30 #if defined(ARM_COMPUTE_ENABLE_SVE)
35 void sve_u8q_nhwc_3x3_s2_output2x2_mla_depthfirst_impl(
36 const unsigned int n_channels,
37 const uint8_t *
const *
const inptrs,
38 const uint8_t *
const weights,
39 const int32_t *
const bias,
43 uint8_t *
const *
const outptrs
48 long unsigned int n_channels;
54 uint8_t *
const *
const outptrs;
55 const uint8_t *inptrs[25];
58 long unsigned int n_channels,
59 const uint8_t *
const *inptrs_raw,
60 const void *
const weights,
61 const int32_t *
const bias,
65 uint8_t *
const *outptrs
66 ) : n_channels(n_channels), weights(weights),
bias(
bias),
70 inptrs[0] = inptrs_raw[12];
71 inptrs[1] = inptrs_raw[0];
72 inptrs[2] = inptrs_raw[1];
73 inptrs[3] = inptrs_raw[3];
74 inptrs[4] = inptrs_raw[4];
75 inptrs[5] = inptrs_raw[5];
76 inptrs[6] = inptrs_raw[6];
77 inptrs[7] = inptrs_raw[2];
78 inptrs[8] = inptrs_raw[8];
79 inptrs[9] = inptrs_raw[9];
80 inptrs[10] = inptrs_raw[7];
81 inptrs[11] = inptrs_raw[15];
82 inptrs[12] = inptrs_raw[10];
83 inptrs[13] = inptrs_raw[16];
84 inptrs[14] = inptrs_raw[11];
85 inptrs[15] = inptrs_raw[18];
86 inptrs[16] = inptrs_raw[13];
87 inptrs[17] = inptrs_raw[19];
88 inptrs[18] = inptrs_raw[20];
89 inptrs[19] = inptrs_raw[14];
90 inptrs[20] = inptrs_raw[21];
91 inptrs[21] = inptrs_raw[17];
92 inptrs[22] = inptrs_raw[23];
93 inptrs[23] = inptrs_raw[22];
94 inptrs[24] = inptrs_raw[24];
99 const Params params(n_channels, inptrs, weights,
bias, qp,
102 __asm__ __volatile__(
104 "ldr x25, [%x[params], %[offsetof_Params_requant]]\n"
106 "ldr x24, [%x[params], %[offsetof_Params_outptrs]]\n"
108 "add x21, x25, %[offsetof_Requantize32_a_offset]\n"
109 "ldr x8, [%x[params], %[offsetof_Params_n_channels]]\n"
110 "ldr x17, [%x[params], %[offsetof_Params_weights]]\n"
111 "add x20, x25, %[offsetof_Requantize32_b_offset]\n"
112 "add x22, x25, %[offsetof_Requantize32_c_offset]\n"
113 "ld1rb { z26.b }, p4/Z, [x21]\n"
114 "ld1rb { z13.b }, p4/Z, [x20]\n"
115 "add x21, x25, %[offsetof_Requantize32_minval]\n"
116 "add x20, x25, %[offsetof_Requantize32_maxval]\n"
117 "ld1rh { z19.h }, p4/Z, [x22]\n"
118 "ld1rh { z12.h }, p4/Z, [x21]\n"
119 "ld1rh { z9.h }, p4/Z, [x20]\n"
120 "ldp x16, x15, [x24, #0x0]\n"
122 "whilelt p3.h, x7, x8\n"
123 "ldp x14, x13, [x24, #0x10]\n"
124 "whilelt p2.s, x7, x8\n"
125 "whilelt p1.s, x23, x8\n"
126 "ldr x12, [%x[params], %[offsetof_Params_bias]]\n"
127 "ld1b { z25.h }, p4/Z, [x17]\n"
128 "ld1b { z30.h }, p4/Z, [x17, #1, MUL VL]\n"
129 "add x11, %x[params], %[offsetof_Params_inptrs]\n"
131 "ld1b { z14.h }, p4/Z, [x17, #2, MUL VL]\n"
132 "ld1b { z4.h }, p4/Z, [x17, #3, MUL VL]\n"
133 ".inst 0x454d1b39 // usublb z25.h, z25.b, z13.b\n"
134 ".inst 0x454d1bde // usublb z30.h, z30.b, z13.b\n"
135 "ld1b { z10.h }, p4/Z, [x17, #4, MUL VL]\n"
136 "ld1b { z3.h }, p4/Z, [x17, #5, MUL VL]\n"
137 ".inst 0x454d19ce // usublb z14.h, z14.b, z13.b\n"
138 ".inst 0x454d1884 // usublb z4.h, z4.b, z13.b\n"
139 "ld1b { z23.h }, p4/Z, [x17, #6, MUL VL]\n"
140 "ld1b { z7.h }, p4/Z, [x17, #7, MUL VL]\n"
141 "inch x17, ALL, MUL #8\n"
142 ".inst 0x454d194a // usublb z10.h, z10.b, z13.b\n"
143 "ld1w { z17.s }, p2/Z, [x12]\n"
144 "ld1w { z16.s }, p1/Z, [x12, #1, MUL VL]\n"
145 "uzp1 z8.s, z17.s, z16.s\n"
146 "uzp2 z24.s, z17.s, z16.s\n"
147 "ld1b { z2.h }, p4/Z, [x17]\n"
148 "ldp x27, x26, [x11, #0x0]\n"
149 "addvl x12, x12, #2\n"
151 "ldp x25, x24, [x11, #0x10]\n"
152 "ldp x23, x22, [x11, #0x20]\n"
155 "ldp x21, x20, [x11, #0x30]\n"
156 "ld1b { z21.h }, p3/Z, [x27, x7]\n"
159 "ld1b { z22.h }, p3/Z, [x26, x7]\n"
160 "ld1b { z11.h }, p3/Z, [x25, x7]\n"
162 ".inst 0x454d1863 // usublb z3.h, z3.b, z13.b\n"
163 "ld1b { z20.h }, p3/Z, [x24, x7]\n"
164 "ld1b { z27.h }, p3/Z, [x23, x7]\n"
165 ".inst 0x454d1af7 // usublb z23.h, z23.b, z13.b\n"
166 ".inst 0x454d18e7 // usublb z7.h, z7.b, z13.b\n"
167 "ld1b { z28.h }, p3/Z, [x22, x7]\n"
168 "ld1b { z16.h }, p3/Z, [x21, x7]\n"
169 ".inst 0x454d1842 // usublb z2.h, z2.b, z13.b\n"
170 ".inst 0x455a1ab5 // usublb z21.h, z21.b, z26.b\n"
171 "ld1b { z31.h }, p3/Z, [x20, x7]\n"
172 "ldr x9, [%x[params], %[offsetof_Params_requant_muls]]\n"
173 ".inst 0x455a1ad6 // usublb z22.h, z22.b, z26.b\n"
174 ".inst 0x455a196b // usublb z11.h, z11.b, z26.b\n"
175 "ldr x28, [%x[params], %[offsetof_Params_requant_shifts]]\n"
176 "str x12, [%x[params], %[offsetof_Params_bias]]\n"
177 ".inst 0x455a1a94 // usublb z20.h, z20.b, z26.b\n"
178 ".inst 0x455a1b7b // usublb z27.h, z27.b, z26.b\n"
179 ".inst 0x455a1b9c // usublb z28.h, z28.b, z26.b\n"
180 ".inst 0x455a1a10 // usublb z16.h, z16.b, z26.b\n"
181 ".inst 0x455a1bff // usublb z31.h, z31.b, z26.b\n"
183 ".inst 0x448242a8 // smlalb z8.s, p4/M, z21.h, z2.h\n"
184 "ldr x21, [x11, #0x58]\n"
185 "ldr x20, [x11, #0x78]\n"
186 ".inst 0x448246b8 // smlalt z24.s, p4/M, z21.h, z2.h\n"
187 ".inst 0x449942c8 // smlalb z8.s, p4/M, z22.h, z25.h\n"
188 "ld1b { z17.h }, p3/Z, [x21, x7]\n"
189 "ld1b { z29.h }, p3/Z, [x20, x7]\n"
190 ".inst 0x449742b2 // smlalb z18.s, p4/M, z21.h, z23.h\n"
191 "ldr x21, [x11, #0x60]\n"
192 "ldr x20, [x11, #0x80]\n"
193 ".inst 0x448e42af // smlalb z15.s, p4/M, z21.h, z14.h\n"
194 ".inst 0x449942a5 // smlalb z5.s, p4/M, z21.h, z25.h\n"
195 ".inst 0x449946d8 // smlalt z24.s, p4/M, z22.h, z25.h\n"
196 ".inst 0x455a1a31 // usublb z17.h, z17.b, z26.b\n"
197 ".inst 0x449e4168 // smlalb z8.s, p4/M, z11.h, z30.h\n"
198 "ld1b { z22.h }, p3/Z, [x21, x7]\n"
199 ".inst 0x455a1bbd // usublb z29.h, z29.b, z26.b\n"
200 ".inst 0x449746a0 // smlalt z0.s, p4/M, z21.h, z23.h\n"
201 ".inst 0x448e46a1 // smlalt z1.s, p4/M, z21.h, z14.h\n"
202 "ldr x21, [x11, #0x68]\n"
203 ".inst 0x449946a6 // smlalt z6.s, p4/M, z21.h, z25.h\n"
204 "ld1b { z21.h }, p3/Z, [x20, x7]\n"
205 "ldr x20, [x11, #0x88]\n"
206 ".inst 0x449e4292 // smlalb z18.s, p4/M, z20.h, z30.h\n"
207 ".inst 0x4484422f // smlalb z15.s, p4/M, z17.h, z4.h\n"
208 ".inst 0x448a43a5 // smlalb z5.s, p4/M, z29.h, z10.h\n"
209 ".inst 0x455a1ad6 // usublb z22.h, z22.b, z26.b\n"
210 "ldr x22, [x11, #0x40]\n"
211 ".inst 0x449e4578 // smlalt z24.s, p4/M, z11.h, z30.h\n"
212 ".inst 0x455a1ab5 // usublb z21.h, z21.b, z26.b\n"
213 ".inst 0x44844388 // smlalb z8.s, p4/M, z28.h, z4.h\n"
214 "ld1b { z11.h }, p3/Z, [x21, x7]\n"
215 ".inst 0x449e4680 // smlalt z0.s, p4/M, z20.h, z30.h\n"
216 "ld1b { z20.h }, p3/Z, [x20, x7]\n"
217 ".inst 0x44844621 // smlalt z1.s, p4/M, z17.h, z4.h\n"
218 "ldr x21, [x11, #0x70]\n"
219 ".inst 0x448a47a6 // smlalt z6.s, p4/M, z29.h, z10.h\n"
220 "ldr x20, [x11, #0x98]\n"
221 ".inst 0x448e4372 // smlalb z18.s, p4/M, z27.h, z14.h\n"
222 "ldr x23, [x11, #0x50]\n"
223 ".inst 0x449942cf // smlalb z15.s, p4/M, z22.h, z25.h\n"
224 ".inst 0x449e42a5 // smlalb z5.s, p4/M, z21.h, z30.h\n"
225 ".inst 0x455a196b // usublb z11.h, z11.b, z26.b\n"
226 "ld1b { z17.h }, p3/Z, [x22, x7]\n"
227 ".inst 0x44844798 // smlalt z24.s, p4/M, z28.h, z4.h\n"
228 ".inst 0x455a1a94 // usublb z20.h, z20.b, z26.b\n"
229 ".inst 0x448a4208 // smlalb z8.s, p4/M, z16.h, z10.h\n"
230 "ld1b { z29.h }, p3/Z, [x21, x7]\n"
231 "ld1b { z28.h }, p3/Z, [x20, x7]\n"
232 ".inst 0x448e4760 // smlalt z0.s, p4/M, z27.h, z14.h\n"
233 "ldr x22, [x11, #0x48]\n"
234 ".inst 0x449946c1 // smlalt z1.s, p4/M, z22.h, z25.h\n"
235 ".inst 0x449e46a6 // smlalt z6.s, p4/M, z21.h, z30.h\n"
236 "ldr x21, [x11, #0x90]\n"
237 "ldr x20, [x11, #0xa8]\n"
238 ".inst 0x449943f2 // smlalb z18.s, p4/M, z31.h, z25.h\n"
239 "ld1b { z27.h }, p3/Z, [x23, x7]\n"
240 ".inst 0x448a416f // smlalb z15.s, p4/M, z11.h, z10.h\n"
241 ".inst 0x44834285 // smlalb z5.s, p4/M, z20.h, z3.h\n"
242 ".inst 0x455a1a31 // usublb z17.h, z17.b, z26.b\n"
243 ".inst 0x448a4618 // smlalt z24.s, p4/M, z16.h, z10.h\n"
244 ".inst 0x455a1bbd // usublb z29.h, z29.b, z26.b\n"
245 ".inst 0x448e43e8 // smlalb z8.s, p4/M, z31.h, z14.h\n"
246 "ld1b { z16.h }, p3/Z, [x22, x7]\n"
247 ".inst 0x455a1b9c // usublb z28.h, z28.b, z26.b\n"
248 ".inst 0x449947e0 // smlalt z0.s, p4/M, z31.h, z25.h\n"
249 "ld1b { z25.h }, p3/Z, [x21, x7]\n"
250 ".inst 0x448a4561 // smlalt z1.s, p4/M, z11.h, z10.h\n"
251 "ld1b { z11.h }, p3/Z, [x20, x7]\n"
252 ".inst 0x455a1b7b // usublb z27.h, z27.b, z26.b\n"
253 ".inst 0x44834686 // smlalt z6.s, p4/M, z20.h, z3.h\n"
254 "ldr x21, [x11, #0xa0]\n"
255 "ldr x20, [x11, #0xb0]\n"
256 ".inst 0x448a4232 // smlalb z18.s, p4/M, z17.h, z10.h\n"
257 ".inst 0x449e43af // smlalb z15.s, p4/M, z29.h, z30.h\n"
258 ".inst 0x455a1a10 // usublb z16.h, z16.b, z26.b\n"
259 ".inst 0x448e4385 // smlalb z5.s, p4/M, z28.h, z14.h\n"
260 ".inst 0x448e47f8 // smlalt z24.s, p4/M, z31.h, z14.h\n"
261 ".inst 0x455a1b39 // usublb z25.h, z25.b, z26.b\n"
262 "ld1b { z20.h }, p3/Z, [x21, x7]\n"
263 ".inst 0x455a196b // usublb z11.h, z11.b, z26.b\n"
264 ".inst 0x44834368 // smlalb z8.s, p4/M, z27.h, z3.h\n"
265 "ld1b { z31.h }, p3/Z, [x20, x7]\n"
266 ".inst 0x448a4620 // smlalt z0.s, p4/M, z17.h, z10.h\n"
267 ".inst 0x449e47a1 // smlalt z1.s, p4/M, z29.h, z30.h\n"
268 ".inst 0x448e4786 // smlalt z6.s, p4/M, z28.h, z14.h\n"
269 "ldr x20, [x11, #0xb8]\n"
270 ".inst 0x455a1a94 // usublb z20.h, z20.b, z26.b\n"
271 ".inst 0x44834212 // smlalb z18.s, p4/M, z16.h, z3.h\n"
272 ".inst 0x4497432f // smlalb z15.s, p4/M, z25.h, z23.h\n"
273 ".inst 0x455a1bff // usublb z31.h, z31.b, z26.b\n"
274 "ld1b { z30.h }, p3/Z, [x20, x7]\n"
275 ".inst 0x44844165 // smlalb z5.s, p4/M, z11.h, z4.h\n"
276 ".inst 0x44834778 // smlalt z24.s, p4/M, z27.h, z3.h\n"
277 "ldr x20, [x11, #0xc0]\n"
278 "ld1w { z17.s }, p2/Z, [x9]\n"
279 ".inst 0x449742c8 // smlalb z8.s, p4/M, z22.h, z23.h\n"
280 ".inst 0x44834600 // smlalt z0.s, p4/M, z16.h, z3.h\n"
281 "ld1w { z14.s }, p1/Z, [x9, #1, MUL VL]\n"
282 ".inst 0x455a1bde // usublb z30.h, z30.b, z26.b\n"
283 ".inst 0x44974721 // smlalt z1.s, p4/M, z25.h, z23.h\n"
284 ".inst 0x44844566 // smlalt z6.s, p4/M, z11.h, z4.h\n"
285 "ld1b { z25.h }, p3/Z, [x20, x7]\n"
286 "uzp1 z10.s, z17.s, z14.s\n"
287 ".inst 0x44844372 // smlalb z18.s, p4/M, z27.h, z4.h\n"
288 ".inst 0x4487428f // smlalb z15.s, p4/M, z20.h, z7.h\n"
289 "uzp2 z14.s, z17.s, z14.s\n"
290 "ld1w { z17.s }, p2/Z, [x28]\n"
291 ".inst 0x448743e5 // smlalb z5.s, p4/M, z31.h, z7.h\n"
292 ".inst 0x449746d8 // smlalt z24.s, p4/M, z22.h, z23.h\n"
293 "ld1w { z16.s }, p1/Z, [x28, #1, MUL VL]\n"
294 ".inst 0x455a1b39 // usublb z25.h, z25.b, z26.b\n"
295 ".inst 0x448743a8 // smlalb z8.s, p4/M, z29.h, z7.h\n"
296 ".inst 0x44844760 // smlalt z0.s, p4/M, z27.h, z4.h\n"
297 "uzp1 z4.s, z17.s, z16.s\n"
299 ".inst 0x44874681 // smlalt z1.s, p4/M, z20.h, z7.h\n"
300 ".inst 0x448747e6 // smlalt z6.s, p4/M, z31.h, z7.h\n"
301 ".inst 0x04aa7508 // sqrdmulh z8.s, z8.s, z10.s\n"
302 "whilelt p0.h, x10, x8\n"
303 ".inst 0x448742b2 // smlalb z18.s, p4/M, z21.h, z7.h\n"
304 ".inst 0x4483416f // smlalb z15.s, p4/M, z11.h, z3.h\n"
305 "uzp2 z22.s, z17.s, z16.s\n"
307 ".inst 0x449743c5 // smlalb z5.s, p4/M, z30.h, z23.h\n"
308 ".inst 0x448747b8 // smlalt z24.s, p4/M, z29.h, z7.h\n"
309 "and z17.d, z8.d, z4.d\n"
311 ".inst 0x448746a0 // smlalt z0.s, p4/M, z21.h, z7.h\n"
312 ".inst 0x44834561 // smlalt z1.s, p4/M, z11.h, z3.h\n"
313 ".inst 0x04ae7718 // sqrdmulh z24.s, z24.s, z14.s\n"
315 ".inst 0x449747c6 // smlalt z6.s, p4/M, z30.h, z23.h\n"
316 ".inst 0x44824392 // smlalb z18.s, p4/M, z28.h, z2.h\n"
317 "asr z17.s, z17.s, #0x1f\n"
318 "whilelt p2.s, x7, x8\n"
319 ".inst 0x448243cf // smlalb z15.s, p4/M, z30.h, z2.h\n"
320 ".inst 0x44824325 // smlalb z5.s, p4/M, z25.h, z2.h\n"
321 "and z16.d, z24.d, z22.d\n"
322 "whilelt p1.s, x20, x8\n"
323 ".inst 0x44824780 // smlalt z0.s, p4/M, z28.h, z2.h\n"
324 ".inst 0x448247c1 // smlalt z1.s, p4/M, z30.h, z2.h\n"
325 ".inst 0x04aa7652 // sqrdmulh z18.s, z18.s, z10.s\n"
326 "ldr x20, [%x[params], %[offsetof_Params_bias]]\n"
327 ".inst 0x44824726 // smlalt z6.s, p4/M, z25.h, z2.h\n"
328 ".inst 0x04aa75ef // sqrdmulh z15.s, z15.s, z10.s\n"
329 "whilelt p3.h, x7, x8\n"
331 ".inst 0x04aa74a5 // sqrdmulh z5.s, z5.s, z10.s\n"
332 "sqadd z8.s, z8.s, z17.s\n"
333 ".inst 0x44829088 // srshl z8.s, p4/M, z8.s, z4.s\n"
334 "addvl x28, x28, #2\n"
335 "asr z16.s, z16.s, #0x1f\n"
336 "and z21.d, z18.d, z4.d\n"
337 ".inst 0x04ae7400 // sqrdmulh z0.s, z0.s, z14.s\n"
338 "and z20.d, z15.d, z4.d\n"
339 ".inst 0x04ae7421 // sqrdmulh z1.s, z1.s, z14.s\n"
340 "and z28.d, z5.d, z4.d\n"
341 ".inst 0x04ae74c6 // sqrdmulh z6.s, z6.s, z14.s\n"
342 "sqadd z24.s, z24.s, z16.s\n"
343 ".inst 0x448292d8 // srshl z24.s, p4/M, z24.s, z22.s\n"
344 "asr z21.s, z21.s, #0x1f\n"
345 "and z25.d, z0.d, z22.d\n"
346 "asr z20.s, z20.s, #0x1f\n"
347 "and z17.d, z1.d, z22.d\n"
348 "asr z28.s, z28.s, #0x1f\n"
349 "and z16.d, z6.d, z22.d\n"
350 "sqadd z18.s, z18.s, z21.s\n"
351 "asr z25.s, z25.s, #0x1f\n"
352 ".inst 0x44829092 // srshl z18.s, p4/M, z18.s, z4.s\n"
353 "sqadd z15.s, z15.s, z20.s\n"
354 "asr z17.s, z17.s, #0x1f\n"
355 ".inst 0x4482908f // srshl z15.s, p4/M, z15.s, z4.s\n"
356 "sqadd z5.s, z5.s, z28.s\n"
357 "asr z16.s, z16.s, #0x1f\n"
358 ".inst 0x44829085 // srshl z5.s, p4/M, z5.s, z4.s\n"
359 "sqadd z0.s, z0.s, z25.s\n"
360 "sqadd z1.s, z1.s, z17.s\n"
361 ".inst 0x448292c0 // srshl z0.s, p4/M, z0.s, z22.s\n"
362 ".inst 0x448292c1 // srshl z1.s, p4/M, z1.s, z22.s\n"
363 "sqadd z6.s, z6.s, z16.s\n"
364 ".inst 0x45304108 // sqxtnb z8.h, z8.s\n"
365 ".inst 0x448292c6 // srshl z6.s, p4/M, z6.s, z22.s\n"
366 ".inst 0x45304252 // sqxtnb z18.h, z18.s\n"
367 ".inst 0x453041ef // sqxtnb z15.h, z15.s\n"
368 ".inst 0x453040a5 // sqxtnb z5.h, z5.s\n"
369 ".inst 0x45304708 // sqxtnt z8.h, z24.s\n"
370 ".inst 0x45304412 // sqxtnt z18.h, z0.s\n"
371 ".inst 0x4530442f // sqxtnt z15.h, z1.s\n"
372 ".inst 0x453044c5 // sqxtnt z5.h, z6.s\n"
373 "sqadd z8.h, z8.h, z19.h\n"
374 "smax z8.h, p4/M, z8.h, z12.h\n"
375 "smin z8.h, p4/M, z8.h, z9.h\n"
376 "sqadd z18.h, z18.h, z19.h\n"
377 "sqadd z15.h, z15.h, z19.h\n"
378 "smax z18.h, p4/M, z18.h, z12.h\n"
379 "smax z15.h, p4/M, z15.h, z12.h\n"
380 "sqadd z5.h, z5.h, z19.h\n"
381 "smax z5.h, p4/M, z5.h, z12.h\n"
382 "smin z18.h, p4/M, z18.h, z9.h\n"
383 "st1b { z8.h }, p0, [x16, x10]\n"
384 "smin z15.h, p4/M, z15.h, z9.h\n"
385 "smin z5.h, p4/M, z5.h, z9.h\n"
386 "st1b { z18.h }, p0, [x15, x10]\n"
387 "st1b { z15.h }, p0, [x14, x10]\n"
388 "st1b { z5.h }, p0, [x13, x10]\n"
389 "ld1b { z25.h }, p4/Z, [x17]\n"
390 "ld1b { z30.h }, p4/Z, [x17, #1, MUL VL]\n"
392 "ld1b { z14.h }, p4/Z, [x17, #2, MUL VL]\n"
393 "ld1b { z4.h }, p4/Z, [x17, #3, MUL VL]\n"
394 ".inst 0x454d1b39 // usublb z25.h, z25.b, z13.b\n"
395 ".inst 0x454d1bde // usublb z30.h, z30.b, z13.b\n"
396 "ld1b { z10.h }, p4/Z, [x17, #4, MUL VL]\n"
397 "ld1b { z3.h }, p4/Z, [x17, #5, MUL VL]\n"
398 ".inst 0x454d19ce // usublb z14.h, z14.b, z13.b\n"
399 ".inst 0x454d1884 // usublb z4.h, z4.b, z13.b\n"
400 "ld1b { z23.h }, p4/Z, [x17, #6, MUL VL]\n"
401 "ld1b { z7.h }, p4/Z, [x17, #7, MUL VL]\n"
402 "inch x17, ALL, MUL #8\n"
403 ".inst 0x454d194a // usublb z10.h, z10.b, z13.b\n"
404 "ld1w { z17.s }, p2/Z, [x20]\n"
405 "ld1w { z16.s }, p1/Z, [x20, #1, MUL VL]\n"
406 "uzp1 z8.s, z17.s, z16.s\n"
407 "uzp2 z24.s, z17.s, z16.s\n"
408 "ld1b { z2.h }, p4/Z, [x17]\n"
409 "ldp x27, x26, [x11, #0x0]\n"
410 "addvl x20, x20, #2\n"
411 "str x20, [%x[params], %[offsetof_Params_bias]]\n"
412 "ldp x25, x24, [x11, #0x10]\n"
413 "ldp x23, x22, [x11, #0x20]\n"
416 "ldp x21, x20, [x11, #0x30]\n"
417 "ld1b { z21.h }, p3/Z, [x27, x7]\n"
420 "ld1b { z22.h }, p3/Z, [x26, x7]\n"
421 "ld1b { z11.h }, p3/Z, [x25, x7]\n"
424 "ld1b { z20.h }, p3/Z, [x24, x7]\n"
425 "ld1b { z27.h }, p3/Z, [x23, x7]\n"
426 ".inst 0x454d1863 // usublb z3.h, z3.b, z13.b\n"
427 ".inst 0x454d1af7 // usublb z23.h, z23.b, z13.b\n"
428 "ld1b { z28.h }, p3/Z, [x22, x7]\n"
429 "ld1b { z16.h }, p3/Z, [x21, x7]\n"
430 ".inst 0x454d18e7 // usublb z7.h, z7.b, z13.b\n"
431 ".inst 0x454d1842 // usublb z2.h, z2.b, z13.b\n"
432 "ld1b { z31.h }, p3/Z, [x20, x7]\n"
433 ".inst 0x455a1ab5 // usublb z21.h, z21.b, z26.b\n"
434 ".inst 0x455a1ad6 // usublb z22.h, z22.b, z26.b\n"
435 ".inst 0x455a196b // usublb z11.h, z11.b, z26.b\n"
436 ".inst 0x455a1a94 // usublb z20.h, z20.b, z26.b\n"
437 ".inst 0x455a1b7b // usublb z27.h, z27.b, z26.b\n"
438 ".inst 0x455a1b9c // usublb z28.h, z28.b, z26.b\n"
439 ".inst 0x455a1a10 // usublb z16.h, z16.b, z26.b\n"
440 ".inst 0x455a1bff // usublb z31.h, z31.b, z26.b\n"
443 : [offsetof_Params_bias]
"I" (offsetof(Params,
bias)), [offsetof_Params_inptrs]
"I" (offsetof(Params, inptrs)), [offsetof_Params_n_channels]
"I" (offsetof(Params, n_channels)), [offsetof_Params_outptrs]
"I" (offsetof(Params, outptrs)), [offsetof_Params_requant]
"I" (offsetof(Params, requant)), [offsetof_Params_requant_muls]
"I" (offsetof(Params,
requant_muls)), [offsetof_Params_requant_shifts]
"I" (offsetof(Params,
requant_shifts)), [offsetof_Params_weights]
"I" (offsetof(Params, weights)), [offsetof_Requantize32_a_offset]
"I" (offsetof(
arm_gemm::Requantize32, a_offset)), [offsetof_Requantize32_b_offset]
"I" (offsetof(
arm_gemm::Requantize32, b_offset)), [offsetof_Requantize32_c_offset]
"I" (offsetof(
arm_gemm::Requantize32, c_offset)), [offsetof_Requantize32_maxval]
"I" (offsetof(
arm_gemm::Requantize32, maxval)), [offsetof_Requantize32_minval]
"I" (offsetof(
arm_gemm::Requantize32, minval)), [params]
"r" (¶ms)
444 :
"cc",
"memory",
"p0",
"p1",
"p2",
"p3",
"p4",
"x7",
"x8",
"x9",
"x10",
"x11",
"x12",
"x13",
"x14",
"x15",
"x16",
"x17",
"x20",
"x21",
"x22",
"x23",
"x24",
"x25",
"x26",
"x27",
"x28",
"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"
451 #endif // defined(ARM_COMPUTE_ENABLE_SVE)