1 // Copyright 2015, VIXL authors
2 // All rights reserved.
3 //
4 // Redistribution and use in source and binary forms, with or without
5 // modification, are permitted provided that the following conditions are met:
6 //
7 // * Redistributions of source code must retain the above copyright notice,
8 // this list of conditions and the following disclaimer.
9 // * Redistributions in binary form must reproduce the above copyright notice,
10 // this list of conditions and the following disclaimer in the documentation
11 // and/or other materials provided with the distribution.
12 // * Neither the name of ARM Limited nor the names of its contributors may be
13 // used to endorse or promote products derived from this software without
14 // specific prior written permission.
15 //
16 // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS CONTRIBUTORS "AS IS" AND
17 // ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
18 // WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
19 // DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE
20 // FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 // DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
22 // SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
23 // CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
24 // OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
25 // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26
27 #include "examples.h"
28
29 using namespace vixl;
30 using namespace vixl::aarch64;
31
32 #define __ masm->
33
34 // A vector by scalar multiply helper routine to generate code for
35 // the multiplication of each column of the resulting 4x4 matrix.
36 // This function provides a template for the following pattern:
37 //
38 // __ Fmul(v<v_out>.V4S(), v4.V4S(), v<s_column>.S(), 0);
39 // __ Fmla(v<v_out>.V4S(), v5.V4S(), v<s_column>.S(), 1);
40 // __ Fmla(v<v_out>.V4S(), v6.V4S(), v<s_column>.S(), 2);
41 // __ Fmla(v<v_out>.V4S(), v7.V4S(), v<s_column>.S(), 3);
42 //
43 // v<v_out> corresponds to a column of the output matrix (v0, v1, v2 or v3).
44 // v<s_column> corresponds to a column of the 2nd input (v16, v17, v18 or v19).
45 //
GenerateMultiplyColumn(MacroAssembler * masm,unsigned out_column,unsigned in_column)46 static void GenerateMultiplyColumn(MacroAssembler* masm,
47 unsigned out_column,
48 unsigned in_column) {
49 // 'v_out' splits a Q register into 4 lanes of 32 bits each.
50 VRegister v_out = VRegister(out_column, kQRegSize, 4);
51 // 'v_in' refers to a single 32 bit 'S' lane.
52 VRegister v_in = VRegister(in_column, kSRegSize);
53
54 __ Fmul(v_out, v4.V4S(), v_in, 0); // e.g. (v0.V4S(), v4.V4S(), v8.S(), 0).
55 __ Fmla(v_out, v5.V4S(), v_in, 1);
56 __ Fmla(v_out, v6.V4S(), v_in, 2);
57 __ Fmla(v_out, v7.V4S(), v_in, 3);
58 }
59
GenerateNEONMatrixMultiply(MacroAssembler * masm)60 void GenerateNEONMatrixMultiply(MacroAssembler* masm) {
61 // Argument location:
62 // dst -> x0
63 // mat1 -> x1
64 // mat2 -> x2
65
66 Label end;
67
68 __ And(x3, x0, x1);
69 __ And(x3, x3, x2);
70 __ Cbz(x3, &end); // Nothing to do if an input is null.
71
72 // Load the first matrix into v4, v5, v6 and v7.
73 __ Ld1(v4.V4S(), v5.V4S(), v6.V4S(), v7.V4S(), MemOperand(x1));
74 // Load the first matrix into v16, v17, v18 and v19.
75 __ Ld1(v16.V4S(), v17.V4S(), v18.V4S(), v19.V4S(), MemOperand(x2));
76
77 // Initialise vectors of the output matrix with zeros.
78 // This is only for the purposes of showing how this can be achived
79 // but technically this is not required because we overwrite all lanes
80 // of the output vectors.
81 __ Movi(v0.V16B(), 0);
82 __ Movi(v1.V16B(), 0);
83 __ Movi(v2.V16B(), 0);
84 __ Movi(v3.V16B(), 0);
85
86 GenerateMultiplyColumn(masm, 0, 16);
87 GenerateMultiplyColumn(masm, 1, 17);
88 GenerateMultiplyColumn(masm, 2, 18);
89 GenerateMultiplyColumn(masm, 3, 19);
90
91 // Store the resulting matrix from v0, v1, v2 and v3.
92 __ St1(v0.V4S(), v1.V4S(), v2.V4S(), v3.V4S(), MemOperand(x0));
93
94 __ Bind(&end);
95 __ Ret();
96 }
97
98
99 #ifndef TEST_EXAMPLES
100 #ifdef VIXL_INCLUDE_SIMULATOR_AARCH64
main(void)101 int main(void) {
102 MacroAssembler masm;
103 Decoder decoder;
104 Simulator simulator(&decoder);
105
106 // Generate the code for the example function.
107 Label neon_matrix_multiply;
108 masm.Bind(&neon_matrix_multiply);
109 GenerateNEONMatrixMultiply(&masm);
110 masm.FinalizeCode();
111
112 // Define the required variables and run the example function.
113 const int kRowSize = 4;
114 const int kColSize = 4;
115 const int kLength = kRowSize * kColSize;
116
117 float mat1[kLength], mat2[kLength], output[kLength];
118
119 // Initialise the output matrix to the zero matrix.
120 memset(output, 0, sizeof(output[0]) * kLength);
121
122 // Fill the two input matrices with some 32 bit floating point values.
123 // Array initialisation using curly brackets is also possible like so:
124 // float mat1[kLength] = { 1.0f, 52.03f, 4.43f, ... };
125 // However, the following way better shows the "column-major" arrangement.
126
127 mat1[0] = 1.0f;
128 mat1[4] = 2.0f;
129 mat1[8] = 3.0f;
130 mat1[12] = 4.0f;
131 mat1[1] = 52.03f;
132 mat1[5] = 12.24f;
133 mat1[9] = 53.56f;
134 mat1[13] = 22.22f;
135 mat1[2] = 4.43f;
136 mat1[6] = 5.00f;
137 mat1[10] = 7.00f;
138 mat1[14] = 3.11f;
139 mat1[3] = 43.47f;
140 mat1[7] = 10.97f;
141 mat1[11] = 37.78f;
142 mat1[15] = 90.91f;
143
144 mat2[0] = 1.0f;
145 mat2[4] = 11.24f;
146 mat2[8] = 21.00f;
147 mat2[12] = 21.31f;
148 mat2[1] = 2.0f;
149 mat2[5] = 2.24f;
150 mat2[9] = 8.56f;
151 mat2[13] = 52.03f;
152 mat2[2] = 3.0f;
153 mat2[6] = 51.00f;
154 mat2[10] = 21.00f;
155 mat2[14] = 33.11f;
156 mat2[3] = 4.0f;
157 mat2[7] = 0.00f;
158 mat2[11] = 84.00f;
159 mat2[15] = 1.97f;
160
161 simulator.ResetState();
162 simulator.WriteXRegister(0, reinterpret_cast<uintptr_t>(output));
163 simulator.WriteXRegister(1, reinterpret_cast<uintptr_t>(mat1));
164 simulator.WriteXRegister(2, reinterpret_cast<uintptr_t>(mat2));
165 simulator.RunFrom(masm.GetLabelAddress<Instruction*>(&neon_matrix_multiply));
166
167 // Print the 4x4 output matrix along with both 4x4 input matrices.
168 for (int i = 0; i < kRowSize; i++) {
169 printf(
170 "| %8.2f %8.2f %8.2f %8.2f | "
171 "| %8.2f %8.2f %8.2f %8.2f | "
172 "| %8.2f %8.2f %8.2f %8.2f |\n",
173 mat1[i],
174 mat1[4 + i],
175 mat1[8 + i],
176 mat1[12 + i],
177 mat2[i],
178 mat2[4 + i],
179 mat2[8 + i],
180 mat2[12 + i],
181 output[i],
182 output[4 + i],
183 output[8 + i],
184 output[12 + i]);
185 if (i == 0 || i == 2) {
186 printf(
187 "| | "
188 "| | "
189 "| |\n");
190 } else if (i == 1) {
191 printf(
192 "| | x "
193 "| | = "
194 "| |\n");
195 }
196 }
197
198 return 0;
199 }
200 #else
201 // Without the simulator there is nothing to test.
main(void)202 int main(void) { return 0; }
203 #endif // VIXL_INCLUDE_SIMULATOR_AARCH64
204 #endif // TEST_EXAMPLES
205