| 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(__ARM_FEATURE_MATMUL_INT8) | ||
| 11 | #error "i8mm extension required to compile this micro-kernel" | ||
| 12 | #else | ||
| 13 | #include "kai_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm.h" | ||
| 14 | |||
| 15 | #include <arm_neon.h> | ||
| 16 | #include <stddef.h> | ||
| 17 | #include <stdint.h> | ||
| 18 | |||
| 19 | #include "kai/kai_common.h" | ||
| 20 | |||
| 21 | static const size_t kai_m_step = 16; | ||
| 22 | static const size_t kai_n_step = 4; | ||
| 23 | static const size_t kai_mr = 4; | ||
| 24 | static const size_t kai_nr = 4; | ||
| 25 | static const size_t kai_kr = 16; | ||
| 26 | static const size_t kai_sr = 2; | ||
| 27 | static const size_t kai_bl = 32; | ||
| 28 | static const size_t kai_num_bytes_multiplier = sizeof(uint16_t); | ||
| 29 | |||
| 30 | 92 | inline static size_t kai_num_bytes_per_block_lhs(void) { | |
| 31 | 92 | return kai_bl * sizeof(int8_t) + kai_num_bytes_multiplier; | |
| 32 | } | ||
| 33 | |||
| 34 | 92 | inline static size_t kai_num_bytes_per_block_rhs(void) { | |
| 35 | 92 | return (kai_bl / 2) * sizeof(int8_t) + kai_num_bytes_multiplier; | |
| 36 | } | ||
| 37 | |||
| 38 | 184 | inline static size_t kai_num_blocks_per_row(size_t k) { | |
| 39 | − | KAI_ASSUME((k % kai_bl) == 0); | |
| 40 | 184 | return k / kai_bl; | |
| 41 | } | ||
| 42 | |||
| 43 | 92 | inline static size_t kai_lhs_packed_stride(size_t k) { | |
| 44 | 92 | return kai_mr * kai_num_blocks_per_row(k) * kai_num_bytes_per_block_lhs(); | |
| 45 | } | ||
| 46 | |||
| 47 | 92 | inline static size_t kai_rhs_packed_stride(size_t k) { | |
| 48 | − | KAI_ASSUME((k % 2) == 0); | |
| 49 | − | KAI_ASSUME((k % kai_kr) == 0); | |
| 50 | − | KAI_ASSUME((k % kai_bl) == 0); | |
| 51 | |||
| 52 | 92 | const size_t num_blocks_per_row = kai_num_blocks_per_row(k); | |
| 53 | 92 | const size_t num_bytes_per_block = kai_num_bytes_per_block_rhs(); | |
| 54 | |||
| 55 | 184 | return kai_nr * (num_bytes_per_block * num_blocks_per_row); | |
| 56 | 92 | } | |
| 57 | |||
| 58 | 144 | size_t kai_get_m_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm(void) { | |
| 59 | 144 | return kai_m_step; | |
| 60 | } | ||
| 61 | |||
| 62 | 144 | size_t kai_get_n_step_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm(void) { | |
| 63 | 144 | return kai_n_step; | |
| 64 | } | ||
| 65 | |||
| 66 | 96 | size_t kai_get_mr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm(void) { | |
| 67 | 96 | return kai_mr; | |
| 68 | } | ||
| 69 | |||
| 70 | 96 | size_t kai_get_nr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm(void) { | |
| 71 | 96 | return kai_nr; | |
| 72 | } | ||
| 73 | |||
| 74 | 144 | size_t kai_get_kr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm(void) { | |
| 75 | 144 | return kai_kr; | |
| 76 | } | ||
| 77 | |||
| 78 | 96 | size_t kai_get_sr_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm(void) { | |
| 79 | 96 | return kai_sr; | |
| 80 | } | ||
| 81 | |||
| 82 | 92 | size_t kai_get_lhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm( | |
| 83 | size_t m_idx, size_t k, size_t bl) { | ||
| 84 | − | KAI_ASSUME(bl == kai_bl); | |
| 85 | − | KAI_ASSUME((k % 2) == 0); | |
| 86 | − | KAI_ASSUME((k % kai_kr) == 0); | |
| 87 | − | KAI_ASSUME((k % bl) == 0); | |
| 88 | − | KAI_ASSUME((m_idx % kai_m_step) == 0); | |
| 89 | |||
| 90 | 92 | return (m_idx / kai_mr) * kai_lhs_packed_stride(k); | |
| 91 | } | ||
| 92 | |||
| 93 | 92 | size_t kai_get_rhs_packed_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm( | |
| 94 | size_t n_idx, size_t k, size_t bl) { | ||
| 95 | − | KAI_ASSUME(bl == kai_bl); | |
| 96 | − | KAI_ASSUME((k % 2) == 0); | |
| 97 | − | KAI_ASSUME((k % kai_kr) == 0); | |
| 98 | − | KAI_ASSUME((k % bl) == 0); | |
| 99 | − | KAI_ASSUME((n_idx % kai_n_step) == 0); | |
| 100 | |||
| 101 | 92 | return (n_idx / kai_nr) * kai_rhs_packed_stride(k); | |
| 102 | } | ||
| 103 | |||
| 104 | 46 | size_t kai_get_dst_offset_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm( | |
| 105 | size_t m_idx, size_t n_idx, size_t dst_stride) { | ||
| 106 | − | KAI_ASSUME((m_idx % kai_m_step) == 0); | |
| 107 | − | KAI_ASSUME((n_idx % kai_n_step) == 0); | |
| 108 | |||
| 109 | 46 | return (n_idx * sizeof(float)) + m_idx * dst_stride; | |
| 110 | } | ||
| 111 | |||
| 112 | 46 | size_t kai_get_dst_size_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm(size_t m, size_t n) { | |
| 113 | 46 | return m * n * sizeof(float); | |
| 114 | } | ||
| 115 | |||
| 116 | 47 | void kai_run_matmul_clamp_f32_qsi8d32p4x8_qsi4c32p4x8_16x4_neon_i8mm( | |
| 117 | size_t m, size_t n, size_t k, size_t bl, const void* lhs_packed, const void* rhs_packed, | ||
| 118 | float* dst, // NOLINT(readability-non-const-parameter) | ||
| 119 | size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max) { | ||
| 120 | − | KAI_ASSUME(bl == kai_bl); | |
| 121 | − | KAI_ASSUME(k % kai_bl == 0); | |
| 122 | − | KAI_ASSUME(dst_stride_col == sizeof(float)); | |
| 123 | |||
| 124 |
1/2✓ Branch 0 taken 47 times.
✗ Branch 1 not taken.
|
47 | if (m == 0) { |
| 125 | ✗ | return; | |
| 126 | } | ||
| 127 | |||
| 128 | 47 | const size_t num_blocks = k / kai_bl; | |
| 129 | 47 | float clamp_vals[2] = {scalar_min, scalar_max}; | |
| 130 | |||
| 131 | 94 | __asm__ __volatile__( | |
| 132 | "mov x13, %x[m]\n" | ||
| 133 | "mov x12, #0x88\n" | ||
| 134 | "cmp x13, #0x10\n" | ||
| 135 | "mul x12, %x[num_blocks], x12\n" | ||
| 136 | "blt 14f\n" | ||
| 137 | "1:" // Row loop | ||
| 138 | "mov x11, %x[rhs_packed]\n" | ||
| 139 | "mov x10, %x[n]\n" | ||
| 140 | "add x9, %x[dst], %x[dst_stride_row], LSL #4\n" | ||
| 141 | "2:" // Column loop | ||
| 142 | "mov x27, %x[lhs_packed]\n" | ||
| 143 | "movi v31.16b, #0x0\n" | ||
| 144 | "movi v30.16b, #0x0\n" | ||
| 145 | "mov x23, %x[num_blocks]\n" | ||
| 146 | "movi v29.16b, #0x0\n" | ||
| 147 | "movi v28.16b, #0x0\n" | ||
| 148 | "movi v27.16b, #0x0\n" | ||
| 149 | "movi v26.16b, #0x0\n" | ||
| 150 | "add x22, x27, x12\n" | ||
| 151 | "add x21, x22, x12\n" | ||
| 152 | "movi v25.16b, #0x0\n" | ||
| 153 | "movi v24.16b, #0x0\n" | ||
| 154 | "add x20, x21, x12\n" | ||
| 155 | "movi v23.16b, #0x0\n" | ||
| 156 | "movi v22.16b, #0x0\n" | ||
| 157 | "movi v21.16b, #0x0\n" | ||
| 158 | "movi v20.16b, #0x0\n" | ||
| 159 | "movi v19.16b, #0x0\n" | ||
| 160 | "movi v18.16b, #0x0\n" | ||
| 161 | "movi v17.16b, #0x0\n" | ||
| 162 | "movi v16.16b, #0x0\n" | ||
| 163 | "3:" // Block loop | ||
| 164 | "ldr d0, [x11, #0x0]\n" | ||
| 165 | "ldr d3, [x27, #0x0]\n" | ||
| 166 | "add x11, x11, #0x8\n" | ||
| 167 | "add x27, x27, #0x8\n" | ||
| 168 | "ldr q12, [x11, #0x0]\n" | ||
| 169 | "ldr q4, [x11, #0x10]\n" | ||
| 170 | "movi v5.4s, #0x0\n" | ||
| 171 | "movi v14.4s, #0x0\n" | ||
| 172 | "ldr q9, [x27, #0x0]\n" | ||
| 173 | "ldr q10, [x27, #0x10]\n" | ||
| 174 | "movi v7.4s, #0x0\n" | ||
| 175 | "movi v8.4s, #0x0\n" | ||
| 176 | "ldr q2, [x11, #0x20]\n" | ||
| 177 | "ldr q11, [x11, #0x30]\n" | ||
| 178 | "movi v1.16b, #0xf0\n" | ||
| 179 | "fcvtl v6.4s, v0.4h\n" | ||
| 180 | "ldr q15, [x27, #0x20]\n" | ||
| 181 | "shl v13.16b, v12.16b, #0x4\n" | ||
| 182 | "shl v0.16b, v4.16b, #0x4\n" | ||
| 183 | "add x11, x11, #0x40\n" | ||
| 184 | "and v12.16b, v12.16b, v1.16b\n" | ||
| 185 | "and v4.16b, v4.16b, v1.16b\n" | ||
| 186 | "fcvtl v3.4s, v3.4h\n" | ||
| 187 | ".inst 0x4e8da525 // smmla v5.4s, v9.16b, v13.16b\n" | ||
| 188 | ".inst 0x4e80a52e // smmla v14.4s, v9.16b, v0.16b\n" | ||
| 189 | ".inst 0x4e8da547 // smmla v7.4s, v10.16b, v13.16b\n" | ||
| 190 | ".inst 0x4e80a548 // smmla v8.4s, v10.16b, v0.16b\n" | ||
| 191 | "shl v10.16b, v2.16b, #0x4\n" | ||
| 192 | "shl v9.16b, v11.16b, #0x4\n" | ||
| 193 | "and v2.16b, v2.16b, v1.16b\n" | ||
| 194 | "and v11.16b, v11.16b, v1.16b\n" | ||
| 195 | "ldr q1, [x27, #0x30]\n" | ||
| 196 | ".inst 0x4e8aa5e5 // smmla v5.4s, v15.16b, v10.16b\n" | ||
| 197 | ".inst 0x4e89a5ee // smmla v14.4s, v15.16b, v9.16b\n" | ||
| 198 | "ldr q15, [x27, #0x40]\n" | ||
| 199 | ".inst 0x4e8aa427 // smmla v7.4s, v1.16b, v10.16b\n" | ||
| 200 | ".inst 0x4e89a428 // smmla v8.4s, v1.16b, v9.16b\n" | ||
| 201 | "ldr q1, [x27, #0x50]\n" | ||
| 202 | ".inst 0x4e8ca5e5 // smmla v5.4s, v15.16b, v12.16b\n" | ||
| 203 | ".inst 0x4e84a5ee // smmla v14.4s, v15.16b, v4.16b\n" | ||
| 204 | "ldr q15, [x27, #0x60]\n" | ||
| 205 | ".inst 0x4e8ca427 // smmla v7.4s, v1.16b, v12.16b\n" | ||
| 206 | ".inst 0x4e84a428 // smmla v8.4s, v1.16b, v4.16b\n" | ||
| 207 | "ldr q1, [x27, #0x70]\n" | ||
| 208 | "add x27, x27, #0x80\n" | ||
| 209 | ".inst 0x4e82a5e5 // smmla v5.4s, v15.16b, v2.16b\n" | ||
| 210 | ".inst 0x4e8ba5ee // smmla v14.4s, v15.16b, v11.16b\n" | ||
| 211 | "fmul v15.4s, v6.4s, v3.s[0]\n" | ||
| 212 | ".inst 0x4e82a427 // smmla v7.4s, v1.16b, v2.16b\n" | ||
| 213 | ".inst 0x4e8ba428 // smmla v8.4s, v1.16b, v11.16b\n" | ||
| 214 | "uzp1 v1.2d, v5.2d, v14.2d\n" | ||
| 215 | "uzp2 v5.2d, v5.2d, v14.2d\n" | ||
| 216 | "fmul v14.4s, v6.4s, v3.s[1]\n" | ||
| 217 | "scvtf v1.4s, v1.4s, #0x4\n" | ||
| 218 | "scvtf v5.4s, v5.4s, #0x4\n" | ||
| 219 | "fmla v31.4s, v1.4s, v15.4s\n" | ||
| 220 | "fmul v15.4s, v6.4s, v3.s[2]\n" | ||
| 221 | "fmul v3.4s, v6.4s, v3.s[3]\n" | ||
| 222 | "uzp1 v1.2d, v7.2d, v8.2d\n" | ||
| 223 | "uzp2 v8.2d, v7.2d, v8.2d\n" | ||
| 224 | "fmla v30.4s, v5.4s, v14.4s\n" | ||
| 225 | "scvtf v1.4s, v1.4s, #0x4\n" | ||
| 226 | "scvtf v8.4s, v8.4s, #0x4\n" | ||
| 227 | "fmla v29.4s, v1.4s, v15.4s\n" | ||
| 228 | "fmla v28.4s, v8.4s, v3.4s\n" | ||
| 229 | "ldr d5, [x22, #0x0]\n" | ||
| 230 | "add x22, x22, #0x8\n" | ||
| 231 | "movi v3.4s, #0x0\n" | ||
| 232 | "movi v1.4s, #0x0\n" | ||
| 233 | "ldr q15, [x22, #0x0]\n" | ||
| 234 | "ldr q7, [x22, #0x10]\n" | ||
| 235 | "movi v14.4s, #0x0\n" | ||
| 236 | "movi v8.4s, #0x0\n" | ||
| 237 | "fcvtl v5.4s, v5.4h\n" | ||
| 238 | ".inst 0x4e8da5e3 // smmla v3.4s, v15.16b, v13.16b\n" | ||
| 239 | ".inst 0x4e80a5e1 // smmla v1.4s, v15.16b, v0.16b\n" | ||
| 240 | "ldr q15, [x22, #0x20]\n" | ||
| 241 | ".inst 0x4e8da4ee // smmla v14.4s, v7.16b, v13.16b\n" | ||
| 242 | ".inst 0x4e80a4e8 // smmla v8.4s, v7.16b, v0.16b\n" | ||
| 243 | "ldr q7, [x22, #0x30]\n" | ||
| 244 | ".inst 0x4e8aa5e3 // smmla v3.4s, v15.16b, v10.16b\n" | ||
| 245 | ".inst 0x4e89a5e1 // smmla v1.4s, v15.16b, v9.16b\n" | ||
| 246 | "ldr q15, [x22, #0x40]\n" | ||
| 247 | ".inst 0x4e8aa4ee // smmla v14.4s, v7.16b, v10.16b\n" | ||
| 248 | ".inst 0x4e89a4e8 // smmla v8.4s, v7.16b, v9.16b\n" | ||
| 249 | "ldr q7, [x22, #0x50]\n" | ||
| 250 | ".inst 0x4e8ca5e3 // smmla v3.4s, v15.16b, v12.16b\n" | ||
| 251 | ".inst 0x4e84a5e1 // smmla v1.4s, v15.16b, v4.16b\n" | ||
| 252 | "ldr q15, [x22, #0x60]\n" | ||
| 253 | ".inst 0x4e8ca4ee // smmla v14.4s, v7.16b, v12.16b\n" | ||
| 254 | ".inst 0x4e84a4e8 // smmla v8.4s, v7.16b, v4.16b\n" | ||
| 255 | "ldr q7, [x22, #0x70]\n" | ||
| 256 | "add x22, x22, #0x80\n" | ||
| 257 | ".inst 0x4e82a5e3 // smmla v3.4s, v15.16b, v2.16b\n" | ||
| 258 | ".inst 0x4e8ba5e1 // smmla v1.4s, v15.16b, v11.16b\n" | ||
| 259 | "fmul v15.4s, v6.4s, v5.s[0]\n" | ||
| 260 | ".inst 0x4e82a4ee // smmla v14.4s, v7.16b, v2.16b\n" | ||
| 261 | ".inst 0x4e8ba4e8 // smmla v8.4s, v7.16b, v11.16b\n" | ||
| 262 | "uzp1 v7.2d, v3.2d, v1.2d\n" | ||
| 263 | "uzp2 v1.2d, v3.2d, v1.2d\n" | ||
| 264 | "fmul v3.4s, v6.4s, v5.s[1]\n" | ||
| 265 | "scvtf v7.4s, v7.4s, #0x4\n" | ||
| 266 | "scvtf v1.4s, v1.4s, #0x4\n" | ||
| 267 | "fmla v27.4s, v7.4s, v15.4s\n" | ||
| 268 | "fmul v15.4s, v6.4s, v5.s[2]\n" | ||
| 269 | "fmul v7.4s, v6.4s, v5.s[3]\n" | ||
| 270 | "uzp1 v5.2d, v14.2d, v8.2d\n" | ||
| 271 | "uzp2 v14.2d, v14.2d, v8.2d\n" | ||
| 272 | "fmla v26.4s, v1.4s, v3.4s\n" | ||
| 273 | "scvtf v5.4s, v5.4s, #0x4\n" | ||
| 274 | "scvtf v14.4s, v14.4s, #0x4\n" | ||
| 275 | "fmla v25.4s, v5.4s, v15.4s\n" | ||
| 276 | "fmla v24.4s, v14.4s, v7.4s\n" | ||
| 277 | "ldr d1, [x21, #0x0]\n" | ||
| 278 | "add x21, x21, #0x8\n" | ||
| 279 | "movi v8.4s, #0x0\n" | ||
| 280 | "movi v5.4s, #0x0\n" | ||
| 281 | "ldr q3, [x21, #0x0]\n" | ||
| 282 | "ldr q7, [x21, #0x10]\n" | ||
| 283 | "movi v14.4s, #0x0\n" | ||
| 284 | "movi v15.4s, #0x0\n" | ||
| 285 | "fcvtl v1.4s, v1.4h\n" | ||
| 286 | ".inst 0x4e8da468 // smmla v8.4s, v3.16b, v13.16b\n" | ||
| 287 | ".inst 0x4e80a465 // smmla v5.4s, v3.16b, v0.16b\n" | ||
| 288 | "ldr q3, [x21, #0x20]\n" | ||
| 289 | ".inst 0x4e8da4ee // smmla v14.4s, v7.16b, v13.16b\n" | ||
| 290 | ".inst 0x4e80a4ef // smmla v15.4s, v7.16b, v0.16b\n" | ||
| 291 | "ldr q7, [x21, #0x30]\n" | ||
| 292 | ".inst 0x4e8aa468 // smmla v8.4s, v3.16b, v10.16b\n" | ||
| 293 | ".inst 0x4e89a465 // smmla v5.4s, v3.16b, v9.16b\n" | ||
| 294 | "ldr q3, [x21, #0x40]\n" | ||
| 295 | ".inst 0x4e8aa4ee // smmla v14.4s, v7.16b, v10.16b\n" | ||
| 296 | ".inst 0x4e89a4ef // smmla v15.4s, v7.16b, v9.16b\n" | ||
| 297 | "ldr q7, [x21, #0x50]\n" | ||
| 298 | ".inst 0x4e8ca468 // smmla v8.4s, v3.16b, v12.16b\n" | ||
| 299 | ".inst 0x4e84a465 // smmla v5.4s, v3.16b, v4.16b\n" | ||
| 300 | "ldr q3, [x21, #0x60]\n" | ||
| 301 | ".inst 0x4e8ca4ee // smmla v14.4s, v7.16b, v12.16b\n" | ||
| 302 | ".inst 0x4e84a4ef // smmla v15.4s, v7.16b, v4.16b\n" | ||
| 303 | "ldr q7, [x21, #0x70]\n" | ||
| 304 | "add x21, x21, #0x80\n" | ||
| 305 | ".inst 0x4e82a468 // smmla v8.4s, v3.16b, v2.16b\n" | ||
| 306 | ".inst 0x4e8ba465 // smmla v5.4s, v3.16b, v11.16b\n" | ||
| 307 | "fmul v3.4s, v6.4s, v1.s[0]\n" | ||
| 308 | ".inst 0x4e82a4ee // smmla v14.4s, v7.16b, v2.16b\n" | ||
| 309 | ".inst 0x4e8ba4ef // smmla v15.4s, v7.16b, v11.16b\n" | ||
| 310 | "uzp1 v7.2d, v8.2d, v5.2d\n" | ||
| 311 | "uzp2 v8.2d, v8.2d, v5.2d\n" | ||
| 312 | "fmul v5.4s, v6.4s, v1.s[1]\n" | ||
| 313 | "scvtf v7.4s, v7.4s, #0x4\n" | ||
| 314 | "scvtf v8.4s, v8.4s, #0x4\n" | ||
| 315 | "fmla v23.4s, v7.4s, v3.4s\n" | ||
| 316 | "fmul v3.4s, v6.4s, v1.s[2]\n" | ||
| 317 | "fmul v1.4s, v6.4s, v1.s[3]\n" | ||
| 318 | "uzp1 v7.2d, v14.2d, v15.2d\n" | ||
| 319 | "uzp2 v14.2d, v14.2d, v15.2d\n" | ||
| 320 | "fmla v22.4s, v8.4s, v5.4s\n" | ||
| 321 | "scvtf v7.4s, v7.4s, #0x4\n" | ||
| 322 | "scvtf v14.4s, v14.4s, #0x4\n" | ||
| 323 | "fmla v21.4s, v7.4s, v3.4s\n" | ||
| 324 | "fmla v20.4s, v14.4s, v1.4s\n" | ||
| 325 | "ldr d3, [x20, #0x0]\n" | ||
| 326 | "add x20, x20, #0x8\n" | ||
| 327 | "movi v15.4s, #0x0\n" | ||
| 328 | "movi v8.4s, #0x0\n" | ||
| 329 | "ldr q5, [x20, #0x0]\n" | ||
| 330 | "ldr q14, [x20, #0x10]\n" | ||
| 331 | "movi v1.4s, #0x0\n" | ||
| 332 | "movi v7.4s, #0x0\n" | ||
| 333 | "fcvtl v3.4s, v3.4h\n" | ||
| 334 | ".inst 0x4e8da4af // smmla v15.4s, v5.16b, v13.16b\n" | ||
| 335 | ".inst 0x4e80a4a8 // smmla v8.4s, v5.16b, v0.16b\n" | ||
| 336 | "ldr q5, [x20, #0x20]\n" | ||
| 337 | ".inst 0x4e8da5c1 // smmla v1.4s, v14.16b, v13.16b\n" | ||
| 338 | "ldr q13, [x20, #0x30]\n" | ||
| 339 | ".inst 0x4e80a5c7 // smmla v7.4s, v14.16b, v0.16b\n" | ||
| 340 | "ldr q14, [x20, #0x40]\n" | ||
| 341 | "ldr q0, [x20, #0x50]\n" | ||
| 342 | ".inst 0x4e8aa4af // smmla v15.4s, v5.16b, v10.16b\n" | ||
| 343 | ".inst 0x4e89a4a8 // smmla v8.4s, v5.16b, v9.16b\n" | ||
| 344 | "ldr q5, [x20, #0x60]\n" | ||
| 345 | ".inst 0x4e8aa5a1 // smmla v1.4s, v13.16b, v10.16b\n" | ||
| 346 | "ldr q10, [x20, #0x70]\n" | ||
| 347 | "add x20, x20, #0x80\n" | ||
| 348 | ".inst 0x4e89a5a7 // smmla v7.4s, v13.16b, v9.16b\n" | ||
| 349 | "fmul v13.4s, v6.4s, v3.s[0]\n" | ||
| 350 | "fmul v9.4s, v6.4s, v3.s[1]\n" | ||
| 351 | ".inst 0x4e8ca5cf // smmla v15.4s, v14.16b, v12.16b\n" | ||
| 352 | ".inst 0x4e84a5c8 // smmla v8.4s, v14.16b, v4.16b\n" | ||
| 353 | "fmul v14.4s, v6.4s, v3.s[2]\n" | ||
| 354 | "fmul v6.4s, v6.4s, v3.s[3]\n" | ||
| 355 | ".inst 0x4e8ca401 // smmla v1.4s, v0.16b, v12.16b\n" | ||
| 356 | ".inst 0x4e84a407 // smmla v7.4s, v0.16b, v4.16b\n" | ||
| 357 | ".inst 0x4e82a4af // smmla v15.4s, v5.16b, v2.16b\n" | ||
| 358 | ".inst 0x4e8ba4a8 // smmla v8.4s, v5.16b, v11.16b\n" | ||
| 359 | ".inst 0x4e82a541 // smmla v1.4s, v10.16b, v2.16b\n" | ||
| 360 | ".inst 0x4e8ba547 // smmla v7.4s, v10.16b, v11.16b\n" | ||
| 361 | "uzp1 v4.2d, v15.2d, v8.2d\n" | ||
| 362 | "uzp2 v2.2d, v15.2d, v8.2d\n" | ||
| 363 | "scvtf v4.4s, v4.4s, #0x4\n" | ||
| 364 | "uzp1 v8.2d, v1.2d, v7.2d\n" | ||
| 365 | "uzp2 v0.2d, v1.2d, v7.2d\n" | ||
| 366 | "scvtf v2.4s, v2.4s, #0x4\n" | ||
| 367 | "fmla v19.4s, v4.4s, v13.4s\n" | ||
| 368 | "scvtf v8.4s, v8.4s, #0x4\n" | ||
| 369 | "scvtf v0.4s, v0.4s, #0x4\n" | ||
| 370 | "fmla v18.4s, v2.4s, v9.4s\n" | ||
| 371 | "fmla v17.4s, v8.4s, v14.4s\n" | ||
| 372 | "fmla v16.4s, v0.4s, v6.4s\n" | ||
| 373 | "subs x23, x23, #0x1\n" | ||
| 374 | "bgt 3b\n" | ||
| 375 | "ld1r { v1.4s }, [%x[clamp_vals]]\n" | ||
| 376 | "add x20, %x[clamp_vals], #0x4\n" | ||
| 377 | "cmp x10, #0x4\n" | ||
| 378 | "ld1r { v0.4s }, [x20]\n" | ||
| 379 | "fmax v31.4s, v31.4s, v1.4s\n" | ||
| 380 | "fmax v30.4s, v30.4s, v1.4s\n" | ||
| 381 | "fmax v29.4s, v29.4s, v1.4s\n" | ||
| 382 | "fmax v28.4s, v28.4s, v1.4s\n" | ||
| 383 | "fmax v27.4s, v27.4s, v1.4s\n" | ||
| 384 | "fmax v26.4s, v26.4s, v1.4s\n" | ||
| 385 | "fmax v25.4s, v25.4s, v1.4s\n" | ||
| 386 | "fmax v24.4s, v24.4s, v1.4s\n" | ||
| 387 | "fmax v23.4s, v23.4s, v1.4s\n" | ||
| 388 | "fmax v22.4s, v22.4s, v1.4s\n" | ||
| 389 | "fmax v21.4s, v21.4s, v1.4s\n" | ||
| 390 | "fmax v20.4s, v20.4s, v1.4s\n" | ||
| 391 | "fmax v19.4s, v19.4s, v1.4s\n" | ||
| 392 | "fmax v18.4s, v18.4s, v1.4s\n" | ||
| 393 | "fmax v17.4s, v17.4s, v1.4s\n" | ||
| 394 | "fmax v16.4s, v16.4s, v1.4s\n" | ||
| 395 | "fmin v31.4s, v31.4s, v0.4s\n" | ||
| 396 | "fmin v30.4s, v30.4s, v0.4s\n" | ||
| 397 | "fmin v29.4s, v29.4s, v0.4s\n" | ||
| 398 | "fmin v28.4s, v28.4s, v0.4s\n" | ||
| 399 | "fmin v27.4s, v27.4s, v0.4s\n" | ||
| 400 | "fmin v26.4s, v26.4s, v0.4s\n" | ||
| 401 | "fmin v25.4s, v25.4s, v0.4s\n" | ||
| 402 | "fmin v24.4s, v24.4s, v0.4s\n" | ||
| 403 | "fmin v23.4s, v23.4s, v0.4s\n" | ||
| 404 | "fmin v22.4s, v22.4s, v0.4s\n" | ||
| 405 | "fmin v21.4s, v21.4s, v0.4s\n" | ||
| 406 | "fmin v20.4s, v20.4s, v0.4s\n" | ||
| 407 | "fmin v19.4s, v19.4s, v0.4s\n" | ||
| 408 | "fmin v18.4s, v18.4s, v0.4s\n" | ||
| 409 | "fmin v17.4s, v17.4s, v0.4s\n" | ||
| 410 | "fmin v16.4s, v16.4s, v0.4s\n" | ||
| 411 | "blt 8f\n" | ||
| 412 | "mov x20, %x[dst]\n" | ||
| 413 | "str q31, [x20, #0x0]\n" | ||
| 414 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 415 | "str q30, [x20, #0x0]\n" | ||
| 416 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 417 | "str q29, [x20, #0x0]\n" | ||
| 418 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 419 | "str q28, [x20, #0x0]\n" | ||
| 420 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 421 | "str q27, [x20, #0x0]\n" | ||
| 422 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 423 | "str q26, [x20, #0x0]\n" | ||
| 424 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 425 | "str q25, [x20, #0x0]\n" | ||
| 426 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 427 | "str q24, [x20, #0x0]\n" | ||
| 428 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 429 | "str q23, [x20, #0x0]\n" | ||
| 430 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 431 | "str q22, [x20, #0x0]\n" | ||
| 432 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 433 | "str q21, [x20, #0x0]\n" | ||
| 434 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 435 | "str q20, [x20, #0x0]\n" | ||
| 436 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 437 | "str q19, [x20, #0x0]\n" | ||
| 438 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 439 | "str q18, [x20, #0x0]\n" | ||
| 440 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 441 | "str q17, [x20, #0x0]\n" | ||
| 442 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 443 | "str q16, [x20, #0x0]\n" | ||
| 444 | "b 13f\n" | ||
| 445 | "8:" // Partial output | ||
| 446 | "mov x28, %x[dst]\n" | ||
| 447 | "add x26, x28, %x[dst_stride_row], LSL #2\n" | ||
| 448 | "add x25, x26, %x[dst_stride_row], LSL #1\n" | ||
| 449 | "add x24, x26, %x[dst_stride_row]\n" | ||
| 450 | "add x23, x25, %x[dst_stride_row]\n" | ||
| 451 | "add x22, x28, %x[dst_stride_row], LSL #1\n" | ||
| 452 | "add x21, x28, %x[dst_stride_row]\n" | ||
| 453 | "add x20, x22, %x[dst_stride_row]\n" | ||
| 454 | "add x27, x23, %x[dst_stride_row]\n" | ||
| 455 | "tbz x10, #1, 9f\n" | ||
| 456 | "st1 { v24.d }[0], [x23], #0x8\n" | ||
| 457 | "st1 { v25.d }[0], [x25], #0x8\n" | ||
| 458 | "st1 { v26.d }[0], [x24], #0x8\n" | ||
| 459 | "st1 { v27.d }[0], [x26], #0x8\n" | ||
| 460 | "st1 { v28.d }[0], [x20], #0x8\n" | ||
| 461 | "st1 { v29.d }[0], [x22], #0x8\n" | ||
| 462 | "st1 { v30.d }[0], [x21], #0x8\n" | ||
| 463 | "st1 { v31.d }[0], [x28], #0x8\n" | ||
| 464 | "tbz x10, #0, 10f\n" | ||
| 465 | "st1 { v24.s }[2], [x23]\n" | ||
| 466 | "st1 { v25.s }[2], [x25]\n" | ||
| 467 | "st1 { v26.s }[2], [x24]\n" | ||
| 468 | "st1 { v27.s }[2], [x26]\n" | ||
| 469 | "st1 { v28.s }[2], [x20]\n" | ||
| 470 | "st1 { v29.s }[2], [x22]\n" | ||
| 471 | "st1 { v30.s }[2], [x21]\n" | ||
| 472 | "st1 { v31.s }[2], [x28]\n" | ||
| 473 | "b 10f\n" | ||
| 474 | "9:" // Output block 0: partial_1_0 | ||
| 475 | "st1 { v24.s }[0], [x23]\n" | ||
| 476 | "st1 { v25.s }[0], [x25]\n" | ||
| 477 | "st1 { v26.s }[0], [x24]\n" | ||
| 478 | "st1 { v27.s }[0], [x26]\n" | ||
| 479 | "st1 { v28.s }[0], [x20]\n" | ||
| 480 | "st1 { v29.s }[0], [x22]\n" | ||
| 481 | "st1 { v30.s }[0], [x21]\n" | ||
| 482 | "st1 { v31.s }[0], [x28]\n" | ||
| 483 | "10:" // Output block 0: Done | ||
| 484 | "add x26, x27, %x[dst_stride_row], LSL #2\n" | ||
| 485 | "add x25, x27, %x[dst_stride_row], LSL #1\n" | ||
| 486 | "add x24, x26, %x[dst_stride_row], LSL #1\n" | ||
| 487 | "add x23, x27, %x[dst_stride_row]\n" | ||
| 488 | "add x22, x25, %x[dst_stride_row]\n" | ||
| 489 | "add x21, x26, %x[dst_stride_row]\n" | ||
| 490 | "add x20, x24, %x[dst_stride_row]\n" | ||
| 491 | "tbz x10, #1, 11f\n" | ||
| 492 | "st1 { v16.d }[0], [x20], #0x8\n" | ||
| 493 | "st1 { v17.d }[0], [x24], #0x8\n" | ||
| 494 | "st1 { v18.d }[0], [x21], #0x8\n" | ||
| 495 | "st1 { v19.d }[0], [x26], #0x8\n" | ||
| 496 | "st1 { v20.d }[0], [x22], #0x8\n" | ||
| 497 | "st1 { v21.d }[0], [x25], #0x8\n" | ||
| 498 | "st1 { v22.d }[0], [x23], #0x8\n" | ||
| 499 | "st1 { v23.d }[0], [x27], #0x8\n" | ||
| 500 | "tbz x10, #0, 12f\n" | ||
| 501 | "st1 { v16.s }[2], [x20]\n" | ||
| 502 | "st1 { v17.s }[2], [x24]\n" | ||
| 503 | "st1 { v18.s }[2], [x21]\n" | ||
| 504 | "st1 { v19.s }[2], [x26]\n" | ||
| 505 | "st1 { v20.s }[2], [x22]\n" | ||
| 506 | "st1 { v21.s }[2], [x25]\n" | ||
| 507 | "st1 { v22.s }[2], [x23]\n" | ||
| 508 | "st1 { v23.s }[2], [x27]\n" | ||
| 509 | "b 12f\n" | ||
| 510 | "11:" // Output block 1: partial_1_0 | ||
| 511 | "st1 { v16.s }[0], [x20]\n" | ||
| 512 | "st1 { v17.s }[0], [x24]\n" | ||
| 513 | "st1 { v18.s }[0], [x21]\n" | ||
| 514 | "st1 { v19.s }[0], [x26]\n" | ||
| 515 | "st1 { v20.s }[0], [x22]\n" | ||
| 516 | "st1 { v21.s }[0], [x25]\n" | ||
| 517 | "st1 { v22.s }[0], [x23]\n" | ||
| 518 | "st1 { v23.s }[0], [x27]\n" | ||
| 519 | "12:" // Output block 1: Done | ||
| 520 | "13:" // Output stage exit | ||
| 521 | "subs x10, x10, #0x4\n" | ||
| 522 | "add %x[dst], %x[dst], #0x10\n" | ||
| 523 | "bgt 2b\n" | ||
| 524 | "mov x20, #0x4\n" | ||
| 525 | "sub x13, x13, #0x10\n" | ||
| 526 | "cmp x13, #0x10\n" | ||
| 527 | "mov %x[dst], x9\n" | ||
| 528 | "madd %x[lhs_packed], x20, x12, %x[lhs_packed]\n" | ||
| 529 | "bge 1b\n" | ||
| 530 | "14:" // Row loop skip | ||
| 531 | "cbz x13, 23f\n" | ||
| 532 | "15:" // Row tail: Row loop | ||
| 533 | "mov x26, %x[rhs_packed]\n" | ||
| 534 | "mov x25, %x[n]\n" | ||
| 535 | "add x24, %x[dst], %x[dst_stride_row], LSL #2\n" | ||
| 536 | "16:" // Row tail: Column loop | ||
| 537 | "movi v31.16b, #0x0\n" | ||
| 538 | "movi v30.16b, #0x0\n" | ||
| 539 | "mov x27, %x[lhs_packed]\n" | ||
| 540 | "mov x20, %x[num_blocks]\n" | ||
| 541 | "movi v29.16b, #0x0\n" | ||
| 542 | "movi v28.16b, #0x0\n" | ||
| 543 | "17:" // Row tail: Block loop | ||
| 544 | "ldr d16, [x26, #0x0]\n" | ||
| 545 | "ldr d10, [x27, #0x0]\n" | ||
| 546 | "add x26, x26, #0x8\n" | ||
| 547 | "add x27, x27, #0x8\n" | ||
| 548 | "ldr q9, [x26, #0x0]\n" | ||
| 549 | "ldr q8, [x26, #0x10]\n" | ||
| 550 | "movi v7.4s, #0x0\n" | ||
| 551 | "movi v6.4s, #0x0\n" | ||
| 552 | "ldr q5, [x27, #0x0]\n" | ||
| 553 | "ldr q4, [x27, #0x10]\n" | ||
| 554 | "movi v3.4s, #0x0\n" | ||
| 555 | "movi v2.4s, #0x0\n" | ||
| 556 | "ldr q1, [x26, #0x20]\n" | ||
| 557 | "ldr q0, [x26, #0x30]\n" | ||
| 558 | "movi v27.16b, #0xf0\n" | ||
| 559 | "fcvtl v26.4s, v16.4h\n" | ||
| 560 | "ldr q23, [x27, #0x20]\n" | ||
| 561 | "ldr q22, [x27, #0x30]\n" | ||
| 562 | "shl v21.16b, v9.16b, #0x4\n" | ||
| 563 | "shl v20.16b, v8.16b, #0x4\n" | ||
| 564 | "ldr q25, [x27, #0x40]\n" | ||
| 565 | "ldr q24, [x27, #0x50]\n" | ||
| 566 | "and v9.16b, v9.16b, v27.16b\n" | ||
| 567 | "and v8.16b, v8.16b, v27.16b\n" | ||
| 568 | "ldr q19, [x27, #0x60]\n" | ||
| 569 | "ldr q18, [x27, #0x70]\n" | ||
| 570 | "shl v17.16b, v1.16b, #0x4\n" | ||
| 571 | "shl v16.16b, v0.16b, #0x4\n" | ||
| 572 | ".inst 0x4e95a4a7 // smmla v7.4s, v5.16b, v21.16b\n" | ||
| 573 | ".inst 0x4e94a4a6 // smmla v6.4s, v5.16b, v20.16b\n" | ||
| 574 | "and v1.16b, v1.16b, v27.16b\n" | ||
| 575 | "add x26, x26, #0x40\n" | ||
| 576 | ".inst 0x4e95a483 // smmla v3.4s, v4.16b, v21.16b\n" | ||
| 577 | ".inst 0x4e94a482 // smmla v2.4s, v4.16b, v20.16b\n" | ||
| 578 | "and v0.16b, v0.16b, v27.16b\n" | ||
| 579 | "add x27, x27, #0x80\n" | ||
| 580 | "fcvtl v10.4s, v10.4h\n" | ||
| 581 | ".inst 0x4e91a6e7 // smmla v7.4s, v23.16b, v17.16b\n" | ||
| 582 | ".inst 0x4e90a6e6 // smmla v6.4s, v23.16b, v16.16b\n" | ||
| 583 | ".inst 0x4e91a6c3 // smmla v3.4s, v22.16b, v17.16b\n" | ||
| 584 | ".inst 0x4e90a6c2 // smmla v2.4s, v22.16b, v16.16b\n" | ||
| 585 | "fmul v23.4s, v26.4s, v10.s[0]\n" | ||
| 586 | "fmul v22.4s, v26.4s, v10.s[1]\n" | ||
| 587 | "fmul v21.4s, v26.4s, v10.s[2]\n" | ||
| 588 | "fmul v20.4s, v26.4s, v10.s[3]\n" | ||
| 589 | ".inst 0x4e89a727 // smmla v7.4s, v25.16b, v9.16b\n" | ||
| 590 | ".inst 0x4e88a726 // smmla v6.4s, v25.16b, v8.16b\n" | ||
| 591 | ".inst 0x4e89a703 // smmla v3.4s, v24.16b, v9.16b\n" | ||
| 592 | ".inst 0x4e88a702 // smmla v2.4s, v24.16b, v8.16b\n" | ||
| 593 | ".inst 0x4e81a667 // smmla v7.4s, v19.16b, v1.16b\n" | ||
| 594 | ".inst 0x4e80a666 // smmla v6.4s, v19.16b, v0.16b\n" | ||
| 595 | ".inst 0x4e81a643 // smmla v3.4s, v18.16b, v1.16b\n" | ||
| 596 | ".inst 0x4e80a642 // smmla v2.4s, v18.16b, v0.16b\n" | ||
| 597 | "uzp1 v19.2d, v7.2d, v6.2d\n" | ||
| 598 | "uzp2 v18.2d, v7.2d, v6.2d\n" | ||
| 599 | "scvtf v19.4s, v19.4s, #0x4\n" | ||
| 600 | "uzp1 v17.2d, v3.2d, v2.2d\n" | ||
| 601 | "uzp2 v16.2d, v3.2d, v2.2d\n" | ||
| 602 | "scvtf v18.4s, v18.4s, #0x4\n" | ||
| 603 | "fmla v31.4s, v19.4s, v23.4s\n" | ||
| 604 | "scvtf v17.4s, v17.4s, #0x4\n" | ||
| 605 | "scvtf v16.4s, v16.4s, #0x4\n" | ||
| 606 | "fmla v30.4s, v18.4s, v22.4s\n" | ||
| 607 | "fmla v29.4s, v17.4s, v21.4s\n" | ||
| 608 | "fmla v28.4s, v16.4s, v20.4s\n" | ||
| 609 | "subs x20, x20, #0x1\n" | ||
| 610 | "bgt 17b\n" | ||
| 611 | "ld1r { v17.4s }, [%x[clamp_vals]]\n" | ||
| 612 | "add x20, %x[clamp_vals], #0x4\n" | ||
| 613 | "cmp x25, #0x4\n" | ||
| 614 | "ld1r { v16.4s }, [x20]\n" | ||
| 615 | "fmax v31.4s, v31.4s, v17.4s\n" | ||
| 616 | "fmax v30.4s, v30.4s, v17.4s\n" | ||
| 617 | "fmax v29.4s, v29.4s, v17.4s\n" | ||
| 618 | "fmax v28.4s, v28.4s, v17.4s\n" | ||
| 619 | "fmin v31.4s, v31.4s, v16.4s\n" | ||
| 620 | "fmin v30.4s, v30.4s, v16.4s\n" | ||
| 621 | "fmin v29.4s, v29.4s, v16.4s\n" | ||
| 622 | "fmin v28.4s, v28.4s, v16.4s\n" | ||
| 623 | "blt 19f\n" | ||
| 624 | "mov x20, %x[dst]\n" | ||
| 625 | "cmp x13, #0x1\n" | ||
| 626 | "str q31, [x20, #0x0]\n" | ||
| 627 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 628 | "ble 22f\n" | ||
| 629 | "cmp x13, #0x2\n" | ||
| 630 | "str q30, [x20, #0x0]\n" | ||
| 631 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 632 | "ble 22f\n" | ||
| 633 | "cmp x13, #0x3\n" | ||
| 634 | "str q29, [x20, #0x0]\n" | ||
| 635 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 636 | "ble 22f\n" | ||
| 637 | "str q28, [x20, #0x0]\n" | ||
| 638 | "b 22f\n" | ||
| 639 | "19:" // Row tail: Partial output | ||
| 640 | "mov x23, %x[dst]\n" | ||
| 641 | "cmp x13, #0x1\n" | ||
| 642 | "add x22, x23, %x[dst_stride_row]\n" | ||
| 643 | "csel x22, x22, x23, GT\n" | ||
| 644 | "cmp x13, #0x2\n" | ||
| 645 | "add x21, x23, %x[dst_stride_row], LSL #1\n" | ||
| 646 | "csel x21, x21, x22, GT\n" | ||
| 647 | "cmp x13, #0x3\n" | ||
| 648 | "add x20, x21, %x[dst_stride_row]\n" | ||
| 649 | "csel x20, x20, x21, GT\n" | ||
| 650 | "tbz x25, #1, 20f\n" | ||
| 651 | "st1 { v28.d }[0], [x20], #0x8\n" | ||
| 652 | "st1 { v29.d }[0], [x21], #0x8\n" | ||
| 653 | "st1 { v30.d }[0], [x22], #0x8\n" | ||
| 654 | "st1 { v31.d }[0], [x23], #0x8\n" | ||
| 655 | "tbz x25, #0, 21f\n" | ||
| 656 | "st1 { v28.s }[2], [x20]\n" | ||
| 657 | "st1 { v29.s }[2], [x21]\n" | ||
| 658 | "st1 { v30.s }[2], [x22]\n" | ||
| 659 | "st1 { v31.s }[2], [x23]\n" | ||
| 660 | "b 21f\n" | ||
| 661 | "20:" // Row tail: Output block 0: partial_1_0 | ||
| 662 | "st1 { v28.s }[0], [x20]\n" | ||
| 663 | "st1 { v29.s }[0], [x21]\n" | ||
| 664 | "st1 { v30.s }[0], [x22]\n" | ||
| 665 | "st1 { v31.s }[0], [x23]\n" | ||
| 666 | "21:" // Row tail: Output block 0: Done | ||
| 667 | "22:" // Row tail: Output stage exit | ||
| 668 | "subs x25, x25, #0x4\n" | ||
| 669 | "add %x[dst], %x[dst], #0x10\n" | ||
| 670 | "bgt 16b\n" | ||
| 671 | "subs x13, x13, #0x4\n" | ||
| 672 | "add %x[lhs_packed], %x[lhs_packed], x12\n" | ||
| 673 | "mov %x[dst], x24\n" | ||
| 674 | "bgt 15b\n" | ||
| 675 | "23:" // Row tail: Row loop skip | ||
| 676 | : [dst] "+&r"(dst), [lhs_packed] "+&r"(lhs_packed) | ||
| 677 | 47 | : [clamp_vals] "r"(clamp_vals), [dst_stride_row] "r"(dst_stride_row), [m] "r"(m), [n] "r"(n), | |
| 678 | 47 | [num_blocks] "r"(num_blocks), [rhs_packed] "r"(rhs_packed) | |
| 679 | : "cc", "memory", "v0", "v1", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v2", "v20", | ||
| 680 | "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v3", "v30", "v31", "v4", "v5", "v6", "v7", | ||
| 681 | "v8", "v9", "x10", "x11", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "x9"); | ||
| 682 | 47 | } | |
| 683 | |||
| 684 | #endif // Architectural feature check | ||
| 685 |