diff mbox series

fix inaccuracy of j0f for x >= 2^127 when sin(x)+cos(x) is tiny

Message ID mwmu3knale.fsf@tomate.loria.fr
State New
Headers show
Series fix inaccuracy of j0f for x >= 2^127 when sin(x)+cos(x) is tiny | expand

Commit Message

Paul Zimmermann July 27, 2020, 5:15 p.m. UTC
Hi,

before the patch below, the maximum ulp error for j0 in the whole binary32
range is 6177902 ulps (for x = 3.153646966e+38).

After this patch, it is 900691 ulps (for x = 2.404825449e+00).

The patch fixes the case where x >= 2^127 and tiny sin(x)+cos(x).

Large remaining errors are due to a cancellation in another branch of the code.

Paul

PS: the same method can be applied to j1 and y1.
PS2: this can wait for 2.33 of course.

From 6b731f36b1a5badf4704645d0dda40957cedd0db Mon Sep 17 00:00:00 2001
From: Paul Zimmermann <Paul.Zimmermann@inria.fr>
Date: Mon, 27 Jul 2020 19:01:18 +0200
Subject: [PATCH] fix inaccuracy of j0f for x >= 2^127 when sin(x)+cos(x) is
 tiny

---
 sysdeps/ieee754/flt-32/e_j0f.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

Comments

Joseph Myers July 27, 2020, 9:35 p.m. UTC | #1
On Mon, 27 Jul 2020, Paul Zimmermann wrote:

> +                   float x0 = 3.153646966e+38f;
> +                   float y = x - x0; /* exact */
> +                   /* sin(y) = sin(x)*cos(x0)-cos(x)*sin(x0) */
> +                   z = __sinf (y);
> +                   float eps = 8.17583368e-8f;
> +                   /* cos(x0) ~ -sin(x0) + eps */
> +                   z += eps * __cosf (x);
> +                   /* now z ~ (sin(x)-cos(x))*cos(x0) */
> +                   float cosx0 = -0.707106740f;

In new code we generally prefer to use hex float constants in such cases 
where a specific floating-point value is wanted.
Paul Zimmermann July 28, 2020, 8:23 a.m. UTC | #2
> In new code we generally prefer to use hex float constants in such cases 
> where a specific floating-point value is wanted.

thank you Joseph. Here is a new version. The maximal error for x >= 2^127
is now 4 ulps (attained for x=1.740713465e+38).

Total: errors=4220511 (0.10%) errors2=393216 maxerr=4 ulp(s)

Paul

From 6b731f36b1a5badf4704645d0dda40957cedd0db Mon Sep 17 00:00:00 2001
From: Paul Zimmermann <Paul.Zimmermann@inria.fr>
Date: Mon, 27 Jul 2020 19:01:18 +0200
Subject: [PATCH 1/2] fix inaccuracy of j0f for x >= 2^127 when sin(x)+cos(x)
 is tiny

---
 sysdeps/ieee754/flt-32/e_j0f.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/sysdeps/ieee754/flt-32/e_j0f.c b/sysdeps/ieee754/flt-32/e_j0f.c
index c89b9f2688..f85d8a59e0 100644
--- a/sysdeps/ieee754/flt-32/e_j0f.c
+++ b/sysdeps/ieee754/flt-32/e_j0f.c
@@ -56,6 +56,22 @@ __ieee754_j0f(float x)
 		    if ((s*c)<zero) cc = z/ss;
 		    else	    ss = z/cc;
 		}
+                else {
+                  /* we subtract (exactly) a value x0 such that cos(x0)+sin(x0)
+                     is very near from 0, and use the identity
+                     sin(x-x0) = sin(x)*cos(x0)-cos(x)*sin(x0) to get
+                     sin(x) + cos(x) with extra accuracy */
+                   float x0 = 3.153646966e+38f;
+                   float y = x - x0; /* exact */
+                   /* sin(y) = sin(x)*cos(x0)-cos(x)*sin(x0) */
+                   z = __sinf (y);
+                   float eps = 8.17583368e-8f;
+                   /* cos(x0) ~ -sin(x0) + eps */
+                   z += eps * __cosf (x);
+                   /* now z ~ (sin(x)-cos(x))*cos(x0) */
+                   float cosx0 = -0.707106740f;
+                   cc = z / cosx0;
+                }
 	/*
 	 * j0(x) = 1/sqrt(pi) * (P(0,x)*cc - Q(0,x)*ss) / sqrt(x)
 	 * y0(x) = 1/sqrt(pi) * (P(0,x)*ss + Q(0,x)*cc) / sqrt(x)
Andreas Schwab July 28, 2020, 9:19 a.m. UTC | #3
On Jul 28 2020, Paul Zimmermann wrote:

> +                  /* we subtract (exactly) a value x0 such that cos(x0)+sin(x0)
> +                     is very near from 0, and use the identity

Did you mean "near to"?

Andreas.
Paul Zimmermann July 28, 2020, 10:50 a.m. UTC | #4
Dear Andreas,

yes thanks. Sorry my english is not perfect.

Paul

From 6b731f36b1a5badf4704645d0dda40957cedd0db Mon Sep 17 00:00:00 2001
From: Paul Zimmermann <Paul.Zimmermann@inria.fr>
Date: Mon, 27 Jul 2020 19:01:18 +0200
Subject: [PATCH 1/3] fix inaccuracy of j0f for x >= 2^127 when sin(x)+cos(x)
 is tiny

---
 sysdeps/ieee754/flt-32/e_j0f.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/sysdeps/ieee754/flt-32/e_j0f.c b/sysdeps/ieee754/flt-32/e_j0f.c
index c89b9f2688..f85d8a59e0 100644
--- a/sysdeps/ieee754/flt-32/e_j0f.c
+++ b/sysdeps/ieee754/flt-32/e_j0f.c
@@ -56,6 +56,22 @@ __ieee754_j0f(float x)
 		    if ((s*c)<zero) cc = z/ss;
 		    else	    ss = z/cc;
 		}
+                else {
+                  /* we subtract (exactly) a value x0 such that cos(x0)+sin(x0)
+                     is very near from 0, and use the identity
+                     sin(x-x0) = sin(x)*cos(x0)-cos(x)*sin(x0) to get
+                     sin(x) + cos(x) with extra accuracy */
+                   float x0 = 3.153646966e+38f;
+                   float y = x - x0; /* exact */
+                   /* sin(y) = sin(x)*cos(x0)-cos(x)*sin(x0) */
+                   z = __sinf (y);
+                   float eps = 8.17583368e-8f;
+                   /* cos(x0) ~ -sin(x0) + eps */
+                   z += eps * __cosf (x);
+                   /* now z ~ (sin(x)-cos(x))*cos(x0) */
+                   float cosx0 = -0.707106740f;
+                   cc = z / cosx0;
+                }
 	/*
 	 * j0(x) = 1/sqrt(pi) * (P(0,x)*cc - Q(0,x)*ss) / sqrt(x)
 	 * y0(x) = 1/sqrt(pi) * (P(0,x)*ss + Q(0,x)*cc) / sqrt(x)
Adhemerval Zanella Netto July 28, 2020, 6:09 p.m. UTC | #5
On 28/07/2020 07:50, Paul Zimmermann wrote:
>        Dear Andreas,
> 
> yes thanks. Sorry my english is not perfect.
> 
> Paul

Could you send v2 patch with all the fixes indicated by Joseph and Andreas
(this change from a change format is confusing)?  Also please fix the
indentation issue and the open brackets on next line. 

I also think this fix should also add an entry on math/auto-libm-test-out-y0
that exercises this code path and with a check if the ULPs file require
some adjustments as well. 

> 
> From 6b731f36b1a5badf4704645d0dda40957cedd0db Mon Sep 17 00:00:00 2001
> From: Paul Zimmermann <Paul.Zimmermann@inria.fr>
> Date: Mon, 27 Jul 2020 19:01:18 +0200
> Subject: [PATCH 1/3] fix inaccuracy of j0f for x >= 2^127 when sin(x)+cos(x)
>  is tiny
> 
> ---
>  sysdeps/ieee754/flt-32/e_j0f.c | 16 ++++++++++++++++
>  1 file changed, 16 insertions(+)
> 
> diff --git a/sysdeps/ieee754/flt-32/e_j0f.c b/sysdeps/ieee754/flt-32/e_j0f.c
> index c89b9f2688..f85d8a59e0 100644
> --- a/sysdeps/ieee754/flt-32/e_j0f.c
> +++ b/sysdeps/ieee754/flt-32/e_j0f.c
> @@ -56,6 +56,22 @@ __ieee754_j0f(float x)
>  		    if ((s*c)<zero) cc = z/ss;
>  		    else	    ss = z/cc;
>  		}
> +                else {
> +                  /* we subtract (exactly) a value x0 such that cos(x0)+sin(x0)
> +                     is very near from 0, and use the identity
> +                     sin(x-x0) = sin(x)*cos(x0)-cos(x)*sin(x0) to get
> +                     sin(x) + cos(x) with extra accuracy */
> +                   float x0 = 3.153646966e+38f;
> +                   float y = x - x0; /* exact */
> +                   /* sin(y) = sin(x)*cos(x0)-cos(x)*sin(x0) */
> +                   z = __sinf (y);
> +                   float eps = 8.17583368e-8f;
> +                   /* cos(x0) ~ -sin(x0) + eps */
> +                   z += eps * __cosf (x);
> +                   /* now z ~ (sin(x)-cos(x))*cos(x0) */
> +                   float cosx0 = -0.707106740f;
> +                   cc = z / cosx0;
> +                }
>  	/*
>  	 * j0(x) = 1/sqrt(pi) * (P(0,x)*cc - Q(0,x)*ss) / sqrt(x)
>  	 * y0(x) = 1/sqrt(pi) * (P(0,x)*ss + Q(0,x)*cc) / sqrt(x)
>
diff mbox series

Patch

diff --git a/sysdeps/ieee754/flt-32/e_j0f.c b/sysdeps/ieee754/flt-32/e_j0f.c
index c89b9f2688..f85d8a59e0 100644
--- a/sysdeps/ieee754/flt-32/e_j0f.c
+++ b/sysdeps/ieee754/flt-32/e_j0f.c
@@ -56,6 +56,22 @@  __ieee754_j0f(float x)
 		    if ((s*c)<zero) cc = z/ss;
 		    else	    ss = z/cc;
 		}
+                else {
+                  /* we subtract (exactly) a value x0 such that cos(x0)+sin(x0)
+                     is very near from 0, and use the identity
+                     sin(x-x0) = sin(x)*cos(x0)-cos(x)*sin(x0) to get
+                     sin(x) + cos(x) with extra accuracy */
+                   float x0 = 3.153646966e+38f;
+                   float y = x - x0; /* exact */
+                   /* sin(y) = sin(x)*cos(x0)-cos(x)*sin(x0) */
+                   z = __sinf (y);
+                   float eps = 8.17583368e-8f;
+                   /* cos(x0) ~ -sin(x0) + eps */
+                   z += eps * __cosf (x);
+                   /* now z ~ (sin(x)-cos(x))*cos(x0) */
+                   float cosx0 = -0.707106740f;
+                   cc = z / cosx0;
+                }
 	/*
 	 * j0(x) = 1/sqrt(pi) * (P(0,x)*cc - Q(0,x)*ss) / sqrt(x)
 	 * y0(x) = 1/sqrt(pi) * (P(0,x)*ss + Q(0,x)*cc) / sqrt(x)