156/// \param src0157///    The 1st source tile. Max size is 1024 Bytes.158/// \param src1159///    The 2nd source tile. Max size is 1024 Bytes.160static __inline__ void __DEFAULT_FN_ATTRS_COMPLEX161__tile_cmmrlfp16ps(__tile1024i *dst, __tile1024i src0, __tile1024i src1) {162  dst->tile = _tile_cmmrlfp16ps_internal(src0.row, src1.col, src0.col,163                                         dst->tile, src0.tile, src1.tile);164}165 166#endif // __x86_64__167#endif // __AMX_COMPLEXINTRIN_H168