| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // | ||
| 2 | // SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com> | ||
| 3 | // | ||
| 4 | // SPDX-License-Identifier: Apache-2.0 | ||
| 5 | // | ||
| 6 | |||
| 7 | // Do not flag up inline assembly blocks | ||
| 8 | #pragma GCC diagnostic ignored "-Woverlength-strings" | ||
| 9 | |||
| 10 | #if !defined(__aarch64__) || !defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) || \ | ||
| 11 | !defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) | ||
| 12 | #error This file must be compiled for AArch64, FEAT_BF16, FEAT_FP16. | ||
| 13 | #else // Architectural features check. | ||
| 14 | |||
| 15 | #include "kai_rhs_pack_kxn_bf16p12x4biasf16_f16_neon.h" | ||
| 16 | |||
| 17 | #include <stddef.h> | ||
| 18 | #include <stdint.h> | ||
| 19 | #include <string.h> | ||
| 20 | |||
| 21 | #include "kai/kai_common.h" | ||
| 22 | |||
| 23 | static const size_t kai_nr = 12; | ||
| 24 | static const size_t kai_kr = 4; | ||
| 25 | static const size_t kai_num_bytes_input = 2; | ||
| 26 | static const size_t kai_num_bytes_output = 2; | ||
| 27 | static const size_t kai_num_bytes_bias = 2; | ||
| 28 | |||
| 29 | ✗ | size_t kai_get_n_step_rhs_pack_kxn_bf16p12x4biasf16_f16_neon(void) { | |
| 30 | ✗ | return kai_nr; | |
| 31 | } | ||
| 32 | |||
| 33 | 84 | size_t kai_get_rhs_offset_rhs_pack_kxn_bf16p12x4biasf16_f16_neon(size_t n_idx) { | |
| 34 | − | KAI_ASSUME(n_idx % kai_nr == 0); | |
| 35 | |||
| 36 | 84 | return n_idx * kai_num_bytes_input; | |
| 37 | } | ||
| 38 | |||
| 39 | ✗ | size_t kai_get_bias_offset_rhs_pack_kxn_bf16p12x4biasf16_f16_neon(size_t n_idx) { | |
| 40 | ✗ | return n_idx * kai_num_bytes_bias; | |
| 41 | } | ||
| 42 | |||
| 43 | 102 | size_t kai_get_rhs_packed_stride_rhs_pack_kxn_bf16p12x4biasf16_f16_neon(size_t k) { | |
| 44 | 102 | return kai_nr * (kai_num_bytes_bias + kai_roundup(k, kai_kr) * kai_num_bytes_output); | |
| 45 | } | ||
| 46 | |||
| 47 | 102 | size_t kai_get_rhs_packed_offset_rhs_pack_kxn_bf16p12x4biasf16_f16_neon(size_t n_idx, size_t k) { | |
| 48 | − | KAI_ASSUME(n_idx % kai_nr == 0); | |
| 49 | |||
| 50 | 102 | return n_idx * (kai_num_bytes_bias + kai_roundup(k, kai_kr) * kai_num_bytes_output); | |
| 51 | } | ||
| 52 | |||
| 53 | 102 | size_t kai_get_rhs_packed_size_rhs_pack_kxn_bf16p12x4biasf16_f16_neon(size_t n, size_t k) { | |
| 54 | 102 | return kai_get_rhs_packed_offset_rhs_pack_kxn_bf16p12x4biasf16_f16_neon(kai_roundup(n, kai_nr), k); | |
| 55 | } | ||
| 56 | |||
| 57 | 102 | void kai_run_rhs_pack_kxn_bf16p12x4biasf16_f16_neon( | |
| 58 | size_t num_groups, size_t n, size_t k, size_t nr, size_t kr, size_t sr, size_t rhs_stride, const void* rhs, | ||
| 59 | const void* bias, const void* scale, void* rhs_packed, size_t extra_bytes, const void* params) { | ||
| 60 | − | KAI_ASSUME(num_groups == 1); | |
| 61 | − | KAI_ASSUME(nr == kai_nr); | |
| 62 | − | KAI_ASSUME(kr == kai_kr); | |
| 63 | − | KAI_ASSUME(sr == 1); | |
| 64 | − | KAI_ASSUME(rhs != NULL); | |
| 65 | − | KAI_ASSUME(scale == NULL); | |
| 66 | − | KAI_ASSUME(rhs_packed != NULL); | |
| 67 | − | KAI_ASSUME(extra_bytes == 0); | |
| 68 | − | KAI_ASSUME(params == NULL); | |
| 69 | |||
| 70 | 102 | size_t height = k; | |
| 71 | 102 | const size_t width = n; | |
| 72 | 102 | const void* in = rhs; | |
| 73 | 102 | void* out = rhs_packed; | |
| 74 | 102 | const size_t in_stride = rhs_stride; | |
| 75 | 102 | const uint16_t* pad_row = rhs; | |
| 76 | |||
| 77 | // Fill zeros if bias is nullptr | ||
| 78 | 102 | size_t bias_step = nr * sizeof(uint16_t); | |
| 79 | 102 | uint8_t zero_bias[kai_nr * sizeof(uint16_t)]; | |
| 80 | |||
| 81 |
2/2✓ Branch 0 taken 51 times.
✓ Branch 1 taken 51 times.
|
102 | if (bias == NULL) { |
| 82 | 51 | memset(zero_bias, 0, kai_nr * sizeof(uint16_t)); | |
| 83 | 51 | bias_step = 0; | |
| 84 | 51 | } | |
| 85 | |||
| 86 |
2/2✓ Branch 0 taken 51 times.
✓ Branch 1 taken 51 times.
|
102 | const void* bias_ptr = bias == NULL ? (const void*)zero_bias : bias; |
| 87 | |||
| 88 | 102 | size_t out_stride = kai_get_rhs_packed_stride_rhs_pack_kxn_bf16p12x4biasf16_f16_neon(height); | |
| 89 | |||
| 90 | 204 | __asm__ __volatile__( | |
| 91 | "mov x22, %x[width]\n" | ||
| 92 | "mov x21, %x[out]\n" | ||
| 93 | "cmp x22, #0xc\n" | ||
| 94 | "blt 2f\n" | ||
| 95 | "1:" // Bias: Full loop | ||
| 96 | "ldr q17, [%x[bias], #0x0]\n" | ||
| 97 | "ldr d16, [%x[bias], #0x10]\n" | ||
| 98 | "sub x22, x22, #0xc\n" | ||
| 99 | "add %x[bias], %x[bias], %x[bias_step]\n" | ||
| 100 | "cmp x22, #0xc\n" | ||
| 101 | "str q17, [x21, #0x0]\n" | ||
| 102 | "str d16, [x21, #0x10]\n" | ||
| 103 | "add x21, x21, %x[out_stride]\n" | ||
| 104 | "bge 1b\n" | ||
| 105 | "cbz x22, 3f\n" | ||
| 106 | "2:" // Bias: Tail loop | ||
| 107 | "ldr h20, [%x[bias], #0x0]\n" | ||
| 108 | "sub x22, x22, #0x1\n" | ||
| 109 | "add %x[bias], %x[bias], #0x2\n" | ||
| 110 | "cmp x22, #0x0\n" | ||
| 111 | "str h20, [x21]\n" | ||
| 112 | "add x21, x21, #0x2\n" | ||
| 113 | "bgt 2b\n" | ||
| 114 | "3:" // Bias: Done | ||
| 115 | "cmp %x[height], #0x8\n" | ||
| 116 | "add %x[out], %x[out], #0x18\n" | ||
| 117 | "blt 12f\n" | ||
| 118 | "4:" // Main row loop: Head | ||
| 119 | "mov x9, %x[in]\n" | ||
| 120 | "mov x28, %x[width]\n" | ||
| 121 | "mov x27, %x[out]\n" | ||
| 122 | "sub %x[height], %x[height], #0x8\n" | ||
| 123 | "add x26, x9, %x[in_stride]\n" | ||
| 124 | "add x25, x26, %x[in_stride]\n" | ||
| 125 | "add x24, x25, %x[in_stride]\n" | ||
| 126 | "cmp x28, #0xc\n" | ||
| 127 | "add x23, x24, %x[in_stride]\n" | ||
| 128 | "add x22, x23, %x[in_stride]\n" | ||
| 129 | "add x21, x22, %x[in_stride]\n" | ||
| 130 | "add x20, x21, %x[in_stride]\n" | ||
| 131 | "add %x[in], x20, %x[in_stride]\n" | ||
| 132 | "blt 6f\n" | ||
| 133 | "5:" // Main row loop: Column loop | ||
| 134 | "ldr q23, [x9], #0x10\n" | ||
| 135 | "ldr q22, [x26], #0x10\n" | ||
| 136 | "sub x28, x28, #0xc\n" | ||
| 137 | "ldr q17, [x25], #0x10\n" | ||
| 138 | "ldr q16, [x24], #0x10\n" | ||
| 139 | "cmp x28, #0xc\n" | ||
| 140 | "ldr q1, [x23], #0x10\n" | ||
| 141 | "ldr q0, [x22], #0x10\n" | ||
| 142 | "ldr q21, [x21], #0x10\n" | ||
| 143 | "ldr q20, [x20], #0x10\n" | ||
| 144 | "ldr d31, [x9], #0x8\n" | ||
| 145 | "ldr d30, [x26], #0x8\n" | ||
| 146 | "zip1 v29.8h, v23.8h, v17.8h\n" | ||
| 147 | "zip1 v28.8h, v22.8h, v16.8h\n" | ||
| 148 | "ldr d19, [x25], #0x8\n" | ||
| 149 | "ldr d18, [x24], #0x8\n" | ||
| 150 | "zip2 v27.8h, v23.8h, v17.8h\n" | ||
| 151 | "zip2 v26.8h, v22.8h, v16.8h\n" | ||
| 152 | "ldr d25, [x23], #0x8\n" | ||
| 153 | "ldr d24, [x22], #0x8\n" | ||
| 154 | "zip1 v23.8h, v1.8h, v21.8h\n" | ||
| 155 | "zip1 v22.8h, v0.8h, v20.8h\n" | ||
| 156 | "ldr d17, [x21], #0x8\n" | ||
| 157 | "ldr d16, [x20], #0x8\n" | ||
| 158 | "zip2 v21.8h, v1.8h, v21.8h\n" | ||
| 159 | "zip2 v20.8h, v0.8h, v20.8h\n" | ||
| 160 | "zip1 v19.8h, v31.8h, v19.8h\n" | ||
| 161 | "zip1 v18.8h, v30.8h, v18.8h\n" | ||
| 162 | "zip1 v1.8h, v29.8h, v28.8h\n" | ||
| 163 | "zip2 v0.8h, v29.8h, v28.8h\n" | ||
| 164 | "zip1 v17.8h, v25.8h, v17.8h\n" | ||
| 165 | "zip1 v16.8h, v24.8h, v16.8h\n" | ||
| 166 | "zip1 v31.8h, v27.8h, v26.8h\n" | ||
| 167 | "zip2 v30.8h, v27.8h, v26.8h\n" | ||
| 168 | "zip1 v29.8h, v19.8h, v18.8h\n" | ||
| 169 | "zip2 v28.8h, v19.8h, v18.8h\n" | ||
| 170 | "zip1 v13.8h, v23.8h, v22.8h\n" | ||
| 171 | "zip2 v12.8h, v23.8h, v22.8h\n" | ||
| 172 | "zip1 v11.8h, v21.8h, v20.8h\n" | ||
| 173 | "zip2 v10.8h, v21.8h, v20.8h\n" | ||
| 174 | "zip1 v9.8h, v17.8h, v16.8h\n" | ||
| 175 | "zip2 v8.8h, v17.8h, v16.8h\n" | ||
| 176 | "fcvtl v27.4s, v1.4h\n" | ||
| 177 | "fcvtl v26.4s, v0.4h\n" | ||
| 178 | "fcvtl v25.4s, v31.4h\n" | ||
| 179 | "fcvtl v24.4s, v30.4h\n" | ||
| 180 | "fcvtl v23.4s, v29.4h\n" | ||
| 181 | "fcvtl v22.4s, v28.4h\n" | ||
| 182 | "fcvtl v21.4s, v13.4h\n" | ||
| 183 | "fcvtl v20.4s, v12.4h\n" | ||
| 184 | "fcvtl v19.4s, v11.4h\n" | ||
| 185 | "fcvtl v18.4s, v10.4h\n" | ||
| 186 | "fcvtl v17.4s, v9.4h\n" | ||
| 187 | "fcvtl v16.4s, v8.4h\n" | ||
| 188 | "fcvtl2 v7.4s, v1.8h\n" | ||
| 189 | ".inst 0x0ea16b66 // bfcvtn v6.4h, v27.4s\n" | ||
| 190 | "fcvtl2 v5.4s, v0.8h\n" | ||
| 191 | ".inst 0x0ea16b44 // bfcvtn v4.4h, v26.4s\n" | ||
| 192 | "fcvtl2 v3.4s, v31.8h\n" | ||
| 193 | ".inst 0x0ea16b22 // bfcvtn v2.4h, v25.4s\n" | ||
| 194 | "fcvtl2 v1.4s, v30.8h\n" | ||
| 195 | ".inst 0x0ea16b00 // bfcvtn v0.4h, v24.4s\n" | ||
| 196 | "fcvtl2 v31.4s, v29.8h\n" | ||
| 197 | ".inst 0x0ea16afe // bfcvtn v30.4h, v23.4s\n" | ||
| 198 | "fcvtl2 v29.4s, v28.8h\n" | ||
| 199 | ".inst 0x0ea16adc // bfcvtn v28.4h, v22.4s\n" | ||
| 200 | "fcvtl2 v27.4s, v13.8h\n" | ||
| 201 | ".inst 0x0ea16aba // bfcvtn v26.4h, v21.4s\n" | ||
| 202 | "fcvtl2 v25.4s, v12.8h\n" | ||
| 203 | ".inst 0x0ea16a98 // bfcvtn v24.4h, v20.4s\n" | ||
| 204 | "fcvtl2 v23.4s, v11.8h\n" | ||
| 205 | ".inst 0x0ea16a76 // bfcvtn v22.4h, v19.4s\n" | ||
| 206 | "fcvtl2 v21.4s, v10.8h\n" | ||
| 207 | ".inst 0x0ea16a54 // bfcvtn v20.4h, v18.4s\n" | ||
| 208 | "fcvtl2 v19.4s, v9.8h\n" | ||
| 209 | ".inst 0x0ea16a32 // bfcvtn v18.4h, v17.4s\n" | ||
| 210 | "fcvtl2 v17.4s, v8.8h\n" | ||
| 211 | ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n" | ||
| 212 | ".inst 0x4ea168e6 // bfcvtn2 v6.8h, v7.4s\n" | ||
| 213 | ".inst 0x4ea168a4 // bfcvtn2 v4.8h, v5.4s\n" | ||
| 214 | ".inst 0x4ea16862 // bfcvtn2 v2.8h, v3.4s\n" | ||
| 215 | ".inst 0x4ea16820 // bfcvtn2 v0.8h, v1.4s\n" | ||
| 216 | ".inst 0x4ea16bfe // bfcvtn2 v30.8h, v31.4s\n" | ||
| 217 | ".inst 0x4ea16bbc // bfcvtn2 v28.8h, v29.4s\n" | ||
| 218 | ".inst 0x4ea16b7a // bfcvtn2 v26.8h, v27.4s\n" | ||
| 219 | ".inst 0x4ea16b38 // bfcvtn2 v24.8h, v25.4s\n" | ||
| 220 | "str q6, [x27, #0x0]\n" | ||
| 221 | ".inst 0x4ea16af6 // bfcvtn2 v22.8h, v23.4s\n" | ||
| 222 | ".inst 0x4ea16ab4 // bfcvtn2 v20.8h, v21.4s\n" | ||
| 223 | "str q4, [x27, #0x10]\n" | ||
| 224 | ".inst 0x4ea16a72 // bfcvtn2 v18.8h, v19.4s\n" | ||
| 225 | ".inst 0x4ea16a30 // bfcvtn2 v16.8h, v17.4s\n" | ||
| 226 | "str q2, [x27, #0x20]\n" | ||
| 227 | "str q0, [x27, #0x30]\n" | ||
| 228 | "str q30, [x27, #0x40]\n" | ||
| 229 | "str q28, [x27, #0x50]\n" | ||
| 230 | "str q26, [x27, #0x60]\n" | ||
| 231 | "str q24, [x27, #0x70]\n" | ||
| 232 | "str q22, [x27, #0x80]\n" | ||
| 233 | "str q20, [x27, #0x90]\n" | ||
| 234 | "str q18, [x27, #0xa0]\n" | ||
| 235 | "str q16, [x27, #0xb0]\n" | ||
| 236 | "add x27, x27, %x[out_stride]\n" | ||
| 237 | "bge 5b\n" | ||
| 238 | "6:" // Main row loop: Column loop skip | ||
| 239 | "cbz x28, 11f\n" | ||
| 240 | "cmp x28, #0x4\n" | ||
| 241 | "movi v16.16b, #0x0\n" | ||
| 242 | "str q16, [x27, #0x0]\n" | ||
| 243 | "str q16, [x27, #0x10]\n" | ||
| 244 | "str q16, [x27, #0x20]\n" | ||
| 245 | "str q16, [x27, #0x30]\n" | ||
| 246 | "str q16, [x27, #0x40]\n" | ||
| 247 | "str q16, [x27, #0x50]\n" | ||
| 248 | "str q16, [x27, #0x60]\n" | ||
| 249 | "str q16, [x27, #0x70]\n" | ||
| 250 | "str q16, [x27, #0x80]\n" | ||
| 251 | "str q16, [x27, #0x90]\n" | ||
| 252 | "str q16, [x27, #0xa0]\n" | ||
| 253 | "str q16, [x27, #0xb0]\n" | ||
| 254 | "blt 8f\n" | ||
| 255 | "7:" // Main row loop: width 4 loop: loop | ||
| 256 | "ldr d23, [x9], #0x8\n" | ||
| 257 | "ldr d22, [x26], #0x8\n" | ||
| 258 | "sub x28, x28, #0x4\n" | ||
| 259 | "ldr d20, [x25], #0x8\n" | ||
| 260 | "ldr d16, [x24], #0x8\n" | ||
| 261 | "cmp x28, #0x4\n" | ||
| 262 | "ldr d19, [x23], #0x8\n" | ||
| 263 | "ldr d21, [x22], #0x8\n" | ||
| 264 | "ldr d18, [x21], #0x8\n" | ||
| 265 | "ldr d17, [x20], #0x8\n" | ||
| 266 | "zip1 v20.8h, v23.8h, v20.8h\n" | ||
| 267 | "zip1 v16.8h, v22.8h, v16.8h\n" | ||
| 268 | "zip1 v19.8h, v19.8h, v18.8h\n" | ||
| 269 | "zip1 v18.8h, v21.8h, v17.8h\n" | ||
| 270 | "zip1 v17.8h, v20.8h, v16.8h\n" | ||
| 271 | "zip2 v16.8h, v20.8h, v16.8h\n" | ||
| 272 | "zip1 v25.8h, v19.8h, v18.8h\n" | ||
| 273 | "zip2 v24.8h, v19.8h, v18.8h\n" | ||
| 274 | "fcvtl v19.4s, v17.4h\n" | ||
| 275 | "fcvtl v18.4s, v16.4h\n" | ||
| 276 | "fcvtl2 v23.4s, v17.8h\n" | ||
| 277 | "fcvtl2 v22.4s, v16.8h\n" | ||
| 278 | "fcvtl v17.4s, v25.4h\n" | ||
| 279 | "fcvtl v16.4s, v24.4h\n" | ||
| 280 | ".inst 0x0ea16a75 // bfcvtn v21.4h, v19.4s\n" | ||
| 281 | ".inst 0x0ea16a54 // bfcvtn v20.4h, v18.4s\n" | ||
| 282 | "fcvtl2 v19.4s, v25.8h\n" | ||
| 283 | "fcvtl2 v18.4s, v24.8h\n" | ||
| 284 | ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n" | ||
| 285 | ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n" | ||
| 286 | ".inst 0x4ea16af5 // bfcvtn2 v21.8h, v23.4s\n" | ||
| 287 | ".inst 0x4ea16ad4 // bfcvtn2 v20.8h, v22.4s\n" | ||
| 288 | ".inst 0x4ea16a71 // bfcvtn2 v17.8h, v19.4s\n" | ||
| 289 | ".inst 0x4ea16a50 // bfcvtn2 v16.8h, v18.4s\n" | ||
| 290 | "str q21, [x27, #0x0]\n" | ||
| 291 | "str q20, [x27, #0x10]\n" | ||
| 292 | "str q17, [x27, #0x60]\n" | ||
| 293 | "str q16, [x27, #0x70]\n" | ||
| 294 | "add x27, x27, #0x20\n" | ||
| 295 | "bge 7b\n" | ||
| 296 | "8:" // Main row loop: width 4 loop: skip | ||
| 297 | "cmp x28, #0x1\n" | ||
| 298 | "blt 10f\n" | ||
| 299 | "9:" // Main row loop: width 1 loop: loop | ||
| 300 | "ldr h23, [x9], #0x2\n" | ||
| 301 | "ldr h22, [x26], #0x2\n" | ||
| 302 | "sub x28, x28, #0x1\n" | ||
| 303 | "ldr h19, [x25], #0x2\n" | ||
| 304 | "ldr h18, [x24], #0x2\n" | ||
| 305 | "cmp x28, #0x1\n" | ||
| 306 | "ldr h21, [x23], #0x2\n" | ||
| 307 | "ldr h20, [x22], #0x2\n" | ||
| 308 | "ldr h17, [x21], #0x2\n" | ||
| 309 | "ldr h16, [x20], #0x2\n" | ||
| 310 | "zip1 v19.8h, v23.8h, v19.8h\n" | ||
| 311 | "zip1 v18.8h, v22.8h, v18.8h\n" | ||
| 312 | "zip1 v17.8h, v21.8h, v17.8h\n" | ||
| 313 | "zip1 v16.8h, v20.8h, v16.8h\n" | ||
| 314 | "zip1 v19.8h, v19.8h, v18.8h\n" | ||
| 315 | "zip1 v18.8h, v17.8h, v16.8h\n" | ||
| 316 | "fcvtl v17.4s, v19.4h\n" | ||
| 317 | "fcvtl v16.4s, v18.4h\n" | ||
| 318 | "fcvtl2 v19.4s, v19.8h\n" | ||
| 319 | "fcvtl2 v18.4s, v18.8h\n" | ||
| 320 | ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n" | ||
| 321 | ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n" | ||
| 322 | ".inst 0x4ea16a71 // bfcvtn2 v17.8h, v19.4s\n" | ||
| 323 | ".inst 0x4ea16a50 // bfcvtn2 v16.8h, v18.4s\n" | ||
| 324 | "str d17, [x27, #0x0]\n" | ||
| 325 | "str d16, [x27, #0x60]\n" | ||
| 326 | "add x27, x27, #0x8\n" | ||
| 327 | "bge 9b\n" | ||
| 328 | "10:" // Main row loop: width 1 loop: skip | ||
| 329 | "11:" // Main row loop: odd col skip | ||
| 330 | "cmp %x[height], #0x8\n" | ||
| 331 | "add %x[out], %x[out], #0xc0\n" | ||
| 332 | "bge 4b\n" | ||
| 333 | "cbz %x[height], 21f\n" | ||
| 334 | "12:" // Main loop skip | ||
| 335 | "13:" // Tail row loop: Head | ||
| 336 | "mov x9, %x[in]\n" | ||
| 337 | "mov x20, %x[width]\n" | ||
| 338 | "cmp %x[height], #0x3\n" | ||
| 339 | "mov x27, %x[out]\n" | ||
| 340 | "add x26, x9, %x[in_stride]\n" | ||
| 341 | "add x25, x26, %x[in_stride]\n" | ||
| 342 | "add x24, x25, %x[in_stride]\n" | ||
| 343 | "csel x25, x25, %x[pad_row], GE\n" | ||
| 344 | "add %x[in], x24, %x[in_stride]\n" | ||
| 345 | "csel x24, x24, %x[pad_row], GT\n" | ||
| 346 | "cmp %x[height], #0x1\n" | ||
| 347 | "sub %x[height], %x[height], #0x4\n" | ||
| 348 | "csel x26, x26, %x[pad_row], GT\n" | ||
| 349 | "cmp x20, #0xc\n" | ||
| 350 | "blt 15f\n" | ||
| 351 | "14:" // Tail row loop: Column loop | ||
| 352 | "ldr q25, [x9], #0x10\n" | ||
| 353 | "ldr q24, [x26], #0x10\n" | ||
| 354 | "sub x20, x20, #0xc\n" | ||
| 355 | "ldr q19, [x25], #0x10\n" | ||
| 356 | "ldr q18, [x24], #0x10\n" | ||
| 357 | "cmp x20, #0xc\n" | ||
| 358 | "ldr d23, [x9], #0x8\n" | ||
| 359 | "ldr d22, [x26], #0x8\n" | ||
| 360 | "ldr d17, [x25], #0x8\n" | ||
| 361 | "ldr d16, [x24], #0x8\n" | ||
| 362 | "zip1 v21.8h, v25.8h, v19.8h\n" | ||
| 363 | "zip1 v20.8h, v24.8h, v18.8h\n" | ||
| 364 | "zip2 v19.8h, v25.8h, v19.8h\n" | ||
| 365 | "zip2 v18.8h, v24.8h, v18.8h\n" | ||
| 366 | "zip1 v17.8h, v23.8h, v17.8h\n" | ||
| 367 | "zip1 v16.8h, v22.8h, v16.8h\n" | ||
| 368 | "zip1 v24.8h, v21.8h, v20.8h\n" | ||
| 369 | "zip2 v23.8h, v21.8h, v20.8h\n" | ||
| 370 | "zip1 v22.8h, v19.8h, v18.8h\n" | ||
| 371 | "zip2 v30.8h, v19.8h, v18.8h\n" | ||
| 372 | "zip1 v29.8h, v17.8h, v16.8h\n" | ||
| 373 | "zip2 v28.8h, v17.8h, v16.8h\n" | ||
| 374 | "fcvtl v21.4s, v24.4h\n" | ||
| 375 | "fcvtl v20.4s, v23.4h\n" | ||
| 376 | "fcvtl v19.4s, v22.4h\n" | ||
| 377 | "fcvtl v18.4s, v30.4h\n" | ||
| 378 | "fcvtl v17.4s, v29.4h\n" | ||
| 379 | "fcvtl v16.4s, v28.4h\n" | ||
| 380 | "fcvtl2 v27.4s, v24.8h\n" | ||
| 381 | ".inst 0x0ea16aba // bfcvtn v26.4h, v21.4s\n" | ||
| 382 | "fcvtl2 v25.4s, v23.8h\n" | ||
| 383 | ".inst 0x0ea16a98 // bfcvtn v24.4h, v20.4s\n" | ||
| 384 | "fcvtl2 v23.4s, v22.8h\n" | ||
| 385 | ".inst 0x0ea16a76 // bfcvtn v22.4h, v19.4s\n" | ||
| 386 | "fcvtl2 v21.4s, v30.8h\n" | ||
| 387 | ".inst 0x0ea16a54 // bfcvtn v20.4h, v18.4s\n" | ||
| 388 | "fcvtl2 v19.4s, v29.8h\n" | ||
| 389 | ".inst 0x0ea16a32 // bfcvtn v18.4h, v17.4s\n" | ||
| 390 | "fcvtl2 v17.4s, v28.8h\n" | ||
| 391 | ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n" | ||
| 392 | ".inst 0x4ea16b7a // bfcvtn2 v26.8h, v27.4s\n" | ||
| 393 | ".inst 0x4ea16b38 // bfcvtn2 v24.8h, v25.4s\n" | ||
| 394 | ".inst 0x4ea16af6 // bfcvtn2 v22.8h, v23.4s\n" | ||
| 395 | ".inst 0x4ea16ab4 // bfcvtn2 v20.8h, v21.4s\n" | ||
| 396 | ".inst 0x4ea16a72 // bfcvtn2 v18.8h, v19.4s\n" | ||
| 397 | ".inst 0x4ea16a30 // bfcvtn2 v16.8h, v17.4s\n" | ||
| 398 | "str q26, [x27, #0x0]\n" | ||
| 399 | "str q24, [x27, #0x10]\n" | ||
| 400 | "str q22, [x27, #0x20]\n" | ||
| 401 | "str q20, [x27, #0x30]\n" | ||
| 402 | "str q18, [x27, #0x40]\n" | ||
| 403 | "str q16, [x27, #0x50]\n" | ||
| 404 | "add x27, x27, %x[out_stride]\n" | ||
| 405 | "bge 14b\n" | ||
| 406 | "15:" // Tail row loop: Column loop skip | ||
| 407 | "cbz x20, 20f\n" | ||
| 408 | "cmp x20, #0x4\n" | ||
| 409 | "movi v16.16b, #0x0\n" | ||
| 410 | "str q16, [x27, #0x0]\n" | ||
| 411 | "str q16, [x27, #0x10]\n" | ||
| 412 | "str q16, [x27, #0x20]\n" | ||
| 413 | "str q16, [x27, #0x30]\n" | ||
| 414 | "str q16, [x27, #0x40]\n" | ||
| 415 | "str q16, [x27, #0x50]\n" | ||
| 416 | "blt 17f\n" | ||
| 417 | "16:" // Tail row loop: width 4 loop: loop | ||
| 418 | "ldr d18, [x9], #0x8\n" | ||
| 419 | "ldr d19, [x26], #0x8\n" | ||
| 420 | "sub x20, x20, #0x4\n" | ||
| 421 | "ldr d17, [x25], #0x8\n" | ||
| 422 | "ldr d16, [x24], #0x8\n" | ||
| 423 | "cmp x20, #0x4\n" | ||
| 424 | "zip1 v18.8h, v18.8h, v17.8h\n" | ||
| 425 | "zip1 v17.8h, v19.8h, v16.8h\n" | ||
| 426 | "zip1 v16.8h, v18.8h, v17.8h\n" | ||
| 427 | "zip2 v20.8h, v18.8h, v17.8h\n" | ||
| 428 | "fcvtl v17.4s, v16.4h\n" | ||
| 429 | "fcvtl2 v19.4s, v16.8h\n" | ||
| 430 | "fcvtl v16.4s, v20.4h\n" | ||
| 431 | ".inst 0x0ea16a32 // bfcvtn v18.4h, v17.4s\n" | ||
| 432 | "fcvtl2 v17.4s, v20.8h\n" | ||
| 433 | ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n" | ||
| 434 | ".inst 0x4ea16a72 // bfcvtn2 v18.8h, v19.4s\n" | ||
| 435 | ".inst 0x4ea16a30 // bfcvtn2 v16.8h, v17.4s\n" | ||
| 436 | "str q18, [x27, #0x0]\n" | ||
| 437 | "str q16, [x27, #0x10]\n" | ||
| 438 | "add x27, x27, #0x20\n" | ||
| 439 | "bge 16b\n" | ||
| 440 | "17:" // Tail row loop: width 4 loop: skip | ||
| 441 | "cmp x20, #0x1\n" | ||
| 442 | "blt 19f\n" | ||
| 443 | "18:" // Tail row loop: width 1 loop: loop | ||
| 444 | "ldr h19, [x9], #0x2\n" | ||
| 445 | "ldr h18, [x26], #0x2\n" | ||
| 446 | "sub x20, x20, #0x1\n" | ||
| 447 | "ldr h17, [x25], #0x2\n" | ||
| 448 | "ldr h16, [x24], #0x2\n" | ||
| 449 | "cmp x20, #0x1\n" | ||
| 450 | "zip1 v17.8h, v19.8h, v17.8h\n" | ||
| 451 | "zip1 v16.8h, v18.8h, v16.8h\n" | ||
| 452 | "zip1 v17.8h, v17.8h, v16.8h\n" | ||
| 453 | "fcvtl v16.4s, v17.4h\n" | ||
| 454 | "fcvtl2 v17.4s, v17.8h\n" | ||
| 455 | ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n" | ||
| 456 | ".inst 0x4ea16a30 // bfcvtn2 v16.8h, v17.4s\n" | ||
| 457 | "str d16, [x27, #0x0]\n" | ||
| 458 | "add x27, x27, #0x8\n" | ||
| 459 | "bge 18b\n" | ||
| 460 | "19:" // Tail row loop: width 1 loop: skip | ||
| 461 | "20:" // Tail row loop: odd col skip | ||
| 462 | "cmp %x[height], #0x1\n" | ||
| 463 | "add %x[out], %x[out], #0x60\n" | ||
| 464 | "bge 13b\n" | ||
| 465 | "21:" // Done | ||
| 466 | : [bias] "+&r"(bias_ptr), [height] "+&r"(height), [in] "+&r"(in), [out] "+&r"(out) | ||
| 467 | 102 | : [bias_step] "r"(bias_step), [in_stride] "r"(in_stride), [out_stride] "r"(out_stride), [pad_row] "r"(pad_row), | |
| 468 | 102 | [width] "r"(width) | |
| 469 | : "cc", "memory", "v0", "v1", "v10", "v11", "v12", "v13", "v16", "v17", "v18", "v19", "v2", "v20", "v21", "v22", | ||
| 470 | "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v3", "v30", "v31", "v4", "v5", "v6", "v7", "v8", "v9", | ||
| 471 | "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "x9"); | ||
| 472 | 102 | } | |
| 473 | |||
| 474 | #endif // Architectural features check. | ||
| 475 |