nv30: Emit fragment program using state objects

6 files changed, 56 insertions, 48 deletions

diff --git a/src/gallium/drivers/nv30/nv30_context.h b/src/gallium/drivers/nv30/nv30_context.h
index 5d7080a555..0c900c5598 100644
--- a/src/gallium/drivers/nv30/nv30_context.h
+++ b/src/gallium/drivers/nv30/nv30_context.h
@@ -125,6 +125,7 @@ struct nv30_context {
 	struct pipe_scissor_state scissor;
 	unsigned stipple[32];
 	struct nv30_vertex_program *vertprog;
+	struct nv30_fragment_program *fragprog;
 	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 {
 		unsigned delta;
 	} vb[16];
 
-	struct {
-		struct nv30_fragment_program *active;
-		struct nv30_fragment_program *current;
-	} fragprog;
-
 	struct pipe_vertex_buffer  vtxbuf[PIPE_MAX_ATTRIBS];
 	struct pipe_vertex_element vtxelt[PIPE_MAX_ATTRIBS];
 };
@@ -188,10 +184,6 @@ extern void nv30_vertprog_destroy(struct nv30_context *,
 				  struct nv30_vertex_program *);
 
 /* nv30_fragprog.c */
-extern void nv30_fragprog_translate(struct nv30_context *,
-				    struct nv30_fragment_program *);
-extern void nv30_fragprog_bind(struct nv30_context *,
-			       struct nv30_fragment_program *);
 extern void nv30_fragprog_destroy(struct nv30_context *,
 				  struct nv30_fragment_program *);
 
@@ -205,6 +197,7 @@ 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_fragprog;
 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;
diff --git a/src/gallium/drivers/nv30/nv30_fragprog.c b/src/gallium/drivers/nv30/nv30_fragprog.c
index b560455a02..06ed04cbfe 100644
--- a/src/gallium/drivers/nv30/nv30_fragprog.c
+++ b/src/gallium/drivers/nv30/nv30_fragprog.c
@@ -723,7 +723,7 @@ out_err:
 	return FALSE;
 }
 
-void
+static void
 nv30_fragprog_translate(struct nv30_context *nv30,
 			struct nv30_fragment_program *fp)
 {
@@ -817,23 +817,46 @@ nv30_fragprog_upload(struct nv30_context *nv30,
 	ws->buffer_unmap(ws, fp->buffer);
 }
 
-void
-nv30_fragprog_bind(struct nv30_context *nv30, struct nv30_fragment_program *fp)
+static boolean
+nv30_fragprog_validate(struct nv30_context *nv30)
 {
+	struct nv30_fragment_program *fp = nv30->fragprog;
 	struct pipe_buffer *constbuf =
 		nv30->constbuf[PIPE_SHADER_FRAGMENT];
 	struct pipe_winsys *ws = nv30->pipe.winsys;
+	struct nouveau_stateobj *so;
+	boolean new_consts = FALSE;
 	int i;
 
+	if (fp->translated)
+		goto update_constants;
+
+	/*nv30->fallback_swrast &= ~NV30_NEW_FRAGPROG;*/
+	nv30_fragprog_translate(nv30, fp);
 	if (!fp->translated) {
-		nv30_fragprog_translate(nv30, fp);
-		if (!fp->translated)
-			assert(0);
+		/*nv30->fallback_swrast |= NV30_NEW_FRAGPROG;*/
+		return FALSE;
 	}
 
+	fp->buffer = ws->buffer_create(ws, 0x100, 0, fp->insn_len * 4);
+	nv30_fragprog_upload(nv30, fp);
+
+	so = so_new(4, 1);
+	so_method(so, nv30->screen->rankine, NV34TCL_FP_ACTIVE_PROGRAM, 1);
+	so_reloc (so, fp->buffer, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_GART |
+		  NOUVEAU_BO_RD | NOUVEAU_BO_LOW | NOUVEAU_BO_OR,
+		  NV34TCL_FP_ACTIVE_PROGRAM_DMA0, NV34TCL_FP_ACTIVE_PROGRAM_DMA1);
+	so_method(so, nv30->screen->rankine, NV34TCL_FP_CONTROL, 1);
+	so_data  (so, fp->fp_control);
+	/* FIXME: Add these, and you'll have big slowdown */
+	/*so_method(so, nv30->screen->rankine, NV34TCL_FP_REG_CONTROL, 1);
+	so_data  (so, fp->fp_control);*/
+	so_ref(so, &fp->so);
+
+update_constants:
 	if (fp->nr_consts) {
 		float *map;
-
+		
 		map = ws->buffer_map(ws, constbuf, PIPE_BUFFER_USAGE_CPU_READ);
 		for (i = 0; i < fp->nr_consts; i++) {
 			struct nv30_fragment_program_data *fpd = &fp->consts[i];
@@ -843,25 +866,20 @@ nv30_fragprog_bind(struct nv30_context *nv30, struct nv30_fragment_program *fp)
 			if (!memcmp(p, cb, 4 * sizeof(float)))
 				continue;
 			memcpy(p, cb, 4 * sizeof(float));
-			fp->on_hw = 0;
+			new_consts = TRUE;
 		}
 		ws->buffer_unmap(ws, constbuf);
-	}
 
-	if (!fp->on_hw) {
-		if (!fp->buffer)
-			fp->buffer = ws->buffer_create(ws, 0x100, 0,
-						       fp->insn_len * 4);
-		nv30_fragprog_upload(nv30, fp);
-		fp->on_hw = TRUE;
+		if (new_consts)
+			nv30_fragprog_upload(nv30, fp);
 	}
 
-	BEGIN_RING(rankine, NV34TCL_FP_CONTROL, 1);
-	OUT_RING  (fp->fp_control);
-	BEGIN_RING(rankine, NV34TCL_FP_REG_CONTROL, 1);
-	OUT_RING  (fp->fp_reg_control);
+	if (new_consts || fp->so != nv30->state.hw[NV30_STATE_FRAGPROG]) {
+		so_ref(fp->so, &nv30->state.hw[NV30_STATE_FRAGPROG]);
+		return TRUE;
+	}
 
-	nv30->fragprog.active = fp;
+	return FALSE;
 }
 
 void
@@ -872,3 +890,10 @@ nv30_fragprog_destroy(struct nv30_context *nv30,
 		FREE(fp->insn);
 }
 
+struct nv30_state_entry nv30_state_fragprog = {
+	.validate = nv30_fragprog_validate,
+	.dirty = {
+		.pipe = NV30_NEW_FRAGPROG,
+		.hw = NV30_STATE_FRAGPROG
+	}
+};
diff --git a/src/gallium/drivers/nv30/nv30_fragtex.c b/src/gallium/drivers/nv30/nv30_fragtex.c
index abe77b51df..8c5e88ea1d 100644
--- a/src/gallium/drivers/nv30/nv30_fragtex.c
+++ b/src/gallium/drivers/nv30/nv30_fragtex.c
@@ -143,7 +143,7 @@ nv30_fragtex_build(struct nv30_context *nv30, int unit)
 static boolean
 nv30_fragtex_validate(struct nv30_context *nv30)
 {
-	struct nv30_fragment_program *fp = nv30->fragprog.current;
+	struct nv30_fragment_program *fp = nv30->fragprog;
 	struct nv30_state *state = &nv30->state;
 	struct nouveau_stateobj *so;
 	unsigned samplers, unit;
diff --git a/src/gallium/drivers/nv30/nv30_state.c b/src/gallium/drivers/nv30/nv30_state.c
index 92695ce23c..8dc16d361d 100644
--- a/src/gallium/drivers/nv30/nv30_state.c
+++ b/src/gallium/drivers/nv30/nv30_state.c
@@ -539,6 +539,8 @@ nv30_fp_state_create(struct pipe_context *pipe,
 	fp = CALLOC(1, sizeof(struct nv30_fragment_program));
 	fp->pipe = *cso;
 
+	tgsi_scan_shader(fp->pipe.tokens, &fp->info);
+
 	return (void *)fp;
 }
 
@@ -546,13 +548,8 @@ static void
 nv30_fp_state_bind(struct pipe_context *pipe, void *hwcso)
 {
 	struct nv30_context *nv30 = nv30_context(pipe);
-	struct nv30_fragment_program *fp = hwcso;
-
-	if (!hwcso) {
-		return;
-	}
 
-	nv30->fragprog.current = fp;
+	nv30->fragprog = hwcso;
 	nv30->dirty |= NV30_NEW_FRAGPROG;
 }
 
diff --git a/src/gallium/drivers/nv30/nv30_state.h b/src/gallium/drivers/nv30/nv30_state.h
index a897bc5068..20c5176160 100644
--- a/src/gallium/drivers/nv30/nv30_state.h
+++ b/src/gallium/drivers/nv30/nv30_state.h
@@ -2,6 +2,7 @@
 #define __NV30_STATE_H__
 
 #include "pipe/p_state.h"
+#include "tgsi/util/tgsi_scan.h"
 
 struct nv30_sampler_state {
 	uint32_t fmt;
@@ -50,6 +51,7 @@ struct nv30_fragment_program_data {
 
 struct nv30_fragment_program {
 	struct pipe_shader_state pipe;
+	struct tgsi_shader_info info;
 
 	boolean translated;
 	boolean on_hw;
@@ -65,6 +67,7 @@ struct nv30_fragment_program {
 
 	uint32_t fp_control;
 	uint32_t fp_reg_control;
+	struct nouveau_stateobj *so;
 };
 
 struct nv30_miptree {
diff --git a/src/gallium/drivers/nv30/nv30_state_emit.c b/src/gallium/drivers/nv30/nv30_state_emit.c
index 4ab62ddc01..c4ccc9422b 100644
--- a/src/gallium/drivers/nv30/nv30_state_emit.c
+++ b/src/gallium/drivers/nv30/nv30_state_emit.c
@@ -6,6 +6,7 @@ static struct nv30_state_entry *render_states[] = {
 	&nv30_state_rasterizer,
 	&nv30_state_scissor,
 	&nv30_state_stipple,
+	&nv30_state_fragprog,
 	&nv30_state_fragtex,
 	&nv30_state_vertprog,
 	&nv30_state_blend,
@@ -61,11 +62,6 @@ nv30_emit_hw_state(struct nv30_context *nv30)
 		screen->cur_pctx = nv30->pctx_id;
 	}
 
-	if (nv30->dirty & NV30_NEW_FRAGPROG) {
-		nv30_fragprog_bind(nv30, nv30->fragprog.current);
-		/*XXX: clear NV30_NEW_FRAGPROG if no new program uploaded */
-	}
-
 	for (i = 0, states = state->dirty; states; i++) {
 		if (!(states & (1ULL << i)))
 			continue;
@@ -83,13 +79,7 @@ nv30_emit_hw_state(struct nv30_context *nv30)
 				      state->hw[NV30_STATE_FRAGTEX0+i]);
 		samplers &= ~(1ULL << i);
 	}
-
-	/* Fragment program */
-	BEGIN_RING(rankine, NV34TCL_FP_ACTIVE_PROGRAM, 1);
-	OUT_RELOC (nv30->fragprog.active->buffer, 0, NOUVEAU_BO_VRAM |
-	           NOUVEAU_BO_GART | NOUVEAU_BO_RD | NOUVEAU_BO_LOW |
-		   NOUVEAU_BO_OR, NV34TCL_FP_ACTIVE_PROGRAM_DMA0,
-		   NV34TCL_FP_ACTIVE_PROGRAM_DMA1);
+	so_emit_reloc_markers(nv30->nvws, state->hw[NV30_STATE_FRAGPROG]);
 }
 
 boolean