added GL_ARB_point_sprite.

color masking correctly implemented for 15bpp.
small fixes for textures and tri engine.

5 files changed, 119 insertions, 44 deletions

diff --git a/src/mesa/drivers/glide/fxdd.c b/src/mesa/drivers/glide/fxdd.c
index 606c7f158f..a08ad97fde 100644
--- a/src/mesa/drivers/glide/fxdd.c
+++ b/src/mesa/drivers/glide/fxdd.c
@@ -90,11 +90,11 @@ GLuint FX_rgb_scale_6[64] = {
  */
 static void fxDisableColor (fxMesaContext fxMesa)
 {
- if (fxMesa->colDepth != 16) {
-    /* 32bpp mode or 15bpp mode */
+ if (fxMesa->colDepth == 32) {
+    /* 32bpp mode */
     fxMesa->Glide.grColorMaskExt(FXFALSE, FXFALSE, FXFALSE, FXFALSE);
  } else {
-    /* 16 bpp mode */
+    /* 15/16 bpp mode */
     grColorMask(FXFALSE, FXFALSE);
  }
 }
@@ -185,7 +185,7 @@ static void fxDDClear( GLcontext *ctx,
    /*
     * As per GL spec, color masking should be obeyed when clearing
     */
-   if (ctx->Visual.greenBits != 8 && ctx->Visual.greenBits != 5) {
+   if (ctx->Visual.greenBits != 8) {
       /* can only do color masking if running in 24/32bpp on Napalm */
       if (ctx->Color.ColorMask[RCOMP] != ctx->Color.ColorMask[GCOMP] ||
           ctx->Color.ColorMask[GCOMP] != ctx->Color.ColorMask[BCOMP]) {
@@ -1113,15 +1113,15 @@ fxDDDrawPixels8888 (GLcontext * ctx, GLint x, GLint y,
        ctx->Pixel.ZoomY != 1.0F ||
        (ctx->_ImageTransferState & (IMAGE_SCALE_BIAS_BIT|
 				    IMAGE_MAP_COLOR_BIT)) ||
-       ctx->Color.AlphaEnabled ||
+       /*ctx->Color.AlphaEnabled ||*/
        ctx->Depth.Test ||
        ctx->Fog.Enabled ||
        ctx->Scissor.Enabled ||
        ctx->Stencil.Enabled ||
-       !ctx->Color.ColorMask[0] ||
+       /*!ctx->Color.ColorMask[0] ||
        !ctx->Color.ColorMask[1] ||
        !ctx->Color.ColorMask[2] ||
-       !ctx->Color.ColorMask[3] ||
+       !ctx->Color.ColorMask[3] ||*/
        ctx->Color.ColorLogicOpEnabled ||
        ctx->Texture._EnabledUnits ||
        ctx->Depth.OcclusionTest ||
@@ -1303,11 +1303,11 @@ fxDDInitFxMesaContext(fxMesaContext fxMesa)
 
    fxMesa->unitsState.stencilWriteMask = 0xff;
 
-   if (fxMesa->colDepth != 16) {
-      /* 32bpp mode or 15bpp mode */
+   if (fxMesa->colDepth == 32) {
+      /* 32bpp */
       fxMesa->Glide.grColorMaskExt(FXTRUE, FXTRUE, FXTRUE, fxMesa->haveHwAlpha);
    } else {
-      /* 16 bpp mode */
+      /* 15/16 bpp mode */
       grColorMask(FXTRUE, fxMesa->haveHwAlpha);
    }
 
@@ -1357,6 +1357,7 @@ fxDDInitFxMesaContext(fxMesaContext fxMesa)
         textureLevels++;
     } while ((textureSize >>= 0x1) & 0x7ff);
     ctx->Const.MaxTextureLevels = textureLevels;
+    ctx->Const.MaxTextureLodBias = /*textureLevels - 1*/8; /* Glide bug */
 #if FX_RESCALE_BIG_TEXURES_HACK
     fxMesa->textureMaxLod = textureLevels - 1;
     if ((env = getenv("MESA_FX_MAXLOD")) != NULL) {
@@ -1448,6 +1449,7 @@ fxDDInitExtensions(GLcontext * ctx)
    _mesa_enable_extension(ctx, "GL_EXT_secondary_color");
 #endif
 
+   _mesa_enable_extension(ctx, "GL_ARB_point_sprite");
    _mesa_enable_extension(ctx, "GL_EXT_point_parameters");
    _mesa_enable_extension(ctx, "GL_EXT_paletted_texture");
    _mesa_enable_extension(ctx, "GL_EXT_texture_lod_bias");
@@ -1582,7 +1584,7 @@ fx_check_IsInHardware(GLcontext * ctx)
    }
 #endif
 
-   if ((fxMesa->colDepth == 16) &&
+   if ((fxMesa->colDepth != 32) &&
        ((ctx->Color.ColorMask[RCOMP] != ctx->Color.ColorMask[GCOMP]) ||
         (ctx->Color.ColorMask[GCOMP] != ctx->Color.ColorMask[BCOMP]))) {
       return FX_FALLBACK_COLORMASK;
diff --git a/src/mesa/drivers/glide/fxddtex.c b/src/mesa/drivers/glide/fxddtex.c
index f433c573d4..4c498b59cc 100644
--- a/src/mesa/drivers/glide/fxddtex.c
+++ b/src/mesa/drivers/glide/fxddtex.c
@@ -171,8 +171,9 @@ fxDDTexEnv(GLcontext * ctx, GLenum target, GLenum pname,
    /* apply any lod biasing right now */
    if (pname == GL_TEXTURE_LOD_BIAS_EXT) {
       GLfloat bias = *param;
-      CLAMP_SELF(bias, -8.0, 7.75);
-      
+      CLAMP_SELF(bias, -ctx->Const.MaxTextureLodBias,
+                        ctx->Const.MaxTextureLodBias - 0.25);
+
       grTexLodBiasValue(GR_TMU0, bias);
 
       if (fxMesa->haveTwoTMUs) {
@@ -296,6 +297,7 @@ fxDDTexParam(GLcontext * ctx, GLenum target, struct gl_texture_object *tObj,
       case GL_MIRRORED_REPEAT:
          ti->sClamp = GR_TEXTURECLAMP_MIRROR_EXT;
          break;
+      case GL_CLAMP_TO_BORDER: /* no-no, but don't REPEAT, either */
       case GL_CLAMP_TO_EDGE: /* CLAMP discarding border */
       case GL_CLAMP:
 	 ti->sClamp = GR_TEXTURECLAMP_CLAMP;
@@ -314,6 +316,7 @@ fxDDTexParam(GLcontext * ctx, GLenum target, struct gl_texture_object *tObj,
       case GL_MIRRORED_REPEAT:
          ti->tClamp = GR_TEXTURECLAMP_MIRROR_EXT;
          break;
+      case GL_CLAMP_TO_BORDER: /* no-no, but don't REPEAT, either */
       case GL_CLAMP_TO_EDGE: /* CLAMP discarding border */
       case GL_CLAMP:
 	 ti->tClamp = GR_TEXTURECLAMP_CLAMP;
@@ -1591,6 +1594,7 @@ fxDDCompressedTexImage2D (GLcontext *ctx, GLenum target,
        *    our data aligned inside a 8:1 rectangle.
        * 3) just in case if MIN("s", "t") gets overflowed with GL_REPEAT,
        *    we replicate the data over the padded area.
+       * For now, we take 2) + 3) but texelfetchers will be wrong!
        */
       GLuint srcRowStride = _mesa_compressed_row_stride(internalFormat, width);
 
@@ -1601,7 +1605,7 @@ fxDDCompressedTexImage2D (GLcontext *ctx, GLenum target,
                                destRowStride, (mml->height+3) / 4,
                                1, data, srcRowStride,
                                texImage->Data);
-
+      ti->padded = GL_TRUE;
    } else {
       MEMCPY(texImage->Data, data, texImage->CompressedSize);
    }
diff --git a/src/mesa/drivers/glide/fxdrv.h b/src/mesa/drivers/glide/fxdrv.h
index af4765d7a9..379851131f 100644
--- a/src/mesa/drivers/glide/fxdrv.h
+++ b/src/mesa/drivers/glide/fxdrv.h
@@ -294,6 +294,8 @@ typedef struct tfxTexInfo_t
 
    GLboolean fixedPalette;
    GLboolean validated;
+
+   GLboolean padded;
 }
 tfxTexInfo;
 
diff --git a/src/mesa/drivers/glide/fxsetup.c b/src/mesa/drivers/glide/fxsetup.c
index 0d900d29f4..d486a5eaf1 100644
--- a/src/mesa/drivers/glide/fxsetup.c
+++ b/src/mesa/drivers/glide/fxsetup.c
@@ -156,10 +156,7 @@ fxTexValidate(GLcontext * ctx, struct gl_texture_object *tObj)
       FX_smallLodLog2(ti->info) = FX_largeLodLog2(ti->info);
 
    /* this is necessary because of fxDDCompressedTexImage2D */
-   if ((ti->info.format == GR_TEXFMT_ARGB_CMP_FXT1) ||
-       (ti->info.format == GR_TEXFMT_ARGB_CMP_DXT1) ||
-       (ti->info.format == GR_TEXFMT_ARGB_CMP_DXT3) ||
-       (ti->info.format == GR_TEXFMT_ARGB_CMP_DXT5)) {
+   if (ti->padded) {
       struct gl_texture_image *texImage = tObj->Image[0][minl];
       tfxMipMapLevel *mml = FX_MIPMAP_DATA(texImage);
       if (mml->wScale != 1 || mml->hScale != 1) {
@@ -1784,15 +1781,15 @@ fxSetupColorMask(GLcontext * ctx)
 {
    fxMesaContext fxMesa = FX_CONTEXT(ctx);
 
-   if (fxMesa->colDepth != 16) {
-      /* 32bpp mode or 15bpp mode */
+   if (fxMesa->colDepth == 32) {
+      /* 32bpp mode */
       fxMesa->Glide.grColorMaskExt(ctx->Color.ColorMask[RCOMP],
                                    ctx->Color.ColorMask[GCOMP],
                                    ctx->Color.ColorMask[BCOMP],
                                    ctx->Color.ColorMask[ACOMP] && fxMesa->haveHwAlpha);
    }
    else {
-      /* 16 bpp mode */
+      /* 15/16 bpp mode */
       grColorMask(ctx->Color.ColorMask[RCOMP] |
                   ctx->Color.ColorMask[GCOMP] |
                   ctx->Color.ColorMask[BCOMP],
diff --git a/src/mesa/drivers/glide/fxtris.c b/src/mesa/drivers/glide/fxtris.c
index 6620eb678a..e2b6c45a42 100644
--- a/src/mesa/drivers/glide/fxtris.c
+++ b/src/mesa/drivers/glide/fxtris.c
@@ -274,6 +274,63 @@ static void fx_draw_point( fxMesaContext fxMesa,
 #endif
 #define __GL_COSF cos
 #define __GL_SINF sin
+static void fx_draw_point_sprite ( fxMesaContext fxMesa,
+				   GrVertex *v0, GLfloat psize )
+{
+ const GLcontext *ctx = fxMesa->glCtx;
+
+ GLfloat radius;
+ GrVertex _v_[4];
+ GLuint ts0 = fxMesa->tmu_source[0];
+ GLuint ts1 = fxMesa->tmu_source[1];
+ GLfloat w = v0->oow;
+ GLfloat u0scale = fxMesa->s0scale * w;
+ GLfloat v0scale = fxMesa->t0scale * w;
+ GLfloat u1scale = fxMesa->s1scale * w;
+ GLfloat v1scale = fxMesa->t1scale * w;
+
+ radius = psize / 2.;
+ _v_[0] = *v0;
+ _v_[1] = *v0;
+ _v_[2] = *v0;
+ _v_[3] = *v0;
+ /* CLIP_LOOP ?!? */
+ /* point coverage? */
+ /* we don't care about culling here (see fxSetupCull) */
+
+ _v_[0].x -= radius;
+ _v_[0].y += radius;
+ _v_[1].x += radius;
+ _v_[1].y += radius;
+ _v_[2].x += radius;
+ _v_[2].y -= radius;
+ _v_[3].x -= radius;
+ _v_[3].y -= radius;
+
+ if (ctx->Point.CoordReplace[ts0]) {
+    _v_[0].tmuvtx[0].sow = 0;
+    _v_[0].tmuvtx[0].tow = 0;
+    _v_[1].tmuvtx[0].sow = u0scale;
+    _v_[1].tmuvtx[0].tow = 0;
+    _v_[2].tmuvtx[0].sow = u0scale;
+    _v_[2].tmuvtx[0].tow = v0scale;
+    _v_[3].tmuvtx[0].sow = 0;
+    _v_[3].tmuvtx[0].tow = v0scale;
+ }
+ if (ctx->Point.CoordReplace[ts1]) {
+    _v_[0].tmuvtx[1].sow = 0;
+    _v_[0].tmuvtx[1].tow = 0;
+    _v_[1].tmuvtx[1].sow = u1scale;
+    _v_[1].tmuvtx[1].tow = 0;
+    _v_[2].tmuvtx[1].sow = u1scale;
+    _v_[2].tmuvtx[1].tow = v1scale;
+    _v_[3].tmuvtx[1].sow = 0;
+    _v_[3].tmuvtx[1].tow = v1scale;
+ }
+
+ grDrawVertexArrayContiguous(GR_TRIANGLE_FAN, 4, _v_, sizeof(GrVertex));
+}
+
 static void fx_draw_point_wide ( fxMesaContext fxMesa,
 			         GrVertex *v0 )
 {
@@ -285,6 +342,11 @@ static void fx_draw_point_wide ( fxMesaContext fxMesa,
  const GLcontext *ctx = fxMesa->glCtx;
  const GLfloat psize = (ctx->_TriangleCaps & DD_POINT_ATTEN) ? v0->psize : ctx->Point.Size;
 
+ if (ctx->Point.PointSprite) {
+    fx_draw_point_sprite(fxMesa, v0, psize);
+    return;
+ }
+
  _v_[0] = v0;
  _v_[1] = &vtxB;
  _v_[2] = &vtxC;
@@ -355,6 +417,11 @@ static void fx_draw_point_wide_aa ( fxMesaContext fxMesa,
  const GLcontext *ctx = fxMesa->glCtx;
  const GLfloat psize = (ctx->_TriangleCaps & DD_POINT_ATTEN) ? v0->psize : ctx->Point.Size;
 
+ if (ctx->Point.PointSprite) {
+    fx_draw_point_sprite(fxMesa, v0, psize);
+    return;
+ }
+
  radius = psize / 2.;
  n = IROUND(psize * 2); /* radius x 4 */
  if (n < 4) n = 4;
@@ -905,27 +972,16 @@ static void fx_render_vb_quads( GLcontext *ctx,
 
    INIT(GL_QUADS);
 
-#if 1
-   /* [dBorca] Hack alert:
-    * since VTX-0-2 we get here with start = 0, count = 2,
-    * causing around 4 billion triangles. Revise this after
-    * a while
-    */
-   if (count < 3) {
-      return;
-   }
-#endif
-
-   for (i = start ; i < count-3 ; i += 4 ) {
+   for (i = start + 3 ; i < count ; i += 4 ) {
 #define VERT(x) (fxVB + (x))
       GrVertex *_v_[4];
-      _v_[0] = VERT(i+3);
-      _v_[1] = VERT(i+0);
-      _v_[2] = VERT(i+1);
-      _v_[3] = VERT(i+2);
+      _v_[0] = VERT(i);
+      _v_[1] = VERT(i-3);
+      _v_[2] = VERT(i-2);
+      _v_[3] = VERT(i-1);
       grDrawVertexArray(GR_TRIANGLE_FAN, 4, _v_);
-      /*grDrawTriangle( VERT(i),   VERT(i+1), VERT(i+3) );*/
-      /*grDrawTriangle( VERT(i+1), VERT(i+2), VERT(i+3) );*/
+      /*grDrawTriangle( VERT(i-3), VERT(i-2), VERT(i) );*/
+      /*grDrawTriangle( VERT(i-2), VERT(i-1), VERT(i) );*/
 #undef VERT
    }
 }
@@ -1092,12 +1148,22 @@ static void fxRenderClippedPoly( GLcontext *ctx, const GLuint *elts,
 static void fxFastRenderClippedPoly( GLcontext *ctx, const GLuint *elts, 
 				       GLuint n )
 {
+   int i;
    fxMesaContext fxMesa = FX_CONTEXT( ctx );
    GrVertex *vertptr = fxMesa->verts;
-   const GrVertex *start = VERT(elts[0]);
-   int i;
-   for (i = 2 ; i < n ; i++) {
-      grDrawTriangle( start, VERT(elts[i-1]), VERT(elts[i]) );
+   if (n == 3) {
+      grDrawTriangle( VERT(elts[0]), VERT(elts[1]), VERT(elts[2]) );
+   } else if (n <= 32) {
+      GrVertex *newvptr[32];
+      for (i = 0 ; i < n ; i++) {
+         newvptr[i] = VERT(elts[i]);
+      }
+      grDrawVertexArray(GR_TRIANGLE_FAN, n, newvptr);
+   } else {
+      const GrVertex *start = VERT(elts[0]);
+      for (i = 2 ; i < n ; i++) {
+         grDrawTriangle( start, VERT(elts[i-1]), VERT(elts[i]) );
+      }
    }
 }
 
@@ -1196,7 +1262,7 @@ void fxDDChooseRenderState(GLcontext *ctx)
          fx_render_tab_verts[0] = fx_render_pw_verts;
          fx_render_tab_elts[0] = fx_render_pw_elts;
       }
-      tnl->Driver.Render.Points = rast_tab[FX_FALLBACK_BIT].points;
+      tnl->Driver.Render.Points = rast_tab[index|FX_FALLBACK_BIT].points;
    } else {
       fx_render_tab_verts[0] = fx_render_vb_points;
       fx_render_tab_elts[0] = fx_render_points_elts;
@@ -1431,6 +1497,10 @@ void fxDDInitTriFuncs( GLcontext *ctx )
 }
 
 
+/* [dBorca] Hack alert:
+ * doesn't work with blending.
+ * need to take care of stencil.
+ */
 GLboolean fxMultipass_ColorSum (GLcontext *ctx, GLuint pass)
 {
  fxMesaContext fxMesa = FX_CONTEXT(ctx);