nv40: delay all state emit until before draw

5 files changed, 66 insertions, 14 deletions

diff --git a/src/mesa/pipe/nouveau/nouveau_stateobj.h b/src/mesa/pipe/nouveau/nouveau_stateobj.h
index 8dfc0e9e9a..58167a24de 100644
--- a/src/mesa/pipe/nouveau/nouveau_stateobj.h
+++ b/src/mesa/pipe/nouveau/nouveau_stateobj.h
@@ -30,7 +30,7 @@ so_new(unsigned push, unsigned reloc)
 	struct nouveau_stateobj *so;
 
 	so = malloc(sizeof(struct nouveau_stateobj));
-	so->refcount = 0;
+	so->refcount = 1;
 	so->push = malloc(sizeof(unsigned) * push);
 	so->reloc = malloc(sizeof(struct nouveau_stateobj_reloc) * reloc);
 
diff --git a/src/mesa/pipe/nv40/nv40_context.h b/src/mesa/pipe/nv40/nv40_context.h
index 1a31f00ad6..4aa34847e8 100644
--- a/src/mesa/pipe/nv40/nv40_context.h
+++ b/src/mesa/pipe/nv40/nv40_context.h
@@ -22,9 +22,18 @@
 #define NOUVEAU_MSG(fmt, args...) \
 	fprintf(stderr, "nouveau: "fmt, ##args);
 
-#define NV40_NEW_VERTPROG	(1 << 1)
-#define NV40_NEW_FRAGPROG	(1 << 2)
-#define NV40_NEW_ARRAYS		(1 << 3)
+#define NV40_NEW_BLEND		(1 <<  0)
+#define NV40_NEW_RAST		(1 <<  1)
+#define NV40_NEW_ZSA		(1 <<  2)
+#define NV40_NEW_SAMPLER	(1 <<  3)
+#define NV40_NEW_FB		(1 <<  4)
+#define NV40_NEW_STIPPLE	(1 <<  5)
+#define NV40_NEW_SCISSOR	(1 <<  6)
+#define NV40_NEW_VIEWPORT	(1 <<  7)
+#define NV40_NEW_BCOL		(1 <<  8)
+#define NV40_NEW_VERTPROG	(1 <<  9)
+#define NV40_NEW_FRAGPROG	(1 << 10)
+#define NV40_NEW_ARRAYS		(1 << 11)
 
 struct nv40_context {
 	struct pipe_context pipe;
@@ -51,6 +60,13 @@ struct nv40_context {
 	struct nouveau_stateobj *so_framebuffer;
 	struct nouveau_stateobj *so_fragtex[16];
 	struct nouveau_stateobj *so_vtxbuf;
+	struct nouveau_stateobj *so_blend;
+	struct nouveau_stateobj *so_rast;
+	struct nouveau_stateobj *so_zsa;
+	struct nouveau_stateobj *so_bcol;
+	struct nouveau_stateobj *so_scissor;
+	struct nouveau_stateobj *so_viewport;
+	struct nouveau_stateobj *so_stipple;
 
 	struct {
 		struct nouveau_resource *exec_heap;
diff --git a/src/mesa/pipe/nv40/nv40_state.c b/src/mesa/pipe/nv40/nv40_state.c
index 125134afdc..ab53b03cb0 100644
--- a/src/mesa/pipe/nv40/nv40_state.c
+++ b/src/mesa/pipe/nv40/nv40_state.c
@@ -53,7 +53,8 @@ nv40_blend_state_bind(struct pipe_context *pipe, void *hwcso)
 {
 	struct nv40_context *nv40 = nv40_context(pipe);
 
-	so_emit(nv40->nvws, hwcso);
+	so_ref(hwcso, &nv40->so_blend);
+	nv40->dirty |= NV40_NEW_BLEND;
 }
 
 static void
@@ -354,7 +355,8 @@ nv40_rasterizer_state_bind(struct pipe_context *pipe, void *hwcso)
 {
 	struct nv40_context *nv40 = nv40_context(pipe);
 
-	so_emit(nv40->nvws, hwcso);
+	so_ref(hwcso, &nv40->so_rast);
+	nv40->dirty |= NV40_NEW_RAST;
 }
 
 static void
@@ -420,7 +422,8 @@ nv40_depth_stencil_alpha_state_bind(struct pipe_context *pipe, void *hwcso)
 {
 	struct nv40_context *nv40 = nv40_context(pipe);
 
-	so_emit(nv40->nvws, hwcso);
+	so_ref(hwcso, &nv40->so_zsa);
+	nv40->dirty |= NV40_NEW_ZSA;
 }
 
 static void
@@ -508,8 +511,9 @@ nv40_set_blend_color(struct pipe_context *pipe,
 		       (float_to_ubyte(bcol->color[1]) <<  8) |
 		       (float_to_ubyte(bcol->color[2]) <<  0)));
 
-	so_emit(nv40->nvws, so);
+	so_ref(so, &nv40->so_bcol);
 	so_ref(NULL, &so);
+	nv40->dirty |= NV40_NEW_BCOL;
 }
 
 static void
@@ -677,8 +681,9 @@ nv40_set_framebuffer_state(struct pipe_context *pipe,
 	so_data  (so, ((w - 1) << 16) | 0);
 	so_data  (so, ((h - 1) << 16) | 0);
 
-	so_emit(nv40->nvws, so);
-	so_ref (so, &nv40->so_framebuffer);
+	so_ref(so, &nv40->so_framebuffer);
+	so_ref(NULL, &so);
+	nv40->dirty |= NV40_NEW_FB;
 }
 
 static void
@@ -693,8 +698,9 @@ nv40_set_polygon_stipple(struct pipe_context *pipe,
 	for (i = 0; i < 32; i++)
 		so_data(so, stipple->stipple[i]);
 
-	so_emit(nv40->nvws, so);
+	so_ref(so, &nv40->so_stipple);
 	so_ref(NULL, &so);
+	nv40->dirty |= NV40_NEW_STIPPLE;
 }
 
 static void
@@ -708,8 +714,9 @@ nv40_set_scissor_state(struct pipe_context *pipe,
 	so_data  (so, ((s->maxx - s->minx) << 16) | s->minx);
 	so_data  (so, ((s->maxy - s->miny) << 16) | s->miny);
 
-	so_emit(nv40->nvws, so);
+	so_ref(so, &nv40->so_scissor);
 	so_ref(NULL, &so);
+	nv40->dirty |= NV40_NEW_SCISSOR;
 }
 
 static void
@@ -729,8 +736,9 @@ nv40_set_viewport_state(struct pipe_context *pipe,
 	so_data  (so, fui(vpt->scale[2]));
 	so_data  (so, fui(vpt->scale[3]));
 
-	so_emit(nv40->nvws, so);
+	so_ref(so, &nv40->so_viewport);
 	so_ref(NULL, &so);
+	nv40->dirty |= NV40_NEW_VIEWPORT;
 }
 
 static void
diff --git a/src/mesa/pipe/nv40/nv40_state_emit.c b/src/mesa/pipe/nv40/nv40_state_emit.c
index 3a22cd4bd5..a10c995548 100644
--- a/src/mesa/pipe/nv40/nv40_state_emit.c
+++ b/src/mesa/pipe/nv40/nv40_state_emit.c
@@ -25,6 +25,30 @@ nv40_state_emit_dummy_relocs(struct nv40_context *nv40)
 void
 nv40_emit_hw_state(struct nv40_context *nv40)
 {
+	if (nv40->dirty & NV40_NEW_FB)
+		so_emit(nv40->nvws, nv40->so_framebuffer);
+
+	if (nv40->dirty & NV40_NEW_BLEND)
+		so_emit(nv40->nvws, nv40->so_blend);
+
+	if (nv40->dirty & NV40_NEW_RAST)
+		so_emit(nv40->nvws, nv40->so_rast);
+
+	if (nv40->dirty & NV40_NEW_ZSA)
+		so_emit(nv40->nvws, nv40->so_zsa);
+
+	if (nv40->dirty & NV40_NEW_BCOL)
+		so_emit(nv40->nvws, nv40->so_bcol);
+
+	if (nv40->dirty & NV40_NEW_SCISSOR)
+		so_emit(nv40->nvws, nv40->so_scissor);
+
+	if (nv40->dirty & NV40_NEW_VIEWPORT)
+		so_emit(nv40->nvws, nv40->so_viewport);
+
+	if (nv40->dirty & NV40_NEW_STIPPLE)
+		so_emit(nv40->nvws, nv40->so_stipple);
+
 	if (nv40->dirty & NV40_NEW_FRAGPROG) {
 		nv40_fragprog_bind(nv40, nv40->fragprog.current);
 		/*XXX: clear NV40_NEW_FRAGPROG if no new program uploaded */
@@ -46,6 +70,7 @@ nv40_emit_hw_state(struct nv40_context *nv40)
 	}
 
 	nv40->dirty_samplers = 0;
+	nv40->dirty = 0;
 
 	nv40_state_emit_dummy_relocs(nv40);
 }
diff --git a/src/mesa/pipe/nv40/nv40_vbo.c b/src/mesa/pipe/nv40/nv40_vbo.c
index e2cb3fda8f..fd1d884193 100644
--- a/src/mesa/pipe/nv40/nv40_vbo.c
+++ b/src/mesa/pipe/nv40/nv40_vbo.c
@@ -158,6 +158,7 @@ nv40_vbo_arrays_update(struct nv40_context *nv40, struct pipe_buffer *ib,
 	so_emit(nv40->nvws, vtxfmt);
 	so_emit(nv40->nvws, vtxbuf);
 	so_ref (vtxbuf, &nv40->so_vtxbuf);
+	so_ref (NULL, &vtxbuf);
 	so_ref (NULL, &vtxfmt);
 }
 
@@ -165,8 +166,10 @@ static boolean
 nv40_vbo_validate_state(struct nv40_context *nv40,
 			struct pipe_buffer *ib, unsigned ib_format)
 {
+	unsigned vdn = nv40->dirty & NV40_NEW_ARRAYS;
+
 	nv40_emit_hw_state(nv40);
-	if (nv40->dirty & NV40_NEW_ARRAYS || ib) {
+	if (vdn || ib) {
 		nv40_vbo_arrays_update(nv40, ib, ib_format);
 		nv40->dirty &= ~NV40_NEW_ARRAYS;
 	}