kai/ukernels/matmul/matmul_clamp_f32_bf16p_bf16p/kai_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot.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 | #include "kai_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot.h" | ||
| 15 | |||
| 16 | #include <stddef.h> | ||
| 17 | #include <stdint.h> | ||
| 18 | |||
| 19 | #include "kai/kai_common.h" | ||
| 20 | |||
| 21 | static const size_t kai_mr = 1; | ||
| 22 | static const size_t kai_nr = 12; | ||
| 23 | static const size_t kai_kr = 4; | ||
| 24 | static const size_t kai_sr = 1; | ||
| 25 | |||
| 26 | static const size_t kai_m_step = 1; | ||
| 27 | static const size_t kai_n_step = 36; | ||
| 28 | |||
| 29 | 1152 | size_t kai_get_m_step_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot(void) { | |
| 30 | 1152 | return kai_m_step; | |
| 31 | } | ||
| 32 | |||
| 33 | 1728 | size_t kai_get_n_step_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot(void) { | |
| 34 | 1728 | return kai_n_step; | |
| 35 | } | ||
| 36 | |||
| 37 | ✗ | size_t kai_get_mr_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot(void) { | |
| 38 | ✗ | return kai_mr; | |
| 39 | } | ||
| 40 | |||
| 41 | 144 | size_t kai_get_nr_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot(void) { | |
| 42 | 144 | return kai_nr; | |
| 43 | } | ||
| 44 | |||
| 45 | 144 | size_t kai_get_kr_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot(void) { | |
| 46 | 144 | return kai_kr; | |
| 47 | } | ||
| 48 | |||
| 49 | 144 | size_t kai_get_sr_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot(void) { | |
| 50 | 144 | return kai_sr; | |
| 51 | } | ||
| 52 | |||
| 53 | 576 | size_t kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot(size_t m_idx, size_t k) { | |
| 54 | − | KAI_ASSUME(m_idx == 0); | |
| 55 | |||
| 56 | 576 | return m_idx * kai_roundup(k, kai_kr) * sizeof(uint16_t); | |
| 57 | } | ||
| 58 | |||
| 59 | 576 | size_t kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot(size_t n_idx, size_t k) { | |
| 60 | − | KAI_ASSUME(n_idx % kai_get_n_step_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot() == 0); | |
| 61 | 576 | return n_idx * (kai_roundup(k, kai_kr) * sizeof(uint16_t) + sizeof(float)); | |
| 62 | } | ||
| 63 | |||
| 64 | 576 | size_t kai_get_dst_offset_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot( | |
| 65 | size_t m_idx, size_t n_idx, size_t dst_stride) { | ||
| 66 | − | KAI_ASSUME(m_idx % kai_get_m_step_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot() == 0); | |
| 67 | − | KAI_ASSUME(n_idx % kai_get_n_step_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot() == 0); | |
| 68 | |||
| 69 | 576 | return (m_idx * dst_stride) + (n_idx * sizeof(float)); | |
| 70 | } | ||
| 71 | |||
| 72 | 576 | size_t kai_get_dst_size_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot(size_t m, size_t n) { | |
| 73 | 576 | return m * n * sizeof(float); | |
| 74 | } | ||
| 75 | |||
| 76 | 582 | void kai_run_matmul_clamp_f32_bf16p1x4_bf16p12x4b_1x36_neon_dot( | |
| 77 | size_t m, size_t n, size_t k, const void* lhs_packed, const void* rhs_packed, void* dst, size_t dst_stride_row, | ||
| 78 | size_t dst_stride_col, float clamp_min, float clamp_max) { | ||
| 79 | 582 | KAI_UNUSED(dst_stride_row); | |
| 80 | 582 | KAI_UNUSED(dst_stride_col); | |
| 81 | |||
| 82 | − | KAI_ASSUME(m == 1); | |
| 83 | |||
| 84 | typedef struct { | ||
| 85 | float maxval; | ||
| 86 | float minval; | ||
| 87 | } KernelArgs; | ||
| 88 | |||
| 89 | 582 | KernelArgs ka; | |
| 90 | 582 | ka.maxval = clamp_max; | |
| 91 | 582 | ka.minval = clamp_min; | |
| 92 | |||
| 93 | 582 | size_t N = n; | |
| 94 | 582 | size_t K = k; | |
| 95 | |||
| 96 | 582 | const void* A_ptr = lhs_packed; | |
| 97 | 582 | const void* B_ptr = rhs_packed; | |
| 98 | 582 | void* output_ptr = dst; | |
| 99 | |||
| 100 | // Add clamping flag | ||
| 101 | 582 | uint64_t flags = 0x2; | |
| 102 | |||
| 103 | 1164 | __asm__ __volatile__( | |
| 104 | "add x26, %x[K], #0x3\n" | ||
| 105 | "mov x20, #0xc\n" | ||
| 106 | "bic x26, x26, #0x3\n" | ||
| 107 | "add x25, %x[N], #0x3\n" | ||
| 108 | "lsr x25, x25, #0x2\n" | ||
| 109 | "lsl x26, x26, #0x1\n" | ||
| 110 | "add x26, x26, #0x4\n" | ||
| 111 | "mul x26, x26, x20\n" | ||
| 112 | "1:" // Column loop | ||
| 113 | "cmp x25, #0x9\n" | ||
| 114 | "bge 89f\n" | ||
| 115 | "cmp x25, #0x7\n" | ||
| 116 | "bgt 78f\n" | ||
| 117 | "beq 67f\n" | ||
| 118 | "cmp x25, #0x5\n" | ||
| 119 | "bgt 56f\n" | ||
| 120 | "beq 45f\n" | ||
| 121 | "cmp x25, #0x3\n" | ||
| 122 | "bgt 34f\n" | ||
| 123 | "beq 23f\n" | ||
| 124 | "cmp x25, #0x1\n" | ||
| 125 | "bgt 12f\n" | ||
| 126 | "ldr q14, [%x[B_ptr], #0x0]\n" | ||
| 127 | "mov x24, %x[K]\n" | ||
| 128 | "movi v16.16b, #0x0\n" | ||
| 129 | "mov x23, %x[A_ptr]\n" | ||
| 130 | "add %x[B_ptr], %x[B_ptr], #0x30\n" | ||
| 131 | "cmp x24, #0x4\n" | ||
| 132 | "zip2 v15.4s, v14.4s, v16.4s\n" | ||
| 133 | "zip1 v14.4s, v14.4s, v16.4s\n" | ||
| 134 | "blt 4f\n" | ||
| 135 | "cmp x24, #0x8\n" | ||
| 136 | "blt 3f\n" | ||
| 137 | "2:" // Width 1: Multiply loop: Main loop head | ||
| 138 | "ld1r { v0.2d }, [x23]\n" | ||
| 139 | "ldr q1, [%x[B_ptr], #0x0]\n" | ||
| 140 | "sub x24, x24, #0x4\n" | ||
| 141 | "add x23, x23, #0x8\n" | ||
| 142 | "ldr q2, [%x[B_ptr], #0x10]\n" | ||
| 143 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 144 | "cmp x24, #0x8\n" | ||
| 145 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 146 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 147 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 148 | ".inst 0x6e40fc2e // bfdot v14.4s, v1.8h, v0.8h\n" | ||
| 149 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 150 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 151 | ".inst 0x6e40fc4f // bfdot v15.4s, v2.8h, v0.8h\n" | ||
| 152 | "bge 2b\n" | ||
| 153 | "3:" // Width 1: Multiply loop: Single iteration only | ||
| 154 | "ld1r { v0.2d }, [x23]\n" | ||
| 155 | "ldr q3, [%x[B_ptr], #0x0]\n" | ||
| 156 | "add x23, x23, #0x8\n" | ||
| 157 | "sub x24, x24, #0x4\n" | ||
| 158 | "ldr q4, [%x[B_ptr], #0x10]\n" | ||
| 159 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 160 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 161 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 162 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 163 | ".inst 0x6e40fc6e // bfdot v14.4s, v3.8h, v0.8h\n" | ||
| 164 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 165 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 166 | ".inst 0x6e40fc8f // bfdot v15.4s, v4.8h, v0.8h\n" | ||
| 167 | "4:" // Width 1: Multiply loop: Main loop skip | ||
| 168 | "cbz x24, 7f\n" | ||
| 169 | "tbz x24, #1, 5f\n" | ||
| 170 | "ldr s0, [x23], #0x4\n" | ||
| 171 | "tbz x24, #0, 6f\n" | ||
| 172 | "ld1 { v0.h }[2], [x23]\n" | ||
| 173 | "b 6f\n" | ||
| 174 | "5:" // Width 1: Multiply loop: Ragged operand read: partial_1_0 | ||
| 175 | "ldr h0, [x23, #0x0]\n" | ||
| 176 | "6:" // Width 1: Multiply loop: Ragged operand read: Done | ||
| 177 | "ldr q5, [%x[B_ptr], #0x0]\n" | ||
| 178 | "ldr q6, [%x[B_ptr], #0x10]\n" | ||
| 179 | "dup v0.2d, v0.d[0]\n" | ||
| 180 | "add %x[B_ptr], %x[B_ptr], #0x90\n" | ||
| 181 | ".inst 0x6e40fcae // bfdot v14.4s, v5.8h, v0.8h\n" | ||
| 182 | ".inst 0x6e40fccf // bfdot v15.4s, v6.8h, v0.8h\n" | ||
| 183 | "7:" // Width 1: Multiply loop: No odd multiplies | ||
| 184 | "prfm pstl1keep, [%x[output_ptr], #0x0]\n" | ||
| 185 | "faddp v14.4s, v14.4s, v15.4s\n" | ||
| 186 | "tbz %x[flags], #1, 8f\n" | ||
| 187 | "add x21, %x[args_ptr], %[offset_max]\n" | ||
| 188 | "add x20, %x[args_ptr], %[offset_min]\n" | ||
| 189 | "ld1r { v17.4s }, [x21]\n" | ||
| 190 | "ld1r { v16.4s }, [x20]\n" | ||
| 191 | "fmin v14.4s, v14.4s, v17.4s\n" | ||
| 192 | "fmax v14.4s, v14.4s, v16.4s\n" | ||
| 193 | "8:" // Width 1: No activation | ||
| 194 | "cmp %x[N], #0x4\n" | ||
| 195 | "blt 9f\n" | ||
| 196 | "str q14, [%x[output_ptr], #0x0]\n" | ||
| 197 | "add %x[output_ptr], %x[output_ptr], #0x10\n" | ||
| 198 | "b 11f\n" | ||
| 199 | "9:" // Width 1: Partial writeback | ||
| 200 | "tbz %x[N], #1, 10f\n" | ||
| 201 | "str d14, [%x[output_ptr]], #0x8\n" | ||
| 202 | "tbz %x[N], #0, 11f\n" | ||
| 203 | "st1 { v14.s }[2], [%x[output_ptr]]\n" | ||
| 204 | "b 11f\n" | ||
| 205 | "10:" // Width 1: Partial direct writeback: partial_1_0 | ||
| 206 | "str s14, [%x[output_ptr], #0x0]\n" | ||
| 207 | "11:" // Width 1: Writeback done | ||
| 208 | "b 100f\n" | ||
| 209 | "12:" // Width 2 | ||
| 210 | "ldr q14, [%x[B_ptr], #0x0]\n" | ||
| 211 | "ldr q15, [%x[B_ptr], #0x10]\n" | ||
| 212 | "mov x24, %x[K]\n" | ||
| 213 | "movi v18.16b, #0x0\n" | ||
| 214 | "mov x23, %x[A_ptr]\n" | ||
| 215 | "add %x[B_ptr], %x[B_ptr], #0x30\n" | ||
| 216 | "cmp x24, #0x4\n" | ||
| 217 | "zip2 v17.4s, v15.4s, v18.4s\n" | ||
| 218 | "zip1 v16.4s, v15.4s, v18.4s\n" | ||
| 219 | "zip2 v15.4s, v14.4s, v18.4s\n" | ||
| 220 | "zip1 v14.4s, v14.4s, v18.4s\n" | ||
| 221 | "blt 15f\n" | ||
| 222 | "cmp x24, #0x8\n" | ||
| 223 | "blt 14f\n" | ||
| 224 | "13:" // Width 2: Multiply loop: Main loop head | ||
| 225 | "ld1r { v0.2d }, [x23]\n" | ||
| 226 | "ldr q1, [%x[B_ptr], #0x0]\n" | ||
| 227 | "sub x24, x24, #0x4\n" | ||
| 228 | "add x23, x23, #0x8\n" | ||
| 229 | "ldr q2, [%x[B_ptr], #0x10]\n" | ||
| 230 | "ldr q3, [%x[B_ptr], #0x20]\n" | ||
| 231 | "cmp x24, #0x8\n" | ||
| 232 | "ldr q4, [%x[B_ptr], #0x30]\n" | ||
| 233 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 234 | ".inst 0x6e40fc2e // bfdot v14.4s, v1.8h, v0.8h\n" | ||
| 235 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 236 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 237 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 238 | ".inst 0x6e40fc4f // bfdot v15.4s, v2.8h, v0.8h\n" | ||
| 239 | ".inst 0x6e40fc70 // bfdot v16.4s, v3.8h, v0.8h\n" | ||
| 240 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 241 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 242 | ".inst 0x6e40fc91 // bfdot v17.4s, v4.8h, v0.8h\n" | ||
| 243 | "bge 13b\n" | ||
| 244 | "14:" // Width 2: Multiply loop: Single iteration only | ||
| 245 | "ld1r { v0.2d }, [x23]\n" | ||
| 246 | "ldr q5, [%x[B_ptr], #0x0]\n" | ||
| 247 | "add x23, x23, #0x8\n" | ||
| 248 | "sub x24, x24, #0x4\n" | ||
| 249 | "ldr q6, [%x[B_ptr], #0x10]\n" | ||
| 250 | "ldr q7, [%x[B_ptr], #0x20]\n" | ||
| 251 | "ldr q8, [%x[B_ptr], #0x30]\n" | ||
| 252 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 253 | ".inst 0x6e40fcae // bfdot v14.4s, v5.8h, v0.8h\n" | ||
| 254 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 255 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 256 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 257 | ".inst 0x6e40fccf // bfdot v15.4s, v6.8h, v0.8h\n" | ||
| 258 | ".inst 0x6e40fcf0 // bfdot v16.4s, v7.8h, v0.8h\n" | ||
| 259 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 260 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 261 | ".inst 0x6e40fd11 // bfdot v17.4s, v8.8h, v0.8h\n" | ||
| 262 | "15:" // Width 2: Multiply loop: Main loop skip | ||
| 263 | "cbz x24, 18f\n" | ||
| 264 | "tbz x24, #1, 16f\n" | ||
| 265 | "ldr s0, [x23], #0x4\n" | ||
| 266 | "tbz x24, #0, 17f\n" | ||
| 267 | "ld1 { v0.h }[2], [x23]\n" | ||
| 268 | "b 17f\n" | ||
| 269 | "16:" // Width 2: Multiply loop: Ragged operand read: partial_1_0 | ||
| 270 | "ldr h0, [x23, #0x0]\n" | ||
| 271 | "17:" // Width 2: Multiply loop: Ragged operand read: Done | ||
| 272 | "ldr q9, [%x[B_ptr], #0x0]\n" | ||
| 273 | "ldr q10, [%x[B_ptr], #0x10]\n" | ||
| 274 | "dup v0.2d, v0.d[0]\n" | ||
| 275 | "ldr q11, [%x[B_ptr], #0x20]\n" | ||
| 276 | "ldr q12, [%x[B_ptr], #0x30]\n" | ||
| 277 | "add %x[B_ptr], %x[B_ptr], #0x90\n" | ||
| 278 | ".inst 0x6e40fd2e // bfdot v14.4s, v9.8h, v0.8h\n" | ||
| 279 | ".inst 0x6e40fd4f // bfdot v15.4s, v10.8h, v0.8h\n" | ||
| 280 | ".inst 0x6e40fd70 // bfdot v16.4s, v11.8h, v0.8h\n" | ||
| 281 | ".inst 0x6e40fd91 // bfdot v17.4s, v12.8h, v0.8h\n" | ||
| 282 | "18:" // Width 2: Multiply loop: No odd multiplies | ||
| 283 | "prfm pstl1keep, [%x[output_ptr], #0x0]\n" | ||
| 284 | "faddp v14.4s, v14.4s, v15.4s\n" | ||
| 285 | "faddp v15.4s, v16.4s, v17.4s\n" | ||
| 286 | "tbz %x[flags], #1, 19f\n" | ||
| 287 | "add x21, %x[args_ptr], %[offset_max]\n" | ||
| 288 | "add x20, %x[args_ptr], %[offset_min]\n" | ||
| 289 | "ld1r { v17.4s }, [x21]\n" | ||
| 290 | "ld1r { v16.4s }, [x20]\n" | ||
| 291 | "fmin v14.4s, v14.4s, v17.4s\n" | ||
| 292 | "fmin v15.4s, v15.4s, v17.4s\n" | ||
| 293 | "fmax v14.4s, v14.4s, v16.4s\n" | ||
| 294 | "fmax v15.4s, v15.4s, v16.4s\n" | ||
| 295 | "19:" // Width 2: No activation | ||
| 296 | "cmp %x[N], #0x8\n" | ||
| 297 | "str q14, [%x[output_ptr], #0x0]\n" | ||
| 298 | "add %x[output_ptr], %x[output_ptr], #0x10\n" | ||
| 299 | "blt 20f\n" | ||
| 300 | "str q15, [%x[output_ptr], #0x0]\n" | ||
| 301 | "add %x[output_ptr], %x[output_ptr], #0x10\n" | ||
| 302 | "b 22f\n" | ||
| 303 | "20:" // Width 2: Partial writeback | ||
| 304 | "tbz %x[N], #1, 21f\n" | ||
| 305 | "str d15, [%x[output_ptr]], #0x8\n" | ||
| 306 | "tbz %x[N], #0, 22f\n" | ||
| 307 | "st1 { v15.s }[2], [%x[output_ptr]]\n" | ||
| 308 | "b 22f\n" | ||
| 309 | "21:" // Width 2: Partial direct writeback: partial_1_4 | ||
| 310 | "tbz %x[N], #0, 22f\n" | ||
| 311 | "str s15, [%x[output_ptr], #0x0]\n" | ||
| 312 | "22:" // Width 2: Writeback done | ||
| 313 | "b 100f\n" | ||
| 314 | "23:" // Width 3 | ||
| 315 | "ldr q14, [%x[B_ptr], #0x0]\n" | ||
| 316 | "ldr q15, [%x[B_ptr], #0x10]\n" | ||
| 317 | "mov x24, %x[K]\n" | ||
| 318 | "movi v20.16b, #0x0\n" | ||
| 319 | "ldr q16, [%x[B_ptr], #0x20]\n" | ||
| 320 | "mov x23, %x[A_ptr]\n" | ||
| 321 | "add %x[B_ptr], %x[B_ptr], #0x30\n" | ||
| 322 | "cmp x24, #0x4\n" | ||
| 323 | "zip2 v17.4s, v15.4s, v20.4s\n" | ||
| 324 | "zip2 v19.4s, v16.4s, v20.4s\n" | ||
| 325 | "zip1 v18.4s, v16.4s, v20.4s\n" | ||
| 326 | "zip1 v16.4s, v15.4s, v20.4s\n" | ||
| 327 | "zip2 v15.4s, v14.4s, v20.4s\n" | ||
| 328 | "zip1 v14.4s, v14.4s, v20.4s\n" | ||
| 329 | "blt 26f\n" | ||
| 330 | "cmp x24, #0x8\n" | ||
| 331 | "blt 25f\n" | ||
| 332 | "24:" // Width 3: Multiply loop: Main loop head | ||
| 333 | "ld1r { v0.2d }, [x23]\n" | ||
| 334 | "ldr q1, [%x[B_ptr], #0x0]\n" | ||
| 335 | "sub x24, x24, #0x4\n" | ||
| 336 | "add x23, x23, #0x8\n" | ||
| 337 | "ldr q2, [%x[B_ptr], #0x10]\n" | ||
| 338 | "ldr q3, [%x[B_ptr], #0x20]\n" | ||
| 339 | "cmp x24, #0x8\n" | ||
| 340 | "ldr q4, [%x[B_ptr], #0x30]\n" | ||
| 341 | "ldr q5, [%x[B_ptr], #0x40]\n" | ||
| 342 | "ldr q6, [%x[B_ptr], #0x50]\n" | ||
| 343 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 344 | ".inst 0x6e40fc2e // bfdot v14.4s, v1.8h, v0.8h\n" | ||
| 345 | ".inst 0x6e40fc4f // bfdot v15.4s, v2.8h, v0.8h\n" | ||
| 346 | ".inst 0x6e40fc70 // bfdot v16.4s, v3.8h, v0.8h\n" | ||
| 347 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 348 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 349 | ".inst 0x6e40fc91 // bfdot v17.4s, v4.8h, v0.8h\n" | ||
| 350 | ".inst 0x6e40fcb2 // bfdot v18.4s, v5.8h, v0.8h\n" | ||
| 351 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 352 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 353 | ".inst 0x6e40fcd3 // bfdot v19.4s, v6.8h, v0.8h\n" | ||
| 354 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 355 | "bge 24b\n" | ||
| 356 | "25:" // Width 3: Multiply loop: Single iteration only | ||
| 357 | "ld1r { v0.2d }, [x23]\n" | ||
| 358 | "ldr q7, [%x[B_ptr], #0x0]\n" | ||
| 359 | "add x23, x23, #0x8\n" | ||
| 360 | "sub x24, x24, #0x4\n" | ||
| 361 | "ldr q8, [%x[B_ptr], #0x10]\n" | ||
| 362 | "ldr q9, [%x[B_ptr], #0x20]\n" | ||
| 363 | "ldr q10, [%x[B_ptr], #0x30]\n" | ||
| 364 | "ldr q11, [%x[B_ptr], #0x40]\n" | ||
| 365 | "ldr q12, [%x[B_ptr], #0x50]\n" | ||
| 366 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 367 | ".inst 0x6e40fcee // bfdot v14.4s, v7.8h, v0.8h\n" | ||
| 368 | ".inst 0x6e40fd0f // bfdot v15.4s, v8.8h, v0.8h\n" | ||
| 369 | ".inst 0x6e40fd30 // bfdot v16.4s, v9.8h, v0.8h\n" | ||
| 370 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 371 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 372 | ".inst 0x6e40fd51 // bfdot v17.4s, v10.8h, v0.8h\n" | ||
| 373 | ".inst 0x6e40fd72 // bfdot v18.4s, v11.8h, v0.8h\n" | ||
| 374 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 375 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 376 | ".inst 0x6e40fd93 // bfdot v19.4s, v12.8h, v0.8h\n" | ||
| 377 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 378 | "26:" // Width 3: Multiply loop: Main loop skip | ||
| 379 | "cbz x24, 29f\n" | ||
| 380 | "tbz x24, #1, 27f\n" | ||
| 381 | "ldr s0, [x23], #0x4\n" | ||
| 382 | "tbz x24, #0, 28f\n" | ||
| 383 | "ld1 { v0.h }[2], [x23]\n" | ||
| 384 | "b 28f\n" | ||
| 385 | "27:" // Width 3: Multiply loop: Ragged operand read: partial_1_0 | ||
| 386 | "ldr h0, [x23, #0x0]\n" | ||
| 387 | "28:" // Width 3: Multiply loop: Ragged operand read: Done | ||
| 388 | "ldr q13, [%x[B_ptr], #0x0]\n" | ||
| 389 | "ldr q1, [%x[B_ptr], #0x10]\n" | ||
| 390 | "dup v0.2d, v0.d[0]\n" | ||
| 391 | "ldr q2, [%x[B_ptr], #0x20]\n" | ||
| 392 | "ldr q3, [%x[B_ptr], #0x30]\n" | ||
| 393 | "ldr q4, [%x[B_ptr], #0x40]\n" | ||
| 394 | "ldr q5, [%x[B_ptr], #0x50]\n" | ||
| 395 | "add %x[B_ptr], %x[B_ptr], #0x90\n" | ||
| 396 | ".inst 0x6e40fdae // bfdot v14.4s, v13.8h, v0.8h\n" | ||
| 397 | ".inst 0x6e40fc2f // bfdot v15.4s, v1.8h, v0.8h\n" | ||
| 398 | ".inst 0x6e40fc50 // bfdot v16.4s, v2.8h, v0.8h\n" | ||
| 399 | ".inst 0x6e40fc71 // bfdot v17.4s, v3.8h, v0.8h\n" | ||
| 400 | ".inst 0x6e40fc92 // bfdot v18.4s, v4.8h, v0.8h\n" | ||
| 401 | ".inst 0x6e40fcb3 // bfdot v19.4s, v5.8h, v0.8h\n" | ||
| 402 | "29:" // Width 3: Multiply loop: No odd multiplies | ||
| 403 | "prfm pstl1keep, [%x[output_ptr], #0x0]\n" | ||
| 404 | "faddp v14.4s, v14.4s, v15.4s\n" | ||
| 405 | "faddp v15.4s, v16.4s, v17.4s\n" | ||
| 406 | "faddp v16.4s, v18.4s, v19.4s\n" | ||
| 407 | "tbz %x[flags], #1, 30f\n" | ||
| 408 | "add x21, %x[args_ptr], %[offset_max]\n" | ||
| 409 | "add x20, %x[args_ptr], %[offset_min]\n" | ||
| 410 | "ld1r { v18.4s }, [x21]\n" | ||
| 411 | "ld1r { v17.4s }, [x20]\n" | ||
| 412 | "fmin v14.4s, v14.4s, v18.4s\n" | ||
| 413 | "fmin v15.4s, v15.4s, v18.4s\n" | ||
| 414 | "fmin v16.4s, v16.4s, v18.4s\n" | ||
| 415 | "fmax v14.4s, v14.4s, v17.4s\n" | ||
| 416 | "fmax v15.4s, v15.4s, v17.4s\n" | ||
| 417 | "fmax v16.4s, v16.4s, v17.4s\n" | ||
| 418 | "30:" // Width 3: No activation | ||
| 419 | "cmp %x[N], #0xc\n" | ||
| 420 | "str q14, [%x[output_ptr], #0x0]\n" | ||
| 421 | "str q15, [%x[output_ptr], #0x10]\n" | ||
| 422 | "add %x[output_ptr], %x[output_ptr], #0x20\n" | ||
| 423 | "blt 31f\n" | ||
| 424 | "str q16, [%x[output_ptr], #0x0]\n" | ||
| 425 | "add %x[output_ptr], %x[output_ptr], #0x10\n" | ||
| 426 | "b 33f\n" | ||
| 427 | "31:" // Width 3: Partial writeback | ||
| 428 | "tbz %x[N], #1, 32f\n" | ||
| 429 | "str d16, [%x[output_ptr]], #0x8\n" | ||
| 430 | "tbz %x[N], #0, 33f\n" | ||
| 431 | "st1 { v16.s }[2], [%x[output_ptr]]\n" | ||
| 432 | "b 33f\n" | ||
| 433 | "32:" // Width 3: Partial direct writeback: partial_1_8 | ||
| 434 | "tbz %x[N], #0, 33f\n" | ||
| 435 | "str s16, [%x[output_ptr], #0x0]\n" | ||
| 436 | "33:" // Width 3: Writeback done | ||
| 437 | "b 100f\n" | ||
| 438 | "34:" // Width 4 | ||
| 439 | "ldr q14, [%x[B_ptr], #0x0]\n" | ||
| 440 | "ldr q15, [%x[B_ptr], #0x10]\n" | ||
| 441 | "mov x24, %x[K]\n" | ||
| 442 | "add x20, %x[B_ptr], x26\n" | ||
| 443 | "ldr q16, [%x[B_ptr], #0x20]\n" | ||
| 444 | "ldr q17, [x20, #0x0]\n" | ||
| 445 | "movi v22.16b, #0x0\n" | ||
| 446 | "mov x23, %x[A_ptr]\n" | ||
| 447 | "add %x[B_ptr], %x[B_ptr], #0x30\n" | ||
| 448 | "add x20, x20, #0x30\n" | ||
| 449 | "cmp x24, #0x4\n" | ||
| 450 | "zip2 v21.4s, v17.4s, v22.4s\n" | ||
| 451 | "zip1 v20.4s, v17.4s, v22.4s\n" | ||
| 452 | "zip2 v19.4s, v16.4s, v22.4s\n" | ||
| 453 | "zip1 v18.4s, v16.4s, v22.4s\n" | ||
| 454 | "zip2 v17.4s, v15.4s, v22.4s\n" | ||
| 455 | "zip1 v16.4s, v15.4s, v22.4s\n" | ||
| 456 | "zip2 v15.4s, v14.4s, v22.4s\n" | ||
| 457 | "zip1 v14.4s, v14.4s, v22.4s\n" | ||
| 458 | "blt 37f\n" | ||
| 459 | "cmp x24, #0x8\n" | ||
| 460 | "blt 36f\n" | ||
| 461 | "35:" // Width 4: Multiply loop: Main loop head | ||
| 462 | "ld1r { v0.2d }, [x23]\n" | ||
| 463 | "ldr q1, [%x[B_ptr], #0x0]\n" | ||
| 464 | "sub x24, x24, #0x4\n" | ||
| 465 | "add x23, x23, #0x8\n" | ||
| 466 | "ldr q2, [%x[B_ptr], #0x10]\n" | ||
| 467 | "ldr q3, [%x[B_ptr], #0x20]\n" | ||
| 468 | "cmp x24, #0x8\n" | ||
| 469 | "ldr q4, [%x[B_ptr], #0x30]\n" | ||
| 470 | "ldr q5, [%x[B_ptr], #0x40]\n" | ||
| 471 | "ldr q6, [%x[B_ptr], #0x50]\n" | ||
| 472 | "ldr q7, [x20, #0x0]\n" | ||
| 473 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 474 | ".inst 0x6e40fc2e // bfdot v14.4s, v1.8h, v0.8h\n" | ||
| 475 | "ldr q8, [x20, #0x10]\n" | ||
| 476 | ".inst 0x6e40fc4f // bfdot v15.4s, v2.8h, v0.8h\n" | ||
| 477 | ".inst 0x6e40fc70 // bfdot v16.4s, v3.8h, v0.8h\n" | ||
| 478 | "add x20, x20, #0x60\n" | ||
| 479 | ".inst 0x6e40fc91 // bfdot v17.4s, v4.8h, v0.8h\n" | ||
| 480 | ".inst 0x6e40fcb2 // bfdot v18.4s, v5.8h, v0.8h\n" | ||
| 481 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 482 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 483 | ".inst 0x6e40fcd3 // bfdot v19.4s, v6.8h, v0.8h\n" | ||
| 484 | ".inst 0x6e40fcf4 // bfdot v20.4s, v7.8h, v0.8h\n" | ||
| 485 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 486 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 487 | ".inst 0x6e40fd15 // bfdot v21.4s, v8.8h, v0.8h\n" | ||
| 488 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 489 | "bge 35b\n" | ||
| 490 | "36:" // Width 4: Multiply loop: Single iteration only | ||
| 491 | "ld1r { v0.2d }, [x23]\n" | ||
| 492 | "ldr q9, [%x[B_ptr], #0x0]\n" | ||
| 493 | "add x23, x23, #0x8\n" | ||
| 494 | "sub x24, x24, #0x4\n" | ||
| 495 | "ldr q10, [%x[B_ptr], #0x10]\n" | ||
| 496 | "ldr q11, [%x[B_ptr], #0x20]\n" | ||
| 497 | "ldr q12, [%x[B_ptr], #0x30]\n" | ||
| 498 | "ldr q13, [%x[B_ptr], #0x40]\n" | ||
| 499 | "ldr q1, [%x[B_ptr], #0x50]\n" | ||
| 500 | "ldr q2, [x20, #0x0]\n" | ||
| 501 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 502 | ".inst 0x6e40fd2e // bfdot v14.4s, v9.8h, v0.8h\n" | ||
| 503 | "ldr q3, [x20, #0x10]\n" | ||
| 504 | ".inst 0x6e40fd4f // bfdot v15.4s, v10.8h, v0.8h\n" | ||
| 505 | ".inst 0x6e40fd70 // bfdot v16.4s, v11.8h, v0.8h\n" | ||
| 506 | "add x20, x20, #0x60\n" | ||
| 507 | ".inst 0x6e40fd91 // bfdot v17.4s, v12.8h, v0.8h\n" | ||
| 508 | ".inst 0x6e40fdb2 // bfdot v18.4s, v13.8h, v0.8h\n" | ||
| 509 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 510 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 511 | ".inst 0x6e40fc33 // bfdot v19.4s, v1.8h, v0.8h\n" | ||
| 512 | ".inst 0x6e40fc54 // bfdot v20.4s, v2.8h, v0.8h\n" | ||
| 513 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 514 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 515 | ".inst 0x6e40fc75 // bfdot v21.4s, v3.8h, v0.8h\n" | ||
| 516 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 517 | "37:" // Width 4: Multiply loop: Main loop skip | ||
| 518 | "cbz x24, 40f\n" | ||
| 519 | "tbz x24, #1, 38f\n" | ||
| 520 | "ldr s0, [x23], #0x4\n" | ||
| 521 | "tbz x24, #0, 39f\n" | ||
| 522 | "ld1 { v0.h }[2], [x23]\n" | ||
| 523 | "b 39f\n" | ||
| 524 | "38:" // Width 4: Multiply loop: Ragged operand read: partial_1_0 | ||
| 525 | "ldr h0, [x23, #0x0]\n" | ||
| 526 | "39:" // Width 4: Multiply loop: Ragged operand read: Done | ||
| 527 | "ldr q4, [%x[B_ptr], #0x0]\n" | ||
| 528 | "ldr q5, [%x[B_ptr], #0x10]\n" | ||
| 529 | "dup v0.2d, v0.d[0]\n" | ||
| 530 | "ldr q6, [%x[B_ptr], #0x20]\n" | ||
| 531 | "ldr q7, [%x[B_ptr], #0x30]\n" | ||
| 532 | "ldr q8, [%x[B_ptr], #0x40]\n" | ||
| 533 | "ldr q9, [%x[B_ptr], #0x50]\n" | ||
| 534 | "add %x[B_ptr], %x[B_ptr], #0x90\n" | ||
| 535 | "ldr q10, [x20, #0x0]\n" | ||
| 536 | "ldr q11, [x20, #0x10]\n" | ||
| 537 | ".inst 0x6e40fc8e // bfdot v14.4s, v4.8h, v0.8h\n" | ||
| 538 | ".inst 0x6e40fcaf // bfdot v15.4s, v5.8h, v0.8h\n" | ||
| 539 | ".inst 0x6e40fcd0 // bfdot v16.4s, v6.8h, v0.8h\n" | ||
| 540 | ".inst 0x6e40fcf1 // bfdot v17.4s, v7.8h, v0.8h\n" | ||
| 541 | ".inst 0x6e40fd12 // bfdot v18.4s, v8.8h, v0.8h\n" | ||
| 542 | ".inst 0x6e40fd33 // bfdot v19.4s, v9.8h, v0.8h\n" | ||
| 543 | ".inst 0x6e40fd54 // bfdot v20.4s, v10.8h, v0.8h\n" | ||
| 544 | ".inst 0x6e40fd75 // bfdot v21.4s, v11.8h, v0.8h\n" | ||
| 545 | "40:" // Width 4: Multiply loop: No odd multiplies | ||
| 546 | "prfm pstl1keep, [%x[output_ptr], #0x0]\n" | ||
| 547 | "faddp v14.4s, v14.4s, v15.4s\n" | ||
| 548 | "faddp v15.4s, v16.4s, v17.4s\n" | ||
| 549 | "faddp v16.4s, v18.4s, v19.4s\n" | ||
| 550 | "faddp v17.4s, v20.4s, v21.4s\n" | ||
| 551 | "tbz %x[flags], #1, 41f\n" | ||
| 552 | "add x21, %x[args_ptr], %[offset_max]\n" | ||
| 553 | "add x20, %x[args_ptr], %[offset_min]\n" | ||
| 554 | "ld1r { v19.4s }, [x21]\n" | ||
| 555 | "ld1r { v18.4s }, [x20]\n" | ||
| 556 | "fmin v14.4s, v14.4s, v19.4s\n" | ||
| 557 | "fmin v15.4s, v15.4s, v19.4s\n" | ||
| 558 | "fmin v16.4s, v16.4s, v19.4s\n" | ||
| 559 | "fmin v17.4s, v17.4s, v19.4s\n" | ||
| 560 | "fmax v14.4s, v14.4s, v18.4s\n" | ||
| 561 | "fmax v15.4s, v15.4s, v18.4s\n" | ||
| 562 | "fmax v16.4s, v16.4s, v18.4s\n" | ||
| 563 | "fmax v17.4s, v17.4s, v18.4s\n" | ||
| 564 | "41:" // Width 4: No activation | ||
| 565 | "cmp %x[N], #0x10\n" | ||
| 566 | "str q14, [%x[output_ptr], #0x0]\n" | ||
| 567 | "str q15, [%x[output_ptr], #0x10]\n" | ||
| 568 | "str q16, [%x[output_ptr], #0x20]\n" | ||
| 569 | "add %x[output_ptr], %x[output_ptr], #0x30\n" | ||
| 570 | "blt 42f\n" | ||
| 571 | "str q17, [%x[output_ptr], #0x0]\n" | ||
| 572 | "add %x[output_ptr], %x[output_ptr], #0x10\n" | ||
| 573 | "b 44f\n" | ||
| 574 | "42:" // Width 4: Partial writeback | ||
| 575 | "tbz %x[N], #1, 43f\n" | ||
| 576 | "str d17, [%x[output_ptr]], #0x8\n" | ||
| 577 | "tbz %x[N], #0, 44f\n" | ||
| 578 | "st1 { v17.s }[2], [%x[output_ptr]]\n" | ||
| 579 | "b 44f\n" | ||
| 580 | "43:" // Width 4: Partial direct writeback: partial_1_12 | ||
| 581 | "tbz %x[N], #0, 44f\n" | ||
| 582 | "str s17, [%x[output_ptr], #0x0]\n" | ||
| 583 | "44:" // Width 4: Writeback done | ||
| 584 | "b 100f\n" | ||
| 585 | "45:" // Width 5 | ||
| 586 | "ldr q14, [%x[B_ptr], #0x0]\n" | ||
| 587 | "ldr q15, [%x[B_ptr], #0x10]\n" | ||
| 588 | "mov x24, %x[K]\n" | ||
| 589 | "add x20, %x[B_ptr], x26\n" | ||
| 590 | "ldr q16, [%x[B_ptr], #0x20]\n" | ||
| 591 | "ldr q17, [x20, #0x0]\n" | ||
| 592 | "movi v24.16b, #0x0\n" | ||
| 593 | "mov x23, %x[A_ptr]\n" | ||
| 594 | "ldr q18, [x20, #0x10]\n" | ||
| 595 | "add %x[B_ptr], %x[B_ptr], #0x30\n" | ||
| 596 | "add x20, x20, #0x30\n" | ||
| 597 | "cmp x24, #0x4\n" | ||
| 598 | "zip2 v21.4s, v17.4s, v24.4s\n" | ||
| 599 | "zip1 v20.4s, v17.4s, v24.4s\n" | ||
| 600 | "zip2 v19.4s, v16.4s, v24.4s\n" | ||
| 601 | "zip2 v17.4s, v15.4s, v24.4s\n" | ||
| 602 | "zip2 v23.4s, v18.4s, v24.4s\n" | ||
| 603 | "zip1 v22.4s, v18.4s, v24.4s\n" | ||
| 604 | "zip1 v18.4s, v16.4s, v24.4s\n" | ||
| 605 | "zip1 v16.4s, v15.4s, v24.4s\n" | ||
| 606 | "zip2 v15.4s, v14.4s, v24.4s\n" | ||
| 607 | "zip1 v14.4s, v14.4s, v24.4s\n" | ||
| 608 | "blt 48f\n" | ||
| 609 | "cmp x24, #0x8\n" | ||
| 610 | "blt 47f\n" | ||
| 611 | "46:" // Width 5: Multiply loop: Main loop head | ||
| 612 | "ld1r { v0.2d }, [x23]\n" | ||
| 613 | "ldr q1, [%x[B_ptr], #0x0]\n" | ||
| 614 | "sub x24, x24, #0x4\n" | ||
| 615 | "add x23, x23, #0x8\n" | ||
| 616 | "ldr q2, [%x[B_ptr], #0x10]\n" | ||
| 617 | "ldr q3, [%x[B_ptr], #0x20]\n" | ||
| 618 | "cmp x24, #0x8\n" | ||
| 619 | "ldr q4, [%x[B_ptr], #0x30]\n" | ||
| 620 | "ldr q5, [%x[B_ptr], #0x40]\n" | ||
| 621 | "ldr q6, [%x[B_ptr], #0x50]\n" | ||
| 622 | "ldr q7, [x20, #0x0]\n" | ||
| 623 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 624 | ".inst 0x6e40fc2e // bfdot v14.4s, v1.8h, v0.8h\n" | ||
| 625 | "ldr q8, [x20, #0x10]\n" | ||
| 626 | "ldr q9, [x20, #0x20]\n" | ||
| 627 | ".inst 0x6e40fc4f // bfdot v15.4s, v2.8h, v0.8h\n" | ||
| 628 | ".inst 0x6e40fc70 // bfdot v16.4s, v3.8h, v0.8h\n" | ||
| 629 | "ldr q10, [x20, #0x30]\n" | ||
| 630 | ".inst 0x6e40fc91 // bfdot v17.4s, v4.8h, v0.8h\n" | ||
| 631 | ".inst 0x6e40fcb2 // bfdot v18.4s, v5.8h, v0.8h\n" | ||
| 632 | "add x20, x20, #0x60\n" | ||
| 633 | ".inst 0x6e40fcd3 // bfdot v19.4s, v6.8h, v0.8h\n" | ||
| 634 | ".inst 0x6e40fcf4 // bfdot v20.4s, v7.8h, v0.8h\n" | ||
| 635 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 636 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 637 | ".inst 0x6e40fd15 // bfdot v21.4s, v8.8h, v0.8h\n" | ||
| 638 | ".inst 0x6e40fd36 // bfdot v22.4s, v9.8h, v0.8h\n" | ||
| 639 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 640 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 641 | ".inst 0x6e40fd57 // bfdot v23.4s, v10.8h, v0.8h\n" | ||
| 642 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 643 | "bge 46b\n" | ||
| 644 | "47:" // Width 5: Multiply loop: Single iteration only | ||
| 645 | "ld1r { v0.2d }, [x23]\n" | ||
| 646 | "ldr q11, [%x[B_ptr], #0x0]\n" | ||
| 647 | "add x23, x23, #0x8\n" | ||
| 648 | "sub x24, x24, #0x4\n" | ||
| 649 | "ldr q12, [%x[B_ptr], #0x10]\n" | ||
| 650 | "ldr q13, [%x[B_ptr], #0x20]\n" | ||
| 651 | "ldr q1, [%x[B_ptr], #0x30]\n" | ||
| 652 | "ldr q2, [%x[B_ptr], #0x40]\n" | ||
| 653 | "ldr q3, [%x[B_ptr], #0x50]\n" | ||
| 654 | "ldr q4, [x20, #0x0]\n" | ||
| 655 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 656 | ".inst 0x6e40fd6e // bfdot v14.4s, v11.8h, v0.8h\n" | ||
| 657 | "ldr q5, [x20, #0x10]\n" | ||
| 658 | "ldr q6, [x20, #0x20]\n" | ||
| 659 | ".inst 0x6e40fd8f // bfdot v15.4s, v12.8h, v0.8h\n" | ||
| 660 | ".inst 0x6e40fdb0 // bfdot v16.4s, v13.8h, v0.8h\n" | ||
| 661 | "ldr q7, [x20, #0x30]\n" | ||
| 662 | ".inst 0x6e40fc31 // bfdot v17.4s, v1.8h, v0.8h\n" | ||
| 663 | ".inst 0x6e40fc52 // bfdot v18.4s, v2.8h, v0.8h\n" | ||
| 664 | "add x20, x20, #0x60\n" | ||
| 665 | ".inst 0x6e40fc73 // bfdot v19.4s, v3.8h, v0.8h\n" | ||
| 666 | ".inst 0x6e40fc94 // bfdot v20.4s, v4.8h, v0.8h\n" | ||
| 667 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 668 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 669 | ".inst 0x6e40fcb5 // bfdot v21.4s, v5.8h, v0.8h\n" | ||
| 670 | ".inst 0x6e40fcd6 // bfdot v22.4s, v6.8h, v0.8h\n" | ||
| 671 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 672 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 673 | ".inst 0x6e40fcf7 // bfdot v23.4s, v7.8h, v0.8h\n" | ||
| 674 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 675 | "48:" // Width 5: Multiply loop: Main loop skip | ||
| 676 | "cbz x24, 51f\n" | ||
| 677 | "tbz x24, #1, 49f\n" | ||
| 678 | "ldr s0, [x23], #0x4\n" | ||
| 679 | "tbz x24, #0, 50f\n" | ||
| 680 | "ld1 { v0.h }[2], [x23]\n" | ||
| 681 | "b 50f\n" | ||
| 682 | "49:" // Width 5: Multiply loop: Ragged operand read: partial_1_0 | ||
| 683 | "ldr h0, [x23, #0x0]\n" | ||
| 684 | "50:" // Width 5: Multiply loop: Ragged operand read: Done | ||
| 685 | "ldr q8, [%x[B_ptr], #0x0]\n" | ||
| 686 | "ldr q9, [%x[B_ptr], #0x10]\n" | ||
| 687 | "dup v0.2d, v0.d[0]\n" | ||
| 688 | "ldr q10, [%x[B_ptr], #0x20]\n" | ||
| 689 | "ldr q11, [%x[B_ptr], #0x30]\n" | ||
| 690 | "ldr q12, [%x[B_ptr], #0x40]\n" | ||
| 691 | "ldr q13, [%x[B_ptr], #0x50]\n" | ||
| 692 | "add %x[B_ptr], %x[B_ptr], #0x90\n" | ||
| 693 | "ldr q1, [x20, #0x0]\n" | ||
| 694 | "ldr q2, [x20, #0x10]\n" | ||
| 695 | ".inst 0x6e40fd0e // bfdot v14.4s, v8.8h, v0.8h\n" | ||
| 696 | ".inst 0x6e40fd2f // bfdot v15.4s, v9.8h, v0.8h\n" | ||
| 697 | "ldr q3, [x20, #0x20]\n" | ||
| 698 | "ldr q4, [x20, #0x30]\n" | ||
| 699 | ".inst 0x6e40fd50 // bfdot v16.4s, v10.8h, v0.8h\n" | ||
| 700 | ".inst 0x6e40fd71 // bfdot v17.4s, v11.8h, v0.8h\n" | ||
| 701 | ".inst 0x6e40fd92 // bfdot v18.4s, v12.8h, v0.8h\n" | ||
| 702 | ".inst 0x6e40fdb3 // bfdot v19.4s, v13.8h, v0.8h\n" | ||
| 703 | ".inst 0x6e40fc34 // bfdot v20.4s, v1.8h, v0.8h\n" | ||
| 704 | ".inst 0x6e40fc55 // bfdot v21.4s, v2.8h, v0.8h\n" | ||
| 705 | ".inst 0x6e40fc76 // bfdot v22.4s, v3.8h, v0.8h\n" | ||
| 706 | ".inst 0x6e40fc97 // bfdot v23.4s, v4.8h, v0.8h\n" | ||
| 707 | "51:" // Width 5: Multiply loop: No odd multiplies | ||
| 708 | "prfm pstl1keep, [%x[output_ptr], #0x0]\n" | ||
| 709 | "faddp v14.4s, v14.4s, v15.4s\n" | ||
| 710 | "faddp v15.4s, v16.4s, v17.4s\n" | ||
| 711 | "faddp v16.4s, v18.4s, v19.4s\n" | ||
| 712 | "faddp v17.4s, v20.4s, v21.4s\n" | ||
| 713 | "faddp v18.4s, v22.4s, v23.4s\n" | ||
| 714 | "tbz %x[flags], #1, 52f\n" | ||
| 715 | "add x21, %x[args_ptr], %[offset_max]\n" | ||
| 716 | "add x20, %x[args_ptr], %[offset_min]\n" | ||
| 717 | "ld1r { v20.4s }, [x21]\n" | ||
| 718 | "ld1r { v19.4s }, [x20]\n" | ||
| 719 | "fmin v14.4s, v14.4s, v20.4s\n" | ||
| 720 | "fmin v15.4s, v15.4s, v20.4s\n" | ||
| 721 | "fmin v16.4s, v16.4s, v20.4s\n" | ||
| 722 | "fmin v17.4s, v17.4s, v20.4s\n" | ||
| 723 | "fmin v18.4s, v18.4s, v20.4s\n" | ||
| 724 | "fmax v14.4s, v14.4s, v19.4s\n" | ||
| 725 | "fmax v15.4s, v15.4s, v19.4s\n" | ||
| 726 | "fmax v16.4s, v16.4s, v19.4s\n" | ||
| 727 | "fmax v17.4s, v17.4s, v19.4s\n" | ||
| 728 | "fmax v18.4s, v18.4s, v19.4s\n" | ||
| 729 | "52:" // Width 5: No activation | ||
| 730 | "cmp %x[N], #0x14\n" | ||
| 731 | "str q14, [%x[output_ptr], #0x0]\n" | ||
| 732 | "str q15, [%x[output_ptr], #0x10]\n" | ||
| 733 | "str q16, [%x[output_ptr], #0x20]\n" | ||
| 734 | "str q17, [%x[output_ptr], #0x30]\n" | ||
| 735 | "add %x[output_ptr], %x[output_ptr], #0x40\n" | ||
| 736 | "blt 53f\n" | ||
| 737 | "str q18, [%x[output_ptr], #0x0]\n" | ||
| 738 | "add %x[output_ptr], %x[output_ptr], #0x10\n" | ||
| 739 | "b 55f\n" | ||
| 740 | "53:" // Width 5: Partial writeback | ||
| 741 | "tbz %x[N], #1, 54f\n" | ||
| 742 | "str d18, [%x[output_ptr]], #0x8\n" | ||
| 743 | "tbz %x[N], #0, 55f\n" | ||
| 744 | "st1 { v18.s }[2], [%x[output_ptr]]\n" | ||
| 745 | "b 55f\n" | ||
| 746 | "54:" // Width 5: Partial direct writeback: partial_1_16 | ||
| 747 | "tbz %x[N], #0, 55f\n" | ||
| 748 | "str s18, [%x[output_ptr], #0x0]\n" | ||
| 749 | "55:" // Width 5: Writeback done | ||
| 750 | "b 100f\n" | ||
| 751 | "56:" // Width 6 | ||
| 752 | "ldr q14, [%x[B_ptr], #0x0]\n" | ||
| 753 | "ldr q15, [%x[B_ptr], #0x10]\n" | ||
| 754 | "mov x24, %x[K]\n" | ||
| 755 | "add x20, %x[B_ptr], x26\n" | ||
| 756 | "ldr q16, [%x[B_ptr], #0x20]\n" | ||
| 757 | "ldr q17, [x20, #0x0]\n" | ||
| 758 | "movi v26.16b, #0x0\n" | ||
| 759 | "mov x23, %x[A_ptr]\n" | ||
| 760 | "ldr q18, [x20, #0x10]\n" | ||
| 761 | "ldr q19, [x20, #0x20]\n" | ||
| 762 | "add %x[B_ptr], %x[B_ptr], #0x30\n" | ||
| 763 | "add x20, x20, #0x30\n" | ||
| 764 | "cmp x24, #0x4\n" | ||
| 765 | "zip2 v21.4s, v17.4s, v26.4s\n" | ||
| 766 | "zip1 v20.4s, v17.4s, v26.4s\n" | ||
| 767 | "zip2 v17.4s, v15.4s, v26.4s\n" | ||
| 768 | "zip2 v25.4s, v19.4s, v26.4s\n" | ||
| 769 | "zip1 v24.4s, v19.4s, v26.4s\n" | ||
| 770 | "zip2 v23.4s, v18.4s, v26.4s\n" | ||
| 771 | "zip1 v22.4s, v18.4s, v26.4s\n" | ||
| 772 | "zip2 v19.4s, v16.4s, v26.4s\n" | ||
| 773 | "zip1 v18.4s, v16.4s, v26.4s\n" | ||
| 774 | "zip1 v16.4s, v15.4s, v26.4s\n" | ||
| 775 | "zip2 v15.4s, v14.4s, v26.4s\n" | ||
| 776 | "zip1 v14.4s, v14.4s, v26.4s\n" | ||
| 777 | "blt 59f\n" | ||
| 778 | "cmp x24, #0x8\n" | ||
| 779 | "blt 58f\n" | ||
| 780 | "57:" // Width 6: Multiply loop: Main loop head | ||
| 781 | "ld1r { v0.2d }, [x23]\n" | ||
| 782 | "ldr q1, [%x[B_ptr], #0x0]\n" | ||
| 783 | "sub x24, x24, #0x4\n" | ||
| 784 | "add x23, x23, #0x8\n" | ||
| 785 | "ldr q2, [%x[B_ptr], #0x10]\n" | ||
| 786 | "ldr q3, [%x[B_ptr], #0x20]\n" | ||
| 787 | "cmp x24, #0x8\n" | ||
| 788 | "ldr q4, [%x[B_ptr], #0x30]\n" | ||
| 789 | "ldr q5, [%x[B_ptr], #0x40]\n" | ||
| 790 | "ldr q6, [%x[B_ptr], #0x50]\n" | ||
| 791 | "ldr q7, [x20, #0x0]\n" | ||
| 792 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 793 | ".inst 0x6e40fc2e // bfdot v14.4s, v1.8h, v0.8h\n" | ||
| 794 | "ldr q8, [x20, #0x10]\n" | ||
| 795 | "ldr q9, [x20, #0x20]\n" | ||
| 796 | ".inst 0x6e40fc4f // bfdot v15.4s, v2.8h, v0.8h\n" | ||
| 797 | ".inst 0x6e40fc70 // bfdot v16.4s, v3.8h, v0.8h\n" | ||
| 798 | "ldr q10, [x20, #0x30]\n" | ||
| 799 | "ldr q11, [x20, #0x40]\n" | ||
| 800 | ".inst 0x6e40fc91 // bfdot v17.4s, v4.8h, v0.8h\n" | ||
| 801 | ".inst 0x6e40fcb2 // bfdot v18.4s, v5.8h, v0.8h\n" | ||
| 802 | "ldr q12, [x20, #0x50]\n" | ||
| 803 | ".inst 0x6e40fcd3 // bfdot v19.4s, v6.8h, v0.8h\n" | ||
| 804 | ".inst 0x6e40fcf4 // bfdot v20.4s, v7.8h, v0.8h\n" | ||
| 805 | "add x20, x20, #0x60\n" | ||
| 806 | ".inst 0x6e40fd15 // bfdot v21.4s, v8.8h, v0.8h\n" | ||
| 807 | ".inst 0x6e40fd36 // bfdot v22.4s, v9.8h, v0.8h\n" | ||
| 808 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 809 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 810 | ".inst 0x6e40fd57 // bfdot v23.4s, v10.8h, v0.8h\n" | ||
| 811 | ".inst 0x6e40fd78 // bfdot v24.4s, v11.8h, v0.8h\n" | ||
| 812 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 813 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 814 | ".inst 0x6e40fd99 // bfdot v25.4s, v12.8h, v0.8h\n" | ||
| 815 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 816 | "bge 57b\n" | ||
| 817 | "58:" // Width 6: Multiply loop: Single iteration only | ||
| 818 | "ld1r { v0.2d }, [x23]\n" | ||
| 819 | "ldr q13, [%x[B_ptr], #0x0]\n" | ||
| 820 | "add x23, x23, #0x8\n" | ||
| 821 | "sub x24, x24, #0x4\n" | ||
| 822 | "ldr q1, [%x[B_ptr], #0x10]\n" | ||
| 823 | "ldr q2, [%x[B_ptr], #0x20]\n" | ||
| 824 | "ldr q3, [%x[B_ptr], #0x30]\n" | ||
| 825 | "ldr q4, [%x[B_ptr], #0x40]\n" | ||
| 826 | "ldr q5, [%x[B_ptr], #0x50]\n" | ||
| 827 | "ldr q6, [x20, #0x0]\n" | ||
| 828 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 829 | ".inst 0x6e40fdae // bfdot v14.4s, v13.8h, v0.8h\n" | ||
| 830 | "ldr q7, [x20, #0x10]\n" | ||
| 831 | "ldr q8, [x20, #0x20]\n" | ||
| 832 | ".inst 0x6e40fc2f // bfdot v15.4s, v1.8h, v0.8h\n" | ||
| 833 | ".inst 0x6e40fc50 // bfdot v16.4s, v2.8h, v0.8h\n" | ||
| 834 | "ldr q9, [x20, #0x30]\n" | ||
| 835 | "ldr q10, [x20, #0x40]\n" | ||
| 836 | ".inst 0x6e40fc71 // bfdot v17.4s, v3.8h, v0.8h\n" | ||
| 837 | ".inst 0x6e40fc92 // bfdot v18.4s, v4.8h, v0.8h\n" | ||
| 838 | "ldr q11, [x20, #0x50]\n" | ||
| 839 | ".inst 0x6e40fcb3 // bfdot v19.4s, v5.8h, v0.8h\n" | ||
| 840 | ".inst 0x6e40fcd4 // bfdot v20.4s, v6.8h, v0.8h\n" | ||
| 841 | "add x20, x20, #0x60\n" | ||
| 842 | ".inst 0x6e40fcf5 // bfdot v21.4s, v7.8h, v0.8h\n" | ||
| 843 | ".inst 0x6e40fd16 // bfdot v22.4s, v8.8h, v0.8h\n" | ||
| 844 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 845 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 846 | ".inst 0x6e40fd37 // bfdot v23.4s, v9.8h, v0.8h\n" | ||
| 847 | ".inst 0x6e40fd58 // bfdot v24.4s, v10.8h, v0.8h\n" | ||
| 848 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 849 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 850 | ".inst 0x6e40fd79 // bfdot v25.4s, v11.8h, v0.8h\n" | ||
| 851 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 852 | "59:" // Width 6: Multiply loop: Main loop skip | ||
| 853 | "cbz x24, 62f\n" | ||
| 854 | "tbz x24, #1, 60f\n" | ||
| 855 | "ldr s0, [x23], #0x4\n" | ||
| 856 | "tbz x24, #0, 61f\n" | ||
| 857 | "ld1 { v0.h }[2], [x23]\n" | ||
| 858 | "b 61f\n" | ||
| 859 | "60:" // Width 6: Multiply loop: Ragged operand read: partial_1_0 | ||
| 860 | "ldr h0, [x23, #0x0]\n" | ||
| 861 | "61:" // Width 6: Multiply loop: Ragged operand read: Done | ||
| 862 | "ldr q12, [%x[B_ptr], #0x0]\n" | ||
| 863 | "ldr q13, [%x[B_ptr], #0x10]\n" | ||
| 864 | "dup v0.2d, v0.d[0]\n" | ||
| 865 | "ldr q1, [%x[B_ptr], #0x20]\n" | ||
| 866 | "ldr q2, [%x[B_ptr], #0x30]\n" | ||
| 867 | "ldr q3, [%x[B_ptr], #0x40]\n" | ||
| 868 | "ldr q4, [%x[B_ptr], #0x50]\n" | ||
| 869 | "add %x[B_ptr], %x[B_ptr], #0x90\n" | ||
| 870 | "ldr q5, [x20, #0x0]\n" | ||
| 871 | "ldr q6, [x20, #0x10]\n" | ||
| 872 | ".inst 0x6e40fd8e // bfdot v14.4s, v12.8h, v0.8h\n" | ||
| 873 | ".inst 0x6e40fdaf // bfdot v15.4s, v13.8h, v0.8h\n" | ||
| 874 | "ldr q7, [x20, #0x20]\n" | ||
| 875 | "ldr q8, [x20, #0x30]\n" | ||
| 876 | ".inst 0x6e40fc30 // bfdot v16.4s, v1.8h, v0.8h\n" | ||
| 877 | ".inst 0x6e40fc51 // bfdot v17.4s, v2.8h, v0.8h\n" | ||
| 878 | "ldr q9, [x20, #0x40]\n" | ||
| 879 | "ldr q10, [x20, #0x50]\n" | ||
| 880 | ".inst 0x6e40fc72 // bfdot v18.4s, v3.8h, v0.8h\n" | ||
| 881 | ".inst 0x6e40fc93 // bfdot v19.4s, v4.8h, v0.8h\n" | ||
| 882 | ".inst 0x6e40fcb4 // bfdot v20.4s, v5.8h, v0.8h\n" | ||
| 883 | ".inst 0x6e40fcd5 // bfdot v21.4s, v6.8h, v0.8h\n" | ||
| 884 | ".inst 0x6e40fcf6 // bfdot v22.4s, v7.8h, v0.8h\n" | ||
| 885 | ".inst 0x6e40fd17 // bfdot v23.4s, v8.8h, v0.8h\n" | ||
| 886 | ".inst 0x6e40fd38 // bfdot v24.4s, v9.8h, v0.8h\n" | ||
| 887 | ".inst 0x6e40fd59 // bfdot v25.4s, v10.8h, v0.8h\n" | ||
| 888 | "62:" // Width 6: Multiply loop: No odd multiplies | ||
| 889 | "prfm pstl1keep, [%x[output_ptr], #0x0]\n" | ||
| 890 | "faddp v14.4s, v14.4s, v15.4s\n" | ||
| 891 | "faddp v15.4s, v16.4s, v17.4s\n" | ||
| 892 | "faddp v16.4s, v18.4s, v19.4s\n" | ||
| 893 | "faddp v17.4s, v20.4s, v21.4s\n" | ||
| 894 | "faddp v18.4s, v22.4s, v23.4s\n" | ||
| 895 | "faddp v19.4s, v24.4s, v25.4s\n" | ||
| 896 | "tbz %x[flags], #1, 63f\n" | ||
| 897 | "add x21, %x[args_ptr], %[offset_max]\n" | ||
| 898 | "add x20, %x[args_ptr], %[offset_min]\n" | ||
| 899 | "ld1r { v21.4s }, [x21]\n" | ||
| 900 | "ld1r { v20.4s }, [x20]\n" | ||
| 901 | "fmin v14.4s, v14.4s, v21.4s\n" | ||
| 902 | "fmin v15.4s, v15.4s, v21.4s\n" | ||
| 903 | "fmin v16.4s, v16.4s, v21.4s\n" | ||
| 904 | "fmin v17.4s, v17.4s, v21.4s\n" | ||
| 905 | "fmin v18.4s, v18.4s, v21.4s\n" | ||
| 906 | "fmin v19.4s, v19.4s, v21.4s\n" | ||
| 907 | "fmax v14.4s, v14.4s, v20.4s\n" | ||
| 908 | "fmax v15.4s, v15.4s, v20.4s\n" | ||
| 909 | "fmax v16.4s, v16.4s, v20.4s\n" | ||
| 910 | "fmax v17.4s, v17.4s, v20.4s\n" | ||
| 911 | "fmax v18.4s, v18.4s, v20.4s\n" | ||
| 912 | "fmax v19.4s, v19.4s, v20.4s\n" | ||
| 913 | "63:" // Width 6: No activation | ||
| 914 | "cmp %x[N], #0x18\n" | ||
| 915 | "str q14, [%x[output_ptr], #0x0]\n" | ||
| 916 | "str q15, [%x[output_ptr], #0x10]\n" | ||
| 917 | "str q16, [%x[output_ptr], #0x20]\n" | ||
| 918 | "str q17, [%x[output_ptr], #0x30]\n" | ||
| 919 | "str q18, [%x[output_ptr], #0x40]\n" | ||
| 920 | "add %x[output_ptr], %x[output_ptr], #0x50\n" | ||
| 921 | "blt 64f\n" | ||
| 922 | "str q19, [%x[output_ptr], #0x0]\n" | ||
| 923 | "add %x[output_ptr], %x[output_ptr], #0x10\n" | ||
| 924 | "b 66f\n" | ||
| 925 | "64:" // Width 6: Partial writeback | ||
| 926 | "tbz %x[N], #1, 65f\n" | ||
| 927 | "str d19, [%x[output_ptr]], #0x8\n" | ||
| 928 | "tbz %x[N], #0, 66f\n" | ||
| 929 | "st1 { v19.s }[2], [%x[output_ptr]]\n" | ||
| 930 | "b 66f\n" | ||
| 931 | "65:" // Width 6: Partial direct writeback: partial_1_20 | ||
| 932 | "tbz %x[N], #0, 66f\n" | ||
| 933 | "str s19, [%x[output_ptr], #0x0]\n" | ||
| 934 | "66:" // Width 6: Writeback done | ||
| 935 | "b 100f\n" | ||
| 936 | "67:" // Width 7 | ||
| 937 | "ldr q14, [%x[B_ptr], #0x0]\n" | ||
| 938 | "ldr q15, [%x[B_ptr], #0x10]\n" | ||
| 939 | "mov x24, %x[K]\n" | ||
| 940 | "add x21, %x[B_ptr], x26\n" | ||
| 941 | "ldr q16, [%x[B_ptr], #0x20]\n" | ||
| 942 | "ldr q17, [x21, #0x0]\n" | ||
| 943 | "add x20, %x[B_ptr], x26, LSL #1\n" | ||
| 944 | "movi v28.16b, #0x0\n" | ||
| 945 | "ldr q18, [x21, #0x10]\n" | ||
| 946 | "ldr q19, [x21, #0x20]\n" | ||
| 947 | "mov x23, %x[A_ptr]\n" | ||
| 948 | "add %x[B_ptr], %x[B_ptr], #0x30\n" | ||
| 949 | "ldr q20, [x20, #0x0]\n" | ||
| 950 | "cmp x24, #0x4\n" | ||
| 951 | "add x21, x21, #0x30\n" | ||
| 952 | "add x20, x20, #0x30\n" | ||
| 953 | "zip2 v21.4s, v17.4s, v28.4s\n" | ||
| 954 | "zip2 v25.4s, v19.4s, v28.4s\n" | ||
| 955 | "zip1 v24.4s, v19.4s, v28.4s\n" | ||
| 956 | "zip2 v23.4s, v18.4s, v28.4s\n" | ||
| 957 | "zip2 v27.4s, v20.4s, v28.4s\n" | ||
| 958 | "zip1 v26.4s, v20.4s, v28.4s\n" | ||
| 959 | "zip1 v22.4s, v18.4s, v28.4s\n" | ||
| 960 | "zip1 v20.4s, v17.4s, v28.4s\n" | ||
| 961 | "zip2 v19.4s, v16.4s, v28.4s\n" | ||
| 962 | "zip1 v18.4s, v16.4s, v28.4s\n" | ||
| 963 | "zip2 v17.4s, v15.4s, v28.4s\n" | ||
| 964 | "zip1 v16.4s, v15.4s, v28.4s\n" | ||
| 965 | "zip2 v15.4s, v14.4s, v28.4s\n" | ||
| 966 | "zip1 v14.4s, v14.4s, v28.4s\n" | ||
| 967 | "blt 70f\n" | ||
| 968 | "cmp x24, #0x8\n" | ||
| 969 | "blt 69f\n" | ||
| 970 | "68:" // Width 7: Multiply loop: Main loop head | ||
| 971 | "ld1r { v0.2d }, [x23]\n" | ||
| 972 | "ldr q1, [%x[B_ptr], #0x0]\n" | ||
| 973 | "sub x24, x24, #0x4\n" | ||
| 974 | "add x23, x23, #0x8\n" | ||
| 975 | "ldr q2, [%x[B_ptr], #0x10]\n" | ||
| 976 | "ldr q3, [%x[B_ptr], #0x20]\n" | ||
| 977 | "cmp x24, #0x8\n" | ||
| 978 | "ldr q4, [%x[B_ptr], #0x30]\n" | ||
| 979 | "ldr q5, [%x[B_ptr], #0x40]\n" | ||
| 980 | "ldr q6, [%x[B_ptr], #0x50]\n" | ||
| 981 | "ldr q7, [x21, #0x0]\n" | ||
| 982 | ".inst 0x6e40fc2e // bfdot v14.4s, v1.8h, v0.8h\n" | ||
| 983 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 984 | "ldr q8, [x21, #0x10]\n" | ||
| 985 | "ldr q9, [x21, #0x20]\n" | ||
| 986 | ".inst 0x6e40fc4f // bfdot v15.4s, v2.8h, v0.8h\n" | ||
| 987 | ".inst 0x6e40fc70 // bfdot v16.4s, v3.8h, v0.8h\n" | ||
| 988 | "ldr q10, [x21, #0x30]\n" | ||
| 989 | "ldr q11, [x21, #0x40]\n" | ||
| 990 | ".inst 0x6e40fc91 // bfdot v17.4s, v4.8h, v0.8h\n" | ||
| 991 | ".inst 0x6e40fcb2 // bfdot v18.4s, v5.8h, v0.8h\n" | ||
| 992 | "ldr q12, [x21, #0x50]\n" | ||
| 993 | "ldr q13, [x20, #0x0]\n" | ||
| 994 | ".inst 0x6e40fcd3 // bfdot v19.4s, v6.8h, v0.8h\n" | ||
| 995 | ".inst 0x6e40fcf4 // bfdot v20.4s, v7.8h, v0.8h\n" | ||
| 996 | "ldr q1, [x20, #0x10]\n" | ||
| 997 | ".inst 0x6e40fd15 // bfdot v21.4s, v8.8h, v0.8h\n" | ||
| 998 | ".inst 0x6e40fd36 // bfdot v22.4s, v9.8h, v0.8h\n" | ||
| 999 | "add x21, x21, #0x60\n" | ||
| 1000 | ".inst 0x6e40fd57 // bfdot v23.4s, v10.8h, v0.8h\n" | ||
| 1001 | ".inst 0x6e40fd78 // bfdot v24.4s, v11.8h, v0.8h\n" | ||
| 1002 | "add x20, x20, #0x60\n" | ||
| 1003 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 1004 | ".inst 0x6e40fd99 // bfdot v25.4s, v12.8h, v0.8h\n" | ||
| 1005 | ".inst 0x6e40fdba // bfdot v26.4s, v13.8h, v0.8h\n" | ||
| 1006 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 1007 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 1008 | ".inst 0x6e40fc3b // bfdot v27.4s, v1.8h, v0.8h\n" | ||
| 1009 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 1010 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 1011 | "bge 68b\n" | ||
| 1012 | "69:" // Width 7: Multiply loop: Single iteration only | ||
| 1013 | "ld1r { v0.2d }, [x23]\n" | ||
| 1014 | "ldr q2, [%x[B_ptr], #0x0]\n" | ||
| 1015 | "add x23, x23, #0x8\n" | ||
| 1016 | "sub x24, x24, #0x4\n" | ||
| 1017 | "ldr q3, [%x[B_ptr], #0x10]\n" | ||
| 1018 | "ldr q4, [%x[B_ptr], #0x20]\n" | ||
| 1019 | "ldr q5, [%x[B_ptr], #0x30]\n" | ||
| 1020 | "ldr q6, [%x[B_ptr], #0x40]\n" | ||
| 1021 | "ldr q7, [%x[B_ptr], #0x50]\n" | ||
| 1022 | "ldr q8, [x21, #0x0]\n" | ||
| 1023 | ".inst 0x6e40fc4e // bfdot v14.4s, v2.8h, v0.8h\n" | ||
| 1024 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 1025 | "ldr q9, [x21, #0x10]\n" | ||
| 1026 | "ldr q10, [x21, #0x20]\n" | ||
| 1027 | ".inst 0x6e40fc6f // bfdot v15.4s, v3.8h, v0.8h\n" | ||
| 1028 | ".inst 0x6e40fc90 // bfdot v16.4s, v4.8h, v0.8h\n" | ||
| 1029 | "ldr q11, [x21, #0x30]\n" | ||
| 1030 | "ldr q12, [x21, #0x40]\n" | ||
| 1031 | ".inst 0x6e40fcb1 // bfdot v17.4s, v5.8h, v0.8h\n" | ||
| 1032 | ".inst 0x6e40fcd2 // bfdot v18.4s, v6.8h, v0.8h\n" | ||
| 1033 | "ldr q13, [x21, #0x50]\n" | ||
| 1034 | "ldr q1, [x20, #0x0]\n" | ||
| 1035 | ".inst 0x6e40fcf3 // bfdot v19.4s, v7.8h, v0.8h\n" | ||
| 1036 | ".inst 0x6e40fd14 // bfdot v20.4s, v8.8h, v0.8h\n" | ||
| 1037 | "ldr q2, [x20, #0x10]\n" | ||
| 1038 | ".inst 0x6e40fd35 // bfdot v21.4s, v9.8h, v0.8h\n" | ||
| 1039 | ".inst 0x6e40fd56 // bfdot v22.4s, v10.8h, v0.8h\n" | ||
| 1040 | "add x21, x21, #0x60\n" | ||
| 1041 | ".inst 0x6e40fd77 // bfdot v23.4s, v11.8h, v0.8h\n" | ||
| 1042 | ".inst 0x6e40fd98 // bfdot v24.4s, v12.8h, v0.8h\n" | ||
| 1043 | "add x20, x20, #0x60\n" | ||
| 1044 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 1045 | ".inst 0x6e40fdb9 // bfdot v25.4s, v13.8h, v0.8h\n" | ||
| 1046 | ".inst 0x6e40fc3a // bfdot v26.4s, v1.8h, v0.8h\n" | ||
| 1047 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 1048 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 1049 | ".inst 0x6e40fc5b // bfdot v27.4s, v2.8h, v0.8h\n" | ||
| 1050 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 1051 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 1052 | "70:" // Width 7: Multiply loop: Main loop skip | ||
| 1053 | "cbz x24, 73f\n" | ||
| 1054 | "tbz x24, #1, 71f\n" | ||
| 1055 | "ldr s0, [x23], #0x4\n" | ||
| 1056 | "tbz x24, #0, 72f\n" | ||
| 1057 | "ld1 { v0.h }[2], [x23]\n" | ||
| 1058 | "b 72f\n" | ||
| 1059 | "71:" // Width 7: Multiply loop: Ragged operand read: partial_1_0 | ||
| 1060 | "ldr h0, [x23, #0x0]\n" | ||
| 1061 | "72:" // Width 7: Multiply loop: Ragged operand read: Done | ||
| 1062 | "ldr q3, [%x[B_ptr], #0x0]\n" | ||
| 1063 | "ldr q4, [%x[B_ptr], #0x10]\n" | ||
| 1064 | "dup v0.2d, v0.d[0]\n" | ||
| 1065 | "ldr q5, [%x[B_ptr], #0x20]\n" | ||
| 1066 | "ldr q6, [%x[B_ptr], #0x30]\n" | ||
| 1067 | "ldr q7, [%x[B_ptr], #0x40]\n" | ||
| 1068 | "ldr q8, [%x[B_ptr], #0x50]\n" | ||
| 1069 | "add %x[B_ptr], %x[B_ptr], #0x90\n" | ||
| 1070 | "ldr q9, [x21, #0x0]\n" | ||
| 1071 | "ldr q10, [x21, #0x10]\n" | ||
| 1072 | ".inst 0x6e40fc6e // bfdot v14.4s, v3.8h, v0.8h\n" | ||
| 1073 | ".inst 0x6e40fc8f // bfdot v15.4s, v4.8h, v0.8h\n" | ||
| 1074 | "ldr q11, [x21, #0x20]\n" | ||
| 1075 | "ldr q12, [x21, #0x30]\n" | ||
| 1076 | ".inst 0x6e40fcb0 // bfdot v16.4s, v5.8h, v0.8h\n" | ||
| 1077 | ".inst 0x6e40fcd1 // bfdot v17.4s, v6.8h, v0.8h\n" | ||
| 1078 | "ldr q13, [x21, #0x40]\n" | ||
| 1079 | "ldr q1, [x21, #0x50]\n" | ||
| 1080 | ".inst 0x6e40fcf2 // bfdot v18.4s, v7.8h, v0.8h\n" | ||
| 1081 | ".inst 0x6e40fd13 // bfdot v19.4s, v8.8h, v0.8h\n" | ||
| 1082 | "ldr q2, [x20, #0x0]\n" | ||
| 1083 | "ldr q3, [x20, #0x10]\n" | ||
| 1084 | ".inst 0x6e40fd34 // bfdot v20.4s, v9.8h, v0.8h\n" | ||
| 1085 | ".inst 0x6e40fd55 // bfdot v21.4s, v10.8h, v0.8h\n" | ||
| 1086 | ".inst 0x6e40fd76 // bfdot v22.4s, v11.8h, v0.8h\n" | ||
| 1087 | ".inst 0x6e40fd97 // bfdot v23.4s, v12.8h, v0.8h\n" | ||
| 1088 | ".inst 0x6e40fdb8 // bfdot v24.4s, v13.8h, v0.8h\n" | ||
| 1089 | ".inst 0x6e40fc39 // bfdot v25.4s, v1.8h, v0.8h\n" | ||
| 1090 | ".inst 0x6e40fc5a // bfdot v26.4s, v2.8h, v0.8h\n" | ||
| 1091 | ".inst 0x6e40fc7b // bfdot v27.4s, v3.8h, v0.8h\n" | ||
| 1092 | "73:" // Width 7: Multiply loop: No odd multiplies | ||
| 1093 | "prfm pstl1keep, [%x[output_ptr], #0x0]\n" | ||
| 1094 | "faddp v14.4s, v14.4s, v15.4s\n" | ||
| 1095 | "faddp v15.4s, v16.4s, v17.4s\n" | ||
| 1096 | "faddp v16.4s, v18.4s, v19.4s\n" | ||
| 1097 | "faddp v17.4s, v20.4s, v21.4s\n" | ||
| 1098 | "faddp v18.4s, v22.4s, v23.4s\n" | ||
| 1099 | "faddp v19.4s, v24.4s, v25.4s\n" | ||
| 1100 | "faddp v20.4s, v26.4s, v27.4s\n" | ||
| 1101 | "tbz %x[flags], #1, 74f\n" | ||
| 1102 | "add x21, %x[args_ptr], %[offset_max]\n" | ||
| 1103 | "add x20, %x[args_ptr], %[offset_min]\n" | ||
| 1104 | "ld1r { v22.4s }, [x21]\n" | ||
| 1105 | "ld1r { v21.4s }, [x20]\n" | ||
| 1106 | "fmin v14.4s, v14.4s, v22.4s\n" | ||
| 1107 | "fmin v15.4s, v15.4s, v22.4s\n" | ||
| 1108 | "fmin v16.4s, v16.4s, v22.4s\n" | ||
| 1109 | "fmin v17.4s, v17.4s, v22.4s\n" | ||
| 1110 | "fmin v18.4s, v18.4s, v22.4s\n" | ||
| 1111 | "fmin v19.4s, v19.4s, v22.4s\n" | ||
| 1112 | "fmin v20.4s, v20.4s, v22.4s\n" | ||
| 1113 | "fmax v14.4s, v14.4s, v21.4s\n" | ||
| 1114 | "fmax v15.4s, v15.4s, v21.4s\n" | ||
| 1115 | "fmax v16.4s, v16.4s, v21.4s\n" | ||
| 1116 | "fmax v17.4s, v17.4s, v21.4s\n" | ||
| 1117 | "fmax v18.4s, v18.4s, v21.4s\n" | ||
| 1118 | "fmax v19.4s, v19.4s, v21.4s\n" | ||
| 1119 | "fmax v20.4s, v20.4s, v21.4s\n" | ||
| 1120 | "74:" // Width 7: No activation | ||
| 1121 | "cmp %x[N], #0x1c\n" | ||
| 1122 | "str q14, [%x[output_ptr], #0x0]\n" | ||
| 1123 | "str q15, [%x[output_ptr], #0x10]\n" | ||
| 1124 | "str q16, [%x[output_ptr], #0x20]\n" | ||
| 1125 | "str q17, [%x[output_ptr], #0x30]\n" | ||
| 1126 | "str q18, [%x[output_ptr], #0x40]\n" | ||
| 1127 | "str q19, [%x[output_ptr], #0x50]\n" | ||
| 1128 | "add %x[output_ptr], %x[output_ptr], #0x60\n" | ||
| 1129 | "blt 75f\n" | ||
| 1130 | "str q20, [%x[output_ptr], #0x0]\n" | ||
| 1131 | "add %x[output_ptr], %x[output_ptr], #0x10\n" | ||
| 1132 | "b 77f\n" | ||
| 1133 | "75:" // Width 7: Partial writeback | ||
| 1134 | "tbz %x[N], #1, 76f\n" | ||
| 1135 | "str d20, [%x[output_ptr]], #0x8\n" | ||
| 1136 | "tbz %x[N], #0, 77f\n" | ||
| 1137 | "st1 { v20.s }[2], [%x[output_ptr]]\n" | ||
| 1138 | "b 77f\n" | ||
| 1139 | "76:" // Width 7: Partial direct writeback: partial_1_24 | ||
| 1140 | "tbz %x[N], #0, 77f\n" | ||
| 1141 | "str s20, [%x[output_ptr], #0x0]\n" | ||
| 1142 | "77:" // Width 7: Writeback done | ||
| 1143 | "b 100f\n" | ||
| 1144 | "78:" // Width 8 | ||
| 1145 | "ldr q14, [%x[B_ptr], #0x0]\n" | ||
| 1146 | "ldr q15, [%x[B_ptr], #0x10]\n" | ||
| 1147 | "mov x24, %x[K]\n" | ||
| 1148 | "add x21, %x[B_ptr], x26\n" | ||
| 1149 | "ldr q16, [%x[B_ptr], #0x20]\n" | ||
| 1150 | "ldr q17, [x21, #0x0]\n" | ||
| 1151 | "add x20, %x[B_ptr], x26, LSL #1\n" | ||
| 1152 | "movi v30.16b, #0x0\n" | ||
| 1153 | "ldr q18, [x21, #0x10]\n" | ||
| 1154 | "ldr q19, [x21, #0x20]\n" | ||
| 1155 | "mov x23, %x[A_ptr]\n" | ||
| 1156 | "add %x[B_ptr], %x[B_ptr], #0x30\n" | ||
| 1157 | "ldr q20, [x20, #0x0]\n" | ||
| 1158 | "ldr q21, [x20, #0x10]\n" | ||
| 1159 | "cmp x24, #0x4\n" | ||
| 1160 | "add x21, x21, #0x30\n" | ||
| 1161 | "add x20, x20, #0x30\n" | ||
| 1162 | "zip2 v25.4s, v19.4s, v30.4s\n" | ||
| 1163 | "zip1 v24.4s, v19.4s, v30.4s\n" | ||
| 1164 | "zip2 v23.4s, v18.4s, v30.4s\n" | ||
| 1165 | "zip1 v22.4s, v18.4s, v30.4s\n" | ||
| 1166 | "zip2 v29.4s, v21.4s, v30.4s\n" | ||
| 1167 | "zip1 v28.4s, v21.4s, v30.4s\n" | ||
| 1168 | "zip2 v27.4s, v20.4s, v30.4s\n" | ||
| 1169 | "zip1 v26.4s, v20.4s, v30.4s\n" | ||
| 1170 | "zip2 v21.4s, v17.4s, v30.4s\n" | ||
| 1171 | "zip1 v20.4s, v17.4s, v30.4s\n" | ||
| 1172 | "zip2 v19.4s, v16.4s, v30.4s\n" | ||
| 1173 | "zip1 v18.4s, v16.4s, v30.4s\n" | ||
| 1174 | "zip2 v17.4s, v15.4s, v30.4s\n" | ||
| 1175 | "zip1 v16.4s, v15.4s, v30.4s\n" | ||
| 1176 | "zip2 v15.4s, v14.4s, v30.4s\n" | ||
| 1177 | "zip1 v14.4s, v14.4s, v30.4s\n" | ||
| 1178 | "blt 81f\n" | ||
| 1179 | "cmp x24, #0x8\n" | ||
| 1180 | "blt 80f\n" | ||
| 1181 | "79:" // Width 8: Multiply loop: Main loop head | ||
| 1182 | "ld1r { v0.2d }, [x23]\n" | ||
| 1183 | "ldr q1, [%x[B_ptr], #0x0]\n" | ||
| 1184 | "sub x24, x24, #0x4\n" | ||
| 1185 | "add x23, x23, #0x8\n" | ||
| 1186 | "ldr q2, [%x[B_ptr], #0x10]\n" | ||
| 1187 | "ldr q3, [%x[B_ptr], #0x20]\n" | ||
| 1188 | "cmp x24, #0x8\n" | ||
| 1189 | "ldr q4, [%x[B_ptr], #0x30]\n" | ||
| 1190 | "ldr q5, [%x[B_ptr], #0x40]\n" | ||
| 1191 | "ldr q6, [%x[B_ptr], #0x50]\n" | ||
| 1192 | "ldr q7, [x21, #0x0]\n" | ||
| 1193 | ".inst 0x6e40fc2e // bfdot v14.4s, v1.8h, v0.8h\n" | ||
| 1194 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 1195 | "ldr q8, [x21, #0x10]\n" | ||
| 1196 | "ldr q9, [x21, #0x20]\n" | ||
| 1197 | ".inst 0x6e40fc4f // bfdot v15.4s, v2.8h, v0.8h\n" | ||
| 1198 | ".inst 0x6e40fc70 // bfdot v16.4s, v3.8h, v0.8h\n" | ||
| 1199 | "ldr q10, [x21, #0x30]\n" | ||
| 1200 | "ldr q11, [x21, #0x40]\n" | ||
| 1201 | ".inst 0x6e40fc91 // bfdot v17.4s, v4.8h, v0.8h\n" | ||
| 1202 | ".inst 0x6e40fcb2 // bfdot v18.4s, v5.8h, v0.8h\n" | ||
| 1203 | "ldr q12, [x21, #0x50]\n" | ||
| 1204 | "ldr q13, [x20, #0x0]\n" | ||
| 1205 | ".inst 0x6e40fcd3 // bfdot v19.4s, v6.8h, v0.8h\n" | ||
| 1206 | ".inst 0x6e40fcf4 // bfdot v20.4s, v7.8h, v0.8h\n" | ||
| 1207 | "ldr q1, [x20, #0x10]\n" | ||
| 1208 | "ldr q2, [x20, #0x20]\n" | ||
| 1209 | ".inst 0x6e40fd15 // bfdot v21.4s, v8.8h, v0.8h\n" | ||
| 1210 | ".inst 0x6e40fd36 // bfdot v22.4s, v9.8h, v0.8h\n" | ||
| 1211 | "ldr q3, [x20, #0x30]\n" | ||
| 1212 | ".inst 0x6e40fd57 // bfdot v23.4s, v10.8h, v0.8h\n" | ||
| 1213 | ".inst 0x6e40fd78 // bfdot v24.4s, v11.8h, v0.8h\n" | ||
| 1214 | "add x21, x21, #0x60\n" | ||
| 1215 | ".inst 0x6e40fd99 // bfdot v25.4s, v12.8h, v0.8h\n" | ||
| 1216 | ".inst 0x6e40fdba // bfdot v26.4s, v13.8h, v0.8h\n" | ||
| 1217 | "add x20, x20, #0x60\n" | ||
| 1218 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 1219 | ".inst 0x6e40fc3b // bfdot v27.4s, v1.8h, v0.8h\n" | ||
| 1220 | ".inst 0x6e40fc5c // bfdot v28.4s, v2.8h, v0.8h\n" | ||
| 1221 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 1222 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 1223 | ".inst 0x6e40fc7d // bfdot v29.4s, v3.8h, v0.8h\n" | ||
| 1224 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 1225 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 1226 | "bge 79b\n" | ||
| 1227 | "80:" // Width 8: Multiply loop: Single iteration only | ||
| 1228 | "ld1r { v0.2d }, [x23]\n" | ||
| 1229 | "ldr q4, [%x[B_ptr], #0x0]\n" | ||
| 1230 | "add x23, x23, #0x8\n" | ||
| 1231 | "sub x24, x24, #0x4\n" | ||
| 1232 | "ldr q5, [%x[B_ptr], #0x10]\n" | ||
| 1233 | "ldr q6, [%x[B_ptr], #0x20]\n" | ||
| 1234 | "ldr q7, [%x[B_ptr], #0x30]\n" | ||
| 1235 | "ldr q8, [%x[B_ptr], #0x40]\n" | ||
| 1236 | "ldr q9, [%x[B_ptr], #0x50]\n" | ||
| 1237 | "ldr q10, [x21, #0x0]\n" | ||
| 1238 | ".inst 0x6e40fc8e // bfdot v14.4s, v4.8h, v0.8h\n" | ||
| 1239 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 1240 | "ldr q11, [x21, #0x10]\n" | ||
| 1241 | "ldr q12, [x21, #0x20]\n" | ||
| 1242 | ".inst 0x6e40fcaf // bfdot v15.4s, v5.8h, v0.8h\n" | ||
| 1243 | ".inst 0x6e40fcd0 // bfdot v16.4s, v6.8h, v0.8h\n" | ||
| 1244 | "ldr q13, [x21, #0x30]\n" | ||
| 1245 | "ldr q1, [x21, #0x40]\n" | ||
| 1246 | ".inst 0x6e40fcf1 // bfdot v17.4s, v7.8h, v0.8h\n" | ||
| 1247 | ".inst 0x6e40fd12 // bfdot v18.4s, v8.8h, v0.8h\n" | ||
| 1248 | "ldr q2, [x21, #0x50]\n" | ||
| 1249 | "ldr q3, [x20, #0x0]\n" | ||
| 1250 | ".inst 0x6e40fd33 // bfdot v19.4s, v9.8h, v0.8h\n" | ||
| 1251 | ".inst 0x6e40fd54 // bfdot v20.4s, v10.8h, v0.8h\n" | ||
| 1252 | "ldr q4, [x20, #0x10]\n" | ||
| 1253 | "ldr q5, [x20, #0x20]\n" | ||
| 1254 | ".inst 0x6e40fd75 // bfdot v21.4s, v11.8h, v0.8h\n" | ||
| 1255 | ".inst 0x6e40fd96 // bfdot v22.4s, v12.8h, v0.8h\n" | ||
| 1256 | "ldr q6, [x20, #0x30]\n" | ||
| 1257 | ".inst 0x6e40fdb7 // bfdot v23.4s, v13.8h, v0.8h\n" | ||
| 1258 | ".inst 0x6e40fc38 // bfdot v24.4s, v1.8h, v0.8h\n" | ||
| 1259 | "add x21, x21, #0x60\n" | ||
| 1260 | ".inst 0x6e40fc59 // bfdot v25.4s, v2.8h, v0.8h\n" | ||
| 1261 | ".inst 0x6e40fc7a // bfdot v26.4s, v3.8h, v0.8h\n" | ||
| 1262 | "add x20, x20, #0x60\n" | ||
| 1263 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 1264 | ".inst 0x6e40fc9b // bfdot v27.4s, v4.8h, v0.8h\n" | ||
| 1265 | ".inst 0x6e40fcbc // bfdot v28.4s, v5.8h, v0.8h\n" | ||
| 1266 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 1267 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 1268 | ".inst 0x6e40fcdd // bfdot v29.4s, v6.8h, v0.8h\n" | ||
| 1269 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 1270 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 1271 | "81:" // Width 8: Multiply loop: Main loop skip | ||
| 1272 | "cbz x24, 84f\n" | ||
| 1273 | "tbz x24, #1, 82f\n" | ||
| 1274 | "ldr s0, [x23], #0x4\n" | ||
| 1275 | "tbz x24, #0, 83f\n" | ||
| 1276 | "ld1 { v0.h }[2], [x23]\n" | ||
| 1277 | "b 83f\n" | ||
| 1278 | "82:" // Width 8: Multiply loop: Ragged operand read: partial_1_0 | ||
| 1279 | "ldr h0, [x23, #0x0]\n" | ||
| 1280 | "83:" // Width 8: Multiply loop: Ragged operand read: Done | ||
| 1281 | "ldr q7, [%x[B_ptr], #0x0]\n" | ||
| 1282 | "ldr q8, [%x[B_ptr], #0x10]\n" | ||
| 1283 | "dup v0.2d, v0.d[0]\n" | ||
| 1284 | "ldr q9, [%x[B_ptr], #0x20]\n" | ||
| 1285 | "ldr q10, [%x[B_ptr], #0x30]\n" | ||
| 1286 | "ldr q11, [%x[B_ptr], #0x40]\n" | ||
| 1287 | "ldr q12, [%x[B_ptr], #0x50]\n" | ||
| 1288 | "add %x[B_ptr], %x[B_ptr], #0x90\n" | ||
| 1289 | "ldr q13, [x21, #0x0]\n" | ||
| 1290 | "ldr q1, [x21, #0x10]\n" | ||
| 1291 | ".inst 0x6e40fcee // bfdot v14.4s, v7.8h, v0.8h\n" | ||
| 1292 | ".inst 0x6e40fd0f // bfdot v15.4s, v8.8h, v0.8h\n" | ||
| 1293 | "ldr q2, [x21, #0x20]\n" | ||
| 1294 | "ldr q3, [x21, #0x30]\n" | ||
| 1295 | ".inst 0x6e40fd30 // bfdot v16.4s, v9.8h, v0.8h\n" | ||
| 1296 | ".inst 0x6e40fd51 // bfdot v17.4s, v10.8h, v0.8h\n" | ||
| 1297 | "ldr q4, [x21, #0x40]\n" | ||
| 1298 | "ldr q5, [x21, #0x50]\n" | ||
| 1299 | ".inst 0x6e40fd72 // bfdot v18.4s, v11.8h, v0.8h\n" | ||
| 1300 | ".inst 0x6e40fd93 // bfdot v19.4s, v12.8h, v0.8h\n" | ||
| 1301 | "ldr q6, [x20, #0x0]\n" | ||
| 1302 | "ldr q7, [x20, #0x10]\n" | ||
| 1303 | ".inst 0x6e40fdb4 // bfdot v20.4s, v13.8h, v0.8h\n" | ||
| 1304 | ".inst 0x6e40fc35 // bfdot v21.4s, v1.8h, v0.8h\n" | ||
| 1305 | "ldr q8, [x20, #0x20]\n" | ||
| 1306 | "ldr q9, [x20, #0x30]\n" | ||
| 1307 | ".inst 0x6e40fc56 // bfdot v22.4s, v2.8h, v0.8h\n" | ||
| 1308 | ".inst 0x6e40fc77 // bfdot v23.4s, v3.8h, v0.8h\n" | ||
| 1309 | ".inst 0x6e40fc98 // bfdot v24.4s, v4.8h, v0.8h\n" | ||
| 1310 | ".inst 0x6e40fcb9 // bfdot v25.4s, v5.8h, v0.8h\n" | ||
| 1311 | ".inst 0x6e40fcda // bfdot v26.4s, v6.8h, v0.8h\n" | ||
| 1312 | ".inst 0x6e40fcfb // bfdot v27.4s, v7.8h, v0.8h\n" | ||
| 1313 | ".inst 0x6e40fd1c // bfdot v28.4s, v8.8h, v0.8h\n" | ||
| 1314 | ".inst 0x6e40fd3d // bfdot v29.4s, v9.8h, v0.8h\n" | ||
| 1315 | "84:" // Width 8: Multiply loop: No odd multiplies | ||
| 1316 | "prfm pstl1keep, [%x[output_ptr], #0x0]\n" | ||
| 1317 | "faddp v14.4s, v14.4s, v15.4s\n" | ||
| 1318 | "faddp v15.4s, v16.4s, v17.4s\n" | ||
| 1319 | "faddp v16.4s, v18.4s, v19.4s\n" | ||
| 1320 | "faddp v17.4s, v20.4s, v21.4s\n" | ||
| 1321 | "faddp v18.4s, v22.4s, v23.4s\n" | ||
| 1322 | "faddp v19.4s, v24.4s, v25.4s\n" | ||
| 1323 | "faddp v20.4s, v26.4s, v27.4s\n" | ||
| 1324 | "faddp v21.4s, v28.4s, v29.4s\n" | ||
| 1325 | "tbz %x[flags], #1, 85f\n" | ||
| 1326 | "add x21, %x[args_ptr], %[offset_max]\n" | ||
| 1327 | "add x20, %x[args_ptr], %[offset_min]\n" | ||
| 1328 | "ld1r { v23.4s }, [x21]\n" | ||
| 1329 | "ld1r { v22.4s }, [x20]\n" | ||
| 1330 | "fmin v14.4s, v14.4s, v23.4s\n" | ||
| 1331 | "fmin v15.4s, v15.4s, v23.4s\n" | ||
| 1332 | "fmin v16.4s, v16.4s, v23.4s\n" | ||
| 1333 | "fmin v17.4s, v17.4s, v23.4s\n" | ||
| 1334 | "fmin v18.4s, v18.4s, v23.4s\n" | ||
| 1335 | "fmin v19.4s, v19.4s, v23.4s\n" | ||
| 1336 | "fmin v20.4s, v20.4s, v23.4s\n" | ||
| 1337 | "fmin v21.4s, v21.4s, v23.4s\n" | ||
| 1338 | "fmax v14.4s, v14.4s, v22.4s\n" | ||
| 1339 | "fmax v15.4s, v15.4s, v22.4s\n" | ||
| 1340 | "fmax v16.4s, v16.4s, v22.4s\n" | ||
| 1341 | "fmax v17.4s, v17.4s, v22.4s\n" | ||
| 1342 | "fmax v18.4s, v18.4s, v22.4s\n" | ||
| 1343 | "fmax v19.4s, v19.4s, v22.4s\n" | ||
| 1344 | "fmax v20.4s, v20.4s, v22.4s\n" | ||
| 1345 | "fmax v21.4s, v21.4s, v22.4s\n" | ||
| 1346 | "85:" // Width 8: No activation | ||
| 1347 | "cmp %x[N], #0x20\n" | ||
| 1348 | "str q14, [%x[output_ptr], #0x0]\n" | ||
| 1349 | "str q15, [%x[output_ptr], #0x10]\n" | ||
| 1350 | "str q16, [%x[output_ptr], #0x20]\n" | ||
| 1351 | "str q17, [%x[output_ptr], #0x30]\n" | ||
| 1352 | "str q18, [%x[output_ptr], #0x40]\n" | ||
| 1353 | "str q19, [%x[output_ptr], #0x50]\n" | ||
| 1354 | "str q20, [%x[output_ptr], #0x60]\n" | ||
| 1355 | "add %x[output_ptr], %x[output_ptr], #0x70\n" | ||
| 1356 | "blt 86f\n" | ||
| 1357 | "str q21, [%x[output_ptr], #0x0]\n" | ||
| 1358 | "add %x[output_ptr], %x[output_ptr], #0x10\n" | ||
| 1359 | "b 88f\n" | ||
| 1360 | "86:" // Width 8: Partial writeback | ||
| 1361 | "tbz %x[N], #1, 87f\n" | ||
| 1362 | "str d21, [%x[output_ptr]], #0x8\n" | ||
| 1363 | "tbz %x[N], #0, 88f\n" | ||
| 1364 | "st1 { v21.s }[2], [%x[output_ptr]]\n" | ||
| 1365 | "b 88f\n" | ||
| 1366 | "87:" // Width 8: Partial direct writeback: partial_1_28 | ||
| 1367 | "tbz %x[N], #0, 88f\n" | ||
| 1368 | "str s21, [%x[output_ptr], #0x0]\n" | ||
| 1369 | "88:" // Width 8: Writeback done | ||
| 1370 | "b 100f\n" | ||
| 1371 | "89:" // Width 9 | ||
| 1372 | "ldr q14, [%x[B_ptr], #0x0]\n" | ||
| 1373 | "ldr q15, [%x[B_ptr], #0x10]\n" | ||
| 1374 | "mov x24, %x[K]\n" | ||
| 1375 | "add x21, %x[B_ptr], x26\n" | ||
| 1376 | "ldr q16, [%x[B_ptr], #0x20]\n" | ||
| 1377 | "ldr q17, [x21, #0x0]\n" | ||
| 1378 | "add x20, %x[B_ptr], x26, LSL #1\n" | ||
| 1379 | "movi v0.16b, #0x0\n" | ||
| 1380 | "ldr q18, [x21, #0x10]\n" | ||
| 1381 | "ldr q19, [x21, #0x20]\n" | ||
| 1382 | "mov x23, %x[A_ptr]\n" | ||
| 1383 | "add x22, x20, x26\n" | ||
| 1384 | "ldr q20, [x20, #0x0]\n" | ||
| 1385 | "ldr q21, [x20, #0x10]\n" | ||
| 1386 | "cmp x24, #0x4\n" | ||
| 1387 | "add %x[B_ptr], %x[B_ptr], #0x30\n" | ||
| 1388 | "ldr q22, [x20, #0x20]\n" | ||
| 1389 | "add x21, x21, #0x30\n" | ||
| 1390 | "add x20, x20, #0x30\n" | ||
| 1391 | "zip2 v25.4s, v19.4s, v0.4s\n" | ||
| 1392 | "zip1 v24.4s, v19.4s, v0.4s\n" | ||
| 1393 | "zip2 v23.4s, v18.4s, v0.4s\n" | ||
| 1394 | "zip2 v29.4s, v21.4s, v0.4s\n" | ||
| 1395 | "zip1 v28.4s, v21.4s, v0.4s\n" | ||
| 1396 | "zip2 v31.4s, v22.4s, v0.4s\n" | ||
| 1397 | "zip1 v30.4s, v22.4s, v0.4s\n" | ||
| 1398 | "zip2 v27.4s, v20.4s, v0.4s\n" | ||
| 1399 | "zip1 v26.4s, v20.4s, v0.4s\n" | ||
| 1400 | "zip1 v22.4s, v18.4s, v0.4s\n" | ||
| 1401 | "zip2 v21.4s, v17.4s, v0.4s\n" | ||
| 1402 | "zip1 v20.4s, v17.4s, v0.4s\n" | ||
| 1403 | "zip2 v19.4s, v16.4s, v0.4s\n" | ||
| 1404 | "zip1 v18.4s, v16.4s, v0.4s\n" | ||
| 1405 | "zip2 v17.4s, v15.4s, v0.4s\n" | ||
| 1406 | "zip1 v16.4s, v15.4s, v0.4s\n" | ||
| 1407 | "zip2 v15.4s, v14.4s, v0.4s\n" | ||
| 1408 | "zip1 v14.4s, v14.4s, v0.4s\n" | ||
| 1409 | "blt 92f\n" | ||
| 1410 | "cmp x24, #0x8\n" | ||
| 1411 | "blt 91f\n" | ||
| 1412 | "90:" // Width 9: Multiply loop: Main loop head | ||
| 1413 | "ld1r { v0.2d }, [x23]\n" | ||
| 1414 | "ldr q1, [%x[B_ptr], #0x0]\n" | ||
| 1415 | "sub x24, x24, #0x4\n" | ||
| 1416 | "add x23, x23, #0x8\n" | ||
| 1417 | "ldr q2, [%x[B_ptr], #0x10]\n" | ||
| 1418 | "ldr q3, [%x[B_ptr], #0x20]\n" | ||
| 1419 | "cmp x24, #0x8\n" | ||
| 1420 | "ldr q4, [%x[B_ptr], #0x30]\n" | ||
| 1421 | "ldr q5, [%x[B_ptr], #0x40]\n" | ||
| 1422 | "ldr q6, [%x[B_ptr], #0x50]\n" | ||
| 1423 | "ldr q7, [x21, #0x0]\n" | ||
| 1424 | ".inst 0x6e40fc2e // bfdot v14.4s, v1.8h, v0.8h\n" | ||
| 1425 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 1426 | "ldr q8, [x21, #0x10]\n" | ||
| 1427 | "ldr q9, [x21, #0x20]\n" | ||
| 1428 | ".inst 0x6e40fc4f // bfdot v15.4s, v2.8h, v0.8h\n" | ||
| 1429 | ".inst 0x6e40fc70 // bfdot v16.4s, v3.8h, v0.8h\n" | ||
| 1430 | "ldr q10, [x21, #0x30]\n" | ||
| 1431 | "ldr q11, [x21, #0x40]\n" | ||
| 1432 | ".inst 0x6e40fc91 // bfdot v17.4s, v4.8h, v0.8h\n" | ||
| 1433 | ".inst 0x6e40fcb2 // bfdot v18.4s, v5.8h, v0.8h\n" | ||
| 1434 | "ldr q12, [x21, #0x50]\n" | ||
| 1435 | "ldr q13, [x20, #0x0]\n" | ||
| 1436 | ".inst 0x6e40fcd3 // bfdot v19.4s, v6.8h, v0.8h\n" | ||
| 1437 | ".inst 0x6e40fcf4 // bfdot v20.4s, v7.8h, v0.8h\n" | ||
| 1438 | "ldr q1, [x20, #0x10]\n" | ||
| 1439 | "ldr q2, [x20, #0x20]\n" | ||
| 1440 | ".inst 0x6e40fd15 // bfdot v21.4s, v8.8h, v0.8h\n" | ||
| 1441 | ".inst 0x6e40fd36 // bfdot v22.4s, v9.8h, v0.8h\n" | ||
| 1442 | "ldr q3, [x20, #0x30]\n" | ||
| 1443 | "ldr q4, [x20, #0x40]\n" | ||
| 1444 | ".inst 0x6e40fd57 // bfdot v23.4s, v10.8h, v0.8h\n" | ||
| 1445 | ".inst 0x6e40fd78 // bfdot v24.4s, v11.8h, v0.8h\n" | ||
| 1446 | "ldr q5, [x20, #0x50]\n" | ||
| 1447 | ".inst 0x6e40fd99 // bfdot v25.4s, v12.8h, v0.8h\n" | ||
| 1448 | ".inst 0x6e40fdba // bfdot v26.4s, v13.8h, v0.8h\n" | ||
| 1449 | "add x21, x21, #0x60\n" | ||
| 1450 | ".inst 0x6e40fc3b // bfdot v27.4s, v1.8h, v0.8h\n" | ||
| 1451 | ".inst 0x6e40fc5c // bfdot v28.4s, v2.8h, v0.8h\n" | ||
| 1452 | "add x20, x20, #0x60\n" | ||
| 1453 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 1454 | ".inst 0x6e40fc7d // bfdot v29.4s, v3.8h, v0.8h\n" | ||
| 1455 | ".inst 0x6e40fc9e // bfdot v30.4s, v4.8h, v0.8h\n" | ||
| 1456 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 1457 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 1458 | ".inst 0x6e40fcbf // bfdot v31.4s, v5.8h, v0.8h\n" | ||
| 1459 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 1460 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 1461 | "bge 90b\n" | ||
| 1462 | "91:" // Width 9: Multiply loop: Single iteration only | ||
| 1463 | "ld1r { v0.2d }, [x23]\n" | ||
| 1464 | "ldr q6, [%x[B_ptr], #0x0]\n" | ||
| 1465 | "add x23, x23, #0x8\n" | ||
| 1466 | "sub x24, x24, #0x4\n" | ||
| 1467 | "ldr q7, [%x[B_ptr], #0x10]\n" | ||
| 1468 | "ldr q8, [%x[B_ptr], #0x20]\n" | ||
| 1469 | "ldr q9, [%x[B_ptr], #0x30]\n" | ||
| 1470 | "ldr q10, [%x[B_ptr], #0x40]\n" | ||
| 1471 | "ldr q11, [%x[B_ptr], #0x50]\n" | ||
| 1472 | "ldr q12, [x21, #0x0]\n" | ||
| 1473 | ".inst 0x6e40fcce // bfdot v14.4s, v6.8h, v0.8h\n" | ||
| 1474 | "add %x[B_ptr], %x[B_ptr], #0x60\n" | ||
| 1475 | "ldr q13, [x21, #0x10]\n" | ||
| 1476 | "ldr q1, [x21, #0x20]\n" | ||
| 1477 | ".inst 0x6e40fcef // bfdot v15.4s, v7.8h, v0.8h\n" | ||
| 1478 | ".inst 0x6e40fd10 // bfdot v16.4s, v8.8h, v0.8h\n" | ||
| 1479 | "ldr q2, [x21, #0x30]\n" | ||
| 1480 | "ldr q3, [x21, #0x40]\n" | ||
| 1481 | ".inst 0x6e40fd31 // bfdot v17.4s, v9.8h, v0.8h\n" | ||
| 1482 | ".inst 0x6e40fd52 // bfdot v18.4s, v10.8h, v0.8h\n" | ||
| 1483 | "ldr q4, [x21, #0x50]\n" | ||
| 1484 | "ldr q5, [x20, #0x0]\n" | ||
| 1485 | ".inst 0x6e40fd73 // bfdot v19.4s, v11.8h, v0.8h\n" | ||
| 1486 | ".inst 0x6e40fd94 // bfdot v20.4s, v12.8h, v0.8h\n" | ||
| 1487 | "ldr q6, [x20, #0x10]\n" | ||
| 1488 | "ldr q7, [x20, #0x20]\n" | ||
| 1489 | ".inst 0x6e40fdb5 // bfdot v21.4s, v13.8h, v0.8h\n" | ||
| 1490 | ".inst 0x6e40fc36 // bfdot v22.4s, v1.8h, v0.8h\n" | ||
| 1491 | "ldr q8, [x20, #0x30]\n" | ||
| 1492 | "ldr q9, [x20, #0x40]\n" | ||
| 1493 | ".inst 0x6e40fc57 // bfdot v23.4s, v2.8h, v0.8h\n" | ||
| 1494 | ".inst 0x6e40fc78 // bfdot v24.4s, v3.8h, v0.8h\n" | ||
| 1495 | "ldr q10, [x20, #0x50]\n" | ||
| 1496 | ".inst 0x6e40fc99 // bfdot v25.4s, v4.8h, v0.8h\n" | ||
| 1497 | ".inst 0x6e40fcba // bfdot v26.4s, v5.8h, v0.8h\n" | ||
| 1498 | "add x21, x21, #0x60\n" | ||
| 1499 | ".inst 0x6e40fcdb // bfdot v27.4s, v6.8h, v0.8h\n" | ||
| 1500 | ".inst 0x6e40fcfc // bfdot v28.4s, v7.8h, v0.8h\n" | ||
| 1501 | "add x20, x20, #0x60\n" | ||
| 1502 | "prfm pldl1keep, [%x[B_ptr], #0x400]\n" | ||
| 1503 | ".inst 0x6e40fd1d // bfdot v29.4s, v8.8h, v0.8h\n" | ||
| 1504 | ".inst 0x6e40fd3e // bfdot v30.4s, v9.8h, v0.8h\n" | ||
| 1505 | "prfm pldl1keep, [%x[B_ptr], #0x440]\n" | ||
| 1506 | "prfm pldl1keep, [%x[B_ptr], #0x480]\n" | ||
| 1507 | ".inst 0x6e40fd5f // bfdot v31.4s, v10.8h, v0.8h\n" | ||
| 1508 | "prfm pldl1keep, [%x[B_ptr], #0x4c0]\n" | ||
| 1509 | "prfm pldl1keep, [x23, #0x80]\n" | ||
| 1510 | "92:" // Width 9: Multiply loop: Main loop skip | ||
| 1511 | "cbz x24, 95f\n" | ||
| 1512 | "tbz x24, #1, 93f\n" | ||
| 1513 | "ldr s0, [x23], #0x4\n" | ||
| 1514 | "tbz x24, #0, 94f\n" | ||
| 1515 | "ld1 { v0.h }[2], [x23]\n" | ||
| 1516 | "b 94f\n" | ||
| 1517 | "93:" // Width 9: Multiply loop: Ragged operand read: partial_1_0 | ||
| 1518 | "ldr h0, [x23, #0x0]\n" | ||
| 1519 | "94:" // Width 9: Multiply loop: Ragged operand read: Done | ||
| 1520 | "ldr q11, [%x[B_ptr], #0x0]\n" | ||
| 1521 | "ldr q12, [%x[B_ptr], #0x10]\n" | ||
| 1522 | "dup v0.2d, v0.d[0]\n" | ||
| 1523 | "ldr q13, [%x[B_ptr], #0x20]\n" | ||
| 1524 | "ldr q1, [%x[B_ptr], #0x30]\n" | ||
| 1525 | "ldr q2, [%x[B_ptr], #0x40]\n" | ||
| 1526 | "ldr q3, [%x[B_ptr], #0x50]\n" | ||
| 1527 | "add %x[B_ptr], %x[B_ptr], #0x90\n" | ||
| 1528 | "ldr q4, [x21, #0x0]\n" | ||
| 1529 | "ldr q5, [x21, #0x10]\n" | ||
| 1530 | ".inst 0x6e40fd6e // bfdot v14.4s, v11.8h, v0.8h\n" | ||
| 1531 | ".inst 0x6e40fd8f // bfdot v15.4s, v12.8h, v0.8h\n" | ||
| 1532 | "ldr q6, [x21, #0x20]\n" | ||
| 1533 | "ldr q7, [x21, #0x30]\n" | ||
| 1534 | ".inst 0x6e40fdb0 // bfdot v16.4s, v13.8h, v0.8h\n" | ||
| 1535 | ".inst 0x6e40fc31 // bfdot v17.4s, v1.8h, v0.8h\n" | ||
| 1536 | "ldr q8, [x21, #0x40]\n" | ||
| 1537 | "ldr q9, [x21, #0x50]\n" | ||
| 1538 | ".inst 0x6e40fc52 // bfdot v18.4s, v2.8h, v0.8h\n" | ||
| 1539 | ".inst 0x6e40fc73 // bfdot v19.4s, v3.8h, v0.8h\n" | ||
| 1540 | "ldr q10, [x20, #0x0]\n" | ||
| 1541 | "ldr q11, [x20, #0x10]\n" | ||
| 1542 | ".inst 0x6e40fc94 // bfdot v20.4s, v4.8h, v0.8h\n" | ||
| 1543 | ".inst 0x6e40fcb5 // bfdot v21.4s, v5.8h, v0.8h\n" | ||
| 1544 | "ldr q12, [x20, #0x20]\n" | ||
| 1545 | "ldr q13, [x20, #0x30]\n" | ||
| 1546 | ".inst 0x6e40fcd6 // bfdot v22.4s, v6.8h, v0.8h\n" | ||
| 1547 | ".inst 0x6e40fcf7 // bfdot v23.4s, v7.8h, v0.8h\n" | ||
| 1548 | "ldr q1, [x20, #0x40]\n" | ||
| 1549 | "ldr q2, [x20, #0x50]\n" | ||
| 1550 | ".inst 0x6e40fd18 // bfdot v24.4s, v8.8h, v0.8h\n" | ||
| 1551 | ".inst 0x6e40fd39 // bfdot v25.4s, v9.8h, v0.8h\n" | ||
| 1552 | ".inst 0x6e40fd5a // bfdot v26.4s, v10.8h, v0.8h\n" | ||
| 1553 | ".inst 0x6e40fd7b // bfdot v27.4s, v11.8h, v0.8h\n" | ||
| 1554 | ".inst 0x6e40fd9c // bfdot v28.4s, v12.8h, v0.8h\n" | ||
| 1555 | ".inst 0x6e40fdbd // bfdot v29.4s, v13.8h, v0.8h\n" | ||
| 1556 | ".inst 0x6e40fc3e // bfdot v30.4s, v1.8h, v0.8h\n" | ||
| 1557 | ".inst 0x6e40fc5f // bfdot v31.4s, v2.8h, v0.8h\n" | ||
| 1558 | "95:" // Width 9: Multiply loop: No odd multiplies | ||
| 1559 | "prfm pstl1keep, [%x[output_ptr], #0x0]\n" | ||
| 1560 | "faddp v14.4s, v14.4s, v15.4s\n" | ||
| 1561 | "faddp v15.4s, v16.4s, v17.4s\n" | ||
| 1562 | "faddp v16.4s, v18.4s, v19.4s\n" | ||
| 1563 | "faddp v17.4s, v20.4s, v21.4s\n" | ||
| 1564 | "faddp v18.4s, v22.4s, v23.4s\n" | ||
| 1565 | "faddp v19.4s, v24.4s, v25.4s\n" | ||
| 1566 | "faddp v20.4s, v26.4s, v27.4s\n" | ||
| 1567 | "faddp v21.4s, v28.4s, v29.4s\n" | ||
| 1568 | "faddp v22.4s, v30.4s, v31.4s\n" | ||
| 1569 | "tbz %x[flags], #1, 96f\n" | ||
| 1570 | "add x21, %x[args_ptr], %[offset_max]\n" | ||
| 1571 | "add x20, %x[args_ptr], %[offset_min]\n" | ||
| 1572 | "ld1r { v24.4s }, [x21]\n" | ||
| 1573 | "ld1r { v23.4s }, [x20]\n" | ||
| 1574 | "fmin v14.4s, v14.4s, v24.4s\n" | ||
| 1575 | "fmin v15.4s, v15.4s, v24.4s\n" | ||
| 1576 | "fmin v16.4s, v16.4s, v24.4s\n" | ||
| 1577 | "fmin v17.4s, v17.4s, v24.4s\n" | ||
| 1578 | "fmin v18.4s, v18.4s, v24.4s\n" | ||
| 1579 | "fmin v19.4s, v19.4s, v24.4s\n" | ||
| 1580 | "fmin v20.4s, v20.4s, v24.4s\n" | ||
| 1581 | "fmin v21.4s, v21.4s, v24.4s\n" | ||
| 1582 | "fmin v22.4s, v22.4s, v24.4s\n" | ||
| 1583 | "fmax v14.4s, v14.4s, v23.4s\n" | ||
| 1584 | "fmax v15.4s, v15.4s, v23.4s\n" | ||
| 1585 | "fmax v16.4s, v16.4s, v23.4s\n" | ||
| 1586 | "fmax v17.4s, v17.4s, v23.4s\n" | ||
| 1587 | "fmax v18.4s, v18.4s, v23.4s\n" | ||
| 1588 | "fmax v19.4s, v19.4s, v23.4s\n" | ||
| 1589 | "fmax v20.4s, v20.4s, v23.4s\n" | ||
| 1590 | "fmax v21.4s, v21.4s, v23.4s\n" | ||
| 1591 | "fmax v22.4s, v22.4s, v23.4s\n" | ||
| 1592 | "96:" // Width 9: No activation | ||
| 1593 | "cmp %x[N], #0x24\n" | ||
| 1594 | "str q14, [%x[output_ptr], #0x0]\n" | ||
| 1595 | "str q15, [%x[output_ptr], #0x10]\n" | ||
| 1596 | "str q16, [%x[output_ptr], #0x20]\n" | ||
| 1597 | "str q17, [%x[output_ptr], #0x30]\n" | ||
| 1598 | "str q18, [%x[output_ptr], #0x40]\n" | ||
| 1599 | "str q19, [%x[output_ptr], #0x50]\n" | ||
| 1600 | "str q20, [%x[output_ptr], #0x60]\n" | ||
| 1601 | "str q21, [%x[output_ptr], #0x70]\n" | ||
| 1602 | "add %x[output_ptr], %x[output_ptr], #0x80\n" | ||
| 1603 | "blt 97f\n" | ||
| 1604 | "str q22, [%x[output_ptr], #0x0]\n" | ||
| 1605 | "add %x[output_ptr], %x[output_ptr], #0x10\n" | ||
| 1606 | "b 99f\n" | ||
| 1607 | "97:" // Width 9: Partial writeback | ||
| 1608 | "tbz %x[N], #1, 98f\n" | ||
| 1609 | "str d22, [%x[output_ptr]], #0x8\n" | ||
| 1610 | "tbz %x[N], #0, 99f\n" | ||
| 1611 | "st1 { v22.s }[2], [%x[output_ptr]]\n" | ||
| 1612 | "b 99f\n" | ||
| 1613 | "98:" // Width 9: Partial direct writeback: partial_1_32 | ||
| 1614 | "tbz %x[N], #0, 99f\n" | ||
| 1615 | "str s22, [%x[output_ptr], #0x0]\n" | ||
| 1616 | "99:" // Width 9: Writeback done | ||
| 1617 | "subs x25, x25, #0x9\n" | ||
| 1618 | "mov %x[B_ptr], x22\n" | ||
| 1619 | "sub %x[N], %x[N], #0x24\n" | ||
| 1620 | "bgt 1b\n" | ||
| 1621 | "100:" // Exit | ||
| 1622 | : [B_ptr] "+&r"(B_ptr), [N] "+&r"(N), [output_ptr] "+&r"(output_ptr) | ||
| 1623 | 582 | : [A_ptr] "r"(A_ptr), [K] "r"(K), [args_ptr] "r"(&ka), [flags] "r"(flags), | |
| 1624 | [offset_max] "I"(offsetof(KernelArgs, maxval)), [offset_min] "I"(offsetof(KernelArgs, minval)) | ||
| 1625 | : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", | ||
| 1626 | "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", | ||
| 1627 | "v30", "v31", "x20", "x21", "x22", "x23", "x24", "x25", "x26"); | ||
| 1628 | 582 | } | |
| 1629 | |||
| 1630 | #endif // Architectural features check. | ||
| 1631 |