6 files changed, 42 insertions, 30 deletions

diff --git a/src/gallium/drivers/nv30/nv30_context.h b/src/gallium/drivers/nv30/nv30_context.h
index c3c8b733090..5d7080a555d 100644
--- a/src/gallium/drivers/nv30/nv30_context.h
+++ b/src/gallium/drivers/nv30/nv30_context.h
@@ -124,6 +124,7 @@ struct nv30_context {
 	unsigned dirty;
 	struct pipe_scissor_state scissor;
 	unsigned stipple[32];
+	struct nv30_vertex_program *vertprog;
 	struct pipe_buffer *constbuf[PIPE_SHADER_TYPES];
 	unsigned constbuf_nr[PIPE_SHADER_TYPES];
 	struct nv30_rasterizer_state *rasterizer;
@@ -151,11 +152,6 @@ struct nv30_context {
 	} vb[16];
 
 	struct {
-		struct nv30_vertex_program *active;
-		struct nv30_vertex_program *current;
-	} vertprog;
-
-	struct {
 		struct nv30_fragment_program *active;
 		struct nv30_fragment_program *current;
 	} fragprog;
@@ -188,10 +184,6 @@ extern void nv30_screen_init_miptree_functions(struct pipe_screen *pscreen);
 extern struct draw_stage *nv30_draw_render_stage(struct nv30_context *nv30);
 
 /* nv30_vertprog.c */
-extern void nv30_vertprog_translate(struct nv30_context *,
-				    struct nv30_vertex_program *);
-extern void nv30_vertprog_bind(struct nv30_context *,
-			       struct nv30_vertex_program *);
 extern void nv30_vertprog_destroy(struct nv30_context *,
 				  struct nv30_vertex_program *);
 
@@ -213,6 +205,7 @@ extern void nv30_state_tex_update(struct nv30_context *nv30);
 extern struct nv30_state_entry nv30_state_rasterizer;
 extern struct nv30_state_entry nv30_state_scissor;
 extern struct nv30_state_entry nv30_state_stipple;
+extern struct nv30_state_entry nv30_state_vertprog;
 extern struct nv30_state_entry nv30_state_blend;
 extern struct nv30_state_entry nv30_state_blend_colour;
 extern struct nv30_state_entry nv30_state_zsa;
diff --git a/src/gallium/drivers/nv30/nv30_state.c b/src/gallium/drivers/nv30/nv30_state.c
index aec6fa2b155..92695ce23c0 100644
--- a/src/gallium/drivers/nv30/nv30_state.c
+++ b/src/gallium/drivers/nv30/nv30_state.c
@@ -499,10 +499,12 @@ static void *
 nv30_vp_state_create(struct pipe_context *pipe,
 		     const struct pipe_shader_state *cso)
 {
+	/*struct nv30_context *nv30 = nv30_context(pipe);*/
 	struct nv30_vertex_program *vp;
 
 	vp = CALLOC(1, sizeof(struct nv30_vertex_program));
 	vp->pipe = *cso;
+	/*vp->draw = draw_create_vertex_shader(nv30->draw, &vp->pipe);*/
 
 	return (void *)vp;
 }
@@ -511,14 +513,10 @@ static void
 nv30_vp_state_bind(struct pipe_context *pipe, void *hwcso)
 {
 	struct nv30_context *nv30 = nv30_context(pipe);
-	struct nv30_vertex_program *vp = hwcso;
-
-	if (!hwcso) {
-		return;
-	}
 
-	nv30->vertprog.current = vp;
+	nv30->vertprog = hwcso;
 	nv30->dirty |= NV30_NEW_VERTPROG;
+	/*nv30->draw_dirty |= NV30_NEW_VERTPROG;*/
 }
 
 static void
@@ -527,6 +525,7 @@ nv30_vp_state_delete(struct pipe_context *pipe, void *hwcso)
 	struct nv30_context *nv30 = nv30_context(pipe);
 	struct nv30_vertex_program *vp = hwcso;
 
+	/*draw_delete_vertex_shader(nv30->draw, vp->draw);*/
 	nv30_vertprog_destroy(nv30, vp);
 	FREE(vp);
 }
diff --git a/src/gallium/drivers/nv30/nv30_state.h b/src/gallium/drivers/nv30/nv30_state.h
index c65a937467e..a897bc50687 100644
--- a/src/gallium/drivers/nv30/nv30_state.h
+++ b/src/gallium/drivers/nv30/nv30_state.h
@@ -26,6 +26,7 @@ struct nv30_vertex_program {
 	struct pipe_shader_state pipe;
 
 	boolean translated;
+
 	struct nv30_vertex_program_exec *insns;
 	unsigned nr_insns;
 	struct nv30_vertex_program_data *consts;
@@ -39,6 +40,7 @@ struct nv30_vertex_program {
 
 	uint32_t ir;
 	uint32_t or;
+	struct nouveau_stateobj *so;
 };
 
 struct nv30_fragment_program_data {
diff --git a/src/gallium/drivers/nv30/nv30_state_emit.c b/src/gallium/drivers/nv30/nv30_state_emit.c
index cf10ddd4b63..4ab62ddc016 100644
--- a/src/gallium/drivers/nv30/nv30_state_emit.c
+++ b/src/gallium/drivers/nv30/nv30_state_emit.c
@@ -7,6 +7,7 @@ static struct nv30_state_entry *render_states[] = {
 	&nv30_state_scissor,
 	&nv30_state_stipple,
 	&nv30_state_fragtex,
+	&nv30_state_vertprog,
 	&nv30_state_blend,
 	&nv30_state_blend_colour,
 	&nv30_state_zsa,
@@ -74,11 +75,6 @@ nv30_emit_hw_state(struct nv30_context *nv30)
 		states &= ~(1ULL << i);
 	}
 
-	if (nv30->dirty & NV30_NEW_VERTPROG) {
-		nv30_vertprog_bind(nv30, nv30->vertprog.current);
-		nv30->dirty &= ~NV30_NEW_VERTPROG;
-	}
-
 	so_emit_reloc_markers(nv30->nvws, state->hw[NV30_STATE_FB]);
 	for (i = 0, samplers = state->fp_samplers; i < 16 && samplers; i++) {
 		if (!(samplers & (1 << i)))
diff --git a/src/gallium/drivers/nv30/nv30_vbo.c b/src/gallium/drivers/nv30/nv30_vbo.c
index d23164eb485..8e4ee7a8742 100644
--- a/src/gallium/drivers/nv30/nv30_vbo.c
+++ b/src/gallium/drivers/nv30/nv30_vbo.c
@@ -100,7 +100,7 @@ nv30_vbo_static_attrib(struct nv30_context *nv30, int attrib,
 static void
 nv30_vbo_arrays_update(struct nv30_context *nv30)
 {
-	struct nv30_vertex_program *vp = nv30->vertprog.active;
+	struct nv30_vertex_program *vp = nv30->vertprog;
 	uint32_t inputs, vtxfmt[16];
 	int hw, num_hw = 0;
 
diff --git a/src/gallium/drivers/nv30/nv30_vertprog.c b/src/gallium/drivers/nv30/nv30_vertprog.c
index 90756febd28..65e00132055 100644
--- a/src/gallium/drivers/nv30/nv30_vertprog.c
+++ b/src/gallium/drivers/nv30/nv30_vertprog.c
@@ -563,7 +563,7 @@ nv30_vertprog_prepare(struct nv30_vpc *vpc)
 	return TRUE;
 }
 
-void
+static void
 nv30_vertprog_translate(struct nv30_context *nv30,
 			struct nv30_vertex_program *vp)
 {
@@ -636,27 +636,31 @@ out_err:
 	FREE(vpc);
 }
 
-void
-nv30_vertprog_bind(struct nv30_context *nv30, struct nv30_vertex_program *vp)
+static boolean
+nv30_vertprog_validate(struct nv30_context *nv30)
 { 
 	struct nouveau_winsys *nvws = nv30->nvws;
 	struct pipe_winsys *ws = nv30->pipe.winsys;
+	struct nouveau_grobj *rankine = nv30->screen->rankine;
+	struct nv30_vertex_program *vp;
 	struct pipe_buffer *constbuf;
 	boolean upload_code = FALSE, upload_data = FALSE;
 	int i;
 
+	vp = nv30->vertprog;
 	constbuf = nv30->constbuf[PIPE_SHADER_VERTEX];
 
 	/* Translate TGSI shader into hw bytecode */
 	if (!vp->translated) {
 		nv30_vertprog_translate(nv30, vp);
 		if (!vp->translated)
-			assert(0);
+			return FALSE;
 	}
 
 	/* Allocate hw vtxprog exec slots */
 	if (!vp->exec) {
 		struct nouveau_resource *heap = nv30->screen->vp_exec_heap;
+		struct nouveau_stateobj *so;
 		uint vplen = vp->nr_insns;
 
 		if (nvws->res_alloc(heap, vplen, vp, &vp->exec)) {
@@ -671,6 +675,15 @@ nv30_vertprog_bind(struct nv30_context *nv30, struct nv30_vertex_program *vp)
 				assert(0);
 		}
 
+		so = so_new(2, 0);
+		so_method(so, rankine, NV34TCL_VP_START_FROM_ID, 1);
+		so_data  (so, vp->exec->start);
+		/* Add these, and you'll go 1/3 speed */
+		/*so_method(so, rankine, NV34TCL_VP_ATTRIB_EN, 2);
+		so_data  (so, vp->ir);
+		so_data  (so, vp->or);*/
+		so_ref(so, &vp->so);
+
 		upload_code = TRUE;
 	}
 
@@ -777,10 +790,12 @@ nv30_vertprog_bind(struct nv30_context *nv30, struct nv30_vertex_program *vp)
 		}
 	}
 
-	BEGIN_RING(rankine, NV34TCL_VP_START_FROM_ID, 1);
-	OUT_RING  (vp->exec->start);
+	if (vp->so != nv30->state.hw[NV30_STATE_VERTPROG]) {
+		so_ref(vp->so, &nv30->state.hw[NV30_STATE_VERTPROG]);
+		return TRUE;
+	}
 
-	nv30->vertprog.active = vp;
+	return FALSE;
 }
 
 void
@@ -808,7 +823,14 @@ nv30_vertprog_destroy(struct nv30_context *nv30, struct nv30_vertex_program *vp)
 	vp->data_start = 0;
 	vp->data_start_min = 0;
 
-	/* vp->ir = vp->or = vp->clip_ctrl = 0;
-	so_ref(NULL, &vp->so); */
+	vp->ir = vp->or = 0;
+	so_ref(NULL, &vp->so);
 }
 
+struct nv30_state_entry nv30_state_vertprog = {
+	.validate = nv30_vertprog_validate,
+	.dirty = {
+		.pipe = NV30_NEW_VERTPROG /*| NV30_NEW_UCP*/,
+		.hw = NV30_STATE_VERTPROG,
+	}
+};