Frank Barchard | baa9ead | 2019-10-18 18:06:41 -0700 | [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. |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 5 | |
| 6 | #include <xnnpack/assembly.h> |
| 7 | |
Frank Barchard | a2f1891 | 2021-12-28 14:17:05 -0800 | [diff] [blame] | 8 | # void xnn_f32_gemm${"inc" if INC else ""}_minmax_ukernel_6x8__aarch64_neonfma${"_prfm" if PREFETCH else ""}_cortex_a75( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 9 | # size_t mr, x0 |
| 10 | # size_t nc, x1 |
| 11 | # size_t kc, x2 / x0 |
| 12 | # const uint8_t*restrict a, x3 |
| 13 | # size_t a_stride, x4 |
| 14 | # const void*restrict w, x5 |
| 15 | # uint8_t*restrict c, x6 |
| 16 | # size_t cm_stride, x7 |
Frank Barchard | 6724218 | 2020-06-11 11:12:50 -0700 | [diff] [blame] | 17 | # size_t cn_stride, [sp] -> (x0) |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 18 | $if INC: |
| 19 | # const float*restrict acc, [sp + 8] -> x15 |
Marat Dukhan | f196d01 | 2020-04-15 11:50:03 -0700 | [diff] [blame] | 20 | # const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) [sp + 16] -> x8 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 21 | $else: |
Marat Dukhan | f196d01 | 2020-04-15 11:50:03 -0700 | [diff] [blame] | 22 | # const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) [sp + 8] -> x8 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 23 | |
Frank Barchard | 909564c | 2020-06-09 03:54:33 -0700 | [diff] [blame] | 24 | # d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS. |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 25 | |
| 26 | # A pointers |
| 27 | # x3 a0 |
| 28 | # x9 a1 |
| 29 | # x10 a2 |
| 30 | # x11 a3 |
| 31 | # x12 a4 |
| 32 | # x4 a5 |
| 33 | |
| 34 | # C pointers |
| 35 | # x6 c0 |
| 36 | # x16 c1 |
| 37 | # x17 c2 |
Frank Barchard | 6724218 | 2020-06-11 11:12:50 -0700 | [diff] [blame] | 38 | # x14 c3 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 39 | # x13 c4 |
| 40 | # x7 c5 |
| 41 | |
| 42 | # Vector register usage |
| 43 | # A0 v0 v6 |
| 44 | # A1 v1 v7 |
| 45 | # A2 v2 v8 |
| 46 | # A3 v3 v9 |
| 47 | # A4 v4 v10 |
| 48 | # A5 v5 v11 |
| 49 | # B v12 v13 v14 v15 |
| 50 | # B v16 v17 v18 v19 |
| 51 | # C v20 v21 |
| 52 | # C v22 v23 |
| 53 | # C v24 v25 |
| 54 | # C v26 v27 |
| 55 | # C v28 v29 |
| 56 | # C v30 v31 |
| 57 | # Clamp v6 v7 |
| 58 | |
Frank Barchard | a2f1891 | 2021-12-28 14:17:05 -0800 | [diff] [blame] | 59 | BEGIN_FUNCTION xnn_f32_gemm${"inc" if INC else ""}_minmax_ukernel_6x8__aarch64_neonfma${"_prfm" if PREFETCH else ""}_cortex_a75 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 60 | |
Frank Barchard | f5cc7e7 | 2020-04-20 11:35:48 -0700 | [diff] [blame] | 61 | $if INC: |
Frank Barchard | 6724218 | 2020-06-11 11:12:50 -0700 | [diff] [blame] | 62 | # Load acc, params pointer |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 63 | LDP x15, x8, [sp, 8] |
Frank Barchard | f5cc7e7 | 2020-04-20 11:35:48 -0700 | [diff] [blame] | 64 | $else: |
Frank Barchard | 6724218 | 2020-06-11 11:12:50 -0700 | [diff] [blame] | 65 | # Load params pointer |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 66 | LDR x8, [sp, 8] |
Frank Barchard | f5cc7e7 | 2020-04-20 11:35:48 -0700 | [diff] [blame] | 67 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 68 | # Clamp A and C pointers / Save d8-d15 on stack |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 69 | STP d8, d9, [sp, -64]! |
| 70 | CMP x0, 2 // if mr < 2 |
| 71 | ADD x9, x3, x4 // a1 = a0 + a_stride |
| 72 | ADD x16, x6, x7 // c1 = c0 + cm_stride |
| 73 | CSEL x9, x3, x9, LO // a1 = a0 |
| 74 | CSEL x16, x6, x16, LO // c1 = c0 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 75 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 76 | STP d10, d11, [sp, 16] |
| 77 | ADD x10, x9, x4 // a2 = a1 + a_stride |
| 78 | ADD x17, x16, x7 // c2 = c1 + cm_stride |
Frank Barchard | 7c9f1f9 | 2021-06-04 14:38:55 -0700 | [diff] [blame] | 79 | // if mr <= 2 |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 80 | CSEL x10, x9, x10, LS // a2 = a1 |
| 81 | CSEL x17, x16, x17, LS // c2 = c1 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 82 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 83 | STP d12, d13, [sp, 32] |
| 84 | CMP x0, 4 // if mr < 4 |
| 85 | ADD x11, x10, x4 // a3 = a2 + a_stride |
| 86 | ADD x14, x17, x7 // c3 = c2 + cm_stride |
| 87 | CSEL x11, x10, x11, LO // a3 = a2 |
| 88 | CSEL x14, x17, x14, LO // c3 = c2 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 89 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 90 | STP d14, d15, [sp, 48] |
| 91 | ADD x12, x11, x4 // a4 = a3 + a_stride |
| 92 | ADD x13, x14, x7 // c4 = c3 + cm_stride |
Frank Barchard | 7c9f1f9 | 2021-06-04 14:38:55 -0700 | [diff] [blame] | 93 | // if mr <= 4 |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 94 | CSEL x12, x11, x12, LS // a4 = a3 |
| 95 | CSEL x13, x14, x13, LS // c4 = c3 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 96 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 97 | CMP x0, 6 // if mr < 6 |
| 98 | ADD x4, x12, x4 // a5 = a4 + a_stride |
| 99 | ADD x7, x13, x7 // c5 = c4 + cm_stride |
| 100 | CSEL x4, x12, x4, LO // a5 = a4 |
| 101 | CSEL x7, x13, x7, LO // c5 = c4 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 102 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 103 | 0: |
| 104 | $if INC: |
| 105 | # Load initial accumulators |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 106 | LDP q20, q21, [x15], 32 |
| 107 | LDP q22, q23, [x15], 32 |
| 108 | LDP q24, q25, [x15], 32 |
| 109 | LDP q26, q27, [x15], 32 |
| 110 | LDP q28, q29, [x15], 32 |
| 111 | LDP q30, q31, [x15], 32 |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 112 | $if PREFETCH: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 113 | PRFM PLDL1KEEP, [x5, 0] // Prefetch B |
Frank Barchard | cbfa338 | 2021-05-07 10:30:05 -0700 | [diff] [blame] | 114 | PRFM PLDL1KEEP, [x5, 64] |
| 115 | PRFM PLDL1KEEP, [x5, 128] |
| 116 | PRFM PLDL1KEEP, [x5, 192] |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 117 | PRFM PLDL1KEEP, [x3] // Prefetch A |
Frank Barchard | cbfa338 | 2021-05-07 10:30:05 -0700 | [diff] [blame] | 118 | PRFM PLDL1KEEP, [x9] |
| 119 | PRFM PLDL1KEEP, [x10] |
| 120 | PRFM PLDL1KEEP, [x11] |
| 121 | PRFM PLDL1KEEP, [x12] |
| 122 | PRFM PLDL1KEEP, [x4] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 123 | $else: |
| 124 | # Load initial bias from w into accumulators |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 125 | LDP q20, q21, [x5], 32 |
| 126 | MOV v22.16b, v20.16b |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 127 | $if PREFETCH: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 128 | PRFM PLDL1KEEP, [x5, 0] // Prefetch B |
| 129 | MOV v23.16b, v21.16b |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 130 | $if PREFETCH: |
Frank Barchard | cbfa338 | 2021-05-07 10:30:05 -0700 | [diff] [blame] | 131 | PRFM PLDL1KEEP, [x5, 64] |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 132 | MOV v24.16b, v20.16b |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 133 | $if PREFETCH: |
Frank Barchard | cbfa338 | 2021-05-07 10:30:05 -0700 | [diff] [blame] | 134 | PRFM PLDL1KEEP, [x5, 128] |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 135 | MOV v25.16b, v21.16b |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 136 | $if PREFETCH: |
Frank Barchard | cbfa338 | 2021-05-07 10:30:05 -0700 | [diff] [blame] | 137 | PRFM PLDL1KEEP, [x5, 192] |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 138 | MOV v26.16b, v20.16b |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 139 | $if PREFETCH: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 140 | PRFM PLDL1KEEP, [x3] // Prefetch A |
| 141 | MOV v27.16b, v21.16b |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 142 | $if PREFETCH: |
Frank Barchard | cbfa338 | 2021-05-07 10:30:05 -0700 | [diff] [blame] | 143 | PRFM PLDL1KEEP, [x9] |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 144 | MOV v28.16b, v20.16b |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 145 | $if PREFETCH: |
Frank Barchard | cbfa338 | 2021-05-07 10:30:05 -0700 | [diff] [blame] | 146 | PRFM PLDL1KEEP, [x10] |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 147 | MOV v29.16b, v21.16b |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 148 | $if PREFETCH: |
Frank Barchard | cbfa338 | 2021-05-07 10:30:05 -0700 | [diff] [blame] | 149 | PRFM PLDL1KEEP, [x11] |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 150 | MOV v30.16b, v20.16b |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 151 | $if PREFETCH: |
Frank Barchard | cbfa338 | 2021-05-07 10:30:05 -0700 | [diff] [blame] | 152 | PRFM PLDL1KEEP, [x12] |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 153 | MOV v31.16b, v21.16b |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 154 | $if PREFETCH: |
Frank Barchard | cbfa338 | 2021-05-07 10:30:05 -0700 | [diff] [blame] | 155 | PRFM PLDL1KEEP, [x4] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 156 | |
| 157 | # Is there at least 8 floats (32 bytes) for prologue + epilogue? |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 158 | SUBS x0, x2, 32 // k = kc - 32 |
| 159 | B.LO 4f |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 160 | |
| 161 | # Prologue - loads for main loop of 96 FMA |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 162 | LDR q0, [x3], 16 |
| 163 | LDR q1, [x9], 16 |
| 164 | LDR q2, [x10], 16 |
| 165 | LDR q3, [x11], 16 |
| 166 | LDR q4, [x12], 16 |
| 167 | LDR q5, [x4], 16 |
| 168 | LDP q12, q13, [x5], 32 // Fetch 3 B (4th deferred) |
| 169 | LDP q14, q15, [x5], 32 |
| 170 | LDP q16, q17, [x5], 32 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 171 | |
| 172 | # Is there at least 8 floats (32 bytes) for main loop? |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 173 | SUBS x0, x0, 32 |
| 174 | B.LO 2f |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 175 | |
| 176 | # Main loop - 8 floats of A (32 bytes) |
| 177 | # 96 FMA + 6 LDP A + 8 LDP B |
| 178 | 1: |
| 179 | # First group of 4 A. 48 FMA. |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 180 | FMLA v20.4s, v12.4s, v0.s[0] |
| 181 | LDP q18, q19, [x5], 32 // Load last B |
| 182 | FMLA v22.4s, v12.4s, v1.s[0] |
| 183 | FMLA v24.4s, v12.4s, v2.s[0] |
| 184 | FMLA v26.4s, v12.4s, v3.s[0] |
| 185 | FMLA v28.4s, v12.4s, v4.s[0] |
| 186 | FMLA v30.4s, v12.4s, v5.s[0] |
| 187 | FMLA v21.4s, v13.4s, v0.s[0] |
| 188 | FMLA v23.4s, v13.4s, v1.s[0] |
| 189 | FMLA v25.4s, v13.4s, v2.s[0] |
| 190 | FMLA v27.4s, v13.4s, v3.s[0] |
| 191 | FMLA v29.4s, v13.4s, v4.s[0] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 192 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 193 | FMLA v31.4s, v13.4s, v5.s[0] |
| 194 | FMLA v20.4s, v14.4s, v0.s[1] |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 195 | $if PREFETCH: |
Frank Barchard | cbfa338 | 2021-05-07 10:30:05 -0700 | [diff] [blame] | 196 | PRFM PLDL1KEEP, [x5, 128] // Prefetch B |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 197 | FMLA v22.4s, v14.4s, v1.s[1] |
| 198 | FMLA v24.4s, v14.4s, v2.s[1] |
| 199 | FMLA v26.4s, v14.4s, v3.s[1] |
| 200 | FMLA v28.4s, v14.4s, v4.s[1] |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 201 | $if PREFETCH: |
Frank Barchard | cbfa338 | 2021-05-07 10:30:05 -0700 | [diff] [blame] | 202 | PRFM PLDL1KEEP, [x5, 256] |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 203 | FMLA v30.4s, v14.4s, v5.s[1] |
| 204 | FMLA v21.4s, v15.4s, v0.s[1] |
| 205 | FMLA v23.4s, v15.4s, v1.s[1] |
| 206 | FMLA v25.4s, v15.4s, v2.s[1] |
| 207 | LDR q6, [x3], 16 // Load next 6 A |
| 208 | FMLA v27.4s, v15.4s, v3.s[1] |
| 209 | FMLA v29.4s, v15.4s, v4.s[1] |
| 210 | FMLA v31.4s, v15.4s, v5.s[1] |
| 211 | LDR q7, [x9], 16 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 212 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 213 | FMLA v20.4s, v16.4s, v0.s[2] |
| 214 | FMLA v22.4s, v16.4s, v1.s[2] |
| 215 | FMLA v24.4s, v16.4s, v2.s[2] |
| 216 | LDR q8, [x10], 16 |
| 217 | FMLA v26.4s, v16.4s, v3.s[2] |
| 218 | FMLA v28.4s, v16.4s, v4.s[2] |
| 219 | FMLA v30.4s, v16.4s, v5.s[2] |
| 220 | LDR q9, [x11], 16 |
| 221 | FMLA v21.4s, v17.4s, v0.s[2] |
| 222 | FMLA v23.4s, v17.4s, v1.s[2] |
| 223 | FMLA v25.4s, v17.4s, v2.s[2] |
| 224 | LDR q10, [x12], 16 |
| 225 | FMLA v27.4s, v17.4s, v3.s[2] |
| 226 | FMLA v29.4s, v17.4s, v4.s[2] |
| 227 | FMLA v31.4s, v17.4s, v5.s[2] |
| 228 | LDR q11, [x4], 16 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 229 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 230 | FMLA v20.4s, v18.4s, v0.s[3] |
| 231 | FMLA v22.4s, v18.4s, v1.s[3] |
| 232 | FMLA v24.4s, v18.4s, v2.s[3] |
| 233 | LDP q12, q13, [x5], 32 // Load 4 B |
| 234 | FMLA v26.4s, v18.4s, v3.s[3] |
| 235 | FMLA v28.4s, v18.4s, v4.s[3] |
| 236 | FMLA v30.4s, v18.4s, v5.s[3] |
| 237 | LDP q14, q15, [x5], 32 |
| 238 | FMLA v21.4s, v19.4s, v0.s[3] |
| 239 | FMLA v23.4s, v19.4s, v1.s[3] |
| 240 | FMLA v25.4s, v19.4s, v2.s[3] |
| 241 | LDP q16, q17, [x5], 32 |
| 242 | FMLA v27.4s, v19.4s, v3.s[3] |
| 243 | FMLA v29.4s, v19.4s, v4.s[3] |
| 244 | FMLA v31.4s, v19.4s, v5.s[3] |
| 245 | LDP q18, q19, [x5], 32 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 246 | |
| 247 | # Second group of 4 A. 48 FMA. |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 248 | FMLA v20.4s, v12.4s, v6.s[0] |
| 249 | FMLA v22.4s, v12.4s, v7.s[0] |
| 250 | FMLA v24.4s, v12.4s, v8.s[0] |
| 251 | LDR q0, [x3], 16 // Load next 6 A |
| 252 | FMLA v26.4s, v12.4s, v9.s[0] |
| 253 | FMLA v28.4s, v12.4s, v10.s[0] |
| 254 | FMLA v30.4s, v12.4s, v11.s[0] |
| 255 | LDR q1, [x9], 16 |
| 256 | FMLA v21.4s, v13.4s, v6.s[0] |
| 257 | FMLA v23.4s, v13.4s, v7.s[0] |
| 258 | FMLA v25.4s, v13.4s, v8.s[0] |
| 259 | LDR q2, [x10], 16 |
| 260 | FMLA v27.4s, v13.4s, v9.s[0] |
| 261 | FMLA v29.4s, v13.4s, v10.s[0] |
| 262 | FMLA v31.4s, v13.4s, v11.s[0] |
| 263 | LDR q3, [x11], 16 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 264 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 265 | FMLA v20.4s, v14.4s, v6.s[1] |
| 266 | FMLA v22.4s, v14.4s, v7.s[1] |
| 267 | FMLA v24.4s, v14.4s, v8.s[1] |
| 268 | LDR q4, [x12], 16 |
| 269 | FMLA v26.4s, v14.4s, v9.s[1] |
| 270 | FMLA v28.4s, v14.4s, v10.s[1] |
| 271 | FMLA v30.4s, v14.4s, v11.s[1] |
| 272 | LDR q5, [x4], 16 |
| 273 | FMLA v21.4s, v15.4s, v6.s[1] |
| 274 | FMLA v23.4s, v15.4s, v7.s[1] |
| 275 | FMLA v25.4s, v15.4s, v8.s[1] |
| 276 | LDP q12, q13, [x5], 32 // Load next 3 B (not last) |
| 277 | FMLA v27.4s, v15.4s, v9.s[1] |
| 278 | FMLA v29.4s, v15.4s, v10.s[1] |
| 279 | FMLA v31.4s, v15.4s, v11.s[1] |
| 280 | LDP q14, q15, [x5], 32 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 281 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 282 | FMLA v20.4s, v16.4s, v6.s[2] |
| 283 | FMLA v22.4s, v16.4s, v7.s[2] |
| 284 | FMLA v24.4s, v16.4s, v8.s[2] |
| 285 | FMLA v26.4s, v16.4s, v9.s[2] |
| 286 | FMLA v28.4s, v16.4s, v10.s[2] |
| 287 | FMLA v30.4s, v16.4s, v11.s[2] |
| 288 | FMLA v21.4s, v17.4s, v6.s[2] |
| 289 | FMLA v23.4s, v17.4s, v7.s[2] |
| 290 | FMLA v25.4s, v17.4s, v8.s[2] |
| 291 | FMLA v27.4s, v17.4s, v9.s[2] |
| 292 | FMLA v29.4s, v17.4s, v10.s[2] |
| 293 | FMLA v31.4s, v17.4s, v11.s[2] |
| 294 | LDP q16, q17, [x5], 32 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 295 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 296 | FMLA v20.4s, v18.4s, v6.s[3] |
| 297 | FMLA v22.4s, v18.4s, v7.s[3] |
| 298 | SUBS x0, x0, 32 |
| 299 | FMLA v24.4s, v18.4s, v8.s[3] |
| 300 | FMLA v26.4s, v18.4s, v9.s[3] |
| 301 | FMLA v28.4s, v18.4s, v10.s[3] |
| 302 | FMLA v30.4s, v18.4s, v11.s[3] |
| 303 | FMLA v21.4s, v19.4s, v6.s[3] |
| 304 | FMLA v23.4s, v19.4s, v7.s[3] |
| 305 | FMLA v25.4s, v19.4s, v8.s[3] |
| 306 | FMLA v27.4s, v19.4s, v9.s[3] |
| 307 | FMLA v29.4s, v19.4s, v10.s[3] |
| 308 | FMLA v31.4s, v19.4s, v11.s[3] |
| 309 | B.HS 1b |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 310 | |
| 311 | # Epilogue - 8 floats of A (32 bytes) |
| 312 | # 96 FMA + 6 LDP A + 8 LDP B |
| 313 | # First block same as main loop. Second block has no preloads. |
| 314 | 2: |
| 315 | # First group of 4 A. 48 FMA. |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 316 | FMLA v20.4s, v12.4s, v0.s[0] |
| 317 | LDP q18, q19, [x5], 32 // Load last B |
| 318 | FMLA v22.4s, v12.4s, v1.s[0] |
| 319 | FMLA v24.4s, v12.4s, v2.s[0] |
| 320 | FMLA v26.4s, v12.4s, v3.s[0] |
| 321 | FMLA v28.4s, v12.4s, v4.s[0] |
| 322 | FMLA v30.4s, v12.4s, v5.s[0] |
| 323 | FMLA v21.4s, v13.4s, v0.s[0] |
| 324 | FMLA v23.4s, v13.4s, v1.s[0] |
| 325 | FMLA v25.4s, v13.4s, v2.s[0] |
| 326 | FMLA v27.4s, v13.4s, v3.s[0] |
| 327 | FMLA v29.4s, v13.4s, v4.s[0] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 328 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 329 | FMLA v31.4s, v13.4s, v5.s[0] |
| 330 | FMLA v20.4s, v14.4s, v0.s[1] |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 331 | $if PREFETCH: |
Frank Barchard | cbfa338 | 2021-05-07 10:30:05 -0700 | [diff] [blame] | 332 | PRFM PLDL1KEEP, [x5, 128] // Prefetch B |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 333 | FMLA v22.4s, v14.4s, v1.s[1] |
| 334 | FMLA v24.4s, v14.4s, v2.s[1] |
| 335 | FMLA v26.4s, v14.4s, v3.s[1] |
| 336 | FMLA v28.4s, v14.4s, v4.s[1] |
Frank Barchard | 387c2d1 | 2019-12-16 19:14:07 -0800 | [diff] [blame] | 337 | $if PREFETCH: |
Frank Barchard | cbfa338 | 2021-05-07 10:30:05 -0700 | [diff] [blame] | 338 | PRFM PLDL1KEEP, [x5, 256] |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 339 | FMLA v30.4s, v14.4s, v5.s[1] |
| 340 | FMLA v21.4s, v15.4s, v0.s[1] |
| 341 | FMLA v23.4s, v15.4s, v1.s[1] |
| 342 | FMLA v25.4s, v15.4s, v2.s[1] |
| 343 | LDR q6, [x3], 16 // Load next 6 A |
| 344 | FMLA v27.4s, v15.4s, v3.s[1] |
| 345 | FMLA v29.4s, v15.4s, v4.s[1] |
| 346 | FMLA v31.4s, v15.4s, v5.s[1] |
| 347 | LDR q7, [x9], 16 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 348 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 349 | FMLA v20.4s, v16.4s, v0.s[2] |
| 350 | FMLA v22.4s, v16.4s, v1.s[2] |
| 351 | FMLA v24.4s, v16.4s, v2.s[2] |
| 352 | LDR q8, [x10], 16 |
| 353 | FMLA v26.4s, v16.4s, v3.s[2] |
| 354 | FMLA v28.4s, v16.4s, v4.s[2] |
| 355 | FMLA v30.4s, v16.4s, v5.s[2] |
| 356 | LDR q9, [x11], 16 |
| 357 | FMLA v21.4s, v17.4s, v0.s[2] |
| 358 | FMLA v23.4s, v17.4s, v1.s[2] |
| 359 | FMLA v25.4s, v17.4s, v2.s[2] |
| 360 | LDR q10, [x12], 16 |
| 361 | FMLA v27.4s, v17.4s, v3.s[2] |
| 362 | FMLA v29.4s, v17.4s, v4.s[2] |
| 363 | FMLA v31.4s, v17.4s, v5.s[2] |
| 364 | LDR q11, [x4], 16 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 365 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 366 | FMLA v20.4s, v18.4s, v0.s[3] |
| 367 | FMLA v22.4s, v18.4s, v1.s[3] |
| 368 | FMLA v24.4s, v18.4s, v2.s[3] |
| 369 | LDP q12, q13, [x5], 32 // Load 4 B |
| 370 | FMLA v26.4s, v18.4s, v3.s[3] |
| 371 | FMLA v28.4s, v18.4s, v4.s[3] |
| 372 | FMLA v30.4s, v18.4s, v5.s[3] |
| 373 | LDP q14, q15, [x5], 32 |
| 374 | FMLA v21.4s, v19.4s, v0.s[3] |
| 375 | FMLA v23.4s, v19.4s, v1.s[3] |
| 376 | FMLA v25.4s, v19.4s, v2.s[3] |
| 377 | LDP q16, q17, [x5], 32 |
| 378 | FMLA v27.4s, v19.4s, v3.s[3] |
| 379 | FMLA v29.4s, v19.4s, v4.s[3] |
| 380 | FMLA v31.4s, v19.4s, v5.s[3] |
| 381 | LDP q18, q19, [x5], 32 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 382 | |
| 383 | # Second group of 4 A. 48 FMA. |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 384 | FMLA v20.4s, v12.4s, v6.s[0] |
| 385 | FMLA v22.4s, v12.4s, v7.s[0] |
| 386 | FMLA v24.4s, v12.4s, v8.s[0] |
| 387 | FMLA v26.4s, v12.4s, v9.s[0] |
| 388 | FMLA v28.4s, v12.4s, v10.s[0] |
| 389 | FMLA v30.4s, v12.4s, v11.s[0] |
| 390 | FMLA v21.4s, v13.4s, v6.s[0] |
| 391 | FMLA v23.4s, v13.4s, v7.s[0] |
| 392 | FMLA v25.4s, v13.4s, v8.s[0] |
| 393 | FMLA v27.4s, v13.4s, v9.s[0] |
| 394 | FMLA v29.4s, v13.4s, v10.s[0] |
| 395 | FMLA v31.4s, v13.4s, v11.s[0] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 396 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 397 | FMLA v20.4s, v14.4s, v6.s[1] |
| 398 | FMLA v22.4s, v14.4s, v7.s[1] |
| 399 | FMLA v24.4s, v14.4s, v8.s[1] |
| 400 | FMLA v26.4s, v14.4s, v9.s[1] |
| 401 | FMLA v28.4s, v14.4s, v10.s[1] |
| 402 | FMLA v30.4s, v14.4s, v11.s[1] |
| 403 | FMLA v21.4s, v15.4s, v6.s[1] |
| 404 | FMLA v23.4s, v15.4s, v7.s[1] |
| 405 | FMLA v25.4s, v15.4s, v8.s[1] |
| 406 | FMLA v27.4s, v15.4s, v9.s[1] |
| 407 | FMLA v29.4s, v15.4s, v10.s[1] |
| 408 | FMLA v31.4s, v15.4s, v11.s[1] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 409 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 410 | FMLA v20.4s, v16.4s, v6.s[2] |
| 411 | FMLA v22.4s, v16.4s, v7.s[2] |
| 412 | FMLA v24.4s, v16.4s, v8.s[2] |
| 413 | FMLA v26.4s, v16.4s, v9.s[2] |
| 414 | FMLA v28.4s, v16.4s, v10.s[2] |
| 415 | FMLA v30.4s, v16.4s, v11.s[2] |
| 416 | FMLA v21.4s, v17.4s, v6.s[2] |
| 417 | FMLA v23.4s, v17.4s, v7.s[2] |
| 418 | FMLA v25.4s, v17.4s, v8.s[2] |
| 419 | FMLA v27.4s, v17.4s, v9.s[2] |
| 420 | FMLA v29.4s, v17.4s, v10.s[2] |
| 421 | FMLA v31.4s, v17.4s, v11.s[2] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 422 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 423 | FMLA v20.4s, v18.4s, v6.s[3] |
| 424 | FMLA v22.4s, v18.4s, v7.s[3] |
| 425 | FMLA v24.4s, v18.4s, v8.s[3] |
| 426 | FMLA v26.4s, v18.4s, v9.s[3] |
| 427 | FMLA v28.4s, v18.4s, v10.s[3] |
| 428 | FMLA v30.4s, v18.4s, v11.s[3] |
| 429 | FMLA v21.4s, v19.4s, v6.s[3] |
| 430 | FMLA v23.4s, v19.4s, v7.s[3] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 431 | |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 432 | # Load min/max values |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 433 | LD2R {v6.4s, v7.4s}, [x8] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 434 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 435 | FMLA v25.4s, v19.4s, v8.s[3] |
| 436 | FMLA v27.4s, v19.4s, v9.s[3] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 437 | # Is there a remainder?- 4 floats of A (16 bytes) or less |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 438 | TST x0, 31 |
| 439 | FMLA v29.4s, v19.4s, v10.s[3] |
| 440 | FMLA v31.4s, v19.4s, v11.s[3] |
| 441 | B.NE 4f |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 442 | |
| 443 | # Clamp |
| 444 | 3: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 445 | FMAX v20.4s, v20.4s, v6.4s |
Frank Barchard | 6724218 | 2020-06-11 11:12:50 -0700 | [diff] [blame] | 446 | # Load cn_stride |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 447 | LDR x0, [sp, 64] |
| 448 | FMAX v21.4s, v21.4s, v6.4s |
| 449 | FMAX v22.4s, v22.4s, v6.4s |
| 450 | FMAX v23.4s, v23.4s, v6.4s |
| 451 | FMAX v24.4s, v24.4s, v6.4s |
| 452 | FMAX v25.4s, v25.4s, v6.4s |
| 453 | FMAX v26.4s, v26.4s, v6.4s |
| 454 | FMAX v27.4s, v27.4s, v6.4s |
| 455 | FMAX v28.4s, v28.4s, v6.4s |
| 456 | FMAX v29.4s, v29.4s, v6.4s |
| 457 | FMAX v30.4s, v30.4s, v6.4s |
| 458 | FMAX v31.4s, v31.4s, v6.4s |
| 459 | SUBS x1, x1, 8 |
| 460 | FMIN v20.4s, v20.4s, v7.4s |
| 461 | FMIN v21.4s, v21.4s, v7.4s |
| 462 | FMIN v22.4s, v22.4s, v7.4s |
| 463 | FMIN v23.4s, v23.4s, v7.4s |
| 464 | FMIN v24.4s, v24.4s, v7.4s |
| 465 | FMIN v25.4s, v25.4s, v7.4s |
| 466 | FMIN v26.4s, v26.4s, v7.4s |
| 467 | FMIN v27.4s, v27.4s, v7.4s |
| 468 | FMIN v28.4s, v28.4s, v7.4s |
| 469 | FMIN v29.4s, v29.4s, v7.4s |
| 470 | FMIN v30.4s, v30.4s, v7.4s |
| 471 | FMIN v31.4s, v31.4s, v7.4s |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 472 | |
| 473 | # Store full 6 x 8 |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 474 | B.LO 7f |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 475 | |
| 476 | $if INC: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 477 | STP q30, q31, [x7] |
| 478 | ADD x7, x7, x0 |
| 479 | SUB x3, x3, x2 // a0 -= kc |
| 480 | STP q28, q29, [x13] |
| 481 | ADD x13, x13, x0 |
| 482 | SUB x9, x9, x2 // a1 -= kc |
| 483 | STP q26, q27, [x14] |
| 484 | ADD x14, x14, x0 |
| 485 | SUB x10, x10, x2 // a2 -= kc |
| 486 | STP q24, q25, [x17] |
| 487 | ADD x17, x17, x0 |
| 488 | SUB x11, x11, x2 // a3 -= kc |
| 489 | STP q22, q23, [x16] |
| 490 | ADD x16, x16, x0 |
| 491 | SUB x12, x12, x2 // a4 -= kc |
| 492 | STP q20, q21, [x6] |
| 493 | ADD x6, x6, x0 |
| 494 | SUB x4, x4, x2 // a5 -= kc |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 495 | $else: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 496 | STP q20, q21, [x6] |
| 497 | ADD x6, x6, x0 |
| 498 | SUB x3, x3, x2 // a0 -= kc |
| 499 | STP q22, q23, [x16] |
| 500 | ADD x16, x16, x0 |
| 501 | SUB x9, x9, x2 // a1 -= kc |
| 502 | STP q24, q25, [x17] |
| 503 | ADD x17, x17, x0 |
| 504 | SUB x10, x10, x2 // a2 -= kc |
| 505 | STP q26, q27, [x14] |
| 506 | ADD x14, x14, x0 |
| 507 | SUB x11, x11, x2 // a3 -= kc |
| 508 | STP q28, q29, [x13] |
| 509 | ADD x13, x13, x0 |
| 510 | SUB x12, x12, x2 // a4 -= kc |
| 511 | STP q30, q31, [x7] |
| 512 | ADD x7, x7, x0 |
| 513 | SUB x4, x4, x2 // a5 -= kc |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 514 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 515 | B.HI 0b |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 516 | |
| 517 | # Restore d8-d15 from stack |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 518 | LDP d14, d15, [sp, 48] |
| 519 | LDP d12, d13, [sp, 32] |
| 520 | LDP d10, d11, [sp, 16] |
| 521 | LDP d8, d9, [sp], 64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 522 | RET |
| 523 | |
| 524 | 4: |
Marat Dukhan | eb09a6b | 2020-04-08 17:34:32 -0700 | [diff] [blame] | 525 | # Load min/max values |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 526 | LD2R {v6.4s, v7.4s}, [x8] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 527 | |
| 528 | # Is there a remainder?- 4 floats of A (16 bytes) |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 529 | TBZ x0, 4, 5f |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 530 | |
| 531 | # Remainder- 4 floats of A (16 bytes) |
| 532 | # Load A |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 533 | LDR q0, [x3], 16 |
| 534 | LDR q1, [x9], 16 |
| 535 | LDR q2, [x10], 16 |
| 536 | LDR q3, [x11], 16 |
| 537 | LDR q4, [x12], 16 |
| 538 | LDR q5, [x4], 16 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 539 | # Load B |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 540 | LDP q12, q13, [x5], 32 |
| 541 | LDP q14, q15, [x5], 32 |
| 542 | LDP q16, q17, [x5], 32 |
| 543 | LDP q18, q19, [x5], 32 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 544 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 545 | FMLA v20.4s, v12.4s, v0.s[0] |
| 546 | FMLA v22.4s, v12.4s, v1.s[0] |
| 547 | FMLA v24.4s, v12.4s, v2.s[0] |
| 548 | FMLA v26.4s, v12.4s, v3.s[0] |
| 549 | FMLA v28.4s, v12.4s, v4.s[0] |
| 550 | FMLA v30.4s, v12.4s, v5.s[0] |
| 551 | FMLA v21.4s, v13.4s, v0.s[0] |
| 552 | FMLA v23.4s, v13.4s, v1.s[0] |
| 553 | FMLA v25.4s, v13.4s, v2.s[0] |
| 554 | FMLA v27.4s, v13.4s, v3.s[0] |
| 555 | FMLA v29.4s, v13.4s, v4.s[0] |
| 556 | FMLA v31.4s, v13.4s, v5.s[0] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 557 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 558 | FMLA v20.4s, v14.4s, v0.s[1] |
| 559 | FMLA v22.4s, v14.4s, v1.s[1] |
| 560 | FMLA v24.4s, v14.4s, v2.s[1] |
| 561 | FMLA v26.4s, v14.4s, v3.s[1] |
| 562 | FMLA v28.4s, v14.4s, v4.s[1] |
| 563 | FMLA v30.4s, v14.4s, v5.s[1] |
| 564 | FMLA v21.4s, v15.4s, v0.s[1] |
| 565 | FMLA v23.4s, v15.4s, v1.s[1] |
| 566 | FMLA v25.4s, v15.4s, v2.s[1] |
| 567 | FMLA v27.4s, v15.4s, v3.s[1] |
| 568 | FMLA v29.4s, v15.4s, v4.s[1] |
| 569 | FMLA v31.4s, v15.4s, v5.s[1] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 570 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 571 | FMLA v20.4s, v16.4s, v0.s[2] |
| 572 | FMLA v22.4s, v16.4s, v1.s[2] |
| 573 | FMLA v24.4s, v16.4s, v2.s[2] |
| 574 | FMLA v26.4s, v16.4s, v3.s[2] |
| 575 | FMLA v28.4s, v16.4s, v4.s[2] |
| 576 | FMLA v30.4s, v16.4s, v5.s[2] |
| 577 | FMLA v21.4s, v17.4s, v0.s[2] |
| 578 | FMLA v23.4s, v17.4s, v1.s[2] |
| 579 | FMLA v25.4s, v17.4s, v2.s[2] |
| 580 | FMLA v27.4s, v17.4s, v3.s[2] |
| 581 | FMLA v29.4s, v17.4s, v4.s[2] |
| 582 | FMLA v31.4s, v17.4s, v5.s[2] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 583 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 584 | FMLA v20.4s, v18.4s, v0.s[3] |
| 585 | FMLA v22.4s, v18.4s, v1.s[3] |
| 586 | FMLA v24.4s, v18.4s, v2.s[3] |
| 587 | FMLA v26.4s, v18.4s, v3.s[3] |
| 588 | FMLA v28.4s, v18.4s, v4.s[3] |
| 589 | FMLA v30.4s, v18.4s, v5.s[3] |
| 590 | FMLA v21.4s, v19.4s, v0.s[3] |
| 591 | FMLA v23.4s, v19.4s, v1.s[3] |
| 592 | FMLA v25.4s, v19.4s, v2.s[3] |
| 593 | FMLA v27.4s, v19.4s, v3.s[3] |
| 594 | FMLA v29.4s, v19.4s, v4.s[3] |
| 595 | FMLA v31.4s, v19.4s, v5.s[3] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 596 | |
| 597 | # Is there a remainder?- 2 floats of A (8 bytes) |
| 598 | 5: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 599 | TBZ x0, 3, 6f |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 600 | |
| 601 | # Remainder- 2 floats of A (8 bytes) |
| 602 | # Load A |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 603 | LDR d0, [x3], 8 |
| 604 | LDR d1, [x9], 8 |
| 605 | LDR d2, [x10], 8 |
| 606 | LDR d3, [x11], 8 |
| 607 | LDR d4, [x12], 8 |
| 608 | LDR d5, [x4], 8 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 609 | # Load B |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 610 | LDP q12, q13, [x5], 32 |
| 611 | LDP q14, q15, [x5], 32 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 612 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 613 | FMLA v20.4s, v12.4s, v0.s[0] |
| 614 | FMLA v22.4s, v12.4s, v1.s[0] |
| 615 | FMLA v24.4s, v12.4s, v2.s[0] |
| 616 | FMLA v26.4s, v12.4s, v3.s[0] |
| 617 | FMLA v28.4s, v12.4s, v4.s[0] |
| 618 | FMLA v30.4s, v12.4s, v5.s[0] |
| 619 | FMLA v21.4s, v13.4s, v0.s[0] |
| 620 | FMLA v23.4s, v13.4s, v1.s[0] |
| 621 | FMLA v25.4s, v13.4s, v2.s[0] |
| 622 | FMLA v27.4s, v13.4s, v3.s[0] |
| 623 | FMLA v29.4s, v13.4s, v4.s[0] |
| 624 | FMLA v31.4s, v13.4s, v5.s[0] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 625 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 626 | FMLA v20.4s, v14.4s, v0.s[1] |
| 627 | FMLA v22.4s, v14.4s, v1.s[1] |
| 628 | FMLA v24.4s, v14.4s, v2.s[1] |
| 629 | FMLA v26.4s, v14.4s, v3.s[1] |
| 630 | FMLA v28.4s, v14.4s, v4.s[1] |
| 631 | FMLA v30.4s, v14.4s, v5.s[1] |
| 632 | FMLA v21.4s, v15.4s, v0.s[1] |
| 633 | FMLA v23.4s, v15.4s, v1.s[1] |
| 634 | FMLA v25.4s, v15.4s, v2.s[1] |
| 635 | FMLA v27.4s, v15.4s, v3.s[1] |
| 636 | FMLA v29.4s, v15.4s, v4.s[1] |
| 637 | FMLA v31.4s, v15.4s, v5.s[1] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 638 | |
| 639 | # Is there a remainder?- 1 float of A (4 bytes) |
| 640 | 6: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 641 | TBZ x0, 2, 3b |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 642 | |
| 643 | # Remainder- 1 float of A (4 bytes) |
| 644 | # Load A |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 645 | LDR s0, [x3], 4 |
| 646 | LDR s1, [x9], 4 |
| 647 | LDR s2, [x10], 4 |
| 648 | LDR s3, [x11], 4 |
| 649 | LDR s4, [x12], 4 |
| 650 | LDR s5, [x4], 4 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 651 | # Load B |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 652 | LDP q12, q13, [x5], 32 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 653 | |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 654 | FMLA v20.4s, v12.4s, v0.s[0] |
| 655 | FMLA v22.4s, v12.4s, v1.s[0] |
| 656 | FMLA v24.4s, v12.4s, v2.s[0] |
| 657 | FMLA v26.4s, v12.4s, v3.s[0] |
| 658 | FMLA v28.4s, v12.4s, v4.s[0] |
| 659 | FMLA v30.4s, v12.4s, v5.s[0] |
| 660 | FMLA v21.4s, v13.4s, v0.s[0] |
| 661 | FMLA v23.4s, v13.4s, v1.s[0] |
| 662 | FMLA v25.4s, v13.4s, v2.s[0] |
| 663 | FMLA v27.4s, v13.4s, v3.s[0] |
| 664 | FMLA v29.4s, v13.4s, v4.s[0] |
| 665 | FMLA v31.4s, v13.4s, v5.s[0] |
| 666 | B 3b |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 667 | |
| 668 | # Store odd width |
| 669 | 7: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 670 | TBZ x1, 2, 8f |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 671 | $if INC: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 672 | STR q30, [x7], 16 |
| 673 | MOV v30.16b, v31.16b |
| 674 | STR q28, [x13], 16 |
| 675 | MOV v28.16b, v29.16b |
| 676 | STR q26, [x14], 16 |
| 677 | MOV v26.16b, v27.16b |
| 678 | STR q24, [x17], 16 |
| 679 | MOV v24.16b, v25.16b |
| 680 | STR q22, [x16], 16 |
| 681 | MOV v22.16b, v23.16b |
| 682 | STR q20, [x6], 16 |
| 683 | MOV v20.16b, v21.16b |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 684 | $else: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 685 | STR q20, [x6], 16 |
| 686 | MOV v20.16b, v21.16b |
| 687 | STR q22, [x16], 16 |
| 688 | MOV v22.16b, v23.16b |
| 689 | STR q24, [x17], 16 |
| 690 | MOV v24.16b, v25.16b |
| 691 | STR q26, [x14], 16 |
| 692 | MOV v26.16b, v27.16b |
| 693 | STR q28, [x13], 16 |
| 694 | MOV v28.16b, v29.16b |
| 695 | STR q30, [x7], 16 |
| 696 | MOV v30.16b, v31.16b |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 697 | 8: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 698 | TBZ x1, 1, 9f |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 699 | $if INC: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 700 | STR d30, [x7], 8 |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 701 | STR d28, [x13], 8 |
Frank Barchard | b7a7c30 | 2021-09-01 16:24:25 -0700 | [diff] [blame] | 702 | DUP d30, v30.d[1] |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 703 | DUP d28, v28.d[1] |
| 704 | STR d26, [x14], 8 |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 705 | STR d24, [x17], 8 |
Frank Barchard | b7a7c30 | 2021-09-01 16:24:25 -0700 | [diff] [blame] | 706 | DUP d26, v26.d[1] |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 707 | DUP d24, v24.d[1] |
| 708 | STR d22, [x16], 8 |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 709 | STR d20, [x6], 8 |
Frank Barchard | b7a7c30 | 2021-09-01 16:24:25 -0700 | [diff] [blame] | 710 | DUP d22, v22.d[1] |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 711 | DUP d20, v20.d[1] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 712 | $else: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 713 | STR d20, [x6], 8 |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 714 | STR d22, [x16], 8 |
Frank Barchard | b7a7c30 | 2021-09-01 16:24:25 -0700 | [diff] [blame] | 715 | DUP d20, v20.d[1] |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 716 | DUP d22, v22.d[1] |
| 717 | STR d24, [x17], 8 |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 718 | STR d26, [x14], 8 |
Frank Barchard | b7a7c30 | 2021-09-01 16:24:25 -0700 | [diff] [blame] | 719 | DUP d24, v24.d[1] |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 720 | DUP d26, v26.d[1] |
| 721 | STR d28, [x13], 8 |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 722 | STR d30, [x7], 8 |
Frank Barchard | b7a7c30 | 2021-09-01 16:24:25 -0700 | [diff] [blame] | 723 | DUP d28, v28.d[1] |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 724 | DUP d30, v30.d[1] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 725 | |
| 726 | 9: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 727 | TBZ x1, 0, 10f |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 728 | $if INC: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 729 | STR s30, [x7] |
| 730 | STR s28, [x13] |
| 731 | STR s26, [x14] |
| 732 | STR s24, [x17] |
| 733 | STR s22, [x16] |
| 734 | STR s20, [x6] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 735 | $else: |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 736 | STR s20, [x6] |
| 737 | STR s22, [x16] |
| 738 | STR s24, [x17] |
| 739 | STR s26, [x14] |
| 740 | STR s28, [x13] |
| 741 | STR s30, [x7] |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 742 | 10: |
| 743 | # Restore d8-d15 from stack |
Frank Barchard | 76f43f0 | 2021-05-12 14:52:01 -0700 | [diff] [blame] | 744 | LDP d14, d15, [sp, 48] |
| 745 | LDP d12, d13, [sp, 32] |
| 746 | LDP d10, d11, [sp, 16] |
| 747 | LDP d8, d9, [sp], 64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 748 | RET |
| 749 | |
Frank Barchard | a2f1891 | 2021-12-28 14:17:05 -0800 | [diff] [blame] | 750 | END_FUNCTION xnn_f32_gemm${"inc" if INC else ""}_minmax_ukernel_6x8__aarch64_neonfma${"_prfm" if PREFETCH else ""}_cortex_a75 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 751 | |
| 752 | #ifdef __ELF__ |
| 753 | .section ".note.GNU-stack","",%progbits |
| 754 | #endif |