1 /*
2 * Copyright (c) 2022 Arm Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in
14 * all copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22 * IN THE SOFTWARE.
23 */
24 #ifdef __ARM_FEATURE_SVE
25 #ifdef ARM_COMPUTE_ENABLE_SME2
26
27 #include "arm_gemm.hpp"
28 #include "../../utils.hpp"
29
30 #include <cassert>
31 #include <limits>
32
33 namespace arm_gemm {
34
sme2_gemv_fp32_mla_16VL(const float * A_ptr,const float * B_ptr,float * output_ptr,size_t N,size_t K,const float * bias,Activation act,bool)35 void sme2_gemv_fp32_mla_16VL (
36 const float *A_ptr, const float *B_ptr, float *output_ptr,
37 size_t N, size_t K,
38 const float *bias, Activation act, bool
39 )
40 {
41 struct KernelArgs {
42 float maxval = static_cast<float>(std::numeric_limits<float>::infinity());
43 float minval = - static_cast<float>(std::numeric_limits<float>::infinity());
44 const float *B_ptr = {};
45 size_t output_offset = {};
46 unsigned int input_initial_col = {};
47 } ka;
48
49 unsigned long flags=0;
50 ka.B_ptr = B_ptr;
51 switch(act.type) {
52 default:
53 case Activation::Type::None:
54 break;
55 case Activation::Type::BoundedReLU:
56 ka.maxval = static_cast<float>(act.param1);
57 /* fall through */
58 case Activation::Type::ReLU:
59 ka.minval = 0;
60 flags |= 0x2;
61 break;
62 }
63 __asm__ __volatile__(
64 "ptrue p1.b\n"
65 ".inst 0xd503477f // SMSTART ZA\n"
66 "cntw x27, ALL, MUL #4\n"
67 "add x26, %x[N], x27\n"
68 "sub x26, x26, #0x1\n"
69 "udiv x26, x26, x27\n"
70 "add x21, x26, #0x3\n"
71 "and x21, x21, #0xfffffffffffffffc\n"
72 "mul x21, x21, x27\n"
73 "mul x21, x21, %x[K]\n"
74 "mov x9, #0x0\n"
75 "mov x25, %x[B_ptr]\n"
76 "mov x24, %x[output_ptr]\n"
77 "ptrue p1.b\n"
78 ".inst 0x25207811 // ptrue pn9.b\n"
79 "lsl x21, x21, #0x2\n"
80 "mov x20, #0x1\n"
81 "1:" // RHS size check loop
82 "cmp x21, #0x200000\n"
83 "blt 2f\n"
84 "tbnz x21, #0, 3f\n"
85 "lsr x21, x21, #0x1\n"
86 "lsl x20, x20, #0x1\n"
87 "b 1b\n"
88 "2:" // RHS do prefetch
89 "lsl x19, x21, #0x26\n"
90 "sub x20, x20, #0x1\n"
91 "lsl x20, x20, #0x16\n"
92 "orr x21, x21, x19\n"
93 "orr x21, x21, x20\n"
94 ".inst 0xf8b54b3a // rprfm pldonce, x21, [x25]\n"
95 "3:" // RHS prefetch exit
96 "mov x23, %x[bias]\n"
97 "4:" // Column loop
98 "cmp x26, #0x4\n"
99 "bge 28f\n"
100 "cmp x26, #0x2\n"
101 "bgt 20f\n"
102 "beq 12f\n"
103 "mov x22, %x[A_ptr]\n"
104 "lsl x21, %x[K], #0x2\n"
105 "mov x19, %x[N]\n"
106 "mov x20, %x[K]\n"
107 ".inst 0xf8b54ad8 // rprfm pldmany, x21, [x22]\n"
108 ".inst 0x25b367f0 // whilelt p8.s, XZR, x19, VLx4\n"
109 "cbz x23, 5f\n"
110 ".inst 0xa040c6e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x23]\n"
111 ".inst 0xc0042c00 // mova za.d[x9, #0], { z0.d-z3.d }\n"
112 "b 6f\n"
113 "5:" // Width 1: no bias
114 ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
115 "6:" // Width 1: setup done
116 "cmp x20, #0x4\n"
117 "ble 8f\n"
118 "7:" // Width 1: Multiply loop: Main loop head
119 "whilelt p0.s, XZR, x20\n"
120 "ld1rqw { z10.s }, p0/Z, [x22]\n"
121 "sub x20, x20, #0x4\n"
122 ".inst 0xa040c721 // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x25]\n"
123 ".inst 0xc15aa000 // fmla za.s[x9, 0], { z0.s-z3.s }, z10.s[0]\n"
124 "addvl x25, x25, #16\n"
125 "cmp x20, #0x4\n"
126 ".inst 0xa040c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25]\n"
127 ".inst 0xc15aa700 // fmla za.s[x9, 0], { z24.s-z27.s }, z10.s[1]\n"
128 "addvl x25, x25, #16\n"
129 "add x22, x22, #0x10\n"
130 ".inst 0xa040c72d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x25]\n"
131 ".inst 0xc15aa980 // fmla za.s[x9, 0], { z12.s-z15.s }, z10.s[2]\n"
132 "addvl x25, x25, #16\n"
133 ".inst 0xa040c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25]\n"
134 ".inst 0xc15aae00 // fmla za.s[x9, 0], { z16.s-z19.s }, z10.s[3]\n"
135 "addvl x25, x25, #16\n"
136 "bgt 7b\n"
137 "8:" // Width 1: Multiply loop: Single iteration only
138 "whilelt p0.s, XZR, x20\n"
139 "ld1rqw { z10.s }, p0/Z, [x22]\n"
140 "subs x20, x20, #0x1\n"
141 ".inst 0xa040c721 // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x25]\n"
142 "add x22, x22, #0x10\n"
143 ".inst 0xc15aa000 // fmla za.s[x9, 0], { z0.s-z3.s }, z10.s[0]\n"
144 "addvl x25, x25, #16\n"
145 "ble 9f\n"
146 ".inst 0xa040c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25]\n"
147 "subs x20, x20, #0x1\n"
148 ".inst 0xc15aa700 // fmla za.s[x9, 0], { z24.s-z27.s }, z10.s[1]\n"
149 "addvl x25, x25, #16\n"
150 "ble 9f\n"
151 ".inst 0xa040c72d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x25]\n"
152 "subs x20, x20, #0x1\n"
153 ".inst 0xc15aa980 // fmla za.s[x9, 0], { z12.s-z15.s }, z10.s[2]\n"
154 "addvl x25, x25, #16\n"
155 "ble 9f\n"
156 ".inst 0xa040c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25]\n"
157 ".inst 0xc15aae00 // fmla za.s[x9, 0], { z16.s-z19.s }, z10.s[3]\n"
158 "addvl x25, x25, #16\n"
159 "9:" // Width 1: Multiply loop: multiply skip
160 "tbz %x[flags], #1, 10f\n"
161 "add x20, %x[args_ptr], %[offset_min]\n"
162 "add x19, %x[args_ptr], %[offset_max]\n"
163 ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n"
164 "ld1rw { z0.s }, p1/Z, [x20]\n"
165 "ld1rw { z6.s }, p1/Z, [x19]\n"
166 ".inst 0xc1a6c808 // fclamp { z8.s-z11.s }, z0.s, z6.s\n"
167 ".inst 0xa060c308 // st1w { z8.s-z11.s }, p8, [x24]\n"
168 "addvl x24, x24, #4\n"
169 "b 11f\n"
170 "10:" // Width 1: No activation
171 ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n"
172 ".inst 0xa060c308 // st1w { z8.s-z11.s }, p8, [x24]\n"
173 "addvl x24, x24, #4\n"
174 "11:" // Width 1: Output done
175 "b 36f\n"
176 "12:" // Width 2
177 "mov x22, %x[A_ptr]\n"
178 "lsl x21, %x[K], #0x2\n"
179 "sub x19, %x[N], x27\n"
180 "mov x20, %x[K]\n"
181 ".inst 0xf8b54ad8 // rprfm pldmany, x21, [x22]\n"
182 ".inst 0x25b367f0 // whilelt p8.s, XZR, x19, VLx4\n"
183 "cbz x23, 13f\n"
184 ".inst 0xa040c6e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x23]\n"
185 ".inst 0xc0042c00 // mova za.d[x9, #0], { z0.d-z3.d }\n"
186 ".inst 0xa041c6e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
187 ".inst 0xc0042d01 // mova za.d[x9, #1], { z8.d-z11.d }\n"
188 "b 14f\n"
189 "13:" // Width 2: no bias
190 ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
191 "14:" // Width 2: setup done
192 "cmp x20, #0x4\n"
193 "ble 16f\n"
194 "15:" // Width 2: Multiply loop: Main loop head
195 "whilelt p0.s, XZR, x20\n"
196 "ld1rqw { z10.s }, p0/Z, [x22]\n"
197 "sub x20, x20, #0x4\n"
198 ".inst 0xa040c721 // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x25]\n"
199 ".inst 0xc15aa000 // fmla za.s[x9, 0], { z0.s-z3.s }, z10.s[0]\n"
200 "cmp x20, #0x4\n"
201 "add x22, x22, #0x10\n"
202 ".inst 0xa041c725 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
203 ".inst 0xc15aa081 // fmla za.s[x9, 1], { z4.s-z7.s }, z10.s[0]\n"
204 "addvl x25, x25, #16\n"
205 ".inst 0xa040c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25]\n"
206 ".inst 0xc15aa700 // fmla za.s[x9, 0], { z24.s-z27.s }, z10.s[1]\n"
207 ".inst 0xa041c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
208 ".inst 0xc15aa601 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[1]\n"
209 "addvl x25, x25, #16\n"
210 ".inst 0xa040c72d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x25]\n"
211 ".inst 0xc15aa980 // fmla za.s[x9, 0], { z12.s-z15.s }, z10.s[2]\n"
212 ".inst 0xa041c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
213 ".inst 0xc15aaa01 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[2]\n"
214 "addvl x25, x25, #16\n"
215 ".inst 0xa040c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25]\n"
216 ".inst 0xc15aae00 // fmla za.s[x9, 0], { z16.s-z19.s }, z10.s[3]\n"
217 ".inst 0xa041c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
218 ".inst 0xc15aaf01 // fmla za.s[x9, 1], { z24.s-z27.s }, z10.s[3]\n"
219 "addvl x25, x25, #16\n"
220 "bgt 15b\n"
221 "16:" // Width 2: Multiply loop: Single iteration only
222 "whilelt p0.s, XZR, x20\n"
223 "ld1rqw { z10.s }, p0/Z, [x22]\n"
224 "subs x20, x20, #0x1\n"
225 ".inst 0xa040c721 // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x25]\n"
226 "add x22, x22, #0x10\n"
227 ".inst 0xc15aa000 // fmla za.s[x9, 0], { z0.s-z3.s }, z10.s[0]\n"
228 ".inst 0xa041c725 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
229 ".inst 0xc15aa081 // fmla za.s[x9, 1], { z4.s-z7.s }, z10.s[0]\n"
230 "addvl x25, x25, #16\n"
231 "ble 17f\n"
232 ".inst 0xa040c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25]\n"
233 "subs x20, x20, #0x1\n"
234 ".inst 0xc15aa700 // fmla za.s[x9, 0], { z24.s-z27.s }, z10.s[1]\n"
235 ".inst 0xa041c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
236 ".inst 0xc15aa601 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[1]\n"
237 "addvl x25, x25, #16\n"
238 "ble 17f\n"
239 ".inst 0xa040c72d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x25]\n"
240 "subs x20, x20, #0x1\n"
241 ".inst 0xc15aa980 // fmla za.s[x9, 0], { z12.s-z15.s }, z10.s[2]\n"
242 ".inst 0xa041c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
243 ".inst 0xc15aaa01 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[2]\n"
244 "addvl x25, x25, #16\n"
245 "ble 17f\n"
246 ".inst 0xa040c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25]\n"
247 ".inst 0xc15aae00 // fmla za.s[x9, 0], { z16.s-z19.s }, z10.s[3]\n"
248 ".inst 0xa041c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
249 ".inst 0xc15aaf01 // fmla za.s[x9, 1], { z24.s-z27.s }, z10.s[3]\n"
250 "addvl x25, x25, #16\n"
251 "17:" // Width 2: Multiply loop: multiply skip
252 "tbz %x[flags], #1, 18f\n"
253 "add x20, %x[args_ptr], %[offset_min]\n"
254 "add x19, %x[args_ptr], %[offset_max]\n"
255 ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n"
256 "ld1rw { z0.s }, p1/Z, [x20]\n"
257 ".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n"
258 "ld1rw { z6.s }, p1/Z, [x19]\n"
259 ".inst 0xc1a6c808 // fclamp { z8.s-z11.s }, z0.s, z6.s\n"
260 ".inst 0xa060c708 // st1w { z8.s-z11.s }, pn9.b, [x24]\n"
261 ".inst 0xc1a6c814 // fclamp { z20.s-z23.s }, z0.s, z6.s\n"
262 ".inst 0xa061c314 // st1w { z20.s-z23.s }, p8, [x24, #0x4, MUL VL]\n"
263 "addvl x24, x24, #8\n"
264 "b 19f\n"
265 "18:" // Width 2: No activation
266 ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n"
267 ".inst 0xa060c708 // st1w { z8.s-z11.s }, pn9.b, [x24]\n"
268 ".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n"
269 ".inst 0xa061c314 // st1w { z20.s-z23.s }, p8, [x24, #0x4, MUL VL]\n"
270 "addvl x24, x24, #8\n"
271 "19:" // Width 2: Output done
272 "b 36f\n"
273 "20:" // Width 3
274 "mov x19, #0x2\n"
275 "mov x22, %x[A_ptr]\n"
276 "lsl x21, %x[K], #0x2\n"
277 "msub x19, x27, x19, %x[N]\n"
278 "mov x20, %x[K]\n"
279 ".inst 0xf8b54ad8 // rprfm pldmany, x21, [x22]\n"
280 ".inst 0x25b367f0 // whilelt p8.s, XZR, x19, VLx4\n"
281 "cbz x23, 21f\n"
282 ".inst 0xa040c6e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x23]\n"
283 ".inst 0xc0042c00 // mova za.d[x9, #0], { z0.d-z3.d }\n"
284 ".inst 0xa041c6e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
285 ".inst 0xc0042d01 // mova za.d[x9, #1], { z8.d-z11.d }\n"
286 ".inst 0xa042c6e4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
287 ".inst 0xc0042c82 // mova za.d[x9, #2], { z4.d-z7.d }\n"
288 "b 22f\n"
289 "21:" // Width 3: no bias
290 ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
291 "22:" // Width 3: setup done
292 "cmp x20, #0x4\n"
293 "ble 24f\n"
294 "23:" // Width 3: Multiply loop: Main loop head
295 "whilelt p0.s, XZR, x20\n"
296 "ld1rqw { z10.s }, p0/Z, [x22]\n"
297 "sub x20, x20, #0x4\n"
298 ".inst 0xa040c721 // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x25]\n"
299 ".inst 0xc15aa000 // fmla za.s[x9, 0], { z0.s-z3.s }, z10.s[0]\n"
300 "cmp x20, #0x4\n"
301 "add x22, x22, #0x10\n"
302 ".inst 0xa041c725 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
303 ".inst 0xc15aa081 // fmla za.s[x9, 1], { z4.s-z7.s }, z10.s[0]\n"
304 ".inst 0xa042c735 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x25, #0x8, MUL VL]\n"
305 ".inst 0xc15aa282 // fmla za.s[x9, 2], { z20.s-z23.s }, z10.s[0]\n"
306 "addvl x25, x25, #16\n"
307 ".inst 0xa040c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25]\n"
308 ".inst 0xc15aa700 // fmla za.s[x9, 0], { z24.s-z27.s }, z10.s[1]\n"
309 ".inst 0xa041c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
310 ".inst 0xc15aa601 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[1]\n"
311 ".inst 0xa042c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25, #0x8, MUL VL]\n"
312 ".inst 0xc15aa702 // fmla za.s[x9, 2], { z24.s-z27.s }, z10.s[1]\n"
313 "addvl x25, x25, #16\n"
314 ".inst 0xa040c72d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x25]\n"
315 ".inst 0xc15aa980 // fmla za.s[x9, 0], { z12.s-z15.s }, z10.s[2]\n"
316 ".inst 0xa041c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
317 ".inst 0xc15aaa01 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[2]\n"
318 ".inst 0xa042c73d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x25, #0x8, MUL VL]\n"
319 ".inst 0xc15aab82 // fmla za.s[x9, 2], { z28.s-z31.s }, z10.s[2]\n"
320 "addvl x25, x25, #16\n"
321 ".inst 0xa040c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25]\n"
322 ".inst 0xc15aae00 // fmla za.s[x9, 0], { z16.s-z19.s }, z10.s[3]\n"
323 ".inst 0xa041c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
324 ".inst 0xc15aaf01 // fmla za.s[x9, 1], { z24.s-z27.s }, z10.s[3]\n"
325 ".inst 0xa042c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0x8, MUL VL]\n"
326 ".inst 0xc15aae02 // fmla za.s[x9, 2], { z16.s-z19.s }, z10.s[3]\n"
327 "addvl x25, x25, #16\n"
328 "bgt 23b\n"
329 "24:" // Width 3: Multiply loop: Single iteration only
330 "whilelt p0.s, XZR, x20\n"
331 "ld1rqw { z10.s }, p0/Z, [x22]\n"
332 "subs x20, x20, #0x1\n"
333 ".inst 0xa040c721 // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x25]\n"
334 "add x22, x22, #0x10\n"
335 ".inst 0xc15aa000 // fmla za.s[x9, 0], { z0.s-z3.s }, z10.s[0]\n"
336 ".inst 0xa041c725 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
337 ".inst 0xc15aa081 // fmla za.s[x9, 1], { z4.s-z7.s }, z10.s[0]\n"
338 ".inst 0xa042c735 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x25, #0x8, MUL VL]\n"
339 ".inst 0xc15aa282 // fmla za.s[x9, 2], { z20.s-z23.s }, z10.s[0]\n"
340 "addvl x25, x25, #16\n"
341 "ble 25f\n"
342 ".inst 0xa040c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25]\n"
343 "subs x20, x20, #0x1\n"
344 ".inst 0xc15aa700 // fmla za.s[x9, 0], { z24.s-z27.s }, z10.s[1]\n"
345 ".inst 0xa041c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
346 ".inst 0xc15aa601 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[1]\n"
347 ".inst 0xa042c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25, #0x8, MUL VL]\n"
348 ".inst 0xc15aa702 // fmla za.s[x9, 2], { z24.s-z27.s }, z10.s[1]\n"
349 "addvl x25, x25, #16\n"
350 "ble 25f\n"
351 ".inst 0xa040c72d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x25]\n"
352 "subs x20, x20, #0x1\n"
353 ".inst 0xc15aa980 // fmla za.s[x9, 0], { z12.s-z15.s }, z10.s[2]\n"
354 ".inst 0xa041c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
355 ".inst 0xc15aaa01 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[2]\n"
356 ".inst 0xa042c73d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x25, #0x8, MUL VL]\n"
357 ".inst 0xc15aab82 // fmla za.s[x9, 2], { z28.s-z31.s }, z10.s[2]\n"
358 "addvl x25, x25, #16\n"
359 "ble 25f\n"
360 ".inst 0xa040c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25]\n"
361 ".inst 0xc15aae00 // fmla za.s[x9, 0], { z16.s-z19.s }, z10.s[3]\n"
362 ".inst 0xa041c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
363 ".inst 0xc15aaf01 // fmla za.s[x9, 1], { z24.s-z27.s }, z10.s[3]\n"
364 ".inst 0xa042c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0x8, MUL VL]\n"
365 ".inst 0xc15aae02 // fmla za.s[x9, 2], { z16.s-z19.s }, z10.s[3]\n"
366 "addvl x25, x25, #16\n"
367 "25:" // Width 3: Multiply loop: multiply skip
368 "tbz %x[flags], #1, 26f\n"
369 "add x20, %x[args_ptr], %[offset_min]\n"
370 "add x19, %x[args_ptr], %[offset_max]\n"
371 ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n"
372 "ld1rw { z0.s }, p1/Z, [x20]\n"
373 ".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n"
374 "ld1rw { z6.s }, p1/Z, [x19]\n"
375 ".inst 0xc1a6c808 // fclamp { z8.s-z11.s }, z0.s, z6.s\n"
376 ".inst 0xc0062c50 // mova { z16.d-z19.d }, za.d[x9, #2]\n"
377 ".inst 0xa060c708 // st1w { z8.s-z11.s }, pn9.b, [x24]\n"
378 ".inst 0xc1a6c814 // fclamp { z20.s-z23.s }, z0.s, z6.s\n"
379 ".inst 0xa061c714 // st1w { z20.s-z23.s }, pn9.b, [x24, #0x4, MUL VL]\n"
380 ".inst 0xc1a6c810 // fclamp { z16.s-z19.s }, z0.s, z6.s\n"
381 ".inst 0xa062c310 // st1w { z16.s-z19.s }, p8, [x24, #0x8, MUL VL]\n"
382 "addvl x24, x24, #12\n"
383 "b 27f\n"
384 "26:" // Width 3: No activation
385 ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n"
386 ".inst 0xa060c708 // st1w { z8.s-z11.s }, pn9.b, [x24]\n"
387 ".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n"
388 ".inst 0xa061c714 // st1w { z20.s-z23.s }, pn9.b, [x24, #0x4, MUL VL]\n"
389 ".inst 0xc0062c50 // mova { z16.d-z19.d }, za.d[x9, #2]\n"
390 ".inst 0xa062c310 // st1w { z16.s-z19.s }, p8, [x24, #0x8, MUL VL]\n"
391 "addvl x24, x24, #12\n"
392 "27:" // Width 3: Output done
393 "b 36f\n"
394 "28:" // Width 4
395 "mov x19, #0x3\n"
396 "mov x22, %x[A_ptr]\n"
397 "lsl x21, %x[K], #0x2\n"
398 "msub x19, x27, x19, %x[N]\n"
399 "mov x20, %x[K]\n"
400 ".inst 0xf8b54ad8 // rprfm pldmany, x21, [x22]\n"
401 ".inst 0x25b367f0 // whilelt p8.s, XZR, x19, VLx4\n"
402 "cbz x23, 29f\n"
403 ".inst 0xa040c6e0 // ld1w { z0.s-z3.s }, pn9.b/Z, [x23]\n"
404 ".inst 0xc0042c00 // mova za.d[x9, #0], { z0.d-z3.d }\n"
405 ".inst 0xa041c6e8 // ld1w { z8.s-z11.s }, pn9.b/Z, [x23, #0x4, MUL VL]\n"
406 ".inst 0xc0042d01 // mova za.d[x9, #1], { z8.d-z11.d }\n"
407 ".inst 0xa042c6e4 // ld1w { z4.s-z7.s }, pn9.b/Z, [x23, #0x8, MUL VL]\n"
408 ".inst 0xc0042c82 // mova za.d[x9, #2], { z4.d-z7.d }\n"
409 ".inst 0xa043c6f0 // ld1w { z16.s-z19.s }, pn9.b/Z, [x23, #0xc, MUL VL]\n"
410 ".inst 0xc0042e03 // mova za.d[x9, #3], { z16.d-z19.d }\n"
411 "addvl x23, x23, #16\n"
412 "b 30f\n"
413 "29:" // Width 4: no bias
414 ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
415 "30:" // Width 4: setup done
416 "cmp x20, #0x4\n"
417 "ble 32f\n"
418 "31:" // Width 4: Multiply loop: Main loop head
419 "whilelt p0.s, XZR, x20\n"
420 "ld1rqw { z10.s }, p0/Z, [x22]\n"
421 "sub x20, x20, #0x4\n"
422 ".inst 0xa040c721 // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x25]\n"
423 ".inst 0xc15aa000 // fmla za.s[x9, 0], { z0.s-z3.s }, z10.s[0]\n"
424 "cmp x20, #0x4\n"
425 "add x22, x22, #0x10\n"
426 ".inst 0xa041c725 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
427 ".inst 0xc15aa081 // fmla za.s[x9, 1], { z4.s-z7.s }, z10.s[0]\n"
428 ".inst 0xa042c735 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x25, #0x8, MUL VL]\n"
429 ".inst 0xc15aa282 // fmla za.s[x9, 2], { z20.s-z23.s }, z10.s[0]\n"
430 ".inst 0xa043c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0xc, MUL VL]\n"
431 ".inst 0xc15aa203 // fmla za.s[x9, 3], { z16.s-z19.s }, z10.s[0]\n"
432 "addvl x25, x25, #16\n"
433 ".inst 0xa040c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25]\n"
434 ".inst 0xc15aa700 // fmla za.s[x9, 0], { z24.s-z27.s }, z10.s[1]\n"
435 ".inst 0xa041c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
436 ".inst 0xc15aa601 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[1]\n"
437 ".inst 0xa042c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25, #0x8, MUL VL]\n"
438 ".inst 0xc15aa702 // fmla za.s[x9, 2], { z24.s-z27.s }, z10.s[1]\n"
439 ".inst 0xa043c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0xc, MUL VL]\n"
440 ".inst 0xc15aa603 // fmla za.s[x9, 3], { z16.s-z19.s }, z10.s[1]\n"
441 "addvl x25, x25, #16\n"
442 ".inst 0xa040c72d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x25]\n"
443 ".inst 0xc15aa980 // fmla za.s[x9, 0], { z12.s-z15.s }, z10.s[2]\n"
444 ".inst 0xa041c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
445 ".inst 0xc15aaa01 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[2]\n"
446 ".inst 0xa042c73d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x25, #0x8, MUL VL]\n"
447 ".inst 0xc15aab82 // fmla za.s[x9, 2], { z28.s-z31.s }, z10.s[2]\n"
448 ".inst 0xa043c735 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x25, #0xc, MUL VL]\n"
449 ".inst 0xc15aaa83 // fmla za.s[x9, 3], { z20.s-z23.s }, z10.s[2]\n"
450 "addvl x25, x25, #16\n"
451 ".inst 0xa040c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25]\n"
452 ".inst 0xc15aae00 // fmla za.s[x9, 0], { z16.s-z19.s }, z10.s[3]\n"
453 ".inst 0xa041c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
454 ".inst 0xc15aaf01 // fmla za.s[x9, 1], { z24.s-z27.s }, z10.s[3]\n"
455 ".inst 0xa042c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0x8, MUL VL]\n"
456 ".inst 0xc15aae02 // fmla za.s[x9, 2], { z16.s-z19.s }, z10.s[3]\n"
457 ".inst 0xa043c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0xc, MUL VL]\n"
458 ".inst 0xc15aae03 // fmla za.s[x9, 3], { z16.s-z19.s }, z10.s[3]\n"
459 "addvl x25, x25, #16\n"
460 "bgt 31b\n"
461 "32:" // Width 4: Multiply loop: Single iteration only
462 "whilelt p0.s, XZR, x20\n"
463 "ld1rqw { z10.s }, p0/Z, [x22]\n"
464 "subs x20, x20, #0x1\n"
465 ".inst 0xa040c721 // ldnt1w { z0.s-z3.s }, pn9.b/Z, [x25]\n"
466 "add x22, x22, #0x10\n"
467 ".inst 0xc15aa000 // fmla za.s[x9, 0], { z0.s-z3.s }, z10.s[0]\n"
468 ".inst 0xa041c725 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
469 ".inst 0xc15aa081 // fmla za.s[x9, 1], { z4.s-z7.s }, z10.s[0]\n"
470 ".inst 0xa042c735 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x25, #0x8, MUL VL]\n"
471 ".inst 0xc15aa282 // fmla za.s[x9, 2], { z20.s-z23.s }, z10.s[0]\n"
472 ".inst 0xa043c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0xc, MUL VL]\n"
473 ".inst 0xc15aa203 // fmla za.s[x9, 3], { z16.s-z19.s }, z10.s[0]\n"
474 "addvl x25, x25, #16\n"
475 "ble 33f\n"
476 ".inst 0xa040c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25]\n"
477 "subs x20, x20, #0x1\n"
478 ".inst 0xc15aa700 // fmla za.s[x9, 0], { z24.s-z27.s }, z10.s[1]\n"
479 ".inst 0xa041c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
480 ".inst 0xc15aa601 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[1]\n"
481 ".inst 0xa042c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25, #0x8, MUL VL]\n"
482 ".inst 0xc15aa702 // fmla za.s[x9, 2], { z24.s-z27.s }, z10.s[1]\n"
483 ".inst 0xa043c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0xc, MUL VL]\n"
484 ".inst 0xc15aa603 // fmla za.s[x9, 3], { z16.s-z19.s }, z10.s[1]\n"
485 "addvl x25, x25, #16\n"
486 "ble 33f\n"
487 ".inst 0xa040c72d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x25]\n"
488 "subs x20, x20, #0x1\n"
489 ".inst 0xc15aa980 // fmla za.s[x9, 0], { z12.s-z15.s }, z10.s[2]\n"
490 ".inst 0xa041c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
491 ".inst 0xc15aaa01 // fmla za.s[x9, 1], { z16.s-z19.s }, z10.s[2]\n"
492 ".inst 0xa042c73d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x25, #0x8, MUL VL]\n"
493 ".inst 0xc15aab82 // fmla za.s[x9, 2], { z28.s-z31.s }, z10.s[2]\n"
494 ".inst 0xa043c735 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x25, #0xc, MUL VL]\n"
495 ".inst 0xc15aaa83 // fmla za.s[x9, 3], { z20.s-z23.s }, z10.s[2]\n"
496 "addvl x25, x25, #16\n"
497 "ble 33f\n"
498 ".inst 0xa040c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25]\n"
499 ".inst 0xc15aae00 // fmla za.s[x9, 0], { z16.s-z19.s }, z10.s[3]\n"
500 ".inst 0xa041c739 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x25, #0x4, MUL VL]\n"
501 ".inst 0xc15aaf01 // fmla za.s[x9, 1], { z24.s-z27.s }, z10.s[3]\n"
502 ".inst 0xa042c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0x8, MUL VL]\n"
503 ".inst 0xc15aae02 // fmla za.s[x9, 2], { z16.s-z19.s }, z10.s[3]\n"
504 ".inst 0xa043c731 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x25, #0xc, MUL VL]\n"
505 ".inst 0xc15aae03 // fmla za.s[x9, 3], { z16.s-z19.s }, z10.s[3]\n"
506 "addvl x25, x25, #16\n"
507 "33:" // Width 4: Multiply loop: multiply skip
508 "tbz %x[flags], #1, 34f\n"
509 "add x20, %x[args_ptr], %[offset_min]\n"
510 "add x19, %x[args_ptr], %[offset_max]\n"
511 ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n"
512 "ld1rw { z0.s }, p1/Z, [x20]\n"
513 ".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n"
514 "ld1rw { z6.s }, p1/Z, [x19]\n"
515 ".inst 0xc1a6c808 // fclamp { z8.s-z11.s }, z0.s, z6.s\n"
516 ".inst 0xc0062c50 // mova { z16.d-z19.d }, za.d[x9, #2]\n"
517 ".inst 0xa060c708 // st1w { z8.s-z11.s }, pn9.b, [x24]\n"
518 ".inst 0xc1a6c814 // fclamp { z20.s-z23.s }, z0.s, z6.s\n"
519 ".inst 0xc0062c78 // mova { z24.d-z27.d }, za.d[x9, #3]\n"
520 ".inst 0xa061c714 // st1w { z20.s-z23.s }, pn9.b, [x24, #0x4, MUL VL]\n"
521 ".inst 0xc1a6c810 // fclamp { z16.s-z19.s }, z0.s, z6.s\n"
522 ".inst 0xa062c710 // st1w { z16.s-z19.s }, pn9.b, [x24, #0x8, MUL VL]\n"
523 ".inst 0xc1a6c818 // fclamp { z24.s-z27.s }, z0.s, z6.s\n"
524 ".inst 0xa063c318 // st1w { z24.s-z27.s }, p8, [x24, #0xc, MUL VL]\n"
525 "addvl x24, x24, #16\n"
526 "b 35f\n"
527 "34:" // Width 4: No activation
528 ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n"
529 ".inst 0xa060c708 // st1w { z8.s-z11.s }, pn9.b, [x24]\n"
530 ".inst 0xc0062c34 // mova { z20.d-z23.d }, za.d[x9, #1]\n"
531 ".inst 0xa061c714 // st1w { z20.s-z23.s }, pn9.b, [x24, #0x4, MUL VL]\n"
532 ".inst 0xc0062c50 // mova { z16.d-z19.d }, za.d[x9, #2]\n"
533 ".inst 0xa062c710 // st1w { z16.s-z19.s }, pn9.b, [x24, #0x8, MUL VL]\n"
534 ".inst 0xc0062c78 // mova { z24.d-z27.d }, za.d[x9, #3]\n"
535 ".inst 0xa063c318 // st1w { z24.s-z27.s }, p8, [x24, #0xc, MUL VL]\n"
536 "addvl x24, x24, #16\n"
537 "35:" // Width 4: Output done
538 "subs x26, x26, #0x4\n"
539 "sub %x[N], %x[N], x27, LSL #2\n"
540 "bgt 4b\n"
541 "36:" // Exit
542 ".inst 0xd503467f // SMSTOP\n"
543 "ptrue p1.b\n"
544 : [N] "+&r" (N)
545 : [A_ptr] "r" (A_ptr), [B_ptr] "r" (B_ptr), [K] "r" (K), [args_ptr] "r" (&ka), [bias] "r" (bias), [flags] "r" (flags), [offset_max] "I" (offsetof(KernelArgs, maxval)), [offset_min] "I" (offsetof(KernelArgs, minval)), [output_ptr] "r" (output_ptr)
546 : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15", "x9", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
547 );
548 }
549
550 } // namespace arm_gemm
551
552 #endif // ARM_COMPUTE_ENABLE_SME2
553 #endif
554