nv50,nvc0: implement blit

16 files changed, 1398 insertions, 715 deletions

diff --git a/src/gallium/drivers/nv50/nv50_blit.h b/src/gallium/drivers/nv50/nv50_blit.h
new file mode 100644
index 00000000000..0a1c2ae6773
--- /dev/null
+++ b/src/gallium/drivers/nv50/nv50_blit.h
@@ -0,0 +1,182 @@
+
+#ifndef __NV50_BLIT_H__
+#define __NV50_BLIT_H__
+
+#include "util/u_inlines.h"
+#include "util/u_format.h"
+
+void *
+nv50_blitter_make_fp(struct pipe_context *,
+                     unsigned mode,
+                     enum pipe_texture_target);
+
+unsigned
+nv50_blit_select_mode(const struct pipe_blit_info *);
+
+/* Converted to a pipe->blit. */
+void
+nv50_resource_resolve(struct pipe_context *, const struct pipe_resolve_info *);
+
+#define NV50_BLIT_MODE_PASS  0 /* pass through TEX $t0/$s0 output */
+#define NV50_BLIT_MODE_Z24S8 1 /* encode ZS values for RGBA unorm8 */
+#define NV50_BLIT_MODE_S8Z24 2
+#define NV50_BLIT_MODE_X24S8 3
+#define NV50_BLIT_MODE_S8X24 4
+#define NV50_BLIT_MODE_Z24X8 5
+#define NV50_BLIT_MODE_X8Z24 6
+#define NV50_BLIT_MODE_ZS    7 /* put $t0/$s0 into R, $t1/$s1 into G */
+#define NV50_BLIT_MODE_XS    8 /* put $t1/$s1 into G */
+#define NV50_BLIT_MODES      9
+
+/* CUBE and RECT textures are reinterpreted as 2D(_ARRAY) */
+#define NV50_BLIT_TEXTURE_BUFFER    0
+#define NV50_BLIT_TEXTURE_1D        1
+#define NV50_BLIT_TEXTURE_2D        2
+#define NV50_BLIT_TEXTURE_3D        3
+#define NV50_BLIT_TEXTURE_1D_ARRAY  4
+#define NV50_BLIT_TEXTURE_2D_ARRAY  5
+#define NV50_BLIT_MAX_TEXTURE_TYPES 6
+
+static INLINE unsigned
+nv50_blit_texture_type(enum pipe_texture_target target)
+{
+   switch (target) {
+   case PIPE_TEXTURE_1D: return NV50_BLIT_TEXTURE_1D;
+   case PIPE_TEXTURE_2D: return NV50_BLIT_TEXTURE_2D;
+   case PIPE_TEXTURE_3D: return NV50_BLIT_TEXTURE_3D;
+   case PIPE_TEXTURE_1D_ARRAY: return NV50_BLIT_TEXTURE_1D_ARRAY;
+   case PIPE_TEXTURE_2D_ARRAY: return NV50_BLIT_TEXTURE_2D_ARRAY;
+   default:
+      assert(target == PIPE_BUFFER);
+      return NV50_BLIT_TEXTURE_BUFFER;
+   }
+}
+
+static INLINE unsigned
+nv50_blit_get_tgsi_texture_target(enum pipe_texture_target target)
+{
+   switch (target) {
+   case PIPE_TEXTURE_1D: return TGSI_TEXTURE_1D;
+   case PIPE_TEXTURE_2D: return TGSI_TEXTURE_2D;
+   case PIPE_TEXTURE_3D: return TGSI_TEXTURE_3D;
+   case PIPE_TEXTURE_1D_ARRAY: return TGSI_TEXTURE_1D_ARRAY;
+   case PIPE_TEXTURE_2D_ARRAY: return TGSI_TEXTURE_2D_ARRAY;
+   default:
+      assert(target == PIPE_BUFFER);
+      return TGSI_TEXTURE_BUFFER;
+   }
+}
+
+static INLINE enum pipe_texture_target
+nv50_blit_reinterpret_pipe_texture_target(enum pipe_texture_target target)
+{
+   switch (target) {
+   case PIPE_TEXTURE_CUBE:
+      return PIPE_TEXTURE_2D_ARRAY;
+   case PIPE_TEXTURE_RECT:
+      return PIPE_TEXTURE_2D;
+   default:
+      return target;
+   }
+}
+
+static INLINE unsigned
+nv50_blit_get_filter(const struct pipe_blit_info *info)
+{
+   if (info->dst.resource->nr_samples < info->src.resource->nr_samples)
+      return util_format_is_depth_or_stencil(info->src.format) ? 0 : 1;
+
+   if (info->filter != PIPE_TEX_FILTER_LINEAR)
+      return 0;
+
+   if ((info->dst.box.width ==  info->src.box.width ||
+        info->dst.box.width == -info->src.box.width) &&
+       (info->dst.box.height ==  info->src.box.height ||
+        info->dst.box.height == -info->src.box.height))
+      return 0;
+
+   return 1;
+}
+
+/* Since shaders cannot export stencil, we cannot copy stencil values when
+ * rendering to ZETA, so we attach the ZS surface to a colour render target.
+ */
+static INLINE enum pipe_format
+nv50_blit_zeta_to_colour_format(enum pipe_format format)
+{
+   switch (format) {
+   case PIPE_FORMAT_Z16_UNORM:
+      return PIPE_FORMAT_R16_UNORM;
+   case PIPE_FORMAT_Z24_UNORM_S8_UINT:
+   case PIPE_FORMAT_S8_UINT_Z24_UNORM:
+   case PIPE_FORMAT_Z24X8_UNORM:
+      return PIPE_FORMAT_R8G8B8A8_UNORM;
+   case PIPE_FORMAT_Z32_FLOAT:
+      return PIPE_FORMAT_R32_FLOAT;
+   case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
+      return PIPE_FORMAT_R32G32_FLOAT;
+   default:
+      assert(0);
+      return PIPE_FORMAT_NONE;
+   }
+}
+
+
+static INLINE uint16_t
+nv50_blit_derive_color_mask(const struct pipe_blit_info *info)
+{
+   const unsigned mask = info->mask;
+
+   uint16_t color_mask = 0;
+
+   switch (info->dst.format) {
+   case PIPE_FORMAT_Z24_UNORM_S8_UINT:
+      if (mask & PIPE_MASK_S)
+         color_mask |= 0x1000;
+      /* fall through */
+   case PIPE_FORMAT_Z24X8_UNORM:
+      if (mask & PIPE_MASK_Z)
+         color_mask |= 0x0111;
+      break;
+   case PIPE_FORMAT_S8_UINT_Z24_UNORM:
+      if (mask & PIPE_MASK_Z)
+         color_mask |= 0x1110;
+      if (mask & PIPE_MASK_S)
+         color_mask |= 0x0001;
+      break;
+   default:
+      if (mask & (PIPE_MASK_R | PIPE_MASK_Z)) color_mask |= 0x0001;
+      if (mask & (PIPE_MASK_G | PIPE_MASK_S)) color_mask |= 0x0010;
+      if (mask & PIPE_MASK_B) color_mask |= 0x0100;
+      if (mask & PIPE_MASK_A) color_mask |= 0x1000;
+      break;
+   }
+
+   return color_mask;
+}
+
+static INLINE uint32_t
+nv50_blit_eng2d_get_mask(const struct pipe_blit_info *info)
+{
+   uint32_t mask = 0;
+
+   switch (info->dst.format) {
+   case PIPE_FORMAT_Z24_UNORM_S8_UINT:
+      if (info->mask & PIPE_MASK_Z) mask |= 0x00ffffff;
+      if (info->mask & PIPE_MASK_S) mask |= 0xff000000;
+      break;
+   case PIPE_FORMAT_S8_UINT_Z24_UNORM:
+      if (info->mask & PIPE_MASK_Z) mask |= 0xffffff00;
+      if (info->mask & PIPE_MASK_S) mask |= 0x000000ff;
+      break;
+   case PIPE_FORMAT_X8Z24_UNORM:
+      if (info->mask & PIPE_MASK_Z) mask = 0x00ffffff;
+      break;
+   default:
+      mask = 0xffffffff;
+      break;
+   }
+   return mask;
+}
+
+#endif /* __NV50_BLIT_H__ */
diff --git a/src/gallium/drivers/nv50/nv50_context.c b/src/gallium/drivers/nv50/nv50_context.c
index cb9dd413a2d..24ca9f478f0 100644
--- a/src/gallium/drivers/nv50/nv50_context.c
+++ b/src/gallium/drivers/nv50/nv50_context.c
@@ -128,6 +128,9 @@ nv50_create(struct pipe_screen *pscreen, void *priv)
       return NULL;
    pipe = &nv50->base.pipe;
 
+   if (!nv50_blitctx_create(nv50))
+      goto out_err;
+
    nv50->base.pushbuf = screen->base.pushbuf;
 
    ret = nouveau_bufctx_new(screen->base.client, NV50_BIND_COUNT,
@@ -195,6 +198,8 @@ out_err:
          nouveau_bufctx_del(&nv50->bufctx_3d);
       if (nv50->bufctx)
          nouveau_bufctx_del(&nv50->bufctx);
+      if (nv50->blit)
+         FREE(nv50->blit);
       FREE(nv50);
    }
    return NULL;
diff --git a/src/gallium/drivers/nv50/nv50_context.h b/src/gallium/drivers/nv50/nv50_context.h
index 818c3bb15fe..043ed898913 100644
--- a/src/gallium/drivers/nv50/nv50_context.h
+++ b/src/gallium/drivers/nv50/nv50_context.h
@@ -73,6 +73,10 @@
 #define NV50_CB_AUX 127
 
 
+struct nv50_blitctx;
+
+boolean nv50_blitctx_create(struct nv50_context *);
+
 struct nv50_context {
    struct nouveau_context base;
 
@@ -151,6 +155,11 @@ struct nv50_context {
 
    boolean vbo_push_hint;
 
+   struct pipe_query *cond_query;
+   uint cond_mode;
+
+   struct nv50_blitctx *blit;
+
 #ifdef NV50_WITH_DRAW_MODULE
    struct draw_context *draw;
 #endif
@@ -233,6 +242,12 @@ void nv50_validate_textures(struct nv50_context *);
 void nv50_validate_samplers(struct nv50_context *);
 
 struct pipe_sampler_view *
+nv50_create_texture_view(struct pipe_context *,
+                         struct pipe_resource *,
+                         const struct pipe_sampler_view *,
+                         uint32_t flags,
+                         enum pipe_texture_target);
+struct pipe_sampler_view *
 nv50_create_sampler_view(struct pipe_context *,
                          struct pipe_resource *,
                          const struct pipe_sampler_view *);
diff --git a/src/gallium/drivers/nv50/nv50_query.c b/src/gallium/drivers/nv50/nv50_query.c
index f9ac33d2623..e68537cd7e4 100644
--- a/src/gallium/drivers/nv50/nv50_query.c
+++ b/src/gallium/drivers/nv50/nv50_query.c
@@ -327,6 +327,9 @@ nv50_render_condition(struct pipe_context *pipe,
    struct nouveau_pushbuf *push = nv50->base.pushbuf;
    struct nv50_query *q;
 
+   nv50->cond_query = pq;
+   nv50->cond_mode = mode;
+
    PUSH_SPACE(push, 6);
 
    if (!pq) {
diff --git a/src/gallium/drivers/nv50/nv50_resource.h b/src/gallium/drivers/nv50/nv50_resource.h
index 920fa0a4a2b..c1b159e1244 100644
--- a/src/gallium/drivers/nv50/nv50_resource.h
+++ b/src/gallium/drivers/nv50/nv50_resource.h
@@ -60,6 +60,11 @@ nv50_miptree(struct pipe_resource *pt)
    return (struct nv50_miptree *)pt;
 }
 
+
+#define NV50_TEXVIEW_SCALED_COORDS     (1 << 0)
+#define NV50_TEXVIEW_FILTER_MSAA8      (1 << 1)
+
+
 /* Internal functions:
  */
 boolean
diff --git a/src/gallium/drivers/nv50/nv50_screen.c b/src/gallium/drivers/nv50/nv50_screen.c
index e13c29427d8..ec61e36cb14 100644
--- a/src/gallium/drivers/nv50/nv50_screen.c
+++ b/src/gallium/drivers/nv50/nv50_screen.c
@@ -268,7 +268,8 @@ nv50_screen_destroy(struct pipe_screen *pscreen)
    if (screen->base.pushbuf)
       screen->base.pushbuf->user_priv = NULL;
 
-   FREE(screen->blitctx);
+   if (screen->blitter)
+      nv50_blitter_destroy(screen);
 
    nouveau_bo_ref(NULL, &screen->code);
    nouveau_bo_ref(NULL, &screen->tls_bo);
@@ -750,7 +751,7 @@ nv50_screen_create(struct nouveau_device *dev)
    screen->tic.entries = CALLOC(4096, sizeof(void *));
    screen->tsc.entries = screen->tic.entries + 2048;
 
-   if (!nv50_blitctx_create(screen))
+   if (!nv50_blitter_create(screen))
       goto fail;
 
    nv50_screen_init_hwctx(screen);
diff --git a/src/gallium/drivers/nv50/nv50_screen.h b/src/gallium/drivers/nv50/nv50_screen.h
index 3ecf20c1e03..8784f36d417 100644
--- a/src/gallium/drivers/nv50/nv50_screen.h
+++ b/src/gallium/drivers/nv50/nv50_screen.h
@@ -21,7 +21,7 @@ struct nv50_context;
 
 #define NV50_SCREEN_RESIDENT_BO_COUNT 5
 
-struct nv50_blitctx;
+struct nv50_blitter;
 
 struct nv50_screen {
    struct nouveau_screen base;
@@ -43,7 +43,7 @@ struct nv50_screen {
    struct nouveau_heap *gp_code_heap;
    struct nouveau_heap *fp_code_heap;
 
-   struct nv50_blitctx *blitctx;
+   struct nv50_blitter *blitter;
 
    struct {
       void **entries;
@@ -75,7 +75,8 @@ nv50_screen(struct pipe_screen *screen)
    return (struct nv50_screen *)screen;
 }
 
-boolean nv50_blitctx_create(struct nv50_screen *);
+boolean nv50_blitter_create(struct nv50_screen *);
+void nv50_blitter_destroy(struct nv50_screen *);
 
 int nv50_screen_tic_alloc(struct nv50_screen *, void *);
 int nv50_screen_tsc_alloc(struct nv50_screen *, void *);
diff --git a/src/gallium/drivers/nv50/nv50_surface.c b/src/gallium/drivers/nv50/nv50_surface.c
index 68809ec181b..e64a2badd77 100644
--- a/src/gallium/drivers/nv50/nv50_surface.c
+++ b/src/gallium/drivers/nv50/nv50_surface.c
@@ -29,8 +29,13 @@
 #include "util/u_format.h"
 #include "util/u_surface.h"
 
+#include "tgsi/tgsi_ureg.h"
+
+#include "os/os_thread.h"
+
 #include "nv50_context.h"
 #include "nv50_resource.h"
+#include "nv50_blit.h"
 
 #include "nv50_defs.xml.h"
 #include "nv50_texture.xml.h"
@@ -72,7 +77,8 @@ nv50_2d_format(enum pipe_format format)
 
 static int
 nv50_2d_texture_set(struct nouveau_pushbuf *push, int dst,
-                    struct nv50_miptree *mt, unsigned level, unsigned layer)
+                    struct nv50_miptree *mt, unsigned level, unsigned layer,
+                    enum pipe_format pformat)
 {
    struct nouveau_bo *bo = mt->base.bo;
    uint32_t width, height, depth;
@@ -80,23 +86,26 @@ nv50_2d_texture_set(struct nouveau_pushbuf *push, int dst,
    uint32_t mthd = dst ? NV50_2D_DST_FORMAT : NV50_2D_SRC_FORMAT;
    uint32_t offset = mt->level[level].offset;
 
-   format = nv50_2d_format(mt->base.base.format);
+   format = nv50_2d_format(pformat);
    if (!format) {
       NOUVEAU_ERR("invalid/unsupported surface format: %s\n",
-                  util_format_name(mt->base.base.format));
+                  util_format_name(pformat));
       return 1;
    }
 
    width = u_minify(mt->base.base.width0, level) << mt->ms_x;
    height = u_minify(mt->base.base.height0, level) << mt->ms_y;
+   depth = u_minify(mt->base.base.depth0, level);
 
    offset = mt->level[level].offset;
    if (!mt->layout_3d) {
       offset += mt->layer_stride * layer;
       depth = 1;
       layer = 0;
-   } else {
-      depth = u_minify(mt->base.base.depth0, level);
+   } else
+   if (!dst) {
+      offset += nv50_mt_zslice_offset(mt, level, layer);
+      layer = 0;
    }
 
    if (!nouveau_bo_memtype(bo)) {
@@ -143,43 +152,34 @@ nv50_2d_texture_do_copy(struct nouveau_pushbuf *push,
                         unsigned sx, unsigned sy, unsigned sz,
                         unsigned w, unsigned h)
 {
-   static const uint32_t duvdxy[5] =
-   {
-      0x40000000, 0x80000000, 0x00000001, 0x00000002, 0x00000004
-   };
-
+   const enum pipe_format dfmt = dst->base.base.format;
+   const enum pipe_format sfmt = src->base.base.format;
    int ret;
-   uint32_t ctrl;
-#if 0
-   ret = MARK_RING(chan, 2 * 16 + 32, 4);
+
+   ret = PUSH_SPACE(push, 2 * 16 + 32);
    if (ret)
       return ret;
-#endif
-   ret = nv50_2d_texture_set(push, 1, dst, dst_level, dz);
+
+   ret = nv50_2d_texture_set(push, 1, dst, dst_level, dz, dfmt);
    if (ret)
       return ret;
 
-   ret = nv50_2d_texture_set(push, 0, src, src_level, sz);
+   ret = nv50_2d_texture_set(push, 0, src, src_level, sz, sfmt);
    if (ret)
       return ret;
 
-   /* NOTE: 2D engine doesn't work for MS8 */
-   if (src->ms_x)
-      ctrl = 0x11;
-
-   /* 0/1 = CENTER/CORNER, 00/10 = POINT/BILINEAR */
    BEGIN_NV04(push, NV50_2D(BLIT_CONTROL), 1);
-   PUSH_DATA (push, ctrl);
+   PUSH_DATA (push, NV50_2D_BLIT_CONTROL_FILTER_POINT_SAMPLE);
    BEGIN_NV04(push, NV50_2D(BLIT_DST_X), 4);
    PUSH_DATA (push, dx << dst->ms_x);
    PUSH_DATA (push, dy << dst->ms_y);
    PUSH_DATA (push, w << dst->ms_x);
    PUSH_DATA (push, h << dst->ms_y);
    BEGIN_NV04(push, NV50_2D(BLIT_DU_DX_FRACT), 4);
-   PUSH_DATA (push, duvdxy[2 + ((int)src->ms_x - (int)dst->ms_x)] & 0xf0000000);
-   PUSH_DATA (push, duvdxy[2 + ((int)src->ms_x - (int)dst->ms_x)] & 0x0000000f);
-   PUSH_DATA (push, duvdxy[2 + ((int)src->ms_y - (int)dst->ms_y)] & 0xf0000000);
-   PUSH_DATA (push, duvdxy[2 + ((int)src->ms_y - (int)dst->ms_y)] & 0x0000000f);
+   PUSH_DATA (push, 0);
+   PUSH_DATA (push, 1);
+   PUSH_DATA (push, 0);
+   PUSH_DATA (push, 1);
    BEGIN_NV04(push, NV50_2D(BLIT_SRC_X_FRACT), 4);
    PUSH_DATA (push, 0);
    PUSH_DATA (push, sx << src->ms_x);
@@ -425,9 +425,27 @@ nv50_clear(struct pipe_context *pipe, unsigned buffers,
 }
 
 
+/* =============================== BLIT CODE ===================================
+ */
+
+struct nv50_blitter
+{
+   struct nv50_program *fp[NV50_BLIT_MAX_TEXTURE_TYPES][NV50_BLIT_MODES];
+   struct nv50_program vp;
+
+   struct nv50_tsc_entry sampler[2]; /* nearest, bilinear */
+
+   pipe_mutex mutex;
+};
+
 struct nv50_blitctx
 {
-   struct nv50_screen *screen;
+   struct nv50_context *nv50;
+   struct nv50_program *fp;
+   uint8_t mode;
+   uint16_t color_mask;
+   uint8_t filter;
+   enum pipe_texture_target target;
    struct {
       struct pipe_framebuffer_state fb;
       struct nv50_program *vp;
@@ -437,31 +455,20 @@ struct nv50_blitctx
       unsigned num_samplers[3];
       struct pipe_sampler_view *texture[2];
       struct nv50_tsc_entry *sampler[2];
-      unsigned dirty;
+      uint32_t dirty;
    } saved;
-   struct nv50_program vp;
-   struct nv50_program fp;
-   struct nv50_tsc_entry sampler[2]; /* nearest, bilinear */
-   uint32_t fp_offset;
-   uint16_t color_mask;
-   uint8_t filter;
 };
 
 static void
-nv50_blitctx_make_vp(struct nv50_blitctx *blit)
+nv50_blitter_make_vp(struct nv50_blitter *blit)
 {
    static const uint32_t code[] =
    {
-      0x10000001, /* mov b32 o[0x00] s[0x00] */ /* HPOS.x */
-      0x0423c788,
-      0x10000205, /* mov b32 o[0x04] s[0x04] */ /* HPOS.y */
-      0x0423c788,
-      0x10000409, /* mov b32 o[0x08] s[0x08] */ /* TEXC.x */
-      0x0423c788,
-      0x1000060d, /* mov b32 o[0x0c] s[0x0c] */ /* TEXC.y */
-      0x0423c788,
-      0x10000811, /* exit mov b32 o[0x10] s[0x10] */ /* TEXC.z */
-      0x0423c789,
+      0x10000001, 0x0423c788, /* mov b32 o[0x00] s[0x00] */ /* HPOS.x */
+      0x10000205, 0x0423c788, /* mov b32 o[0x04] s[0x04] */ /* HPOS.y */
+      0x10000409, 0x0423c788, /* mov b32 o[0x08] s[0x08] */ /* TEXC.x */
+      0x1000060d, 0x0423c788, /* mov b32 o[0x0c] s[0x0c] */ /* TEXC.y */
+      0x10000811, 0x0423c789, /* mov b32 o[0x10] s[0x10] */ /* TEXC.z */
    };
 
    blit->vp.type = PIPE_SHADER_VERTEX;
@@ -481,119 +488,123 @@ nv50_blitctx_make_vp(struct nv50_blitctx *blit)
    blit->vp.vp.edgeflag = 0x40;
 }
 
-static void
-nv50_blitctx_make_fp(struct nv50_blitctx *blit)
+void *
+nv50_blitter_make_fp(struct pipe_context *pipe,
+                     unsigned mode,
+                     enum pipe_texture_target ptarg)
 {
-   static const uint32_t code[] =
-   {
-      /* 3 coords RGBA in, RGBA out, also for Z32_FLOAT(_S8X24_UINT) */
-      0x80000000, /* interp $r0 v[0x0] */
-      0x80010004, /* interp $r1 v[0x4] */
-      0x80020009, /* interp $r2 flat v[0x8] */
-      0x00040780,
-      0xf6800001, /* texauto live { $r0,1,2,3 } $t0 $s0 { $r0,1,2 } */
-      0x0000c785, /* exit */
-
-      /* 3 coords ZS in, S encoded in R, Z encoded in GBA (8_UNORM) */
-      0x80000000, /* interp $r0 v[0x00] */
-      0x80010004, /* interp $r1 v[0x04] */
-      0x80020108, /* interp $r2 flat v[0x8] */
-      0x10008010, /* mov b32 $r4 $r0 */
-      0xf2820201, /* texauto live { $r0,#,#,# } $t1 $s1 { $r0,1,2 } */
-      0x00000784,
-      0xa000000d, /* cvt f32 $r3 s32 $r0 */
-      0x44014780,
-      0x10000801, /* mov b32 $r0 $r4 */
-      0x0403c780,
-      0xf2800001, /* texauto live { $r0,#,#,# } $t0 $s0 { $r0,1,2 } */
-      0x00000784,
-      0xc03f0009, /* mul f32 $r2 $r0 (2^24 - 1) */
-      0x04b7ffff,
-      0xa0000409, /* cvt rni s32 $r2 f32 $r2 */
-      0x8c004780,
-      0xc0010601, /* mul f32 $r0 $r3 1/0xff */
-      0x03b8080b,
-      0xd03f0405, /* and b32 $r1 $r2 0x0000ff */
-      0x0000000f,
-      0xd000040d, /* and b32 $r3 $r2 0xff0000 */
-      0x000ff003,
-      0xd0000409, /* and b32 $r2 $r2 0x00ff00 */
-      0x00000ff3,
-      0xa0000205, /* cvt f32 $r1 s32 $r1 */
-      0x44014780,
-      0xa000060d, /* cvt f32 $r3 s32 $r3 */
-      0x44014780,
-      0xa0000409, /* cvt f32 $r2 s32 $r2 */
-      0x44014780,
-      0xc0010205, /* mul f32 $r1 $r1 1/0x0000ff */
-      0x03b8080b,
-      0xc001060d, /* mul f32 $r3 $r3 1/0x00ff00 */
-      0x0338080b,
-      0xc0010409, /* mul f32 $r2 $r2 1/0xff0000 */
-      0x0378080b,
-      0xf0000001, /* exit never nop */
-      0xe0000001,
-
-      /* 3 coords ZS in, Z encoded in RGB, S encoded in A (U8_UNORM) */
-      0x80000000, /* interp $r0 v[0x00] */
-      0x80010004, /* interp $r1 v[0x04] */
-      0x80020108, /* interp $r2 flat v[0x8] */
-      0x10008010, /* mov b32 $r4 $r0 */
-      0xf2820201, /* texauto live { $r0,#,#,# } $t1 $s1 { $r0,1,2 } */
-      0x00000784,
-      0xa000000d, /* cvt f32 $r3 s32 $r0 */
-      0x44014780,
-      0x10000801, /* mov b32 $r0 $r4 */
-      0x0403c780,
-      0xf2800001, /* texauto live { $r0,#,#,# } $t0 $s0 { $r0,1,2 } */
-      0x00000784,
-      0xc03f0009, /* mul f32 $r2 $r0 (2^24 - 1) */
-      0x04b7ffff,
-      0xa0000409, /* cvt rni s32 $r2 f32 $r2 */
-      0x8c004780,
-      0xc001060d, /* mul f32 $r3 $r3 1/0xff */
-      0x03b8080b,
-      0xd03f0401, /* and b32 $r0 $r2 0x0000ff */
-      0x0000000f,
-      0xd0000405, /* and b32 $r1 $r2 0x00ff00 */
-      0x00000ff3,
-      0xd0000409, /* and b32 $r2 $r2 0xff0000 */
-      0x000ff003,
-      0xa0000001, /* cvt f32 $r0 s32 $r0 */
-      0x44014780,
-      0xa0000205, /* cvt f32 $r1 s32 $r1 */
-      0x44014780,
-      0xa0000409, /* cvt f32 $r2 s32 $r2 */
-      0x44014780,
-      0xc0010001, /* mul f32 $r0 $r0 1/0x0000ff */
-      0x03b8080b,
-      0xc0010205, /* mul f32 $r1 $r1 1/0x00ff00 */
-      0x0378080b,
-      0xc0010409, /* mul f32 $r2 $r2 1/0xff0000 */
-      0x0338080b,
-      0xf0000001, /* exit never nop */
-      0xe0000001
-   };
+   struct ureg_program *ureg;
+   struct ureg_src tc;
+   struct ureg_dst out;
+   struct ureg_dst data;
+
+   const unsigned target = nv50_blit_get_tgsi_texture_target(ptarg);
+
+   boolean tex_rgbaz = FALSE;
+   boolean tex_s = FALSE;
+   boolean cvt_un8 = FALSE;
+
+   if (mode != NV50_BLIT_MODE_PASS &&
+       mode != NV50_BLIT_MODE_Z24X8 &&
+       mode != NV50_BLIT_MODE_X8Z24)
+      tex_s = TRUE;
+
+   if (mode != NV50_BLIT_MODE_X24S8 &&
+       mode != NV50_BLIT_MODE_S8X24 &&
+       mode != NV50_BLIT_MODE_XS)
+      tex_rgbaz = TRUE;
+
+   if (mode != NV50_BLIT_MODE_PASS &&
+       mode != NV50_BLIT_MODE_ZS &&
+       mode != NV50_BLIT_MODE_XS)
+      cvt_un8 = TRUE;
+
+   ureg = ureg_create(TGSI_PROCESSOR_FRAGMENT);
+   if (!ureg)
+      return NULL;
+
+   out = ureg_DECL_output(ureg, TGSI_SEMANTIC_COLOR, 0);
+   tc = ureg_DECL_fs_input(
+      ureg, TGSI_SEMANTIC_GENERIC, 0, TGSI_INTERPOLATE_LINEAR);
+
+   data = ureg_DECL_temporary(ureg);
+
+   if (tex_s) {
+      ureg_TEX(ureg, ureg_writemask(data, TGSI_WRITEMASK_X),
+               target, tc, ureg_DECL_sampler(ureg, 1));
+      ureg_MOV(ureg, ureg_writemask(data, TGSI_WRITEMASK_Y),
+               ureg_scalar(ureg_src(data), TGSI_SWIZZLE_X));
+   }
+   if (tex_rgbaz) {
+      const unsigned mask = (mode == NV50_BLIT_MODE_PASS) ?
+         TGSI_WRITEMASK_XYZW : TGSI_WRITEMASK_X;
+      ureg_TEX(ureg, ureg_writemask(data, mask),
+               target, tc, ureg_DECL_sampler(ureg, 0));
+   }
+
+   if (cvt_un8) {
+      struct ureg_src mask;
+      struct ureg_src scale;
+      struct ureg_dst outz;
+      struct ureg_dst outs;
+      struct ureg_dst zdst3 = ureg_writemask(data, TGSI_WRITEMASK_XYZ);
+      struct ureg_dst zdst = ureg_writemask(data, TGSI_WRITEMASK_X);
+      struct ureg_dst sdst = ureg_writemask(data, TGSI_WRITEMASK_Y);
+      struct ureg_src zsrc3 = ureg_src(data);
+      struct ureg_src zsrc = ureg_scalar(zsrc3, TGSI_SWIZZLE_X);
+      struct ureg_src ssrc = ureg_scalar(zsrc3, TGSI_SWIZZLE_Y);
+      struct ureg_src zshuf;
+
+      mask = ureg_imm3u(ureg, 0x0000ff, 0x00ff00, 0xff0000);
+      scale = ureg_imm4f(ureg,
+                         1.0f / 0x0000ff, 1.0f / 0x00ff00, 1.0f / 0xff0000,
+                         (1 << 24) - 1);
+
+      if (mode == NV50_BLIT_MODE_Z24S8 ||
+          mode == NV50_BLIT_MODE_X24S8 ||
+          mode == NV50_BLIT_MODE_Z24X8) {
+         outz = ureg_writemask(out, TGSI_WRITEMASK_XYZ);
+         outs = ureg_writemask(out, TGSI_WRITEMASK_W);
+         zshuf = ureg_src(data);
+      } else {
+         outz = ureg_writemask(out, TGSI_WRITEMASK_YZW);
+         outs = ureg_writemask(out, TGSI_WRITEMASK_X);
+         zshuf = ureg_swizzle(zsrc3, TGSI_SWIZZLE_W,
+                              TGSI_SWIZZLE_X, TGSI_SWIZZLE_Y, TGSI_SWIZZLE_Z);
+      }
 
-   blit->fp.type = PIPE_SHADER_FRAGMENT;
-   blit->fp.translated = TRUE;
-   blit->fp.code = (uint32_t *)code; /* const_cast */
-   blit->fp.code_size = sizeof(code);
-   blit->fp.max_gpr = 5;
-   blit->fp.max_out = 4;
-   blit->fp.in_nr = 1;
-   blit->fp.in[0].mask = 0x7; /* last component flat */
-   blit->fp.in[0].linear = 1;
-   blit->fp.in[0].sn = TGSI_SEMANTIC_GENERIC;
-   blit->fp.out_nr = 1;
-   blit->fp.out[0].mask = 0xf;
-   blit->fp.out[0].sn = TGSI_SEMANTIC_COLOR;
-   blit->fp.fp.interp = 0x00020403;
-   blit->fp.gp.primid = 0x80;
+      if (tex_s) {
+         ureg_I2F(ureg, sdst, ssrc);
+         ureg_MUL(ureg, outs, ssrc, ureg_scalar(scale, TGSI_SWIZZLE_X));
+      }
+
+      if (tex_rgbaz) {
+         ureg_MUL(ureg, zdst, zsrc, ureg_scalar(scale, TGSI_SWIZZLE_W));
+         ureg_F2I(ureg, zdst, zsrc);
+         ureg_AND(ureg, zdst3, zsrc, mask);
+         ureg_I2F(ureg, zdst3, zsrc3);
+         ureg_MUL(ureg, zdst3, zsrc3, scale);
+         ureg_MOV(ureg, outz, zshuf);
+      }
+   } else {
+      unsigned mask = TGSI_WRITEMASK_XYZW;
+
+      if (mode != NV50_BLIT_MODE_PASS) {
+         mask &= ~TGSI_WRITEMASK_ZW;
+         if (!tex_s)
+            mask = TGSI_WRITEMASK_X;
+         if (!tex_rgbaz)
+            mask = TGSI_WRITEMASK_Y;
+      }
+      ureg_MOV(ureg, ureg_writemask(out, mask), ureg_src(data));
+   }
+   ureg_END(ureg);
+
+   return ureg_create_shader_and_destroy(ureg, pipe);
 }
 
 static void
-nv50_blitctx_make_sampler(struct nv50_blitctx *blit)
+nv50_blitter_make_sampler(struct nv50_blitter *blit)
 {
    /* clamp to edge, min/max lod = 0, nearest filtering */
 
@@ -615,73 +626,85 @@ nv50_blitctx_make_sampler(struct nv50_blitctx *blit)
       NV50_TSC_1_MAGF_LINEAR | NV50_TSC_1_MINF_LINEAR | NV50_TSC_1_MIPF_NONE;
 }
 
-/* Since shaders cannot export stencil, we cannot copy stencil values when
- * rendering to ZETA, so we attach the ZS surface to a colour render target.
- */
-static INLINE enum pipe_format
-nv50_blit_zeta_to_colour_format(enum pipe_format format)
+unsigned
+nv50_blit_select_mode(const struct pipe_blit_info *info)
 {
-   switch (format) {
-   case PIPE_FORMAT_Z16_UNORM:               return PIPE_FORMAT_R16_UNORM;
+   const unsigned mask = info->mask;
+
+   switch (info->dst.resource->format) {
    case PIPE_FORMAT_Z24_UNORM_S8_UINT:
+   case PIPE_FORMAT_Z24X8_UNORM:
+      switch (mask & PIPE_MASK_ZS) {
+      case PIPE_MASK_ZS: return NV50_BLIT_MODE_Z24S8;
+      case PIPE_MASK_Z:  return NV50_BLIT_MODE_Z24X8;
+      default:
+         return NV50_BLIT_MODE_X24S8;
+      }
    case PIPE_FORMAT_S8_UINT_Z24_UNORM:
-   case PIPE_FORMAT_Z24X8_UNORM:             return PIPE_FORMAT_R8G8B8A8_UNORM;
-   case PIPE_FORMAT_Z32_FLOAT:               return PIPE_FORMAT_R32_FLOAT;
-   case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT: return PIPE_FORMAT_R32G32_FLOAT;
+      switch (mask & PIPE_MASK_ZS) {
+      case PIPE_MASK_ZS: return NV50_BLIT_MODE_S8Z24;
+      case PIPE_MASK_Z:  return NV50_BLIT_MODE_X8Z24;
+      default:
+         return NV50_BLIT_MODE_S8X24;
+      }
+   case PIPE_FORMAT_Z32_FLOAT:
+   case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:
+      switch (mask & PIPE_MASK_ZS) {
+      case PIPE_MASK_ZS: return NV50_BLIT_MODE_ZS;
+      case PIPE_MASK_Z:  return NV50_BLIT_MODE_PASS;
+      default:
+         return NV50_BLIT_MODE_XS;
+      }
    default:
-      assert(0);
-      return PIPE_FORMAT_NONE;
+      return NV50_BLIT_MODE_PASS;
    }
 }
 
 static void
-nv50_blitctx_get_color_mask_and_fp(struct nv50_blitctx *blit,
-                                   enum pipe_format format, uint8_t mask)
+nv50_blit_select_fp(struct nv50_blitctx *ctx, const struct pipe_blit_info *info)
 {
-   blit->color_mask = 0;
+   struct nv50_blitter *blitter = ctx->nv50->screen->blitter;
 
-   switch (format) {
-   case PIPE_FORMAT_Z24X8_UNORM:
-   case PIPE_FORMAT_Z24_UNORM_S8_UINT:
-      blit->fp_offset = 0xb0;
-      if (mask & PIPE_MASK_Z)
-         blit->color_mask |= 0x0111;
-      if (mask & PIPE_MASK_S)
-         blit->color_mask |= 0x1000;
-      break;
-   case PIPE_FORMAT_S8_UINT_Z24_UNORM:
-      blit->fp_offset = 0x18;
-      if (mask & PIPE_MASK_Z)
-         blit->color_mask |= 0x1110;
-      if (mask & PIPE_MASK_S)
-         blit->color_mask |= 0x0001;
-      break;
-   default:
-      blit->fp_offset = 0;
-      if (mask & (PIPE_MASK_R | PIPE_MASK_Z)) blit->color_mask |= 0x0001;
-      if (mask & (PIPE_MASK_G | PIPE_MASK_S)) blit->color_mask |= 0x0010;
-      if (mask & PIPE_MASK_B) blit->color_mask |= 0x0100;
-      if (mask & PIPE_MASK_A) blit->color_mask |= 0x1000;
-      break;
+   const enum pipe_texture_target ptarg =
+      nv50_blit_reinterpret_pipe_texture_target(info->src.resource->target);
+
+   const unsigned targ = nv50_blit_texture_type(ptarg);
+   const unsigned mode = ctx->mode;
+
+   if (!blitter->fp[targ][mode]) {
+      pipe_mutex_lock(blitter->mutex);
+      if (!blitter->fp[targ][mode])
+         blitter->fp[targ][mode] =
+            nv50_blitter_make_fp(&ctx->nv50->base.pipe, mode, ptarg);
+      pipe_mutex_unlock(blitter->mutex);
    }
+   ctx->fp = blitter->fp[targ][mode];
 }
 
 static void
-nv50_blit_set_dst(struct nv50_context *nv50,
-                  struct pipe_resource *res, unsigned level, unsigned layer)
+nv50_blit_set_dst(struct nv50_blitctx *ctx,
+                  struct pipe_resource *res, unsigned level, unsigned layer,
+                  enum pipe_format format)
 {
+   struct nv50_context *nv50 = ctx->nv50;
    struct pipe_context *pipe = &nv50->base.pipe;
    struct pipe_surface templ;
 
-   if (util_format_is_depth_or_stencil(res->format))
-      templ.format = nv50_blit_zeta_to_colour_format(res->format);
+   if (util_format_is_depth_or_stencil(format))
+      templ.format = nv50_blit_zeta_to_colour_format(format);
    else
-      templ.format = res->format;
+      templ.format = format;
 
    templ.usage = PIPE_USAGE_STREAM;
    templ.u.tex.level = level;
    templ.u.tex.first_layer = templ.u.tex.last_layer = layer;
 
+   if (layer == -1) {
+      templ.u.tex.first_layer = 0;
+      templ.u.tex.last_layer =
+         (res->target == PIPE_TEXTURE_3D ? res->depth0 : res->array_size) - 1;
+   }
+
    nv50->framebuffer.cbufs[0] = nv50_miptree_surface_new(pipe, res, &templ);
    nv50->framebuffer.nr_cbufs = 1;
    nv50->framebuffer.zsbuf = NULL;
@@ -689,45 +712,48 @@ nv50_blit_set_dst(struct nv50_context *nv50,
    nv50->framebuffer.height = nv50->framebuffer.cbufs[0]->height;
 }
 
-static INLINE void
-nv50_blit_fixup_tic_entry(struct pipe_sampler_view *view)
-{
-   struct nv50_tic_entry *ent = nv50_tic_entry(view);
-
-   ent->tic[2] &= ~(1 << 31); /* scaled coordinates, ok with 3d textures ? */
-
-   /* magic: */
-
-   ent->tic[3] = 0x20000000; /* affects quality of near vertical edges in MS8 */
-}
-
 static void
-nv50_blit_set_src(struct nv50_context *nv50,
-                  struct pipe_resource *res, unsigned level, unsigned layer)
+nv50_blit_set_src(struct nv50_blitctx *blit,
+                  struct pipe_resource *res, unsigned level, unsigned layer,
+                  enum pipe_format format, const uint8_t filter)
 {
+   struct nv50_context *nv50 = blit->nv50;
    struct pipe_context *pipe = &nv50->base.pipe;
    struct pipe_sampler_view templ;
+   uint32_t flags;
+   enum pipe_texture_target target;
 
-   templ.format = res->format;
-   templ.u.tex.first_layer = templ.u.tex.last_layer = layer;
+   target = nv50_blit_reinterpret_pipe_texture_target(res->target);
+
+   templ.format = format;
    templ.u.tex.first_level = templ.u.tex.last_level = level;
+   templ.u.tex.first_layer = templ.u.tex.last_layer = layer;
    templ.swizzle_r = PIPE_SWIZZLE_RED;
    templ.swizzle_g = PIPE_SWIZZLE_GREEN;
    templ.swizzle_b = PIPE_SWIZZLE_BLUE;
    templ.swizzle_a = PIPE_SWIZZLE_ALPHA;
 
-   nv50->textures[2][0] = nv50_create_sampler_view(pipe, res, &templ);
-   nv50->textures[2][1] = NULL;
+   if (layer == -1) {
+      templ.u.tex.first_layer = 0;
+      templ.u.tex.last_layer =
+         (res->target == PIPE_TEXTURE_3D ? res->depth0 : res->array_size) - 1;
+   }
 
-   nv50_blit_fixup_tic_entry(nv50->textures[2][0]);
+   flags = res->last_level ? 0 : NV50_TEXVIEW_SCALED_COORDS;
+   if (filter && res->nr_samples == 8)
+      flags |= NV50_TEXVIEW_FILTER_MSAA8;
+
+   nv50->textures[2][0] = nv50_create_texture_view(
+      pipe, res, &templ, flags, target);
+   nv50->textures[2][1] = NULL;
 
    nv50->num_textures[0] = nv50->num_textures[1] = 0;
    nv50->num_textures[2] = 1;
 
-   templ.format = nv50_zs_to_s_format(res->format);
+   templ.format = nv50_zs_to_s_format(format);
    if (templ.format != res->format) {
-      nv50->textures[2][1] = nv50_create_sampler_view(pipe, res, &templ);
-      nv50_blit_fixup_tic_entry(nv50->textures[2][1]);
+      nv50->textures[2][1] = nv50_create_texture_view(
+         pipe, res, &templ, flags, target);
       nv50->num_textures[2] = 2;
    }
 }
@@ -735,7 +761,12 @@ nv50_blit_set_src(struct nv50_context *nv50,
 static void
 nv50_blitctx_prepare_state(struct nv50_blitctx *blit)
 {
-   struct nouveau_pushbuf *push = blit->screen->base.pushbuf;
+   struct nouveau_pushbuf *push = blit->nv50->base.pushbuf;
+
+   if (blit->nv50->cond_query) {
+      BEGIN_NV04(push, NV50_3D(COND_MODE), 1);
+      PUSH_DATA (push, NV50_3D_COND_MODE_ALWAYS);
+   }
 
    /* blend state */
    BEGIN_NV04(push, NV50_3D(COLOR_MASK(0)), 1);
@@ -782,40 +813,45 @@ nv50_blitctx_prepare_state(struct nv50_blitctx *blit)
 }
 
 static void
-nv50_blitctx_pre_blit(struct nv50_blitctx *blit, struct nv50_context *nv50)
+nv50_blitctx_pre_blit(struct nv50_blitctx *ctx)
 {
+   struct nv50_context *nv50 = ctx->nv50;
+   struct nv50_blitter *blitter = nv50->screen->blitter;
    int s;
 
-   blit->saved.fb.width = nv50->framebuffer.width;
-   blit->saved.fb.height = nv50->framebuffer.height;
-   blit->saved.fb.nr_cbufs = nv50->framebuffer.nr_cbufs;
-   blit->saved.fb.cbufs[0] = nv50->framebuffer.cbufs[0];
-   blit->saved.fb.zsbuf = nv50->framebuffer.zsbuf;
+   ctx->saved.fb.width = nv50->framebuffer.width;
+   ctx->saved.fb.height = nv50->framebuffer.height;
+   ctx->saved.fb.nr_cbufs = nv50->framebuffer.nr_cbufs;
+   ctx->saved.fb.cbufs[0] = nv50->framebuffer.cbufs[0];
+   ctx->saved.fb.zsbuf = nv50->framebuffer.zsbuf;
 
-   blit->saved.vp = nv50->vertprog;
-   blit->saved.gp = nv50->gmtyprog;
-   blit->saved.fp = nv50->fragprog;
+   ctx->saved.vp = nv50->vertprog;
+   ctx->saved.gp = nv50->gmtyprog;
+   ctx->saved.fp = nv50->fragprog;
 
-   nv50->vertprog = &blit->vp;
+   nv50->vertprog = &blitter->vp;
    nv50->gmtyprog = NULL;
-   nv50->fragprog = &blit->fp;
+   nv50->fragprog = ctx->fp;
 
    for (s = 0; s < 3; ++s) {
-      blit->saved.num_textures[s] = nv50->num_textures[s];
-      blit->saved.num_samplers[s] = nv50->num_samplers[s];
+      ctx->saved.num_textures[s] = nv50->num_textures[s];
+      ctx->saved.num_samplers[s] = nv50->num_samplers[s];
    }
-   blit->saved.texture[0] = nv50->textures[2][0];
-   blit->saved.texture[1] = nv50->textures[2][1];
-   blit->saved.sampler[0] = nv50->samplers[2][0];
-   blit->saved.sampler[1] = nv50->samplers[2][1];
+   ctx->saved.texture[0] = nv50->textures[2][0];
+   ctx->saved.texture[1] = nv50->textures[2][1];
+   ctx->saved.sampler[0] = nv50->samplers[2][0];
+   ctx->saved.sampler[1] = nv50->samplers[2][1];
 
-   nv50->samplers[2][0] = &blit->sampler[blit->filter];
-   nv50->samplers[2][1] = &blit->sampler[blit->filter];
+   nv50->samplers[2][0] = &blitter->sampler[ctx->filter];
+   nv50->samplers[2][1] = &blitter->sampler[ctx->filter];
 
    nv50->num_samplers[0] = nv50->num_samplers[1] = 0;
    nv50->num_samplers[2] = 2;
 
-   blit->saved.dirty = nv50->dirty;
+   ctx->saved.dirty = nv50->dirty;
+
+   nouveau_bufctx_reset(nv50->bufctx_3d, NV50_BIND_FB);
+   nouveau_bufctx_reset(nv50->bufctx_3d, NV50_BIND_TEXTURES);
 
    nv50->dirty =
       NV50_NEW_FRAMEBUFFER |
@@ -824,8 +860,9 @@ nv50_blitctx_pre_blit(struct nv50_blitctx *blit, struct nv50_context *nv50)
 }
 
 static void
-nv50_blitctx_post_blit(struct nv50_context *nv50, struct nv50_blitctx *blit)
+nv50_blitctx_post_blit(struct nv50_blitctx *blit)
 {
+   struct nv50_context *nv50 = blit->nv50;
    int s;
 
    pipe_surface_reference(&nv50->framebuffer.cbufs[0], NULL);
@@ -852,6 +889,13 @@ nv50_blitctx_post_blit(struct nv50_context *nv50, struct nv50_blitctx *blit)
    nv50->samplers[2][0] = blit->saved.sampler[0];
    nv50->samplers[2][1] = blit->saved.sampler[1];
 
+   if (nv50->cond_query)
+      nv50->base.pipe.render_condition(&nv50->base.pipe, nv50->cond_query,
+                                       nv50->cond_mode);
+
+   nouveau_bufctx_reset(nv50->bufctx_3d, NV50_BIND_FB);
+   nouveau_bufctx_reset(nv50->bufctx_3d, NV50_BIND_TEXTURES);
+
    nv50->dirty = blit->saved.dirty |
       (NV50_NEW_FRAMEBUFFER | NV50_NEW_SCISSOR | NV50_NEW_SAMPLE_MASK |
        NV50_NEW_RASTERIZER | NV50_NEW_ZSA | NV50_NEW_BLEND |
@@ -859,42 +903,40 @@ nv50_blitctx_post_blit(struct nv50_context *nv50, struct nv50_blitctx *blit)
        NV50_NEW_VERTPROG | NV50_NEW_GMTYPROG | NV50_NEW_FRAGPROG);
 }
 
-#if 0
+
 static void
-nv50_resource_resolve(struct pipe_context *pipe,
-                      const struct pipe_resolve_info *info)
+nv50_blit_3d(struct nv50_context *nv50, const struct pipe_blit_info *info)
 {
-   struct nv50_context *nv50 = nv50_context(pipe);
-   struct nv50_screen *screen = nv50->screen;
-   struct nv50_blitctx *blit = screen->blitctx;
+   struct nv50_blitctx *blit = nv50->blit;
    struct nouveau_pushbuf *push = nv50->base.pushbuf;
-   struct pipe_resource *src = info->src.res;
-   struct pipe_resource *dst = info->dst.res;
+   struct pipe_resource *src = info->src.resource;
+   struct pipe_resource *dst = info->dst.resource;
+   int32_t minx, maxx, miny, maxy;
+   int32_t i;
    float x0, x1, y0, y1, z;
+   float dz;
    float x_range, y_range;
 
-   nv50_blitctx_get_color_mask_and_fp(blit, dst->format, info->mask);
+   blit->mode = nv50_blit_select_mode(info);
+   blit->color_mask = nv50_blit_derive_color_mask(info);
+   blit->filter = nv50_blit_get_filter(info);
 
-   blit->filter = util_format_is_depth_or_stencil(dst->format) ? 0 : 1;
+   nv50_blit_select_fp(blit, info);
+   nv50_blitctx_pre_blit(blit);
 
-   nv50_blitctx_pre_blit(blit, nv50);
-
-   nv50_blit_set_dst(nv50, dst, info->dst.level, info->dst.layer);
-   nv50_blit_set_src(nv50, src, 0,               info->src.layer);
+   nv50_blit_set_dst(blit, dst, info->dst.level,  0, info->dst.format);
+   nv50_blit_set_src(blit, src, info->src.level, -1, info->src.format,
+                     blit->filter);
 
    nv50_blitctx_prepare_state(blit);
 
    nv50_state_validate(nv50, ~0, 36);
 
-   x_range =
-      (float)(info->src.x1 - info->src.x0) /
-      (float)(info->dst.x1 - info->dst.x0);
-   y_range =
-      (float)(info->src.y1 - info->src.y0) /
-      (float)(info->dst.y1 - info->dst.y0);
+   x_range = (float)info->src.box.width / (float)info->dst.box.width;
+   y_range = (float)info->src.box.height / (float)info->dst.box.height;
 
-   x0 = (float)info->src.x0 - x_range * (float)info->dst.x0;
-   y0 = (float)info->src.y0 - y_range * (float)info->dst.y0;
+   x0 = (float)info->src.box.x - x_range * (float)info->dst.box.x;
+   y0 = (float)info->src.box.y - y_range * (float)info->dst.box.y;
 
    x1 = x0 + 16384.0f * x_range;
    y1 = y0 + 16384.0f * y_range;
@@ -904,11 +946,21 @@ nv50_resource_resolve(struct pipe_context *pipe,
    y0 *= (float)(1 << nv50_miptree(src)->ms_y);
    y1 *= (float)(1 << nv50_miptree(src)->ms_y);
 
-   z = (float)info->src.layer;
+   if (src->last_level > 0) {
+      /* If there are mip maps, GPU always assumes normalized coordinates. */
+      const unsigned l = info->src.level;
+      const float fh = u_minify(src->width0 << nv50_miptree(src)->ms_x, l);
+      const float fv = u_minify(src->height0 << nv50_miptree(src)->ms_y, l);
+      x0 /= fh;
+      x1 /= fh;
+      y0 /= fv;
+      y1 /= fv;
+   }
 
-   BEGIN_NV04(push, NV50_3D(FP_START_ID), 1);
-   PUSH_DATA (push,
-              blit->fp.code_base + blit->fp_offset);
+   dz = (float)info->src.box.depth / (float)info->dst.box.depth;
+   z = (float)info->src.box.z;
+   if (nv50_miptree(src)->layout_3d)
+      z += 0.5f * dz;
 
    BEGIN_NV04(push, NV50_3D(VIEWPORT_TRANSFORM_EN), 1);
    PUSH_DATA (push, 0);
@@ -919,62 +971,318 @@ nv50_resource_resolve(struct pipe_context *pipe,
     * arranged in a way to yield the desired offset and scale.
     */
 
+   minx = info->dst.box.x;
+   maxx = info->dst.box.x + info->dst.box.width;
+   miny = info->dst.box.y;
+   maxy = info->dst.box.y + info->dst.box.height;
+   if (info->scissor_enable) {
+      minx = MAX2(minx, info->scissor.minx);
+      maxx = MIN2(maxx, info->scissor.maxx);
+      miny = MAX2(miny, info->scissor.miny);
+      maxy = MIN2(maxy, info->scissor.maxy);
+   }
    BEGIN_NV04(push, NV50_3D(SCISSOR_HORIZ(0)), 2);
-   PUSH_DATA (push, (info->dst.x1 << 16) | info->dst.x0);
-   PUSH_DATA (push, (info->dst.y1 << 16) | info->dst.y0);
-
-   BEGIN_NV04(push, NV50_3D(VERTEX_BEGIN_GL), 1);
-   PUSH_DATA (push, NV50_3D_VERTEX_BEGIN_GL_PRIMITIVE_TRIANGLES);
-   BEGIN_NV04(push, NV50_3D(VTX_ATTR_3F_X(1)), 3);
-   PUSH_DATAf(push, x0);
-   PUSH_DATAf(push, y0);
-   PUSH_DATAf(push, z);
-   BEGIN_NV04(push, NV50_3D(VTX_ATTR_2F_X(0)), 2);
-   PUSH_DATAf(push, 0.0f);
-   PUSH_DATAf(push, 0.0f);
-   BEGIN_NV04(push, NV50_3D(VTX_ATTR_3F_X(1)), 3);
-   PUSH_DATAf(push, x1);
-   PUSH_DATAf(push, y0);
-   PUSH_DATAf(push, z);
-   BEGIN_NV04(push, NV50_3D(VTX_ATTR_2F_X(0)), 2);
-   PUSH_DATAf(push, 16384 << nv50_miptree(dst)->ms_x);
-   PUSH_DATAf(push, 0.0f);
-   BEGIN_NV04(push, NV50_3D(VTX_ATTR_3F_X(1)), 3);
-   PUSH_DATAf(push, x0);
-   PUSH_DATAf(push, y1);
-   PUSH_DATAf(push, z);
-   BEGIN_NV04(push, NV50_3D(VTX_ATTR_2F_X(0)), 2);
-   PUSH_DATAf(push, 0.0f);
-   PUSH_DATAf(push, 16384 << nv50_miptree(dst)->ms_y);
-   BEGIN_NV04(push, NV50_3D(VERTEX_END_GL), 1);
-   PUSH_DATA (push, 0);
+   PUSH_DATA (push, (maxx << 16) | minx);
+   PUSH_DATA (push, (maxy << 16) | miny);
+
+   for (i = 0; i < info->dst.box.depth; ++i, z += dz) {
+      if (info->dst.box.z + i) {
+         BEGIN_NV04(push, NV50_3D(LAYER), 1);
+         PUSH_DATA (push, info->dst.box.z + i);
+      }
+      PUSH_SPACE(push, 32);
+      BEGIN_NV04(push, NV50_3D(VERTEX_BEGIN_GL), 1);
+      PUSH_DATA (push, NV50_3D_VERTEX_BEGIN_GL_PRIMITIVE_TRIANGLES);
+      BEGIN_NV04(push, NV50_3D(VTX_ATTR_3F_X(1)), 3);
+      PUSH_DATAf(push, x0);
+      PUSH_DATAf(push, y0);
+      PUSH_DATAf(push, z);
+      BEGIN_NV04(push, NV50_3D(VTX_ATTR_2F_X(0)), 2);
+      PUSH_DATAf(push, 0.0f);
+      PUSH_DATAf(push, 0.0f);
+      BEGIN_NV04(push, NV50_3D(VTX_ATTR_3F_X(1)), 3);
+      PUSH_DATAf(push, x1);
+      PUSH_DATAf(push, y0);
+      PUSH_DATAf(push, z);
+      BEGIN_NV04(push, NV50_3D(VTX_ATTR_2F_X(0)), 2);
+      PUSH_DATAf(push, 16384 << nv50_miptree(dst)->ms_x);
+      PUSH_DATAf(push, 0.0f);
+      BEGIN_NV04(push, NV50_3D(VTX_ATTR_3F_X(1)), 3);
+      PUSH_DATAf(push, x0);
+      PUSH_DATAf(push, y1);
+      PUSH_DATAf(push, z);
+      BEGIN_NV04(push, NV50_3D(VTX_ATTR_2F_X(0)), 2);
+      PUSH_DATAf(push, 0.0f);
+      PUSH_DATAf(push, 16384 << nv50_miptree(dst)->ms_y);
+      BEGIN_NV04(push, NV50_3D(VERTEX_END_GL), 1);
+      PUSH_DATA (push, 0);
+   }
+   if (info->dst.box.z + info->dst.box.depth - 1) {
+      BEGIN_NV04(push, NV50_3D(LAYER), 1);
+      PUSH_DATA (push, 0);
+   }
 
    /* re-enable normally constant state */
 
    BEGIN_NV04(push, NV50_3D(VIEWPORT_TRANSFORM_EN), 1);
    PUSH_DATA (push, 1);
 
-   nv50_blitctx_post_blit(nv50, blit);
+   nv50_blitctx_post_blit(blit);
+}
+
+static void
+nv50_blit_eng2d(struct nv50_context *nv50, const struct pipe_blit_info *info)
+{
+   struct nouveau_pushbuf *push = nv50->base.pushbuf;
+   struct nv50_miptree *dst = nv50_miptree(info->dst.resource);
+   struct nv50_miptree *src = nv50_miptree(info->src.resource);
+   const int32_t srcx_adj = info->src.box.width < 0 ? -1 : 0;
+   const int32_t srcy_adj = info->src.box.height < 0 ? -1 : 0;
+   const int32_t dz = info->dst.box.z;
+   const int32_t sz = info->src.box.z;
+   uint32_t dstw, dsth;
+   int32_t dstx, dsty;
+   int64_t srcx, srcy;
+   int64_t du_dx, dv_dy;
+   int i;
+   uint32_t mode;
+   const uint32_t mask = nv50_blit_eng2d_get_mask(info);
+
+   mode = nv50_blit_get_filter(info) ?
+      NV50_2D_BLIT_CONTROL_FILTER_BILINEAR :
+      NV50_2D_BLIT_CONTROL_FILTER_POINT_SAMPLE;
+   mode |= (src->base.base.nr_samples > dst->base.base.nr_samples) ?
+      NV50_2D_BLIT_CONTROL_ORIGIN_CORNER : NV50_2D_BLIT_CONTROL_ORIGIN_CENTER;
+
+   du_dx = ((int64_t)info->src.box.width << 32) / info->dst.box.width;
+   dv_dy = ((int64_t)info->src.box.height << 32) / info->dst.box.height;
+
+   nv50_2d_texture_set(push, 1, dst, info->dst.level, dz, info->dst.format);
+   nv50_2d_texture_set(push, 0, src, info->src.level, sz, info->src.format);
+
+   if (info->scissor_enable) {
+      BEGIN_NV04(push, NV50_2D(CLIP_X), 5);
+      PUSH_DATA (push, info->scissor.minx << dst->ms_x);
+      PUSH_DATA (push, info->scissor.miny << dst->ms_y);
+      PUSH_DATA (push, (info->scissor.maxx - info->scissor.minx) << dst->ms_x);
+      PUSH_DATA (push, (info->scissor.maxy - info->scissor.miny) << dst->ms_y);
+      PUSH_DATA (push, 1); /* enable */
+   }
+
+   if (mask != 0xffffffff) {
+      BEGIN_NV04(push, NV50_2D(PATTERN_COLOR_FORMAT), 1);
+      PUSH_DATA (push, NV50_2D_PATTERN_COLOR_FORMAT_32BPP);
+      BEGIN_NV04(push, NV50_2D(PATTERN_COLOR(0)), 4);
+      PUSH_DATA (push, 0x00000000);
+      PUSH_DATA (push, mask);
+      PUSH_DATA (push, 0xffffffff);
+      PUSH_DATA (push, 0xffffffff);
+      BEGIN_NV04(push, NV50_2D(OPERATION), 1);
+      PUSH_DATA (push, NV50_2D_OPERATION_SRCCOPY_PREMULT);
+   }
+
+   if (src->ms_x > dst->ms_x || src->ms_y > dst->ms_y) {
+      /* ms_x is always >= ms_y */
+      du_dx <<= src->ms_x - dst->ms_x;
+      dv_dy <<= src->ms_y - dst->ms_y;
+   } else {
+      du_dx >>= dst->ms_x - src->ms_x;
+      dv_dy >>= dst->ms_y - src->ms_y;
+   }
+
+   srcx = (int64_t)(info->src.box.x + srcx_adj) << (src->ms_x + 32);
+   srcy = (int64_t)(info->src.box.y + srcy_adj) << (src->ms_y + 32);
+
+   if (src->base.base.nr_samples > dst->base.base.nr_samples) {
+      /* center src coorinates for proper MS resolve filtering */
+      srcx += (int64_t)src->ms_x << 32;
+      srcy += (int64_t)src->ms_y << 32;
+   }
+
+   dstx = info->dst.box.x << dst->ms_x;
+   dsty = info->dst.box.y << dst->ms_y;
+
+   dstw = info->dst.box.width << dst->ms_x;
+   dsth = info->dst.box.height << dst->ms_y;
+
+   if (dstx < 0) {
+      dstw += dstx;
+      srcx -= du_dx * dstx;
+      dstx = 0;
+   }
+   if (dsty < 0) {
+      dsth += dsty;
+      srcy -= dv_dy * dsty;
+      dsty = 0;
+   }
+
+   BEGIN_NV04(push, NV50_2D(BLIT_CONTROL), 1);
+   PUSH_DATA (push, mode);
+   BEGIN_NV04(push, NV50_2D(BLIT_DST_X), 4);
+   PUSH_DATA (push, dstx);
+   PUSH_DATA (push, dsty);
+   PUSH_DATA (push, dstw);
+   PUSH_DATA (push, dsth);
+   BEGIN_NV04(push, NV50_2D(BLIT_DU_DX_FRACT), 4);
+   PUSH_DATA (push, du_dx);
+   PUSH_DATA (push, du_dx >> 32);
+   PUSH_DATA (push, dv_dy);
+   PUSH_DATA (push, dv_dy >> 32);
+
+   BCTX_REFN(nv50->bufctx, 2D, &dst->base, WR);
+   BCTX_REFN(nv50->bufctx, 2D, &src->base, RD);
+   nouveau_pushbuf_bufctx(nv50->base.pushbuf, nv50->bufctx);
+   if (nouveau_pushbuf_validate(nv50->base.pushbuf))
+      return;
+
+   for (i = 0; i < info->dst.box.depth; ++i) {
+      if (i > 0) {
+         /* no scaling in z-direction possible for eng2d blits */
+         if (dst->layout_3d) {
+            BEGIN_NV04(push, NV50_2D(DST_LAYER), 1);
+            PUSH_DATA (push, info->dst.box.z + i);
+         } else {
+            const unsigned z = info->dst.box.z + i;
+            BEGIN_NV04(push, NV50_2D(DST_ADDRESS_HIGH), 2);
+            PUSH_DATAh(push, dst->base.address + z * dst->layer_stride);
+            PUSH_DATA (push, dst->base.address + z * dst->layer_stride);
+         }
+         if (src->layout_3d) {
+            /* not possible because of depth tiling */
+            assert(0);
+         } else {
+            const unsigned z = info->src.box.z + i;
+            BEGIN_NV04(push, NV50_2D(SRC_ADDRESS_HIGH), 2);
+            PUSH_DATAh(push, src->base.address + z * src->layer_stride);
+            PUSH_DATA (push, src->base.address + z * src->layer_stride);
+         }
+         BEGIN_NV04(push, NV50_2D(BLIT_SRC_Y_INT), 1); /* trigger */
+         PUSH_DATA (push, srcy >> 32);
+      } else {
+         BEGIN_NV04(push, NV50_2D(BLIT_SRC_X_FRACT), 4);
+         PUSH_DATA (push, srcx);
+         PUSH_DATA (push, srcx >> 32);
+         PUSH_DATA (push, srcy);
+         PUSH_DATA (push, srcy >> 32);
+      }
+   }
+   nv50_bufctx_fence(nv50->bufctx, FALSE);
+
+   nouveau_bufctx_reset(nv50->bufctx, NV50_BIND_2D);
+
+   if (info->scissor_enable) {
+      BEGIN_NV04(push, NV50_2D(CLIP_ENABLE), 1);
+      PUSH_DATA (push, 0);
+   }
+   if (mask != 0xffffffff) {
+      BEGIN_NV04(push, NV50_2D(OPERATION), 1);
+      PUSH_DATA (push, NV50_2D_OPERATION_SRCCOPY);
+   }
+}
+
+static void
+nv50_blit(struct pipe_context *pipe, const struct pipe_blit_info *info)
+{
+   struct nv50_context *nv50 = nv50_context(pipe);
+   boolean eng3d = FALSE;
+
+   if (util_format_is_depth_or_stencil(info->dst.resource->format)) {
+      if (!(info->mask & PIPE_MASK_ZS))
+         return;
+      if (info->dst.resource->format == PIPE_FORMAT_Z32_FLOAT ||
+          info->dst.resource->format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT)
+         eng3d = TRUE;
+      if (info->filter != PIPE_TEX_FILTER_NEAREST)
+         eng3d = TRUE;
+   } else {
+      if (!(info->mask & PIPE_MASK_RGBA))
+         return;
+      if (info->mask != PIPE_MASK_RGBA)
+         eng3d = TRUE;
+   }
+
+   if (nv50_miptree(info->src.resource)->layout_3d) {
+      eng3d = TRUE;
+   } else
+   if (info->src.box.depth != info->dst.box.depth) {
+      eng3d = TRUE;
+      debug_printf("blit: cannot filter array or cube textures in z direction");
+   }
+
+   if (!eng3d && info->dst.format != info->src.format)
+      if (!nv50_2d_format_faithful(info->dst.format) ||
+          !nv50_2d_format_faithful(info->src.format))
+         eng3d = TRUE;
+
+   if (info->src.resource->nr_samples == 8 &&
+       info->dst.resource->nr_samples <= 1)
+      eng3d = TRUE;
+
+   /* FIXME: can't make this work with eng2d anymore */
+   if (info->src.resource->nr_samples > 1 ||
+       info->dst.resource->nr_samples > 1)
+      eng3d = TRUE;
+
+   /* FIXME: find correct src coordinate adjustments */
+   if ((info->src.box.width !=  info->dst.box.width &&
+        info->src.box.width != -info->dst.box.width) ||
+       (info->src.box.height !=  info->dst.box.height &&
+        info->src.box.height != -info->dst.box.height))
+      eng3d = TRUE;
+
+   if (!eng3d)
+      nv50_blit_eng2d(nv50, info);
+   else
+      nv50_blit_3d(nv50, info);
 }
-#endif
 
 boolean
-nv50_blitctx_create(struct nv50_screen *screen)
+nv50_blitter_create(struct nv50_screen *screen)
 {
-   screen->blitctx = CALLOC_STRUCT(nv50_blitctx);
-   if (!screen->blitctx) {
-      NOUVEAU_ERR("failed to allocate blit context\n");
+   screen->blitter = CALLOC_STRUCT(nv50_blitter);
+   if (!screen->blitter) {
+      NOUVEAU_ERR("failed to allocate blitter struct\n");
       return FALSE;
    }
 
-   screen->blitctx->screen = screen;
+   pipe_mutex_init(screen->blitter->mutex);
+
+   nv50_blitter_make_vp(screen->blitter);
+   nv50_blitter_make_sampler(screen->blitter);
 
-   nv50_blitctx_make_vp(screen->blitctx);
-   nv50_blitctx_make_fp(screen->blitctx);
+   return TRUE;
+}
 
-   nv50_blitctx_make_sampler(screen->blitctx);
+void
+nv50_blitter_destroy(struct nv50_screen *screen)
+{
+   struct nv50_blitter *blitter = screen->blitter;
+   unsigned i, m;
+
+   for (i = 0; i < NV50_BLIT_MAX_TEXTURE_TYPES; ++i) {
+      for (m = 0; m < NV50_BLIT_MODES; ++m) {
+         struct nv50_program *prog = blitter->fp[i][m];
+         if (prog) {
+            nv50_program_destroy(NULL, prog);
+            FREE((void *)prog->pipe.tokens);
+            FREE(prog);
+         }
+      }
+   }
+
+   FREE(blitter);
+}
+
+boolean
+nv50_blitctx_create(struct nv50_context *nv50)
+{
+   nv50->blit = CALLOC_STRUCT(nv50_blitctx);
+   if (!nv50->blit) {
+      NOUVEAU_ERR("failed to allocate blit context\n");
+      return FALSE;
+   }
 
-   screen->blitctx->color_mask = 0x1111;
+   nv50->blit->nv50 = nv50;
 
    return TRUE;
 }
@@ -985,6 +1293,7 @@ nv50_init_surface_functions(struct nv50_context *nv50)
    struct pipe_context *pipe = &nv50->base.pipe;
 
    pipe->resource_copy_region = nv50_resource_copy_region;
+   pipe->blit = nv50_blit;
    pipe->clear_render_target = nv50_clear_render_target;
    pipe->clear_depth_stencil = nv50_clear_depth_stencil;
 }
diff --git a/src/gallium/drivers/nv50/nv50_tex.c b/src/gallium/drivers/nv50/nv50_tex.c
index 4a769d5b810..992ac4b760c 100644
--- a/src/gallium/drivers/nv50/nv50_tex.c
+++ b/src/gallium/drivers/nv50/nv50_tex.c
@@ -71,9 +71,24 @@ nv50_init_tic_entry_linear(uint32_t *tic, struct pipe_resource *res)
 
 struct pipe_sampler_view *
 nv50_create_sampler_view(struct pipe_context *pipe,
-                         struct pipe_resource *texture,
+                         struct pipe_resource *res,
                          const struct pipe_sampler_view *templ)
 {
+   uint32_t flags = 0;
+
+   if (res->target == PIPE_TEXTURE_RECT)
+      flags |= NV50_TEXVIEW_SCALED_COORDS;
+
+   return nv50_create_texture_view(pipe, res, templ, flags, res->target);
+}
+
+struct pipe_sampler_view *
+nv50_create_texture_view(struct pipe_context *pipe,
+                         struct pipe_resource *texture,
+                         const struct pipe_sampler_view *templ,
+                         uint32_t flags,
+                         enum pipe_texture_target target)
+{
    const struct util_format_description *desc;
    uint64_t addr;
    uint32_t *tic;
@@ -139,14 +154,14 @@ nv50_create_sampler_view(struct pipe_context *pipe,
       return &view->pipe;
    }
 
-   if (mt->base.base.target != PIPE_TEXTURE_RECT)
+   if (!(flags & NV50_TEXVIEW_SCALED_COORDS))
       tic[2] |= NV50_TIC_2_NORMALIZED_COORDS;
 
    tic[2] |=
       ((mt->level[0].tile_mode & 0x0f0) << (22 - 4)) |
       ((mt->level[0].tile_mode & 0xf00) << (25 - 8));
 
-   switch (mt->base.base.target) {
+   switch (target) {
    case PIPE_TEXTURE_1D:
       tic[2] |= NV50_TIC_2_TARGET_1D;
       break;
@@ -181,18 +196,22 @@ nv50_create_sampler_view(struct pipe_context *pipe,
       return FALSE;
    }
 
-   tic[3] = 0x00300000;
+   tic[3] = (flags & NV50_TEXVIEW_FILTER_MSAA8) ? 0x20000000 : 0x00300000;
 
    tic[4] = (1 << 31) | (mt->base.base.width0 << mt->ms_x);
 
    tic[5] = (mt->base.base.height0 << mt->ms_y) & 0xffff;
    tic[5] |= depth << 16;
-   tic[5] |= mt->base.base.last_level << 28;
+   tic[5] |= mt->base.base.last_level << NV50_TIC_5_LAST_LEVEL__SHIFT;
 
    tic[6] = (mt->ms_x > 1) ? 0x88000000 : 0x03000000; /* sampling points */
 
    tic[7] = (view->pipe.u.tex.last_level << 4) | view->pipe.u.tex.first_level;
 
+   if (unlikely(!(tic[2] & NV50_TIC_2_NORMALIZED_COORDS)))
+      if (mt->base.base.last_level)
+         tic[5] &= ~NV50_TIC_5_LAST_LEVEL__MASK;
+
    return &view->pipe;
 }
 
diff --git a/src/gallium/drivers/nvc0/nvc0_context.c b/src/gallium/drivers/nvc0/nvc0_context.c
index b021f2e99dc..ee131d29bbf 100644
--- a/src/gallium/drivers/nvc0/nvc0_context.c
+++ b/src/gallium/drivers/nvc0/nvc0_context.c
@@ -129,6 +129,9 @@ nvc0_create(struct pipe_screen *pscreen, void *priv)
       return NULL;
    pipe = &nvc0->base.pipe;
 
+   if (!nvc0_blitctx_create(nvc0))
+      goto out_err;
+
    nvc0->base.pushbuf = screen->base.pushbuf;
 
    ret = nouveau_bufctx_new(screen->base.client, NVC0_BIND_COUNT,
@@ -199,6 +202,8 @@ out_err:
          nouveau_bufctx_del(&nvc0->bufctx_3d);
       if (nvc0->bufctx)
          nouveau_bufctx_del(&nvc0->bufctx);
+      if (nvc0->blit)
+         FREE(nvc0->blit);
       FREE(nvc0);
    }
    return NULL;
diff --git a/src/gallium/drivers/nvc0/nvc0_context.h b/src/gallium/drivers/nvc0/nvc0_context.h
index ef792227b34..8e2aaff08a7 100644
--- a/src/gallium/drivers/nvc0/nvc0_context.h
+++ b/src/gallium/drivers/nvc0/nvc0_context.h
@@ -71,6 +71,11 @@
 #define NVC0_BIND_M2MF          0
 #define NVC0_BIND_FENCE         1
 
+
+struct nvc0_blitctx;
+
+boolean nvc0_blitctx_create(struct nvc0_context *);
+
 struct nvc0_context {
    struct nouveau_context base;
 
@@ -159,6 +164,11 @@ struct nvc0_context {
    struct pipe_stream_output_target *tfbbuf[4];
    unsigned num_tfbbufs;
 
+   struct pipe_query *cond_query;
+   uint cond_mode;
+
+   struct nvc0_blitctx *blit;
+
 #ifdef NVC0_WITH_DRAW_MODULE
    struct draw_context *draw;
 #endif
@@ -241,6 +251,12 @@ void nvc0_validate_samplers(struct nvc0_context *);
 void nve4_set_tex_handles(struct nvc0_context *);
 
 struct pipe_sampler_view *
+nvc0_create_texture_view(struct pipe_context *,
+                         struct pipe_resource *,
+                         const struct pipe_sampler_view *,
+                         uint32_t flags,
+                         enum pipe_texture_target);
+struct pipe_sampler_view *
 nvc0_create_sampler_view(struct pipe_context *,
                          struct pipe_resource *,
                          const struct pipe_sampler_view *);
diff --git a/src/gallium/drivers/nvc0/nvc0_query.c b/src/gallium/drivers/nvc0/nvc0_query.c
index 8dfda378374..d329148de62 100644
--- a/src/gallium/drivers/nvc0/nvc0_query.c
+++ b/src/gallium/drivers/nvc0/nvc0_query.c
@@ -436,6 +436,9 @@ nvc0_render_condition(struct pipe_context *pipe,
       mode != PIPE_RENDER_COND_NO_WAIT &&
       mode != PIPE_RENDER_COND_BY_REGION_NO_WAIT;
 
+   nvc0->cond_query = pq;
+   nvc0->cond_mode = mode;
+
    if (!pq) {
       PUSH_SPACE(push, 1);
       IMMED_NVC0(push, NVC0_3D(COND_MODE), NVC0_3D_COND_MODE_ALWAYS);
diff --git a/src/gallium/drivers/nvc0/nvc0_screen.c b/src/gallium/drivers/nvc0/nvc0_screen.c
index e26313548e3..99bd4e731b2 100644
--- a/src/gallium/drivers/nvc0/nvc0_screen.c
+++ b/src/gallium/drivers/nvc0/nvc0_screen.c
@@ -257,7 +257,8 @@ nvc0_screen_destroy(struct pipe_screen *pscreen)
    if (screen->base.pushbuf)
       screen->base.pushbuf->user_priv = NULL;
 
-   FREE(screen->blitctx);
+   if (screen->blitter)
+      nvc0_blitter_destroy(screen);
 
    nouveau_bo_ref(NULL, &screen->text);
    nouveau_bo_ref(NULL, &screen->uniform_bo);
@@ -746,7 +747,7 @@ nvc0_screen_create(struct nouveau_device *dev)
    mm_config.nvc0.memtype = 0xfe0;
    screen->mm_VRAM_fe0 = nouveau_mm_create(dev, NOUVEAU_BO_VRAM, &mm_config);
 
-   if (!nvc0_blitctx_create(screen))
+   if (!nvc0_blitter_create(screen))
       goto fail;
 
    nouveau_fence_new(&screen->base, &screen->base.fence.current, FALSE);
diff --git a/src/gallium/drivers/nvc0/nvc0_screen.h b/src/gallium/drivers/nvc0/nvc0_screen.h
index 69cddf2c9b0..1387d67a273 100644
--- a/src/gallium/drivers/nvc0/nvc0_screen.h
+++ b/src/gallium/drivers/nvc0/nvc0_screen.h
@@ -19,7 +19,7 @@
 
 struct nvc0_context;
 
-struct nvc0_blitctx;
+struct nvc0_blitter;
 
 struct nvc0_screen {
    struct nouveau_screen base;
@@ -39,7 +39,7 @@ struct nvc0_screen {
    struct nouveau_heap *text_heap;
    struct nouveau_heap *lib_code; /* allocated from text_heap */
 
-   struct nvc0_blitctx *blitctx;
+   struct nvc0_blitter *blitter;
 
    struct {
       void **entries;
@@ -72,7 +72,8 @@ nvc0_screen(struct pipe_screen *screen)
    return (struct nvc0_screen *)screen;
 }
 
-boolean nvc0_blitctx_create(struct nvc0_screen *);
+boolean nvc0_blitter_create(struct nvc0_screen *);
+void nvc0_blitter_destroy(struct nvc0_screen *);
 
 void nvc0_screen_make_buffers_resident(struct nvc0_screen *);
 
diff --git a/src/gallium/drivers/nvc0/nvc0_surface.c b/src/gallium/drivers/nvc0/nvc0_surface.c
index d95124dec10..dcf008537ec 100644
--- a/src/gallium/drivers/nvc0/nvc0_surface.c
+++ b/src/gallium/drivers/nvc0/nvc0_surface.c
@@ -29,11 +29,14 @@
 #include "util/u_format.h"
 #include "util/u_surface.h"
 
+#include "os/os_thread.h"
+
 #include "nvc0_context.h"
 #include "nvc0_resource.h"
 
 #include "nv50/nv50_defs.xml.h"
 #include "nv50/nv50_texture.xml.h"
+#include "nv50/nv50_blit.h"
 
 #define NVC0_ENG2D_SUPPORTED_FORMATS 0xff9ccfe1cce3ccc9ULL
 
@@ -75,7 +78,8 @@ nvc0_2d_format(enum pipe_format format)
 
 static int
 nvc0_2d_texture_set(struct nouveau_pushbuf *push, boolean dst,
-                    struct nv50_miptree *mt, unsigned level, unsigned layer)
+                    struct nv50_miptree *mt, unsigned level, unsigned layer,
+                    enum pipe_format pformat)
 {
    struct nouveau_bo *bo = mt->base.bo;
    uint32_t width, height, depth;
@@ -83,10 +87,10 @@ nvc0_2d_texture_set(struct nouveau_pushbuf *push, boolean dst,
    uint32_t mthd = dst ? NVC0_2D_DST_FORMAT : NVC0_2D_SRC_FORMAT;
    uint32_t offset = mt->level[level].offset;
 
-   format = nvc0_2d_format(mt->base.base.format);
+   format = nvc0_2d_format(pformat);
    if (!format) {
       NOUVEAU_ERR("invalid/unsupported surface format: %s\n",
-                  util_format_name(mt->base.base.format));
+                  util_format_name(pformat));
       return 1;
    }
 
@@ -106,7 +110,7 @@ nvc0_2d_texture_set(struct nouveau_pushbuf *push, boolean dst,
       layer = 0;
    }
 
-   if (nouveau_bo_memtype(bo)) {
+   if (!nouveau_bo_memtype(bo)) {
       BEGIN_NVC0(push, SUBC_2D(mthd), 2);
       PUSH_DATA (push, format);
       PUSH_DATA (push, 1);
@@ -150,43 +154,33 @@ nvc0_2d_texture_do_copy(struct nouveau_pushbuf *push,
                         unsigned sx, unsigned sy, unsigned sz,
                         unsigned w, unsigned h)
 {
-   static const uint32_t duvdxy[5] =
-   {
-      0x40000000, 0x80000000, 0x00000001, 0x00000002, 0x00000004
-   };
-
+   const enum pipe_format dfmt = dst->base.base.format;
+   const enum pipe_format sfmt = src->base.base.format;
    int ret;
-   uint32_t ctrl = 0x00;
 
    ret = PUSH_SPACE(push, 2 * 16 + 32);
    if (ret)
       return ret;
 
-   ret = nvc0_2d_texture_set(push, TRUE, dst, dst_level, dz);
+   ret = nvc0_2d_texture_set(push, TRUE, dst, dst_level, dz, dfmt);
    if (ret)
       return ret;
 
-   ret = nvc0_2d_texture_set(push, FALSE, src, src_level, sz);
+   ret = nvc0_2d_texture_set(push, FALSE, src, src_level, sz, sfmt);
    if (ret)
       return ret;
 
-   /* NOTE: 2D engine doesn't work for MS8 */
-   if (src->ms_x)
-      ctrl = 0x11;
-
-   /* 0/1 = CENTER/CORNER, 00/10 = POINT/BILINEAR */
-   BEGIN_NVC0(push, NVC0_2D(BLIT_CONTROL), 1);
-   PUSH_DATA (push, ctrl);
+   IMMED_NVC0(push, NVC0_2D(BLIT_CONTROL), 0x00);
    BEGIN_NVC0(push, NVC0_2D(BLIT_DST_X), 4);
    PUSH_DATA (push, dx << dst->ms_x);
    PUSH_DATA (push, dy << dst->ms_y);
    PUSH_DATA (push, w << dst->ms_x);
    PUSH_DATA (push, h << dst->ms_y);
    BEGIN_NVC0(push, NVC0_2D(BLIT_DU_DX_FRACT), 4);
-   PUSH_DATA (push, duvdxy[2 + ((int)src->ms_x - (int)dst->ms_x)] & 0xf0000000);
-   PUSH_DATA (push, duvdxy[2 + ((int)src->ms_x - (int)dst->ms_x)] & 0x0000000f);
-   PUSH_DATA (push, duvdxy[2 + ((int)src->ms_y - (int)dst->ms_y)] & 0xf0000000);
-   PUSH_DATA (push, duvdxy[2 + ((int)src->ms_y - (int)dst->ms_y)] & 0x0000000f);
+   PUSH_DATA (push, 0);
+   PUSH_DATA (push, 1);
+   PUSH_DATA (push, 0);
+   PUSH_DATA (push, 1);
    BEGIN_NVC0(push, NVC0_2D(BLIT_SRC_X_FRACT), 4);
    PUSH_DATA (push, 0);
    PUSH_DATA (push, sx << src->ms_x);
@@ -444,9 +438,29 @@ nvc0_clear(struct pipe_context *pipe, unsigned buffers,
 }
 
 
-struct nvc0_blitctx
+/* =============================== BLIT CODE ===================================
+ */
+
+struct nvc0_blitter
 {
+   struct nvc0_program *fp[NV50_BLIT_MAX_TEXTURE_TYPES][NV50_BLIT_MODES];
+   struct nvc0_program vp;
+
+   struct nv50_tsc_entry sampler[2]; /* nearest, bilinear */
+
+   pipe_mutex mutex;
+
    struct nvc0_screen *screen;
+};
+
+struct nvc0_blitctx
+{
+   struct nvc0_context *nvc0;
+   struct nvc0_program *fp;
+   uint8_t mode;
+   uint16_t color_mask;
+   uint8_t filter;
+   enum pipe_texture_target target;
    struct {
       struct pipe_framebuffer_state fb;
       struct nvc0_program *vp;
@@ -458,33 +472,41 @@ struct nvc0_blitctx
       unsigned num_samplers[5];
       struct pipe_sampler_view *texture[2];
       struct nv50_tsc_entry *sampler[2];
-      unsigned dirty;
+      uint32_t dirty;
    } saved;
-   struct nvc0_program vp;
-   struct nvc0_program fp;
-   struct nv50_tsc_entry sampler[2]; /* nearest, bilinear */
-   uint32_t fp_offset;
-   uint16_t color_mask;
-   uint8_t filter;
 };
 
 static void
-nvc0_blitctx_make_vp(struct nvc0_blitctx *blit)
+nvc0_blitter_make_vp(struct nvc0_blitter *blit)
 {
-   static const uint32_t code[] =
+   static const uint32_t code_nvc0[] =
+   {
+      0xfff01c66, 0x06000080, /* vfetch b128 { $r0 $r1 $r2 $r3 } a[0x80] */
+      0xfff11c26, 0x06000090, /* vfetch b96 { $r4 $r5 $r6 } a[0x90]*/
+      0x03f01c66, 0x0a7e0070, /* export b128 o[0x70] { $r0 $r1 $r2 $r3 } */
+      0x13f01c26, 0x0a7e0080, /* export b96 o[0x80] { $r4 $r5 $r6 } */
+      0x00001de7, 0x80000000, /* exit */
+   };
+   static const uint32_t code_nve4[] =
    {
+      0x00000007, 0x20000000, /* sched */
       0xfff01c66, 0x06000080, /* vfetch b128 { $r0 $r1 $r2 $r3 } a[0x80] */
-      0xfff11c26, 0x06000090, /* vfetch b64 { $r4 $r5 } a[0x90]*/
+      0xfff11c46, 0x06000090, /* vfetch b96 { $r4 $r5 $r6 } a[0x90]*/
       0x03f01c66, 0x0a7e0070, /* export b128 o[0x70] { $r0 $r1 $r2 $r3 } */
-      0x13f01c26, 0x0a7e0080, /* export b64 o[0x80] { $r4 $r5 } */
+      0x13f01c46, 0x0a7e0080, /* export b96 o[0x80] { $r4 $r5 $r6 } */
       0x00001de7, 0x80000000, /* exit */
    };
 
    blit->vp.type = PIPE_SHADER_VERTEX;
    blit->vp.translated = TRUE;
-   blit->vp.code = (uint32_t *)code; /* no relocations -> no modification */
-   blit->vp.code_size = sizeof(code);
-   blit->vp.max_gpr = 6;
+   if (blit->screen->base.class_3d >= NVE4_3D_CLASS) {
+      blit->vp.code = (uint32_t *)code_nve4; /* const_cast */
+      blit->vp.code_size = sizeof(code_nve4);
+   } else {
+      blit->vp.code = (uint32_t *)code_nvc0; /* const_cast */
+      blit->vp.code_size = sizeof(code_nvc0);
+   }
+   blit->vp.max_gpr = 7;
    blit->vp.vp.edgeflag = PIPE_MAX_ATTRIBS;
 
    blit->vp.hdr[0]  = 0x00020461; /* vertprog magic */
@@ -494,172 +516,7 @@ nvc0_blitctx_make_vp(struct nvc0_blitctx *blit)
 }
 
 static void
-nvc0_blitctx_make_fp(struct nvc0_blitctx *blit)
-{
-   static const uint32_t code_nvc0[] = /* use nvc0dis */
-   {
-      /* 2 coords RGBA in, RGBA out, also for Z32_FLOAT(_S8X24_UINT)
-       * NOTE:
-       * NVC0 doesn't like tex 3d on non-3d textures, but there should
-       * only be 2d and 2d-array MS resources anyway.
-       */
-      0xfff01c00, 0xc07e0080,
-      0xfff05c00, 0xc07e0084,
-      0x00001e86, 0x8013c000,
-      0x00001de7, 0x80000000,
-      /* size: 0x70 + padding  */
-      0, 0, 0, 0,
-
-      /* 2 coords ZS in, S encoded in R, Z encoded in GBA (8_UNORM)
-       * Setup float outputs in a way that conversion to UNORM yields the
-       * desired byte value.
-       */
-      /* NOTE: need to repeat header */
-      0x00021462, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
-      0x80000000, 0x0000000f, 0x00000000, 0x00000000, 0x00000000,
-      0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
-      0x00000000, 0x00000000, 0x00000000, 0x0000000f, 0x00000000,
-      0xfff09c00, 0xc07e0080,
-      0xfff0dc00, 0xc07e0084,
-      0x00201e86, 0x80104000,
-      0x00205f06, 0x80104101,
-      0xfc009c02, 0x312dffff,
-      0x05001c88,
-      0x09009e88,
-      0x04001c02, 0x30ee0202,
-      0xfc205c02, 0x38000003,
-      0x0020dc02, 0x3803fc00,
-      0x00209c02, 0x380003fc,
-      0x05005c88,
-      0x0d00dc88,
-      0x09209e04, 0x18000000,
-      0x04105c02, 0x30ee0202,
-      0x0430dc02, 0x30ce0202,
-      0x04209c02, 0x30de0202,
-      0x00001de7, 0x80000000,
-      /* size: 0xd0 + padding */
-      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-
-      /* 2 coords ZS in, Z encoded in RGB, S encoded in A (U8_UNORM) */
-      0x00021462, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
-      0x80000000, 0x0000000f, 0x00000000, 0x00000000, 0x00000000,
-      0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
-      0x00000000, 0x00000000, 0x00000000, 0x0000000f, 0x00000000,
-      0xfff09c00, 0xc07e0080,
-      0xfff0dc00, 0xc07e0084,
-      0x00201e86, 0x80104000,
-      0x00205f06, 0x80104101,
-      0xfc009c02, 0x312dffff,
-      0x0500dc88,
-      0x09009e88,
-      0x0430dc02, 0x30ee0202,
-      0xfc201c02, 0x38000003,
-      0x00205c02, 0x380003fc,
-      0x00209c02, 0x3803fc00,
-      0x01001c88,
-      0x05005c88,
-      0x09209e04, 0x18000000,
-      0x04001c02, 0x30ee0202,
-      0x04105c02, 0x30de0202,
-      0x04209c02, 0x30ce0202,
-      0x00001de7, 0x80000000,
-   };
-   static const uint32_t code_nve4[] = /* use nvc0dis */
-   {
-      /* 2 coords RGBA in, RGBA out, also for Z32_FLOAT(_S8X24_UINT)
-       * NOTE:
-       * NVC0 doesn't like tex 3d on non-3d textures, but there should
-       * only be 2d and 2d-array MS resources anyway.
-       */
-      0x2202e237, 0x200002ec,
-      0xfff01c00, 0xc07e0080,
-      0xfff05c00, 0xc07e0084,
-      0x00001e86, 0x8013c000,
-      0x00001de6, 0xf0000000,
-      0x00001de7, 0x80000000,
-      /* size: 0x80 */
-
-      /* 2 coords ZS in, S encoded in R, Z encoded in GBA (8_UNORM)
-       * Setup float outputs in a way that conversion to UNORM yields the
-       * desired byte value.
-       */
-      /* NOTE: need to repeat header */
-      0x00021462, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
-      0x80000000, 0x0000000f, 0x00000000, 0x00000000, 0x00000000,
-      0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
-      0x00000000, 0x00000000, 0x00000000, 0x0000000f, 0x00000000,
-      0x0202e237, 0x22804c22,
-      0xfff09c00, 0xc07e0080,
-      0xfff0dc00, 0xc07e0084,
-      0x00201e86, 0x80104008,
-      0x00205f06, 0x80104009,
-      0x00001de6, 0xf0000000,
-      0xfc009c02, 0x312dffff,
-      0x05201e04, 0x18000000,
-      0x00428047, 0x22020272,
-      0x09209c84, 0x14000000,
-      0x04001c02, 0x30ee0202,
-      0xfc205c02, 0x38000003,
-      0x0020dc02, 0x3803fc00,
-      0x00209c02, 0x380003fc,
-      0x05205e04, 0x18000000,
-      0x0d20de04, 0x18000000,
-      0x42004277, 0x200002e0,
-      0x09209e04, 0x18000000,
-      0x04105c02, 0x30ee0202,
-      0x0430dc02, 0x30ce0202,
-      0x04209c02, 0x30de0202,
-      0x00001de7, 0x80000000,
-      /* size: 0x100 */
-
-      /* 2 coords ZS in, Z encoded in RGB, S encoded in A (U8_UNORM) */
-      0x00021462, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
-      0x80000000, 0x0000000f, 0x00000000, 0x00000000, 0x00000000,
-      0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000,
-      0x00000000, 0x00000000, 0x00000000, 0x0000000f, 0x00000000,
-      0x0202e237, 0x22804c22,
-      0xfff09c00, 0xc07e0080,
-      0xfff0dc00, 0xc07e0084,
-      0x00201e86, 0x80104008,
-      0x00205f06, 0x80104009,
-      0x00001de6, 0xf0000000,
-      0xfc009c02, 0x312dffff,
-      0x0520de04, 0x18000000,
-      0x00428047, 0x22020272,
-      0x09209c84, 0x14000000,
-      0x0430dc02, 0x30ee0202,
-      0xfc201c02, 0x38000003,
-      0x00205c02, 0x380003fc,
-      0x00209c02, 0x3803fc00,
-      0x01201e04, 0x18000000,
-      0x05205e04, 0x18000000,
-      0x42004277, 0x200002e0,
-      0x09209e04, 0x18000000,
-      0x04001c02, 0x30ee0202,
-      0x04105c02, 0x30de0202,
-      0x04209c02, 0x30ce0202,
-      0x00001de7, 0x80000000,
-   };
-
-   blit->fp.type = PIPE_SHADER_FRAGMENT;
-   blit->fp.translated = TRUE;
-   if (blit->screen->base.class_3d >= NVE4_3D_CLASS) {
-      blit->fp.code = (uint32_t *)code_nve4; /* const_cast */
-      blit->fp.code_size = sizeof(code_nve4);
-   } else {
-      blit->fp.code = (uint32_t *)code_nvc0; /* const_cast */
-      blit->fp.code_size = sizeof(code_nvc0);
-   }
-   blit->fp.max_gpr = 4;
-
-   blit->fp.hdr[0]  = 0x00021462; /* fragprog magic */
-   blit->fp.hdr[5]  = 0x80000000;
-   blit->fp.hdr[6]  = 0x0000000f; /* 2 linear */
-   blit->fp.hdr[18] = 0x0000000f; /* 1 colour output */
-}
-
-static void
-nvc0_blitctx_make_sampler(struct nvc0_blitctx *blit)
+nvc0_blitter_make_sampler(struct nvc0_blitter *blit)
 {
    /* clamp to edge, min/max lod = 0, nearest filtering */
 
@@ -681,73 +538,51 @@ nvc0_blitctx_make_sampler(struct nvc0_blitctx *blit)
       NV50_TSC_1_MAGF_LINEAR | NV50_TSC_1_MINF_LINEAR | NV50_TSC_1_MIPF_NONE;
 }
 
-/* Since shaders cannot export stencil, we cannot copy stencil values when
- * rendering to ZETA, so we attach the ZS surface to a colour render target.
- */
-static INLINE enum pipe_format
-nvc0_blit_zeta_to_colour_format(enum pipe_format format)
-{
-   switch (format) {
-   case PIPE_FORMAT_Z16_UNORM:               return PIPE_FORMAT_R16_UNORM;
-   case PIPE_FORMAT_Z24_UNORM_S8_UINT:
-   case PIPE_FORMAT_S8_UINT_Z24_UNORM:
-   case PIPE_FORMAT_Z24X8_UNORM:             return PIPE_FORMAT_R8G8B8A8_UNORM;
-   case PIPE_FORMAT_Z32_FLOAT:               return PIPE_FORMAT_R32_FLOAT;
-   case PIPE_FORMAT_Z32_FLOAT_S8X24_UINT:    return PIPE_FORMAT_R32G32_FLOAT;
-   default:
-      assert(0);
-      return PIPE_FORMAT_NONE;
-   }
-}
-
 static void
-nvc0_blitctx_get_color_mask_and_fp(struct nvc0_blitctx *blit,
-                                   enum pipe_format format, uint8_t mask)
+nvc0_blit_select_fp(struct nvc0_blitctx *ctx, const struct pipe_blit_info *info)
 {
-   blit->color_mask = 0;
-
-   switch (format) {
-   case PIPE_FORMAT_Z24X8_UNORM:
-   case PIPE_FORMAT_Z24_UNORM_S8_UINT:
-      blit->fp_offset = 0x180;
-      if (mask & PIPE_MASK_Z)
-         blit->color_mask |= 0x0111;
-      if (mask & PIPE_MASK_S)
-         blit->color_mask |= 0x1000;
-      break;
-   case PIPE_FORMAT_S8_UINT_Z24_UNORM:
-      blit->fp_offset = 0x80;
-      if (mask & PIPE_MASK_Z)
-         blit->color_mask |= 0x1110;
-      if (mask & PIPE_MASK_S)
-         blit->color_mask |= 0x0001;
-      break;
-   default:
-      blit->fp_offset = 0;
-      if (mask & (PIPE_MASK_R | PIPE_MASK_Z)) blit->color_mask |= 0x0001;
-      if (mask & (PIPE_MASK_G | PIPE_MASK_S)) blit->color_mask |= 0x0010;
-      if (mask & PIPE_MASK_B) blit->color_mask |= 0x0100;
-      if (mask & PIPE_MASK_A) blit->color_mask |= 0x1000;
-      break;
+   struct nvc0_blitter *blitter = ctx->nvc0->screen->blitter;
+
+   const enum pipe_texture_target ptarg =
+      nv50_blit_reinterpret_pipe_texture_target(info->dst.resource->target);
+
+   const unsigned targ = nv50_blit_texture_type(ptarg);
+   const unsigned mode = ctx->mode;
+
+   if (!blitter->fp[targ][mode]) {
+      pipe_mutex_lock(blitter->mutex);
+      if (!blitter->fp[targ][mode])
+         blitter->fp[targ][mode] =
+            nv50_blitter_make_fp(&ctx->nvc0->base.pipe, mode, targ);
+      pipe_mutex_unlock(blitter->mutex);
    }
+   ctx->fp = blitter->fp[targ][mode];
 }
 
 static void
-nvc0_blit_set_dst(struct nvc0_context *nvc0,
-                  struct pipe_resource *res, unsigned level, unsigned layer)
+nvc0_blit_set_dst(struct nvc0_blitctx *ctx,
+                  struct pipe_resource *res, unsigned level, unsigned layer,
+                  enum pipe_format format)
 {
+   struct nvc0_context *nvc0 = ctx->nvc0;
    struct pipe_context *pipe = &nvc0->base.pipe;
    struct pipe_surface templ;
 
-   if (util_format_is_depth_or_stencil(res->format))
-      templ.format = nvc0_blit_zeta_to_colour_format(res->format);
+   if (util_format_is_depth_or_stencil(format))
+      templ.format = nv50_blit_zeta_to_colour_format(format);
    else
-      templ.format = res->format;
+      templ.format = format;
 
    templ.usage = PIPE_USAGE_STREAM;
    templ.u.tex.level = level;
    templ.u.tex.first_layer = templ.u.tex.last_layer = layer;
 
+   if (layer == -1) {
+      templ.u.tex.first_layer = 0;
+      templ.u.tex.last_layer =
+         (res->target == PIPE_TEXTURE_3D ? res->depth0 : res->array_size) - 1;
+   }
+
    nvc0->framebuffer.cbufs[0] = nvc0_miptree_surface_new(pipe, res, &templ);
    nvc0->framebuffer.nr_cbufs = 1;
    nvc0->framebuffer.zsbuf = NULL;
@@ -755,34 +590,21 @@ nvc0_blit_set_dst(struct nvc0_context *nvc0,
    nvc0->framebuffer.height = nvc0->framebuffer.cbufs[0]->height;
 }
 
-static INLINE void
-nvc0_blit_fixup_tic_entry(struct pipe_sampler_view *view, const boolean filter)
-{
-   struct nv50_tic_entry *ent = nv50_tic_entry(view);
-
-   ent->tic[2] &= ~(1 << 31); /* scaled coordinates, ok with 3d textures ? */
-
-   /* magic: */
-
-   if (filter) {
-      /* affects quality of near vertical edges in MS8: */
-      ent->tic[3] = 0x20000000;
-   } else {
-      ent->tic[3] = 0;
-      ent->tic[6] = 0;
-   }
-}
-
 static void
-nvc0_blit_set_src(struct nvc0_context *nvc0,
+nvc0_blit_set_src(struct nvc0_blitctx *ctx,
                   struct pipe_resource *res, unsigned level, unsigned layer,
-                  const boolean filter)
+                  enum pipe_format format, const uint8_t filter)
 {
+   struct nvc0_context *nvc0 = ctx->nvc0;
    struct pipe_context *pipe = &nvc0->base.pipe;
    struct pipe_sampler_view templ;
-   int s;
+   uint32_t flags;
+   unsigned s;
+   enum pipe_texture_target target;
 
-   templ.format = res->format;
+   target = nv50_blit_reinterpret_pipe_texture_target(res->target);
+
+   templ.format = format;
    templ.u.tex.first_layer = templ.u.tex.last_layer = layer;
    templ.u.tex.first_level = templ.u.tex.last_level = level;
    templ.swizzle_r = PIPE_SWIZZLE_RED;
@@ -790,19 +612,28 @@ nvc0_blit_set_src(struct nvc0_context *nvc0,
    templ.swizzle_b = PIPE_SWIZZLE_BLUE;
    templ.swizzle_a = PIPE_SWIZZLE_ALPHA;
 
-   nvc0->textures[4][0] = nvc0_create_sampler_view(pipe, res, &templ);
-   nvc0->textures[4][1] = NULL;
+   if (layer == -1) {
+      templ.u.tex.first_layer = 0;
+      templ.u.tex.last_layer =
+         (res->target == PIPE_TEXTURE_3D ? res->depth0 : res->array_size) - 1;
+   }
 
-   nvc0_blit_fixup_tic_entry(nvc0->textures[4][0], filter);
+   flags = NV50_TEXVIEW_SCALED_COORDS;
+   if (filter && res->nr_samples == 8)
+      flags |= NV50_TEXVIEW_FILTER_MSAA8;
+
+   nvc0->textures[4][0] = nvc0_create_texture_view(
+      pipe, res, &templ, flags, target);
+   nvc0->textures[4][1] = NULL;
 
    for (s = 0; s <= 3; ++s)
       nvc0->num_textures[s] = 0;
    nvc0->num_textures[4] = 1;
 
-   templ.format = nv50_zs_to_s_format(res->format);
-   if (templ.format != res->format) {
-      nvc0->textures[4][1] = nvc0_create_sampler_view(pipe, res, &templ);
-      nvc0_blit_fixup_tic_entry(nvc0->textures[4][1], filter);
+   templ.format = nv50_zs_to_s_format(format);
+   if (templ.format != format) {
+      nvc0->textures[4][1] = nvc0_create_texture_view(
+         pipe, res, &templ, flags, target);
       nvc0->num_textures[4] = 2;
    }
 }
@@ -810,20 +641,21 @@ nvc0_blit_set_src(struct nvc0_context *nvc0,
 static void
 nvc0_blitctx_prepare_state(struct nvc0_blitctx *blit)
 {
-   struct nouveau_pushbuf *push = blit->screen->base.pushbuf;
+   struct nouveau_pushbuf *push = blit->nvc0->base.pushbuf;
 
    /* TODO: maybe make this a MACRO (if we need more logic) ? */
 
+   if (blit->nvc0->cond_query)
+      IMMED_NVC0(push, NVC0_3D(COND_MODE), NVC0_3D_COND_MODE_ALWAYS);
+
    /* blend state */
    BEGIN_NVC0(push, NVC0_3D(COLOR_MASK(0)), 1);
    PUSH_DATA (push, blit->color_mask);
-   BEGIN_NVC0(push, NVC0_3D(BLEND_ENABLE(0)), 1);
-   PUSH_DATA (push, 0);
+   IMMED_NVC0(push, NVC0_3D(BLEND_ENABLE(0)), 0);
    IMMED_NVC0(push, NVC0_3D(LOGIC_OP_ENABLE), 0);
 
    /* rasterizer state */
-   BEGIN_NVC0(push, NVC0_3D(FRAG_COLOR_CLAMP_EN), 1);
-   PUSH_DATA (push, 0);
+   IMMED_NVC0(push, NVC0_3D(FRAG_COLOR_CLAMP_EN), 0);
    IMMED_NVC0(push, NVC0_3D(MULTISAMPLE_ENABLE), 0);
    BEGIN_NVC0(push, NVC0_3D(MSAA_MASK(0)), 4);
    PUSH_DATA (push, 0xffff);
@@ -849,51 +681,57 @@ nvc0_blitctx_prepare_state(struct nvc0_blitctx *blit)
 }
 
 static void
-nvc0_blitctx_pre_blit(struct nvc0_blitctx *blit, struct nvc0_context *nvc0)
+nvc0_blitctx_pre_blit(struct nvc0_blitctx *ctx)
 {
+   struct nvc0_context *nvc0 = ctx->nvc0;
+   struct nvc0_blitter *blitter = nvc0->screen->blitter;
    int s;
 
-   blit->saved.fb.width = nvc0->framebuffer.width;
-   blit->saved.fb.height = nvc0->framebuffer.height;
-   blit->saved.fb.nr_cbufs = nvc0->framebuffer.nr_cbufs;
-   blit->saved.fb.cbufs[0] = nvc0->framebuffer.cbufs[0];
-   blit->saved.fb.zsbuf = nvc0->framebuffer.zsbuf;
+   ctx->saved.fb.width = nvc0->framebuffer.width;
+   ctx->saved.fb.height = nvc0->framebuffer.height;
+   ctx->saved.fb.nr_cbufs = nvc0->framebuffer.nr_cbufs;
+   ctx->saved.fb.cbufs[0] = nvc0->framebuffer.cbufs[0];
+   ctx->saved.fb.zsbuf = nvc0->framebuffer.zsbuf;
 
-   blit->saved.vp = nvc0->vertprog;
-   blit->saved.tcp = nvc0->tctlprog;
-   blit->saved.tep = nvc0->tevlprog;
-   blit->saved.gp = nvc0->gmtyprog;
-   blit->saved.fp = nvc0->fragprog;
+   ctx->saved.vp = nvc0->vertprog;
+   ctx->saved.tcp = nvc0->tctlprog;
+   ctx->saved.tep = nvc0->tevlprog;
+   ctx->saved.gp = nvc0->gmtyprog;
+   ctx->saved.fp = nvc0->fragprog;
 
-   nvc0->vertprog = &blit->vp;
-   nvc0->fragprog = &blit->fp;
+   nvc0->vertprog = &blitter->vp;
    nvc0->tctlprog = NULL;
    nvc0->tevlprog = NULL;
    nvc0->gmtyprog = NULL;
+   nvc0->fragprog = ctx->fp;
 
    for (s = 0; s <= 4; ++s) {
-      blit->saved.num_textures[s] = nvc0->num_textures[s];
-      blit->saved.num_samplers[s] = nvc0->num_samplers[s];
+      ctx->saved.num_textures[s] = nvc0->num_textures[s];
+      ctx->saved.num_samplers[s] = nvc0->num_samplers[s];
       nvc0->textures_dirty[s] = (1 << nvc0->num_textures[s]) - 1;
       nvc0->samplers_dirty[s] = (1 << nvc0->num_samplers[s]) - 1;
    }
-   blit->saved.texture[0] = nvc0->textures[4][0];
-   blit->saved.texture[1] = nvc0->textures[4][1];
-   blit->saved.sampler[0] = nvc0->samplers[4][0];
-   blit->saved.sampler[1] = nvc0->samplers[4][1];
+   ctx->saved.texture[0] = nvc0->textures[4][0];
+   ctx->saved.texture[1] = nvc0->textures[4][1];
+   ctx->saved.sampler[0] = nvc0->samplers[4][0];
+   ctx->saved.sampler[1] = nvc0->samplers[4][1];
 
-   nvc0->samplers[4][0] = &blit->sampler[blit->filter];
-   nvc0->samplers[4][1] = &blit->sampler[blit->filter];
+   nvc0->samplers[4][0] = &blitter->sampler[ctx->filter];
+   nvc0->samplers[4][1] = &blitter->sampler[ctx->filter];
 
    for (s = 0; s <= 3; ++s)
       nvc0->num_samplers[s] = 0;
    nvc0->num_samplers[4] = 2;
 
-   blit->saved.dirty = nvc0->dirty;
+   ctx->saved.dirty = nvc0->dirty;
 
    nvc0->textures_dirty[4] |= 3;
    nvc0->samplers_dirty[4] |= 3;
 
+   nouveau_bufctx_reset(nvc0->bufctx_3d, NVC0_BIND_FB);
+   nouveau_bufctx_reset(nvc0->bufctx_3d, NVC0_BIND_TEX(4, 0));
+   nouveau_bufctx_reset(nvc0->bufctx_3d, NVC0_BIND_TEX(4, 1));
+
    nvc0->dirty = NVC0_NEW_FRAMEBUFFER |
       NVC0_NEW_VERTPROG | NVC0_NEW_FRAGPROG |
       NVC0_NEW_TCTLPROG | NVC0_NEW_TEVLPROG | NVC0_NEW_GMTYPROG |
@@ -901,8 +739,9 @@ nvc0_blitctx_pre_blit(struct nvc0_blitctx *blit, struct nvc0_context *nvc0)
 }
 
 static void
-nvc0_blitctx_post_blit(struct nvc0_context *nvc0, struct nvc0_blitctx *blit)
+nvc0_blitctx_post_blit(struct nvc0_blitctx *blit)
 {
+   struct nvc0_context *nvc0 = blit->nvc0;
    int s;
 
    pipe_surface_reference(&nvc0->framebuffer.cbufs[0], NULL);
@@ -936,6 +775,14 @@ nvc0_blitctx_post_blit(struct nvc0_context *nvc0, struct nvc0_blitctx *blit)
    nvc0->textures_dirty[4] |= 3;
    nvc0->samplers_dirty[4] |= 3;
 
+   if (nvc0->cond_query)
+      nvc0->base.pipe.render_condition(&nvc0->base.pipe, nvc0->cond_query,
+                                       nvc0->cond_mode);
+
+   nouveau_bufctx_reset(nvc0->bufctx_3d, NVC0_BIND_FB);
+   nouveau_bufctx_reset(nvc0->bufctx_3d, NVC0_BIND_TEX(4, 0));
+   nouveau_bufctx_reset(nvc0->bufctx_3d, NVC0_BIND_TEX(4, 1));
+
    nvc0->dirty = blit->saved.dirty |
       (NVC0_NEW_FRAMEBUFFER | NVC0_NEW_SCISSOR | NVC0_NEW_SAMPLE_MASK |
        NVC0_NEW_RASTERIZER | NVC0_NEW_ZSA | NVC0_NEW_BLEND |
@@ -945,60 +792,63 @@ nvc0_blitctx_post_blit(struct nvc0_context *nvc0, struct nvc0_blitctx *blit)
        NVC0_NEW_TFB_TARGETS);
 }
 
-#if 0
 static void
-nvc0_resource_resolve(struct pipe_context *pipe,
-                      const struct pipe_resolve_info *info)
+nvc0_blit_3d(struct nvc0_context *nvc0, const struct pipe_blit_info *info)
 {
-   struct nvc0_context *nvc0 = nvc0_context(pipe);
-   struct nvc0_screen *screen = nvc0->screen;
-   struct nvc0_blitctx *blit = screen->blitctx;
-   struct nouveau_pushbuf *push = screen->base.pushbuf;
-   struct pipe_resource *src = info->src.res;
-   struct pipe_resource *dst = info->dst.res;
-   float x0, x1, y0, y1;
+   struct nvc0_blitctx *blit = nvc0->blit;
+   struct nouveau_pushbuf *push = nvc0->base.pushbuf;
+   struct pipe_resource *src = info->src.resource;
+   struct pipe_resource *dst = info->dst.resource;
+   int32_t minx, maxx, miny, maxy;
+   int32_t i;
+   float x0, x1, y0, y1, z;
+   float dz;
    float x_range, y_range;
 
-   /* Would need more shader variants or, better, just change the TIC target.
-    * But no API creates 3D MS textures ...
-    */
-   if (src->target == PIPE_TEXTURE_3D)
-      return;
-
-   nvc0_blitctx_get_color_mask_and_fp(blit, dst->format, info->mask);
+   blit->mode = nv50_blit_select_mode(info);
+   blit->color_mask = nv50_blit_derive_color_mask(info);
+   blit->filter = nv50_blit_get_filter(info);
 
-   blit->filter = util_format_is_depth_or_stencil(dst->format) ? 0 : 1;
+   nvc0_blit_select_fp(blit, info);
+   nvc0_blitctx_pre_blit(blit);
 
-   nvc0_blitctx_pre_blit(blit, nvc0);
-
-   nvc0_blit_set_dst(nvc0, dst, info->dst.level, info->dst.layer);
-   nvc0_blit_set_src(nvc0, src, 0,               info->src.layer, blit->filter);
+   nvc0_blit_set_dst(blit, dst, info->dst.level,  0, info->dst.format);
+   nvc0_blit_set_src(blit, src, info->src.level, -1, info->src.format,
+                     blit->filter);
 
    nvc0_blitctx_prepare_state(blit);
 
-   nvc0_state_validate(nvc0, ~0, 36);
+   nvc0_state_validate(nvc0, ~0, 48);
 
-   x_range =
-      (float)(info->src.x1 - info->src.x0) /
-      (float)(info->dst.x1 - info->dst.x0);
-   y_range =
-      (float)(info->src.y1 - info->src.y0) /
-      (float)(info->dst.y1 - info->dst.y0);
+   x_range = (float)info->src.box.width / (float)info->dst.box.width;
+   y_range = (float)info->src.box.height / (float)info->dst.box.height;
 
-   x0 = (float)info->src.x0 - x_range * (float)info->dst.x0;
-   y0 = (float)info->src.y0 - y_range * (float)info->dst.y0;
+   x0 = (float)info->src.box.x - x_range * (float)info->dst.box.x;
+   y0 = (float)info->src.box.y - y_range * (float)info->dst.box.y;
 
-   x1 = x0 + 16384.0f * x_range;
-   y1 = y0 + 16384.0f * y_range;
+   x1 = info->src.box.x + 16384.0f * x_range;
+   y1 = info->src.box.y + 16384.0f * y_range;
 
    x0 *= (float)(1 << nv50_miptree(src)->ms_x);
    x1 *= (float)(1 << nv50_miptree(src)->ms_x);
    y0 *= (float)(1 << nv50_miptree(src)->ms_y);
    y1 *= (float)(1 << nv50_miptree(src)->ms_y);
 
-   BEGIN_NVC0(push, NVC0_3D(SP_START_ID(5)), 1);
-   PUSH_DATA (push,
-              blit->fp.code_base + blit->fp_offset);
+   if (src->last_level > 0) {
+      /* If there are mip maps, GPU always assumes normalized coordinates. */
+      const unsigned l = info->src.level;
+      const float fh = u_minify(src->width0 << nv50_miptree(src)->ms_x, l);
+      const float fv = u_minify(src->height0 << nv50_miptree(src)->ms_y, l);
+      x0 /= fh;
+      x1 /= fh;
+      y0 /= fv;
+      y1 /= fv;
+   }
+
+   dz = (float)info->src.box.depth / (float)info->dst.box.depth;
+   z = (float)info->src.box.z;
+   if (nv50_miptree(src)->layout_3d)
+      z += 0.5f * dz;
 
    IMMED_NVC0(push, NVC0_3D(VIEWPORT_TRANSFORM_EN), 0);
 
@@ -1008,69 +858,320 @@ nvc0_resource_resolve(struct pipe_context *pipe,
     * arranged in a way to yield the desired offset and scale.
     */
 
+   minx = info->dst.box.x;
+   maxx = info->dst.box.x + info->dst.box.width;
+   miny = info->dst.box.y;
+   maxy = info->dst.box.y + info->dst.box.height;
+   if (info->scissor_enable) {
+      minx = MAX2(minx, info->scissor.minx);
+      maxx = MIN2(maxx, info->scissor.maxx);
+      miny = MAX2(miny, info->scissor.miny);
+      maxy = MIN2(maxy, info->scissor.maxy);
+   }
    BEGIN_NVC0(push, NVC0_3D(SCISSOR_HORIZ(0)), 2);
-   PUSH_DATA (push, (info->dst.x1 << 16) | info->dst.x0);
-   PUSH_DATA (push, (info->dst.y1 << 16) | info->dst.y0);
-
-   IMMED_NVC0(push, NVC0_3D(VERTEX_BEGIN_GL),
-              NVC0_3D_VERTEX_BEGIN_GL_PRIMITIVE_TRIANGLES);
-
-   BEGIN_NVC0(push, NVC0_3D(VTX_ATTR_DEFINE), 3);
-   PUSH_DATA (push, 0x74201);
-   PUSH_DATAf(push, x0);
-   PUSH_DATAf(push, y0);
-   BEGIN_NVC0(push, NVC0_3D(VTX_ATTR_DEFINE), 3);
-   PUSH_DATA (push, 0x74200);
-   PUSH_DATAf(push, 0.0f);
-   PUSH_DATAf(push, 0.0f);
-   BEGIN_NVC0(push, NVC0_3D(VTX_ATTR_DEFINE), 3);
-   PUSH_DATA (push, 0x74201);
-   PUSH_DATAf(push, x1);
-   PUSH_DATAf(push, y0);
-   BEGIN_NVC0(push, NVC0_3D(VTX_ATTR_DEFINE), 3);
-   PUSH_DATA (push, 0x74200);
-   PUSH_DATAf(push, 16384 << nv50_miptree(dst)->ms_x);
-   PUSH_DATAf(push, 0.0f);
-   BEGIN_NVC0(push, NVC0_3D(VTX_ATTR_DEFINE), 3);
-   PUSH_DATA (push, 0x74201);
-   PUSH_DATAf(push, x0);
-   PUSH_DATAf(push, y1);
-   BEGIN_NVC0(push, NVC0_3D(VTX_ATTR_DEFINE), 3);
-   PUSH_DATA (push, 0x74200);
-   PUSH_DATAf(push, 0.0f);
-   PUSH_DATAf(push, 16384 << nv50_miptree(dst)->ms_y);
-
-   IMMED_NVC0(push, NVC0_3D(VERTEX_END_GL), 0);
+   PUSH_DATA (push, (maxx << 16) | minx);
+   PUSH_DATA (push, (maxy << 16) | miny);
+
+   for (i = 0; i < info->dst.box.depth; ++i, z += dz) {
+      if (info->dst.box.z + i) {
+         BEGIN_NVC0(push, NVC0_3D(LAYER), 1);
+         PUSH_DATA (push, info->dst.box.z + i);
+      }
+
+      IMMED_NVC0(push, NVC0_3D(VERTEX_BEGIN_GL),
+                       NVC0_3D_VERTEX_BEGIN_GL_PRIMITIVE_TRIANGLES);
+
+      BEGIN_NVC0(push, NVC0_3D(VTX_ATTR_DEFINE), 4);
+      PUSH_DATA (push, 0x74301);
+      PUSH_DATAf(push, x0);
+      PUSH_DATAf(push, y0);
+      PUSH_DATAf(push, z);
+      BEGIN_NVC0(push, NVC0_3D(VTX_ATTR_DEFINE), 3);
+      PUSH_DATA (push, 0x74200);
+      PUSH_DATAf(push, 0.0f);
+      PUSH_DATAf(push, 0.0f);
+      BEGIN_NVC0(push, NVC0_3D(VTX_ATTR_DEFINE), 4);
+      PUSH_DATA (push, 0x74301);
+      PUSH_DATAf(push, x1);
+      PUSH_DATAf(push, y0);
+      PUSH_DATAf(push, z);
+      BEGIN_NVC0(push, NVC0_3D(VTX_ATTR_DEFINE), 3);
+      PUSH_DATA (push, 0x74200);
+      PUSH_DATAf(push, 16384 << nv50_miptree(dst)->ms_x);
+      PUSH_DATAf(push, 0.0f);
+      BEGIN_NVC0(push, NVC0_3D(VTX_ATTR_DEFINE), 4);
+      PUSH_DATA (push, 0x74301);
+      PUSH_DATAf(push, x0);
+      PUSH_DATAf(push, y1);
+      PUSH_DATAf(push, z);
+      BEGIN_NVC0(push, NVC0_3D(VTX_ATTR_DEFINE), 3);
+      PUSH_DATA (push, 0x74200);
+      PUSH_DATAf(push, 0.0f);
+      PUSH_DATAf(push, 16384 << nv50_miptree(dst)->ms_y);
+
+      IMMED_NVC0(push, NVC0_3D(VERTEX_END_GL), 0);
+   }
+   if (info->dst.box.z + info->dst.box.depth - 1)
+      IMMED_NVC0(push, NVC0_3D(LAYER), 0);
 
    /* re-enable normally constant state */
 
    IMMED_NVC0(push, NVC0_3D(VIEWPORT_TRANSFORM_EN), 1);
 
-   nvc0_blitctx_post_blit(nvc0, blit);
+   nvc0_blitctx_post_blit(blit);
+}
+
+static void
+nvc0_blit_eng2d(struct nvc0_context *nvc0, const struct pipe_blit_info *info)
+{
+   struct nouveau_pushbuf *push = nvc0->base.pushbuf;
+   struct nv50_miptree *dst = nv50_miptree(info->dst.resource);
+   struct nv50_miptree *src = nv50_miptree(info->src.resource);
+   const int32_t srcx_adj = info->src.box.width < 0 ? -1 : 0;
+   const int32_t srcy_adj = info->src.box.height < 0 ? -1 : 0;
+   const int dz = info->dst.box.z;
+   const int sz = info->src.box.z;
+   uint32_t dstw, dsth;
+   int32_t dstx, dsty;
+   int64_t srcx, srcy;
+   int64_t du_dx, dv_dy;
+   int i;
+   uint32_t mode;
+   const uint32_t mask = nv50_blit_eng2d_get_mask(info);
+
+   mode = nv50_blit_get_filter(info) ?
+      NVC0_2D_BLIT_CONTROL_FILTER_BILINEAR :
+      NVC0_2D_BLIT_CONTROL_FILTER_POINT_SAMPLE;
+   mode |= (src->base.base.nr_samples > dst->base.base.nr_samples) ?
+      NVC0_2D_BLIT_CONTROL_ORIGIN_CORNER : NVC0_2D_BLIT_CONTROL_ORIGIN_CENTER;
+
+   du_dx = ((int64_t)info->src.box.width << 32) / info->dst.box.width;
+   dv_dy = ((int64_t)info->src.box.height << 32) / info->dst.box.height;
+
+   nvc0_2d_texture_set(push, 1, dst, info->dst.level, dz, info->dst.format);
+   nvc0_2d_texture_set(push, 0, src, info->src.level, sz, info->src.format);
+
+   if (info->scissor_enable) {
+      BEGIN_NVC0(push, NVC0_2D(CLIP_X), 5);
+      PUSH_DATA (push, info->scissor.minx << dst->ms_x);
+      PUSH_DATA (push, info->scissor.miny << dst->ms_y);
+      PUSH_DATA (push, (info->scissor.maxx - info->scissor.minx) << dst->ms_x);
+      PUSH_DATA (push, (info->scissor.maxy - info->scissor.miny) << dst->ms_y);
+      PUSH_DATA (push, 1); /* enable */
+   }
+
+   if (mask != 0xffffffff) {
+      IMMED_NVC0(push, NVC0_2D(PATTERN_COLOR_FORMAT),
+                       NVC0_2D_PATTERN_COLOR_FORMAT_32BPP);
+      BEGIN_NVC0(push, NVC0_2D(PATTERN_COLOR(0)), 4);
+      PUSH_DATA (push, 0x00000000);
+      PUSH_DATA (push, mask);
+      PUSH_DATA (push, 0xffffffff);
+      PUSH_DATA (push, 0xffffffff);
+      IMMED_NVC0(push, NVC0_2D(OPERATION), NVC0_2D_OPERATION_SRCCOPY_PREMULT);
+   }
+
+   if (src->ms_x > dst->ms_x || src->ms_y > dst->ms_y) {
+      /* ms_x is always >= ms_y */
+      du_dx <<= src->ms_x - dst->ms_x;
+      dv_dy <<= src->ms_y - dst->ms_y;
+   } else {
+      du_dx >>= dst->ms_x - src->ms_x;
+      dv_dy >>= dst->ms_y - src->ms_y;
+   }
+
+   srcx = (int64_t)(info->src.box.x + srcx_adj) << (src->ms_x + 32);
+   srcy = (int64_t)(info->src.box.y + srcy_adj) << (src->ms_y + 32);
+
+   if (src->base.base.nr_samples > dst->base.base.nr_samples) {
+      /* center src coorinates for proper MS resolve filtering */
+      srcx += (int64_t)src->ms_x << 32;
+      srcy += (int64_t)src->ms_y << 32;
+   }
+
+   dstx = info->dst.box.x << dst->ms_x;
+   dsty = info->dst.box.y << dst->ms_y;
+
+   dstw = info->dst.box.width << dst->ms_x;
+   dsth = info->dst.box.height << dst->ms_y;
+
+   if (dstx < 0) {
+      dstw += dstx;
+      srcx -= du_dx * dstx;
+      dstx = 0;
+   }
+   if (dsty < 0) {
+      dsth += dsty;
+      srcy -= dv_dy * dsty;
+      dsty = 0;
+   }
+
+   IMMED_NVC0(push, NVC0_2D(BLIT_CONTROL), mode);
+   BEGIN_NVC0(push, NVC0_2D(BLIT_DST_X), 4);
+   PUSH_DATA (push, dstx);
+   PUSH_DATA (push, dsty);
+   PUSH_DATA (push, dstw);
+   PUSH_DATA (push, dsth);
+   BEGIN_NVC0(push, NVC0_2D(BLIT_DU_DX_FRACT), 4);
+   PUSH_DATA (push, du_dx);
+   PUSH_DATA (push, du_dx >> 32);
+   PUSH_DATA (push, dv_dy);
+   PUSH_DATA (push, dv_dy >> 32);
+
+   BCTX_REFN(nvc0->bufctx, 2D, &dst->base, WR);
+   BCTX_REFN(nvc0->bufctx, 2D, &src->base, RD);
+   nouveau_pushbuf_bufctx(nvc0->base.pushbuf, nvc0->bufctx);
+   if (nouveau_pushbuf_validate(nvc0->base.pushbuf))
+      return;
+
+   for (i = 0; i < info->dst.box.depth; ++i) {
+      if (i > 0) {
+         /* no scaling in z-direction possible for eng2d blits */
+         if (dst->layout_3d) {
+            BEGIN_NVC0(push, NVC0_2D(DST_LAYER), 1);
+            PUSH_DATA (push, info->dst.box.z + i);
+         } else {
+            const unsigned z = info->dst.box.z + i;
+            BEGIN_NVC0(push, NVC0_2D(DST_ADDRESS_HIGH), 2);
+            PUSH_DATAh(push, dst->base.address + z * dst->layer_stride);
+            PUSH_DATA (push, dst->base.address + z * dst->layer_stride);
+         }
+         if (src->layout_3d) {
+            /* not possible because of depth tiling */
+            assert(0);
+         } else {
+            const unsigned z = info->src.box.z + i;
+            BEGIN_NVC0(push, NVC0_2D(SRC_ADDRESS_HIGH), 2);
+            PUSH_DATAh(push, src->base.address + z * src->layer_stride);
+            PUSH_DATA (push, src->base.address + z * src->layer_stride);
+         }
+         BEGIN_NVC0(push, NVC0_2D(BLIT_SRC_Y_INT), 1); /* trigger */
+         PUSH_DATA (push, srcy >> 32);
+      } else {
+         BEGIN_NVC0(push, NVC0_2D(BLIT_SRC_X_FRACT), 4);
+         PUSH_DATA (push, srcx);
+         PUSH_DATA (push, srcx >> 32);
+         PUSH_DATA (push, srcy);
+         PUSH_DATA (push, srcy >> 32);
+      }
+   }
+   nvc0_bufctx_fence(nvc0, nvc0->bufctx, FALSE);
+
+   nouveau_bufctx_reset(nvc0->bufctx, NVC0_BIND_2D);
+
+   if (info->scissor_enable)
+      IMMED_NVC0(push, NVC0_2D(CLIP_ENABLE), 0);
+   if (mask != 0xffffffff)
+      IMMED_NVC0(push, NVC0_2D(OPERATION), NVC0_2D_OPERATION_SRCCOPY);
 }
+
+static void
+nvc0_blit(struct pipe_context *pipe, const struct pipe_blit_info *info)
+{
+   struct nvc0_context *nvc0 = nvc0_context(pipe);
+   boolean eng3d = FALSE;
+
+   if (util_format_is_depth_or_stencil(info->dst.resource->format)) {
+      if (!(info->mask & PIPE_MASK_ZS))
+         return;
+      if (info->dst.resource->format == PIPE_FORMAT_Z32_FLOAT ||
+          info->dst.resource->format == PIPE_FORMAT_Z32_FLOAT_S8X24_UINT)
+         eng3d = TRUE;
+      if (info->filter != PIPE_TEX_FILTER_NEAREST)
+         eng3d = TRUE;
+   } else {
+      if (!(info->mask & PIPE_MASK_RGBA))
+         return;
+      if (info->mask != PIPE_MASK_RGBA)
+         eng3d = TRUE;
+   }
+
+   if (nv50_miptree(info->src.resource)->layout_3d) {
+      eng3d = TRUE;
+   } else
+   if (info->src.box.depth != info->dst.box.depth) {
+      eng3d = TRUE;
+      debug_printf("blit: cannot filter array or cube textures in z direction");
+   }
+
+   if (!eng3d && info->dst.format != info->src.format)
+      if (!nvc0_2d_format_faithful(info->dst.format) ||
+          !nvc0_2d_format_faithful(info->src.format))
+         eng3d = TRUE;
+
+   if (info->src.resource->nr_samples == 8 &&
+       info->dst.resource->nr_samples <= 1)
+      eng3d = TRUE;
+#if 0
+   /* FIXME: can't make this work with eng2d anymore, at least not on nv50 */
+   if (info->src.resource->nr_samples > 1 ||
+       info->dst.resource->nr_samples > 1)
+      eng3d = TRUE;
 #endif
+   /* FIXME: find correct src coordinates adjustments */
+   if ((info->src.box.width !=  info->dst.box.width &&
+        info->src.box.width != -info->dst.box.width) ||
+       (info->src.box.height !=  info->dst.box.height &&
+        info->src.box.height != -info->dst.box.height))
+      eng3d = TRUE;
+
+   if (!eng3d)
+      nvc0_blit_eng2d(nvc0, info);
+   else
+      nvc0_blit_3d(nvc0, info);
+}
 
 boolean
-nvc0_blitctx_create(struct nvc0_screen *screen)
+nvc0_blitter_create(struct nvc0_screen *screen)
 {
-   screen->blitctx = CALLOC_STRUCT(nvc0_blitctx);
-   if (!screen->blitctx) {
-      NOUVEAU_ERR("failed to allocate blit context\n");
+   screen->blitter = CALLOC_STRUCT(nvc0_blitter);
+   if (!screen->blitter) {
+      NOUVEAU_ERR("failed to allocate blitter struct\n");
       return FALSE;
    }
+   screen->blitter->screen = screen;
 
-   screen->blitctx->screen = screen;
+   pipe_mutex_init(screen->blitter->mutex);
 
-   nvc0_blitctx_make_vp(screen->blitctx);
-   nvc0_blitctx_make_fp(screen->blitctx);
+   nvc0_blitter_make_vp(screen->blitter);
+   nvc0_blitter_make_sampler(screen->blitter);
 
-   nvc0_blitctx_make_sampler(screen->blitctx);
+   return TRUE;
+}
 
-   screen->blitctx->color_mask = 0x1111;
+void
+nvc0_blitter_destroy(struct nvc0_screen *screen)
+{
+   struct nvc0_blitter *blitter = screen->blitter;
+   unsigned i, m;
+
+   for (i = 0; i < NV50_BLIT_MAX_TEXTURE_TYPES; ++i) {
+      for (m = 0; m < NV50_BLIT_MODES; ++m) {
+         struct nvc0_program *prog = blitter->fp[i][m];
+         if (prog) {
+            nvc0_program_destroy(NULL, prog);
+            FREE((void *)prog->pipe.tokens);
+            FREE(prog);
+         }
+      }
+   }
 
-   return TRUE;
+   FREE(blitter);
 }
 
+boolean
+nvc0_blitctx_create(struct nvc0_context *nvc0)
+{
+   nvc0->blit = CALLOC_STRUCT(nvc0_blitctx);
+   if (!nvc0->blit) {
+      NOUVEAU_ERR("failed to allocate blit context\n");
+      return FALSE;
+   }
+
+   nvc0->blit->nvc0 = nvc0;
+
+   return TRUE;
+}
 
 void
 nvc0_init_surface_functions(struct nvc0_context *nvc0)
@@ -1078,6 +1179,7 @@ nvc0_init_surface_functions(struct nvc0_context *nvc0)
    struct pipe_context *pipe = &nvc0->base.pipe;
 
    pipe->resource_copy_region = nvc0_resource_copy_region;
+   pipe->blit = nvc0_blit;
    pipe->clear_render_target = nvc0_clear_render_target;
    pipe->clear_depth_stencil = nvc0_clear_depth_stencil;
 }
diff --git a/src/gallium/drivers/nvc0/nvc0_tex.c b/src/gallium/drivers/nvc0/nvc0_tex.c
index 8dd7185bcdf..0930212a350 100644
--- a/src/gallium/drivers/nvc0/nvc0_tex.c
+++ b/src/gallium/drivers/nvc0/nvc0_tex.c
@@ -55,9 +55,24 @@ nv50_tic_swizzle(uint32_t tc, unsigned swz, boolean tex_int)
 
 struct pipe_sampler_view *
 nvc0_create_sampler_view(struct pipe_context *pipe,
-                         struct pipe_resource *texture,
+                         struct pipe_resource *res,
                          const struct pipe_sampler_view *templ)
 {
+   uint32_t flags = 0;
+
+   if (res->target == PIPE_TEXTURE_RECT)
+      flags |= NV50_TEXVIEW_SCALED_COORDS;
+
+   return nvc0_create_texture_view(pipe, res, templ, flags, res->target);
+}
+
+struct pipe_sampler_view *
+nvc0_create_texture_view(struct pipe_context *pipe,
+                         struct pipe_resource *texture,
+                         const struct pipe_sampler_view *templ,
+                         uint32_t flags,
+                         enum pipe_texture_target target)
+{
    const struct util_format_description *desc;
    uint64_t address;
    uint32_t *tic;
@@ -132,7 +147,7 @@ nvc0_create_sampler_view(struct pipe_context *pipe,
       return &view->pipe;
    }
 
-   if (mt->base.base.target != PIPE_TEXTURE_RECT)
+   if (!(flags & NV50_TEXVIEW_SCALED_COORDS))
       tic[2] |= NV50_TIC_2_NORMALIZED_COORDS;
 
    tic[2] |=
@@ -149,7 +164,7 @@ nvc0_create_sampler_view(struct pipe_context *pipe,
    tic[1] = address;
    tic[2] |= address >> 32;
 
-   switch (mt->base.base.target) {
+   switch (target) {
    case PIPE_TEXTURE_1D:
       tic[2] |= NV50_TIC_2_TARGET_1D;
       break;
@@ -185,7 +200,7 @@ nvc0_create_sampler_view(struct pipe_context *pipe,
    if (mt->base.base.target == PIPE_BUFFER)
       tic[3] = mt->base.base.width0;
    else
-      tic[3] = 0x00300000;
+      tic[3] = (flags & NV50_TEXVIEW_FILTER_MSAA8) ? 0x20000000 : 0x00300000;
 
    tic[4] = (1 << 31) | (mt->base.base.width0 << mt->ms_x);