2. #ifdef __aarch64__
    3. #include "MNNAsmGlobal.h"
    5. .text // 代码段
    6. .align 5 // 地址对齐 2^5 = 32
    8. asm_function MNNAddBias
    9. // void MNNAddBias(float* dst, const float* bias, size_t planeNumber, size_t biasNumber) {
    10. // x0: dst, x1: bias, x2: planeNumber, x3: biasNumber
    12. cmp x3, #0
    13. beq End
    15. cmp x2, #0
    16. beq End
    18. // Relu 以及 Relu6
    19. movi v22.4s, #0 // 立即数存入寄存器
    20. movi v21.4s, #6 // 6存入寄存器
    21. scvtf v21.4s, v21.4s // 整型转为浮点型
    23. LoopBias:
    24. ld1 {v31.4s}, [x1], #16 // 读取128位到v31寄存器，v31视为4个32位寄存器
    26. mov x4, x2
    28. // 循环展开，每次计算四个循环
    29. L4:
    30. cmp x4, #3
    31. ble L1
    32. Loop4:
    33. mov x5, x0
    34. // 指令重排，减少指令依赖
    35. ld1 {v0.4s, v1.4s}, [x5], #32
    36. fadd v0.4s, v0.4s, v31.4s
    37. ld1 {v2.4s, v3.4s}, [x5]
    38. fadd v1.4s, v1.4s, v31.4s
    39. fadd v2.4s, v2.4s, v31.4s
    40. st1 {v0.4s, v1.4s}, [x0], #32
    41. fadd v3.4s, v3.4s, v31.4s
    42. st1 {v1.4s, v3.4s}, [x0], #32
    43. sub x4, x4, #4
    44. cmp x4, #4
    45. bge Loop4
    47. // 剩余循环计算
    48. L1:
    49. cmp x4, #0
    50. beq EndLoopPlane
    51. Loop1:
    52. ld1 {v0.4s}, [x0]
    53. fadd v0.4s, v0.4s, v31.4s
    54. subs x4, x4, #1
    55. st1 {v0.4s}, [x0], #16
    56. bne Loop1
    58. EndLoopPlane:
    59. subs x3, x3, 1
    60. bne LoopBias
    62. End:
    64. ret
    66. #endif