7 files changed, 55 insertions, 37 deletions

diff --git a/src/gallium/drivers/nv40/nv40_context.h b/src/gallium/drivers/nv40/nv40_context.h
index ce0933fea4a..28a0274d771 100644
--- a/src/gallium/drivers/nv40/nv40_context.h
+++ b/src/gallium/drivers/nv40/nv40_context.h
@@ -75,6 +75,7 @@ struct nv40_state {
 	} stipple;
 
 	struct nouveau_stateobj *fragprog;
+	struct nouveau_stateobj *vertprog;
 };
 
 struct nv40_context {
@@ -99,6 +100,7 @@ struct nv40_context {
 		struct pipe_scissor_state scissor;
 		unsigned stipple[32];
 		struct pipe_clip_state clip;
+		struct nv40_vertex_program *vertprog;
 		struct nv40_fragment_program *fragprog;
 		struct pipe_buffer *constbuf[PIPE_SHADER_TYPES];
 	} pipe_state;
@@ -117,13 +119,6 @@ struct nv40_context {
 	struct nouveau_stateobj *so_viewport;
 	struct nouveau_stateobj *so_stipple;
 
-	struct {
-		struct nv40_vertex_program *active;
-
-		struct nv40_vertex_program *current;
-		struct pipe_buffer *constant_buf;
-	} vertprog;
-
 	struct pipe_vertex_buffer  vtxbuf[PIPE_ATTRIB_MAX];
 	struct pipe_vertex_element vtxelt[PIPE_ATTRIB_MAX];
 };
@@ -151,10 +146,6 @@ extern void nv40_init_query_functions(struct nv40_context *nv40);
 extern struct draw_stage *nv40_draw_render_stage(struct nv40_context *nv40);
 
 /* nv40_vertprog.c */
-extern void nv40_vertprog_translate(struct nv40_context *,
-				    struct nv40_vertex_program *);
-extern void nv40_vertprog_bind(struct nv40_context *,
-			       struct nv40_vertex_program *);
 extern void nv40_vertprog_destroy(struct nv40_context *,
 				  struct nv40_vertex_program *);
 
@@ -172,6 +163,7 @@ extern struct nv40_state_entry nv40_state_clip;
 extern struct nv40_state_entry nv40_state_scissor;
 extern struct nv40_state_entry nv40_state_stipple;
 extern struct nv40_state_entry nv40_state_fragprog;
+extern struct nv40_state_entry nv40_state_vertprog;
 
 /* nv40_vbo.c */
 extern boolean nv40_draw_arrays(struct pipe_context *, unsigned mode,
diff --git a/src/gallium/drivers/nv40/nv40_fragprog.c b/src/gallium/drivers/nv40/nv40_fragprog.c
index bfc75eb4620..77ac8ab2c61 100644
--- a/src/gallium/drivers/nv40/nv40_fragprog.c
+++ b/src/gallium/drivers/nv40/nv40_fragprog.c
@@ -795,7 +795,6 @@ nv40_fragprog_validate(struct nv40_context *nv40)
 		nv40->pipe_state.constbuf[PIPE_SHADER_FRAGMENT];
 	struct pipe_winsys *ws = nv40->pipe.winsys;
 	struct nouveau_stateobj *so;
-	unsigned new_program = FALSE;
 	int i;
 
 	if (fp->translated)
@@ -806,7 +805,6 @@ nv40_fragprog_validate(struct nv40_context *nv40)
 		nv40->fallback |= NV40_FALLBACK_RAST;
 		return FALSE;
 	}
-	new_program = TRUE;
 
 	fp->buffer = ws->buffer_create(ws, 0x100, 0, fp->insn_len * 4);
 	nv40_fragprog_upload(nv40, fp);
@@ -843,8 +841,12 @@ update_constants:
 			nv40_fragprog_upload(nv40, fp);
 	}
 
-	so_ref(fp->so, &nv40->state.fragprog);
-	return new_program;
+	if (fp->so != nv40->state.fragprog) {
+		so_ref(fp->so, &nv40->state.fragprog);
+		return TRUE;
+	}
+
+	return FALSE;
 }
 
 void
diff --git a/src/gallium/drivers/nv40/nv40_state.c b/src/gallium/drivers/nv40/nv40_state.c
index 2886c6b0dc4..8ffbb131f73 100644
--- a/src/gallium/drivers/nv40/nv40_state.c
+++ b/src/gallium/drivers/nv40/nv40_state.c
@@ -491,9 +491,8 @@ static void
 nv40_vp_state_bind(struct pipe_context *pipe, void *hwcso)
 {
 	struct nv40_context *nv40 = nv40_context(pipe);
-	struct nv40_vertex_program *vp = hwcso;
 
-	nv40->vertprog.current = vp;
+	nv40->pipe_state.vertprog = hwcso;
 	nv40->dirty |= NV40_NEW_VERTPROG;
 }
 
@@ -573,7 +572,7 @@ nv40_set_constant_buffer(struct pipe_context *pipe, uint shader, uint index,
 	struct nv40_context *nv40 = nv40_context(pipe);
 
 	if (shader == PIPE_SHADER_VERTEX) {
-		nv40->vertprog.constant_buf = buf->buffer;
+		nv40->pipe_state.constbuf[PIPE_SHADER_VERTEX] = buf->buffer;
 		nv40->dirty |= NV40_NEW_VERTPROG;
 	} else
 	if (shader == PIPE_SHADER_FRAGMENT) {
diff --git a/src/gallium/drivers/nv40/nv40_state.h b/src/gallium/drivers/nv40/nv40_state.h
index 2701294a077..e5217fe91c4 100644
--- a/src/gallium/drivers/nv40/nv40_state.h
+++ b/src/gallium/drivers/nv40/nv40_state.h
@@ -39,6 +39,7 @@ struct nv40_vertex_program {
 
 	uint32_t ir;
 	uint32_t or;
+	struct nouveau_stateobj *so;
 };
 
 struct nv40_fragment_program_data {
diff --git a/src/gallium/drivers/nv40/nv40_state_emit.c b/src/gallium/drivers/nv40/nv40_state_emit.c
index e10e178432e..e702b103236 100644
--- a/src/gallium/drivers/nv40/nv40_state_emit.c
+++ b/src/gallium/drivers/nv40/nv40_state_emit.c
@@ -27,6 +27,7 @@ static struct nv40_state_entry *render_states[] = {
 	&nv40_state_scissor,
 	&nv40_state_stipple,
 	&nv40_state_fragprog,
+	&nv40_state_vertprog,
 	NULL
 };
 
@@ -116,7 +117,7 @@ nv40_emit_hw_state(struct nv40_context *nv40)
 	}
 
 	if (nv40->dirty & NV40_NEW_VERTPROG) {
-		nv40_vertprog_bind(nv40, nv40->vertprog.current);
+		so_emit(nv40->nvws, nv40->state.vertprog);
 		nv40->dirty &= ~NV40_NEW_VERTPROG;
 	}
 
diff --git a/src/gallium/drivers/nv40/nv40_vbo.c b/src/gallium/drivers/nv40/nv40_vbo.c
index fa827ef0c5d..3bfcb264db1 100644
--- a/src/gallium/drivers/nv40/nv40_vbo.c
+++ b/src/gallium/drivers/nv40/nv40_vbo.c
@@ -101,7 +101,7 @@ static void
 nv40_vbo_arrays_update(struct nv40_context *nv40, struct pipe_buffer *ib,
 		       unsigned ib_format)
 {
-	struct nv40_vertex_program *vp = nv40->vertprog.active;
+	struct nv40_vertex_program *vp = nv40->pipe_state.vertprog;
 	struct nouveau_stateobj *vtxbuf, *vtxfmt;
 	unsigned inputs, hw, num_hw;
 	unsigned vb_flags = NOUVEAU_BO_VRAM | NOUVEAU_BO_GART | NOUVEAU_BO_RD;
diff --git a/src/gallium/drivers/nv40/nv40_vertprog.c b/src/gallium/drivers/nv40/nv40_vertprog.c
index 9f4738b8308..4a15e51eb33 100644
--- a/src/gallium/drivers/nv40/nv40_vertprog.c
+++ b/src/gallium/drivers/nv40/nv40_vertprog.c
@@ -558,7 +558,7 @@ nv40_vertprog_prepare(struct nv40_vpc *vpc)
 	return TRUE;
 }
 
-void
+static void
 nv40_vertprog_translate(struct nv40_context *nv40,
 			struct nv40_vertex_program *vp)
 {
@@ -631,24 +631,32 @@ out_err:
 	free(vpc);
 }
 
-void
-nv40_vertprog_bind(struct nv40_context *nv40, struct nv40_vertex_program *vp)
+static boolean
+nv40_vertprog_validate(struct nv40_context *nv40)
 { 
+	struct nv40_vertex_program *vp = nv40->pipe_state.vertprog;
+	struct pipe_buffer *constbuf =
+		nv40->pipe_state.constbuf[PIPE_SHADER_VERTEX];
 	struct nouveau_winsys *nvws = nv40->nvws;
 	struct pipe_winsys *ws = nv40->pipe.winsys;
 	boolean upload_code = FALSE, upload_data = FALSE;
 	int i;
 
 	/* Translate TGSI shader into hw bytecode */
+	if (vp->translated)
+		goto check_gpu_resources;
+
+	nv40_vertprog_translate(nv40, vp);
 	if (!vp->translated) {
-		nv40_vertprog_translate(nv40, vp);
-		if (!vp->translated)
-			assert(0);
+		nv40->fallback |= NV40_FALLBACK_TNL;
+		return FALSE;
 	}
 
+check_gpu_resources:
 	/* Allocate hw vtxprog exec slots */
 	if (!vp->exec) {
 		struct nouveau_resource *heap = nv40->hw->vp_exec_heap;
+		struct nouveau_stateobj *so;
 		uint vplen = vp->nr_insns;
 
 		if (nvws->res_alloc(heap, vplen, vp, &vp->exec)) {
@@ -663,6 +671,15 @@ nv40_vertprog_bind(struct nv40_context *nv40, struct nv40_vertex_program *vp)
 				assert(0);
 		}
 
+		so = so_new(5, 0);
+		so_method(so, nv40->hw->curie, NV40TCL_VP_START_FROM_ID, 1);
+		so_data  (so, vp->exec->start);
+		so_method(so, nv40->hw->curie, NV40TCL_VP_ATTRIB_EN, 2);
+		so_data  (so, vp->ir);
+		so_data  (so, vp->or);
+		so_ref(so, &vp->so);
+		so_ref(NULL, &so);
+
 		upload_code = TRUE;
 	}
 
@@ -725,8 +742,8 @@ nv40_vertprog_bind(struct nv40_context *nv40, struct nv40_vertex_program *vp)
 	if (vp->nr_consts) {
 		float *map = NULL;
 
-		if (nv40->vertprog.constant_buf) {
-			map = ws->buffer_map(ws, nv40->vertprog.constant_buf,
+		if (constbuf) {
+			map = ws->buffer_map(ws, constbuf,
 					     PIPE_BUFFER_USAGE_CPU_READ);
 		}
 
@@ -747,9 +764,8 @@ nv40_vertprog_bind(struct nv40_context *nv40, struct nv40_vertex_program *vp)
 			OUT_RINGp ((uint32_t *)vpd->value, 4);
 		}
 
-		if (map) {
-			ws->buffer_unmap(ws, nv40->vertprog.constant_buf);
-		}
+		if (constbuf)
+			ws->buffer_unmap(ws, constbuf);
 	}
 
 	/* Upload vtxprog */
@@ -770,13 +786,12 @@ nv40_vertprog_bind(struct nv40_context *nv40, struct nv40_vertex_program *vp)
 		}
 	}
 
-	BEGIN_RING(curie, NV40TCL_VP_START_FROM_ID, 1);
-	OUT_RING  (vp->exec->start);
-	BEGIN_RING(curie, NV40TCL_VP_ATTRIB_EN, 2);
-	OUT_RING  (vp->ir);
-	OUT_RING  (vp->or);
+	if (vp->so != nv40->state.vertprog) {
+		so_ref(vp->so, &nv40->state.vertprog);
+		return TRUE;
+	}
 
-	nv40->vertprog.active = vp;
+	return FALSE;
 }
 
 void
@@ -788,3 +803,11 @@ nv40_vertprog_destroy(struct nv40_context *nv40, struct nv40_vertex_program *vp)
 		free(vp->insns);
 }
 
+struct nv40_state_entry nv40_state_vertprog = {
+	.validate = nv40_vertprog_validate,
+	.dirty = {
+		.pipe = NV40_NEW_VERTPROG,
+		.hw = NV40_NEW_VERTPROG
+	}
+};
+