Frank Barchard | 1391604 | 2019-12-11 10:56:34 -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 | |
| 10 | // void xnn_f32_gemm_ukernel_4x8__aarch32_neon_cortex_a53( |
| 11 | // size_t mr, r0 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 12 | // size_t nc, r1 |
| 13 | // size_t kc, r2 -> r5 -> sp + 0 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 14 | // const uint8_t*restrict a, r3 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 15 | // size_t a_stride, sp + 100 -> (r7) |
| 16 | // const void*restrict w, sp + 104 -> r9 |
| 17 | // uint8_t*restrict c, sp + 108 -> r11 |
| 18 | // size_t cm_stride, sp + 112 -> (r6) |
| 19 | // size_t cn_stride, sp + 116 -> (r0) |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 20 | // const union xnn_f32_minmax_params params[restrict static 1]) sp + 120 -> (r5) |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 21 | |
| 22 | |
| 23 | // inner loop registers |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 24 | // r0, r2 scratch temporaries for loads |
| 25 | // r14 (lr) unused |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 26 | |
| 27 | // A0 r3 d0 |
| 28 | // A1 r12 d1 |
| 29 | // A2 r10 d2 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 30 | // A3 r7 d3 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 31 | |
| 32 | // B r9 d8, d9, d10, d11 |
| 33 | // B d12, d13, d14, d15 |
| 34 | |
| 35 | // C0 r11 d16-d17 q8 d18-d19 q9 |
| 36 | // C1 r4 d20-d21 q10 d22-d23 q11 |
| 37 | // C2 r8 d24-d25 q12 d26-d27 q13 |
| 38 | // C3 r6 d28-d29 q14 d30-d31 q15 |
| 39 | |
| 40 | // Clamp (r5) d4 d5 d6 d7 |
| 41 | |
| 42 | BEGIN_FUNCTION xnn_f32_gemm_ukernel_4x8__aarch32_neon_cortex_a53 |
| 43 | .arm |
| 44 | #ifndef __APPLE__ |
| 45 | .arch armv7-a |
| 46 | .fpu neon |
| 47 | #endif |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 48 | // Push 100 bytes |
| 49 | // r2 will be reloaded in outer loop |
| 50 | VPUSH {d8-d15} // 64 |
| 51 | PUSH {r2, r4, r5, r6, r7, r8, r9, r10, r11} // +36 = 100 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 52 | |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 53 | LDR r7, [sp, 100] // a_stride |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 54 | LDR r11, [sp, 108] // c |
| 55 | LDR r6, [sp, 112] // cm_stride |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 56 | LDR r9, [sp, 104] // w |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 57 | |
| 58 | // Clamp A and C pointers |
| 59 | CMP r0, 2 // if mr >= 2 |
| 60 | ADD r12, r3, r7 // a1 = a0 + a_stride |
| 61 | ADD r4, r11, r6 // c1 = c0 + cm_stride |
| 62 | MOVLO r12, r3 // a1 |
| 63 | MOVLO r4, r11 // c1 |
| 64 | // if mr > 2 |
| 65 | ADD r10, r12, r7 // a2 = a1 + a_stride |
| 66 | ADD r8, r4, r6 // c2 = c1 + cm_stride |
| 67 | MOVLS r10, r12 // a2 |
| 68 | MOVLS r8, r4 // c2 |
| 69 | |
| 70 | CMP r0, 4 // if mr >=4 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 71 | ADD r7, r10, r7 // a3 = a2 + a_stride |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 72 | ADD r6, r8, r6 // c3 = c2 + cm_stride |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 73 | MOVLO r7, r10 // a3 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 74 | MOVLO r6, r8 // c3 |
| 75 | |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 76 | .p2align 3 |
| 77 | 1: |
| 78 | # Load initial bias from w into accumulators |
| 79 | VLDM r9!, {d16-d19} // Bias |
Frank Barchard | 005feb8 | 2019-12-16 18:16:55 -0800 | [diff] [blame] | 80 | |
| 81 | SUBS r5, r2, 16 // kc - 16 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 82 | PLD [r3, 0] // Prefetch A |
| 83 | PLD [r3, 64] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 84 | VMOV q10, q8 |
| 85 | PLD [r12, 0] |
| 86 | PLD [r12, 64] |
| 87 | VMOV q11, q9 |
| 88 | PLD [r10, 0] |
| 89 | PLD [r10, 64] |
| 90 | VMOV q12, q8 |
| 91 | PLD [r7, 0] |
| 92 | PLD [r7, 64] |
| 93 | VMOV q13, q9 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 94 | PLD [r9, 0] // Prefetch B |
| 95 | PLD [r9, 64] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 96 | VMOV q14, q8 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 97 | PLD [r9, 128] |
| 98 | PLD [r9, 192] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 99 | VMOV q15, q9 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 100 | PLD [r9, 256] |
| 101 | PLD [r9, 320] |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 102 | BLO 5f // less than 4 channels? |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 103 | |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 104 | // Prologue |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 105 | VLD1.32 {d0}, [r3]! // A0 |
| 106 | VLD1.32 {d1}, [r12]! // A1 |
| 107 | VLD1.32 {d2}, [r10]! // A2 |
| 108 | VLD1.32 {d3}, [r7]! // A3 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 109 | SUBS r5, r5, 16 |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 110 | VLDM r9, {d8-d11} // B0 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 111 | LDR r0, [r9, 56] // B1 low VMOV is in BLOCK 0 |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 112 | LDR r2, [r9, 60] // B1 high |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 113 | VLDR d13, [r9, 40] // B1 |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 114 | |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 115 | BLO 3f // less than 4 channels? skip main loop |
| 116 | |
Frank Barchard | 005feb8 | 2019-12-16 18:16:55 -0800 | [diff] [blame] | 117 | # Main loop - 4 floats of A (16 bytes) |
| 118 | # 32 FMA + 8 LD64 A + 8 LDR B |
| 119 | .p2align 3 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 120 | 2: |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 121 | # First group of 16 FMA, Second group loads |
| 122 | // BLOCK 0 |
Frank Barchard | 005feb8 | 2019-12-16 18:16:55 -0800 | [diff] [blame] | 123 | VLD1.32 {d4}, [r3]! // A0 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 124 | VMOV d15, r0, r2 // b1 VMOV b from second group |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 125 | VMLA.F32 q8, q4, d0[0] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 126 | LDR r0, [r12] // A1 low |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 127 | VMLA.F32 q10, q4, d1[0] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 128 | LDR r2, [r12, 4] // A1 high |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 129 | VMLA.F32 q12, q4, d2[0] |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 130 | PLD [r3, 128] // Prefetch A0 |
Frank Barchard | 005feb8 | 2019-12-16 18:16:55 -0800 | [diff] [blame] | 131 | |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 132 | // BLOCK 1 |
| 133 | VLDR d12, [r9, 32] // B1 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 134 | VMOV d5, r0, r2 // a1 VMOV |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 135 | VMLA.F32 q14, q4, d3[0] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 136 | LDR r0, [r9, 72] // B0 low |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 137 | VMLA.F32 q9, q5, d0[0] |
| 138 | LDR r2, [r9, 76] // B0 high |
| 139 | VMLA.F32 q11, q5, d1[0] |
| 140 | PLD [r12, 128] // Prefetch A1 |
| 141 | |
| 142 | // BLOCK 2 |
| 143 | VLD1.32 {d6}, [r10]! // A2 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 144 | VMOV d9, r0, r2 // b0 VMOV |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 145 | VMLA.F32 q13, q5, d2[0] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 146 | LDR r0, [r7] // A3 low |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 147 | VMLA.F32 q15, q5, d3[0] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 148 | LDR r2, [r7, 4] // A3 high |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 149 | VMLA.F32 q8, q6, d0[1] |
| 150 | PLD [r10, 128] // Prefetch A2 |
| 151 | |
| 152 | // BLOCK 3 |
| 153 | VLDR d14, [r9, 48] // B1 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 154 | VMOV d7, r0, r2 // a3 VMOV |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 155 | VMLA.F32 q10, q6, d1[1] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 156 | LDR r0, [r9, 88] // B0 low |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 157 | VMLA.F32 q12, q6, d2[1] |
| 158 | LDR r2, [r9, 92] // B0 high |
| 159 | VMLA.F32 q14, q6, d3[1] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 160 | PLD [r7, 128] // Prefetch A3 |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 161 | |
| 162 | // BLOCK 4 |
| 163 | VLDR d8, [r9, 64] // B0 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 164 | VMOV d11, r0, r2 // B0 VMOV |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 165 | VMLA.F32 q9, q7, d0[1] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 166 | LDR r0, [r9, 104] // B1 low VMOV is in BLOCK 0 |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 167 | VMLA.F32 q11, q7, d1[1] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 168 | LDR r2, [r9, 108] // B1 high |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 169 | VMLA.F32 q13, q7, d2[1] |
| 170 | PLD [r9, 384] // Prefetch B |
| 171 | |
| 172 | // BLOCK 5 |
| 173 | VLDR d10, [r9, 80] // B0 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 174 | VMOV d13, r0, r2 // b1 VMOV b from second group |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 175 | VMLA.F32 q15, q7, d3[1] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 176 | LDR r0, [r9, 120] // B1 low VMOV is in BLOCK 0 |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 177 | NOP |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 178 | LDR r2, [r9, 124] // B1 high |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 179 | NOP |
| 180 | PLD [r9, 448] // Prefetch B |
| 181 | |
| 182 | # Second group of 16 FMA, First group of loads |
| 183 | // BLOCK 0 |
| 184 | VLD1.32 {d0}, [r3]! // A0 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 185 | VMOV d15, r0, r2 // b1 VMOV b from second group |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 186 | VMLA.F32 q8, q4, d4[0] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 187 | LDR r0, [r12, 8] // A1 low |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 188 | VMLA.F32 q10, q4, d5[0] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 189 | LDR r2, [r12, 12] // A1 high |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 190 | VMLA.F32 q12, q4, d6[0] |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 191 | // NOP |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 192 | |
| 193 | // BLOCK 1 |
| 194 | VLDR d12, [r9, 96] // B1 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 195 | VMOV d1, r0, r2 // a1 VMOV |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 196 | VMLA.F32 q14, q4, d7[0] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 197 | LDR r0, [r9, 136] // B0 low |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 198 | VMLA.F32 q9, q5, d4[0] |
| 199 | LDR r2, [r9, 140] // B0 high |
| 200 | VMLA.F32 q11, q5, d5[0] |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 201 | // NOP |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 202 | |
| 203 | // BLOCK 2 |
| 204 | VLD1.32 {d2}, [r10]! // A2 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 205 | VMOV d9, r0, r2 // b0 VMOV |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 206 | VMLA.F32 q13, q5, d6[0] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 207 | LDR r0, [r7, 8] // A3 low |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 208 | VMLA.F32 q15, q5, d7[0] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 209 | LDR r2, [r7, 12] // A3 high |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 210 | VMLA.F32 q8, q6, d4[1] |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 211 | // NOP |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 212 | |
| 213 | // BLOCK 3 |
| 214 | VLDR d14, [r9, 112] // B1 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 215 | VMOV d3, r0, r2 // a3 VMOV |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 216 | VMLA.F32 q10, q6, d5[1] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 217 | LDR r0, [r9, 152] // B0 low |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 218 | VMLA.F32 q12, q6, d6[1] |
| 219 | LDR r2, [r9, 156] // B0 high |
| 220 | VMLA.F32 q14, q6, d7[1] |
| 221 | ADD r12, r12, 16 // A1++ |
| 222 | |
| 223 | // BLOCK 4 |
| 224 | VLDR d8, [r9, 128] // B0 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 225 | VMOV d11, r0, r2 // B0 VMOV |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 226 | VMLA.F32 q9, q7, d4[1] |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 227 | LDR r0, [r9, 168] // B1 low |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 228 | VMLA.F32 q11, q7, d5[1] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 229 | LDR r2, [r9, 172] // B1 high |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 230 | VMLA.F32 q13, q7, d6[1] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 231 | ADD r7, r7, 16 // A3++ |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 232 | |
| 233 | // BLOCK 5 |
| 234 | VLDR d10, [r9, 144] // B0 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 235 | VMOV d13, r0, r2 // b1 VMOV b |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 236 | VMLA.F32 q15, q7, d7[1] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 237 | LDR r0, [r9, 184] // B1 low VMOV is in BLOCK 0 |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 238 | SUBS r5, r5, 16 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 239 | LDR r2, [r9, 188] // B1 high |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 240 | ADD r9, r9, 128 // B++ |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 241 | BHS 2b |
| 242 | |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 243 | # Epilogue - 4 floats of A (16 bytes) |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 244 | 3: |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 245 | # First group of 16 FMA, Second group loads |
| 246 | // BLOCK 0 |
Frank Barchard | 279908a | 2019-12-30 17:07:46 -0800 | [diff] [blame] | 247 | VLD1.32 {d4}, [r3]! // A0 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 248 | VMOV d15, r0, r2 // b1 VMOV b from second group |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 249 | VMLA.F32 q8, q4, d0[0] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 250 | LDR r0, [r12] // A1 low |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 251 | VMLA.F32 q10, q4, d1[0] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 252 | LDR r2, [r12, 4] // A1 high |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 253 | VMLA.F32 q12, q4, d2[0] |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 254 | // NOP |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 255 | |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 256 | // BLOCK 1 |
| 257 | VLDR d12, [r9, 32] // B1 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 258 | VMOV d5, r0, r2 // a1 VMOV |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 259 | VMLA.F32 q14, q4, d3[0] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 260 | LDR r0, [r9, 72] // B0 low |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 261 | VMLA.F32 q9, q5, d0[0] |
| 262 | LDR r2, [r9, 76] // B0 high |
| 263 | VMLA.F32 q11, q5, d1[0] |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 264 | // NOP |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 265 | |
| 266 | // BLOCK 2 |
| 267 | VLD1.32 {d6}, [r10]! // A2 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 268 | VMOV d9, r0, r2 // b0 VMOV |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 269 | VMLA.F32 q13, q5, d2[0] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 270 | LDR r0, [r7] // A3 low |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 271 | VMLA.F32 q15, q5, d3[0] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 272 | LDR r2, [r7, 4] // A3 high |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 273 | VMLA.F32 q8, q6, d0[1] |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 274 | // NOP |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 275 | |
| 276 | // BLOCK 3 |
| 277 | VLDR d14, [r9, 48] // B1 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 278 | VMOV d7, r0, r2 // a3 VMOV |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 279 | VMLA.F32 q10, q6, d1[1] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 280 | LDR r0, [r9, 88] // B0 low |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 281 | VMLA.F32 q12, q6, d2[1] |
| 282 | LDR r2, [r9, 92] // B0 high |
| 283 | VMLA.F32 q14, q6, d3[1] |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 284 | // NOP |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 285 | |
| 286 | // BLOCK 4 |
| 287 | VLDR d8, [r9, 64] // B0 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 288 | VMOV d11, r0, r2 // B0 VMOV |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 289 | VMLA.F32 q9, q7, d0[1] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 290 | LDR r0, [r9, 104] // B1 low |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 291 | VMLA.F32 q11, q7, d1[1] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 292 | LDR r2, [r9, 108] // B1 high |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 293 | VMLA.F32 q13, q7, d2[1] |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 294 | // NOP |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 295 | |
| 296 | // BLOCK 5 |
| 297 | VLDR d10, [r9, 80] // B0 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 298 | VMOV d13, r0, r2 // b1 VMOV b |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 299 | VMLA.F32 q15, q7, d3[1] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 300 | LDR r0, [r9, 120] // B1 low VMOV is in BLOCK 0 |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 301 | NOP |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 302 | LDR r2, [r9, 124] // B1 high |
| 303 | NOP |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 304 | NOP |
| 305 | |
| 306 | # Second group of 16 FMA, First group of loads |
| 307 | // BLOCK 0 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 308 | VLDR d12, [r9, 96] // B1 |
| 309 | VMOV d15, r0, r2 // b1 VMOV b from second group |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 310 | VMLA.F32 q8, q4, d4[0] |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 311 | VMLA.F32 q10, q4, d5[0] |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 312 | VMLA.F32 q12, q4, d6[0] |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 313 | |
| 314 | // BLOCK 1 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 315 | VLDR d14, [r9, 112] // B1 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 316 | VMLA.F32 q14, q4, d7[0] |
Frank Barchard | 279908a | 2019-12-30 17:07:46 -0800 | [diff] [blame] | 317 | VMLA.F32 q9, q5, d4[0] |
| 318 | VMLA.F32 q11, q5, d5[0] |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 319 | ADD r12, r12, 8 // A1++ |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 320 | |
| 321 | // BLOCK 2 |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 322 | ADD r7, r7, 8 // A3++ VLDR B1 lands here |
| 323 | ADD r9, r9, 128 // B++ |
Frank Barchard | 279908a | 2019-12-30 17:07:46 -0800 | [diff] [blame] | 324 | VMLA.F32 q13, q5, d6[0] |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 325 | VMLA.F32 q15, q5, d7[0] |
| 326 | VMLA.F32 q8, q6, d4[1] |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 327 | |
| 328 | // BLOCK 3 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 329 | VMLA.F32 q10, q6, d5[1] |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 330 | VMLA.F32 q12, q6, d6[1] |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 331 | VMLA.F32 q14, q6, d7[1] |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 332 | TST r5, 15 |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 333 | |
| 334 | // BLOCK 4 |
Frank Barchard | 279908a | 2019-12-30 17:07:46 -0800 | [diff] [blame] | 335 | VMLA.F32 q9, q7, d4[1] |
| 336 | VMLA.F32 q11, q7, d5[1] |
| 337 | VMLA.F32 q13, q7, d6[1] |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 338 | |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 339 | // BLOCK 5 |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 340 | VMLA.F32 q15, q7, d7[1] |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 341 | |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 342 | // Is there a remainder?- 1 to 3 floats of A (4, 8 or 12 bytes) |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 343 | BNE 5f |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 344 | |
| 345 | .p2align 3 |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 346 | 4: |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 347 | // Load params pointer |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 348 | LDR r0, [sp, 116] // cn_stride |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 349 | LDR r5, [sp, 120] // params |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 350 | LDR r2, [sp, 0] // kc |
| 351 | SUBS r1, r1, 8 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 352 | |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 353 | // Load min/max values |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 354 | VLD1.32 {d4[],d5[]}, [r5]! |
| 355 | VLD1.32 {d6[],d7[]}, [r5] |
| 356 | |
| 357 | // Clamp |
Marat Dukhan | a51cf48 | 2020-04-08 16:16:19 -0700 | [diff] [blame] | 358 | VMAX.F32 q8, q8, q2 |
| 359 | VMAX.F32 q9, q9, q2 |
| 360 | VMAX.F32 q10, q10, q2 |
| 361 | VMAX.F32 q11, q11, q2 |
| 362 | VMAX.F32 q12, q12, q2 |
| 363 | VMAX.F32 q13, q13, q2 |
| 364 | VMAX.F32 q14, q14, q2 |
| 365 | VMAX.F32 q15, q15, q2 |
| 366 | VMIN.F32 q8, q8, q3 |
| 367 | VMIN.F32 q9, q9, q3 |
| 368 | VMIN.F32 q10, q10, q3 |
| 369 | VMIN.F32 q11, q11, q3 |
| 370 | VMIN.F32 q12, q12, q3 |
| 371 | VMIN.F32 q13, q13, q3 |
| 372 | VMIN.F32 q14, q14, q3 |
| 373 | VMIN.F32 q15, q15, q3 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 374 | |
| 375 | // Store full 4 x 8 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 376 | BLO 10f |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 377 | VST1.32 {d16-d19}, [r11], r0 |
| 378 | SUB r7, r7, r2 |
| 379 | VST1.32 {d20-d23}, [r4], r0 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 380 | SUB r10, r10, r2 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 381 | VST1.32 {d24-d27}, [r8], r0 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 382 | SUB r12, r12, r2 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 383 | VST1.32 {d28-d31}, [r6], r0 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 384 | SUB r3, r3, r2 |
| 385 | BHI 1b |
| 386 | |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 387 | ADD sp, sp, 4 |
| 388 | POP {r4, r5, r6, r7, r8, r9, r10, r11} |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 389 | VPOP {d8-d15} |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 390 | BX lr |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 391 | |
| 392 | .p2align 3 |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 393 | 5: |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 394 | // Is there a remainder?- 2 floats of A (8 bytes) |
| 395 | TST r5, 8 |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 396 | BEQ 6f |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 397 | |
| 398 | // Remainder - 2 floats of A (8 bytes) |
| 399 | VLD1.32 {d0}, [r3]! // A0 |
| 400 | VLDM r9!, {d8-d11} // B0 |
| 401 | VLD1.32 {d1}, [r12]! // A1 |
| 402 | VLD1.32 {d2}, [r10]! // A2 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 403 | VLD1.32 {d3}, [ r7]! // A3 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 404 | |
| 405 | VMLA.F32 q8, q4, d0[0] |
| 406 | VMLA.F32 q9, q5, d0[0] |
| 407 | VMLA.F32 q10, q4, d1[0] |
| 408 | VMLA.F32 q11, q5, d1[0] |
| 409 | VLDM r9!, {d12-d15} // B1 |
| 410 | VMLA.F32 q12, q4, d2[0] |
| 411 | VMLA.F32 q13, q5, d2[0] |
| 412 | VMLA.F32 q14, q4, d3[0] |
| 413 | VMLA.F32 q15, q5, d3[0] |
| 414 | VMLA.F32 q8, q6, d0[1] |
| 415 | VMLA.F32 q9, q7, d0[1] |
| 416 | VMLA.F32 q10, q6, d1[1] |
| 417 | VMLA.F32 q11, q7, d1[1] |
| 418 | VMLA.F32 q12, q6, d2[1] |
| 419 | VMLA.F32 q13, q7, d2[1] |
| 420 | VMLA.F32 q14, q6, d3[1] |
| 421 | VMLA.F32 q15, q7, d3[1] |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 422 | |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 423 | // Is there a remainder?- 1 floats of A (4 bytes) |
| 424 | TST r5, 4 |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 425 | BEQ 4b |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 426 | |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 427 | 6: |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 428 | // Remainder- 1 floats of A (4 bytes) |
| 429 | VLDM r3!, {s0} // A0 |
| 430 | VLDM r9!, {d8-d11} // B0 |
| 431 | VLDM r12!, {s2} // A1 |
| 432 | VLDM r10!, {s4} // A2 |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 433 | VLDM r7!, {s6} // A3 |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 434 | VMLA.F32 q8, q4, d0[0] |
| 435 | VMLA.F32 q9, q5, d0[0] |
| 436 | VMLA.F32 q10, q4, d1[0] |
| 437 | VMLA.F32 q11, q5, d1[0] |
| 438 | VMLA.F32 q12, q4, d2[0] |
| 439 | VMLA.F32 q13, q5, d2[0] |
| 440 | VMLA.F32 q14, q4, d3[0] |
| 441 | VMLA.F32 q15, q5, d3[0] |
Frank Barchard | 9b499d6 | 2020-02-12 11:05:00 -0800 | [diff] [blame] | 442 | B 4b |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 443 | |
Frank Barchard | c01d8a4 | 2020-01-08 16:33:27 -0800 | [diff] [blame] | 444 | // Store odd width |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 445 | 10: |
| 446 | TST r1, 4 |
| 447 | BEQ 11f |
| 448 | VST1.32 {d16-d17}, [r11]! |
| 449 | VMOV q8, q9 |
| 450 | VST1.32 {d20-d21}, [r4]! |
| 451 | VMOV q10, q11 |
| 452 | VST1.32 {d24-d25}, [r8]! |
| 453 | VMOV q12, q13 |
| 454 | VST1.32 {d28-d29}, [r6]! |
| 455 | VMOV q14, q15 |
| 456 | |
| 457 | 11: |
| 458 | TST r1, 2 |
| 459 | BEQ 12f |
| 460 | VST1.32 {d16}, [r11]! |
| 461 | VMOV d16, d17 |
| 462 | VST1.32 {d20}, [r4]! |
| 463 | VMOV d20, d21 |
| 464 | VST1.32 {d24}, [r8]! |
| 465 | VMOV d24, d25 |
| 466 | VST1.32 {d28}, [r6]! |
| 467 | VMOV d28, d29 |
| 468 | |
| 469 | 12: |
| 470 | TST r1, 1 |
| 471 | BEQ 13f |
| 472 | VST1.32 {d16[0]}, [r11] |
| 473 | VST1.32 {d20[0]}, [r4] |
| 474 | VST1.32 {d24[0]}, [r8] |
| 475 | VST1.32 {d28[0]}, [r6] |
| 476 | |
| 477 | 13: |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 478 | ADD sp, sp, 4 |
| 479 | POP {r4, r5, r6, r7, r8, r9, r10, r11} |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 480 | VPOP {d8-d15} |
Frank Barchard | 54afb13 | 2020-01-10 19:12:28 -0800 | [diff] [blame] | 481 | BX lr |
Frank Barchard | 1391604 | 2019-12-11 10:56:34 -0800 | [diff] [blame] | 482 | |
| 483 | END_FUNCTION xnn_f32_gemm_ukernel_4x8__aarch32_neon_cortex_a53 |
| 484 | |
| 485 | #ifdef __ELF__ |
| 486 | .section ".note.GNU-stack","",%progbits |
Frank Barchard | 005feb8 | 2019-12-16 18:16:55 -0800 | [diff] [blame] | 487 | #endif |