diff --git a/.gitignore b/.gitignore
index 91154068..af3b5d66 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,4 +1,8 @@
 a.out
+*.log
+*.trs
+tmp
+test/variables.sh
 scan
 compile
 po/*.pot
diff --git a/Makefile.am b/Makefile.am
index a173ba1e..b361cae6 100644
--- a/Makefile.am
+++ b/Makefile.am
@@ -42,6 +42,7 @@ SUBDIRS = \
 	po \
 	man \
 	doc \
+	test \
 	$(C_COMPILE_DIR) \
 	$(P_COMPILE_DIR) \
 	$(P8_COMPILE_DIR) \
diff --git a/TODO b/TODO
index 8c44765f..c0ddb796 100644
--- a/TODO
+++ b/TODO
@@ -1,21 +1,14 @@
-- use vips_gaussblur() in vips_resize(), and vipsthumbnail.c
+- getting an error in test_formats.sh
 
+	$ ./test_formats.sh 
+	testing IMG_4618.jpg v ... ok
+	testing IMG_4618.jpg tif ... ok
+	testing IMG_4618.jpg tif[compression=jpeg] ... ok
+	testing IMG_4618.jpg tif[compression=deflate] ... ok
+	testing IMG_4618.jpg tif[compression=packbits] ... ok
+	testing IMG_4618.jpg tif[compression=jpeg,tile] ... vips warning: linecache:
+	error reading tile 0x128: VipsJpeg: out of order read at line 384
 
-- test building without cpp
-
-	need to add introspection.m4 to m4/ dir?
-
-	configure.ac has 
-
-		GOBJECT_INTROSPECTION_CHECK([1.30.0])
-
-	libvips/Makefile.am has
-
-		if HAVE_INTROSPECTION
-
-	what defines that? is it a left-over?
-
-- test building without python
 
 - use vips_resize() in vipsthumbnail?
 
@@ -39,8 +32,6 @@
 	output imagevec
 	output blob
 
-- use vips_autorot() in jpegload and vipsthumbnail
-
 - fix up aconv
 
 - more python tests
diff --git a/configure.ac b/configure.ac
index 049ad5bd..e51c7467 100644
--- a/configure.ac
+++ b/configure.ac
@@ -791,6 +791,11 @@ AC_SUBST(VIPS_LIBS)
 AC_SUBST(VIPS_CXX_LIBS)
 AC_SUBST(PACKAGES_USED)
 
+# needed by test/variables.sh.in
+# :( what's a better way to do this, argh
+TOP_SRCDIR=$ac_pwd
+AC_SUBST(TOP_SRCDIR)
+
 AC_OUTPUT([
 	vips.pc
 	vipsCC.pc
@@ -828,6 +833,8 @@ AC_OUTPUT([
 	tools/light_correct 
 	tools/shrink_width 
 	python/Makefile 
+	test/Makefile 
+	test/variables.sh
 	swig/Makefile 
 	swig/vipsCC/Makefile 
 	man/Makefile
diff --git a/libvips/conversion/autorot.c b/libvips/conversion/autorot.c
index 90b3ab1e..db9c032c 100644
--- a/libvips/conversion/autorot.c
+++ b/libvips/conversion/autorot.c
@@ -55,6 +55,17 @@ G_DEFINE_TYPE( VipsAutorot, vips_autorot, VIPS_TYPE_CONVERSION );
 
 #define ORIENTATION ("exif-ifd0-Orientation")
 
+/**
+ * vips_autorot_get_angle:
+ * @im: image to fetch orientation from
+ *
+ * Examine the metadata on @im and return the #VipsAngle to rotate by to turn
+ * the image upright. 
+ *
+ * See also: vips_autorot(). 
+ *
+ * Returns: the #VipsAngle to rotate by to make the image upright.
+ */
 VipsAngle
 vips_autorot_get_angle( VipsImage *im )
 {
diff --git a/python/test/test_all.py b/python/test/test_all.py
deleted file mode 100755
index ffc0b69c..00000000
--- a/python/test/test_all.py
+++ /dev/null
@@ -1,10 +0,0 @@
-#!/usr/bin/python
-
-import unittest
-
-from test_arithmetic import TestArithmetic
-from test_colour import TestColour
-from test_conversion import TestConversion
-
-if __name__ == '__main__':
-    unittest.main()
diff --git a/test/Makefile.am b/test/Makefile.am
new file mode 100644
index 00000000..fea31bc6
--- /dev/null
+++ b/test/Makefile.am
@@ -0,0 +1,14 @@
+
+EXTRA_DIST = \
+	images \
+	variables.sh.in
+
+TESTS = \
+	test_cli.sh \
+	test_formats.sh \
+	test_seq.sh \
+	test_threading.sh \
+	test_arithmetic.py \
+	test_colour.py \
+	test_conversion.py 
+
diff --git a/python/test/images/IMG_4618.jpg b/test/images/IMG_4618.jpg
similarity index 100%
rename from python/test/images/IMG_4618.jpg
rename to test/images/IMG_4618.jpg
diff --git a/python/test/images/sRGB.icm b/test/images/sRGB.icm
similarity index 100%
rename from python/test/images/sRGB.icm
rename to test/images/sRGB.icm
diff --git a/test/images/sample.mat b/test/images/sample.mat
new file mode 100644
index 00000000..b809dee6
Binary files /dev/null and b/test/images/sample.mat differ
diff --git a/test/images/sample.png b/test/images/sample.png
new file mode 100644
index 00000000..ede09131
Binary files /dev/null and b/test/images/sample.png differ
diff --git a/test/images/t00740_tr1_segm.hdr b/test/images/t00740_tr1_segm.hdr
new file mode 100644
index 00000000..6ea6141b
Binary files /dev/null and b/test/images/t00740_tr1_segm.hdr differ
diff --git a/test/images/t00740_tr1_segm.img b/test/images/t00740_tr1_segm.img
new file mode 100644
index 00000000..1101323b
Binary files /dev/null and b/test/images/t00740_tr1_segm.img differ
diff --git a/python/test/test_arithmetic.py b/test/test_arithmetic.py
similarity index 100%
rename from python/test/test_arithmetic.py
rename to test/test_arithmetic.py
diff --git a/test/test_cli.sh b/test/test_cli.sh
new file mode 100755
index 00000000..0eed7c75
--- /dev/null
+++ b/test/test_cli.sh
@@ -0,0 +1,57 @@
+#!/bin/sh
+
+#set -x
+
+. ./variables.sh
+
+# is a difference beyond a threshold? return 0 (meaning all ok) or 1 (meaning
+# error, or outside threshold)
+# 
+# use bc since bash does not support fp math
+break_threshold() {
+	diff=$1
+	threshold=$2
+	return $(echo "$diff <= $threshold" | bc -l)
+}
+
+# subtract, look for max difference less than a threshold
+test_difference() {
+	before=$1
+	after=$2
+	threshold=$3
+
+	vips subtract $before $after $tmp/difference.v
+	vips abs $tmp/difference.v $tmp/abs.v 
+	dif=$(vips max $tmp/abs.v)
+
+	if break_threshold $dif $threshold; then
+		echo "difference is $dif"
+		exit 1
+	fi
+}
+
+test_rotate() {
+	im=$1
+	inter=$2
+
+	echo -n "testing $inter ... "
+
+	# 90 degree clockwise rotate 
+	trn="0 1 1 0"
+
+	vips affine $im $tmp/t1.v "$trn" --interpolate $inter
+	vips affine $tmp/t1.v $tmp/t2.v "$trn" --interpolate $inter
+	vips affine $tmp/t2.v $tmp/t1.v "$trn" --interpolate $inter
+	vips affine $tmp/t1.v $tmp/t2.v "$trn" --interpolate $inter
+
+	test_difference $im $tmp/t2.v 1
+
+	echo "ok"
+}
+
+# vsqbs is non-interpolatory, don't test this way
+
+echo "testing with $(basename $image)"
+for i in nearest bicubic bilinear nohalo lbb; do
+	test_rotate $image $i
+done
diff --git a/python/test/test_colour.py b/test/test_colour.py
similarity index 100%
rename from python/test/test_colour.py
rename to test/test_colour.py
diff --git a/python/test/test_conversion.py b/test/test_conversion.py
similarity index 100%
rename from python/test/test_conversion.py
rename to test/test_conversion.py
diff --git a/python/test/test_convolution.py b/test/test_convolution.py
similarity index 100%
rename from python/test/test_convolution.py
rename to test/test_convolution.py
diff --git a/test/test_formats.sh b/test/test_formats.sh
new file mode 100755
index 00000000..3d384db9
--- /dev/null
+++ b/test/test_formats.sh
@@ -0,0 +1,170 @@
+#!/bin/sh
+
+# set -x
+
+. ./variables.sh
+
+# the matlab image and reference image
+matlab=$test_images/sample.mat
+matlab_ref=$test_images/sample.png
+
+# make a mono image
+vips extract_band $image $tmp/mono.v 1
+mono=$tmp/mono.v
+
+# make a radiance image
+vips float2rad $image $tmp/rad.v 
+rad=$tmp/rad.v
+
+# make a cmyk image
+vips bandjoin "$image $tmp/mono.v" $tmp/t1.v
+vips copy $tmp/t1.v $tmp/cmyk.v --interpretation cmyk
+cmyk=$tmp/cmyk.v
+
+# save to t1.format, load as back.v
+save_load() {
+	in=$1
+	format=$2
+	mode=$3
+
+	if ! vips copy $in $tmp/t1.$format$mode ; then
+		echo "write to $out failed"
+		exit 1
+	fi
+
+	if ! vips copy $tmp/t1.$format $tmp/back.v ; then
+		echo "read from $out failed"
+		exit 1
+	fi
+}
+
+# is a difference beyond a threshold? return 0 (meaning all ok) or 1 (meaning
+# error, or outside threshold)
+# 
+# use bc since bash does not support fp math
+break_threshold() {
+	diff=$1
+	threshold=$2
+	return $(echo "$diff <= $threshold" | bc -l)
+}
+
+# subtract, look for max difference less than a threshold
+test_difference() {
+	before=$1
+	after=$2
+	threshold=$3
+
+	vips subtract $before $after $tmp/difference.v
+	vips abs $tmp/difference.v $tmp/abs.v 
+	dif=`vips max $tmp/abs.v`
+
+	if break_threshold $dif $threshold; then
+		echo "save / load difference is $dif"
+		exit 1
+	fi
+}
+
+# save to the named file in tmp, convert back to vips again, subtract, look
+# for max difference less than a threshold
+test_format() {
+	in=$1
+	format=$2
+	threshold=$3
+	mode=$4
+
+	echo -n "testing $(basename $in) $format$mode ... "
+
+	save_load $in $format $mode
+	test_difference $in $tmp/back.v $threshold
+
+	echo "ok"
+}
+
+# as above, but hdr format
+# this is a coded format, so we need to rad2float before we can test for
+# differences
+test_rad() {
+	in=$1
+
+	echo -n "testing $(basename $in) hdr ... "
+
+	save_load $in hdr
+
+	vips rad2float $in $tmp/before.v
+	vips rad2float $tmp/back.v $tmp/after.v
+
+	test_difference $tmp/before.v $tmp/after.v 0
+
+	echo "ok"
+}
+
+# as above, but raw format
+# we can't use suffix stuff to pick the load/save
+test_raw() {
+	in=$1
+
+	echo -n "testing $(basename $in) raw ... "
+
+	vips copy $in $tmp/before.v
+	width=`vipsheader -f Xsize $tmp/before.v`
+	height=`vipsheader -f Ysize $tmp/before.v`
+	bands=`vipsheader -f Bands $tmp/before.v`
+
+	vips rawsave $tmp/before.v $tmp/raw
+	vips rawload $tmp/raw $tmp/after.v $width $height $bands
+
+	test_difference $tmp/before.v $tmp/after.v 0
+
+	echo "ok"
+}
+
+# a format for which we only have a load (eg. matlab)
+# pass in a reference file as well and compare to that
+test_loader() {
+	ref=$1
+	in=$2
+	format=$3
+
+	echo -n "testing $(basename $in) $format ... "
+
+	vips copy $ref $tmp/before.v
+	vips copy $in $tmp/after.v
+
+	test_difference $tmp/before.v $tmp/after.v 0
+
+	echo "ok"
+}
+
+test_format $image v 0
+test_format $image tif 0
+test_format $image tif 70 [compression=jpeg]
+test_format $image tif 0 [compression=deflate]
+test_format $image tif 0 [compression=packbits]
+test_format $image tif 70 [compression=jpeg,tile]
+test_format $image tif 70 [compression=jpeg,tile,pyramid]
+test_format $image png 0
+test_format $image png 0 [compression=9,interlace=1]
+test_format $image jpg 15
+test_format $image ppm 0
+test_format $image pfm 0
+test_format $image fits 0
+
+# csv can only do mono
+test_format $mono csv 0
+
+# cmyk jpg is a special path
+test_format $cmyk jpg 15
+test_format $cmyk tif 0
+test_format $cmyk tif 15 [compression=jpeg]
+test_format $cmyk tif 15 [compression=jpeg,tile]
+test_format $cmyk tif 15 [compression=jpeg,tile,pyramid]
+
+test_rad $rad 
+
+test_raw $mono 
+test_raw $image 
+
+test_loader $matlab_ref $matlab matlab
+
+# we have loaders but not savers for other formats, add tests here
+
diff --git a/test/test_seq.sh b/test/test_seq.sh
new file mode 100755
index 00000000..3023685d
--- /dev/null
+++ b/test/test_seq.sh
@@ -0,0 +1,45 @@
+#!/bin/sh
+
+# set -x
+
+. ./variables.sh
+
+# make a large PNG, roughly the size of Chicago.png
+echo -n "building huge test PNG image ... "
+vips replicate $image $tmp/huge.png 30 5
+echo "ok"
+
+huge=$tmp/huge.png
+
+echo -n "testing vipsthumbnail ... "
+rm -f $tmp/x.png
+vipsthumbnail $huge -o $tmp/x.png
+if ! vipsheader $tmp/x.png > /dev/null 2>&1 ; then
+	echo "vipsthumbnail failed in basic mode"
+	exit 1
+fi
+echo "ok"
+
+if [ ! -d $tmp/readonly ] ; then
+	mkdir $tmp/readonly
+	chmod ugo-wx $tmp/readonly
+fi 
+export TMPDIR=$tmp/readonly
+
+echo -n "testing vipsthumbnail does not make temps ... "
+rm -f $tmp/x.png
+vipsthumbnail $huge -o $tmp/x.png
+if ! vipsheader $tmp/x.png > /dev/null 2>&1 ; then
+	echo "vipsthumbnail made a temp"
+	exit 1
+fi
+echo "ok"
+
+echo -n "testing shrink does not make temps ... "
+rm -f $tmp/x.png
+vips shrink $huge $tmp/x.png 230 230 
+if ! vipsheader $tmp/x.png > /dev/null 2>&1 ; then
+	echo "shrink made a temp"
+	exit 1
+fi
+echo "ok"
diff --git a/test/test_threading.sh b/test/test_threading.sh
new file mode 100755
index 00000000..75a83f5b
--- /dev/null
+++ b/test/test_threading.sh
@@ -0,0 +1,43 @@
+#!/bin/sh
+
+# set -x
+
+source variables.sh 
+
+chain=1
+
+# im_benchmark needs a labq
+vips im_sRGB2XYZ $image $tmp/t1.v
+vips im_XYZ2Lab $tmp/t1.v $tmp/t2.v
+vips im_Lab2LabQ $tmp/t2.v $tmp/t3.v
+
+for tile in 10 64 128 512; do
+	# benchmark includes a dither which will vary with tile size
+	vips --vips-concurrency=1 \
+		--vips-tile-width=$tile --vips-tile-height=$tile \
+		im_benchmarkn $tmp/t3.v $tmp/t5.v $chain
+
+	for cpus in 2 3 4 5 6 7 8 99 1024; do
+		echo trying cpus = $cpus, tile = $tile ...
+		vips --vips-concurrency=$cpus \
+			--vips-tile-width=$tile --vips-tile-height=$tile \
+			im_benchmarkn $tmp/t3.v $tmp/t7.v $chain
+		vips im_subtract $tmp/t5.v $tmp/t7.v $tmp/t8.v
+		vips im_abs $tmp/t8.v $tmp/t9.v
+		max=`vips im_max $tmp/t9.v`
+		if [ $max -gt 0 ]; then
+			break
+		fi
+	done
+	if [ $max -gt 0 ]; then
+		break
+	fi
+done
+
+if [ $max -gt 0 ]; then
+	echo error, max == $max
+	exit 1
+else
+	echo all threading tests passed
+fi
+
diff --git a/test/variables.sh.in b/test/variables.sh.in
new file mode 100644
index 00000000..148b838e
--- /dev/null
+++ b/test/variables.sh.in
@@ -0,0 +1,4 @@
+top_srcdir=@TOP_SRCDIR@
+tmp=$top_srcdir/test/tmp
+test_images=$top_srcdir/test/images
+image=$test_images/IMG_4618.jpg
diff --git a/tools/vipsthumbnail.c b/tools/vipsthumbnail.c
index 08e5741c..aff2dd7d 100644
--- a/tools/vipsthumbnail.c
+++ b/tools/vipsthumbnail.c
@@ -151,35 +151,6 @@ static GOptionEntry options[] = {
 	{ NULL }
 };
 
-static VipsAngle 
-get_angle( VipsImage *im )
-{
-	VipsAngle angle;
-	const char *orientation;
-
-	angle = VIPS_ANGLE_D0;
-
-	if( vips_image_get_typeof( im, ORIENTATION ) && 
-		!vips_image_get_string( im, ORIENTATION, &orientation ) ) {
-		if( vips_isprefix( "6", orientation ) )
-			angle = VIPS_ANGLE_D90;
-		else if( vips_isprefix( "8", orientation ) )
-			angle = VIPS_ANGLE_D270;
-		else if( vips_isprefix( "3", orientation ) )
-			angle = VIPS_ANGLE_D180;
-
-		/* Other values do rotate + mirror, don't bother handling them
-		 * though, how common can mirroring be. 
-		 *
-		 * See:
-		 *
-		 * http://www.80sidea.com/archives/2316
-		 */
-	}
-
-	return( angle );
-}
-
 /* Calculate the shrink factors. 
  *
  * We shrink in two stages: first, a shrink with a block average. This can
@@ -193,7 +164,7 @@ static int
 calculate_shrink( VipsImage *im, double *residual, 
 	VipsInterpolate *interp )
 {
-	VipsAngle angle = get_angle( im ); 
+	VipsAngle angle = vips_autorot_get_angle( im ); 
 	gboolean rotate = angle == VIPS_ANGLE_D90 || angle == VIPS_ANGLE_D270;
 	int width = rotate_image && rotate ? im->Ysize : im->Xsize;
 	int height = rotate_image && rotate ? im->Xsize : im->Ysize;
@@ -518,18 +489,10 @@ thumbnail_shrink( VipsObject *process, VipsImage *in,
 	sigma = ((1.0 / residual) - 0.5) / 1.5;
 	if( residual < 1.0 &&
 		sigma > 0.1 ) { 
-		if( vips_gaussmat( &t[9], sigma, 0.2,
-			"separable", TRUE,
-			"integer", TRUE,
-			NULL ) ||
-			vips_convsep( in, &t[5], t[9], NULL ) )
+		if( vips_gaussblur( in, &t[5], sigma, NULL ) )
 			return( NULL );
 		vips_info( "vipsthumbnail", "anti-alias, sigma %g",
 			sigma );
-#ifdef DEBUG
-		printf( "anti-alias blur matrix is:\n" ); 
-		vips_matrixprint( t[9], NULL ); 
-#endif /*DEBUG*/
 		in = t[5];
 	}
 
@@ -664,7 +627,7 @@ static VipsImage *
 thumbnail_rotate( VipsObject *process, VipsImage *im )
 {
 	VipsImage **t = (VipsImage **) vips_object_local_array( process, 2 );
-	VipsAngle angle = get_angle( im );
+	VipsAngle angle = vips_autorot_get_angle( im );
 
 	if( rotate_image &&
 		angle != VIPS_ANGLE_D0 ) {