[Beignet] [PATCH] [PATCH]GBE: improve precision of rootn

Zhigang Gong zhigang.gong at linux.intel.com
Thu Feb 20 19:05:17 PST 2014


LGTM, will push latter. Thanks.

On Tue, Jan 14, 2014 at 11:04:57AM +0800, Lv Meng wrote:
> 
> Signed-off-by: Lv Meng <meng.lv at intel.com>
> ---
>  backend/src/ocl_stdlib.tmpl.h |  257 ++++++++++++++++++++++++++++++++++++++++-
>  1 file changed, 256 insertions(+), 1 deletion(-)
> 
> diff --git a/backend/src/ocl_stdlib.tmpl.h b/backend/src/ocl_stdlib.tmpl.h
> index 4a4d321..6a88df6 100755
> --- a/backend/src/ocl_stdlib.tmpl.h
> +++ b/backend/src/ocl_stdlib.tmpl.h
> @@ -2642,6 +2642,232 @@ INLINE_OVERLOADABLE float __gen_ocl_internal_fdim(float x, float y) {
>      return y;
>    return x > y ? (x - y) : +0.f;
>  }
> +
> +INLINE_OVERLOADABLE float __gen_ocl_internal_pow(float x, float y) {
> +  float z,ax,z_h,z_l,p_h,p_l;
> +  float y1,t1,t2,r,s,sn,t,u,v,w;
> +  int i,j,k,yisint,n;
> +  int hx,hy,ix,iy,is;
> +  float bp[2],dp_h[2],dp_l[2],
> +  zero    =  0.0,
> +  one	=  1.0,
> +  two	=  2.0,
> +  two24	=  16777216.0,	/* 0x4b800000 */
> +  huge	=  1.0e30,
> +  tiny    =  1.0e-30,
> +  /* poly coefs for (3/2)*(log(x)-2s-2/3*s**3 */
> +  L1  =  6.0000002384e-01, /* 0x3f19999a */
> +  L2  =  4.2857143283e-01, /* 0x3edb6db7 */
> +  L3  =  3.3333334327e-01, /* 0x3eaaaaab */
> +  L4  =  2.7272811532e-01, /* 0x3e8ba305 */
> +  L5  =  2.3066075146e-01, /* 0x3e6c3255 */
> +  L6  =  2.0697501302e-01, /* 0x3e53f142 */
> +  P1   =  1.6666667163e-01, /* 0x3e2aaaab */
> +  P2   = -2.7777778450e-03, /* 0xbb360b61 */
> +  P3   =  6.6137559770e-05, /* 0x388ab355 */
> +  P4   = -1.6533901999e-06, /* 0xb5ddea0e */
> +  P5   =  4.1381369442e-08, /* 0x3331bb4c */
> +  lg2  =  6.9314718246e-01, /* 0x3f317218 */
> +  lg2_h  =  6.93145752e-01, /* 0x3f317200 */
> +  lg2_l  =  1.42860654e-06, /* 0x35bfbe8c */
> +  ovt =  4.2995665694e-08, /* -(128-log2(ovfl+.5ulp)) */
> +  cp    =  9.6179670095e-01, /* 0x3f76384f =2/(3ln2) */
> +  cp_h  =  9.6179199219e-01, /* 0x3f763800 =head of cp */
> +  cp_l  =  4.7017383622e-06, /* 0x369dc3a0 =tail of cp_h */
> +  ivln2    =  1.4426950216e+00, /* 0x3fb8aa3b =1/ln2 */
> +  ivln2_h  =  1.4426879883e+00, /* 0x3fb8aa00 =16b 1/ln2*/
> +  ivln2_l  =  7.0526075433e-06; /* 0x36eca570 =1/ln2 tail*/
> +  bp[0] = 1.0,bp[1] = 1.5,
> +  dp_h[0] = 0.0,dp_h[1] = 5.84960938e-01,
> +  dp_l[0] = 0.0,dp_l[1] = 1.56322085e-06;
> +  GEN_OCL_GET_FLOAT_WORD(hx,x);
> +  GEN_OCL_GET_FLOAT_WORD(hy,y);
> +  ix = hx&0x7fffffff;  iy = hy&0x7fffffff;
> +  if (ix < 0x00800000) {	   /* x < 2**-126  */
> +    ix = 0;/* Gen does not support subnormal number now */
> +  }
> +  if (iy < 0x00800000) {	  /* y < 2**-126  */
> +    iy = 0;/* Gen does not support subnormal number now */
> +  }
> +   /* y==zero: x**0 = 1 */
> +  if(iy==0) return one;
> +  if(hx==0x3f800000) return one;
> +  /* +-NaN return x+y */
> +  if(ix > 0x7f800000 || iy > 0x7f800000)
> +    return (x+0.0f)+y+(0.0f);
> +  /* determine if y is an odd int when x < 0
> +     * yisint = 0	... y is not an integer
> +     * yisint = 1	... y is an odd int
> +     * yisint = 2	... y is an even int
> +     */
> +  yisint  = 0;
> +  if(hx<0) {
> +    if(iy>=0x4b800000) yisint = 2; /* even integer y */
> +    else if(iy>=0x3f800000) {
> +      k = (iy>>23)-0x7f;	   /* exponent */
> +      j = iy>>(23-k);
> +      if((j<<(23-k))==iy) yisint = 2-(j&1);
> +    }
> +  }
> +  /* special value of y */
> +  if (iy==0x7f800000) {	/* y is +-inf */
> +    if (ix==0x3f800000)
> +      //return  y - y;	/* inf**+-1 is NaN */
> +      return one;
> +    else if (ix > 0x3f800000)/* (|x|>1)**+-inf = inf,0 */
> +      return (hy>=0)? y: zero;
> +    else			/* (|x|<1)**-,+inf = inf,0 */
> +      return (hy<0)?-y: zero;
> +  }
> +  if(iy==0x3f800000) {	/* y is  +-1 */
> +    if(hy<0) return one/x; else return x;
> +  }
> +  if(hy==0x40000000) return x*x; /* y is  2 */
> +  if(hy==0x3f000000) {	/* y is  0.5 */
> +    if(hx>=0)return __gen_ocl_sqrt(x);
> +  }
> +
> +  ax   = __gen_ocl_fabs(x);
> +    /* special value of x */
> +  if(ix==0x7f800000||ix==0||ix==0x3f800000){
> +    z = ax;			/*x is +-0,+-inf,+-1*/
> +    if(hy<0) z = one/z;	/* z = (1/|x|) */
> +    if(hx<0) {
> +      if(((ix-0x3f800000)|yisint)==0) {
> +        z = (z-z)/(z-z); /* (-1)**non-int is NaN */
> +      } else if(yisint==1)
> +        z = -z;		/* (x<0)**odd = -(|x|**odd) */
> +    }
> +    return z;
> +  }
> +  n = ((uint)hx>>31)-1;
> +
> +  /* (x<0)**(non-int) is NaN */
> +  if((n|yisint)==0) return (x-x)/(x-x);
> +
> +  sn = one; /* s (sign of result -ve**odd) = -1 else = 1 */
> +  if((n|(yisint-1))==0) sn = -one;/* (-ve)**(odd int) */
> +
> +  /* |y| is huge */
> +  if(iy>0x4d000000) { /* if |y| > 2**27 */
> +    /* over/underflow if x is not close to one */
> +    if(ix<0x3f7ffff8) return (hy<0)? sn*huge*huge:sn*tiny*tiny;
> +    if(ix>0x3f800007) return (hy>0)? sn*huge*huge:sn*tiny*tiny;
> +    /* now |1-x| is tiny <= 2**-20, suffice to compute
> +          log(x) by x-x^2/2+x^3/3-x^4/4 */
> +    t = ax-1;		/* t has 20 trailing zeros */
> +    w = (t*t)*((float)0.5-t*(0.333333333333f-t*0.25f));
> +    u = ivln2_h*t;	/* ivln2_h has 16 sig. bits */
> +    v = t*ivln2_l-w*ivln2;
> +    t1 = u+v;
> +    GEN_OCL_GET_FLOAT_WORD(is,t1);
> +    GEN_OCL_SET_FLOAT_WORD(t1,is&0xfffff000);
> +    t2 = v-(t1-u);
> +  } else {
> +    float s2,s_h,s_l,t_h,t_l;
> +    n = 0;
> +	/* take care subnormal number */
> +    //if(ix<0x00800000)
> +      //{ax *= two24; n -= 24; GEN_OCL_GET_FLOAT_WORD(ix,ax); }
> +    n  += ((ix)>>23)-0x7f;
> +    j  = ix&0x007fffff;
> +	/* determine interval */
> +    ix = j|0x3f800000;		/* normalize ix */
> +    if(j<=0x1cc471) k=0;	/* |x|<sqrt(3/2) */
> +    else if(j<0x5db3d7) k=1;	/* |x|<sqrt(3)   */
> +    else {k=0;n+=1;ix -= 0x00800000;}
> +    GEN_OCL_SET_FLOAT_WORD(ax,ix);
> +
> +	/* compute s = s_h+s_l = (x-1)/(x+1) or (x-1.5)/(x+1.5) */
> +    u = ax-bp[k];		/* bp[0]=1.0, bp[1]=1.5 */
> +    v = one/(ax+bp[k]);
> +    s = u*v;
> +    s_h = s;
> +    GEN_OCL_GET_FLOAT_WORD(is,s_h);
> +    GEN_OCL_SET_FLOAT_WORD(s_h,is&0xfffff000);
> +    /* t_h=ax+bp[k] High */
> +    is = ((ix>>1)&0xfffff000)|0x20000000;
> +    GEN_OCL_SET_FLOAT_WORD(t_h,is+0x00400000+(k<<21));
> +    t_l = ax - (t_h-bp[k]);
> +    s_l = v*((u-s_h*t_h)-s_h*t_l);
> +    /* compute log(ax) */
> +    s2 = s*s;
> +    r = s2*s2*(L1+s2*(L2+s2*(L3+s2*(L4+s2*(L5+s2*L6)))));
> +    r += s_l*(s_h+s);
> +    s2  = s_h*s_h;
> +    t_h = 3.0f+s2+r;
> +    GEN_OCL_GET_FLOAT_WORD(is,t_h);
> +    GEN_OCL_SET_FLOAT_WORD(t_h,is&0xfffff000);
> +    t_l = r-((t_h-3.0f)-s2);
> +    /* u+v = s*(1+...) */
> +    u = s_h*t_h;
> +    v = s_l*t_h+t_l*s;
> +    /* 2/(3log2)*(s+...) */
> +    p_h = u+v;
> +    GEN_OCL_GET_FLOAT_WORD(is,p_h);
> +    GEN_OCL_SET_FLOAT_WORD(p_h,is&0xfffff000);
> +    p_l = v-(p_h-u);
> +    z_h = cp_h*p_h;		/* cp_h+cp_l = 2/(3*log2) */
> +    z_l = cp_l*p_h+p_l*cp+dp_l[k];
> +    /* log2(ax) = (s+..)*2/(3*log2) = n + dp_h + z_h + z_l */
> +    t = (float)n;
> +    t1 = (((z_h+z_l)+dp_h[k])+t);
> +    GEN_OCL_GET_FLOAT_WORD(is,t1);
> +    GEN_OCL_SET_FLOAT_WORD(t1,is&0xfffff000);
> +    t2 = z_l-(((t1-t)-dp_h[k])-z_h);
> +  }
> +
> +  /* split up y into y1+y2 and compute (y1+y2)*(t1+t2) */
> +  GEN_OCL_GET_FLOAT_WORD(is,y);
> +  GEN_OCL_SET_FLOAT_WORD(y1,is&0xfffff000);
> +  p_l = (y-y1)*t1+y*t2;
> +  p_h = y1*t1;
> +  z = p_l+p_h;
> +  GEN_OCL_GET_FLOAT_WORD(j,z);
> +  if (j>0x43000000)				/* if z > 128 */
> +    return sn*huge*huge;			/* overflow */
> +  else if (j==0x43000000) {			/* if z == 128 */
> +    if(p_l+ovt>z-p_h) return sn*huge*huge;	/* overflow */
> +  }
> +  else if ((j&0x7fffffff)>0x43160000)		/* z <= -150 */
> +    return sn*tiny*tiny;			/* underflow */
> +  else if (j==0xc3160000){			/* z == -150 */
> +    if(p_l<=z-p_h) return sn*tiny*tiny;		/* underflow */
> +  }
> +
> +  /*
> +    * compute 2**(p_h+p_l)
> +    */
> +  i = j&0x7fffffff;
> +  k = (i>>23)-0x7f;
> +  n = 0;
> +  if(i>0x3f000000) {		/* if |z| > 0.5, set n = [z+0.5] */
> +    n = j+(0x00800000>>(k+1));
> +    k = ((n&0x7fffffff)>>23)-0x7f;	/* new k for n */
> +    GEN_OCL_SET_FLOAT_WORD(t,n&~(0x007fffff>>k));
> +    n = ((n&0x007fffff)|0x00800000)>>(23-k);
> +    if(j<0) n = -n;
> +    p_h -= t;
> +  }
> +  t = p_l+p_h;
> +  GEN_OCL_GET_FLOAT_WORD(is,t);
> +  GEN_OCL_SET_FLOAT_WORD(t,is&0xffff8000);
> +  u = t*lg2_h;
> +  v = (p_l-(t-p_h))*lg2+t*lg2_l;
> +  z = u+v;
> +  w = v-(z-u);
> +  t  = z*z;
> +  t1  = z - t*(P1+t*(P2+t*(P3+t*(P4+t*P5))));
> +  r  = (z*t1)/(t1-two)-(w+z*w);
> +  z  = one-(r-z);
> +  GEN_OCL_GET_FLOAT_WORD(j,z);
> +  j += (n<<23);
> +  if((j>>23)<=0) z = __gen_ocl_scalbnf(z,n);	/* subnormal output */
> +  else GEN_OCL_SET_FLOAT_WORD(z,j);
> +  return sn*z;
> +}
> +
> +
>  INLINE_OVERLOADABLE float hypot(float x, float y) {
>    //return __gen_ocl_sqrt(x*x + y*y);
>    float a,b,an,bn,cn;
> @@ -2772,7 +2998,36 @@ INLINE_OVERLOADABLE float pown(float x, int n) {
>    return powr(x, n);
>  }
>  INLINE_OVERLOADABLE float rootn(float x, int n) {
> -  return powr(x, 1.f / n);
> +  float ax,re;
> +  int sign = 0;
> +  if( n == 0 )return NAN;
> +  //rootn ( x, n )  returns a NaN for x < 0 and n is even.
> +  if( x < 0 && 0 == (n&1) )
> +    return NAN;
> +  if( x == 0.0 ){
> +    switch( n & 0x80000001 ){
> +      //rootn ( +-0,  n ) is +0 for even n > 0.
> +      case 0:
> +        return 0.0f;
> +      //rootn ( +-0,  n ) is +-0 for odd n > 0.
> +      case 1:
> +        return x;
> +      //rootn ( +-0,  n ) is +inf for even n < 0.
> +      case 0x80000000:
> +        return INFINITY;
> +
> +      //rootn ( +-0,  n ) is +-inf for odd n < 0.
> +      case 0x80000001:
> +        return __gen_ocl_internal_copysign(INFINITY, x);
> +    }
> +  }
> +  ax = __gen_ocl_fabs(x);
> +  if(x <0.0f && (n&1))
> +    sign = 1;
> +  re = __gen_ocl_internal_pow(ax,1.f/n);
> +  if(sign)
> +    re = -re;
> +  return re;
>  }
>  
>  /////////////////////////////////////////////////////////////////////////////
> -- 
> 1.7.10.4
> 
> _______________________________________________
> Beignet mailing list
> Beignet at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/beignet


More information about the Beignet mailing list