| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | // | ||
| 2 | // SPDX-FileCopyrightText: Copyright 2024-2025 Arm Limited and/or its affiliates <open-source-office@arm.com> | ||
| 3 | // | ||
| 4 | // SPDX-License-Identifier: Apache-2.0 | ||
| 5 | // | ||
| 6 | |||
| 7 | // Do not flag up inline assembly blocks | ||
| 8 | #pragma GCC diagnostic ignored "-Woverlength-strings" | ||
| 9 | |||
| 10 | #if !defined(__ARM_FEATURE_MATMUL_INT8) | ||
| 11 | #error "I8mm extension required to compile this micro-kernel" | ||
| 12 | #else | ||
| 13 | #include "kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm.h" | ||
| 14 | |||
| 15 | #include <arm_neon.h> | ||
| 16 | #include <stddef.h> | ||
| 17 | #include <stdint.h> | ||
| 18 | |||
| 19 | #include "kai/kai_common.h" | ||
| 20 | |||
| 21 | static const size_t kai_m_step = 4; | ||
| 22 | static const size_t kai_n_step = 8; | ||
| 23 | static const size_t kai_mr = 4; | ||
| 24 | static const size_t kai_nr = 8; | ||
| 25 | static const size_t kai_kr = 16; | ||
| 26 | static const size_t kai_sr = 2; | ||
| 27 | static const size_t kai_num_bytes_multiplier_lhs = sizeof(float); | ||
| 28 | static const size_t kai_num_bytes_multiplier_rhs = sizeof(float); | ||
| 29 | static const size_t kai_num_bytes_offset_lhs = sizeof(int32_t); | ||
| 30 | static const size_t kai_num_bytes_sum_rhs = sizeof(int32_t); | ||
| 31 | static const size_t kai_num_bytes_bias = sizeof(float); | ||
| 32 | |||
| 33 | 641 | inline static size_t kai_k_roundedup(size_t k) { | |
| 34 | // Since we pack a float and int32 value at the end of the row, | ||
| 35 | // we must make sure that k is a multiple of 4 for alignment | ||
| 36 | 641 | size_t kr_sr_roundedup4 = kai_roundup(kai_kr * kai_sr, 4); | |
| 37 | 1282 | return kai_roundup(k, kr_sr_roundedup4); | |
| 38 | 641 | } | |
| 39 | |||
| 40 | 240 | inline static size_t kai_lhs_packed_stride(size_t k) { | |
| 41 | 240 | const size_t k_internal = kai_k_roundedup(k); | |
| 42 | |||
| 43 | − | KAI_ASSERT((k_internal % 2) == 0); | |
| 44 | |||
| 45 | 480 | return kai_mr * (k_internal * sizeof(int8_t) + kai_num_bytes_multiplier_lhs + kai_num_bytes_offset_lhs); | |
| 46 | 240 | } | |
| 47 | |||
| 48 | 240 | inline static size_t kai_rhs_packed_stride(size_t k) { | |
| 49 | 240 | const size_t k_internal = kai_k_roundedup(k); | |
| 50 | |||
| 51 | − | KAI_ASSERT((k_internal % 2) == 0); | |
| 52 | |||
| 53 | 480 | return kai_nr * ((k_internal / 2) + kai_num_bytes_multiplier_rhs + kai_num_bytes_sum_rhs + kai_num_bytes_bias); | |
| 54 | 240 | } | |
| 55 | |||
| 56 | 320 | size_t kai_get_m_step_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm(void) { | |
| 57 | 320 | return kai_m_step; | |
| 58 | } | ||
| 59 | |||
| 60 | 320 | size_t kai_get_n_step_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm(void) { | |
| 61 | 320 | return kai_n_step; | |
| 62 | } | ||
| 63 | |||
| 64 | 240 | size_t kai_get_mr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm(void) { | |
| 65 | 240 | return kai_mr; | |
| 66 | } | ||
| 67 | |||
| 68 | 240 | size_t kai_get_nr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm(void) { | |
| 69 | 240 | return kai_nr; | |
| 70 | } | ||
| 71 | |||
| 72 | 320 | size_t kai_get_kr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm(void) { | |
| 73 | 320 | return kai_kr; | |
| 74 | } | ||
| 75 | |||
| 76 | 320 | size_t kai_get_sr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm(void) { | |
| 77 | 320 | return kai_sr; | |
| 78 | } | ||
| 79 | |||
| 80 | 240 | size_t kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm(size_t m_idx, size_t k) { | |
| 81 | − | KAI_ASSERT((m_idx % kai_m_step) == 0); | |
| 82 | |||
| 83 | 240 | return (m_idx / kai_mr) * kai_lhs_packed_stride(k); | |
| 84 | } | ||
| 85 | |||
| 86 | 240 | size_t kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm(size_t n_idx, size_t k) { | |
| 87 | − | KAI_ASSERT((n_idx % kai_n_step) == 0); | |
| 88 | |||
| 89 | 240 | return (n_idx / kai_nr) * kai_rhs_packed_stride(k); | |
| 90 | } | ||
| 91 | |||
| 92 | 160 | size_t kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm( | |
| 93 | size_t m_idx, size_t n_idx, size_t dst_stride) { | ||
| 94 | − | KAI_ASSERT((m_idx % kai_m_step) == 0); | |
| 95 | − | KAI_ASSERT((n_idx % kai_n_step) == 0); | |
| 96 | |||
| 97 | 160 | return (n_idx * sizeof(float)) + m_idx * dst_stride; | |
| 98 | } | ||
| 99 | |||
| 100 | 160 | size_t kai_get_dst_size_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm(size_t m, size_t n) { | |
| 101 | 160 | return m * n * sizeof(float); | |
| 102 | } | ||
| 103 | |||
| 104 | 161 | void kai_run_matmul_clamp_f32_qai8dxp4x8_qsi4cxp8x8_4x8x32_neon_i8mm( | |
| 105 | size_t m, size_t n, size_t k, const void* restrict lhs_packed, const void* restrict rhs_packed, | ||
| 106 | float* restrict dst, // NOLINT(readability-non-const-parameter) | ||
| 107 | size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max) { | ||
| 108 | − | KAI_ASSERT(dst_stride_col == sizeof(float)); | |
| 109 | |||
| 110 |
1/2✓ Branch 0 taken 161 times.
✗ Branch 1 not taken.
|
161 | if (m == 0) { |
| 111 | ✗ | return; | |
| 112 | } | ||
| 113 | |||
| 114 | 161 | const size_t k_internal = kai_k_roundedup(k); | |
| 115 | |||
| 116 | 161 | size_t num_blocks = k_internal / 32; | |
| 117 | |||
| 118 | 161 | float clamp_vals[2] = {scalar_min, scalar_max}; | |
| 119 | |||
| 120 | 322 | __asm__ __volatile__( | |
| 121 | "mov x28, #0x80\n" | ||
| 122 | "mov x20, #0x20\n" | ||
| 123 | "movi v12.16b, #0xf0\n" | ||
| 124 | "mov x27, %x[m]\n" | ||
| 125 | "madd x28, %x[num_blocks], x28, x20\n" | ||
| 126 | "cbz x27, 11f\n" | ||
| 127 | "1:" // Row loop | ||
| 128 | "mov x26, %x[rhs_packed]\n" | ||
| 129 | "mov x25, %x[n]\n" | ||
| 130 | "add x24, %x[dst], %x[dst_stride_row], LSL #2\n" | ||
| 131 | "2:" // Column loop | ||
| 132 | "mov x21, %x[lhs_packed]\n" | ||
| 133 | "movi v11.4s, #0x0\n" | ||
| 134 | "movi v10.4s, #0x0\n" | ||
| 135 | "mov x20, %x[num_blocks]\n" | ||
| 136 | "movi v9.4s, #0x0\n" | ||
| 137 | "movi v8.4s, #0x0\n" | ||
| 138 | "movi v7.4s, #0x0\n" | ||
| 139 | "movi v6.4s, #0x0\n" | ||
| 140 | "movi v5.4s, #0x0\n" | ||
| 141 | "movi v4.4s, #0x0\n" | ||
| 142 | "3:" // Sub block loop | ||
| 143 | "ldr q3, [x26, #0x0]\n" | ||
| 144 | "ldr q2, [x26, #0x10]\n" | ||
| 145 | "subs x20, x20, #0x1\n" | ||
| 146 | "ldr q1, [x26, #0x20]\n" | ||
| 147 | "ldr q0, [x26, #0x30]\n" | ||
| 148 | "ldr q31, [x21, #0x0]\n" | ||
| 149 | "ldr q30, [x21, #0x10]\n" | ||
| 150 | "ldr q29, [x26, #0x40]\n" | ||
| 151 | "ldr q28, [x26, #0x50]\n" | ||
| 152 | "shl v19.16b, v3.16b, #0x4\n" | ||
| 153 | "shl v18.16b, v2.16b, #0x4\n" | ||
| 154 | "ldr q27, [x26, #0x60]\n" | ||
| 155 | "ldr q26, [x26, #0x70]\n" | ||
| 156 | "shl v17.16b, v1.16b, #0x4\n" | ||
| 157 | "shl v16.16b, v0.16b, #0x4\n" | ||
| 158 | "ldr q25, [x21, #0x20]\n" | ||
| 159 | "ldr q24, [x21, #0x30]\n" | ||
| 160 | "and v3.16b, v3.16b, v12.16b\n" | ||
| 161 | "and v2.16b, v2.16b, v12.16b\n" | ||
| 162 | "ldr q23, [x21, #0x40]\n" | ||
| 163 | "ldr q22, [x21, #0x50]\n" | ||
| 164 | ".inst 0x4e93a7eb // smmla v11.4s, v31.16b, v19.16b\n" | ||
| 165 | ".inst 0x4e92a7e9 // smmla v9.4s, v31.16b, v18.16b\n" | ||
| 166 | "ldr q21, [x21, #0x60]\n" | ||
| 167 | "ldr q20, [x21, #0x70]\n" | ||
| 168 | ".inst 0x4e91a7ea // smmla v10.4s, v31.16b, v17.16b\n" | ||
| 169 | ".inst 0x4e90a7e8 // smmla v8.4s, v31.16b, v16.16b\n" | ||
| 170 | ".inst 0x4e93a7c7 // smmla v7.4s, v30.16b, v19.16b\n" | ||
| 171 | ".inst 0x4e92a7c5 // smmla v5.4s, v30.16b, v18.16b\n" | ||
| 172 | "shl v19.16b, v29.16b, #0x4\n" | ||
| 173 | "add x26, x26, #0x80\n" | ||
| 174 | ".inst 0x4e91a7c6 // smmla v6.4s, v30.16b, v17.16b\n" | ||
| 175 | ".inst 0x4e90a7c4 // smmla v4.4s, v30.16b, v16.16b\n" | ||
| 176 | "shl v18.16b, v28.16b, #0x4\n" | ||
| 177 | "add x21, x21, #0x80\n" | ||
| 178 | "shl v17.16b, v27.16b, #0x4\n" | ||
| 179 | "shl v16.16b, v26.16b, #0x4\n" | ||
| 180 | ".inst 0x4e93a72b // smmla v11.4s, v25.16b, v19.16b\n" | ||
| 181 | "and v1.16b, v1.16b, v12.16b\n" | ||
| 182 | "and v0.16b, v0.16b, v12.16b\n" | ||
| 183 | ".inst 0x4e92a729 // smmla v9.4s, v25.16b, v18.16b\n" | ||
| 184 | ".inst 0x4e93a707 // smmla v7.4s, v24.16b, v19.16b\n" | ||
| 185 | ".inst 0x4e92a705 // smmla v5.4s, v24.16b, v18.16b\n" | ||
| 186 | "and v29.16b, v29.16b, v12.16b\n" | ||
| 187 | ".inst 0x4e91a72a // smmla v10.4s, v25.16b, v17.16b\n" | ||
| 188 | ".inst 0x4e90a728 // smmla v8.4s, v25.16b, v16.16b\n" | ||
| 189 | "and v28.16b, v28.16b, v12.16b\n" | ||
| 190 | ".inst 0x4e91a706 // smmla v6.4s, v24.16b, v17.16b\n" | ||
| 191 | ".inst 0x4e90a704 // smmla v4.4s, v24.16b, v16.16b\n" | ||
| 192 | "and v27.16b, v27.16b, v12.16b\n" | ||
| 193 | ".inst 0x4e83a6eb // smmla v11.4s, v23.16b, v3.16b\n" | ||
| 194 | ".inst 0x4e82a6e9 // smmla v9.4s, v23.16b, v2.16b\n" | ||
| 195 | "and v26.16b, v26.16b, v12.16b\n" | ||
| 196 | ".inst 0x4e83a6c7 // smmla v7.4s, v22.16b, v3.16b\n" | ||
| 197 | ".inst 0x4e82a6c5 // smmla v5.4s, v22.16b, v2.16b\n" | ||
| 198 | ".inst 0x4e81a6ea // smmla v10.4s, v23.16b, v1.16b\n" | ||
| 199 | ".inst 0x4e80a6e8 // smmla v8.4s, v23.16b, v0.16b\n" | ||
| 200 | ".inst 0x4e81a6c6 // smmla v6.4s, v22.16b, v1.16b\n" | ||
| 201 | ".inst 0x4e80a6c4 // smmla v4.4s, v22.16b, v0.16b\n" | ||
| 202 | ".inst 0x4e9da6ab // smmla v11.4s, v21.16b, v29.16b\n" | ||
| 203 | ".inst 0x4e9ca6a9 // smmla v9.4s, v21.16b, v28.16b\n" | ||
| 204 | ".inst 0x4e9da687 // smmla v7.4s, v20.16b, v29.16b\n" | ||
| 205 | ".inst 0x4e9ca685 // smmla v5.4s, v20.16b, v28.16b\n" | ||
| 206 | ".inst 0x4e9ba6aa // smmla v10.4s, v21.16b, v27.16b\n" | ||
| 207 | ".inst 0x4e9aa6a8 // smmla v8.4s, v21.16b, v26.16b\n" | ||
| 208 | ".inst 0x4e9ba686 // smmla v6.4s, v20.16b, v27.16b\n" | ||
| 209 | ".inst 0x4e9aa684 // smmla v4.4s, v20.16b, v26.16b\n" | ||
| 210 | "bgt 3b\n" | ||
| 211 | "ldr q20, [x26, #0x0]\n" | ||
| 212 | "ldr q19, [x26, #0x10]\n" | ||
| 213 | "uzp1 v0.2d, v11.2d, v9.2d\n" | ||
| 214 | "uzp2 v31.2d, v11.2d, v9.2d\n" | ||
| 215 | "ld1 { v18.4s }, [x21]\n" | ||
| 216 | "ldr q17, [x26, #0x20]\n" | ||
| 217 | "uzp1 v30.2d, v10.2d, v8.2d\n" | ||
| 218 | "uzp2 v29.2d, v10.2d, v8.2d\n" | ||
| 219 | "ldr q28, [x26, #0x30]\n" | ||
| 220 | "uzp1 v27.2d, v7.2d, v5.2d\n" | ||
| 221 | "uzp2 v26.2d, v7.2d, v5.2d\n" | ||
| 222 | "add x21, x21, #0x10\n" | ||
| 223 | "ldr q16, [x21, #0x0]\n" | ||
| 224 | "uzp1 v25.2d, v6.2d, v4.2d\n" | ||
| 225 | "uzp2 v24.2d, v6.2d, v4.2d\n" | ||
| 226 | "add x26, x26, #0x40\n" | ||
| 227 | "mla v0.4s, v20.4s, v18.s[0]\n" | ||
| 228 | "mla v30.4s, v19.4s, v18.s[0]\n" | ||
| 229 | "mla v31.4s, v20.4s, v18.s[1]\n" | ||
| 230 | "mla v29.4s, v19.4s, v18.s[1]\n" | ||
| 231 | "mla v27.4s, v20.4s, v18.s[2]\n" | ||
| 232 | "mla v25.4s, v19.4s, v18.s[2]\n" | ||
| 233 | "fmul v23.4s, v17.4s, v16.s[0]\n" | ||
| 234 | "mla v26.4s, v20.4s, v18.s[3]\n" | ||
| 235 | "mla v24.4s, v19.4s, v18.s[3]\n" | ||
| 236 | "fmul v22.4s, v28.4s, v16.s[0]\n" | ||
| 237 | "scvtf v0.4s, v0.4s\n" | ||
| 238 | "scvtf v30.4s, v30.4s\n" | ||
| 239 | "fmul v21.4s, v17.4s, v16.s[1]\n" | ||
| 240 | "scvtf v31.4s, v31.4s\n" | ||
| 241 | "fmul v20.4s, v28.4s, v16.s[1]\n" | ||
| 242 | "scvtf v29.4s, v29.4s\n" | ||
| 243 | "fmul v19.4s, v17.4s, v16.s[2]\n" | ||
| 244 | "scvtf v27.4s, v27.4s\n" | ||
| 245 | "fmul v18.4s, v28.4s, v16.s[2]\n" | ||
| 246 | "scvtf v25.4s, v25.4s\n" | ||
| 247 | "fmul v17.4s, v17.4s, v16.s[3]\n" | ||
| 248 | "scvtf v26.4s, v26.4s\n" | ||
| 249 | "fmul v16.4s, v28.4s, v16.s[3]\n" | ||
| 250 | "scvtf v24.4s, v24.4s\n" | ||
| 251 | "fmul v11.4s, v0.4s, v23.4s\n" | ||
| 252 | "fmul v10.4s, v30.4s, v22.4s\n" | ||
| 253 | "fmul v9.4s, v31.4s, v21.4s\n" | ||
| 254 | "fmul v8.4s, v29.4s, v20.4s\n" | ||
| 255 | "fmul v7.4s, v27.4s, v19.4s\n" | ||
| 256 | "fmul v6.4s, v25.4s, v18.4s\n" | ||
| 257 | "fmul v5.4s, v26.4s, v17.4s\n" | ||
| 258 | "fmul v4.4s, v24.4s, v16.4s\n" | ||
| 259 | "ldr q19, [x26, #0x0]\n" | ||
| 260 | "ldr q18, [x26, #0x10]\n" | ||
| 261 | "add x20, %x[clamp_vals], #0x4\n" | ||
| 262 | "cmp x25, #0x8\n" | ||
| 263 | "ld1r { v17.4s }, [%x[clamp_vals]]\n" | ||
| 264 | "ld1r { v16.4s }, [x20]\n" | ||
| 265 | "add x26, x26, #0x20\n" | ||
| 266 | "fadd v11.4s, v11.4s, v19.4s\n" | ||
| 267 | "fadd v10.4s, v10.4s, v18.4s\n" | ||
| 268 | "fadd v9.4s, v9.4s, v19.4s\n" | ||
| 269 | "fadd v8.4s, v8.4s, v18.4s\n" | ||
| 270 | "fadd v7.4s, v7.4s, v19.4s\n" | ||
| 271 | "fadd v6.4s, v6.4s, v18.4s\n" | ||
| 272 | "fadd v5.4s, v5.4s, v19.4s\n" | ||
| 273 | "fadd v4.4s, v4.4s, v18.4s\n" | ||
| 274 | "fmax v11.4s, v11.4s, v17.4s\n" | ||
| 275 | "fmax v10.4s, v10.4s, v17.4s\n" | ||
| 276 | "fmax v9.4s, v9.4s, v17.4s\n" | ||
| 277 | "fmax v8.4s, v8.4s, v17.4s\n" | ||
| 278 | "fmax v7.4s, v7.4s, v17.4s\n" | ||
| 279 | "fmax v6.4s, v6.4s, v17.4s\n" | ||
| 280 | "fmax v5.4s, v5.4s, v17.4s\n" | ||
| 281 | "fmax v4.4s, v4.4s, v17.4s\n" | ||
| 282 | "fmin v11.4s, v11.4s, v16.4s\n" | ||
| 283 | "fmin v10.4s, v10.4s, v16.4s\n" | ||
| 284 | "fmin v9.4s, v9.4s, v16.4s\n" | ||
| 285 | "fmin v8.4s, v8.4s, v16.4s\n" | ||
| 286 | "fmin v7.4s, v7.4s, v16.4s\n" | ||
| 287 | "fmin v6.4s, v6.4s, v16.4s\n" | ||
| 288 | "fmin v5.4s, v5.4s, v16.4s\n" | ||
| 289 | "fmin v4.4s, v4.4s, v16.4s\n" | ||
| 290 | "blt 5f\n" | ||
| 291 | "mov x20, %x[dst]\n" | ||
| 292 | "cmp x27, #0x1\n" | ||
| 293 | "str q11, [x20, #0x0]\n" | ||
| 294 | "str q10, [x20, #0x10]\n" | ||
| 295 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 296 | "ble 10f\n" | ||
| 297 | "cmp x27, #0x2\n" | ||
| 298 | "str q9, [x20, #0x0]\n" | ||
| 299 | "str q8, [x20, #0x10]\n" | ||
| 300 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 301 | "ble 10f\n" | ||
| 302 | "cmp x27, #0x3\n" | ||
| 303 | "str q7, [x20, #0x0]\n" | ||
| 304 | "str q6, [x20, #0x10]\n" | ||
| 305 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 306 | "ble 10f\n" | ||
| 307 | "str q5, [x20, #0x0]\n" | ||
| 308 | "str q4, [x20, #0x10]\n" | ||
| 309 | "b 10f\n" | ||
| 310 | "5:" // Partial output | ||
| 311 | "mov x23, %x[dst]\n" | ||
| 312 | "cmp x27, #0x1\n" | ||
| 313 | "add x22, x23, %x[dst_stride_row]\n" | ||
| 314 | "csel x22, x22, x23, GT\n" | ||
| 315 | "cmp x27, #0x2\n" | ||
| 316 | "add x21, x23, %x[dst_stride_row], LSL #1\n" | ||
| 317 | "csel x21, x21, x22, GT\n" | ||
| 318 | "cmp x27, #0x3\n" | ||
| 319 | "add x20, x21, %x[dst_stride_row]\n" | ||
| 320 | "csel x20, x20, x21, GT\n" | ||
| 321 | "tbz x25, #2, 7f\n" | ||
| 322 | "st1 { v5.4s }, [x20], #0x10\n" | ||
| 323 | "st1 { v7.4s }, [x21], #0x10\n" | ||
| 324 | "st1 { v9.4s }, [x22], #0x10\n" | ||
| 325 | "st1 { v11.4s }, [x23], #0x10\n" | ||
| 326 | "tbz x25, #1, 6f\n" | ||
| 327 | "st1 { v4.d }[0], [x20], #0x8\n" | ||
| 328 | "st1 { v6.d }[0], [x21], #0x8\n" | ||
| 329 | "st1 { v8.d }[0], [x22], #0x8\n" | ||
| 330 | "st1 { v10.d }[0], [x23], #0x8\n" | ||
| 331 | "tbz x25, #0, 9f\n" | ||
| 332 | "st1 { v4.s }[2], [x20]\n" | ||
| 333 | "st1 { v6.s }[2], [x21]\n" | ||
| 334 | "st1 { v8.s }[2], [x22]\n" | ||
| 335 | "st1 { v10.s }[2], [x23]\n" | ||
| 336 | "b 9f\n" | ||
| 337 | "6:" // Output block 0: partial_1_4 | ||
| 338 | "tbz x25, #0, 9f\n" | ||
| 339 | "st1 { v4.s }[0], [x20]\n" | ||
| 340 | "st1 { v6.s }[0], [x21]\n" | ||
| 341 | "st1 { v8.s }[0], [x22]\n" | ||
| 342 | "st1 { v10.s }[0], [x23]\n" | ||
| 343 | "b 9f\n" | ||
| 344 | "7:" // Output block 0: partial_2_0 | ||
| 345 | "tbz x25, #1, 8f\n" | ||
| 346 | "st1 { v5.d }[0], [x20], #0x8\n" | ||
| 347 | "st1 { v7.d }[0], [x21], #0x8\n" | ||
| 348 | "st1 { v9.d }[0], [x22], #0x8\n" | ||
| 349 | "st1 { v11.d }[0], [x23], #0x8\n" | ||
| 350 | "tbz x25, #0, 9f\n" | ||
| 351 | "st1 { v5.s }[2], [x20]\n" | ||
| 352 | "st1 { v7.s }[2], [x21]\n" | ||
| 353 | "st1 { v9.s }[2], [x22]\n" | ||
| 354 | "st1 { v11.s }[2], [x23]\n" | ||
| 355 | "b 9f\n" | ||
| 356 | "8:" // Output block 0: partial_1_0 | ||
| 357 | "st1 { v5.s }[0], [x20]\n" | ||
| 358 | "st1 { v7.s }[0], [x21]\n" | ||
| 359 | "st1 { v9.s }[0], [x22]\n" | ||
| 360 | "st1 { v11.s }[0], [x23]\n" | ||
| 361 | "9:" // Output block 0: Done | ||
| 362 | "10:" // Output stage exit | ||
| 363 | "subs x25, x25, #0x8\n" | ||
| 364 | "add %x[dst], %x[dst], #0x20\n" | ||
| 365 | "bgt 2b\n" | ||
| 366 | "subs x27, x27, #0x4\n" | ||
| 367 | "add %x[lhs_packed], %x[lhs_packed], x28\n" | ||
| 368 | "mov %x[dst], x24\n" | ||
| 369 | "bgt 1b\n" | ||
| 370 | "11:" // Row loop skip | ||
| 371 | : [dst] "+&r"(dst), [lhs_packed] "+&r"(lhs_packed) | ||
| 372 | 161 | : [clamp_vals] "r"(clamp_vals), [dst_stride_row] "r"(dst_stride_row), [m] "r"(m), [n] "r"(n), | |
| 373 | 161 | [num_blocks] "r"(num_blocks), [rhs_packed] "r"(rhs_packed) | |
| 374 | : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v16", "v17", | ||
| 375 | "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", | ||
| 376 | "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"); | ||
| 377 | 161 | } | |
| 378 | #endif // Architectural feature check | ||
| 379 |