| 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_DOTPROD) | ||
| 11 | #error "Dotprod extension required to compile this micro-kernel" | ||
| 12 | #else // Architectural features check. | ||
| 13 | #include "kai_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod.h" | ||
| 14 | |||
| 15 | #include <arm_neon.h> | ||
| 16 | #include <stddef.h> | ||
| 17 | #include <stdint.h> | ||
| 18 | |||
| 19 | #include "kai/kai_common.h" | ||
| 20 | |||
| 21 | static const size_t kai_m_step = 16; | ||
| 22 | static const size_t kai_n_step = 4; | ||
| 23 | static const size_t kai_mr = 4; | ||
| 24 | static const size_t kai_nr = 4; | ||
| 25 | static const size_t kai_kr = 8; | ||
| 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 | // Round up k to be a multiple of 32. | ||
| 35 | 641 | size_t kai_k_multiple_of = 32; | |
| 36 | 1282 | return kai_roundup(k, kai_k_multiple_of); | |
| 37 | 641 | } | |
| 38 | |||
| 39 | 240 | inline static size_t kai_lhs_packed_stride(size_t k) { | |
| 40 | 240 | const size_t k_internal = kai_k_roundedup(k); | |
| 41 | |||
| 42 | − | KAI_ASSERT((k_internal % 2) == 0); | |
| 43 | |||
| 44 | 480 | return kai_mr * (k_internal * sizeof(int8_t) + kai_num_bytes_multiplier_lhs + kai_num_bytes_offset_lhs); | |
| 45 | 240 | } | |
| 46 | |||
| 47 | 240 | inline static size_t kai_rhs_packed_stride(size_t k) { | |
| 48 | 240 | const size_t k_internal = kai_k_roundedup(k); | |
| 49 | |||
| 50 | − | KAI_ASSERT((k_internal % 2) == 0); | |
| 51 | |||
| 52 | 480 | return kai_nr * ((k_internal / 2) + kai_num_bytes_multiplier_rhs + kai_num_bytes_sum_rhs + kai_num_bytes_bias); | |
| 53 | 240 | } | |
| 54 | |||
| 55 | 320 | size_t kai_get_m_step_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod(void) { | |
| 56 | 320 | return kai_m_step; | |
| 57 | } | ||
| 58 | |||
| 59 | 320 | size_t kai_get_n_step_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod(void) { | |
| 60 | 320 | return kai_n_step; | |
| 61 | } | ||
| 62 | |||
| 63 | 240 | size_t kai_get_mr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod(void) { | |
| 64 | 240 | return kai_mr; | |
| 65 | } | ||
| 66 | |||
| 67 | 240 | size_t kai_get_nr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod(void) { | |
| 68 | 240 | return kai_nr; | |
| 69 | } | ||
| 70 | |||
| 71 | 320 | size_t kai_get_kr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod(void) { | |
| 72 | 320 | return kai_kr; | |
| 73 | } | ||
| 74 | |||
| 75 | 320 | size_t kai_get_sr_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod(void) { | |
| 76 | 320 | return kai_sr; | |
| 77 | } | ||
| 78 | |||
| 79 | 240 | size_t kai_get_lhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod(size_t m_idx, size_t k) { | |
| 80 | − | KAI_ASSERT((m_idx % kai_m_step) == 0); | |
| 81 | |||
| 82 | 240 | return (m_idx / kai_mr) * kai_lhs_packed_stride(k); | |
| 83 | } | ||
| 84 | |||
| 85 | 240 | size_t kai_get_rhs_packed_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod(size_t n_idx, size_t k) { | |
| 86 | − | KAI_ASSERT((n_idx % kai_n_step) == 0); | |
| 87 | |||
| 88 | 240 | return (n_idx / kai_nr) * kai_rhs_packed_stride(k); | |
| 89 | } | ||
| 90 | |||
| 91 | 160 | size_t kai_get_dst_offset_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod( | |
| 92 | size_t m_idx, size_t n_idx, size_t dst_stride) { | ||
| 93 | − | KAI_ASSERT((m_idx % kai_m_step) == 0); | |
| 94 | − | KAI_ASSERT((n_idx % kai_n_step) == 0); | |
| 95 | |||
| 96 | 160 | return (n_idx * sizeof(float)) + m_idx * dst_stride; | |
| 97 | } | ||
| 98 | |||
| 99 | 160 | size_t kai_get_dst_size_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod(size_t m, size_t n) { | |
| 100 | 160 | return m * n * sizeof(float); | |
| 101 | } | ||
| 102 | |||
| 103 | 161 | void kai_run_matmul_clamp_f32_qai8dxp4x8_qsi4cxp4x4_16x4x32_neon_dotprod( | |
| 104 | size_t m, size_t n, size_t k, const void* restrict lhs_packed, const void* restrict rhs_packed, | ||
| 105 | float* restrict dst, // NOLINT(readability-non-const-parameter) | ||
| 106 | size_t dst_stride_row, size_t dst_stride_col, float scalar_min, float scalar_max) { | ||
| 107 | − | KAI_ASSERT(dst_stride_col == sizeof(float)); | |
| 108 | |||
| 109 |
1/2✓ Branch 0 taken 161 times.
✗ Branch 1 not taken.
|
161 | if (m == 0) { |
| 110 | ✗ | return; | |
| 111 | } | ||
| 112 | |||
| 113 | 161 | const size_t k_internal = kai_k_roundedup(k); | |
| 114 | |||
| 115 | 161 | size_t num_blocks = k_internal / 32; | |
| 116 | |||
| 117 | 161 | float clamp_vals[2] = {scalar_min, scalar_max}; | |
| 118 | 322 | __asm__ __volatile__( | |
| 119 | "mov x13, %x[m]\n" | ||
| 120 | "mov x12, #0x80\n" | ||
| 121 | "mov x20, #0x20\n" | ||
| 122 | "cmp x13, #0x10\n" | ||
| 123 | "madd x12, %x[num_blocks], x12, x20\n" | ||
| 124 | "blt 14f\n" | ||
| 125 | "1:" // Row loop | ||
| 126 | "mov x11, %x[rhs_packed]\n" | ||
| 127 | "mov x10, %x[n]\n" | ||
| 128 | "add x9, %x[dst], %x[dst_stride_row], LSL #4\n" | ||
| 129 | "2:" // Column loop | ||
| 130 | "mov x27, %x[lhs_packed]\n" | ||
| 131 | "movi v31.4s, #0x0\n" | ||
| 132 | "movi v30.4s, #0x0\n" | ||
| 133 | "mov x23, %x[num_blocks]\n" | ||
| 134 | "movi v29.4s, #0x0\n" | ||
| 135 | "movi v28.4s, #0x0\n" | ||
| 136 | "movi v27.4s, #0x0\n" | ||
| 137 | "movi v26.4s, #0x0\n" | ||
| 138 | "add x22, x27, x12\n" | ||
| 139 | "add x21, x22, x12\n" | ||
| 140 | "add x20, x21, x12\n" | ||
| 141 | "movi v25.4s, #0x0\n" | ||
| 142 | "movi v24.4s, #0x0\n" | ||
| 143 | "movi v23.4s, #0x0\n" | ||
| 144 | "movi v22.4s, #0x0\n" | ||
| 145 | "movi v21.4s, #0x0\n" | ||
| 146 | "movi v20.4s, #0x0\n" | ||
| 147 | "movi v19.4s, #0x0\n" | ||
| 148 | "movi v18.4s, #0x0\n" | ||
| 149 | "movi v17.4s, #0x0\n" | ||
| 150 | "movi v16.4s, #0x0\n" | ||
| 151 | "3:" // Sub block loop | ||
| 152 | "ldr q13, [x11, #0x0]\n" | ||
| 153 | "ldr q14, [x27, #0x0]\n" | ||
| 154 | "movi v10.16b, #0xf0\n" | ||
| 155 | "subs x23, x23, #0x1\n" | ||
| 156 | "ldr q6, [x22, #0x0]\n" | ||
| 157 | "ldr q15, [x21, #0x0]\n" | ||
| 158 | "ldr q3, [x20, #0x0]\n" | ||
| 159 | "ldr q12, [x11, #0x10]\n" | ||
| 160 | "ldr q8, [x27, #0x10]\n" | ||
| 161 | "ldr q4, [x22, #0x10]\n" | ||
| 162 | "shl v9.16b, v13.16b, #0x4\n" | ||
| 163 | "and v13.16b, v13.16b, v10.16b\n" | ||
| 164 | "ldr q0, [x21, #0x10]\n" | ||
| 165 | "ldr q1, [x20, #0x10]\n" | ||
| 166 | "ldr q5, [x11, #0x20]\n" | ||
| 167 | "ldr q2, [x27, #0x20]\n" | ||
| 168 | "shl v7.16b, v12.16b, #0x4\n" | ||
| 169 | "and v12.16b, v12.16b, v10.16b\n" | ||
| 170 | "ldr q11, [x22, #0x20]\n" | ||
| 171 | ".inst 0x4f8ee13f // sdot v31.4s, v9.16b, v14.4b[0]\n" | ||
| 172 | ".inst 0x4faee13e // sdot v30.4s, v9.16b, v14.4b[1]\n" | ||
| 173 | ".inst 0x4f8ee93d // sdot v29.4s, v9.16b, v14.4b[2]\n" | ||
| 174 | ".inst 0x4faee93c // sdot v28.4s, v9.16b, v14.4b[3]\n" | ||
| 175 | "ldr q14, [x21, #0x20]\n" | ||
| 176 | ".inst 0x4f86e13b // sdot v27.4s, v9.16b, v6.4b[0]\n" | ||
| 177 | ".inst 0x4fa6e13a // sdot v26.4s, v9.16b, v6.4b[1]\n" | ||
| 178 | ".inst 0x4f86e939 // sdot v25.4s, v9.16b, v6.4b[2]\n" | ||
| 179 | ".inst 0x4fa6e938 // sdot v24.4s, v9.16b, v6.4b[3]\n" | ||
| 180 | "ldr q6, [x20, #0x20]\n" | ||
| 181 | ".inst 0x4f8fe137 // sdot v23.4s, v9.16b, v15.4b[0]\n" | ||
| 182 | ".inst 0x4fafe136 // sdot v22.4s, v9.16b, v15.4b[1]\n" | ||
| 183 | ".inst 0x4f8fe935 // sdot v21.4s, v9.16b, v15.4b[2]\n" | ||
| 184 | ".inst 0x4fafe934 // sdot v20.4s, v9.16b, v15.4b[3]\n" | ||
| 185 | "ldr q15, [x11, #0x30]\n" | ||
| 186 | "add x11, x11, #0x40\n" | ||
| 187 | ".inst 0x4f83e133 // sdot v19.4s, v9.16b, v3.4b[0]\n" | ||
| 188 | ".inst 0x4fa3e132 // sdot v18.4s, v9.16b, v3.4b[1]\n" | ||
| 189 | ".inst 0x4f83e931 // sdot v17.4s, v9.16b, v3.4b[2]\n" | ||
| 190 | ".inst 0x4fa3e930 // sdot v16.4s, v9.16b, v3.4b[3]\n" | ||
| 191 | "ldr q9, [x27, #0x30]\n" | ||
| 192 | "ldr q3, [x22, #0x30]\n" | ||
| 193 | ".inst 0x4f88e0ff // sdot v31.4s, v7.16b, v8.4b[0]\n" | ||
| 194 | ".inst 0x4fa8e0fe // sdot v30.4s, v7.16b, v8.4b[1]\n" | ||
| 195 | ".inst 0x4f88e8fd // sdot v29.4s, v7.16b, v8.4b[2]\n" | ||
| 196 | ".inst 0x4fa8e8fc // sdot v28.4s, v7.16b, v8.4b[3]\n" | ||
| 197 | "ldr q8, [x21, #0x30]\n" | ||
| 198 | ".inst 0x4f84e0fb // sdot v27.4s, v7.16b, v4.4b[0]\n" | ||
| 199 | ".inst 0x4fa4e0fa // sdot v26.4s, v7.16b, v4.4b[1]\n" | ||
| 200 | ".inst 0x4f84e8f9 // sdot v25.4s, v7.16b, v4.4b[2]\n" | ||
| 201 | ".inst 0x4fa4e8f8 // sdot v24.4s, v7.16b, v4.4b[3]\n" | ||
| 202 | "ldr q4, [x20, #0x30]\n" | ||
| 203 | ".inst 0x4f80e0f7 // sdot v23.4s, v7.16b, v0.4b[0]\n" | ||
| 204 | ".inst 0x4fa0e0f6 // sdot v22.4s, v7.16b, v0.4b[1]\n" | ||
| 205 | ".inst 0x4f80e8f5 // sdot v21.4s, v7.16b, v0.4b[2]\n" | ||
| 206 | ".inst 0x4fa0e8f4 // sdot v20.4s, v7.16b, v0.4b[3]\n" | ||
| 207 | "ldr q0, [x27, #0x40]\n" | ||
| 208 | ".inst 0x4f81e0f3 // sdot v19.4s, v7.16b, v1.4b[0]\n" | ||
| 209 | ".inst 0x4fa1e0f2 // sdot v18.4s, v7.16b, v1.4b[1]\n" | ||
| 210 | ".inst 0x4f81e8f1 // sdot v17.4s, v7.16b, v1.4b[2]\n" | ||
| 211 | ".inst 0x4fa1e8f0 // sdot v16.4s, v7.16b, v1.4b[3]\n" | ||
| 212 | "ldr q1, [x22, #0x40]\n" | ||
| 213 | "shl v7.16b, v5.16b, #0x4\n" | ||
| 214 | "and v5.16b, v5.16b, v10.16b\n" | ||
| 215 | ".inst 0x4f82e0ff // sdot v31.4s, v7.16b, v2.4b[0]\n" | ||
| 216 | ".inst 0x4fa2e0fe // sdot v30.4s, v7.16b, v2.4b[1]\n" | ||
| 217 | ".inst 0x4f82e8fd // sdot v29.4s, v7.16b, v2.4b[2]\n" | ||
| 218 | ".inst 0x4fa2e8fc // sdot v28.4s, v7.16b, v2.4b[3]\n" | ||
| 219 | "ldr q2, [x21, #0x40]\n" | ||
| 220 | ".inst 0x4f8be0fb // sdot v27.4s, v7.16b, v11.4b[0]\n" | ||
| 221 | ".inst 0x4fabe0fa // sdot v26.4s, v7.16b, v11.4b[1]\n" | ||
| 222 | ".inst 0x4f8be8f9 // sdot v25.4s, v7.16b, v11.4b[2]\n" | ||
| 223 | ".inst 0x4fabe8f8 // sdot v24.4s, v7.16b, v11.4b[3]\n" | ||
| 224 | "ldr q11, [x20, #0x40]\n" | ||
| 225 | ".inst 0x4f8ee0f7 // sdot v23.4s, v7.16b, v14.4b[0]\n" | ||
| 226 | ".inst 0x4faee0f6 // sdot v22.4s, v7.16b, v14.4b[1]\n" | ||
| 227 | ".inst 0x4f8ee8f5 // sdot v21.4s, v7.16b, v14.4b[2]\n" | ||
| 228 | ".inst 0x4faee8f4 // sdot v20.4s, v7.16b, v14.4b[3]\n" | ||
| 229 | "ldr q14, [x27, #0x50]\n" | ||
| 230 | ".inst 0x4f86e0f3 // sdot v19.4s, v7.16b, v6.4b[0]\n" | ||
| 231 | ".inst 0x4fa6e0f2 // sdot v18.4s, v7.16b, v6.4b[1]\n" | ||
| 232 | ".inst 0x4f86e8f1 // sdot v17.4s, v7.16b, v6.4b[2]\n" | ||
| 233 | ".inst 0x4fa6e8f0 // sdot v16.4s, v7.16b, v6.4b[3]\n" | ||
| 234 | "ldr q6, [x22, #0x50]\n" | ||
| 235 | "shl v7.16b, v15.16b, #0x4\n" | ||
| 236 | "and v15.16b, v15.16b, v10.16b\n" | ||
| 237 | "ldr q10, [x21, #0x50]\n" | ||
| 238 | ".inst 0x4f89e0ff // sdot v31.4s, v7.16b, v9.4b[0]\n" | ||
| 239 | ".inst 0x4fa9e0fe // sdot v30.4s, v7.16b, v9.4b[1]\n" | ||
| 240 | ".inst 0x4f89e8fd // sdot v29.4s, v7.16b, v9.4b[2]\n" | ||
| 241 | ".inst 0x4fa9e8fc // sdot v28.4s, v7.16b, v9.4b[3]\n" | ||
| 242 | "ldr q9, [x20, #0x50]\n" | ||
| 243 | ".inst 0x4f83e0fb // sdot v27.4s, v7.16b, v3.4b[0]\n" | ||
| 244 | ".inst 0x4fa3e0fa // sdot v26.4s, v7.16b, v3.4b[1]\n" | ||
| 245 | ".inst 0x4f83e8f9 // sdot v25.4s, v7.16b, v3.4b[2]\n" | ||
| 246 | ".inst 0x4fa3e8f8 // sdot v24.4s, v7.16b, v3.4b[3]\n" | ||
| 247 | "ldr q3, [x27, #0x60]\n" | ||
| 248 | ".inst 0x4f88e0f7 // sdot v23.4s, v7.16b, v8.4b[0]\n" | ||
| 249 | ".inst 0x4fa8e0f6 // sdot v22.4s, v7.16b, v8.4b[1]\n" | ||
| 250 | ".inst 0x4f88e8f5 // sdot v21.4s, v7.16b, v8.4b[2]\n" | ||
| 251 | ".inst 0x4fa8e8f4 // sdot v20.4s, v7.16b, v8.4b[3]\n" | ||
| 252 | "ldr q8, [x22, #0x60]\n" | ||
| 253 | ".inst 0x4f84e0f3 // sdot v19.4s, v7.16b, v4.4b[0]\n" | ||
| 254 | ".inst 0x4fa4e0f2 // sdot v18.4s, v7.16b, v4.4b[1]\n" | ||
| 255 | ".inst 0x4f84e8f1 // sdot v17.4s, v7.16b, v4.4b[2]\n" | ||
| 256 | ".inst 0x4fa4e8f0 // sdot v16.4s, v7.16b, v4.4b[3]\n" | ||
| 257 | "ldr q7, [x21, #0x60]\n" | ||
| 258 | "ldr q4, [x20, #0x60]\n" | ||
| 259 | ".inst 0x4f80e1bf // sdot v31.4s, v13.16b, v0.4b[0]\n" | ||
| 260 | ".inst 0x4fa0e1be // sdot v30.4s, v13.16b, v0.4b[1]\n" | ||
| 261 | ".inst 0x4f80e9bd // sdot v29.4s, v13.16b, v0.4b[2]\n" | ||
| 262 | ".inst 0x4fa0e9bc // sdot v28.4s, v13.16b, v0.4b[3]\n" | ||
| 263 | "ldr q0, [x27, #0x70]\n" | ||
| 264 | "add x27, x27, #0x80\n" | ||
| 265 | ".inst 0x4f81e1bb // sdot v27.4s, v13.16b, v1.4b[0]\n" | ||
| 266 | ".inst 0x4fa1e1ba // sdot v26.4s, v13.16b, v1.4b[1]\n" | ||
| 267 | ".inst 0x4f81e9b9 // sdot v25.4s, v13.16b, v1.4b[2]\n" | ||
| 268 | ".inst 0x4fa1e9b8 // sdot v24.4s, v13.16b, v1.4b[3]\n" | ||
| 269 | "ldr q1, [x22, #0x70]\n" | ||
| 270 | "add x22, x22, #0x80\n" | ||
| 271 | ".inst 0x4f82e1b7 // sdot v23.4s, v13.16b, v2.4b[0]\n" | ||
| 272 | ".inst 0x4fa2e1b6 // sdot v22.4s, v13.16b, v2.4b[1]\n" | ||
| 273 | ".inst 0x4f82e9b5 // sdot v21.4s, v13.16b, v2.4b[2]\n" | ||
| 274 | ".inst 0x4fa2e9b4 // sdot v20.4s, v13.16b, v2.4b[3]\n" | ||
| 275 | "ldr q2, [x21, #0x70]\n" | ||
| 276 | "add x21, x21, #0x80\n" | ||
| 277 | ".inst 0x4f8be1b3 // sdot v19.4s, v13.16b, v11.4b[0]\n" | ||
| 278 | ".inst 0x4fabe1b2 // sdot v18.4s, v13.16b, v11.4b[1]\n" | ||
| 279 | ".inst 0x4f8be9b1 // sdot v17.4s, v13.16b, v11.4b[2]\n" | ||
| 280 | ".inst 0x4fabe9b0 // sdot v16.4s, v13.16b, v11.4b[3]\n" | ||
| 281 | "ldr q11, [x20, #0x70]\n" | ||
| 282 | "add x20, x20, #0x80\n" | ||
| 283 | ".inst 0x4f8ee19f // sdot v31.4s, v12.16b, v14.4b[0]\n" | ||
| 284 | ".inst 0x4faee19e // sdot v30.4s, v12.16b, v14.4b[1]\n" | ||
| 285 | ".inst 0x4f8ee99d // sdot v29.4s, v12.16b, v14.4b[2]\n" | ||
| 286 | ".inst 0x4faee99c // sdot v28.4s, v12.16b, v14.4b[3]\n" | ||
| 287 | ".inst 0x4f86e19b // sdot v27.4s, v12.16b, v6.4b[0]\n" | ||
| 288 | ".inst 0x4fa6e19a // sdot v26.4s, v12.16b, v6.4b[1]\n" | ||
| 289 | ".inst 0x4f86e999 // sdot v25.4s, v12.16b, v6.4b[2]\n" | ||
| 290 | ".inst 0x4fa6e998 // sdot v24.4s, v12.16b, v6.4b[3]\n" | ||
| 291 | ".inst 0x4f8ae197 // sdot v23.4s, v12.16b, v10.4b[0]\n" | ||
| 292 | ".inst 0x4faae196 // sdot v22.4s, v12.16b, v10.4b[1]\n" | ||
| 293 | ".inst 0x4f8ae995 // sdot v21.4s, v12.16b, v10.4b[2]\n" | ||
| 294 | ".inst 0x4faae994 // sdot v20.4s, v12.16b, v10.4b[3]\n" | ||
| 295 | ".inst 0x4f89e193 // sdot v19.4s, v12.16b, v9.4b[0]\n" | ||
| 296 | ".inst 0x4fa9e192 // sdot v18.4s, v12.16b, v9.4b[1]\n" | ||
| 297 | ".inst 0x4f89e991 // sdot v17.4s, v12.16b, v9.4b[2]\n" | ||
| 298 | ".inst 0x4fa9e990 // sdot v16.4s, v12.16b, v9.4b[3]\n" | ||
| 299 | ".inst 0x4f83e0bf // sdot v31.4s, v5.16b, v3.4b[0]\n" | ||
| 300 | ".inst 0x4fa3e0be // sdot v30.4s, v5.16b, v3.4b[1]\n" | ||
| 301 | ".inst 0x4f83e8bd // sdot v29.4s, v5.16b, v3.4b[2]\n" | ||
| 302 | ".inst 0x4fa3e8bc // sdot v28.4s, v5.16b, v3.4b[3]\n" | ||
| 303 | ".inst 0x4f88e0bb // sdot v27.4s, v5.16b, v8.4b[0]\n" | ||
| 304 | ".inst 0x4fa8e0ba // sdot v26.4s, v5.16b, v8.4b[1]\n" | ||
| 305 | ".inst 0x4f88e8b9 // sdot v25.4s, v5.16b, v8.4b[2]\n" | ||
| 306 | ".inst 0x4fa8e8b8 // sdot v24.4s, v5.16b, v8.4b[3]\n" | ||
| 307 | ".inst 0x4f87e0b7 // sdot v23.4s, v5.16b, v7.4b[0]\n" | ||
| 308 | ".inst 0x4fa7e0b6 // sdot v22.4s, v5.16b, v7.4b[1]\n" | ||
| 309 | ".inst 0x4f87e8b5 // sdot v21.4s, v5.16b, v7.4b[2]\n" | ||
| 310 | ".inst 0x4fa7e8b4 // sdot v20.4s, v5.16b, v7.4b[3]\n" | ||
| 311 | ".inst 0x4f84e0b3 // sdot v19.4s, v5.16b, v4.4b[0]\n" | ||
| 312 | ".inst 0x4fa4e0b2 // sdot v18.4s, v5.16b, v4.4b[1]\n" | ||
| 313 | ".inst 0x4f84e8b1 // sdot v17.4s, v5.16b, v4.4b[2]\n" | ||
| 314 | ".inst 0x4fa4e8b0 // sdot v16.4s, v5.16b, v4.4b[3]\n" | ||
| 315 | ".inst 0x4f80e1ff // sdot v31.4s, v15.16b, v0.4b[0]\n" | ||
| 316 | ".inst 0x4fa0e1fe // sdot v30.4s, v15.16b, v0.4b[1]\n" | ||
| 317 | ".inst 0x4f80e9fd // sdot v29.4s, v15.16b, v0.4b[2]\n" | ||
| 318 | ".inst 0x4fa0e9fc // sdot v28.4s, v15.16b, v0.4b[3]\n" | ||
| 319 | ".inst 0x4f81e1fb // sdot v27.4s, v15.16b, v1.4b[0]\n" | ||
| 320 | ".inst 0x4fa1e1fa // sdot v26.4s, v15.16b, v1.4b[1]\n" | ||
| 321 | ".inst 0x4f81e9f9 // sdot v25.4s, v15.16b, v1.4b[2]\n" | ||
| 322 | ".inst 0x4fa1e9f8 // sdot v24.4s, v15.16b, v1.4b[3]\n" | ||
| 323 | ".inst 0x4f82e1f7 // sdot v23.4s, v15.16b, v2.4b[0]\n" | ||
| 324 | ".inst 0x4fa2e1f6 // sdot v22.4s, v15.16b, v2.4b[1]\n" | ||
| 325 | ".inst 0x4f82e9f5 // sdot v21.4s, v15.16b, v2.4b[2]\n" | ||
| 326 | ".inst 0x4fa2e9f4 // sdot v20.4s, v15.16b, v2.4b[3]\n" | ||
| 327 | ".inst 0x4f8be1f3 // sdot v19.4s, v15.16b, v11.4b[0]\n" | ||
| 328 | ".inst 0x4fabe1f2 // sdot v18.4s, v15.16b, v11.4b[1]\n" | ||
| 329 | ".inst 0x4f8be9f1 // sdot v17.4s, v15.16b, v11.4b[2]\n" | ||
| 330 | ".inst 0x4fabe9f0 // sdot v16.4s, v15.16b, v11.4b[3]\n" | ||
| 331 | "bgt 3b\n" | ||
| 332 | "ldr q5, [x11, #0x0]\n" | ||
| 333 | "ld1 { v1.4s }, [x27]\n" | ||
| 334 | "add x27, x27, #0x10\n" | ||
| 335 | "ldr q4, [x11, #0x10]\n" | ||
| 336 | "ldr q0, [x27, #0x0]\n" | ||
| 337 | "add x11, x11, #0x20\n" | ||
| 338 | "mla v31.4s, v5.4s, v1.s[0]\n" | ||
| 339 | "mla v30.4s, v5.4s, v1.s[1]\n" | ||
| 340 | "mla v29.4s, v5.4s, v1.s[2]\n" | ||
| 341 | "mla v28.4s, v5.4s, v1.s[3]\n" | ||
| 342 | "fmul v3.4s, v4.4s, v0.s[0]\n" | ||
| 343 | "fmul v2.4s, v4.4s, v0.s[1]\n" | ||
| 344 | "fmul v1.4s, v4.4s, v0.s[2]\n" | ||
| 345 | "scvtf v31.4s, v31.4s\n" | ||
| 346 | "fmul v0.4s, v4.4s, v0.s[3]\n" | ||
| 347 | "scvtf v30.4s, v30.4s\n" | ||
| 348 | "scvtf v29.4s, v29.4s\n" | ||
| 349 | "scvtf v28.4s, v28.4s\n" | ||
| 350 | "fmul v31.4s, v31.4s, v3.4s\n" | ||
| 351 | "fmul v30.4s, v30.4s, v2.4s\n" | ||
| 352 | "fmul v29.4s, v29.4s, v1.4s\n" | ||
| 353 | "fmul v28.4s, v28.4s, v0.4s\n" | ||
| 354 | "ld1 { v1.4s }, [x22]\n" | ||
| 355 | "add x22, x22, #0x10\n" | ||
| 356 | "ldr q0, [x22, #0x0]\n" | ||
| 357 | "mla v27.4s, v5.4s, v1.s[0]\n" | ||
| 358 | "mla v26.4s, v5.4s, v1.s[1]\n" | ||
| 359 | "mla v25.4s, v5.4s, v1.s[2]\n" | ||
| 360 | "mla v24.4s, v5.4s, v1.s[3]\n" | ||
| 361 | "fmul v3.4s, v4.4s, v0.s[0]\n" | ||
| 362 | "fmul v2.4s, v4.4s, v0.s[1]\n" | ||
| 363 | "fmul v1.4s, v4.4s, v0.s[2]\n" | ||
| 364 | "scvtf v27.4s, v27.4s\n" | ||
| 365 | "fmul v0.4s, v4.4s, v0.s[3]\n" | ||
| 366 | "scvtf v26.4s, v26.4s\n" | ||
| 367 | "scvtf v25.4s, v25.4s\n" | ||
| 368 | "scvtf v24.4s, v24.4s\n" | ||
| 369 | "fmul v27.4s, v27.4s, v3.4s\n" | ||
| 370 | "fmul v26.4s, v26.4s, v2.4s\n" | ||
| 371 | "fmul v25.4s, v25.4s, v1.4s\n" | ||
| 372 | "fmul v24.4s, v24.4s, v0.4s\n" | ||
| 373 | "ld1 { v1.4s }, [x21]\n" | ||
| 374 | "add x21, x21, #0x10\n" | ||
| 375 | "ldr q0, [x21, #0x0]\n" | ||
| 376 | "mla v23.4s, v5.4s, v1.s[0]\n" | ||
| 377 | "mla v22.4s, v5.4s, v1.s[1]\n" | ||
| 378 | "mla v21.4s, v5.4s, v1.s[2]\n" | ||
| 379 | "mla v20.4s, v5.4s, v1.s[3]\n" | ||
| 380 | "fmul v3.4s, v4.4s, v0.s[0]\n" | ||
| 381 | "fmul v2.4s, v4.4s, v0.s[1]\n" | ||
| 382 | "fmul v1.4s, v4.4s, v0.s[2]\n" | ||
| 383 | "scvtf v23.4s, v23.4s\n" | ||
| 384 | "fmul v0.4s, v4.4s, v0.s[3]\n" | ||
| 385 | "scvtf v22.4s, v22.4s\n" | ||
| 386 | "scvtf v21.4s, v21.4s\n" | ||
| 387 | "scvtf v20.4s, v20.4s\n" | ||
| 388 | "fmul v23.4s, v23.4s, v3.4s\n" | ||
| 389 | "fmul v22.4s, v22.4s, v2.4s\n" | ||
| 390 | "fmul v21.4s, v21.4s, v1.4s\n" | ||
| 391 | "fmul v20.4s, v20.4s, v0.4s\n" | ||
| 392 | "ld1 { v1.4s }, [x20]\n" | ||
| 393 | "add x20, x20, #0x10\n" | ||
| 394 | "ldr q0, [x20, #0x0]\n" | ||
| 395 | "mla v19.4s, v5.4s, v1.s[0]\n" | ||
| 396 | "mla v18.4s, v5.4s, v1.s[1]\n" | ||
| 397 | "mla v17.4s, v5.4s, v1.s[2]\n" | ||
| 398 | "mla v16.4s, v5.4s, v1.s[3]\n" | ||
| 399 | "fmul v3.4s, v4.4s, v0.s[0]\n" | ||
| 400 | "fmul v2.4s, v4.4s, v0.s[1]\n" | ||
| 401 | "fmul v1.4s, v4.4s, v0.s[2]\n" | ||
| 402 | "scvtf v19.4s, v19.4s\n" | ||
| 403 | "fmul v0.4s, v4.4s, v0.s[3]\n" | ||
| 404 | "scvtf v18.4s, v18.4s\n" | ||
| 405 | "scvtf v17.4s, v17.4s\n" | ||
| 406 | "scvtf v16.4s, v16.4s\n" | ||
| 407 | "fmul v19.4s, v19.4s, v3.4s\n" | ||
| 408 | "fmul v18.4s, v18.4s, v2.4s\n" | ||
| 409 | "fmul v17.4s, v17.4s, v1.4s\n" | ||
| 410 | "fmul v16.4s, v16.4s, v0.4s\n" | ||
| 411 | "ldr q2, [x11, #0x0]\n" | ||
| 412 | "ld1r { v1.4s }, [%x[clamp_vals]]\n" | ||
| 413 | "add x20, %x[clamp_vals], #0x4\n" | ||
| 414 | "cmp x10, #0x4\n" | ||
| 415 | "ld1r { v0.4s }, [x20]\n" | ||
| 416 | "add x11, x11, #0x10\n" | ||
| 417 | "fadd v31.4s, v31.4s, v2.4s\n" | ||
| 418 | "fadd v30.4s, v30.4s, v2.4s\n" | ||
| 419 | "fadd v29.4s, v29.4s, v2.4s\n" | ||
| 420 | "fadd v28.4s, v28.4s, v2.4s\n" | ||
| 421 | "fadd v27.4s, v27.4s, v2.4s\n" | ||
| 422 | "fadd v26.4s, v26.4s, v2.4s\n" | ||
| 423 | "fadd v25.4s, v25.4s, v2.4s\n" | ||
| 424 | "fadd v24.4s, v24.4s, v2.4s\n" | ||
| 425 | "fadd v23.4s, v23.4s, v2.4s\n" | ||
| 426 | "fadd v22.4s, v22.4s, v2.4s\n" | ||
| 427 | "fadd v21.4s, v21.4s, v2.4s\n" | ||
| 428 | "fadd v20.4s, v20.4s, v2.4s\n" | ||
| 429 | "fadd v19.4s, v19.4s, v2.4s\n" | ||
| 430 | "fadd v18.4s, v18.4s, v2.4s\n" | ||
| 431 | "fadd v17.4s, v17.4s, v2.4s\n" | ||
| 432 | "fadd v16.4s, v16.4s, v2.4s\n" | ||
| 433 | "fmax v31.4s, v31.4s, v1.4s\n" | ||
| 434 | "fmax v30.4s, v30.4s, v1.4s\n" | ||
| 435 | "fmax v29.4s, v29.4s, v1.4s\n" | ||
| 436 | "fmax v28.4s, v28.4s, v1.4s\n" | ||
| 437 | "fmax v27.4s, v27.4s, v1.4s\n" | ||
| 438 | "fmax v26.4s, v26.4s, v1.4s\n" | ||
| 439 | "fmax v25.4s, v25.4s, v1.4s\n" | ||
| 440 | "fmax v24.4s, v24.4s, v1.4s\n" | ||
| 441 | "fmax v23.4s, v23.4s, v1.4s\n" | ||
| 442 | "fmax v22.4s, v22.4s, v1.4s\n" | ||
| 443 | "fmax v21.4s, v21.4s, v1.4s\n" | ||
| 444 | "fmax v20.4s, v20.4s, v1.4s\n" | ||
| 445 | "fmax v19.4s, v19.4s, v1.4s\n" | ||
| 446 | "fmax v18.4s, v18.4s, v1.4s\n" | ||
| 447 | "fmax v17.4s, v17.4s, v1.4s\n" | ||
| 448 | "fmax v16.4s, v16.4s, v1.4s\n" | ||
| 449 | "fmin v31.4s, v31.4s, v0.4s\n" | ||
| 450 | "fmin v30.4s, v30.4s, v0.4s\n" | ||
| 451 | "fmin v29.4s, v29.4s, v0.4s\n" | ||
| 452 | "fmin v28.4s, v28.4s, v0.4s\n" | ||
| 453 | "fmin v27.4s, v27.4s, v0.4s\n" | ||
| 454 | "fmin v26.4s, v26.4s, v0.4s\n" | ||
| 455 | "fmin v25.4s, v25.4s, v0.4s\n" | ||
| 456 | "fmin v24.4s, v24.4s, v0.4s\n" | ||
| 457 | "fmin v23.4s, v23.4s, v0.4s\n" | ||
| 458 | "fmin v22.4s, v22.4s, v0.4s\n" | ||
| 459 | "fmin v21.4s, v21.4s, v0.4s\n" | ||
| 460 | "fmin v20.4s, v20.4s, v0.4s\n" | ||
| 461 | "fmin v19.4s, v19.4s, v0.4s\n" | ||
| 462 | "fmin v18.4s, v18.4s, v0.4s\n" | ||
| 463 | "fmin v17.4s, v17.4s, v0.4s\n" | ||
| 464 | "fmin v16.4s, v16.4s, v0.4s\n" | ||
| 465 | "blt 8f\n" | ||
| 466 | "mov x20, %x[dst]\n" | ||
| 467 | "str q31, [x20, #0x0]\n" | ||
| 468 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 469 | "str q30, [x20, #0x0]\n" | ||
| 470 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 471 | "str q29, [x20, #0x0]\n" | ||
| 472 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 473 | "str q28, [x20, #0x0]\n" | ||
| 474 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 475 | "str q27, [x20, #0x0]\n" | ||
| 476 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 477 | "str q26, [x20, #0x0]\n" | ||
| 478 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 479 | "str q25, [x20, #0x0]\n" | ||
| 480 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 481 | "str q24, [x20, #0x0]\n" | ||
| 482 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 483 | "str q23, [x20, #0x0]\n" | ||
| 484 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 485 | "str q22, [x20, #0x0]\n" | ||
| 486 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 487 | "str q21, [x20, #0x0]\n" | ||
| 488 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 489 | "str q20, [x20, #0x0]\n" | ||
| 490 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 491 | "str q19, [x20, #0x0]\n" | ||
| 492 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 493 | "str q18, [x20, #0x0]\n" | ||
| 494 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 495 | "str q17, [x20, #0x0]\n" | ||
| 496 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 497 | "str q16, [x20, #0x0]\n" | ||
| 498 | "b 13f\n" | ||
| 499 | "8:" // Partial output | ||
| 500 | "mov x28, %x[dst]\n" | ||
| 501 | "add x26, x28, %x[dst_stride_row], LSL #2\n" | ||
| 502 | "add x25, x26, %x[dst_stride_row], LSL #1\n" | ||
| 503 | "add x24, x26, %x[dst_stride_row]\n" | ||
| 504 | "add x23, x25, %x[dst_stride_row]\n" | ||
| 505 | "add x22, x28, %x[dst_stride_row], LSL #1\n" | ||
| 506 | "add x21, x28, %x[dst_stride_row]\n" | ||
| 507 | "add x20, x22, %x[dst_stride_row]\n" | ||
| 508 | "add x27, x23, %x[dst_stride_row]\n" | ||
| 509 | "tbz x10, #1, 9f\n" | ||
| 510 | "st1 { v24.d }[0], [x23], #0x8\n" | ||
| 511 | "st1 { v25.d }[0], [x25], #0x8\n" | ||
| 512 | "st1 { v26.d }[0], [x24], #0x8\n" | ||
| 513 | "st1 { v27.d }[0], [x26], #0x8\n" | ||
| 514 | "st1 { v28.d }[0], [x20], #0x8\n" | ||
| 515 | "st1 { v29.d }[0], [x22], #0x8\n" | ||
| 516 | "st1 { v30.d }[0], [x21], #0x8\n" | ||
| 517 | "st1 { v31.d }[0], [x28], #0x8\n" | ||
| 518 | "tbz x10, #0, 10f\n" | ||
| 519 | "st1 { v24.s }[2], [x23]\n" | ||
| 520 | "st1 { v25.s }[2], [x25]\n" | ||
| 521 | "st1 { v26.s }[2], [x24]\n" | ||
| 522 | "st1 { v27.s }[2], [x26]\n" | ||
| 523 | "st1 { v28.s }[2], [x20]\n" | ||
| 524 | "st1 { v29.s }[2], [x22]\n" | ||
| 525 | "st1 { v30.s }[2], [x21]\n" | ||
| 526 | "st1 { v31.s }[2], [x28]\n" | ||
| 527 | "b 10f\n" | ||
| 528 | "9:" // Output block 0: partial_1_0 | ||
| 529 | "st1 { v24.s }[0], [x23]\n" | ||
| 530 | "st1 { v25.s }[0], [x25]\n" | ||
| 531 | "st1 { v26.s }[0], [x24]\n" | ||
| 532 | "st1 { v27.s }[0], [x26]\n" | ||
| 533 | "st1 { v28.s }[0], [x20]\n" | ||
| 534 | "st1 { v29.s }[0], [x22]\n" | ||
| 535 | "st1 { v30.s }[0], [x21]\n" | ||
| 536 | "st1 { v31.s }[0], [x28]\n" | ||
| 537 | "10:" // Output block 0: Done | ||
| 538 | "add x26, x27, %x[dst_stride_row], LSL #2\n" | ||
| 539 | "add x25, x27, %x[dst_stride_row], LSL #1\n" | ||
| 540 | "add x24, x26, %x[dst_stride_row], LSL #1\n" | ||
| 541 | "add x23, x27, %x[dst_stride_row]\n" | ||
| 542 | "add x22, x25, %x[dst_stride_row]\n" | ||
| 543 | "add x21, x26, %x[dst_stride_row]\n" | ||
| 544 | "add x20, x24, %x[dst_stride_row]\n" | ||
| 545 | "tbz x10, #1, 11f\n" | ||
| 546 | "st1 { v16.d }[0], [x20], #0x8\n" | ||
| 547 | "st1 { v17.d }[0], [x24], #0x8\n" | ||
| 548 | "st1 { v18.d }[0], [x21], #0x8\n" | ||
| 549 | "st1 { v19.d }[0], [x26], #0x8\n" | ||
| 550 | "st1 { v20.d }[0], [x22], #0x8\n" | ||
| 551 | "st1 { v21.d }[0], [x25], #0x8\n" | ||
| 552 | "st1 { v22.d }[0], [x23], #0x8\n" | ||
| 553 | "st1 { v23.d }[0], [x27], #0x8\n" | ||
| 554 | "tbz x10, #0, 12f\n" | ||
| 555 | "st1 { v16.s }[2], [x20]\n" | ||
| 556 | "st1 { v17.s }[2], [x24]\n" | ||
| 557 | "st1 { v18.s }[2], [x21]\n" | ||
| 558 | "st1 { v19.s }[2], [x26]\n" | ||
| 559 | "st1 { v20.s }[2], [x22]\n" | ||
| 560 | "st1 { v21.s }[2], [x25]\n" | ||
| 561 | "st1 { v22.s }[2], [x23]\n" | ||
| 562 | "st1 { v23.s }[2], [x27]\n" | ||
| 563 | "b 12f\n" | ||
| 564 | "11:" // Output block 1: partial_1_0 | ||
| 565 | "st1 { v16.s }[0], [x20]\n" | ||
| 566 | "st1 { v17.s }[0], [x24]\n" | ||
| 567 | "st1 { v18.s }[0], [x21]\n" | ||
| 568 | "st1 { v19.s }[0], [x26]\n" | ||
| 569 | "st1 { v20.s }[0], [x22]\n" | ||
| 570 | "st1 { v21.s }[0], [x25]\n" | ||
| 571 | "st1 { v22.s }[0], [x23]\n" | ||
| 572 | "st1 { v23.s }[0], [x27]\n" | ||
| 573 | "12:" // Output block 1: Done | ||
| 574 | "13:" // Output stage exit | ||
| 575 | "subs x10, x10, #0x4\n" | ||
| 576 | "add %x[dst], %x[dst], #0x10\n" | ||
| 577 | "bgt 2b\n" | ||
| 578 | "mov x20, #0x4\n" | ||
| 579 | "sub x13, x13, #0x10\n" | ||
| 580 | "cmp x13, #0x10\n" | ||
| 581 | "mov %x[dst], x9\n" | ||
| 582 | "madd %x[lhs_packed], x20, x12, %x[lhs_packed]\n" | ||
| 583 | "bge 1b\n" | ||
| 584 | "14:" // Row loop skip | ||
| 585 | "cbz x13, 23f\n" | ||
| 586 | "15:" // Row tail: Row loop | ||
| 587 | "mov x26, %x[rhs_packed]\n" | ||
| 588 | "mov x25, %x[n]\n" | ||
| 589 | "add x24, %x[dst], %x[dst_stride_row], LSL #2\n" | ||
| 590 | "16:" // Row tail: Column loop | ||
| 591 | "mov x27, %x[lhs_packed]\n" | ||
| 592 | "movi v31.4s, #0x0\n" | ||
| 593 | "movi v30.4s, #0x0\n" | ||
| 594 | "mov x20, %x[num_blocks]\n" | ||
| 595 | "movi v29.4s, #0x0\n" | ||
| 596 | "movi v28.4s, #0x0\n" | ||
| 597 | "17:" // Row tail: Sub block loop | ||
| 598 | "ldr q4, [x26, #0x0]\n" | ||
| 599 | "ldr q3, [x27, #0x0]\n" | ||
| 600 | "movi v2.16b, #0xf0\n" | ||
| 601 | "subs x20, x20, #0x1\n" | ||
| 602 | "ldr q1, [x26, #0x10]\n" | ||
| 603 | "ldr q0, [x27, #0x10]\n" | ||
| 604 | "ldr q27, [x26, #0x20]\n" | ||
| 605 | "ldr q26, [x27, #0x20]\n" | ||
| 606 | "ldr q25, [x26, #0x30]\n" | ||
| 607 | "ldr q24, [x27, #0x30]\n" | ||
| 608 | "shl v23.16b, v4.16b, #0x4\n" | ||
| 609 | "and v4.16b, v4.16b, v2.16b\n" | ||
| 610 | "ldr q22, [x27, #0x40]\n" | ||
| 611 | "ldr q21, [x27, #0x50]\n" | ||
| 612 | "shl v20.16b, v1.16b, #0x4\n" | ||
| 613 | "and v1.16b, v1.16b, v2.16b\n" | ||
| 614 | "ldr q19, [x27, #0x60]\n" | ||
| 615 | "ldr q18, [x27, #0x70]\n" | ||
| 616 | "shl v17.16b, v27.16b, #0x4\n" | ||
| 617 | "and v27.16b, v27.16b, v2.16b\n" | ||
| 618 | ".inst 0x4f83e2ff // sdot v31.4s, v23.16b, v3.4b[0]\n" | ||
| 619 | ".inst 0x4fa3e2fe // sdot v30.4s, v23.16b, v3.4b[1]\n" | ||
| 620 | "shl v16.16b, v25.16b, #0x4\n" | ||
| 621 | "add x26, x26, #0x40\n" | ||
| 622 | ".inst 0x4f83eafd // sdot v29.4s, v23.16b, v3.4b[2]\n" | ||
| 623 | ".inst 0x4fa3eafc // sdot v28.4s, v23.16b, v3.4b[3]\n" | ||
| 624 | "and v25.16b, v25.16b, v2.16b\n" | ||
| 625 | "add x27, x27, #0x80\n" | ||
| 626 | ".inst 0x4f80e29f // sdot v31.4s, v20.16b, v0.4b[0]\n" | ||
| 627 | ".inst 0x4fa0e29e // sdot v30.4s, v20.16b, v0.4b[1]\n" | ||
| 628 | ".inst 0x4f80ea9d // sdot v29.4s, v20.16b, v0.4b[2]\n" | ||
| 629 | ".inst 0x4fa0ea9c // sdot v28.4s, v20.16b, v0.4b[3]\n" | ||
| 630 | ".inst 0x4f9ae23f // sdot v31.4s, v17.16b, v26.4b[0]\n" | ||
| 631 | ".inst 0x4fbae23e // sdot v30.4s, v17.16b, v26.4b[1]\n" | ||
| 632 | ".inst 0x4f9aea3d // sdot v29.4s, v17.16b, v26.4b[2]\n" | ||
| 633 | ".inst 0x4fbaea3c // sdot v28.4s, v17.16b, v26.4b[3]\n" | ||
| 634 | ".inst 0x4f98e21f // sdot v31.4s, v16.16b, v24.4b[0]\n" | ||
| 635 | ".inst 0x4fb8e21e // sdot v30.4s, v16.16b, v24.4b[1]\n" | ||
| 636 | ".inst 0x4f98ea1d // sdot v29.4s, v16.16b, v24.4b[2]\n" | ||
| 637 | ".inst 0x4fb8ea1c // sdot v28.4s, v16.16b, v24.4b[3]\n" | ||
| 638 | ".inst 0x4f96e09f // sdot v31.4s, v4.16b, v22.4b[0]\n" | ||
| 639 | ".inst 0x4fb6e09e // sdot v30.4s, v4.16b, v22.4b[1]\n" | ||
| 640 | ".inst 0x4f96e89d // sdot v29.4s, v4.16b, v22.4b[2]\n" | ||
| 641 | ".inst 0x4fb6e89c // sdot v28.4s, v4.16b, v22.4b[3]\n" | ||
| 642 | ".inst 0x4f95e03f // sdot v31.4s, v1.16b, v21.4b[0]\n" | ||
| 643 | ".inst 0x4fb5e03e // sdot v30.4s, v1.16b, v21.4b[1]\n" | ||
| 644 | ".inst 0x4f95e83d // sdot v29.4s, v1.16b, v21.4b[2]\n" | ||
| 645 | ".inst 0x4fb5e83c // sdot v28.4s, v1.16b, v21.4b[3]\n" | ||
| 646 | ".inst 0x4f93e37f // sdot v31.4s, v27.16b, v19.4b[0]\n" | ||
| 647 | ".inst 0x4fb3e37e // sdot v30.4s, v27.16b, v19.4b[1]\n" | ||
| 648 | ".inst 0x4f93eb7d // sdot v29.4s, v27.16b, v19.4b[2]\n" | ||
| 649 | ".inst 0x4fb3eb7c // sdot v28.4s, v27.16b, v19.4b[3]\n" | ||
| 650 | ".inst 0x4f92e33f // sdot v31.4s, v25.16b, v18.4b[0]\n" | ||
| 651 | ".inst 0x4fb2e33e // sdot v30.4s, v25.16b, v18.4b[1]\n" | ||
| 652 | ".inst 0x4f92eb3d // sdot v29.4s, v25.16b, v18.4b[2]\n" | ||
| 653 | ".inst 0x4fb2eb3c // sdot v28.4s, v25.16b, v18.4b[3]\n" | ||
| 654 | "bgt 17b\n" | ||
| 655 | "ldr q18, [x26, #0x0]\n" | ||
| 656 | "ld1 { v17.4s }, [x27]\n" | ||
| 657 | "add x27, x27, #0x10\n" | ||
| 658 | "ldr q20, [x26, #0x10]\n" | ||
| 659 | "ldr q16, [x27, #0x0]\n" | ||
| 660 | "add x26, x26, #0x20\n" | ||
| 661 | "mla v31.4s, v18.4s, v17.s[0]\n" | ||
| 662 | "mla v30.4s, v18.4s, v17.s[1]\n" | ||
| 663 | "mla v29.4s, v18.4s, v17.s[2]\n" | ||
| 664 | "mla v28.4s, v18.4s, v17.s[3]\n" | ||
| 665 | "fmul v19.4s, v20.4s, v16.s[0]\n" | ||
| 666 | "fmul v18.4s, v20.4s, v16.s[1]\n" | ||
| 667 | "fmul v17.4s, v20.4s, v16.s[2]\n" | ||
| 668 | "scvtf v31.4s, v31.4s\n" | ||
| 669 | "fmul v16.4s, v20.4s, v16.s[3]\n" | ||
| 670 | "scvtf v30.4s, v30.4s\n" | ||
| 671 | "scvtf v29.4s, v29.4s\n" | ||
| 672 | "scvtf v28.4s, v28.4s\n" | ||
| 673 | "fmul v31.4s, v31.4s, v19.4s\n" | ||
| 674 | "fmul v30.4s, v30.4s, v18.4s\n" | ||
| 675 | "fmul v29.4s, v29.4s, v17.4s\n" | ||
| 676 | "fmul v28.4s, v28.4s, v16.4s\n" | ||
| 677 | "ldr q18, [x26, #0x0]\n" | ||
| 678 | "ld1r { v17.4s }, [%x[clamp_vals]]\n" | ||
| 679 | "add x20, %x[clamp_vals], #0x4\n" | ||
| 680 | "cmp x25, #0x4\n" | ||
| 681 | "ld1r { v16.4s }, [x20]\n" | ||
| 682 | "add x26, x26, #0x10\n" | ||
| 683 | "fadd v31.4s, v31.4s, v18.4s\n" | ||
| 684 | "fadd v30.4s, v30.4s, v18.4s\n" | ||
| 685 | "fadd v29.4s, v29.4s, v18.4s\n" | ||
| 686 | "fadd v28.4s, v28.4s, v18.4s\n" | ||
| 687 | "fmax v31.4s, v31.4s, v17.4s\n" | ||
| 688 | "fmax v30.4s, v30.4s, v17.4s\n" | ||
| 689 | "fmax v29.4s, v29.4s, v17.4s\n" | ||
| 690 | "fmax v28.4s, v28.4s, v17.4s\n" | ||
| 691 | "fmin v31.4s, v31.4s, v16.4s\n" | ||
| 692 | "fmin v30.4s, v30.4s, v16.4s\n" | ||
| 693 | "fmin v29.4s, v29.4s, v16.4s\n" | ||
| 694 | "fmin v28.4s, v28.4s, v16.4s\n" | ||
| 695 | "blt 19f\n" | ||
| 696 | "mov x20, %x[dst]\n" | ||
| 697 | "cmp x13, #0x1\n" | ||
| 698 | "str q31, [x20, #0x0]\n" | ||
| 699 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 700 | "ble 22f\n" | ||
| 701 | "cmp x13, #0x2\n" | ||
| 702 | "str q30, [x20, #0x0]\n" | ||
| 703 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 704 | "ble 22f\n" | ||
| 705 | "cmp x13, #0x3\n" | ||
| 706 | "str q29, [x20, #0x0]\n" | ||
| 707 | "add x20, x20, %x[dst_stride_row]\n" | ||
| 708 | "ble 22f\n" | ||
| 709 | "str q28, [x20, #0x0]\n" | ||
| 710 | "b 22f\n" | ||
| 711 | "19:" // Row tail: Partial output | ||
| 712 | "mov x23, %x[dst]\n" | ||
| 713 | "cmp x13, #0x1\n" | ||
| 714 | "add x22, x23, %x[dst_stride_row]\n" | ||
| 715 | "csel x22, x22, x23, GT\n" | ||
| 716 | "cmp x13, #0x2\n" | ||
| 717 | "add x21, x23, %x[dst_stride_row], LSL #1\n" | ||
| 718 | "csel x21, x21, x22, GT\n" | ||
| 719 | "cmp x13, #0x3\n" | ||
| 720 | "add x20, x21, %x[dst_stride_row]\n" | ||
| 721 | "csel x20, x20, x21, GT\n" | ||
| 722 | "tbz x25, #1, 20f\n" | ||
| 723 | "st1 { v28.d }[0], [x20], #0x8\n" | ||
| 724 | "st1 { v29.d }[0], [x21], #0x8\n" | ||
| 725 | "st1 { v30.d }[0], [x22], #0x8\n" | ||
| 726 | "st1 { v31.d }[0], [x23], #0x8\n" | ||
| 727 | "tbz x25, #0, 21f\n" | ||
| 728 | "st1 { v28.s }[2], [x20]\n" | ||
| 729 | "st1 { v29.s }[2], [x21]\n" | ||
| 730 | "st1 { v30.s }[2], [x22]\n" | ||
| 731 | "st1 { v31.s }[2], [x23]\n" | ||
| 732 | "b 21f\n" | ||
| 733 | "20:" // Row tail: Output block 0: partial_1_0 | ||
| 734 | "st1 { v28.s }[0], [x20]\n" | ||
| 735 | "st1 { v29.s }[0], [x21]\n" | ||
| 736 | "st1 { v30.s }[0], [x22]\n" | ||
| 737 | "st1 { v31.s }[0], [x23]\n" | ||
| 738 | "21:" // Row tail: Output block 0: Done | ||
| 739 | "22:" // Row tail: Output stage exit | ||
| 740 | "subs x25, x25, #0x4\n" | ||
| 741 | "add %x[dst], %x[dst], #0x10\n" | ||
| 742 | "bgt 16b\n" | ||
| 743 | "subs x13, x13, #0x4\n" | ||
| 744 | "add %x[lhs_packed], %x[lhs_packed], x12\n" | ||
| 745 | "mov %x[dst], x24\n" | ||
| 746 | "bgt 15b\n" | ||
| 747 | "23:" // Row tail: Row loop skip | ||
| 748 | : [dst] "+&r"(dst), [lhs_packed] "+&r"(lhs_packed) | ||
| 749 | 161 | : [clamp_vals] "r"(clamp_vals), [dst_stride_row] "r"(dst_stride_row), [m] "r"(m), [n] "r"(n), | |
| 750 | 161 | [num_blocks] "r"(num_blocks), [rhs_packed] "r"(rhs_packed) | |
| 751 | : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", | ||
| 752 | "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", | ||
| 753 | "v30", "v31", "x9", "x10", "x11", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", | ||
| 754 | "x28"); | ||
| 755 | 161 | } | |
| 756 | |||
| 757 | #endif // Architectural features check. | ||
| 758 |