xref: /aosp_15_r20/external/gemmlowp/internal/simd_wrappers_msa.h (revision 5f39d1b313f0528e11bae88b3029b54b9e1033e7)
1 // Copyright 2018 The Gemmlowp Authors. All Rights Reserved.
2 //
3 // Licensed under the Apache License, Version 2.0 (the "License");
4 // you may not use this file except in compliance with the License.
5 // You may obtain a copy of the License at
6 //
7 //     http://www.apache.org/licenses/LICENSE-2.0
8 //
9 // Unless required by applicable law or agreed to in writing, software
10 // distributed under the License is distributed on an "AS IS" BASIS,
11 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 // See the License for the specific language governing permissions and
13 // limitations under the License.
14 
15 // simd_wrappers_msa.h: MSA specialization of simd_wrappers.h
16 
17 #ifndef GEMMLOWP_INTERNAL_SIMD_WRAPPERS_MSA_H_
18 #define GEMMLOWP_INTERNAL_SIMD_WRAPPERS_MSA_H_
19 
20 #include <msa.h>
21 
22 namespace gemmlowp {
23 
24 using Int32x4 = v4i32;
25 using Int16x8 = v8i16;
26 using Uint8x16 = v16i8;
27 
28 template <int ScalarCount>
29 struct RegisterType<std::int32_t, ScalarCount> {
30   using Type =
31       typename std::conditional<ScalarCount >= 4, Int32x4, std::int32_t>::type;
32 };
33 
34 template <int ScalarCount>
35 struct RegisterType<std::int16_t, ScalarCount> {
36   using Type = typename std::conditional<ScalarCount >= 8, Int16x8, std::int16_t>::type;
37 };
38 
39 template <int ScalarCount>
40 struct RegisterType<std::uint8_t, ScalarCount> {
41   using Type = typename std::conditional<
42       ScalarCount >= 16, Uint8x16,
43       typename std::conditional<ScalarCount >= 4, std::uint32_t,
44                                 std::uint8_t>::type>::type;
45 };
46 
47 inline Int32x4 LoadInt32x4(const std::int32_t* src) {
48   return __builtin_msa_ld_w(const_cast<std::int32_t*>(src), 0);
49 }
50 
51 inline Int32x4 LoadInt32x4(const Int32x4* src) {
52   return __builtin_msa_ld_w(const_cast<Int32x4*>(src), 0);
53 }
54 
55 inline void StoreInt32x4(std::int32_t* dst, Int32x4 value) {
56   __builtin_msa_st_w(value, dst, 0);
57 }
58 
59 inline void StoreInt32x4(Int32x4* dst, Int32x4 value) {
60   __builtin_msa_st_w(value, dst, 0);
61 }
62 
63 inline Int16x8 LoadInt16x8(const std::int16_t* src) {
64   return __builtin_msa_ld_h(const_cast<std::int16_t*>(src), 0);
65 }
66 
67 inline Int16x8 LoadInt16x8(const Int16x8* src) {
68   return __builtin_msa_ld_h(const_cast<Int16x8*>(src), 0);
69 }
70 
71 inline void StoreInt16x8(std::int16_t* dst, Int16x8 value) { __builtin_msa_st_h(value, dst, 0); }
72 
73 inline void StoreInt16x8(Int16x8* dst, Int16x8 value) { __builtin_msa_st_h(value, dst, 0); }
74 
75 inline Uint8x16 LoadUint8x16(const std::uint8_t* src) {
76   return __builtin_msa_ld_b(const_cast<std::uint8_t*>(src), 0);
77 }
78 
79 inline Uint8x16 LoadUint8x16(const Uint8x16* src) {
80   return __builtin_msa_ld_b(const_cast<Uint8x16*>(src), 0);
81 }
82 
83 inline void StoreUint8x16(std::uint8_t* dst, Uint8x16 value) {
84   __builtin_msa_st_b(value, dst, 0);
85 }
86 
87 inline void StoreUint8x16(Uint8x16* dst, Uint8x16 value) {
88   __builtin_msa_st_b(value, dst, 0);
89 }
90 
91 template <int Lane>
92 std::int32_t GetLane(Int32x4 value) {
93   return __builtin_msa_copy_s_w(value, Lane);
94 }
95 
96 template <int Lane>
97 Int32x4 DupLane(Int32x4 value) {
98   static_assert(Lane >= 0 && Lane <= 3, "");
99   return __builtin_msa_splati_w(value, Lane);
100 }
101 
102 inline Int32x4 Mul(Int32x4 a, std::int32_t b) {
103   return __builtin_msa_mulv_w(a, __builtin_msa_fill_w(b));
104 }
105 
106 inline Int32x4 Min(Int32x4 a, Int32x4 b) { return __builtin_msa_min_s_w(a, b); }
107 
108 inline Int32x4 Max(Int32x4 a, Int32x4 b) { return __builtin_msa_max_s_w(a, b); }
109 
110 inline Int32x4 SaturatingRoundingDoublingHighMul(Int32x4 a, std::int32_t b) {
111   return __builtin_msa_mulr_q_w(a, __builtin_msa_fill_w(b));
112 }
113 
114 template <int Lane>
115 Int32x4 MulByRhsLane(Int32x4 a, Int32x4 b) {
116   static_assert(Lane >= 0 && Lane <= 3, "");
117   return __builtin_msa_mulv_w(a, __builtin_msa_splati_w(b, Lane));
118 }
119 
120 static inline v4i32 workaround_msa_maddv_w(v4i32 a, v4i32 b, v4i32 c) {
121   // Workaround for incorrect encoding of maddv.df in gcc (a exchanged with c).
122 #if 0
123   return __builtin_msa_maddv_w(a, b, c);
124 #else
125   asm volatile("maddv.w %w[a], %w[b], %w[c]\n"
126                // Outputs
127                : [a] "+f"(a)
128                // Inputs
129                : [b] "f"(b), [c] "f"(c));
130   return a;
131 #endif
132 }
133 
134 inline void MulAdd(Int32x4 lhs, Int32x4 rhs, Int32x4* acc) {
135   Int32x4 tmp = LoadInt32x4(acc);
136   tmp = workaround_msa_maddv_w(tmp, lhs, rhs);
137   StoreInt32x4(acc, tmp);
138 }
139 
140 inline void MulAdd(Int32x4 lhs, std::int32_t rhs, Int32x4* acc) {
141   Int32x4 tmp = LoadInt32x4(acc);
142   tmp = workaround_msa_maddv_w(tmp, lhs, __builtin_msa_fill_w(rhs));
143   StoreInt32x4(acc, tmp);
144 }
145 
146 template <int Lane>
147 inline void MulAddByRhsLane(Int32x4 lhs, Int32x4 rhs, Int32x4* acc) {
148   static_assert(Lane >= 0 && Lane <= 3, "");
149   Int32x4 tmp = LoadInt32x4(acc);
150   tmp = workaround_msa_maddv_w(tmp, lhs, __builtin_msa_splati_w(rhs, Lane));
151   StoreInt32x4(acc, tmp);
152 }
153 
154 template <>
155 struct LoadContiguousImpl<RegBlockUint8<8, 8>> {
156   static RegBlockUint8<8, 8> Run(const std::uint8_t* src) {
157     RegBlockUint8<8, 8> result;
158     for (int i = 0; i < 4; i++) {
159       result.buf.reg[i] = LoadUint8x16(src + 16 * i);
160     }
161     return result;
162   }
163 };
164 
165 template <>
166 struct LoadContiguousImpl<RegBlockInt32<8, 8>> {
167   static RegBlockInt32<8, 8> Run(const std::int32_t* src) {
168     RegBlockInt32<8, 8> result;
169     for (int i = 0; i < 16; i++) {
170       result.buf.reg[i] = LoadInt32x4(src + 4 * i);
171     }
172     return result;
173   }
174 };
175 
176 template <>
177 struct LoadContiguousImpl<RegBlockInt16<8, 8>> {
178   static RegBlockInt16<8, 8> Run(const std::int16_t* src) {
179     RegBlockInt16<8, 8> result;
180     for (int i = 0; i < 8; i++) {
181       result.buf.reg[i] = LoadInt16x8(src + 8 * i);
182     }
183     return result;
184   }
185 };
186 
187 }  // end namespace gemmlowp
188 
189 #include "simd_wrappers_common_neon_sse.h"
190 
191 #endif  // GEMMLOWP_INTERNAL_SIMD_WRAPPERS_MSA_H_
192