Debian Codegraph @

src/f32-ppmm/gen/f32-ppmm-8x8-minmax-asm-aarch64-neonfma-cortex-a75.S

raw
1 // Auto-generated file. Do not edit!
2 // Template: src/f32-ppmm/8x8-aarch64-neonfma-cortex-a75.S.in
3 // Generator: tools/xngen
4 //
5 // Copyright 2023 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9
10 #include "xnnpack/assembly.h"
11
12 # void xnn_f32_ppmm_minmax_ukernel_8x8__asm_aarch64_neonfma_cortex_a75(
13 # size_t mr, x0
14 # size_t nc, x1
15 # size_t kc, x2 / x0
16 # const uint8_t* restrict a, x3
17 # const void* restrict w, x4
18 # uint8_t* restrict c, x5
19 # size_t cm_stride, x6
20 # size_t cn_stride, x7
21 # const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) [sp] -> x8
22
23 # d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS.
24
25 # Vector register usage
26 # A x3 v0 v1
27 # B x4 v4 v5
28 # C0 x5 v16 v17
29 # C1 x12 v18 v19
30 # C2 x13 v20 v21
31 # C3 x14 v22 v23
32 # C4 x15 v24 v25
33 # C5 x16 v26 v27
34 # C6 x17 v28 v29
35 # C7 x6 v30 v31
36 # Clamp v6 v7
37
38 BEGIN_FUNCTION xnn_f32_ppmm_minmax_ukernel_8x8__asm_aarch64_neonfma_cortex_a75
39
40 # Load params pointer
41 LDR x8, [sp]
42
43 # Clamp A and C pointers
44 CMP x0, 2 // if mr < 2
45 ADD x12, x5, x6 // c1 = c0 + cm_stride
46 CSEL x12, x5, x12, LO // c1 = c0
47 ADD x13, x12, x6 // c2 = c1 + cm_stride
48 // if mr <= 2
49 CSEL x13, x12, x13, LS // c2 = c1
50 CMP x0, 4 // if mr < 4
51 ADD x14, x13, x6 // c3 = c2 + cm_stride
52 CSEL x14, x13, x14, LO // c3 = c2
53
54 # Load min/max values
55 LD2R {v6.4s, v7.4s}, [x8]
56
57 ADD x15, x14, x6 // c4 = c3 + cm_stride
58 // if mr <= 4
59 CSEL x15, x14, x15, LS // c4 = c3
60 CMP x0, 6 // if mr < 6
61 ADD x16, x15, x6 // c5 = c4 + cm_stride
62 CSEL x16, x15, x16, LO // c5 = c4
63 ADD x17, x16, x6 // c6 = c5 + cm_stride
64 // if mr <= 4
65 CSEL x17, x16, x17, LS // c6 = c5
66 CMP x0, 8 // if mr < 8
67 ADD x6, x17, x6 // c7 = c6 + cm_stride
68 CSEL x6, x17, x6, LO // c7 = c6
69
70 0:
71 # Load initial bias from w into accumulators
72 LDR q16, [x4], 16
73 LDR q17, [x4], 16
74 MOV v18.16b, v16.16b
75 MOV v20.16b, v16.16b
76 LDR q0, [x3], 16 // Prologue - loads for main loop
77 MOV v22.16b, v16.16b
78 MOV v24.16b, v16.16b
79 LDR q4, [x4], 16
80 MOV v26.16b, v16.16b
81 MOV v28.16b, v16.16b
82 MOV v30.16b, v16.16b
83 MOV v19.16b, v17.16b
84 MOV v21.16b, v17.16b
85 MOV v23.16b, v17.16b
86 SUBS x0, x2, 8 // k = kc - 8
87 MOV v25.16b, v17.16b
88 MOV v27.16b, v17.16b
89 MOV v29.16b, v17.16b
90 MOV v31.16b, v17.16b
91
92 # Is there at least 2 floats (8 bytes) for mainloop + epilogue?
93 B.LO 2f
94
95 # Main loop - 2 float of A (8 bytes)
96 1:
97 LDR q1, [x3], 16
98 FMLA v16.4s, v4.4s, v0.s[0]
99 LDR q5, [x4], 16
100 FMLA v18.4s, v4.4s, v0.s[1]
101 FMLA v20.4s, v4.4s, v0.s[2]
102 FMLA v22.4s, v4.4s, v0.s[3]
103 FMLA v24.4s, v4.4s, v1.s[0]
104 FMLA v26.4s, v4.4s, v1.s[1]
105 FMLA v28.4s, v4.4s, v1.s[2]
106 FMLA v30.4s, v4.4s, v1.s[3]
107 FMLA v17.4s, v5.4s, v0.s[0]
108 LDR q4, [x4], 16
109 FMLA v19.4s, v5.4s, v0.s[1]
110 FMLA v21.4s, v5.4s, v0.s[2]
111 FMLA v23.4s, v5.4s, v0.s[3]
112 FMLA v25.4s, v5.4s, v1.s[0]
113 LDR q0, [x3], 16
114 FMLA v27.4s, v5.4s, v1.s[1]
115 SUBS x0, x0, 4
116 FMLA v29.4s, v5.4s, v1.s[2]
117 FMLA v31.4s, v5.4s, v1.s[3]
118 B.HS 1b
119
120 # Epilogue
121 2:
122 LDR q1, [x3], 16
123 FMLA v16.4s, v4.4s, v0.s[0]
124 LDR q5, [x4], 16
125 FMLA v18.4s, v4.4s, v0.s[1]
126 FMLA v20.4s, v4.4s, v0.s[2]
127 FMLA v22.4s, v4.4s, v0.s[3]
128 FMLA v24.4s, v4.4s, v1.s[0]
129 FMLA v26.4s, v4.4s, v1.s[1]
130 FMLA v28.4s, v4.4s, v1.s[2]
131 FMLA v30.4s, v4.4s, v1.s[3]
132 FMLA v17.4s, v5.4s, v0.s[0]
133 FMLA v19.4s, v5.4s, v0.s[1]
134 FMLA v21.4s, v5.4s, v0.s[2]
135 FMLA v23.4s, v5.4s, v0.s[3]
136 FMLA v25.4s, v5.4s, v1.s[0]
137 FMLA v27.4s, v5.4s, v1.s[1]
138 FMLA v29.4s, v5.4s, v1.s[2]
139 FMLA v31.4s, v5.4s, v1.s[3]
140
141 # Clamp
142 FMAX v16.4s, v16.4s, v6.4s
143 FMAX v17.4s, v17.4s, v6.4s
144 FMAX v18.4s, v18.4s, v6.4s
145 FMAX v19.4s, v19.4s, v6.4s
146 FMAX v20.4s, v20.4s, v6.4s
147 FMAX v21.4s, v21.4s, v6.4s
148 FMAX v22.4s, v22.4s, v6.4s
149 FMAX v23.4s, v23.4s, v6.4s
150 FMAX v24.4s, v24.4s, v6.4s
151 FMAX v25.4s, v25.4s, v6.4s
152 FMAX v26.4s, v26.4s, v6.4s
153 FMAX v27.4s, v27.4s, v6.4s
154 FMAX v28.4s, v28.4s, v6.4s
155 FMAX v29.4s, v29.4s, v6.4s
156 FMAX v30.4s, v30.4s, v6.4s
157 FMAX v31.4s, v31.4s, v6.4s
158 SUBS x1, x1, 8
159 FMIN v16.4s, v16.4s, v7.4s
160 FMIN v17.4s, v17.4s, v7.4s
161 FMIN v18.4s, v18.4s, v7.4s
162 FMIN v19.4s, v19.4s, v7.4s
163 FMIN v20.4s, v20.4s, v7.4s
164 FMIN v21.4s, v21.4s, v7.4s
165 FMIN v22.4s, v22.4s, v7.4s
166 FMIN v23.4s, v23.4s, v7.4s
167 FMIN v24.4s, v24.4s, v7.4s
168 FMIN v25.4s, v25.4s, v7.4s
169 FMIN v26.4s, v26.4s, v7.4s
170 FMIN v27.4s, v27.4s, v7.4s
171 FMIN v28.4s, v28.4s, v7.4s
172 FMIN v29.4s, v29.4s, v7.4s
173 FMIN v30.4s, v30.4s, v7.4s
174 FMIN v31.4s, v31.4s, v7.4s
175
176 # Store full 8 x 8
177 B.LO 3f
178
179 ST1 {v16.16b, v17.16b}, [x5], x7
180 ST1 {v18.16b, v19.16b}, [x12], x7
181 ST1 {v20.16b, v21.16b}, [x13], x7
182 ST1 {v22.16b, v23.16b}, [x14], x7
183 SUB x3, x3, x2, lsl #3 // a0 -= kc * 8
184 ST1 {v24.16b, v25.16b}, [x15], x7
185 ST1 {v26.16b, v27.16b}, [x16], x7
186 ST1 {v28.16b, v29.16b}, [x17], x7
187 ST1 {v30.16b, v31.16b}, [x6], x7
188 B.HI 0b
189 RET
190
191 # Store odd width
192 3:
193 TBZ x1, 2, 4f
194 STR q16, [x5], 16
195 STR q18, [x12], 16
196 MOV v16.16b, v17.16b
197 MOV v18.16b, v19.16b
198 STR q20, [x13], 16
199 STR q22, [x14], 16
200 MOV v20.16b, v21.16b
201 MOV v22.16b, v23.16b
202 STR q24, [x15], 16
203 STR q26, [x16], 16
204 MOV v24.16b, v25.16b
205 MOV v26.16b, v27.16b
206 STR q28, [x17], 16
207 STR q30, [x6], 16
208 MOV v28.16b, v29.16b
209 MOV v30.16b, v31.16b
210 4:
211 TBZ x1, 1, 5f
212 STR d16, [x5], 8
213 STR d18, [x12], 8
214 DUP d16, v16.d[1]
215 DUP d18, v18.d[1]
216 STR d20, [x13], 8
217 STR d22, [x14], 8
218 DUP d20, v20.d[1]
219 DUP d22, v22.d[1]
220 STR d24, [x15], 8
221 STR d26, [x16], 8
222 DUP d24, v24.d[1]
223 DUP d26, v26.d[1]
224 STR d28, [x17], 8
225 STR d30, [x6], 8
226 DUP d28, v28.d[1]
227 DUP d30, v30.d[1]
228
229 5:
230 TBZ x1, 0, 6f
231 STR s16, [x5]
232 STR s18, [x12]
233 STR s20, [x13]
234 STR s22, [x14]
235 STR s24, [x15]
236 STR s26, [x16]
237 STR s28, [x17]
238 STR s30, [x6]
239
240 6:
241 RET
242
243 END_FUNCTION xnn_f32_ppmm_minmax_ukernel_8x8__asm_aarch64_neonfma_cortex_a75
244
245 #ifdef __ELF__
246 .section ".note.GNU-stack","",%progbits
247 #endif