• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 =
37       typename std::conditional<ScalarCount >= 8, Int16x8, std::int16_t>::type;
38 };
39 
40 template <int ScalarCount>
41 struct RegisterType<std::uint8_t, ScalarCount> {
42   using Type = typename std::conditional<
43       ScalarCount >= 16, Uint8x16,
44       typename std::conditional<ScalarCount >= 4, std::uint32_t,
45                                 std::uint8_t>::type>::type;
46 };
47 
48 inline Int32x4 LoadInt32x4(const std::int32_t* src) {
49   return __builtin_msa_ld_w(const_cast<std::int32_t*>(src), 0);
50 }
51 
52 inline Int32x4 LoadInt32x4(const Int32x4* src) {
53   return __builtin_msa_ld_w(const_cast<Int32x4*>(src), 0);
54 }
55 
56 inline void StoreInt32x4(std::int32_t* dst, Int32x4 value) {
57   __builtin_msa_st_w(value, dst, 0);
58 }
59 
60 inline void StoreInt32x4(Int32x4* dst, Int32x4 value) {
61   __builtin_msa_st_w(value, dst, 0);
62 }
63 
64 inline Int16x8 LoadInt16x8(const std::int16_t* src) {
65   return __builtin_msa_ld_h(const_cast<std::int16_t*>(src), 0);
66 }
67 
68 inline Int16x8 LoadInt16x8(const Int16x8* src) {
69   return __builtin_msa_ld_h(const_cast<Int16x8*>(src), 0);
70 }
71 
72 inline void StoreInt16x8(std::int16_t* dst, Int16x8 value) {
73   __builtin_msa_st_h(value, dst, 0);
74 }
75 
76 inline void StoreInt16x8(Int16x8* dst, Int16x8 value) {
77   __builtin_msa_st_h(value, dst, 0);
78 }
79 
80 inline Uint8x16 LoadUint8x16(const std::uint8_t* src) {
81   return __builtin_msa_ld_b(const_cast<std::uint8_t*>(src), 0);
82 }
83 
84 inline Uint8x16 LoadUint8x16(const Uint8x16* src) {
85   return __builtin_msa_ld_b(const_cast<Uint8x16*>(src), 0);
86 }
87 
88 inline void StoreUint8x16(std::uint8_t* dst, Uint8x16 value) {
89   __builtin_msa_st_b(value, dst, 0);
90 }
91 
92 inline void StoreUint8x16(Uint8x16* dst, Uint8x16 value) {
93   __builtin_msa_st_b(value, dst, 0);
94 }
95 
96 template <int Lane>
97 std::int32_t GetLane(Int32x4 value) {
98   return __builtin_msa_copy_s_w(value, Lane);
99 }
100 
101 template <int Lane>
102 Int32x4 DupLane(Int32x4 value) {
103   static_assert(Lane >= 0 && Lane <= 3, "");
104   return __builtin_msa_splati_w(value, Lane);
105 }
106 
107 inline Int32x4 Mul(Int32x4 a, std::int32_t b) {
108   return __builtin_msa_mulv_w(a, __builtin_msa_fill_w(b));
109 }
110 
111 inline Int32x4 Min(Int32x4 a, Int32x4 b) { return __builtin_msa_min_s_w(a, b); }
112 
113 inline Int32x4 Max(Int32x4 a, Int32x4 b) { return __builtin_msa_max_s_w(a, b); }
114 
115 inline Int32x4 SaturatingRoundingDoublingHighMul(Int32x4 a, std::int32_t b) {
116   return __builtin_msa_mulr_q_w(a, __builtin_msa_fill_w(b));
117 }
118 
119 template <int Lane>
120 Int32x4 MulByRhsLane(Int32x4 a, Int32x4 b) {
121   static_assert(Lane >= 0 && Lane <= 3, "");
122   return __builtin_msa_mulv_w(a, __builtin_msa_splati_w(b, Lane));
123 }
124 
125 static inline v4i32 workaround_msa_maddv_w(v4i32 a, v4i32 b, v4i32 c) {
126   // Workaround for incorrect encoding of maddv.df in gcc (a exchanged with c).
127 #if 0
128   return __builtin_msa_maddv_w(a, b, c);
129 #else
130   asm volatile("maddv.w %w[a], %w[b], %w[c]\n"
131                // Outputs
132                : [a] "+f"(a)
133                // Inputs
134                : [b] "f"(b), [c] "f"(c));
135   return a;
136 #endif
137 }
138 
139 inline void MulAdd(Int32x4 lhs, Int32x4 rhs, Int32x4* acc) {
140   Int32x4 tmp = LoadInt32x4(acc);
141   tmp = workaround_msa_maddv_w(tmp, lhs, rhs);
142   StoreInt32x4(acc, tmp);
143 }
144 
145 inline void MulAdd(Int32x4 lhs, std::int32_t rhs, Int32x4* acc) {
146   Int32x4 tmp = LoadInt32x4(acc);
147   tmp = workaround_msa_maddv_w(tmp, lhs, __builtin_msa_fill_w(rhs));
148   StoreInt32x4(acc, tmp);
149 }
150 
151 template <int Lane>
152 inline void MulAddByRhsLane(Int32x4 lhs, Int32x4 rhs, Int32x4* acc) {
153   static_assert(Lane >= 0 && Lane <= 3, "");
154   Int32x4 tmp = LoadInt32x4(acc);
155   tmp = workaround_msa_maddv_w(tmp, lhs, __builtin_msa_splati_w(rhs, Lane));
156   StoreInt32x4(acc, tmp);
157 }
158 
159 template <>
160 struct LoadContiguousImpl<RegBlockUint8<8, 8>> {
161   static RegBlockUint8<8, 8> Run(const std::uint8_t* src) {
162     RegBlockUint8<8, 8> result;
163     for (int i = 0; i < 4; i++) {
164       result.buf.reg[i] = LoadUint8x16(src + 16 * i);
165     }
166     return result;
167   }
168 };
169 
170 template <>
171 struct LoadContiguousImpl<RegBlockInt32<8, 8>> {
172   static RegBlockInt32<8, 8> Run(const std::int32_t* src) {
173     RegBlockInt32<8, 8> result;
174     for (int i = 0; i < 16; i++) {
175       result.buf.reg[i] = LoadInt32x4(src + 4 * i);
176     }
177     return result;
178   }
179 };
180 
181 template <>
182 struct LoadContiguousImpl<RegBlockInt16<8, 8>> {
183   static RegBlockInt16<8, 8> Run(const std::int16_t* src) {
184     RegBlockInt16<8, 8> result;
185     for (int i = 0; i < 8; i++) {
186       result.buf.reg[i] = LoadInt16x8(src + 8 * i);
187     }
188     return result;
189   }
190 };
191 
192 }  // end namespace gemmlowp
193 
194 #include "simd_wrappers_common_neon_sse.h"
195 
196 #endif  // GEMMLOWP_INTERNAL_SIMD_WRAPPERS_MSA_H_
197