diff --git a/src/mesa/drivers/dri/nouveau/Makefile b/src/mesa/drivers/dri/nouveau/Makefile
index 49e8933..7be19b2 100644
--- a/src/mesa/drivers/dri/nouveau/Makefile
+++ b/src/mesa/drivers/dri/nouveau/Makefile
@@ -8,8 +8,6 @@ DRI_LIB_DEPS += $(shell pkg-config libdrm_nouveau --libs)
 
 LIBNAME = nouveau_vieux_dri.so
 
-MINIGLX_SOURCES =
-
 DRIVER_SOURCES = \
 	nouveau_screen.c \
 	nouveau_context.c \
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_context.c b/src/mesa/drivers/dri/nouveau/nouveau_context.c
index 52185a2..42bec65 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_context.c
+++ b/src/mesa/drivers/dri/nouveau/nouveau_context.c
@@ -43,19 +43,23 @@
 
 #define need_GL_EXT_framebuffer_object
 #define need_GL_EXT_fog_coord
+#define need_GL_EXT_secondary_color
 
 #include "main/remap_helper.h"
 
 static const struct dri_extension nouveau_extensions[] = {
 	{ "GL_ARB_multitexture",	NULL },
+	{ "GL_ARB_texture_env_add",	NULL },
 	{ "GL_ARB_texture_env_combine",	NULL },
 	{ "GL_ARB_texture_env_dot3",	NULL },
-	{ "GL_ARB_texture_env_add",	NULL },
-	{ "GL_EXT_texture_lod_bias",	NULL },
-	{ "GL_EXT_framebuffer_object",	GL_EXT_framebuffer_object_functions },
 	{ "GL_ARB_texture_mirrored_repeat", NULL },
-	{ "GL_EXT_stencil_wrap",	NULL },
 	{ "GL_EXT_fog_coord",		GL_EXT_fog_coord_functions },
+	{ "GL_EXT_framebuffer_blit",	NULL },
+	{ "GL_EXT_framebuffer_object",	GL_EXT_framebuffer_object_functions },
+	{ "GL_EXT_secondary_color",	GL_EXT_secondary_color_functions },
+	{ "GL_EXT_stencil_wrap",	NULL },
+	{ "GL_EXT_texture_lod_bias",	NULL },
+	{ "GL_NV_blend_square",         NULL },
 	{ "GL_SGIS_generate_mipmap",	NULL },
 	{ NULL,				NULL }
 };
@@ -66,8 +70,8 @@ nouveau_channel_flush_notify(struct nouveau_channel *chan)
 	struct nouveau_context *nctx = chan->user_private;
 	GLcontext *ctx = &nctx->base;
 
-	if (nctx->fallback < SWRAST && ctx->DrawBuffer)
-		nouveau_state_emit(&nctx->base);
+	if (nctx->fallback < SWRAST)
+		nouveau_bo_state_emit(ctx);
 }
 
 GLboolean
@@ -334,6 +338,8 @@ nouveau_validate_framebuffer(GLcontext *ctx)
 		update_framebuffer(dri_ctx, dri_read,
 				   &dri_ctx->dri2.read_stamp);
 
-	if (nouveau_next_dirty_state(ctx) >= 0)
+	if (nouveau_next_dirty_state(ctx) >= 0) {
+		nouveau_state_emit(ctx);
 		FIRE_RING(context_chan(ctx));
+	}
 }
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_context.h b/src/mesa/drivers/dri/nouveau/nouveau_context.h
index 682f8a4..fe64fec 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_context.h
+++ b/src/mesa/drivers/dri/nouveau/nouveau_context.h
@@ -85,6 +85,8 @@ struct nouveau_context {
 	BITSET_SET(to_nouveau_context(ctx)->dirty, NOUVEAU_STATE_##s)
 #define context_dirty_i(ctx, s, i) \
 	BITSET_SET(to_nouveau_context(ctx)->dirty, NOUVEAU_STATE_##s##0 + i)
+#define context_emit(ctx, s) \
+	context_drv(ctx)->emit[NOUVEAU_STATE_##s](ctx, NOUVEAU_STATE_##s)
 
 GLboolean
 nouveau_context_create(const __GLcontextModes *visual, __DRIcontext *dri_ctx,
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_driver.c b/src/mesa/drivers/dri/nouveau/nouveau_driver.c
index 1d12f43..4ec864c 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_driver.c
+++ b/src/mesa/drivers/dri/nouveau/nouveau_driver.c
@@ -135,4 +135,8 @@ nouveau_driver_functions_init(struct dd_function_table *functions)
 	functions->Flush = nouveau_flush;
 	functions->Finish = nouveau_finish;
 	functions->Clear = nouveau_clear;
+	functions->DrawPixels = _mesa_meta_DrawPixels;
+	functions->CopyPixels = _mesa_meta_CopyPixels;
+	functions->Bitmap = _mesa_meta_Bitmap;
+	functions->BlitFramebuffer = _mesa_meta_BlitFramebuffer;
 }
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_fbo.c b/src/mesa/drivers/dri/nouveau/nouveau_fbo.c
index 2ec3dc9..8be7edb 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_fbo.c
+++ b/src/mesa/drivers/dri/nouveau/nouveau_fbo.c
@@ -236,7 +236,7 @@ nouveau_render_texture(GLcontext *ctx, struct gl_framebuffer *fb,
 	/* Allocate a renderbuffer object for the texture if we
 	 * haven't already done so. */
 	if (!rb) {
-		rb = nouveau_renderbuffer_new(ctx, 0);
+		rb = nouveau_renderbuffer_new(ctx, ~0);
 		assert(rb);
 
 		rb->AllocStorage = NULL;
@@ -259,11 +259,7 @@ static void
 nouveau_finish_render_texture(GLcontext *ctx,
 			      struct gl_renderbuffer_attachment *att)
 {
-	struct nouveau_renderbuffer *nrb
-		= to_nouveau_renderbuffer(att->Renderbuffer);
-
 	texture_dirty(att->Texture);
-	nouveau_surface_ref(NULL, &nrb->surface);
 }
 
 void
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_gldefs.h b/src/mesa/drivers/dri/nouveau/nouveau_gldefs.h
index 00007a9..fbeed3b 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_gldefs.h
+++ b/src/mesa/drivers/dri/nouveau/nouveau_gldefs.h
@@ -260,4 +260,23 @@ nvgl_filter_mode(unsigned filter)
 	}
 }
 
+static inline unsigned
+nvgl_texgen_mode(unsigned mode)
+{
+	switch (mode) {
+	case GL_EYE_LINEAR:
+		return 0x2400;
+	case GL_OBJECT_LINEAR:
+		return 0x2401;
+	case GL_SPHERE_MAP:
+		return 0x2402;
+	case GL_NORMAL_MAP:
+		return 0x8511;
+	case GL_REFLECTION_MAP:
+		return 0x8512;
+	default:
+		assert(0);
+	}
+}
+
 #endif
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_render.h b/src/mesa/drivers/dri/nouveau/nouveau_render.h
index bff0ccf..923b79b 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_render.h
+++ b/src/mesa/drivers/dri/nouveau/nouveau_render.h
@@ -32,8 +32,8 @@
 struct nouveau_array_state;
 
 typedef void (*dispatch_t)(GLcontext *, unsigned int, int, unsigned int);
-typedef unsigned (*extract_u_t)(struct nouveau_array_state *a, int i, int j);
-typedef float (*extract_f_t)(struct nouveau_array_state *a, int i, int j);
+typedef unsigned (*extract_u_t)(struct nouveau_array_state *, int, int);
+typedef float (*extract_f_t)(struct nouveau_array_state *, int, int);
 
 struct nouveau_attr_info {
 	int vbo_index;
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_render_t.c b/src/mesa/drivers/dri/nouveau/nouveau_render_t.c
index c050578..7ccd7e6 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_render_t.c
+++ b/src/mesa/drivers/dri/nouveau/nouveau_render_t.c
@@ -254,7 +254,7 @@ get_scratch_vbo(GLcontext *ctx, unsigned size, struct nouveau_bo **bo,
  */
 static inline unsigned
 get_max_vertices(GLcontext *ctx, const struct _mesa_index_buffer *ib,
-		 unsigned n)
+		 int n)
 {
 	struct nouveau_render_state *render = to_render_state(ctx);
 
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_span.c b/src/mesa/drivers/dri/nouveau/nouveau_span.c
index f1a56dd..1bfdecc 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_span.c
+++ b/src/mesa/drivers/dri/nouveau/nouveau_span.c
@@ -32,7 +32,6 @@
 #include "swrast/swrast.h"
 
 #define LOCAL_VARS							\
-	struct gl_framebuffer *fb = ctx->DrawBuffer;			\
 	struct nouveau_surface *s = &to_nouveau_renderbuffer(rb)->surface; \
 	GLuint p;							\
 	(void)p;
@@ -45,12 +44,12 @@
 #define HW_CLIPLOOP() {							\
 	int minx = 0;							\
 	int miny = 0;							\
-	int maxx = fb->Width;						\
-	int maxy = fb->Height;
+	int maxx = rb->Width;						\
+	int maxy = rb->Height;
 
 #define HW_ENDCLIPLOOP() }
 
-#define Y_FLIP(y) (fb->Name ? (y) : rb->Height - 1 - (y))
+#define Y_FLIP(y) (rb->Name ? (y) : rb->Height - 1 - (y))
 
 /* RGB565 span functions */
 #define SPANTMP_PIXEL_FMT GL_RGB
@@ -144,17 +143,28 @@ texture_unit_map_unmap(GLcontext *ctx, struct gl_texture_unit *u, GLboolean map)
 }
 
 static void
-span_map_unmap(GLcontext *ctx, GLboolean map)
+framebuffer_map_unmap(struct gl_framebuffer *fb, GLboolean map)
 {
 	int i;
 
-	for (i = 0; i < ctx->DrawBuffer->_NumColorDrawBuffers; i++)
-		renderbuffer_map_unmap(ctx->DrawBuffer->_ColorDrawBuffers[i], map);
+	for (i = 0; i < fb->_NumColorDrawBuffers; i++)
+		renderbuffer_map_unmap(fb->_ColorDrawBuffers[i], map);
+
+	renderbuffer_map_unmap(fb->_ColorReadBuffer, map);
+
+	if (fb->_DepthBuffer)
+		renderbuffer_map_unmap(fb->_DepthBuffer->Wrapped, map);
+}
+
+static void
+span_map_unmap(GLcontext *ctx, GLboolean map)
+{
+	int i;
 
-	renderbuffer_map_unmap(ctx->DrawBuffer->_ColorReadBuffer, map);
+	framebuffer_map_unmap(ctx->DrawBuffer, map);
 
-	if (ctx->DrawBuffer->_DepthBuffer)
-		renderbuffer_map_unmap(ctx->DrawBuffer->_DepthBuffer->Wrapped, map);
+	if (ctx->ReadBuffer != ctx->DrawBuffer)
+		framebuffer_map_unmap(ctx->ReadBuffer, map);
 
 	for (i = 0; i < ctx->Const.MaxTextureUnits; i++)
 		texture_unit_map_unmap(ctx, &ctx->Texture.Unit[i], map);
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_state.c b/src/mesa/drivers/dri/nouveau/nouveau_state.c
index e1871db..a57df2d 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_state.c
+++ b/src/mesa/drivers/dri/nouveau/nouveau_state.c
@@ -150,6 +150,7 @@ nouveau_enable(GLcontext *ctx, GLenum cap, GLboolean state)
 		break;
 	case GL_COLOR_SUM_EXT:
 		context_dirty(ctx, FRAG);
+		context_dirty(ctx, LIGHT_MODEL);
 		break;
 	case GL_CULL_FACE:
 		context_dirty(ctx, CULL_FACE);
@@ -188,6 +189,7 @@ nouveau_enable(GLcontext *ctx, GLenum cap, GLboolean state)
 	case GL_LIGHTING:
 		context_dirty(ctx, FRAG);
 		context_dirty(ctx, MODELVIEW);
+		context_dirty(ctx, LIGHT_MODEL);
 		context_dirty(ctx, LIGHT_ENABLE);
 
 		for (i = 0; i < MAX_LIGHTS; i++) {
@@ -230,9 +232,17 @@ nouveau_enable(GLcontext *ctx, GLenum cap, GLboolean state)
 	case GL_TEXTURE_1D:
 	case GL_TEXTURE_2D:
 	case GL_TEXTURE_3D:
+	case GL_TEXTURE_RECTANGLE:
 		context_dirty_i(ctx, TEX_ENV, ctx->Texture.CurrentUnit);
 		context_dirty_i(ctx, TEX_OBJ, ctx->Texture.CurrentUnit);
 		break;
+	case GL_TEXTURE_GEN_S:
+	case GL_TEXTURE_GEN_T:
+	case GL_TEXTURE_GEN_R:
+	case GL_TEXTURE_GEN_Q:
+		context_dirty_i(ctx, TEX_GEN, ctx->Texture.CurrentUnit);
+		context_dirty(ctx, MODELVIEW);
+		break;
 	}
 }
 
@@ -367,7 +377,15 @@ static void
 nouveau_tex_gen(GLcontext *ctx, GLenum coord, GLenum pname,
 		const GLfloat *params)
 {
-	context_dirty_i(ctx, TEX_GEN, ctx->Texture.CurrentUnit);
+	switch (pname) {
+	case GL_TEXTURE_GEN_MODE:
+		context_dirty_i(ctx, TEX_GEN, ctx->Texture.CurrentUnit);
+		context_dirty(ctx, MODELVIEW);
+		break;
+	default:
+		context_dirty_i(ctx, TEX_GEN, ctx->Texture.CurrentUnit);
+		break;
+	}
 }
 
 static void
@@ -453,12 +471,19 @@ nouveau_state_emit(GLcontext *ctx)
 static void
 nouveau_update_state(GLcontext *ctx, GLbitfield new_state)
 {
+	int i;
+
 	if (new_state & (_NEW_PROJECTION | _NEW_MODELVIEW))
 		context_dirty(ctx, PROJECTION);
 
 	if (new_state & _NEW_MODELVIEW)
 		context_dirty(ctx, MODELVIEW);
 
+	if (new_state & _NEW_TEXTURE_MATRIX) {
+		for (i = 0; i < ctx->Const.MaxTextureCoordUnits; i++)
+			context_dirty_i(ctx, TEX_MAT, i);
+	}
+
 	if (new_state & _NEW_CURRENT_ATTRIB &&
 	    new_state & _NEW_LIGHT) {
 		context_dirty(ctx, MATERIAL_FRONT_AMBIENT);
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_state.h b/src/mesa/drivers/dri/nouveau/nouveau_state.h
index d01d962..38ac975 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_state.h
+++ b/src/mesa/drivers/dri/nouveau/nouveau_state.h
@@ -89,6 +89,10 @@ enum {
 	NOUVEAU_STATE_TEX_GEN1,
 	NOUVEAU_STATE_TEX_GEN2,
 	NOUVEAU_STATE_TEX_GEN3,
+	NOUVEAU_STATE_TEX_MAT0,
+	NOUVEAU_STATE_TEX_MAT1,
+	NOUVEAU_STATE_TEX_MAT2,
+	NOUVEAU_STATE_TEX_MAT3,
 	NOUVEAU_STATE_TEX_OBJ0,
 	NOUVEAU_STATE_TEX_OBJ1,
 	NOUVEAU_STATE_TEX_OBJ2,
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_texture.c b/src/mesa/drivers/dri/nouveau/nouveau_texture.c
index bf365bf..dbf9a5c 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_texture.c
+++ b/src/mesa/drivers/dri/nouveau/nouveau_texture.c
@@ -177,15 +177,15 @@ nouveau_choose_tex_format(GLcontext *ctx, GLint internalFormat,
 }
 
 static GLboolean
-teximage_fits(struct gl_texture_object *t, int level,
-	      struct gl_texture_image *ti)
+teximage_fits(struct gl_texture_object *t, int level)
 {
 	struct nouveau_surface *s = &to_nouveau_texture(t)->surfaces[level];
+	struct gl_texture_image *ti = t->Image[0][level];
 
-	return t->Target == GL_TEXTURE_RECTANGLE ||
-		(s->bo && s->width == ti->Width &&
-		 s->height == ti->Height &&
-		 s->format == ti->TexFormat);
+	return ti && (t->Target == GL_TEXTURE_RECTANGLE ||
+		      (s->bo && s->width == ti->Width &&
+		       s->height == ti->Height &&
+		       s->format == ti->TexFormat));
 }
 
 static GLboolean
@@ -195,7 +195,7 @@ validate_teximage(GLcontext *ctx, struct gl_texture_object *t,
 {
 	struct gl_texture_image *ti = t->Image[0][level];
 
-	if (ti && teximage_fits(t, level, ti)) {
+	if (teximage_fits(t, level)) {
 		struct nouveau_surface *ss = to_nouveau_texture(t)->surfaces;
 		struct nouveau_surface *s = &to_nouveau_teximage(ti)->surface;
 
@@ -283,7 +283,8 @@ nouveau_texture_validate(GLcontext *ctx, struct gl_texture_object *t)
 	struct nouveau_texture *nt = to_nouveau_texture(t);
 	int i, last = get_last_level(t);
 
-	if (!nt->surfaces[last].bo)
+	if (!teximage_fits(t, t->BaseLevel) ||
+	    !teximage_fits(t, last))
 		return GL_FALSE;
 
 	if (nt->dirty) {
@@ -296,6 +297,8 @@ nouveau_texture_validate(GLcontext *ctx, struct gl_texture_object *t)
 			validate_teximage(ctx, t, i, 0, 0, 0,
 					  s->width, s->height, 1);
 		}
+
+		FIRE_RING(context_chan(ctx));
 	}
 
 	return GL_TRUE;
@@ -304,9 +307,12 @@ nouveau_texture_validate(GLcontext *ctx, struct gl_texture_object *t)
 void
 nouveau_texture_reallocate(GLcontext *ctx, struct gl_texture_object *t)
 {
-	texture_dirty(t);
-	relayout_texture(ctx, t);
-	nouveau_texture_validate(ctx, t);
+	if (!teximage_fits(t, t->BaseLevel) ||
+	    !teximage_fits(t, get_last_level(t))) {
+		texture_dirty(t);
+		relayout_texture(ctx, t);
+		nouveau_texture_validate(ctx, t);
+	}
 }
 
 static unsigned
@@ -364,7 +370,7 @@ nouveau_teximage(GLcontext *ctx, GLint dims, GLenum target, GLint level,
 	}
 
 	if (level == t->BaseLevel) {
-		if (!teximage_fits(t, level, ti))
+		if (!teximage_fits(t, level))
 			relayout_texture(ctx, t);
 		nouveau_texture_validate(ctx, t);
 	}
@@ -416,6 +422,40 @@ nouveau_teximage_3d(GLcontext *ctx, GLenum target, GLint level,
 }
 
 static void
+nouveau_texsubimage(GLcontext *ctx, GLint dims, GLenum target, GLint level,
+		    GLint xoffset, GLint yoffset, GLint zoffset,
+		    GLint width, GLint height, GLint depth,
+		    GLenum format, GLenum type, const void *pixels,
+		    const struct gl_pixelstore_attrib *packing,
+		    struct gl_texture_object *t,
+		    struct gl_texture_image *ti)
+{
+	struct nouveau_surface *s = &to_nouveau_teximage(ti)->surface;
+	int ret;
+
+	pixels = _mesa_validate_pbo_teximage(ctx, dims, width, height, depth,
+					     format, type, pixels, packing,
+					     "glTexSubImage");
+	if (pixels) {
+		nouveau_teximage_map(ctx, ti);
+
+		ret = _mesa_texstore(ctx, 3, ti->_BaseFormat, ti->TexFormat,
+				     ti->Data, xoffset, yoffset, zoffset,
+				     s->pitch, ti->ImageOffsets,
+				     width, height, depth, format, type,
+				     pixels, packing);
+		assert(ret);
+
+		nouveau_teximage_unmap(ctx, ti);
+		_mesa_unmap_teximage_pbo(ctx, packing);
+	}
+
+	if (!to_nouveau_texture(t)->dirty)
+		validate_teximage(ctx, t, level, xoffset, yoffset, zoffset,
+				  width, height, depth);
+}
+
+static void
 nouveau_texsubimage_3d(GLcontext *ctx, GLenum target, GLint level,
 		       GLint xoffset, GLint yoffset, GLint zoffset,
 		       GLint width, GLint height, GLint depth,
@@ -424,15 +464,9 @@ nouveau_texsubimage_3d(GLcontext *ctx, GLenum target, GLint level,
 		       struct gl_texture_object *t,
 		       struct gl_texture_image *ti)
 {
-	nouveau_teximage_map(ctx, ti);
-	_mesa_store_texsubimage3d(ctx, target, level, xoffset, yoffset, zoffset,
-				  width, height, depth, format, type, pixels,
-				  packing, t, ti);
-	nouveau_teximage_unmap(ctx, ti);
-
-	if (!to_nouveau_texture(t)->dirty)
-		validate_teximage(ctx, t, level, xoffset, yoffset, zoffset,
-				  width, height, depth);
+	nouveau_texsubimage(ctx, 3, target, level, xoffset, yoffset, zoffset,
+			    width, height, depth, format, type, pixels,
+			    packing, t, ti);
 }
 
 static void
@@ -444,15 +478,9 @@ nouveau_texsubimage_2d(GLcontext *ctx, GLenum target, GLint level,
 		       struct gl_texture_object *t,
 		       struct gl_texture_image *ti)
 {
-	nouveau_teximage_map(ctx, ti);
-	_mesa_store_texsubimage2d(ctx, target, level, xoffset, yoffset,
-				  width, height, format, type, pixels,
-				  packing, t, ti);
-	nouveau_teximage_unmap(ctx, ti);
-
-	if (!to_nouveau_texture(t)->dirty)
-		validate_teximage(ctx, t, level, xoffset, yoffset, 0,
-				  width, height, 1);
+	nouveau_texsubimage(ctx, 2, target, level, xoffset, yoffset, 0,
+			    width, height, 1, format, type, pixels,
+			    packing, t, ti);
 }
 
 static void
@@ -463,15 +491,9 @@ nouveau_texsubimage_1d(GLcontext *ctx, GLenum target, GLint level,
 		       struct gl_texture_object *t,
 		       struct gl_texture_image *ti)
 {
-	nouveau_teximage_map(ctx, ti);
-	_mesa_store_texsubimage1d(ctx, target, level, xoffset,
-				  width, format, type, pixels,
-				  packing, t, ti);
-	nouveau_teximage_unmap(ctx, ti);
-
-	if (!to_nouveau_texture(t)->dirty)
-		validate_teximage(ctx, t, level, xoffset, 0, 0,
-				  width, 1, 1);
+	nouveau_texsubimage(ctx, 1, target, level, xoffset, 0, 0,
+			    width, 1, 1, format, type, pixels,
+			    packing, t, ti);
 }
 
 static void
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_texture.h b/src/mesa/drivers/dri/nouveau/nouveau_texture.h
index b91facb..251f537 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_texture.h
+++ b/src/mesa/drivers/dri/nouveau/nouveau_texture.h
@@ -41,7 +41,7 @@ struct nouveau_texture {
 #define to_nouveau_texture(x) ((struct nouveau_texture *)(x))
 
 #define texture_dirty(t) \
-	to_nouveau_texture(t)->dirty = GL_TRUE;
+	to_nouveau_texture(t)->dirty = GL_TRUE
 
 void
 nouveau_set_texbuffer(__DRIcontext *dri_ctx,
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_util.h b/src/mesa/drivers/dri/nouveau/nouveau_util.h
index d6007ab..584cb80 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_util.h
+++ b/src/mesa/drivers/dri/nouveau/nouveau_util.h
@@ -191,4 +191,22 @@ is_texture_source(int s)
 	return s == GL_TEXTURE || (s >= GL_TEXTURE0 && s <= GL_TEXTURE31);
 }
 
+static inline struct gl_texgen *
+get_texgen_coord(struct gl_texture_unit *u, int i)
+{
+	return ((struct gl_texgen *[])
+		{ &u->GenS, &u->GenT, &u->GenR, &u->GenQ }) [i];
+}
+
+static inline float *
+get_texgen_coeff(struct gl_texgen *c)
+{
+	if (c->Mode == GL_OBJECT_LINEAR)
+		return c->ObjectPlane;
+	else if (c->Mode == GL_EYE_LINEAR)
+		return c->EyePlane;
+	else
+		return NULL;
+}
+
 #endif
diff --git a/src/mesa/drivers/dri/nouveau/nouveau_vbo_t.c b/src/mesa/drivers/dri/nouveau/nouveau_vbo_t.c
index 69a9b96..e5858f8 100644
--- a/src/mesa/drivers/dri/nouveau/nouveau_vbo_t.c
+++ b/src/mesa/drivers/dri/nouveau/nouveau_vbo_t.c
@@ -85,6 +85,18 @@ vbo_deinit_array(struct nouveau_array_state *a)
 	a->fields = 0;
 }
 
+static int
+get_array_stride(GLcontext *ctx, const struct gl_client_array *a)
+{
+	struct nouveau_render_state *render = to_render_state(ctx);
+
+	if (render->mode == VBO && !_mesa_is_bufferobj(a->BufferObj))
+		/* Pack client buffers. */
+		return align(_mesa_sizeof_type(a->Type) * a->Size, 4);
+	else
+		return a->StrideB;
+}
+
 static void
 vbo_init_arrays(GLcontext *ctx, const struct _mesa_index_buffer *ib,
 		const struct gl_client_array **arrays)
@@ -101,18 +113,10 @@ vbo_init_arrays(GLcontext *ctx, const struct _mesa_index_buffer *ib,
 
 		if (attr >= 0) {
 			const struct gl_client_array *array = arrays[attr];
-			int stride;
-
-			if (render->mode == VBO &&
-			    !_mesa_is_bufferobj(array->BufferObj))
-				/* Pack client buffers. */
-				stride = align(_mesa_sizeof_type(array->Type)
-					       * array->Size, 4);
-			else
-				stride = array->StrideB;
 
 			vbo_init_array(&render->attrs[attr], attr,
-				       stride, array->Size, array->Type,
+				       get_array_stride(ctx, array),
+				       array->Size, array->Type,
 				       array->BufferObj, array->Ptr,
 				       render->mode == IMM);
 		}
@@ -224,9 +228,11 @@ vbo_choose_attrs(GLcontext *ctx, const struct gl_client_array **arrays)
 	if (ctx->Fog.Enabled && ctx->Fog.FogCoordinateSource == GL_FOG_COORD)
 		vbo_emit_attr(ctx, arrays, VERT_ATTRIB_FOG);
 
-	if (ctx->Light.Enabled) {
+	if (ctx->Light.Enabled ||
+	    (ctx->Texture._GenFlags & TEXGEN_NEED_NORMALS))
 		vbo_emit_attr(ctx, arrays, VERT_ATTRIB_NORMAL);
 
+	if (ctx->Light.Enabled) {
 		vbo_emit_attr(ctx, arrays, MAT(FRONT_AMBIENT));
 		vbo_emit_attr(ctx, arrays, MAT(FRONT_DIFFUSE));
 		vbo_emit_attr(ctx, arrays, MAT(FRONT_SPECULAR));
@@ -243,18 +249,21 @@ vbo_choose_attrs(GLcontext *ctx, const struct gl_client_array **arrays)
 	vbo_emit_attr(ctx, arrays, VERT_ATTRIB_POS);
 }
 
-static unsigned
-get_max_client_stride(GLcontext *ctx)
+static int
+get_max_client_stride(GLcontext *ctx, const struct gl_client_array **arrays)
 {
 	struct nouveau_render_state *render = to_render_state(ctx);
 	int i, s = 0;
 
 	for (i = 0; i < render->attr_count; i++) {
 		int attr = render->map[i];
-		struct nouveau_array_state *a = &render->attrs[attr];
 
-		if (attr >= 0 && !a->bo)
-			s = MAX2(a->stride, s);
+		if (attr >= 0) {
+			const struct gl_client_array *a = arrays[attr];
+
+			if (!_mesa_is_bufferobj(a->BufferObj))
+				s = MAX2(s, get_array_stride(ctx, a));
+		}
 	}
 
 	return s;
@@ -275,20 +284,20 @@ vbo_maybe_split(GLcontext *ctx, const struct gl_client_array **arrays,
 {
 	struct nouveau_context *nctx = to_nouveau_context(ctx);
 	struct nouveau_render_state *render = to_render_state(ctx);
-	unsigned pushbuf_avail = PUSHBUF_DWORDS - 2 * nctx->bo.count,
+	unsigned pushbuf_avail = PUSHBUF_DWORDS - 2 * (nctx->bo.count +
+						       render->attr_count),
 		vert_avail = get_max_vertices(ctx, NULL, pushbuf_avail),
 		idx_avail = get_max_vertices(ctx, ib, pushbuf_avail);
 	int stride;
 
 	/* Try to keep client buffers smaller than the scratch BOs. */
-	if (!ib && render->mode == VBO &&
-	    (stride = get_max_client_stride(ctx)))
+	if (render->mode == VBO &&
+	    (stride = get_max_client_stride(ctx, arrays)))
 		    vert_avail = MIN2(vert_avail,
 				      RENDER_SCRATCH_SIZE / stride);
 
-
-	if ((ib && ib->count > idx_avail) ||
-	    (!ib && max_index - min_index > vert_avail)) {
+	if (max_index - min_index > vert_avail ||
+	    (ib && ib->count > idx_avail)) {
 		struct split_limits limits = {
 			.max_verts = vert_avail,
 			.max_indices = idx_avail,
@@ -322,6 +331,7 @@ vbo_bind_vertices(GLcontext *ctx, const struct gl_client_array **arrays,
 				* array->StrideB;
 
 			if (a->bo) {
+				/* Array in a buffer obj. */
 				a->offset = (intptr_t)array->Ptr + delta;
 			} else {
 				int j, n = max_index - min_index + 1;
@@ -329,6 +339,8 @@ vbo_bind_vertices(GLcontext *ctx, const struct gl_client_array **arrays,
 				char *dp = get_scratch_vbo(ctx, n * a->stride,
 							   &a->bo, &a->offset);
 
+				/* Array in client memory, move it to
+				 * a scratch buffer obj. */
 				for (j = 0; j < n; j++)
 					memcpy(dp + j * a->stride,
 					       sp + j * array->StrideB,
@@ -372,8 +384,6 @@ vbo_draw_vbo(GLcontext *ctx, const struct gl_client_array **arrays,
 		dispatch(ctx, start, delta, count);
 		BATCH_END();
 	}
-
-	FIRE_RING(chan);
 }
 
 /* Immediate rendering path. */
@@ -417,8 +427,6 @@ vbo_draw_imm(GLcontext *ctx, const struct gl_client_array **arrays,
 
 		BATCH_END();
 	}
-
-	FIRE_RING(chan);
 }
 
 /* draw_prims entry point when we're doing hw-tnl. */
diff --git a/src/mesa/drivers/dri/nouveau/nv04_context.c b/src/mesa/drivers/dri/nouveau/nv04_context.c
index a442425..6834f7c 100644
--- a/src/mesa/drivers/dri/nouveau/nv04_context.c
+++ b/src/mesa/drivers/dri/nouveau/nv04_context.c
@@ -75,18 +75,16 @@ nv04_channel_flush_notify(struct nouveau_channel *chan)
 	struct nouveau_context *nctx = chan->user_private;
 	GLcontext *ctx = &nctx->base;
 
-	if (nctx->fallback < SWRAST && ctx->DrawBuffer) {
-		GLcontext *ctx = &nctx->base;
-
+	if (nctx->fallback < SWRAST) {
 		/* Flushing seems to clobber the engine context. */
-		context_dirty_i(ctx, TEX_OBJ, 0);
-		context_dirty_i(ctx, TEX_OBJ, 1);
-		context_dirty_i(ctx, TEX_ENV, 0);
-		context_dirty_i(ctx, TEX_ENV, 1);
-		context_dirty(ctx, CONTROL);
-		context_dirty(ctx, BLEND);
-
-		nouveau_state_emit(ctx);
+		context_emit(ctx, TEX_OBJ0);
+		context_emit(ctx, TEX_OBJ1);
+		context_emit(ctx, TEX_ENV0);
+		context_emit(ctx, TEX_ENV1);
+		context_emit(ctx, CONTROL);
+		context_emit(ctx, BLEND);
+
+		nouveau_bo_state_emit(ctx);
 	}
 }
 
@@ -200,9 +198,9 @@ nv04_context_create(struct nouveau_screen *screen, const GLvisual *visual,
 	if (ret)
 		goto fail;
 
+	init_dummy_texture(ctx);
 	nv04_hwctx_init(ctx);
 	nv04_render_init(ctx);
-	init_dummy_texture(ctx);
 
 	return ctx;
 
@@ -278,6 +276,10 @@ const struct nouveau_driver nv04_driver = {
 		nouveau_emit_nothing,
 		nouveau_emit_nothing,
 		nouveau_emit_nothing,
+		nouveau_emit_nothing,
+		nouveau_emit_nothing,
+		nouveau_emit_nothing,
+		nouveau_emit_nothing,
 		nv04_emit_tex_obj,
 		nv04_emit_tex_obj,
 		nouveau_emit_nothing,
diff --git a/src/mesa/drivers/dri/nouveau/nv04_state_fb.c b/src/mesa/drivers/dri/nouveau/nv04_state_fb.c
index aad1e49..5e5e0c5 100644
--- a/src/mesa/drivers/dri/nouveau/nv04_state_fb.c
+++ b/src/mesa/drivers/dri/nouveau/nv04_state_fb.c
@@ -63,7 +63,7 @@ nv04_emit_framebuffer(GLcontext *ctx, int emit)
 		return;
 
 	/* Render target */
-	if (fb->_NumColorDrawBuffers) {
+	if (fb->_ColorDrawBuffers[0]) {
 		s = &to_nouveau_renderbuffer(
 			fb->_ColorDrawBuffers[0])->surface;
 
diff --git a/src/mesa/drivers/dri/nouveau/nv04_state_raster.c b/src/mesa/drivers/dri/nouveau/nv04_state_raster.c
index 4314fc3..c191571 100644
--- a/src/mesa/drivers/dri/nouveau/nv04_state_raster.c
+++ b/src/mesa/drivers/dri/nouveau/nv04_state_raster.c
@@ -275,6 +275,10 @@ nv04_emit_blend(GLcontext *ctx, int emit)
 		else
 			blend |= NV04_MULTITEX_TRIANGLE_BLEND_SHADE_MODE_FLAT;
 
+		/* Secondary color */
+		if (NEED_SECONDARY_COLOR(ctx))
+			blend |= NV04_MULTITEX_TRIANGLE_BLEND_SPECULAR_ENABLE;
+
 		/* Fog. */
 		if (ctx->Fog.Enabled)
 			blend |= NV04_MULTITEX_TRIANGLE_BLEND_FOG_ENABLE;
@@ -309,6 +313,10 @@ nv04_emit_blend(GLcontext *ctx, int emit)
 		else
 			blend |= get_texenv_mode(GL_MODULATE);
 
+		/* Secondary color */
+		if (NEED_SECONDARY_COLOR(ctx))
+			blend |= NV04_TEXTURED_TRIANGLE_BLEND_SPECULAR_ENABLE;
+
 		/* Fog. */
 		if (ctx->Fog.Enabled)
 			blend |= NV04_TEXTURED_TRIANGLE_BLEND_FOG_ENABLE;
diff --git a/src/mesa/drivers/dri/nouveau/nv10_context.c b/src/mesa/drivers/dri/nouveau/nv10_context.c
index 860d0ae..b6d1036 100644
--- a/src/mesa/drivers/dri/nouveau/nv10_context.c
+++ b/src/mesa/drivers/dri/nouveau/nv10_context.c
@@ -212,7 +212,7 @@ nv10_hwctx_init(GLcontext *ctx)
 	OUT_RING(chan, 0);
 	BEGIN_RING(chan, celsius, NV10TCL_CULL_FACE_ENABLE, 1);
 	OUT_RING(chan, 0);
-	BEGIN_RING(chan, celsius, NV10TCL_TX_GEN_S(0), 8);
+	BEGIN_RING(chan, celsius, NV10TCL_TX_GEN_MODE_S(0), 8);
 	for (i = 0; i < 8; i++)
 		OUT_RING(chan, 0);
 
@@ -412,6 +412,10 @@ const struct nouveau_driver nv10_driver = {
 		nv10_emit_tex_gen,
 		nouveau_emit_nothing,
 		nouveau_emit_nothing,
+		nv10_emit_tex_mat,
+		nv10_emit_tex_mat,
+		nouveau_emit_nothing,
+		nouveau_emit_nothing,
 		nv10_emit_tex_obj,
 		nv10_emit_tex_obj,
 		nouveau_emit_nothing,
diff --git a/src/mesa/drivers/dri/nouveau/nv10_driver.h b/src/mesa/drivers/dri/nouveau/nv10_driver.h
index d662712..cefd6c6 100644
--- a/src/mesa/drivers/dri/nouveau/nv10_driver.h
+++ b/src/mesa/drivers/dri/nouveau/nv10_driver.h
@@ -134,6 +134,9 @@ void
 nv10_emit_tex_gen(GLcontext *ctx, int emit);
 
 void
+nv10_emit_tex_mat(GLcontext *ctx, int emit);
+
+void
 nv10_emit_tex_obj(GLcontext *ctx, int emit);
 
 /* nv10_state_tnl.c */
diff --git a/src/mesa/drivers/dri/nouveau/nv10_state_fb.c b/src/mesa/drivers/dri/nouveau/nv10_state_fb.c
index 05c36b4..a2fcb6b 100644
--- a/src/mesa/drivers/dri/nouveau/nv10_state_fb.c
+++ b/src/mesa/drivers/dri/nouveau/nv10_state_fb.c
@@ -71,6 +71,7 @@ setup_lma_buffer(GLcontext *ctx)
 	nouveau_bo_markl(bctx, celsius, NV17TCL_LMA_DEPTH_BUFFER_OFFSET,
 			 nfb->lma_bo, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RDWR);
 
+	WAIT_RING(chan, 9);
 	BEGIN_RING(chan, celsius, NV17TCL_LMA_DEPTH_WINDOW_X, 4);
 	OUT_RINGf(chan, - 1792);
 	OUT_RINGf(chan, - 2304 + fb->Height);
@@ -111,7 +112,7 @@ nv10_emit_framebuffer(GLcontext *ctx, int emit)
 	}
 
 	/* Render target */
-	if (fb->_NumColorDrawBuffers) {
+	if (fb->_ColorDrawBuffers[0]) {
 		s = &to_nouveau_renderbuffer(
 			fb->_ColorDrawBuffers[0])->surface;
 
@@ -171,15 +172,13 @@ nv10_emit_viewport(GLcontext *ctx, int emit)
 	struct nouveau_grobj *celsius = context_eng3d(ctx);
 	struct gl_framebuffer *fb = ctx->DrawBuffer;
 	float a[4] = {};
-	int i;
 
 	get_viewport_translate(ctx, a);
 	a[0] -= 2048;
 	a[1] -= 2048;
 
 	BEGIN_RING(chan, celsius, NV10TCL_VIEWPORT_TRANSLATE_X, 4);
-	for (i = 0; i < 4; i++)
-		OUT_RINGf(chan, a[i]);
+	OUT_RINGp(chan, a, 4);
 
 	BEGIN_RING(chan, celsius, NV10TCL_VIEWPORT_CLIP_HORIZ(0), 1);
 	OUT_RING(chan, (fb->Width - 1) << 16 | 0x08000800);
diff --git a/src/mesa/drivers/dri/nouveau/nv10_state_tex.c b/src/mesa/drivers/dri/nouveau/nv10_state_tex.c
index 02a5ca7..6dedb18 100644
--- a/src/mesa/drivers/dri/nouveau/nv10_state_tex.c
+++ b/src/mesa/drivers/dri/nouveau/nv10_state_tex.c
@@ -32,9 +32,64 @@
 #include "nouveau_util.h"
 #include "nv10_driver.h"
 
+#define TX_GEN_MODE(i, j) (NV10TCL_TX_GEN_MODE_S(i) + 4 * (j))
+#define TX_GEN_COEFF(i, j) (NV10TCL_TX_GEN_COEFF_S_A(i) + 16 * (j))
+#define TX_MATRIX(i) (NV10TCL_TX0_MATRIX(0) + 64 * (i))
+
 void
 nv10_emit_tex_gen(GLcontext *ctx, int emit)
 {
+	const int i = emit - NOUVEAU_STATE_TEX_GEN0;
+	struct nouveau_context *nctx = to_nouveau_context(ctx);
+	struct nouveau_channel *chan = context_chan(ctx);
+	struct nouveau_grobj *celsius = context_eng3d(ctx);
+	struct gl_texture_unit *unit = &ctx->Texture.Unit[i];
+	int j;
+
+	for (j = 0; j < 4; j++) {
+		if (nctx->fallback == HWTNL && (unit->TexGenEnabled & 1 << j)) {
+			struct gl_texgen *coord = get_texgen_coord(unit, j);
+			float *k = get_texgen_coeff(coord);
+
+			if (k) {
+				BEGIN_RING(chan, celsius,
+					   TX_GEN_COEFF(i, j), 4);
+				OUT_RINGp(chan, k, 4);
+			}
+
+			BEGIN_RING(chan, celsius, TX_GEN_MODE(i, j), 1);
+			OUT_RING(chan, nvgl_texgen_mode(coord->Mode));
+
+		} else {
+			BEGIN_RING(chan, celsius, TX_GEN_MODE(i, j), 1);
+			OUT_RING(chan, 0);
+		}
+	}
+
+	context_dirty_i(ctx, TEX_MAT, i);
+}
+
+void
+nv10_emit_tex_mat(GLcontext *ctx, int emit)
+{
+	const int i = emit - NOUVEAU_STATE_TEX_MAT0;
+	struct nouveau_context *nctx = to_nouveau_context(ctx);
+	struct nouveau_channel *chan = context_chan(ctx);
+	struct nouveau_grobj *celsius = context_eng3d(ctx);
+
+	if (nctx->fallback == HWTNL &&
+	    ((ctx->Texture._TexMatEnabled & 1 << i) ||
+	     ctx->Texture.Unit[i]._GenFlags)) {
+		BEGIN_RING(chan, celsius, NV10TCL_TX_MATRIX_ENABLE(i), 1);
+		OUT_RING(chan, 1);
+
+		BEGIN_RING(chan, celsius, TX_MATRIX(i), 16);
+		OUT_RINGm(chan, ctx->TextureMatrixStack[i].Top->m);
+
+	} else {
+		BEGIN_RING(chan, celsius, NV10TCL_TX_MATRIX_ENABLE(i), 1);
+		OUT_RING(chan, 0);
+	}
 }
 
 static uint32_t
diff --git a/src/mesa/drivers/dri/nouveau/nv10_state_tnl.c b/src/mesa/drivers/dri/nouveau/nv10_state_tnl.c
index 6db14d8..0e592a1 100644
--- a/src/mesa/drivers/dri/nouveau/nv10_state_tnl.c
+++ b/src/mesa/drivers/dri/nouveau/nv10_state_tnl.c
@@ -140,9 +140,7 @@ nv10_emit_fog(GLcontext *ctx, int emit)
 	OUT_RING(chan, pack_rgba_f(MESA_FORMAT_RGBA8888_REV, f->Color));
 
 	BEGIN_RING(chan, celsius, NV10TCL_FOG_EQUATION_CONSTANT, 3);
-	OUT_RINGf(chan, k[0]);
-	OUT_RINGf(chan, k[1]);
-	OUT_RINGf(chan, k[2]);
+	OUT_RINGp(chan, k, 3);
 
 	context_dirty(ctx, FRAG);
 }
@@ -201,8 +199,10 @@ nv10_emit_light_model(GLcontext *ctx, int emit)
 	BEGIN_RING(chan, celsius, NV10TCL_LIGHT_MODEL, 1);
 	OUT_RING(chan, ((m->LocalViewer ?
 			 NV10TCL_LIGHT_MODEL_LOCAL_VIEWER : 0) |
-			(m->ColorControl == GL_SEPARATE_SPECULAR_COLOR ?
-			 NV10TCL_LIGHT_MODEL_SEPARATE_SPECULAR : 0)));
+			(NEED_SECONDARY_COLOR(ctx) ?
+			 NV10TCL_LIGHT_MODEL_SEPARATE_SPECULAR : 0) |
+			(!ctx->Light.Enabled && ctx->Fog.ColorSumEnabled ?
+			 NV10TCL_LIGHT_MODEL_VERTEX_SPECULAR : 0)));
 }
 
 static float
@@ -282,9 +282,7 @@ nv10_emit_light_source(GLcontext *ctx, int emit)
 
 	if (l->_Flags & LIGHT_POSITIONAL) {
 		BEGIN_RING(chan, celsius, NV10TCL_LIGHT_POSITION_X(i), 3);
-		OUT_RINGf(chan, l->_Position[0]);
-		OUT_RINGf(chan, l->_Position[1]);
-		OUT_RINGf(chan, l->_Position[2]);
+		OUT_RINGp(chan, l->_Position, 3);
 
 		BEGIN_RING(chan, celsius,
 			   NV10TCL_LIGHT_ATTENUATION_CONSTANT(i), 3);
@@ -294,14 +292,10 @@ nv10_emit_light_source(GLcontext *ctx, int emit)
 
 	} else {
 		BEGIN_RING(chan, celsius, NV10TCL_LIGHT_DIRECTION_X(i), 3);
-		OUT_RINGf(chan, l->_VP_inf_norm[0]);
-		OUT_RINGf(chan, l->_VP_inf_norm[1]);
-		OUT_RINGf(chan, l->_VP_inf_norm[2]);
+		OUT_RINGp(chan, l->_VP_inf_norm, 3);
 
 		BEGIN_RING(chan, celsius, NV10TCL_LIGHT_HALF_VECTOR_X(i), 3);
-		OUT_RINGf(chan, l->_h_inf_norm[0]);
-		OUT_RINGf(chan, l->_h_inf_norm[1]);
-		OUT_RINGf(chan, l->_h_inf_norm[2]);
+		OUT_RINGp(chan, l->_h_inf_norm, 3);
 	}
 
 	if (l->_Flags & LIGHT_SPOT) {
@@ -310,13 +304,7 @@ nv10_emit_light_source(GLcontext *ctx, int emit)
 		nv10_get_spot_coeff(l, k);
 
 		BEGIN_RING(chan, celsius, NV10TCL_LIGHT_SPOT_CUTOFF_A(i), 7);
-		OUT_RINGf(chan, k[0]);
-		OUT_RINGf(chan, k[1]);
-		OUT_RINGf(chan, k[2]);
-		OUT_RINGf(chan, k[3]);
-		OUT_RINGf(chan, k[4]);
-		OUT_RINGf(chan, k[5]);
-		OUT_RINGf(chan, k[6]);
+		OUT_RINGp(chan, k, 7);
 	}
 }
 
@@ -348,15 +336,11 @@ nv10_emit_material_ambient(GLcontext *ctx, int emit)
 	}
 
 	BEGIN_RING(chan, celsius, NV10TCL_LIGHT_MODEL_AMBIENT_R, 3);
-	OUT_RINGf(chan, c_scene[0]);
-	OUT_RINGf(chan, c_scene[1]);
-	OUT_RINGf(chan, c_scene[2]);
+	OUT_RINGp(chan, c_scene, 3);
 
 	if (ctx->Light.ColorMaterialEnabled) {
 		BEGIN_RING(chan, celsius, NV10TCL_MATERIAL_FACTOR_R, 3);
-		OUT_RINGf(chan, c_factor[0]);
-		OUT_RINGf(chan, c_factor[1]);
-		OUT_RINGf(chan, c_factor[2]);
+		OUT_RINGp(chan, c_factor, 3);
 	}
 
 	foreach(l, &ctx->Light.EnabledList) {
@@ -366,9 +350,7 @@ nv10_emit_material_ambient(GLcontext *ctx, int emit)
 				  l->_MatAmbient[0]);
 
 		BEGIN_RING(chan, celsius, NV10TCL_LIGHT_AMBIENT_R(i), 3);
-		OUT_RINGf(chan, c_light[0]);
-		OUT_RINGf(chan, c_light[1]);
-		OUT_RINGf(chan, c_light[2]);
+		OUT_RINGp(chan, c_light, 3);
 	}
 }
 
@@ -390,9 +372,7 @@ nv10_emit_material_diffuse(GLcontext *ctx, int emit)
 				  l->_MatDiffuse[0]);
 
 		BEGIN_RING(chan, celsius, NV10TCL_LIGHT_DIFFUSE_R(i), 3);
-		OUT_RINGf(chan, c_light[0]);
-		OUT_RINGf(chan, c_light[1]);
-		OUT_RINGf(chan, c_light[2]);
+		OUT_RINGp(chan, c_light, 3);
 	}
 }
 
@@ -410,9 +390,7 @@ nv10_emit_material_specular(GLcontext *ctx, int emit)
 				  l->_MatSpecular[0]);
 
 		BEGIN_RING(chan, celsius, NV10TCL_LIGHT_SPECULAR_R(i), 3);
-		OUT_RINGf(chan, c_light[0]);
-		OUT_RINGf(chan, c_light[1]);
-		OUT_RINGf(chan, c_light[2]);
+		OUT_RINGp(chan, c_light, 3);
 	}
 }
 
@@ -453,12 +431,7 @@ nv10_emit_material_shininess(GLcontext *ctx, int emit)
 		k);
 
 	BEGIN_RING(chan, celsius, NV10TCL_MATERIAL_SHININESS(0), 6);
-	OUT_RINGf(chan, k[0]);
-	OUT_RINGf(chan, k[1]);
-	OUT_RINGf(chan, k[2]);
-	OUT_RINGf(chan, k[3]);
-	OUT_RINGf(chan, k[4]);
-	OUT_RINGf(chan, k[5]);
+	OUT_RINGp(chan, k, 6);
 }
 
 void
@@ -472,12 +445,14 @@ nv10_emit_modelview(GLcontext *ctx, int emit)
 	if (nctx->fallback != HWTNL)
 		return;
 
-	if (ctx->Light._NeedEyeCoords || ctx->Fog.Enabled) {
+	if (ctx->Light._NeedEyeCoords || ctx->Fog.Enabled ||
+	    (ctx->Texture._GenFlags & TEXGEN_NEED_EYE_COORD)) {
 		BEGIN_RING(chan, celsius, NV10TCL_MODELVIEW0_MATRIX(0), 16);
 		OUT_RINGm(chan, m->m);
 	}
 
-	if (ctx->Light.Enabled) {
+	if (ctx->Light.Enabled ||
+	    (ctx->Texture._GenFlags & TEXGEN_NEED_EYE_COORD)) {
 		int i, j;
 
 		BEGIN_RING(chan, celsius,
diff --git a/src/mesa/drivers/dri/nouveau/nv20_context.c b/src/mesa/drivers/dri/nouveau/nv20_context.c
index db39ef7..789dcaa 100644
--- a/src/mesa/drivers/dri/nouveau/nv20_context.c
+++ b/src/mesa/drivers/dri/nouveau/nv20_context.c
@@ -297,9 +297,9 @@ nv20_hwctx_init(GLcontext *ctx)
 	BEGIN_RING(chan, kelvin, NV20TCL_POLYGON_STIPPLE_ENABLE, 1);
 	OUT_RING  (chan, 0);
 
-	BEGIN_RING(chan, kelvin, NV20TCL_TX_GEN_S(0),
-		   4 * NV20TCL_TX_GEN_S__SIZE);
-	for (i=0; i < 4 * NV20TCL_TX_GEN_S__SIZE; i++)
+	BEGIN_RING(chan, kelvin, NV20TCL_TX_GEN_MODE_S(0),
+		   4 * NV20TCL_TX_GEN_MODE_S__SIZE);
+	for (i=0; i < 4 * NV20TCL_TX_GEN_MODE_S__SIZE; i++)
 		OUT_RING(chan, 0);
 
 	BEGIN_RING(chan, kelvin, NV20TCL_FOG_EQUATION_CONSTANT, 3);
@@ -497,10 +497,14 @@ const struct nouveau_driver nv20_driver = {
 		nv20_emit_tex_env,
 		nv20_emit_tex_env,
 		nv20_emit_tex_env,
-		nv10_emit_tex_gen,
-		nv10_emit_tex_gen,
-		nv10_emit_tex_gen,
-		nv10_emit_tex_gen,
+		nv20_emit_tex_gen,
+		nv20_emit_tex_gen,
+		nv20_emit_tex_gen,
+		nv20_emit_tex_gen,
+		nv20_emit_tex_mat,
+		nv20_emit_tex_mat,
+		nv20_emit_tex_mat,
+		nv20_emit_tex_mat,
 		nv20_emit_tex_obj,
 		nv20_emit_tex_obj,
 		nv20_emit_tex_obj,
diff --git a/src/mesa/drivers/dri/nouveau/nv20_driver.h b/src/mesa/drivers/dri/nouveau/nv20_driver.h
index 18574e9..8adecef 100644
--- a/src/mesa/drivers/dri/nouveau/nv20_driver.h
+++ b/src/mesa/drivers/dri/nouveau/nv20_driver.h
@@ -68,6 +68,12 @@ nv20_emit_frag(GLcontext *ctx, int emit);
 
 /* nv20_state_tex.c */
 void
+nv20_emit_tex_gen(GLcontext *ctx, int emit);
+
+void
+nv20_emit_tex_mat(GLcontext *ctx, int emit);
+
+void
 nv20_emit_tex_obj(GLcontext *ctx, int emit);
 
 void
diff --git a/src/mesa/drivers/dri/nouveau/nv20_state_fb.c b/src/mesa/drivers/dri/nouveau/nv20_state_fb.c
index 869acd6..21da4f7 100644
--- a/src/mesa/drivers/dri/nouveau/nv20_state_fb.c
+++ b/src/mesa/drivers/dri/nouveau/nv20_state_fb.c
@@ -67,7 +67,7 @@ nv20_emit_framebuffer(GLcontext *ctx, int emit)
 		return;
 
 	/* Render target */
-	if (fb->_NumColorDrawBuffers) {
+	if (fb->_ColorDrawBuffers[0]) {
 		s = &to_nouveau_renderbuffer(
 			fb->_ColorDrawBuffers[0])->surface;
 
@@ -106,13 +106,11 @@ nv20_emit_viewport(GLcontext *ctx, int emit)
 	struct nouveau_grobj *kelvin = context_eng3d(ctx);
 	struct gl_framebuffer *fb = ctx->DrawBuffer;
 	float a[4] = {};
-	int i;
 
 	get_viewport_translate(ctx, a);
 
 	BEGIN_RING(chan, kelvin, NV20TCL_VIEWPORT_TRANSLATE_X, 4);
-	for (i = 0; i < 4; i++)
-		OUT_RINGf(chan, a[i]);
+	OUT_RINGp(chan, a, 4);
 
 	BEGIN_RING(chan, kelvin, NV20TCL_VIEWPORT_CLIP_HORIZ(0), 1);
 	OUT_RING(chan, (fb->Width - 1) << 16);
diff --git a/src/mesa/drivers/dri/nouveau/nv20_state_tex.c b/src/mesa/drivers/dri/nouveau/nv20_state_tex.c
index 9287010..e46118e 100644
--- a/src/mesa/drivers/dri/nouveau/nv20_state_tex.c
+++ b/src/mesa/drivers/dri/nouveau/nv20_state_tex.c
@@ -32,6 +32,62 @@
 #include "nouveau_util.h"
 #include "nv20_driver.h"
 
+#define TX_GEN_MODE(i, j) (NV20TCL_TX_GEN_MODE_S(i) + 4 * (j))
+#define TX_GEN_COEFF(i, j) (NV20TCL_TX_GEN_COEFF_S_A(i) + 16 * (j))
+#define TX_MATRIX(i) (NV20TCL_TX0_MATRIX(0) + 64 * (i))
+
+void
+nv20_emit_tex_gen(GLcontext *ctx, int emit)
+{
+	const int i = emit - NOUVEAU_STATE_TEX_GEN0;
+	struct nouveau_context *nctx = to_nouveau_context(ctx);
+	struct nouveau_channel *chan = context_chan(ctx);
+	struct nouveau_grobj *kelvin = context_eng3d(ctx);
+	struct gl_texture_unit *unit = &ctx->Texture.Unit[i];
+	int j;
+
+	for (j = 0; j < 4; j++) {
+		if (nctx->fallback == HWTNL && (unit->TexGenEnabled & 1 << j)) {
+			struct gl_texgen *coord = get_texgen_coord(unit, j);
+			float *k = get_texgen_coeff(coord);
+
+			if (k) {
+				BEGIN_RING(chan, kelvin, TX_GEN_COEFF(i, j), 4);
+				OUT_RINGp(chan, k, 4);
+			}
+
+			BEGIN_RING(chan, kelvin, TX_GEN_MODE(i, j), 1);
+			OUT_RING(chan, nvgl_texgen_mode(coord->Mode));
+
+		} else {
+			BEGIN_RING(chan, kelvin, TX_GEN_MODE(i, j), 1);
+			OUT_RING(chan, 0);
+		}
+	}
+}
+
+void
+nv20_emit_tex_mat(GLcontext *ctx, int emit)
+{
+	const int i = emit - NOUVEAU_STATE_TEX_MAT0;
+	struct nouveau_context *nctx = to_nouveau_context(ctx);
+	struct nouveau_channel *chan = context_chan(ctx);
+	struct nouveau_grobj *kelvin = context_eng3d(ctx);
+
+	if (nctx->fallback == HWTNL &&
+	    (ctx->Texture._TexMatEnabled & 1 << i)) {
+		BEGIN_RING(chan, kelvin, NV20TCL_TX_MATRIX_ENABLE(i), 1);
+		OUT_RING(chan, 1);
+
+		BEGIN_RING(chan, kelvin, TX_MATRIX(i), 16);
+		OUT_RINGm(chan, ctx->TextureMatrixStack[i].Top->m);
+
+	} else {
+		BEGIN_RING(chan, kelvin, NV20TCL_TX_MATRIX_ENABLE(i), 1);
+		OUT_RING(chan, 0);
+	}
+}
+
 static uint32_t
 get_tex_format_pot(struct gl_texture_image *ti)
 {
diff --git a/src/mesa/drivers/dri/nouveau/nv20_state_tnl.c b/src/mesa/drivers/dri/nouveau/nv20_state_tnl.c
index 0d56606..62efe80 100644
--- a/src/mesa/drivers/dri/nouveau/nv20_state_tnl.c
+++ b/src/mesa/drivers/dri/nouveau/nv20_state_tnl.c
@@ -139,9 +139,7 @@ nv20_emit_fog(GLcontext *ctx, int emit)
 	OUT_RING(chan, pack_rgba_f(MESA_FORMAT_RGBA8888_REV, f->Color));
 
 	BEGIN_RING(chan, kelvin, NV20TCL_FOG_EQUATION_CONSTANT, 3);
-	OUT_RINGf(chan, k[0]);
-	OUT_RINGf(chan, k[1]);
-	OUT_RINGf(chan, k[2]);
+	OUT_RINGp(chan, k, 3);
 }
 
 void
@@ -158,7 +156,7 @@ nv20_emit_light_model(GLcontext *ctx, int emit)
 	OUT_RING(chan, ((m->LocalViewer ?
 			 NV20TCL_LIGHT_MODEL_VIEWER_LOCAL :
 			 NV20TCL_LIGHT_MODEL_VIEWER_NONLOCAL) |
-			(m->ColorControl == GL_SEPARATE_SPECULAR_COLOR ?
+			(NEED_SECONDARY_COLOR(ctx) ?
 			 NV20TCL_LIGHT_MODEL_SEPARATE_SPECULAR :
 			 0)));
 
@@ -176,9 +174,7 @@ nv20_emit_light_source(GLcontext *ctx, int emit)
 
 	if (l->_Flags & LIGHT_POSITIONAL) {
 		BEGIN_RING(chan, kelvin, NV20TCL_LIGHT_POSITION_X(i), 3);
-		OUT_RINGf(chan, l->_Position[0]);
-		OUT_RINGf(chan, l->_Position[1]);
-		OUT_RINGf(chan, l->_Position[2]);
+		OUT_RINGp(chan, l->_Position, 3);
 
 		BEGIN_RING(chan, kelvin, NV20TCL_LIGHT_ATTENUATION_CONSTANT(i), 3);
 		OUT_RINGf(chan, l->ConstantAttenuation);
@@ -187,14 +183,10 @@ nv20_emit_light_source(GLcontext *ctx, int emit)
 
 	} else {
 		BEGIN_RING(chan, kelvin, NV20TCL_LIGHT_DIRECTION_X(i), 3);
-		OUT_RINGf(chan, l->_VP_inf_norm[0]);
-		OUT_RINGf(chan, l->_VP_inf_norm[1]);
-		OUT_RINGf(chan, l->_VP_inf_norm[2]);
+		OUT_RINGp(chan, l->_VP_inf_norm, 3);
 
 		BEGIN_RING(chan, kelvin, NV20TCL_LIGHT_HALF_VECTOR_X(i), 3);
-		OUT_RINGf(chan, l->_h_inf_norm[0]);
-		OUT_RINGf(chan, l->_h_inf_norm[1]);
-		OUT_RINGf(chan, l->_h_inf_norm[2]);
+		OUT_RINGp(chan, l->_h_inf_norm, 3);
 	}
 
 	if (l->_Flags & LIGHT_SPOT) {
@@ -203,13 +195,7 @@ nv20_emit_light_source(GLcontext *ctx, int emit)
 		nv10_get_spot_coeff(l, k);
 
 		BEGIN_RING(chan, kelvin, NV20TCL_LIGHT_SPOT_CUTOFF_A(i), 7);
-		OUT_RINGf(chan, k[0]);
-		OUT_RINGf(chan, k[1]);
-		OUT_RINGf(chan, k[2]);
-		OUT_RINGf(chan, k[3]);
-		OUT_RINGf(chan, k[4]);
-		OUT_RINGf(chan, k[5]);
-		OUT_RINGf(chan, k[6]);
+		OUT_RINGp(chan, k, 7);
 	}
 }
 
@@ -246,15 +232,11 @@ nv20_emit_material_ambient(GLcontext *ctx, int emit)
 	}
 
 	BEGIN_RING(chan, kelvin, m_scene[side], 3);
-	OUT_RINGf(chan, c_scene[0]);
-	OUT_RINGf(chan, c_scene[1]);
-	OUT_RINGf(chan, c_scene[2]);
+	OUT_RINGp(chan, c_scene, 3);
 
 	if (ctx->Light.ColorMaterialEnabled) {
 		BEGIN_RING(chan, kelvin, m_factor[side], 3);
-		OUT_RINGf(chan, c_factor[0]);
-		OUT_RINGf(chan, c_factor[1]);
-		OUT_RINGf(chan, c_factor[2]);
+		OUT_RINGp(chan, c_factor, 3);
 	}
 
 	foreach(l, &ctx->Light.EnabledList) {
@@ -266,9 +248,7 @@ nv20_emit_material_ambient(GLcontext *ctx, int emit)
 				  l->_MatAmbient[side]);
 
 		BEGIN_RING(chan, kelvin, m_light[side], 3);
-		OUT_RINGf(chan, c_light[0]);
-		OUT_RINGf(chan, c_light[1]);
-		OUT_RINGf(chan, c_light[2]);
+		OUT_RINGp(chan, c_light, 3);
 	}
 }
 
@@ -295,9 +275,7 @@ nv20_emit_material_diffuse(GLcontext *ctx, int emit)
 				  l->_MatDiffuse[side]);
 
 		BEGIN_RING(chan, kelvin, m_light[side], 3);
-		OUT_RINGf(chan, c_light[0]);
-		OUT_RINGf(chan, c_light[1]);
-		OUT_RINGf(chan, c_light[2]);
+		OUT_RINGp(chan, c_light, 3);
 	}
 }
 
@@ -318,9 +296,7 @@ nv20_emit_material_specular(GLcontext *ctx, int emit)
 				  l->_MatSpecular[side]);
 
 		BEGIN_RING(chan, kelvin, m_light[side], 3);
-		OUT_RINGf(chan, c_light[0]);
-		OUT_RINGf(chan, c_light[1]);
-		OUT_RINGf(chan, c_light[2]);
+		OUT_RINGp(chan, c_light, 3);
 	}
 }
 
@@ -340,12 +316,7 @@ nv20_emit_material_shininess(GLcontext *ctx, int emit)
 		k);
 
 	BEGIN_RING(chan, kelvin, mthd[side], 6);
-	OUT_RINGf(chan, k[0]);
-	OUT_RINGf(chan, k[1]);
-	OUT_RINGf(chan, k[2]);
-	OUT_RINGf(chan, k[3]);
-	OUT_RINGf(chan, k[4]);
-	OUT_RINGf(chan, k[5]);
+	OUT_RINGp(chan, k, 6);
 }
 
 void
@@ -359,12 +330,14 @@ nv20_emit_modelview(GLcontext *ctx, int emit)
 	if (nctx->fallback != HWTNL)
 		return;
 
-	if (ctx->Light._NeedEyeCoords || ctx->Fog.Enabled) {
+	if (ctx->Light._NeedEyeCoords || ctx->Fog.Enabled ||
+	    (ctx->Texture._GenFlags & TEXGEN_NEED_EYE_COORD)) {
 		BEGIN_RING(chan, kelvin, NV20TCL_MODELVIEW0_MATRIX(0), 16);
 		OUT_RINGm(chan, m->m);
 	}
 
-	if (ctx->Light.Enabled) {
+	if (ctx->Light.Enabled ||
+	    (ctx->Texture._GenFlags & TEXGEN_NEED_EYE_COORD)) {
 		int i, j;
 
 		BEGIN_RING(chan, kelvin,