Added optional support for the Katmai assembly. It doesn't work so me, so I'mtdfx-2-1-20000711-freeze

not turning it on.

Improved the lock handling in the DRI cases. This results in ~10% speedup.

3 files changed, 166 insertions, 32 deletions

diff --git a/xc/extras/Mesa/src/FX/fxglidew.c b/xc/extras/Mesa/src/FX/fxglidew.c
index 790e5c923..57b35ad4d 100644
--- a/xc/extras/Mesa/src/FX/fxglidew.c
+++ b/xc/extras/Mesa/src/FX/fxglidew.c
@@ -55,6 +55,13 @@
 #include <stdlib.h>
 #include <string.h>
 
+
+grStencilFunc_t grStencilFuncPtr = NULL;
+grStencilMask_t grStencilMaskPtr = NULL;
+grStencilOp_t grStencilOpPtr = NULL;
+grBufferClearExt_t grBufferClearExtPtr = NULL;
+
+
 FxI32 FX_grGetInteger_NoLock(FxU32 pname)
 {
 #if !defined(FX_GLIDE3)
@@ -71,6 +78,8 @@ FxI32 FX_grGetInteger_NoLock(FxU32 pname)
     case FX_TEXTURE_ALIGN:
         /* This is a guess from reading the glide3 docs */
         return 8;
+    case FX_ZDEPTH_MAX:
+      return 0xFFFF;
     default:
        if (MESA_VERBOSE&VERBOSE_DRIVER) {
           fprintf(stderr,"Wrong parameter in FX_grGetInteger!\n");
@@ -90,6 +99,12 @@ FxI32 FX_grGetInteger_NoLock(FxU32 pname)
      case FX_TEXTURE_ALIGN:
        grname = pname;
        break;
+     case FX_ZDEPTH_MAX: {
+       int zvals[2];
+
+       grGet(GR_ZDEPTH_MIN_MAX, 8, zvals);
+       return zvals[0];
+     }
      default:
        if (MESA_VERBOSE&VERBOSE_DRIVER) {
           fprintf(stderr,"Wrong parameter in FX_grGetInteger!\n");
@@ -154,7 +169,17 @@ extern FxU32 FX_grTexMaxAddress(GrChipID_t tmu) {
 FxBool FX_grSstControl(FxU32 code)
 {
 #if defined(FX_GLIDE3)
-  (void) code;
+  /* The glide 3 sources call for grEnable/grDisable to be called in exchange
+   * for grSstControl. */
+  switch(code) {
+    case GR_CONTROL_ACTIVATE:
+      grEnable(GR_PASSTHRU);
+      break;
+    case GR_CONTROL_DEACTIVATE:
+      grDisable(GR_PASSTHRU);
+      break;
+  }
+  /* Appearently GR_CONTROL_RESIZE can be ignored. */
   return 1;  /* OK? */
 #else
   FxU32 result;
@@ -225,12 +250,12 @@ void FX_grGlideGetVersion(char *buf)
 
 void FX_grSstPerfStats(GrSstPerfStats_t *st)
 {
-  /* ToDo */
-  st->pixelsIn = 0;
-  st->chromaFail = 0;
-  st->zFuncFail = 0;
-  st->aFuncFail = 0;
-  st->pixelsOut = 0;
+  FxI32 n;
+  grGet(GR_STATS_PIXELS_IN, 4, &n);              st->pixelsIn = n;
+  grGet(GR_STATS_PIXELS_CHROMA_FAIL, 4, &n);     st->chromaFail = n;
+  grGet(GR_STATS_PIXELS_DEPTHFUNC_FAIL, 4, &n);  st->zFuncFail = n;
+  grGet(GR_STATS_PIXELS_AFUNC_FAIL, 4, &n);      st->aFuncFail = n;
+  grGet(GR_STATS_PIXELS_OUT, 4, &n);             st->pixelsOut = n;
 }
 
 void FX_grAADrawLine(GrVertex *a,GrVertex *b)
@@ -385,6 +410,10 @@ int FX_grSstQueryHardware(GrHwConfiguration *c)
    return i;
 } 
 
+
+#endif  /* FX_GLIDE3 */
+
+/* It appears to me that this function is needed either way. */
 FX_GrContext_t FX_grSstWinOpen( FxU32                hWnd,
                                 GrScreenResolution_t screen_resolution,
                                 GrScreenRefresh_t    refresh_rate,
@@ -403,6 +432,7 @@ FX_GrContext_t FX_grSstWinOpen( FxU32                hWnd,
                      nColBuffers,
                      nAuxBuffers );
    
+   /*
    fprintf(stderr, 
            "grSstWinOpen( win %d res %d ref %d fmt %d\n"
            "              org %d ncol %d naux %d )\n"
@@ -415,13 +445,11 @@ FX_GrContext_t FX_grSstWinOpen( FxU32                hWnd,
            nColBuffers,
            nAuxBuffers,
            i);
+   */
    END_BOARD_LOCK();
    return i;
 }
 
-
-
-#endif 
 #else
 
 /*
diff --git a/xc/extras/Mesa/src/FX/fxglidew.h b/xc/extras/Mesa/src/FX/fxglidew.h
index 28556d8c5..f188c263d 100644
--- a/xc/extras/Mesa/src/FX/fxglidew.h
+++ b/xc/extras/Mesa/src/FX/fxglidew.h
@@ -47,7 +47,51 @@
 #define __FX_GLIDE_WARPER__
 
 #include <glide.h>
+#include <g3ext.h>
 
+/*
+ * These are glide extension definitions.  These are not
+ * defined in glide.h.  They should really be defined in
+ * g3ext.h, but they are not.
+ */
+#if 0
+FX_ENTRY void FX_CALL 
+grStencilFunc(GrCmpFnc_t fnc, GrStencil_t ref, GrStencil_t mask);
+
+FX_ENTRY void FX_CALL 
+grStencilMask(GrStencil_t write_mask);
+
+FX_ENTRY void FX_CALL 
+grStencilOp(
+            GrStencilOp_t stencil_fail,
+            GrStencilOp_t depth_fail,
+            GrStencilOp_t depth_pass);
+
+FX_ENTRY void FX_CALL 
+grBufferClearExt(
+              GrColor_t color,
+              GrAlpha_t alpha,
+              FxU32     depth,
+              GrStencil_t stencil);
+#endif
+
+
+typedef void (*grStencilFunc_t)(GrCmpFnc_t fnc, GrStencil_t ref, GrStencil_t mask);
+typedef void (*grStencilMask_t)(GrStencil_t write_mask);
+typedef void (*grStencilOp_t)(GrStencilOp_t stencil_fail, GrStencilOp_t depth_fail, GrStencilOp_t depth_pass);
+typedef void (*grBufferClearExt_t)(GrColor_t color, GrAlpha_t alpha, FxU32 depth, GrStencil_t stencil);
+
+extern grStencilFunc_t grStencilFuncPtr;
+extern grStencilMask_t grStencilMaskPtr;
+extern grStencilOp_t grStencilOpPtr;
+extern grBufferClearExt_t grBufferClearExtPtr;
+
+
+FX_ENTRY void FX_CALL
+grEnable(GrEnableMode_t mode);
+
+FX_ENTRY void FX_CALL
+grEnable(GrEnableMode_t mode);
 /* 
  * General context: 
  */
@@ -74,12 +118,13 @@
 	#define FX_PENDING_BUFFERSWAPS          GR_PENDING_BUFFERSWAPS
 	#define FX_TEXTURE_ALIGN		GR_TEXTURE_ALIGN
 #endif
+#define FX_ZDEPTH_MAX 0x100
 
 /*
  * Genral warper functions for Glide2/Glide3:
  */ 
-extern FxI32 FX_grGetInteger(FxU32 pname);
 extern FxI32 FX_grGetInteger_NoLock(FxU32 pname);
+extern FxI32 FX_grGetInteger(FxU32 pname);
 
 /*
  * Glide2 emulation on Glide3:
@@ -253,7 +298,6 @@ typedef struct
 
 #endif
 
-
 /*
  * Glide2 functions for Glide3
  */
@@ -334,10 +378,11 @@ typedef struct
 /*
  * Draw triangle
  */
+#define FX_grDrawTriangle_NoLock(a,b,c) grDrawTriangle(a,b,c)
 #define FX_grDrawTriangle(a,b,c)	\
   do {					\
     BEGIN_CLIP_LOOP();			\
-    grDrawTriangle(a,b,c);		\
+    FX_grDrawTriangle_NoLock(a,b,c);	\
     END_CLIP_LOOP();			\
   } while (0)
 
@@ -521,17 +566,19 @@ extern FX_GrContext_t FX_grSstWinOpen( FxU32                hWnd,
                                        int                  nAuxBuffers);
 
 
+#define FX_grDrawLine_NoLock(v1, v2) grDrawLine(v1, v2)
 #define FX_grDrawLine(v1, v2)	\
   do {				\
     BEGIN_CLIP_LOOP();		\
-    grDrawLine(v1, v2);		\
+    FX_grDrawLine_NoLock(v1, v2);\
     END_CLIP_LOOP();		\
   } while (0)
 
+#define FX_grDrawPoint_NoLock(p) grDrawPoint(p)
 #define FX_grDrawPoint(p)	\
   do {				\
     BEGIN_CLIP_LOOP();		\
-    grDrawPoint(p);		\
+    FX_grDrawPoint_NoLock(p);	\
     END_CLIP_LOOP();		\
   } while (0)
 
@@ -567,6 +614,55 @@ extern void FX_grDrawPolygonVertexList(int n, GrVertex *v);
     END_CLIP_LOOP();			\
   } while (0)
 
+#define FX_grBufferClearExt(c, a, d, s)	\
+  do {					\
+    BEGIN_CLIP_LOOP();			\
+    (*grBufferClearExtPtr)(c, a, d, s);	\
+    END_CLIP_LOOP();			\
+  } while (0)
+
+/*
+ * Enable/Disable
+ */
+#define FX_grEnable(m)                   \
+  do {				         \
+    BEGIN_BOARD_LOCK();		         \
+    grEnable(m);                         \
+    END_BOARD_LOCK();		         \
+  } while (0)
+
+#define FX_grDisable(m)                  \
+  do {				         \
+    BEGIN_BOARD_LOCK();		         \
+    grDisable(m);                        \
+    END_BOARD_LOCK();		         \
+  } while (0)
+
+/*
+ * Stencil operations.
+ */
+#define FX_grStencilFunc(fnc, ref, mask) \
+  do {				         \
+    BEGIN_BOARD_LOCK();		         \
+    (*grStencilFuncPtr)((fnc), (ref), (mask)); \
+    END_BOARD_LOCK();		         \
+  } while (0)
+
+#define FX_grStencilMask(write_mask)     \
+  do {				         \
+    BEGIN_BOARD_LOCK();		         \
+    (*grStencilMaskPtr)(write_mask);           \
+    END_BOARD_LOCK();		         \
+  } while (0)
+
+
+#define FX_grStencilOp(stencil_fail, depth_fail, depth_pass) \
+  do {				                             \
+    BEGIN_BOARD_LOCK();		                             \
+    (*grStencilOpPtr)((stencil_fail), (depth_fail), (depth_pass)); \
+    END_BOARD_LOCK();		                             \
+  } while (0)
+
 #define FX_grDepthMask(m)	\
   do {				\
     BEGIN_BOARD_LOCK();		\
@@ -821,9 +917,18 @@ extern FxU32 FX_grTexTextureMemRequired(FxU32 evenOdd, GrTexInfo *info);
 
 #define FX_grGlideShutdown()		\
   do {					\
-    BEGIN_CLIP_LOOP();			\
+    BEGIN_BOARD_LOCK();			\
     grGlideShutdown();			\
-    END_CLIP_LOOP();			\
+    END_BOARD_LOCK();			\
+  } while (0)
+
+#define FX_grTexLodBiasValue_NoLock(t, v) grTexLodBiasValue(t, v)
+
+#define FX_grTexLodBiasValue(t, v)	\
+  do {					\
+    BEGIN_BOARD_LOCK();			\
+    grTexLodBiasValue(t, v);	\
+    END_BOARD_LOCK();			\
   } while (0)
 
 #define FX_grGlideInit_NoLock grGlideInit
diff --git a/xc/extras/Mesa/src/FX/fxrender.c b/xc/extras/Mesa/src/FX/fxrender.c
index 8ce4b0618..9c5329fff 100644
--- a/xc/extras/Mesa/src/FX/fxrender.c
+++ b/xc/extras/Mesa/src/FX/fxrender.c
@@ -533,43 +533,44 @@ static INLINE void fxSafeClippedTriangle3( struct vertex_buffer *VB,
   do {							\
     GLuint e;						\
     for(e=start;e<=count;e++) {				\
-      FX_grDrawPoint((GrVertex *)gWin[elt[e]].f);	\
+      FX_grDrawPoint_NoLock((GrVertex *)gWin[elt[e]].f);\
     }							\
   } while (0)
 
 #define RENDER_LINE( i1, i )						\
   do {									\
     GLuint e1 = elt[i1], e = elt[i];					\
-    FX_grDrawLine((GrVertex *)gWin[e1].f, (GrVertex *)gWin[e].f);	\
+    FX_grDrawLine_NoLock((GrVertex *)gWin[e1].f, (GrVertex *)gWin[e].f);\
   } while (0)
 
 
 #define RENDER_TRI( i2, i1, i, pv, parity)		\
   do {							\
     GLuint e2 = elt[i2], e1 = elt[i1], e = elt[i];	\
-    if (parity) {GLuint tmp = e2; e2 = e1; e1 = tmp;}	\
-    FX_grDrawTriangle((GrVertex *)gWin[e2].f,		\
-		 (GrVertex *)gWin[e1].f,		\
-		 (GrVertex *)gWin[e].f);		\
+    if (parity) {GLuint tmp = e2; e2 = e1; e1 = tmp;} 	\
+    FX_grDrawTriangle_NoLock((GrVertex *)gWin[e2].f,	\
+		             (GrVertex *)gWin[e1].f,	\
+		             (GrVertex *)gWin[e].f);	\
   } while (0)
 
 
 #define RENDER_QUAD( i3, i2, i1, i, pv)				\
   do {								\
     GLuint e3 = elt[i3], e2 = elt[i2], e1 = elt[i1], e = elt[i];\
-    FX_grDrawTriangle((GrVertex *)gWin[e3].f,			\
-		 (GrVertex *)gWin[e2].f,			\
-		 (GrVertex *)gWin[e].f);			\
-    FX_grDrawTriangle((GrVertex *)gWin[e2].f,			\
-		 (GrVertex *)gWin[e1].f,			\
-		 (GrVertex *)gWin[e].f);			\
+    FX_grDrawTriangle_NoLock((GrVertex *)gWin[e3].f,		\
+		             (GrVertex *)gWin[e2].f,		\
+		             (GrVertex *)gWin[e].f);		\
+    FX_grDrawTriangle_NoLock((GrVertex *)gWin[e2].f,		\
+		             (GrVertex *)gWin[e1].f,		\
+		             (GrVertex *)gWin[e].f);		\
   } while (0)
 
 #define LOCAL_VARS				\
    fxVertex *gWin = FX_DRIVER_DATA(VB)->verts;	\
    const GLuint *elt = VB->EltPtr->data;
 
-#define INIT(x)  
+#define INIT(x)  BEGIN_CLIP_LOOP();
+#define POSTFIX  END_CLIP_LOOP();
 
 #define TAG(x) x##_fx_smooth_indirect
 #undef  RVB_COLOR
@@ -722,8 +723,8 @@ render_func **fxDDChooseRenderVBTables(GLcontext *ctx)
     return null_tables;  
 
   switch (fxMesa->render_index) {
-  case FX_FLAT:
-    return fxDDRenderVBFlat_tables;
+/*    case FX_FLAT: */
+/*      return fxDDRenderVBFlat_tables; */
   case 0:
     return fxDDRenderVBSmooth_tables; 
   default: