SSE SIMD enhanced code 4x slower than regular code

SSE SIMD enhanced code 4x slower than regular code

[Boris Hollas]

Hello,

I have a function iter1 that iterates a sequence of complex numbers. I
redesigned this function, using SSE intrinsics such as _mm_mul_pd, to obtain
iter0. Nonetheless, iter0 is 4x slower than iter1:

iter0 (with SSE intrinsics):
$ gcc -O -march=core2 t.c && time ./a.out
257829745

real	0m7.912s
user	0m7.908s
sys	0m0.000s

iter1 (w/o SSE intrinsics):
$ gcc -O -march=core2 t.c && time ./a.out
257829745

real	0m2.075s
user	0m2.076s
sys	0m0.000s

The size of a.out ist 7.1K in both cases. I use gcc version 4.4.5 and the
CPU is an Intel Core 2 Duo.

The code is below. iter0 and iter1 give the same numerical results.


#include <pmmintrin.h>
#include <stdio.h>
#define sqr(x) ((x)*(x))

typedef union {
  __m128d m;
  double v[2]; // v[0] low, v[1] up
} v2df;

int iter0(v2df z, v2df c, int n, int bound) {
  v2df z2, z2r, z2r_addsub, z_;
  z2.m = _mm_mul_pd(z.m, z.m);  // z_re^2, z_im^2
  z2r.v[1] = z2.v[0];
  z2r.v[0] = z2.v[1];
  z2r_addsub.m = _mm_addsub_pd(z2r.m, z2.m); // z_re^2 + z_im^2, z_re^2 -
z_im^2

  if(z2r_addsub.v[1] > 4.0 || n == bound) return n;
  else {
	z_.v[1] = z2r_addsub.v[0];
	z_.v[0] = 2.0 * z.v[1] * z.v[0];
	z_.m = _mm_add_pd(z_.m, c.m); // z_re^2 - z_im^2 + c_re, 2 * z_re * z_im +
c_im
	return iter0(z_, c, n+1, bound);
  }
}


int iter1(double z_re, double z_im, double c_re, double c_im, int n, int
bound) {
  double zre2 = sqr(z_re);
  double zim2 = sqr(z_im);

  if(zre2 + zim2 > 4.0 || n == bound) return n;
  else return iter1(zre2 - zim2 + c_re, 2.0 * z_re * z_im + c_im, c_re,
c_im, n+1, bound);
}

#define sse

int main() {
  v2df z, c;
  long n = 0;
  z.v[1] = 0.0; z.v[0] = 0.0;

  for(c.v[1] = -2.0; c.v[1] < 1.0; c.v[1] += 3.0/1000.0) {
    for(c.v[0] = -1.0; c.v[0] < 1.0; c.v[0] += 2.0/1000.0) {
#ifdef sse
  	  n += iter0(z, c, 0, 1000);
#else
	  n += iter1(0.0, 0.0, c.v[1], c.v[0], 0, 1000);
#endif
    }
  }
  printf("%ld\n", n);
  return 0;
}



-- 
View this message in context: http://old.nabble.com/SSE-SIMD-enhanced-code-4x-slower-than-regular-code-tp33159404p33159404.html
Sent from the gcc - Help mailing list archive at Nabble.com.

Re: SSE SIMD enhanced code 4x slower than regular code

[Marc Glisse]

On Tue, 17 Jan 2012, Boris Hollas wrote:

> I have a function iter1 that iterates a sequence of complex numbers. I
> redesigned this function, using SSE intrinsics such as _mm_mul_pd, to obtain
> iter0. Nonetheless, iter0 is 4x slower than iter1:

That is not surprising at all and happens to most codes using double when 
people try converting them to SSE.

> $ gcc -O -march=core2 t.c && time ./a.out

Maybe use at least -O2 ?

> The size of a.out ist 7.1K in both cases. I use gcc version 4.4.5 and the
> CPU is an Intel Core 2 Duo.

You may want to try newer versions of gcc (note the plural, results 
between 4.4, 4.5, 4.6 and the future 4.7 can vary a lot, and not always in 
the right direction).

> #include <pmmintrin.h>
> #include <stdio.h>
> #define sqr(x) ((x)*(x))
>
> typedef union {
>  __m128d m;
>  double v[2]; // v[0] low, v[1] up
> } v2df;
>
> int iter0(v2df z, v2df c, int n, int bound) {
>  v2df z2, z2r, z2r_addsub, z_;
>  z2.m = _mm_mul_pd(z.m, z.m);  // z_re^2, z_im^2
>  z2r.v[1] = z2.v[0];
>  z2r.v[0] = z2.v[1];

You may want to try _mm_shuffle_pd or __builtin_shuffle.

>  z2r_addsub.m = _mm_addsub_pd(z2r.m, z2.m); // z_re^2 + z_im^2, z_re^2 -
> z_im^2
>
>  if(z2r_addsub.v[1] > 4.0 || n == bound) return n;
>  else {
> 	z_.v[1] = z2r_addsub.v[0];
> 	z_.v[0] = 2.0 * z.v[1] * z.v[0];
> 	z_.m = _mm_add_pd(z_.m, c.m); // z_re^2 - z_im^2 + c_re, 2 * z_re * z_im +
> c_im
> 	return iter0(z_, c, n+1, bound);
>  }
> }

Did you take a look at the generated code (use flag -S and read the 
generated t.s)? Going back and forth between packed and unpacked through a 
union often generates plenty of mov instructions. If you manually use 
_mm_cvtsd_f64 and _mm_unpackhi_pd you may be able to save a bit. Note that 
with the latest gcc, you can use the [] notation directly on your __m128d.

> int iter1(double z_re, double z_im, double c_re, double c_im, int n, int
> bound) {
>  double zre2 = sqr(z_re);
>  double zim2 = sqr(z_im);
>
>  if(zre2 + zim2 > 4.0 || n == bound) return n;
>  else return iter1(zre2 - zim2 + c_re, 2.0 * z_re * z_im + c_im, c_re,
> c_im, n+1, bound);
> }
>
> #define sse
>
> int main() {
>  v2df z, c;
>  long n = 0;
>  z.v[1] = 0.0; z.v[0] = 0.0;
>
>  for(c.v[1] = -2.0; c.v[1] < 1.0; c.v[1] += 3.0/1000.0) {
>    for(c.v[0] = -1.0; c.v[0] < 1.0; c.v[0] += 2.0/1000.0) {
> #ifdef sse
>  	  n += iter0(z, c, 0, 1000);
> #else
> 	  n += iter1(0.0, 0.0, c.v[1], c.v[0], 0, 1000);
> #endif
>    }
>  }
>  printf("%ld\n", n);
>  return 0;
> }

I'd be surprised if you managed any gain on this thanks to __m128d.

-- 
Marc Glisse

Re: SSE SIMD enhanced code 4x slower than regular code

[Boris Hollas]

>Maybe use at least -O2 ?

no difference.

> You may want to try _mm_shuffle_pd or __builtin_shuffle.

indeed, this reduces the runtime by 2 s. Still 3x slower than iter1.

>Did you take a look at the generated code (use flag -S and read the 
>generated t.s)? Going back and forth between packed and unpacked through a 
>union often generates plenty of mov instructions. If you manually use 
>_mm_cvtsd_f64 and _mm_unpackhi_pd you may be able to save a bit. Note that 
>with the latest gcc, you can use the [] notation directly on your __m128d.

yes, there are lots of move-instructions. It seems this adds a lot of
runtime.

> I'd be surprised if you managed any gain on this thanks to __m128d.

well, maybe SIMD isn't very well suited for these calculations. So, I'll use
iter1 until I have a better idea.

thanks for your hints!

-Boris
-- 
View this message in context: http://old.nabble.com/SSE-SIMD-enhanced-code-4x-slower-than-regular-code-tp33159404p33161396.html
Sent from the gcc - Help mailing list archive at Nabble.com.