AVX scalar operations are much faster

Question

I test the following simple function

void mul(double *a, double *b) {
  for (int i = 0; i<N; i++) a[i] *= b[i];
}

with very large arrays so that it is memory bandwidth bound. The test code I use is below. When I compile with -O2 it takes 1.7 seconds. When I compile with -O2 -mavx it takes only 1.0 seconds. The non vex-encoded scalar operations are 70% slower! Why is this?

Here is the the assembly for -O2 and -O2 -mavx.

https://godbolt.org/g/w4p60f

System: [email protected] (Skylake) 32 GB mem, Ubuntu 16.10, GCC 6.3

Test code

//gcc -O2 -fopenmp test.c
//or
//gcc -O2 -mavx -fopenmp test.c
#include <string.h>
#include <stdio.h>
#include <x86intrin.h>
#include <omp.h>

#define N 1000000
#define R 1000

void mul(double *a, double *b) {
  for (int i = 0; i<N; i++) a[i] *= b[i];
}

int main() {
  double *a = (double*)_mm_malloc(sizeof *a * N, 32);
  double *b = (double*)_mm_malloc(sizeof *b * N, 32);

  //b must be initialized to get the correct bandwidth!!!
  memset(a, 1, sizeof *a * N);
  memset(b, 1, sizeof *b * N);

  double dtime;
  const double mem = 3*sizeof(double)*N*R/1024/1024/1024;
  const double maxbw = 34.1;
  dtime = -omp_get_wtime();
  for(int i=0; i<R; i++) mul(a,b);
  dtime += omp_get_wtime();
  printf("time %.2f s, %.1f GB/s, efficency %.1f%%\n", dtime, mem/dtime, 100*mem/dtime/maxbw);

  _mm_free(a), _mm_free(b);
}

Answer 1

The problem is related to a dirty upper half of an AVX register after calling omp_get_wtime(). This is a problem particularly for Skylake processors.

The first time I read about this problem was here. Since then other people have observed this problem: here and here.

Using gdb I found that omp_get_wtime() calls clock_gettime. I rewrote my code to use clock_gettime() and I see the same problem.

void fix_avx() { __asm__ __volatile__ ( "vzeroupper" : : : ); }
void fix_sse() { }
void (*fix)();

double get_wtime() {
  struct timespec time;
  clock_gettime(CLOCK_MONOTONIC, &time);
  #ifndef  __AVX__ 
  fix();
  #endif
  return time.tv_sec + 1E-9*time.tv_nsec;
}

void dispatch() {
  fix = fix_sse;
  #if defined(__INTEL_COMPILER)
  if (_may_i_use_cpu_feature (_FEATURE_AVX)) fix = fix_avx;
  #else
  #if defined(__GNUC__) && !defined(__clang__)
  __builtin_cpu_init();
  #endif
  if(__builtin_cpu_supports("avx")) fix = fix_avx;
  #endif
}

Stepping through code with gdb I see that the first time clock_gettime is called it calls _dl_runtime_resolve_avx(). I believe the problem is in this function based on this comment. This function appears to only be called the first time clock_gettime is called.

With GCC the problem goes away using //__asm__ __volatile__ ( "vzeroupper" : : : ); after the first call with clock_gettime however with Clang (using clang -O2 -fno-vectorize since Clang vectorizes even at -O2) it only goes away using it after every call to clock_gettime.

Here is the code I used to test this (with GCC 6.3 and Clang 3.8)

#include <string.h>
#include <stdio.h>
#include <x86intrin.h>
#include <time.h>

void fix_avx() { __asm__ __volatile__ ( "vzeroupper" : : : ); }
void fix_sse() { }
void (*fix)();

double get_wtime() {
  struct timespec time;
  clock_gettime(CLOCK_MONOTONIC, &time);
  #ifndef  __AVX__ 
  fix();
  #endif
  return time.tv_sec + 1E-9*time.tv_nsec;
}

void dispatch() {
  fix = fix_sse;
  #if defined(__INTEL_COMPILER)
  if (_may_i_use_cpu_feature (_FEATURE_AVX)) fix = fix_avx;
  #else
  #if defined(__GNUC__) && !defined(__clang__)
  __builtin_cpu_init();
  #endif
  if(__builtin_cpu_supports("avx")) fix = fix_avx;
  #endif
}

#define N 1000000
#define R 1000

void mul(double *a, double *b) {
  for (int i = 0; i<N; i++) a[i] *= b[i];
}

int main() {
  dispatch();
  const double mem = 3*sizeof(double)*N*R/1024/1024/1024;
  const double maxbw = 34.1;

  double *a = (double*)_mm_malloc(sizeof *a * N, 32);
  double *b = (double*)_mm_malloc(sizeof *b * N, 32);

  //b must be initialized to get the correct bandwidth!!!
  memset(a, 1, sizeof *a * N);
  memset(b, 1, sizeof *b * N);

  double dtime;
  //dtime = get_wtime(); // call once to fix GCC
  //printf("%f\n", dtime);
  //fix = fix_sse;

  dtime = -get_wtime();
  for(int i=0; i<R; i++) mul(a,b);
  dtime += get_wtime();
  printf("time %.2f s, %.1f GB/s, efficency %.1f%%\n", dtime, mem/dtime, 100*mem/dtime/maxbw);

  _mm_free(a), _mm_free(b);
}

---

If I disable lazy function call resolution with -z now (e.g. clang -O2 -fno-vectorize -z now foo.c) then Clang only needs __asm__ __volatile__ ( "vzeroupper" : : : ); after the first call to clock_gettime just like GCC.

I expected that with -z now I would only need __asm__ __volatile__ ( "vzeroupper" : : : ); right after main() but I still need it after the first call to clock_gettime.