How Can SIMD Instructions Be Used to Implement a High-Performance atoi Function?-C++-php.cn

How Can SIMD Instructions Be Used to Implement a High-Performance atoi Function?

DDD

Release： 2024-12-01 08:05:16

Original

766 people have browsed it

How Can SIMD Instructions Be Used to Implement a High-Performance atoi Function?

SIMD Implementation of the atoi Function

Introduction:

atoi is a function that converts a string representation of an integer to its numeric value. This article explores how to implement atoi using SIMD instructions.

Algorithm:

Initialize a vector V with values 10^0, 10^1, ..., 10^N.
Convert each character in the input string to an integer and store it in vector S.
Multiply each element of S by the corresponding element of V and store the results in a new vector P.
Perform a series of horizontal adds and multiplies on P to obtain the final result.

Implementation in GNU Assembler:

.intel_syntax noprefix
.data
  .align 64
    ddqDigitRange: .byte  '0','9',0,0,0,0,0,0,0,0,0,0,0,0,0,0
    ddqShuffleMask:.byte  15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0 
    ddqFactor1:    .word  1,10,100,1000, 1,10,100,1000  
    ddqFactor2:    .long  1,10000,100000000,0

Copy after login

.text    
_start:
   mov   esi, lpInputNumberString
   /* (**A**) indicate negative number in EDX */
   mov   eax, -1
   xor   ecx, ecx
   xor   edx, edx
   mov   bl,  byte ptr [esi]
   cmp   bl,  '-'
   cmove edx, eax
   cmp   bl,  '+'
   cmove ecx, eax
   sub   esi, edx
   sub   esi, ecx
   /* (**B**)remove leading zeros */
   xor   eax,eax               /* return value ZERO */
  remove_leading_zeros:
   inc   esi
   cmp   byte ptr [esi-1], '0'  /* skip leading zeros */
  je remove_leading_zeros
   cmp   byte ptr [esi-1], 0    /* catch empty string/number */
  je FINISH             /* if first char is invalid return 0 - prevent processing empty string - 0 is still in EAX */
   dec   esi
   /* check for valid digit-chars and invert from front to back */
   pxor      xmm2, xmm2         
   movdqa    xmm0, xmmword ptr [ddqDigitRange]
   movdqu    xmm1, xmmword ptr [esi]
   pcmpistri xmm0, xmm1, 0b00010100 /* (**C**) iim8=Unsigned bytes, Ranges, Negative Polarity(-), returns strlen() in ECX */
  jo FINISH             /* if first char is invalid return 0 - prevent processing empty string - 0 is still in EAX */
   mov al , '0'         /* value to subtract from chars */
   sub ecx, 16          /* len-16=negative to zero for shuffle mask */
   movd      xmm0, ecx
   pshufb    xmm0, xmm2 /* broadcast CL to all 16 BYTEs */
   paddb     xmm0, xmmword ptr [ddqShuffleMask] /* Generate permute mask for PSHUFB - all bytes < 0 have highest bit set means place gets zeroed */
   pshufb    xmm1, xmm0 /* (**D**) permute - now from highest to lowest BYTE are factors 10^0, 10^1, 10^2, ... */
   movd      xmm0, eax                         /* AL='0' from above */
   pshufb    xmm0, xmm2                        /* broadcast AL to XMM0 */
   psubusb   xmm1, xmm0                        /* (**1**) */
   movdqa    xmm0, xmm1
   punpcklbw xmm0, xmm2                        /* (**2**) */
   punpckhbw xmm1, xmm2
   pmaddwd   xmm0, xmmword ptr [ddqFactor1]    /* (**3**) */
   pmaddwd   xmm1, xmmword ptr [ddqFactor1]
   phaddd    xmm0, xmm1                        /* (**4**) */
   pmulld    xmm0, xmmword ptr [ddqFactor2]    /* (**5**) */
   pshufd    xmm1, xmm0, 0b11101110            /* (**6**) */
   paddd     xmm0, xmm1
   pshufd    xmm1, xmm0, 0b01010101            /* (**7**) */
   paddd     xmm0, xmm1
   movd      eax, xmm0
   /* negate if negative number */              
   add       eax, edx                          /* (**8**) */
   xor       eax, edx
  FINISH:
   /* EAX is return (u)int value */

Copy after login

Advantages of SIMD Implementation:

Increased performance for processing large strings of numbers.
Feasible for x86 and x86-64 architectures.
Supports multiple simultaneous atoi operations.

Limitations:

Requires specific SSE4.2 instructions.
May not be suitable for small strings or strings with mixed characters.

Conclusion:

The SIMD implementation of atoi offers significant speedup for processing large integer strings compared to traditional methods. The algorithm is optimized for x86 and x86-64 architectures and can perform multiple atoi operations in parallel. While it has limitations in handling small and mixed-character strings, it remains a valuable technique for numerical computations.

The above is the detailed content of How Can SIMD Instructions Be Used to Implement a High-Performance atoi Function?. For more information, please follow other related articles on the PHP Chinese website!