nv40: keep track of generated context state vs current channel state

6 files changed, 53 insertions, 68 deletions

diff --git a/src/gallium/drivers/nv40/nv40_context.h b/src/gallium/drivers/nv40/nv40_context.h
index 28a0274d77..2dd137ead0 100644
--- a/src/gallium/drivers/nv40/nv40_context.h
+++ b/src/gallium/drivers/nv40/nv40_context.h
@@ -22,6 +22,15 @@
 #define NOUVEAU_MSG(fmt, args...) \
 	fprintf(stderr, "nouveau: "fmt, ##args);
 
+enum nv40_state_index {
+	NV40_STATE_CLIP = 1ULL,
+	NV40_STATE_SCISSOR = 2ULL,
+	NV40_STATE_STIPPLE = 3ULL,
+	NV40_STATE_FRAGPROG = 4ULL,
+	NV40_STATE_VERTPROG = 5ULL,
+	NV40_STATE_MAX = 6ULL
+};
+
 #define NV40_NEW_BLEND		(1 <<  0)
 #define NV40_NEW_RAST		(1 <<  1)
 #define NV40_NEW_ZSA		(1 <<  2)
@@ -56,6 +65,9 @@ struct nv40_channel_context {
 	/* Vtxprog resources */
 	struct nouveau_resource *vp_exec_heap;
 	struct nouveau_resource *vp_data_heap;
+
+	/* Current 3D state of channel */
+	struct nouveau_stateobj *state[NV40_STATE_MAX];
 };
 
 struct nv40_rasterizer_state {
@@ -64,18 +76,10 @@ struct nv40_rasterizer_state {
 };
 
 struct nv40_state {
-	struct {
-		unsigned enabled;
-		struct nouveau_stateobj *so;
-	} scissor;
-
-	struct {
-		unsigned enabled;
-		struct nouveau_stateobj *so;
-	} stipple;
+	unsigned scissor_enabled;
+	unsigned stipple_enabled;
 
-	struct nouveau_stateobj *fragprog;
-	struct nouveau_stateobj *vertprog;
+	struct nouveau_stateobj *hw[NV40_STATE_MAX];
 };
 
 struct nv40_context {
diff --git a/src/gallium/drivers/nv40/nv40_fragprog.c b/src/gallium/drivers/nv40/nv40_fragprog.c
index 77ac8ab2c6..db2613ef8b 100644
--- a/src/gallium/drivers/nv40/nv40_fragprog.c
+++ b/src/gallium/drivers/nv40/nv40_fragprog.c
@@ -841,8 +841,8 @@ update_constants:
 			nv40_fragprog_upload(nv40, fp);
 	}
 
-	if (fp->so != nv40->state.fragprog) {
-		so_ref(fp->so, &nv40->state.fragprog);
+	if (fp->so != nv40->state.hw[NV40_STATE_FRAGPROG]) {
+		so_ref(fp->so, &nv40->state.hw[NV40_STATE_FRAGPROG]);
 		return TRUE;
 	}
 
@@ -861,7 +861,7 @@ struct nv40_state_entry nv40_state_fragprog = {
 	.validate = nv40_fragprog_validate,
 	.dirty = {
 		.pipe = NV40_NEW_FRAGPROG,
-		.hw = NV40_NEW_FRAGPROG
+		.hw = NV40_STATE_FRAGPROG
 	}
 };
 
diff --git a/src/gallium/drivers/nv40/nv40_state_emit.c b/src/gallium/drivers/nv40/nv40_state_emit.c
index e8230111bb..58beb72389 100644
--- a/src/gallium/drivers/nv40/nv40_state_emit.c
+++ b/src/gallium/drivers/nv40/nv40_state_emit.c
@@ -1,27 +1,6 @@
 #include "nv40_context.h"
 #include "nv40_state.h"
 
-/* Emit relocs for every referenced buffer.
- *
- * This is to ensure the bufmgr has an accurate idea of how
- * the buffer is used.  These relocs appear in the push buffer as
- * NOPs, and will only be turned into state changes if a buffer
- * actually moves.
- */
-static void
-nv40_state_emit_dummy_relocs(struct nv40_context *nv40)
-{
-	unsigned i;	
-	
-	so_emit_reloc_markers(nv40->nvws, nv40->so_framebuffer);
-	for (i = 0; i < 16; i++) {
-		if (!(nv40->fp_samplers & (1 << i)))
-			continue;
-		so_emit_reloc_markers(nv40->nvws, nv40->so_fragtex[i]);
-	}
-	so_emit_reloc_markers(nv40->nvws, nv40->state.fragprog);
-}
-
 static struct nv40_state_entry *render_states[] = {
 	&nv40_state_clip,
 	&nv40_state_scissor,
@@ -45,7 +24,7 @@ nv40_state_validate(struct nv40_context *nv40)
 
 		if (nv40->dirty & e->dirty.pipe) {
 			if (e->validate(nv40))
-				nv40->hw_dirty |= e->dirty.hw;
+				nv40->hw_dirty |= (1 << e->dirty.hw);
 		}
 
 		states++;
@@ -70,6 +49,28 @@ nv40_state_validate(struct nv40_context *nv40)
 	}
 }
 
+static void
+nv40_state_emit(struct nv40_context *nv40)
+{
+	unsigned i;
+
+	while (nv40->hw_dirty) {
+		unsigned idx = ffs(nv40->hw_dirty) - 1;
+		nv40->hw_dirty &= ~(1 << idx);
+
+		so_ref (nv40->state.hw[idx], &nv40->hw->state[idx]);
+		so_emit(nv40->nvws, nv40->hw->state[idx]);
+	}
+
+	so_emit_reloc_markers(nv40->nvws, nv40->so_framebuffer);
+	for (i = 0; i < 16; i++) {
+		if (!(nv40->fp_samplers & (1 << i)))
+			continue;
+		so_emit_reloc_markers(nv40->nvws, nv40->so_fragtex[i]);
+	}
+	so_emit_reloc_markers(nv40->nvws, nv40->state.hw[NV40_STATE_FRAGPROG]);
+}
+
 void
 nv40_emit_hw_state(struct nv40_context *nv40)
 {
@@ -90,24 +91,9 @@ nv40_emit_hw_state(struct nv40_context *nv40)
 	if (nv40->dirty & NV40_NEW_BCOL)
 		so_emit(nv40->nvws, nv40->so_bcol);
 
-	if (nv40->hw_dirty & NV40_NEW_SCISSOR) {
-		so_emit(nv40->nvws, nv40->state.scissor.so);
-		nv40->hw_dirty &= ~NV40_NEW_SCISSOR;
-	}
-
 	if (nv40->dirty & NV40_NEW_VIEWPORT)
 		so_emit(nv40->nvws, nv40->so_viewport);
 
-	if (nv40->hw_dirty & NV40_NEW_STIPPLE) {
-		so_emit(nv40->nvws, nv40->state.stipple.so);
-		nv40->hw_dirty &= ~NV40_NEW_STIPPLE;
-	}
-
-	if (nv40->hw_dirty & NV40_NEW_FRAGPROG) {
-		so_emit(nv40->nvws, nv40->state.fragprog);
-		nv40->hw_dirty &= ~NV40_NEW_FRAGPROG;
-	}
-
 	if (nv40->dirty_samplers || (nv40->dirty & NV40_NEW_FRAGPROG)) {
 		nv40_fragtex_bind(nv40);
 
@@ -118,14 +104,9 @@ nv40_emit_hw_state(struct nv40_context *nv40)
 		nv40->dirty &= ~NV40_NEW_FRAGPROG;
 	}
 
-	if (nv40->hw_dirty & NV40_NEW_VERTPROG) {
-		so_emit(nv40->nvws, nv40->state.vertprog);
-		nv40->hw_dirty &= ~NV40_NEW_VERTPROG;
-	}
+	nv40_state_emit(nv40);
 
 	nv40->dirty_samplers = 0;
 	nv40->dirty = 0;
-
-	nv40_state_emit_dummy_relocs(nv40);
 }
 
diff --git a/src/gallium/drivers/nv40/nv40_state_scissor.c b/src/gallium/drivers/nv40/nv40_state_scissor.c
index 556b820e58..dc7b6d3a9d 100644
--- a/src/gallium/drivers/nv40/nv40_state_scissor.c
+++ b/src/gallium/drivers/nv40/nv40_state_scissor.c
@@ -7,8 +7,8 @@ nv40_state_scissor_validate(struct nv40_context *nv40)
 	struct pipe_scissor_state *s = &nv40->pipe_state.scissor;
 	struct nouveau_stateobj *so;
 
-	if (nv40->state.scissor.so &&
-	    (rast->scissor == 0 && nv40->state.scissor.enabled == 0))
+	if (nv40->state.hw[NV40_STATE_SCISSOR] &&
+	    (rast->scissor == 0 && nv40->state.scissor_enabled == 0))
 		return FALSE;
 
 	so = so_new(3, 0);
@@ -21,7 +21,7 @@ nv40_state_scissor_validate(struct nv40_context *nv40)
 		so_data  (so, 4096 << 16);
 	}
 
-	so_ref(so, &nv40->state.scissor.so);
+	so_ref(so, &nv40->state.hw[NV40_STATE_SCISSOR]);
 	so_ref(NULL, &so);
 	return TRUE;
 }
@@ -30,6 +30,6 @@ struct nv40_state_entry nv40_state_scissor = {
 	.validate = nv40_state_scissor_validate,
 	.dirty = {
 		.pipe = NV40_NEW_SCISSOR | NV40_NEW_RAST,
-		.hw = NV40_NEW_SCISSOR
+		.hw = NV40_STATE_SCISSOR
 	}
 };
diff --git a/src/gallium/drivers/nv40/nv40_state_stipple.c b/src/gallium/drivers/nv40/nv40_state_stipple.c
index 52462a0b50..1b0b194432 100644
--- a/src/gallium/drivers/nv40/nv40_state_stipple.c
+++ b/src/gallium/drivers/nv40/nv40_state_stipple.c
@@ -7,8 +7,8 @@ nv40_state_stipple_validate(struct nv40_context *nv40)
 	struct nouveau_grobj *curie = nv40->hw->curie;
 	struct nouveau_stateobj *so;
 
-	if (nv40->state.stipple.so && (rast->poly_stipple_enable == 0 &&
-				       nv40->state.stipple.enabled == 0))
+	if (nv40->state.hw[NV40_STATE_STIPPLE] &&
+	   (rast->poly_stipple_enable == 0 && nv40->state.stipple_enabled == 0))
 		return FALSE;
 
 	if (rast->poly_stipple_enable) {
@@ -26,7 +26,7 @@ nv40_state_stipple_validate(struct nv40_context *nv40)
 		so_data  (so, 0);
 	}
 
-	so_ref(so, &nv40->state.stipple.so);
+	so_ref(so, &nv40->state.hw[NV40_STATE_STIPPLE]);
 	so_ref(NULL, &so);
 	return TRUE;
 }
@@ -35,6 +35,6 @@ struct nv40_state_entry nv40_state_stipple = {
 	.validate = nv40_state_stipple_validate,
 	.dirty = {
 		.pipe = NV40_NEW_STIPPLE | NV40_NEW_RAST,
-		.hw = NV40_NEW_STIPPLE
+		.hw = NV40_STATE_STIPPLE,
 	}
 };
diff --git a/src/gallium/drivers/nv40/nv40_vertprog.c b/src/gallium/drivers/nv40/nv40_vertprog.c
index 4a15e51eb3..8a2d233697 100644
--- a/src/gallium/drivers/nv40/nv40_vertprog.c
+++ b/src/gallium/drivers/nv40/nv40_vertprog.c
@@ -786,8 +786,8 @@ check_gpu_resources:
 		}
 	}
 
-	if (vp->so != nv40->state.vertprog) {
-		so_ref(vp->so, &nv40->state.vertprog);
+	if (vp->so != nv40->state.hw[NV40_STATE_VERTPROG]) {
+		so_ref(vp->so, &nv40->state.hw[NV40_STATE_VERTPROG]);
 		return TRUE;
 	}
 
@@ -807,7 +807,7 @@ struct nv40_state_entry nv40_state_vertprog = {
 	.validate = nv40_vertprog_validate,
 	.dirty = {
 		.pipe = NV40_NEW_VERTPROG,
-		.hw = NV40_NEW_VERTPROG
+		.hw = NV40_STATE_VERTPROG,
 	}
 };