nvfx: new 2D: new render temporaries with resources

This patch adds support for creating temporary surfaces to allow
rendering to surfaces that cannot be rendered to.
It uses the _second_ version of the render temporary infrastructure.

This is necessary for swizzled 3D textures and small mipmaps of
swizzled 2D textures.

This version of the patch creates a resource to use as a temporary
instead of a raw BO, making the code simpler.

8 files changed, 408 insertions, 146 deletions

diff --git a/src/gallium/drivers/nvfx/nvfx_context.c b/src/gallium/drivers/nvfx/nvfx_context.c
index 3d45f5f0ba..7ab81de7dd 100644
--- a/src/gallium/drivers/nvfx/nvfx_context.c
+++ b/src/gallium/drivers/nvfx/nvfx_context.c
@@ -15,6 +15,7 @@ nvfx_flush(struct pipe_context *pipe, unsigned flags,
 	struct nouveau_channel *chan = screen->base.channel;
 	struct nouveau_grobj *eng3d = screen->eng3d;
 
+	/* XXX: we need to actually be intelligent here */
 	if (flags & PIPE_FLUSH_TEXTURE_CACHE) {
 		BEGIN_RING(chan, eng3d, 0x1fd8, 1);
 		OUT_RING  (chan, 2);
@@ -87,5 +88,8 @@ nvfx_create(struct pipe_screen *pscreen, void *priv)
 	/* set these to that we init them on first validation */
 	nvfx->state.scissor_enabled = ~0;
 	nvfx->state.stipple_enabled = ~0;
+
+	LIST_INITHEAD(&nvfx->render_cache);
+
 	return &nvfx->pipe;
 }
diff --git a/src/gallium/drivers/nvfx/nvfx_context.h b/src/gallium/drivers/nvfx/nvfx_context.h
index 278be94d52..a6ea913967 100644
--- a/src/gallium/drivers/nvfx/nvfx_context.h
+++ b/src/gallium/drivers/nvfx/nvfx_context.h
@@ -11,6 +11,7 @@
 #include "util/u_memory.h"
 #include "util/u_math.h"
 #include "util/u_inlines.h"
+#include "util/u_double_list.h"
 
 #include "draw/draw_vertex.h"
 #include "util/u_blitter.h"
@@ -67,6 +68,7 @@ struct nvfx_state {
 	unsigned scissor_enabled;
 	unsigned stipple_enabled;
 	unsigned fp_samplers;
+	unsigned render_temps;
 };
 
 struct nvfx_vtxelt_state {
@@ -90,6 +92,7 @@ struct nvfx_context {
 
 	struct draw_context *draw;
 	struct blitter_context* blitter;
+	struct list_head render_cache;
 
 	/* HW state derived from pipe states */
 	struct nvfx_state state;
@@ -185,7 +188,8 @@ extern void nvfx_draw_elements_swtnl(struct pipe_context *pipe,
 extern void nvfx_vtxfmt_validate(struct nvfx_context *nvfx);
 
 /* nvfx_fb.c */
-extern void nvfx_state_framebuffer_validate(struct nvfx_context *nvfx);
+extern int nvfx_framebuffer_prepare(struct nvfx_context *nvfx);
+extern void nvfx_framebuffer_validate(struct nvfx_context *nvfx, unsigned prepare_result);
 void
 nvfx_framebuffer_relocate(struct nvfx_context *nvfx);
 
diff --git a/src/gallium/drivers/nvfx/nvfx_fragtex.c b/src/gallium/drivers/nvfx/nvfx_fragtex.c
index 0b4a434fec..6605745433 100644
--- a/src/gallium/drivers/nvfx/nvfx_fragtex.c
+++ b/src/gallium/drivers/nvfx/nvfx_fragtex.c
@@ -16,6 +16,10 @@ nvfx_fragtex_validate(struct nvfx_context *nvfx)
 		samplers &= ~(1 << unit);
 
 		if(nvfx->fragment_sampler_views[unit] && nvfx->tex_sampler[unit]) {
+			util_dirty_surfaces_use_for_sampling(&nvfx->pipe,
+					&((struct nvfx_miptree*)nvfx->fragment_sampler_views[unit]->texture)->dirty_surfaces,
+					nvfx_surface_flush);
+
 			if(!nvfx->is_nv4x)
 				nv30_fragtex_set(nvfx, unit);
 			else
diff --git a/src/gallium/drivers/nvfx/nvfx_miptree.c b/src/gallium/drivers/nvfx/nvfx_miptree.c
index 7deb9d7b9a..530d705e13 100644
--- a/src/gallium/drivers/nvfx/nvfx_miptree.c
+++ b/src/gallium/drivers/nvfx/nvfx_miptree.c
@@ -11,6 +11,7 @@
 #include "nvfx_screen.h"
 #include "nvfx_resource.h"
 #include "nvfx_transfer.h"
+#include "nv04_2d.h"
 
 static void
 nvfx_miptree_choose_format(struct nvfx_miptree *mt)
@@ -115,16 +116,23 @@ nvfx_miptree_get_handle(struct pipe_screen *pscreen,
 
 
 static void
+nvfx_miptree_surface_final_destroy(struct pipe_surface* ps)
+{
+	struct nvfx_surface* ns = (struct nvfx_surface*)ps;
+	pipe_resource_reference(&ps->texture, 0);
+	pipe_resource_reference((struct pipe_resource**)&ns->temp, 0);
+	FREE(ps);
+}
+
+static void
 nvfx_miptree_destroy(struct pipe_screen *screen, struct pipe_resource *pt)
 {
 	struct nvfx_miptree *mt = (struct nvfx_miptree *)pt;
+	util_surfaces_destroy(&mt->surfaces, pt, nvfx_miptree_surface_final_destroy);
 	nouveau_screen_bo_release(screen, mt->base.bo);
 	FREE(mt);
 }
 
-
-
-
 struct u_resource_vtbl nvfx_miptree_vtbl = 
 {
    nvfx_miptree_get_handle,	      /* get_handle */
@@ -152,6 +160,8 @@ nvfx_miptree_create_skeleton(struct pipe_screen *pscreen, const struct pipe_reso
 
         mt->base.base = *pt;
         mt->base.vtbl = &nvfx_miptree_vtbl;
+        util_dirty_surfaces_init(&mt->dirty_surfaces);
+
         pipe_reference_init(&mt->base.base.reference, 1);
         mt->base.base.screen = pscreen;
 
@@ -218,29 +228,28 @@ nvfx_miptree_surface_new(struct pipe_screen *pscreen, struct pipe_resource *pt,
 			 unsigned face, unsigned level, unsigned zslice,
 			 unsigned flags)
 {
+	struct nvfx_miptree* mt = (struct nvfx_miptree*)pt;
 	struct nvfx_surface *ns;
 
-	ns = CALLOC_STRUCT(nvfx_surface);
-	if (!ns)
-		return NULL;
-	pipe_resource_reference(&ns->base.texture, pt);
-	ns->base.format = pt->format;
-	ns->base.width = u_minify(pt->width0, level);
-	ns->base.height = u_minify(pt->height0, level);
-	ns->base.usage = flags;
-	pipe_reference_init(&ns->base.reference, 1);
-	ns->base.face = face;
-	ns->base.level = level;
-	ns->base.zslice = zslice;
-	ns->pitch = nvfx_subresource_pitch(pt, level);
-	ns->base.offset = nvfx_subresource_offset(pt, face, level, zslice);
-
-	return &ns->base;
+	ns = (struct nvfx_surface*)util_surfaces_get(&mt->surfaces, sizeof(struct nvfx_surface), pscreen, pt, face, level, zslice, flags);
+	if(ns->base.base.offset == ~0) {
+		util_dirty_surface_init(&ns->base);
+		ns->pitch = nvfx_subresource_pitch(pt, level);
+		ns->base.base.offset = nvfx_subresource_offset(pt, face, level, zslice);
+	}
+
+	return &ns->base.base;
 }
 
 void
 nvfx_miptree_surface_del(struct pipe_surface *ps)
 {
-	pipe_resource_reference(&ps->texture, NULL);
-	FREE(ps);
+	struct nvfx_surface* ns = (struct nvfx_surface*)ps;
+
+	if(!ns->temp)
+	{
+		util_surfaces_detach(&((struct nvfx_miptree*)ps->texture)->surfaces, ps);
+		pipe_resource_reference(&ps->texture, 0);
+		FREE(ps);
+	}
 }
diff --git a/src/gallium/drivers/nvfx/nvfx_resource.h b/src/gallium/drivers/nvfx/nvfx_resource.h
index 42d04ebb37..be1845dd9c 100644
--- a/src/gallium/drivers/nvfx/nvfx_resource.h
+++ b/src/gallium/drivers/nvfx/nvfx_resource.h
@@ -1,13 +1,16 @@
-
 #ifndef NVFX_RESOURCE_H
 #define NVFX_RESOURCE_H
 
 #include "util/u_transfer.h"
 #include "util/u_format.h"
 #include "util/u_math.h"
+#include "util/u_double_list.h"
+#include "util/u_surfaces.h"
+#include "util/u_dirty_surfaces.h"
 #include <nouveau/nouveau_bo.h>
 
 struct pipe_resource;
+struct nv04_region;
 
 
 /* This gets further specialized into either buffer or texture
@@ -38,17 +41,34 @@ nvfx_resource_on_gpu(struct pipe_resource* pr)
 
 #define NVFX_MAX_TEXTURE_LEVELS  16
 
+/* We have the following invariants for render temporaries
+ *
+ * 1. Render temporaries are always linear
+ * 2. Render temporaries are always up to date
+ * 3. Currently, render temporaries are destroyed when the resource is used for sampling, but kept for any other use
+ *
+ * Also, we do NOT flush temporaries on any pipe->flush().
+ * This is fine, as long as scanout targets and shared resources never need temps.
+ *
+ * TODO: we may want to also support swizzled temporaries to improve performance in some cases.
+ */
+
 struct nvfx_miptree {
         struct nvfx_resource base;
 
         unsigned linear_pitch; /* for linear textures, 0 for swizzled and compressed textures with level-dependent minimal pitch */
         unsigned face_size; /* 128-byte aligned face/total size */
         unsigned level_offset[NVFX_MAX_TEXTURE_LEVELS];
+
+        struct util_surfaces surfaces;
+        struct util_dirty_surfaces dirty_surfaces;
 };
 
 struct nvfx_surface {
-	struct pipe_surface base;
+	struct util_dirty_surface base;
 	unsigned pitch;
+
+	struct nvfx_miptree* temp;
 };
 
 static INLINE 
@@ -65,6 +85,12 @@ nvfx_surface_buffer(struct pipe_surface *surf)
 	return mt->bo;
 }
 
+static INLINE struct util_dirty_surfaces*
+nvfx_surface_get_dirty_surfaces(struct pipe_surface* surf)
+{
+	struct nvfx_miptree *mt = (struct nvfx_miptree *)surf->texture;
+	return &mt->dirty_surfaces;
+}
 
 void
 nvfx_init_resource_functions(struct pipe_context *pipe);
@@ -141,4 +167,10 @@ nvfx_subresource_pitch(struct pipe_resource* pt, unsigned level)
 	}
 }
 
+void
+nvfx_surface_create_temp(struct pipe_context* pipe, struct pipe_surface* surf);
+
+void
+nvfx_surface_flush(struct pipe_context* pipe, struct pipe_surface* surf);
+
 #endif
diff --git a/src/gallium/drivers/nvfx/nvfx_state_emit.c b/src/gallium/drivers/nvfx/nvfx_state_emit.c
index f91ae19ecd..dc70f3de87 100644
--- a/src/gallium/drivers/nvfx/nvfx_state_emit.c
+++ b/src/gallium/drivers/nvfx/nvfx_state_emit.c
@@ -1,15 +1,48 @@
 #include "nvfx_context.h"
 #include "nvfx_state.h"
+#include "nvfx_resource.h"
 #include "draw/draw_context.h"
 
 static boolean
 nvfx_state_validate_common(struct nvfx_context *nvfx)
 {
 	struct nouveau_channel* chan = nvfx->screen->base.channel;
-	unsigned dirty = nvfx->dirty;
+	unsigned dirty;
+	int all_swizzled = -1;
+	boolean flush_tex_cache = FALSE;
 
 	if(nvfx != nvfx->screen->cur_ctx)
-		dirty = ~0;
+	{
+		nvfx->dirty = ~0;
+		nvfx->screen->cur_ctx = nvfx;
+	}
+
+	/* These can trigger use the of 3D engine to copy temporaries.
+	 * That will recurse here and thus dirty all 3D state, so we need to this before anything else, and in a loop..
+	 * This converges to having clean temps, then binding both fragtexes and framebuffers.
+	 */
+	while(nvfx->dirty & (NVFX_NEW_FB | NVFX_NEW_SAMPLER))
+	{
+		if(nvfx->dirty & NVFX_NEW_SAMPLER)
+		{
+			nvfx->dirty &=~ NVFX_NEW_SAMPLER;
+			nvfx_fragtex_validate(nvfx);
+
+			// TODO: only set this if really necessary
+			flush_tex_cache = TRUE;
+		}
+
+		if(nvfx->dirty & NVFX_NEW_FB)
+		{
+			nvfx->dirty &=~ NVFX_NEW_FB;
+			all_swizzled = nvfx_framebuffer_prepare(nvfx);
+
+			// TODO: make sure this doesn't happen, i.e. fbs have matching formats
+			assert(all_swizzled >= 0);
+		}
+	}
+
+	dirty = nvfx->dirty;
 
 	if(nvfx->render_mode == HW)
 	{
@@ -35,9 +68,6 @@ nvfx_state_validate_common(struct nvfx_context *nvfx)
 			nvfx_vtxfmt_validate(nvfx);
 	}
 
-	if(dirty & NVFX_NEW_FB)
-		nvfx_state_framebuffer_validate(nvfx);
-
 	if(dirty & NVFX_NEW_RAST)
 		sb_emit(chan, nvfx->rasterizer->sb, nvfx->rasterizer->sb_len);
 
@@ -48,10 +78,14 @@ nvfx_state_validate_common(struct nvfx_context *nvfx)
 		nvfx_state_stipple_validate(nvfx);
 
 	if(dirty & (NVFX_NEW_FRAGPROG | NVFX_NEW_FRAGCONST))
+	{
 		nvfx_fragprog_validate(nvfx);
+		if(dirty & NVFX_NEW_FRAGPROG)
+			flush_tex_cache = TRUE; // TODO: do we need this?
+	}
 
-	if(dirty & NVFX_NEW_SAMPLER)
-		nvfx_fragtex_validate(nvfx);
+	if(all_swizzled >= 0)
+		nvfx_framebuffer_validate(nvfx, all_swizzled);
 
 	if(dirty & NVFX_NEW_BLEND)
 		sb_emit(chan, nvfx->blend->sb, nvfx->blend->sb_len);
@@ -72,13 +106,17 @@ nvfx_state_validate_common(struct nvfx_context *nvfx)
 	if(dirty & (NVFX_NEW_VIEWPORT | NVFX_NEW_FB))
 		nvfx_state_viewport_validate(nvfx);
 
-	/* TODO: could nv30 need this or something similar too? */
-	if((dirty & (NVFX_NEW_FRAGPROG | NVFX_NEW_SAMPLER)) && nvfx->is_nv4x) {
-		WAIT_RING(chan, 4);
-		OUT_RING(chan, RING_3D(NV40TCL_TEX_CACHE_CTL, 1));
-		OUT_RING(chan, 2);
-		OUT_RING(chan, RING_3D(NV40TCL_TEX_CACHE_CTL, 1));
-		OUT_RING(chan, 1);
+	if(flush_tex_cache)
+	{
+		// TODO: what about nv30?
+		if(nvfx->is_nv4x)
+		{
+			WAIT_RING(chan, 4);
+			OUT_RING(chan, RING_3D(NV40TCL_TEX_CACHE_CTL, 1));
+			OUT_RING(chan, 2);
+			OUT_RING(chan, RING_3D(NV40TCL_TEX_CACHE_CTL, 1));
+			OUT_RING(chan, 1);
+		}
 	}
 	nvfx->dirty = 0;
 	return TRUE;
@@ -99,6 +137,21 @@ nvfx_state_emit(struct nvfx_context *nvfx)
 	      ;
 	MARK_RING(chan, max_relocs * 2, max_relocs * 2);
 	nvfx_state_relocate(nvfx);
+
+	unsigned render_temps = nvfx->state.render_temps;
+	if(render_temps)
+	{
+		for(int i = 0; i < nvfx->framebuffer.nr_cbufs; ++i)
+		{
+			if(render_temps & (1 << i))
+				util_dirty_surface_set_dirty(nvfx_surface_get_dirty_surfaces(nvfx->framebuffer.cbufs[i]),
+						(struct util_dirty_surface*)nvfx->framebuffer.cbufs[i]);
+		}
+
+		if(render_temps & 0x80)
+			util_dirty_surface_set_dirty(nvfx_surface_get_dirty_surfaces(nvfx->framebuffer.zsbuf),
+					(struct util_dirty_surface*)nvfx->framebuffer.zsbuf);
+	}
 }
 
 void
diff --git a/src/gallium/drivers/nvfx/nvfx_state_fb.c b/src/gallium/drivers/nvfx/nvfx_state_fb.c
index e111d11627..80b0f21575 100644
--- a/src/gallium/drivers/nvfx/nvfx_state_fb.c
+++ b/src/gallium/drivers/nvfx/nvfx_state_fb.c
@@ -1,19 +1,56 @@
 #include "nvfx_context.h"
 #include "nvfx_resource.h"
 #include "nouveau/nouveau_util.h"
+#include "util/u_format.h"
 
-void
-nvfx_state_framebuffer_validate(struct nvfx_context *nvfx)
+static inline boolean
+nvfx_surface_linear_renderable(struct pipe_surface* surf)
+{
+	return (surf->texture->flags & NVFX_RESOURCE_FLAG_LINEAR)
+		&& !(surf->offset & 63)
+		&& !(((struct nvfx_surface*)surf)->pitch & 63);
+}
+
+static inline boolean
+nvfx_surface_swizzled_renderable(struct pipe_framebuffer_state* fb, struct pipe_surface* surf)
+{
+	/* TODO: return FALSE if we have a format not supporting swizzled rendering (e.g. r8); currently those are not supported at all */
+	return !((struct nvfx_miptree*)surf->texture)->linear_pitch
+		&& (surf->texture->target != PIPE_TEXTURE_3D || u_minify(surf->texture->depth0, surf->level) <= 1)
+		&& !(surf->offset & 127)
+		&& (surf->width == fb->width)
+		&& (surf->height == fb->height)
+		&& !((struct nvfx_surface*)surf)->temp;
+}
+
+static boolean
+nvfx_surface_get_render_target(struct pipe_surface* surf, int all_swizzled, struct nvfx_render_target* target)
+{
+	struct nvfx_surface* ns = (struct nvfx_surface*)surf;
+	if(!ns->temp)
+	{
+		target->bo = ((struct nvfx_miptree*)surf->texture)->base.bo;
+		target->offset = surf->offset;
+		target->pitch = align(ns->pitch, 64);
+		assert(target->pitch);
+		return FALSE;
+	}
+	else
+	{
+		target->offset = 0;
+		target->pitch = ns->temp->linear_pitch;
+		target->bo = ns->temp->base.bo;
+		assert(target->pitch);
+		return TRUE;
+	}
+}
+
+int
+nvfx_framebuffer_prepare(struct nvfx_context *nvfx)
 {
 	struct pipe_framebuffer_state *fb = &nvfx->framebuffer;
-	struct nouveau_channel *chan = nvfx->screen->base.channel;
-	uint32_t rt_enable = 0, rt_format = 0;
-	int i, colour_format = 0, zeta_format = 0;
-	int depth_only = 0;
-	unsigned rt_flags = NOUVEAU_BO_RDWR | NOUVEAU_BO_VRAM;
-	unsigned w = fb->width;
-	unsigned h = fb->height;
-	int colour_bits = 32, zeta_bits = 32;
+	int i, color_format = 0, zeta_format = 0;
+	int all_swizzled = 1;
 
 	if(!nvfx->is_nv4x)
 		assert(fb->nr_cbufs <= 2);
@@ -21,113 +58,135 @@ nvfx_state_framebuffer_validate(struct nvfx_context *nvfx)
 		assert(fb->nr_cbufs <= 4);
 
 	for (i = 0; i < fb->nr_cbufs; i++) {
-		if (colour_format)
-			assert(colour_format == fb->cbufs[i]->format);
-		else
-			colour_format = fb->cbufs[i]->format;
-
-		rt_enable |= (NV34TCL_RT_ENABLE_COLOR0 << i);
-		nvfx->hw_rt[i].bo = ((struct nvfx_miptree*)fb->cbufs[i]->texture)->base.bo;
-		nvfx->hw_rt[i].offset = fb->cbufs[i]->offset;
-		nvfx->hw_rt[i].pitch = ((struct nvfx_surface *)fb->cbufs[i])->pitch;
+		if (color_format) {
+			if(color_format != fb->cbufs[i]->format)
+				return -1;
+		} else
+			color_format = fb->cbufs[i]->format;
+
+		if(!nvfx_surface_swizzled_renderable(fb, fb->cbufs[i]))
+			all_swizzled = 0;
 	}
-	for(; i < 4; ++i)
-		nvfx->hw_rt[i].bo = 0;
 
+	if (fb->zsbuf) {
+		/* TODO: return FALSE if we have a format not supporting a depth buffer (e.g. r8); currently those are not supported at all */
+		if(!nvfx_surface_swizzled_renderable(fb, fb->zsbuf))
+			all_swizzled = 0;
+
+		if(all_swizzled && util_format_get_blocksize(color_format) != util_format_get_blocksize(zeta_format))
+			all_swizzled = 0;
+	}
+
+	for (i = 0; i < fb->nr_cbufs; i++) {
+		if(!((struct nvfx_surface*)fb->cbufs[i])->temp && !all_swizzled && !nvfx_surface_linear_renderable(fb->cbufs[i]))
+			nvfx_surface_create_temp(&nvfx->pipe, fb->cbufs[i]);
+	}
+
+	if(fb->zsbuf) {
+		if(!((struct nvfx_surface*)fb->zsbuf)->temp && !all_swizzled && !nvfx_surface_linear_renderable(fb->zsbuf))
+			nvfx_surface_create_temp(&nvfx->pipe, fb->zsbuf);
+	}
+
+	return all_swizzled;
+}
+
+void
+nvfx_framebuffer_validate(struct nvfx_context *nvfx, unsigned prepare_result)
+{
+	struct pipe_framebuffer_state *fb = &nvfx->framebuffer;
+	struct nouveau_channel *chan = nvfx->screen->base.channel;
+	uint32_t rt_enable, rt_format;
+	int i;
+	unsigned rt_flags = NOUVEAU_BO_RDWR | NOUVEAU_BO_VRAM;
+	unsigned w = fb->width;
+	unsigned h = fb->height;
+
+	rt_enable = (NV34TCL_RT_ENABLE_COLOR0 << fb->nr_cbufs) - 1;
 	if (rt_enable & (NV34TCL_RT_ENABLE_COLOR1 |
 			 NV40TCL_RT_ENABLE_COLOR2 | NV40TCL_RT_ENABLE_COLOR3))
 		rt_enable |= NV34TCL_RT_ENABLE_MRT;
 
+	nvfx->state.render_temps = 0;
+
+	for (i = 0; i < fb->nr_cbufs; i++)
+		nvfx->state.render_temps |= nvfx_surface_get_render_target(fb->cbufs[i], prepare_result, &nvfx->hw_rt[i]) << i;
+
+	for(; i < 4; ++i)
+		nvfx->hw_rt[i].bo = 0;
+
 	if (fb->zsbuf) {
-		zeta_format = fb->zsbuf->format;
-		nvfx->hw_zeta.bo = ((struct nvfx_miptree*)fb->zsbuf->texture)->base.bo;
-		nvfx->hw_zeta.offset = fb->zsbuf->offset;
-		nvfx->hw_zeta.pitch = ((struct nvfx_surface *)fb->zsbuf)->pitch;
-	}
-	else
-		nvfx->hw_zeta.bo = 0;
-
-	if (rt_enable & (NV34TCL_RT_ENABLE_COLOR0 | NV34TCL_RT_ENABLE_COLOR1 |
-		NV40TCL_RT_ENABLE_COLOR2 | NV40TCL_RT_ENABLE_COLOR3)) {
-		/* Render to at least a colour buffer */
-		if (!(fb->cbufs[0]->texture->flags & NVFX_RESOURCE_FLAG_LINEAR)) {
-			assert(!(fb->width & (fb->width - 1)) && !(fb->height & (fb->height - 1)));
-			for (i = 1; i < fb->nr_cbufs; i++)
-				assert(!(fb->cbufs[i]->texture->flags & NVFX_RESOURCE_FLAG_LINEAR));
-
-			rt_format = NV34TCL_RT_FORMAT_TYPE_SWIZZLED |
-				(log2i(fb->cbufs[0]->width) << NV34TCL_RT_FORMAT_LOG2_WIDTH_SHIFT) |
-				(log2i(fb->cbufs[0]->height) << NV34TCL_RT_FORMAT_LOG2_HEIGHT_SHIFT);
-		}
-		else
-			rt_format = NV34TCL_RT_FORMAT_TYPE_LINEAR;
-	} else if (fb->zsbuf) {
-		depth_only = 1;
-
-		/* Render to depth buffer only */
-		if (!(fb->zsbuf->texture->flags & NVFX_RESOURCE_FLAG_LINEAR)) {
-			assert(!(fb->width & (fb->width - 1)) && !(fb->height & (fb->height - 1)));
-
-			rt_format = NV34TCL_RT_FORMAT_TYPE_SWIZZLED |
-				(log2i(fb->zsbuf->width) << NV34TCL_RT_FORMAT_LOG2_WIDTH_SHIFT) |
-				(log2i(fb->zsbuf->height) << NV34TCL_RT_FORMAT_LOG2_HEIGHT_SHIFT);
-		}
-		else
-			rt_format = NV34TCL_RT_FORMAT_TYPE_LINEAR;
-	} else {
-		return;
+		nvfx->state.render_temps |= nvfx_surface_get_render_target(fb->zsbuf, prepare_result, &nvfx->hw_zeta) << 7;
+
+		assert(util_format_get_stride(fb->zsbuf->format, fb->width) <= nvfx->hw_zeta.pitch);
+		assert(nvfx->hw_zeta.offset + nvfx->hw_zeta.pitch * fb->height <= nvfx->hw_zeta.bo->size);
 	}
 
-	switch (colour_format) {
-	case PIPE_FORMAT_B8G8R8X8_UNORM:
-		rt_format |= NV34TCL_RT_FORMAT_COLOR_X8R8G8B8;
-		break;
-	case PIPE_FORMAT_B8G8R8A8_UNORM:
-	case 0:
-		rt_format |= NV34TCL_RT_FORMAT_COLOR_A8R8G8B8;
-		break;
-	case PIPE_FORMAT_B5G6R5_UNORM:
+	if (prepare_result) {
+		assert(!(fb->width & (fb->width - 1)) && !(fb->height & (fb->height - 1)));
+
+		rt_format = NV34TCL_RT_FORMAT_TYPE_SWIZZLED |
+			(log2i(fb->width) << NV34TCL_RT_FORMAT_LOG2_WIDTH_SHIFT) |
+			(log2i(fb->height) << NV34TCL_RT_FORMAT_LOG2_HEIGHT_SHIFT);
+	} else
+		rt_format = NV34TCL_RT_FORMAT_TYPE_LINEAR;
+
+	if(fb->nr_cbufs > 0) {
+		switch (fb->cbufs[0]->format) {
+		case PIPE_FORMAT_B8G8R8X8_UNORM:
+			rt_format |= NV34TCL_RT_FORMAT_COLOR_X8R8G8B8;
+			break;
+		case PIPE_FORMAT_B8G8R8A8_UNORM:
+		case 0:
+			rt_format |= NV34TCL_RT_FORMAT_COLOR_A8R8G8B8;
+			break;
+		case PIPE_FORMAT_B5G6R5_UNORM:
+			rt_format |= NV34TCL_RT_FORMAT_COLOR_R5G6B5;
+			break;
+		default:
+			assert(0);
+		}
+	} else if(fb->zsbuf && util_format_get_blocksize(fb->zsbuf->format) == 2)
 		rt_format |= NV34TCL_RT_FORMAT_COLOR_R5G6B5;
-		colour_bits = 16;
-		break;
-	default:
-		assert(0);
-	}
+	else
+		rt_format |= NV34TCL_RT_FORMAT_COLOR_A8R8G8B8;
 
-	switch (zeta_format) {
-	case PIPE_FORMAT_Z16_UNORM:
+	if(fb->zsbuf) {
+		switch (fb->zsbuf->format) {
+		case PIPE_FORMAT_Z16_UNORM:
+			rt_format |= NV34TCL_RT_FORMAT_ZETA_Z16;
+			break;
+		case PIPE_FORMAT_S8_USCALED_Z24_UNORM:
+		case PIPE_FORMAT_X8Z24_UNORM:
+		case 0:
+			rt_format |= NV34TCL_RT_FORMAT_ZETA_Z24S8;
+			break;
+		default:
+			assert(0);
+		}
+	} else if(fb->nr_cbufs && util_format_get_blocksize(fb->cbufs[0]->format) == 2)
 		rt_format |= NV34TCL_RT_FORMAT_ZETA_Z16;
-		zeta_bits = 16;
-		break;
-	case PIPE_FORMAT_S8_USCALED_Z24_UNORM:
-	case PIPE_FORMAT_X8Z24_UNORM:
-	case 0:
+	else
 		rt_format |= NV34TCL_RT_FORMAT_ZETA_Z24S8;
-		break;
-	default:
-		assert(0);
-	}
 
-	if ((!nvfx->is_nv4x) && colour_bits > zeta_bits) {
-		/* TODO: does this limitation really exist?
-		   TODO: can it be worked around somehow? */
-		assert(0);
-	}
+	if ((rt_enable & NV34TCL_RT_ENABLE_COLOR0) || fb->zsbuf) {
+		struct nvfx_render_target *rt0 = &nvfx->hw_rt[0];
+		uint32_t pitch;
+
+		if(!(rt_enable & NV34TCL_RT_ENABLE_COLOR0))
+			rt0 = &nvfx->hw_zeta;
 
-	if ((rt_enable & NV34TCL_RT_ENABLE_COLOR0)
-		|| ((!nvfx->is_nv4x) && depth_only)) {
-		struct nvfx_render_target *rt0 = (depth_only ? &nvfx->hw_zeta : &nvfx->hw_rt[0]);
-		uint32_t pitch = rt0->pitch;
+		pitch = rt0->pitch;
 
 		if(!nvfx->is_nv4x)
 		{
-			if (nvfx->hw_zeta.bo) {
+			if (nvfx->hw_zeta.bo)
 				pitch |= (nvfx->hw_zeta.pitch << 16);
-			} else {
+			else
 				pitch |= (pitch << 16);
-			}
 		}
 
+		//printf("rendering to bo %p [%i] at offset %i with pitch %i\n", rt0->bo, rt0->bo->handle, rt0->offset, pitch);
+
 		OUT_RING(chan, RING_3D(NV34TCL_DMA_COLOR0, 1));
 		OUT_RELOC(chan, rt0->bo, 0,
 			      rt_flags | NOUVEAU_BO_OR,
@@ -180,7 +239,7 @@ nvfx_state_framebuffer_validate(struct nvfx_context *nvfx)
 		}
 	}
 
-	if (zeta_format) {
+	if (fb->zsbuf) {
 		OUT_RING(chan, RING_3D(NV34TCL_DMA_ZETA, 1));
 		OUT_RELOC(chan, nvfx->hw_zeta.bo, 0,
 			      rt_flags | NOUVEAU_BO_OR,
diff --git a/src/gallium/drivers/nvfx/nvfx_surface.c b/src/gallium/drivers/nvfx/nvfx_surface.c
index a97f342c64..8208c67f2a 100644
--- a/src/gallium/drivers/nvfx/nvfx_surface.c
+++ b/src/gallium/drivers/nvfx/nvfx_surface.c
@@ -94,23 +94,44 @@ nvfx_region_fixup_swizzled(struct nv04_region* rgn, unsigned zslice, unsigned wi
 }
 
 static INLINE void
-nvfx_region_init_for_surface(struct nv04_region* rgn, struct nvfx_surface* surf, unsigned x, unsigned y)
+nvfx_region_init_for_surface(struct nv04_region* rgn, struct nvfx_surface* surf, unsigned x, unsigned y, bool for_write)
 {
-	rgn->bo = ((struct nvfx_resource*)surf->base.texture)->bo;
-	rgn->offset = surf->base.offset;
-	rgn->pitch = surf->pitch;
 	rgn->x = x;
 	rgn->y = y;
 	rgn->z = 0;
+	nvfx_region_set_format(rgn, surf->base.base.format);
 
-	nvfx_region_set_format(rgn, surf->base.format);
-	if(!(surf->base.texture->flags & NVFX_RESOURCE_FLAG_LINEAR))
-		nvfx_region_fixup_swizzled(rgn, surf->base.zslice, surf->base.width, surf->base.height, u_minify(surf->base.texture->depth0, surf->base.level));
+	if(surf->temp)
+	{
+		rgn->bo = surf->temp->base.bo;
+		rgn->offset = 0;
+		rgn->pitch = surf->temp->linear_pitch;
+
+		if(for_write)
+			util_dirty_surface_set_dirty(nvfx_surface_get_dirty_surfaces(&surf->base.base), &surf->base);
+	} else {
+		rgn->bo = ((struct nvfx_resource*)surf->base.base.texture)->bo;
+		rgn->offset = surf->base.base.offset;
+		rgn->pitch = surf->pitch;
+
+	        if(!(surf->base.base.texture->flags & NVFX_RESOURCE_FLAG_LINEAR))
+		        nvfx_region_fixup_swizzled(rgn, surf->base.base.zslice, surf->base.base.width, surf->base.base.height, u_minify(surf->base.base.texture->depth0, surf->base.base.level));
+	}
 }
 
 static INLINE void
-nvfx_region_init_for_subresource(struct nv04_region* rgn, struct pipe_resource* pt, struct pipe_subresource sub, unsigned x, unsigned y, unsigned z)
+nvfx_region_init_for_subresource(struct nv04_region* rgn, struct pipe_resource* pt, struct pipe_subresource sub, unsigned x, unsigned y, unsigned z, bool for_write)
 {
+	if(pt->target != PIPE_BUFFER)
+	{
+		struct nvfx_surface* ns = (struct nvfx_surface*)util_surfaces_peek(&((struct nvfx_miptree*)pt)->surfaces, pt, sub.face, sub.level, z);
+		if(ns && util_dirty_surface_is_dirty(&ns->base))
+		{
+			nvfx_region_init_for_surface(rgn, ns, x, y, for_write);
+			return;
+		}
+	}
+
 	rgn->bo = ((struct nvfx_resource*)pt)->bo;
 	rgn->offset = nvfx_subresource_offset(pt, sub.face, sub.level, z);
 	rgn->pitch = nvfx_subresource_pitch(pt, sub.level);
@@ -165,6 +186,7 @@ nv04_scaled_image_format(enum pipe_format format)
 	}
 }
 
+// XXX: must save index buffer too!
 static struct blitter_context*
 nvfx_get_blitter(struct pipe_context* pipe, int copy)
 {
@@ -237,8 +259,8 @@ nvfx_resource_copy_region(struct pipe_context *pipe,
 	int dst_to_gpu = dstr->usage != PIPE_USAGE_DYNAMIC && dstr->usage != PIPE_USAGE_STAGING;
 	int src_on_gpu = nvfx_resource_on_gpu(srcr);
 
-	nvfx_region_init_for_subresource(&dst, dstr, subdst, dstx, dsty, dstz);
-	nvfx_region_init_for_subresource(&src, srcr, subsrc, srcx, srcy, srcz);
+	nvfx_region_init_for_subresource(&dst, dstr, subdst, dstx, dsty, dstz, TRUE);
+	nvfx_region_init_for_subresource(&src, srcr, subsrc, srcx, srcy, srcz, FALSE);
 	w = util_format_get_stride(dstr->format, w) >> dst.bpps;
 	h = util_format_get_nblocksy(dstr->format, h);
 
@@ -293,10 +315,11 @@ nvfx_surface_fill(struct pipe_context* pipe, struct pipe_surface *dsts,
 	struct nv04_2d_context *ctx = nvfx_screen(pipe->screen)->eng2d;
 	struct nv04_region dst;
 	/* Always try to use the GPU right now, if possible
-	 * If the user wanted the surface data on the CPU, he would have cleared with memset */
+	 * If the user wanted the surface data on the CPU, he would have cleared with memset (hopefully) */
 
 	// we don't care about interior pixel order since we set all them to the same value
-	nvfx_region_init_for_surface(&dst, (struct nvfx_surface*)dsts, dx, dy);
+	nvfx_region_init_for_surface(&dst, (struct nvfx_surface*)dsts, dx, dy, TRUE);
+
 	w = util_format_get_stride(dsts->format, w) >> dst.bpps;
 	h = util_format_get_nblocksy(dsts->format, h);
 
@@ -342,6 +365,80 @@ nvfx_screen_surface_init(struct pipe_screen *pscreen)
 }
 
 static void
+nvfx_surface_copy_temp(struct pipe_context* pipe, struct pipe_surface* surf, int to_temp)
+{
+	struct nvfx_surface* ns = (struct nvfx_surface*)surf;
+	struct pipe_subresource tempsr, surfsr;
+	struct pipe_resource *idxbuf_buffer;
+	unsigned idxbuf_format;
+
+	tempsr.face = 0;
+	tempsr.level = 0;
+	surfsr.face = surf->face;
+	surfsr.level = surf->level;
+
+	// TODO: do this properly, in blitter save
+	idxbuf_buffer = ((struct nvfx_context*)pipe)->idxbuf_buffer;
+	idxbuf_format = ((struct nvfx_context*)pipe)->idxbuf_format;
+
+	if(to_temp)
+		nvfx_resource_copy_region(pipe, &ns->temp->base.base, tempsr, 0, 0, 0, surf->texture, surfsr, 0, 0, surf->zslice, surf->width, surf->height);
+	else
+		nvfx_resource_copy_region(pipe, surf->texture, surfsr, 0, 0, surf->zslice, &ns->temp->base.base, tempsr, 0, 0, 0, surf->width, surf->height);
+
+	((struct nvfx_context*)pipe)->idxbuf_buffer = idxbuf_buffer;
+	((struct nvfx_context*)pipe)->idxbuf_format = idxbuf_format;
+}
+
+void
+nvfx_surface_create_temp(struct pipe_context* pipe, struct pipe_surface* surf)
+{
+	struct nvfx_surface* ns = (struct nvfx_surface*)surf;
+	struct pipe_resource template;
+	memset(&template, 0, sizeof(struct pipe_resource));
+	template.target = PIPE_TEXTURE_2D;
+	template.format = surf->format;
+	template.width0 = surf->width;
+	template.height0 = surf->height;
+	template.depth0 = 1;
+	template.nr_samples = surf->texture->nr_samples;
+	template.flags = NVFX_RESOURCE_FLAG_LINEAR;
+
+	ns->temp = (struct nvfx_miptree*)nvfx_miptree_create(pipe->screen, &template);
+	nvfx_surface_copy_temp(pipe, surf, 1);
+}
+
+void
+nvfx_surface_flush(struct pipe_context* pipe, struct pipe_surface* surf)
+{
+	struct nvfx_context* nvfx = (struct nvfx_context*)pipe;
+	struct nvfx_surface* ns = (struct nvfx_surface*)surf;
+	boolean bound = FALSE;
+
+	/* must be done before the copy, otherwise the copy will use the temp as destination */
+	util_dirty_surface_set_clean(nvfx_surface_get_dirty_surfaces(surf), &ns->base);
+
+	nvfx_surface_copy_temp(pipe, surf, 0);
+
+	if(nvfx->framebuffer.zsbuf == surf)
+		bound = TRUE;
+	else
+	{
+		for(unsigned i = 0; i < nvfx->framebuffer.nr_cbufs; ++i)
+		{
+			if(nvfx->framebuffer.cbufs[i] == surf)
+			{
+				bound = TRUE;
+				break;
+			}
+		}
+	}
+
+	if(!bound)
+		pipe_resource_reference((struct pipe_resource**)&ns->temp, 0);
+}
+
+static void
 nvfx_clear_render_target(struct pipe_context *pipe,
 			 struct pipe_surface *dst,
 			 const float *rgba,