1// Auto-generated file. Do not edit! 2// Template: src/f32-igemm/4x2-aarch64-neonfma-cortex-a75.S.in 3// Generator: tools/xngen 4// 5// Copyright 2019 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_igemm_minmax_ukernel_4x2__aarch64_neonfma_prfm_cortex_a75( 13# size_t mr, x0 14# size_t nc, x1 15# size_t kc, x2 / x0 16# size_t ks, x3 / x9 17# const float**restrict a, x4 18# const float*restrict w, x5 19# float*restrict c, x6 20# size_t cm_stride, x7 21# size_t cn_stride, [sp] -> x10 22# size_t a_offset, [sp + 8] -> x11 23# const float* zero, [sp + 16] -> x12 24# const xnn_f32_minmax_params params [sp + 24] -> x8 25 26# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS. 27 28# Register usage 29# A0 x20 v0 v4 30# A1 x13 v1 v5 31# A2 x14 v2 v6 32# A3 x15 v3 v7 33 34# B x5 v16 v17 v18 v19 v20 v21 v22 v23 35 36# C0 x6 v24 v25 37# C1 x16 v26 v27 38# C2 x17 v28 v29 39# C3 x7 v30 v31 40 41# Clamp v4 v5 42 43 44BEGIN_FUNCTION xnn_f32_igemm_minmax_ukernel_4x2__aarch64_neonfma_prfm_cortex_a75 45 46 # Load cn_stride, a_offset 47 LDP x10, x11, [sp] 48 49 # Load zero, params pointer 50 LDP x12, x8, [sp, 16] 51 52 # Load min/max values 53 LD2R {v4.2s, v5.2s}, [x8] 54 55 # Save x20 on stack 56 STR x20, [sp, -16]! 57 58 # Clamp C pointers 59 CMP x0, 2 // if mr < 2 60 ADD x16, x6, x7 // c1 = c0 + cm_stride 61 CSEL x16, x6, x16, LO // c1 = c0 62 63 ADD x17, x16, x7 // c2 = c1 + cm_stride 64 // if mr <= 2 65 CSEL x17, x16, x17, LS // c2 = c1 66 67 CMP x0, 4 // if mr < 4 68 ADD x7, x17, x7 // c3 = c2 + cm_stride 69 CSEL x7, x17, x7, LO // c3 = c2 70 710: 72 # Load initial bias from w into accumulators 73 LDR d24, [x5], 8 74 MOV v26.8b, v24.8b 75 MOV v28.8b, v24.8b 76 MOV v30.8b, v24.8b 77 MOVI v25.2s, 0 78 PRFM PLDL1KEEP, [x5, 64] 79 MOVI v27.2s, 0 80 PRFM PLDL1KEEP, [x5, 128] 81 MOVI v29.2s, 0 82 PRFM PLDL1KEEP, [x5, 192] 83 MOVI v31.2s, 0 84 PRFM PLDL1KEEP, [x5, 256] 85 86 MOV x9, x3 // p = ks 87 881: 89 # Load next 4 A pointers 90 LDP x20, x13, [x4], 16 91 LDP x14, x15, [x4], 16 92 93 CMP x20, x12 // if a0 == zero 94 ADD x20, x20, x11 // a0 += a_offset 95 CSEL x20, x12, x20, EQ // a0 = zero, else += a0 + a_offset 96 CMP x13, x12 // if a1 == zero 97 ADD x13, x13, x11 // a1 += a_offset 98 CSEL x13, x12, x13, EQ // a1 = zero, else += a1 + a_offset 99 CMP x14, x12 // if a2 == zero 100 ADD x14, x14, x11 // a2 += a_offset 101 CSEL x14, x12, x14, EQ // a2 = zero, else += a2 + a_offset 102 CMP x15, x12 // if a3 == zero 103 ADD x15, x15, x11 // a3 += a_offset 104 CSEL x15, x12, x15, EQ // a3 = zero, else += a3 + a_offset 105 106 # Is there at least 8 floats (32 bytes) for prologue + epilogue? 107 SUBS x0, x2, 32 // k = kc - 32 108 B.LO 5f 109 110 # Prologue 111 # Read first block of 4 A and B. 112 LDR q0, [x20], 16 113 LDP d20, d21, [x5], 16 114 LDR q1, [x13], 16 115 LDR q2, [x14], 16 116 LDR q3, [x15], 16 117 LDP d22, d23, [x5], 16 118 119 # Is there at least 32. yes do main loop 120 SUBS x0, x0, 32 121 B.LO 3f 122 123 # Main loop - 8 floats of A (32 bytes) 1242: 125 # First block of 4. FMA for first 4, loads for 2nd block of 4. 126 FMLA v24.2s, v20.2s, v0.s[0] 127 LDR q4, [x20], 16 128 FMLA v26.2s, v20.2s, v1.s[0] 129 FMLA v28.2s, v20.2s, v2.s[0] 130 LDR d16, [x5, 0] 131 FMLA v30.2s, v20.2s, v3.s[0] 132 FMLA v25.2s, v21.2s, v0.s[1] 133 LDR q5, [x13], 16 134 FMLA v27.2s, v21.2s, v1.s[1] 135 FMLA v29.2s, v21.2s, v2.s[1] 136 LDR q6, [x14], 16 137 FMLA v31.2s, v21.2s, v3.s[1] 138 FMLA v24.2s, v22.2s, v0.s[2] 139 LDR q7, [x15], 16 140 FMLA v26.2s, v22.2s, v1.s[2] 141 FMLA v28.2s, v22.2s, v2.s[2] 142 LDR d17, [x5, 8] 143 FMLA v30.2s, v22.2s, v3.s[2] 144 FMLA v25.2s, v23.2s, v0.s[3] 145 LDR d18, [x5, 16] 146 FMLA v27.2s, v23.2s, v1.s[3] 147 FMLA v29.2s, v23.2s, v2.s[3] 148 LDR d19, [x5, 24] 149 FMLA v31.2s, v23.2s, v3.s[3] 150 PRFM PLDL1KEEP, [x5, 320] 151 152 # Second block of 4. FMA for second 4, loads for 1st block of 4. 153 FMLA v24.2s, v16.2s, v4.s[0] 154 LDR q0, [x20], 16 155 FMLA v26.2s, v16.2s, v5.s[0] 156 FMLA v28.2s, v16.2s, v6.s[0] 157 LDR d20, [x5, 32] 158 FMLA v30.2s, v16.2s, v7.s[0] 159 FMLA v25.2s, v17.2s, v4.s[1] 160 LDR q1, [x13], 16 161 FMLA v27.2s, v17.2s, v5.s[1] 162 FMLA v29.2s, v17.2s, v6.s[1] 163 LDR q2, [x14], 16 164 FMLA v31.2s, v17.2s, v7.s[1] 165 FMLA v24.2s, v18.2s, v4.s[2] 166 LDR q3, [x15], 16 167 FMLA v26.2s, v18.2s, v5.s[2] 168 FMLA v28.2s, v18.2s, v6.s[2] 169 LDR d21, [x5, 40] 170 FMLA v30.2s, v18.2s, v7.s[2] 171 SUBS x0, x0, 32 172 FMLA v25.2s, v19.2s, v4.s[3] 173 LDR d22, [x5, 48] 174 FMLA v27.2s, v19.2s, v5.s[3] 175 LDR d23, [x5, 56] 176 FMLA v29.2s, v19.2s, v6.s[3] 177 ADD x5, x5, 64 178 FMLA v31.2s, v19.2s, v7.s[3] 179 B.HS 2b 180 1813: 182 # Epilogue 183 # First block of 4. FMA for first 4, loads for 2nd block of 4. 184 FMLA v24.2s, v20.2s, v0.s[0] 185 LDR q4, [x20], 16 186 FMLA v26.2s, v20.2s, v1.s[0] 187 FMLA v28.2s, v20.2s, v2.s[0] 188 LDR d16, [x5, 0] 189 FMLA v30.2s, v20.2s, v3.s[0] 190 FMLA v25.2s, v21.2s, v0.s[1] 191 LDR q5, [x13], 16 192 FMLA v27.2s, v21.2s, v1.s[1] 193 FMLA v29.2s, v21.2s, v2.s[1] 194 LDR q6, [x14], 16 195 FMLA v31.2s, v21.2s, v3.s[1] 196 FMLA v24.2s, v22.2s, v0.s[2] 197 LDR q7, [x15], 16 198 FMLA v26.2s, v22.2s, v1.s[2] 199 FMLA v28.2s, v22.2s, v2.s[2] 200 LDR d17, [x5, 8] 201 FMLA v30.2s, v22.2s, v3.s[2] 202 FMLA v25.2s, v23.2s, v0.s[3] 203 LDR d18, [x5, 16] 204 FMLA v27.2s, v23.2s, v1.s[3] 205 FMLA v29.2s, v23.2s, v2.s[3] 206 LDR d19, [x5, 24] 207 FMLA v31.2s, v23.2s, v3.s[3] 208 PRFM PLDL1KEEP, [x5, 320] 209 210 # Second block of 4. FMA for second 4, no loads 211 FMLA v24.2s, v16.2s, v4.s[0] 212 FMLA v26.2s, v16.2s, v5.s[0] 213 FMLA v28.2s, v16.2s, v6.s[0] 214 FMLA v30.2s, v16.2s, v7.s[0] 215 FMLA v25.2s, v17.2s, v4.s[1] 216 FMLA v27.2s, v17.2s, v5.s[1] 217 FMLA v29.2s, v17.2s, v6.s[1] 218 FMLA v31.2s, v17.2s, v7.s[1] 219 FMLA v24.2s, v18.2s, v4.s[2] 220 FMLA v26.2s, v18.2s, v5.s[2] 221 FMLA v28.2s, v18.2s, v6.s[2] 222 ADDS x0, x0, 32 223 FMLA v30.2s, v18.2s, v7.s[2] 224 FMLA v25.2s, v19.2s, v4.s[3] 225 ADD x5, x5, 32 226 FMLA v27.2s, v19.2s, v5.s[3] 227 FMLA v29.2s, v19.2s, v6.s[3] 228 LD2R {v4.2s, v5.2s}, [x8] // Load min/max values 229 FMLA v31.2s, v19.2s, v7.s[3] 230 231 # Is there a remainder? up to 8 floats (32 bytes) 232 B.NE 5f 233 2344: 235 # ks loop 236 SUBS x9, x9, 32 // ks -= MR * sizeof(void*) 237 B.HI 1b 238 239 FADD v24.2s, v24.2s, v25.2s 240 FADD v26.2s, v26.2s, v27.2s 241 FADD v28.2s, v28.2s, v29.2s 242 FADD v30.2s, v30.2s, v31.2s 243 244 # Clamp 245 FMAX v24.2s, v24.2s, v4.2s 246 FMAX v26.2s, v26.2s, v4.2s 247 FMAX v28.2s, v28.2s, v4.2s 248 FMAX v30.2s, v30.2s, v4.2s 249 SUBS x1, x1, 2 250 FMIN v24.2s, v24.2s, v5.2s 251 FMIN v26.2s, v26.2s, v5.2s 252 FMIN v28.2s, v28.2s, v5.2s 253 FMIN v30.2s, v30.2s, v5.2s 254 255 # Store full 4 x 2 256 B.LO 8f 257 258 STR d30, [x7] 259 ADD x7, x7, x10 260 STR d28, [x17] 261 ADD x17, x17, x10 262 STR d26, [x16] 263 ADD x16, x16, x10 264 STR d24, [x6] 265 ADD x6, x6, x10 266 SUB x4, x4, x3 // a -= ks 267 268 # nc loop 269 B.HI 0b 270 271 # Restore x20 from stack 272 LDR x20, [sp], 16 273 RET 274 2755: 276 # Remainder- 4 floats of A (16 bytes) 277 TBZ x0, 4, 6f 278 279 LDR q0, [x20], 16 280 LDP d20, d21, [x5], 16 281 LDR q1, [x13], 16 282 LDR q2, [x14], 16 283 LDR q3, [x15], 16 284 LDP d22, d23, [x5], 16 285 FMLA v24.2s, v20.2s, v0.s[0] 286 FMLA v26.2s, v20.2s, v1.s[0] 287 FMLA v28.2s, v20.2s, v2.s[0] 288 FMLA v30.2s, v20.2s, v3.s[0] 289 FMLA v25.2s, v21.2s, v0.s[1] 290 FMLA v27.2s, v21.2s, v1.s[1] 291 FMLA v29.2s, v21.2s, v2.s[1] 292 FMLA v31.2s, v21.2s, v3.s[1] 293 FMLA v24.2s, v22.2s, v0.s[2] 294 FMLA v26.2s, v22.2s, v1.s[2] 295 FMLA v28.2s, v22.2s, v2.s[2] 296 FMLA v30.2s, v22.2s, v3.s[2] 297 FMLA v25.2s, v23.2s, v0.s[3] 298 FMLA v27.2s, v23.2s, v1.s[3] 299 FMLA v29.2s, v23.2s, v2.s[3] 300 FMLA v31.2s, v23.2s, v3.s[3] 301 3026: 303 # Remainder- 2 floats of A (8 bytes) 304 TBZ x0, 3, 7f 305 306 LDR d0, [x20], 8 307 LDP d20, d21, [x5], 16 308 LDR d1, [x13], 8 309 LDR d2, [x14], 8 310 LDR d3, [x15], 8 311 FMLA v24.2s, v20.2s, v0.s[0] 312 FMLA v26.2s, v20.2s, v1.s[0] 313 FMLA v28.2s, v20.2s, v2.s[0] 314 FMLA v30.2s, v20.2s, v3.s[0] 315 FMLA v25.2s, v21.2s, v0.s[1] 316 FMLA v27.2s, v21.2s, v1.s[1] 317 FMLA v29.2s, v21.2s, v2.s[1] 318 FMLA v31.2s, v21.2s, v3.s[1] 319 3207: 321 # Remainder- 1 float of A (4 bytes) 322 TBZ x0, 2, 4b 323 324 LDR s0, [x20], 4 325 LDR d20, [x5], 8 326 LDR s1, [x13], 4 327 LDR s2, [x14], 4 328 LDR s3, [x15], 4 329 FMLA v24.2s, v20.2s, v0.s[0] 330 FMLA v26.2s, v20.2s, v1.s[0] 331 FMLA v28.2s, v20.2s, v2.s[0] 332 FMLA v30.2s, v20.2s, v3.s[0] 333 B 4b 334 335 # Store odd width 3368: 337 STR s30, [x7] 338 STR s28, [x17] 339 STR s26, [x16] 340 STR s24, [x6] 341 342 # Restore x20 from stack 343 LDR x20, [sp], 16 344 RET 345 346END_FUNCTION xnn_f32_igemm_minmax_ukernel_4x2__aarch64_neonfma_prfm_cortex_a75 347 348#ifdef __ELF__ 349.section ".note.GNU-stack","",%progbits 350#endif 351