Add LOCAL_TRIG_FNS option.

If this is defined during building, we use our own sinf/cosf/atan2f
functions during shading. This is set automatically if CLUSTER is defined
too, so this should remove the cross-platform differences seen during
cluster rendering.

1 files changed, 150 insertions, 4 deletions

diff --git a/draw/draw_mesh.c b/draw/draw_mesh.c
index 7579ad60..025ec366 100644
--- a/draw/draw_mesh.c
+++ b/draw/draw_mesh.c
@@ -1,5 +1,151 @@
 #include "fitz-internal.h"
 
+#ifdef CLUSTER
+#define LOCAL_TRIG_FNS
+#endif
+
+#ifdef LOCAL_TRIG_FNS
+/*
+ * Trig functions
+ */
+static float
+my_atan_table[258] =
+{
+0.0000000000f, 0.00390623013f,0.00781234106f,0.0117182136f,
+0.0156237286f, 0.0195287670f, 0.0234332099f, 0.0273369383f,
+0.0312398334f, 0.0351417768f, 0.0390426500f, 0.0429423347f,
+0.0468407129f, 0.0507376669f, 0.0546330792f, 0.0585268326f,
+0.0624188100f, 0.0663088949f, 0.0701969711f, 0.0740829225f,
+0.0779666338f, 0.0818479898f, 0.0857268758f, 0.0896031775f,
+0.0934767812f, 0.0973475735f, 0.1012154420f, 0.1050802730f,
+0.1089419570f, 0.1128003810f, 0.1166554350f, 0.1205070100f,
+0.1243549950f, 0.1281992810f, 0.1320397620f, 0.1358763280f,
+0.1397088740f, 0.1435372940f, 0.1473614810f, 0.1511813320f,
+0.1549967420f, 0.1588076080f, 0.1626138290f, 0.1664153010f,
+0.1702119250f, 0.1740036010f, 0.1777902290f, 0.1815717110f,
+0.1853479500f, 0.1891188490f, 0.1928843120f, 0.1966442450f,
+0.2003985540f, 0.2041471450f, 0.2078899270f, 0.2116268090f,
+0.2153577000f, 0.2190825110f, 0.2228011540f, 0.2265135410f,
+0.2302195870f, 0.2339192060f, 0.2376123140f, 0.2412988270f,
+0.2449786630f, 0.2486517410f, 0.2523179810f, 0.2559773030f,
+0.2596296290f, 0.2632748830f, 0.2669129880f, 0.2705438680f,
+0.2741674510f, 0.2777836630f, 0.2813924330f, 0.2849936890f,
+0.2885873620f, 0.2921733830f, 0.2957516860f, 0.2993222020f,
+0.3028848680f, 0.3064396190f, 0.3099863910f, 0.3135251230f,
+0.3170557530f, 0.3205782220f, 0.3240924700f, 0.3275984410f,
+0.3310960770f, 0.3345853220f, 0.3380661230f, 0.3415384250f,
+0.3450021770f, 0.3484573270f, 0.3519038250f, 0.3553416220f,
+0.3587706700f, 0.3621909220f, 0.3656023320f, 0.3690048540f,
+0.3723984470f, 0.3757830650f, 0.3791586690f, 0.3825252170f,
+0.3858826690f, 0.3892309880f, 0.3925701350f, 0.3959000740f,
+0.3992207700f, 0.4025321870f, 0.4058342930f, 0.4091270550f,
+0.4124104420f, 0.4156844220f, 0.4189489670f, 0.4222040480f,
+0.4254496370f, 0.4286857080f, 0.4319122350f, 0.4351291940f,
+0.4383365600f, 0.4415343100f, 0.4447224240f, 0.4479008790f,
+0.4510696560f, 0.4542287350f, 0.4573780990f, 0.4605177290f,
+0.4636476090f, 0.4667677240f, 0.4698780580f, 0.4729785980f,
+0.4760693300f, 0.4791502430f, 0.4822213240f, 0.4852825630f,
+0.4883339510f, 0.4913754780f, 0.4944071350f, 0.4974289160f,
+0.5004408130f, 0.5034428210f, 0.5064349340f, 0.5094171490f,
+0.5123894600f, 0.5153518660f, 0.5183043630f, 0.5212469510f,
+0.5241796290f, 0.5271023950f, 0.5300152510f, 0.5329181980f,
+0.5358112380f, 0.5386943730f, 0.5415676050f, 0.5444309400f,
+0.5472843810f, 0.5501279330f, 0.5529616020f, 0.5557853940f,
+0.5585993150f, 0.5614033740f, 0.5641975770f, 0.5669819340f,
+0.5697564530f, 0.5725211450f, 0.5752760180f, 0.5780210840f,
+0.5807563530f, 0.5834818390f, 0.5861975510f, 0.5889035040f,
+0.5915997100f, 0.5942861830f, 0.5969629370f, 0.5996299860f,
+0.6022873460f, 0.6049350310f, 0.6075730580f, 0.6102014430f,
+0.6128202020f, 0.6154293530f, 0.6180289120f, 0.6206188990f,
+0.6231993300f, 0.6257702250f, 0.6283316020f, 0.6308834820f,
+0.6334258830f, 0.6359588250f, 0.6384823300f, 0.6409964180f,
+0.6435011090f, 0.6459964250f, 0.6484823880f, 0.6509590190f,
+0.6534263410f, 0.6558843770f, 0.6583331480f, 0.6607726790f,
+0.6632029930f, 0.6656241120f, 0.6680360620f, 0.6704388650f,
+0.6728325470f, 0.6752171330f, 0.6775926450f, 0.6799591110f,
+0.6823165550f, 0.6846650020f, 0.6870044780f, 0.6893350100f,
+0.6916566220f, 0.6939693410f, 0.6962731940f, 0.6985682070f,
+0.7008544080f, 0.7031318220f, 0.7054004770f, 0.7076604000f,
+0.7099116190f, 0.7121541600f, 0.7143880520f, 0.7166133230f,
+0.7188300000f, 0.7210381110f, 0.7232376840f, 0.7254287490f,
+0.7276113330f, 0.7297854640f, 0.7319511710f, 0.7341084830f,
+0.7362574290f, 0.7383980370f, 0.7405303370f, 0.7426543560f,
+0.7447701260f, 0.7468776740f, 0.7489770290f, 0.7510682220f,
+0.7531512810f, 0.7552262360f, 0.7572931160f, 0.7593519510f,
+0.7614027700f, 0.7634456020f, 0.7654804790f, 0.7675074280f,
+0.7695264800f, 0.7715376650f, 0.7735410110f, 0.7755365500f,
+0.7775243100f, 0.7795043220f, 0.7814766150f, 0.7834412190f,
+0.7853981630f, 0.7853981630f /* Extended by 1 for interpolation */
+};
+
+float my_sinf(float x)
+{
+	float x2, xn;
+	int i;
+	/* Map x into the -PI to PI range. We could do this using:
+	 * x = fmodf(x, (float)(2.0 * M_PI));
+	 * but that's C99, and seems to misbehave with negative numbers
+	 * on some platforms. */
+	x -= (float)M_PI;
+	i = x / (float)(2.0f * M_PI);
+	x -= i * (float)(2.0f * M_PI);
+	if (x < 0.0f)
+		x += (float)(2.0f * M_PI);
+	x -= (float)M_PI;
+	if (x <= (float)(-M_PI/2.0))
+		x = -(float)M_PI-x;
+	else if (x >= (float)(M_PI/2.0))
+		x = (float)M_PI-x;
+	x2 = x*x;
+	xn = x*x2/6.0f;
+	x -= xn;
+	xn *= x2/20.0f;
+	x += xn;
+	xn *= x2/42.0f;
+	x -= xn;
+	xn *= x2/72.0f;
+	x += xn;
+	return x;
+}
+
+float my_atan2f(float o, float a)
+{
+	int negate = 0, flip = 0, i;
+	float r, s;
+	if (o == 0.0f)
+	{
+		if (a > 0)
+			return 0.0f;
+		else
+			return (float)M_PI;
+	}
+	if (o < 0)
+		o = -o, negate = 1;
+	if (a < 0)
+		a = -a, flip = 1;
+	if (o < a)
+		i = (int)(65536.0f*o/a + 0.5f);
+	else
+		i = (int)(65536.0f*a/o + 0.5f);
+	r = my_atan_table[i>>8];
+	s = my_atan_table[(i>>8)+1];
+	r += (s-r)*(i&255)/256.0f;
+	if (o >= a)
+		r = (float)(M_PI/2.0f) - r;
+	if (flip)
+		r = (float)M_PI - r;
+	if (negate)
+		r = -r;
+	return r;
+}
+
+#define my_cosf(x) my_sinf(((float)(M_PI/2.0f)) + (x))
+#else
+#define my_sinf(x) sinf(x)
+#define my_cosf(x) cosf(x)
+#define my_atan2f(x,y) atan2f(x,y)
+#endif
+
 /*
  * polygon clipping
  */
@@ -340,8 +486,8 @@ fz_paint_quad(fz_pixmap *pix,
 static fz_point
 fz_point_on_circle(fz_point p, float r, float theta)
 {
-	p.x = p.x + cosf(theta) * r;
-	p.y = p.y + sinf(theta) * r;
+	p.x = p.x + my_cosf(theta) * r;
+	p.y = p.y + my_sinf(theta) * r;
 
 	return p;
 }
@@ -362,7 +508,7 @@ fz_paint_linear(fz_shade *shade, fz_matrix ctm, fz_pixmap *dest, fz_bbox bbox)
 	p1.y = shade->mesh[4];
 	p1 = fz_transform_point(ctm, p1);
 
-	theta = atan2f(p1.y - p0.y, p1.x - p0.x);
+	theta = my_atan2f(p1.y - p0.y, p1.x - p0.x);
 	theta += (float)M_PI * 0.5f;
 
 	v0 = fz_point_on_circle(p0, HUGENUM, theta);
@@ -405,7 +551,7 @@ fz_paint_annulus(fz_matrix ctm,
 	float theta, step;
 	int i;
 
-	theta = atan2f(p1.y - p0.y, p1.x - p0.x);
+	theta = my_atan2f(p1.y - p0.y, p1.x - p0.x);
 	step = (float)M_PI * 2 / RADSEGS;
 
 	for (i = 0; i < RADSEGS / 2; i++)