28 #if defined(ARM_COMPUTE_ENABLE_SME2)
33 void sme2_fp32_nhwc_3x3_s2_output2x2_mla_depthfirst_indirect_impl(
34 const float *
const *
const input_ptrs,
35 float *
const *
const outptrs,
37 unsigned int n_channels,
44 float *
const *outptrs;
47 const float *inptrs[25];
50 const float *
const *
const input_ptrs,
51 float *
const *
const outptrs,
52 const void *
const params,
55 ) : outptrs(outptrs), params(params), min(min), max(max)
57 inptrs[0] = input_ptrs[12];
58 inptrs[1] = input_ptrs[0];
59 inptrs[2] = input_ptrs[1];
60 inptrs[3] = input_ptrs[3];
61 inptrs[4] = input_ptrs[4];
62 inptrs[5] = input_ptrs[5];
63 inptrs[6] = input_ptrs[6];
64 inptrs[7] = input_ptrs[2];
65 inptrs[8] = input_ptrs[8];
66 inptrs[9] = input_ptrs[9];
67 inptrs[10] = input_ptrs[7];
68 inptrs[11] = input_ptrs[15];
69 inptrs[12] = input_ptrs[10];
70 inptrs[13] = input_ptrs[16];
71 inptrs[14] = input_ptrs[11];
72 inptrs[15] = input_ptrs[18];
73 inptrs[16] = input_ptrs[13];
74 inptrs[17] = input_ptrs[19];
75 inptrs[18] = input_ptrs[20];
76 inptrs[19] = input_ptrs[14];
77 inptrs[20] = input_ptrs[21];
78 inptrs[21] = input_ptrs[17];
79 inptrs[22] = input_ptrs[23];
80 inptrs[23] = input_ptrs[22];
81 inptrs[24] = input_ptrs[24];
86 Args params_struct(input_ptrs, outptrs, params,
90 "ldr x20, [%x[params_struct], %[offsetof_args_outptrs]]\n"
91 ".inst 0xd503477f // SMSTART ZA\n"
92 "add x16, %x[params_struct], %[offsetof_Args_inptrs]\n"
94 "ldr x15, [%x[params_struct], %[offsetof_args_params]]\n"
95 ".inst 0x25207810 // ptrue pn8.b\n"
96 "ld1w { z26.s }, p3/Z, [x15]\n"
97 "addvl x15, x15, #1\n"
98 "ldp x14, x13, [x20, #0x0]\n"
100 ".inst 0xa040c1e0 // ld1w { z0.s-z3.s }, pn8.b/Z, [x15]\n"
101 "addvl x15, x15, #4\n"
102 "ldp x11, x10, [x20, #0x10]\n"
104 "whilelt p2.s, XZR, %x[n_channels]\n"
105 ".inst 0xa040c1e4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x15]\n"
106 "ldp x28, x26, [x16, #0x0]\n"
107 "addvl x15, x15, #4\n"
108 "cmp x12, %x[n_channels]\n"
109 "ld1rw { z25.s }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
110 "ldp x25, x24, [x16, #0x10]\n"
111 "ld1rw { z24.s }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
112 "sub x27, XZR, x12\n"
113 "ldp x23, x22, [x16, #0x20]\n"
114 "ld1w { z8.s }, p3/Z, [x15]\n"
115 "addvl x15, x15, #1\n"
116 "ldp x21, x20, [x16, #0x30]\n"
117 "ld1w { z9.s }, p2/Z, [x28, x9, LSL #2]\n"
118 "ld1w { z10.s }, p2/Z, [x26, x9, LSL #2]\n"
119 "ld1w { z11.s }, p2/Z, [x25, x9, LSL #2]\n"
120 "ld1w { z12.s }, p2/Z, [x24, x9, LSL #2]\n"
121 "ld1w { z13.s }, p2/Z, [x23, x9, LSL #2]\n"
122 "ld1w { z14.s }, p2/Z, [x22, x9, LSL #2]\n"
123 "ld1w { z15.s }, p2/Z, [x21, x9, LSL #2]\n"
124 "ld1w { z16.s }, p2/Z, [x20, x9, LSL #2]\n"
127 "movprfx z28, z26\n fmla z28.s, p3/M, z8.s, z9.s\n"
128 "movprfx z29, z26\n fmla z29.s, p3/M, z6.s, z9.s\n"
129 "ldr x21, [x16, #0x40]\n"
130 "whilelt p1.s, x12, %x[n_channels]\n"
131 "fmla z28.s, p3/M, z0.s, z10.s\n"
132 "fmla z29.s, p3/M, z1.s, z12.s\n"
133 "ldr x20, [x16, #0x48]\n"
134 "ld1w { z18.s }, p2/Z, [x20, x9, LSL #2]\n"
135 "fmla z28.s, p3/M, z1.s, z11.s\n"
136 "fmla z29.s, p3/M, z2.s, z13.s\n"
137 "ld1w { z22.s }, p2/Z, [x21, x9, LSL #2]\n"
138 "ldr x20, [x16, #0x50]\n"
139 "fmla z28.s, p3/M, z3.s, z14.s\n"
140 "fmla z29.s, p3/M, z0.s, z16.s\n"
141 "ld1w { z17.s }, p2/Z, [x20, x9, LSL #2]\n"
142 "ldr x20, [x16, #0x58]\n"
143 "fmla z28.s, p3/M, z4.s, z15.s\n"
144 "fmla z29.s, p3/M, z4.s, z22.s\n"
145 "ldr x21, [x16, #0x78]\n"
146 "ld1w { z23.s }, p2/Z, [x20, x9, LSL #2]\n"
147 "fmla z28.s, p3/M, z2.s, z16.s\n"
148 "fmla z29.s, p3/M, z5.s, z18.s\n"
149 "ldr x20, [x16, #0x60]\n"
150 "ld1w { z13.s }, p2/Z, [x20, x9, LSL #2]\n"
151 "movprfx z30, z26\n fmla z30.s, p3/M, z2.s, z9.s\n"
152 "movprfx z31, z26\n fmla z31.s, p3/M, z0.s, z9.s\n"
153 "ldr x20, [x16, #0x80]\n"
154 "ld1w { z18.s }, p2/Z, [x20, x9, LSL #2]\n"
155 "fmla z28.s, p3/M, z5.s, z17.s\n"
156 "fmla z29.s, p3/M, z3.s, z17.s\n"
157 "ld1w { z16.s }, p2/Z, [x21, x9, LSL #2]\n"
158 "ldr x21, [x16, #0x68]\n"
159 "fmla z30.s, p3/M, z3.s, z23.s\n"
160 "fmla z31.s, p3/M, z4.s, z16.s\n"
161 "ldr x20, [x16, #0x88]\n"
162 "ld1w { z17.s }, p2/Z, [x21, x9, LSL #2]\n"
163 "fmla z30.s, p3/M, z0.s, z13.s\n"
164 "fmla z31.s, p3/M, z1.s, z18.s\n"
165 "ld1w { z16.s }, p2/Z, [x20, x9, LSL #2]\n"
166 "ldr x21, [x16, #0x70]\n"
167 "ldr x20, [x16, #0x98]\n"
168 "fmla z30.s, p3/M, z4.s, z17.s\n"
169 "fmla z31.s, p3/M, z5.s, z16.s\n"
170 "ld1w { z16.s }, p2/Z, [x21, x9, LSL #2]\n"
171 "fmla z28.s, p3/M, z6.s, z13.s\n"
172 "ld1w { z4.s }, p2/Z, [x20, x9, LSL #2]\n"
173 "ldr x21, [x16, #0x90]\n"
174 "fmla z30.s, p3/M, z1.s, z16.s\n"
175 "ldr x20, [x16, #0xa8]\n"
176 "fmla z31.s, p3/M, z2.s, z4.s\n"
177 "fmla z28.s, p3/M, z7.s, z16.s\n"
178 "ld1w { z16.s }, p2/Z, [x21, x9, LSL #2]\n"
179 "ld1w { z17.s }, p2/Z, [x20, x9, LSL #2]\n"
180 "ldr x21, [x16, #0xa0]\n"
181 "fmla z30.s, p3/M, z6.s, z16.s\n"
182 "fmla z31.s, p3/M, z3.s, z17.s\n"
183 "ldr x20, [x16, #0xb0]\n"
184 "ld1w { z16.s }, p2/Z, [x21, x9, LSL #2]\n"
185 "fmla z30.s, p3/M, z7.s, z16.s\n"
186 "fmla z29.s, p3/M, z7.s, z18.s\n"
187 "ld1w { z16.s }, p2/Z, [x20, x9, LSL #2]\n"
188 "ldr x20, [x16, #0xb8]\n"
189 "fmla z31.s, p3/M, z7.s, z16.s\n"
190 "fmla z30.s, p3/M, z5.s, z17.s\n"
191 "ld1w { z17.s }, p2/Z, [x20, x9, LSL #2]\n"
192 "ldr x20, [x16, #0xc0]\n"
193 "fmla z31.s, p3/M, z6.s, z17.s\n"
194 "fmla z29.s, p3/M, z8.s, z4.s\n"
195 "ld1w { z16.s }, p2/Z, [x20, x9, LSL #2]\n"
196 "ldp x20, x26, [x16, #0x0]\n"
197 "fmla z30.s, p3/M, z8.s, z17.s\n"
198 "fmla z31.s, p3/M, z8.s, z16.s\n"
199 "ldp x25, x24, [x16, #0x10]\n"
200 "ld1w { z26.s }, p3/Z, [x15]\n"
201 "addvl x15, x15, #1\n"
203 "ldp x23, x22, [x16, #0x20]\n"
204 "ld1w { z9.s }, p1/Z, [x20, x12, LSL #2]\n"
207 "ldp x21, x20, [x16, #0x30]\n"
208 "ld1w { z10.s }, p1/Z, [x26, x12, LSL #2]\n"
209 "whilelt p2.s, x9, %x[n_channels]\n"
210 ".inst 0xc1b8cb3c // fclamp { z28.s-z31.s }, z25.s, z24.s\n"
211 "ld1w { z11.s }, p1/Z, [x25, x12, LSL #2]\n"
212 "st1w { z28.s }, p0, [x14, x27, LSL #2]\n"
213 "ld1w { z12.s }, p1/Z, [x24, x12, LSL #2]\n"
214 "st1w { z29.s }, p0, [x13, x27, LSL #2]\n"
215 "ld1w { z13.s }, p1/Z, [x23, x12, LSL #2]\n"
216 "st1w { z30.s }, p0, [x11, x27, LSL #2]\n"
217 "ld1w { z14.s }, p1/Z, [x22, x12, LSL #2]\n"
218 "st1w { z31.s }, p0, [x10, x27, LSL #2]\n"
219 "ld1w { z15.s }, p1/Z, [x21, x12, LSL #2]\n"
220 "ld1w { z16.s }, p1/Z, [x20, x12, LSL #2]\n"
222 "cmp x12, %x[n_channels]\n"
223 ".inst 0xa040c1e0 // ld1w { z0.s-z3.s }, pn8.b/Z, [x15]\n"
224 "addvl x15, x15, #4\n"
225 ".inst 0xa040c1e4 // ld1w { z4.s-z7.s }, pn8.b/Z, [x15]\n"
226 "addvl x15, x15, #4\n"
227 "ld1w { z8.s }, p3/Z, [x15]\n"
228 "addvl x15, x15, #1\n"
231 "movprfx z28, z26\n fmla z28.s, p3/M, z8.s, z9.s\n"
232 "movprfx z29, z26\n fmla z29.s, p3/M, z6.s, z9.s\n"
233 "ldr x21, [x16, #0x40]\n"
235 "fmla z28.s, p3/M, z0.s, z10.s\n"
236 "fmla z29.s, p3/M, z1.s, z12.s\n"
237 "ldr x20, [x16, #0x48]\n"
238 "ld1w { z18.s }, p2/Z, [x20, x9, LSL #2]\n"
239 "fmla z28.s, p3/M, z1.s, z11.s\n"
240 "fmla z29.s, p3/M, z2.s, z13.s\n"
241 "ld1w { z17.s }, p2/Z, [x21, x9, LSL #2]\n"
242 "ldr x20, [x16, #0x50]\n"
243 "fmla z28.s, p3/M, z3.s, z14.s\n"
244 "fmla z29.s, p3/M, z0.s, z16.s\n"
245 "ld1w { z20.s }, p2/Z, [x20, x9, LSL #2]\n"
246 "ldr x20, [x16, #0x58]\n"
247 "fmla z28.s, p3/M, z4.s, z15.s\n"
248 "fmla z29.s, p3/M, z4.s, z17.s\n"
249 "ldr x21, [x16, #0x78]\n"
250 "ld1w { z17.s }, p2/Z, [x20, x9, LSL #2]\n"
251 "fmla z28.s, p3/M, z2.s, z16.s\n"
252 "fmla z29.s, p3/M, z5.s, z18.s\n"
253 "ldr x20, [x16, #0x60]\n"
254 "ld1w { z18.s }, p2/Z, [x20, x9, LSL #2]\n"
255 "movprfx z30, z26\n fmla z30.s, p3/M, z2.s, z9.s\n"
256 "movprfx z31, z26\n fmla z31.s, p3/M, z0.s, z9.s\n"
257 "ldr x20, [x16, #0x80]\n"
258 "ld1w { z19.s }, p2/Z, [x20, x9, LSL #2]\n"
259 "fmla z28.s, p3/M, z5.s, z20.s\n"
260 "fmla z29.s, p3/M, z3.s, z20.s\n"
261 "ld1w { z16.s }, p2/Z, [x21, x9, LSL #2]\n"
262 "ldr x21, [x16, #0x68]\n"
263 "fmla z30.s, p3/M, z3.s, z17.s\n"
264 "fmla z31.s, p3/M, z4.s, z16.s\n"
265 "ldr x20, [x16, #0x88]\n"
266 "ld1w { z17.s }, p2/Z, [x21, x9, LSL #2]\n"
267 "fmla z30.s, p3/M, z0.s, z18.s\n"
268 "fmla z31.s, p3/M, z1.s, z19.s\n"
269 "ld1w { z16.s }, p2/Z, [x20, x9, LSL #2]\n"
270 "ldr x21, [x16, #0x70]\n"
271 "ldr x20, [x16, #0x98]\n"
272 "fmla z30.s, p3/M, z4.s, z17.s\n"
273 "fmla z31.s, p3/M, z5.s, z16.s\n"
274 "ld1w { z16.s }, p2/Z, [x21, x9, LSL #2]\n"
275 "fmla z28.s, p3/M, z6.s, z18.s\n"
276 "ld1w { z18.s }, p2/Z, [x20, x9, LSL #2]\n"
277 "ldr x21, [x16, #0x90]\n"
278 "fmla z30.s, p3/M, z1.s, z16.s\n"
279 "ldr x20, [x16, #0xa8]\n"
280 "fmla z31.s, p3/M, z2.s, z18.s\n"
281 "fmla z28.s, p3/M, z7.s, z16.s\n"
282 "ld1w { z16.s }, p2/Z, [x21, x9, LSL #2]\n"
283 "ld1w { z17.s }, p2/Z, [x20, x9, LSL #2]\n"
284 "ldr x21, [x16, #0xa0]\n"
285 "fmla z30.s, p3/M, z6.s, z16.s\n"
286 "fmla z31.s, p3/M, z3.s, z17.s\n"
287 "ldr x20, [x16, #0xb0]\n"
288 "ld1w { z16.s }, p2/Z, [x21, x9, LSL #2]\n"
289 "fmla z30.s, p3/M, z7.s, z16.s\n"
290 "fmla z29.s, p3/M, z7.s, z19.s\n"
291 "ld1w { z16.s }, p2/Z, [x20, x9, LSL #2]\n"
292 "ldr x20, [x16, #0xb8]\n"
293 "fmla z31.s, p3/M, z7.s, z16.s\n"
294 "fmla z30.s, p3/M, z5.s, z17.s\n"
295 "ld1w { z17.s }, p2/Z, [x20, x9, LSL #2]\n"
296 "ldr x20, [x16, #0xc0]\n"
297 "fmla z31.s, p3/M, z6.s, z17.s\n"
298 "fmla z29.s, p3/M, z8.s, z18.s\n"
299 "ld1w { z16.s }, p2/Z, [x20, x9, LSL #2]\n"
300 "fmla z30.s, p3/M, z8.s, z17.s\n"
301 "fmla z31.s, p3/M, z8.s, z16.s\n"
303 ".inst 0xc1b8cb3c // fclamp { z28.s-z31.s }, z25.s, z24.s\n"
304 "st1w { z28.s }, p0, [x14, x27, LSL #2]\n"
305 "st1w { z29.s }, p0, [x13, x27, LSL #2]\n"
306 "st1w { z30.s }, p0, [x11, x27, LSL #2]\n"
307 "st1w { z31.s }, p0, [x10, x27, LSL #2]\n"
308 ".inst 0xd503467f // SMSTOP\n"
310 : [n_channels]
"r" ((
unsigned long) n_channels), [offsetof_Args_inptrs]
"I" (offsetof(Args, inptrs)), [offsetof_args_max]
"I" (offsetof(Args, max)), [offsetof_args_min]
"I" (offsetof(Args, min)), [offsetof_args_outptrs]
"I" (offsetof(Args, outptrs)), [offsetof_args_params]
"I" (offsetof(Args, params)), [params_struct]
"r" (¶ms_struct)
311 :
"cc",
"memory",
"p0",
"p1",
"p2",
"p3",
"p4",
"p5",
"p6",
"p7",
"p8",
"p9",
"p10",
"p11",
"p12",
"p13",
"p14",
"p15",
"x9",
"x10",
"x11",
"x12",
"x13",
"x14",
"x15",
"x16",
"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"
318 #endif // defined(ARM_COMPUTE_ENABLE_SME2)