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-exception
6  *
7  *===------------------------------------------------------------------------===
8  */
9 
10 #ifndef __IMMINTRIN_H
11 #error "Never use <amxcomplexintrin.h> directly; include <immintrin.h> instead."
12 #endif // __IMMINTRIN_H
13 
14 #ifndef __AMX_COMPLEXINTRIN_H
15 #define __AMX_COMPLEXINTRIN_H
16 #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 and
22 ///    accumulate the results into a packed single precision tile. Each dword
23 ///    element in input tiles \a a and \a b is interpreted as a complex number
24 ///    with FP16 real part and FP16 imaginary part.
25 /// Calculates the imaginary part of the result. For each possible combination
26 ///    of (row of \a a, column of \a b), it performs a set of multiplication
27 ///    and accumulations on all corresponding complex numbers (one from \a a
28 ///    and one from \a b). The imaginary part of the \a a element is multiplied
29 ///    with the real part of the corresponding \a b element, and the real part
30 ///    of the \a a element is multiplied with the imaginary part of the
31 ///    corresponding \a b elements. The two accumulated results are added, and
32 ///    then accumulated into the corresponding row and column of \a dst.
33 ///
34 /// \headerfile <x86intrin.h>
35 ///
36 /// \code
37 /// void _tile_cmmimfp16ps(__tile dst, __tile a, __tile b);
38 /// \endcode
39 ///
40 /// \code{.operation}
41 /// FOR m := 0 TO dst.rows - 1
42 ///	tmp := dst.row[m]
43 ///	FOR k := 0 TO (a.colsb / 4) - 1
44 ///		FOR n := 0 TO (dst.colsb / 4) - 1
45 ///			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 ///		ENDFOR
48 ///	ENDFOR
49 ///	write_row_and_zero(dst, m, tmp, dst.colsb)
50 /// ENDFOR
51 /// zero_upper_rows(dst, dst.rows)
52 /// zero_tileconfig_start()
53 /// \endcode
54 ///
55 /// This intrinsic corresponds to the \c TCMMIMFP16PS instruction.
56 ///
57 /// \param dst
58 ///    The destination tile. Max size is 1024 Bytes.
59 /// \param a
60 ///    The 1st source tile. Max size is 1024 Bytes.
61 /// \param b
62 ///    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 and
66 ///    accumulate the results into a packed single precision tile. Each dword
67 ///    element in input tiles \a a and \a b is interpreted as a complex number
68 ///    with FP16 real part and FP16 imaginary part.
69 /// Calculates the real part of the result. For each possible combination
70 ///    of (row of \a a, column of \a b), it performs a set of multiplication
71 ///    and accumulations on all corresponding complex numbers (one from \a a
72 ///    and one from \a b). The real part of the \a a element is multiplied
73 ///    with the real part of the corresponding \a b element, and the negated
74 ///    imaginary part of the \a a element is multiplied with the imaginary
75 ///    part of the corresponding \a b elements. The two accumulated results
76 ///    are added, and then accumulated into the corresponding row and column
77 ///    of \a dst.
78 ///
79 /// \headerfile <x86intrin.h>
80 ///
81 /// \code
82 /// void _tile_cmmrlfp16ps(__tile dst, __tile a, __tile b);
83 /// \endcode
84 ///
85 /// \code{.operation}
86 /// FOR m := 0 TO dst.rows - 1
87 ///	tmp := dst.row[m]
88 ///	FOR k := 0 TO (a.colsb / 4) - 1
89 ///		FOR n := 0 TO (dst.colsb / 4) - 1
90 ///			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 ///		ENDFOR
93 ///	ENDFOR
94 ///	write_row_and_zero(dst, m, tmp, dst.colsb)
95 /// ENDFOR
96 /// zero_upper_rows(dst, dst.rows)
97 /// zero_tileconfig_start()
98 /// \endcode
99 ///
100 /// This intrinsic corresponds to the \c TCMMIMFP16PS instruction.
101 ///
102 /// \param dst
103 ///    The destination tile. Max size is 1024 Bytes.
104 /// \param a
105 ///    The 1st source tile. Max size is 1024 Bytes.
106 /// \param b
107 ///    The 2nd source tile. Max size is 1024 Bytes.
108 #define _tile_cmmrlfp16ps(dst, a, b) __builtin_ia32_tcmmrlfp16ps(dst, a, b)
109 
110 static __inline__ _tile1024i __DEFAULT_FN_ATTRS_COMPLEX
111 _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 
116 static __inline__ _tile1024i __DEFAULT_FN_ATTRS_COMPLEX
117 _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 and
123 /// accumulate the results into a packed single precision tile. Each dword
124 /// element in input tiles src0 and src1 is interpreted as a complex number with
125 /// 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 dst
133 ///    The destination tile. Max size is 1024 Bytes.
134 /// \param src0
135 ///    The 1st source tile. Max size is 1024 Bytes.
136 /// \param src1
137 ///    The 2nd source tile. Max size is 1024 Bytes.
138 __DEFAULT_FN_ATTRS_COMPLEX
139 static void __tile_cmmimfp16ps(__tile1024i *dst, __tile1024i src0,
140                                __tile1024i src1) {
141   dst->tile = _tile_cmmimfp16ps_internal(src0.row, src1.col, src0.col,
142                                          dst->tile, src0.tile, src1.tile);
143 }
144 
145 /// Perform matrix multiplication of two tiles containing complex elements and
146 /// accumulate the results into a packed single precision tile. Each dword
147 /// element in input tiles src0 and src1 is interpreted as a complex number with
148 /// FP16 real part and FP16 imaginary part.
149 /// This function calculates the real part of the result.
150 ///
151 /// \headerfile <immintrin.h>
152 ///
153 /// This intrinsic corresponds to the <c> TCMMRLFP16PS </c> instruction.
154 ///
155 /// \param dst
156 ///    The destination tile. Max size is 1024 Bytes.
157 /// \param src0
158 ///    The 1st source tile. Max size is 1024 Bytes.
159 /// \param src1
160 ///    The 2nd source tile. Max size is 1024 Bytes.
161 __DEFAULT_FN_ATTRS_COMPLEX
162 static void __tile_cmmrlfp16ps(__tile1024i *dst, __tile1024i src0,
163                                __tile1024i src1) {
164   dst->tile = _tile_cmmrlfp16ps_internal(src0.row, src1.col, src0.col,
165                                          dst->tile, src0.tile, src1.tile);
166 }
167 
168 #endif // __x86_64__
169 #endif // __AMX_COMPLEXINTRIN_H
170