diff --git a/clients/screenshot.c b/clients/screenshot.c
index 328d106c..a23ef13b 100644
--- a/clients/screenshot.c
+++ b/clients/screenshot.c
@@ -47,6 +47,7 @@ struct screenshooter_output {
 	struct wl_output *output;
 	struct wl_buffer *buffer;
 	int width, height, offset_x, offset_y;
+	void *data;
 	struct wl_list link;
 };
 
@@ -165,23 +166,45 @@ create_shm_buffer(int width, int height, void **data_out)
 }
 
 static void
-write_png(int width, int height, void *data)
+write_png(int width, int height)
 {
+	int output_stride, buffer_stride, i;
 	cairo_surface_t *surface;
+	void *data, *d, *s;
+	struct screenshooter_output *output, *next;
+
+	buffer_stride = width * 4;
+
+	data = malloc(buffer_stride * height);
+	if (!data)
+		return;
+
+	wl_list_for_each_safe(output, next, &output_list, link) {
+		output_stride = output->width * 4;
+		s = output->data;
+		d = data + output->offset_y * buffer_stride + output->offset_x * 4;
+
+		for (i = 0; i < output->height; i++) {
+			memcpy(d, s, output_stride);
+			d += buffer_stride;
+			s += output_stride;
+		}
+
+		free(output);
+	}
 
 	surface = cairo_image_surface_create_for_data(data,
 						      CAIRO_FORMAT_ARGB32,
-						      width, height, width * 4);
+						      width, height, buffer_stride);
 	cairo_surface_write_to_png(surface, "wayland-screenshot.png");
 	cairo_surface_destroy(surface);
+	free(data);
 }
 
 int main(int argc, char *argv[])
 {
 	struct wl_display *display;
-	struct wl_buffer *buffer;
-	void *data = NULL;
-	struct screenshooter_output *output, *next;
+	struct screenshooter_output *output;
 	int width = 0, height = 0;
 
 	display = wl_display_connect(NULL);
@@ -203,21 +226,16 @@ int main(int argc, char *argv[])
 
 
 	wl_list_for_each(output, &output_list, link) {
+		output->buffer = create_shm_buffer(output->width, output->height, &output->data);
+		screenshooter_shoot(screenshooter, output->output, output->buffer);
 		width = MAX(width, output->offset_x + output->width);
 		height = MAX(height, output->offset_y + output->height);
-	}
-
-	buffer = create_shm_buffer(width, height, &data);
-
-	wl_list_for_each_safe(output, next, &output_list, link) {
-		screenshooter_shoot(screenshooter, output->output, buffer);
 		buffer_copy_done = 0;
 		while (!buffer_copy_done)
 			wl_display_roundtrip(display);
-		free(output);
 	}
 
-	write_png(width, height, data);
+	write_png(width, height);
 
 	return 0;
 }
diff --git a/src/screenshooter.c b/src/screenshooter.c
index 7ba475d9..90dd4979 100644
--- a/src/screenshooter.c
+++ b/src/screenshooter.c
@@ -43,16 +43,15 @@ struct screenshooter_read_pixels {
 };
 
 static void
-copy_bgra_yflip(uint8_t *dst, uint8_t *src, int height,
-		int dst_stride, int src_stride)
+copy_bgra_yflip(uint8_t *dst, uint8_t *src, int height, int stride)
 {
 	uint8_t *end;
 
-	end = dst + height * dst_stride;
+	end = dst + height * stride;
 	while (dst < end) {
-		memcpy(dst, src, src_stride);
-		dst += dst_stride;
-		src -= src_stride;
+		memcpy(dst, src, stride);
+		dst += stride;
+		src -= stride;
 	}
 }
 
@@ -74,16 +73,15 @@ copy_row_swap_RB(void *vdst, void *vsrc, int bytes)
 }
 
 static void
-copy_rgba_yflip(uint8_t *dst, uint8_t *src, int height,
-		int dst_stride, int src_stride)
+copy_rgba_yflip(uint8_t *dst, uint8_t *src, int height, int stride)
 {
 	uint8_t *end;
 
-	end = dst + height * dst_stride;
+	end = dst + height * stride;
 	while (dst < end) {
-		copy_row_swap_RB(dst, src, src_stride);
-		dst += dst_stride;
-		src -= src_stride;
+		copy_row_swap_RB(dst, src, stride);
+		dst += stride;
+		src -= stride;
 	}
 }
 
@@ -93,24 +91,20 @@ screenshooter_read_pixels_done(struct weston_read_pixels *base,
 {
 	struct screenshooter_read_pixels *r =
 		(struct screenshooter_read_pixels *) base;
-	int32_t buffer_stride, output_stride;
+	int32_t stride;
 	uint8_t *d, *s;
 
-	buffer_stride = wl_shm_buffer_get_stride(r->buffer);
-	output_stride = output->current->width * 4;
+	stride = wl_shm_buffer_get_stride(r->buffer);
 
-	d = wl_shm_buffer_get_data(r->buffer) +
-		output->y * buffer_stride + output->x * 4;
-	s = r->base.data + output_stride * (output->current->height - 1);
+	d = wl_shm_buffer_get_data(r->buffer);
+	s = r->base.data + stride * (r->buffer->height - 1);
 
 	switch (output->compositor->read_format) {
 	case GL_BGRA_EXT:
-		copy_bgra_yflip(d, s, output->current->height,
-				buffer_stride, output_stride);
+		copy_bgra_yflip(d, s, output->current->height, stride);
 		break;
 	case GL_RGBA:
-		copy_rgba_yflip(d, s, output->current->height,
-				buffer_stride, output_stride);
+		copy_rgba_yflip(d, s, output->current->height, stride);
 		break;
 	default:
 		break;