Commit | Line | Data |
---|---|---|
51d543ed MW |
1 | /* simdhp.s Test file for AArch64 half-precision floating-point |
2 | vector instructions. */ | |
3 | ||
4 | /* Vector three-same. */ | |
5 | ||
6 | .macro three_same, op | |
7 | \op v1.2d, v2.2d, v3.2d | |
8 | \op v1.2s, v2.2s, v3.2s | |
9 | \op v1.4s, v2.4s, v3.4s | |
10 | \op v0.4h, v0.4h, v0.4h | |
11 | \op v1.4h, v2.4h, v3.4h | |
12 | \op v0.8h, v0.8h, v0.8h | |
13 | \op v1.8h, v2.8h, v3.8h | |
14 | .endm | |
15 | ||
16 | .text | |
17 | ||
18 | three_same fmaxnm | |
19 | three_same fmaxnmp | |
20 | three_same fminnm | |
21 | three_same fminnmp | |
22 | three_same fmla | |
23 | three_same fmls | |
24 | three_same fadd | |
25 | three_same faddp | |
26 | three_same fsub | |
27 | three_same fmulx | |
28 | three_same fmul | |
29 | three_same fcmeq | |
30 | three_same fcmge | |
31 | three_same fcmgt | |
32 | three_same facge | |
33 | three_same facgt | |
34 | three_same fmax | |
35 | three_same fmaxp | |
36 | three_same fmin | |
37 | three_same fminp | |
38 | three_same frecps | |
39 | three_same fdiv | |
40 | three_same frsqrts | |
6b4680fb MW |
41 | |
42 | /* Scalar three-same. */ | |
43 | ||
44 | .macro sthree_same, op | |
45 | \op d0, d1, d2 | |
46 | \op s0, s1, s2 | |
47 | \op h0, h1, h2 | |
48 | \op h0, h0, h0 | |
49 | .endm | |
50 | ||
51 | sthree_same fabd | |
52 | sthree_same fmulx | |
53 | sthree_same fcmeq | |
54 | sthree_same fcmgt | |
55 | sthree_same fcmge | |
56 | sthree_same facge | |
57 | sthree_same facgt | |
58 | sthree_same frecps | |
59 | sthree_same frsqrts | |
f3aa142b MW |
60 | |
61 | /* Vector two-register misc. */ | |
62 | ||
63 | .macro tworeg_zero, op | |
64 | \op v0.2d, v1.2d, #0.0 | |
65 | \op v0.2s, v1.2s, #0.0 | |
66 | \op v0.4s, v1.4s, #0.0 | |
67 | \op v0.4h, v1.4h, #0.0 | |
68 | \op v0.8h, v1.8h, #0.0 | |
69 | .endm | |
70 | ||
71 | tworeg_zero fcmgt | |
72 | tworeg_zero fcmge | |
73 | tworeg_zero fcmeq | |
74 | tworeg_zero fcmle | |
75 | tworeg_zero fcmlt | |
76 | ||
77 | .macro tworeg_misc, op | |
78 | \op v0.2d, v1.2d | |
79 | \op v0.2s, v1.2s | |
80 | \op v0.4s, v1.4s | |
81 | \op v0.4h, v1.4h | |
82 | \op v0.8h, v1.8h | |
83 | .endm | |
84 | ||
85 | tworeg_misc fabs | |
86 | tworeg_misc fneg | |
87 | ||
88 | tworeg_misc frintn | |
89 | tworeg_misc frinta | |
90 | tworeg_misc frintp | |
91 | ||
92 | tworeg_misc frintm | |
93 | tworeg_misc frintx | |
94 | tworeg_misc frintz | |
95 | tworeg_misc frinti | |
96 | ||
97 | tworeg_misc fcvtns | |
98 | tworeg_misc fcvtnu | |
99 | tworeg_misc fcvtps | |
100 | tworeg_misc fcvtpu | |
101 | ||
102 | tworeg_misc fcvtms | |
103 | tworeg_misc fcvtmu | |
104 | tworeg_misc fcvtzs | |
105 | tworeg_misc fcvtzu | |
106 | ||
107 | tworeg_misc fcvtas | |
108 | tworeg_misc fcvtau | |
109 | ||
110 | tworeg_misc scvtf | |
111 | tworeg_misc ucvtf | |
112 | tworeg_misc frecpe | |
113 | tworeg_misc frsqrte | |
114 | tworeg_misc fsqrt | |
80776b29 MW |
115 | |
116 | /* Scalar two-register misc. */ | |
117 | ||
118 | .macro stworeg_zero, op | |
119 | \op d0, d1, #0.0 | |
120 | \op s0, s1, #0.0 | |
121 | \op h0, h1, #0.0 | |
122 | \op h0, h0, #0.0 | |
123 | .endm | |
124 | ||
125 | stworeg_zero fcmgt | |
126 | stworeg_zero fcmge | |
127 | stworeg_zero fcmeq | |
128 | stworeg_zero fcmle | |
129 | stworeg_zero fcmlt | |
130 | ||
131 | .macro stworeg_misc, op | |
132 | \op d0, d1 | |
133 | \op s0, s1 | |
134 | \op h0, h1 | |
135 | \op h0, h0 | |
136 | .endm | |
137 | ||
138 | stworeg_misc fcvtns | |
139 | stworeg_misc fcvtnu | |
140 | stworeg_misc fcvtps | |
141 | stworeg_misc fcvtpu | |
142 | ||
143 | stworeg_misc fcvtms | |
144 | stworeg_misc fcvtmu | |
145 | stworeg_misc fcvtzs | |
146 | stworeg_misc fcvtzu | |
147 | ||
148 | stworeg_misc fcvtas | |
149 | stworeg_misc fcvtau | |
150 | ||
151 | stworeg_misc scvtf | |
152 | stworeg_misc ucvtf | |
153 | ||
154 | stworeg_misc frecpe | |
155 | stworeg_misc frsqrte | |
156 | stworeg_misc frecpx | |
42f23f62 MW |
157 | |
158 | /* Vector indexed element. */ | |
159 | ||
160 | .macro indexed_elem, op | |
161 | \op v1.2d, v2.2d, v3.d[1] | |
162 | \op v1.2s, v2.2s, v3.s[2] | |
163 | \op v1.4s, v2.4s, v3.s[1] | |
164 | \op v0.4h, v0.4h, v0.h[0] | |
165 | \op v1.4h, v2.4h, v3.h[0] | |
166 | \op v0.8h, v0.8h, v0.h[0] | |
167 | \op v1.8h, v2.8h, v3.h[0] | |
168 | .endm | |
169 | ||
170 | indexed_elem fmla | |
171 | indexed_elem fmls | |
172 | ||
173 | indexed_elem fmul | |
174 | indexed_elem fmulx | |
5f7728b7 MW |
175 | |
176 | /* Scalar indexed element. */ | |
177 | ||
178 | .macro sindexed_elem, op | |
179 | \op d1, d2, v3.d[1] | |
180 | \op s1, s2, v3.s[1] | |
181 | \op h1, h2, v3.h[1] | |
182 | \op h0, h0, v0.h[0] | |
183 | .endm | |
184 | ||
185 | sindexed_elem fmla | |
186 | sindexed_elem fmls | |
187 | ||
188 | sindexed_elem fmul | |
189 | sindexed_elem fmulx | |
bb515fea MW |
190 | |
191 | /* Adv.SIMD across lanes. */ | |
192 | ||
193 | .macro across_lanes, op | |
194 | \op s1, v2.4s | |
195 | \op h1, v2.4h | |
196 | \op h1, v2.8h | |
197 | \op h0, v0.4h | |
198 | \op h0, v0.8h | |
199 | .endm | |
200 | ||
201 | across_lanes fmaxnmv | |
202 | across_lanes fmaxv | |
203 | across_lanes fminnmv | |
204 | across_lanes fminv | |
205 | ||
4b5fc357 MW |
206 | /* Adv.SIMD modified immediate. */ |
207 | ||
208 | fmov v1.2d, #2.0 | |
209 | fmov v1.2s, #2.0 | |
210 | fmov v1.4s, #2.0 | |
211 | fmov v1.4h, #2.0 | |
212 | fmov v1.8h, #2.0 | |
213 | fmov v0.4h, #1.0 | |
214 | fmov v0.8h, #1.0 | |
b195470d | 215 | |
b5b0f34c | 216 | /* Adv.SIMD scalar pairwise. */ |
b195470d MW |
217 | |
218 | .macro scalar_pairwise, op | |
219 | \op d1, v2.2d | |
220 | \op s1, v2.2s | |
221 | \op h1, v2.2h | |
222 | \op h0, v0.2h | |
223 | .endm | |
224 | ||
225 | scalar_pairwise fmaxnmp | |
226 | scalar_pairwise faddp | |
227 | scalar_pairwise fmaxp | |
228 | scalar_pairwise fminnmp | |
229 | scalar_pairwise fminp | |
230 | ||
b5b0f34c MW |
231 | /* Adv.SIMD shift by immediate. */ |
232 | ||
233 | .macro shift_imm, op | |
234 | \op v1.2d, v2.2d, #3 | |
235 | \op v1.2s, v2.2s, #3 | |
236 | \op v1.4s, v2.4s, #3 | |
237 | \op v1.4h, v2.4h, #3 | |
238 | \op v1.8h, v2.8h, #3 | |
239 | \op v0.4h, v0.4h, #1 | |
240 | \op v0.8h, v0.8h, #1 | |
241 | .endm | |
242 | ||
243 | shift_imm scvtf | |
244 | shift_imm fcvtzs | |
245 | shift_imm ucvtf | |
246 | shift_imm fcvtzu |