| 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_bf16p8x4_bf16p12x4b_8x12_neon_mmla.h" | ||
| 15 | |||
| 16 | #include <arm_bf16.h> | ||
| 17 | #include <arm_neon.h> | ||
| 18 | #include <stddef.h> | ||
| 19 | #include <stdint.h> | ||
| 20 | |||
| 21 | #include "kai/kai_common.h" | ||
| 22 | |||
| 23 | static const size_t kai_mr = 8; | ||
| 24 | static const size_t kai_nr = 12; | ||
| 25 | static const size_t kai_kr = 4; | ||
| 26 | static const size_t kai_sr = 1; | ||
| 27 | |||
| 28 | 200 | size_t kai_get_m_step_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla(void) { | |
| 29 | 200 | return kai_mr; | |
| 30 | } | ||
| 31 | |||
| 32 | 200 | size_t kai_get_n_step_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla(void) { | |
| 33 | 200 | return kai_nr; | |
| 34 | } | ||
| 35 | |||
| 36 | ✗ | size_t kai_get_mr_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla(void) { | |
| 37 | ✗ | return kai_mr; | |
| 38 | } | ||
| 39 | |||
| 40 | 40 | size_t kai_get_nr_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla(void) { | |
| 41 | 40 | return kai_nr; | |
| 42 | } | ||
| 43 | |||
| 44 | 40 | size_t kai_get_kr_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla(void) { | |
| 45 | 40 | return kai_kr; | |
| 46 | } | ||
| 47 | |||
| 48 | 40 | size_t kai_get_sr_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla(void) { | |
| 49 | 40 | return kai_sr; | |
| 50 | } | ||
| 51 | |||
| 52 | 184 | size_t kai_get_lhs_packed_offset_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla(size_t m_idx, size_t k) { | |
| 53 | − | KAI_ASSUME(m_idx % kai_mr == 0); | |
| 54 | |||
| 55 | 184 | return m_idx * kai_roundup(k, kai_kr) * sizeof(uint16_t); | |
| 56 | } | ||
| 57 | |||
| 58 | 184 | size_t kai_get_rhs_packed_offset_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla(size_t n_idx, size_t k) { | |
| 59 | − | KAI_ASSUME(n_idx % kai_nr == 0); | |
| 60 | |||
| 61 | 184 | return n_idx * (sizeof(float) + kai_roundup(k, kai_kr) * sizeof(uint16_t)); | |
| 62 | } | ||
| 63 | |||
| 64 | 184 | size_t kai_get_dst_offset_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla( | |
| 65 | size_t m_idx, size_t n_idx, size_t stride) { | ||
| 66 | − | KAI_ASSUME(m_idx % kai_mr == 0); | |
| 67 | − | KAI_ASSUME(n_idx % kai_nr == 0); | |
| 68 | |||
| 69 | 184 | return m_idx * stride + n_idx * sizeof(float); | |
| 70 | } | ||
| 71 | |||
| 72 | 184 | size_t kai_get_dst_size_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla(size_t m, size_t n) { | |
| 73 | 184 | return m * n * sizeof(float); | |
| 74 | } | ||
| 75 | |||
| 76 | 185 | void kai_run_matmul_clamp_f32_bf16p8x4_bf16p12x4b_8x12_neon_mmla( | |
| 77 | size_t m, size_t n, size_t k, // | ||
| 78 | const void* lhs_packed, // | ||
| 79 | const void* rhs_packed, // | ||
| 80 | void* dst, size_t dst_stride_row, size_t dst_stride_col, // | ||
| 81 | float clamp_min, float clamp_max) { | ||
| 82 | − | KAI_ASSERT(dst_stride_col == sizeof(float)); | |
| 83 | |||
| 84 | 185 | const void* Apanel = lhs_packed; | |
| 85 | 185 | void* Cpanel = dst; | |
| 86 | 185 | size_t ldc = dst_stride_row / sizeof(float); | |
| 87 | |||
| 88 | 185 | size_t M = m; | |
| 89 | |||
| 90 | typedef struct { | ||
| 91 | float maxval; | ||
| 92 | float minval; | ||
| 93 | size_t N; | ||
| 94 | size_t K; | ||
| 95 | const void* Bpanel; | ||
| 96 | void* output_ptr; | ||
| 97 | } KernelArgs; | ||
| 98 | |||
| 99 | 185 | KernelArgs ka; | |
| 100 | |||
| 101 | 185 | ka.N = n; | |
| 102 | 185 | ka.K = kai_roundup(k, kai_kr) / kai_kr - 1; | |
| 103 | |||
| 104 | 185 | ka.Bpanel = rhs_packed; | |
| 105 | |||
| 106 | // Direct output. | ||
| 107 | 185 | ka.output_ptr = dst; | |
| 108 | |||
| 109 | // Clamping output. | ||
| 110 | 185 | ka.maxval = clamp_max; | |
| 111 | 185 | ka.minval = clamp_min; | |
| 112 | |||
| 113 | 370 | __asm__ __volatile__( | |
| 114 | "1:" // Height loop | ||
| 115 | "add x11, %x[Cpanel], %x[ldc], LSL #2\n" | ||
| 116 | "add x10, %x[Cpanel], %x[ldc], LSL #1\n" | ||
| 117 | "add x9, x11, %x[ldc], LSL #1\n" | ||
| 118 | "cmp %x[M], #0x8\n" | ||
| 119 | "add x28, %x[Cpanel], %x[ldc], LSL #3\n" | ||
| 120 | "add x27, %x[Cpanel], %x[ldc]\n" | ||
| 121 | "add x26, x10, %x[ldc]\n" | ||
| 122 | "add x25, x11, %x[ldc]\n" | ||
| 123 | "add x24, x9, %x[ldc]\n" | ||
| 124 | "bge 2f\n" | ||
| 125 | "cmp %x[M], #0x2\n" | ||
| 126 | "mov x24, %x[Cpanel]\n" | ||
| 127 | "csel x27, x27, %x[Cpanel], GE\n" | ||
| 128 | "csel x10, x10, %x[Cpanel], GT\n" | ||
| 129 | "cmp %x[M], #0x4\n" | ||
| 130 | "csel x26, x26, %x[Cpanel], GE\n" | ||
| 131 | "csel x11, x11, %x[Cpanel], GT\n" | ||
| 132 | "cmp %x[M], #0x6\n" | ||
| 133 | "csel x25, x25, %x[Cpanel], GE\n" | ||
| 134 | "csel x9, x9, %x[Cpanel], GT\n" | ||
| 135 | "2:" // all rows valid | ||
| 136 | "ldr x23, [%x[args_ptr], %[offsetof_N]]\n" | ||
| 137 | "ldr x22, [%x[args_ptr], %[offsetof_Bpanel]]\n" | ||
| 138 | "mov x21, %x[Apanel]\n" | ||
| 139 | "3:" // Width loop | ||
| 140 | "ldr q4, [x22, #0x0]\n" | ||
| 141 | "ldr q5, [x22, #0x10]\n" | ||
| 142 | "mov %x[Apanel], x21\n" | ||
| 143 | "ldr q6, [x22, #0x20]\n" | ||
| 144 | "ldr x20, [%x[args_ptr], %[offsetof_K]]\n" | ||
| 145 | "add x22, x22, #0x30\n" | ||
| 146 | "ldr q7, [x22, #0x0]\n" | ||
| 147 | "ldr q0, [%x[Apanel], #0x0]\n" | ||
| 148 | "ldr q1, [%x[Apanel], #0x10]\n" | ||
| 149 | "zip1 v8.2d, v4.2d, v4.2d\n" | ||
| 150 | "ldr q2, [%x[Apanel], #0x20]\n" | ||
| 151 | "zip2 v11.2d, v4.2d, v4.2d\n" | ||
| 152 | "ldr q4, [x22, #0x10]\n" | ||
| 153 | "zip1 v9.2d, v5.2d, v5.2d\n" | ||
| 154 | "zip2 v12.2d, v5.2d, v5.2d\n" | ||
| 155 | "cmp x20, #0x2\n" | ||
| 156 | "zip1 v10.2d, v6.2d, v6.2d\n" | ||
| 157 | "zip2 v13.2d, v6.2d, v6.2d\n" | ||
| 158 | "prfm pldl1keep, [%x[Apanel], #0x0]\n" | ||
| 159 | "mov v14.16b, v8.16b\n" | ||
| 160 | "mov v17.16b, v11.16b\n" | ||
| 161 | "prfm pldl1keep, [x22, #0x0]\n" | ||
| 162 | "mov v15.16b, v9.16b\n" | ||
| 163 | "mov v18.16b, v12.16b\n" | ||
| 164 | "prfm pldl1keep, [x22, #0x40]\n" | ||
| 165 | "mov v16.16b, v10.16b\n" | ||
| 166 | "mov v19.16b, v13.16b\n" | ||
| 167 | "prfm pldl1keep, [%x[Apanel], #0x40]\n" | ||
| 168 | "mov v20.16b, v8.16b\n" | ||
| 169 | "mov v21.16b, v9.16b\n" | ||
| 170 | "prfm pldl1keep, [x22, #0x80]\n" | ||
| 171 | "mov v22.16b, v10.16b\n" | ||
| 172 | "mov v23.16b, v11.16b\n" | ||
| 173 | "prfm pldl1keep, [%x[Apanel], #0x80]\n" | ||
| 174 | "mov v24.16b, v12.16b\n" | ||
| 175 | "mov v25.16b, v13.16b\n" | ||
| 176 | "prfm pldl1keep, [x22, #0xc0]\n" | ||
| 177 | "mov v26.16b, v8.16b\n" | ||
| 178 | "mov v27.16b, v9.16b\n" | ||
| 179 | "prfm pldl1keep, [x22, #0x100]\n" | ||
| 180 | "mov v28.16b, v10.16b\n" | ||
| 181 | "mov v29.16b, v11.16b\n" | ||
| 182 | "prfm pldl1keep, [%x[Apanel], #0xc0]\n" | ||
| 183 | "mov v30.16b, v12.16b\n" | ||
| 184 | "mov v31.16b, v13.16b\n" | ||
| 185 | "prfm pldl1keep, [x22, #0x140]\n" | ||
| 186 | "add x22, x22, #0x20\n" | ||
| 187 | "add %x[Apanel], %x[Apanel], #0x30\n" | ||
| 188 | "blt 5f\n" | ||
| 189 | "4:" // main loop head | ||
| 190 | "ldr q3, [%x[Apanel], #0x0]\n" | ||
| 191 | "ldr q5, [x22, #0x0]\n" | ||
| 192 | ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" | ||
| 193 | "ldr q6, [x22, #0x10]\n" | ||
| 194 | ".inst 0x6e44ec0b // bfmmla v11.4s, v0.8h, v4.8h\n" | ||
| 195 | ".inst 0x6e47ec2e // bfmmla v14.4s, v1.8h, v7.8h\n" | ||
| 196 | ".inst 0x6e44ec31 // bfmmla v17.4s, v1.8h, v4.8h\n" | ||
| 197 | ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n" | ||
| 198 | "sub x20, x20, #0x2\n" | ||
| 199 | ".inst 0x6e44ec57 // bfmmla v23.4s, v2.8h, v4.8h\n" | ||
| 200 | ".inst 0x6e47ec7a // bfmmla v26.4s, v3.8h, v7.8h\n" | ||
| 201 | "ldr q7, [x22, #0x20]\n" | ||
| 202 | ".inst 0x6e44ec7d // bfmmla v29.4s, v3.8h, v4.8h\n" | ||
| 203 | "ldr q4, [x22, #0x30]\n" | ||
| 204 | ".inst 0x6e45ec09 // bfmmla v9.4s, v0.8h, v5.8h\n" | ||
| 205 | ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" | ||
| 206 | ".inst 0x6e45ec2f // bfmmla v15.4s, v1.8h, v5.8h\n" | ||
| 207 | "cmp x20, #0x2\n" | ||
| 208 | ".inst 0x6e46ec32 // bfmmla v18.4s, v1.8h, v6.8h\n" | ||
| 209 | ".inst 0x6e45ec55 // bfmmla v21.4s, v2.8h, v5.8h\n" | ||
| 210 | "prfm pldl1keep, [%x[Apanel], #0x100]\n" | ||
| 211 | ".inst 0x6e46ec58 // bfmmla v24.4s, v2.8h, v6.8h\n" | ||
| 212 | ".inst 0x6e45ec7b // bfmmla v27.4s, v3.8h, v5.8h\n" | ||
| 213 | "ldr q5, [x22, #0x40]\n" | ||
| 214 | ".inst 0x6e46ec7e // bfmmla v30.4s, v3.8h, v6.8h\n" | ||
| 215 | "ldr q6, [x22, #0x50]\n" | ||
| 216 | ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" | ||
| 217 | ".inst 0x6e44ec0d // bfmmla v13.4s, v0.8h, v4.8h\n" | ||
| 218 | "ldr q0, [%x[Apanel], #0x10]\n" | ||
| 219 | ".inst 0x6e47ec30 // bfmmla v16.4s, v1.8h, v7.8h\n" | ||
| 220 | ".inst 0x6e44ec33 // bfmmla v19.4s, v1.8h, v4.8h\n" | ||
| 221 | "ldr q1, [%x[Apanel], #0x20]\n" | ||
| 222 | ".inst 0x6e47ec56 // bfmmla v22.4s, v2.8h, v7.8h\n" | ||
| 223 | ".inst 0x6e44ec59 // bfmmla v25.4s, v2.8h, v4.8h\n" | ||
| 224 | "ldr q2, [%x[Apanel], #0x30]\n" | ||
| 225 | ".inst 0x6e47ec7c // bfmmla v28.4s, v3.8h, v7.8h\n" | ||
| 226 | "ldr q7, [x22, #0x60]\n" | ||
| 227 | ".inst 0x6e44ec7f // bfmmla v31.4s, v3.8h, v4.8h\n" | ||
| 228 | "ldr q3, [%x[Apanel], #0x40]\n" | ||
| 229 | "ldr q4, [x22, #0x70]\n" | ||
| 230 | ".inst 0x6e45ec08 // bfmmla v8.4s, v0.8h, v5.8h\n" | ||
| 231 | ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" | ||
| 232 | ".inst 0x6e45ec2e // bfmmla v14.4s, v1.8h, v5.8h\n" | ||
| 233 | ".inst 0x6e46ec31 // bfmmla v17.4s, v1.8h, v6.8h\n" | ||
| 234 | "prfm pldl1keep, [x22, #0x180]\n" | ||
| 235 | ".inst 0x6e45ec54 // bfmmla v20.4s, v2.8h, v5.8h\n" | ||
| 236 | ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" | ||
| 237 | "prfm pldl1keep, [x22, #0x1c0]\n" | ||
| 238 | ".inst 0x6e45ec7a // bfmmla v26.4s, v3.8h, v5.8h\n" | ||
| 239 | "ldr q5, [x22, #0x80]\n" | ||
| 240 | ".inst 0x6e46ec7d // bfmmla v29.4s, v3.8h, v6.8h\n" | ||
| 241 | "ldr q6, [x22, #0x90]\n" | ||
| 242 | "prfm pldl1keep, [%x[Apanel], #0x140]\n" | ||
| 243 | ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" | ||
| 244 | "prfm pldl1keep, [x22, #0x200]\n" | ||
| 245 | ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" | ||
| 246 | ".inst 0x6e47ec2f // bfmmla v15.4s, v1.8h, v7.8h\n" | ||
| 247 | ".inst 0x6e44ec32 // bfmmla v18.4s, v1.8h, v4.8h\n" | ||
| 248 | ".inst 0x6e47ec55 // bfmmla v21.4s, v2.8h, v7.8h\n" | ||
| 249 | ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n" | ||
| 250 | ".inst 0x6e47ec7b // bfmmla v27.4s, v3.8h, v7.8h\n" | ||
| 251 | "ldr q7, [x22, #0xa0]\n" | ||
| 252 | ".inst 0x6e44ec7e // bfmmla v30.4s, v3.8h, v4.8h\n" | ||
| 253 | "ldr q4, [x22, #0xb0]\n" | ||
| 254 | ".inst 0x6e45ec0a // bfmmla v10.4s, v0.8h, v5.8h\n" | ||
| 255 | ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" | ||
| 256 | "ldr q0, [%x[Apanel], #0x50]\n" | ||
| 257 | ".inst 0x6e45ec30 // bfmmla v16.4s, v1.8h, v5.8h\n" | ||
| 258 | ".inst 0x6e46ec33 // bfmmla v19.4s, v1.8h, v6.8h\n" | ||
| 259 | "ldr q1, [%x[Apanel], #0x60]\n" | ||
| 260 | ".inst 0x6e45ec56 // bfmmla v22.4s, v2.8h, v5.8h\n" | ||
| 261 | ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n" | ||
| 262 | "ldr q2, [%x[Apanel], #0x70]\n" | ||
| 263 | ".inst 0x6e45ec7c // bfmmla v28.4s, v3.8h, v5.8h\n" | ||
| 264 | ".inst 0x6e46ec7f // bfmmla v31.4s, v3.8h, v6.8h\n" | ||
| 265 | "add %x[Apanel], %x[Apanel], #0x80\n" | ||
| 266 | "add x22, x22, #0xc0\n" | ||
| 267 | "bge 4b\n" | ||
| 268 | "5:" // main loop skip | ||
| 269 | "ldr q3, [%x[Apanel], #0x0]\n" | ||
| 270 | "ldr q5, [x22, #0x0]\n" | ||
| 271 | ".inst 0x6e47ec08 // bfmmla v8.4s, v0.8h, v7.8h\n" | ||
| 272 | "ldr q6, [x22, #0x10]\n" | ||
| 273 | ".inst 0x6e44ec0b // bfmmla v11.4s, v0.8h, v4.8h\n" | ||
| 274 | ".inst 0x6e47ec2e // bfmmla v14.4s, v1.8h, v7.8h\n" | ||
| 275 | ".inst 0x6e44ec31 // bfmmla v17.4s, v1.8h, v4.8h\n" | ||
| 276 | ".inst 0x6e47ec54 // bfmmla v20.4s, v2.8h, v7.8h\n" | ||
| 277 | "add %x[Apanel], %x[Apanel], #0x10\n" | ||
| 278 | ".inst 0x6e44ec57 // bfmmla v23.4s, v2.8h, v4.8h\n" | ||
| 279 | ".inst 0x6e47ec7a // bfmmla v26.4s, v3.8h, v7.8h\n" | ||
| 280 | "ldr q7, [x22, #0x20]\n" | ||
| 281 | ".inst 0x6e44ec7d // bfmmla v29.4s, v3.8h, v4.8h\n" | ||
| 282 | "ldr q4, [x22, #0x30]\n" | ||
| 283 | ".inst 0x6e45ec09 // bfmmla v9.4s, v0.8h, v5.8h\n" | ||
| 284 | ".inst 0x6e46ec0c // bfmmla v12.4s, v0.8h, v6.8h\n" | ||
| 285 | ".inst 0x6e45ec2f // bfmmla v15.4s, v1.8h, v5.8h\n" | ||
| 286 | "add x22, x22, #0x40\n" | ||
| 287 | ".inst 0x6e46ec32 // bfmmla v18.4s, v1.8h, v6.8h\n" | ||
| 288 | ".inst 0x6e45ec55 // bfmmla v21.4s, v2.8h, v5.8h\n" | ||
| 289 | ".inst 0x6e46ec58 // bfmmla v24.4s, v2.8h, v6.8h\n" | ||
| 290 | ".inst 0x6e45ec7b // bfmmla v27.4s, v3.8h, v5.8h\n" | ||
| 291 | ".inst 0x6e46ec7e // bfmmla v30.4s, v3.8h, v6.8h\n" | ||
| 292 | ".inst 0x6e47ec0a // bfmmla v10.4s, v0.8h, v7.8h\n" | ||
| 293 | ".inst 0x6e44ec0d // bfmmla v13.4s, v0.8h, v4.8h\n" | ||
| 294 | ".inst 0x6e47ec30 // bfmmla v16.4s, v1.8h, v7.8h\n" | ||
| 295 | ".inst 0x6e44ec33 // bfmmla v19.4s, v1.8h, v4.8h\n" | ||
| 296 | ".inst 0x6e47ec56 // bfmmla v22.4s, v2.8h, v7.8h\n" | ||
| 297 | ".inst 0x6e44ec59 // bfmmla v25.4s, v2.8h, v4.8h\n" | ||
| 298 | ".inst 0x6e47ec7c // bfmmla v28.4s, v3.8h, v7.8h\n" | ||
| 299 | ".inst 0x6e44ec7f // bfmmla v31.4s, v3.8h, v4.8h\n" | ||
| 300 | "cbz x20, 6f\n" | ||
| 301 | "ldr q5, [x22, #0x0]\n" | ||
| 302 | "ldr q0, [%x[Apanel], #0x0]\n" | ||
| 303 | "ldr q1, [%x[Apanel], #0x10]\n" | ||
| 304 | "ldr q6, [x22, #0x10]\n" | ||
| 305 | "ldr q2, [%x[Apanel], #0x20]\n" | ||
| 306 | "ldr q3, [%x[Apanel], #0x30]\n" | ||
| 307 | "add %x[Apanel], %x[Apanel], #0x40\n" | ||
| 308 | "ldr q7, [x22, #0x20]\n" | ||
| 309 | "ldr q4, [x22, #0x30]\n" | ||
| 310 | ".inst 0x6e45ec08 // bfmmla v8.4s, v0.8h, v5.8h\n" | ||
| 311 | ".inst 0x6e46ec0b // bfmmla v11.4s, v0.8h, v6.8h\n" | ||
| 312 | ".inst 0x6e45ec2e // bfmmla v14.4s, v1.8h, v5.8h\n" | ||
| 313 | ".inst 0x6e46ec31 // bfmmla v17.4s, v1.8h, v6.8h\n" | ||
| 314 | ".inst 0x6e45ec54 // bfmmla v20.4s, v2.8h, v5.8h\n" | ||
| 315 | ".inst 0x6e46ec57 // bfmmla v23.4s, v2.8h, v6.8h\n" | ||
| 316 | ".inst 0x6e45ec7a // bfmmla v26.4s, v3.8h, v5.8h\n" | ||
| 317 | "ldr q5, [x22, #0x40]\n" | ||
| 318 | ".inst 0x6e46ec7d // bfmmla v29.4s, v3.8h, v6.8h\n" | ||
| 319 | "ldr q6, [x22, #0x50]\n" | ||
| 320 | ".inst 0x6e47ec09 // bfmmla v9.4s, v0.8h, v7.8h\n" | ||
| 321 | ".inst 0x6e44ec0c // bfmmla v12.4s, v0.8h, v4.8h\n" | ||
| 322 | ".inst 0x6e47ec2f // bfmmla v15.4s, v1.8h, v7.8h\n" | ||
| 323 | "add x22, x22, #0x60\n" | ||
| 324 | ".inst 0x6e44ec32 // bfmmla v18.4s, v1.8h, v4.8h\n" | ||
| 325 | ".inst 0x6e47ec55 // bfmmla v21.4s, v2.8h, v7.8h\n" | ||
| 326 | ".inst 0x6e44ec58 // bfmmla v24.4s, v2.8h, v4.8h\n" | ||
| 327 | ".inst 0x6e47ec7b // bfmmla v27.4s, v3.8h, v7.8h\n" | ||
| 328 | ".inst 0x6e44ec7e // bfmmla v30.4s, v3.8h, v4.8h\n" | ||
| 329 | ".inst 0x6e45ec0a // bfmmla v10.4s, v0.8h, v5.8h\n" | ||
| 330 | ".inst 0x6e46ec0d // bfmmla v13.4s, v0.8h, v6.8h\n" | ||
| 331 | ".inst 0x6e45ec30 // bfmmla v16.4s, v1.8h, v5.8h\n" | ||
| 332 | ".inst 0x6e46ec33 // bfmmla v19.4s, v1.8h, v6.8h\n" | ||
| 333 | ".inst 0x6e45ec56 // bfmmla v22.4s, v2.8h, v5.8h\n" | ||
| 334 | ".inst 0x6e46ec59 // bfmmla v25.4s, v2.8h, v6.8h\n" | ||
| 335 | ".inst 0x6e45ec7c // bfmmla v28.4s, v3.8h, v5.8h\n" | ||
| 336 | ".inst 0x6e46ec7f // bfmmla v31.4s, v3.8h, v6.8h\n" | ||
| 337 | "6:" // multiply loop done | ||
| 338 | "add x20, %x[args_ptr], %[offset_max]\n" | ||
| 339 | "uzp1 v7.2d, v8.2d, v11.2d\n" | ||
| 340 | "uzp2 v8.2d, v8.2d, v11.2d\n" | ||
| 341 | "ld1r { v1.4s }, [x20]\n" | ||
| 342 | "uzp1 v11.2d, v9.2d, v12.2d\n" | ||
| 343 | "uzp2 v9.2d, v9.2d, v12.2d\n" | ||
| 344 | "uzp1 v12.2d, v10.2d, v13.2d\n" | ||
| 345 | "uzp2 v10.2d, v10.2d, v13.2d\n" | ||
| 346 | "add x20, %x[args_ptr], %[offset_min]\n" | ||
| 347 | "ld1r { v0.4s }, [x20]\n" | ||
| 348 | "uzp1 v13.2d, v14.2d, v17.2d\n" | ||
| 349 | "uzp2 v14.2d, v14.2d, v17.2d\n" | ||
| 350 | "uzp1 v17.2d, v15.2d, v18.2d\n" | ||
| 351 | "uzp2 v15.2d, v15.2d, v18.2d\n" | ||
| 352 | "cmp x23, #0xc\n" | ||
| 353 | "uzp1 v18.2d, v16.2d, v19.2d\n" | ||
| 354 | "uzp2 v16.2d, v16.2d, v19.2d\n" | ||
| 355 | "uzp1 v19.2d, v20.2d, v23.2d\n" | ||
| 356 | "uzp2 v20.2d, v20.2d, v23.2d\n" | ||
| 357 | "uzp1 v23.2d, v21.2d, v24.2d\n" | ||
| 358 | "uzp2 v21.2d, v21.2d, v24.2d\n" | ||
| 359 | "uzp1 v24.2d, v22.2d, v25.2d\n" | ||
| 360 | "uzp2 v22.2d, v22.2d, v25.2d\n" | ||
| 361 | "uzp1 v25.2d, v26.2d, v29.2d\n" | ||
| 362 | "uzp2 v26.2d, v26.2d, v29.2d\n" | ||
| 363 | "uzp1 v29.2d, v27.2d, v30.2d\n" | ||
| 364 | "uzp2 v27.2d, v27.2d, v30.2d\n" | ||
| 365 | "uzp1 v30.2d, v28.2d, v31.2d\n" | ||
| 366 | "uzp2 v28.2d, v28.2d, v31.2d\n" | ||
| 367 | "fmin v7.4s, v7.4s, v1.4s\n" | ||
| 368 | "fmin v11.4s, v11.4s, v1.4s\n" | ||
| 369 | "fmin v12.4s, v12.4s, v1.4s\n" | ||
| 370 | "fmin v8.4s, v8.4s, v1.4s\n" | ||
| 371 | "fmin v9.4s, v9.4s, v1.4s\n" | ||
| 372 | "fmin v10.4s, v10.4s, v1.4s\n" | ||
| 373 | "fmin v13.4s, v13.4s, v1.4s\n" | ||
| 374 | "fmin v17.4s, v17.4s, v1.4s\n" | ||
| 375 | "fmin v18.4s, v18.4s, v1.4s\n" | ||
| 376 | "fmin v14.4s, v14.4s, v1.4s\n" | ||
| 377 | "fmin v15.4s, v15.4s, v1.4s\n" | ||
| 378 | "fmin v16.4s, v16.4s, v1.4s\n" | ||
| 379 | "fmin v19.4s, v19.4s, v1.4s\n" | ||
| 380 | "fmin v23.4s, v23.4s, v1.4s\n" | ||
| 381 | "fmin v24.4s, v24.4s, v1.4s\n" | ||
| 382 | "fmin v20.4s, v20.4s, v1.4s\n" | ||
| 383 | "fmin v21.4s, v21.4s, v1.4s\n" | ||
| 384 | "fmin v22.4s, v22.4s, v1.4s\n" | ||
| 385 | "fmin v25.4s, v25.4s, v1.4s\n" | ||
| 386 | "fmin v29.4s, v29.4s, v1.4s\n" | ||
| 387 | "fmin v30.4s, v30.4s, v1.4s\n" | ||
| 388 | "fmin v26.4s, v26.4s, v1.4s\n" | ||
| 389 | "fmin v27.4s, v27.4s, v1.4s\n" | ||
| 390 | "fmin v28.4s, v28.4s, v1.4s\n" | ||
| 391 | "fmax v7.4s, v7.4s, v0.4s\n" | ||
| 392 | "fmax v11.4s, v11.4s, v0.4s\n" | ||
| 393 | "fmax v12.4s, v12.4s, v0.4s\n" | ||
| 394 | "fmax v8.4s, v8.4s, v0.4s\n" | ||
| 395 | "fmax v9.4s, v9.4s, v0.4s\n" | ||
| 396 | "fmax v10.4s, v10.4s, v0.4s\n" | ||
| 397 | "fmax v13.4s, v13.4s, v0.4s\n" | ||
| 398 | "fmax v17.4s, v17.4s, v0.4s\n" | ||
| 399 | "fmax v18.4s, v18.4s, v0.4s\n" | ||
| 400 | "fmax v14.4s, v14.4s, v0.4s\n" | ||
| 401 | "fmax v15.4s, v15.4s, v0.4s\n" | ||
| 402 | "fmax v16.4s, v16.4s, v0.4s\n" | ||
| 403 | "fmax v19.4s, v19.4s, v0.4s\n" | ||
| 404 | "fmax v23.4s, v23.4s, v0.4s\n" | ||
| 405 | "fmax v24.4s, v24.4s, v0.4s\n" | ||
| 406 | "fmax v20.4s, v20.4s, v0.4s\n" | ||
| 407 | "fmax v21.4s, v21.4s, v0.4s\n" | ||
| 408 | "fmax v22.4s, v22.4s, v0.4s\n" | ||
| 409 | "fmax v25.4s, v25.4s, v0.4s\n" | ||
| 410 | "fmax v29.4s, v29.4s, v0.4s\n" | ||
| 411 | "fmax v30.4s, v30.4s, v0.4s\n" | ||
| 412 | "fmax v26.4s, v26.4s, v0.4s\n" | ||
| 413 | "fmax v27.4s, v27.4s, v0.4s\n" | ||
| 414 | "fmax v28.4s, v28.4s, v0.4s\n" | ||
| 415 | "blt 7f\n" | ||
| 416 | "str q26, [x24, #0x0]\n" | ||
| 417 | "str q27, [x24, #0x10]\n" | ||
| 418 | "str q28, [x24, #0x20]\n" | ||
| 419 | "add x24, x24, #0x30\n" | ||
| 420 | "str q25, [x9, #0x0]\n" | ||
| 421 | "str q29, [x9, #0x10]\n" | ||
| 422 | "str q30, [x9, #0x20]\n" | ||
| 423 | "add x9, x9, #0x30\n" | ||
| 424 | "str q20, [x25, #0x0]\n" | ||
| 425 | "str q21, [x25, #0x10]\n" | ||
| 426 | "str q22, [x25, #0x20]\n" | ||
| 427 | "add x25, x25, #0x30\n" | ||
| 428 | "str q19, [x11, #0x0]\n" | ||
| 429 | "str q23, [x11, #0x10]\n" | ||
| 430 | "str q24, [x11, #0x20]\n" | ||
| 431 | "add x11, x11, #0x30\n" | ||
| 432 | "str q14, [x26, #0x0]\n" | ||
| 433 | "str q15, [x26, #0x10]\n" | ||
| 434 | "str q16, [x26, #0x20]\n" | ||
| 435 | "add x26, x26, #0x30\n" | ||
| 436 | "str q13, [x10, #0x0]\n" | ||
| 437 | "str q17, [x10, #0x10]\n" | ||
| 438 | "str q18, [x10, #0x20]\n" | ||
| 439 | "add x10, x10, #0x30\n" | ||
| 440 | "str q8, [x27, #0x0]\n" | ||
| 441 | "str q9, [x27, #0x10]\n" | ||
| 442 | "str q10, [x27, #0x20]\n" | ||
| 443 | "add x27, x27, #0x30\n" | ||
| 444 | "str q7, [%x[Cpanel], #0x0]\n" | ||
| 445 | "str q11, [%x[Cpanel], #0x10]\n" | ||
| 446 | "str q12, [%x[Cpanel], #0x20]\n" | ||
| 447 | "add %x[Cpanel], %x[Cpanel], #0x30\n" | ||
| 448 | "b 14f\n" | ||
| 449 | "7:" // partial output | ||
| 450 | "tbz x23, #3, 9f\n" | ||
| 451 | "st1 { v26.4s }, [x24], #0x10\n" | ||
| 452 | "st1 { v27.4s }, [x24], #0x10\n" | ||
| 453 | "st1 { v25.4s }, [x9], #0x10\n" | ||
| 454 | "st1 { v29.4s }, [x9], #0x10\n" | ||
| 455 | "st1 { v20.4s }, [x25], #0x10\n" | ||
| 456 | "st1 { v21.4s }, [x25], #0x10\n" | ||
| 457 | "st1 { v19.4s }, [x11], #0x10\n" | ||
| 458 | "st1 { v23.4s }, [x11], #0x10\n" | ||
| 459 | "st1 { v14.4s }, [x26], #0x10\n" | ||
| 460 | "st1 { v15.4s }, [x26], #0x10\n" | ||
| 461 | "st1 { v13.4s }, [x10], #0x10\n" | ||
| 462 | "st1 { v17.4s }, [x10], #0x10\n" | ||
| 463 | "st1 { v8.4s }, [x27], #0x10\n" | ||
| 464 | "st1 { v9.4s }, [x27], #0x10\n" | ||
| 465 | "st1 { v7.4s }, [%x[Cpanel]], #0x10\n" | ||
| 466 | "st1 { v11.4s }, [%x[Cpanel]], #0x10\n" | ||
| 467 | "tbz x23, #1, 8f\n" | ||
| 468 | "str d28, [x24], #0x8\n" | ||
| 469 | "str d30, [x9], #0x8\n" | ||
| 470 | "str d22, [x25], #0x8\n" | ||
| 471 | "str d24, [x11], #0x8\n" | ||
| 472 | "str d16, [x26], #0x8\n" | ||
| 473 | "str d18, [x10], #0x8\n" | ||
| 474 | "str d10, [x27], #0x8\n" | ||
| 475 | "str d12, [%x[Cpanel]], #0x8\n" | ||
| 476 | "tbz x23, #0, 13f\n" | ||
| 477 | "st1 { v28.s }[2], [x24]\n" | ||
| 478 | "st1 { v30.s }[2], [x9]\n" | ||
| 479 | "st1 { v22.s }[2], [x25]\n" | ||
| 480 | "st1 { v24.s }[2], [x11]\n" | ||
| 481 | "st1 { v16.s }[2], [x26]\n" | ||
| 482 | "st1 { v18.s }[2], [x10]\n" | ||
| 483 | "st1 { v10.s }[2], [x27]\n" | ||
| 484 | "st1 { v12.s }[2], [%x[Cpanel]]\n" | ||
| 485 | "b 13f\n" | ||
| 486 | "8:" // partial result store: partial_1_8 | ||
| 487 | "tbz x23, #0, 13f\n" | ||
| 488 | "str s28, [x24, #0x0]\n" | ||
| 489 | "str s30, [x9, #0x0]\n" | ||
| 490 | "str s22, [x25, #0x0]\n" | ||
| 491 | "str s24, [x11, #0x0]\n" | ||
| 492 | "str s16, [x26, #0x0]\n" | ||
| 493 | "str s18, [x10, #0x0]\n" | ||
| 494 | "str s10, [x27, #0x0]\n" | ||
| 495 | "str s12, [%x[Cpanel], #0x0]\n" | ||
| 496 | "b 13f\n" | ||
| 497 | "9:" // partial result store: partial_4_0 | ||
| 498 | "tbz x23, #2, 11f\n" | ||
| 499 | "st1 { v26.4s }, [x24], #0x10\n" | ||
| 500 | "st1 { v25.4s }, [x9], #0x10\n" | ||
| 501 | "st1 { v20.4s }, [x25], #0x10\n" | ||
| 502 | "st1 { v19.4s }, [x11], #0x10\n" | ||
| 503 | "st1 { v14.4s }, [x26], #0x10\n" | ||
| 504 | "st1 { v13.4s }, [x10], #0x10\n" | ||
| 505 | "st1 { v8.4s }, [x27], #0x10\n" | ||
| 506 | "st1 { v7.4s }, [%x[Cpanel]], #0x10\n" | ||
| 507 | "tbz x23, #1, 10f\n" | ||
| 508 | "str d27, [x24], #0x8\n" | ||
| 509 | "str d29, [x9], #0x8\n" | ||
| 510 | "str d21, [x25], #0x8\n" | ||
| 511 | "str d23, [x11], #0x8\n" | ||
| 512 | "str d15, [x26], #0x8\n" | ||
| 513 | "str d17, [x10], #0x8\n" | ||
| 514 | "str d9, [x27], #0x8\n" | ||
| 515 | "str d11, [%x[Cpanel]], #0x8\n" | ||
| 516 | "tbz x23, #0, 13f\n" | ||
| 517 | "st1 { v27.s }[2], [x24]\n" | ||
| 518 | "st1 { v29.s }[2], [x9]\n" | ||
| 519 | "st1 { v21.s }[2], [x25]\n" | ||
| 520 | "st1 { v23.s }[2], [x11]\n" | ||
| 521 | "st1 { v15.s }[2], [x26]\n" | ||
| 522 | "st1 { v17.s }[2], [x10]\n" | ||
| 523 | "st1 { v9.s }[2], [x27]\n" | ||
| 524 | "st1 { v11.s }[2], [%x[Cpanel]]\n" | ||
| 525 | "b 13f\n" | ||
| 526 | "10:" // partial result store: partial_1_4 | ||
| 527 | "tbz x23, #0, 13f\n" | ||
| 528 | "str s27, [x24, #0x0]\n" | ||
| 529 | "str s29, [x9, #0x0]\n" | ||
| 530 | "str s21, [x25, #0x0]\n" | ||
| 531 | "str s23, [x11, #0x0]\n" | ||
| 532 | "str s15, [x26, #0x0]\n" | ||
| 533 | "str s17, [x10, #0x0]\n" | ||
| 534 | "str s9, [x27, #0x0]\n" | ||
| 535 | "str s11, [%x[Cpanel], #0x0]\n" | ||
| 536 | "b 13f\n" | ||
| 537 | "11:" // partial result store: partial_2_0 | ||
| 538 | "tbz x23, #1, 12f\n" | ||
| 539 | "str d26, [x24], #0x8\n" | ||
| 540 | "str d25, [x9], #0x8\n" | ||
| 541 | "str d20, [x25], #0x8\n" | ||
| 542 | "str d19, [x11], #0x8\n" | ||
| 543 | "str d14, [x26], #0x8\n" | ||
| 544 | "str d13, [x10], #0x8\n" | ||
| 545 | "str d8, [x27], #0x8\n" | ||
| 546 | "str d7, [%x[Cpanel]], #0x8\n" | ||
| 547 | "tbz x23, #0, 13f\n" | ||
| 548 | "st1 { v26.s }[2], [x24]\n" | ||
| 549 | "st1 { v25.s }[2], [x9]\n" | ||
| 550 | "st1 { v20.s }[2], [x25]\n" | ||
| 551 | "st1 { v19.s }[2], [x11]\n" | ||
| 552 | "st1 { v14.s }[2], [x26]\n" | ||
| 553 | "st1 { v13.s }[2], [x10]\n" | ||
| 554 | "st1 { v8.s }[2], [x27]\n" | ||
| 555 | "st1 { v7.s }[2], [%x[Cpanel]]\n" | ||
| 556 | "b 13f\n" | ||
| 557 | "12:" // partial result store: partial_1_0 | ||
| 558 | "str s26, [x24, #0x0]\n" | ||
| 559 | "str s25, [x9, #0x0]\n" | ||
| 560 | "str s20, [x25, #0x0]\n" | ||
| 561 | "str s19, [x11, #0x0]\n" | ||
| 562 | "str s14, [x26, #0x0]\n" | ||
| 563 | "str s13, [x10, #0x0]\n" | ||
| 564 | "str s8, [x27, #0x0]\n" | ||
| 565 | "str s7, [%x[Cpanel], #0x0]\n" | ||
| 566 | "13:" // partial result store: Done | ||
| 567 | "14:" // store done | ||
| 568 | "subs x23, x23, #0xc\n" | ||
| 569 | "bgt 3b\n" | ||
| 570 | "subs %x[M], %x[M], #0x8\n" | ||
| 571 | "mov %x[Cpanel], x28\n" | ||
| 572 | "bgt 1b\n" | ||
| 573 | : [Apanel] "+&r"(Apanel), [Cpanel] "+&r"(Cpanel), [M] "+&r"(M) | ||
| 574 | 185 | : [args_ptr] "r"(&ka), [ldc] "r"(ldc * sizeof(float)), [offset_max] "I"(offsetof(KernelArgs, maxval)), | |
| 575 | [offset_min] "I"(offsetof(KernelArgs, minval)), [offsetof_Bpanel] "I"(offsetof(KernelArgs, Bpanel)), | ||
| 576 | [offsetof_K] "I"(offsetof(KernelArgs, K)), [offsetof_N] "I"(offsetof(KernelArgs, N)) | ||
| 577 | : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", | ||
| 578 | "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", | ||
| 579 | "v30", "v31", "x9", "x10", "x11", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"); | ||
| 580 | 185 | } | |
| 581 | |||
| 582 | #endif // Architectural features check. | ||
| 583 |