From 84cc07dc89c0ebce4ad55b4b3684d4420a202683 Mon Sep 17 00:00:00 2001
From: Ben Skeggs <skeggsb@gmail.com>
Date: Fri, 29 Feb 2008 15:03:57 +1100
Subject: nouveau: implement pipe_screen

Untested on NV3x/NV5x.  Quite possibly broken.
---
 src/gallium/drivers/nv50/nv50_screen.h | 20 ++++++++++++++++++++
 1 file changed, 20 insertions(+)
 create mode 100644 src/gallium/drivers/nv50/nv50_screen.h

(limited to 'src/gallium/drivers/nv50/nv50_screen.h')

diff --git a/src/gallium/drivers/nv50/nv50_screen.h b/src/gallium/drivers/nv50/nv50_screen.h
new file mode 100644
index 0000000000..45ebbb8051
--- /dev/null
+++ b/src/gallium/drivers/nv50/nv50_screen.h
@@ -0,0 +1,20 @@
+#ifndef __NV50_SCREEN_H__
+#define __NV50_SCREEN_H__
+
+#include "pipe/p_screen.h"
+
+struct nv50_screen {
+	struct pipe_screen pipe;
+	unsigned chipset;
+};
+
+static INLINE struct nv50_screen *
+nv50_screen(struct pipe_screen *screen)
+{
+	return (struct nv50_screen *)screen;
+}
+
+extern struct pipe_screen *
+nv50_screen_create(struct pipe_winsys *winsys, unsigned chipset);
+
+#endif
-- 
cgit v1.2.3


From baaae562f02563c5966b857c61b3eae7341950e3 Mon Sep 17 00:00:00 2001
From: Ben Skeggs <skeggsb@gmail.com>
Date: Fri, 29 Feb 2008 22:54:40 +1100
Subject: nouveau: hand nouveau_winsys in with pipe_screen init

---
 src/gallium/drivers/nouveau/nouveau_winsys.h    | 15 +++++++++------
 src/gallium/drivers/nv30/nv30_context.c         |  3 ++-
 src/gallium/drivers/nv30/nv30_screen.c          |  4 +++-
 src/gallium/drivers/nv30/nv30_screen.h          |  5 ++---
 src/gallium/drivers/nv40/nv40_context.c         |  3 ++-
 src/gallium/drivers/nv40/nv40_screen.c          |  4 +++-
 src/gallium/drivers/nv40/nv40_screen.h          |  5 ++---
 src/gallium/drivers/nv50/nv50_context.c         |  3 ++-
 src/gallium/drivers/nv50/nv50_screen.c          |  4 +++-
 src/gallium/drivers/nv50/nv50_screen.h          |  5 ++---
 src/gallium/winsys/dri/nouveau/nouveau_winsys.c |  8 ++++----
 11 files changed, 34 insertions(+), 25 deletions(-)

(limited to 'src/gallium/drivers/nv50/nv50_screen.h')

diff --git a/src/gallium/drivers/nouveau/nouveau_winsys.h b/src/gallium/drivers/nouveau/nouveau_winsys.h
index 98d95e94a5..11ca7e80dd 100644
--- a/src/gallium/drivers/nouveau/nouveau_winsys.h
+++ b/src/gallium/drivers/nouveau/nouveau_winsys.h
@@ -50,21 +50,24 @@ struct nouveau_winsys {
 };
 
 extern struct pipe_screen *
-nv30_screen_create(struct pipe_winsys *ws, unsigned chipset);
+nv30_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *,
+		   unsigned chipset);
 
 extern struct pipe_context *
-nv30_create(struct pipe_screen *, struct nouveau_winsys *);
+nv30_create(struct pipe_screen *);
 
 extern struct pipe_screen *
-nv40_screen_create(struct pipe_winsys *ws, unsigned chipset);
+nv40_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *,
+		   unsigned chipset);
 
 extern struct pipe_context *
-nv40_create(struct pipe_screen *, struct nouveau_winsys *);
+nv40_create(struct pipe_screen *);
 
 extern struct pipe_screen *
-nv50_screen_create(struct pipe_winsys *ws, unsigned chipset);
+nv50_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *,
+		   unsigned chipset);
 
 extern struct pipe_context *
-nv50_create(struct pipe_screen *, struct nouveau_winsys *);
+nv50_create(struct pipe_screen *);
 
 #endif
diff --git a/src/gallium/drivers/nv30/nv30_context.c b/src/gallium/drivers/nv30/nv30_context.c
index b8452e23b1..522fb13226 100644
--- a/src/gallium/drivers/nv30/nv30_context.c
+++ b/src/gallium/drivers/nv30/nv30_context.c
@@ -265,9 +265,10 @@ nv30_init_hwctx(struct nv30_context *nv30, int rankine_class)
 #define NV35TCL_CHIPSET_3X_MASK 0x000001e0
 
 struct pipe_context *
-nv30_create(struct pipe_screen *screen, struct nouveau_winsys *nvws)
+nv30_create(struct pipe_screen *screen)
 {
 	struct pipe_winsys *pipe_winsys = screen->winsys;
+	struct nouveau_winsys *nvws = nv30_screen(screen)->nvws;
 	unsigned chipset = nv30_screen(screen)->chipset;
 	struct nv30_context *nv30;
 	int rankine_class = 0, ret;
diff --git a/src/gallium/drivers/nv30/nv30_screen.c b/src/gallium/drivers/nv30/nv30_screen.c
index 6d64025528..39f2ac1af5 100644
--- a/src/gallium/drivers/nv30/nv30_screen.c
+++ b/src/gallium/drivers/nv30/nv30_screen.c
@@ -125,7 +125,8 @@ nv30_screen_destroy(struct pipe_screen *screen)
 }
 
 struct pipe_screen *
-nv30_screen_create(struct pipe_winsys *winsys, unsigned chipset)
+nv30_screen_create(struct pipe_winsys *winsys, struct nouveau_winsys *nvws,
+		   unsigned chipset)
 {
 	struct nv30_screen *nv30screen = CALLOC_STRUCT(nv30_screen);
 
@@ -133,6 +134,7 @@ nv30_screen_create(struct pipe_winsys *winsys, unsigned chipset)
 		return NULL;
 
 	nv30screen->chipset = chipset;
+	nv30screen->nvws = nvws;
 
 	nv30screen->screen.winsys = winsys;
 
diff --git a/src/gallium/drivers/nv30/nv30_screen.h b/src/gallium/drivers/nv30/nv30_screen.h
index e55242fbf7..f878f81e11 100644
--- a/src/gallium/drivers/nv30/nv30_screen.h
+++ b/src/gallium/drivers/nv30/nv30_screen.h
@@ -5,6 +5,8 @@
 
 struct nv30_screen {
 	struct pipe_screen screen;
+
+	struct nouveau_winsys *nvws;
 	unsigned chipset;
 };
 
@@ -14,7 +16,4 @@ nv30_screen(struct pipe_screen *screen)
 	return (struct nv30_screen *)screen;
 }
 
-extern struct pipe_screen *
-nv30_screen_create(struct pipe_winsys *winsys, unsigned chipset);
-
 #endif
diff --git a/src/gallium/drivers/nv40/nv40_context.c b/src/gallium/drivers/nv40/nv40_context.c
index a7f64c6e9e..679c2ddc6b 100644
--- a/src/gallium/drivers/nv40/nv40_context.c
+++ b/src/gallium/drivers/nv40/nv40_context.c
@@ -196,11 +196,12 @@ nv40_destroy(struct pipe_context *pipe)
 }
 
 struct pipe_context *
-nv40_create(struct pipe_screen *pscreen, struct nouveau_winsys *nvws)
+nv40_create(struct pipe_screen *pscreen)
 {
 	struct pipe_winsys *ws = pscreen->winsys;
 	struct nv40_context *nv40;
 	unsigned chipset = nv40_screen(pscreen)->chipset;
+	struct nouveau_winsys *nvws = nv40_screen(pscreen)->nvws;
 
 	nv40 = CALLOC(1, sizeof(struct nv40_context));
 	if (!nv40)
diff --git a/src/gallium/drivers/nv40/nv40_screen.c b/src/gallium/drivers/nv40/nv40_screen.c
index 1941598c64..66e84b6890 100644
--- a/src/gallium/drivers/nv40/nv40_screen.c
+++ b/src/gallium/drivers/nv40/nv40_screen.c
@@ -125,7 +125,8 @@ nv40_screen_destroy(struct pipe_screen *pscreen)
 }
 
 struct pipe_screen *
-nv40_screen_create(struct pipe_winsys *ws, unsigned chipset)
+nv40_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws,
+		   unsigned chipset)
 {
 	struct nv40_screen *screen = CALLOC_STRUCT(nv40_screen);
 
@@ -133,6 +134,7 @@ nv40_screen_create(struct pipe_winsys *ws, unsigned chipset)
 		return NULL;
 
 	screen->chipset = chipset;
+	screen->nvws = nvws;
 
 	screen->pipe.winsys = ws;
 	screen->pipe.destroy = nv40_screen_destroy;
diff --git a/src/gallium/drivers/nv40/nv40_screen.h b/src/gallium/drivers/nv40/nv40_screen.h
index b30a6c5ad5..88b8fed26c 100644
--- a/src/gallium/drivers/nv40/nv40_screen.h
+++ b/src/gallium/drivers/nv40/nv40_screen.h
@@ -5,6 +5,8 @@
 
 struct nv40_screen {
 	struct pipe_screen pipe;
+
+	struct nouveau_winsys *nvws;
 	unsigned chipset;
 };
 
@@ -14,7 +16,4 @@ nv40_screen(struct pipe_screen *screen)
 	return (struct nv40_screen *)screen;
 }
 
-extern struct pipe_screen *
-nv40_screen_create(struct pipe_winsys *winsys, unsigned chipset);
-
 #endif
diff --git a/src/gallium/drivers/nv50/nv50_context.c b/src/gallium/drivers/nv50/nv50_context.c
index 98022809a6..e5054e34f6 100644
--- a/src/gallium/drivers/nv50/nv50_context.c
+++ b/src/gallium/drivers/nv50/nv50_context.c
@@ -56,9 +56,10 @@ nv50_init_hwctx(struct nv50_context *nv50, int tesla_class)
 #define GRCLASS5097_CHIPSETS 0x00000000
 #define GRCLASS8297_CHIPSETS 0x00000010
 struct pipe_context *
-nv50_create(struct pipe_screen *pscreen, struct nouveau_winsys *nvws)
+nv50_create(struct pipe_screen *pscreen)
 {
 	struct pipe_winsys *pipe_winsys = pscreen->winsys;
+	struct nouveau_winsys *nvws = nv50_screen(pscreen)->nvws;
 	unsigned chipset = nv50_screen(pscreen)->chipset;
 	struct nv50_context *nv50;
 	int tesla_class, ret;
diff --git a/src/gallium/drivers/nv50/nv50_screen.c b/src/gallium/drivers/nv50/nv50_screen.c
index 8bf82eb0bc..f091779e3b 100644
--- a/src/gallium/drivers/nv50/nv50_screen.c
+++ b/src/gallium/drivers/nv50/nv50_screen.c
@@ -90,7 +90,8 @@ nv50_screen_destroy(struct pipe_screen *pscreen)
 }
 
 struct pipe_screen *
-nv50_screen_create(struct pipe_winsys *ws, unsigned chipset)
+nv50_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws,
+		   unsigned chipset)
 {
 	struct nv50_screen *screen = CALLOC_STRUCT(nv50_screen);
 
@@ -98,6 +99,7 @@ nv50_screen_create(struct pipe_winsys *ws, unsigned chipset)
 		return NULL;
 
 	screen->chipset = chipset;
+	screen->nvws = nvws;
 
 	screen->pipe.winsys = ws;
 
diff --git a/src/gallium/drivers/nv50/nv50_screen.h b/src/gallium/drivers/nv50/nv50_screen.h
index 45ebbb8051..d664816a03 100644
--- a/src/gallium/drivers/nv50/nv50_screen.h
+++ b/src/gallium/drivers/nv50/nv50_screen.h
@@ -5,6 +5,8 @@
 
 struct nv50_screen {
 	struct pipe_screen pipe;
+
+	struct nouveau_winsys *nvws;
 	unsigned chipset;
 };
 
@@ -14,7 +16,4 @@ nv50_screen(struct pipe_screen *screen)
 	return (struct nv50_screen *)screen;
 }
 
-extern struct pipe_screen *
-nv50_screen_create(struct pipe_winsys *winsys, unsigned chipset);
-
 #endif
diff --git a/src/gallium/winsys/dri/nouveau/nouveau_winsys.c b/src/gallium/winsys/dri/nouveau/nouveau_winsys.c
index 1d758e29e7..529f577181 100644
--- a/src/gallium/winsys/dri/nouveau/nouveau_winsys.c
+++ b/src/gallium/winsys/dri/nouveau/nouveau_winsys.c
@@ -73,9 +73,9 @@ nouveau_pipe_create(struct nouveau_context *nv)
 {
 	struct nouveau_winsys *nvws = CALLOC_STRUCT(nouveau_winsys);
 	struct pipe_screen *(*hws_create)(struct pipe_winsys *,
+					  struct nouveau_winsys *,
 					  unsigned chipset);
-	struct pipe_context *(*hw_create)(struct pipe_screen *,
-					  struct nouveau_winsys *);
+	struct pipe_context *(*hw_create)(struct pipe_screen *);
 	struct pipe_winsys *ws;
 	struct pipe_screen *pscreen;
 
@@ -126,7 +126,7 @@ nouveau_pipe_create(struct nouveau_context *nv)
 	nvws->surface_fill	= nouveau_pipe_surface_fill;
 
 	ws = nouveau_create_pipe_winsys(nv);
-	pscreen = hws_create(ws, nv->chipset);
-	return hw_create(pscreen, nvws);
+	pscreen = hws_create(ws, nvws, nv->chipset);
+	return hw_create(pscreen);
 }
 
-- 
cgit v1.2.3


From b2e48f848496d5e315e536688c8c33dfb1fab7eb Mon Sep 17 00:00:00 2001
From: Ben Skeggs <skeggsb@gmail.com>
Date: Wed, 12 Mar 2008 02:39:13 +1100
Subject: nv50: convert to hwctx-in-screen as nv40 is

---
 src/gallium/drivers/nv50/nv50_context.c | 61 ++++-----------------------------
 src/gallium/drivers/nv50/nv50_context.h | 11 +++---
 src/gallium/drivers/nv50/nv50_screen.c  | 49 +++++++++++++++++++++++++-
 src/gallium/drivers/nv50/nv50_screen.h  |  5 +++
 src/gallium/drivers/nv50/nv50_surface.c |  4 +--
 5 files changed, 67 insertions(+), 63 deletions(-)

(limited to 'src/gallium/drivers/nv50/nv50_screen.h')

diff --git a/src/gallium/drivers/nv50/nv50_context.c b/src/gallium/drivers/nv50/nv50_context.c
index c937b8de6d..980d066c84 100644
--- a/src/gallium/drivers/nv50/nv50_context.c
+++ b/src/gallium/drivers/nv50/nv50_context.c
@@ -10,10 +10,11 @@ static void
 nv50_flush(struct pipe_context *pipe, unsigned flags)
 {
 	struct nv50_context *nv50 = (struct nv50_context *)pipe;
-	struct nouveau_winsys *nvws = nv50->nvws;
+	struct nv50_screen *screen = nv50->screen;
+	struct nouveau_winsys *nvws = screen->nvws;
 	
 	if (flags & PIPE_FLUSH_WAIT) {
-		nvws->notifier_reset(nv50->sync, 0);
+		nvws->notifier_reset(screen->sync, 0);
 		BEGIN_RING(tesla, 0x104, 1);
 		OUT_RING  (0);
 		BEGIN_RING(tesla, 0x100, 1);
@@ -23,7 +24,7 @@ nv50_flush(struct pipe_context *pipe, unsigned flags)
 	FIRE_RING();
 
 	if (flags & PIPE_FLUSH_WAIT)
-		nvws->notifier_wait(nv50->sync, 0, 0, 2000);
+		nvws->notifier_wait(screen->sync, 0, 0, 2000);
 }
 
 static void
@@ -35,66 +36,18 @@ nv50_destroy(struct pipe_context *pipe)
 	free(nv50);
 }
 
-static boolean
-nv50_init_hwctx(struct nv50_context *nv50, int tesla_class)
-{
-	struct nouveau_winsys *nvws = nv50->nvws;
-	int ret;
-
-	if ((ret = nvws->grobj_alloc(nvws, tesla_class, &nv50->tesla))) {
-		NOUVEAU_ERR("Error creating 3D object: %d\n", ret);
-		return FALSE;
-	}
-
-	BEGIN_RING(tesla, NV50TCL_DMA_NOTIFY, 1);
-	OUT_RING  (nv50->sync->handle);
-
-	FIRE_RING ();
-	return TRUE;
-}
-
-#define GRCLASS5097_CHIPSETS 0x00000000
-#define GRCLASS8297_CHIPSETS 0x00000010
 struct pipe_context *
 nv50_create(struct pipe_screen *pscreen, unsigned pctx_id)
 {
 	struct pipe_winsys *pipe_winsys = pscreen->winsys;
-	struct nouveau_winsys *nvws = nv50_screen(pscreen)->nvws;
-	unsigned chipset = nv50_screen(pscreen)->chipset;
+	struct nv50_screen *screen = nv50_screen(pscreen);
 	struct nv50_context *nv50;
-	int tesla_class, ret;
-
-	if ((chipset & 0xf0) != 0x50 && (chipset & 0xf0) != 0x80) {
-		NOUVEAU_ERR("Not a G8x chipset\n");
-		return NULL;
-	}
-
-	if (GRCLASS5097_CHIPSETS & (1 << (chipset & 0x0f))) {
-		tesla_class = 0x5097;
-	} else
-	if (GRCLASS8297_CHIPSETS & (1 << (chipset & 0x0f))) {
-		tesla_class = 0x8297;
-	} else {
-		NOUVEAU_ERR("Unknown G8x chipset: NV%02x\n", chipset);
-		return NULL;
-	}
 
 	nv50 = CALLOC_STRUCT(nv50_context);
 	if (!nv50)
 		return NULL;
-	nv50->chipset = chipset;
-	nv50->nvws = nvws;
-
-	if ((ret = nvws->notifier_alloc(nvws, 1, &nv50->sync))) {
-		NOUVEAU_ERR("Error creating notifier object: %d\n", ret);
-		free(nv50);
-		return NULL;
-	}
-
-	if (!nv50_init_hwctx(nv50, tesla_class)) {
-		free(nv50);
-		return NULL;
-	}
+	nv50->screen = screen;
+	nv50->pctx_id = pctx_id;
 
 	nv50->pipe.winsys = pipe_winsys;
 	nv50->pipe.screen = pscreen;
diff --git a/src/gallium/drivers/nv50/nv50_context.h b/src/gallium/drivers/nv50/nv50_context.h
index a529bf3c3e..e2656bd539 100644
--- a/src/gallium/drivers/nv50/nv50_context.h
+++ b/src/gallium/drivers/nv50/nv50_context.h
@@ -11,10 +11,11 @@
 #include "nouveau/nouveau_gldefs.h"
 
 #define NOUVEAU_PUSH_CONTEXT(ctx)                                              \
-	struct nv50_context *ctx = nv50
+	struct nv50_screen *ctx = nv50->screen
 #include "nouveau/nouveau_push.h"
 
 #include "nv50_state.h"
+#include "nv50_screen.h"
 
 #define NOUVEAU_ERR(fmt, args...) \
 	fprintf(stderr, "%s:%d -  "fmt, __func__, __LINE__, ##args);
@@ -23,13 +24,11 @@
 
 struct nv50_context {
 	struct pipe_context pipe;
-	struct nouveau_winsys *nvws;
 
-	struct draw_context *draw;
+	struct nv50_screen *screen;
+	unsigned pctx_id;
 
-	int chipset;
-	struct nouveau_grobj *tesla;
-	struct nouveau_notifier *sync;
+	struct draw_context *draw;
 };
 
 
diff --git a/src/gallium/drivers/nv50/nv50_screen.c b/src/gallium/drivers/nv50/nv50_screen.c
index 77ceb678f2..f8fd11dc5f 100644
--- a/src/gallium/drivers/nv50/nv50_screen.c
+++ b/src/gallium/drivers/nv50/nv50_screen.c
@@ -4,6 +4,11 @@
 #include "nv50_context.h"
 #include "nv50_screen.h"
 
+#include "nouveau/nouveau_stateobj.h"
+
+#define GRCLASS5097_CHIPSETS 0x00000000
+#define GRCLASS8297_CHIPSETS 0x00000010
+
 static boolean
 nv50_screen_is_format_supported(struct pipe_screen *pscreen,
 				enum pipe_format format, uint type)
@@ -96,13 +101,55 @@ nv50_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws,
 		   unsigned chipset)
 {
 	struct nv50_screen *screen = CALLOC_STRUCT(nv50_screen);
+	struct nouveau_stateobj *so;
+	unsigned tesla_class = 0, ret;
 
 	if (!screen)
 		return NULL;
-
 	screen->chipset = chipset;
 	screen->nvws = nvws;
 
+	/* 3D object */
+	if ((chipset & 0xf0) != 0x50 && (chipset & 0xf0) != 0x80) {
+		NOUVEAU_ERR("Not a G8x chipset\n");
+		nv50_screen_destroy(&screen->pipe);
+		return NULL;
+	}
+
+	if (GRCLASS5097_CHIPSETS & (1 << (chipset & 0x0f))) {
+		tesla_class = 0x5097;
+	} else
+	if (GRCLASS8297_CHIPSETS & (1 << (chipset & 0x0f))) {
+		tesla_class = 0x8297;
+	} else {
+		NOUVEAU_ERR("Unknown G8x chipset: NV%02x\n", chipset);
+		nv50_screen_destroy(&screen->pipe);
+		return NULL;
+	}
+
+	ret = nvws->grobj_alloc(nvws, tesla_class, &screen->tesla);
+	if (ret) {
+		NOUVEAU_ERR("Error creating 3D object: %d\n", ret);
+		nv50_screen_destroy(&screen->pipe);
+		return NULL;
+	}
+
+	/* Sync notifier */
+	ret = nvws->notifier_alloc(nvws, 1, &screen->sync);
+	if (ret) {
+		NOUVEAU_ERR("Error creating notifier object: %d\n", ret);
+		nv50_screen_destroy(&screen->pipe);
+		return NULL;
+	}
+
+	/* Static tesla init */
+	so = so_new(128, 0);
+	so_method(so, screen->tesla, NV50TCL_DMA_NOTIFY, 1);
+	so_data  (so, screen->sync->handle);
+	so_emit(nvws, so);
+	so_ref(NULL, &so);
+	nvws->push_flush(nvws->channel, 0);
+
 	screen->pipe.winsys = ws;
 
 	screen->pipe.destroy = nv50_screen_destroy;
diff --git a/src/gallium/drivers/nv50/nv50_screen.h b/src/gallium/drivers/nv50/nv50_screen.h
index d664816a03..6e4120d669 100644
--- a/src/gallium/drivers/nv50/nv50_screen.h
+++ b/src/gallium/drivers/nv50/nv50_screen.h
@@ -8,6 +8,11 @@ struct nv50_screen {
 
 	struct nouveau_winsys *nvws;
 	unsigned chipset;
+
+	unsigned cur_pctx;
+
+	struct nouveau_grobj *tesla;
+	struct nouveau_notifier *sync;
 };
 
 static INLINE struct nv50_screen *
diff --git a/src/gallium/drivers/nv50/nv50_surface.c b/src/gallium/drivers/nv50/nv50_surface.c
index 39cf675a57..4e294198b0 100644
--- a/src/gallium/drivers/nv50/nv50_surface.c
+++ b/src/gallium/drivers/nv50/nv50_surface.c
@@ -40,7 +40,7 @@ nv50_surface_copy(struct pipe_context *pipe, unsigned flip,
 		  unsigned width, unsigned height)
 {
 	struct nv50_context *nv50 = (struct nv50_context *)pipe;
-	struct nouveau_winsys *nvws = nv50->nvws;
+	struct nouveau_winsys *nvws = nv50->screen->nvws;
 
 	nvws->surface_copy(nvws, dest, destx, desty, src, srcx, srcy,
 			   width, height);
@@ -52,7 +52,7 @@ nv50_surface_fill(struct pipe_context *pipe, struct pipe_surface *dest,
 		  unsigned height, unsigned value)
 {
 	struct nv50_context *nv50 = (struct nv50_context *)pipe;
-	struct nouveau_winsys *nvws = nv50->nvws;
+	struct nouveau_winsys *nvws = nv50->screen->nvws;
 
 	nvws->surface_fill(nvws, dest, destx, desty, width, height, value);
 }
-- 
cgit v1.2.3


From 1ef08564d2a201a422db772a6bb23d1129888304 Mon Sep 17 00:00:00 2001
From: Ben Skeggs <skeggsb@gmail.com>
Date: Tue, 13 May 2008 12:16:35 +1000
Subject: nouveau: remove chipset fields in all nv pipe driver context/screen
 structs.

---
 src/gallium/drivers/nouveau/nouveau_winsys.h    | 13 +++++--------
 src/gallium/drivers/nv10/nv10_context.c         |  2 --
 src/gallium/drivers/nv10/nv10_context.h         |  2 --
 src/gallium/drivers/nv10/nv10_screen.c          |  8 ++++----
 src/gallium/drivers/nv10/nv10_screen.h          |  1 -
 src/gallium/drivers/nv30/nv30_context.c         |  2 --
 src/gallium/drivers/nv30/nv30_context.h         |  2 --
 src/gallium/drivers/nv30/nv30_screen.c          |  8 ++++----
 src/gallium/drivers/nv30/nv30_screen.h          |  1 -
 src/gallium/drivers/nv40/nv40_context.c         |  2 --
 src/gallium/drivers/nv40/nv40_context.h         |  2 --
 src/gallium/drivers/nv40/nv40_screen.c          |  8 ++++----
 src/gallium/drivers/nv40/nv40_screen.h          |  1 -
 src/gallium/drivers/nv50/nv50_screen.c          |  8 ++++----
 src/gallium/drivers/nv50/nv50_screen.h          |  1 -
 src/gallium/winsys/dri/nouveau/nouveau_winsys.c |  5 ++---
 16 files changed, 23 insertions(+), 43 deletions(-)

(limited to 'src/gallium/drivers/nv50/nv50_screen.h')

diff --git a/src/gallium/drivers/nouveau/nouveau_winsys.h b/src/gallium/drivers/nouveau/nouveau_winsys.h
index fbde7adc6a..07a41dcf4a 100644
--- a/src/gallium/drivers/nouveau/nouveau_winsys.h
+++ b/src/gallium/drivers/nouveau/nouveau_winsys.h
@@ -8,6 +8,7 @@
 #include "nouveau/nouveau_bo.h"
 #include "nouveau/nouveau_channel.h"
 #include "nouveau/nouveau_class.h"
+#include "nouveau/nouveau_device.h"
 #include "nouveau/nouveau_grobj.h"
 #include "nouveau/nouveau_notifier.h"
 #include "nouveau/nouveau_resource.h"
@@ -56,29 +57,25 @@ struct nouveau_winsys {
 };
 
 extern struct pipe_screen *
-nv10_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *,
-		   unsigned chipset);
+nv10_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *);
 
 extern struct pipe_context *
 nv10_create(struct pipe_screen *, unsigned pctx_id);
 
 extern struct pipe_screen *
-nv30_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *,
-		   unsigned chipset);
+nv30_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *);
 
 extern struct pipe_context *
 nv30_create(struct pipe_screen *, unsigned pctx_id);
 
 extern struct pipe_screen *
-nv40_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *,
-		   unsigned chipset);
+nv40_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *);
 
 extern struct pipe_context *
 nv40_create(struct pipe_screen *, unsigned pctx_id);
 
 extern struct pipe_screen *
-nv50_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *,
-		   unsigned chipset);
+nv50_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *);
 
 extern struct pipe_context *
 nv50_create(struct pipe_screen *, unsigned pctx_id);
diff --git a/src/gallium/drivers/nv10/nv10_context.c b/src/gallium/drivers/nv10/nv10_context.c
index a25c082a5d..79253f8a2b 100644
--- a/src/gallium/drivers/nv10/nv10_context.c
+++ b/src/gallium/drivers/nv10/nv10_context.c
@@ -264,7 +264,6 @@ nv10_create(struct pipe_screen *pscreen, unsigned pctx_id)
 	struct nv10_screen *screen = nv10_screen(pscreen);
 	struct pipe_winsys *ws = pscreen->winsys;
 	struct nv10_context *nv10;
-	unsigned chipset = screen->chipset;
 	struct nouveau_winsys *nvws = screen->nvws;
 
 	nv10 = CALLOC(1, sizeof(struct nv10_context));
@@ -273,7 +272,6 @@ nv10_create(struct pipe_screen *pscreen, unsigned pctx_id)
 	nv10->screen = screen;
 	nv10->pctx_id = pctx_id;
 
-	nv10->chipset = chipset;
 	nv10->nvws = nvws;
 
 	nv10->pipe.winsys = ws;
diff --git a/src/gallium/drivers/nv10/nv10_context.h b/src/gallium/drivers/nv10/nv10_context.h
index 1b794c1872..433d04dc2a 100644
--- a/src/gallium/drivers/nv10/nv10_context.h
+++ b/src/gallium/drivers/nv10/nv10_context.h
@@ -43,8 +43,6 @@ struct nv10_context {
 
 	struct draw_context *draw;
 
-	int chipset;
-
 	uint32_t dirty;
 
 	struct nv10_sampler_state *tex_sampler[PIPE_MAX_SAMPLERS];
diff --git a/src/gallium/drivers/nv10/nv10_screen.c b/src/gallium/drivers/nv10/nv10_screen.c
index cf6b2fac84..fad96058b8 100644
--- a/src/gallium/drivers/nv10/nv10_screen.c
+++ b/src/gallium/drivers/nv10/nv10_screen.c
@@ -8,9 +8,10 @@ static const char *
 nv10_screen_get_name(struct pipe_screen *screen)
 {
 	struct nv10_screen *nv10screen = nv10_screen(screen);
+	struct nouveau_device *dev = nv10screen->nvws->channel->device;
 	static char buffer[128];
 
-	snprintf(buffer, sizeof(buffer), "NV%02X", nv10screen->chipset);
+	snprintf(buffer, sizeof(buffer), "NV%02X", dev->chipset);
 	return buffer;
 }
 
@@ -129,16 +130,15 @@ nv10_screen_destroy(struct pipe_screen *pscreen)
 }
 
 struct pipe_screen *
-nv10_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws,
-		   unsigned chipset)
+nv10_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws)
 {
 	struct nv10_screen *screen = CALLOC_STRUCT(nv10_screen);
 	unsigned celsius_class;
+	unsigned chipset = nvws->channel->device->chipset;
 	int ret;
 
 	if (!screen)
 		return NULL;
-	screen->chipset = chipset;
 	screen->nvws = nvws;
 
 	/* 3D object */
diff --git a/src/gallium/drivers/nv10/nv10_screen.h b/src/gallium/drivers/nv10/nv10_screen.h
index 4192fe11ef..3f8750a13f 100644
--- a/src/gallium/drivers/nv10/nv10_screen.h
+++ b/src/gallium/drivers/nv10/nv10_screen.h
@@ -7,7 +7,6 @@ struct nv10_screen {
 	struct pipe_screen pipe;
 
 	struct nouveau_winsys *nvws;
-	unsigned chipset;
 
 	/* HW graphics objects */
 	struct nouveau_grobj *celsius;
diff --git a/src/gallium/drivers/nv30/nv30_context.c b/src/gallium/drivers/nv30/nv30_context.c
index 4b0892971e..7a2fee7875 100644
--- a/src/gallium/drivers/nv30/nv30_context.c
+++ b/src/gallium/drivers/nv30/nv30_context.c
@@ -44,7 +44,6 @@ nv30_create(struct pipe_screen *pscreen, unsigned pctx_id)
 	struct nv30_screen *screen = nv30_screen(pscreen);
 	struct pipe_winsys *ws = pscreen->winsys;
 	struct nv30_context *nv30;
-	unsigned chipset = screen->chipset;
 	struct nouveau_winsys *nvws = screen->nvws;
 
 	nv30 = CALLOC(1, sizeof(struct nv30_context));
@@ -53,7 +52,6 @@ nv30_create(struct pipe_screen *pscreen, unsigned pctx_id)
 	nv30->screen = screen;
 	nv30->pctx_id = pctx_id;
 
-	nv30->chipset = chipset;
 	nv30->nvws = nvws;
 
 	nv30->pipe.winsys = ws;
diff --git a/src/gallium/drivers/nv30/nv30_context.h b/src/gallium/drivers/nv30/nv30_context.h
index 1a016405e6..f49450f82d 100644
--- a/src/gallium/drivers/nv30/nv30_context.h
+++ b/src/gallium/drivers/nv30/nv30_context.h
@@ -85,8 +85,6 @@ struct nv30_context {
 
 	struct draw_context *draw;
 
-	int chipset;
-
 	uint32_t dirty;
 
 	struct nv30_sampler_state *tex_sampler[PIPE_MAX_SAMPLERS];
diff --git a/src/gallium/drivers/nv30/nv30_screen.c b/src/gallium/drivers/nv30/nv30_screen.c
index 1de4507904..9b0ac55ff8 100644
--- a/src/gallium/drivers/nv30/nv30_screen.c
+++ b/src/gallium/drivers/nv30/nv30_screen.c
@@ -12,9 +12,10 @@ static const char *
 nv30_screen_get_name(struct pipe_screen *pscreen)
 {
 	struct nv30_screen *screen = nv30_screen(pscreen);
+	struct nouveau_device *dev = screen->nvws->channel->device;
 	static char buffer[128];
 
-	snprintf(buffer, sizeof(buffer), "NV%02X", screen->chipset);
+	snprintf(buffer, sizeof(buffer), "NV%02X", dev->chipset);
 	return buffer;
 }
 
@@ -140,17 +141,16 @@ nv30_screen_destroy(struct pipe_screen *pscreen)
 }
 
 struct pipe_screen *
-nv30_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws,
-		   unsigned chipset)
+nv30_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws)
 {
 	struct nv30_screen *screen = CALLOC_STRUCT(nv30_screen);
 	struct nouveau_stateobj *so;
 	unsigned rankine_class = 0;
+	unsigned chipset = nvws->channel->device->chipset;
 	int ret, i;
 
 	if (!screen)
 		return NULL;
-	screen->chipset = chipset;
 	screen->nvws = nvws;
 
 	/* 3D object */
diff --git a/src/gallium/drivers/nv30/nv30_screen.h b/src/gallium/drivers/nv30/nv30_screen.h
index 56f8776a17..816ece94c4 100644
--- a/src/gallium/drivers/nv30/nv30_screen.h
+++ b/src/gallium/drivers/nv30/nv30_screen.h
@@ -7,7 +7,6 @@ struct nv30_screen {
 	struct pipe_screen pipe;
 
 	struct nouveau_winsys *nvws;
-	unsigned chipset;
 
 	/* HW graphics objects */
 	struct nouveau_grobj *rankine;
diff --git a/src/gallium/drivers/nv40/nv40_context.c b/src/gallium/drivers/nv40/nv40_context.c
index f9c93f7a2d..d9d9accea8 100644
--- a/src/gallium/drivers/nv40/nv40_context.c
+++ b/src/gallium/drivers/nv40/nv40_context.c
@@ -38,7 +38,6 @@ nv40_create(struct pipe_screen *pscreen, unsigned pctx_id)
 	struct nv40_screen *screen = nv40_screen(pscreen);
 	struct pipe_winsys *ws = pscreen->winsys;
 	struct nv40_context *nv40;
-	unsigned chipset = screen->chipset;
 	struct nouveau_winsys *nvws = screen->nvws;
 
 	nv40 = CALLOC(1, sizeof(struct nv40_context));
@@ -47,7 +46,6 @@ nv40_create(struct pipe_screen *pscreen, unsigned pctx_id)
 	nv40->screen = screen;
 	nv40->pctx_id = pctx_id;
 
-	nv40->chipset = chipset;
 	nv40->nvws = nvws;
 
 	nv40->pipe.winsys = ws;
diff --git a/src/gallium/drivers/nv40/nv40_context.h b/src/gallium/drivers/nv40/nv40_context.h
index 24e8cd2337..77b3da0ab9 100644
--- a/src/gallium/drivers/nv40/nv40_context.h
+++ b/src/gallium/drivers/nv40/nv40_context.h
@@ -111,8 +111,6 @@ struct nv40_context {
 
 	struct draw_context *draw;
 
-	int chipset;
-
 	/* HW state derived from pipe states */
 	struct nv40_state state;
 	struct {
diff --git a/src/gallium/drivers/nv40/nv40_screen.c b/src/gallium/drivers/nv40/nv40_screen.c
index 7f68539a85..1bf0931e4d 100644
--- a/src/gallium/drivers/nv40/nv40_screen.c
+++ b/src/gallium/drivers/nv40/nv40_screen.c
@@ -12,9 +12,10 @@ static const char *
 nv40_screen_get_name(struct pipe_screen *pscreen)
 {
 	struct nv40_screen *screen = nv40_screen(pscreen);
+	struct nouveau_device *dev = screen->nvws->channel->device;
 	static char buffer[128];
 
-	snprintf(buffer, sizeof(buffer), "NV%02X", screen->chipset);
+	snprintf(buffer, sizeof(buffer), "NV%02X", dev->chipset);
 	return buffer;
 }
 
@@ -149,17 +150,16 @@ nv40_screen_destroy(struct pipe_screen *pscreen)
 }
 
 struct pipe_screen *
-nv40_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws,
-		   unsigned chipset)
+nv40_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws)
 {
 	struct nv40_screen *screen = CALLOC_STRUCT(nv40_screen);
 	struct nouveau_stateobj *so;
 	unsigned curie_class;
+	unsigned chipset = nvws->channel->device;
 	int ret;
 
 	if (!screen)
 		return NULL;
-	screen->chipset = chipset;
 	screen->nvws = nvws;
 
 	/* 3D object */
diff --git a/src/gallium/drivers/nv40/nv40_screen.h b/src/gallium/drivers/nv40/nv40_screen.h
index 3ea78aadfd..c04a1275a0 100644
--- a/src/gallium/drivers/nv40/nv40_screen.h
+++ b/src/gallium/drivers/nv40/nv40_screen.h
@@ -7,7 +7,6 @@ struct nv40_screen {
 	struct pipe_screen pipe;
 
 	struct nouveau_winsys *nvws;
-	unsigned chipset;
 
 	unsigned cur_pctx;
 
diff --git a/src/gallium/drivers/nv50/nv50_screen.c b/src/gallium/drivers/nv50/nv50_screen.c
index e4ca72c717..d069639dc4 100644
--- a/src/gallium/drivers/nv50/nv50_screen.c
+++ b/src/gallium/drivers/nv50/nv50_screen.c
@@ -45,9 +45,10 @@ static const char *
 nv50_screen_get_name(struct pipe_screen *pscreen)
 {
 	struct nv50_screen *screen = nv50_screen(pscreen);
+	struct nouveau_device *dev = screen->nvws->channel->device;
 	static char buffer[128];
 
-	snprintf(buffer, sizeof(buffer), "NV%02X", screen->chipset);
+	snprintf(buffer, sizeof(buffer), "NV%02X", dev->chipset);
 	return buffer;
 }
 
@@ -123,17 +124,16 @@ nv50_screen_destroy(struct pipe_screen *pscreen)
 }
 
 struct pipe_screen *
-nv50_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws,
-		   unsigned chipset)
+nv50_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws)
 {
 	struct nv50_screen *screen = CALLOC_STRUCT(nv50_screen);
 	struct nouveau_stateobj *so;
 	unsigned tesla_class = 0, ret;
+	unsigned chipset = nvws->channel->device->chipset;
 	int i;
 
 	if (!screen)
 		return NULL;
-	screen->chipset = chipset;
 	screen->nvws = nvws;
 
 	/* 3D object */
diff --git a/src/gallium/drivers/nv50/nv50_screen.h b/src/gallium/drivers/nv50/nv50_screen.h
index 6e4120d669..5a2732e37f 100644
--- a/src/gallium/drivers/nv50/nv50_screen.h
+++ b/src/gallium/drivers/nv50/nv50_screen.h
@@ -7,7 +7,6 @@ struct nv50_screen {
 	struct pipe_screen pipe;
 
 	struct nouveau_winsys *nvws;
-	unsigned chipset;
 
 	unsigned cur_pctx;
 
diff --git a/src/gallium/winsys/dri/nouveau/nouveau_winsys.c b/src/gallium/winsys/dri/nouveau/nouveau_winsys.c
index 635fd478a9..5eabbc8893 100644
--- a/src/gallium/winsys/dri/nouveau/nouveau_winsys.c
+++ b/src/gallium/winsys/dri/nouveau/nouveau_winsys.c
@@ -91,8 +91,7 @@ nouveau_pipe_create(struct nouveau_context *nv)
 	struct nouveau_channel_context *nvc = nv->nvc;
 	struct nouveau_winsys *nvws = CALLOC_STRUCT(nouveau_winsys);
 	struct pipe_screen *(*hws_create)(struct pipe_winsys *,
-					  struct nouveau_winsys *,
-					  unsigned chipset);
+					  struct nouveau_winsys *);
 	struct pipe_context *(*hw_create)(struct pipe_screen *, unsigned);
 	struct pipe_winsys *ws;
 	unsigned chipset = nv->nv_screen->device->chipset;
@@ -152,7 +151,7 @@ nouveau_pipe_create(struct nouveau_context *nv)
 	ws = nouveau_create_pipe_winsys(nv);
 
 	if (!nvc->pscreen)
-		nvc->pscreen = hws_create(ws, nvws, chipset);
+		nvc->pscreen = hws_create(ws, nvws);
 	nvc->pctx[nv->pctx_id] = hw_create(nvc->pscreen, nv->pctx_id);
 	return nvc->pctx[nv->pctx_id];
 }
-- 
cgit v1.2.3


From f722fd937db2f3cacf1947d538c66528fd16eb89 Mon Sep 17 00:00:00 2001
From: Ben Skeggs <skeggsb@gmail.com>
Date: Sun, 1 Jun 2008 22:41:40 +1000
Subject: nv50: import current "state of the art" nv50 code

---
 src/gallium/drivers/nv50/Makefile              |   1 +
 src/gallium/drivers/nv50/nv50_context.h        |  18 +
 src/gallium/drivers/nv50/nv50_program.c        | 742 +++++++++++++++++++++++++
 src/gallium/drivers/nv50/nv50_screen.c         |  68 ++-
 src/gallium/drivers/nv50/nv50_screen.h         |   3 +
 src/gallium/drivers/nv50/nv50_state.c          |  54 +-
 src/gallium/drivers/nv50/nv50_state.h          |  30 +
 src/gallium/drivers/nv50/nv50_state_validate.c |  29 +-
 src/gallium/drivers/nv50/nv50_vbo.c            |  85 ++-
 9 files changed, 1017 insertions(+), 13 deletions(-)
 create mode 100644 src/gallium/drivers/nv50/nv50_program.c

(limited to 'src/gallium/drivers/nv50/nv50_screen.h')

diff --git a/src/gallium/drivers/nv50/Makefile b/src/gallium/drivers/nv50/Makefile
index a62a4d961b..2cc6e9de28 100644
--- a/src/gallium/drivers/nv50/Makefile
+++ b/src/gallium/drivers/nv50/Makefile
@@ -9,6 +9,7 @@ DRIVER_SOURCES = \
 	nv50_draw.c \
 	nv50_miptree.c \
 	nv50_query.c \
+	nv50_program.c \
 	nv50_screen.c \
 	nv50_state.c \
 	nv50_state_validate.c \
diff --git a/src/gallium/drivers/nv50/nv50_context.h b/src/gallium/drivers/nv50/nv50_context.h
index e68c702dea..d4d716b94b 100644
--- a/src/gallium/drivers/nv50/nv50_context.h
+++ b/src/gallium/drivers/nv50/nv50_context.h
@@ -9,6 +9,7 @@
 
 #include "nouveau/nouveau_winsys.h"
 #include "nouveau/nouveau_gldefs.h"
+#include "nouveau/nouveau_stateobj.h"
 
 #define NOUVEAU_PUSH_CONTEXT(ctx)                                              \
 	struct nv50_screen *ctx = nv50->screen
@@ -30,6 +31,9 @@
 #define NV50_NEW_VIEWPORT	(1 << 5)
 #define NV50_NEW_RASTERIZER	(1 << 6)
 #define NV50_NEW_FRAMEBUFFER	(1 << 7)
+#define NV50_NEW_VERTPROG	(1 << 8)
+#define NV50_NEW_FRAGPROG	(1 << 9)
+#define NV50_NEW_ARRAYS		(1 << 10)
 
 struct nv50_blend_stateobj {
 	struct pipe_blend_state pipe;
@@ -63,6 +67,13 @@ struct nv50_context {
 	struct pipe_scissor_state scissor;
 	struct pipe_viewport_state viewport;
 	struct pipe_framebuffer_state framebuffer;
+	struct nv50_program *vertprog;
+	struct nv50_program *fragprog;
+	struct pipe_buffer *constbuf[PIPE_SHADER_TYPES];
+	struct pipe_vertex_buffer vtxbuf[PIPE_MAX_ATTRIBS];
+	unsigned vtxbuf_nr;
+	struct pipe_vertex_element vtxelt[PIPE_MAX_ATTRIBS];
+	unsigned vtxelt_nr;
 };
 
 static INLINE struct nv50_context *
@@ -88,11 +99,18 @@ extern boolean nv50_draw_elements(struct pipe_context *pipe,
 				  unsigned indexSize,
 				  unsigned mode, unsigned start,
 				  unsigned count);
+extern void nv50_vbo_validate(struct nv50_context *nv50);
 
 /* nv50_clear.c */
 extern void nv50_clear(struct pipe_context *pipe, struct pipe_surface *ps,
 		       unsigned clearValue);
 
+/* nv50_program.c */
+extern void nv50_vertprog_validate(struct nv50_context *nv50);
+extern void nv50_fragprog_validate(struct nv50_context *nv50);
+extern void nv50_program_destroy(struct nv50_context *nv50, struct nv50_program *p);
+
+/* nv50_state_validate.c */
 extern boolean nv50_state_validate(struct nv50_context *nv50);
 
 #endif
diff --git a/src/gallium/drivers/nv50/nv50_program.c b/src/gallium/drivers/nv50/nv50_program.c
new file mode 100644
index 0000000000..30953b7d8a
--- /dev/null
+++ b/src/gallium/drivers/nv50/nv50_program.c
@@ -0,0 +1,742 @@
+#include "pipe/p_context.h"
+#include "pipe/p_defines.h"
+#include "pipe/p_state.h"
+#include "pipe/p_inlines.h"
+
+#include "pipe/p_shader_tokens.h"
+#include "tgsi/util/tgsi_parse.h"
+#include "tgsi/util/tgsi_util.h"
+
+#include "nv50_context.h"
+#include "nv50_state.h"
+
+#define OP_MOV 0x001
+#define OP_RCP 0x009
+#define OP_ADD 0x00b
+#define OP_MUL 0x00c
+#define NV50_SU_MAX_TEMP 64
+
+struct nv50_reg {
+	enum {
+		P_TEMP,
+		P_ATTR,
+		P_RESULT,
+		P_CONST,
+		P_IMMD
+	} type;
+	int index;
+
+	int hw;
+};
+
+struct nv50_pc {
+	struct nv50_program *p;
+
+	/* hw resources */
+	struct nv50_reg *r_temp[NV50_SU_MAX_TEMP];
+
+	/* tgsi resources */
+	struct nv50_reg *temp;
+	int temp_nr;
+	struct nv50_reg *attr;
+	int attr_nr;
+	struct nv50_reg *result;
+	int result_nr;
+	struct nv50_reg *param;
+	int param_nr;
+	struct nv50_reg *immd;
+	float *immd_buf;
+	int immd_nr;
+};
+
+static void
+alloc_reg(struct nv50_pc *pc, struct nv50_reg *reg)
+{
+	int i;
+
+	if (reg->type != P_TEMP || reg->hw >= 0)
+		return;
+
+	for (i = 0; i < NV50_SU_MAX_TEMP; i++) {
+		if (!(pc->r_temp[i])) {
+			pc->r_temp[i] = reg;
+			reg->hw = i;
+			if (pc->p->cfg.vp.high_temp < (i + 1))
+				pc->p->cfg.vp.high_temp = i + 1;
+			return;
+		}
+	}
+
+	assert(0);
+}
+
+static struct nv50_reg *
+alloc_temp(struct nv50_pc *pc, struct nv50_reg *dst)
+{
+	struct nv50_reg *r;
+	int i;
+
+	if (dst && dst->type == P_TEMP && dst->hw == -1)
+		return dst;
+
+	for (i = 0; i < NV50_SU_MAX_TEMP; i++) {
+		if (!pc->r_temp[i]) {
+			r = CALLOC_STRUCT(nv50_reg);
+			r->type = P_TEMP;
+			r->index = -1;
+			r->hw = i;
+			pc->r_temp[i] = r;
+			return r;
+		}
+	}
+
+	assert(0);
+	return NULL;
+}
+
+static void
+free_temp(struct nv50_pc *pc, struct nv50_reg *r)
+{
+	if (r->index == -1) {
+		FREE(pc->r_temp[r->hw]);
+		pc->r_temp[r->hw] = NULL;
+	}
+}
+
+#if 0
+static struct nv50_reg *
+constant(struct nv50_pc *pc, int pipe, int c, float v)
+{
+	struct nv50_reg *r = CALLOC_STRUCT(nv50_reg);
+	struct nv50_program *p = pc->p;
+	struct nv50_program_data *pd;
+	int idx;
+
+	if (pipe >= 0) {
+		for (idx = 0; idx < p->nr_consts; idx++) {
+			if (p->consts[idx].index == pipe)
+				return nv40_sr(NV40SR_CONST, idx);
+		}
+	}
+
+	idx = p->nr_consts++;
+	p->consts = realloc(p->consts, sizeof(*pd) * p->nr_consts);
+	pd = &p->consts[idx];
+
+	pd->index = pipe;
+	pd->component = c;
+	pd->value = v;
+	return nv40_sr(NV40SR_CONST, idx);
+}
+#endif
+
+static void
+emit(struct nv50_pc *pc, unsigned op, struct nv50_reg *dst,
+     struct nv50_reg *src0, struct nv50_reg *src1, struct nv50_reg *src2)
+{
+	struct nv50_program *p = pc->p;
+	struct nv50_reg *tmp = NULL, *tmp2 = NULL;
+	unsigned inst[2] = { 0, 0 };
+
+	/* Grr.. Fun restrictions on where attribs can be sourced from.. */
+	if (src1 && src1->type == P_ATTR) {
+		tmp = alloc_temp(pc, dst);
+		emit(pc, 1, tmp, src1, NULL, NULL);
+		src1 = tmp;
+	}
+
+	if (src2 && src2->type == P_ATTR) {
+		tmp2 = alloc_temp(pc, dst);
+		emit(pc, 1, tmp2, src2, NULL, NULL);
+		src2 = tmp2;
+	}
+
+	/* Get this out of the way first.  What type of opcode do we
+	 * want/need to build?
+	 */
+	if ((op & 0x3f0) || dst->type == P_RESULT ||
+	    (src0 && src0->type == P_ATTR) || src1 || src2)
+		inst[0] |= 0x00000001;
+
+	if (inst[0] & 0x00000001) {
+		inst[0] |= ((op & 0xf) << 28);
+		inst[1] |= ((op >> 4) << 26);
+
+		alloc_reg(pc, dst);
+		if (dst->type == P_RESULT)
+			inst[1] |= 0x00000008;
+		inst[0] |= (dst->hw << 2);
+
+		if (src0) {
+			if (src0->type == P_ATTR)
+				inst[1] |= 0x00200000;
+			else
+			if (src0->type == P_CONST || src0->type == P_IMMD)
+				assert(0);
+			alloc_reg(pc, src0);
+			inst[0] |= (src0->hw << 9);
+		}
+
+		if (src1) {
+			if (src1->type == P_CONST || src1->type == P_IMMD) {
+				inst[0] |= 0x00800000; /* src1 is const */
+				/*XXX: does src1 come from "src2" now? */
+				alloc_reg(pc, src1);
+				inst[0] |= (src1->hw << 16);
+			} else {
+				alloc_reg(pc, src1);
+				if (op == 0xc || op == 0xe)
+					inst[0] |= (src1->hw << 16);
+				else
+					inst[1] |= (src1->hw << 14);
+			}
+		} else {
+			inst[1] |= 0x0003c000; /*XXX FIXME */
+		}
+
+		if (src2) {
+			if (src2->type == P_CONST || src2->type == P_IMMD) {
+				inst[0] |= 0x01000000; /* src2 is const */
+				inst[1] |= (src2->hw << 14);
+			} else {
+				alloc_reg(pc, src2);
+				if (inst[0] & 0x00800000 || op ==0xe)
+					inst[1] |= (src2->hw << 14);
+				else
+					inst[0] |= (src2->hw << 16);
+			}
+		}
+
+		/*XXX: FIXME */
+		if (op == 0xb || op == 0xc || op == 0x9 || op == 0xe) {
+			/* 0x04000000 negates arg0 */
+			/* 0x08000000 negates arg1 */
+			/*XXX: true for !0xb also ? */
+			inst[1] |= 0x00000780;
+		} else {
+			/* 0x04000000 == arg0 32 bit, otherwise 16 bit */
+			inst[1] |= 0x04000780;
+		}
+	} else {
+		inst[0] |= ((op & 0xf) << 28);
+
+		alloc_reg(pc, dst);
+		inst[0] |= (dst->hw << 2);
+
+		if (src0) {
+			alloc_reg(pc, src0);
+			inst[0] |= (src0->hw << 9);
+		}
+
+		/*XXX: NFI if this even works - probably not.. */
+		if (src1) {
+			alloc_reg(pc, src1);
+			inst[0] |= (src1->hw << 16);
+		}
+	}
+
+	if (tmp) free_temp(pc, tmp);
+	if (tmp2) free_temp(pc, tmp2);
+
+	if (inst[0] & 1) {
+		p->insns_nr += 2;
+		p->insns = realloc(p->insns, sizeof(unsigned) * p->insns_nr);
+		memcpy(p->insns + (p->insns_nr - 2), inst, sizeof(unsigned)*2);
+	} else {
+		p->insns_nr += 1;
+		p->insns = realloc(p->insns, sizeof(unsigned) * p->insns_nr);
+		memcpy(p->insns + (p->insns_nr - 1), inst, sizeof(unsigned));
+	}
+}
+
+static struct nv50_reg *
+tgsi_dst(struct nv50_pc *pc, int c, const struct tgsi_full_dst_register *dst)
+{
+	switch (dst->DstRegister.File) {
+	case TGSI_FILE_TEMPORARY:
+		return &pc->temp[dst->DstRegister.Index * 4 + c];
+	case TGSI_FILE_OUTPUT:
+		return &pc->result[dst->DstRegister.Index * 4 + c];
+	case TGSI_FILE_NULL:
+		return NULL;
+	default:
+		break;
+	}
+
+	return NULL;
+}
+
+static struct nv50_reg *
+tgsi_src(struct nv50_pc *pc, int c, const struct tgsi_full_src_register *src)
+{
+	/* Handle swizzling */
+	switch (c) {
+	case 0: c = src->SrcRegister.SwizzleX; break;
+	case 1: c = src->SrcRegister.SwizzleY; break;
+	case 2: c = src->SrcRegister.SwizzleZ; break;
+	case 3: c = src->SrcRegister.SwizzleW; break;
+	default:
+		assert(0);
+	}
+
+	switch (src->SrcRegister.File) {
+	case TGSI_FILE_INPUT:
+		return &pc->attr[src->SrcRegister.Index * 4 + c];
+	case TGSI_FILE_TEMPORARY:
+		return &pc->temp[src->SrcRegister.Index * 4 + c];
+	case TGSI_FILE_CONSTANT:
+		return &pc->param[src->SrcRegister.Index * 4 + c];
+	case TGSI_FILE_IMMEDIATE:
+		return &pc->immd[src->SrcRegister.Index * 4 + c];
+	default:
+		break;
+	}
+
+	return NULL;
+}
+
+static boolean
+nv50_program_tx_insn(struct nv50_pc *pc, const union tgsi_full_token *tok)
+{
+	const struct tgsi_full_instruction *inst = &tok->FullInstruction;
+	struct nv50_reg *dst[4], *src[3][4], *none = NULL, *tmp;
+	unsigned mask;
+	int i, c;
+
+	NOUVEAU_ERR("insn %p\n", tok);
+
+	mask = inst->FullDstRegisters[0].DstRegister.WriteMask;
+
+	for (c = 0; c < 4; c++) {
+		if (mask & (1 << c))
+			dst[c] = tgsi_dst(pc, c, &inst->FullDstRegisters[0]);
+		else
+			dst[c] = NULL;
+	}
+
+	for (i = 0; i < inst->Instruction.NumSrcRegs; i++) {
+		for (c = 0; c < 4; c++)
+			src[i][c] = tgsi_src(pc, c, &inst->FullSrcRegisters[i]);
+	}
+
+	switch (inst->Instruction.Opcode) {
+	case TGSI_OPCODE_ADD:
+		for (c = 0; c < 4; c++) {
+			if (mask & (1 << c)) {
+				emit(pc, 0x0b, dst[c],
+				     src[0][c], src[1][c], none);
+			}
+		}
+		break;
+	case TGSI_OPCODE_DP3:
+		tmp = alloc_temp(pc, NULL);
+		emit(pc, 0x0c, tmp, src[0][0], src[1][0], NULL);
+		emit(pc, 0x0e, tmp, src[0][1], src[1][1], tmp);
+		emit(pc, 0x0e, tmp, src[0][2], src[1][2], tmp);
+		for (c = 0; c < 4; c++) {
+			if (mask & (1 << c))
+				emit(pc, 0x01, dst[c], tmp, none, none);
+		}
+		free_temp(pc, tmp);
+		break;
+	case TGSI_OPCODE_DP4:
+		tmp = alloc_temp(pc, NULL);
+		emit(pc, 0x0c, tmp, src[0][0], src[1][0], NULL);
+		emit(pc, 0x0e, tmp, src[0][1], src[1][1], tmp);
+		emit(pc, 0x0e, tmp, src[0][2], src[1][2], tmp);
+		emit(pc, 0x0e, tmp, src[0][3], src[1][3], tmp);
+		for (c = 0; c < 4; c++) {
+			if (mask & (1 << c))
+				emit(pc, 0x01, dst[c], tmp, none, none);
+		}
+		free_temp(pc, tmp);
+		break;
+	case TGSI_OPCODE_MAD:
+		for (c = 0; c < 4; c++) {
+			if (mask & (1 << c))
+				emit(pc, 0x0e, dst[c],
+				     src[0][c], src[1][c], src[2][c]);
+		}
+		break;
+	case TGSI_OPCODE_MOV:
+		for (c = 0; c < 4; c++) {
+			if (mask & (1 << c))
+				emit(pc, 0x01, dst[c], src[0][c], none, none);
+		}
+		break;
+	case TGSI_OPCODE_MUL:
+		for (c = 0; c < 4; c++) {
+			if (mask & (1 << c))
+				emit(pc, 0x0c, dst[c],
+				     src[0][c], src[1][c], none);
+		}
+		break;
+	case TGSI_OPCODE_RCP:
+		for (c = 0; c < 4; c++) {
+			if (mask & (1 << c))
+				emit(pc, 0x09, dst[c],
+				     src[0][c], none, none);
+		}
+		break;
+	case TGSI_OPCODE_END:
+		break;
+	default:
+		NOUVEAU_ERR("invalid opcode %d\n", inst->Instruction.Opcode);
+		return FALSE;
+	}
+
+	return TRUE;
+}
+
+static boolean
+nv50_program_tx_prep(struct nv50_pc *pc)
+{
+	struct tgsi_parse_context p;
+	boolean ret = FALSE;
+	unsigned i, c;
+
+	tgsi_parse_init(&p, pc->p->pipe.tokens);
+	while (!tgsi_parse_end_of_tokens(&p)) {
+		const union tgsi_full_token *tok = &p.FullToken;
+
+		tgsi_parse_token(&p);
+		switch (tok->Token.Type) {
+		case TGSI_TOKEN_TYPE_IMMEDIATE:
+		{
+			const struct tgsi_full_immediate *imm =
+				&p.FullToken.FullImmediate;
+
+			pc->immd_nr++;
+			pc->immd_buf = realloc(pc->immd_buf, 4 * pc->immd_nr *
+							     sizeof(float));
+			pc->immd_buf[4 * (pc->immd_nr - 1) + 0] =
+				imm->u.ImmediateFloat32[0].Float;
+			pc->immd_buf[4 * (pc->immd_nr - 1) + 1] =
+				imm->u.ImmediateFloat32[1].Float;
+			pc->immd_buf[4 * (pc->immd_nr - 1) + 2] =
+				imm->u.ImmediateFloat32[2].Float;
+			pc->immd_buf[4 * (pc->immd_nr - 1) + 3] =
+				imm->u.ImmediateFloat32[3].Float;
+		}
+			break;
+		case TGSI_TOKEN_TYPE_DECLARATION:
+		{
+			const struct tgsi_full_declaration *d;
+			unsigned last;
+
+			d = &p.FullToken.FullDeclaration;
+			last = d->u.DeclarationRange.Last;
+
+			switch (d->Declaration.File) {
+			case TGSI_FILE_TEMPORARY:
+				if (pc->temp_nr < (last + 1))
+					pc->temp_nr = last + 1;
+				break;
+			case TGSI_FILE_OUTPUT:
+				if (pc->result_nr < (last + 1))
+					pc->result_nr = last + 1;
+				break;
+			case TGSI_FILE_INPUT:
+				if (pc->attr_nr < (last + 1))
+					pc->attr_nr = last + 1;
+				break;
+			case TGSI_FILE_CONSTANT:
+				if (pc->param_nr < (last + 1))
+					pc->param_nr = last + 1;
+				break;
+			default:
+				NOUVEAU_ERR("bad decl file %d\n",
+					    d->Declaration.File);
+				goto out_err;
+			}
+		}
+			break;
+		case TGSI_TOKEN_TYPE_INSTRUCTION:
+			break;
+		default:
+			break;
+		}
+	}
+
+	NOUVEAU_ERR("%d temps\n", pc->temp_nr);
+	if (pc->temp_nr) {
+		pc->temp = calloc(pc->temp_nr * 4, sizeof(struct nv50_reg));
+		if (!pc->temp)
+			goto out_err;
+
+		for (i = 0; i < pc->temp_nr; i++) {
+			for (c = 0; c < 4; c++) {
+				pc->temp[i*4+c].type = P_TEMP;
+				pc->temp[i*4+c].hw = -1;
+				pc->temp[i*4+c].index = i;
+			}
+		}
+	}
+
+	NOUVEAU_ERR("%d attrib regs\n", pc->attr_nr);
+	if (pc->attr_nr) {
+		int aid = 0;
+
+		pc->attr = calloc(pc->attr_nr * 4, sizeof(struct nv50_reg));
+		if (!pc->attr)
+			goto out_err;
+
+		for (i = 0; i < pc->attr_nr; i++) {
+			for (c = 0; c < 4; c++) {
+				pc->p->cfg.vp.attr[aid/32] |= (1 << (aid % 32));
+				pc->attr[i*4+c].type = P_ATTR;
+				pc->attr[i*4+c].hw = aid++;
+				pc->attr[i*4+c].index = i;
+			}
+		}
+	}
+
+	NOUVEAU_ERR("%d result regs\n", pc->result_nr);
+	if (pc->result_nr) {
+		int rid = 0;
+
+		pc->result = calloc(pc->result_nr * 4, sizeof(struct nv50_reg));
+		if (!pc->result)
+			goto out_err;
+
+		for (i = 0; i < pc->result_nr; i++) {
+			for (c = 0; c < 4; c++) {
+				pc->result[i*4+c].type = P_RESULT;
+				pc->result[i*4+c].hw = rid++;
+				pc->result[i*4+c].index = i;
+			}
+		}
+	}
+
+	NOUVEAU_ERR("%d param regs\n", pc->param_nr);
+	if (pc->param_nr) {
+		int rid = 0;
+
+		pc->param = calloc(pc->param_nr * 4, sizeof(struct nv50_reg));
+		if (!pc->param)
+			goto out_err;
+
+		for (i = 0; i < pc->param_nr; i++) {
+			for (c = 0; c < 4; c++) {
+				pc->param[i*4+c].type = P_CONST;
+				pc->param[i*4+c].hw = rid++;
+				pc->param[i*4+c].index = i;
+			}
+		}
+	}
+
+	if (pc->immd_nr) {
+		int rid = pc->param_nr * 4;
+
+		pc->immd = calloc(pc->immd_nr * 4, sizeof(struct nv50_reg));
+		if (!pc->immd)
+			goto out_err;
+
+		for (i = 0; i < pc->immd_nr; i++) {
+			for (c = 0; c < 4; c++) {
+				pc->immd[i*4+c].type = P_IMMD;
+				pc->immd[i*4+c].hw = rid++;
+				pc->immd[i*4+c].index = i;
+			}
+		}
+	}
+
+	ret = TRUE;
+out_err:
+	tgsi_parse_free(&p);
+	return ret;
+}
+
+static boolean
+nv50_program_tx(struct nv50_program *p)
+{
+	struct tgsi_parse_context parse;
+	struct nv50_pc *pc;
+	boolean ret;
+
+	pc = CALLOC_STRUCT(nv50_pc);
+	if (!pc)
+		return FALSE;
+	pc->p = p;
+	pc->p->cfg.vp.high_temp = 4;
+
+	ret = nv50_program_tx_prep(pc);
+	if (ret == FALSE)
+		goto out_cleanup;
+
+	tgsi_parse_init(&parse, pc->p->pipe.tokens);
+	while (!tgsi_parse_end_of_tokens(&parse)) {
+		const union tgsi_full_token *tok = &parse.FullToken;
+
+		tgsi_parse_token(&parse);
+
+		switch (tok->Token.Type) {
+		case TGSI_TOKEN_TYPE_INSTRUCTION:
+			ret = nv50_program_tx_insn(pc, tok);
+			if (ret == FALSE)
+				goto out_err;
+			break;
+		default:
+			break;
+		}
+	}
+
+	p->param_nr = pc->param_nr * 4;
+	p->immd_nr = pc->immd_nr * 4;
+	p->immd = pc->immd_buf;
+
+out_err:
+	tgsi_parse_free(&parse);
+
+out_cleanup:
+	return ret;
+}
+
+static void
+nv50_program_validate(struct nv50_context *nv50, struct nv50_program *p)
+{
+	struct tgsi_parse_context pc;
+
+	tgsi_parse_init(&pc, p->pipe.tokens);
+
+	if (pc.FullHeader.Processor.Processor == TGSI_PROCESSOR_FRAGMENT) {
+		p->insns_nr = 8;
+		p->insns = malloc(p->insns_nr * sizeof(unsigned));
+		p->insns[0] = 0x80000000;
+		p->insns[1] = 0x9000000c;
+		p->insns[2] = 0x82010600;
+		p->insns[3] = 0x82020604;
+		p->insns[4] = 0x80030609;
+		p->insns[5] = 0x00020780;
+		p->insns[6] = 0x8004060d;
+		p->insns[7] = 0x00020781;
+	} else
+	if (pc.FullHeader.Processor.Processor == TGSI_PROCESSOR_VERTEX) {
+		int i;
+
+		if (nv50_program_tx(p) == FALSE)
+			assert(0);
+		p->insns[p->insns_nr - 1] |= 0x00000001;
+
+		for (i = 0; i < p->insns_nr; i++)
+			NOUVEAU_ERR("%d 0x%08x\n", i, p->insns[i]);
+	} else {
+		NOUVEAU_ERR("invalid TGSI processor\n");
+		tgsi_parse_free(&pc);
+		return;
+	}
+
+	tgsi_parse_free(&pc);
+
+	p->translated = TRUE;
+}
+
+void
+nv50_vertprog_validate(struct nv50_context *nv50)
+{
+	struct pipe_winsys *ws = nv50->pipe.winsys;
+	struct nouveau_winsys *nvws = nv50->screen->nvws;
+	struct nouveau_grobj *tesla = nv50->screen->tesla;
+	struct nv50_program *p = nv50->vertprog;
+	struct nouveau_stateobj *so;
+	void *map;
+	int i;
+
+	if (!p->translated) {
+		nv50_program_validate(nv50, p);
+		if (!p->translated)
+			assert(0);
+	}
+
+	if (!p->buffer)
+		p->buffer = ws->buffer_create(ws, 0x100, 0, p->insns_nr * 4);
+	map = ws->buffer_map(ws, p->buffer, PIPE_BUFFER_USAGE_CPU_WRITE);
+	memcpy(map, p->insns, p->insns_nr * 4);
+	ws->buffer_unmap(ws, p->buffer);
+
+	if (p->param_nr) {
+		float *cb;
+
+		cb = ws->buffer_map(ws, nv50->constbuf[PIPE_SHADER_VERTEX],
+				    PIPE_BUFFER_USAGE_CPU_READ);
+		for (i = 0; i < p->param_nr; i++) {
+			BEGIN_RING(tesla, 0x0f00, 2);
+			OUT_RING  (i << 8);
+			OUT_RING  (fui(cb[i]));
+		}
+		ws->buffer_unmap(ws, nv50->constbuf[PIPE_SHADER_VERTEX]);
+	}
+
+
+	for (i = 0; i < p->immd_nr; i++) {
+		BEGIN_RING(tesla, 0x0f00, 2);
+		OUT_RING  ((p->param_nr + i) << 8);
+		OUT_RING  (fui(p->immd[i]));
+	}
+
+	so = so_new(11, 2);
+	so_method(so, tesla, NV50TCL_VP_ADDRESS_HIGH, 2);
+	so_reloc (so, p->buffer, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
+		  NOUVEAU_BO_HIGH, 0, 0);
+	so_reloc (so, p->buffer, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
+		  NOUVEAU_BO_LOW, 0, 0);
+	so_method(so, tesla, 0x1650, 2);
+	so_data  (so, p->cfg.vp.attr[0]);
+	so_data  (so, p->cfg.vp.attr[1]);
+	so_method(so, tesla, 0x16ac, 2);
+	so_data  (so, 8);
+	so_data  (so, p->cfg.vp.high_temp);
+	so_method(so, tesla, 0x140c, 1);
+	so_data  (so, 0); /* program start offset */
+	so_emit(nv50->screen->nvws, so);
+	so_ref(NULL, &so);
+}
+
+void
+nv50_fragprog_validate(struct nv50_context *nv50)
+{
+	struct pipe_winsys *ws = nv50->pipe.winsys;
+	struct nouveau_grobj *tesla = nv50->screen->tesla;
+	struct nv50_program *p = nv50->fragprog;
+	struct nouveau_stateobj *so;
+	void *map;
+
+	if (!p->translated) {
+		nv50_program_validate(nv50, p);
+		if (!p->translated)
+			assert(0);
+	}
+
+	if (!p->buffer)
+		p->buffer = ws->buffer_create(ws, 0x100, 0, p->insns_nr * 4);
+	map = ws->buffer_map(ws, p->buffer, PIPE_BUFFER_USAGE_CPU_WRITE);
+	memcpy(map, p->insns, p->insns_nr * 4);
+	ws->buffer_unmap(ws, p->buffer);
+
+	so = so_new(3, 2);
+	so_method(so, tesla, NV50TCL_FP_ADDRESS_HIGH, 2);
+	so_reloc (so, p->buffer, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
+		  NOUVEAU_BO_HIGH, 0, 0);
+	so_reloc (so, p->buffer, 0, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
+		  NOUVEAU_BO_LOW, 0, 0);
+	so_emit(nv50->screen->nvws, so);
+	so_ref(NULL, &so);
+}
+
+void
+nv50_program_destroy(struct nv50_context *nv50, struct nv50_program *p)
+{
+	struct pipe_winsys *ws = nv50->pipe.winsys;
+
+	if (p->insns_nr) {
+		if (p->insns)
+			FREE(p->insns);
+		p->insns_nr = 0;
+	}
+
+	if (p->buffer)
+		pipe_buffer_reference(ws, &p->buffer, NULL);
+
+	p->translated = 0;
+}
+
diff --git a/src/gallium/drivers/nv50/nv50_screen.c b/src/gallium/drivers/nv50/nv50_screen.c
index 29c057a145..fc3eeed913 100644
--- a/src/gallium/drivers/nv50/nv50_screen.c
+++ b/src/gallium/drivers/nv50/nv50_screen.c
@@ -203,8 +203,19 @@ nv50_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws)
 		return NULL;
 	}
 
+	/* Static constant buffer */
+	screen->constbuf = ws->buffer_create(ws, 0, 0, 256 * 4 * 4);
+	if (nvws->res_init(&screen->vp_data_heap, 0, 256)) {
+		NOUVEAU_ERR("Error initialising constant buffer\n");
+		nv50_screen_destroy(&screen->pipe);
+		return NULL;
+	}
+
 	/* Static tesla init */
-	so = so_new(128, 0);
+	so = so_new(256, 20);
+
+	so_method(so, screen->tesla, 0x1558, 1);
+	so_data  (so, 1);
 	so_method(so, screen->tesla, NV50TCL_DMA_NOTIFY, 1);
 	so_data  (so, screen->sync->handle);
 	so_method(so, screen->tesla, NV50TCL_DMA_IN_MEMORY0(0),
@@ -215,6 +226,61 @@ nv50_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws)
 		  NV50TCL_DMA_IN_MEMORY1__SIZE);
 	for (i = 0; i < NV50TCL_DMA_IN_MEMORY1__SIZE; i++)
 		so_data(so, nvws->channel->vram->handle);
+	so_method(so, screen->tesla, 0x121c, 1);
+	so_data  (so, 1);
+
+	so_method(so, screen->tesla, 0x13bc, 1);
+	so_data  (so, 0x54);
+	so_method(so, screen->tesla, 0x13ac, 1);
+	so_data  (so, 1);
+	so_method(so, screen->tesla, 0x16bc, 2);
+	so_data  (so, 0x03020100);
+	so_data  (so, 0x07060504);
+	so_method(so, screen->tesla, 0x1988, 2);
+	so_data  (so, 0x08040404);
+	so_data  (so, 0x00000004);
+	so_method(so, screen->tesla, 0x16ac, 2);
+	so_data  (so, 8);
+	so_data  (so, 4);
+	so_method(so, screen->tesla, 0x16b8, 1);
+	so_data  (so, 8);
+
+	so_method(so, screen->tesla, 0x1280, 3);
+	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+		  NOUVEAU_BO_RD | NOUVEAU_BO_HIGH, 0, 0);
+	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+		  NOUVEAU_BO_RD | NOUVEAU_BO_LOW, 0, 0);
+	so_data  (so, 0x00001000);
+	so_method(so, screen->tesla, 0x1280, 3);
+	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+		  NOUVEAU_BO_RD | NOUVEAU_BO_HIGH, 0, 0);
+	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+		  NOUVEAU_BO_RD | NOUVEAU_BO_LOW, 0, 0);
+	so_data  (so, 0x00014000);
+	so_method(so, screen->tesla, 0x1280, 3);
+	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+		  NOUVEAU_BO_RD | NOUVEAU_BO_HIGH, 0, 0);
+	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+		  NOUVEAU_BO_RD | NOUVEAU_BO_LOW, 0, 0);
+	so_data  (so, 0x00024000);
+	so_method(so, screen->tesla, 0x1280, 3);
+	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+		  NOUVEAU_BO_RD | NOUVEAU_BO_HIGH, 0, 0);
+	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+		  NOUVEAU_BO_RD | NOUVEAU_BO_LOW, 0, 0);
+	so_data  (so, 0x00034000);
+	so_method(so, screen->tesla, 0x1280, 3);
+	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+		  NOUVEAU_BO_RD | NOUVEAU_BO_HIGH, 0, 0);
+	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+		  NOUVEAU_BO_RD | NOUVEAU_BO_LOW, 0, 0);
+	so_data  (so, 0x00040100);
+
+	for (i = 0; i < 16; i++) {
+		so_method(so, screen->tesla, 0x1080 + (i * 8), 2);
+		so_data  (so, 0x000000ff);
+		so_data  (so, 0xffffffff);
+	}
 
 	so_emit(nvws, so);
 	so_ref(NULL, &so);
diff --git a/src/gallium/drivers/nv50/nv50_screen.h b/src/gallium/drivers/nv50/nv50_screen.h
index 5a2732e37f..d63dd48508 100644
--- a/src/gallium/drivers/nv50/nv50_screen.h
+++ b/src/gallium/drivers/nv50/nv50_screen.h
@@ -12,6 +12,9 @@ struct nv50_screen {
 
 	struct nouveau_grobj *tesla;
 	struct nouveau_notifier *sync;
+
+	struct pipe_buffer *constbuf;
+	struct nouveau_resource *vp_data_heap;
 };
 
 static INLINE struct nv50_screen *
diff --git a/src/gallium/drivers/nv50/nv50_state.c b/src/gallium/drivers/nv50/nv50_state.c
index f42bae0c28..fd10a38378 100644
--- a/src/gallium/drivers/nv50/nv50_state.c
+++ b/src/gallium/drivers/nv50/nv50_state.c
@@ -248,8 +248,8 @@ nv50_depth_stencil_alpha_state_create(struct pipe_context *pipe,
 	struct nouveau_stateobj *so = so_new(64, 0);
 
 	so_method(so, tesla, NV50TCL_DEPTH_WRITE_ENABLE, 1);
-	so_data  (so, cso->depth.writemask ? 1 : 0);
-	if (cso->depth.enabled) {
+	so_data  (so, 0); //cso->depth.writemask ? 1 : 0);
+	if (0 && cso->depth.enabled) {
 		so_method(so, tesla, NV50TCL_DEPTH_TEST_ENABLE, 1);
 		so_data  (so, 1);
 		so_method(so, tesla, NV50TCL_DEPTH_TEST_FUNC, 1);
@@ -328,34 +328,58 @@ static void *
 nv50_vp_state_create(struct pipe_context *pipe,
 		     const struct pipe_shader_state *cso)
 {
-	return NULL;
+	struct nv50_program *p = CALLOC_STRUCT(nv50_program);
+
+	p->pipe = *cso;
+	tgsi_scan_shader(p->pipe.tokens, &p->info);
+	return (void *)p;
 }
 
 static void
 nv50_vp_state_bind(struct pipe_context *pipe, void *hwcso)
 {
+	struct nv50_context *nv50 = nv50_context(pipe);
+
+	nv50->vertprog = hwcso;
+	nv50->dirty |= NV50_NEW_VERTPROG;
 }
 
 static void
 nv50_vp_state_delete(struct pipe_context *pipe, void *hwcso)
 {
+	struct nv50_context *nv50 = nv50_context(pipe);
+
+	nv50_program_destroy(nv50, hwcso);
+	FREE(hwcso);
 }
 
 static void *
 nv50_fp_state_create(struct pipe_context *pipe,
 		     const struct pipe_shader_state *cso)
 {
-	return NULL;
+	struct nv50_program *p = CALLOC_STRUCT(nv50_program);
+
+	p->pipe = *cso;
+	tgsi_scan_shader(p->pipe.tokens, &p->info);
+	return (void *)p;
 }
 
 static void
 nv50_fp_state_bind(struct pipe_context *pipe, void *hwcso)
 {
+	struct nv50_context *nv50 = nv50_context(pipe);
+
+	nv50->fragprog = hwcso;
+	nv50->dirty |= NV50_NEW_FRAGPROG;
 }
 
 static void
 nv50_fp_state_delete(struct pipe_context *pipe, void *hwcso)
 {
+	struct nv50_context *nv50 = nv50_context(pipe);
+
+	nv50_program_destroy(nv50, hwcso);
+	FREE(hwcso);
 }
 
 static void
@@ -378,6 +402,16 @@ static void
 nv50_set_constant_buffer(struct pipe_context *pipe, uint shader, uint index,
 			 const struct pipe_constant_buffer *buf )
 {
+	struct nv50_context *nv50 = nv50_context(pipe);
+
+	if (shader == PIPE_SHADER_VERTEX) {
+		nv50->constbuf[PIPE_SHADER_VERTEX] = buf->buffer;
+		nv50->dirty |= NV50_NEW_VERTPROG;
+	} else
+	if (shader == PIPE_SHADER_FRAGMENT) {
+		nv50->constbuf[PIPE_SHADER_FRAGMENT] = buf->buffer;
+		nv50->dirty |= NV50_NEW_FRAGPROG;
+	}
 }
 
 static void
@@ -424,12 +458,24 @@ static void
 nv50_set_vertex_buffers(struct pipe_context *pipe, unsigned count,
 			const struct pipe_vertex_buffer *vb)
 {
+	struct nv50_context *nv50 = nv50_context(pipe);
+
+	memcpy(nv50->vtxbuf, vb, sizeof(*vb) * count);
+	nv50->vtxbuf_nr = count;
+
+	nv50->dirty |= NV50_NEW_ARRAYS;
 }
 
 static void
 nv50_set_vertex_elements(struct pipe_context *pipe, unsigned count,
 			 const struct pipe_vertex_element *ve)
 {
+	struct nv50_context *nv50 = nv50_context(pipe);
+
+	memcpy(nv50->vtxelt, ve, sizeof(*ve) * count);
+	nv50->vtxelt_nr = count;
+
+	nv50->dirty |= NV50_NEW_ARRAYS;
 }
 
 void
diff --git a/src/gallium/drivers/nv50/nv50_state.h b/src/gallium/drivers/nv50/nv50_state.h
index a3b781d4c6..be0c75ad6e 100644
--- a/src/gallium/drivers/nv50/nv50_state.h
+++ b/src/gallium/drivers/nv50/nv50_state.h
@@ -2,6 +2,36 @@
 #define __NV50_STATE_H__
 
 #include "pipe/p_state.h"
+#include "tgsi/util/tgsi_scan.h"
 
+struct nv50_program_data {
+	int index;
+	int component;
+	float value;
+};
+
+struct nv50_program {
+	struct pipe_shader_state pipe;
+	struct tgsi_shader_info info;
+	boolean translated;
+
+	unsigned *insns;
+	unsigned insns_nr;
+
+	struct pipe_buffer *buffer;
+
+	unsigned param_nr;
+	float *immd;
+	unsigned immd_nr;
+
+	struct nouveau_resource *data;
+
+	union {
+		struct {
+			unsigned high_temp;
+			unsigned attr[2];
+		} vp;
+	} cfg;
+};
 
 #endif
diff --git a/src/gallium/drivers/nv50/nv50_state_validate.c b/src/gallium/drivers/nv50/nv50_state_validate.c
index 68d419f9fc..4a548378b7 100644
--- a/src/gallium/drivers/nv50/nv50_state_validate.c
+++ b/src/gallium/drivers/nv50/nv50_state_validate.c
@@ -15,8 +15,8 @@ nv50_state_validate_fb(struct nv50_context *nv50)
 			h = fb->cbufs[i]->height;
 			gw = 1;
 		} else {
-			assert(w != fb->cbufs[i]->width);
-			assert(h != fb->cbufs[i]->height);
+			assert(w == fb->cbufs[i]->width);
+			assert(h == fb->cbufs[i]->height);
 		}
 
 		so_method(so, tesla, NV50TCL_RT_HORIZ(i), 2);
@@ -46,6 +46,9 @@ nv50_state_validate_fb(struct nv50_context *nv50)
 		}
 		so_data(so, 0x00000040);
 		so_data(so, 0x00000000);
+
+		so_method(so, tesla, 0x1224, 1);
+		so_data  (so, 1);
 	}
 
 	if (fb->zsbuf) {
@@ -54,11 +57,11 @@ nv50_state_validate_fb(struct nv50_context *nv50)
 			h = fb->zsbuf->height;
 			gw = 1;
 		} else {
-			assert(w != fb->zsbuf->width);
-			assert(h != fb->zsbuf->height);
+			assert(w == fb->zsbuf->width);
+			assert(h == fb->zsbuf->height);
 		}
 
-		so_method(so, tesla, NV50TCL_RT_ADDRESS_HIGH(i), 5);
+		so_method(so, tesla, NV50TCL_ZETA_ADDRESS_HIGH, 5);
 		so_reloc (so, fb->zsbuf->buffer, fb->zsbuf->offset,
 			  NOUVEAU_BO_VRAM | NOUVEAU_BO_HIGH, 0, 0);
 		so_reloc (so, fb->zsbuf->buffer, fb->zsbuf->offset,
@@ -83,9 +86,12 @@ nv50_state_validate_fb(struct nv50_context *nv50)
 	so_method(so, tesla, NV50TCL_VIEWPORT_HORIZ, 2);
 	so_data  (so, w << 16);
 	so_data  (so, h << 16);
-	so_method(so, tesla, 0xff0, 2);
+	so_method(so, tesla, 0xff4, 2);
 	so_data  (so, w << 16);
 	so_data  (so, h << 16);
+	so_method(so, tesla, 0xdf8, 2);
+	so_data  (so, 0);
+	so_data  (so, h);
 
 	so_emit(nv50->screen->nvws, so);
 	so_ref(NULL, &so);
@@ -108,6 +114,12 @@ nv50_state_validate(struct nv50_context *nv50)
 	if (nv50->dirty & NV50_NEW_ZSA)
 		so_emit(nvws, nv50->zsa->so);
 
+	if (nv50->dirty & NV50_NEW_VERTPROG)
+		nv50_vertprog_validate(nv50);
+
+	if (nv50->dirty & NV50_NEW_FRAGPROG)
+		nv50_fragprog_validate(nv50);
+
 	if (nv50->dirty & NV50_NEW_RASTERIZER)
 		so_emit(nvws, nv50->rasterizer->so);
 
@@ -150,12 +162,15 @@ nv50_state_validate(struct nv50_context *nv50)
 		so_data  (so, fui(nv50->viewport.translate[2]));
 		so_method(so, tesla, NV50TCL_VIEWPORT_UNK1(0), 3);
 		so_data  (so, fui(nv50->viewport.scale[0]));
-		so_data  (so, fui(nv50->viewport.scale[1]));
+		so_data  (so, fui(-nv50->viewport.scale[1]));
 		so_data  (so, fui(nv50->viewport.scale[2]));
 		so_emit(nvws, so);
 		so_ref(NULL, &so);
 	}
 
+	if (nv50->dirty & NV50_NEW_ARRAYS)
+		nv50_vbo_validate(nv50);
+
 	nv50->dirty = 0;
 	return TRUE;
 }
diff --git a/src/gallium/drivers/nv50/nv50_vbo.c b/src/gallium/drivers/nv50/nv50_vbo.c
index b01ce1d42c..f086c76258 100644
--- a/src/gallium/drivers/nv50/nv50_vbo.c
+++ b/src/gallium/drivers/nv50/nv50_vbo.c
@@ -5,6 +5,29 @@
 #include "nv50_context.h"
 #include "nv50_state.h"
 
+static INLINE unsigned
+nv50_prim(unsigned mode)
+{
+	switch (mode) {
+	case PIPE_PRIM_POINTS: return NV50TCL_VERTEX_BEGIN_POINTS;
+	case PIPE_PRIM_LINES: return NV50TCL_VERTEX_BEGIN_LINES;
+	case PIPE_PRIM_LINE_LOOP: return NV50TCL_VERTEX_BEGIN_LINE_LOOP;
+	case PIPE_PRIM_LINE_STRIP: return NV50TCL_VERTEX_BEGIN_LINE_STRIP;
+	case PIPE_PRIM_TRIANGLES: return NV50TCL_VERTEX_BEGIN_TRIANGLES;
+	case PIPE_PRIM_TRIANGLE_STRIP:
+		return NV50TCL_VERTEX_BEGIN_TRIANGLE_STRIP;
+	case PIPE_PRIM_TRIANGLE_FAN: return NV50TCL_VERTEX_BEGIN_TRIANGLE_FAN;
+	case PIPE_PRIM_QUADS: return NV50TCL_VERTEX_BEGIN_QUADS;
+	case PIPE_PRIM_QUAD_STRIP: return NV50TCL_VERTEX_BEGIN_QUAD_STRIP;
+	case PIPE_PRIM_POLYGON: return NV50TCL_VERTEX_BEGIN_POLYGON;
+	default:
+		break;
+	}
+
+	NOUVEAU_ERR("invalid primitive type %d\n", mode);
+	return NV50TCL_VERTEX_BEGIN_POINTS;
+}
+
 boolean
 nv50_draw_arrays(struct pipe_context *pipe, unsigned mode, unsigned start,
 		 unsigned count)
@@ -12,8 +35,22 @@ nv50_draw_arrays(struct pipe_context *pipe, unsigned mode, unsigned start,
 	struct nv50_context *nv50 = nv50_context(pipe);
 
 	nv50_state_validate(nv50);
-
 	NOUVEAU_ERR("unimplemented\n");
+
+	BEGIN_RING(tesla, 0x142c, 1);
+	OUT_RING  (0);
+	BEGIN_RING(tesla, 0x142c, 1);
+	OUT_RING  (0);
+
+	BEGIN_RING(tesla, NV50TCL_VERTEX_BEGIN, 1);
+	OUT_RING  (nv50_prim(mode));
+	BEGIN_RING(tesla, NV50TCL_VERTEX_BUFFER_FIRST, 2);
+	OUT_RING  (start);
+	OUT_RING  (count);
+	BEGIN_RING(tesla, NV50TCL_VERTEX_END, 1);
+	OUT_RING  (0);
+
+	pipe->flush(pipe, 0, NULL);
 	return TRUE;
 }
 
@@ -30,3 +67,49 @@ nv50_draw_elements(struct pipe_context *pipe,
 	return TRUE;
 }
 
+void
+nv50_vbo_validate(struct nv50_context *nv50)
+{
+	struct nouveau_grobj *tesla = nv50->screen->tesla;
+	struct nouveau_stateobj *vtxbuf, *vtxfmt;
+	int i, vpi = 0;
+
+	vtxbuf = so_new(nv50->vtxelt_nr * 4, nv50->vtxelt_nr * 2);
+	vtxfmt = so_new(nv50->vtxelt_nr + 1, 0);
+	so_method(vtxfmt, tesla, 0x1ac0, nv50->vtxelt_nr);
+
+	for (i = 0; i < nv50->vtxelt_nr; i++) {
+		struct pipe_vertex_element *ve = &nv50->vtxelt[i];
+		struct pipe_vertex_buffer *vb =
+			&nv50->vtxbuf[ve->vertex_buffer_index];
+
+		switch (ve->src_format) {
+		case PIPE_FORMAT_R32G32B32_FLOAT:
+			so_data(vtxfmt, 0x7e100000 | i);
+			break;
+		default:
+		{
+			char fmt[128];
+			pf_sprint_name(fmt, ve->src_format);
+			NOUVEAU_ERR("invalid vbo format %s\n", fmt);
+			assert(0);
+			return;
+		}
+		}
+
+		so_method(vtxbuf, tesla, 0x900 + (i * 16), 3);
+		so_data  (vtxbuf, 0x20000000 | vb->pitch);
+		so_reloc (vtxbuf, vb->buffer, vb->buffer_offset +
+			  ve->src_offset, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
+			  NOUVEAU_BO_HIGH, 0, 0);
+		so_reloc (vtxbuf, vb->buffer, vb->buffer_offset +
+			  ve->src_offset, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD |
+			  NOUVEAU_BO_LOW, 0, 0);
+	}
+
+	so_emit(nv50->screen->nvws, vtxfmt);
+	so_ref (NULL, &vtxfmt);
+	so_emit(nv50->screen->nvws, vtxbuf);
+	so_ref (NULL, &vtxbuf);
+}
+
-- 
cgit v1.2.3


From 716c1cd2ecbc1e86c0fd747c9fa9e095ded5fd5d Mon Sep 17 00:00:00 2001
From: Ben Skeggs <skeggsb@gmail.com>
Date: Sun, 1 Jun 2008 23:10:31 +1000
Subject: nv50: use "real" constbufs for shaders + tcb uploads

---
 src/gallium/drivers/nv50/nv50_context.h        | 14 ++++++-
 src/gallium/drivers/nv50/nv50_program.c        | 27 +++++-------
 src/gallium/drivers/nv50/nv50_screen.c         | 57 +++++++++++++++-----------
 src/gallium/drivers/nv50/nv50_screen.h         |  3 ++
 src/gallium/drivers/nv50/nv50_state.c          |  4 +-
 src/gallium/drivers/nv50/nv50_state.h          |  1 -
 src/gallium/drivers/nv50/nv50_state_validate.c | 29 +++++++++++++
 7 files changed, 90 insertions(+), 45 deletions(-)

(limited to 'src/gallium/drivers/nv50/nv50_screen.h')

diff --git a/src/gallium/drivers/nv50/nv50_context.h b/src/gallium/drivers/nv50/nv50_context.h
index d4d716b94b..c4a8a4c064 100644
--- a/src/gallium/drivers/nv50/nv50_context.h
+++ b/src/gallium/drivers/nv50/nv50_context.h
@@ -23,6 +23,14 @@
 #define NOUVEAU_MSG(fmt, args...) \
 	fprintf(stderr, "nouveau: "fmt, ##args);
 
+/* Constant buffer assignment */
+#define NV50_CB_PMISC		0
+#define NV50_CB_PVP		1
+#define NV50_CB_PFP		2
+#define NV50_CB_PGP		3
+#define NV50_CB_TIC		4
+#define NV50_CB_TSC		5
+
 #define NV50_NEW_BLEND		(1 << 0)
 #define NV50_NEW_ZSA		(1 << 1)
 #define NV50_NEW_BLEND_COLOUR	(1 << 2)
@@ -32,8 +40,10 @@
 #define NV50_NEW_RASTERIZER	(1 << 6)
 #define NV50_NEW_FRAMEBUFFER	(1 << 7)
 #define NV50_NEW_VERTPROG	(1 << 8)
-#define NV50_NEW_FRAGPROG	(1 << 9)
-#define NV50_NEW_ARRAYS		(1 << 10)
+#define NV50_NEW_VERTPROG_CB	(1 << 9)
+#define NV50_NEW_FRAGPROG	(1 << 10)
+#define NV50_NEW_FRAGPROG_CB	(1 << 11)
+#define NV50_NEW_ARRAYS		(1 << 12)
 
 struct nv50_blend_stateobj {
 	struct pipe_blend_state pipe;
diff --git a/src/gallium/drivers/nv50/nv50_program.c b/src/gallium/drivers/nv50/nv50_program.c
index 30953b7d8a..0a43646923 100644
--- a/src/gallium/drivers/nv50/nv50_program.c
+++ b/src/gallium/drivers/nv50/nv50_program.c
@@ -179,6 +179,10 @@ emit(struct nv50_pc *pc, unsigned op, struct nv50_reg *dst,
 
 		if (src1) {
 			if (src1->type == P_CONST || src1->type == P_IMMD) {
+				if (src1->type == P_IMMD)
+					inst[1] |= (NV50_CB_PMISC << 22);
+				else
+					inst[1] |= (NV50_CB_PVP << 22);
 				inst[0] |= 0x00800000; /* src1 is const */
 				/*XXX: does src1 come from "src2" now? */
 				alloc_reg(pc, src1);
@@ -196,6 +200,10 @@ emit(struct nv50_pc *pc, unsigned op, struct nv50_reg *dst,
 
 		if (src2) {
 			if (src2->type == P_CONST || src2->type == P_IMMD) {
+				if (src2->type == P_IMMD)
+					inst[1] |= (NV50_CB_PMISC << 22);
+				else
+					inst[1] |= (NV50_CB_PVP << 22);
 				inst[0] |= 0x01000000; /* src2 is const */
 				inst[1] |= (src2->hw << 14);
 			} else {
@@ -526,7 +534,7 @@ nv50_program_tx_prep(struct nv50_pc *pc)
 	}
 
 	if (pc->immd_nr) {
-		int rid = pc->param_nr * 4;
+		int rid = 0;
 
 		pc->immd = calloc(pc->immd_nr * 4, sizeof(struct nv50_reg));
 		if (!pc->immd)
@@ -581,7 +589,6 @@ nv50_program_tx(struct nv50_program *p)
 		}
 	}
 
-	p->param_nr = pc->param_nr * 4;
 	p->immd_nr = pc->immd_nr * 4;
 	p->immd = pc->immd_buf;
 
@@ -654,23 +661,9 @@ nv50_vertprog_validate(struct nv50_context *nv50)
 	memcpy(map, p->insns, p->insns_nr * 4);
 	ws->buffer_unmap(ws, p->buffer);
 
-	if (p->param_nr) {
-		float *cb;
-
-		cb = ws->buffer_map(ws, nv50->constbuf[PIPE_SHADER_VERTEX],
-				    PIPE_BUFFER_USAGE_CPU_READ);
-		for (i = 0; i < p->param_nr; i++) {
-			BEGIN_RING(tesla, 0x0f00, 2);
-			OUT_RING  (i << 8);
-			OUT_RING  (fui(cb[i]));
-		}
-		ws->buffer_unmap(ws, nv50->constbuf[PIPE_SHADER_VERTEX]);
-	}
-
-
 	for (i = 0; i < p->immd_nr; i++) {
 		BEGIN_RING(tesla, 0x0f00, 2);
-		OUT_RING  ((p->param_nr + i) << 8);
+		OUT_RING  ((NV50_CB_PMISC << 16) | (i << 8));
 		OUT_RING  (fui(p->immd[i]));
 	}
 
diff --git a/src/gallium/drivers/nv50/nv50_screen.c b/src/gallium/drivers/nv50/nv50_screen.c
index fc3eeed913..6c0810a9cf 100644
--- a/src/gallium/drivers/nv50/nv50_screen.c
+++ b/src/gallium/drivers/nv50/nv50_screen.c
@@ -203,14 +203,6 @@ nv50_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws)
 		return NULL;
 	}
 
-	/* Static constant buffer */
-	screen->constbuf = ws->buffer_create(ws, 0, 0, 256 * 4 * 4);
-	if (nvws->res_init(&screen->vp_data_heap, 0, 256)) {
-		NOUVEAU_ERR("Error initialising constant buffer\n");
-		nv50_screen_destroy(&screen->pipe);
-		return NULL;
-	}
-
 	/* Static tesla init */
 	so = so_new(256, 20);
 
@@ -245,37 +237,56 @@ nv50_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws)
 	so_method(so, screen->tesla, 0x16b8, 1);
 	so_data  (so, 8);
 
+	/* Shared constant buffer */
+	screen->constbuf = ws->buffer_create(ws, 0, 0, 256 * 4 * 4);
+	if (nvws->res_init(&screen->vp_data_heap, 0, 256)) {
+		NOUVEAU_ERR("Error initialising constant buffer\n");
+		nv50_screen_destroy(&screen->pipe);
+		return NULL;
+	}
+
 	so_method(so, screen->tesla, 0x1280, 3);
 	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
 		  NOUVEAU_BO_RD | NOUVEAU_BO_HIGH, 0, 0);
 	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
 		  NOUVEAU_BO_RD | NOUVEAU_BO_LOW, 0, 0);
-	so_data  (so, 0x00001000);
+	so_data  (so, (NV50_CB_PMISC << 16) | 0x00001000);
+
+	/* Texture sampler/image unit setup - we abuse the constant buffer
+	 * upload mechanism for the moment to upload data to the tex config
+	 * blocks.  At some point we *may* want to go the NVIDIA way of doing
+	 * things?
+	 */
+	screen->tic = ws->buffer_create(ws, 0, 0, 32 * 8 * 4);
 	so_method(so, screen->tesla, 0x1280, 3);
-	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+	so_reloc (so, screen->tic, 0, NOUVEAU_BO_VRAM |
 		  NOUVEAU_BO_RD | NOUVEAU_BO_HIGH, 0, 0);
-	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+	so_reloc (so, screen->tic, 0, NOUVEAU_BO_VRAM |
 		  NOUVEAU_BO_RD | NOUVEAU_BO_LOW, 0, 0);
-	so_data  (so, 0x00014000);
-	so_method(so, screen->tesla, 0x1280, 3);
-	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+	so_data  (so, (NV50_CB_TIC << 16) | 0x0800);
+	so_method(so, screen->tesla, 0x1574, 3);
+	so_reloc (so, screen->tic, 0, NOUVEAU_BO_VRAM |
 		  NOUVEAU_BO_RD | NOUVEAU_BO_HIGH, 0, 0);
-	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+	so_reloc (so, screen->tic, 0, NOUVEAU_BO_VRAM |
 		  NOUVEAU_BO_RD | NOUVEAU_BO_LOW, 0, 0);
-	so_data  (so, 0x00024000);
+	so_data  (so, 0x00000800);
+
+	screen->tsc = ws->buffer_create(ws, 0, 0, 32 * 8 * 4);
 	so_method(so, screen->tesla, 0x1280, 3);
-	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+	so_reloc (so, screen->tsc, 0, NOUVEAU_BO_VRAM |
 		  NOUVEAU_BO_RD | NOUVEAU_BO_HIGH, 0, 0);
-	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+	so_reloc (so, screen->tsc, 0, NOUVEAU_BO_VRAM |
 		  NOUVEAU_BO_RD | NOUVEAU_BO_LOW, 0, 0);
-	so_data  (so, 0x00034000);
-	so_method(so, screen->tesla, 0x1280, 3);
-	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+	so_data  (so, (NV50_CB_TSC << 16) | 0x0800);
+	so_method(so, screen->tesla, 0x155c, 3);
+	so_reloc (so, screen->tsc, 0, NOUVEAU_BO_VRAM |
 		  NOUVEAU_BO_RD | NOUVEAU_BO_HIGH, 0, 0);
-	so_reloc (so, screen->constbuf, 0, NOUVEAU_BO_VRAM |
+	so_reloc (so, screen->tsc, 0, NOUVEAU_BO_VRAM |
 		  NOUVEAU_BO_RD | NOUVEAU_BO_LOW, 0, 0);
-	so_data  (so, 0x00040100);
+	so_data  (so, 0x00000800);
+
 
+	/* Vertex array limits - max them out */
 	for (i = 0; i < 16; i++) {
 		so_method(so, screen->tesla, 0x1080 + (i * 8), 2);
 		so_data  (so, 0x000000ff);
diff --git a/src/gallium/drivers/nv50/nv50_screen.h b/src/gallium/drivers/nv50/nv50_screen.h
index d63dd48508..5acb5003ba 100644
--- a/src/gallium/drivers/nv50/nv50_screen.h
+++ b/src/gallium/drivers/nv50/nv50_screen.h
@@ -15,6 +15,9 @@ struct nv50_screen {
 
 	struct pipe_buffer *constbuf;
 	struct nouveau_resource *vp_data_heap;
+
+	struct pipe_buffer *tic;
+	struct pipe_buffer *tsc;
 };
 
 static INLINE struct nv50_screen *
diff --git a/src/gallium/drivers/nv50/nv50_state.c b/src/gallium/drivers/nv50/nv50_state.c
index fd10a38378..ba3d04cede 100644
--- a/src/gallium/drivers/nv50/nv50_state.c
+++ b/src/gallium/drivers/nv50/nv50_state.c
@@ -406,11 +406,11 @@ nv50_set_constant_buffer(struct pipe_context *pipe, uint shader, uint index,
 
 	if (shader == PIPE_SHADER_VERTEX) {
 		nv50->constbuf[PIPE_SHADER_VERTEX] = buf->buffer;
-		nv50->dirty |= NV50_NEW_VERTPROG;
+		nv50->dirty |= NV50_NEW_VERTPROG_CB;
 	} else
 	if (shader == PIPE_SHADER_FRAGMENT) {
 		nv50->constbuf[PIPE_SHADER_FRAGMENT] = buf->buffer;
-		nv50->dirty |= NV50_NEW_FRAGPROG;
+		nv50->dirty |= NV50_NEW_FRAGPROG_CB;
 	}
 }
 
diff --git a/src/gallium/drivers/nv50/nv50_state.h b/src/gallium/drivers/nv50/nv50_state.h
index be0c75ad6e..9e3876871b 100644
--- a/src/gallium/drivers/nv50/nv50_state.h
+++ b/src/gallium/drivers/nv50/nv50_state.h
@@ -20,7 +20,6 @@ struct nv50_program {
 
 	struct pipe_buffer *buffer;
 
-	unsigned param_nr;
 	float *immd;
 	unsigned immd_nr;
 
diff --git a/src/gallium/drivers/nv50/nv50_state_validate.c b/src/gallium/drivers/nv50/nv50_state_validate.c
index 4a548378b7..05395c6df7 100644
--- a/src/gallium/drivers/nv50/nv50_state_validate.c
+++ b/src/gallium/drivers/nv50/nv50_state_validate.c
@@ -168,6 +168,35 @@ nv50_state_validate(struct nv50_context *nv50)
 		so_ref(NULL, &so);
 	}
 
+	if (nv50->dirty & NV50_NEW_VERTPROG_CB) {
+		so = so_new(4, 2);
+		so_method(so, tesla, 0x1280, 3);
+		so_reloc (so, nv50->constbuf[PIPE_SHADER_VERTEX], 0,
+			  NOUVEAU_BO_HIGH | NOUVEAU_BO_RD | NOUVEAU_BO_VRAM,
+			  0, 0);
+		so_reloc (so, nv50->constbuf[PIPE_SHADER_VERTEX], 0,
+			  NOUVEAU_BO_LOW | NOUVEAU_BO_RD | NOUVEAU_BO_VRAM,
+			  0, 0);
+		so_data  (so, (NV50_CB_PVP << 16) | 0x1000);
+		so_emit(nvws, so);
+		so_ref(NULL, &so);
+	}
+
+	if (nv50->dirty & NV50_NEW_FRAGPROG_CB) {
+		so = so_new(4, 2);
+		so_method(so, tesla, 0x1280, 3);
+		so_reloc (so, nv50->constbuf[PIPE_SHADER_FRAGMENT], 0,
+			  NOUVEAU_BO_HIGH | NOUVEAU_BO_RD | NOUVEAU_BO_VRAM,
+			  0, 0);
+		so_reloc (so, nv50->constbuf[PIPE_SHADER_FRAGMENT], 0,
+			  NOUVEAU_BO_LOW | NOUVEAU_BO_RD | NOUVEAU_BO_VRAM,
+			  0, 0);
+		so_data  (so, (NV50_CB_PFP << 16) | 0x1000);
+		so_emit(nvws, so);
+		so_ref(NULL, &so);
+	}
+
+
 	if (nv50->dirty & NV50_NEW_ARRAYS)
 		nv50_vbo_validate(nv50);
 
-- 
cgit v1.2.3


From 65ad8176ca91b5ed2a01b1b3ee145cfdce369419 Mon Sep 17 00:00:00 2001
From: Ben Skeggs <skeggsb@gmail.com>
Date: Tue, 17 Jun 2008 23:43:38 +1000
Subject: nv50: move surface_map/unmap into nv50_surface.c

---
 src/gallium/drivers/nv50/nv50_screen.c  | 26 +------------------------
 src/gallium/drivers/nv50/nv50_screen.h  |  2 ++
 src/gallium/drivers/nv50/nv50_surface.c | 34 +++++++++++++++++++++++++++++++--
 3 files changed, 35 insertions(+), 27 deletions(-)

(limited to 'src/gallium/drivers/nv50/nv50_screen.h')

diff --git a/src/gallium/drivers/nv50/nv50_screen.c b/src/gallium/drivers/nv50/nv50_screen.c
index 8affb0f073..18f22c5960 100644
--- a/src/gallium/drivers/nv50/nv50_screen.c
+++ b/src/gallium/drivers/nv50/nv50_screen.c
@@ -118,28 +118,6 @@ nv50_screen_get_paramf(struct pipe_screen *pscreen, int param)
 	}
 }
 
-static void *
-nv50_surface_map(struct pipe_screen *screen, struct pipe_surface *surface,
-		 unsigned flags )
-{
-	struct pipe_winsys *ws = screen->winsys;
-	void *map;
-
-	map = ws->buffer_map(ws, surface->buffer, flags);
-	if (!map)
-		return NULL;
-
-	return map + surface->offset;
-}
-
-static void
-nv50_surface_unmap(struct pipe_screen *screen, struct pipe_surface *surface)
-{
-	struct pipe_winsys *ws = screen->winsys;
-
-	ws->buffer_unmap(ws, surface->buffer);
-}
-
 static void
 nv50_screen_destroy(struct pipe_screen *pscreen)
 {
@@ -309,10 +287,8 @@ nv50_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws)
 
 	screen->pipe.is_format_supported = nv50_screen_is_format_supported;
 
-	screen->pipe.surface_map = nv50_surface_map;
-	screen->pipe.surface_unmap = nv50_surface_unmap;
-
 	nv50_screen_init_miptree_functions(&screen->pipe);
+	nv50_surface_init_screen_functions(&screen->pipe);
 
 	return &screen->pipe;
 }
diff --git a/src/gallium/drivers/nv50/nv50_screen.h b/src/gallium/drivers/nv50/nv50_screen.h
index 5acb5003ba..08d1f45089 100644
--- a/src/gallium/drivers/nv50/nv50_screen.h
+++ b/src/gallium/drivers/nv50/nv50_screen.h
@@ -26,4 +26,6 @@ nv50_screen(struct pipe_screen *screen)
 	return (struct nv50_screen *)screen;
 }
 
+void nv50_surface_init_screen_functions(struct pipe_screen *);
+
 #endif
diff --git a/src/gallium/drivers/nv50/nv50_surface.c b/src/gallium/drivers/nv50/nv50_surface.c
index 4e294198b0..acd7b501ef 100644
--- a/src/gallium/drivers/nv50/nv50_surface.c
+++ b/src/gallium/drivers/nv50/nv50_surface.c
@@ -57,9 +57,39 @@ nv50_surface_fill(struct pipe_context *pipe, struct pipe_surface *dest,
 	nvws->surface_fill(nvws, dest, destx, desty, width, height, value);
 }
 
+static void *
+nv50_surface_map(struct pipe_screen *screen, struct pipe_surface *surface,
+		 unsigned flags )
+{
+	struct pipe_winsys *ws = screen->winsys;
+	void *map;
+
+	map = ws->buffer_map(ws, surface->buffer, flags);
+	if (!map)
+		return NULL;
+
+	return map + surface->offset;
+}
+
+static void
+nv50_surface_unmap(struct pipe_screen *screen, struct pipe_surface *surface)
+{
+	struct pipe_winsys *ws = screen->winsys;
+
+	ws->buffer_unmap(ws, surface->buffer);
+}
+
 void
 nv50_init_surface_functions(struct nv50_context *nv50)
 {
-   nv50->pipe.surface_copy = nv50_surface_copy;
-   nv50->pipe.surface_fill = nv50_surface_fill;
+	nv50->pipe.surface_copy = nv50_surface_copy;
+	nv50->pipe.surface_fill = nv50_surface_fill;
 }
+
+void
+nv50_surface_init_screen_functions(struct pipe_screen *pscreen)
+{
+	pscreen->surface_map = nv50_surface_map;
+	pscreen->surface_unmap = nv50_surface_unmap;
+}
+
-- 
cgit v1.2.3


From 714cb4a86c1f503334b37ca6c24272fa1bdf7899 Mon Sep 17 00:00:00 2001
From: Ben Skeggs <skeggsb@gmail.com>
Date: Tue, 17 Jun 2008 23:55:23 +1000
Subject: nv50: make sure static buffers (constbuf, tex control etc) get on
 reloc list

---
 src/gallium/drivers/nv50/nv50_screen.c         | 2 +-
 src/gallium/drivers/nv50/nv50_screen.h         | 2 ++
 src/gallium/drivers/nv50/nv50_state_validate.c | 1 +
 3 files changed, 4 insertions(+), 1 deletion(-)

(limited to 'src/gallium/drivers/nv50/nv50_screen.h')

diff --git a/src/gallium/drivers/nv50/nv50_screen.c b/src/gallium/drivers/nv50/nv50_screen.c
index 18f22c5960..5b4c5f96ac 100644
--- a/src/gallium/drivers/nv50/nv50_screen.c
+++ b/src/gallium/drivers/nv50/nv50_screen.c
@@ -273,7 +273,7 @@ nv50_screen_create(struct pipe_winsys *ws, struct nouveau_winsys *nvws)
 	so_data  (so, 1);
 
 	so_emit(nvws, so);
-	so_ref(NULL, &so);
+	so_ref(so, &screen->static_init);
 	nvws->push_flush(nvws, 0, NULL);
 
 	screen->pipe.winsys = ws;
diff --git a/src/gallium/drivers/nv50/nv50_screen.h b/src/gallium/drivers/nv50/nv50_screen.h
index 08d1f45089..400ddcef06 100644
--- a/src/gallium/drivers/nv50/nv50_screen.h
+++ b/src/gallium/drivers/nv50/nv50_screen.h
@@ -18,6 +18,8 @@ struct nv50_screen {
 
 	struct pipe_buffer *tic;
 	struct pipe_buffer *tsc;
+
+	struct nouveau_stateobj *static_init;
 };
 
 static INLINE struct nv50_screen *
diff --git a/src/gallium/drivers/nv50/nv50_state_validate.c b/src/gallium/drivers/nv50/nv50_state_validate.c
index d99cdc73ca..8229bce89e 100644
--- a/src/gallium/drivers/nv50/nv50_state_validate.c
+++ b/src/gallium/drivers/nv50/nv50_state_validate.c
@@ -151,6 +151,7 @@ nv50_state_emit(struct nv50_context *nv50)
 	so_emit_reloc_markers(nvws, nv50->state.vertprog);
 	so_emit_reloc_markers(nvws, nv50->state.fragprog);
 	so_emit_reloc_markers(nvws, nv50->state.vtxbuf);
+	so_emit_reloc_markers(nvws, nv50->screen->static_init);
 }
 
 boolean
-- 
cgit v1.2.3