added _mesa_inv_sqrtf() and INV_SQRTF() (Josh Vanderhoof)

9 files changed, 152 insertions, 30 deletions

diff --git a/src/mesa/main/imports.c b/src/mesa/main/imports.c
index e9d579e7640..8474ed4abcf 100644
--- a/src/mesa/main/imports.c
+++ b/src/mesa/main/imports.c
@@ -1,4 +1,4 @@
-/* $Id: imports.c,v 1.32 2003/03/01 01:50:21 brianp Exp $ */
+/* $Id: imports.c,v 1.33 2003/03/04 16:33:53 brianp Exp $ */
 
 /*
  * Mesa 3-D graphics library
@@ -346,6 +346,7 @@ _mesa_sqrtf( float x )
     * then reconstruct the result back into a float
     */
    num.i = ((sqrttab[num.i >> 16]) << 16) | ((e + 127) << 23);
+
    return num.f;
 #else
    return (float) _mesa_sqrtd((double) x);
@@ -353,6 +354,115 @@ _mesa_sqrtf( float x )
 }
 
 
+/**
+ inv_sqrt - A single precision 1/sqrt routine for IEEE format floats.
+ written by Josh Vanderhoof, based on newsgroup posts by James Van Buskirk
+ and Vesa Karvonen.
+*/
+float
+_mesa_inv_sqrtf(float n)
+{
+#if defined(USE_IEEE) && !defined(DEBUG)
+        float r0, x0, y0;
+        float r1, x1, y1;
+        float r2, x2, y2;
+#if 0 /* not used, see below -BP */
+        float r3, x3, y3;
+#endif
+        union { float f; unsigned int i; } u;
+        unsigned int magic;
+
+        /*
+         Exponent part of the magic number -
+
+         We want to:
+         1. subtract the bias from the exponent,
+         2. negate it
+         3. divide by two (rounding towards -inf)
+         4. add the bias back
+
+         Which is the same as subtracting the exponent from 381 and dividing
+         by 2.
+
+         floor(-(x - 127) / 2) + 127 = floor((381 - x) / 2)
+        */
+
+        magic = 381 << 23;
+
+        /*
+         Significand part of magic number -
+
+         With the current magic number, "(magic - u.i) >> 1" will give you:
+
+         for 1 <= u.f <= 2: 1.25 - u.f / 4
+         for 2 <= u.f <= 4: 1.00 - u.f / 8
+
+         This isn't a bad approximation of 1/sqrt.  The maximum difference from
+         1/sqrt will be around .06.  After three Newton-Raphson iterations, the
+         maximum difference is less than 4.5e-8.  (Which is actually close
+         enough to make the following bias academic...)
+
+         To get a better approximation you can add a bias to the magic
+         number.  For example, if you subtract 1/2 of the maximum difference in
+         the first approximation (.03), you will get the following function:
+
+         for 1 <= u.f <= 2:    1.22 - u.f / 4
+         for 2 <= u.f <= 3.76: 0.97 - u.f / 8
+         for 3.76 <= u.f <= 4: 0.72 - u.f / 16
+         (The 3.76 to 4 range is where the result is < .5.)
+
+         This is the closest possible initial approximation, but with a maximum
+         error of 8e-11 after three NR iterations, it is still not perfect.  If
+         you subtract 0.0332281 instead of .03, the maximum error will be
+         2.5e-11 after three NR iterations, which should be about as close as
+         is possible.
+
+         for 1 <= u.f <= 2:    1.2167719 - u.f / 4
+         for 2 <= u.f <= 3.73: 0.9667719 - u.f / 8
+         for 3.73 <= u.f <= 4: 0.7167719 - u.f / 16
+
+        */
+
+        magic -= (int)(0.0332281 * (1 << 25));
+
+        u.f = n;
+        u.i = (magic - u.i) >> 1;
+
+        /*
+         Instead of Newton-Raphson, we use Goldschmidt's algorithm, which
+         allows more parallelism.  From what I understand, the parallelism
+         comes at the cost of less precision, because it lets error
+         accumulate across iterations.
+        */
+        x0 = 1.0f;
+        y0 = 0.5f * n;
+        r0 = u.f;
+
+        x1 = x0 * r0;
+        y1 = y0 * r0 * r0;
+        r1 = 1.5f - y1;
+
+        x2 = x1 * r1;
+        y2 = y1 * r1 * r1;
+        r2 = 1.5f - y2;
+
+#if 1
+        return x2 * r2;  /* we can stop here, and be conformant -BP */
+#else
+        x3 = x2 * r2;
+        y3 = y2 * r2 * r2;
+        r3 = 1.5f - y3;
+
+        return x3 * r3;
+#endif
+#elif defined(XFree86LOADER) && defined(IN_MODULE)
+        return 1.0F / xf86sqrt(n);
+#else
+        return 1.0F / sqrt(n);
+#endif
+}
+
+
 double
 _mesa_pow(double x, double y)
 {
diff --git a/src/mesa/main/imports.h b/src/mesa/main/imports.h
index 354819f43f0..73c3b6eb427 100644
--- a/src/mesa/main/imports.h
+++ b/src/mesa/main/imports.h
@@ -1,4 +1,4 @@
-/* $Id: imports.h,v 1.16 2003/03/03 21:44:39 brianp Exp $ */
+/* $Id: imports.h,v 1.17 2003/03/04 16:33:53 brianp Exp $ */
 
 /*
  * Mesa 3-D graphics library
@@ -169,6 +169,16 @@ float asm_sqrt (float x);
 
 
 /***
+ *** INV_SQRTF: single-precision inverse square root
+ ***/
+#if 0
+#define INV_SQRTF(X) _mesa_inv_sqrt(X)
+#else
+#define INV_SQRTF(X) (1.0F / SQRTF(X))  /* this is faster on a P4 */
+#endif
+
+
+/***
  *** LOG2: Log base 2 of float
  ***/
 #ifdef USE_IEEE
@@ -588,6 +598,9 @@ _mesa_sqrtd(double x);
 extern float
 _mesa_sqrtf(float x);
 
+extern float
+_mesa_inv_sqrtf(float x);
+
 extern double
 _mesa_pow(double x, double y);
 
diff --git a/src/mesa/main/macros.h b/src/mesa/main/macros.h
index 0be95109957..63bd02bdff8 100644
--- a/src/mesa/main/macros.h
+++ b/src/mesa/main/macros.h
@@ -1,4 +1,4 @@
-/* $Id: macros.h,v 1.31 2003/03/01 01:50:21 brianp Exp $ */
+/* $Id: macros.h,v 1.32 2003/03/04 16:33:54 brianp Exp $ */
 
 /*
  * Mesa 3-D graphics library
@@ -572,7 +572,7 @@ do {						\
 do {						\
    GLfloat len = (GLfloat) LEN_SQUARED_3FV(V);	\
    if (len) {					\
-      len = (GLfloat) (1.0 / SQRTF(len));	\
+      len = INV_SQRTF(len);			\
       (V)[0] = (GLfloat) ((V)[0] * len);	\
       (V)[1] = (GLfloat) ((V)[1] * len);	\
       (V)[2] = (GLfloat) ((V)[2] * len);	\
diff --git a/src/mesa/main/nvvertexec.c b/src/mesa/main/nvvertexec.c
index 72ebcffca9a..07029273544 100644
--- a/src/mesa/main/nvvertexec.c
+++ b/src/mesa/main/nvvertexec.c
@@ -1,4 +1,4 @@
-/* $Id: nvvertexec.c,v 1.2 2003/03/01 01:50:22 brianp Exp $ */
+/* $Id: nvvertexec.c,v 1.3 2003/03/04 16:33:55 brianp Exp $ */
 
 /*
  * Mesa 3-D graphics library
@@ -380,7 +380,7 @@ _mesa_exec_vertex_program(GLcontext *ctx, const struct vertex_program *program)
             {
                GLfloat t[4];
                fetch_vector1( &inst->SrcReg[0], machine, t );
-               t[0] = (float) (1.0 / sqrt(fabs(t[0])));
+               t[0] = INV_SQRTF(FABSF(t[0]));
                t[1] = t[2] = t[3] = t[0];
                store_vector4( &inst->DstReg, machine, t );
             }
diff --git a/src/mesa/math/m_norm_tmp.h b/src/mesa/math/m_norm_tmp.h
index 2da7bbd6761..5f24f5a2a27 100644
--- a/src/mesa/math/m_norm_tmp.h
+++ b/src/mesa/math/m_norm_tmp.h
@@ -1,4 +1,4 @@
-/* $Id: m_norm_tmp.h,v 1.13 2003/03/01 01:50:24 brianp Exp $ */
+/* $Id: m_norm_tmp.h,v 1.14 2003/03/04 16:34:01 brianp Exp $ */
 
 /*
  * Mesa 3-D graphics library
@@ -69,10 +69,10 @@ TAG(transform_normalize_normals)( const GLmatrix *mat,
 	 {
 	    GLdouble len = tx*tx + ty*ty + tz*tz;
 	    if (len > 1e-20) {
-	       GLdouble scale = 1.0F / SQRTF(len);
-	       out[i][0] = (GLfloat) (tx * scale);
-	       out[i][1] = (GLfloat) (ty * scale);
-	       out[i][2] = (GLfloat) (tz * scale);
+	       GLfloat scale = INV_SQRTF(len);
+	       out[i][0] = tx * scale;
+	       out[i][1] = ty * scale;
+	       out[i][2] = tz * scale;
 	    }
 	    else {
 	       out[i][0] = out[i][1] = out[i][2] = 0;
@@ -136,10 +136,10 @@ TAG(transform_normalize_normals_no_rot)( const GLmatrix *mat,
 	 {
 	    GLdouble len = tx*tx + ty*ty + tz*tz;
 	    if (len > 1e-20) {
-	       GLdouble scale = 1.0F / SQRTF(len);
-	       out[i][0] = (GLfloat) (tx * scale);
-	       out[i][1] = (GLfloat) (ty * scale);
-	       out[i][2] = (GLfloat) (tz * scale);
+	       GLfloat scale = INV_SQRTF(len);
+	       out[i][0] = tx * scale;
+	       out[i][1] = ty * scale;
+	       out[i][2] = tz * scale;
 	    }
 	    else {
 	       out[i][0] = out[i][1] = out[i][2] = 0;
@@ -323,10 +323,10 @@ TAG(normalize_normals)( const GLmatrix *mat,
 	 const GLfloat x = from[0], y = from[1], z = from[2];
 	 GLdouble len = x * x + y * y + z * z;
 	 if (len > 1e-50) {
-	    len = 1.0F / SQRTF(len);
-	    out[i][0] = (GLfloat) (x * len);
-	    out[i][1] = (GLfloat) (y * len);
-	    out[i][2] = (GLfloat) (z * len);
+	    len = INV_SQRTF(len);
+	    out[i][0] = x * len;
+	    out[i][1] = y * len;
+	    out[i][2] = z * len;
 	 }
 	 else {
 	    out[i][0] = x;
diff --git a/src/mesa/swrast/s_aalinetemp.h b/src/mesa/swrast/s_aalinetemp.h
index 529da481103..9850ff9ac5a 100644
--- a/src/mesa/swrast/s_aalinetemp.h
+++ b/src/mesa/swrast/s_aalinetemp.h
@@ -1,4 +1,4 @@
-/* $Id: s_aalinetemp.h,v 1.22 2003/02/21 21:00:27 brianp Exp $ */
+/* $Id: s_aalinetemp.h,v 1.23 2003/03/04 16:34:02 brianp Exp $ */
 
 /*
  * Mesa 3-D graphics library
@@ -132,7 +132,7 @@ NAME(line)(GLcontext *ctx, const SWvertex *v0, const SWvertex *v1)
    line.y1 = v1->win[1];
    line.dx = line.x1 - line.x0;
    line.dy = line.y1 - line.y0;
-   line.len = (GLfloat) sqrt(line.dx * line.dx + line.dy * line.dy);
+   line.len = SQRTF(line.dx * line.dx + line.dy * line.dy);
    line.halfWidth = 0.5F * ctx->Line.Width;
 
    if (line.len == 0.0 || IS_INF_OR_NAN(line.len))
diff --git a/src/mesa/swrast/s_nvfragprog.c b/src/mesa/swrast/s_nvfragprog.c
index 97f65c04794..147b5248d4c 100644
--- a/src/mesa/swrast/s_nvfragprog.c
+++ b/src/mesa/swrast/s_nvfragprog.c
@@ -1,4 +1,4 @@
-/* $Id: s_nvfragprog.c,v 1.5 2003/03/01 01:50:26 brianp Exp $ */
+/* $Id: s_nvfragprog.c,v 1.6 2003/03/04 16:34:03 brianp Exp $ */
 
 /*
  * Mesa 3-D graphics library
@@ -582,8 +582,7 @@ execute_program(GLcontext *ctx, const struct fragment_program *program)
             {
                GLfloat a[4], result[4];
                fetch_vector1( &inst->SrcReg[0], machine, a );
-               result[0] = result[1] = result[2] = result[3]
-                  = 1.0F / SQRTF(a[0]);
+               result[0] = result[1] = result[2] = result[3] = INV_SQRTF(a[0]);
                store_vector4( inst, machine, result );
             }
             break;
diff --git a/src/mesa/swrast/s_span.c b/src/mesa/swrast/s_span.c
index d4081518036..2106b87c5b6 100644
--- a/src/mesa/swrast/s_span.c
+++ b/src/mesa/swrast/s_span.c
@@ -1,4 +1,4 @@
-/* $Id: s_span.c,v 1.56 2003/03/01 01:50:26 brianp Exp $ */
+/* $Id: s_span.c,v 1.57 2003/03/04 16:34:03 brianp Exp $ */
 
 /*
  * Mesa 3-D graphics library
@@ -309,8 +309,8 @@ compute_lambda(GLfloat dsdx, GLfloat dsdy, GLfloat dtdx, GLfloat dtdy,
    GLfloat dvdx = texH * ((t + dtdx) / (q + dqdx) - t * invQ);
    GLfloat dudy = texW * ((s + dsdy) / (q + dqdy) - s * invQ);
    GLfloat dvdy = texH * ((t + dtdy) / (q + dqdy) - t * invQ);
-   GLfloat x = sqrt(dudx * dudx + dvdx * dvdx);
-   GLfloat y = sqrt(dudy * dudy + dvdy * dvdy);
+   GLfloat x = SQRTF(dudx * dudx + dvdx * dvdx);
+   GLfloat y = SQRTF(dudy * dudy + dvdy * dvdy);
    GLfloat rho = MAX2(x, y);
    GLfloat lambda = LOG2(rho);
    return lambda;
diff --git a/src/mesa/tnl/t_vb_texgen.c b/src/mesa/tnl/t_vb_texgen.c
index 74ad17356b0..f2454e6ffbc 100644
--- a/src/mesa/tnl/t_vb_texgen.c
+++ b/src/mesa/tnl/t_vb_texgen.c
@@ -1,4 +1,4 @@
-/* $Id: t_vb_texgen.c,v 1.17 2003/03/01 01:50:27 brianp Exp $ */
+/* $Id: t_vb_texgen.c,v 1.18 2003/03/04 16:34:04 brianp Exp $ */
 
 /*
  * Mesa 3-D graphics library
@@ -113,7 +113,7 @@ static void build_m3( GLfloat f[][3], GLfloat m[],
       fz = f[i][2] = u[2] - norm[2] * two_nu;
       m[i] = fx * fx + fy * fy + (fz + 1.0F) * (fz + 1.0F);
       if (m[i] != 0.0F) {
-	 m[i] = 0.5F / SQRTF(m[i]);
+	 m[i] = 0.5F * _mesa_inv_sqrtf(m[i]);
       }
    }
 }
@@ -142,7 +142,7 @@ static void build_m2( GLfloat f[][3], GLfloat m[],
       fz = f[i][2] = u[2] - norm[2] * two_nu;
       m[i] = fx * fx + fy * fy + (fz + 1.0F) * (fz + 1.0F);
       if (m[i] != 0.0F) {
-	 m[i] = 0.5F / SQRTF(m[i]);
+	 m[i] = 0.5F * _mesa_inv_sqrtf(m[i]);
       }
    }
 }