nouveau: remove useless NOUVEAU_BO_SWIZZLED flag, copy/paste nv40 work to swizzle textures

6 files changed, 167 insertions, 42 deletions

diff --git a/src/gallium/drivers/nouveau/nouveau_bo.h b/src/gallium/drivers/nouveau/nouveau_bo.h
index 0ed3367815..65b138283c 100644
--- a/src/gallium/drivers/nouveau/nouveau_bo.h
+++ b/src/gallium/drivers/nouveau/nouveau_bo.h
@@ -37,7 +37,6 @@
 #define NOUVEAU_BO_LOCAL (1 << 9)
 #define NOUVEAU_BO_TILED (1 << 10)
 #define NOUVEAU_BO_ZTILE (1 << 11)
-#define NOUVEAU_BO_SWIZZLED (1 << 12)
 #define NOUVEAU_BO_DUMMY (1 << 31)
 
 struct nouveau_bo {
diff --git a/src/gallium/drivers/nv30/nv30_fragtex.c b/src/gallium/drivers/nv30/nv30_fragtex.c
index 9e6f746a42..efba8db86d 100644
--- a/src/gallium/drivers/nv30/nv30_fragtex.c
+++ b/src/gallium/drivers/nv30/nv30_fragtex.c
@@ -26,7 +26,7 @@ static INLINE int log2i(int i)
 	return r;
 }
 
-#define _(m,tf,ts0x,ts0y,ts0z,ts0w,ts1x,ts1y,ts1z,ts1w,swsurf)                 \
+#define _(m,tf,ts0x,ts0y,ts0z,ts0w,ts1x,ts1y,ts1z,ts1w)                        \
 {                                                                              \
   TRUE,                                                                        \
   PIPE_FORMAT_##m,                                                             \
@@ -34,8 +34,7 @@ static INLINE int log2i(int i)
   (NV34TCL_TX_SWIZZLE_S0_X_##ts0x | NV34TCL_TX_SWIZZLE_S0_Y_##ts0y |           \
    NV34TCL_TX_SWIZZLE_S0_Z_##ts0z | NV34TCL_TX_SWIZZLE_S0_W_##ts0w |           \
    NV34TCL_TX_SWIZZLE_S1_X_##ts1x | NV34TCL_TX_SWIZZLE_S1_Y_##ts1y |           \
-   NV34TCL_TX_SWIZZLE_S1_Z_##ts1z | NV34TCL_TX_SWIZZLE_S1_W_##ts1w),           \
-  swsurf                                                                       \
+   NV34TCL_TX_SWIZZLE_S1_Z_##ts1z | NV34TCL_TX_SWIZZLE_S1_W_##ts1w)            \
 }
 
 struct nv30_texture_format {
@@ -43,25 +42,24 @@ struct nv30_texture_format {
 	uint	pipe;
 	int     format;
 	int     swizzle;
-	int	swizzled_surface;
 };
 
 static struct nv30_texture_format
 nv30_texture_formats[] = {
-	_(A8R8G8B8_UNORM, A8R8G8B8,   S1,   S1,   S1,   S1, X, Y, Z, W, 1),
-	_(A1R5G5B5_UNORM, A1R5G5B5,   S1,   S1,   S1,   S1, X, Y, Z, W, 1),
-	_(A4R4G4B4_UNORM, A4R4G4B4,   S1,   S1,   S1,   S1, X, Y, Z, W, 1),
-	_(R5G6B5_UNORM  , R5G6B5  ,   S1,   S1,   S1,  ONE, X, Y, Z, W, 1),
-	_(L8_UNORM      , L8      ,   S1,   S1,   S1,  ONE, X, X, X, X, 1),
-	_(A8_UNORM      , L8      , ZERO, ZERO, ZERO,   S1, X, X, X, X, 1),
-	_(I8_UNORM      , L8      ,   S1,   S1,   S1,   S1, X, X, X, X, 1),
-	_(A8L8_UNORM    , A8L8    ,   S1,   S1,   S1,   S1, X, X, X, Y, 1),
-//	_(Z16_UNORM     , Z16     ,   S1,   S1,   S1,  ONE, X, X, X, X, 0),
-//	_(Z24S8_UNORM   , Z24     ,   S1,   S1,   S1,  ONE, X, X, X, X, 0),
-	_(DXT1_RGB      , DXT1    ,   S1,   S1,   S1,  ONE, X, Y, Z, W, 0),
-	_(DXT1_RGBA     , DXT1    ,   S1,   S1,   S1,   S1, X, Y, Z, W, 0),
-	_(DXT3_RGBA     , DXT3    ,   S1,   S1,   S1,   S1, X, Y, Z, W, 0),
-	_(DXT5_RGBA     , DXT5    ,   S1,   S1,   S1,   S1, X, Y, Z, W, 0),
+	_(A8R8G8B8_UNORM, A8R8G8B8,   S1,   S1,   S1,   S1, X, Y, Z, W),
+	_(A1R5G5B5_UNORM, A1R5G5B5,   S1,   S1,   S1,   S1, X, Y, Z, W),
+	_(A4R4G4B4_UNORM, A4R4G4B4,   S1,   S1,   S1,   S1, X, Y, Z, W),
+	_(R5G6B5_UNORM  , R5G6B5  ,   S1,   S1,   S1,  ONE, X, Y, Z, W),
+	_(L8_UNORM      , L8      ,   S1,   S1,   S1,  ONE, X, X, X, X),
+	_(A8_UNORM      , L8      , ZERO, ZERO, ZERO,   S1, X, X, X, X),
+	_(I8_UNORM      , L8      ,   S1,   S1,   S1,   S1, X, X, X, X),
+	_(A8L8_UNORM    , A8L8    ,   S1,   S1,   S1,   S1, X, X, X, Y),
+//	_(Z16_UNORM     , Z16     ,   S1,   S1,   S1,  ONE, X, X, X, X),
+//	_(Z24S8_UNORM   , Z24     ,   S1,   S1,   S1,  ONE, X, X, X, X),
+	_(DXT1_RGB      , DXT1    ,   S1,   S1,   S1,  ONE, X, Y, Z, W),
+	_(DXT1_RGBA     , DXT1    ,   S1,   S1,   S1,   S1, X, Y, Z, W),
+	_(DXT3_RGBA     , DXT3    ,   S1,   S1,   S1,   S1, X, Y, Z, W),
+	_(DXT5_RGBA     , DXT5    ,   S1,   S1,   S1,   S1, X, Y, Z, W),
 	{},
 };
 
@@ -90,16 +88,13 @@ nv30_fragtex_build(struct nv30_context *nv30, int unit)
 	struct pipe_texture *pt = &nv30mt->base;
 	struct nv30_texture_format *tf;
 	struct nouveau_stateobj *so;
-	uint32_t txf, txs /*, txp*/;
-	/*int swizzled = 0;*/ /*XXX: implement in region code? */
+	uint32_t txf, txs , txp;
 	unsigned tex_flags = NOUVEAU_BO_VRAM | NOUVEAU_BO_GART | NOUVEAU_BO_RD;
 
 	tf = nv30_fragtex_format(pt->format);
 	if (!tf)
 		assert(0);
 
-	tex_flags |= (tf->swizzled_surface ? NOUVEAU_BO_SWIZZLED : 0);
-
 	txf  = tf->format;
 	txf |= ((pt->last_level>0) ? NV34TCL_TX_FORMAT_MIPMAP : 0);
 	txf |= log2i(pt->width[0]) << 20;
@@ -125,6 +120,13 @@ nv30_fragtex_build(struct nv30_context *nv30, int unit)
 		return NULL;
 	}
 
+	if (!(pt->tex_usage & NOUVEAU_TEXTURE_USAGE_LINEAR)) {
+		txp = 0;
+	} else {
+		txp  = nv30mt->level[0].pitch;
+		txf |= (1<<13) /*FIXME: NV34TCL_TX_FORMAT_LINEAR ? */;
+	}
+
 	txs = tf->swizzle;
 
 	so = so_new(16, 2);
diff --git a/src/gallium/drivers/nv30/nv30_miptree.c b/src/gallium/drivers/nv30/nv30_miptree.c
index 9124db03e8..aa670b9a45 100644
--- a/src/gallium/drivers/nv30/nv30_miptree.c
+++ b/src/gallium/drivers/nv30/nv30_miptree.c
@@ -65,6 +65,29 @@ nv30_miptree_create(struct pipe_screen *pscreen, const struct pipe_texture *pt)
 	mt->base = *pt;
 	mt->base.refcount = 1;
 	mt->base.screen = pscreen;
+	mt->shadow_tex = NULL;
+	mt->shadow_surface = NULL;
+
+	/* Swizzled textures must be POT */
+	if (pt->width[0] & (pt->width[0] - 1) ||
+	    pt->height[0] & (pt->height[0] - 1))
+		mt->base.tex_usage |= NOUVEAU_TEXTURE_USAGE_LINEAR;
+	else
+	if (pt->tex_usage & (PIPE_TEXTURE_USAGE_PRIMARY |
+	                     PIPE_TEXTURE_USAGE_DISPLAY_TARGET))
+		mt->base.tex_usage |= NOUVEAU_TEXTURE_USAGE_LINEAR;
+	else {
+		switch (pt->format) {
+		/* TODO: Figure out which formats can be swizzled */
+		case PIPE_FORMAT_A8R8G8B8_UNORM:
+		case PIPE_FORMAT_X8R8G8B8_UNORM:
+		/* XXX: Re-enable when SIFM size limits are fixed */
+		/*case PIPE_FORMAT_R16_SNORM:*/
+			break;
+		default:
+			mt->base.tex_usage |= NOUVEAU_TEXTURE_USAGE_LINEAR;
+		}
+	}
 
 	nv30_miptree_layout(mt);
 
@@ -81,22 +104,29 @@ nv30_miptree_create(struct pipe_screen *pscreen, const struct pipe_texture *pt)
 }
 
 static void
-nv30_miptree_release(struct pipe_screen *pscreen, struct pipe_texture **pt)
+nv30_miptree_release(struct pipe_screen *pscreen, struct pipe_texture **ppt)
 {
-	struct pipe_texture *mt = *pt;
+	struct pipe_texture *pt = *ppt;
+	struct nv30_miptree *mt = (struct nv30_miptree *)pt;
+	int l;
 
-	*pt = NULL;
-	if (--mt->refcount <= 0) {
-		struct nv30_miptree *nv30mt = (struct nv30_miptree *)mt;
-		int l;
+	*ppt = NULL;
+	if (--pt->refcount)
+		return;
 
-		pipe_buffer_reference(pscreen, &nv30mt->buffer, NULL);
-		for (l = 0; l <= mt->last_level; l++) {
-			if (nv30mt->level[l].image_offset)
-				FREE(nv30mt->level[l].image_offset);
-		}
-		FREE(nv30mt);
+	pipe_buffer_reference(pscreen, &mt->buffer, NULL);
+	for (l = 0; l <= pt->last_level; l++) {
+		if (mt->level[l].image_offset)
+			FREE(mt->level[l].image_offset);
+	}
+
+	if (mt->shadow_tex) {
+		assert(mt->shadow_surface);
+		pscreen->tex_surface_release(pscreen, &mt->shadow_surface);
+		nv30_miptree_release(pscreen, &mt->shadow_tex);
 	}
+
+	FREE(mt);
 }
 
 static struct pipe_surface *
@@ -123,6 +153,9 @@ nv30_miptree_surface_new(struct pipe_screen *pscreen, struct pipe_texture *pt,
 	ps->status = PIPE_SURFACE_STATUS_DEFINED;
 	ps->refcount = 1;
 	ps->winsys = pscreen->winsys;
+	ps->face = face;
+	ps->level = level;
+	ps->zslice = zslice;
 
 	if (pt->target == PIPE_TEXTURE_CUBE) {
 		ps->offset = nv30mt->level[level].image_offset[face];
diff --git a/src/gallium/drivers/nv30/nv30_screen.c b/src/gallium/drivers/nv30/nv30_screen.c
index a595e2eb22..910a3c456d 100644
--- a/src/gallium/drivers/nv30/nv30_screen.c
+++ b/src/gallium/drivers/nv30/nv30_screen.c
@@ -128,22 +128,73 @@ static void *
 nv30_surface_map(struct pipe_screen *screen, struct pipe_surface *surface,
 		 unsigned flags )
 {
-	struct pipe_winsys *ws = screen->winsys;
-	void *map;
+	struct pipe_winsys	*ws = screen->winsys;
+	struct pipe_surface	*surface_to_map;
+	void			*map;
+
+	if (!(surface->texture->tex_usage & NOUVEAU_TEXTURE_USAGE_LINEAR)) {
+		struct nv30_miptree *mt = (struct nv30_miptree *)surface->texture;
+
+		if (!mt->shadow_tex) {
+			unsigned old_tex_usage = surface->texture->tex_usage;
+			surface->texture->tex_usage = NOUVEAU_TEXTURE_USAGE_LINEAR;
+			mt->shadow_tex = screen->texture_create(screen, surface->texture);
+			surface->texture->tex_usage = old_tex_usage;
+
+			assert(mt->shadow_tex->tex_usage & NOUVEAU_TEXTURE_USAGE_LINEAR);
+			mt->shadow_surface = screen->get_tex_surface
+			(
+				screen, mt->shadow_tex,
+				surface->face, surface->level, surface->zslice,
+				surface->usage
+			);
+		}
+
+		surface_to_map = mt->shadow_surface;
+	}
+	else
+		surface_to_map = surface;
 
-	map = ws->buffer_map(ws, surface->buffer, flags);
+	assert(surface_to_map);
+
+	map = ws->buffer_map(ws, surface_to_map->buffer, flags);
 	if (!map)
 		return NULL;
 
-	return map + surface->offset;
+	return map + surface_to_map->offset;
 }
 
 static void
 nv30_surface_unmap(struct pipe_screen *screen, struct pipe_surface *surface)
 {
-	struct pipe_winsys *ws = screen->winsys;
+	struct pipe_winsys	*ws = screen->winsys;
+	struct pipe_surface	*surface_to_unmap;
+
+	/* TODO: Copy from shadow just before push buffer is flushed instead.
+	         There are probably some programs that map/unmap excessively
+	         before rendering. */
+	if (!(surface->texture->tex_usage & NOUVEAU_TEXTURE_USAGE_LINEAR)) {
+		struct nv30_miptree *mt = (struct nv30_miptree *)surface->texture;
+
+		assert(mt->shadow_tex);
 
-	ws->buffer_unmap(ws, surface->buffer);
+		surface_to_unmap = mt->shadow_surface;
+	}
+	else
+		surface_to_unmap = surface;
+
+	assert(surface_to_unmap);
+
+	ws->buffer_unmap(ws, surface_to_unmap->buffer);
+
+	if (surface_to_unmap != surface) {
+		struct nv30_screen *nvscreen = nv30_screen(screen);
+
+		nvscreen->nvws->surface_copy(nvscreen->nvws,
+		                             surface, 0, 0,
+		                             surface_to_unmap, 0, 0,
+		                             surface->width, surface->height);
+	}
 }
 
 static void
diff --git a/src/gallium/drivers/nv30/nv30_state.h b/src/gallium/drivers/nv30/nv30_state.h
index e6f23bf166..2023278e37 100644
--- a/src/gallium/drivers/nv30/nv30_state.h
+++ b/src/gallium/drivers/nv30/nv30_state.h
@@ -76,6 +76,9 @@ struct nv30_miptree {
 	struct pipe_buffer *buffer;
 	uint total_size;
 
+	struct pipe_texture *shadow_tex;
+	struct pipe_surface *shadow_surface;
+
 	struct {
 		uint pitch;
 		uint *image_offset;
diff --git a/src/gallium/drivers/nv30/nv30_state_fb.c b/src/gallium/drivers/nv30/nv30_state_fb.c
index d93e2bb5c8..c549b17656 100644
--- a/src/gallium/drivers/nv30/nv30_state_fb.c
+++ b/src/gallium/drivers/nv30/nv30_state_fb.c
@@ -1,5 +1,31 @@
 #include "nv30_context.h"
 
+static INLINE int log2i(int i)
+{
+	int r = 0;
+
+	if (i & 0xffff0000) {
+		i >>= 16;
+		r += 16;
+	}
+	if (i & 0x0000ff00) {
+		i >>= 8;
+		r += 8;
+	}
+	if (i & 0x000000f0) {
+		i >>= 4;
+		r += 4;
+	}
+	if (i & 0x0000000c) {
+		i >>= 2;
+		r += 2;
+	}
+	if (i & 0x00000002) {
+		r += 1;
+	}
+	return r;
+}
+
 static boolean
 nv30_state_framebuffer_validate(struct nv30_context *nv30)
 {
@@ -31,7 +57,18 @@ nv30_state_framebuffer_validate(struct nv30_context *nv30)
 		zeta = fb->zsbuf;
 	}
 
-	rt_format = NV34TCL_RT_FORMAT_TYPE_LINEAR;
+	if (!(rt[0]->texture->tex_usage & NOUVEAU_TEXTURE_USAGE_LINEAR)) {
+		assert(!(fb->width & (fb->width - 1)) && !(fb->height & (fb->height - 1)));
+		for (i = 1; i < fb->num_cbufs; i++)
+			assert(!(rt[i]->texture->tex_usage & NOUVEAU_TEXTURE_USAGE_LINEAR));
+
+		/* FIXME: NV34TCL_RT_FORMAT_LOG2_[WIDTH/HEIGHT] */
+		rt_format = NV34TCL_RT_FORMAT_TYPE_SWIZZLED |
+		log2i(fb->width) << 16 /*NV34TCL_RT_FORMAT_LOG2_WIDTH_SHIFT*/ |
+		log2i(fb->height) << 24 /*NV34TCL_RT_FORMAT_LOG2_HEIGHT_SHIFT*/;
+	}
+	else
+		rt_format = NV34TCL_RT_FORMAT_TYPE_LINEAR;
 
 	switch (colour_format) {
 	case PIPE_FORMAT_A8R8G8B8_UNORM: