kai/ukernels/matmul/pack/kai_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon.c
| 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 | #error This file must be compiled for AArch64, FEAT_BF16. | ||
| 12 | #else // Architectural features check. | ||
| 13 | |||
| 14 | #define MAX_NR 12 | ||
| 15 | |||
| 16 | #include "kai_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon.h" | ||
| 17 | |||
| 18 | #include <arm_neon.h> | ||
| 19 | #include <stddef.h> | ||
| 20 | #include <stdint.h> | ||
| 21 | #include <string.h> | ||
| 22 | |||
| 23 | #include "kai/kai_common.h" | ||
| 24 | |||
| 25 | static const size_t kai_nr = 12; | ||
| 26 | static const size_t kai_kr = 4; | ||
| 27 | static const size_t kai_sr = 1; | ||
| 28 | |||
| 29 | ✗ | size_t kai_get_n_step_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon(void) { | |
| 30 | ✗ | return kai_nr; | |
| 31 | } | ||
| 32 | |||
| 33 | 1176 | size_t kai_get_rhs_offset_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon(size_t n_idx) { | |
| 34 | − | KAI_ASSUME(n_idx % kai_nr == 0); | |
| 35 | 1176 | return n_idx * sizeof(float); | |
| 36 | } | ||
| 37 | |||
| 38 | ✗ | size_t kai_get_bias_offset_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon(size_t n_idx) { | |
| 39 | − | KAI_ASSUME(n_idx % kai_nr == 0); | |
| 40 | ✗ | return n_idx * sizeof(uint32_t); | |
| 41 | } | ||
| 42 | |||
| 43 | 1440 | size_t kai_get_rhs_packed_offset_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon( | |
| 44 | size_t n_idx, size_t k, size_t nr, size_t kr) { | ||
| 45 | − | KAI_ASSUME(n_idx % nr == 0); | |
| 46 | − | KAI_ASSUME(kai_nr == nr); | |
| 47 | − | KAI_ASSUME(kai_kr == kr); | |
| 48 | |||
| 49 | 1440 | return n_idx * (sizeof(uint32_t) + kai_roundup(k, kr) * sizeof(uint16_t)); | |
| 50 | } | ||
| 51 | |||
| 52 | 1440 | size_t kai_get_rhs_packed_size_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon(size_t n, size_t k, size_t nr, size_t kr) { | |
| 53 | 1440 | return kai_get_rhs_packed_offset_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon(kai_roundup(n, nr), k, nr, kr); | |
| 54 | } | ||
| 55 | |||
| 56 | 1440 | void kai_run_rhs_quant_pack_kxn_bf16p12x4biasf32_f32_neon( | |
| 57 | 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, | ||
| 58 | const void* bias, const void* scale, void* rhs_packed, size_t extra_bytes, const void* params) { | ||
| 59 | − | KAI_ASSUME(num_groups == 1); | |
| 60 | − | KAI_ASSUME(kai_nr == nr); | |
| 61 | − | KAI_ASSUME(kai_kr == kr); | |
| 62 | − | KAI_ASSUME(kai_sr == sr); | |
| 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 | − | KAI_ASSUME(nr <= MAX_NR); | |
| 70 | |||
| 71 | 1440 | size_t height = k; | |
| 72 | 1440 | const size_t width = n; | |
| 73 | 1440 | const void* in = rhs; | |
| 74 | 1440 | void* out = rhs_packed; | |
| 75 | 1440 | const size_t in_stride = rhs_stride; | |
| 76 | 1440 | const float* pad_row = rhs; | |
| 77 | |||
| 78 | // Fill zeros if bias is nullptr | ||
| 79 | 1440 | size_t bias_step = nr * sizeof(float); | |
| 80 | 1440 | uint8_t zero_bias[MAX_NR * sizeof(float)]; | |
| 81 | |||
| 82 |
2/2✓ Branch 0 taken 720 times.
✓ Branch 1 taken 720 times.
|
1440 | if (bias == NULL) { |
| 83 | 720 | memset(zero_bias, 0, MAX_NR * sizeof(float)); | |
| 84 | 720 | bias_step = 0; | |
| 85 | 720 | } | |
| 86 | |||
| 87 |
2/2✓ Branch 0 taken 720 times.
✓ Branch 1 taken 720 times.
|
1440 | const void* bias_ptr = bias == NULL ? (const void*)zero_bias : bias; |
| 88 | |||
| 89 | 1440 | const size_t out_stride = nr * kai_roundup(height, kr) * sizeof(uint16_t) + nr * sizeof(uint32_t); | |
| 90 | |||
| 91 | 2880 | __asm__ __volatile__( | |
| 92 | "mov x22, %x[width]\n" | ||
| 93 | "mov x21, %x[out]\n" | ||
| 94 | "cmp x22, #0xc\n" | ||
| 95 | "blt 2f\n" | ||
| 96 | "1:" // Bias: Full loop | ||
| 97 | "ldr q16, [%x[bias], #0x0]\n" | ||
| 98 | "ldr q26, [%x[bias], #0x10]\n" | ||
| 99 | "sub x22, x22, #0xc\n" | ||
| 100 | "ldr q8, [%x[bias], #0x20]\n" | ||
| 101 | "cmp x22, #0xc\n" | ||
| 102 | "add %x[bias], %x[bias], %x[bias_step]\n" | ||
| 103 | "str q16, [x21, #0x0]\n" | ||
| 104 | "str q26, [x21, #0x10]\n" | ||
| 105 | "str q8, [x21, #0x20]\n" | ||
| 106 | "add x21, x21, %x[out_stride]\n" | ||
| 107 | "bge 1b\n" | ||
| 108 | "cbz x22, 3f\n" | ||
| 109 | "2:" // Bias: Tail loop | ||
| 110 | "ldr w20, [%x[bias], #0x0]\n" | ||
| 111 | "sub x22, x22, #0x1\n" | ||
| 112 | "add %x[bias], %x[bias], #0x4\n" | ||
| 113 | "cmp x22, #0x0\n" | ||
| 114 | "str w20, [x21]\n" | ||
| 115 | "add x21, x21, #0x4\n" | ||
| 116 | "bgt 2b\n" | ||
| 117 | "3:" // Bias: Done | ||
| 118 | "cmp %x[height], #0x8\n" | ||
| 119 | "add %x[out], %x[out], #0x30\n" | ||
| 120 | "blt 12f\n" | ||
| 121 | "4:" // Main row loop: Head | ||
| 122 | "mov x9, %x[in]\n" | ||
| 123 | "mov x28, %x[width]\n" | ||
| 124 | "mov x27, %x[out]\n" | ||
| 125 | "sub %x[height], %x[height], #0x8\n" | ||
| 126 | "add x26, x9, %x[in_stride]\n" | ||
| 127 | "add x25, x26, %x[in_stride]\n" | ||
| 128 | "add x24, x25, %x[in_stride]\n" | ||
| 129 | "cmp x28, #0xc\n" | ||
| 130 | "add x23, x24, %x[in_stride]\n" | ||
| 131 | "add x22, x23, %x[in_stride]\n" | ||
| 132 | "add x21, x22, %x[in_stride]\n" | ||
| 133 | "add x20, x21, %x[in_stride]\n" | ||
| 134 | "add %x[in], x20, %x[in_stride]\n" | ||
| 135 | "blt 6f\n" | ||
| 136 | "5:" // Main row loop: Column loop | ||
| 137 | "ldr q28, [x9], #0x10\n" | ||
| 138 | "ldr q27, [x26], #0x10\n" | ||
| 139 | "sub x28, x28, #0xc\n" | ||
| 140 | "ldr q11, [x25], #0x10\n" | ||
| 141 | "ldr q5, [x24], #0x10\n" | ||
| 142 | "cmp x28, #0xc\n" | ||
| 143 | "ldr q14, [x23], #0x10\n" | ||
| 144 | "ldr q6, [x22], #0x10\n" | ||
| 145 | "ldr q2, [x21], #0x10\n" | ||
| 146 | "ldr q18, [x20], #0x10\n" | ||
| 147 | "ldr q1, [x9], #0x10\n" | ||
| 148 | "ldr q7, [x26], #0x10\n" | ||
| 149 | "zip1 v15.4s, v28.4s, v11.4s\n" | ||
| 150 | "zip1 v8.4s, v27.4s, v5.4s\n" | ||
| 151 | "ldr q3, [x25], #0x10\n" | ||
| 152 | "ldr q23, [x24], #0x10\n" | ||
| 153 | "zip2 v17.4s, v28.4s, v11.4s\n" | ||
| 154 | "zip2 v27.4s, v27.4s, v5.4s\n" | ||
| 155 | "ldr q5, [x23], #0x10\n" | ||
| 156 | "ldr q30, [x22], #0x10\n" | ||
| 157 | "zip1 v26.4s, v14.4s, v2.4s\n" | ||
| 158 | "zip1 v31.4s, v6.4s, v18.4s\n" | ||
| 159 | "ldr q20, [x21], #0x10\n" | ||
| 160 | "ldr q16, [x20], #0x10\n" | ||
| 161 | "zip2 v12.4s, v14.4s, v2.4s\n" | ||
| 162 | "zip2 v24.4s, v6.4s, v18.4s\n" | ||
| 163 | "ldr q29, [x9], #0x10\n" | ||
| 164 | "ldr q6, [x26], #0x10\n" | ||
| 165 | "zip1 v18.4s, v1.4s, v3.4s\n" | ||
| 166 | "zip1 v4.4s, v7.4s, v23.4s\n" | ||
| 167 | "ldr q22, [x25], #0x10\n" | ||
| 168 | "ldr q0, [x24], #0x10\n" | ||
| 169 | "zip2 v3.4s, v1.4s, v3.4s\n" | ||
| 170 | "zip2 v1.4s, v7.4s, v23.4s\n" | ||
| 171 | "ldr q2, [x23], #0x10\n" | ||
| 172 | "ldr q10, [x22], #0x10\n" | ||
| 173 | "zip1 v28.4s, v5.4s, v20.4s\n" | ||
| 174 | "zip1 v14.4s, v30.4s, v16.4s\n" | ||
| 175 | "ldr q9, [x21], #0x10\n" | ||
| 176 | "ldr q23, [x20], #0x10\n" | ||
| 177 | "zip2 v13.4s, v5.4s, v20.4s\n" | ||
| 178 | "zip2 v30.4s, v30.4s, v16.4s\n" | ||
| 179 | "zip1 v16.4s, v29.4s, v22.4s\n" | ||
| 180 | "zip1 v5.4s, v6.4s, v0.4s\n" | ||
| 181 | "zip2 v22.4s, v29.4s, v22.4s\n" | ||
| 182 | "zip2 v0.4s, v6.4s, v0.4s\n" | ||
| 183 | "zip1 v7.4s, v2.4s, v9.4s\n" | ||
| 184 | "zip1 v19.4s, v10.4s, v23.4s\n" | ||
| 185 | "zip2 v21.4s, v2.4s, v9.4s\n" | ||
| 186 | "zip2 v25.4s, v10.4s, v23.4s\n" | ||
| 187 | "zip1 v11.4s, v15.4s, v8.4s\n" | ||
| 188 | "zip1 v9.4s, v17.4s, v27.4s\n" | ||
| 189 | "zip1 v6.4s, v18.4s, v4.4s\n" | ||
| 190 | "zip1 v2.4s, v3.4s, v1.4s\n" | ||
| 191 | "zip1 v29.4s, v16.4s, v5.4s\n" | ||
| 192 | "zip1 v20.4s, v22.4s, v0.4s\n" | ||
| 193 | "zip1 v10.4s, v26.4s, v31.4s\n" | ||
| 194 | "zip1 v23.4s, v12.4s, v24.4s\n" | ||
| 195 | ".inst 0x0ea1696b // bfcvtn v11.4h, v11.4s\n" | ||
| 196 | "zip2 v8.4s, v15.4s, v8.4s\n" | ||
| 197 | "zip1 v15.4s, v28.4s, v14.4s\n" | ||
| 198 | ".inst 0x0ea16929 // bfcvtn v9.4h, v9.4s\n" | ||
| 199 | "zip2 v27.4s, v17.4s, v27.4s\n" | ||
| 200 | "zip1 v17.4s, v13.4s, v30.4s\n" | ||
| 201 | ".inst 0x0ea168c6 // bfcvtn v6.4h, v6.4s\n" | ||
| 202 | "zip2 v4.4s, v18.4s, v4.4s\n" | ||
| 203 | "zip1 v18.4s, v7.4s, v19.4s\n" | ||
| 204 | ".inst 0x0ea16842 // bfcvtn v2.4h, v2.4s\n" | ||
| 205 | "zip2 v1.4s, v3.4s, v1.4s\n" | ||
| 206 | "zip1 v3.4s, v21.4s, v25.4s\n" | ||
| 207 | ".inst 0x0ea16bbd // bfcvtn v29.4h, v29.4s\n" | ||
| 208 | "zip2 v5.4s, v16.4s, v5.4s\n" | ||
| 209 | ".inst 0x0ea16a94 // bfcvtn v20.4h, v20.4s\n" | ||
| 210 | "zip2 v0.4s, v22.4s, v0.4s\n" | ||
| 211 | ".inst 0x0ea16956 // bfcvtn v22.4h, v10.4s\n" | ||
| 212 | "zip2 v31.4s, v26.4s, v31.4s\n" | ||
| 213 | ".inst 0x0ea16aea // bfcvtn v10.4h, v23.4s\n" | ||
| 214 | "zip2 v26.4s, v12.4s, v24.4s\n" | ||
| 215 | ".inst 0x0ea169ef // bfcvtn v15.4h, v15.4s\n" | ||
| 216 | "zip2 v12.4s, v28.4s, v14.4s\n" | ||
| 217 | ".inst 0x0ea16a2e // bfcvtn v14.4h, v17.4s\n" | ||
| 218 | "zip2 v24.4s, v13.4s, v30.4s\n" | ||
| 219 | ".inst 0x0ea16a57 // bfcvtn v23.4h, v18.4s\n" | ||
| 220 | "zip2 v18.4s, v7.4s, v19.4s\n" | ||
| 221 | ".inst 0x0ea16871 // bfcvtn v17.4h, v3.4s\n" | ||
| 222 | "zip2 v16.4s, v21.4s, v25.4s\n" | ||
| 223 | ".inst 0x4ea1690b // bfcvtn2 v11.8h, v8.4s\n" | ||
| 224 | ".inst 0x4ea16b69 // bfcvtn2 v9.8h, v27.4s\n" | ||
| 225 | ".inst 0x4ea16886 // bfcvtn2 v6.8h, v4.4s\n" | ||
| 226 | ".inst 0x4ea16822 // bfcvtn2 v2.8h, v1.4s\n" | ||
| 227 | ".inst 0x4ea168bd // bfcvtn2 v29.8h, v5.4s\n" | ||
| 228 | ".inst 0x4ea16814 // bfcvtn2 v20.8h, v0.4s\n" | ||
| 229 | ".inst 0x4ea16bf6 // bfcvtn2 v22.8h, v31.4s\n" | ||
| 230 | ".inst 0x4ea16b4a // bfcvtn2 v10.8h, v26.4s\n" | ||
| 231 | "str q11, [x27, #0x0]\n" | ||
| 232 | ".inst 0x4ea1698f // bfcvtn2 v15.8h, v12.4s\n" | ||
| 233 | ".inst 0x4ea16b0e // bfcvtn2 v14.8h, v24.4s\n" | ||
| 234 | "str q9, [x27, #0x10]\n" | ||
| 235 | ".inst 0x4ea16a57 // bfcvtn2 v23.8h, v18.4s\n" | ||
| 236 | ".inst 0x4ea16a11 // bfcvtn2 v17.8h, v16.4s\n" | ||
| 237 | "str q6, [x27, #0x20]\n" | ||
| 238 | "str q2, [x27, #0x30]\n" | ||
| 239 | "str q29, [x27, #0x40]\n" | ||
| 240 | "str q20, [x27, #0x50]\n" | ||
| 241 | "str q22, [x27, #0x60]\n" | ||
| 242 | "str q10, [x27, #0x70]\n" | ||
| 243 | "str q15, [x27, #0x80]\n" | ||
| 244 | "str q14, [x27, #0x90]\n" | ||
| 245 | "str q23, [x27, #0xa0]\n" | ||
| 246 | "str q17, [x27, #0xb0]\n" | ||
| 247 | "add x27, x27, %x[out_stride]\n" | ||
| 248 | "bge 5b\n" | ||
| 249 | "6:" // Main row loop: Column loop skip | ||
| 250 | "cbz x28, 11f\n" | ||
| 251 | "cmp x28, #0x4\n" | ||
| 252 | "movi v16.16b, #0x0\n" | ||
| 253 | "str q16, [x27, #0x0]\n" | ||
| 254 | "str q16, [x27, #0x10]\n" | ||
| 255 | "str q16, [x27, #0x20]\n" | ||
| 256 | "str q16, [x27, #0x30]\n" | ||
| 257 | "str q16, [x27, #0x40]\n" | ||
| 258 | "str q16, [x27, #0x50]\n" | ||
| 259 | "str q16, [x27, #0x60]\n" | ||
| 260 | "str q16, [x27, #0x70]\n" | ||
| 261 | "str q16, [x27, #0x80]\n" | ||
| 262 | "str q16, [x27, #0x90]\n" | ||
| 263 | "str q16, [x27, #0xa0]\n" | ||
| 264 | "str q16, [x27, #0xb0]\n" | ||
| 265 | "blt 8f\n" | ||
| 266 | "7:" // Main row loop: width 4 loop: loop | ||
| 267 | "ldr q25, [x9], #0x10\n" | ||
| 268 | "ldr q24, [x26], #0x10\n" | ||
| 269 | "sub x28, x28, #0x4\n" | ||
| 270 | "ldr q21, [x25], #0x10\n" | ||
| 271 | "ldr q20, [x24], #0x10\n" | ||
| 272 | "cmp x28, #0x4\n" | ||
| 273 | "ldr q23, [x23], #0x10\n" | ||
| 274 | "ldr q19, [x22], #0x10\n" | ||
| 275 | "ldr q18, [x21], #0x10\n" | ||
| 276 | "ldr q17, [x20], #0x10\n" | ||
| 277 | "zip1 v22.4s, v25.4s, v21.4s\n" | ||
| 278 | "zip1 v16.4s, v24.4s, v20.4s\n" | ||
| 279 | "zip2 v21.4s, v25.4s, v21.4s\n" | ||
| 280 | "zip2 v20.4s, v24.4s, v20.4s\n" | ||
| 281 | "zip1 v27.4s, v23.4s, v18.4s\n" | ||
| 282 | "zip1 v26.4s, v19.4s, v17.4s\n" | ||
| 283 | "zip2 v25.4s, v23.4s, v18.4s\n" | ||
| 284 | "zip2 v24.4s, v19.4s, v17.4s\n" | ||
| 285 | "zip1 v19.4s, v22.4s, v16.4s\n" | ||
| 286 | "zip1 v18.4s, v21.4s, v20.4s\n" | ||
| 287 | "zip1 v17.4s, v27.4s, v26.4s\n" | ||
| 288 | "zip2 v23.4s, v22.4s, v16.4s\n" | ||
| 289 | "zip1 v16.4s, v25.4s, v24.4s\n" | ||
| 290 | "zip2 v22.4s, v21.4s, v20.4s\n" | ||
| 291 | ".inst 0x0ea16a75 // bfcvtn v21.4h, v19.4s\n" | ||
| 292 | ".inst 0x0ea16a54 // bfcvtn v20.4h, v18.4s\n" | ||
| 293 | ".inst 0x0ea16a33 // bfcvtn v19.4h, v17.4s\n" | ||
| 294 | "zip2 v18.4s, v27.4s, v26.4s\n" | ||
| 295 | ".inst 0x0ea16a11 // bfcvtn v17.4h, v16.4s\n" | ||
| 296 | "zip2 v16.4s, v25.4s, v24.4s\n" | ||
| 297 | ".inst 0x4ea16af5 // bfcvtn2 v21.8h, v23.4s\n" | ||
| 298 | ".inst 0x4ea16ad4 // bfcvtn2 v20.8h, v22.4s\n" | ||
| 299 | ".inst 0x4ea16a53 // bfcvtn2 v19.8h, v18.4s\n" | ||
| 300 | ".inst 0x4ea16a11 // bfcvtn2 v17.8h, v16.4s\n" | ||
| 301 | "str q21, [x27, #0x0]\n" | ||
| 302 | "str q20, [x27, #0x10]\n" | ||
| 303 | "str q19, [x27, #0x60]\n" | ||
| 304 | "str q17, [x27, #0x70]\n" | ||
| 305 | "add x27, x27, #0x20\n" | ||
| 306 | "bge 7b\n" | ||
| 307 | "8:" // Main row loop: width 4 loop: skip | ||
| 308 | "cmp x28, #0x1\n" | ||
| 309 | "blt 10f\n" | ||
| 310 | "9:" // Main row loop: width 1 loop: loop | ||
| 311 | "ldr s23, [x9], #0x4\n" | ||
| 312 | "ldr s22, [x26], #0x4\n" | ||
| 313 | "sub x28, x28, #0x1\n" | ||
| 314 | "ldr s19, [x25], #0x4\n" | ||
| 315 | "ldr s17, [x24], #0x4\n" | ||
| 316 | "cmp x28, #0x1\n" | ||
| 317 | "ldr s21, [x23], #0x4\n" | ||
| 318 | "ldr s20, [x22], #0x4\n" | ||
| 319 | "ldr s18, [x21], #0x4\n" | ||
| 320 | "ldr s16, [x20], #0x4\n" | ||
| 321 | "zip1 v19.4s, v23.4s, v19.4s\n" | ||
| 322 | "zip1 v17.4s, v22.4s, v17.4s\n" | ||
| 323 | "zip1 v18.4s, v21.4s, v18.4s\n" | ||
| 324 | "zip1 v16.4s, v20.4s, v16.4s\n" | ||
| 325 | "zip1 v17.4s, v19.4s, v17.4s\n" | ||
| 326 | "zip1 v16.4s, v18.4s, v16.4s\n" | ||
| 327 | ".inst 0x0ea16a31 // bfcvtn v17.4h, v17.4s\n" | ||
| 328 | ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n" | ||
| 329 | "str d17, [x27, #0x0]\n" | ||
| 330 | "str d16, [x27, #0x60]\n" | ||
| 331 | "add x27, x27, #0x8\n" | ||
| 332 | "bge 9b\n" | ||
| 333 | "10:" // Main row loop: width 1 loop: skip | ||
| 334 | "11:" // Main row loop: odd col skip | ||
| 335 | "cmp %x[height], #0x8\n" | ||
| 336 | "add %x[out], %x[out], #0xc0\n" | ||
| 337 | "bge 4b\n" | ||
| 338 | "cbz %x[height], 21f\n" | ||
| 339 | "12:" // Main loop skip | ||
| 340 | "13:" // Tail row loop: Head | ||
| 341 | "mov x9, %x[in]\n" | ||
| 342 | "mov x20, %x[width]\n" | ||
| 343 | "cmp %x[height], #0x3\n" | ||
| 344 | "mov x27, %x[out]\n" | ||
| 345 | "add x26, x9, %x[in_stride]\n" | ||
| 346 | "add x25, x26, %x[in_stride]\n" | ||
| 347 | "add x24, x25, %x[in_stride]\n" | ||
| 348 | "csel x25, x25, %x[pad_row], GE\n" | ||
| 349 | "add %x[in], x24, %x[in_stride]\n" | ||
| 350 | "csel x24, x24, %x[pad_row], GT\n" | ||
| 351 | "cmp %x[height], #0x1\n" | ||
| 352 | "sub %x[height], %x[height], #0x4\n" | ||
| 353 | "csel x26, x26, %x[pad_row], GT\n" | ||
| 354 | "cmp x20, #0xc\n" | ||
| 355 | "blt 15f\n" | ||
| 356 | "14:" // Tail row loop: Column loop | ||
| 357 | "ldr q24, [x9], #0x10\n" | ||
| 358 | "ldr q23, [x26], #0x10\n" | ||
| 359 | "sub x20, x20, #0xc\n" | ||
| 360 | "ldr q22, [x25], #0x10\n" | ||
| 361 | "ldr q16, [x24], #0x10\n" | ||
| 362 | "cmp x20, #0xc\n" | ||
| 363 | "ldr q28, [x9], #0x10\n" | ||
| 364 | "ldr q27, [x26], #0x10\n" | ||
| 365 | "ldr q21, [x25], #0x10\n" | ||
| 366 | "ldr q20, [x24], #0x10\n" | ||
| 367 | "ldr q19, [x9], #0x10\n" | ||
| 368 | "zip1 v26.4s, v24.4s, v22.4s\n" | ||
| 369 | "zip1 v25.4s, v23.4s, v16.4s\n" | ||
| 370 | "ldr q18, [x26], #0x10\n" | ||
| 371 | "ldr q17, [x25], #0x10\n" | ||
| 372 | "zip2 v24.4s, v24.4s, v22.4s\n" | ||
| 373 | "zip2 v23.4s, v23.4s, v16.4s\n" | ||
| 374 | "ldr q16, [x24], #0x10\n" | ||
| 375 | "zip1 v2.4s, v28.4s, v21.4s\n" | ||
| 376 | "zip1 v22.4s, v27.4s, v20.4s\n" | ||
| 377 | "zip2 v1.4s, v28.4s, v21.4s\n" | ||
| 378 | "zip2 v0.4s, v27.4s, v20.4s\n" | ||
| 379 | "zip1 v31.4s, v19.4s, v17.4s\n" | ||
| 380 | "zip1 v30.4s, v18.4s, v16.4s\n" | ||
| 381 | "zip2 v29.4s, v19.4s, v17.4s\n" | ||
| 382 | "zip2 v28.4s, v18.4s, v16.4s\n" | ||
| 383 | "zip1 v21.4s, v26.4s, v25.4s\n" | ||
| 384 | "zip1 v20.4s, v24.4s, v23.4s\n" | ||
| 385 | "zip1 v19.4s, v2.4s, v22.4s\n" | ||
| 386 | "zip1 v18.4s, v1.4s, v0.4s\n" | ||
| 387 | "zip1 v17.4s, v31.4s, v30.4s\n" | ||
| 388 | "zip1 v16.4s, v29.4s, v28.4s\n" | ||
| 389 | ".inst 0x0ea16abb // bfcvtn v27.4h, v21.4s\n" | ||
| 390 | "zip2 v26.4s, v26.4s, v25.4s\n" | ||
| 391 | ".inst 0x0ea16a99 // bfcvtn v25.4h, v20.4s\n" | ||
| 392 | "zip2 v24.4s, v24.4s, v23.4s\n" | ||
| 393 | ".inst 0x0ea16a77 // bfcvtn v23.4h, v19.4s\n" | ||
| 394 | "zip2 v22.4s, v2.4s, v22.4s\n" | ||
| 395 | ".inst 0x0ea16a55 // bfcvtn v21.4h, v18.4s\n" | ||
| 396 | "zip2 v20.4s, v1.4s, v0.4s\n" | ||
| 397 | ".inst 0x0ea16a33 // bfcvtn v19.4h, v17.4s\n" | ||
| 398 | "zip2 v18.4s, v31.4s, v30.4s\n" | ||
| 399 | ".inst 0x0ea16a11 // bfcvtn v17.4h, v16.4s\n" | ||
| 400 | "zip2 v16.4s, v29.4s, v28.4s\n" | ||
| 401 | ".inst 0x4ea16b5b // bfcvtn2 v27.8h, v26.4s\n" | ||
| 402 | ".inst 0x4ea16b19 // bfcvtn2 v25.8h, v24.4s\n" | ||
| 403 | ".inst 0x4ea16ad7 // bfcvtn2 v23.8h, v22.4s\n" | ||
| 404 | ".inst 0x4ea16a95 // bfcvtn2 v21.8h, v20.4s\n" | ||
| 405 | ".inst 0x4ea16a53 // bfcvtn2 v19.8h, v18.4s\n" | ||
| 406 | ".inst 0x4ea16a11 // bfcvtn2 v17.8h, v16.4s\n" | ||
| 407 | "str q27, [x27, #0x0]\n" | ||
| 408 | "str q25, [x27, #0x10]\n" | ||
| 409 | "str q23, [x27, #0x20]\n" | ||
| 410 | "str q21, [x27, #0x30]\n" | ||
| 411 | "str q19, [x27, #0x40]\n" | ||
| 412 | "str q17, [x27, #0x50]\n" | ||
| 413 | "add x27, x27, %x[out_stride]\n" | ||
| 414 | "bge 14b\n" | ||
| 415 | "15:" // Tail row loop: Column loop skip | ||
| 416 | "cbz x20, 20f\n" | ||
| 417 | "cmp x20, #0x4\n" | ||
| 418 | "movi v16.16b, #0x0\n" | ||
| 419 | "str q16, [x27, #0x0]\n" | ||
| 420 | "str q16, [x27, #0x10]\n" | ||
| 421 | "str q16, [x27, #0x20]\n" | ||
| 422 | "str q16, [x27, #0x30]\n" | ||
| 423 | "str q16, [x27, #0x40]\n" | ||
| 424 | "str q16, [x27, #0x50]\n" | ||
| 425 | "blt 17f\n" | ||
| 426 | "16:" // Tail row loop: width 4 loop: loop | ||
| 427 | "ldr q21, [x9], #0x10\n" | ||
| 428 | "ldr q20, [x26], #0x10\n" | ||
| 429 | "sub x20, x20, #0x4\n" | ||
| 430 | "ldr q19, [x25], #0x10\n" | ||
| 431 | "ldr q17, [x24], #0x10\n" | ||
| 432 | "cmp x20, #0x4\n" | ||
| 433 | "zip1 v18.4s, v21.4s, v19.4s\n" | ||
| 434 | "zip1 v16.4s, v20.4s, v17.4s\n" | ||
| 435 | "zip2 v21.4s, v21.4s, v19.4s\n" | ||
| 436 | "zip2 v20.4s, v20.4s, v17.4s\n" | ||
| 437 | "zip1 v17.4s, v18.4s, v16.4s\n" | ||
| 438 | "zip2 v19.4s, v18.4s, v16.4s\n" | ||
| 439 | "zip1 v16.4s, v21.4s, v20.4s\n" | ||
| 440 | ".inst 0x0ea16a32 // bfcvtn v18.4h, v17.4s\n" | ||
| 441 | "zip2 v17.4s, v21.4s, v20.4s\n" | ||
| 442 | ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n" | ||
| 443 | ".inst 0x4ea16a72 // bfcvtn2 v18.8h, v19.4s\n" | ||
| 444 | ".inst 0x4ea16a30 // bfcvtn2 v16.8h, v17.4s\n" | ||
| 445 | "str q18, [x27, #0x0]\n" | ||
| 446 | "str q16, [x27, #0x10]\n" | ||
| 447 | "add x27, x27, #0x20\n" | ||
| 448 | "bge 16b\n" | ||
| 449 | "17:" // Tail row loop: width 4 loop: skip | ||
| 450 | "cmp x20, #0x1\n" | ||
| 451 | "blt 19f\n" | ||
| 452 | "18:" // Tail row loop: width 1 loop: loop | ||
| 453 | "ldr s19, [x9], #0x4\n" | ||
| 454 | "ldr s18, [x26], #0x4\n" | ||
| 455 | "sub x20, x20, #0x1\n" | ||
| 456 | "ldr s17, [x25], #0x4\n" | ||
| 457 | "ldr s16, [x24], #0x4\n" | ||
| 458 | "cmp x20, #0x1\n" | ||
| 459 | "zip1 v17.4s, v19.4s, v17.4s\n" | ||
| 460 | "zip1 v16.4s, v18.4s, v16.4s\n" | ||
| 461 | "zip1 v16.4s, v17.4s, v16.4s\n" | ||
| 462 | ".inst 0x0ea16a10 // bfcvtn v16.4h, v16.4s\n" | ||
| 463 | "str d16, [x27, #0x0]\n" | ||
| 464 | "add x27, x27, #0x8\n" | ||
| 465 | "bge 18b\n" | ||
| 466 | "19:" // Tail row loop: width 1 loop: skip | ||
| 467 | "20:" // Tail row loop: odd col skip | ||
| 468 | "cmp %x[height], #0x1\n" | ||
| 469 | "add %x[out], %x[out], #0x60\n" | ||
| 470 | "bge 13b\n" | ||
| 471 | "21:" // Done | ||
| 472 | : [bias] "+&r"(bias_ptr), [height] "+&r"(height), [in] "+&r"(in), [out] "+&r"(out) | ||
| 473 | 1440 | : [bias_step] "r"(bias_step), [in_stride] "r"(in_stride), [out_stride] "r"(out_stride), [pad_row] "r"(pad_row), | |
| 474 | 1440 | [width] "r"(width) | |
| 475 | : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", | ||
| 476 | "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", | ||
| 477 | "v30", "v31", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"); | ||
| 478 | 1440 | } | |
| 479 | |||
| 480 | #endif // Architectural features check. | ||
| 481 |