sna: Apply any previous transformation when downsampling

In order to handle rotations and fractional offsets produced by the act
of downsampling, we need to compute the full affine transformation and
apply it to the vertices rather than attempt to fudge it with an integer
offset.

References: https://bugs.freedesktop.org/show_bug.cgi?id=45086
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

4 files changed, 121 insertions, 248 deletions

diff --git a/src/sna/kgem.c b/src/sna/kgem.c
index 6eae248a..a2fcefc4 100644
--- a/src/sna/kgem.c
+++ b/src/sna/kgem.c
@@ -3323,67 +3323,6 @@ struct kgem_bo *kgem_upload_source_image(struct kgem *kgem,
 	return bo;
 }
 
-struct kgem_bo *kgem_upload_source_image_halved(struct kgem *kgem,
-						pixman_format_code_t format,
-						const void *data,
-						int x, int y,
-						int width, int height,
-						int stride, int bpp)
-{
-	struct kgem_bo *bo;
-	pixman_image_t *src_image, *dst_image;
-	pixman_transform_t t;
-	int w, h;
-	void *dst;
-
-	DBG(("%s : (%d, %d), (%d, %d), stride=%d, bpp=%d\n",
-	     __FUNCTION__, x, y, width, height, stride, bpp));
-
-	w = (width + 1) / 2;
-	h = (height + 1) / 2;
-
-	bo = kgem_create_buffer_2d(kgem, w, h, bpp,
-				   KGEM_BUFFER_WRITE_INPLACE,
-				   &dst);
-	if (bo == NULL)
-		return NULL;
-
-	dst_image = pixman_image_create_bits(format, w, h, dst, bo->pitch);
-	if (dst_image == NULL)
-		goto cleanup_bo;
-
-	src_image = pixman_image_create_bits(format, width, height,
-					     (uint32_t*)data, stride);
-	if (src_image == NULL)
-		goto cleanup_dst;
-
-	memset(&t, 0, sizeof(t));
-	t.matrix[0][0] = 2 << 16;
-	t.matrix[1][1] = 2 << 16;
-	t.matrix[2][2] = 1 << 16;
-	pixman_image_set_transform(src_image, &t);
-	pixman_image_set_filter(src_image, PIXMAN_FILTER_BILINEAR, NULL, 0);
-	pixman_image_set_repeat(src_image, PIXMAN_REPEAT_PAD);
-
-	pixman_image_composite(PIXMAN_OP_SRC,
-			       src_image, NULL, dst_image,
-			       x, y,
-			       0, 0,
-			       0, 0,
-			       w, h);
-
-	pixman_image_unref(src_image);
-	pixman_image_unref(dst_image);
-
-	return bo;
-
-cleanup_dst:
-	pixman_image_unref(dst_image);
-cleanup_bo:
-	kgem_bo_destroy(kgem, bo);
-	return NULL;
-}
-
 void kgem_buffer_read_sync(struct kgem *kgem, struct kgem_bo *_bo)
 {
 	struct kgem_partial_bo *bo;
diff --git a/src/sna/kgem.h b/src/sna/kgem.h
index fd3aa9de..db4f061b 100644
--- a/src/sna/kgem.h
+++ b/src/sna/kgem.h
@@ -191,12 +191,6 @@ struct kgem_bo *kgem_upload_source_image(struct kgem *kgem,
 					 const void *data,
 					 BoxPtr box,
 					 int stride, int bpp);
-struct kgem_bo *kgem_upload_source_image_halved(struct kgem *kgem,
-						pixman_format_code_t format,
-						const void *data,
-						int x, int y,
-						int width, int height,
-						int stride, int bpp);
 
 int kgem_choose_tiling(struct kgem *kgem,
 		       int tiling, int width, int height, int bpp);
diff --git a/src/sna/sna_render.c b/src/sna/sna_render.c
index 28b93a25..9d7857cf 100644
--- a/src/sna/sna_render.c
+++ b/src/sna/sna_render.c
@@ -615,15 +615,17 @@ static int sna_render_picture_downsample(struct sna *sna,
 					 int16_t w, int16_t h,
 					 int16_t dst_x, int16_t dst_y)
 {
-	struct kgem_bo *bo = NULL;
 	PixmapPtr pixmap = get_drawable_pixmap(picture->pDrawable);
-	int16_t ox, oy, ow, oh;
-	BoxRec box;
-
-	assert(w && h);
-
-	DBG(("%s (%d, %d)x(%d, %d) [dst=(%d, %d)]\n",
-	     __FUNCTION__, x, y, w, h, dst_x, dst_y));
+	ScreenPtr screen = pixmap->drawable.pScreen;
+	PicturePtr tmp_src, tmp_dst;
+	PictFormatPtr format;
+	struct sna_pixmap *priv;
+	pixman_transform_t t;
+	PixmapPtr tmp;
+	int width, height;
+	int sx, sy, ox, oy, ow, oh;
+	int error, ret = 0;
+	BoxRec box, b;
 
 	ow = w;
 	oh = h;
@@ -645,12 +647,6 @@ static int sna_render_picture_downsample(struct sna *sna,
 		oy = v.vector[1] / v.vector[2];
 	}
 
-	/* Align the origin to an even pixel so that the sampling of
-	 * partial images is stable.
-	 */
-	box.x1 &= ~1;
-	box.y1 &= ~1;
-
 	if (channel->repeat == RepeatNone || channel->repeat == RepeatPad) {
 		if (box.x1 < 0)
 			box.x1 = 0;
@@ -684,184 +680,124 @@ static int sna_render_picture_downsample(struct sna *sna,
 
 	w = box.x2 - box.x1;
 	h = box.y2 - box.y1;
-	DBG(("%s: sample area (%d, %d), (%d, %d): %dx%d\n",
-	     __FUNCTION__, box.x1, box.y1, box.x2, box.y2, w, h));
-	assert(w && h);
-	if (w > 2*sna->render.max_3d_size || h > 2*sna->render.max_3d_size) {
-		DBG(("%s: sample size too large for pixman downscaling\n",
-		     __FUNCTION__));
-		goto fixup;
-	}
 
-	if (texture_is_cpu(pixmap, &box) && !move_to_gpu(pixmap, &box)) {
-		DBG(("%s: uploading partial texture\n", __FUNCTION__));
-		bo = kgem_upload_source_image_halved(&sna->kgem,
-						     picture->format,
-						     pixmap->devPrivate.ptr,
-						     box.x1, box.y1, w, h,
-						     pixmap->devKind,
-						     pixmap->drawable.bitsPerPixel);
-		if (!bo) {
-			DBG(("%s: failed to upload source image, using clear\n",
-			     __FUNCTION__));
-			return 0;
-		}
-	} else {
-		ScreenPtr screen = pixmap->drawable.pScreen;
-		PicturePtr tmp_src, tmp_dst;
-		PictFormatPtr format;
-		struct sna_pixmap *priv;
-		pixman_transform_t t;
-		PixmapPtr tmp;
-		int error, i, j, ww, hh, ni, nj;
-
-		if (!sna_pixmap_move_to_gpu(pixmap, MOVE_READ))
-			goto fixup;
-
-		tmp = screen->CreatePixmap(screen,
-					   (w+1)/2, (h+1)/2,
-					   pixmap->drawable.depth,
-					   SNA_CREATE_SCRATCH);
-		if (!tmp)
-			goto fixup;
-
-		priv = sna_pixmap(tmp);
-		if (!priv) {
-			screen->DestroyPixmap(tmp);
-			goto fixup;
-		}
+	DBG(("%s: sample (%d, %d), (%d, %d)\n",
+	     __FUNCTION__, box.x1, box.y1, box.x2, box.y2));
 
-		format = PictureMatchFormat(screen,
-					    pixmap->drawable.depth,
-					    picture->format);
+	sx = (w + sna->render.max_3d_size - 1) / sna->render.max_3d_size;
+	sy = (h + sna->render.max_3d_size - 1) / sna->render.max_3d_size;
 
-		tmp_dst = CreatePicture(0, &tmp->drawable, format, 0, NULL,
-					serverClient, &error);
-		if (!tmp_dst) {
-			screen->DestroyPixmap(tmp);
-			goto fixup;
-		}
+	DBG(("%s: scaling (%d, %d) down by %dx%d\n",
+	     __FUNCTION__, w, h, sx, sy));
 
-		tmp_src = CreatePicture(0, &pixmap->drawable, format, 0, NULL,
-					serverClient, &error);
-		if (!tmp_src) {
-			FreePicture(tmp_dst, 0);
-			screen->DestroyPixmap(tmp);
-			goto fixup;
-		}
+	width  = w / sx;
+	height = h / sy;
 
-		tmp_src->repeat = true;
-		tmp_src->repeatType = RepeatPad;
-		tmp_src->filter = PictFilterBilinear;
-		memset(&t, 0, sizeof(t));
-		t.matrix[0][0] = 2 << 16;
-		t.matrix[1][1] = 2 << 16;
-		t.matrix[2][2] = 1 << 16;
-		tmp_src->transform = &t;
-
-		ValidatePicture(tmp_dst);
-		ValidatePicture(tmp_src);
-
-		if (w > sna->render.max_3d_size) {
-			ww = (w+3)/4;
-			nj = 2;
-		} else {
-			ww = (w+1)/2;
-			nj = 1;
-		}
+	DBG(("%s: creating temporary GPU bo %dx%d\n",
+	     __FUNCTION__, width, height));
 
-		if (h > sna->render.max_3d_size) {
-			hh = (h+3)/4;
-			ni = 2;
-		} else {
-			hh = (h+1)/2;
-			ni = 1;
-		}
+	tmp = screen->CreatePixmap(screen,
+				   width, height,
+				   pixmap->drawable.depth,
+				   SNA_CREATE_SCRATCH);
+	if (!tmp)
+		return 0;
 
-		DBG(("%s %d:%d downsampling using %dx%d GPU tiles\n",
-		     __FUNCTION__, nj, ni, ww, hh));
-
-		for (i = 0; i < ni; i++) {
-			BoxRec b;
-
-			b.y1 = hh*i;
-			if (i == ni - 1)
-				b.y2 = (h+1)/2;
-			else
-				b.y2 = b.y1 + hh;
-
-			for (j = 0; j < nj; j++) {
-				struct sna_composite_op op;
-
-				b.x1 = ww*j;
-				if (j == nj - 1)
-					b.x2 = (w+1)/2;
-				else
-					b.x2 = b.x1 + ww;
-
-				DBG(("%s: tile %d:%d, box=(%d,%d), (%d, %d)\n",
-				     __FUNCTION__, i, j, b.x1, b.y1, b.x2, b.y2));
-
-				memset(&op, 0, sizeof(op));
-				if (!sna->render.composite(sna,
-							   PictOpSrc,
-							   tmp_src, NULL, tmp_dst,
-							   box.x1/2 + b.x1, box.y1/2 + b.y1,
-							   0, 0,
-							   b.x1, b.y1,
-							   b.x2 - b.x1, b.y2 - b.y1,
-							   &op)) {
-					tmp_src->transform = NULL;
-					FreePicture(tmp_src, 0);
-					FreePicture(tmp_dst, 0);
-					screen->DestroyPixmap(tmp);
-					goto fixup;
-				}
-
-				op.boxes(sna, &op, &b, 1);
-				op.done(sna, &op);
-			}
+	priv = sna_pixmap(tmp);
+	if (!priv)
+		goto cleanup_tmp;
+
+	format = PictureMatchFormat(screen,
+				    pixmap->drawable.depth,
+				    picture->format);
+
+	tmp_dst = CreatePicture(0, &tmp->drawable, format, 0, NULL,
+				serverClient, &error);
+	if (!tmp_dst)
+		goto cleanup_tmp;
+
+	tmp_src = CreatePicture(0, &pixmap->drawable, format, 0, NULL,
+				serverClient, &error);
+	if (!tmp_src)
+		goto cleanup_dst;
+
+	tmp_src->repeat = 1;
+	tmp_src->repeatType = RepeatPad;
+	/* Prefer to use nearest as it helps reduce artefacts from
+	 * interpolating and filtering twice.
+	 */
+	tmp_src->filter = PictFilterNearest;
+	memset(&t, 0, sizeof(t));
+	t.matrix[0][0] = (w << 16) / width;
+	t.matrix[0][2] = box.x1 << 16;
+	t.matrix[1][1] = (h << 16) / height;
+	t.matrix[1][2] = box.y1 << 16;
+	t.matrix[2][2] = 1 << 16;
+	tmp_src->transform = &t;
+
+	ValidatePicture(tmp_dst);
+	ValidatePicture(tmp_src);
+
+	w = sna->render.max_3d_size / sx - 2 * sx;
+	h = sna->render.max_3d_size / sy - 2 * sy;
+	DBG(("%s %d:%d downsampling using %dx%d GPU tiles\n",
+	     __FUNCTION__, (width + w-1)/w, (height + h-1)/h, w, h));
+
+	for (b.y1 = 0; b.y1 < height; b.y1 = b.y2) {
+		b.y2 = b.y1 + h;
+		if (b.y2 > height)
+			b.y2 = height;
+
+		for (b.x1 = 0; b.x1 < width; b.x1 = b.x2) {
+			struct sna_composite_op op;
+
+			b.x2 = b.x1 + w;
+			if (b.x2 > width)
+				b.x2 = width;
+
+			DBG(("%s: tile (%d, %d), (%d, %d)\n",
+			     __FUNCTION__, b.x1, b.y1, b.x2, b.y2));
+
+			memset(&op, 0, sizeof(op));
+			if (!sna->render.composite(sna,
+						   PictOpSrc,
+						   tmp_src, NULL, tmp_dst,
+						   b.x1, b.y1,
+						   0, 0,
+						   b.x1, b.y1,
+						   b.x2 - b.x1, b.y2 - b.y1,
+						   &op))
+				goto cleanup_src;
+
+			op.boxes(sna, &op, &b, 1);
+			op.done(sna, &op);
 		}
-
-		bo = kgem_bo_reference(priv->gpu_bo);
-
-		tmp_src->transform = NULL;
-		FreePicture(tmp_src, 0);
-		FreePicture(tmp_dst, 0);
-		screen->DestroyPixmap(tmp);
 	}
 
-	if (ox == x && oy == y) {
-		x = y = 0;
-	} else if (channel->transform) {
-		pixman_vector_t v;
-		pixman_transform_t m;
-
-		v.vector[0] = (ox - box.x1) << 16;
-		v.vector[1] = (oy - box.y1) << 16;
-		v.vector[2] = 1 << 16;
-		pixman_transform_invert(&m, channel->transform);
-		pixman_transform_point(&m, &v);
-		x = v.vector[0] / v.vector[2];
-		y = v.vector[1] / v.vector[2];
-	} else {
-		x = ox - box.x1;
-		y = oy - box.y1;
-	}
+	pixman_transform_invert(&channel->embedded_transform, &t);
+	if (channel->transform)
+		pixman_transform_multiply(&channel->embedded_transform,
+					  &channel->embedded_transform,
+					  channel->transform);
+	channel->transform = &channel->embedded_transform;
 
 	channel->offset[0] = x - dst_x;
 	channel->offset[1] = y - dst_y;
-	channel->scale[0] = 1.f/w;
-	channel->scale[1] = 1.f/h;
-	channel->width  = (w + 1) / 2;
-	channel->height = (h + 1) / 2;
-	channel->bo = bo;
-	return 1;
-
-fixup:
-	return sna_render_picture_fixup(sna, picture, channel,
-					x, y, w, h,
-					dst_x, dst_y);
+	channel->scale[0] = 1.f/width;
+	channel->scale[1] = 1.f/height;
+	channel->width  = width;
+	channel->height = height;
+	channel->bo = kgem_bo_reference(priv->gpu_bo);
+
+	ret = 1;
+cleanup_src:
+	tmp_src->transform = NULL;
+	FreePicture(tmp_src, 0);
+cleanup_dst:
+	FreePicture(tmp_dst, 0);
+cleanup_tmp:
+	screen->DestroyPixmap(tmp);
+	return ret;
 }
 
 int
@@ -965,7 +901,10 @@ sna_render_picture_extract(struct sna *sna,
 	}
 
 	src_bo = use_cpu_bo(sna, pixmap, &box);
-	if (src_bo == NULL) {
+	if (src_bo) {
+		if (!sna_pixmap_move_to_cpu(pixmap, MOVE_READ))
+			return 0;
+	} else {
 		if (texture_is_cpu(pixmap, &box) &&
 		    !move_to_gpu(pixmap, &box)) {
 			bo = kgem_upload_source_image(&sna->kgem,
@@ -981,7 +920,6 @@ sna_render_picture_extract(struct sna *sna,
 				src_bo = priv->gpu_bo;
 		}
 	}
-
 	if (src_bo) {
 		bo = kgem_create_2d(&sna->kgem, w, h,
 				    pixmap->drawable.bitsPerPixel,
diff --git a/src/sna/sna_render.h b/src/sna/sna_render.h
index 4346196d..c4711f4b 100644
--- a/src/sna/sna_render.h
+++ b/src/sna/sna_render.h
@@ -57,6 +57,8 @@ struct sna_composite_op {
 		int16_t offset[2];
 		float scale[2];
 
+		pixman_transform_t embedded_transform;
+
 		union {
 			struct {
 				uint32_t pixel;