• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (C) 2017 The Android Open Source Project
3  *
4  * Licensed under the Apache License, Version 2.0 (the "License");
5  * you may not use this file except in compliance with the License.
6  * You may obtain a copy of the License at
7  *
8  *      http://www.apache.org/licenses/LICENSE-2.0
9  *
10  * Unless required by applicable law or agreed to in writing, software
11  * distributed under the License is distributed on an "AS IS" BASIS,
12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16 
17 /**
18  * Tests for halving-add idiomatic vectorization.
19  */
20 public class Main {
21 
22   private static final int N = 256;
23   private static final int M = N * N + 15;
24 
25   static byte[] sB1 = new byte[M];
26   static byte[] sB2 = new byte[M];
27   static byte[] sBo = new byte[M];
28 
29   /// CHECK-START: void Main.halving_add_signed(byte[], byte[], byte[]) loop_optimization (before)
30   /// CHECK-DAG: <<I1:i\d+>>   IntConstant 1                       loop:none
31   /// CHECK-DAG: <<Phi:i\d+>>  Phi                                 loop:<<Loop:B\d+>> outer_loop:none
32   /// CHECK-DAG: <<Get1:b\d+>> ArrayGet                            loop:<<Loop>>      outer_loop:none
33   /// CHECK-DAG: <<Get2:b\d+>> ArrayGet                            loop:<<Loop>>      outer_loop:none
34   /// CHECK-DAG: <<Add:i\d+>>  Add [<<Get1>>,<<Get2>>]             loop:<<Loop>>      outer_loop:none
35   /// CHECK-DAG: <<Shr:i\d+>>  Shr [<<Add>>,<<I1>>]                loop:<<Loop>>      outer_loop:none
36   /// CHECK-DAG: <<Cnv:b\d+>>  TypeConversion [<<Shr>>]            loop:<<Loop>>      outer_loop:none
37   /// CHECK-DAG:               ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>>      outer_loop:none
38   //
39   /// CHECK-START-ARM64: void Main.halving_add_signed(byte[], byte[], byte[]) loop_optimization (after)
40   /// CHECK-DAG: <<Phi:i\d+>>  Phi                                  loop:<<Loop:B\d+>> outer_loop:none
41   /// CHECK-DAG: <<Get1:d\d+>> VecLoad                              loop:<<Loop>>      outer_loop:none
42   /// CHECK-DAG: <<Get2:d\d+>> VecLoad                              loop:<<Loop>>      outer_loop:none
43   /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:false loop:<<Loop>> outer_loop:none
44   /// CHECK-DAG:               VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>>      outer_loop:none
halving_add_signed(byte[] b1, byte[] b2, byte[] bo)45   private static void halving_add_signed(byte[] b1, byte[] b2, byte[] bo) {
46     int min_length = Math.min(bo.length, Math.min(b1.length, b2.length));
47     for (int i = 0; i < min_length; i++) {
48       bo[i] = (byte) ((b1[i] + b2[i]) >> 1);
49     }
50   }
51 
52   /// CHECK-START: void Main.halving_add_unsigned(byte[], byte[], byte[]) loop_optimization (before)
53   /// CHECK-DAG: <<I1:i\d+>>   IntConstant 1                       loop:none
54   /// CHECK-DAG: <<I255:i\d+>> IntConstant 255                     loop:none
55   /// CHECK-DAG: <<Phi:i\d+>>  Phi                                 loop:<<Loop:B\d+>> outer_loop:none
56   /// CHECK-DAG: <<Get1:b\d+>> ArrayGet                            loop:<<Loop>>      outer_loop:none
57   /// CHECK-DAG: <<Get2:b\d+>> ArrayGet                            loop:<<Loop>>      outer_loop:none
58   /// CHECK-DAG: <<And1:i\d+>> And [<<Get1>>,<<I255>>]             loop:<<Loop>>      outer_loop:none
59   /// CHECK-DAG: <<And2:i\d+>> And [<<Get2>>,<<I255>>]             loop:<<Loop>>      outer_loop:none
60   /// CHECK-DAG: <<Add:i\d+>>  Add [<<And1>>,<<And2>>]             loop:<<Loop>>      outer_loop:none
61   /// CHECK-DAG: <<Shr:i\d+>>  Shr [<<Add>>,<<I1>>]                loop:<<Loop>>      outer_loop:none
62   /// CHECK-DAG: <<Cnv:b\d+>>  TypeConversion [<<Shr>>]            loop:<<Loop>>      outer_loop:none
63   /// CHECK-DAG:               ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>>      outer_loop:none
64   //
65   /// CHECK-START-ARM64: void Main.halving_add_unsigned(byte[], byte[], byte[]) loop_optimization (after)
66   /// CHECK-DAG: <<Phi:i\d+>>  Phi                                  loop:<<Loop:B\d+>> outer_loop:none
67   /// CHECK-DAG: <<Get1:d\d+>> VecLoad                              loop:<<Loop>>      outer_loop:none
68   /// CHECK-DAG: <<Get2:d\d+>> VecLoad                              loop:<<Loop>>      outer_loop:none
69   /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:true rounded:false loop:<<Loop>> outer_loop:none
70   /// CHECK-DAG:               VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>>      outer_loop:none
halving_add_unsigned(byte[] b1, byte[] b2, byte[] bo)71   private static void halving_add_unsigned(byte[] b1, byte[] b2, byte[] bo) {
72     int min_length = Math.min(bo.length, Math.min(b1.length, b2.length));
73     for (int i = 0; i < min_length; i++) {
74       bo[i] = (byte) (((b1[i] & 0xff) + (b2[i] & 0xff)) >> 1);
75     }
76   }
77 
78   /// CHECK-START: void Main.rounding_halving_add_signed(byte[], byte[], byte[]) loop_optimization (before)
79   /// CHECK-DAG: <<I1:i\d+>>   IntConstant 1                       loop:none
80   /// CHECK-DAG: <<Phi:i\d+>>  Phi                                 loop:<<Loop:B\d+>> outer_loop:none
81   /// CHECK-DAG: <<Get1:b\d+>> ArrayGet                            loop:<<Loop>>      outer_loop:none
82   /// CHECK-DAG: <<Get2:b\d+>> ArrayGet                            loop:<<Loop>>      outer_loop:none
83   /// CHECK-DAG: <<Add1:i\d+>> Add [<<Get1>>,<<Get2>>]             loop:<<Loop>>      outer_loop:none
84   /// CHECK-DAG: <<Add2:i\d+>> Add [<<Add1>>,<<I1>>]               loop:<<Loop>>      outer_loop:none
85   /// CHECK-DAG: <<Shr:i\d+>>  Shr [<<Add2>>,<<I1>>]               loop:<<Loop>>      outer_loop:none
86   /// CHECK-DAG: <<Cnv:b\d+>>  TypeConversion [<<Shr>>]            loop:<<Loop>>      outer_loop:none
87   /// CHECK-DAG:               ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>>      outer_loop:none
88   //
89   /// CHECK-START-ARM64: void Main.rounding_halving_add_signed(byte[], byte[], byte[]) loop_optimization (after)
90   /// CHECK-DAG: <<Phi:i\d+>>  Phi                                  loop:<<Loop:B\d+>> outer_loop:none
91   /// CHECK-DAG: <<Get1:d\d+>> VecLoad                              loop:<<Loop>>      outer_loop:none
92   /// CHECK-DAG: <<Get2:d\d+>> VecLoad                              loop:<<Loop>>      outer_loop:none
93   /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>] unsigned:false rounded:true loop:<<Loop>> outer_loop:none
94   /// CHECK-DAG:               VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>>      outer_loop:none
rounding_halving_add_signed(byte[] b1, byte[] b2, byte[] bo)95   private static void rounding_halving_add_signed(byte[] b1, byte[] b2, byte[] bo) {
96     int min_length = Math.min(bo.length, Math.min(b1.length, b2.length));
97     for (int i = 0; i < min_length; i++) {
98       bo[i] = (byte) ((b1[i] + b2[i] + 1) >> 1);
99     }
100   }
101 
102   /// CHECK-START: void Main.rounding_halving_add_unsigned(byte[], byte[], byte[]) loop_optimization (before)
103   /// CHECK-DAG: <<I1:i\d+>>   IntConstant 1                       loop:none
104   /// CHECK-DAG: <<I255:i\d+>> IntConstant 255                     loop:none
105   /// CHECK-DAG: <<Phi:i\d+>>  Phi                                 loop:<<Loop:B\d+>> outer_loop:none
106   /// CHECK-DAG: <<Get1:b\d+>> ArrayGet                            loop:<<Loop>>      outer_loop:none
107   /// CHECK-DAG: <<Get2:b\d+>> ArrayGet                            loop:<<Loop>>      outer_loop:none
108   /// CHECK-DAG: <<And1:i\d+>> And [<<Get1>>,<<I255>>]             loop:<<Loop>>      outer_loop:none
109   /// CHECK-DAG: <<And2:i\d+>> And [<<Get2>>,<<I255>>]             loop:<<Loop>>      outer_loop:none
110   /// CHECK-DAG: <<Add1:i\d+>> Add [<<And1>>,<<And2>>]             loop:<<Loop>>      outer_loop:none
111   /// CHECK-DAG: <<Add2:i\d+>> Add [<<Add1>>,<<I1>>]               loop:<<Loop>>      outer_loop:none
112   /// CHECK-DAG: <<Shr:i\d+>>  Shr [<<Add2>>,<<I1>>]               loop:<<Loop>>      outer_loop:none
113   /// CHECK-DAG: <<Cnv:b\d+>>  TypeConversion [<<Shr>>]            loop:<<Loop>>      outer_loop:none
114   /// CHECK-DAG:               ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>>      outer_loop:none
115   //
116   /// CHECK-START-ARM64: void Main.rounding_halving_add_unsigned(byte[], byte[], byte[]) loop_optimization (after)
117   /// CHECK-DAG: <<Phi:i\d+>>  Phi                                  loop:<<Loop:B\d+>> outer_loop:none
118   /// CHECK-DAG: <<Get1:d\d+>> VecLoad                              loop:<<Loop>>      outer_loop:none
119   /// CHECK-DAG: <<Get2:d\d+>> VecLoad                              loop:<<Loop>>      outer_loop:none
120   /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get1>>,<<Get2>>]  unsigned:true rounded:true loop:<<Loop>> outer_loop:none
121   /// CHECK-DAG:               VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>>      outer_loop:none
rounding_halving_add_unsigned(byte[] b1, byte[] b2, byte[] bo)122   private static void rounding_halving_add_unsigned(byte[] b1, byte[] b2, byte[] bo) {
123     int min_length = Math.min(bo.length, Math.min(b1.length, b2.length));
124     for (int i = 0; i < min_length; i++) {
125       bo[i] = (byte) (((b1[i] & 0xff) + (b2[i] & 0xff) + 1) >> 1);
126     }
127   }
128 
129   /// CHECK-START: void Main.halving_add_signed_constant(byte[], byte[]) loop_optimization (before)
130   /// CHECK-DAG: <<I1:i\d+>>   IntConstant 1                       loop:none
131   /// CHECK-DAG: <<I127:i\d+>> IntConstant 127                     loop:none
132   /// CHECK-DAG: <<Phi:i\d+>>  Phi                                 loop:<<Loop:B\d+>> outer_loop:none
133   /// CHECK-DAG: <<Get:b\d+>>  ArrayGet                            loop:<<Loop>>      outer_loop:none
134   /// CHECK-DAG: <<Add:i\d+>>  Add [<<Get>>,<<I127>>]              loop:<<Loop>>      outer_loop:none
135   /// CHECK-DAG: <<Shr:i\d+>>  Shr [<<Add>>,<<I1>>]                loop:<<Loop>>      outer_loop:none
136   /// CHECK-DAG: <<Cnv:b\d+>>  TypeConversion [<<Shr>>]            loop:<<Loop>>      outer_loop:none
137   /// CHECK-DAG:               ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>>      outer_loop:none
138   //
139   /// CHECK-START-ARM64: void Main.halving_add_signed_constant(byte[], byte[]) loop_optimization (after)
140   /// CHECK-DAG: <<I127:i\d+>> IntConstant 127                      loop:none
141   /// CHECK-DAG: <<Repl:d\d+>> VecReplicateScalar [<<I127>>]        loop:none
142   /// CHECK-DAG: <<Phi:i\d+>>  Phi                                  loop:<<Loop:B\d+>> outer_loop:none
143   /// CHECK-DAG: <<Get:d\d+>>  VecLoad                              loop:<<Loop>>      outer_loop:none
144   /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get>>,<<Repl>>] unsigned:false rounded:false loop:<<Loop>> outer_loop:none
145   /// CHECK-DAG:               VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>>      outer_loop:none
halving_add_signed_constant(byte[] b1, byte[] bo)146   private static void halving_add_signed_constant(byte[] b1, byte[] bo) {
147     int min_length = Math.min(bo.length, b1.length);
148     for (int i = 0; i < min_length; i++) {
149       bo[i] = (byte) ((b1[i] + 0x7f) >> 1);
150     }
151   }
152 
153   /// CHECK-START: void Main.halving_add_unsigned_constant(byte[], byte[]) loop_optimization (before)
154   /// CHECK-DAG: <<I1:i\d+>>   IntConstant 1                       loop:none
155   /// CHECK-DAG: <<I255:i\d+>> IntConstant 255                     loop:none
156   /// CHECK-DAG: <<Phi:i\d+>>  Phi                                 loop:<<Loop:B\d+>> outer_loop:none
157   /// CHECK-DAG: <<Get:b\d+>>  ArrayGet                            loop:<<Loop>>      outer_loop:none
158   /// CHECK-DAG: <<And:i\d+>>  And [<<Get>>,<<I255>>]              loop:<<Loop>>      outer_loop:none
159   /// CHECK-DAG: <<Add:i\d+>>  Add [<<And>>,<<I255>>]              loop:<<Loop>>      outer_loop:none
160   /// CHECK-DAG: <<Shr:i\d+>>  Shr [<<Add>>,<<I1>>]                loop:<<Loop>>      outer_loop:none
161   /// CHECK-DAG: <<Cnv:b\d+>>  TypeConversion [<<Shr>>]            loop:<<Loop>>      outer_loop:none
162   /// CHECK-DAG:               ArraySet [{{l\d+}},<<Phi>>,<<Cnv>>] loop:<<Loop>>      outer_loop:none
163   //
164   /// CHECK-START-ARM64: void Main.halving_add_unsigned_constant(byte[], byte[]) loop_optimization (after)
165   /// CHECK-DAG: <<I255:i\d+>> IntConstant 255                      loop:none
166   /// CHECK-DAG: <<Repl:d\d+>> VecReplicateScalar [<<I255>>]        loop:none
167   /// CHECK-DAG: <<Phi:i\d+>>  Phi                                  loop:<<Loop:B\d+>> outer_loop:none
168   /// CHECK-DAG: <<Get:d\d+>>  VecLoad                              loop:<<Loop>>      outer_loop:none
169   /// CHECK-DAG: <<HAdd:d\d+>> VecHalvingAdd [<<Get>>,<<Repl>>] unsigned:true rounded:false loop:<<Loop>> outer_loop:none
170   /// CHECK-DAG:               VecStore [{{l\d+}},<<Phi>>,<<HAdd>>] loop:<<Loop>>      outer_loop:none
halving_add_unsigned_constant(byte[] b1, byte[] bo)171   private static void halving_add_unsigned_constant(byte[] b1, byte[] bo) {
172     int min_length = Math.min(bo.length, b1.length);
173     for (int i = 0; i < min_length; i++) {
174       bo[i] = (byte) (((b1[i] & 0xff) + 0xff) >> 1);
175     }
176   }
177 
main(String[] args)178   public static void main(String[] args) {
179     // Initialize cross-values to test all cases, and also
180     // set up some extra values to exercise the cleanup loop.
181     int k = 0;
182     for (int i = 0; i < N; i++) {
183       for (int j = 0; j < N; j++) {
184         sB1[k] = (byte) i;
185         sB2[k] = (byte) j;
186         k++;
187       }
188     }
189     for (int i = 0; i < 15; i++) {
190       sB1[k] = (byte) i;
191       sB2[k] = 100;
192       k++;
193     }
194     expectEquals(k, M);
195 
196     // Test halving add idioms.
197     halving_add_signed(sB1, sB2, sBo);
198     for (int i = 0; i < M; i++) {
199       byte e = (byte) ((sB1[i] + sB2[i]) >> 1);
200       expectEquals(e, sBo[i]);
201     }
202     halving_add_unsigned(sB1, sB2, sBo);
203     for (int i = 0; i < M; i++) {
204       byte e = (byte) (((sB1[i] & 0xff) + (sB2[i] & 0xff)) >> 1);
205       expectEquals(e, sBo[i]);
206     }
207     rounding_halving_add_signed(sB1, sB2, sBo);
208     for (int i = 0; i < M; i++) {
209       byte e = (byte) ((sB1[i] + sB2[i] + 1) >> 1);
210       expectEquals(e, sBo[i]);
211     }
212     rounding_halving_add_unsigned(sB1, sB2, sBo);
213     for (int i = 0; i < M; i++) {
214       byte e = (byte) (((sB1[i] & 0xff) + (sB2[i] & 0xff) + 1) >> 1);
215       expectEquals(e, sBo[i]);
216     }
217     halving_add_signed_constant(sB1, sBo);
218     for (int i = 0; i < M; i++) {
219       byte e = (byte) ((sB1[i] + 0x7f) >> 1);
220       expectEquals(e, sBo[i]);
221     }
222     halving_add_unsigned_constant(sB1, sBo);
223     for (int i = 0; i < M; i++) {
224       byte e = (byte) (((sB1[i] & 0xff) + 0xff) >> 1);
225       expectEquals(e, sBo[i]);
226     }
227 
228     System.out.println("passed");
229   }
230 
expectEquals(int expected, int result)231   private static void expectEquals(int expected, int result) {
232     if (expected != result) {
233       throw new Error("Expected: " + expected + ", found: " + result);
234     }
235   }
236 }
237