1 files changed, 84 insertions, 53 deletions

diff --git a/src/gallium/drivers/nv40/nv40_state_emit.c b/src/gallium/drivers/nv40/nv40_state_emit.c
index a10c995548..9f268640e0 100644
--- a/src/gallium/drivers/nv40/nv40_state_emit.c
+++ b/src/gallium/drivers/nv40/nv40_state_emit.c
@@ -1,77 +1,108 @@
 #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 struct nv40_state_entry *render_states[] = {
+	&nv40_state_framebuffer,
+	&nv40_state_rasterizer,
+	&nv40_state_clip,
+	&nv40_state_scissor,
+	&nv40_state_stipple,
+	&nv40_state_fragprog,
+	&nv40_state_fragtex,
+	&nv40_state_vertprog,
+	&nv40_state_blend,
+	&nv40_state_blend_colour,
+	&nv40_state_zsa,
+	&nv40_state_viewport,
+	&nv40_state_vbo,
+	NULL
+};
+
 static void
-nv40_state_emit_dummy_relocs(struct nv40_context *nv40)
+nv40_state_validate(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->fragprog.active->so);
-}
+	struct nv40_state_entry **states = render_states;
+	unsigned last_fallback;
 
-void
-nv40_emit_hw_state(struct nv40_context *nv40)
-{
-	if (nv40->dirty & NV40_NEW_FB)
-		so_emit(nv40->nvws, nv40->so_framebuffer);
+	last_fallback = nv40->fallback;
+	nv40->fallback = 0;
 
-	if (nv40->dirty & NV40_NEW_BLEND)
-		so_emit(nv40->nvws, nv40->so_blend);
+	while (*states) {
+		struct nv40_state_entry *e = *states;
 
-	if (nv40->dirty & NV40_NEW_RAST)
-		so_emit(nv40->nvws, nv40->so_rast);
+		if (nv40->dirty & e->dirty.pipe) {
+			if (e->validate(nv40))
+				nv40->state.dirty |= (1ULL << e->dirty.hw);
+		}
 
-	if (nv40->dirty & NV40_NEW_ZSA)
-		so_emit(nv40->nvws, nv40->so_zsa);
+		states++;
+	}
+	nv40->dirty = 0;
 
-	if (nv40->dirty & NV40_NEW_BCOL)
-		so_emit(nv40->nvws, nv40->so_bcol);
+	if (nv40->fallback & NV40_FALLBACK_TNL &&
+	    !(last_fallback & NV40_FALLBACK_TNL)) {
+		NOUVEAU_ERR("XXX: hwtnl->swtnl\n");
+	} else
+	if (last_fallback & NV40_FALLBACK_TNL &&
+	    !(nv40->fallback & NV40_FALLBACK_TNL)) {
+		NOUVEAU_ERR("XXX: swtnl->hwtnl\n");
+	}
 
-	if (nv40->dirty & NV40_NEW_SCISSOR)
-		so_emit(nv40->nvws, nv40->so_scissor);
+	if (nv40->fallback & NV40_FALLBACK_RAST &&
+	    !(last_fallback & NV40_FALLBACK_RAST)) {
+		NOUVEAU_ERR("XXX: hwrast->swrast\n");
+	} else
+	if (last_fallback & NV40_FALLBACK_RAST &&
+	    !(nv40->fallback & NV40_FALLBACK_RAST)) {
+		NOUVEAU_ERR("XXX: swrast->hwrast\n");
+	}
+}
 
-	if (nv40->dirty & NV40_NEW_VIEWPORT)
-		so_emit(nv40->nvws, nv40->so_viewport);
+static void
+nv40_state_emit(struct nv40_context *nv40)
+{
+	struct nv40_state *state = &nv40->state;
+	struct nv40_screen *screen = nv40->screen;
+	unsigned i, samplers;
 
-	if (nv40->dirty & NV40_NEW_STIPPLE)
-		so_emit(nv40->nvws, nv40->so_stipple);
+	if (nv40->pctx_id != screen->cur_pctx) {
+		for (i = 0; i < NV40_STATE_MAX; i++) {
+			if (state->hw[i] && screen->state[i] != state->hw[i])
+				state->dirty |= (1ULL << i);
+		}
 
-	if (nv40->dirty & NV40_NEW_FRAGPROG) {
-		nv40_fragprog_bind(nv40, nv40->fragprog.current);
-		/*XXX: clear NV40_NEW_FRAGPROG if no new program uploaded */
+		screen->cur_pctx = nv40->pctx_id;
 	}
 
-	if (nv40->dirty_samplers || (nv40->dirty & NV40_NEW_FRAGPROG)) {
-		nv40_fragtex_bind(nv40);
+	while (state->dirty) {
+		unsigned idx = ffsll(state->dirty) - 1;
 
-		BEGIN_RING(curie, NV40TCL_TEX_CACHE_CTL, 1);
-		OUT_RING  (2);
-		BEGIN_RING(curie, NV40TCL_TEX_CACHE_CTL, 1);
-		OUT_RING  (1);
-		nv40->dirty &= ~NV40_NEW_FRAGPROG;
+		so_ref (state->hw[idx], &nv40->screen->state[idx]);
+		so_emit(nv40->nvws, nv40->screen->state[idx]);
+		state->dirty &= ~(1ULL << idx);
 	}
 
-	if (nv40->dirty & NV40_NEW_VERTPROG) {
-		nv40_vertprog_bind(nv40, nv40->vertprog.current);
-		nv40->dirty &= ~NV40_NEW_VERTPROG;
+	so_emit_reloc_markers(nv40->nvws, state->hw[NV40_STATE_FB]);
+	for (i = 0, samplers = state->fp_samplers; i < 16 && samplers; i++) {
+		if (!(samplers & (1 << i)))
+			continue;
+		so_emit_reloc_markers(nv40->nvws,
+				      state->hw[NV40_STATE_FRAGTEX0+i]);
+		samplers &= ~(1ULL << i);
 	}
+	so_emit_reloc_markers(nv40->nvws, state->hw[NV40_STATE_FRAGPROG]);
+	so_emit_reloc_markers(nv40->nvws, state->hw[NV40_STATE_VTXBUF]);
+}
 
-	nv40->dirty_samplers = 0;
-	nv40->dirty = 0;
+void
+nv40_emit_hw_state(struct nv40_context *nv40)
+{
+	nv40_state_validate(nv40);
+	nv40_state_emit(nv40);
 
-	nv40_state_emit_dummy_relocs(nv40);
+	BEGIN_RING(curie, NV40TCL_TEX_CACHE_CTL, 1);
+	OUT_RING  (2);
+	BEGIN_RING(curie, NV40TCL_TEX_CACHE_CTL, 1);
+	OUT_RING  (1);
 }