sna/gen4: Split the have_render flag in separate prefer_gpu hints

The idea is to implement more fine-grained checks as we may want
different heuristics for desktops with GT1s than for mobile GT2s, etc.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

14 files changed, 226 insertions, 166 deletions

diff --git a/src/sna/gen2_render.c b/src/sna/gen2_render.c
index b539e278..4d92adcf 100644
--- a/src/sna/gen2_render.c
+++ b/src/sna/gen2_render.c
@@ -2227,7 +2227,7 @@ gen2_check_composite_spans(struct sna *sna,
 		return false;
 
 	if (need_tiling(sna, width, height)) {
-		if (!is_gpu(dst->pDrawable)) {
+		if (!is_gpu(sna, dst->pDrawable, PREFER_GPU_SPANS)) {
 			DBG(("%s: fallback, tiled operation not on GPU\n",
 			     __FUNCTION__));
 			return false;
@@ -3143,10 +3143,12 @@ bool gen2_render_init(struct sna *sna)
 	 */
 #if !NO_COMPOSITE
 	render->composite = gen2_render_composite;
+	render->prefer_gpu |= PREFER_GPU_RENDER;
 #endif
 #if !NO_COMPOSITE_SPANS
 	render->check_composite_spans = gen2_check_composite_spans;
 	render->composite_spans = gen2_render_composite_spans;
+	render->prefer_gpu |= PREFER_GPU_SPANS;
 #endif
 	render->fill_boxes = gen2_render_fill_boxes;
 	render->fill = gen2_render_fill;
diff --git a/src/sna/gen3_render.c b/src/sna/gen3_render.c
index 6b9a2f81..95d44ab5 100644
--- a/src/sna/gen3_render.c
+++ b/src/sna/gen3_render.c
@@ -2658,7 +2658,7 @@ source_use_blt(struct sna *sna, PicturePtr picture)
 	if (too_large(picture->pDrawable->width, picture->pDrawable->height))
 		return true;
 
-	return !is_gpu(picture->pDrawable);
+	return !is_gpu(sna, picture->pDrawable, PREFER_GPU_RENDER);
 }
 
 static bool
@@ -3884,12 +3884,11 @@ gen3_check_composite_spans(struct sna *sna,
 	if (gen3_composite_fallback(sna, op, src, NULL, dst))
 		return false;
 
-	if (need_tiling(sna, width, height)) {
-		if (!is_gpu(dst->pDrawable)) {
-			DBG(("%s: fallback, tiled operation not on GPU\n",
-			     __FUNCTION__));
-			return false;
-		}
+	if (need_tiling(sna, width, height) &&
+	    !is_gpu(sna, dst->pDrawable, PREFER_GPU_SPANS)) {
+		DBG(("%s: fallback, tiled operation not on GPU\n",
+		     __FUNCTION__));
+		return false;
 	}
 
 	return true;
@@ -5205,10 +5204,12 @@ bool gen3_render_init(struct sna *sna)
 
 #if !NO_COMPOSITE
 	render->composite = gen3_render_composite;
+	render->prefer_gpu |= PREFER_GPU_RENDER;
 #endif
 #if !NO_COMPOSITE_SPANS
 	render->check_composite_spans = gen3_check_composite_spans;
 	render->composite_spans = gen3_render_composite_spans;
+	render->prefer_gpu |= PREFER_GPU_SPANS;
 #endif
 
 	render->video = gen3_render_video;
diff --git a/src/sna/gen4_render.c b/src/sna/gen4_render.c
index cc1778ae..53fe52a9 100644
--- a/src/sna/gen4_render.c
+++ b/src/sna/gen4_render.c
@@ -1580,15 +1580,6 @@ gen4_composite_set_target(struct sna *sna,
 	return true;
 }
 
-static inline bool
-picture_is_cpu(PicturePtr picture)
-{
-	if (!picture->pDrawable)
-		return false;
-
-	return !is_gpu(picture->pDrawable);
-}
-
 static bool
 try_blt(struct sna *sna,
 	PicturePtr dst, PicturePtr src,
@@ -1613,7 +1604,7 @@ try_blt(struct sna *sna,
 		return true;
 
 	/* is the source picture only in cpu memory e.g. a shm pixmap? */
-	return picture_is_cpu(src);
+	return picture_is_cpu(sna, src);
 }
 
 static bool
@@ -1635,9 +1626,10 @@ has_alphamap(PicturePtr p)
 }
 
 static bool
-need_upload(PicturePtr p)
+need_upload(struct sna *sna, PicturePtr p)
 {
-	return p->pDrawable && untransformed(p) && !is_gpu(p->pDrawable);
+	return p->pDrawable && untransformed(p) &&
+		!is_gpu(sna, p->pDrawable, PREFER_GPU_RENDER);
 }
 
 static bool
@@ -1660,7 +1652,7 @@ source_is_busy(PixmapPtr pixmap)
 }
 
 static bool
-source_fallback(PicturePtr p, PixmapPtr pixmap)
+source_fallback(struct sna *sna, PicturePtr p, PixmapPtr pixmap)
 {
 	if (sna_picture_is_solid(p, NULL))
 		return false;
@@ -1675,7 +1667,7 @@ source_fallback(PicturePtr p, PixmapPtr pixmap)
 	if (pixmap && source_is_busy(pixmap))
 		return false;
 
-	return has_alphamap(p) || !gen4_check_filter(p) || need_upload(p);
+	return has_alphamap(p) || !gen4_check_filter(p) || need_upload(sna, p);
 }
 
 static bool
@@ -1698,11 +1690,11 @@ gen4_composite_fallback(struct sna *sna,
 	dst_pixmap = get_drawable_pixmap(dst->pDrawable);
 
 	src_pixmap = src->pDrawable ? get_drawable_pixmap(src->pDrawable) : NULL;
-	src_fallback = source_fallback(src, src_pixmap);
+	src_fallback = source_fallback(sna, src, src_pixmap);
 
 	if (mask) {
 		mask_pixmap = mask->pDrawable ? get_drawable_pixmap(mask->pDrawable) : NULL;
-		mask_fallback = source_fallback(mask, mask_pixmap);
+		mask_fallback = source_fallback(sna, mask, mask_pixmap);
 	} else {
 		mask_pixmap = NULL;
 		mask_fallback = false;
@@ -2098,7 +2090,8 @@ gen4_check_composite_spans(struct sna *sna,
 		return false;
 	}
 
-	if (need_tiling(sna, width, height) && !is_gpu(dst->pDrawable)) {
+	if (need_tiling(sna, width, height) &&
+	    !is_gpu(sna, dst->pDrawable, PREFER_GPU_SPANS)) {
 		DBG(("%s: fallback, tiled operation not on GPU\n",
 		     __FUNCTION__));
 		return false;
@@ -2108,24 +2101,20 @@ gen4_check_composite_spans(struct sna *sna,
 		return FORCE_SPANS > 0;
 
 	if ((flags & COMPOSITE_SPANS_RECTILINEAR) == 0) {
-		struct sna_pixmap *priv;
-
-		if (FORCE_NONRECTILINEAR_SPANS)
-			return FORCE_NONRECTILINEAR_SPANS > 0;
-
-		priv = sna_pixmap_from_drawable(dst->pDrawable);
+		struct sna_pixmap *priv = sna_pixmap_from_drawable(dst->pDrawable);
 		assert(priv);
 
 		if (priv->cpu_bo && kgem_bo_is_busy(priv->cpu_bo))
 			return true;
 
-		if ((flags & COMPOSITE_SPANS_INPLACE_HINT) == 0 &&
-		    priv->gpu_bo && kgem_bo_is_busy(priv->gpu_bo))
-			return true;
+		if (flags & COMPOSITE_SPANS_INPLACE_HINT)
+			return false;
 
-		DBG(("%s: fallback, non-rectilinear spans to idle bo\n",
-		     __FUNCTION__));
-		return false;
+		if ((sna->render.prefer_gpu & PREFER_GPU_SPANS) == 0 &&
+		    dst->format == PICT_a8)
+			return false;
+
+		return priv->gpu_bo && kgem_bo_is_busy(priv->gpu_bo);
 	}
 
 	return true;
@@ -3124,10 +3113,13 @@ bool gen4_render_init(struct sna *sna)
 
 #if !NO_COMPOSITE
 	sna->render.composite = gen4_render_composite;
+	sna->render.prefer_gpu |= PREFER_GPU_RENDER;
 #endif
 #if !NO_COMPOSITE_SPANS
 	sna->render.check_composite_spans = gen4_check_composite_spans;
 	sna->render.composite_spans = gen4_render_composite_spans;
+	if (0)
+		sna->render.prefer_gpu |= PREFER_GPU_SPANS;
 #endif
 
 #if !NO_VIDEO
diff --git a/src/sna/gen5_render.c b/src/sna/gen5_render.c
index 35f70f83..6e119963 100644
--- a/src/sna/gen5_render.c
+++ b/src/sna/gen5_render.c
@@ -1556,15 +1556,6 @@ gen5_composite_set_target(struct sna *sna,
 	return true;
 }
 
-static inline bool
-picture_is_cpu(PicturePtr picture)
-{
-	if (!picture->pDrawable)
-		return false;
-
-	return !is_gpu(picture->pDrawable);
-}
-
 static bool
 try_blt(struct sna *sna,
 	PicturePtr dst, PicturePtr src,
@@ -1589,7 +1580,7 @@ try_blt(struct sna *sna,
 		return true;
 
 	/* is the source picture only in cpu memory e.g. a shm pixmap? */
-	return picture_is_cpu(src);
+	return picture_is_cpu(sna, src);
 }
 
 static bool
@@ -1614,9 +1605,10 @@ has_alphamap(PicturePtr p)
 }
 
 static bool
-need_upload(PicturePtr p)
+need_upload(struct sna *sna, PicturePtr p)
 {
-	return p->pDrawable && untransformed(p) && !is_gpu(p->pDrawable);
+	return p->pDrawable && untransformed(p) &&
+		!is_gpu(sna, p->pDrawable, PREFER_GPU_RENDER);
 }
 
 static bool
@@ -1639,7 +1631,7 @@ source_is_busy(PixmapPtr pixmap)
 }
 
 static bool
-source_fallback(PicturePtr p, PixmapPtr pixmap)
+source_fallback(struct sna *sna, PicturePtr p, PixmapPtr pixmap)
 {
 	if (sna_picture_is_solid(p, NULL))
 		return false;
@@ -1652,7 +1644,7 @@ source_fallback(PicturePtr p, PixmapPtr pixmap)
 	if (pixmap && source_is_busy(pixmap))
 		return false;
 
-	return has_alphamap(p) || !gen5_check_filter(p) || need_upload(p);
+	return has_alphamap(p) || !gen5_check_filter(p) || need_upload(sna, p);
 }
 
 static bool
@@ -1675,11 +1667,11 @@ gen5_composite_fallback(struct sna *sna,
 	dst_pixmap = get_drawable_pixmap(dst->pDrawable);
 
 	src_pixmap = src->pDrawable ? get_drawable_pixmap(src->pDrawable) : NULL;
-	src_fallback = source_fallback(src, src_pixmap);
+	src_fallback = source_fallback(sna, src, src_pixmap);
 
 	if (mask) {
 		mask_pixmap = mask->pDrawable ? get_drawable_pixmap(mask->pDrawable) : NULL;
-		mask_fallback = source_fallback(mask, mask_pixmap);
+		mask_fallback = source_fallback(sna, mask, mask_pixmap);
 	} else {
 		mask_pixmap = NULL;
 		mask_fallback = false;
@@ -2075,7 +2067,8 @@ gen5_check_composite_spans(struct sna *sna,
 		return false;
 	}
 
-	if (need_tiling(sna, width, height) && !is_gpu(dst->pDrawable)) {
+	if (need_tiling(sna, width, height) &&
+	    !is_gpu(sna, dst->pDrawable, PREFER_GPU_SPANS)) {
 		DBG(("%s: fallback, tiled operation not on GPU\n",
 		     __FUNCTION__));
 		return false;
@@ -2088,13 +2081,14 @@ gen5_check_composite_spans(struct sna *sna,
 		if (priv->cpu_bo && kgem_bo_is_busy(priv->cpu_bo))
 			return true;
 
-		if ((flags & COMPOSITE_SPANS_INPLACE_HINT) == 0 &&
-		    priv->gpu_bo && kgem_bo_is_busy(priv->gpu_bo))
-			return true;
+		if (flags & COMPOSITE_SPANS_INPLACE_HINT)
+			return false;
 
-		DBG(("%s: fallback, non-rectilinear spans to idle bo\n",
-		     __FUNCTION__));
-		return false;
+		if ((sna->render.prefer_gpu & PREFER_GPU_SPANS) == 0 &&
+		    dst->format == PICT_a8)
+			return false;
+
+		return priv->gpu_bo && kgem_bo_is_busy(priv->gpu_bo);
 	}
 
 	return true;
@@ -3272,10 +3266,13 @@ bool gen5_render_init(struct sna *sna)
 
 #if !NO_COMPOSITE
 	sna->render.composite = gen5_render_composite;
+	sna->render.prefer_gpu |= PREFER_GPU_RENDER;
 #endif
 #if !NO_COMPOSITE_SPANS
 	sna->render.check_composite_spans = gen5_check_composite_spans;
 	sna->render.composite_spans = gen5_render_composite_spans;
+	if (DEVICE_ID(sna->PciInfo) == 0x0044)
+		sna->render.prefer_gpu |= PREFER_GPU_SPANS;
 #endif
 	sna->render.video = gen5_render_video;
 
diff --git a/src/sna/gen6_render.c b/src/sna/gen6_render.c
index ec642504..76d0c8ca 100644
--- a/src/sna/gen6_render.c
+++ b/src/sna/gen6_render.c
@@ -2422,7 +2422,8 @@ gen6_check_composite_spans(struct sna *sna,
 		return false;
 	}
 
-	if (need_tiling(sna, width, height) && !is_gpu(dst->pDrawable)) {
+	if (need_tiling(sna, width, height) &&
+	    !is_gpu(sna, dst->pDrawable, PREFER_GPU_SPANS)) {
 		DBG(("%s: fallback, tiled operation not on GPU\n",
 		     __FUNCTION__));
 		return false;
@@ -2435,13 +2436,10 @@ gen6_check_composite_spans(struct sna *sna,
 		if (priv->cpu_bo && kgem_bo_is_busy(priv->cpu_bo))
 			return true;
 
-		if ((flags & COMPOSITE_SPANS_INPLACE_HINT) == 0 &&
-		    priv->gpu_bo && kgem_bo_is_busy(priv->gpu_bo))
-			return true;
+		if (flags & COMPOSITE_SPANS_INPLACE_HINT)
+			return false;
 
-		DBG(("%s: fallback, non-rectilinear spans to idle bo\n",
-		     __FUNCTION__));
-		return false;
+		return priv->gpu_bo && kgem_bo_is_busy(priv->gpu_bo);
 	}
 
 	return true;
@@ -3557,6 +3555,16 @@ static void gen6_render_fini(struct sna *sna)
 	kgem_bo_destroy(&sna->kgem, sna->render_state.gen6.general_bo);
 }
 
+static bool is_gt2(struct sna *sna)
+{
+	return DEVICE_ID(sna->PciInfo) & 0x20;
+}
+
+static bool is_mobile(struct sna *sna)
+{
+	return (DEVICE_ID(sna->PciInfo) & 0xf) == 0x6;
+}
+
 static bool gen6_render_setup(struct sna *sna)
 {
 	struct gen6_render_state *state = &sna->render_state.gen6;
@@ -3565,7 +3573,7 @@ static bool gen6_render_setup(struct sna *sna)
 	int i, j, k, l, m;
 
 	state->info = &gt1_info;
-	if (DEVICE_ID(sna->PciInfo) & 0x20)
+	if (is_gt2(sna))
 		state->info = &gt2_info; /* XXX requires GT_MODE WiZ disabled */
 
 	sna_static_stream_init(&general);
@@ -3646,10 +3654,14 @@ bool gen6_render_init(struct sna *sna)
 
 #if !NO_COMPOSITE
 	sna->render.composite = gen6_render_composite;
+	sna->render.prefer_gpu |= PREFER_GPU_RENDER;
+
 #endif
 #if !NO_COMPOSITE_SPANS
 	sna->render.check_composite_spans = gen6_check_composite_spans;
 	sna->render.composite_spans = gen6_render_composite_spans;
+	if (is_mobile(sna))
+		sna->render.prefer_gpu |= PREFER_GPU_SPANS;
 #endif
 	sna->render.video = gen6_render_video;
 
diff --git a/src/sna/gen7_render.c b/src/sna/gen7_render.c
index 59b38f6c..f05d6f92 100644
--- a/src/sna/gen7_render.c
+++ b/src/sna/gen7_render.c
@@ -2553,12 +2553,11 @@ gen7_check_composite_spans(struct sna *sna,
 	if (gen7_composite_fallback(sna, src, NULL, dst))
 		return false;
 
-	if (need_tiling(sna, width, height)) {
-		if (!is_gpu(dst->pDrawable)) {
-			DBG(("%s: fallback, tiled operation not on GPU\n",
-			     __FUNCTION__));
-			return false;
-		}
+	if (need_tiling(sna, width, height) &&
+	    !is_gpu(sna, dst->pDrawable, PREFER_GPU_SPANS)) {
+		DBG(("%s: fallback, tiled operation not on GPU\n",
+		     __FUNCTION__));
+		return false;
 	}
 
 	return true;
@@ -3647,6 +3646,16 @@ static void gen7_render_fini(struct sna *sna)
 	kgem_bo_destroy(&sna->kgem, sna->render_state.gen7.general_bo);
 }
 
+static bool is_gt2(struct sna *sna)
+{
+	return DEVICE_ID(sna->PciInfo) & 0x20;
+}
+
+static bool is_mobile(struct sna *sna)
+{
+	return (DEVICE_ID(sna->PciInfo) & 0xf) == 0x6;
+}
+
 static bool gen7_render_setup(struct sna *sna)
 {
 	struct gen7_render_state *state = &sna->render_state.gen7;
@@ -3658,14 +3667,14 @@ static bool gen7_render_setup(struct sna *sna)
 		state->info = &ivb_gt_info;
 		if (DEVICE_ID(sna->PciInfo) & 0xf) {
 			state->info = &ivb_gt1_info;
-			if (DEVICE_ID(sna->PciInfo) & 0x20)
+			if (is_gt2(sna))
 				state->info = &ivb_gt2_info; /* XXX requires GT_MODE WiZ disabled */
 		}
 	} else if (sna->kgem.gen == 075) {
 		state->info = &hsw_gt_info;
 		if (DEVICE_ID(sna->PciInfo) & 0xf) {
 			state->info = &hsw_gt1_info;
-			if (DEVICE_ID(sna->PciInfo) & 0x20)
+			if (is_gt2(sna))
 				state->info = &hsw_gt2_info;
 		}
 	} else
@@ -3744,10 +3753,13 @@ bool gen7_render_init(struct sna *sna)
 
 #if !NO_COMPOSITE
 	sna->render.composite = gen7_render_composite;
+	sna->render.prefer_gpu |= PREFER_GPU_RENDER;
 #endif
 #if !NO_COMPOSITE_SPANS
 	sna->render.check_composite_spans = gen7_check_composite_spans;
 	sna->render.composite_spans = gen7_render_composite_spans;
+	if (is_mobile(sna))
+		sna->render.prefer_gpu |= PREFER_GPU_SPANS;
 #endif
 	sna->render.video = gen7_render_video;
 
diff --git a/src/sna/sna.h b/src/sna/sna.h
index e4073596..5a57a6e2 100644
--- a/src/sna/sna.h
+++ b/src/sna/sna.h
@@ -256,7 +256,6 @@ struct sna {
 		struct gen6_render_state gen6;
 		struct gen7_render_state gen7;
 	} render_state;
-	uint32_t have_render;
 
 	bool dri_available;
 	bool dri_open;
@@ -672,7 +671,7 @@ static inline bool wedged(struct sna *sna)
 
 static inline bool can_render(struct sna *sna)
 {
-	return likely(!sna->kgem.wedged && sna->have_render);
+	return likely(!sna->kgem.wedged && sna->render.prefer_gpu & PREFER_GPU_RENDER);
 }
 
 static inline uint32_t pixmap_size(PixmapPtr pixmap)
diff --git a/src/sna/sna_accel.c b/src/sna/sna_accel.c
index 7c8a3609..0edc1e8b 100644
--- a/src/sna/sna_accel.c
+++ b/src/sna/sna_accel.c
@@ -879,7 +879,8 @@ sna_pixmap_create_scratch(ScreenPtr screen,
 	     width, height, depth, tiling));
 
 	bpp = bits_per_pixel(depth);
-	if (tiling == I915_TILING_Y && !sna->have_render)
+	if (tiling == I915_TILING_Y &&
+	    (sna->render.prefer_gpu & PREFER_GPU_RENDER) == 0)
 		tiling = I915_TILING_X;
 
 	if (tiling == I915_TILING_Y &&
@@ -1180,7 +1181,7 @@ static PixmapPtr sna_create_pixmap(ScreenPtr screen,
 		goto fallback;
 	}
 
-	if (unlikely(!sna->have_render))
+	if (unlikely((sna->render.prefer_gpu & PREFER_GPU_RENDER) == 0))
 		flags &= ~KGEM_CAN_CREATE_GPU;
 	if (wedged(sna))
 		flags = 0;
@@ -14199,31 +14200,30 @@ bool sna_accel_init(ScreenPtr screen, struct sna *sna)
 		return false;
 
 	backend = "no";
-	sna->have_render = false;
 	no_render_init(sna);
 
 	if (sna_option_accel_blt(sna) || sna->info->gen >= 0100) {
 	} else if (sna->info->gen >= 070) {
-		if ((sna->have_render = gen7_render_init(sna)))
+		if (gen7_render_init(sna))
 			backend = "IvyBridge";
 	} else if (sna->info->gen >= 060) {
-		if ((sna->have_render = gen6_render_init(sna)))
+		if (gen6_render_init(sna))
 			backend = "SandyBridge";
 	} else if (sna->info->gen >= 050) {
-		if ((sna->have_render = gen5_render_init(sna)))
+		if (gen5_render_init(sna))
 			backend = "Ironlake";
 	} else if (sna->info->gen >= 040) {
-		if ((sna->have_render = gen4_render_init(sna)))
+		if (gen4_render_init(sna))
 			backend = "Broadwater/Crestline";
 	} else if (sna->info->gen >= 030) {
-		if ((sna->have_render = gen3_render_init(sna)))
+		if (gen3_render_init(sna))
 			backend = "gen3";
 	} else if (sna->info->gen >= 020) {
-		if ((sna->have_render = gen2_render_init(sna)))
+		if (gen2_render_init(sna))
 			backend = "gen2";
 	}
-	DBG(("%s(backend=%s, have_render=%d)\n",
-	     __FUNCTION__, backend, sna->have_render));
+	DBG(("%s(backend=%s, prefer_gpu=%x)\n",
+	     __FUNCTION__, backend, sna->render.prefer_gpu));
 
 	kgem_reset(&sna->kgem);
 
@@ -14252,7 +14252,6 @@ void sna_accel_create(struct sna *sna)
 fail:
 	xf86DrvMsg(sna->scrn->scrnIndex, X_ERROR,
 		   "Failed to allocate caches, disabling RENDER acceleration\n");
-	sna->have_render = false;
 	no_render_init(sna);
 }
 
diff --git a/src/sna/sna_composite.c b/src/sna/sna_composite.c
index e82d5f49..c329fb8b 100644
--- a/src/sna/sna_composite.c
+++ b/src/sna/sna_composite.c
@@ -584,7 +584,7 @@ sna_composite(CARD8 op,
 	}
 
 	if (use_cpu(pixmap, priv, op, width, height) &&
-	    !picture_is_gpu(src) && !picture_is_gpu(mask)) {
+	    !picture_is_gpu(sna, src) && !picture_is_gpu(sna, mask)) {
 		DBG(("%s: fallback, dst pixmap=%ld is too small (or completely damaged)\n",
 		     __FUNCTION__, pixmap->drawable.serialNumber));
 		goto fallback;
diff --git a/src/sna/sna_glyphs.c b/src/sna/sna_glyphs.c
index 130c1eaa..ad050df6 100644
--- a/src/sna/sna_glyphs.c
+++ b/src/sna/sna_glyphs.c
@@ -193,11 +193,17 @@ bool sna_glyphs_create(struct sna *sna)
 	if (sna->render.white_image == NULL)
 		goto bail;
 
-	if (!can_render(sna))
+	if (!can_render(sna)) {
+		DBG(("%s: no render acceleration, no render glyph caches\n",
+		     __FUNCTION__));
 		return true;
+	}
 
-	if (xf86IsEntityShared(sna->scrn->entityList[0]))
+	if (xf86IsEntityShared(sna->scrn->entityList[0])) {
+		DBG(("%s: shared GlyphPictures, no render glyph caches\n",
+		     __FUNCTION__));
 		return true;
+	}
 
 	for (i = 0; i < ARRAY_SIZE(formats); i++) {
 		struct sna_glyph_cache *cache = &sna->render.glyph[i];
@@ -218,8 +224,11 @@ bool sna_glyphs_create(struct sna *sna)
 					      CACHE_PICTURE_SIZE,
 					      depth,
 					      SNA_CREATE_GLYPHS);
-		if (!pixmap)
+		if (!pixmap) {
+			DBG(("%s: failed to allocate pixmap for Glyph cache\n",
+			     __FUNCTION__));
 			goto bail;
+		}
 
 		priv = sna_pixmap(pixmap);
 		if (priv != NULL) {
@@ -1678,7 +1687,7 @@ sna_glyphs(CARD8 op,
 	}
 
 	if ((too_small(priv) || DAMAGE_IS_ALL(priv->cpu_damage)) &&
-	    !picture_is_gpu(src)) {
+	    !picture_is_gpu(sna, src)) {
 		DBG(("%s: fallback -- too small (%dx%d)\n",
 		     __FUNCTION__, dst->pDrawable->width, dst->pDrawable->height));
 		goto fallback;
@@ -2002,7 +2011,7 @@ sna_glyphs__shared(CARD8 op,
 	}
 
 	if ((too_small(priv) || DAMAGE_IS_ALL(priv->cpu_damage)) &&
-	    !picture_is_gpu(src)) {
+	    !picture_is_gpu(sna, src)) {
 		DBG(("%s: fallback -- too small (%dx%d)\n",
 		     __FUNCTION__, dst->pDrawable->width, dst->pDrawable->height));
 		goto fallback;
diff --git a/src/sna/sna_render.c b/src/sna/sna_render.c
index 7784a5b1..6c5d962d 100644
--- a/src/sna/sna_render.c
+++ b/src/sna/sna_render.c
@@ -87,8 +87,8 @@ no_render_composite(struct sna *sna,
 	if (mask)
 		return false;
 
-	if (!is_gpu(dst->pDrawable) &&
-	    (src->pDrawable == NULL || !is_gpu(src->pDrawable)))
+	if (!is_gpu(sna, dst->pDrawable, PREFER_GPU_BLT) &&
+	    (src->pDrawable == NULL || !is_gpu(sna, src->pDrawable, PREFER_GPU_BLT)))
 		return false;
 
 	return sna_blt_composite(sna,
@@ -279,7 +279,9 @@ void no_render_init(struct sna *sna)
 {
 	struct sna_render *render = &sna->render;
 
-	memset (render,0, sizeof (*render));
+	memset (render, 0, sizeof (*render));
+
+	render->prefer_gpu = PREFER_GPU_BLT;
 
 	render->vertices = render->vertex_data;
 	render->vertex_size = ARRAY_SIZE(render->vertex_data);
@@ -1533,7 +1535,8 @@ sna_render_picture_fixup(struct sna *sna,
 
 	if (picture->alphaMap) {
 		DBG(("%s: alphamap\n", __FUNCTION__));
-		if (is_gpu(picture->pDrawable) || is_gpu(picture->alphaMap->pDrawable)) {
+		if (is_gpu(sna, picture->pDrawable, PREFER_GPU_RENDER) ||
+		    is_gpu(sna, picture->alphaMap->pDrawable, PREFER_GPU_RENDER)) {
 			return sna_render_picture_flatten(sna, picture, channel,
 							  x, y, w, h, dst_x, dst_y);
 		}
@@ -1543,7 +1546,7 @@ sna_render_picture_fixup(struct sna *sna,
 
 	if (picture->filter == PictFilterConvolution) {
 		DBG(("%s: convolution\n", __FUNCTION__));
-		if (is_gpu(picture->pDrawable)) {
+		if (is_gpu(sna, picture->pDrawable, PREFER_GPU_RENDER)) {
 			return sna_render_picture_convolve(sna, picture, channel,
 							   x, y, w, h, dst_x, dst_y);
 		}
@@ -1704,7 +1707,7 @@ sna_render_picture_convert(struct sna *sna,
 		return 0;
 	}
 
-	if (fixup_alpha && is_gpu(&pixmap->drawable)) {
+	if (fixup_alpha && is_gpu(sna, &pixmap->drawable, PREFER_GPU_RENDER)) {
 		ScreenPtr screen = pixmap->drawable.pScreen;
 		PixmapPtr tmp;
 		PicturePtr src, dst;
diff --git a/src/sna/sna_render.h b/src/sna/sna_render.h
index a8084f27..01176c6a 100644
--- a/src/sna/sna_render.h
+++ b/src/sna/sna_render.h
@@ -215,6 +215,11 @@ struct sna_render {
 	int max_3d_size;
 	int max_3d_pitch;
 
+	unsigned prefer_gpu;
+#define PREFER_GPU_BLT 0x1
+#define PREFER_GPU_RENDER 0x2
+#define PREFER_GPU_SPANS 0x4
+
 	bool (*composite)(struct sna *sna, uint8_t op,
 			  PicturePtr dst, PicturePtr src, PicturePtr mask,
 			  int16_t src_x, int16_t src_y,
diff --git a/src/sna/sna_render_inline.h b/src/sna/sna_render_inline.h
index 432201f5..7d9f2cac 100644
--- a/src/sna/sna_render_inline.h
+++ b/src/sna/sna_render_inline.h
@@ -67,7 +67,7 @@ static inline void batch_emit_float(struct sna *sna, float f)
 }
 
 static inline bool
-is_gpu(DrawablePtr drawable)
+is_gpu(struct sna *sna, DrawablePtr drawable, unsigned prefer)
 {
 	struct sna_pixmap *priv = sna_pixmap_from_drawable(drawable);
 
@@ -77,7 +77,8 @@ is_gpu(DrawablePtr drawable)
 	if (priv->cpu_damage == NULL)
 		return true;
 
-	if (priv->gpu_damage && !priv->gpu_bo->proxy)
+	if (priv->gpu_damage && !priv->gpu_bo->proxy &&
+	    (sna->render.prefer_gpu & prefer))
 		return true;
 
 	if (priv->cpu_bo && kgem_bo_is_busy(priv->cpu_bo))
@@ -108,11 +109,20 @@ unattached(DrawablePtr drawable)
 }
 
 static inline bool
-picture_is_gpu(PicturePtr picture)
+picture_is_gpu(struct sna *sna, PicturePtr picture)
 {
 	if (!picture || !picture->pDrawable)
 		return false;
-	return is_gpu(picture->pDrawable);
+	return is_gpu(sna, picture->pDrawable, PREFER_GPU_RENDER);
+}
+
+static inline bool
+picture_is_cpu(struct sna *sna, PicturePtr picture)
+{
+	if (!picture->pDrawable)
+		return false;
+
+	return !is_gpu(sna, picture->pDrawable, PREFER_GPU_RENDER);
 }
 
 static inline bool sna_blt_compare_depth(DrawablePtr src, DrawablePtr dst)
diff --git a/src/sna/sna_trapezoids.c b/src/sna/sna_trapezoids.c
index 547e4f53..406a6b31 100644
--- a/src/sna/sna_trapezoids.c
+++ b/src/sna/sna_trapezoids.c
@@ -2457,7 +2457,8 @@ is_mono(PicturePtr dst, PictFormatPtr mask)
 }
 
 static bool
-trapezoids_inplace_fallback(CARD8 op,
+trapezoids_inplace_fallback(struct sna *sna,
+			    CARD8 op,
 			    PicturePtr src, PicturePtr dst, PictFormatPtr mask,
 			    int ntrap, xTrapezoid *traps)
 {
@@ -2497,7 +2498,7 @@ trapezoids_inplace_fallback(CARD8 op,
 		return false;
 	}
 
-	if (is_gpu(dst->pDrawable)) {
+	if (is_gpu(sna, dst->pDrawable, PREFER_GPU_SPANS)) {
 		DBG(("%s: not performing inplace as dst is already on the GPU\n",
 		     __FUNCTION__));
 		return false;
@@ -2581,7 +2582,8 @@ static void rasterize_traps_thread(void *arg)
 }
 
 static void
-trapezoids_fallback(CARD8 op, PicturePtr src, PicturePtr dst,
+trapezoids_fallback(struct sna *sna,
+		    CARD8 op, PicturePtr src, PicturePtr dst,
 		    PictFormatPtr maskFormat, INT16 xSrc, INT16 ySrc,
 		    int ntrap, xTrapezoid * traps)
 {
@@ -2636,7 +2638,8 @@ trapezoids_fallback(CARD8 op, PicturePtr src, PicturePtr dst,
 
 		DBG(("%s: mask (%dx%d) depth=%d, format=%08x\n",
 		     __FUNCTION__, width, height, depth, format));
-		if (is_gpu(dst->pDrawable) || picture_is_gpu(src)) {
+		if (is_gpu(sna, dst->pDrawable, PREFER_GPU_RENDER) ||
+		    picture_is_gpu(sna, src)) {
 			int num_threads;
 
 			scratch = sna_pixmap_create_upload(screen,
@@ -2760,7 +2763,7 @@ trapezoids_fallback(CARD8 op, PicturePtr src, PicturePtr dst,
 			maskFormat = PictureMatchFormat(screen, 8, PICT_a8);
 
 		for (; ntrap; ntrap--, traps++)
-			trapezoids_fallback(op,
+			trapezoids_fallback(sna, op,
 					    src, dst, maskFormat,
 					    xSrc, ySrc, 1, traps);
 	}
@@ -3420,7 +3423,8 @@ pixsolid_unaligned_box_row(struct pixman_inplace *pi,
 }
 
 static bool
-composite_unaligned_boxes_inplace__solid(CARD8 op, uint32_t color,
+composite_unaligned_boxes_inplace__solid(struct sna *sna,
+					 CARD8 op, uint32_t color,
 					 PicturePtr dst, int n, xTrapezoid *t,
 					 bool force_fallback)
 {
@@ -3428,9 +3432,9 @@ composite_unaligned_boxes_inplace__solid(CARD8 op, uint32_t color,
 	int16_t dx, dy;
 
 	DBG(("%s: force=%d, is_gpu=%d, op=%d, color=%x\n", __FUNCTION__,
-	     force_fallback, is_gpu(dst->pDrawable), op, color));
+	     force_fallback, is_gpu(sna, dst->pDrawable, PREFER_GPU_SPANS), op, color));
 
-	if (!force_fallback && is_gpu(dst->pDrawable)) {
+	if (!force_fallback && is_gpu(sna, dst->pDrawable, PREFER_GPU_SPANS)) {
 		DBG(("%s: fallback -- can not perform operation in place, destination busy\n",
 		     __FUNCTION__));
 
@@ -3744,13 +3748,15 @@ static void rectilinear_inplace_thread(void *arg)
 }
 
 static bool
-composite_unaligned_boxes_inplace(CARD8 op,
+composite_unaligned_boxes_inplace(struct sna *sna,
+				  CARD8 op,
 				  PicturePtr src, int16_t src_x, int16_t src_y,
 				  PicturePtr dst, int n, xTrapezoid *t,
 				  bool force_fallback)
 {
 	if (!force_fallback &&
-	    (is_gpu(dst->pDrawable) || picture_is_gpu(src))) {
+	    (is_gpu(sna, dst->pDrawable, PREFER_GPU_SPANS) ||
+	     picture_is_gpu(sna, src))) {
 		DBG(("%s: fallback -- not forcing\n", __FUNCTION__));
 		return false;
 	}
@@ -3894,7 +3900,8 @@ composite_unaligned_boxes_inplace(CARD8 op,
 }
 
 static bool
-composite_unaligned_boxes_fallback(CARD8 op,
+composite_unaligned_boxes_fallback(struct sna *sna,
+				   CARD8 op,
 				   PicturePtr src,
 				   PicturePtr dst,
 				   INT16 src_x, INT16 src_y,
@@ -3908,12 +3915,12 @@ composite_unaligned_boxes_fallback(CARD8 op,
 	int n;
 
 	if (sna_picture_is_solid(src, &color) &&
-	    composite_unaligned_boxes_inplace__solid(op, color, dst,
+	    composite_unaligned_boxes_inplace__solid(sna, op, color, dst,
 						     ntrap, traps,
 						     force_fallback))
 		return true;
 
-	if (composite_unaligned_boxes_inplace(op, src, src_x, src_y,
+	if (composite_unaligned_boxes_inplace(sna, op, src, src_x, src_y,
 					      dst, ntrap, traps,
 					      force_fallback))
 		return true;
@@ -4037,7 +4044,7 @@ composite_unaligned_boxes(struct sna *sna,
 	    !sna->render.check_composite_spans(sna, op, src, dst, 0, 0,
 					       COMPOSITE_SPANS_RECTILINEAR)) {
 fallback:
-		return composite_unaligned_boxes_fallback(op, src, dst,
+		return composite_unaligned_boxes_fallback(sna, op, src, dst,
 							  src_x, src_y,
 							  ntrap, traps,
 							  force_fallback);
@@ -4285,7 +4292,8 @@ mono_span_thread(void *arg)
 }
 
 static bool
-mono_trapezoids_span_converter(CARD8 op, PicturePtr src, PicturePtr dst,
+mono_trapezoids_span_converter(struct sna *sna,
+			       CARD8 op, PicturePtr src, PicturePtr dst,
 			       INT16 src_x, INT16 src_y,
 			       int ntrap, xTrapezoid *traps)
 {
@@ -4336,7 +4344,7 @@ mono_trapezoids_span_converter(CARD8 op, PicturePtr src, PicturePtr dst,
 	unbounded = (!sna_drawable_is_clear(dst->pDrawable) &&
 		     !operator_is_bounded(op));
 
-	mono.sna = to_sna_from_drawable(dst->pDrawable);
+	mono.sna = sna;
 	if (!mono.sna->render.composite(mono.sna, op, src, NULL, dst,
 				       src_x + mono.clip.extents.x1 - dst_x - dx,
 				       src_y + mono.clip.extents.y1 - dst_y - dy,
@@ -4620,12 +4628,12 @@ span_thread(void *arg)
 }
 
 static bool
-trapezoid_span_converter(CARD8 op, PicturePtr src, PicturePtr dst,
+trapezoid_span_converter(struct sna *sna,
+			 CARD8 op, PicturePtr src, PicturePtr dst,
 			 PictFormatPtr maskFormat, unsigned int flags,
 			 INT16 src_x, INT16 src_y,
 			 int ntrap, xTrapezoid *traps)
 {
-	struct sna *sna;
 	struct sna_composite_spans_op tmp;
 	BoxRec extents;
 	pixman_region16_t clip;
@@ -4638,7 +4646,7 @@ trapezoid_span_converter(CARD8 op, PicturePtr src, PicturePtr dst,
 		return false;
 
 	if (is_mono(dst, maskFormat))
-		return mono_trapezoids_span_converter(op, src, dst,
+		return mono_trapezoids_span_converter(sna, op, src, dst,
 						      src_x, src_y,
 						      ntrap, traps);
 
@@ -4649,7 +4657,6 @@ trapezoid_span_converter(CARD8 op, PicturePtr src, PicturePtr dst,
 		return false;
 	}
 
-	sna = to_sna_from_drawable(dst->pDrawable);
 	if (!sna->render.check_composite_spans(sna, op, src, dst, 0, 0, flags)) {
 		DBG(("%s: fallback -- composite spans not supported\n",
 		     __FUNCTION__));
@@ -5315,7 +5322,8 @@ mono_inplace_composite_boxes(struct sna *sna,
 }
 
 static bool
-trapezoid_spans_maybe_inplace(CARD8 op, PicturePtr src, PicturePtr dst,
+trapezoid_spans_maybe_inplace(struct sna *sna,
+			      CARD8 op, PicturePtr src, PicturePtr dst,
 			      PictFormatPtr maskFormat)
 {
 	struct sna_pixmap *priv;
@@ -5348,7 +5356,7 @@ trapezoid_spans_maybe_inplace(CARD8 op, PicturePtr src, PicturePtr dst,
 
 	case PICT_x8r8g8b8:
 	case PICT_a8r8g8b8:
-		if (picture_is_gpu(src))
+		if (picture_is_gpu(sna, src))
 			return false;
 
 		switch (op) {
@@ -5395,7 +5403,8 @@ out:
 }
 
 static bool
-trapezoid_span_mono_inplace(CARD8 op,
+trapezoid_span_mono_inplace(struct sna *sna,
+			    CARD8 op,
 			    PicturePtr src,
 			    PicturePtr dst,
 			    INT16 src_x, INT16 src_y,
@@ -5441,7 +5450,7 @@ trapezoid_span_mono_inplace(CARD8 op,
 					     MOVE_WRITE | MOVE_READ))
 		return true;
 
-	mono.sna = to_sna_from_drawable(dst->pDrawable);
+	mono.sna = sna;
 	if (!mono_init(&mono, 2*ntrap))
 		return false;
 
@@ -6052,7 +6061,8 @@ static void inplace_thread(void *arg)
 }
 
 static bool
-trapezoid_span_inplace(CARD8 op, PicturePtr src, PicturePtr dst,
+trapezoid_span_inplace(struct sna *sna,
+		       CARD8 op, PicturePtr src, PicturePtr dst,
 		       PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
 		       int ntrap, xTrapezoid *traps,
 		       bool fallback)
@@ -6082,7 +6092,7 @@ trapezoid_span_inplace(CARD8 op, PicturePtr src, PicturePtr dst,
 		return false;
 	}
 
-	if (!fallback && is_gpu(dst->pDrawable)) {
+	if (!fallback && is_gpu(sna, dst->pDrawable, PREFER_GPU_SPANS)) {
 		DBG(("%s: fallback -- can not perform operation in place, destination busy\n",
 		     __FUNCTION__));
 
@@ -6090,7 +6100,7 @@ trapezoid_span_inplace(CARD8 op, PicturePtr src, PicturePtr dst,
 	}
 
 	if (is_mono(dst, maskFormat))
-		return trapezoid_span_mono_inplace(op, src, dst,
+		return trapezoid_span_mono_inplace(sna, op, src, dst,
 						   src_x, src_y, ntrap, traps);
 
 	if (dst->format == PICT_a8r8g8b8 || dst->format == PICT_x8r8g8b8)
@@ -6165,7 +6175,7 @@ trapezoid_span_inplace(CARD8 op, PicturePtr src, PicturePtr dst,
 		     __FUNCTION__));
 		do {
 			/* XXX unwind errors? */
-			if (!trapezoid_span_inplace(op, src, dst, NULL,
+			if (!trapezoid_span_inplace(sna, op, src, dst, NULL,
 						    src_x, src_y, 1, traps++,
 						    fallback))
 				return false;
@@ -6476,14 +6486,14 @@ sna_composite_trapezoids(CARD8 op,
 
 	force_fallback = FORCE_FALLBACK > 0;
 	if ((too_small(priv) || DAMAGE_IS_ALL(priv->cpu_damage)) &&
-	    !picture_is_gpu(src) && untransformed(src)) {
+	    !picture_is_gpu(sna, src) && untransformed(src)) {
 		DBG(("%s: force fallbacks --too small, %dx%d? %d, all-cpu? %d, src-is-cpu? %d\n",
 		     __FUNCTION__,
 		     dst->pDrawable->width,
 		     dst->pDrawable->height,
 		     too_small(priv),
 		     (int)DAMAGE_IS_ALL(priv->cpu_damage),
-		     !picture_is_gpu(src)));
+		     !picture_is_gpu(sna, src)));
 		force_fallback = true;
 	}
 	if (FORCE_FALLBACK < 0)
@@ -6550,24 +6560,24 @@ sna_composite_trapezoids(CARD8 op,
 		goto fallback;
 
 	if (is_mono(dst, maskFormat) &&
-	    mono_trapezoids_span_converter(op, src, dst,
+	    mono_trapezoids_span_converter(sna, op, src, dst,
 					   xSrc, ySrc,
 					   ntrap, traps))
 		return;
 
-	if (trapezoid_spans_maybe_inplace(op, src, dst, maskFormat)) {
+	if (trapezoid_spans_maybe_inplace(sna, op, src, dst, maskFormat)) {
 		flags |= COMPOSITE_SPANS_INPLACE_HINT;
-		if (trapezoid_span_inplace(op, src, dst, maskFormat,
+		if (trapezoid_span_inplace(sna, op, src, dst, maskFormat,
 					   xSrc, ySrc, ntrap, traps,
 					   false))
 			return;
 	}
 
-	if (trapezoid_span_converter(op, src, dst, maskFormat, flags,
+	if (trapezoid_span_converter(sna, op, src, dst, maskFormat, flags,
 				     xSrc, ySrc, ntrap, traps))
 		return;
 
-	if (trapezoid_span_inplace(op, src, dst, maskFormat,
+	if (trapezoid_span_inplace(sna, op, src, dst, maskFormat,
 				   xSrc, ySrc, ntrap, traps,
 				   false))
 		return;
@@ -6577,7 +6587,7 @@ sna_composite_trapezoids(CARD8 op,
 		return;
 
 fallback:
-	if (trapezoid_span_inplace(op, src, dst, maskFormat,
+	if (trapezoid_span_inplace(sna, op, src, dst, maskFormat,
 				   xSrc, ySrc, ntrap, traps,
 				   true))
 		return;
@@ -6586,12 +6596,13 @@ fallback:
 				    xSrc, ySrc, ntrap, traps))
 		return;
 
-	if (trapezoids_inplace_fallback(op, src, dst, maskFormat, ntrap, traps))
+	if (trapezoids_inplace_fallback(sna, op, src, dst, maskFormat,
+					ntrap, traps))
 		return;
 
 	DBG(("%s: fallback mask=%08x, ntrap=%d\n", __FUNCTION__,
 	     maskFormat ? (unsigned)maskFormat->format : 0, ntrap));
-	trapezoids_fallback(op, src, dst, maskFormat,
+	trapezoids_fallback(sna, op, src, dst, maskFormat,
 			    xSrc, ySrc,
 			    ntrap, traps);
 }
@@ -6613,7 +6624,8 @@ project_trap_onto_grid(const xTrap *in,
 }
 
 static bool
-mono_trap_span_converter(PicturePtr dst,
+mono_trap_span_converter(struct sna *sna,
+			 PicturePtr dst,
 			 INT16 x, INT16 y,
 			 int ntrap, xTrap *traps)
 {
@@ -6638,7 +6650,7 @@ mono_trap_span_converter(PicturePtr dst,
 	     mono.clip.extents.x2, mono.clip.extents.y2,
 	     x, y));
 
-	mono.sna = to_sna_from_drawable(dst->pDrawable);
+	mono.sna = sna;
 	if (!mono_init(&mono, 2*ntrap))
 		return false;
 
@@ -6683,11 +6695,11 @@ mono_trap_span_converter(PicturePtr dst,
 }
 
 static bool
-trap_span_converter(PicturePtr dst,
+trap_span_converter(struct sna *sna,
+		    PicturePtr dst,
 		    INT16 src_x, INT16 src_y,
 		    int ntrap, xTrap *trap)
 {
-	struct sna *sna;
 	struct sna_composite_spans_op tmp;
 	struct tor tor;
 	BoxRec extents;
@@ -6701,9 +6713,8 @@ trap_span_converter(PicturePtr dst,
 		return false;
 
 	if (dst->polyEdge == PolyEdgeSharp)
-		return mono_trap_span_converter(dst, src_x, src_y, ntrap, trap);
+		return mono_trap_span_converter(sna, dst, src_x, src_y, ntrap, trap);
 
-	sna = to_sna_from_drawable(dst->pDrawable);
 	if (!sna->render.check_composite_spans(sna, PictOpAdd, sna->render.white_picture, dst,
 					       dst->pCompositeClip->extents.x2 - dst->pCompositeClip->extents.x1,
 					       dst->pCompositeClip->extents.y2 - dst->pCompositeClip->extents.y1,
@@ -6795,11 +6806,11 @@ static void mark_damaged(PixmapPtr pixmap, struct sna_pixmap *priv,
 }
 
 static bool
-trap_mask_converter(PicturePtr picture,
+trap_mask_converter(struct sna *sna,
+		    PicturePtr picture,
 		    INT16 x, INT16 y,
 		    int ntrap, xTrap *trap)
 {
-	struct sna *sna;
 	struct tor tor;
 	ScreenPtr screen = picture->pDrawable->pScreen;
 	PixmapPtr scratch, pixmap;
@@ -6993,13 +7004,18 @@ trap_upload(PicturePtr picture,
 void
 sna_add_traps(PicturePtr picture, INT16 x, INT16 y, int n, xTrap *t)
 {
+	struct sna *sna;
+
 	DBG(("%s (%d, %d) x %d\n", __FUNCTION__, x, y, n));
 
-	if (is_gpu(picture->pDrawable)) {
-		if (trap_span_converter(picture, x, y, n, t))
+	sna = to_sna_from_drawable(picture->pDrawable);
+	if (is_gpu(sna, picture->pDrawable, PREFER_GPU_SPANS)) {
+		if (trap_span_converter(sna, picture, x, y, n, t))
 			return;
+	}
 
-		if (trap_mask_converter(picture, x, y, n, t))
+	if (is_gpu(sna, picture->pDrawable, PREFER_GPU_RENDER)) {
+		if (trap_mask_converter(sna, picture, x, y, n, t))
 			return;
 
 		if (trap_upload(picture, x, y, n, t))
@@ -7064,7 +7080,8 @@ project_triangle_onto_grid(const xTriangle *in,
 }
 
 static bool
-mono_triangles_span_converter(CARD8 op, PicturePtr src, PicturePtr dst,
+mono_triangles_span_converter(struct sna *sna,
+			      CARD8 op, PicturePtr src, PicturePtr dst,
 			      INT16 src_x, INT16 src_y,
 			      int count, xTriangle *tri)
 {
@@ -7075,7 +7092,7 @@ mono_triangles_span_converter(CARD8 op, PicturePtr src, PicturePtr dst,
 	bool was_clear;
 	int n;
 
-	mono.sna = to_sna_from_drawable(dst->pDrawable);
+	mono.sna = sna;
 
 	dst_x = pixman_fixed_to_int(tri[0].p1.x);
 	dst_y = pixman_fixed_to_int(tri[0].p1.y);
@@ -7200,11 +7217,11 @@ mono_triangles_span_converter(CARD8 op, PicturePtr src, PicturePtr dst,
 }
 
 static bool
-triangles_span_converter(CARD8 op, PicturePtr src, PicturePtr dst,
+triangles_span_converter(struct sna *sna,
+			 CARD8 op, PicturePtr src, PicturePtr dst,
 			 PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
 			 int count, xTriangle *tri)
 {
-	struct sna *sna;
 	struct sna_composite_spans_op tmp;
 	struct tor tor;
 	BoxRec extents;
@@ -7217,7 +7234,7 @@ triangles_span_converter(CARD8 op, PicturePtr src, PicturePtr dst,
 		return false;
 
 	if (is_mono(dst, maskFormat))
-		return mono_triangles_span_converter(op, src, dst,
+		return mono_triangles_span_converter(sna, op, src, dst,
 						     src_x, src_y,
 						     count, tri);
 
@@ -7228,7 +7245,6 @@ triangles_span_converter(CARD8 op, PicturePtr src, PicturePtr dst,
 		return false;
 	}
 
-	sna = to_sna_from_drawable(dst->pDrawable);
 	if (!sna->render.check_composite_spans(sna, op, src, dst, 0, 0, 0)) {
 		DBG(("%s: fallback -- composite spans not supported\n",
 		     __FUNCTION__));
@@ -7553,7 +7569,9 @@ sna_composite_triangles(CARD8 op,
 			 INT16 xSrc, INT16 ySrc,
 			 int n, xTriangle *tri)
 {
-	if (triangles_span_converter(op, src, dst, maskFormat,
+	struct sna *sna = to_sna_from_drawable(dst->pDrawable);
+
+	if (triangles_span_converter(sna, op, src, dst, maskFormat,
 				     xSrc, ySrc,
 				     n, tri))
 		return;
@@ -7567,11 +7585,11 @@ sna_composite_triangles(CARD8 op,
 }
 
 static bool
-tristrip_span_converter(CARD8 op, PicturePtr src, PicturePtr dst,
+tristrip_span_converter(struct sna *sna,
+			CARD8 op, PicturePtr src, PicturePtr dst,
 			PictFormatPtr maskFormat, INT16 src_x, INT16 src_y,
 			int count, xPointFixed *points)
 {
-	struct sna *sna;
 	struct sna_composite_spans_op tmp;
 	struct tor tor;
 	BoxRec extents;
@@ -7592,7 +7610,6 @@ tristrip_span_converter(CARD8 op, PicturePtr src, PicturePtr dst,
 		return false;
 	}
 
-	sna = to_sna_from_drawable(dst->pDrawable);
 	if (!sna->render.check_composite_spans(sna, op, src, dst, 0, 0, 0)) {
 		DBG(("%s: fallback -- composite spans not supported\n",
 		     __FUNCTION__));
@@ -7833,7 +7850,9 @@ sna_composite_tristrip(CARD8 op,
 		       INT16 xSrc, INT16 ySrc,
 		       int n, xPointFixed *points)
 {
-	if (tristrip_span_converter(op, src, dst, maskFormat, xSrc, ySrc, n, points))
+	struct sna *sna = to_sna_from_drawable(dst->pDrawable);
+
+	if (tristrip_span_converter(sna, op, src, dst, maskFormat, xSrc, ySrc, n, points))
 		return;
 
 	tristrip_fallback(op, src, dst, maskFormat, xSrc, ySrc, n, points);