Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 1 | // Copyright 2019 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #include <xnnpack/assembly.h> |
| 7 | |
| 8 | .syntax unified |
| 9 | |
Frank Barchard | 569561d | 2020-06-17 13:11:12 -0700 | [diff] [blame] | 10 | // void xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_${"pld_" if PREFETCH else ""}ld64( |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 11 | // size_t mr, r0 |
| 12 | // size_t nc, r1 |
Frank Barchard | fd262e1 | 2020-02-10 23:45:48 -0800 | [diff] [blame] | 13 | // size_t kc, r2 -> r5 -> sp + 68 |
| 14 | // size_t ks, r3 -> sp + 72 -> r14 |
| 15 | // const float**restrict a, sp + 112 -> r2 |
| 16 | // const void*restrict w, sp + 116 -> r9 |
| 17 | // uint8_t*restrict c, sp + 120 -> r11 |
| 18 | // size_t cm_stride, sp + 124 -> (r6) |
| 19 | // size_t cn_stride, sp + 128 -> (r7) |
| 20 | // size_t a_offset, sp + 132 -> (r5) |
| 21 | // const float* zero, sp + 136 -> (r7) |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 22 | // minmax_params*params, sp + 140 -> (r5) |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 23 | |
| 24 | // inner loop registers |
| 25 | |
| 26 | // A0 r3 d0 |
| 27 | // A1 r12 d1 |
| 28 | // A2 r10 d2 |
| 29 | // A3 r0 d3 |
| 30 | |
| 31 | // B r9 d8, d9, d10, d11 |
| 32 | // B d12, d13, d14, d15 |
| 33 | |
| 34 | // C0 r11 d16-d17 q8 d18-d19 q9 |
| 35 | // C1 r4 d20-d21 q10 d22-d23 q11 |
| 36 | // C2 r8 d24-d25 q12 d26-d27 q13 |
| 37 | // C3 r6 d28-d29 q14 d30-d31 q15 |
| 38 | |
| 39 | // Clamp (r5) d4 d5 d6 d7 |
| 40 | |
Frank Barchard | 569561d | 2020-06-17 13:11:12 -0700 | [diff] [blame] | 41 | BEGIN_FUNCTION xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_${"pld_" if PREFETCH else ""}ld64 |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 42 | .arm |
| 43 | #ifndef __APPLE__ |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 44 | .arch armv7-a |
| 45 | .fpu neon |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 46 | #endif |
Frank Barchard | 7c9f1f9 | 2021-06-04 14:38:55 -0700 | [diff] [blame^] | 47 | # Push 112 bytes |
| 48 | # r2 will be reloaded in outer loop. r3 is ks |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 49 | PUSH {r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r14} // +44 |
| 50 | SUB sp, sp, 4 // 4 |
| 51 | VPUSH {d8-d15} // +64 = 112 |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 52 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 53 | LDR r11, [sp, 120] // c |
| 54 | LDR r6, [sp, 124] // cm_stride |
| 55 | LDR r2, [sp, 112] // a |
| 56 | LDR r9, [sp, 116] // w |
| 57 | LDR r5, [sp, 140] // params |
| 58 | MOV r14, r3 // p = ks |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 59 | |
Frank Barchard | 7c9f1f9 | 2021-06-04 14:38:55 -0700 | [diff] [blame^] | 60 | # Clamp C pointers |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 61 | CMP r0, 2 // if mr >= 2 |
| 62 | ADD r4, r11, r6 // c1 = c0 + cm_stride |
| 63 | MOVLO r4, r11 // c1 |
Frank Barchard | 7c9f1f9 | 2021-06-04 14:38:55 -0700 | [diff] [blame^] | 64 | // if mr > 2 |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 65 | ADD r8, r4, r6 // c2 = c1 + cm_stride |
| 66 | MOVLS r8, r4 // c2 |
| 67 | CMP r0, 4 // if mr >=4 |
| 68 | ADD r6, r8, r6 // c3 = c2 + cm_stride |
| 69 | MOVLO r6, r8 // c3 |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 70 | |
Frank Barchard | 7c9f1f9 | 2021-06-04 14:38:55 -0700 | [diff] [blame^] | 71 | # Load min/max values |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 72 | VLD1.32 {d4[], d5[]}, [r5]! |
| 73 | VLD1.32 {d6[], d7[]}, [r5] |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 74 | |
| 75 | 0: |
| 76 | # Load initial bias from w into accumulators |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 77 | VLDM r9!, {d16-d19} // Bias |
| 78 | VMOV q10, q8 |
| 79 | VMOV q11, q9 |
| 80 | VMOV q12, q8 |
| 81 | VMOV q13, q9 |
| 82 | VMOV q14, q8 |
| 83 | VMOV q15, q9 |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 84 | |
Frank Barchard | 569561d | 2020-06-17 13:11:12 -0700 | [diff] [blame] | 85 | $if PREFETCH: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 86 | PLD [r9, 0] // Prefetch B |
| 87 | PLD [r9, 64] |
| 88 | PLD [r9, 128] |
| 89 | PLD [r9, 192] |
| 90 | PLD [r9, 256] |
| 91 | PLD [r9, 320] |
| 92 | PLD [r9, 384] |
| 93 | PLD [r9, 448] |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 94 | 1: |
| 95 | # Load next 4 A pointers |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 96 | LDR r3, [r2, 0] |
| 97 | LDR r12, [r2, 4] |
| 98 | LDR r10, [r2, 8] |
| 99 | LDR r0, [r2, 12] |
| 100 | ADD r2, r2, 16 |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 101 | |
Frank Barchard | 7c9f1f9 | 2021-06-04 14:38:55 -0700 | [diff] [blame^] | 102 | # Add a_offset |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 103 | LDR r5, [sp, 132] // a_offset |
| 104 | LDR r7, [sp, 136] // zero |
| 105 | CMP r3, r7 // if a0 == zero |
| 106 | ADD r3, r3, r5 // a0 += a_offset |
| 107 | MOVEQ r3, r7 // a0 = zero, else += a0 + a_offset |
| 108 | CMP r12, r7 // if a1 == zero |
| 109 | ADD r12, r12, r5 // a1 += a_offset |
| 110 | MOVEQ r12, r7 // a1 = zero, else += a1 + a_offset |
| 111 | CMP r10, r7 // if a2 == zero |
| 112 | ADD r10, r10, r5 // a2 += a_offset |
| 113 | MOVEQ r10, r7 // a2 = zero, else += a2 + a_offset |
| 114 | CMP r0, r7 // if a3 == zero |
| 115 | ADD r0, r0, r5 // a3 += a_offset |
| 116 | LDR r5, [sp, 68] // kc |
| 117 | MOVEQ r0, r7 // a3 = zero, else += a3 + a_offset |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 118 | |
Frank Barchard | 569561d | 2020-06-17 13:11:12 -0700 | [diff] [blame] | 119 | $if PREFETCH: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 120 | PLD [r3, 0] // Prefetch A |
| 121 | PLD [r3, 64] |
| 122 | PLD [r12, 0] |
| 123 | PLD [r12, 64] |
| 124 | PLD [r10, 0] |
| 125 | PLD [r10, 64] |
| 126 | PLD [r0, 0] |
| 127 | PLD [r0, 64] |
Frank Barchard | 569561d | 2020-06-17 13:11:12 -0700 | [diff] [blame] | 128 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 129 | SUBS r5, r5, 8 // kc - 8 |
| 130 | BLO 4f // less than 2 channels? |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 131 | |
Frank Barchard | 7c9f1f9 | 2021-06-04 14:38:55 -0700 | [diff] [blame^] | 132 | # Main loop - 2 floats of A (8 bytes) |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 133 | 2: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 134 | VLD1.32 {d0}, [r3]! // A0 |
| 135 | VLDM r9!, {d8-d11} // B0 |
| 136 | VLD1.32 {d1}, [r12]! // A1 |
| 137 | VLD1.32 {d2}, [r10]! // A2 |
| 138 | VLD1.32 {d3}, [ r0]! // A3 |
| 139 | VLDM r9!, {d12-d15} // B1 |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 140 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 141 | VMLA.F32 q8, q4, d0[0] |
| 142 | VMLA.F32 q9, q5, d0[0] |
Frank Barchard | 569561d | 2020-06-17 13:11:12 -0700 | [diff] [blame] | 143 | $if PREFETCH: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 144 | PLD [r3, 128] // Prefetch A0 |
| 145 | VMLA.F32 q10, q4, d1[0] |
| 146 | VMLA.F32 q11, q5, d1[0] |
Frank Barchard | efc1014 | 2020-07-08 11:46:57 -0700 | [diff] [blame] | 147 | $if PREFETCH: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 148 | PLD [r12, 128] // Prefetch A1 |
| 149 | VMLA.F32 q12, q4, d2[0] |
| 150 | VMLA.F32 q13, q5, d2[0] |
Frank Barchard | 569561d | 2020-06-17 13:11:12 -0700 | [diff] [blame] | 151 | $if PREFETCH: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 152 | PLD [r10, 128] // Prefetch A2 |
| 153 | VMLA.F32 q14, q4, d3[0] |
| 154 | VMLA.F32 q15, q5, d3[0] |
Frank Barchard | 569561d | 2020-06-17 13:11:12 -0700 | [diff] [blame] | 155 | $if PREFETCH: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 156 | PLD [r0, 128] // Prefetch A3 |
| 157 | VMLA.F32 q8, q6, d0[1] |
| 158 | VMLA.F32 q9, q7, d0[1] |
Frank Barchard | 569561d | 2020-06-17 13:11:12 -0700 | [diff] [blame] | 159 | $if PREFETCH: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 160 | PLD [r9, 448] // Prefetch B |
| 161 | VMLA.F32 q10, q6, d1[1] |
| 162 | VMLA.F32 q11, q7, d1[1] |
| 163 | SUBS r5, r5, 8 |
| 164 | VMLA.F32 q12, q6, d2[1] |
| 165 | VMLA.F32 q13, q7, d2[1] |
| 166 | VMLA.F32 q14, q6, d3[1] |
| 167 | VMLA.F32 q15, q7, d3[1] |
| 168 | BHS 2b |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 169 | |
Frank Barchard | 7c9f1f9 | 2021-06-04 14:38:55 -0700 | [diff] [blame^] | 170 | # Is there a remainder?- 1 floats of A (4 bytes) |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 171 | TST r5, 4 |
| 172 | BNE 4f |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 173 | |
Frank Barchard | 3b26206 | 2020-09-30 15:53:17 -0700 | [diff] [blame] | 174 | 3: |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 175 | # ks loop |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 176 | SUBS r14, r14, 16 // ks -= MR * sizeof(void*) |
| 177 | BHI 1b |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 178 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 179 | LDR r7, [sp, 128] // cn_stride |
| 180 | LDR r14, [sp, 72] // p = ks |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 181 | |
Frank Barchard | 7c9f1f9 | 2021-06-04 14:38:55 -0700 | [diff] [blame^] | 182 | # Clamp |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 183 | VMAX.F32 q8, q8, q2 |
| 184 | SUBS r1, r1, 8 |
| 185 | VMAX.F32 q9, q9, q2 |
| 186 | VMAX.F32 q10, q10, q2 |
| 187 | VMAX.F32 q11, q11, q2 |
| 188 | VMAX.F32 q12, q12, q2 |
| 189 | VMAX.F32 q13, q13, q2 |
| 190 | VMAX.F32 q14, q14, q2 |
| 191 | VMAX.F32 q15, q15, q2 |
| 192 | VMIN.F32 q8, q8, q3 |
| 193 | VMIN.F32 q9, q9, q3 |
| 194 | VMIN.F32 q10, q10, q3 |
| 195 | VMIN.F32 q11, q11, q3 |
| 196 | VMIN.F32 q12, q12, q3 |
| 197 | VMIN.F32 q13, q13, q3 |
| 198 | VMIN.F32 q14, q14, q3 |
| 199 | VMIN.F32 q15, q15, q3 |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 200 | |
Frank Barchard | 7c9f1f9 | 2021-06-04 14:38:55 -0700 | [diff] [blame^] | 201 | # Store full 4 x 8 |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 202 | BLO 5f |
| 203 | VST1.32 {d28-d31}, [r6], r7 |
| 204 | VST1.32 {d24-d27}, [r8], r7 |
| 205 | VST1.32 {d20-d23}, [r4], r7 |
| 206 | VST1.32 {d16-d19}, [r11], r7 |
| 207 | SUB r2, r2, r14 // a -= ks |
| 208 | BHI 0b |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 209 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 210 | VPOP {d8-d15} |
| 211 | ADD sp, sp, 12 // skip pad, r2, r3 |
| 212 | POP {r4, r5, r6, r7, r8, r9, r10, r11, pc} |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 213 | |
Frank Barchard | 3b26206 | 2020-09-30 15:53:17 -0700 | [diff] [blame] | 214 | 4: |
Frank Barchard | 7c9f1f9 | 2021-06-04 14:38:55 -0700 | [diff] [blame^] | 215 | # Remainder- 1 floats of A (4 bytes) |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 216 | VLDM r3!, {s0} // A0 |
| 217 | VLDM r9!, {d8-d11} // B0 |
| 218 | VLDM r12!, {s2} // A1 |
| 219 | VLDM r10!, {s4} // A2 |
| 220 | VLDM r0!, {s6} // A3 |
| 221 | VMLA.F32 q8, q4, d0[0] |
| 222 | VMLA.F32 q9, q5, d0[0] |
| 223 | VMLA.F32 q10, q4, d1[0] |
| 224 | VMLA.F32 q11, q5, d1[0] |
| 225 | VMLA.F32 q12, q4, d2[0] |
| 226 | VMLA.F32 q13, q5, d2[0] |
| 227 | VMLA.F32 q14, q4, d3[0] |
| 228 | VMLA.F32 q15, q5, d3[0] |
| 229 | B 3b |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 230 | |
Frank Barchard | 7c9f1f9 | 2021-06-04 14:38:55 -0700 | [diff] [blame^] | 231 | # Store odd width |
Frank Barchard | 3b26206 | 2020-09-30 15:53:17 -0700 | [diff] [blame] | 232 | 5: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 233 | TST r1, 4 |
| 234 | BEQ 6f |
| 235 | VST1.32 {d28-d29}, [r6]! |
| 236 | VMOV q14, q15 |
| 237 | VST1.32 {d24-d25}, [r8]! |
| 238 | VMOV q12, q13 |
| 239 | VST1.32 {d20-d21}, [r4]! |
| 240 | VMOV q10, q11 |
| 241 | VST1.32 {d16-d17}, [r11]! |
| 242 | VMOV q8, q9 |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 243 | |
Frank Barchard | 3b26206 | 2020-09-30 15:53:17 -0700 | [diff] [blame] | 244 | 6: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 245 | TST r1, 2 |
| 246 | BEQ 7f |
| 247 | VST1.32 {d28}, [r6]! |
| 248 | VMOV d28, d29 |
| 249 | VST1.32 {d24}, [r8]! |
| 250 | VMOV d24, d25 |
| 251 | VST1.32 {d20}, [r4]! |
| 252 | VMOV d20, d21 |
| 253 | VST1.32 {d16}, [r11]! |
| 254 | VMOV d16, d17 |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 255 | |
Frank Barchard | 3b26206 | 2020-09-30 15:53:17 -0700 | [diff] [blame] | 256 | 7: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 257 | TST r1, 1 |
| 258 | BEQ 8f |
| 259 | VST1.32 {d28[0]}, [r6]! |
| 260 | VST1.32 {d24[0]}, [r8]! |
| 261 | VST1.32 {d20[0]}, [r4]! |
| 262 | VST1.32 {d16[0]}, [r11]! |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 263 | |
Frank Barchard | 3b26206 | 2020-09-30 15:53:17 -0700 | [diff] [blame] | 264 | 8: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 265 | VPOP {d8-d15} |
| 266 | ADD sp, sp, 12 // skip pad, r2, r3 |
| 267 | POP {r4, r5, r6, r7, r8, r9, r10, r11, pc} |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 268 | |
Frank Barchard | 569561d | 2020-06-17 13:11:12 -0700 | [diff] [blame] | 269 | END_FUNCTION xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_${"pld_" if PREFETCH else ""}ld64 |
Frank Barchard | dc38f07 | 2020-02-10 13:21:42 -0800 | [diff] [blame] | 270 | |
| 271 | #ifdef __ELF__ |
| 272 | .section ".note.GNU-stack","",%progbits |
| 273 | #endif |