nv40: simple swtnl path (half broken, but getting there)

12 files changed, 531 insertions, 91 deletions

diff --git a/src/gallium/drivers/nv40/nv40_context.c b/src/gallium/drivers/nv40/nv40_context.c
index 203c843a01..58627443b8 100644
--- a/src/gallium/drivers/nv40/nv40_context.c
+++ b/src/gallium/drivers/nv40/nv40_context.c
@@ -74,8 +74,12 @@ nv40_create(struct pipe_screen *pscreen, unsigned pctx_id)
 	nv40_init_state_functions(nv40);
 	nv40_init_miptree_functions(nv40);
 
+	/* Create, configure, and install fallback swtnl path */
 	nv40->draw = draw_create();
-	assert(nv40->draw);
+	draw_wide_point_threshold(nv40->draw, 9999999.0);
+	draw_wide_line_threshold(nv40->draw, 9999999.0);
+	draw_enable_line_stipple(nv40->draw, FALSE);
+	draw_enable_point_sprites(nv40->draw, FALSE);
 	draw_set_rasterize_stage(nv40->draw, nv40_draw_render_stage(nv40));
 
 	return &nv40->pipe;
diff --git a/src/gallium/drivers/nv40/nv40_context.h b/src/gallium/drivers/nv40/nv40_context.h
index 100c678187..02ca20b801 100644
--- a/src/gallium/drivers/nv40/nv40_context.h
+++ b/src/gallium/drivers/nv40/nv40_context.h
@@ -116,7 +116,20 @@ struct nv40_context {
 
 	/* HW state derived from pipe states */
 	struct nv40_state state;
-	unsigned fallback;
+	struct {
+		struct nv40_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
+	} render_mode;
+	unsigned fallback_swtnl;
+	unsigned fallback_swrast;
 
 	/* Context state */
 	unsigned dirty;
@@ -166,6 +179,10 @@ extern void nv40_screen_init_miptree_functions(struct pipe_screen *pscreen);
 
 /* nv40_draw.c */
 extern struct draw_stage *nv40_draw_render_stage(struct nv40_context *nv40);
+extern boolean nv40_draw_elements_swtnl(struct pipe_context *pipe,
+					struct pipe_buffer *idxbuf,
+					unsigned ib_size, unsigned mode,
+					unsigned start, unsigned count);
 
 /* nv40_vertprog.c */
 extern void nv40_vertprog_destroy(struct nv40_context *,
@@ -179,8 +196,9 @@ extern void nv40_fragprog_destroy(struct nv40_context *,
 extern void nv40_fragtex_bind(struct nv40_context *);
 
 /* nv40_state.c and friends */
-extern void nv40_emit_hw_state(struct nv40_context *nv40);
-extern void nv40_state_tex_update(struct nv40_context *nv40);
+extern boolean nv40_state_validate(struct nv40_context *nv40);
+extern boolean nv40_state_validate_swtnl(struct nv40_context *nv40);
+extern void nv40_state_emit(struct nv40_context *nv40);
 extern struct nv40_state_entry nv40_state_clip;
 extern struct nv40_state_entry nv40_state_rasterizer;
 extern struct nv40_state_entry nv40_state_scissor;
@@ -194,6 +212,7 @@ extern struct nv40_state_entry nv40_state_viewport;
 extern struct nv40_state_entry nv40_state_framebuffer;
 extern struct nv40_state_entry nv40_state_fragtex;
 extern struct nv40_state_entry nv40_state_vbo;
+extern struct nv40_state_entry nv40_state_vtxfmt;
 
 /* nv40_vbo.c */
 extern boolean nv40_draw_arrays(struct pipe_context *, unsigned mode,
diff --git a/src/gallium/drivers/nv40/nv40_draw.c b/src/gallium/drivers/nv40/nv40_draw.c
index a39bb85e99..ce0e0bc6f2 100644
--- a/src/gallium/drivers/nv40/nv40_draw.c
+++ b/src/gallium/drivers/nv40/nv40_draw.c
@@ -1,62 +1,350 @@
-#include "draw/draw_private.h"
 #include "pipe/p_util.h"
+#include "pipe/p_shader_tokens.h"
+
+#include "draw/draw_context.h"
+#include "draw/draw_vertex.h"
+#include "draw/draw_private.h"
 
 #include "nv40_context.h"
+#define NV40_SHADER_NO_FUCKEDNESS
+#include "nv40_shader.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 nv40_draw_stage {
-	struct draw_stage draw;
+struct nv40_render_stage {
+	struct draw_stage stage;
 	struct nv40_context *nv40;
+	unsigned prim;
 };
 
+static INLINE struct nv40_render_stage *
+nv40_render_stage(struct draw_stage *stage)
+{
+	return (struct nv40_render_stage *)stage;
+}
+
+static INLINE void
+nv40_render_vertex(struct nv40_context *nv40, const struct vertex_header *v)
+{
+	unsigned i;
+
+	for (i = 0; i < nv40->swtnl.nr_attribs; i++) {
+		unsigned idx = nv40->swtnl.draw[i];
+		unsigned hw = nv40->swtnl.hw[i];
+
+		switch (nv40->swtnl.emit[i]) {
+		case EMIT_OMIT:
+			break;
+		case EMIT_1F:
+			BEGIN_RING(curie, 0x1e40 + (hw * 4), 1);
+			OUT_RING  (fui(v->data[idx][0]));
+			break;
+		case EMIT_2F:
+			BEGIN_RING(curie, NV40TCL_VTX_ATTR_2F_X(hw), 2);
+			OUT_RING  (fui(v->data[idx][0]));
+			OUT_RING  (fui(v->data[idx][1]));
+			break;
+		case EMIT_3F:
+			BEGIN_RING(curie, NV40TCL_VTX_ATTR_3F_X(hw), 3);
+			OUT_RING  (fui(v->data[idx][0]));
+			OUT_RING  (fui(v->data[idx][1]));
+			OUT_RING  (fui(v->data[idx][2]));
+			break;
+		case EMIT_4F:
+			BEGIN_RING(curie, NV40TCL_VTX_ATTR_4F_X(hw), 4);
+			OUT_RING  (fui(v->data[idx][0]));
+			OUT_RING  (fui(v->data[idx][1]));
+			OUT_RING  (fui(v->data[idx][2]));
+			OUT_RING  (fui(v->data[idx][3]));
+			break;
+		case EMIT_4UB:
+			BEGIN_RING(curie, 0x1940 + (hw * 4), 1);
+			OUT_RING  (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])));
+			break;
+		default:
+			assert(0);
+			break;
+		}
+	}
+}
+
+static INLINE void
+nv40_render_prim(struct draw_stage *stage, struct prim_header *prim,
+	       unsigned mode, unsigned count)
+{
+	struct nv40_render_stage *rs = nv40_render_stage(stage);
+	struct nv40_context *nv40 = rs->nv40;
+	struct nouveau_pushbuf *pb = nv40->nvws->channel->pushbuf;
+	unsigned i;
+
+	/* Ensure there's room for 4xfloat32 + potentially 3 begin/end */
+	if (pb->remaining < ((count * 20) + 6)) {
+		if (rs->prim != NV40TCL_BEGIN_END_STOP) {
+			NOUVEAU_ERR("AIII, missed flush\n");
+			assert(0);
+		}
+		FIRE_RING();
+		nv40_state_emit(nv40);
+	}
+
+	/* Switch primitive modes if necessary */
+	if (rs->prim != mode) {
+		if (rs->prim != NV40TCL_BEGIN_END_STOP) {
+			BEGIN_RING(curie, NV40TCL_BEGIN_END, 1);
+			OUT_RING  (NV40TCL_BEGIN_END_STOP);	
+		}
+
+		BEGIN_RING(curie, NV40TCL_BEGIN_END, 1);
+		OUT_RING  (mode);
+		rs->prim = mode;
+	}
+
+	/* Emit vertex data */
+	for (i = 0; i < count; i++)
+		nv40_render_vertex(nv40, prim->v[i]);
+
+	/* If it's likely we'll need to empty the push buffer soon, finish
+	 * off the primitive now.
+	 */
+	if (pb->remaining < ((count * 20) + 6)) {
+		BEGIN_RING(curie, NV40TCL_BEGIN_END, 1);
+		OUT_RING  (NV40TCL_BEGIN_END_STOP);
+		rs->prim = NV40TCL_BEGIN_END_STOP;
+	}
+}
+
 static void
-nv40_draw_point(struct draw_stage *draw, struct prim_header *prim)
+nv40_render_point(struct draw_stage *draw, struct prim_header *prim)
 {
-	NOUVEAU_ERR("\n");
+	nv40_render_prim(draw, prim, NV40TCL_BEGIN_END_POINTS, 1);
 }
 
 static void
-nv40_draw_line(struct draw_stage *draw, struct prim_header *prim)
+nv40_render_line(struct draw_stage *draw, struct prim_header *prim)
 {
-	NOUVEAU_ERR("\n");
+	nv40_render_prim(draw, prim, NV40TCL_BEGIN_END_LINES, 2);
 }
 
 static void
-nv40_draw_tri(struct draw_stage *draw, struct prim_header *prim)
+nv40_render_tri(struct draw_stage *draw, struct prim_header *prim)
 {
-	NOUVEAU_ERR("\n");
+	nv40_render_prim(draw, prim, NV40TCL_BEGIN_END_TRIANGLES, 3);
 }
 
 static void
-nv40_draw_flush(struct draw_stage *draw, unsigned flags)
+nv40_render_flush(struct draw_stage *draw, unsigned flags)
 {
+	struct nv40_render_stage *rs = nv40_render_stage(draw);
+	struct nv40_context *nv40 = rs->nv40;
+
+	if (rs->prim != NV40TCL_BEGIN_END_STOP) {
+		BEGIN_RING(curie, NV40TCL_BEGIN_END, 1);
+		OUT_RING  (NV40TCL_BEGIN_END_STOP);
+		rs->prim = NV40TCL_BEGIN_END_STOP;
+	}
 }
 
 static void
-nv40_draw_reset_stipple_counter(struct draw_stage *draw)
+nv40_render_reset_stipple_counter(struct draw_stage *draw)
 {
-	NOUVEAU_ERR("\n");
 }
 
 static void
-nv40_draw_destroy(struct draw_stage *draw)
+nv40_render_destroy(struct draw_stage *draw)
 {
 	free(draw);
 }
 
+static INLINE void
+emit_mov(struct nv40_vertex_program *vp,
+	 unsigned dst, unsigned src, unsigned vor, unsigned mask)
+{
+	struct nv40_vertex_program_exec *inst;
+
+	vp->insns = realloc(vp->insns,
+			    sizeof(struct nv40_vertex_program_exec) *
+			    ++vp->nr_insns);
+	inst = &vp->insns[vp->nr_insns - 1];
+
+	inst->data[0] = 0x401f9c6c;
+	inst->data[1] = 0x0040000d | (src << 8);
+	inst->data[2] = 0x8106c083;
+	inst->data[3] = 0x6041ff80 | (dst << 2) | (mask << 13);
+	inst->const_index = -1;
+	inst->has_branch_offset = FALSE;
+
+	vp->ir |= (1 << src);
+	if (vor != ~0)
+		vp->or |= (1 << vor);
+}
+
+static struct nv40_vertex_program *
+create_drawvp(struct nv40_context *nv40)
+{
+	struct nv40_vertex_program *vp = CALLOC_STRUCT(nv40_vertex_program);
+	unsigned i;
+
+	emit_mov(vp, NV40_VP_INST_DEST_POS, 0, ~0, 0xf);
+	emit_mov(vp, NV40_VP_INST_DEST_COL0, 3, 0, 0xf);
+	emit_mov(vp, NV40_VP_INST_DEST_COL1, 4, 1, 0xf);
+	emit_mov(vp, NV40_VP_INST_DEST_BFC0, 3, 2, 0xf);
+	emit_mov(vp, NV40_VP_INST_DEST_BFC1, 4, 3, 0xf);
+	emit_mov(vp, NV40_VP_INST_DEST_FOGC, 5, 4, 0x8);
+	for (i = 0; i < 8; i++)
+		emit_mov(vp, NV40_VP_INST_DEST_TC(i), 8 + i, 14 + i, 0xf);
+
+	vp->insns[vp->nr_insns - 1].data[3] |= 1;
+	vp->translated = TRUE;
+	return vp;
+}
+
 struct draw_stage *
 nv40_draw_render_stage(struct nv40_context *nv40)
 {
-	struct nv40_draw_stage *nv40draw = CALLOC_STRUCT(nv40_draw_stage);
+	struct nv40_render_stage *render = CALLOC_STRUCT(nv40_render_stage);
+
+	if (!nv40->swtnl.vertprog)
+		nv40->swtnl.vertprog = create_drawvp(nv40);
+
+	render->nv40 = nv40;
+	render->stage.draw = nv40->draw;
+	render->stage.point = nv40_render_point;
+	render->stage.line = nv40_render_line;
+	render->stage.tri = nv40_render_tri;
+	render->stage.flush = nv40_render_flush;
+	render->stage.reset_stipple_counter = nv40_render_reset_stipple_counter;
+	render->stage.destroy = nv40_render_destroy;
+
+	return &render->stage;
+}
+
+boolean
+nv40_draw_elements_swtnl(struct pipe_context *pipe,
+			 struct pipe_buffer *idxbuf, unsigned idxbuf_size,
+			 unsigned mode, unsigned start, unsigned count)
+{
+	struct nv40_context *nv40 = nv40_context(pipe);
+	struct pipe_winsys *ws = pipe->winsys;
+	unsigned i;
+	void *map;
+
+	if (!nv40_state_validate_swtnl(nv40))
+		return FALSE;
+	nv40_state_emit(nv40);
 
-	nv40draw->nv40 = nv40;
-	nv40draw->draw.draw = nv40->draw;
-	nv40draw->draw.point = nv40_draw_point;
-	nv40draw->draw.line = nv40_draw_line;
-	nv40draw->draw.tri = nv40_draw_tri;
-	nv40draw->draw.flush = nv40_draw_flush;
-	nv40draw->draw.reset_stipple_counter = nv40_draw_reset_stipple_counter;
-	nv40draw->draw.destroy = nv40_draw_destroy;
+	for (i = 0; i < PIPE_ATTRIB_MAX; i++) {
+		if (!nv40->vtxbuf[i].buffer)
+			continue;
+		map = ws->buffer_map(ws, nv40->vtxbuf[i].buffer,
+				     PIPE_BUFFER_USAGE_CPU_READ);
+		draw_set_mapped_vertex_buffer(nv40->draw, i, map);
+	}
 
-	return &nv40draw->draw;
+	if (idxbuf) {
+		map = ws->buffer_map(ws, idxbuf, PIPE_BUFFER_USAGE_CPU_READ);
+		draw_set_mapped_element_buffer(nv40->draw, idxbuf_size, map);
+	} else {
+		draw_set_mapped_element_buffer(nv40->draw, 0, NULL);
+	}
+
+	if (nv40->constbuf[PIPE_SHADER_VERTEX]) {
+		map = ws->buffer_map(ws, nv40->constbuf[PIPE_SHADER_VERTEX],
+				     PIPE_BUFFER_USAGE_CPU_READ);
+		draw_set_mapped_constant_buffer(nv40->draw, map);
+	}
+
+	draw_arrays(nv40->draw, mode, start, count);
+
+	for (i = 0; i < PIPE_ATTRIB_MAX; i++) {
+		if (!nv40->vtxbuf[i].buffer)
+			continue;
+		ws->buffer_unmap(ws, nv40->vtxbuf[i].buffer);
+	}
+
+	if (idxbuf)
+		ws->buffer_unmap(ws, idxbuf);
+
+	if (nv40->constbuf[PIPE_SHADER_VERTEX])
+		ws->buffer_unmap(ws, nv40->constbuf[PIPE_SHADER_VERTEX]);
+
+	draw_flush(nv40->draw);
+	pipe->flush(pipe, 0);
+
+	return TRUE;
 }
 
+static INLINE void
+emit_attrib(struct nv40_context *nv40, unsigned hw, unsigned emit,
+	    unsigned semantic, unsigned index)
+{
+	unsigned draw_out = draw_find_vs_output(nv40->draw, semantic, index);
+	unsigned a = nv40->swtnl.nr_attribs++;
+
+	nv40->swtnl.hw[a] = hw;
+	nv40->swtnl.emit[a] = emit;
+	nv40->swtnl.draw[a] = draw_out;
+}
+
+static boolean
+nv40_state_vtxfmt_validate(struct nv40_context *nv40)
+{
+	struct nv40_fragment_program *fp = nv40->fragprog;
+	unsigned colour = 0, texcoords = 0, fog = 0, i;
+
+	/* Determine needed fragprog inputs */
+	for (i = 0; i < fp->info.num_inputs; i++) {
+		switch (fp->info.input_semantic_name[i]) {
+		case TGSI_SEMANTIC_POSITION:
+			break;
+		case TGSI_SEMANTIC_COLOR:
+			colour |= (1 << fp->info.input_semantic_index[i]);
+			break;
+		case TGSI_SEMANTIC_GENERIC:
+			texcoords |= (1 << fp->info.input_semantic_index[i]);
+			break;
+		case TGSI_SEMANTIC_FOG:
+			fog = 1;
+			break;
+		default:
+			assert(0);
+		}
+	}
+
+	nv40->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(nv40, 3 + i, EMIT_4UB, TGSI_SEMANTIC_COLOR, i);
+	}
+
+	for (i = 0; i < 8; i++) {
+		if (!(texcoords & (1 << i)))
+			continue;
+		emit_attrib(nv40, 8 + i, EMIT_4F, TGSI_SEMANTIC_GENERIC, i);
+	}
+
+	if (fog) {
+		emit_attrib(nv40, 5, EMIT_1F, TGSI_SEMANTIC_FOG, 0);
+	}
+
+	emit_attrib(nv40, 0, EMIT_4F, TGSI_SEMANTIC_POSITION, 0);
+
+	return FALSE;
+}
+
+struct nv40_state_entry nv40_state_vtxfmt = {
+	.validate = nv40_state_vtxfmt_validate,
+	.dirty = {
+		.pipe = NV40_NEW_ARRAYS | NV40_NEW_FRAGPROG,
+		.hw = 0
+	}
+};
+
diff --git a/src/gallium/drivers/nv40/nv40_fragprog.c b/src/gallium/drivers/nv40/nv40_fragprog.c
index 953f9cd908..82dbcd3eef 100644
--- a/src/gallium/drivers/nv40/nv40_fragprog.c
+++ b/src/gallium/drivers/nv40/nv40_fragprog.c
@@ -797,9 +797,10 @@ nv40_fragprog_validate(struct nv40_context *nv40)
 	if (fp->translated)
 		goto update_constants;
 
+	nv40->fallback_swrast &= ~NV40_NEW_FRAGPROG;
 	nv40_fragprog_translate(nv40, fp);
 	if (!fp->translated) {
-		nv40->fallback |= NV40_FALLBACK_RAST;
+		nv40->fallback_swrast |= NV40_NEW_FRAGPROG;
 		return FALSE;
 	}
 
diff --git a/src/gallium/drivers/nv40/nv40_shader.h b/src/gallium/drivers/nv40/nv40_shader.h
index 5909c70713..854dccf548 100644
--- a/src/gallium/drivers/nv40/nv40_shader.h
+++ b/src/gallium/drivers/nv40/nv40_shader.h
@@ -476,6 +476,7 @@
 #        define NV40_FP_SWIZZLE_W                                              3
 #define NV40_FP_REG_NEGATE                                             (1 << 17)
 
+#ifndef NV40_SHADER_NO_FUCKEDNESS
 #define NV40SR_NONE	0
 #define NV40SR_OUTPUT	1
 #define NV40SR_INPUT	2
@@ -550,5 +551,6 @@ nv40_sr_scale(struct nv40_sreg src, int scale)
 	src.dst_scale = scale;
 	return src;
 }
+#endif
 
 #endif
diff --git a/src/gallium/drivers/nv40/nv40_state.c b/src/gallium/drivers/nv40/nv40_state.c
index 321d5de041..3eafbece30 100644
--- a/src/gallium/drivers/nv40/nv40_state.c
+++ b/src/gallium/drivers/nv40/nv40_state.c
@@ -3,6 +3,8 @@
 #include "pipe/p_util.h"
 #include "pipe/p_inlines.h"
 
+#include "draw/draw_context.h"
+
 #include "nv40_context.h"
 #include "nv40_state.h"
 
@@ -345,7 +347,7 @@ nv40_rasterizer_state_create(struct pipe_context *pipe,
 			so_data(so, NV40TCL_CULL_FACE_FRONT_AND_BACK);
 			break;
 		default:
-			so_data(so, 0);
+			so_data(so, NV40TCL_CULL_FACE_BACK);
 			break;
 		}
 		so_data(so, NV40TCL_FRONT_FACE_CCW);
@@ -363,13 +365,13 @@ nv40_rasterizer_state_create(struct pipe_context *pipe,
 			so_data(so, NV40TCL_CULL_FACE_FRONT_AND_BACK);
 			break;
 		default:
-			so_data(so, 0);
+			so_data(so, NV40TCL_CULL_FACE_BACK);
 			break;
 		}
 		so_data(so, NV40TCL_FRONT_FACE_CW);
 	}
 	so_data(so, cso->poly_smooth ? 1 : 0);
-	so_data(so, cso->cull_mode != PIPE_WINDING_NONE ? 1 : 0);
+	so_data(so, (cso->cull_mode != PIPE_WINDING_NONE) ? 1 : 0);
 
 	so_method(so, curie, NV40TCL_POLYGON_STIPPLE_ENABLE, 1);
 	so_data  (so, cso->poly_stipple_enable ? 1 : 0);
@@ -419,6 +421,9 @@ static void
 nv40_rasterizer_state_bind(struct pipe_context *pipe, void *hwcso)
 {
 	struct nv40_context *nv40 = nv40_context(pipe);
+	struct nv40_rasterizer_state *rsso = hwcso;
+
+	draw_set_rasterizer_state(nv40->draw, &rsso->pipe);
 
 	nv40->rasterizer = hwcso;
 	nv40->dirty |= NV40_NEW_RAST;
@@ -508,10 +513,12 @@ static void *
 nv40_vp_state_create(struct pipe_context *pipe,
 		     const struct pipe_shader_state *cso)
 {
+	struct nv40_context *nv40 = nv40_context(pipe);
 	struct nv40_vertex_program *vp;
 
 	vp = CALLOC(1, sizeof(struct nv40_vertex_program));
 	vp->pipe = *cso;
+	vp->draw = draw_create_vertex_shader(nv40->draw, &vp->pipe);
 
 	return (void *)vp;
 }
@@ -520,6 +527,9 @@ static void
 nv40_vp_state_bind(struct pipe_context *pipe, void *hwcso)
 {
 	struct nv40_context *nv40 = nv40_context(pipe);
+	struct nv40_vertex_program *vp = hwcso;
+
+	draw_bind_vertex_shader(nv40->draw, vp ? vp->draw : NULL);
 
 	nv40->vertprog = hwcso;
 	nv40->dirty |= NV40_NEW_VERTPROG;
@@ -531,6 +541,7 @@ nv40_vp_state_delete(struct pipe_context *pipe, void *hwcso)
 	struct nv40_context *nv40 = nv40_context(pipe);
 	struct nv40_vertex_program *vp = hwcso;
 
+	draw_delete_vertex_shader(nv40->draw, vp->draw);
 	nv40_vertprog_destroy(nv40, vp);
 	FREE(vp);
 }
@@ -544,6 +555,8 @@ nv40_fp_state_create(struct pipe_context *pipe,
 	fp = CALLOC(1, sizeof(struct nv40_fragment_program));
 	fp->pipe = *cso;
 
+	tgsi_scan_shader(fp->pipe.tokens, &fp->info);
+
 	return (void *)fp;
 }
 
@@ -582,6 +595,8 @@ nv40_set_clip_state(struct pipe_context *pipe,
 {
 	struct nv40_context *nv40 = nv40_context(pipe);
 
+	draw_set_clip_state(nv40->draw, clip);
+
 	nv40->clip = *clip;
 	nv40->dirty |= NV40_NEW_UCP;
 }
@@ -638,6 +653,8 @@ nv40_set_viewport_state(struct pipe_context *pipe,
 {
 	struct nv40_context *nv40 = nv40_context(pipe);
 
+	draw_set_viewport_state(nv40->draw, vpt);
+
 	nv40->viewport = *vpt;
 	nv40->dirty |= NV40_NEW_VIEWPORT;
 }
@@ -648,6 +665,8 @@ nv40_set_vertex_buffer(struct pipe_context *pipe, unsigned index,
 {
 	struct nv40_context *nv40 = nv40_context(pipe);
 
+	draw_set_vertex_buffer(nv40->draw, index, vb);
+
 	nv40->vtxbuf[index] = *vb;
 	nv40->dirty |= NV40_NEW_ARRAYS;
 }
@@ -658,6 +677,8 @@ nv40_set_vertex_element(struct pipe_context *pipe, unsigned index,
 {
 	struct nv40_context *nv40 = nv40_context(pipe);
 
+	draw_set_vertex_element(nv40->draw, index, ve);
+
 	nv40->vtxelt[index] = *ve;
 	nv40->dirty |= NV40_NEW_ARRAYS;
 }
diff --git a/src/gallium/drivers/nv40/nv40_state.h b/src/gallium/drivers/nv40/nv40_state.h
index a02ea0c878..ab2866eb7a 100644
--- a/src/gallium/drivers/nv40/nv40_state.h
+++ b/src/gallium/drivers/nv40/nv40_state.h
@@ -2,6 +2,7 @@
 #define __NV40_STATE_H__
 
 #include "pipe/p_state.h"
+#include "tgsi/util/tgsi_scan.h"
 
 struct nv40_sampler_state {
 	uint32_t fmt;
@@ -25,6 +26,8 @@ struct nv40_vertex_program_data {
 struct nv40_vertex_program {
 	struct pipe_shader_state pipe;
 
+	struct draw_vertex_shader *draw;
+
 	boolean translated;
 	struct nv40_vertex_program_exec *insns;
 	unsigned nr_insns;
@@ -49,6 +52,7 @@ struct nv40_fragment_program_data {
 
 struct nv40_fragment_program {
 	struct pipe_shader_state pipe;
+	struct tgsi_shader_info info;
 
 	boolean translated;
 	unsigned samplers;
diff --git a/src/gallium/drivers/nv40/nv40_state_clip.c b/src/gallium/drivers/nv40/nv40_state_clip.c
index 93e690161f..c52390f9ed 100644
--- a/src/gallium/drivers/nv40/nv40_state_clip.c
+++ b/src/gallium/drivers/nv40/nv40_state_clip.c
@@ -3,8 +3,12 @@
 static boolean
 nv40_state_clip_validate(struct nv40_context *nv40)
 {
-	if (nv40->clip.nr)
-		nv40->fallback |= NV40_FALLBACK_TNL;
+
+	if (nv40->render_mode == HW) {
+		nv40->fallback_swtnl &= ~NV40_NEW_UCP;
+		if (nv40->clip.nr)
+			nv40->fallback_swtnl |= NV40_NEW_UCP;
+	}
 
 	return FALSE;
 }
diff --git a/src/gallium/drivers/nv40/nv40_state_emit.c b/src/gallium/drivers/nv40/nv40_state_emit.c
index 9f268640e0..056238cc83 100644
--- a/src/gallium/drivers/nv40/nv40_state_emit.c
+++ b/src/gallium/drivers/nv40/nv40_state_emit.c
@@ -1,5 +1,6 @@
 #include "nv40_context.h"
 #include "nv40_state.h"
+#include "draw/draw_context.h"
 
 static struct nv40_state_entry *render_states[] = {
 	&nv40_state_framebuffer,
@@ -18,15 +19,27 @@ static struct nv40_state_entry *render_states[] = {
 	NULL
 };
 
+static struct nv40_state_entry *swtnl_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_vtxfmt,
+	NULL
+};
+
 static void
-nv40_state_validate(struct nv40_context *nv40)
+nv40_state_do_validate(struct nv40_context *nv40,
+		       struct nv40_state_entry **states)
 {
-	struct nv40_state_entry **states = render_states;
-	unsigned last_fallback;
-
-	last_fallback = nv40->fallback;
-	nv40->fallback = 0;
-
 	while (*states) {
 		struct nv40_state_entry *e = *states;
 
@@ -38,32 +51,15 @@ nv40_state_validate(struct nv40_context *nv40)
 		states++;
 	}
 	nv40->dirty = 0;
-
-	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->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");
-	}
 }
 
-static void
+void
 nv40_state_emit(struct nv40_context *nv40)
 {
 	struct nv40_state *state = &nv40->state;
 	struct nv40_screen *screen = nv40->screen;
 	unsigned i, samplers;
+	uint64 states;
 
 	if (nv40->pctx_id != screen->cur_pctx) {
 		for (i = 0; i < NV40_STATE_MAX; i++) {
@@ -74,14 +70,24 @@ nv40_state_emit(struct nv40_context *nv40)
 		screen->cur_pctx = nv40->pctx_id;
 	}
 
-	while (state->dirty) {
-		unsigned idx = ffsll(state->dirty) - 1;
+	for (i = 0, states = state->dirty; states; i++) {
+		if (!(states & (1ULL << i)))
+			continue;
+		so_ref (state->hw[i], &nv40->screen->state[i]);
+		so_emit(nv40->nvws, nv40->screen->state[i]);
+		states &= ~(1ULL << i);
+	}
 
-		so_ref (state->hw[idx], &nv40->screen->state[idx]);
-		so_emit(nv40->nvws, nv40->screen->state[idx]);
-		state->dirty &= ~(1ULL << idx);
+	if (state->dirty & ((1ULL << NV40_STATE_FRAGPROG) |
+			    (1ULL << NV40_STATE_FRAGTEX0))) {
+		BEGIN_RING(curie, NV40TCL_TEX_CACHE_CTL, 1);
+		OUT_RING  (2);
+		BEGIN_RING(curie, NV40TCL_TEX_CACHE_CTL, 1);
+		OUT_RING  (1);
 	}
 
+	state->dirty = 0;
+
 	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)))
@@ -91,18 +97,62 @@ nv40_state_emit(struct nv40_context *nv40)
 		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]);
+	if (state->hw[NV40_STATE_VTXBUF] && nv40->render_mode == HW)
+		so_emit_reloc_markers(nv40->nvws, state->hw[NV40_STATE_VTXBUF]);
 }
 
-void
-nv40_emit_hw_state(struct nv40_context *nv40)
+boolean
+nv40_state_validate(struct nv40_context *nv40)
 {
-	nv40_state_validate(nv40);
-	nv40_state_emit(nv40);
+	boolean was_sw = nv40->fallback_swtnl ? TRUE : FALSE;
+
+	if (nv40->render_mode != HW) {
+		/* Don't even bother trying to go back to hw if none
+		 * of the states that caused swtnl previously have changed.
+		 */
+		if ((nv40->fallback_swtnl & nv40->dirty)
+				!= nv40->fallback_swtnl)
+			return FALSE;
+
+		/* Attempt to go to hwtnl again */
+		nv40->pipe.flush(&nv40->pipe, 0);
+		nv40->dirty |= (NV40_NEW_VIEWPORT |
+				NV40_NEW_VERTPROG |
+				NV40_NEW_ARRAYS |
+				NV40_NEW_UCP);
+		nv40->render_mode = HW;
+	}
+
+	nv40_state_do_validate(nv40, render_states);
+	if (nv40->fallback_swtnl || nv40->fallback_swrast)
+		return FALSE;
+	
+	if (was_sw)
+		NOUVEAU_ERR("swtnl->hw\n");
+
+	return TRUE;
+}
+
+boolean
+nv40_state_validate_swtnl(struct nv40_context *nv40)
+{
+	/* Setup for swtnl */
+	if (nv40->render_mode == HW) {
+		NOUVEAU_ERR("hw->swtnl 0x%08x\n", nv40->fallback_swtnl);
+		nv40->pipe.flush(&nv40->pipe, 0);
+		nv40->dirty |= (NV40_NEW_VIEWPORT |
+				NV40_NEW_VERTPROG |
+				NV40_NEW_ARRAYS |
+				NV40_NEW_UCP);
+		nv40->render_mode = SWTNL;
+	}
+
+	nv40_state_do_validate(nv40, swtnl_states);
+	if (nv40->fallback_swrast) {
+		NOUVEAU_ERR("swtnl->swrast 0x%08x\n", nv40->fallback_swrast);
+		return FALSE;
+	}
 
-	BEGIN_RING(curie, NV40TCL_TEX_CACHE_CTL, 1);
-	OUT_RING  (2);
-	BEGIN_RING(curie, NV40TCL_TEX_CACHE_CTL, 1);
-	OUT_RING  (1);
+	return TRUE;
 }
 
diff --git a/src/gallium/drivers/nv40/nv40_state_viewport.c b/src/gallium/drivers/nv40/nv40_state_viewport.c
index 3a32533907..9e5c7a72a7 100644
--- a/src/gallium/drivers/nv40/nv40_state_viewport.c
+++ b/src/gallium/drivers/nv40/nv40_state_viewport.c
@@ -3,18 +3,43 @@
 static boolean
 nv40_state_viewport_validate(struct nv40_context *nv40)
 {
-	struct nouveau_stateobj *so = so_new(9, 0);
+	struct nouveau_stateobj *so = so_new(11, 0);
 	struct pipe_viewport_state *vpt = &nv40->viewport;
 
-	so_method(so, nv40->screen->curie, NV40TCL_VIEWPORT_TRANSLATE_X, 8);
-	so_data  (so, fui(vpt->translate[0]));
-	so_data  (so, fui(vpt->translate[1]));
-	so_data  (so, fui(vpt->translate[2]));
-	so_data  (so, fui(vpt->translate[3]));
-	so_data  (so, fui(vpt->scale[0]));
-	so_data  (so, fui(vpt->scale[1]));
-	so_data  (so, fui(vpt->scale[2]));
-	so_data  (so, fui(vpt->scale[3]));
+	if (nv40->render_mode == HW) {
+		so_method(so, nv40->screen->curie,
+			  NV40TCL_VIEWPORT_TRANSLATE_X, 8);
+		so_data  (so, fui(vpt->translate[0]));
+		so_data  (so, fui(vpt->translate[1]));
+		so_data  (so, fui(vpt->translate[2]));
+		so_data  (so, fui(vpt->translate[3]));
+		so_data  (so, fui(vpt->scale[0]));
+		so_data  (so, fui(vpt->scale[1]));
+		so_data  (so, fui(vpt->scale[2]));
+		so_data  (so, fui(vpt->scale[3]));
+		so_method(so, nv40->screen->curie, 0x1d78, 1);
+		so_data  (so, 1);
+	} else {
+		so_method(so, nv40->screen->curie,
+			  NV40TCL_VIEWPORT_TRANSLATE_X, 8);
+		so_data  (so, fui(0.0));
+		so_data  (so, fui(0.0));
+		so_data  (so, fui(0.0));
+		so_data  (so, fui(0.0));
+		so_data  (so, fui(1.0));
+		so_data  (so, fui(1.0));
+		so_data  (so, fui(1.0));
+		so_data  (so, fui(0.0));
+		/* Not entirely certain what this is yet.  The DDX uses this
+		 * value also as it fixes rendering when you pass
+		 * pre-transformed vertices to the GPU.  My best gusss is that
+		 * this bypasses some culling/clipping stage.  Might be worth
+		 * noting that points/lines are uneffected by whatever this
+		 * value fixes, only filled polygons are effected.
+		 */
+		so_method(so, nv40->screen->curie, 0x1d78, 1);
+		so_data  (so, 0x110);
+	}
 
 	so_ref(so, &nv40->state.hw[NV40_STATE_VIEWPORT]);
 	return TRUE;
diff --git a/src/gallium/drivers/nv40/nv40_vbo.c b/src/gallium/drivers/nv40/nv40_vbo.c
index f16afc23b8..fad423fdf8 100644
--- a/src/gallium/drivers/nv40/nv40_vbo.c
+++ b/src/gallium/drivers/nv40/nv40_vbo.c
@@ -8,6 +8,8 @@
 #include "nouveau/nouveau_channel.h"
 #include "nouveau/nouveau_pushbuf.h"
 
+#define FORCE_SWTNL 0
+
 static INLINE int
 nv40_vbo_format_to_hw(enum pipe_format pipe, unsigned *fmt, unsigned *ncomp)
 {
@@ -165,7 +167,11 @@ nv40_draw_arrays(struct pipe_context *pipe, unsigned mode, unsigned start,
 	unsigned nr;
 
 	nv40_vbo_set_idxbuf(nv40, NULL, 0);
-	nv40_emit_hw_state(nv40);
+	if (FORCE_SWTNL || !nv40_state_validate(nv40)) {
+		return nv40_draw_elements_swtnl(pipe, NULL, 0,
+						mode, start, count);
+	}
+	nv40_state_emit(nv40);
 
 	BEGIN_RING(curie, NV40TCL_BEGIN_END, 1);
 	OUT_RING  (nvgl_primitive(mode));
@@ -274,7 +280,7 @@ nv40_draw_elements_inline(struct pipe_context *pipe,
 	struct pipe_winsys *ws = pipe->winsys;
 	void *map;
 
-	nv40_emit_hw_state(nv40);
+	nv40_state_emit(nv40);
 
 	map = ws->buffer_map(ws, ib, PIPE_BUFFER_USAGE_CPU_READ);
 	if (!ib) {
@@ -315,7 +321,7 @@ nv40_draw_elements_vbo(struct pipe_context *pipe,
 	struct nv40_context *nv40 = nv40_context(pipe);
 	unsigned nr;
 
-	nv40_emit_hw_state(nv40);
+	nv40_state_emit(nv40);
 
 	BEGIN_RING(curie, NV40TCL_BEGIN_END, 1);
 	OUT_RING  (nvgl_primitive(mode));
@@ -352,8 +358,16 @@ nv40_draw_elements(struct pipe_context *pipe,
 		   unsigned mode, unsigned start, unsigned count)
 {
 	struct nv40_context *nv40 = nv40_context(pipe);
+	boolean idxbuf;
+
+	idxbuf = nv40_vbo_set_idxbuf(nv40, indexBuffer, indexSize);
+	if (FORCE_SWTNL || !nv40_state_validate(nv40)) {
+		return nv40_draw_elements_swtnl(pipe, NULL, 0,
+						mode, start, count);
+	}
+	nv40_state_emit(nv40);
 
-	if (nv40_vbo_set_idxbuf(nv40, indexBuffer, indexSize)) {
+	if (idxbuf) {
 		nv40_draw_elements_vbo(pipe, mode, start, count);
 	} else {
 		nv40_draw_elements_inline(pipe, indexBuffer, indexSize,
diff --git a/src/gallium/drivers/nv40/nv40_vertprog.c b/src/gallium/drivers/nv40/nv40_vertprog.c
index 3d730c1a32..9f1ee575ce 100644
--- a/src/gallium/drivers/nv40/nv40_vertprog.c
+++ b/src/gallium/drivers/nv40/nv40_vertprog.c
@@ -634,21 +634,29 @@ out_err:
 static boolean
 nv40_vertprog_validate(struct nv40_context *nv40)
 { 
-	struct nv40_vertex_program *vp = nv40->vertprog;
-	struct pipe_buffer *constbuf =
-		nv40->constbuf[PIPE_SHADER_VERTEX];
 	struct nouveau_winsys *nvws = nv40->nvws;
 	struct pipe_winsys *ws = nv40->pipe.winsys;
+	struct nv40_vertex_program *vp;
+	struct pipe_buffer *constbuf;
 	boolean upload_code = FALSE, upload_data = FALSE;
 	int i;
 
+	if (nv40->render_mode == HW) {
+		vp = nv40->vertprog;
+		constbuf = nv40->constbuf[PIPE_SHADER_VERTEX];
+	} else {
+		vp = nv40->swtnl.vertprog;
+		constbuf = NULL;
+	}
+
 	/* Translate TGSI shader into hw bytecode */
 	if (vp->translated)
 		goto check_gpu_resources;
 
+	nv40->fallback_swtnl &= ~NV40_NEW_VERTPROG;
 	nv40_vertprog_translate(nv40, vp);
 	if (!vp->translated) {
-		nv40->fallback |= NV40_FALLBACK_TNL;
+		nv40->fallback_swtnl |= NV40_NEW_VERTPROG;
 		return FALSE;
 	}