nv50: rework state emission

7 files changed, 249 insertions, 311 deletions

diff --git a/src/gallium/drivers/nv50/nv50_context.c b/src/gallium/drivers/nv50/nv50_context.c
index 7be12fcdef4..eff012c14d3 100644
--- a/src/gallium/drivers/nv50/nv50_context.c
+++ b/src/gallium/drivers/nv50/nv50_context.c
@@ -46,43 +46,13 @@ static void
 nv50_destroy(struct pipe_context *pipe)
 {
 	struct nv50_context *nv50 = nv50_context(pipe);
+	int i;
 
-        if (nv50->state.fb)
-		so_ref(NULL, &nv50->state.fb);
-	if (nv50->state.blend)
-		so_ref(NULL, &nv50->state.blend);
-	if (nv50->state.blend_colour)
-		so_ref(NULL, &nv50->state.blend_colour);
-	if (nv50->state.zsa)
-		so_ref(NULL, &nv50->state.zsa);
-	if (nv50->state.rast)
-		so_ref(NULL, &nv50->state.rast);
-	if (nv50->state.stipple)
-		so_ref(NULL, &nv50->state.stipple);
-	if (nv50->state.scissor)
-		so_ref(NULL, &nv50->state.scissor);
-	if (nv50->state.viewport)
-		so_ref(NULL, &nv50->state.viewport);
-	if (nv50->state.tsc_upload)
-		so_ref(NULL, &nv50->state.tsc_upload);
-	if (nv50->state.tic_upload)
-		so_ref(NULL, &nv50->state.tic_upload);
-	if (nv50->state.vertprog)
-		so_ref(NULL, &nv50->state.vertprog);
-	if (nv50->state.fragprog)
-		so_ref(NULL, &nv50->state.fragprog);
-	if (nv50->state.geomprog)
-		so_ref(NULL, &nv50->state.geomprog);
-	if (nv50->state.fp_linkage)
-		so_ref(NULL, &nv50->state.fp_linkage);
-	if (nv50->state.gp_linkage)
-		so_ref(NULL, &nv50->state.gp_linkage);
-	if (nv50->state.vtxfmt)
-		so_ref(NULL, &nv50->state.vtxfmt);
-	if (nv50->state.vtxbuf)
-		so_ref(NULL, &nv50->state.vtxbuf);
-	if (nv50->state.vtxattr)
-		so_ref(NULL, &nv50->state.vtxattr);
+	for (i = 0; i < 64; i++) {
+		if (!nv50->state.hw[i])
+			continue;
+		so_ref(NULL, &nv50->state.hw[i]);
+	}
 
 	draw_destroy(nv50->draw);
 
diff --git a/src/gallium/drivers/nv50/nv50_context.h b/src/gallium/drivers/nv50/nv50_context.h
index 2cb5538c54a..5fa39111148 100644
--- a/src/gallium/drivers/nv50/nv50_context.h
+++ b/src/gallium/drivers/nv50/nv50_context.h
@@ -123,27 +123,11 @@ nv50_surface(struct pipe_surface *pt)
 }
 
 struct nv50_state {
-	unsigned dirty;
+	struct nouveau_stateobj *hw[64];
+	uint64_t hw_dirty;
 
-	struct nouveau_stateobj *fb;
-	struct nouveau_stateobj *blend;
-	struct nouveau_stateobj *blend_colour;
-	struct nouveau_stateobj *zsa;
-	struct nouveau_stateobj *stencil_ref;
-	struct nouveau_stateobj *rast;
-	struct nouveau_stateobj *stipple;
-	struct nouveau_stateobj *scissor;
 	unsigned scissor_enabled;
-	struct nouveau_stateobj *viewport;
-	struct nouveau_stateobj *tsc_upload;
-	struct nouveau_stateobj *tic_upload;
 	unsigned miptree_nr[PIPE_SHADER_TYPES];
-	struct nouveau_stateobj *vertprog;
-	struct nouveau_stateobj *fragprog;
-	struct nouveau_stateobj *geomprog;
-	struct nouveau_stateobj *fp_linkage;
-	struct nouveau_stateobj *gp_linkage;
-	struct nouveau_stateobj *vtxfmt;
 	struct nouveau_stateobj *vtxbuf;
 	struct nouveau_stateobj *vtxattr;
 	struct nouveau_stateobj *instbuf;
@@ -224,18 +208,23 @@ extern void nv50_draw_elements_instanced(struct pipe_context *pipe,
 					 unsigned startInstance,
 					 unsigned instanceCount);
 extern void nv50_vtxelt_construct(struct nv50_vtxelt_stateobj *cso);
-extern void nv50_vbo_validate(struct nv50_context *nv50);
+extern struct nouveau_stateobj *nv50_vbo_validate(struct nv50_context *nv50);
 
 /* nv50_clear.c */
 extern void nv50_clear(struct pipe_context *pipe, unsigned buffers,
 		       const float *rgba, double depth, unsigned stencil);
 
 /* nv50_program.c */
-extern void nv50_vertprog_validate(struct nv50_context *nv50);
-extern void nv50_fragprog_validate(struct nv50_context *nv50);
-extern void nv50_geomprog_validate(struct nv50_context *nv50);
-extern void nv50_fp_linkage_validate(struct nv50_context *nv50);
-extern void nv50_gp_linkage_validate(struct nv50_context *nv50);
+extern struct nouveau_stateobj *
+nv50_vertprog_validate(struct nv50_context *nv50);
+extern struct nouveau_stateobj *
+nv50_fragprog_validate(struct nv50_context *nv50);
+extern struct nouveau_stateobj *
+nv50_geomprog_validate(struct nv50_context *nv50);
+extern struct nouveau_stateobj *
+nv50_fp_linkage_validate(struct nv50_context *nv50);
+extern struct nouveau_stateobj *
+nv50_gp_linkage_validate(struct nv50_context *nv50);
 extern void nv50_program_destroy(struct nv50_context *nv50,
 				 struct nv50_program *p);
 
@@ -250,7 +239,7 @@ extern void nv50_so_init_sifc(struct nv50_context *nv50,
 
 /* nv50_tex.c */
 extern void nv50_tex_relocs(struct nv50_context *);
-extern void nv50_tex_validate(struct nv50_context *);
+extern struct nouveau_stateobj *nv50_tex_validate(struct nv50_context *);
 
 /* nv50_transfer.c */
 extern void
diff --git a/src/gallium/drivers/nv50/nv50_program.c b/src/gallium/drivers/nv50/nv50_program.c
index 2372cbbef69..50bb982076e 100644
--- a/src/gallium/drivers/nv50/nv50_program.c
+++ b/src/gallium/drivers/nv50/nv50_program.c
@@ -4270,7 +4270,7 @@ nv50_program_validate_code(struct nv50_context *nv50, struct nv50_program *p)
 	FREE(up);
 }
 
-void
+struct nouveau_stateobj *
 nv50_vertprog_validate(struct nv50_context *nv50)
 {
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
@@ -4301,11 +4301,10 @@ nv50_vertprog_validate(struct nv50_context *nv50)
 	so_data  (so, p->cfg.high_temp);
 	so_method(so, tesla, NV50TCL_VP_START_ID, 1);
 	so_data  (so, 0); /* program start offset */
-	so_ref(so, &nv50->state.vertprog);
-	so_ref(NULL, &so);
+	return so;
 }
 
-void
+struct nouveau_stateobj *
 nv50_fragprog_validate(struct nv50_context *nv50)
 {
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
@@ -4337,11 +4336,10 @@ nv50_fragprog_validate(struct nv50_context *nv50)
 	so_data  (so, p->cfg.regs[3]);
 	so_method(so, tesla, NV50TCL_FP_START_ID, 1);
 	so_data  (so, 0); /* program start offset */
-	so_ref(so, &nv50->state.fragprog);
-	so_ref(NULL, &so);
+	return so;
 }
 
-void
+struct nouveau_stateobj *
 nv50_geomprog_validate(struct nv50_context *nv50)
 {
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
@@ -4373,8 +4371,7 @@ nv50_geomprog_validate(struct nv50_context *nv50)
 	so_data  (so, p->cfg.vert_count);
 	so_method(so, tesla, NV50TCL_GP_START_ID, 1);
 	so_data  (so, 0);
-	so_ref(so, &nv50->state.geomprog);
-	so_ref(NULL, &so);
+	return so;
 }
 
 static uint32_t
@@ -4454,7 +4451,7 @@ nv50_vec4_map(uint32_t *map32, int mid, uint8_t zval, uint32_t lin[4],
 	return mid;
 }
 
-void
+struct nouveau_stateobj *
 nv50_fp_linkage_validate(struct nv50_context *nv50)
 {
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
@@ -4580,8 +4577,7 @@ nv50_fp_linkage_validate(struct nv50_context *nv50)
 	so_method(so, tesla, NV50TCL_GP_ENABLE, 1);
 	so_data  (so, (vp->type == PIPE_SHADER_GEOMETRY) ? 1 : 0);
 
-	so_ref(so, &nv50->state.fp_linkage);
-	so_ref(NULL, &so);
+	return so;
 }
 
 static int
@@ -4615,7 +4611,7 @@ construct_vp_gp_mapping(uint32_t *map32, int m,
 	return m;
 }
 
-void
+struct nouveau_stateobj *
 nv50_gp_linkage_validate(struct nv50_context *nv50)
 {
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
@@ -4625,10 +4621,8 @@ nv50_gp_linkage_validate(struct nv50_context *nv50)
 	uint32_t map[16];
 	int m = 0;
 
-	if (!gp) {
-		so_ref(NULL, &nv50->state.gp_linkage);
-		return;
-	}
+	if (!gp)
+		return NULL;
 	memset(map, 0, sizeof(map));
 
 	m = construct_vp_gp_mapping(map, m, vp, gp);
@@ -4646,8 +4640,7 @@ nv50_gp_linkage_validate(struct nv50_context *nv50)
 	so_method(so, tesla, NV50TCL_VP_RESULT_MAP(0), m);
 	so_datap (so, map, m);
 
-	so_ref(so, &nv50->state.gp_linkage);
-	so_ref(NULL, &so);
+	return so;
 }
 
 void
diff --git a/src/gallium/drivers/nv50/nv50_screen.c b/src/gallium/drivers/nv50/nv50_screen.c
index eed6031eafa..b530de64b6e 100644
--- a/src/gallium/drivers/nv50/nv50_screen.c
+++ b/src/gallium/drivers/nv50/nv50_screen.c
@@ -215,9 +215,9 @@ nv50_pre_pipebuffer_map(struct pipe_screen *pscreen, struct pipe_buffer *pb,
 	/* Our vtxbuf got mapped, it can no longer be considered part of current
 	 * state, remove it to avoid emitting reloc markers.
 	 */
-	if (ctx && ctx->state.vtxbuf && so_bo_is_reloc(ctx->state.vtxbuf,
+	if (ctx && ctx->state.hw[17] && so_bo_is_reloc(ctx->state.hw[17],
 			nouveau_bo(pb))) {
-		so_ref(NULL, &ctx->state.vtxbuf);
+		so_ref(NULL, &ctx->state.hw[17]);
 		ctx->dirty |= NV50_NEW_ARRAYS;
 	}
 
diff --git a/src/gallium/drivers/nv50/nv50_state_validate.c b/src/gallium/drivers/nv50/nv50_state_validate.c
index 0d30957ee3c..abdb9a55c85 100644
--- a/src/gallium/drivers/nv50/nv50_state_validate.c
+++ b/src/gallium/drivers/nv50/nv50_state_validate.c
@@ -25,8 +25,8 @@
 #include "nv50_context.h"
 #include "nouveau/nouveau_stateobj.h"
 
-static void
-nv50_state_validate_fb(struct nv50_context *nv50)
+static struct nouveau_stateobj *
+validate_fb(struct nv50_context *nv50)
 {
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
 	struct nouveau_stateobj *so = so_new(32, 79, 18);
@@ -169,10 +169,8 @@ nv50_state_validate_fb(struct nv50_context *nv50)
 
 	/* we set scissors to framebuffer size when they're 'turned off' */
 	nv50->dirty |= NV50_NEW_SCISSOR;
-	so_ref(NULL, &nv50->state.scissor);
-
-	so_ref(so, &nv50->state.fb);
-	so_ref(NULL, &so);
+	so_ref(NULL, &nv50->state.hw[12]);
+	return so;
 }
 
 static void
@@ -199,261 +197,251 @@ nv50_validate_samplers(struct nv50_context *nv50, struct nouveau_stateobj *so,
 	}
 }
 
-static void
-nv50_state_emit(struct nv50_context *nv50)
+static struct nouveau_stateobj *
+validate_blend(struct nv50_context *nv50)
 {
-	struct nv50_screen *screen = nv50->screen;
-	struct nouveau_channel *chan = screen->base.channel;
+	struct nouveau_stateobj *so = NULL;
+	so_ref(nv50->blend->so, &so);
+	return so;
+}
 
-	/* XXX: this is racy for multiple contexts active on separate
-	 * threads.
-	 */
-	if (screen->cur_ctx != nv50) {
-		if (nv50->state.fb)
-			nv50->state.dirty |= NV50_NEW_FRAMEBUFFER;
-		if (nv50->state.blend)
-			nv50->state.dirty |= NV50_NEW_BLEND;
-		if (nv50->state.zsa)
-			nv50->state.dirty |= NV50_NEW_ZSA;
-		if (nv50->state.vertprog)
-			nv50->state.dirty |= NV50_NEW_VERTPROG;
-		if (nv50->state.fragprog)
-			nv50->state.dirty |= NV50_NEW_FRAGPROG;
-		if (nv50->state.geomprog)
-			nv50->state.dirty |= NV50_NEW_GEOMPROG;
-		if (nv50->state.rast)
-			nv50->state.dirty |= NV50_NEW_RASTERIZER;
-		if (nv50->state.blend_colour)
-			nv50->state.dirty |= NV50_NEW_BLEND_COLOUR;
-		if (nv50->state.stencil_ref)
-			nv50->state.dirty |= NV50_NEW_STENCIL_REF;
-		if (nv50->state.stipple)
-			nv50->state.dirty |= NV50_NEW_STIPPLE;
-		if (nv50->state.scissor)
-			nv50->state.dirty |= NV50_NEW_SCISSOR;
-		if (nv50->state.viewport)
-			nv50->state.dirty |= NV50_NEW_VIEWPORT;
-		if (nv50->state.tsc_upload)
-			nv50->state.dirty |= NV50_NEW_SAMPLER;
-		if (nv50->state.tic_upload)
-			nv50->state.dirty |= NV50_NEW_TEXTURE;
-		if (nv50->state.vtxfmt && nv50->state.vtxbuf)
-			nv50->state.dirty |= NV50_NEW_ARRAYS;
-		screen->cur_ctx = nv50;
-	}
+static struct nouveau_stateobj *
+validate_zsa(struct nv50_context *nv50)
+{
+	struct nouveau_stateobj *so = NULL;
+	so_ref(nv50->zsa->so, &so);
+	return so;
+}
 
-	if (nv50->state.dirty & NV50_NEW_FRAMEBUFFER)
-		so_emit(chan, nv50->state.fb);
-	if (nv50->state.dirty & NV50_NEW_BLEND)
-		so_emit(chan, nv50->state.blend);
-	if (nv50->state.dirty & NV50_NEW_ZSA)
-		so_emit(chan, nv50->state.zsa);
-	if (nv50->state.dirty & NV50_NEW_VERTPROG)
-		so_emit(chan, nv50->state.vertprog);
-	if (nv50->state.dirty & NV50_NEW_FRAGPROG)
-		so_emit(chan, nv50->state.fragprog);
-	if (nv50->state.dirty & NV50_NEW_GEOMPROG && nv50->state.geomprog)
-		so_emit(chan, nv50->state.geomprog);
-	if (nv50->state.dirty & (NV50_NEW_FRAGPROG | NV50_NEW_VERTPROG |
-				 NV50_NEW_GEOMPROG | NV50_NEW_RASTERIZER))
-		so_emit(chan, nv50->state.fp_linkage);
-	if ((nv50->state.dirty & (NV50_NEW_VERTPROG | NV50_NEW_GEOMPROG))
-	    && nv50->state.gp_linkage)
-		so_emit(chan, nv50->state.gp_linkage);
-	if (nv50->state.dirty & NV50_NEW_RASTERIZER)
-		so_emit(chan, nv50->state.rast);
-	if (nv50->state.dirty & NV50_NEW_BLEND_COLOUR)
-		so_emit(chan, nv50->state.blend_colour);
-	if (nv50->state.dirty & NV50_NEW_STENCIL_REF)
-		so_emit(chan, nv50->state.stencil_ref);
-	if (nv50->state.dirty & NV50_NEW_STIPPLE)
-		so_emit(chan, nv50->state.stipple);
-	if (nv50->state.dirty & NV50_NEW_SCISSOR)
-		so_emit(chan, nv50->state.scissor);
-	if (nv50->state.dirty & NV50_NEW_VIEWPORT)
-		so_emit(chan, nv50->state.viewport);
-	if (nv50->state.dirty & NV50_NEW_SAMPLER)
-		so_emit(chan, nv50->state.tsc_upload);
-	if (nv50->state.dirty & NV50_NEW_TEXTURE)
-		so_emit(chan, nv50->state.tic_upload);
-	if (nv50->state.dirty & NV50_NEW_ARRAYS) {
-		so_emit(chan, nv50->state.vtxfmt);
-		so_emit(chan, nv50->state.vtxbuf);
-		if (nv50->state.vtxattr)
-			so_emit(chan, nv50->state.vtxattr);
-	}
-	nv50->state.dirty = 0;
+static struct nouveau_stateobj *
+validate_rast(struct nv50_context *nv50)
+{
+	struct nouveau_stateobj *so = NULL;
+	so_ref(nv50->rasterizer->so, &so);
+	return so;
 }
 
-void
-nv50_state_flush_notify(struct nouveau_channel *chan)
+static struct nouveau_stateobj *
+validate_blend_colour(struct nv50_context *nv50)
 {
-	struct nv50_context *nv50 = chan->user_private;
+	struct nouveau_grobj *tesla = nv50->screen->tesla;
+	struct nouveau_stateobj *so = so_new(1, 4, 0);
+
+	so_method(so, tesla, NV50TCL_BLEND_COLOR(0), 4);
+	so_data  (so, fui(nv50->blend_colour.color[0]));
+	so_data  (so, fui(nv50->blend_colour.color[1]));
+	so_data  (so, fui(nv50->blend_colour.color[2]));
+	so_data  (so, fui(nv50->blend_colour.color[3]));
+	return so;
+}
 
-	nv50_tex_relocs(nv50);
+static struct nouveau_stateobj *
+validate_stencil_ref(struct nv50_context *nv50)
+{
+	struct nouveau_grobj *tesla = nv50->screen->tesla;
+	struct nouveau_stateobj *so = so = so_new(2, 2, 0);
 
-	so_emit_reloc_markers(chan, nv50->state.fb);
-	so_emit_reloc_markers(chan, nv50->state.vertprog);
-	so_emit_reloc_markers(chan, nv50->state.fragprog);
-	so_emit_reloc_markers(chan, nv50->state.vtxbuf);
-	so_emit_reloc_markers(chan, nv50->screen->static_init);
+	so_method(so, tesla, NV50TCL_STENCIL_FRONT_FUNC_REF, 1);
+	so_data  (so, nv50->stencil_ref.ref_value[0]);
+	so_method(so, tesla, NV50TCL_STENCIL_BACK_FUNC_REF, 1);
+	so_data  (so, nv50->stencil_ref.ref_value[1]);
+	return so;
+}
 
-	if (nv50->state.instbuf)
-		so_emit_reloc_markers(chan, nv50->state.instbuf);
+static struct nouveau_stateobj *
+validate_stipple(struct nv50_context *nv50)
+{
+	struct nouveau_grobj *tesla = nv50->screen->tesla;
+	struct nouveau_stateobj *so = so_new(1, 32, 0);
+	int i;
+
+	so_method(so, tesla, NV50TCL_POLYGON_STIPPLE_PATTERN(0), 32);
+	for (i = 0; i < 32; i++)
+		so_data(so, util_bswap32(nv50->stipple.stipple[i]));
+	return so;
 }
 
-boolean
-nv50_state_validate(struct nv50_context *nv50)
+static struct nouveau_stateobj *
+validate_scissor(struct nv50_context *nv50)
 {
+	struct pipe_rasterizer_state *rast = &nv50->rasterizer->pipe;
+	struct pipe_scissor_state *s = &nv50->scissor;
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
 	struct nouveau_stateobj *so;
-	unsigned i;
 
-	if (nv50->dirty & NV50_NEW_FRAMEBUFFER)
-		nv50_state_validate_fb(nv50);
+	if (nv50->state.hw[12] &&
+	    (rast->scissor == 0 && nv50->state.scissor_enabled == 0))
+		return NULL;
+	nv50->state.scissor_enabled = rast->scissor;
 
-	if (nv50->dirty & NV50_NEW_BLEND)
-		so_ref(nv50->blend->so, &nv50->state.blend);
+	so = so_new(1, 2, 0);
+	so_method(so, tesla, NV50TCL_SCISSOR_HORIZ(0), 2);
+	if (nv50->state.scissor_enabled) {
+		so_data(so, (s->maxx << 16) | s->minx);
+		so_data(so, (s->maxy << 16) | s->miny);
+	} else {
+		so_data(so, (nv50->framebuffer.width << 16));
+		so_data(so, (nv50->framebuffer.height << 16));
+	}
 
-	if (nv50->dirty & NV50_NEW_ZSA)
-		so_ref(nv50->zsa->so, &nv50->state.zsa);
+	return so;
+}
 
-	if (nv50->dirty & (NV50_NEW_VERTPROG | NV50_NEW_VERTPROG_CB))
-		nv50_vertprog_validate(nv50);
+static struct nouveau_stateobj *
+validate_viewport(struct nv50_context *nv50)
+{
+	struct nouveau_grobj *tesla = nv50->screen->tesla;
+	struct nouveau_stateobj *so = so_new(5, 9, 0);
+
+	so_method(so, tesla, NV50TCL_VIEWPORT_TRANSLATE_X(0), 3);
+	so_data  (so, fui(nv50->viewport.translate[0]));
+	so_data  (so, fui(nv50->viewport.translate[1]));
+	so_data  (so, fui(nv50->viewport.translate[2]));
+	so_method(so, tesla, NV50TCL_VIEWPORT_SCALE_X(0), 3);
+	so_data  (so, fui(nv50->viewport.scale[0]));
+	so_data  (so, fui(nv50->viewport.scale[1]));
+	so_data  (so, fui(nv50->viewport.scale[2]));
+
+	so_method(so, tesla, NV50TCL_VIEWPORT_TRANSFORM_EN, 1);
+	so_data  (so, 1);
+	/* 0x0000 = remove whole primitive only (xyz)
+	 * 0x1018 = remove whole primitive only (xy), clamp z
+	 * 0x1080 = clip primitive (xyz)
+	 * 0x1098 = clip primitive (xy), clamp z
+	 */
+	so_method(so, tesla, NV50TCL_VIEW_VOLUME_CLIP_CTRL, 1);
+	so_data  (so, 0x1080);
+	/* no idea what 0f90 does */
+	so_method(so, tesla, 0x0f90, 1);
+	so_data  (so, 0);
 
-	if (nv50->dirty & (NV50_NEW_FRAGPROG | NV50_NEW_FRAGPROG_CB))
-		nv50_fragprog_validate(nv50);
+	return so;
+}
 
-	if (nv50->dirty & (NV50_NEW_GEOMPROG | NV50_NEW_GEOMPROG_CB))
-		nv50_geomprog_validate(nv50);
+static struct nouveau_stateobj *
+validate_sampler(struct nv50_context *nv50)
+{
+	struct nouveau_grobj *tesla = nv50->screen->tesla;
+	struct nouveau_stateobj *so;
+	unsigned nr = 0, i;
 
-	if (nv50->dirty & (NV50_NEW_FRAGPROG | NV50_NEW_VERTPROG |
-			   NV50_NEW_GEOMPROG | NV50_NEW_RASTERIZER))
-		nv50_fp_linkage_validate(nv50);
+	for (i = 0; i < PIPE_SHADER_TYPES; ++i)
+		nr += nv50->sampler_nr[i];
 
-	if (nv50->dirty & (NV50_NEW_GEOMPROG | NV50_NEW_VERTPROG))
-		nv50_gp_linkage_validate(nv50);
+	so = so_new(1 + 5 * PIPE_SHADER_TYPES,
+		    1 + 19 * PIPE_SHADER_TYPES + nr * 8,
+		    PIPE_SHADER_TYPES * 2);
 
-	if (nv50->dirty & NV50_NEW_RASTERIZER)
-		so_ref(nv50->rasterizer->so, &nv50->state.rast);
+	nv50_validate_samplers(nv50, so, PIPE_SHADER_VERTEX);
+	nv50_validate_samplers(nv50, so, PIPE_SHADER_FRAGMENT);
 
-	if (nv50->dirty & NV50_NEW_BLEND_COLOUR) {
-		so = so_new(1, 4, 0);
-		so_method(so, tesla, NV50TCL_BLEND_COLOR(0), 4);
-		so_data  (so, fui(nv50->blend_colour.color[0]));
-		so_data  (so, fui(nv50->blend_colour.color[1]));
-		so_data  (so, fui(nv50->blend_colour.color[2]));
-		so_data  (so, fui(nv50->blend_colour.color[3]));
-		so_ref(so, &nv50->state.blend_colour);
-		so_ref(NULL, &so);
-	}
-
-	if (nv50->dirty & NV50_NEW_STENCIL_REF) {
-		so = so_new(2, 2, 0);
-		so_method(so, tesla, NV50TCL_STENCIL_FRONT_FUNC_REF, 1);
-		so_data  (so, nv50->stencil_ref.ref_value[0]);
-		so_method(so, tesla, NV50TCL_STENCIL_BACK_FUNC_REF, 1);
-		so_data  (so, nv50->stencil_ref.ref_value[1]);
-		so_ref(so, &nv50->state.stencil_ref);
-		so_ref(NULL, &so);
-	}
+	so_method(so, tesla, 0x1334, 1); /* flush TSC */
+	so_data  (so, 0);
 
-	if (nv50->dirty & NV50_NEW_STIPPLE) {
-		so = so_new(1, 32, 0);
-		so_method(so, tesla, NV50TCL_POLYGON_STIPPLE_PATTERN(0), 32);
-		for (i = 0; i < 32; i++)
-			so_data(so, util_bswap32(nv50->stipple.stipple[i]));
-		so_ref(so, &nv50->state.stipple);
-		so_ref(NULL, &so);
-	}
+	return so;
+}
 
-	if (nv50->dirty & (NV50_NEW_SCISSOR | NV50_NEW_RASTERIZER)) {
-		struct pipe_rasterizer_state *rast = &nv50->rasterizer->pipe;
-		struct pipe_scissor_state *s = &nv50->scissor;
+static struct nouveau_stateobj *
+validate_vtxbuf(struct nv50_context *nv50)
+{
+	struct nouveau_stateobj *so = NULL;
+	so_ref(nv50->state.vtxbuf, &so);
+	return so;
+}
 
-		if (nv50->state.scissor &&
-		    (rast->scissor == 0 && nv50->state.scissor_enabled == 0))
-			goto scissor_uptodate;
-		nv50->state.scissor_enabled = rast->scissor;
+static struct nouveau_stateobj *
+validate_vtxattr(struct nv50_context *nv50)
+{
+	struct nouveau_stateobj *so = NULL;
+	so_ref(nv50->state.vtxattr, &so);
+	return so;
+}
 
-		so = so_new(1, 2, 0);
-		so_method(so, tesla, NV50TCL_SCISSOR_HORIZ(0), 2);
-		if (nv50->state.scissor_enabled) {
-			so_data(so, (s->maxx << 16) | s->minx);
-			so_data(so, (s->maxy << 16) | s->miny);
-		} else {
-			so_data(so, (nv50->framebuffer.width << 16));
-			so_data(so, (nv50->framebuffer.height << 16));
-		}
-		so_ref(so, &nv50->state.scissor);
-		so_ref(NULL, &so);
-		nv50->state.dirty |= NV50_NEW_SCISSOR;
-	}
-scissor_uptodate:
-
-	if (nv50->dirty & (NV50_NEW_VIEWPORT | NV50_NEW_RASTERIZER)) {
-		if (nv50->state.viewport &&
-		    !(nv50->dirty & NV50_NEW_VIEWPORT))
-			goto viewport_uptodate;
-
-		so = so_new(5, 9, 0);
-		so_method(so, tesla, NV50TCL_VIEWPORT_TRANSLATE_X(0), 3);
-		so_data  (so, fui(nv50->viewport.translate[0]));
-		so_data  (so, fui(nv50->viewport.translate[1]));
-		so_data  (so, fui(nv50->viewport.translate[2]));
-		so_method(so, tesla, NV50TCL_VIEWPORT_SCALE_X(0), 3);
-		so_data  (so, fui(nv50->viewport.scale[0]));
-		so_data  (so, fui(nv50->viewport.scale[1]));
-		so_data  (so, fui(nv50->viewport.scale[2]));
-
-		so_method(so, tesla, NV50TCL_VIEWPORT_TRANSFORM_EN, 1);
-		so_data  (so, 1);
-		/* 0x0000 = remove whole primitive only (xyz)
-		 * 0x1018 = remove whole primitive only (xy), clamp z
-		 * 0x1080 = clip primitive (xyz)
-		 * 0x1098 = clip primitive (xy), clamp z
-		 */
-		so_method(so, tesla, NV50TCL_VIEW_VOLUME_CLIP_CTRL, 1);
-		so_data  (so, 0x1080);
-		/* no idea what 0f90 does */
-		so_method(so, tesla, 0x0f90, 1);
-		so_data  (so, 0);
+struct state_validate {
+	struct nouveau_stateobj *(*func)(struct nv50_context *nv50);
+	unsigned states;
+} validate_list[] = {
+	{ validate_fb             , NV50_NEW_FRAMEBUFFER                      },
+	{ validate_blend          , NV50_NEW_BLEND                            },
+	{ validate_zsa            , NV50_NEW_ZSA                              },
+	{ nv50_vertprog_validate  , NV50_NEW_VERTPROG | NV50_NEW_VERTPROG_CB  },
+	{ nv50_fragprog_validate  , NV50_NEW_FRAGPROG | NV50_NEW_FRAGPROG_CB  },
+	{ nv50_geomprog_validate  , NV50_NEW_GEOMPROG | NV50_NEW_GEOMPROG_CB  },
+	{ nv50_fp_linkage_validate, NV50_NEW_VERTPROG | NV50_NEW_GEOMPROG |
+				    NV50_NEW_FRAGPROG | NV50_NEW_RASTERIZER   },
+	{ nv50_gp_linkage_validate, NV50_NEW_VERTPROG | NV50_NEW_GEOMPROG     },
+	{ validate_rast           , NV50_NEW_RASTERIZER                       },
+	{ validate_blend_colour   , NV50_NEW_BLEND_COLOUR                     },
+	{ validate_stencil_ref    , NV50_NEW_STENCIL_REF                      },
+	{ validate_stipple        , NV50_NEW_STIPPLE                          },
+	{ validate_scissor        , NV50_NEW_SCISSOR | NV50_NEW_RASTERIZER    },
+	{ validate_viewport       , NV50_NEW_VIEWPORT                         },
+	{ validate_sampler        , NV50_NEW_SAMPLER                          },
+	{ nv50_tex_validate       , NV50_NEW_TEXTURE | NV50_NEW_SAMPLER       },
+	{ nv50_vbo_validate       , NV50_NEW_ARRAYS                           },
+	{ validate_vtxbuf         , NV50_NEW_ARRAYS                           },
+	{ validate_vtxattr        , NV50_NEW_ARRAYS                           },
+	{}
+};
+#define validate_list_len (sizeof(validate_list) / sizeof(validate_list[0]))
 
-		so_ref(so, &nv50->state.viewport);
-		so_ref(NULL, &so);
-		nv50->state.dirty |= NV50_NEW_VIEWPORT;
-	}
-viewport_uptodate:
+void
+nv50_state_flush_notify(struct nouveau_channel *chan)
+{
+	struct nv50_context *nv50 = chan->user_private;
 
-	if (nv50->dirty & NV50_NEW_SAMPLER) {
-		unsigned nr = 0;
+	nv50_tex_relocs(nv50);
 
-		for (i = 0; i < PIPE_SHADER_TYPES; ++i)
-			nr += nv50->sampler_nr[i];
+	so_emit_reloc_markers(chan, nv50->state.hw[0]); /* fb */
+	so_emit_reloc_markers(chan, nv50->state.hw[3]); /* vp */
+	so_emit_reloc_markers(chan, nv50->state.hw[4]); /* fp */
+	so_emit_reloc_markers(chan, nv50->state.hw[17]); /* vb */
+	so_emit_reloc_markers(chan, nv50->screen->static_init);
 
-		so = so_new(1 + 5 * PIPE_SHADER_TYPES,
-			    1 + 19 * PIPE_SHADER_TYPES + nr * 8,
-			    PIPE_SHADER_TYPES * 2);
+#if 0
+	if (nv50->state.instbuf)
+		so_emit_reloc_markers(chan, nv50->state.instbuf);
+#endif
+}
 
-		nv50_validate_samplers(nv50, so, PIPE_SHADER_VERTEX);
-		nv50_validate_samplers(nv50, so, PIPE_SHADER_FRAGMENT);
+boolean
+nv50_state_validate(struct nv50_context *nv50)
+{
+	struct nouveau_channel *chan = nv50->screen->base.channel;
+	int i;
 
-		so_method(so, tesla, 0x1334, 1); /* flush TSC */
-		so_data  (so, 0);
+	if (nv50->screen->cur_ctx != nv50) {
+		for (i = 0; i < validate_list_len; i++) {
+			if (nv50->state.hw[i])
+				nv50->state.hw_dirty |= (1 << i);
+		}
 
-		so_ref(so, &nv50->state.tsc_upload);
-		so_ref(NULL, &so);
+		nv50->screen->cur_ctx = nv50;
 	}
 
-	if (nv50->dirty & (NV50_NEW_TEXTURE | NV50_NEW_SAMPLER))
-		nv50_tex_validate(nv50);
+	for (i = 0; i < validate_list_len; i++) {
+		struct state_validate *validate = &validate_list[i];
+		struct nouveau_stateobj *so;
+
+		if (!(nv50->dirty & validate->states))
+			continue;
 
-	if (nv50->dirty & NV50_NEW_ARRAYS)
-		nv50_vbo_validate(nv50);
+		so = validate->func(nv50);
+		if (!so)
+			continue;
 
-	nv50->state.dirty |= nv50->dirty;
+		so_ref(so, &nv50->state.hw[i]);
+		so_ref(NULL, &so);
+		nv50->state.hw_dirty |= (1 << i);
+	}
 	nv50->dirty = 0;
-	nv50_state_emit(nv50);
+
+	while (nv50->state.hw_dirty) {
+		i = ffs(nv50->state.hw_dirty) - 1;
+		nv50->state.hw_dirty &= ~(1 << i);
+
+		so_emit(chan, nv50->state.hw[i]);
+	}
 
 	return TRUE;
 }
diff --git a/src/gallium/drivers/nv50/nv50_tex.c b/src/gallium/drivers/nv50/nv50_tex.c
index ed3440fcc19..4c48b12cd87 100644
--- a/src/gallium/drivers/nv50/nv50_tex.c
+++ b/src/gallium/drivers/nv50/nv50_tex.c
@@ -224,7 +224,7 @@ nv50_tex_relocs(struct nv50_context *nv50)
 	}
 }
 
-void
+struct nouveau_stateobj *
 nv50_tex_validate(struct nv50_context *nv50)
 {
 	struct nouveau_stateobj *so;
@@ -247,12 +247,11 @@ nv50_tex_validate(struct nv50_context *nv50)
 		so_ref(NULL, &so);
 
 		NOUVEAU_ERR("failed tex validate\n");
-		return;
+		return NULL;
 	}
 
 	so_method(so, tesla, 0x1330, 1); /* flush TIC */
 	so_data  (so, 0);
 
-	so_ref(so, &nv50->state.tic_upload);
-	so_ref(NULL, &so);
+	return so;
 }
diff --git a/src/gallium/drivers/nv50/nv50_vbo.c b/src/gallium/drivers/nv50/nv50_vbo.c
index c1dcb93b48c..3e540013c42 100644
--- a/src/gallium/drivers/nv50/nv50_vbo.c
+++ b/src/gallium/drivers/nv50/nv50_vbo.c
@@ -792,7 +792,7 @@ nv50_vtxelt_construct(struct nv50_vtxelt_stateobj *cso)
 	}
 }
 
-void
+struct nouveau_stateobj *
 nv50_vbo_validate(struct nv50_context *nv50)
 {
 	struct nouveau_grobj *tesla = nv50->screen->tesla;
@@ -801,7 +801,7 @@ nv50_vbo_validate(struct nv50_context *nv50)
 
 	/* don't validate if Gallium took away our buffers */
 	if (nv50->vtxbuf_nr == 0)
-		return;
+		return NULL;
 	nv50->vbo_fifo = 0;
 
 	for (i = 0; i < nv50->vtxbuf_nr; ++i)
@@ -875,12 +875,11 @@ nv50_vbo_validate(struct nv50_context *nv50)
 	}
 	nv50->state.vtxelt_nr = nv50->vtxelt->num_elements;
 
-	so_ref (vtxfmt, &nv50->state.vtxfmt);
 	so_ref (vtxbuf, &nv50->state.vtxbuf);
 	so_ref (vtxattr, &nv50->state.vtxattr);
 	so_ref (NULL, &vtxbuf);
-	so_ref (NULL, &vtxfmt);
 	so_ref (NULL, &vtxattr);
+	return vtxfmt;
 }
 
 typedef void (*pfn_push)(struct nouveau_channel *, void *);