Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 1 | // Copyright 2019 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #include <xnnpack/assembly.h> |
| 7 | |
| 8 | .syntax unified |
| 9 | |
| 10 | // void xnn_f32_gemm_ukernel_4x8__aarch32_neon_ld64( |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 11 | // size_t mr, r0 |
| 12 | // size_t nc, r1 |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 13 | // size_t kc, r2 -> r7 |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 14 | // const uint8_t*restrict a, r3 |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 15 | // size_t a_stride, sp + 96 -> (r5) |
| 16 | // const void*restrict w, sp + 100 -> r9 |
| 17 | // uint8_t*restrict c, sp + 104 -> r14 |
| 18 | // size_t cm_stride, sp + 108 -> (r6) |
| 19 | // size_t cn_stride, sp + 112 -> r5 |
| 20 | // const union xnn_f32_output_params params[restrict static 1]) sp + 116 -> (r5) |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 21 | |
| 22 | |
| 23 | // inner loop registers |
| 24 | |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 25 | // A0 r3 d0 |
| 26 | // A1 r12 d1 |
| 27 | // A2 r10 d2 |
| 28 | // A3 r0 d3 |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 29 | |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 30 | // B r9 d8, d9, d10, d11 |
| 31 | // B d12, d13, d14, d15 |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 32 | |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 33 | // C0 r14 d16-d17 q8 d18-d19 q9 |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 34 | // C1 r4 d20-d21 q10 d22-d23 q11 |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 35 | // C2 r8 d24-d25 q12 d26-d27 q13 |
| 36 | // C3 r6 d28-d29 q14 d30-d31 q15 |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 37 | |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 38 | // Clamp (r5) d4 d5 d6 d7 |
| 39 | // Unused r11 |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 40 | |
| 41 | BEGIN_FUNCTION xnn_f32_gemm_ukernel_4x8__aarch32_neon_ld64 |
| 42 | .arm |
| 43 | #ifndef __APPLE__ |
| 44 | .arch armv7-a |
| 45 | .fpu neon |
| 46 | #endif |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 47 | // Push 96 bytes |
| 48 | PUSH {r4, r5, r6, r7, r8, r9, r10, r14} // 32 |
| 49 | VPUSH {d8-d15} // +64 = 96 |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 50 | |
| 51 | // Clamp A and C pointers |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 52 | |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 53 | LDR r5, [sp, 96] // a_stride |
| 54 | LDR r14, [sp, 104] // c |
| 55 | LDR r6, [sp, 108] // cm_stride |
| 56 | |
| 57 | CMP r0, 2 // if mr >= 2 |
| 58 | ADDHS r12, r3, r5 // a1 = a0 + a_stride |
| 59 | ADDHS r4, r14, r6 // c1 = c0 + cm_stride |
| 60 | MOVLO r12, r3 // a1 |
| 61 | MOVLO r4, r14 // c1 |
| 62 | // if mr > 2 |
| 63 | ADDHI r10, r12, r5 // a2 = a1 + a_stride |
| 64 | ADDHI r8, r4, r6 // c2 = c1 + cm_stride |
| 65 | MOVLS r10, r12 // a2 |
| 66 | MOVLS r8, r4 // c2 |
| 67 | |
| 68 | CMP r0, 4 // if mr >=4 |
| 69 | ADDHS r0, r10, r5 // a3 = a2 + a_stride |
| 70 | ADDHS r6, r8, r6 // c3 = c2 + cm_stride |
| 71 | MOVLO r0, r10 // a3 |
| 72 | MOVLO r6, r8 // c3 |
| 73 | |
| 74 | // Load params pointer |
| 75 | LDR r5, [sp, 116] // clamping_params |
| 76 | LDR r9, [sp, 100] // W |
| 77 | |
| 78 | // Load clamping_params values |
| 79 | VLD1.32 {d4[]-d5[]}, [r5]! |
| 80 | VLD1.32 {d6[]-d7[]}, [r5] |
| 81 | |
| 82 | LDR r5, [sp, 112] // cn_stride |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 83 | |
| 84 | 1: |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 85 | # Load initial bias from w into accumulators |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 86 | VLDM r9!, {d16-d19} // Bias |
| 87 | SUBS r7, r2, 8 |
| 88 | VMOV q10, q8 |
| 89 | VMOV q11, q9 |
| 90 | VMOV q12, q8 |
| 91 | VMOV q13, q9 |
| 92 | VMOV q14, q8 |
| 93 | VMOV q15, q9 |
| 94 | BLO 3f // less than 2 channels? |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 95 | |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 96 | // Main loop - 2 floats of A (8 bytes) |
| 97 | 2: |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 98 | VLD1.32 {d0}, [r3]! // A0 |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 99 | VLDM r9!, {d8-d11} // B0 |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 100 | VLD1.32 {d1}, [r12]! // A1 |
| 101 | VLD1.32 {d2}, [r10]! // A2 |
| 102 | VLD1.32 {d3}, [ r0]! // A3 |
| 103 | |
| 104 | VMLA.F32 q8, q4, d0[0] |
| 105 | VMLA.F32 q9, q5, d0[0] |
| 106 | VMLA.F32 q10, q4, d1[0] |
| 107 | VMLA.F32 q11, q5, d1[0] |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 108 | VLDM r9!, {d12-d15} // B1 |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 109 | VMLA.F32 q12, q4, d2[0] |
| 110 | VMLA.F32 q13, q5, d2[0] |
| 111 | VMLA.F32 q14, q4, d3[0] |
| 112 | VMLA.F32 q15, q5, d3[0] |
| 113 | VMLA.F32 q8, q6, d0[1] |
| 114 | VMLA.F32 q9, q7, d0[1] |
| 115 | VMLA.F32 q10, q6, d1[1] |
| 116 | VMLA.F32 q11, q7, d1[1] |
| 117 | VMLA.F32 q12, q6, d2[1] |
| 118 | VMLA.F32 q13, q7, d2[1] |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 119 | SUBS r7, r7, 8 |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 120 | VMLA.F32 q14, q6, d3[1] |
| 121 | VMLA.F32 q15, q7, d3[1] |
| 122 | BHS 2b |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 123 | |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 124 | 3: |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 125 | // Is there a remainder?- 1 floats of A (4 bytes) |
| 126 | TST r7, 4 |
| 127 | BNE 7f |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 128 | |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 129 | 4: |
| 130 | // Clamp |
| 131 | VMIN.F32 q8, q8, q2 |
| 132 | VMIN.F32 q9, q9, q2 |
| 133 | VMIN.F32 q10, q10, q2 |
| 134 | VMIN.F32 q11, q11, q2 |
| 135 | VMIN.F32 q12, q12, q2 |
| 136 | VMIN.F32 q13, q13, q2 |
| 137 | VMIN.F32 q14, q14, q2 |
| 138 | VMIN.F32 q15, q15, q2 |
| 139 | VMAX.F32 q8, q8, q3 |
| 140 | VMAX.F32 q9, q9, q3 |
| 141 | VMAX.F32 q10, q10, q3 |
| 142 | VMAX.F32 q11, q11, q3 |
| 143 | VMAX.F32 q12, q12, q3 |
| 144 | VMAX.F32 q13, q13, q3 |
| 145 | VMAX.F32 q14, q14, q3 |
| 146 | VMAX.F32 q15, q15, q3 |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 147 | |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 148 | // Store full 4 x 8 |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 149 | |
| 150 | CMP r1, 8 |
| 151 | BLO 8f |
| 152 | |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 153 | SUBS r1, r1, 8 // Loop counter |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 154 | VST1.32 {d16-d19}, [r14], r5 |
| 155 | SUB r0, r0, r2 |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 156 | VST1.32 {d20-d23}, [r4], r5 |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 157 | SUB r10, r10, r2 |
| 158 | VST1.32 {d24-d27}, [r8], r5 |
| 159 | SUB r12, r12, r2 |
| 160 | VST1.32 {d28-d31}, [r6], r5 |
| 161 | SUB r3, r3, r2 |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 162 | BNE 1b |
| 163 | |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 164 | 6: |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 165 | VPOP {d8-d15} |
| 166 | POP {r4, r5, r6, r7, r8, r9, r10, pc} |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 167 | |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 168 | 7: |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 169 | // Remainder- 1 floats of A (4 bytes) |
| 170 | VLDM r3!, {s0} // A0 |
| 171 | VLDM r9!, {d8-d11} // B0 |
| 172 | VLDM r12!, {s2} // A1 |
| 173 | VLDM r10!, {s4} // A2 |
| 174 | VLDM r0!, {s6} // A3 |
| 175 | VMLA.F32 q8, q4, d0[0] |
| 176 | VMLA.F32 q9, q5, d0[0] |
| 177 | VMLA.F32 q10, q4, d1[0] |
| 178 | VMLA.F32 q11, q5, d1[0] |
| 179 | VMLA.F32 q12, q4, d2[0] |
| 180 | VMLA.F32 q13, q5, d2[0] |
| 181 | VMLA.F32 q14, q4, d3[0] |
| 182 | VMLA.F32 q15, q5, d3[0] |
| 183 | B 4b |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 184 | |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 185 | // Store odd width |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 186 | |
| 187 | 9: |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 188 | VST1.32 {d16-d17}, [r14]! |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 189 | VST1.32 {d20-d21}, [r4]! |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 190 | VST1.32 {d24-d25}, [r8]! |
| 191 | VST1.32 {d28-d29}, [r6]! |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 192 | TST r1, 2 |
| 193 | BNE 11f |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 194 | |
| 195 | 10: |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 196 | VMOV d19, d18 |
| 197 | VMOV d23, d22 |
| 198 | VMOV d27, d26 |
| 199 | VMOV d31, d30 |
| 200 | TST r1, 1 |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 201 | BEQ 6b |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 202 | B 12f |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 203 | |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 204 | 8: |
| 205 | TST r1, 4 |
| 206 | BNE 9b |
| 207 | VMOV q9, q8 |
| 208 | VMOV q11, q10 |
| 209 | VMOV q13, q12 |
| 210 | VMOV q15, q14 |
| 211 | TST r1, 2 |
| 212 | BEQ 10b |
| 213 | |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 214 | 11: |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 215 | VST1.32 {d18}, [r14]! |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 216 | VST1.32 {d22}, [r4]! |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 217 | VST1.32 {d26}, [r8]! |
| 218 | VST1.32 {d30}, [r6]! |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 219 | TST r1, 1 |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 220 | BEQ 6b |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 221 | |
| 222 | 12: |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 223 | VST1.32 {d19[0]}, [r14] |
Frank Barchard | 72d6afb | 2019-11-30 19:21:26 -0800 | [diff] [blame] | 224 | VST1.32 {d23[0]}, [r4] |
Frank Barchard | 61cad89 | 2019-12-02 13:06:15 -0800 | [diff] [blame^] | 225 | VST1.32 {d27[0]}, [r8] |
| 226 | VST1.32 {d31[0]}, [r6] |
| 227 | B 6b |
Frank Barchard | 8b0f026 | 2019-11-27 23:18:40 -0800 | [diff] [blame] | 228 | |
| 229 | END_FUNCTION xnn_f32_gemm_ukernel_4x8__aarch32_neon_ld64 |
| 230 | |
| 231 | #ifdef __ELF__ |
| 232 | .section ".note.GNU-stack","",%progbits |
| 233 | #endif |
| 234 | |
| 235 | |
| 236 | |