File size: 6,301 Bytes
8b7c501 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 |
// Copyright 2020 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
#include <xnnpack/assembly.h>
.syntax unified
// void xnn_f32_gemm_ukernel_4x4__asm_aarch32_vfp_ld64(
// size_t mr, r0
// size_t nc, r1
// size_t kc, r2 -> r5
// const float* a, r3
// size_t a_stride, sp + 96 -> (r11)
// const float* w, sp + 100 -> r9
// float* c, sp + 104 -> r6
// size_t cm_stride, sp + 108 -> (r7)
// size_t cn_stride, sp + 112 -> r11
// const union xnn_f32_default_params params) sp + 116 -> (r11)
// d8-d15, r4-r11,r14(lr) need to be preserved if used. r13(sp),r15(pc) are reserved.
// Register usage
// A0 r3 s0-s1 d0
// A1 r12 s2-s3 d1
// A2 r10 s4-s5 d2
// A3 r0 s6-s7 d3
// B r9 s8, s9, s10, s11 d4-d5
// B s12, s13, s14, s15 d6-d7
// C0 r6 s16-s17 d8 s18-s19 d9
// C1 r4 s20-s21 d10 s22-s23 d11
// C2 r8 s24-s25 d12 s26-s27 d13
// C3 r7 s28-s29 d14 s30-s31 d15
BEGIN_FUNCTION xnn_f32_gemm_ukernel_4x4__asm_aarch32_vfp_ld64
.arm
#ifndef __APPLE__
.arch armv6
.fpu vfp
#endif
# Push 96 bytes
PUSH {r4, r5, r6, r7, r8, r9, r10, r11} // 32
VPUSH {d8-d15} // +64 = 96
LDR r11, [sp, 96] // Load a_stride
LDRD r6, r7, [sp, 104] // Load c and cm_stride
# Clamp A and C pointers
CMP r0, 2 // if mr >= 2
ADD r12, r3, r11 // a1 = a0 + a_stride
ADD r4, r6, r7 // c1 = c0 + cm_stride
MOVLO r12, r3 // a1
MOVLO r4, r6 // c1
LDR r9, [sp, 100] // Load w
// if mr > 2
ADD r10, r12, r11 // a2 = a1 + a_stride
ADD r8, r4, r7 // c2 = c1 + cm_stride
MOVLS r10, r12 // a2
MOVLS r8, r4 // c2
CMP r0, 4 // if mr >=4
ADD r0, r10, r11 // a3 = a2 + a_stride
ADD r7, r8, r7 // c3 = c2 + cm_stride
LDR r11, [sp, 112] // Load cn_stride
MOVLO r0, r10 // a3
MOVLO r7, r8 // c3
0:
# Load initial bias from w into accumulators
VLDM r9!, {d8-d9} // Bias
SUBS r5, r2, 8
VMOV.F64 d10, d8
VMOV.F64 d12, d8
VMOV.F64 d14, d8
VMOV.F64 d11, d9
VMOV.F64 d13, d9
VMOV.F64 d15, d9
BLO 3f // less than 2 channels?
# Main loop - 2 floats of A (8 bytes)
1:
VLDM r3!, {d0} // A0
VLDM r9!, {d4-d5} // B0
VLDM r12!, {d1} // A1
VLDM r10!, {d2} // A2
VLDM r0!, {d3} // A3
VMLA.F32 s16, s8, s0
VMLA.F32 s17, s9, s0
VMLA.F32 s20, s8, s2
VMLA.F32 s21, s9, s2
VMLA.F32 s24, s8, s4
VMLA.F32 s25, s9, s4
VMLA.F32 s28, s8, s6
VMLA.F32 s29, s9, s6
VLDM r9!, {d6-d7} // B1
VMLA.F32 s18, s10, s0
VMLA.F32 s19, s11, s0
VMLA.F32 s22, s10, s2
VMLA.F32 s23, s11, s2
VMLA.F32 s26, s10, s4
VMLA.F32 s27, s11, s4
VMLA.F32 s30, s10, s6
VMLA.F32 s31, s11, s6
VMLA.F32 s16, s12, s1
VMLA.F32 s17, s13, s1
VMLA.F32 s20, s12, s3
VMLA.F32 s21, s13, s3
VMLA.F32 s24, s12, s5
VMLA.F32 s25, s13, s5
VMLA.F32 s28, s12, s7
VMLA.F32 s29, s13, s7
SUBS r5, r5, 8
VMLA.F32 s18, s14, s1
VMLA.F32 s19, s15, s1
VMLA.F32 s22, s14, s3
VMLA.F32 s23, s15, s3
VMLA.F32 s26, s14, s5
VMLA.F32 s27, s15, s5
VMLA.F32 s30, s14, s7
VMLA.F32 s31, s15, s7
BHS 1b
# Is there a remainder?- 1 float of A (4 bytes)
TST r5, 4
BNE 3f
2:
SUBS r1, r1, 4
BLO 4f
# Store full 4 x 4
VSTM r6, {d8-d9}
SUB r0, r0, r2
ADD r6, r11
VSTM r4, {d10-d11}
SUB r10, r10, r2
ADD r4, r11
VSTM r8, {d12-d13}
SUB r12, r12, r2
ADD r8, r11
VSTM r7, {d14-d15}
SUB r3, r3, r2
ADD r7, r11
BHI 0b
VPOP {d8-d15}
POP {r4, r5, r6, r7, r8, r9, r10, r11}
BX lr
3:
# Remainder- 1 float of A (4 bytes)
VLDM r3!, {s0} // A0
VLDM r9!, {d6-d7} // B
VLDM r12!, {s1} // A1
VLDM r10!, {s2} // A2
VLDM r0!, {s3} // A3
VMLA.F32 s16, s12, s0
VMLA.F32 s17, s13, s0
VMLA.F32 s18, s14, s0
VMLA.F32 s19, s15, s0
VMLA.F32 s20, s12, s1
VMLA.F32 s21, s13, s1
VMLA.F32 s22, s14, s1
VMLA.F32 s23, s15, s1
VMLA.F32 s24, s12, s2
VMLA.F32 s25, s13, s2
VMLA.F32 s26, s14, s2
VMLA.F32 s27, s15, s2
VMLA.F32 s28, s12, s3
VMLA.F32 s29, s13, s3
VMLA.F32 s30, s14, s3
VMLA.F32 s31, s15, s3
B 2b
# Store odd width
4:
TST r1, 2
BEQ 5f
VSTM r6!, {d8}
VMOV.F32 s16, s18
VSTM r4!, {d10}
VMOV.F32 s20, s22
VSTM r8!, {d12}
VMOV.F32 s24, s26
VSTM r7!, {d14}
VMOV.F32 s28, s30
5:
TST r1, 1
BEQ 6f
VSTR s16, [r6]
VSTR s20, [r4]
VSTR s24, [r8]
VSTR s28, [r7]
6:
VPOP {d8-d15}
POP {r4, r5, r6, r7, r8, r9, r10, r11}
BX lr
END_FUNCTION xnn_f32_gemm_ukernel_4x4__asm_aarch32_vfp_ld64
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
#endif
|