我想尝试使用SIMD指令编写一个atoi实现,以包含在RapidJSON(一个C JSON阅读器/写入器库)中。它目前在其他地方进行了一些SSE2和SSE4.2优化。
如果是速度增益,可以并行完成多个atoi
结果。这些字符串最初来自JSON数据的缓冲区,因此多原子函数必须执行任何所需的滑动。
我提出的算法如下:
我的目标是x86和x86-64体系结构。
我知道AVX2支持三个操作数融合乘加,所以我将能够执行和=数*有效位数和。
这就是我到目前为止的地方。
我的算法正确吗?有更好的方法吗?
是否有使用任何SIMD指令集的atoi参考实现?
我会这样处理这个问题:
'0'
9
的第一个值
1000
,100
,10
,1
)并与之相乘
9
的值,则转到第二步
您可以通过在步骤5中将所有以错误条目开头的条目归零来简化算法,而不是移位,然后在最后除以适当的10次方。
请记住,该算法读取的数据超过了字符串的末尾,因此不能替代atoi
。
该算法及其实现现已完成。它已经完成并经过(适度)测试(针对较少的恒定内存使用和容忍加号进行了更新)。
此代码的属性如下:
int
和uint
,从MIN_int=-2147483648
到MAX_int=2147483647
和从MIN_uint=0
到MAX_uint=4294967295
'-'
字符表示负数(合理),前导的'
字符被忽略0=-0
关于此实施:
使用GNU Assembler(as
)运行。intel_syntaxnoprefix
开头算法的方法:
- Pointer to number string is expected in ESI
- Check if first char is '-', then indicate if negative number in EDX (**A**)
- Check for leading zeros and EOS (**B**)
- Check string for valid digits and get strlen() of valid chars (**C**)
- Reverse string so that power of
10^0 is always at BYTE 15
10^1 is always at BYTE 14
10^2 is always at BYTE 13
10^3 is always at BYTE 12
10^4 is always at BYTE 11
...
and mask out all following chars (**D**)
- Subtract saturated '0' from each of the 16 possible chars (**1**)
- Take 16 consecutive byte-values and and split them to WORDs
in two XMM-registers (**2**)
P O N M L K J I | H G F E D C B A ->
H G F E | D C B A (XMM0)
P O N M | L K J I (XMM1)
- Multiply each WORD by its place-value modulo 10000 (1,10,100,1000)
(factors smaller then MAX_WORD, 4 factors per QWORD/halfXMM)
(**2**) so we can horizontally combine twice before another multiply.
The PMADDWD instruction can do this and the next step:
- Horizontally add adjacent WORDs to DWORDs (**3**)
H*1000+G*100 F*10+E*1 | D*1000+C*100 B*10+A*1 (XMM0)
P*1000+O*100 N*10+M*1 | L*1000+K*100 J*10+I*1 (XMM1)
- Horizontally add adjacent DWORDs from XMM0 and XMM1 to XMM0 (**4**)
xmmDst[31-0] = xmm0[63-32] + xmm0[31-0]
xmmDst[63-32] = xmm0[127-96] + xmm0[95-64]
xmmDst[95-64] = xmm1[63-32] + xmm1[31-0]
xmmDst[127-96] = xmm1[127-96] + xmm1[95-64]
- Values in XMM0 are multiplied with the factors (**5**)
P*1000+O*100+N*10+M*1 (DWORD factor 1000000000000 = too big for DWORD, but possibly useful for QWORD number strings)
L*1000+K*100+J*10+I*1 (DWORD factor 100000000)
H*1000+G*100+F*10+E*1 (DWORD factor 10000)
D*1000+C*100+B*10+A*1 (DWORD factor 1)
- The last step is adding these four DWORDs together with 2*PHADDD emulated by 2*(PSHUFD+PADDD)
- xmm0[31-0] = xmm0[63-32] + xmm0[31-0] (**6**)
xmm0[63-32] = xmm0[127-96] + xmm0[95-64]
(the upper QWORD contains the same and is ignored)
- xmm0[31-0] = xmm0[63-32] + xmm0[31-0] (**7**)
- If the number is negative (indicated in EDX by 000...0=pos or 111...1=neg), negate it(**8**)
以及GNU汇编程序中使用英特尔语法的示例实现:
.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
.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
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 */
针对Haswell 32位处理器的英特尔IACA吞吐量分析结果:
Throughput Analysis Report
--------------------------
Block Throughput: 16.10 Cycles Throughput Bottleneck: InterIteration
Port Binding In Cycles Per Iteration:
---------------------------------------------------------------------------------------
| Port | 0 - DV | 1 | 2 - D | 3 - D | 4 | 5 | 6 | 7 |
---------------------------------------------------------------------------------------
| Cycles | 9.5 0.0 | 10.0 | 4.5 4.5 | 4.5 4.5 | 0.0 | 11.1 | 11.4 | 0.0 |
---------------------------------------------------------------------------------------
N - port number or number of cycles resource conflict caused delay, DV - Divider pipe (on port 0)
D - Data fetch pipe (on ports 2 and 3), CP - on a critical path
F - Macro Fusion with the previous instruction occurred
* - instruction micro-ops not bound to a port
^ - Micro Fusion happened
# - ESP Tracking sync uop was issued
@ - SSE instruction followed an AVX256 instruction, dozens of cycles penalty is expected
! - instruction not supported, was not accounted in Analysis
| Num Of | Ports pressure in cycles | |
| Uops | 0 - DV | 1 | 2 - D | 3 - D | 4 | 5 | 6 | 7 | |
---------------------------------------------------------------------------------
| 0* | | | | | | | | | | xor eax, eax
| 0* | | | | | | | | | | xor ecx, ecx
| 0* | | | | | | | | | | xor edx, edx
| 1 | | 0.1 | | | | | 0.9 | | | dec eax
| 1 | | | 0.5 0.5 | 0.5 0.5 | | | | | CP | mov bl, byte ptr [esi]
| 1 | | | | | | | 1.0 | | CP | cmp bl, 0x2d
| 2 | 0.1 | 0.2 | | | | | 1.8 | | CP | cmovz edx, eax
| 1 | 0.1 | 0.5 | | | | | 0.4 | | CP | cmp bl, 0x2b
| 2 | 0.5 | 0.2 | | | | | 1.2 | | CP | cmovz ecx, eax
| 1 | 0.2 | 0.5 | | | | | 0.2 | | CP | sub esi, edx
| 1 | 0.2 | 0.5 | | | | | 0.3 | | CP | sub esi, ecx
| 0* | | | | | | | | | | xor eax, eax
| 1 | 0.3 | 0.1 | | | | | 0.6 | | CP | inc esi
| 2^ | 0.3 | | 0.5 0.5 | 0.5 0.5 | | | 0.6 | | | cmp byte ptr [esi-0x1], 0x30
| 0F | | | | | | | | | | jz 0xfffffffb
| 2^ | 0.6 | | 0.5 0.5 | 0.5 0.5 | | | 0.4 | | | cmp byte ptr [esi-0x1], 0x0
| 0F | | | | | | | | | | jz 0x8b
| 1 | 0.1 | 0.9 | | | | | | | CP | dec esi
| 1 | | | 0.5 0.5 | 0.5 0.5 | | | | | | movdqa xmm0, xmmword ptr [0x80492f0]
| 1 | | | 0.5 0.5 | 0.5 0.5 | | | | | CP | movdqu xmm1, xmmword ptr [esi]
| 0* | | | | | | | | | | pxor xmm2, xmm2
| 3 | 2.0 | 1.0 | | | | | | | CP | pcmpistri xmm0, xmm1, 0x14
| 1 | | | | | | | 1.0 | | | jo 0x6e
| 1 | | 0.4 | | | | 0.1 | 0.5 | | | mov al, 0x30
| 1 | 0.1 | 0.5 | | | | 0.1 | 0.3 | | CP | sub ecx, 0x10
| 1 | | | | | | 1.0 | | | CP | movd xmm0, ecx
| 1 | | | | | | 1.0 | | | CP | pshufb xmm0, xmm2
| 2^ | | 1.0 | 0.5 0.5 | 0.5 0.5 | | | | | CP | paddb xmm0, xmmword ptr [0x80492c0]
| 1 | | | | | | 1.0 | | | CP | pshufb xmm1, xmm0
| 1 | | | | | | 1.0 | | | | movd xmm0, eax
| 1 | | | | | | 1.0 | | | | pshufb xmm0, xmm2
| 1 | | 1.0 | | | | | | | CP | psubusb xmm1, xmm0
| 0* | | | | | | | | | CP | movdqa xmm0, xmm1
| 1 | | | | | | 1.0 | | | CP | punpcklbw xmm0, xmm2
| 1 | | | | | | 1.0 | | | | punpckhbw xmm1, xmm2
| 2^ | 1.0 | | 0.5 0.5 | 0.5 0.5 | | | | | CP | pmaddwd xmm0, xmmword ptr [0x80492d0]
| 2^ | 1.0 | | 0.5 0.5 | 0.5 0.5 | | | | | | pmaddwd xmm1, xmmword ptr [0x80492d0]
| 3 | | 1.0 | | | | 2.0 | | | CP | phaddd xmm0, xmm1
| 3^ | 2.0 | | 0.5 0.5 | 0.5 0.5 | | | | | CP | pmulld xmm0, xmmword ptr [0x80492e0]
| 1 | | | | | | 1.0 | | | CP | pshufd xmm1, xmm0, 0xee
| 1 | | 1.0 | | | | | | | CP | paddd xmm0, xmm1
| 1 | | | | | | 1.0 | | | CP | pshufd xmm1, xmm0, 0x55
| 1 | | 1.0 | | | | | | | CP | paddd xmm0, xmm1
| 1 | 1.0 | | | | | | | | CP | movd eax, xmm0
| 1 | | | | | | | 1.0 | | CP | add eax, edx
| 1 | | | | | | | 1.0 | | CP | xor eax, edx
Total Num Of Uops: 51
针对Haswell 32位处理器的英特尔IACA延迟分析结果:
Latency Analysis Report
---------------------------
Latency: 64 Cycles
N - port number or number of cycles resource conflict caused delay, DV - Divider pipe (on port 0)
D - Data fetch pipe (on ports 2 and 3), CP - on a critical path
F - Macro Fusion with the previous instruction occurred
* - instruction micro-ops not bound to a port
^ - Micro Fusion happened
# - ESP Tracking sync uop was issued
@ - Intel(R) AVX to Intel(R) SSE code switch, dozens of cycles penalty is expected
! - instruction not supported, was not accounted in Analysis
The Resource delay is counted since all the sources of the instructions are ready
and until the needed resource becomes available
| Inst | Resource Delay In Cycles | |
| Num | 0 - DV | 1 | 2 - D | 3 - D | 4 | 5 | 6 | 7 | FE | |
-------------------------------------------------------------------------
| 0 | | | | | | | | | | | xor eax, eax
| 1 | | | | | | | | | | | xor ecx, ecx
| 2 | | | | | | | | | | | xor edx, edx
| 3 | | | | | | | | | | | dec eax
| 4 | | | | | | | | | 1 | CP | mov bl, byte ptr [esi]
| 5 | | | | | | | | | | CP | cmp bl, 0x2d
| 6 | | | | | | | | | | CP | cmovz edx, eax
| 7 | | | | | | | | | | CP | cmp bl, 0x2b
| 8 | | | | | | | 1 | | | CP | cmovz ecx, eax
| 9 | | | | | | | | | | CP | sub esi, edx
| 10 | | | | | | | | | | CP | sub esi, ecx
| 11 | | | | | | | | | 3 | | xor eax, eax
| 12 | | | | | | | | | | CP | inc esi
| 13 | | | | | | | | | | | cmp byte ptr [esi-0x1], 0x30
| 14 | | | | | | | | | | | jz 0xfffffffb
| 15 | | | | | | | | | | | cmp byte ptr [esi-0x1], 0x0
| 16 | | | | | | | | | | | jz 0x8b
| 17 | | | | | | | | | | CP | dec esi
| 18 | | | | | | | | | 4 | | movdqa xmm0, xmmword ptr [0x80492f0]
| 19 | | | | | | | | | | CP | movdqu xmm1, xmmword ptr [esi]
| 20 | | | | | | | | | 5 | | pxor xmm2, xmm2
| 21 | | | | | | | | | | CP | pcmpistri xmm0, xmm1, 0x14
| 22 | | | | | | | | | | | jo 0x6e
| 23 | | | | | | | | | 6 | | mov al, 0x30
| 24 | | | | | | | | | | CP | sub ecx, 0x10
| 25 | | | | | | | | | | CP | movd xmm0, ecx
| 26 | | | | | | | | | | CP | pshufb xmm0, xmm2
| 27 | | | | | | | | | 7 | CP | paddb xmm0, xmmword ptr [0x80492c0]
| 28 | | | | | | | | | | CP | pshufb xmm1, xmm0
| 29 | | | | | | 1 | | | | | movd xmm0, eax
| 30 | | | | | | 1 | | | | | pshufb xmm0, xmm2
| 31 | | | | | | | | | | CP | psubusb xmm1, xmm0
| 32 | | | | | | | | | | CP | movdqa xmm0, xmm1
| 33 | | | | | | | | | | CP | punpcklbw xmm0, xmm2
| 34 | | | | | | | | | | | punpckhbw xmm1, xmm2
| 35 | | | | | | | | | 9 | CP | pmaddwd xmm0, xmmword ptr [0x80492d0]
| 36 | | | | | | | | | 9 | | pmaddwd xmm1, xmmword ptr [0x80492d0]
| 37 | | | | | | | | | | CP | phaddd xmm0, xmm1
| 38 | | | | | | | | | 10 | CP | pmulld xmm0, xmmword ptr [0x80492e0]
| 39 | | | | | | | | | | CP | pshufd xmm1, xmm0, 0xee
| 40 | | | | | | | | | | CP | paddd xmm0, xmm1
| 41 | | | | | | | | | | CP | pshufd xmm1, xmm0, 0x55
| 42 | | | | | | | | | | CP | paddd xmm0, xmm1
| 43 | | | | | | | | | | CP | movd eax, xmm0
| 44 | | | | | | | | | | CP | add eax, edx
| 45 | | | | | | | | | | CP | xor eax, edx
Resource Conflict on Critical Paths:
-----------------------------------------------------------------
| Port | 0 - DV | 1 | 2 - D | 3 - D | 4 | 5 | 6 | 7 |
-----------------------------------------------------------------
| Cycles | 0 0 | 0 | 0 0 | 0 0 | 0 | 0 | 1 | 0 |
-----------------------------------------------------------------
List Of Delays On Critical Paths
-------------------------------
6 --> 8 1 Cycles Delay On Port6
Peter Cordes在评论中建议的另一种处理是用imul
替换最后两个add xor
指令。OpCodes的这种集中可能会更好。不幸的是,IACA不支持该指令并抛出一个!-不支持的指令,在Analysis
评论中没有考虑。尽管如此,尽管我喜欢OpCodes的减少和从(2uops,2c延迟)减少到(1 uop,3c延迟-“更差的延迟,但在AMD上仍然有一个m-op”),但我更喜欢让实现者选择哪种方式。我还没有检查以下代码是否足以解析任何数字。只是为了完整性而提到,其他部分的代码修改可能是必要的(尤其是处理正数)。
替代方案可能是将最后两行替换为:
...
/* negate if negative number */
imul eax, edx
FINISH:
/* EAX is return (u)int value */
问题内容: 如何监视进程的SIMD(SSE,AVX,AVX2,AVX-512)指令使用量?例如,可用于监视常规CPU使用率,但不能监视SIMD指令使用率。 问题答案: 我认为,对 所有 SIMD指令(不仅仅是FP数学)进行计数的唯一可靠方法是动态检测(例如,通过诸如Intel PIN / SDE之类的方法)。 请参阅如何通过获取指令类型明细来表征工作负载?以及如何确定在C程序中执行的x86机器指令
问题内容: 我很快将使用Java的log4j类来创建项目。但是我不认为我对此有任何了解。希望有人能启发我这个小问题。 问题答案: Log4j基本上接受您要输出的任何语句,让您为其分配“严重性”级别(警告,错误,严重等),并以多种方式将其注销。(对于文件,字节流等,有很多附加选项可用。) 这是对log4j的简短介绍。 http://www.developer.com/open/article.php
在北网的优化指南中,针对英特尔GPU的OpenCL开源实现 工作组大小应大于16,并且是16的倍数。 Gen上有两个可能的SIMD车道是8或16。为了不浪费SIMD车道,我们需要遵循这个规则。 英特尔处理器图形Gen7.5的计算架构中也提到: 对于Gen7。基于5个产品,每个EU有7个线程,总计28 KB的通用寄存器文件(GRF)。 。。。 在Gen7上。5计算架构,大多数SPMD编程模型都采用这
我正在使用AudioTrack制作Android MP3播放器应用程序。我不能在任何情况下使用MediaPlayer,我使用16位单声道和44100Hz。我听说方法可以使用类实现,但我不太清楚这一点。请帮帮我。
有几种使用SIMD指令优化HOG描述符计算的尝试:OpenCV、Dlib和SIMD。它们都使用标量代码将结果幅值添加到HOG直方图中: 在那里,大小的值取决于实现,但一般意义相同。 我知道使用SIMD计算直方图的问题并没有简单有效的解决方案。但在这种情况下,我们有小尺寸(18)的直方图。它能帮助SIMD优化吗?
我有许多使用. NET会员资格和表单身份验证的遗留. NET Framework Web应用程序。它们都有自己的登录页面,但都在同一个域(例如mycompany.com)上,共享ASP. NET会员数据库并共享机器密钥,以便它们能够识别彼此的cookie以实现SSO;也就是说,如果您登录到一个应用程序,那么您将全部登录到它们,与注销相同。 现在我希望所有未来的应用程序都能在中开发。NET核心,并使