optimize generated vertex programs a bit
[mesa.git] / src / mesa / drivers / glide / fxvb.c
index 2d9f73d6a8f5af6e6fa97bf897ff0198e009c66b..34ada61f4e444d342105e18563a046567b75b34f 100644 (file)
  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
- * KEITH WHITWELL, OR ANY OTHER CONTRIBUTORS BE LIABLE FOR ANY CLAIM, 
- * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 
- * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE 
+ * KEITH WHITWELL, OR ANY OTHER CONTRIBUTORS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE
  * OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  */
 
 /* Authors:
- *   Keith Whitwell
+ *    Keith Whitwell <keith@tungstengraphics.com>
+ *    Daniel Borca <dborca@users.sourceforge.net>
  */
 
 #ifdef HAVE_CONFIG_H
@@ -54,54 +55,46 @@ static void copy_pv( GLcontext *ctx, GLuint edst, GLuint esrc )
    GrVertex *dst = fxMesa->verts + edst;
    GrVertex *src = fxMesa->verts + esrc;
 
+#if FX_PACKEDCOLOR
    *(GLuint *)&dst->pargb = *(GLuint *)&src->pargb;
+#else  /* !FX_PACKEDCOLOR */
+   COPY_FLOAT(dst->r, src->r);
+   COPY_FLOAT(dst->g, src->g);
+   COPY_FLOAT(dst->b, src->b);
+   COPY_FLOAT(dst->a, src->a);
+#endif /* !FX_PACKEDCOLOR */
 }
 
-typedef void (*emit_func)( GLcontext *, GLuint, GLuint, void * );
-
-static struct {
-   emit_func           emit;
-   interp_func         interp;
-   GLboolean           (*check_tex_sizes)( GLcontext *ctx );
-   GLuint               vertex_format;
-} setup_tab[MAX_SETUP];
-
-
-static void import_float_colors( GLcontext *ctx )
+static void copy_pv2( GLcontext *ctx, GLuint edst, GLuint esrc )
 {
-   struct vertex_buffer *VB = &TNL_CONTEXT(ctx)->vb;
-   struct gl_client_array *from = VB->ColorPtr[0];
-   struct gl_client_array *to = &FX_CONTEXT(ctx)->UbyteColor;
-   GLuint count = VB->Count;
-
-   if (!to->Ptr) {
-      to->Ptr = ALIGN_MALLOC( VB->Size * 4 * sizeof(GLubyte), 32 );
-      to->Type = GL_UNSIGNED_BYTE;
-   }
+   fxMesaContext fxMesa = FX_CONTEXT( ctx );
+   GrVertex *dst = fxMesa->verts + edst;
+   GrVertex *src = fxMesa->verts + esrc;
 
-   /* No need to transform the same value 3000 times.
-    */
-   if (!from->StrideB) {
-      to->StrideB = 0;
-      count = 1;
-   }
-   else
-      to->StrideB = 4 * sizeof(GLubyte);
-   
-   _math_trans_4ub( (GLubyte (*)[4]) to->Ptr,
-                   from->Ptr,
-                   from->StrideB,
-                   from->Type,
-                   from->Size,
-                   0,
-                   count);
-
-   VB->ColorPtr[0] = to;
+#if FX_PACKEDCOLOR
+   *(GLuint *)&dst->pargb = *(GLuint *)&src->pargb;
+   *(GLuint *)&dst->pspec = *(GLuint *)&src->pspec;
+#else  /* !FX_PACKEDCOLOR */
+   COPY_FLOAT(dst->r, src->r);
+   COPY_FLOAT(dst->g, src->g);
+   COPY_FLOAT(dst->b, src->b);
+   COPY_FLOAT(dst->a, src->a);
+   COPY_FLOAT(dst->r1, src->r1);
+   COPY_FLOAT(dst->g1, src->g1);
+   COPY_FLOAT(dst->b1, src->b1);
+#endif /* !FX_PACKEDCOLOR */
 }
 
+static struct {
+   void                      (*emit) (GLcontext *ctx, GLuint start, GLuint end, void *dest);
+   tnl_copy_pv_func    copy_pv;
+   tnl_interp_func     interp;
+   GLboolean         (*check_tex_sizes) (GLcontext *ctx);
+   GLuint              vertex_format;
+} setup_tab[MAX_SETUP];
 
-/* Hack alert: assume chan is 8 bits */
-#define GET_COLOR(ptr, idx) (((GLchan (*)[4])((ptr)->Ptr))[idx])
+
+#define GET_COLOR(ptr, idx) ((ptr)->data[idx])
 
 
 static void interp_extras( GLcontext *ctx,
@@ -112,18 +105,24 @@ static void interp_extras( GLcontext *ctx,
    struct vertex_buffer *VB = &TNL_CONTEXT(ctx)->vb;
 
    if (VB->ColorPtr[1]) {
-      INTERP_4CHAN( t,
-                GET_COLOR(VB->ColorPtr[1], dst),
-                GET_COLOR(VB->ColorPtr[1], out),
-                GET_COLOR(VB->ColorPtr[1], in) );
-#if 1 /* [dBorca] GL_EXT_separate_specular_color */
+      /* If stride is zero, ColorPtr[1] is constant across the VB, so
+       * there is no point interpolating between two values as they will
+       * be identical.  This case is handled in t_dd_tritmp.h
+       */
+      if (VB->ColorPtr[1]->stride) {
+        assert(VB->ColorPtr[1]->stride == 4 * sizeof(GLfloat));
+        INTERP_4F( t,
+                   GET_COLOR(VB->ColorPtr[1], dst),
+                   GET_COLOR(VB->ColorPtr[1], out),
+                   GET_COLOR(VB->ColorPtr[1], in) );
+      }
+
       if (VB->SecondaryColorPtr[1]) {
-        INTERP_3CHAN( t,
+        INTERP_3F( t,
                    GET_COLOR(VB->SecondaryColorPtr[1], dst),
                    GET_COLOR(VB->SecondaryColorPtr[1], out),
                    GET_COLOR(VB->SecondaryColorPtr[1], in) );
       }
-#endif
    }
 
    if (VB->EdgeFlag) {
@@ -139,17 +138,16 @@ static void copy_pv_extras( GLcontext *ctx, GLuint dst, GLuint src )
    struct vertex_buffer *VB = &TNL_CONTEXT(ctx)->vb;
 
    if (VB->ColorPtr[1]) {
-        COPY_CHAN4( GET_COLOR(VB->ColorPtr[1], dst),
+        COPY_4FV( GET_COLOR(VB->ColorPtr[1], dst),
                   GET_COLOR(VB->ColorPtr[1], src) );
-#if 1 /* [dBorca] GL_EXT_separate_specular_color */
+
         if (VB->SecondaryColorPtr[1]) {
-           COPY_CHAN4( GET_COLOR(VB->SecondaryColorPtr[1], dst),
+           COPY_3FV( GET_COLOR(VB->SecondaryColorPtr[1], dst),
                      GET_COLOR(VB->SecondaryColorPtr[1], src) );
         }
-#endif
    }
 
-   copy_pv(ctx, dst, src);
+   setup_tab[FX_CONTEXT(ctx)->SetupIndex].copy_pv(ctx, dst, src);
 }
 
 
@@ -174,6 +172,157 @@ static void copy_pv_extras( GLcontext *ctx, GLuint dst, GLuint src )
 #define TAG(x) x##_wgpt0t1
 #include "fxvbtmp.h"
 
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_PSIZ)
+#define TAG(x) x##_wga
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_TMU0|SETUP_PSIZ)
+#define TAG(x) x##_wgt0a
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_TMU0|SETUP_TMU1|SETUP_PSIZ)
+#define TAG(x) x##_wgt0t1a
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_TMU0|SETUP_PTEX|SETUP_PSIZ)
+#define TAG(x) x##_wgpt0a
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_TMU0|SETUP_TMU1|\
+             SETUP_PTEX|SETUP_PSIZ)
+#define TAG(x) x##_wgpt0t1a
+#include "fxvbtmp.h"
+
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC)
+#define TAG(x) x##_2wg
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0)
+#define TAG(x) x##_2wgt0
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_TMU1)
+#define TAG(x) x##_2wgt0t1
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_PTEX)
+#define TAG(x) x##_2wgpt0
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_TMU1|\
+             SETUP_PTEX)
+#define TAG(x) x##_2wgpt0t1
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_PSIZ)
+#define TAG(x) x##_2wga
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_PSIZ)
+#define TAG(x) x##_2wgt0a
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_TMU1|SETUP_PSIZ)
+#define TAG(x) x##_2wgt0t1a
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_PTEX|SETUP_PSIZ)
+#define TAG(x) x##_2wgpt0a
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_TMU1|\
+             SETUP_PTEX|SETUP_PSIZ)
+#define TAG(x) x##_2wgpt0t1a
+#include "fxvbtmp.h"
+
+/* fog { */
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_FOGC)
+#define TAG(x) x##_wgf
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_TMU0|SETUP_FOGC)
+#define TAG(x) x##_wgt0f
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_TMU0|SETUP_TMU1|SETUP_FOGC)
+#define TAG(x) x##_wgt0t1f
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_TMU0|SETUP_PTEX|SETUP_FOGC)
+#define TAG(x) x##_wgpt0f
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_TMU0|SETUP_TMU1|\
+             SETUP_PTEX|SETUP_FOGC)
+#define TAG(x) x##_wgpt0t1f
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_wgaf
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_TMU0|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_wgt0af
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_TMU0|SETUP_TMU1|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_wgt0t1af
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_TMU0|SETUP_PTEX|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_wgpt0af
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_TMU0|SETUP_TMU1|\
+             SETUP_PTEX|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_wgpt0t1af
+#include "fxvbtmp.h"
+
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_FOGC)
+#define TAG(x) x##_2wgf
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_FOGC)
+#define TAG(x) x##_2wgt0f
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_TMU1|SETUP_FOGC)
+#define TAG(x) x##_2wgt0t1f
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_PTEX|SETUP_FOGC)
+#define TAG(x) x##_2wgpt0f
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_TMU1|\
+             SETUP_PTEX|SETUP_FOGC)
+#define TAG(x) x##_2wgpt0t1f
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_2wgaf
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_2wgt0af
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_TMU1|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_2wgt0t1af
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_PTEX|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_2wgpt0af
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_TMU1|\
+             SETUP_PTEX|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_2wgpt0t1af
+#include "fxvbtmp.h"
+/* fog } */
+
 
 /* Snapping for voodoo-1
  */
@@ -200,6 +349,171 @@ static void copy_pv_extras( GLcontext *ctx, GLuint dst, GLuint src )
 #define TAG(x) x##_wsgpt0t1
 #include "fxvbtmp.h"
 
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_PSIZ)
+#define TAG(x) x##_wsga
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_TMU0|SETUP_PSIZ)
+#define TAG(x) x##_wsgt0a
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_TMU0|\
+             SETUP_TMU1|SETUP_PSIZ)
+#define TAG(x) x##_wsgt0t1a
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_TMU0|\
+             SETUP_PTEX|SETUP_PSIZ)
+#define TAG(x) x##_wsgpt0a
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_TMU0|\
+            SETUP_TMU1|SETUP_PTEX|SETUP_PSIZ)
+#define TAG(x) x##_wsgpt0t1a
+#include "fxvbtmp.h"
+
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC)
+#define TAG(x) x##_2wsg
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0)
+#define TAG(x) x##_2wsgt0
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|\
+             SETUP_TMU1)
+#define TAG(x) x##_2wsgt0t1
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|\
+             SETUP_PTEX)
+#define TAG(x) x##_2wsgpt0
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|\
+            SETUP_TMU1|SETUP_PTEX)
+#define TAG(x) x##_2wsgpt0t1
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_PSIZ)
+#define TAG(x) x##_2wsga
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_PSIZ)
+#define TAG(x) x##_2wsgt0a
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|\
+             SETUP_TMU1|SETUP_PSIZ)
+#define TAG(x) x##_2wsgt0t1a
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|\
+             SETUP_PTEX|SETUP_PSIZ)
+#define TAG(x) x##_2wsgpt0a
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|\
+            SETUP_TMU1|SETUP_PTEX|SETUP_PSIZ)
+#define TAG(x) x##_2wsgpt0t1a
+#include "fxvbtmp.h"
+
+/* fog { */
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_FOGC)
+#define TAG(x) x##_wsgf
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_TMU0|SETUP_FOGC)
+#define TAG(x) x##_wsgt0f
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_TMU0|\
+             SETUP_TMU1|SETUP_FOGC)
+#define TAG(x) x##_wsgt0t1f
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_TMU0|\
+             SETUP_PTEX|SETUP_FOGC)
+#define TAG(x) x##_wsgpt0f
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_TMU0|\
+            SETUP_TMU1|SETUP_PTEX|SETUP_FOGC)
+#define TAG(x) x##_wsgpt0t1f
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_wsgaf
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_TMU0|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_wsgt0af
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_TMU0|\
+             SETUP_TMU1|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_wsgt0t1af
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_TMU0|\
+             SETUP_PTEX|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_wsgpt0af
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_TMU0|\
+            SETUP_TMU1|SETUP_PTEX|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_wsgpt0t1af
+#include "fxvbtmp.h"
+
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_FOGC)
+#define TAG(x) x##_2wsgf
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_FOGC)
+#define TAG(x) x##_2wsgt0f
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|\
+             SETUP_TMU1|SETUP_FOGC)
+#define TAG(x) x##_2wsgt0t1f
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|\
+             SETUP_PTEX|SETUP_FOGC)
+#define TAG(x) x##_2wsgpt0f
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|\
+            SETUP_TMU1|SETUP_PTEX|SETUP_FOGC)
+#define TAG(x) x##_2wsgpt0t1f
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_2wsgaf
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_2wsgt0af
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|\
+             SETUP_TMU1|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_2wsgt0t1af
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|\
+             SETUP_PTEX|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_2wsgpt0af
+#include "fxvbtmp.h"
+
+#define IND (SETUP_XYZW|SETUP_SNAP|SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|\
+            SETUP_TMU1|SETUP_PTEX|SETUP_PSIZ|SETUP_FOGC)
+#define TAG(x) x##_2wsgpt0t1af
+#include "fxvbtmp.h"
+/* fog } */
+
 
 /* Vertex repair (multipass rendering)
  */
@@ -224,6 +538,27 @@ static void copy_pv_extras( GLcontext *ctx, GLuint dst, GLuint src )
 #include "fxvbtmp.h"
 
 
+#define IND (SETUP_RGBA|SETUP_SPEC)
+#define TAG(x) x##_2g
+#include "fxvbtmp.h"
+
+#define IND (SETUP_TMU0|SETUP_SPEC)
+#define TAG(x) x##_2t0
+#include "fxvbtmp.h"
+
+#define IND (SETUP_TMU0|SETUP_SPEC|SETUP_TMU1)
+#define TAG(x) x##_2t0t1
+#include "fxvbtmp.h"
+
+#define IND (SETUP_RGBA|SETUP_SPEC|SETUP_TMU0)
+#define TAG(x) x##_2gt0
+#include "fxvbtmp.h"
+
+#define IND (SETUP_RGBA|SETUP_SPEC|SETUP_TMU0|SETUP_TMU1)
+#define TAG(x) x##_2gt0t1
+#include "fxvbtmp.h"
+
+
 
 static void init_setup_tab( void )
 {
@@ -232,31 +567,109 @@ static void init_setup_tab( void )
    init_wgt0t1();
    init_wgpt0();
    init_wgpt0t1();
+   init_wga();
+   init_wgt0a();
+   init_wgt0t1a();
+   init_wgpt0a();
+   init_wgpt0t1a();
+    init_2wg();
+    init_2wgt0();
+    init_2wgt0t1();
+    init_2wgpt0();
+    init_2wgpt0t1();
+    init_2wga();
+    init_2wgt0a();
+    init_2wgt0t1a();
+    init_2wgpt0a();
+    init_2wgpt0t1a();
+   init_wgf();
+   init_wgt0f();
+   init_wgt0t1f();
+   init_wgpt0f();
+   init_wgpt0t1f();
+   init_wgaf();
+   init_wgt0af();
+   init_wgt0t1af();
+   init_wgpt0af();
+   init_wgpt0t1af();
+    init_2wgf();
+    init_2wgt0f();
+    init_2wgt0t1f();
+    init_2wgpt0f();
+    init_2wgpt0t1f();
+    init_2wgaf();
+    init_2wgt0af();
+    init_2wgt0t1af();
+    init_2wgpt0af();
+    init_2wgpt0t1af();
 
    init_wsg();
    init_wsgt0();
    init_wsgt0t1();
    init_wsgpt0();
    init_wsgpt0t1();
+   init_wsga();
+   init_wsgt0a();
+   init_wsgt0t1a();
+   init_wsgpt0a();
+   init_wsgpt0t1a();
+    init_2wsg();
+    init_2wsgt0();
+    init_2wsgt0t1();
+    init_2wsgpt0();
+    init_2wsgpt0t1();
+    init_2wsga();
+    init_2wsgt0a();
+    init_2wsgt0t1a();
+    init_2wsgpt0a();
+    init_2wsgpt0t1a();
+   init_wsgf();
+   init_wsgt0f();
+   init_wsgt0t1f();
+   init_wsgpt0f();
+   init_wsgpt0t1f();
+   init_wsgaf();
+   init_wsgt0af();
+   init_wsgt0t1af();
+   init_wsgpt0af();
+   init_wsgpt0t1af();
+    init_2wsgf();
+    init_2wsgt0f();
+    init_2wsgt0t1f();
+    init_2wsgpt0f();
+    init_2wsgpt0t1f();
+    init_2wsgaf();
+    init_2wsgt0af();
+    init_2wsgt0t1af();
+    init_2wsgpt0af();
+    init_2wsgpt0t1af();
 
    init_g();
    init_t0();
    init_t0t1();
    init_gt0();
    init_gt0t1();
+    init_2g();
+    init_2t0();
+    init_2t0t1();
+    init_2gt0();
+    init_2gt0t1();
 }
 
 
 void fxPrintSetupFlags(char *msg, GLuint flags )
 {
-   fprintf(stderr, "%s(%x): %s%s%s%s%s\n",
+   fprintf(stderr, "%s(%x): %s%s%s%s%s%s%s%s\n",
           msg,
           (int)flags,
-          (flags & SETUP_XYZW)     ? " xyzw," : "", 
-          (flags & SETUP_SNAP)     ? " snap," : "", 
+          (flags & SETUP_XYZW)     ? " xyzw," : "",
+          (flags & SETUP_SNAP)     ? " snap," : "",
           (flags & SETUP_RGBA)     ? " rgba," : "",
           (flags & SETUP_TMU0)     ? " tex-0," : "",
-          (flags & SETUP_TMU1)     ? " tex-1," : "");
+          (flags & SETUP_TMU1)     ? " tex-1," : "",
+          (flags & SETUP_PSIZ)     ? " psiz," : "",
+          (flags & SETUP_SPEC)     ? " spec," : "",
+          (flags & SETUP_FOGC)     ? " fog," : "");
 }
 
 
@@ -290,7 +703,7 @@ void fxCheckTexSizes( GLcontext *ctx )
 }
 
 
-void fxBuildVertices( GLcontext *ctx, GLuint start, GLuint count,
+void fxBuildVertices( GLcontext *ctx, GLuint start, GLuint end,
                        GLuint newinputs )
 {
    fxMesaContext fxMesa = FX_CONTEXT( ctx );
@@ -299,15 +712,21 @@ void fxBuildVertices( GLcontext *ctx, GLuint start, GLuint count,
    if (!newinputs)
       return;
 
-   if (newinputs & VERT_BIT_CLIP) {
-      setup_tab[fxMesa->SetupIndex].emit( ctx, start, count, v );   
+   if (newinputs & VERT_BIT_POS) {
+      setup_tab[fxMesa->SetupIndex].emit( ctx, start, end, v );
    } else {
       GLuint ind = 0;
 
       if (newinputs & VERT_BIT_COLOR0)
         ind |= SETUP_RGBA;
-      
-      if (newinputs & VERT_BIT_TEX0) 
+
+      if (newinputs & VERT_BIT_COLOR1)
+        ind |= SETUP_SPEC;
+
+      if (newinputs & VERT_BIT_FOG)
+        ind |= SETUP_FOGC;
+
+      if (newinputs & VERT_BIT_TEX0)
         ind |= SETUP_TMU0;
 
       if (newinputs & VERT_BIT_TEX1)
@@ -319,7 +738,7 @@ void fxBuildVertices( GLcontext *ctx, GLuint start, GLuint count,
       ind &= fxMesa->SetupIndex;
 
       if (ind) {
-        setup_tab[ind].emit( ctx, start, count, v );   
+        setup_tab[ind].emit( ctx, start, end, v );
       }
    }
 }
@@ -339,18 +758,28 @@ void fxChooseVertexState( GLcontext *ctx )
 
    if (ctx->Texture._EnabledUnits & 0x2) {
       if (ctx->Texture._EnabledUnits & 0x1) {
-        ind |= SETUP_TMU1|SETUP_TMU0;
-      }
-      else {
-        fxMesa->tmu_source[0] = 1;
-        fxMesa->tmu_source[1] = 0;
-        ind |= SETUP_TMU0;
+        ind |= SETUP_TMU1;
       }
+      ind |= SETUP_TMU0;
+      fxMesa->tmu_source[0] = 1;
+      fxMesa->tmu_source[1] = 0;
    }
    else if (ctx->Texture._EnabledUnits & 0x1) {
       ind |= SETUP_TMU0;
    }
-   
+
+   if (ctx->_TriangleCaps & DD_POINT_ATTEN) {
+      ind |= SETUP_PSIZ;
+   }
+
+   if (ctx->_TriangleCaps & DD_SEPARATE_SPECULAR) {
+      ind |= SETUP_SPEC;
+   }
+
+   if (ctx->Fog.FogCoordinateSource == GL_FOG_COORDINATE_EXT) {
+      ind |= SETUP_FOGC;
+   }
+
    fxMesa->SetupIndex = ind;
 
    if (ctx->_TriangleCaps & (DD_TRI_LIGHT_TWOSIDE|DD_TRI_UNFILLED)) {
@@ -358,7 +787,7 @@ void fxChooseVertexState( GLcontext *ctx )
       tnl->Driver.Render.CopyPV = copy_pv_extras;
    } else {
       tnl->Driver.Render.Interp = setup_tab[ind].interp;
-      tnl->Driver.Render.CopyPV = copy_pv;
+      tnl->Driver.Render.CopyPV = setup_tab[ind].copy_pv;
    }
 
    if (setup_tab[ind].vertex_format != fxMesa->stw_hint_state) {
@@ -391,11 +820,6 @@ void fxFreeVB( GLcontext *ctx )
       ALIGN_FREE(fxMesa->verts);
       fxMesa->verts = 0;
    }
-
-   if (fxMesa->UbyteColor.Ptr) {
-      ALIGN_FREE((void *)fxMesa->UbyteColor.Ptr);
-      fxMesa->UbyteColor.Ptr = 0;
-   }
 }
 #else