nvfx: add rewritten swtnl support

The old swtnl code was broken by the new shader linkage support for
GLSL.

This is a rewrite of swtnl support, which should instead work properly,
be faster and more closer to the much more tested hardware pipeline.

8 files changed, 237 insertions, 262 deletions

diff --git a/src/gallium/drivers/nvfx/nvfx_context.h b/src/gallium/drivers/nvfx/nvfx_context.h
index b837437c58..369c216388 100644
--- a/src/gallium/drivers/nvfx/nvfx_context.h
+++ b/src/gallium/drivers/nvfx/nvfx_context.h
@@ -151,14 +151,6 @@ struct nvfx_context {
 
 	/* HW state derived from pipe states */
 	struct nvfx_state state;
-	struct {
-		struct nvfx_vertex_program *vertprog;
-
-		unsigned nr_attribs;
-		unsigned hw[PIPE_MAX_SHADER_INPUTS];
-		unsigned draw[PIPE_MAX_SHADER_INPUTS];
-		unsigned emit[PIPE_MAX_SHADER_INPUTS];
-	} swtnl;
 
 	enum {
 		HW, SWTNL, SWRAST
@@ -170,7 +162,7 @@ struct nvfx_context {
 	struct pipe_scissor_state scissor;
 	unsigned stipple[32];
 	struct pipe_clip_state clip;
-	struct nvfx_vertex_program *vertprog;
+	struct nvfx_pipe_vertex_program *vertprog;
 	struct nvfx_pipe_fragment_program *fragprog;
 	struct pipe_resource *constbuf[PIPE_SHADER_TYPES];
 	unsigned constbuf_nr[PIPE_SHADER_TYPES];
@@ -208,6 +200,7 @@ struct nvfx_context {
 	int hw_pointsprite_control;
 	int hw_vp_output;
 	struct nvfx_fragment_program* hw_fragprog;
+	struct nvfx_vertex_program* hw_vertprog;
 
 	unsigned relocs_needed;
 };
@@ -326,6 +319,7 @@ extern void nvfx_init_transfer_functions(struct pipe_context *pipe);
 
 /* nvfx_vbo.c */
 extern boolean nvfx_vbo_validate(struct nvfx_context *nvfx);
+extern void nvfx_vbo_swtnl_validate(struct nvfx_context *nvfx);
 extern void nvfx_vbo_relocate(struct nvfx_context *nvfx);
 extern void nvfx_idxbuf_validate(struct nvfx_context* nvfx);
 extern void nvfx_idxbuf_relocate(struct nvfx_context* nvfx);
diff --git a/src/gallium/drivers/nvfx/nvfx_draw.c b/src/gallium/drivers/nvfx/nvfx_draw.c
index 2601d5b8e2..4bf38a9c18 100644
--- a/src/gallium/drivers/nvfx/nvfx_draw.c
+++ b/src/gallium/drivers/nvfx/nvfx_draw.c
@@ -1,6 +1,5 @@
 #include "pipe/p_shader_tokens.h"
 #include "util/u_inlines.h"
-#include "tgsi/tgsi_ureg.h"
 
 #include "util/u_pack_color.h"
 
@@ -11,11 +10,6 @@
 #include "nvfx_context.h"
 #include "nvfx_resource.h"
 
-/* Simple, but crappy, swtnl path, hopefully we wont need to hit this very
- * often at all.  Uses "quadro style" vertex submission + a fixed vertex
- * layout to avoid the need to generate a vertex program or vtxfmt.
- */
-
 struct nvfx_render_stage {
 	struct draw_stage stage;
 	struct nvfx_context *nvfx;
@@ -28,58 +22,18 @@ nvfx_render_stage(struct draw_stage *stage)
 	return (struct nvfx_render_stage *)stage;
 }
 
-static INLINE void
-nvfx_render_vertex(struct nvfx_context *nvfx, const struct vertex_header *v)
+static void
+nvfx_render_flush(struct draw_stage *stage, unsigned flags)
 {
-	struct nvfx_screen *screen = nvfx->screen;
-	struct nouveau_channel *chan = screen->base.channel;
-	struct nouveau_grobj *eng3d = screen->eng3d;
-	unsigned i;
-
-	for (i = 0; i < nvfx->swtnl.nr_attribs; i++) {
-		unsigned idx = nvfx->swtnl.draw[i];
-		unsigned hw = nvfx->swtnl.hw[i];
+	struct nvfx_render_stage *rs = nvfx_render_stage(stage);
+	struct nvfx_context *nvfx = rs->nvfx;
+	struct nouveau_channel *chan = nvfx->screen->base.channel;
 
-		WAIT_RING(chan, 5);
-		switch (nvfx->swtnl.emit[i]) {
-		case EMIT_OMIT:
-			break;
-		case EMIT_1F:
-			nvfx_emit_vtx_attr(chan, hw, v->data[idx], 1);
-			break;
-		case EMIT_2F:
-			nvfx_emit_vtx_attr(chan, hw, v->data[idx], 2);
-			break;
-		case EMIT_3F:
-			nvfx_emit_vtx_attr(chan, hw, v->data[idx], 3);
-			break;
-		case EMIT_4F:
-			nvfx_emit_vtx_attr(chan, hw, v->data[idx], 4);
-			break;
-		case 0xff:
-			BEGIN_RING(chan, eng3d, NV34TCL_VTX_ATTR_4F_X(hw), 4);
-			OUT_RING  (chan, fui(v->data[idx][0] / v->data[idx][3]));
-			OUT_RING  (chan, fui(v->data[idx][1] / v->data[idx][3]));
-			OUT_RING  (chan, fui(v->data[idx][2] / v->data[idx][3]));
-			OUT_RING  (chan, fui(1.0f / v->data[idx][3]));
-			break;
-		case EMIT_4UB:
-			BEGIN_RING(chan, eng3d, NV34TCL_VTX_ATTR_4UB(hw), 1);
-			OUT_RING  (chan, pack_ub4(float_to_ubyte(v->data[idx][0]),
-					    float_to_ubyte(v->data[idx][1]),
-					    float_to_ubyte(v->data[idx][2]),
-					    float_to_ubyte(v->data[idx][3])));
-		case EMIT_4UB_BGRA:
-			BEGIN_RING(chan, eng3d, NV34TCL_VTX_ATTR_4UB(hw), 1);
-			OUT_RING  (chan, pack_ub4(float_to_ubyte(v->data[idx][2]),
-					    float_to_ubyte(v->data[idx][1]),
-					    float_to_ubyte(v->data[idx][0]),
-					    float_to_ubyte(v->data[idx][3])));
-			break;
-		default:
-			assert(0);
-			break;
-		}
+	if (rs->prim != NV34TCL_VERTEX_BEGIN_END_STOP) {
+		assert(AVAIL_RING(chan) >= 2);
+		OUT_RING(chan, RING_3D(NV34TCL_VERTEX_BEGIN_END, 1));
+		OUT_RING(chan, NV34TCL_VERTEX_BEGIN_END_STOP);
+		rs->prim = NV34TCL_VERTEX_BEGIN_END_STOP;
 	}
 }
 
@@ -92,42 +46,61 @@ nvfx_render_prim(struct draw_stage *stage, struct prim_header *prim,
 
 	struct nvfx_screen *screen = nvfx->screen;
 	struct nouveau_channel *chan = screen->base.channel;
-	struct nouveau_grobj *eng3d = screen->eng3d;
-	unsigned i;
+	boolean no_elements = nvfx->vertprog->draw_no_elements;
+	unsigned num_attribs = nvfx->vertprog->draw_elements;
 
-	/* Ensure there's room for 4xfloat32 + potentially 3 begin/end */
-	if (AVAIL_RING(chan) < ((count * 20) + 6)) {
-		if (rs->prim != NV34TCL_VERTEX_BEGIN_END_STOP) {
-			NOUVEAU_ERR("AIII, missed flush\n");
-			assert(0);
-		}
+	/* we need to account the flush as well here even if it is done afterthis
+	 * function
+	 */
+	if (AVAIL_RING(chan) < ((1 + count * num_attribs * 4) + 6 + 64)) {
+		nvfx_render_flush(stage, 0);
 		FIRE_RING(chan);
 		nvfx_state_emit(nvfx);
+
+		assert(AVAIL_RING(chan) >= ((1 + count * num_attribs * 4) + 6 + 64));
 	}
 
 	/* Switch primitive modes if necessary */
 	if (rs->prim != mode) {
 		if (rs->prim != NV34TCL_VERTEX_BEGIN_END_STOP) {
-			BEGIN_RING(chan, eng3d, NV34TCL_VERTEX_BEGIN_END, 1);
-			OUT_RING  (chan, NV34TCL_VERTEX_BEGIN_END_STOP);
+			OUT_RING(chan, RING_3D(NV34TCL_VERTEX_BEGIN_END, 1));
+			OUT_RING(chan, NV34TCL_VERTEX_BEGIN_END_STOP);
+		}
+
+		/* XXX: any command a lot of times seems to (mostly) fix corruption that would otherwise happen */
+		/* this seems to cause issues on nv3x, and also be unneeded there */
+		if(nvfx->is_nv4x)
+		{
+			int i;
+			for(i = 0; i < 32; ++i)
+			{
+				OUT_RING(chan, RING_3D(0x1dac, 1));
+				OUT_RING(chan, 0);
+			}
 		}
 
-		BEGIN_RING(chan, eng3d, NV34TCL_VERTEX_BEGIN_END, 1);
+		OUT_RING(chan, RING_3D(NV34TCL_VERTEX_BEGIN_END, 1));
 		OUT_RING  (chan, mode);
 		rs->prim = mode;
 	}
 
-	/* Emit vertex data */
-	for (i = 0; i < count; i++)
-		nvfx_render_vertex(nvfx, prim->v[i]);
-
-	/* If it's likely we'll need to empty the push buffer soon, finish
-	 * off the primitive now.
-	 */
-	if (AVAIL_RING(chan) < ((count * 20) + 6)) {
-		BEGIN_RING(chan, eng3d, NV34TCL_VERTEX_BEGIN_END, 1);
-		OUT_RING  (chan, NV34TCL_VERTEX_BEGIN_END_STOP);
-		rs->prim = NV34TCL_VERTEX_BEGIN_END_STOP;
+	OUT_RING(chan, RING_3D_NI(NV34TCL_VERTEX_DATA, num_attribs * 4 * count));
+	if(no_elements) {
+		OUT_RING(chan, 0);
+		OUT_RING(chan, 0);
+		OUT_RING(chan, 0);
+		OUT_RING(chan, 0);
+	} else {
+		for (unsigned i = 0; i < count; ++i)
+		{
+			struct vertex_header* v = prim->v[i];
+			/* TODO: disable divide where it's causing the problem, and remove this hack */
+			OUT_RING(chan, fui(v->data[0][0] / v->data[0][3]));
+			OUT_RING(chan, fui(v->data[0][1] / v->data[0][3]));
+			OUT_RING(chan, fui(v->data[0][2] / v->data[0][3]));
+			OUT_RING(chan, fui(1.0f / v->data[0][3]));
+			OUT_RINGp(chan, &v->data[1][0], 4 * (num_attribs - 1));
+		}
 	}
 }
 
@@ -150,24 +123,10 @@ nvfx_render_tri(struct draw_stage *draw, struct prim_header *prim)
 }
 
 static void
-nvfx_render_flush(struct draw_stage *draw, unsigned flags)
-{
-	struct nvfx_render_stage *rs = nvfx_render_stage(draw);
-	struct nvfx_context *nvfx = rs->nvfx;
-	struct nvfx_screen *screen = nvfx->screen;
-	struct nouveau_channel *chan = screen->base.channel;
-	struct nouveau_grobj *eng3d = screen->eng3d;
-
-	if (rs->prim != NV34TCL_VERTEX_BEGIN_END_STOP) {
-		BEGIN_RING(chan, eng3d, NV34TCL_VERTEX_BEGIN_END, 1);
-		OUT_RING  (chan, NV34TCL_VERTEX_BEGIN_END_STOP);
-		rs->prim = NV34TCL_VERTEX_BEGIN_END_STOP;
-	}
-}
-
-static void
 nvfx_render_reset_stipple_counter(struct draw_stage *draw)
 {
+	/* this doesn't really seem to work, but it matters rather little */
+	nvfx_render_flush(draw, 0);
 }
 
 static void
@@ -176,40 +135,11 @@ nvfx_render_destroy(struct draw_stage *draw)
 	FREE(draw);
 }
 
-static struct nvfx_vertex_program *
-nvfx_create_drawvp(struct nvfx_context *nvfx)
-{
-	struct ureg_program *ureg;
-	uint i;
-
-	ureg = ureg_create( TGSI_PROCESSOR_VERTEX );
-	if (ureg == NULL)
-		return NULL;
-
-	ureg_MOV(ureg, ureg_DECL_output(ureg, TGSI_SEMANTIC_POSITION, 0), ureg_DECL_vs_input(ureg, 0));
-	ureg_MOV(ureg, ureg_DECL_output(ureg, TGSI_SEMANTIC_COLOR, 0), ureg_DECL_vs_input(ureg, 3));
-	ureg_MOV(ureg, ureg_DECL_output(ureg, TGSI_SEMANTIC_COLOR, 1), ureg_DECL_vs_input(ureg, 4));
-	ureg_MOV(ureg, ureg_DECL_output(ureg, TGSI_SEMANTIC_BCOLOR, 0), ureg_DECL_vs_input(ureg, 3));
-	ureg_MOV(ureg, ureg_DECL_output(ureg, TGSI_SEMANTIC_BCOLOR, 1), ureg_DECL_vs_input(ureg, 4));
-	ureg_MOV(ureg,
-		   ureg_writemask(ureg_DECL_output(ureg, TGSI_SEMANTIC_FOG, 1), TGSI_WRITEMASK_X),
-		   ureg_DECL_vs_input(ureg, 5));
-	for (i = 0; i < 8; ++i)
-		ureg_MOV(ureg, ureg_DECL_output(ureg, TGSI_SEMANTIC_GENERIC, i), ureg_DECL_vs_input(ureg, 8 + i));
-
-	ureg_END( ureg );
-
-	return ureg_create_shader_and_destroy( ureg, &nvfx->pipe );
-}
-
 struct draw_stage *
 nvfx_draw_render_stage(struct nvfx_context *nvfx)
 {
 	struct nvfx_render_stage *render = CALLOC_STRUCT(nvfx_render_stage);
 
-	if (!nvfx->swtnl.vertprog)
-		nvfx->swtnl.vertprog = nvfx_create_drawvp(nvfx);
-
 	render->nvfx = nvfx;
 	render->stage.draw = nvfx->draw;
 	render->stage.point = nvfx_render_point;
@@ -231,6 +161,7 @@ nvfx_draw_vbo_swtnl(struct pipe_context *pipe, const struct pipe_draw_info* info
 
 	if (!nvfx_state_validate_swtnl(nvfx))
 		return;
+
 	nvfx_state_emit(nvfx);
 
 	/* these must be passed without adding the offsets */
@@ -256,62 +187,3 @@ nvfx_draw_vbo_swtnl(struct pipe_context *pipe, const struct pipe_draw_info* info
 
 	draw_flush(nvfx->draw);
 }
-
-static INLINE void
-emit_attrib(struct nvfx_context *nvfx, unsigned hw, unsigned emit,
-	    unsigned semantic, unsigned index)
-{
-	unsigned draw_out = draw_find_shader_output(nvfx->draw, semantic, index);
-	unsigned a = nvfx->swtnl.nr_attribs++;
-
-	nvfx->swtnl.hw[a] = hw;
-	nvfx->swtnl.emit[a] = emit;
-	nvfx->swtnl.draw[a] = draw_out;
-}
-
-void
-nvfx_vtxfmt_validate(struct nvfx_context *nvfx)
-{
-	struct nvfx_pipe_fragment_program *pfp = nvfx->fragprog;
-	unsigned colour = 0, texcoords = 0, fog = 0, i;
-
-	/* Determine needed fragprog inputs */
-	for (i = 0; i < pfp->info.num_inputs; i++) {
-		switch (pfp->info.input_semantic_name[i]) {
-		case TGSI_SEMANTIC_POSITION:
-			break;
-		case TGSI_SEMANTIC_COLOR:
-			colour |= (1 << pfp->info.input_semantic_index[i]);
-			break;
-		case TGSI_SEMANTIC_GENERIC:
-			texcoords |= (1 << pfp->info.input_semantic_index[i]);
-			break;
-		case TGSI_SEMANTIC_FOG:
-			fog = 1;
-			break;
-		default:
-			assert(0);
-		}
-	}
-
-	nvfx->swtnl.nr_attribs = 0;
-
-	/* Map draw vtxprog output to hw attribute IDs */
-	for (i = 0; i < 2; i++) {
-		if (!(colour & (1 << i)))
-			continue;
-		emit_attrib(nvfx, 3 + i, EMIT_4F, TGSI_SEMANTIC_COLOR, i);
-	}
-
-	for (i = 0; i < 8; i++) {
-		if (!(texcoords & (1 << i)))
-			continue;
-		emit_attrib(nvfx, 8 + i, EMIT_4F, TGSI_SEMANTIC_GENERIC, i);
-	}
-
-	if (fog) {
-		emit_attrib(nvfx, 5, EMIT_1F, TGSI_SEMANTIC_FOG, 0);
-	}
-
-	emit_attrib(nvfx, 0, 0xff, TGSI_SEMANTIC_POSITION, 0);
-}
diff --git a/src/gallium/drivers/nvfx/nvfx_fragprog.c b/src/gallium/drivers/nvfx/nvfx_fragprog.c
index 93ba538241..86df7f0049 100644
--- a/src/gallium/drivers/nvfx/nvfx_fragprog.c
+++ b/src/gallium/drivers/nvfx/nvfx_fragprog.c
@@ -1263,7 +1263,7 @@ nvfx_fragprog_validate(struct nvfx_context *nvfx)
 		pfp->fps[key] = fp;
 	}
 
-	vp = nvfx->render_mode == HW ? nvfx->vertprog : nvfx->swtnl.vertprog;
+	vp = nvfx->hw_vertprog;
 
 	if (fp->last_vp_id != vp->id || fp->last_sprite_coord_enable != sprite_coord_enable) {
 		int sprite_real_input = -1;
diff --git a/src/gallium/drivers/nvfx/nvfx_screen.c b/src/gallium/drivers/nvfx/nvfx_screen.c
index ac8053f26b..3900821de4 100644
--- a/src/gallium/drivers/nvfx/nvfx_screen.c
+++ b/src/gallium/drivers/nvfx/nvfx_screen.c
@@ -432,7 +432,7 @@ nvfx_screen_create(struct pipe_winsys *ws, struct nouveau_device *dev)
 		return NULL;
 	}
 
-	screen->force_swtnl = debug_get_bool_option("NOUVEAU_SWTNL", FALSE);
+	screen->force_swtnl = debug_get_bool_option("NVFX_SWTNL", FALSE);
 	screen->trace_draw = debug_get_bool_option("NVFX_TRACE_DRAW", FALSE);
 
 	screen->buffer_allocation_cost = debug_get_num_option("NVFX_BUFFER_ALLOCATION_COST", 16384);
diff --git a/src/gallium/drivers/nvfx/nvfx_state.h b/src/gallium/drivers/nvfx/nvfx_state.h
index 9200f78a54..8fafca1950 100644
--- a/src/gallium/drivers/nvfx/nvfx_state.h
+++ b/src/gallium/drivers/nvfx/nvfx_state.h
@@ -17,13 +17,8 @@ struct nvfx_vertex_program_data {
 };
 
 struct nvfx_vertex_program {
-	struct pipe_shader_state pipe;
 	unsigned long long id;
 
-	struct draw_vertex_shader *draw;
-
-	boolean translated;
-
 	struct nvfx_vertex_program_exec *insns;
 	unsigned nr_insns;
 	struct nvfx_vertex_program_data *consts;
@@ -46,6 +41,20 @@ struct nvfx_vertex_program {
 	struct util_dynarray const_relocs;
 };
 
+#define NVFX_VP_FAILED ((struct nvfx_vertex_program*)-1)
+
+struct nvfx_pipe_vertex_program {
+	struct pipe_shader_state pipe;
+	struct tgsi_shader_info info;
+
+	unsigned draw_elements;
+	boolean draw_no_elements;
+	struct draw_vertex_shader *draw_vs;
+	struct nvfx_vertex_program* draw_vp;
+
+	struct nvfx_vertex_program* vp;
+};
+
 struct nvfx_fragment_program_data {
 	unsigned offset;
 	unsigned index;
diff --git a/src/gallium/drivers/nvfx/nvfx_state_emit.c b/src/gallium/drivers/nvfx/nvfx_state_emit.c
index 308c25fbe1..30ef12a95b 100644
--- a/src/gallium/drivers/nvfx/nvfx_state_emit.c
+++ b/src/gallium/drivers/nvfx/nvfx_state_emit.c
@@ -145,7 +145,7 @@ nvfx_vertprog_ucp_validate(struct nvfx_context* nvfx)
 {
 	struct nouveau_channel* chan = nvfx->screen->base.channel;
 	unsigned i;
-	struct nvfx_vertex_program* vp = nvfx->vertprog;
+	struct nvfx_vertex_program* vp = nvfx->hw_vertprog;
 	if(nvfx->clip.nr != vp->clip_nr)
 	{
 		unsigned idx;
@@ -230,7 +230,7 @@ nvfx_state_validate_common(struct nvfx_context *nvfx)
 
 	if(nvfx->render_mode == HW)
 	{
-		if(dirty & (NVFX_NEW_VERTPROG | NVFX_NEW_VERTCONST | NVFX_NEW_UCP))
+		if(dirty & (NVFX_NEW_VERTPROG | NVFX_NEW_VERTCONST))
 		{
 			if(!nvfx_vertprog_validate(nvfx))
 				return FALSE;
@@ -252,12 +252,10 @@ nvfx_state_validate_common(struct nvfx_context *nvfx)
 	}
 	else
 	{
-		/* TODO: this looks a bit misdesigned */
-		if(dirty & (NVFX_NEW_VERTPROG | NVFX_NEW_UCP))
-			nvfx_vertprog_validate(nvfx);
-
-		if(dirty & (NVFX_NEW_ARRAYS | NVFX_NEW_INDEX | NVFX_NEW_FRAGPROG))
-			nvfx_vtxfmt_validate(nvfx);
+		if(dirty & NVFX_NEW_VERTPROG) {
+			assert(nvfx_vertprog_validate(nvfx));
+			nvfx_vbo_swtnl_validate(nvfx);
+		}
 	}
 
 	if(dirty & NVFX_NEW_RAST)
@@ -284,7 +282,7 @@ nvfx_state_validate_common(struct nvfx_context *nvfx)
 
 	if(nvfx->is_nv4x)
 	{
-		unsigned vp_output = nvfx->vertprog->or | nvfx->hw_fragprog->or;
+		unsigned vp_output = nvfx->hw_vertprog->or | nvfx->hw_fragprog->or;
 		vp_output |= (1 << (nvfx->clip.nr + 6)) - (1 << 6);
 
 		if(vp_output != nvfx->hw_vp_output)
@@ -399,8 +397,6 @@ nvfx_state_relocate(struct nvfx_context *nvfx, unsigned relocs)
 boolean
 nvfx_state_validate(struct nvfx_context *nvfx)
 {
-	boolean was_sw = nvfx->fallback_swtnl ? TRUE : FALSE;
-
 	if (nvfx->render_mode != HW) {
 		/* Don't even bother trying to go back to hw if none
 		 * of the states that caused swtnl previously have changed.
@@ -429,7 +425,11 @@ nvfx_state_validate_swtnl(struct nvfx_context *nvfx)
 
 	/* Setup for swtnl */
 	if (nvfx->render_mode == HW) {
-		NOUVEAU_ERR("hw->swtnl 0x%08x\n", nvfx->fallback_swtnl);
+		static boolean warned = FALSE;
+		if(!warned) {
+			NOUVEAU_ERR("hw->swtnl 0x%08x\n", nvfx->fallback_swtnl);
+			warned = TRUE;
+		}
 		nvfx->pipe.flush(&nvfx->pipe, 0, NULL);
 		nvfx->dirty |= (NVFX_NEW_VIEWPORT |
 				NVFX_NEW_VERTPROG |
@@ -437,8 +437,11 @@ nvfx_state_validate_swtnl(struct nvfx_context *nvfx)
 		nvfx->render_mode = SWTNL;
 	}
 
-	if (nvfx->draw_dirty & NVFX_NEW_VERTPROG)
-		draw_bind_vertex_shader(draw, nvfx->vertprog->draw);
+	if (nvfx->draw_dirty & NVFX_NEW_VERTPROG) {
+		if(!nvfx->vertprog->draw_vs)
+			nvfx->vertprog->draw_vs = draw_create_vertex_shader(draw, &nvfx->vertprog->pipe);
+		draw_bind_vertex_shader(draw, nvfx->vertprog->draw_vs);
+	}
 
 	if (nvfx->draw_dirty & NVFX_NEW_RAST)
            draw_set_rasterizer_state(draw, &nvfx->rasterizer->pipe,
diff --git a/src/gallium/drivers/nvfx/nvfx_vbo.c b/src/gallium/drivers/nvfx/nvfx_vbo.c
index 611de808af..c35e926a7a 100644
--- a/src/gallium/drivers/nvfx/nvfx_vbo.c
+++ b/src/gallium/drivers/nvfx/nvfx_vbo.c
@@ -339,6 +339,44 @@ nvfx_vbo_validate(struct nvfx_context *nvfx)
 }
 
 void
+nvfx_vbo_swtnl_validate(struct nvfx_context *nvfx)
+{
+	struct nouveau_channel* chan = nvfx->screen->base.channel;
+	unsigned num_outputs = nvfx->vertprog->draw_elements;
+	int elements = MAX2(num_outputs, nvfx->hw_vtxelt_nr);
+
+	if (!elements)
+		return;
+
+	WAIT_RING(chan, (1 + 6 + 1 + 2) + elements * 2);
+
+	OUT_RING(chan, RING_3D(NV34TCL_VTXFMT(0), elements));
+	for(unsigned i = 0; i < num_outputs; ++i)
+		OUT_RING(chan, (4 << NV34TCL_VTXFMT_SIZE_SHIFT) | NV34TCL_VTXFMT_TYPE_32_FLOAT);
+	for(unsigned i = num_outputs; i < elements; ++i)
+		OUT_RING(chan, NV34TCL_VTXFMT_TYPE_32_FLOAT);
+
+	if(nvfx->is_nv4x) {
+		unsigned i;
+		/* seems to be some kind of cache flushing */
+		for(i = 0; i < 3; ++i) {
+			OUT_RING(chan, RING_3D(0x1718, 1));
+			OUT_RING(chan, 0);
+		}
+	}
+
+	OUT_RING(chan, RING_3D(NV34TCL_VTXBUF_ADDRESS(0), elements));
+	for (unsigned i = 0; i < elements; i++)
+		OUT_RING(chan, 0);
+
+	OUT_RING(chan, RING_3D(0x1710, 1));
+	OUT_RING(chan, 0);
+
+	nvfx->hw_vtxelt_nr = num_outputs;
+	nvfx->relocs_needed &=~ NVFX_RELOCATE_VTXBUF;
+}
+
+void
 nvfx_vbo_relocate(struct nvfx_context *nvfx)
 {
 	struct nouveau_channel* chan;
diff --git a/src/gallium/drivers/nvfx/nvfx_vertprog.c b/src/gallium/drivers/nvfx/nvfx_vertprog.c
index bc78ed400a..30385b26f7 100644
--- a/src/gallium/drivers/nvfx/nvfx_vertprog.c
+++ b/src/gallium/drivers/nvfx/nvfx_vertprog.c
@@ -8,6 +8,7 @@
 #include "tgsi/tgsi_parse.h"
 #include "tgsi/tgsi_dump.h"
 #include "tgsi/tgsi_util.h"
+#include "tgsi/tgsi_ureg.h"
 
 #include "draw/draw_context.h"
 
@@ -37,6 +38,7 @@ struct nvfx_loop_entry
 
 struct nvfx_vpc {
 	struct nvfx_context* nvfx;
+	struct pipe_shader_state pipe;
 	struct nvfx_vertex_program *vp;
 
 	struct nvfx_vertex_program_exec *vpi;
@@ -813,7 +815,7 @@ nvfx_vertprog_prepare(struct nvfx_context* nvfx, struct nvfx_vpc *vpc)
 	unsigned num_outputs;
 	unsigned num_texcoords = nvfx->is_nv4x ? 10 : 8;
 
-	num_outputs = util_semantic_set_from_program_file(&set, vpc->vp->pipe.tokens, TGSI_FILE_OUTPUT);
+	num_outputs = util_semantic_set_from_program_file(&set, vpc->pipe.tokens, TGSI_FILE_OUTPUT);
 
 	if(num_outputs > num_texcoords) {
 		NOUVEAU_ERR("too many vertex program outputs: %i\n", num_outputs);
@@ -840,7 +842,7 @@ nvfx_vertprog_prepare(struct nvfx_context* nvfx, struct nvfx_vpc *vpc)
 		}
 	}
 
-	tgsi_parse_init(&p, vpc->vp->pipe.tokens);
+	tgsi_parse_init(&p, vpc->pipe.tokens);
 	while (!tgsi_parse_end_of_tokens(&p)) {
 		const union tgsi_full_token *tok = &p.FullToken;
 
@@ -917,21 +919,35 @@ nvfx_vertprog_prepare(struct nvfx_context* nvfx, struct nvfx_vpc *vpc)
 
 DEBUG_GET_ONCE_BOOL_OPTION(nvfx_dump_vp, "NVFX_DUMP_VP", FALSE)
 
-static void
-nvfx_vertprog_translate(struct nvfx_context *nvfx,
-			struct nvfx_vertex_program *vp)
+static struct nvfx_vertex_program*
+nvfx_vertprog_translate(struct nvfx_context *nvfx, const struct pipe_shader_state* vps)
 {
 	struct tgsi_parse_context parse;
+	struct nvfx_vertex_program* vp = NULL;
 	struct nvfx_vpc *vpc = NULL;
 	struct nvfx_src none = nvfx_src(nvfx_reg(NVFXSR_NONE, 0));
 	struct util_dynarray insns;
 	int i;
 
-	vpc = CALLOC(1, sizeof(struct nvfx_vpc));
+	tgsi_parse_init(&parse, vps->tokens);
+
+	vp = CALLOC_STRUCT(nvfx_vertex_program);
+	if(!vp)
+		goto out_err;
+
+	vpc = CALLOC_STRUCT(nvfx_vpc);
 	if (!vpc)
-		return;
+		goto out_err;
+
 	vpc->nvfx = nvfx;
 	vpc->vp = vp;
+	vpc->pipe = *vps;
+
+	{
+		// TODO: use a 64-bit atomic here!
+		static unsigned long long id = 0;
+		vp->id = ++id;
+	}
 
 	/* reserve space for ucps */
 	if(nvfx->use_vp_clipping)
@@ -942,7 +958,7 @@ nvfx_vertprog_translate(struct nvfx_context *nvfx,
 
 	if (!nvfx_vertprog_prepare(nvfx, vpc)) {
 		FREE(vpc);
-		return;
+		return NULL;
 	}
 
 	/* Redirect post-transform vertex position to a temp if user clip
@@ -955,8 +971,6 @@ nvfx_vertprog_translate(struct nvfx_context *nvfx,
 		vpc->r_temps_discard = 0;
 	}
 
-	tgsi_parse_init(&parse, vp->pipe.tokens);
-
 	util_dynarray_init(&insns);
 	while (!tgsi_parse_end_of_tokens(&parse)) {
 		tgsi_parse_token(&parse);
@@ -1058,7 +1072,7 @@ nvfx_vertprog_translate(struct nvfx_context *nvfx,
 	if(debug_get_option_nvfx_dump_vp())
 	{
 		debug_printf("\n");
-		tgsi_dump(vp->pipe.tokens, 0);
+		tgsi_dump(vpc->pipe.tokens, 0);
 
 		debug_printf("\n%s vertex program:\n", nvfx->is_nv4x ? "nv4x" : "nv3x");
 		for (i = 0; i < vp->nr_insns; i++)
@@ -1068,20 +1082,49 @@ nvfx_vertprog_translate(struct nvfx_context *nvfx,
 
 	vp->clip_nr = -1;
 	vp->exec_start = -1;
-	vp->translated = TRUE;
-out_err:
+
+out:
 	tgsi_parse_free(&parse);
-	util_dynarray_fini(&vpc->label_relocs);
-	util_dynarray_fini(&vpc->loop_stack);
-	if (vpc->r_temp)
+	if(vpc) {
+		util_dynarray_fini(&vpc->label_relocs);
+		util_dynarray_fini(&vpc->loop_stack);
 		FREE(vpc->r_temp);
-	if (vpc->r_address)
 		FREE(vpc->r_address);
-	if (vpc->r_const)
 		FREE(vpc->r_const);
-	if (vpc->imm)
 		FREE(vpc->imm);
-	FREE(vpc);
+		FREE(vpc);
+	}
+	return vp;
+
+out_err:
+	FREE(vp);
+	vp = NULL;
+	goto out;
+}
+
+static struct nvfx_vertex_program*
+nvfx_vertprog_translate_draw_vp(struct nvfx_context *nvfx, struct nvfx_pipe_vertex_program* pvp)
+{
+	struct nvfx_vertex_program* vp = NULL;
+	struct pipe_shader_state vps;
+	struct ureg_program *ureg = NULL;
+	unsigned num_outputs = MIN2(pvp->info.num_outputs, 16);
+
+	ureg = ureg_create( TGSI_PROCESSOR_VERTEX );
+	if(ureg == NULL)
+		return 0;
+
+	for (unsigned i = 0; i < num_outputs; i++)
+		ureg_MOV(ureg, ureg_DECL_output(ureg, pvp->info.output_semantic_name[i], pvp->info.output_semantic_index[i]), ureg_DECL_vs_input(ureg, i));
+
+	ureg_END( ureg );
+
+	vps.tokens = ureg_get_tokens(ureg, 0);
+	vp = nvfx_vertprog_translate(nvfx, &vps);
+	ureg_free_tokens(vps.tokens);
+	ureg_destroy(ureg);
+
+	return vp;
 }
 
 boolean
@@ -1090,30 +1133,44 @@ nvfx_vertprog_validate(struct nvfx_context *nvfx)
 	struct nvfx_screen *screen = nvfx->screen;
 	struct nouveau_channel *chan = screen->base.channel;
 	struct nouveau_grobj *eng3d = screen->eng3d;
-	struct nvfx_vertex_program *vp;
+	struct nvfx_pipe_vertex_program *pvp = nvfx->vertprog;
+	struct nvfx_vertex_program* vp;
 	struct pipe_resource *constbuf;
 	boolean upload_code = FALSE, upload_data = FALSE;
 	int i;
 
 	if (nvfx->render_mode == HW) {
-		vp = nvfx->vertprog;
-		constbuf = nvfx->constbuf[PIPE_SHADER_VERTEX];
-	} else {
-		vp = nvfx->swtnl.vertprog;
-		constbuf = NULL;
-	}
-
-	/* Translate TGSI shader into hw bytecode */
-	if (!vp->translated)
-	{
 		nvfx->fallback_swtnl &= ~NVFX_NEW_VERTPROG;
-		nvfx_vertprog_translate(nvfx, vp);
-		if (!vp->translated) {
+		vp = pvp->vp;
+
+		if(!vp) {
+			vp = nvfx_vertprog_translate(nvfx, &pvp->pipe);
+			if(!vp)
+				vp = NVFX_VP_FAILED;
+			pvp->vp = vp;
+		}
+
+		if(vp == NVFX_VP_FAILED) {
 			nvfx->fallback_swtnl |= NVFX_NEW_VERTPROG;
 			return FALSE;
 		}
+
+		constbuf = nvfx->constbuf[PIPE_SHADER_VERTEX];
+	} else {
+		vp = pvp->draw_vp;
+		if(!vp)
+		{
+			pvp->draw_vp = vp = nvfx_vertprog_translate_draw_vp(nvfx, pvp);
+			if(!vp) {
+				_debug_printf("Error: unable to create a swtnl passthrough vertex shader: aborting.");
+				abort();
+			}
+		}
+		constbuf = NULL;
 	}
 
+	nvfx->hw_vertprog = vp;
+
 	/* Allocate hw vtxprog exec slots */
 	if (!vp->exec) {
 		struct nouveau_resource *heap = nvfx->screen->vp_exec_heap;
@@ -1294,24 +1351,22 @@ nvfx_vertprog_destroy(struct nvfx_context *nvfx, struct nvfx_vertex_program *vp)
 
 	util_dynarray_fini(&vp->branch_relocs);
 	util_dynarray_fini(&vp->const_relocs);
+	FREE(vp);
 }
 
 static void *
-nvfx_vp_state_create(struct pipe_context *pipe,
-                     const struct pipe_shader_state *cso)
+nvfx_vp_state_create(struct pipe_context *pipe, const struct pipe_shader_state *cso)
 {
         struct nvfx_context *nvfx = nvfx_context(pipe);
-        struct nvfx_vertex_program *vp;
+        struct nvfx_pipe_vertex_program *pvp;
 
-        // TODO: use a 64-bit atomic here!
-        static unsigned long long id = 0;
+        pvp = CALLOC(1, sizeof(struct nvfx_pipe_vertex_program));
+        pvp->pipe.tokens = tgsi_dup_tokens(cso->tokens);
+        tgsi_scan_shader(pvp->pipe.tokens, &pvp->info);
+        pvp->draw_elements = MAX2(1, MIN2(pvp->info.num_outputs, 16));
+        pvp->draw_no_elements = pvp->info.num_outputs == 0;
 
-        vp = CALLOC(1, sizeof(struct nvfx_vertex_program));
-        vp->pipe.tokens = tgsi_dup_tokens(cso->tokens);
-        vp->draw = draw_create_vertex_shader(nvfx->draw, &vp->pipe);
-        vp->id = ++id;
-
-        return (void *)vp;
+        return (void *)pvp;
 }
 
 static void
@@ -1327,13 +1382,17 @@ nvfx_vp_state_bind(struct pipe_context *pipe, void *hwcso)
 static void
 nvfx_vp_state_delete(struct pipe_context *pipe, void *hwcso)
 {
-        struct nvfx_context *nvfx = nvfx_context(pipe);
-        struct nvfx_vertex_program *vp = hwcso;
+	struct nvfx_context *nvfx = nvfx_context(pipe);
+	struct nvfx_pipe_vertex_program *pvp = hwcso;
 
-        draw_delete_vertex_shader(nvfx->draw, vp->draw);
-        nvfx_vertprog_destroy(nvfx, vp);
-        FREE((void*)vp->pipe.tokens);
-        FREE(vp);
+	if(pvp->draw_vs)
+		draw_delete_vertex_shader(nvfx->draw, pvp->draw_vs);
+	if(pvp->vp && pvp->vp != NVFX_VP_FAILED)
+		nvfx_vertprog_destroy(nvfx, pvp->vp);
+	if(pvp->draw_vp)
+		nvfx_vertprog_destroy(nvfx, pvp->draw_vp);
+	FREE((void*)pvp->pipe.tokens);
+	FREE(pvp);
 }
 
 void