问题
My new PC has a Core i7 CPU and I am running my benchmarks, including newer versions that use AVX instructions. I have installed Visual Studio 2013 to use a newer compiler, as my last one could not fully compile for full SSE SIMD operation. Below is some code used in one of my benchmarks (MPMFLOPS), and compile and link commands used. Tests were run with the first command to use SSE instructions. When xtra is 16 or less, the benchmark produces 24.4 GFLOPS. CPU runs at 3.9 GHz, so result is good at 6.25 calculations per cycle, compared with maximum of four multiples and four adds. Increasing xtra to greater than 16, produces 2.6 GFLOPS. Reducing words to much lower values makes to difference to speed.
/*
Visual Studio 2013
C/C++ Optimizing Compiler Version 18.00.21005.1 for x64
cl /O2 /Oi /MD /W4 /TP /EHsc /Zi /Fa /c mflops.c
cl /O2 /Oi /MD /W4 /TP /EHsc /Zi /Fa /arch:AVX /c mflops.c
link /LARGEADDRESSAWARE mflops.obj CPUasm.obj asmtimeavx.obj
BUFFEROVERFLOWU.LIB
link Includes CPUID information with identification of AVX and timer
*/
#include <stdio.h>
#include <stdlib.h>
#include "asmtimeavx.h"
#include <windows.h>
#include <time.h>
#include <malloc.h>
int main()
{
float *x;
float a = 0.000020f;
float b = 0.999980f;
float c = 0.000011f;
float d = 1.000011f;
float e = 0.000012f;
float f = 0.999992f;
float mflops;
int i, j;
int xtra = 16; // 24447 MFLOPS, > 16 around 2600 MFLOPS
int words = 1000000;
x = (float *)_aligned_malloc(words * 4, 16);
for (i = 0; i < words; i++) x[i] = 0.999999f;
start_time();
for (j = 0; j < xtra; j++)
{
for (i = 0; i < words; i++)
{
x[i] = (x[i] + a)*b - (x[i] + c)*d + (x[i] + e)*f;
}
}
end_time();
mflops = (float)words * (float)xtra * 8.0f / 1000000.0f / (float)secs;
printf("%18.8f, %18.8f, %10.7f secs, %8.2f mflops\n\n", x[0],
x[words-1], secs, mflops);
_aligned_free(x);
return 0;
}
Following shows assembly code generated, where such as mulps is full SIMD with four values in 128 bit registers, with mulss using one floating point number (SISD).
Windows SSE
words = 1000000
xtra = 16 xtra > 16
call start_time
npad 10
$LL6@main:
mov rcx, rsi
mov edx, 125000
npad 8
$LL3@main:
movups xmm1, XMMWORD PTR [rcx]
add rcx, 32 movaps xmm1, xmm2
movaps xmm2, xmm1 movaps xmm0, xmm2
movaps xmm0, xmm1 addss xmm2, xmm8
addps xmm1, xmm10 addss xmm0, xmm6
addps xmm2, xmm6 addss xmm1, xmm4
addps xmm0, xmm8 dec rax
mulps xmm1, xmm11 mulss xmm2, xmm9
mulps xmm2, xmm7 mulss xmm0, xmm7
mulps xmm0, xmm9 mulss xmm1, xmm5
subps xmm2, xmm0 subss xmm1, xmm0
addps xmm2, xmm1 addss xmm1, xmm2
movups XMMWORD PTR [rcx-32], xmm2 movaps xmm2, xmm1
movups xmm1, XMMWORD PTR [rcx-16] movaps xmm0, xmm1
movaps xmm2, xmm1 addss xmm1, xmm8
movaps xmm0, xmm1 addss xmm2, xmm4
addps xmm2, xmm6 addss xmm0, xmm6
addps xmm0, xmm8 mulss xmm1, xmm9
addps xmm1, xmm10 mulss xmm0, xmm7
mulps xmm2, xmm7 mulss xmm2, xmm5
mulps xmm0, xmm9 subss xmm2, xmm0
mulps xmm1, xmm11 addss xmm2, xmm1
subps xmm2, xmm0 movaps xmm3, xmm2
addps xmm2, xmm1 movaps xmm0, xmm2
movups XMMWORD PTR [rcx-16], xmm2 addss xmm2, xmm8
dec rdx
jne SHORT $LL3@main More of the same
dec rbx Loop 82 lines
jne SHORT $LL6@main
call end_time
Next I compiled the program to use AVX instructions, but this produced the same speed as using SSE. Following is assembly code generated. Also shown is much faster code included in that generated via Linux (GCC with Ubuntu 14.04). Linux speeds were reduced somewhat with a greater number of words but the parameters shown produced SISD type results via Windows.
Note that Windows code is using 128 bit xmm registers but Linux employs 256 bit ymm registers. Does anyone have an explanation of what is going on or suggestions to improve performance using this sample program.
Windows AVX Linux AVX
Only uses xmm registers Part can use ymm registers
Othe parts imcludes xmm
xtra = 1000000; words = 10000; 44271 MFLOPS
call start_time xtra = 10000000; words = 1000; 45653 MFLOPS
npad 6 SSE
$LL6@main: xtra = 1000000; words = 10000; 24492 MFLOPS
mov rcx, rsi
mov edx, 125000
npad 8
$LL3@main: .L24:
vmovups xmm0, XMMWORD PTR [rcx] vmovaps (%rcx,%rax), %ymm6
lea rcx, QWORD PTR [rcx+32] addl $1, %edx
vmovups xmm5, xmm0 vaddps %ymm5, %ymm6, %ymm13
vaddps xmm1, xmm0, xmm6 vaddps %ymm3, %ymm6, %ymm7
vaddps xmm0, xmm0, xmm8 vaddps %ymm1, %ymm6, %ymm6
vmulps xmm2, xmm0, xmm9 vmulps %ymm4, %ymm13, %ymm13
vmulps xmm3, xmm1, xmm7 vmulps %ymm2, %ymm7, %ymm7
vsubps xmm4, xmm3, xmm2 vmulps %ymm0, %ymm6, %ymm6
vaddps xmm1, xmm5, xmm10 vsubps %ymm7, %ymm13, %ymm7
vmulps xmm0, xmm1, xmm11 vaddps %ymm6, %ymm7, %ymm6
vaddps xmm2, xmm4, xmm0 vmovaps %ymm6, (%rcx,%rax)
vmovups XMMWORD PTR [rcx-32], xmm2 addq $32, %rax
vmovups xmm0, XMMWORD PTR [rcx-16] cmpl %edx, %esi
vaddps xmm1, xmm0, xmm6 ja .L24
vmovups xmm5, xmm0
vaddps xmm0, xmm0, xmm8
vmulps xmm2, xmm0, xmm9
vmulps xmm3, xmm1, xmm7
vaddps xmm1, xmm5, xmm10
vsubps xmm4, xmm3, xmm2
vmulps xmm0, xmm1, xmm11
vaddps xmm2, xmm4, xmm0
vmovups XMMWORD PTR [rcx-16], xmm2
dec rdx
jne SHORT $LL3@main
dec rbx
jne $LL6@main
call end_time
来源：https://stackoverflow.com/questions/24143828/how-can-i-improve-performance-compiling-for-sse-and-avx