Merge remote branch 'upstream/gallium-0.1' into nouveau-gallium-0.1

6 files changed, 357 insertions, 94 deletions

diff --git a/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c b/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c
index f599bee07e7..ce41418a0f9 100644
--- a/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c
+++ b/src/gallium/auxiliary/pipebuffer/pb_buffer_fenced.c
@@ -142,12 +142,13 @@ _fenced_buffer_destroy(struct fenced_buffer *fenced_buf)
 
 
 static INLINE void
-_fenced_buffer_remove(struct fenced_buffer *fenced_buf)
+_fenced_buffer_remove(struct fenced_buffer_list *fenced_list,
+                      struct fenced_buffer *fenced_buf)
 {
-   struct fenced_buffer_list *fenced_list = fenced_buf->list;
    struct pipe_winsys *winsys = fenced_list->winsys;
 
    assert(fenced_buf->fence);
+   assert(fenced_buf->list == fenced_list);
    
    winsys->fence_reference(winsys, &fenced_buf->fence, NULL);
    fenced_buf->flags &= ~PIPE_BUFFER_USAGE_GPU_READ_WRITE;
@@ -184,7 +185,8 @@ _fenced_buffer_finish(struct fenced_buffer *fenced_buf)
 	 return PIPE_ERROR;
       }
       /* Remove from the fenced list */
-      _fenced_buffer_remove(fenced_buf); /* TODO: remove consequents */
+      /* TODO: remove consequents */
+      _fenced_buffer_remove(fenced_list, fenced_buf);
    }
 
    fenced_buf->flags &= ~PIPE_BUFFER_USAGE_GPU_READ_WRITE;
@@ -223,7 +225,7 @@ _fenced_buffer_list_check_free(struct fenced_buffer_list *fenced_list,
 	 assert(winsys->fence_signalled(winsys, fenced_buf->fence, 0) == 0);
       }
 
-      _fenced_buffer_remove(fenced_buf);
+      _fenced_buffer_remove(fenced_list, fenced_buf);
 
       curr = next; 
       next = curr->next;
@@ -248,7 +250,7 @@ fenced_buffer_destroy(struct pb_buffer *buf)
 	 do {
 	    fenced_buf = LIST_ENTRY(struct fenced_buffer, curr, head);
 	    assert(winsys->fence_signalled(winsys, fenced_buf->fence, 0) == 0);
-	    _fenced_buffer_remove(fenced_buf);
+	    _fenced_buffer_remove(fenced_list, fenced_buf);
 	    curr = prev;
 	    prev = curr->prev;
 	 } while (curr != &fenced_list->delayed);
@@ -395,7 +397,7 @@ buffer_fence(struct pb_buffer *buf,
    
    _glthread_LOCK_MUTEX(fenced_list->mutex);
    if (fenced_buf->fence)
-      _fenced_buffer_remove(fenced_buf);
+      _fenced_buffer_remove(fenced_list, fenced_buf);
    if (fence) {
       winsys->fence_reference(winsys, &fenced_buf->fence, fence);
       fenced_buf->flags |= flags & PIPE_BUFFER_USAGE_GPU_READ_WRITE;
diff --git a/src/gallium/auxiliary/tgsi/tgsi_dump.c b/src/gallium/auxiliary/tgsi/tgsi_dump.c
index 0eedb1c91e3..29bb530b4dd 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_dump.c
+++ b/src/gallium/auxiliary/tgsi/tgsi_dump.c
@@ -26,6 +26,7 @@
  **************************************************************************/
 
 #include "pipe/p_debug.h"
+#include "util/u_string.h"
 #include "tgsi_dump.h"
 #include "tgsi_iterate.h"
 
@@ -34,28 +35,31 @@ struct dump_ctx
    struct tgsi_iterate_context iter;
 
    uint instno;
+   
+   struct util_strbuf *sbuf;
 };
 
 static void
 dump_enum(
+   struct util_strbuf *sbuf,
    uint e,
    const char **enums,
    uint enum_count )
 {
    if (e >= enum_count)
-      debug_printf( "%u", e );
+      util_strbuf_printf( sbuf, "%u", e );
    else
-      debug_printf( "%s", enums[e] );
+      util_strbuf_printf( sbuf, "%s", enums[e] );
 }
 
-#define EOL()           debug_printf( "\n" )
-#define TXT(S)          debug_printf( "%s", S )
-#define CHR(C)          debug_printf( "%c", C )
-#define UIX(I)          debug_printf( "0x%x", I )
-#define UID(I)          debug_printf( "%u", I )
-#define SID(I)          debug_printf( "%d", I )
-#define FLT(F)          debug_printf( "%10.4f", F )
-#define ENM(E,ENUMS)    dump_enum( E, ENUMS, sizeof( ENUMS ) / sizeof( *ENUMS ) )
+#define EOL()           util_strbuf_printf( sbuf, "\n" )
+#define TXT(S)          util_strbuf_printf( sbuf, "%s", S )
+#define CHR(C)          util_strbuf_printf( sbuf, "%c", C )
+#define UIX(I)          util_strbuf_printf( sbuf, "0x%x", I )
+#define UID(I)          util_strbuf_printf( sbuf, "%u", I )
+#define SID(I)          util_strbuf_printf( sbuf, "%d", I )
+#define FLT(F)          util_strbuf_printf( sbuf, "%10.4f", F )
+#define ENM(E,ENUMS)    dump_enum( sbuf, E, ENUMS, sizeof( ENUMS ) / sizeof( *ENUMS ) )
 
 static const char *processor_type_names[] =
 {
@@ -266,6 +270,7 @@ static const char *modulate_names[TGSI_MODULATE_COUNT] =
 
 static void
 _dump_register(
+   struct util_strbuf *sbuf,
    uint file,
    int first,
    int last )
@@ -282,6 +287,7 @@ _dump_register(
 
 static void
 _dump_register_ind(
+   struct util_strbuf *sbuf,
    uint file,
    int index,
    uint ind_file,
@@ -303,6 +309,7 @@ _dump_register_ind(
 
 static void
 _dump_writemask(
+   struct util_strbuf *sbuf,
    uint writemask )
 {
    if (writemask != TGSI_WRITEMASK_XYZW) {
@@ -318,17 +325,23 @@ _dump_writemask(
    }
 }
 
-void
-tgsi_dump_declaration(
-   const struct tgsi_full_declaration *decl )
+static boolean
+iter_declaration(
+   struct tgsi_iterate_context *iter,
+   struct tgsi_full_declaration *decl )
 {
+   struct dump_ctx *ctx = (struct dump_ctx *) iter;
+   struct util_strbuf *sbuf = ctx->sbuf;
+
    TXT( "DCL " );
 
    _dump_register(
+      sbuf,
       decl->Declaration.File,
       decl->DeclarationRange.First,
       decl->DeclarationRange.Last );
    _dump_writemask(
+      sbuf,
       decl->Declaration.UsageMask );
 
    if (decl->Declaration.Semantic) {
@@ -346,21 +359,35 @@ tgsi_dump_declaration(
    ENM( decl->Declaration.Interpolate, interpolate_names );
 
    EOL();
-}
 
-static boolean
-iter_declaration(
-   struct tgsi_iterate_context *iter,
-   struct tgsi_full_declaration *decl )
-{
-   tgsi_dump_declaration( decl );
    return TRUE;
 }
 
 void
-tgsi_dump_immediate(
-   const struct tgsi_full_immediate *imm )
+tgsi_dump_declaration(
+   const struct tgsi_full_declaration *decl )
 {
+   static char str[1024];
+   struct util_strbuf sbuf;
+   struct dump_ctx ctx;
+
+   util_strbuf_init(&sbuf, str, sizeof(str));
+   
+   ctx.sbuf = &sbuf;
+
+   iter_declaration( &ctx.iter, (struct tgsi_full_declaration *)decl );
+   
+   debug_printf("%s", str);
+}
+
+static boolean
+iter_immediate(
+   struct tgsi_iterate_context *iter,
+   struct tgsi_full_immediate *imm )
+{
+   struct dump_ctx *ctx = (struct dump_ctx *) iter;
+   struct util_strbuf *sbuf = ctx->sbuf;
+
    uint i;
 
    TXT( "IMM " );
@@ -382,22 +409,36 @@ tgsi_dump_immediate(
    TXT( " }" );
 
    EOL();
-}
 
-static boolean
-iter_immediate(
-   struct tgsi_iterate_context *iter,
-   struct tgsi_full_immediate *imm )
-{
-   tgsi_dump_immediate( imm );
    return TRUE;
 }
 
 void
-tgsi_dump_instruction(
-   const struct tgsi_full_instruction *inst,
-   uint instno )
+tgsi_dump_immediate(
+   const struct tgsi_full_immediate *imm )
 {
+   static char str[1024];
+   struct util_strbuf sbuf;
+   struct dump_ctx ctx;
+
+   util_strbuf_init(&sbuf, str, sizeof(str));
+   
+   ctx.sbuf = &sbuf;
+
+   iter_immediate( &ctx.iter, (struct tgsi_full_immediate *)imm );
+   
+   debug_printf("%s", str);
+}
+
+static boolean
+iter_instruction(
+   struct tgsi_iterate_context *iter,
+   struct tgsi_full_instruction *inst )
+{
+   struct dump_ctx *ctx = (struct dump_ctx *) iter;
+   struct util_strbuf *sbuf = ctx->sbuf;
+   uint instno = ctx->instno++;
+   
    uint i;
    boolean first_reg = TRUE;
 
@@ -426,11 +467,12 @@ tgsi_dump_instruction(
       CHR( ' ' );
 
       _dump_register(
+         sbuf,
          dst->DstRegister.File,
          dst->DstRegister.Index,
          dst->DstRegister.Index );
       ENM( dst->DstRegisterExtModulate.Modulate, modulate_names );
-      _dump_writemask( dst->DstRegister.WriteMask );
+      _dump_writemask( sbuf, dst->DstRegister.WriteMask );
 
       first_reg = FALSE;
    }
@@ -457,6 +499,7 @@ tgsi_dump_instruction(
 
       if (src->SrcRegister.Indirect) {
          _dump_register_ind(
+            sbuf,
             src->SrcRegister.File,
             src->SrcRegister.Index,
             src->SrcRegisterInd.File,
@@ -464,6 +507,7 @@ tgsi_dump_instruction(
       }
       else {
          _dump_register(
+            sbuf,
             src->SrcRegister.File,
             src->SrcRegister.Index,
             src->SrcRegister.Index );
@@ -529,38 +573,55 @@ tgsi_dump_instruction(
    }
 
    EOL();
+
+   return TRUE;
 }
 
-static boolean
-iter_instruction(
-   struct tgsi_iterate_context *iter,
-   struct tgsi_full_instruction *inst )
+void
+tgsi_dump_instruction(
+   const struct tgsi_full_instruction *inst,
+   uint instno )
 {
-   struct dump_ctx *ctx = (struct dump_ctx *) iter;
+   static char str[1024];
+   struct util_strbuf sbuf;
+   struct dump_ctx ctx;
 
-   tgsi_dump_instruction( inst, ctx->instno++ );
-   return TRUE;
+   util_strbuf_init(&sbuf, str, sizeof(str));
+   
+   ctx.instno = instno;
+   ctx.sbuf = &sbuf;
+
+   iter_instruction( &ctx.iter, (struct tgsi_full_instruction *)inst );
+   
+   debug_printf("%s", str);
 }
 
 static boolean
 prolog(
-   struct tgsi_iterate_context *ctx )
+   struct tgsi_iterate_context *iter )
 {
-   ENM( ctx->processor.Processor, processor_type_names );
-   UID( ctx->version.MajorVersion );
+   struct dump_ctx *ctx = (struct dump_ctx *) iter;
+   struct util_strbuf *sbuf = ctx->sbuf;
+   ENM( iter->processor.Processor, processor_type_names );
+   UID( iter->version.MajorVersion );
    CHR( '.' );
-   UID( ctx->version.MinorVersion );
+   UID( iter->version.MinorVersion );
    EOL();
    return TRUE;
 }
 
 void
-tgsi_dump(
+tgsi_dump_str(
    const struct tgsi_token *tokens,
-   uint flags )
+   uint flags,
+   char *str,
+   size_t size)
 {
+   struct util_strbuf sbuf;
    struct dump_ctx ctx;
 
+   util_strbuf_init(&sbuf, str, size);
+   
    /* sanity checks */
    assert( strcmp( opcode_names[TGSI_OPCODE_CONT], "CONT" ) == 0 );
    assert( strcmp( opcode_names[TGSI_OPCODE_END], "END" ) == 0 );
@@ -572,6 +633,19 @@ tgsi_dump(
    ctx.iter.epilog = NULL;
 
    ctx.instno = 0;
+   ctx.sbuf = &sbuf;
 
    tgsi_iterate_shader( tokens, &ctx.iter );
 }
+
+void
+tgsi_dump(
+   const struct tgsi_token *tokens,
+   uint flags )
+{
+   static char str[4096];
+   
+   tgsi_dump_str(tokens, flags, str, sizeof(str));
+   
+   debug_printf("%s", str);
+}
diff --git a/src/gallium/auxiliary/tgsi/tgsi_dump.h b/src/gallium/auxiliary/tgsi/tgsi_dump.h
index 51c230b5db4..ad1e647ec90 100644
--- a/src/gallium/auxiliary/tgsi/tgsi_dump.h
+++ b/src/gallium/auxiliary/tgsi/tgsi_dump.h
@@ -35,6 +35,13 @@ extern "C" {
 #endif
 
 void
+tgsi_dump_str(
+   const struct tgsi_token *tokens,
+   uint flags,
+   char *str,
+   size_t size);
+
+void
 tgsi_dump(
    const struct tgsi_token *tokens,
    uint flags );
diff --git a/src/gallium/auxiliary/util/p_debug.c b/src/gallium/auxiliary/util/p_debug.c
index cdc7e66361d..082b0e9fb56 100644
--- a/src/gallium/auxiliary/util/p_debug.c
+++ b/src/gallium/auxiliary/util/p_debug.c
@@ -50,12 +50,12 @@
 
 #endif
 
-
-
 #include "pipe/p_compiler.h" 
 #include "pipe/p_util.h" 
 #include "pipe/p_debug.h" 
 #include "pipe/p_format.h" 
+#include "pipe/p_state.h" 
+#include "pipe/p_inlines.h" 
 #include "util/u_string.h" 
 
 
@@ -335,7 +335,7 @@ const char *
 debug_dump_enum(const struct debug_named_value *names, 
                 unsigned long value)
 {
-   static char rest[256];
+   static char rest[64];
    
    while(names->name) {
       if(names->value == value)
@@ -498,10 +498,9 @@ void debug_print_format(const char *msg, unsigned fmt )
 }
 #endif
 
-char *pf_sprint_name( char *str, enum pipe_format format )
+const char *pf_name( enum pipe_format format )
 {
-   strcpy( str, debug_dump_enum(pipe_format_names, format) );
-   return str;
+   return debug_dump_enum(pipe_format_names, format);
 }
 
 
@@ -509,7 +508,7 @@ char *pf_sprint_name( char *str, enum pipe_format format )
 void debug_dump_image(const char *prefix,
                       unsigned format, unsigned cpp,
                       unsigned width, unsigned height,
-                      unsigned pitch,
+                      unsigned stride,
                       const void *data)     
 {
 #ifdef PIPE_SUBSYSTEM_WINDOWS_DISPLAY
@@ -530,7 +529,7 @@ void debug_dump_image(const char *prefix,
    for(i = 0; i < sizeof(filename); ++i)
       wfilename[i] = (WCHAR)filename[i];
    
-   pMap = (unsigned char *)EngMapFile(wfilename, sizeof(header) + cpp*width*height, &iFile);
+   pMap = (unsigned char *)EngMapFile(wfilename, sizeof(header) + height*width*cpp, &iFile);
    if(!pMap)
       return;
    
@@ -542,11 +541,44 @@ void debug_dump_image(const char *prefix,
    pMap += sizeof(header);
    
    for(i = 0; i < height; ++i) {
-      memcpy(pMap, (unsigned char *)data + cpp*pitch*i, cpp*width);
+      memcpy(pMap, (unsigned char *)data + stride*i, cpp*width);
       pMap += cpp*width;
    }
       
    EngUnmapFile(iFile);
 #endif
 }
+
+void debug_dump_surface(const char *prefix,
+                        struct pipe_surface *surface)     
+{
+   unsigned surface_usage;
+   void *data;
+
+   if (!surface)
+      goto error1;
+
+   /* XXX: force mappable surface */
+   surface_usage = surface->usage;
+   surface->usage |= PIPE_BUFFER_USAGE_CPU_READ;
+   
+   data = pipe_surface_map(surface, 
+                           PIPE_BUFFER_USAGE_CPU_READ);
+   if(!data)
+      goto error2;
+   
+   debug_dump_image(prefix, 
+                    surface->format,
+                    surface->block.size, 
+                    surface->nblocksx,
+                    surface->nblocksy,
+                    surface->stride,
+                    data);
+   
+   pipe_surface_unmap(surface);
+error2:
+   surface->usage = surface_usage;
+error1:
+   ;
+}
 #endif
diff --git a/src/gallium/auxiliary/util/p_tile.c b/src/gallium/auxiliary/util/p_tile.c
index 2b6db43bee6..e366039a27c 100644
--- a/src/gallium/auxiliary/util/p_tile.c
+++ b/src/gallium/auxiliary/util/p_tile.c
@@ -51,12 +51,12 @@ pipe_get_tile_raw(struct pipe_surface *ps,
 {
    const void *src;
 
-   if (pipe_clip_tile(x, y, &w, &h, ps))
-      return;
-
    if (dst_stride == 0)
       dst_stride = pf_get_nblocksx(&ps->block, w) * ps->block.size;
 
+   if (pipe_clip_tile(x, y, &w, &h, ps))
+      return;
+
    src = pipe_surface_map(ps, PIPE_BUFFER_USAGE_CPU_READ);
    assert(src);
    if(!src)
@@ -79,12 +79,12 @@ pipe_put_tile_raw(struct pipe_surface *ps,
 {
    void *dst;
 
-   if (pipe_clip_tile(x, y, &w, &h, ps))
-      return;
-
    if (src_stride == 0)
       src_stride = pf_get_nblocksx(&ps->block, w) * ps->block.size;
 
+   if (pipe_clip_tile(x, y, &w, &h, ps))
+      return;
+
    dst = pipe_surface_map(ps, PIPE_BUFFER_USAGE_CPU_WRITE);
    assert(dst);
    if(!dst)
@@ -109,7 +109,7 @@ pipe_put_tile_raw(struct pipe_surface *ps,
 /*** PIPE_FORMAT_A8R8G8B8_UNORM ***/
 
 static void
-a8r8g8b8_get_tile_rgba(unsigned *src,
+a8r8g8b8_get_tile_rgba(const unsigned *src,
                        unsigned w, unsigned h,
                        float *p,
                        unsigned dst_stride)
@@ -156,7 +156,7 @@ a8r8g8b8_put_tile_rgba(unsigned *dst,
 /*** PIPE_FORMAT_A8R8G8B8_UNORM ***/
 
 static void
-x8r8g8b8_get_tile_rgba(unsigned *src,
+x8r8g8b8_get_tile_rgba(const unsigned *src,
                        unsigned w, unsigned h,
                        float *p,
                        unsigned dst_stride)
@@ -202,7 +202,7 @@ x8r8g8b8_put_tile_rgba(unsigned *dst,
 /*** PIPE_FORMAT_B8G8R8A8_UNORM ***/
 
 static void
-b8g8r8a8_get_tile_rgba(unsigned *src,
+b8g8r8a8_get_tile_rgba(const unsigned *src,
                        unsigned w, unsigned h,
                        float *p,
                        unsigned dst_stride)
@@ -249,7 +249,7 @@ b8g8r8a8_put_tile_rgba(unsigned *dst,
 /*** PIPE_FORMAT_A1R5G5B5_UNORM ***/
 
 static void
-a1r5g5b5_get_tile_rgba(ushort *src,
+a1r5g5b5_get_tile_rgba(const ushort *src,
                        unsigned w, unsigned h,
                        float *p,
                        unsigned dst_stride)
@@ -270,10 +270,37 @@ a1r5g5b5_get_tile_rgba(ushort *src,
 }
 
 
+static void
+a1r5g5b5_put_tile_rgba(ushort *dst,
+                       unsigned w, unsigned h,
+                       const float *p,
+                       unsigned src_stride)
+{
+   unsigned i, j;
+
+   for (i = 0; i < h; i++) {
+      const float *pRow = p;
+      for (j = 0; j < w; j++, pRow += 4) {
+         unsigned r, g, b, a;
+         UNCLAMPED_FLOAT_TO_UBYTE(r, pRow[0]);
+         UNCLAMPED_FLOAT_TO_UBYTE(g, pRow[1]);
+         UNCLAMPED_FLOAT_TO_UBYTE(b, pRow[2]);
+         UNCLAMPED_FLOAT_TO_UBYTE(a, pRow[3]);
+         r = r >> 3;  /* 5 bits */
+         g = g >> 3;  /* 5 bits */
+         b = b >> 3;  /* 5 bits */
+         a = a >> 7;  /* 1 bit */
+         *dst++ = (a << 15) | (r << 10) | (g << 5) | b;
+      }
+      p += src_stride;
+   }
+}
+
+
 /*** PIPE_FORMAT_A4R4G4B4_UNORM ***/
 
 static void
-a4r4g4b4_get_tile_rgba(ushort *src,
+a4r4g4b4_get_tile_rgba(const ushort *src,
                        unsigned w, unsigned h,
                        float *p,
                        unsigned dst_stride)
@@ -324,7 +351,7 @@ a4r4g4b4_put_tile_rgba(ushort *dst,
 /*** PIPE_FORMAT_R5G6B5_UNORM ***/
 
 static void
-r5g6b5_get_tile_rgba(ushort *src,
+r5g6b5_get_tile_rgba(const ushort *src,
                      unsigned w, unsigned h,
                      float *p,
                      unsigned dst_stride)
@@ -373,7 +400,7 @@ r5g6b5_put_tile_rgba(ushort *dst,
  * Return each Z value as four floats in [0,1].
  */
 static void
-z16_get_tile_rgba(ushort *src,
+z16_get_tile_rgba(const ushort *src,
                   unsigned w, unsigned h,
                   float *p,
                   unsigned dst_stride)
@@ -399,7 +426,7 @@ z16_get_tile_rgba(ushort *src,
 /*** PIPE_FORMAT_L8_UNORM ***/
 
 static void
-l8_get_tile_rgba(ubyte *src,
+l8_get_tile_rgba(const ubyte *src,
                  unsigned w, unsigned h,
                  float *p,
                  unsigned dst_stride)
@@ -419,10 +446,31 @@ l8_get_tile_rgba(ubyte *src,
 }
 
 
+static void
+l8_put_tile_rgba(ubyte *dst,
+                 unsigned w, unsigned h,
+                 const float *p,
+                 unsigned src_stride)
+{
+   unsigned i, j;
+
+   for (i = 0; i < h; i++) {
+      const float *pRow = p;
+      for (j = 0; j < w; j++, pRow += 4) {
+         unsigned r;
+         UNCLAMPED_FLOAT_TO_UBYTE(r, pRow[0]);
+         *dst++ = r;
+      }
+      p += src_stride;
+   }
+}
+
+
+
 /*** PIPE_FORMAT_A8_UNORM ***/
 
 static void
-a8_get_tile_rgba(ubyte *src,
+a8_get_tile_rgba(const ubyte *src,
                  unsigned w, unsigned h,
                  float *p,
                  unsigned dst_stride)
@@ -441,10 +489,32 @@ a8_get_tile_rgba(ubyte *src,
    }
 }
 
+
+static void
+a8_put_tile_rgba(ubyte *dst,
+                 unsigned w, unsigned h,
+                 const float *p,
+                 unsigned src_stride)
+{
+   unsigned i, j;
+
+   for (i = 0; i < h; i++) {
+      const float *pRow = p;
+      for (j = 0; j < w; j++, pRow += 4) {
+         unsigned a;
+         UNCLAMPED_FLOAT_TO_UBYTE(a, pRow[3]);
+         *dst++ = a;
+      }
+      p += src_stride;
+   }
+}
+
+
+
 /*** PIPE_FORMAT_R16_SNORM ***/
 
 static void
-r16_get_tile_rgba(short *src,
+r16_get_tile_rgba(const short *src,
                   unsigned w, unsigned h,
                   float *p,
                   unsigned dst_stride)
@@ -485,7 +555,7 @@ r16_put_tile_rgba(short *dst,
 /*** PIPE_FORMAT_R16G16B16A16_SNORM ***/
 
 static void
-r16g16b16a16_get_tile_rgba(short *src,
+r16g16b16a16_get_tile_rgba(const short *src,
                            unsigned w, unsigned h,
                            float *p,
                            unsigned dst_stride)
@@ -530,7 +600,7 @@ r16g16b16a16_put_tile_rgba(short *dst,
 /*** PIPE_FORMAT_I8_UNORM ***/
 
 static void
-i8_get_tile_rgba(ubyte *src,
+i8_get_tile_rgba(const ubyte *src,
                  unsigned w, unsigned h,
                  float *p,
                  unsigned dst_stride)
@@ -550,13 +620,33 @@ i8_get_tile_rgba(ubyte *src,
 }
 
 
+static void
+i8_put_tile_rgba(ubyte *dst,
+                 unsigned w, unsigned h,
+                 const float *p,
+                 unsigned src_stride)
+{
+   unsigned i, j;
+
+   for (i = 0; i < h; i++) {
+      const float *pRow = p;
+      for (j = 0; j < w; j++, pRow += 4) {
+         unsigned r;
+         UNCLAMPED_FLOAT_TO_UBYTE(r, pRow[0]);
+         *dst++ = r;
+      }
+      p += src_stride;
+   }
+}
+
+
 /*** PIPE_FORMAT_A8L8_UNORM ***/
 
 static void
-a8_l8_get_tile_rgba(ushort *src,
-                    unsigned w, unsigned h,
-                    float *p,
-                    unsigned dst_stride)
+a8l8_get_tile_rgba(const ushort *src,
+                   unsigned w, unsigned h,
+                   float *p,
+                   unsigned dst_stride)
 {
    unsigned i, j;
 
@@ -574,6 +664,27 @@ a8_l8_get_tile_rgba(ushort *src,
 }
 
 
+static void
+a8l8_put_tile_rgba(ushort *dst,
+                   unsigned w, unsigned h,
+                   const float *p,
+                   unsigned src_stride)
+{
+   unsigned i, j;
+
+   for (i = 0; i < h; i++) {
+      const float *pRow = p;
+      for (j = 0; j < w; j++, pRow += 4) {
+         unsigned r, a;
+         UNCLAMPED_FLOAT_TO_UBYTE(r, pRow[0]);
+         UNCLAMPED_FLOAT_TO_UBYTE(a, pRow[3]);
+         *dst++ = (a << 8) | r;
+      }
+      p += src_stride;
+   }
+}
+
+
 
 
 /*** PIPE_FORMAT_Z32_UNORM ***/
@@ -582,7 +693,7 @@ a8_l8_get_tile_rgba(ushort *src,
  * Return each Z value as four floats in [0,1].
  */
 static void
-z32_get_tile_rgba(unsigned *src,
+z32_get_tile_rgba(const unsigned *src,
                   unsigned w, unsigned h,
                   float *p,
                   unsigned dst_stride)
@@ -609,7 +720,7 @@ z32_get_tile_rgba(unsigned *src,
  * Return Z component as four float in [0,1].  Stencil part ignored.
  */
 static void
-s8z24_get_tile_rgba(unsigned *src,
+s8z24_get_tile_rgba(const unsigned *src,
                     unsigned w, unsigned h,
                     float *p,
                     unsigned dst_stride)
@@ -636,7 +747,7 @@ s8z24_get_tile_rgba(unsigned *src,
  * Return Z component as four float in [0,1].  Stencil part ignored.
  */
 static void
-z24s8_get_tile_rgba(unsigned *src,
+z24s8_get_tile_rgba(const unsigned *src,
                     unsigned w, unsigned h,
                     float *p,
                     unsigned dst_stride)
@@ -663,7 +774,7 @@ z24s8_get_tile_rgba(unsigned *src,
  * Convert YCbCr (or YCrCb) to RGBA.
  */
 static void
-ycbcr_get_tile_rgba(ushort *src,
+ycbcr_get_tile_rgba(const ushort *src,
                     unsigned w, unsigned h,
                     float *p,
                     unsigned dst_stride,
@@ -780,7 +891,7 @@ pipe_tile_raw_to_rgba(enum pipe_format format,
       i8_get_tile_rgba((ubyte *) src, w, h, dst, dst_stride);
       break;
    case PIPE_FORMAT_A8L8_UNORM:
-      a8_l8_get_tile_rgba((ushort *) src, w, h, dst, dst_stride);
+      a8l8_get_tile_rgba((ushort *) src, w, h, dst, dst_stride);
       break;
    case PIPE_FORMAT_R16_SNORM:
       r16_get_tile_rgba((short *) src, w, h, dst, dst_stride);
@@ -867,7 +978,7 @@ pipe_put_tile_rgba(struct pipe_surface *ps,
       b8g8r8a8_put_tile_rgba((unsigned *) packed, w, h, p, src_stride);
       break;
    case PIPE_FORMAT_A1R5G5B5_UNORM:
-      /*a1r5g5b5_put_tile_rgba((ushort *) packed, w, h, p, src_stride);*/
+      a1r5g5b5_put_tile_rgba((ushort *) packed, w, h, p, src_stride);
       break;
    case PIPE_FORMAT_R5G6B5_UNORM:
       r5g6b5_put_tile_rgba((ushort *) packed, w, h, p, src_stride);
@@ -879,19 +990,19 @@ pipe_put_tile_rgba(struct pipe_surface *ps,
       a4r4g4b4_put_tile_rgba((ushort *) packed, w, h, p, src_stride);
       break;
    case PIPE_FORMAT_L8_UNORM:
-      /*l8_put_tile_rgba((ubyte *) packed, w, h, p, src_stride);*/
+      l8_put_tile_rgba((ubyte *) packed, w, h, p, src_stride);
       break;
    case PIPE_FORMAT_A8_UNORM:
-      /*a8_put_tile_rgba((ubyte *) packed, w, h, p, src_stride);*/
+      a8_put_tile_rgba((ubyte *) packed, w, h, p, src_stride);
       break;
    case PIPE_FORMAT_I8_UNORM:
-      /*i8_put_tile_rgba((ubyte *) packed, w, h, p, src_stride);*/
+      i8_put_tile_rgba((ubyte *) packed, w, h, p, src_stride);
       break;
    case PIPE_FORMAT_A8L8_UNORM:
-      /*a8_l8_put_tile_rgba((ushort *) packed, w, h, p, src_stride);*/
+      a8l8_put_tile_rgba((ushort *) packed, w, h, p, src_stride);
       break;
    case PIPE_FORMAT_R16_SNORM:
-      /*r16_put_tile_rgba((short *) packed, w, h, p, src_stride);*/
+      r16_put_tile_rgba((short *) packed, w, h, p, src_stride);
       break;
    case PIPE_FORMAT_R16G16B16A16_SNORM:
       r16g16b16a16_put_tile_rgba((short *) packed, w, h, p, src_stride);
diff --git a/src/gallium/auxiliary/util/u_string.h b/src/gallium/auxiliary/util/u_string.h
index 73c88d87b4c..abc3232b492 100644
--- a/src/gallium/auxiliary/util/u_string.h
+++ b/src/gallium/auxiliary/util/u_string.h
@@ -176,6 +176,43 @@ util_memmove(void *dest, const void *src, size_t n)
 #endif
 
 
+/**
+ * Printable string buffer
+ */
+struct util_strbuf
+{
+   char *str;
+   char *ptr;
+   size_t left;
+};
+
+
+static INLINE void
+util_strbuf_init(struct util_strbuf *sbuf, char *str, size_t size) 
+{
+   sbuf->str = str;
+   sbuf->str[0] = 0;
+   sbuf->ptr = sbuf->str;
+   sbuf->left = size;
+}
+
+
+static INLINE void
+util_strbuf_printf(struct util_strbuf *sbuf, const char *format, ...)
+{
+   if(sbuf->left > 1) {
+      size_t written;
+      va_list ap;
+      va_start(ap, format);
+      written = util_vsnprintf(sbuf->ptr, sbuf->left, format, ap);
+      va_end(ap);
+      sbuf->ptr += written;
+      sbuf->left -= written;
+   }
+}
+
+
+
 #ifdef __cplusplus
 }
 #endif