use builtin floor / ceil

get rid of FAST_PSEUDO_FLOOR, add VIPS_FLOOR and VIPS_CEIL, use them
everywhere

see https://github.com/jcupitt/libvips/pull/372

ChangeLog

@@ -8,6 +8,7 @@
 - tune vips_shrinkh(), 30% faster [Lovell Fuller]
 - remove SEQ hint from vips_subsample(), fixes cli performance [erdmann]
 - fix double free on attach ICC profile from file in tiff write [erdmann]
+- add VIPS_FLOOR()/VIPS_CEIL(), much faster [Lovell Fuller]
 
 1/1/16 started 8.2.1
 - add a compat stub [Benjamin Gilbert]

libvips/arithmetic/remainder.c

@@ -111,7 +111,7 @@ vips_remainder_build( VipsObject *object )
 		double a = p1[x]; \
 		double b = p2[x]; \
 		\
-		q[x] = b ? a - b * floor (a / b) : -1; \
+		q[x] = b ? a - b * VIPS_FLOOR (a / b) : -1; \
 	} \
 }
 

libvips/arithmetic/round.c

@@ -121,8 +121,8 @@ vips_round_buffer( VipsArithmetic *arithmetic,
 
 	switch( round->round ) {
 	case VIPS_OPERATION_ROUND_RINT:		SWITCH( VIPS_RINT ); break;
-	case VIPS_OPERATION_ROUND_CEIL:		SWITCH( ceil ); break;
+	case VIPS_OPERATION_ROUND_CEIL:		SWITCH( VIPS_CEIL ); break;
-	case VIPS_OPERATION_ROUND_FLOOR:	SWITCH( floor ); break;
+	case VIPS_OPERATION_ROUND_FLOOR:	SWITCH( VIPS_FLOOR ); break;
 
 	default: 
 		g_assert( 0 ); 

libvips/colour/UCS2LCh.c

@@ -150,7 +150,7 @@ vips_col_Lcmc2L( float Lcmc )
 {	
 	int known;
 
-	known = floor( Lcmc * 10.0 );
+	known = VIPS_FLOOR( Lcmc * 10.0 );
 	known = VIPS_CLIP( 0, known, 999 );
 
 	return( LI[known] + 
@@ -172,7 +172,7 @@ vips_col_Ccmc2C( float Ccmc )
 {	
 	int known;
 
-	known = floor( Ccmc * 10.0 );
+	known = VIPS_FLOOR( Ccmc * 10.0 );
 	known = VIPS_CLIP( 0, known, 2999 );
 
 	return( CI[known] + 
@@ -201,7 +201,7 @@ vips_col_Chcmc2h( float C, float hcmc )
 	r = (int) ((C + 1.0) / 2.0);
 	r = VIPS_CLIP( 0, r, 99 ); 
 
-	known = floor( hcmc );
+	known = VIPS_FLOOR( hcmc );
 	known = VIPS_CLIP( 0, known, 359 ); 
 
 	return( hI[r][known] + 

libvips/conversion/cast.c

@@ -246,7 +246,7 @@ vips_cast_start( VipsImage *out, void *a, void *b )
 	OTYPE * restrict q = (OTYPE *) out; \
 	\
 	for( x = 0; x < sz; x++ ) { \
-		ITYPE v = floor( p[x] ); \
+		ITYPE v = VIPS_FLOOR( p[x] ); \
 		\
 		VIPS_CLIP( v, seq ); \
 		\
@@ -261,7 +261,7 @@ vips_cast_start( VipsImage *out, void *a, void *b )
 	OTYPE * restrict q = (OTYPE *) out; \
 	\
 	for( x = 0; x < sz; x++ ) { \
-		ITYPE v = floor( p[0] ); \
+		ITYPE v = VIPS_FLOOR( p[0] ); \
 		p += 2; \
 		\
 		VIPS_CLIP( v, seq ); \

libvips/convolution/im_aconv.c

@@ -316,9 +316,9 @@ boxes_break( Boxes *boxes )
 	 * fixed n-lines which includes any negative parts.
 	 */
 	depth = (max - min) / boxes->n_layers;
-	layers_above = ceil( max / depth );
+	layers_above = VIPS_CEIL( max / depth );
 	depth = max / layers_above;
-	layers_below = floor( min / depth );
+	layers_below = VIPS_FLOOR( min / depth );
 
 	boxes->n_layers = layers_above - layers_below;
 

libvips/foreign/csv.c

@@ -541,8 +541,8 @@ vips__matrix_header( char *whitemap, FILE *fp,
 		vips_error( "mask2vips", "%s", _( "no width / height" ) );
 		return( -1 );
 	}
-	if( floor( header[0] ) != header[0] ||
+	if( VIPS_FLOOR( header[0] ) != header[0] ||
-		floor( header[1] ) != header[1] ) {
+		VIPS_FLOOR( header[1] ) != header[1] ) {
 		vips_error( "mask2vips", "%s", _( "width / height not int" ) );
 		return( -1 );
 	}

libvips/histogram/hist_plot.c

@@ -237,7 +237,7 @@ plot( IMAGE *in, IMAGE *out )
 	if( in->BandFmt == IM_BANDFMT_UCHAR )
 		tsize = 256;
 	else
-		tsize = ceil( max );
+		tsize = VIPS_CEIL( max );
 
 	/* Make sure we don't make a zero height image.
 	 */

libvips/include/vips/util.h

@@ -94,9 +94,6 @@ G_STMT_START { \
 	} \
 } G_STMT_END
 
-/* Round a float to the nearest integer. Much faster than rint(). 
- */
-#define VIPS_RINT( R ) ((int) ((R) > 0 ? ((R) + 0.5) : ((R) - 0.5)))
 
 /* Various integer range clips. Record over/under flows.
  */
@@ -164,9 +161,15 @@ G_STMT_START { \
 #if defined(__clang__) || (__GNUC__ >= 4)
 #define VIPS_ISNAN( V ) __builtin_isnan( V )
 #define VIPS_ISINF( V ) __builtin_isinf( V )
+#define VIPS_FLOOR( V ) __builtin_floor( V )
+#define VIPS_CEIL( V ) __builtin_ceil( V )
+#define VIPS_RINT( V ) __builtin_rint( V )
 #else
 #define VIPS_ISNAN( V ) isnan( V )
 #define VIPS_ISINF( V ) isinf( V )
+#define VIPS_FLOOR( V ) floor( V )
+#define VIPS_CEIL( V ) ceil( V )
+#define VIPS_RINT( R ) ((int) ((R) > 0 ? ((R) + 0.5) : ((R) - 0.5)))
 #endif
 
 /* Not all platforms have PATH_MAX (eg. Hurd) and we don't need a platform one

libvips/resample/affine.c

@@ -330,8 +330,8 @@ vips_affine_gen( VipsRegion *or, void *seq, void *a, void *b, gboolean *stop )
 		for( x = le; x < ri; x++ ) {
 			int fx, fy; 	
 
-			fx = FAST_PSEUDO_FLOOR( ix );
+			fx = VIPS_FLOOR( ix );
-			fy = FAST_PSEUDO_FLOOR( iy );
+			fy = VIPS_FLOOR( iy );
 
 			/* Clip against iarea.
 			 */

libvips/resample/presample.h

@@ -50,35 +50,6 @@ extern "C" {
 	(G_TYPE_INSTANCE_GET_CLASS( (obj), \
 		VIPS_TYPE_RESAMPLE, VipsResampleClass ))
 
-/*
- * __builtin_floor is the fastest available floor function, if available.
- *
- * FAST_PSEUDO_FLOOR is a floor and floorf replacement which has been
- * found to be faster on several linux boxes than the library
- * version. It returns the floor of its argument unless the argument
- * is a negative integer, in which case it returns one less than the
- * floor. For example:
- *
- * FAST_PSEUDO_FLOOR(0.5) = 0
- *
- * FAST_PSEUDO_FLOOR(0.) = 0
- *
- * FAST_PSEUDO_FLOOR(-.5) = -1
- *
- * as expected, but
- *
- * FAST_PSEUDO_FLOOR(-1.) = -2
- *
- * The locations of the discontinuities of FAST_PSEUDO_FLOOR are the
- * same as floor and floorf; it is just that at negative integers the
- * function is discontinuous on the right instead of the left.
- */
-#if defined(__clang__) || (__GNUC__ >= 4)
-#define FAST_PSEUDO_FLOOR(x) __builtin_floor( x )
-#else
-#define FAST_PSEUDO_FLOOR(x) ( (int)(x) - ( (x) < 0. ) )
-#endif
-
 typedef struct _VipsResample {
 	VipsOperation parent_instance;
 

libvips/resample/resize.c

@@ -123,8 +123,10 @@ vips_resize_build( VipsObject *object )
 	/* If the factor is > 1.0, we need to zoom rather than shrink.
 	 * Just set the int part to 1 in this case.
 	 */
-	int_hshrink = resize->scale > 1.0 ? 1 : floor( 1.0 / resize->scale );
+	int_hshrink = resize->scale > 1.0 ? 
-	int_vshrink = resize->vscale > 1.0 ? 1 : floor( 1.0 / resize->vscale );
+		1 : VIPS_FLOOR( 1.0 / resize->scale );
+	int_vshrink = resize->vscale > 1.0 ? 
+		1 : VIPS_FLOOR( 1.0 / resize->vscale );
 
 	/* We want to shrink by less for interpolators with larger windows.
 	 */

libvips/resample/shrink2.c

@@ -284,8 +284,8 @@ vips_shrink2_gen( VipsRegion *or, void *vseq, void *a, void *b, gboolean *stop )
 
 		s.left = r->left * shrink->xshrink;
 		s.top = (r->top + y) * shrink->yshrink;
-		s.width = ceil( r->width * shrink->xshrink );
+		s.width = VIPS_CEIL( r->width * shrink->xshrink );
-		s.height = ceil( height * shrink->yshrink );
+		s.height = VIPS_CEIL( height * shrink->yshrink );
 #ifdef DEBUG
 		printf( "shrink_gen: requesting %d x %d at %d x %d\n",
 			s.width, s.height, s.left, s.top ); 
@@ -319,8 +319,8 @@ vips_shrink2_build( VipsObject *object )
 	if( VIPS_OBJECT_CLASS( vips_shrink2_parent_class )->build( object ) )
 		return( -1 );
 
-	shrink->mw = ceil( shrink->xshrink );
+	shrink->mw = VIPS_CEIL( shrink->xshrink );
-	shrink->mh = ceil( shrink->yshrink );
+	shrink->mh = VIPS_CEIL( shrink->yshrink );
 	shrink->np = shrink->mw * shrink->mh;
 
 	in = resample->in;