在M1 Mac上搜索数组的最快方法

huangapple
go评论116阅读模式
英文:

Fastest way to search an array on m1 mac

问题

我正在尝试从内存加载一个u16数组，并尽快在M1 Mac上找到第一个小于某个数字的元素。我已经查看了NEON指令，但无法找到一个好的方法来实现这个目标。虽然有比较向量指令，但它们会将您留下来具有全部为1或0的元素的向量。如何将它们加载到通用寄存器中？并且有没有一种通过测试向量指令来中断循环的方法？

英文:

I am trying to load an array of u16s from memory and find the first element that is less than some number, as fast as possible on an M1 mac. I have been looking through the NEON instructions, but I wasn't able to find a good way to do it. There are comparison vector instructions, but they leave you with vectors of elements of all 1 or 0. How would you get that into a general purpose register? And is there a way to break a loop by testing a vector instruction?

答案1

得分: 1

Here is the translated code without the comments and other non-translatable parts:

  1.     .arch armv8-a
  2.     .global searchArrayU16
  3.     .text
  5.     pArray0 .req x0
  6.     thresh .req w1
  7.     len .req x2
  8.     stride .req x3
  9.     pArray1 .req x4
  10.     count .req w5
  11.     val0 .req x6
  12.     val0w .req w6
  13.     val1 .req x7
  14.     val1w .req w7
  16.     .balign 64
  17.     .func
  18.     searchArrayU16:
  19.         adr x15, 2f
  20.         dup v0.8h, thresh
  21.         ld1r {v1.2d}, [x15]
  22.         mov stride, #128
  23.         add pArray1, pArray0, #64
  24.         mov count, #0
  26.         b 1f
  28.     .balign 8
  29.     2:
  30.     .byte 1, 2, 4, 8, 16, 32, 64, 128
  32.     .balign 64
  33.     1:
  34.         ld1 {v16.8h-v19.8h}, [pArray0], stride
  35.         ld1 {v20.8h-v23.8h}, [pArray1], stride
  36.         ld1 {v24.8h-v27.8h}, [pArray0], stride
  37.         ld1 {v28.8h-v31.8h}, [pArray1], stride
  39.         cmhi v16.8h, v0.8h, v16.8h
  40.         cmhi v17.8h, v0.8h, v17.8h
  41.         cmhi v18.8h, v0.8h, v18.8h
  42.         cmhi v19.8h, v0.8h, v19.8h
  43.         cmhi v20.8h, v0.8h, v20.8h
  44.         cmhi v21.8h, v0.8h, v21.8h
  45.         cmhi v22.8h, v0.8h, v22.8h
  46.         cmhi v23.8h, v0.8h, v23.8h
  47.         cmhi v24.8h, v0.8h, v24.8h
  48.         cmhi v25.8h, v0.8h, v25.8h
  49.         cmhi v26.8h, v0.8h, v26.8h
  50.         cmhi v27.8h, v0.8h, v27.8h
  51.         cmhi v28.8h, v0.8h, v28.8h
  52.         cmhi v29.8h, v0.8h, v29.8h
  53.         cmhi v30.8h, v0.8h, v30.8h
  54.         cmhi v31.8h, v0.8h, v31.8h
  56.         uzp1 v16.16b, v16.16b, v17.16b
  57.         uzp1 v18.16b, v18.16b, v19.16b
  58.         uzp1 v20.16b, v20.16b, v21.16b
  59.         uzp1 v22.16b, v22.16b, v23.16b
  60.         uzp1 v24.16b, v24.16b, v25.16b
  61.         uzp1 v26.16b, v26.16b, v27.16b
  62.         uzp1 v28.16b, v28.16b, v29.16b
  63.         uzp1 v30.16b, v30.16b, v31.16b
  65.         and v16.16b, v16.16b, v1.16b
  66.         and v18.16b, v18.16b, v1.16b
  67.         and v20.16b, v20.16b, v1.16b
  68.         and v22.16b, v22.16b, v1.16b
  69.         and v24.16b, v24.16b, v1.16b
  70.         and v26.16b, v26.16b, v1.16b
  71.         and v28.16b, v28.16b, v1.16b
  72.         and v30.16b, v30.16b, v1.16b
  74.         addp v16.16b, v16.16b, v18.16b
  75.         addp v20.16b, v20.16b, v22.16b
  76.         addp v24.16b, v24.16b, v26.16b
  77.         addp v28.16b, v28.16b, v30.16b
  79.         addp v16.16b, v16.16b, v20.16b
  80.         addp v24.16b, v24.16b, v28.16b
  82.         add count, count, #128
  83.         addp v16.16b, v16.16b, v24.16b
  85.         mov val0, v16.d[0]
  86.         mov val1, v16.d[1]
  88.         orr x15, val0, val1
  89.         cbnz x15, 1f
  91.         cmp len, count, uxtw
  92.         b.hi 1b
  94.     .balign 16
  95.     mov w0, #-1 // no match found
  96.     ret
  98.     .balign 16
  99.     1:
  100.         rbit val0, val0
  101.         rbit val1, val1
  102.         cmp val0, #0
  103.         sub w0, count, #128
  104.         sub w1, count, #64
  105.         clz val0, val0
  106.         clz val1, val1
  107.         add w0, w0, val0w
  108.         add w1, w1, val1w
  109.         csel w0, w0, w1, ne
  110.         ret
  111.     .endfunc
  112.     .end

Please note that this is a direct translation, and the code's functionality is retained as per the original code provided.

英文: 
  1. // int32_t searchArrayU16(uint16_t *pArray, uint16_t threshold, uint32_t len);
  2. // assert(len & 127 == 0);
  3. // assert(len >= 128);
  5. 	.arch armv8-a
  6. 	.global	searchArrayU16
  7. 	.text
  10. pArray0	.req	x0
  11. thresh	.req	w1
  12. len		.req	x2
  13. stride	.req	x3
  14. pArray1	.req	x4
  15. count	.req	w5
  16. val0	.req	x6
  17. val0w	.req	w6
  18. val1	.req	x7
  19. val1w	.req	w7
  21. .balign 64
  22. .func
  23. searchArrayU16:
  24. 	adr		x15, 2f
  25. 	dup		v0.8h, thresh
  26. 	ld1r	{v1.2d}, [x15]
  27. 	mov		stride, #128
  28. 	add		pArray1, pArray0, #64
  29. 	mov		count, #0
  31. 	b		1f
  33. .balign 8
  34. 2:
  35. .byte	1, 2, 4, 8, 16, 32, 64, 128
  37. .balign 64
  38. 1:
  39. 	ld1		{v16.8h-v19.8h}, [pArray0], stride
  40. 	ld1		{v20.8h-v23.8h}, [pArray1], stride
  41. 	ld1		{v24.8h-v27.8h}, [pArray0], stride
  42. 	ld1		{v28.8h-v31.8h}, [pArray1], stride
  44. 	cmhi	v16.8h, v0.8h, v16.8h
  45. 	cmhi	v17.8h, v0.8h, v17.8h
  46. 	cmhi	v18.8h, v0.8h, v18.8h
  47. 	cmhi	v19.8h, v0.8h, v19.8h
  48. 	cmhi	v20.8h, v0.8h, v20.8h
  49. 	cmhi	v21.8h, v0.8h, v21.8h
  50. 	cmhi	v22.8h, v0.8h, v22.8h
  51. 	cmhi	v23.8h, v0.8h, v23.8h
  52. 	cmhi	v24.8h, v0.8h, v24.8h
  53. 	cmhi	v25.8h, v0.8h, v25.8h
  54. 	cmhi	v26.8h, v0.8h, v26.8h
  55. 	cmhi	v27.8h, v0.8h, v27.8h
  56. 	cmhi	v28.8h, v0.8h, v28.8h
  57. 	cmhi	v29.8h, v0.8h, v29.8h
  58. 	cmhi	v30.8h, v0.8h, v30.8h
  59. 	cmhi	v31.8h, v0.8h, v31.8h
  61. 	uzp1	v16.16b, v16.16b, v17.16b
  62. 	uzp1	v18.16b, v18.16b, v19.16b
  63. 	uzp1	v20.16b, v20.16b, v21.16b
  64. 	uzp1	v22.16b, v22.16b, v23.16b
  65. 	uzp1	v24.16b, v24.16b, v25.16b
  66. 	uzp1	v26.16b, v26.16b, v27.16b
  67. 	uzp1	v28.16b, v28.16b, v29.16b
  68. 	uzp1	v30.16b, v30.16b, v31.16b
  70. 	and		v16.16b, v16.16b, v1.16b
  71. 	and		v18.16b, v18.16b, v1.16b
  72. 	and		v20.16b, v20.16b, v1.16b
  73. 	and		v22.16b, v22.16b, v1.16b
  74. 	and		v24.16b, v24.16b, v1.16b
  75. 	and		v26.16b, v26.16b, v1.16b
  76. 	and		v28.16b, v28.16b, v1.16b
  77. 	and		v30.16b, v30.16b, v1.16b
  79. 	addp	v16.16b, v16.16b, v18.16b
  80. 	addp	v20.16b, v20.16b, v22.16b
  81. 	addp	v24.16b, v24.16b, v26.16b
  82. 	addp	v28.16b, v28.16b, v30.16b
  84. 	addp	v16.16b, v16.16b, v20.16b
  85. 	addp	v24.16b, v24.16b, v28.16b
  87. 	add		count, count, #128
  88. 	addp	v16.16b, v16.16b, v24.16b
  90. // total pipeline stall here
  92. 	mov		val0, v16.d[0]
  93. 	mov		val1, v16.d[1]
  95. 	orr		x15, val0, val1
  96. 	cbnz	x15, 1f	// found a match!!!
  98. 	cmp		len, count, uxtw
  99. 	b.hi	1b
  100. .balign 16
  101. 	mov		w0, #-1		// no match found
  102. 	ret
  104. .balign 16
  105. 1:
  106. 	rbit	val0, val0
  107. 	rbit	val1, val1
  108. 	cmp		val0, #0
  109. 	sub		w0, count, #128
  110. 	sub		w1, count, #64
  111. 	clz		val0, val0
  112. 	clz		val1, val1
  113. 	add		w0, w0, val0w
  114. 	add		w1, w1, val1w
  115. 	csel	w0, w0, w1, ne
  116. 	ret
  117. .endfunc
  118. .end

Here you are. It returns -1 when no match is found.
It should work on all armv8-a cores or above.

huangapple
  • 本文由 发表于 2023年1月9日 15:46:24
  • 转载请务必保留本文链接：https://go.coder-hub.com/75054362.html
  • apple-m1
  • arm64
  • assembly
  • neon
  • simd
匿名

发表评论

匿名网友

:?: :razz: :sad: :evil: :!: :smile: :oops: :grin: :eek: :shock: :???: :cool: :lol: :mad: :twisted: :roll: :wink: :idea: :arrow: :neutral: :cry: :mrgreen:

确定