X-Git-Url: https://git.libre-soc.org/?a=blobdiff_plain;ds=sidebyside;f=src%2Fmesa%2Fswrast%2Fs_texcombine.c;h=889164b986132e798d2bb902701b6237394e713a;hb=19119517ce00f7710c6cd627c75e7eef765021c2;hp=27c5c15cf16bdee531fadf4f22be9d4c96b80e38;hpb=933f3b13c34c2ed9223755c0e7c7dc22f09d23e8;p=mesa.git

diff --git a/src/mesa/swrast/s_texcombine.c b/src/mesa/swrast/s_texcombine.c
index 27c5c15cf16..889164b9861 100644
--- a/src/mesa/swrast/s_texcombine.c
+++ b/src/mesa/swrast/s_texcombine.c
@@ -29,7 +29,6 @@
 #include "main/colormac.h"
 #include "main/image.h"
 #include "main/imports.h"
-#include "main/macros.h"
 #include "main/pixel.h"
 #include "shader/prog_instruction.h"
 
@@ -37,46 +36,61 @@
 #include "s_texcombine.h"
 
 
-#define MAX_COMBINER_TERMS 4
+/**
+ * Pointer to array of float[4]
+ * This type makes the code below more concise and avoids a lot of casting.
+ */
+typedef float (*float4_array)[4];
+
+
+/**
+ * Return array of texels for given unit.
+ */
+static INLINE float4_array
+get_texel_array(SWcontext *swrast, GLuint unit)
+{
+   return (float4_array) (swrast->TexelBuffer + unit * MAX_WIDTH * 4);
+}
+
 
 
 /**
- * Do texture application for GL_ARB/EXT_texture_env_combine.
- * This function also supports GL_{EXT,ARB}_texture_env_dot3 and
- * GL_ATI_texture_env_combine3.  Since "classic" texture environments are
- * implemented using GL_ARB_texture_env_combine-like state, this same function
- * is used for classic texture environment application as well.
+ * Do texture application for:
+ *  GL_EXT_texture_env_combine
+ *  GL_ARB_texture_env_combine
+ *  GL_EXT_texture_env_dot3
+ *  GL_ARB_texture_env_dot3
+ *  GL_ATI_texture_env_combine3
+ *  GL_NV_texture_env_combine4
+ *  conventional GL texture env modes
  *
  * \param ctx          rendering context
- * \param textureUnit  the texture unit to apply
+ * \param unit         the texture combiner unit
  * \param n            number of fragments to process (span width)
  * \param primary_rgba incoming fragment color array
  * \param texelBuffer  pointer to texel colors for all texture units
  * 
- * \param rgba         incoming colors, which get modified here
+ * \param rgba         incoming/result fragment colors
  */
 static void
-texture_combine( const GLcontext *ctx, GLuint unit, GLuint n,
-                 CONST GLfloat (*primary_rgba)[4],
-                 CONST GLfloat *texelBuffer,
+texture_combine( GLcontext *ctx, GLuint unit, GLuint n,
+                 const float4_array primary_rgba,
+                 const GLfloat *texelBuffer,
                  GLchan (*rgbaChan)[4] )
 {
+   SWcontext *swrast = SWRAST_CONTEXT(ctx);
    const struct gl_texture_unit *textureUnit = &(ctx->Texture.Unit[unit]);
    const struct gl_tex_env_combine_state *combine = textureUnit->_CurrentCombine;
-   const GLfloat (*argRGB [MAX_COMBINER_TERMS])[4];
-   const GLfloat (*argA [MAX_COMBINER_TERMS])[4];
-   const GLfloat RGBmult = (GLfloat) (1 << combine->ScaleShiftRGB);
-   const GLfloat Amult = (GLfloat) (1 << combine->ScaleShiftA);
-   const GLuint numColorArgs = combine->_NumArgsRGB;
-   const GLuint numAlphaArgs = combine->_NumArgsA;
+   float4_array argRGB[MAX_COMBINER_TERMS];
+   float4_array argA[MAX_COMBINER_TERMS];
+   const GLfloat scaleRGB = (GLfloat) (1 << combine->ScaleShiftRGB);
+   const GLfloat scaleA = (GLfloat) (1 << combine->ScaleShiftA);
+   const GLuint numArgsRGB = combine->_NumArgsRGB;
+   const GLuint numArgsA = combine->_NumArgsA;
    GLfloat ccolor[MAX_COMBINER_TERMS][MAX_WIDTH][4]; /* temp color buffers */
    GLfloat rgba[MAX_WIDTH][4];
    GLuint i, term;
 
-   ASSERT(ctx->Extensions.EXT_texture_env_combine ||
-          ctx->Extensions.ARB_texture_env_combine);
-   ASSERT(CONST_SWRAST_CONTEXT(ctx)->_AnyTextureCombine);
-
    for (i = 0; i < n; i++) {
       rgba[i][RCOMP] = CHAN_TO_FLOAT(rgbaChan[i][RCOMP]);
       rgba[i][GCOMP] = CHAN_TO_FLOAT(rgbaChan[i][GCOMP]);
@@ -97,24 +111,23 @@ texture_combine( const GLcontext *ctx, GLuint unit, GLuint n,
    /*
     * Do operand setup for up to 4 operands.  Loop over the terms.
     */
-   for (term = 0; term < numColorArgs; term++) {
+   for (term = 0; term < numArgsRGB; term++) {
       const GLenum srcRGB = combine->SourceRGB[term];
       const GLenum operandRGB = combine->OperandRGB[term];
 
       switch (srcRGB) {
          case GL_TEXTURE:
-            argRGB[term] = (const GLfloat (*)[4])
-               (texelBuffer + unit * (n * 4 * sizeof(GLfloat)));
+            argRGB[term] = get_texel_array(swrast, unit);
             break;
          case GL_PRIMARY_COLOR:
             argRGB[term] = primary_rgba;
             break;
          case GL_PREVIOUS:
-            argRGB[term] = (const GLfloat (*)[4]) rgba;
+            argRGB[term] = rgba;
             break;
          case GL_CONSTANT:
             {
-               GLfloat (*c)[4] = ccolor[term];
+               float4_array c = ccolor[term];
                GLfloat red   = textureUnit->EnvColor[0];
                GLfloat green = textureUnit->EnvColor[1];
                GLfloat blue  = textureUnit->EnvColor[2];
@@ -122,27 +135,27 @@ texture_combine( const GLcontext *ctx, GLuint unit, GLuint n,
                for (i = 0; i < n; i++) {
                   ASSIGN_4V(c[i], red, green, blue, alpha);
                }
-               argRGB[term] = (const GLfloat (*)[4]) ccolor[term];
+               argRGB[term] = ccolor[term];
             }
             break;
 	 /* GL_ATI_texture_env_combine3 allows GL_ZERO & GL_ONE as sources.
 	  */
 	 case GL_ZERO:
             {
-               GLfloat (*c)[4] = ccolor[term];
+               float4_array c = ccolor[term];
                for (i = 0; i < n; i++) {
                   ASSIGN_4V(c[i], 0.0F, 0.0F, 0.0F, 0.0F);
                }
-               argRGB[term] = (const GLfloat (*)[4]) ccolor[term];
+               argRGB[term] = ccolor[term];
             }
             break;
 	 case GL_ONE:
             {
-               GLfloat (*c)[4] = ccolor[term];
+               float4_array c = ccolor[term];
                for (i = 0; i < n; i++) {
                   ASSIGN_4V(c[i], 1.0F, 1.0F, 1.0F, 1.0F);
                }
-               argRGB[term] = (const GLfloat (*)[4]) ccolor[term];
+               argRGB[term] = ccolor[term];
             }
             break;
          default:
@@ -152,17 +165,16 @@ texture_combine( const GLcontext *ctx, GLuint unit, GLuint n,
                ASSERT(srcUnit < ctx->Const.MaxTextureUnits);
                if (!ctx->Texture.Unit[srcUnit]._ReallyEnabled)
                   return;
-               argRGB[term] = (const GLfloat (*)[4])
-                  (texelBuffer + srcUnit * (n * 4 * sizeof(GLfloat)));
+               argRGB[term] = get_texel_array(swrast, srcUnit);
             }
       }
 
       if (operandRGB != GL_SRC_COLOR) {
-         const GLfloat (*src)[4] = argRGB[term];
-         GLfloat (*dst)[4] = ccolor[term];
+         float4_array src = argRGB[term];
+         float4_array dst = ccolor[term];
 
          /* point to new arg[term] storage */
-         argRGB[term] = (const GLfloat (*)[4]) ccolor[term];
+         argRGB[term] = ccolor[term];
 
          switch (operandRGB) {
          case GL_ONE_MINUS_SRC_COLOR:
@@ -174,15 +186,15 @@ texture_combine( const GLcontext *ctx, GLuint unit, GLuint n,
             break;
          case GL_SRC_ALPHA:
             for (i = 0; i < n; i++) {
-               dst[i][RCOMP] = src[i][ACOMP];
-               dst[i][GCOMP] = src[i][ACOMP];
+               dst[i][RCOMP] =
+               dst[i][GCOMP] =
                dst[i][BCOMP] = src[i][ACOMP];
             }
             break;
          case GL_ONE_MINUS_SRC_ALPHA:
             for (i = 0; i < n; i++) {
-               dst[i][RCOMP] = 1.0F - src[i][ACOMP];
-               dst[i][GCOMP] = 1.0F - src[i][ACOMP];
+               dst[i][RCOMP] =
+               dst[i][GCOMP] =
                dst[i][BCOMP] = 1.0F - src[i][ACOMP];
             }
             break;
@@ -193,48 +205,47 @@ texture_combine( const GLcontext *ctx, GLuint unit, GLuint n,
    }
 
    /*
-    * Set up the argA[i] pointers
+    * Set up the argA[term] pointers
     */
-   for (term = 0; term < numAlphaArgs; term++) {
+   for (term = 0; term < numArgsA; term++) {
       const GLenum srcA = combine->SourceA[term];
       const GLenum operandA = combine->OperandA[term];
 
       switch (srcA) {
          case GL_TEXTURE:
-            argA[term] = (const GLfloat (*)[4])
-               (texelBuffer + unit * (n * 4 * sizeof(GLfloat)));
+            argA[term] = get_texel_array(swrast, unit);
             break;
          case GL_PRIMARY_COLOR:
             argA[term] = primary_rgba;
             break;
          case GL_PREVIOUS:
-            argA[term] = (const GLfloat (*)[4]) rgba;
+            argA[term] = rgba;
             break;
          case GL_CONSTANT:
             {
-               GLfloat (*c)[4] = ccolor[term];
+               float4_array c = ccolor[term];
                GLfloat alpha = textureUnit->EnvColor[3];
                for (i = 0; i < n; i++)
                   c[i][ACOMP] = alpha;
-               argA[term] = (const GLfloat (*)[4]) ccolor[term];
+               argA[term] = ccolor[term];
             }
             break;
 	 /* GL_ATI_texture_env_combine3 allows GL_ZERO & GL_ONE as sources.
 	  */
 	 case GL_ZERO:
             {
-               GLfloat (*c)[4] = ccolor[term];
+               float4_array c = ccolor[term];
                for (i = 0; i < n; i++)
                   c[i][ACOMP] = 0.0F;
-               argA[term] = (const GLfloat (*)[4]) ccolor[term];
+               argA[term] = ccolor[term];
             }
             break;
 	 case GL_ONE:
             {
-               GLfloat (*c)[4] = ccolor[term];
+               float4_array c = ccolor[term];
                for (i = 0; i < n; i++)
                   c[i][ACOMP] = 1.0F;
-               argA[term] = (const GLfloat (*)[4]) ccolor[term];
+               argA[term] = ccolor[term];
             }
             break;
          default:
@@ -244,15 +255,14 @@ texture_combine( const GLcontext *ctx, GLuint unit, GLuint n,
                ASSERT(srcUnit < ctx->Const.MaxTextureUnits);
                if (!ctx->Texture.Unit[srcUnit]._ReallyEnabled)
                   return;
-               argA[term] = (const GLfloat (*)[4])
-                  (texelBuffer + srcUnit * (n * 4 * sizeof(GLfloat)));
+               argA[term] = get_texel_array(swrast, srcUnit);
             }
       }
 
       if (operandA == GL_ONE_MINUS_SRC_ALPHA) {
-         const GLfloat (*src)[4] = argA[term];
-         GLfloat (*dst)[4] = ccolor[term];
-         argA[term] = (const GLfloat (*)[4]) ccolor[term];
+         float4_array src = argA[term];
+         float4_array dst = ccolor[term];
+         argA[term] = ccolor[term];
          for (i = 0; i < n; i++) {
             dst[i][ACOMP] = 1.0F - src[i][ACOMP];
          }
@@ -261,24 +271,24 @@ texture_combine( const GLcontext *ctx, GLuint unit, GLuint n,
 
    /* RGB channel combine */
    {
-      const GLfloat (*arg0)[4] = (const GLfloat (*)[4]) argRGB[0];
-      const GLfloat (*arg1)[4] = (const GLfloat (*)[4]) argRGB[1];
-      const GLfloat (*arg2)[4] = (const GLfloat (*)[4]) argRGB[2];
-      const GLfloat (*arg3)[4] = (const GLfloat (*)[4]) argRGB[3];
+      float4_array arg0 = argRGB[0];
+      float4_array arg1 = argRGB[1];
+      float4_array arg2 = argRGB[2];
+      float4_array arg3 = argRGB[3];
 
       switch (combine->ModeRGB) {
       case GL_REPLACE:
          for (i = 0; i < n; i++) {
-            rgba[i][RCOMP] = arg0[i][RCOMP] * RGBmult;
-            rgba[i][GCOMP] = arg0[i][GCOMP] * RGBmult;
-            rgba[i][BCOMP] = arg0[i][BCOMP] * RGBmult;
+            rgba[i][RCOMP] = arg0[i][RCOMP] * scaleRGB;
+            rgba[i][GCOMP] = arg0[i][GCOMP] * scaleRGB;
+            rgba[i][BCOMP] = arg0[i][BCOMP] * scaleRGB;
          }
          break;
       case GL_MODULATE:
          for (i = 0; i < n; i++) {
-            rgba[i][RCOMP] = arg0[i][RCOMP] * arg1[i][RCOMP] * RGBmult;
-            rgba[i][GCOMP] = arg0[i][GCOMP] * arg1[i][GCOMP] * RGBmult;
-            rgba[i][BCOMP] = arg0[i][BCOMP] * arg1[i][BCOMP] * RGBmult;
+            rgba[i][RCOMP] = arg0[i][RCOMP] * arg1[i][RCOMP] * scaleRGB;
+            rgba[i][GCOMP] = arg0[i][GCOMP] * arg1[i][GCOMP] * scaleRGB;
+            rgba[i][BCOMP] = arg0[i][BCOMP] * arg1[i][BCOMP] * scaleRGB;
          }
          break;
       case GL_ADD:
@@ -286,19 +296,19 @@ texture_combine( const GLcontext *ctx, GLuint unit, GLuint n,
             /* (a * b) + (c * d) */
             for (i = 0; i < n; i++) {
                rgba[i][RCOMP] = (arg0[i][RCOMP] * arg1[i][RCOMP] +
-                                 arg2[i][RCOMP] * arg3[i][RCOMP]) * RGBmult;
+                                 arg2[i][RCOMP] * arg3[i][RCOMP]) * scaleRGB;
                rgba[i][GCOMP] = (arg0[i][GCOMP] * arg1[i][GCOMP] +
-                                 arg2[i][GCOMP] * arg3[i][GCOMP]) * RGBmult;
+                                 arg2[i][GCOMP] * arg3[i][GCOMP]) * scaleRGB;
                rgba[i][BCOMP] = (arg0[i][BCOMP] * arg1[i][BCOMP] +
-                                 arg2[i][BCOMP] * arg3[i][BCOMP]) * RGBmult;
+                                 arg2[i][BCOMP] * arg3[i][BCOMP]) * scaleRGB;
             }
          }
          else {
             /* 2-term addition */
             for (i = 0; i < n; i++) {
-               rgba[i][RCOMP] = (arg0[i][RCOMP] + arg1[i][RCOMP]) * RGBmult;
-               rgba[i][GCOMP] = (arg0[i][GCOMP] + arg1[i][GCOMP]) * RGBmult;
-               rgba[i][BCOMP] = (arg0[i][BCOMP] + arg1[i][BCOMP]) * RGBmult;
+               rgba[i][RCOMP] = (arg0[i][RCOMP] + arg1[i][RCOMP]) * scaleRGB;
+               rgba[i][GCOMP] = (arg0[i][GCOMP] + arg1[i][GCOMP]) * scaleRGB;
+               rgba[i][BCOMP] = (arg0[i][BCOMP] + arg1[i][BCOMP]) * scaleRGB;
             }
          }
          break;
@@ -306,110 +316,101 @@ texture_combine( const GLcontext *ctx, GLuint unit, GLuint n,
          if (textureUnit->EnvMode == GL_COMBINE4_NV) {
             /* (a * b) + (c * d) - 0.5 */
             for (i = 0; i < n; i++) {
-               rgba[i][RCOMP] = (arg0[i][RCOMP] + arg1[i][RCOMP] *
-                                 arg2[i][RCOMP] + arg3[i][RCOMP] - 0.5) * RGBmult;
-               rgba[i][GCOMP] = (arg0[i][GCOMP] + arg1[i][GCOMP] *
-                                 arg2[i][GCOMP] + arg3[i][GCOMP] - 0.5) * RGBmult;
-               rgba[i][BCOMP] = (arg0[i][BCOMP] + arg1[i][BCOMP] *
-                                 arg2[i][BCOMP] + arg3[i][BCOMP] - 0.5) * RGBmult;
+               rgba[i][RCOMP] = (arg0[i][RCOMP] * arg1[i][RCOMP] +
+                                 arg2[i][RCOMP] * arg3[i][RCOMP] - 0.5) * scaleRGB;
+               rgba[i][GCOMP] = (arg0[i][GCOMP] * arg1[i][GCOMP] +
+                                 arg2[i][GCOMP] * arg3[i][GCOMP] - 0.5) * scaleRGB;
+               rgba[i][BCOMP] = (arg0[i][BCOMP] * arg1[i][BCOMP] +
+                                 arg2[i][BCOMP] * arg3[i][BCOMP] - 0.5) * scaleRGB;
             }
          }
          else {
             for (i = 0; i < n; i++) {
-               rgba[i][RCOMP] = (arg0[i][RCOMP] + arg1[i][RCOMP] - 0.5) * RGBmult;
-               rgba[i][GCOMP] = (arg0[i][GCOMP] + arg1[i][GCOMP] - 0.5) * RGBmult;
-               rgba[i][BCOMP] = (arg0[i][BCOMP] + arg1[i][BCOMP] - 0.5) * RGBmult;
+               rgba[i][RCOMP] = (arg0[i][RCOMP] + arg1[i][RCOMP] - 0.5) * scaleRGB;
+               rgba[i][GCOMP] = (arg0[i][GCOMP] + arg1[i][GCOMP] - 0.5) * scaleRGB;
+               rgba[i][BCOMP] = (arg0[i][BCOMP] + arg1[i][BCOMP] - 0.5) * scaleRGB;
             }
          }
          break;
       case GL_INTERPOLATE:
          for (i = 0; i < n; i++) {
             rgba[i][RCOMP] = (arg0[i][RCOMP] * arg2[i][RCOMP] +
-                          arg1[i][RCOMP] * (1.0F - arg2[i][RCOMP])) * RGBmult;
+                          arg1[i][RCOMP] * (1.0F - arg2[i][RCOMP])) * scaleRGB;
             rgba[i][GCOMP] = (arg0[i][GCOMP] * arg2[i][GCOMP] +
-                          arg1[i][GCOMP] * (1.0F - arg2[i][GCOMP])) * RGBmult;
+                          arg1[i][GCOMP] * (1.0F - arg2[i][GCOMP])) * scaleRGB;
             rgba[i][BCOMP] = (arg0[i][BCOMP] * arg2[i][BCOMP] +
-                          arg1[i][BCOMP] * (1.0F - arg2[i][BCOMP])) * RGBmult;
+                          arg1[i][BCOMP] * (1.0F - arg2[i][BCOMP])) * scaleRGB;
          }
          break;
       case GL_SUBTRACT:
          for (i = 0; i < n; i++) {
-            rgba[i][RCOMP] = (arg0[i][RCOMP] - arg1[i][RCOMP]) * RGBmult;
-            rgba[i][GCOMP] = (arg0[i][GCOMP] - arg1[i][GCOMP]) * RGBmult;
-            rgba[i][BCOMP] = (arg0[i][BCOMP] - arg1[i][BCOMP]) * RGBmult;
+            rgba[i][RCOMP] = (arg0[i][RCOMP] - arg1[i][RCOMP]) * scaleRGB;
+            rgba[i][GCOMP] = (arg0[i][GCOMP] - arg1[i][GCOMP]) * scaleRGB;
+            rgba[i][BCOMP] = (arg0[i][BCOMP] - arg1[i][BCOMP]) * scaleRGB;
          }
          break;
       case GL_DOT3_RGB_EXT:
       case GL_DOT3_RGBA_EXT:
          /* Do not scale the result by 1 2 or 4 */
          for (i = 0; i < n; i++) {
-            GLfloat dot = ((arg0[i][RCOMP]-0.5F) * (arg1[i][RCOMP]-0.5F) +
-                           (arg0[i][GCOMP]-0.5F) * (arg1[i][GCOMP]-0.5F) +
-                           (arg0[i][BCOMP]-0.5F) * (arg1[i][BCOMP]-0.5F))
+            GLfloat dot = ((arg0[i][RCOMP] - 0.5F) * (arg1[i][RCOMP] - 0.5F) +
+                           (arg0[i][GCOMP] - 0.5F) * (arg1[i][GCOMP] - 0.5F) +
+                           (arg0[i][BCOMP] - 0.5F) * (arg1[i][BCOMP] - 0.5F))
                * 4.0F;
             dot = CLAMP(dot, 0.0F, 1.0F);
-            rgba[i][RCOMP] = rgba[i][GCOMP] = rgba[i][BCOMP] = (GLfloat) dot;
+            rgba[i][RCOMP] = rgba[i][GCOMP] = rgba[i][BCOMP] = dot;
          }
          break;
       case GL_DOT3_RGB:
       case GL_DOT3_RGBA:
          /* DO scale the result by 1 2 or 4 */
          for (i = 0; i < n; i++) {
-            GLfloat dot = ((arg0[i][RCOMP]-0.5F) * (arg1[i][RCOMP]-0.5F) +
-                           (arg0[i][GCOMP]-0.5F) * (arg1[i][GCOMP]-0.5F) +
-                           (arg0[i][BCOMP]-0.5F) * (arg1[i][BCOMP]-0.5F))
-               * 4.0F * RGBmult;
+            GLfloat dot = ((arg0[i][RCOMP] - 0.5F) * (arg1[i][RCOMP] - 0.5F) +
+                           (arg0[i][GCOMP] - 0.5F) * (arg1[i][GCOMP] - 0.5F) +
+                           (arg0[i][BCOMP] - 0.5F) * (arg1[i][BCOMP] - 0.5F))
+               * 4.0F * scaleRGB;
             dot = CLAMP(dot, 0.0, 1.0F);
-            rgba[i][RCOMP] = rgba[i][GCOMP] = rgba[i][BCOMP] = (GLfloat) dot;
+            rgba[i][RCOMP] = rgba[i][GCOMP] = rgba[i][BCOMP] = dot;
          }
          break;
       case GL_MODULATE_ADD_ATI:
          for (i = 0; i < n; i++) {
             rgba[i][RCOMP] = ((arg0[i][RCOMP] * arg2[i][RCOMP]) +
-                              arg1[i][RCOMP]) * RGBmult;
+                              arg1[i][RCOMP]) * scaleRGB;
             rgba[i][GCOMP] = ((arg0[i][GCOMP] * arg2[i][GCOMP]) +
-                              arg1[i][GCOMP]) * RGBmult;
+                              arg1[i][GCOMP]) * scaleRGB;
             rgba[i][BCOMP] = ((arg0[i][BCOMP] * arg2[i][BCOMP]) +
-                              arg1[i][BCOMP]) * RGBmult;
+                              arg1[i][BCOMP]) * scaleRGB;
 	 }
          break;
       case GL_MODULATE_SIGNED_ADD_ATI:
          for (i = 0; i < n; i++) {
             rgba[i][RCOMP] = ((arg0[i][RCOMP] * arg2[i][RCOMP]) +
-                              arg1[i][RCOMP] - 0.5) * RGBmult;
+                              arg1[i][RCOMP] - 0.5) * scaleRGB;
             rgba[i][GCOMP] = ((arg0[i][GCOMP] * arg2[i][GCOMP]) +
-                              arg1[i][GCOMP] - 0.5) * RGBmult;
+                              arg1[i][GCOMP] - 0.5) * scaleRGB;
             rgba[i][BCOMP] = ((arg0[i][BCOMP] * arg2[i][BCOMP]) +
-                              arg1[i][BCOMP] - 0.5) * RGBmult;
+                              arg1[i][BCOMP] - 0.5) * scaleRGB;
 	 }
          break;
       case GL_MODULATE_SUBTRACT_ATI:
          for (i = 0; i < n; i++) {
             rgba[i][RCOMP] = ((arg0[i][RCOMP] * arg2[i][RCOMP]) -
-                              arg1[i][RCOMP]) * RGBmult;
+                              arg1[i][RCOMP]) * scaleRGB;
             rgba[i][GCOMP] = ((arg0[i][GCOMP] * arg2[i][GCOMP]) -
-                              arg1[i][GCOMP]) * RGBmult;
+                              arg1[i][GCOMP]) * scaleRGB;
             rgba[i][BCOMP] = ((arg0[i][BCOMP] * arg2[i][BCOMP]) -
-                              arg1[i][BCOMP]) * RGBmult;
+                              arg1[i][BCOMP]) * scaleRGB;
 	 }
          break;
       case GL_BUMP_ENVMAP_ATI:
-         {
-            /* this produces a fixed rgba color, and the coord calc is done elsewhere */
-            for (i = 0; i < n; i++) {
+         /* this produces a fixed rgba color, and the coord calc is done elsewhere */
+         for (i = 0; i < n; i++) {
             /* rgba result is 0,0,0,1 */
-#if CHAN_TYPE == GL_FLOAT
-               rgba[i][RCOMP] = 0.0;
-               rgba[i][GCOMP] = 0.0;
-               rgba[i][BCOMP] = 0.0;
-               rgba[i][ACOMP] = 1.0;
-#else
-               rgba[i][RCOMP] = 0;
-               rgba[i][GCOMP] = 0;
-               rgba[i][BCOMP] = 0;
-               rgba[i][ACOMP] = CHAN_MAX;
-#endif
-            }
+            rgba[i][RCOMP] = 0.0;
+            rgba[i][GCOMP] = 0.0;
+            rgba[i][BCOMP] = 0.0;
+            rgba[i][ACOMP] = 1.0;
 	 }
          return; /* no alpha processing */
       default:
@@ -419,21 +420,20 @@ texture_combine( const GLcontext *ctx, GLuint unit, GLuint n,
 
    /* Alpha channel combine */
    {
-      const GLfloat (*arg0)[4] = (const GLfloat (*)[4]) argA[0];
-      const GLfloat (*arg1)[4] = (const GLfloat (*)[4]) argA[1];
-      const GLfloat (*arg2)[4] = (const GLfloat (*)[4]) argA[2];
-      const GLfloat (*arg3)[4] = (const GLfloat (*)[4]) argA[3];
+      float4_array arg0 = argA[0];
+      float4_array arg1 = argA[1];
+      float4_array arg2 = argA[2];
+      float4_array arg3 = argA[3];
 
       switch (combine->ModeA) {
       case GL_REPLACE:
          for (i = 0; i < n; i++) {
-            GLfloat a = arg0[i][ACOMP] * Amult;
-            rgba[i][ACOMP] = (GLfloat) MIN2(a, 1.0F);
+            rgba[i][ACOMP] = arg0[i][ACOMP] * scaleA;
          }
          break;
       case GL_MODULATE:
          for (i = 0; i < n; i++) {
-            rgba[i][ACOMP] = arg0[i][ACOMP] * arg1[i][ACOMP] * Amult;
+            rgba[i][ACOMP] = arg0[i][ACOMP] * arg1[i][ACOMP] * scaleA;
          }
          break;
       case GL_ADD:
@@ -441,13 +441,13 @@ texture_combine( const GLcontext *ctx, GLuint unit, GLuint n,
             /* (a * b) + (c * d) */
             for (i = 0; i < n; i++) {
                rgba[i][ACOMP] = (arg0[i][ACOMP] * arg1[i][ACOMP] +
-                                 arg2[i][ACOMP] * arg3[i][ACOMP]) * Amult;
+                                 arg2[i][ACOMP] * arg3[i][ACOMP]) * scaleA;
             }
          }
          else {
             /* two-term add */
             for (i = 0; i < n; i++) {
-               rgba[i][ACOMP] = (arg0[i][ACOMP] + arg1[i][ACOMP]) * Amult;
+               rgba[i][ACOMP] = (arg0[i][ACOMP] + arg1[i][ACOMP]) * scaleA;
             }
          }
          break;
@@ -457,44 +457,44 @@ texture_combine( const GLcontext *ctx, GLuint unit, GLuint n,
             for (i = 0; i < n; i++) {
                rgba[i][ACOMP] = (arg0[i][ACOMP] * arg1[i][ACOMP] +
                                  arg2[i][ACOMP] * arg3[i][ACOMP] -
-                                 0.5) * Amult;
+                                 0.5) * scaleA;
             }
          }
          else {
             /* a + b - 0.5 */
             for (i = 0; i < n; i++) {
-               rgba[i][ACOMP] = (arg0[i][ACOMP] + arg1[i][ACOMP] - 0.5F) * Amult;
+               rgba[i][ACOMP] = (arg0[i][ACOMP] + arg1[i][ACOMP] - 0.5F) * scaleA;
             }
          }
          break;
       case GL_INTERPOLATE:
-         for (i=0; i<n; i++) {
+         for (i = 0; i < n; i++) {
             rgba[i][ACOMP] = (arg0[i][ACOMP] * arg2[i][ACOMP] +
                               arg1[i][ACOMP] * (1.0F - arg2[i][ACOMP]))
-               * Amult;
+               * scaleA;
          }
          break;
       case GL_SUBTRACT:
          for (i = 0; i < n; i++) {
-            rgba[i][ACOMP] = (arg0[i][ACOMP] - arg1[i][ACOMP]) * Amult;
+            rgba[i][ACOMP] = (arg0[i][ACOMP] - arg1[i][ACOMP]) * scaleA;
          }
          break;
       case GL_MODULATE_ADD_ATI:
          for (i = 0; i < n; i++) {
             rgba[i][ACOMP] = ((arg0[i][ACOMP] * arg2[i][ACOMP])
-                              + arg1[i][ACOMP]) * Amult;
+                              + arg1[i][ACOMP]) * scaleA;
          }
          break;
       case GL_MODULATE_SIGNED_ADD_ATI:
          for (i = 0; i < n; i++) {
             rgba[i][ACOMP] = ((arg0[i][ACOMP] * arg2[i][ACOMP]) +
-                              arg1[i][ACOMP] - 0.5F) * Amult;
+                              arg1[i][ACOMP] - 0.5F) * scaleA;
          }
          break;
       case GL_MODULATE_SUBTRACT_ATI:
          for (i = 0; i < n; i++) {
             rgba[i][ACOMP] = ((arg0[i][ACOMP] * arg2[i][ACOMP])
-                              - arg1[i][ACOMP]) * Amult;
+                              - arg1[i][ACOMP]) * scaleA;
          }
          break;
       default:
@@ -523,13 +523,12 @@ texture_combine( const GLcontext *ctx, GLuint unit, GLuint n,
 }
 
 
-
 /**
  * Apply X/Y/Z/W/0/1 swizzle to an array of colors/texels.
  * See GL_EXT_texture_swizzle.
  */
 static void
-swizzle_texels(GLuint swizzle, GLuint count, GLfloat (*texels)[4])
+swizzle_texels(GLuint swizzle, GLuint count, float4_array texels)
 {
    const GLuint swzR = GET_SWZ(swizzle, 0);
    const GLuint swzG = GET_SWZ(swizzle, 1);
@@ -554,386 +553,6 @@ swizzle_texels(GLuint swizzle, GLuint count, GLfloat (*texels)[4])
 }
 
 
-/**
- * Apply a conventional OpenGL texture env mode (REPLACE, ADD, BLEND,
- * MODULATE, or DECAL) to an array of fragments.
- * Input:  textureUnit - pointer to texture unit to apply
- *         format - base internal texture format
- *         n - number of fragments
- *         primary_rgba - primary colors (may alias rgba for single texture)
- *         texels - array of texel colors
- * InOut:  rgba - incoming fragment colors modified by texel colors
- *                according to the texture environment mode.
- */
-static void
-texture_apply( const GLcontext *ctx,
-               const struct gl_texture_unit *texUnit,
-               GLuint n,
-               CONST GLfloat primary_rgba[][4], CONST GLfloat texel[][4],
-               GLchan rgbaChan[][4] )
-{
-   GLint baseLevel;
-   GLuint i;
-   GLfloat Rc, Gc, Bc, Ac;
-   GLenum format;
-   GLfloat rgba[MAX_WIDTH][4];
-
-   (void) primary_rgba;
-
-   ASSERT(texUnit);
-   ASSERT(texUnit->_Current);
-
-   baseLevel = texUnit->_Current->BaseLevel;
-   ASSERT(texUnit->_Current->Image[0][baseLevel]);
-
-   format = texUnit->_Current->Image[0][baseLevel]->_BaseFormat;
-
-   if (format == GL_COLOR_INDEX || format == GL_YCBCR_MESA) {
-      format = GL_RGBA;  /* a bit of a hack */
-   }
-   else if (format == GL_DEPTH_COMPONENT || format == GL_DEPTH_STENCIL_EXT) {
-      format = texUnit->_Current->DepthMode;
-   }
-
-   if (texUnit->EnvMode != GL_REPLACE) {
-      /* convert GLchan colors to GLfloat */
-      for (i = 0; i < n; i++) {
-         rgba[i][RCOMP] = CHAN_TO_FLOAT(rgbaChan[i][RCOMP]);
-         rgba[i][GCOMP] = CHAN_TO_FLOAT(rgbaChan[i][GCOMP]);
-         rgba[i][BCOMP] = CHAN_TO_FLOAT(rgbaChan[i][BCOMP]);
-         rgba[i][ACOMP] = CHAN_TO_FLOAT(rgbaChan[i][ACOMP]);
-      }
-   }
-
-   switch (texUnit->EnvMode) {
-      case GL_REPLACE:
-	 switch (format) {
-	    case GL_ALPHA:
-	       for (i=0;i<n;i++) {
-		  /* Cv = Cf */
-                  /* Av = At */
-                  rgba[i][ACOMP] = texel[i][ACOMP];
-	       }
-	       break;
-	    case GL_LUMINANCE:
-	       for (i=0;i<n;i++) {
-		  /* Cv = Lt */
-                  GLfloat Lt = texel[i][RCOMP];
-                  rgba[i][RCOMP] = rgba[i][GCOMP] = rgba[i][BCOMP] = Lt;
-                  /* Av = Af */
-	       }
-	       break;
-	    case GL_LUMINANCE_ALPHA:
-	       for (i=0;i<n;i++) {
-                  GLfloat Lt = texel[i][RCOMP];
-		  /* Cv = Lt */
-		  rgba[i][RCOMP] = rgba[i][GCOMP] = rgba[i][BCOMP] = Lt;
-		  /* Av = At */
-		  rgba[i][ACOMP] = texel[i][ACOMP];
-	       }
-	       break;
-	    case GL_INTENSITY:
-	       for (i=0;i<n;i++) {
-		  /* Cv = It */
-                  GLfloat It = texel[i][RCOMP];
-                  rgba[i][RCOMP] = rgba[i][GCOMP] = rgba[i][BCOMP] = It;
-                  /* Av = It */
-                  rgba[i][ACOMP] = It;
-	       }
-	       break;
-	    case GL_RGB:
-	       for (i=0;i<n;i++) {
-		  /* Cv = Ct */
-		  rgba[i][RCOMP] = texel[i][RCOMP];
-		  rgba[i][GCOMP] = texel[i][GCOMP];
-		  rgba[i][BCOMP] = texel[i][BCOMP];
-		  /* Av = Af */
-	       }
-	       break;
-	    case GL_RGBA:
-	       for (i=0;i<n;i++) {
-		  /* Cv = Ct */
-		  rgba[i][RCOMP] = texel[i][RCOMP];
-		  rgba[i][GCOMP] = texel[i][GCOMP];
-		  rgba[i][BCOMP] = texel[i][BCOMP];
-		  /* Av = At */
-		  rgba[i][ACOMP] = texel[i][ACOMP];
-	       }
-	       break;
-            default:
-               _mesa_problem(ctx, "Bad format (GL_REPLACE) in texture_apply");
-               return;
-	 }
-	 break;
-
-      case GL_MODULATE:
-         switch (format) {
-	    case GL_ALPHA:
-	       for (i=0;i<n;i++) {
-		  /* Cv = Cf */
-		  /* Av = AfAt */
-		  rgba[i][ACOMP] = rgba[i][ACOMP] * texel[i][ACOMP];
-	       }
-	       break;
-	    case GL_LUMINANCE:
-	       for (i=0;i<n;i++) {
-		  /* Cv = LtCf */
-                  GLfloat Lt = texel[i][RCOMP];
-		  rgba[i][RCOMP] = rgba[i][RCOMP] * Lt;
-		  rgba[i][GCOMP] = rgba[i][GCOMP] * Lt;
-		  rgba[i][BCOMP] = rgba[i][BCOMP] * Lt;
-		  /* Av = Af */
-	       }
-	       break;
-	    case GL_LUMINANCE_ALPHA:
-	       for (i=0;i<n;i++) {
-		  /* Cv = CfLt */
-                  GLfloat Lt = texel[i][RCOMP];
-		  rgba[i][RCOMP] = rgba[i][RCOMP] * Lt;
-		  rgba[i][GCOMP] = rgba[i][GCOMP] * Lt;
-		  rgba[i][BCOMP] = rgba[i][BCOMP] * Lt;
-		  /* Av = AfAt */
-		  rgba[i][ACOMP] = rgba[i][ACOMP] * texel[i][ACOMP];
-	       }
-	       break;
-	    case GL_INTENSITY:
-	       for (i=0;i<n;i++) {
-		  /* Cv = CfIt */
-                  GLfloat It = texel[i][RCOMP];
-		  rgba[i][RCOMP] = rgba[i][RCOMP] * It;
-		  rgba[i][GCOMP] = rgba[i][GCOMP] * It;
-		  rgba[i][BCOMP] = rgba[i][BCOMP] * It;
-		  /* Av = AfIt */
-		  rgba[i][ACOMP] = rgba[i][ACOMP] * It;
-	       }
-	       break;
-	    case GL_RGB:
-	       for (i=0;i<n;i++) {
-		  /* Cv = CfCt */
-		  rgba[i][RCOMP] = rgba[i][RCOMP] * texel[i][RCOMP];
-		  rgba[i][GCOMP] = rgba[i][GCOMP] * texel[i][GCOMP];
-		  rgba[i][BCOMP] = rgba[i][BCOMP] * texel[i][BCOMP];
-		  /* Av = Af */
-	       }
-	       break;
-	    case GL_RGBA:
-	       for (i=0;i<n;i++) {
-		  /* Cv = CfCt */
-		  rgba[i][RCOMP] = rgba[i][RCOMP] * texel[i][RCOMP];
-		  rgba[i][GCOMP] = rgba[i][GCOMP] * texel[i][GCOMP];
-		  rgba[i][BCOMP] = rgba[i][BCOMP] * texel[i][BCOMP];
-		  /* Av = AfAt */
-		  rgba[i][ACOMP] = rgba[i][ACOMP] * texel[i][ACOMP];
-	       }
-	       break;
-            default:
-               _mesa_problem(ctx, "Bad format (GL_MODULATE) in texture_apply");
-               return;
-	 }
-	 break;
-
-      case GL_DECAL:
-         switch (format) {
-            case GL_ALPHA:
-            case GL_LUMINANCE:
-            case GL_LUMINANCE_ALPHA:
-            case GL_INTENSITY:
-               /* undefined */
-               break;
-	    case GL_RGB:
-	       for (i=0;i<n;i++) {
-		  /* Cv = Ct */
-		  rgba[i][RCOMP] = texel[i][RCOMP];
-		  rgba[i][GCOMP] = texel[i][GCOMP];
-		  rgba[i][BCOMP] = texel[i][BCOMP];
-		  /* Av = Af */
-	       }
-	       break;
-	    case GL_RGBA:
-	       for (i=0;i<n;i++) {
-		  /* Cv = Cf(1-At) + CtAt */
-		  GLfloat t = texel[i][ACOMP], s = 1.0F - t;
-		  rgba[i][RCOMP] = rgba[i][RCOMP] * s + texel[i][RCOMP] * t;
-		  rgba[i][GCOMP] = rgba[i][GCOMP] * s + texel[i][GCOMP] * t;
-		  rgba[i][BCOMP] = rgba[i][BCOMP] * s + texel[i][BCOMP] * t;
-		  /* Av = Af */
-	       }
-	       break;
-            default:
-               _mesa_problem(ctx, "Bad format (GL_DECAL) in texture_apply");
-               return;
-	 }
-	 break;
-
-      case GL_BLEND:
-         Rc = texUnit->EnvColor[0];
-         Gc = texUnit->EnvColor[1];
-         Bc = texUnit->EnvColor[2];
-         Ac = texUnit->EnvColor[3];
-	 switch (format) {
-	    case GL_ALPHA:
-	       for (i=0;i<n;i++) {
-		  /* Cv = Cf */
-		  /* Av = AfAt */
-                  rgba[i][ACOMP] = rgba[i][ACOMP] * texel[i][ACOMP];
-	       }
-	       break;
-            case GL_LUMINANCE:
-	       for (i=0;i<n;i++) {
-		  /* Cv = Cf(1-Lt) + CcLt */
-		  GLfloat Lt = texel[i][RCOMP], s = 1.0F - Lt;
-		  rgba[i][RCOMP] = rgba[i][RCOMP] * s + Rc * Lt;
-		  rgba[i][GCOMP] = rgba[i][GCOMP] * s + Gc * Lt;
-		  rgba[i][BCOMP] = rgba[i][BCOMP] * s + Bc * Lt;
-		  /* Av = Af */
-	       }
-	       break;
-	    case GL_LUMINANCE_ALPHA:
-	       for (i=0;i<n;i++) {
-		  /* Cv = Cf(1-Lt) + CcLt */
-		  GLfloat Lt = texel[i][RCOMP], s = 1.0F - Lt;
-		  rgba[i][RCOMP] = rgba[i][RCOMP] * s + Rc * Lt;
-		  rgba[i][GCOMP] = rgba[i][GCOMP] * s + Gc * Lt;
-		  rgba[i][BCOMP] = rgba[i][BCOMP] * s + Bc * Lt;
-		  /* Av = AfAt */
-		  rgba[i][ACOMP] = rgba[i][ACOMP] * texel[i][ACOMP];
-	       }
-	       break;
-            case GL_INTENSITY:
-	       for (i=0;i<n;i++) {
-		  /* Cv = Cf(1-It) + CcIt */
-		  GLfloat It = texel[i][RCOMP], s = 1.0F - It;
-		  rgba[i][RCOMP] = rgba[i][RCOMP] * s + Rc * It;
-		  rgba[i][GCOMP] = rgba[i][GCOMP] * s + Gc * It;
-		  rgba[i][BCOMP] = rgba[i][BCOMP] * s + Bc * It;
-                  /* Av = Af(1-It) + Ac*It */
-                  rgba[i][ACOMP] = rgba[i][ACOMP] * s + Ac * It;
-               }
-               break;
-	    case GL_RGB:
-	       for (i=0;i<n;i++) {
-		  /* Cv = Cf(1-Ct) + CcCt */
-		  rgba[i][RCOMP] = rgba[i][RCOMP] * (1.0F - texel[i][RCOMP])
-                     + Rc * texel[i][RCOMP];
-		  rgba[i][GCOMP] = rgba[i][GCOMP] * (1.0F - texel[i][GCOMP])
-                     + Gc * texel[i][GCOMP];
-		  rgba[i][BCOMP] = rgba[i][BCOMP] * (1.0F - texel[i][BCOMP])
-                     + Bc * texel[i][BCOMP];
-		  /* Av = Af */
-	       }
-	       break;
-	    case GL_RGBA:
-	       for (i=0;i<n;i++) {
-		  /* Cv = Cf(1-Ct) + CcCt */
-		  rgba[i][RCOMP] = rgba[i][RCOMP] * (1.0F - texel[i][RCOMP])
-                     + Rc * texel[i][RCOMP];
-		  rgba[i][GCOMP] = rgba[i][GCOMP] * (1.0F - texel[i][GCOMP])
-                     + Gc * texel[i][GCOMP];
-		  rgba[i][BCOMP] = rgba[i][BCOMP] * (1.0F - texel[i][BCOMP])
-                     + Bc * texel[i][BCOMP];
-		  /* Av = AfAt */
-		  rgba[i][ACOMP] = rgba[i][ACOMP] * texel[i][ACOMP];
-	       }
-	       break;
-            default:
-               _mesa_problem(ctx, "Bad format (GL_BLEND) in texture_apply");
-               return;
-	 }
-	 break;
-
-     /* XXX don't clamp results if GLfloat is float??? */
-
-      case GL_ADD:  /* GL_EXT_texture_add_env */
-         switch (format) {
-            case GL_ALPHA:
-               for (i=0;i<n;i++) {
-                  /* Rv = Rf */
-                  /* Gv = Gf */
-                  /* Bv = Bf */
-                  rgba[i][ACOMP] = rgba[i][ACOMP] * texel[i][ACOMP];
-               }
-               break;
-            case GL_LUMINANCE:
-               for (i=0;i<n;i++) {
-                  GLfloat Lt = texel[i][RCOMP];
-                  GLfloat r = rgba[i][RCOMP] + Lt;
-                  GLfloat g = rgba[i][GCOMP] + Lt;
-                  GLfloat b = rgba[i][BCOMP] + Lt;
-                  rgba[i][RCOMP] = MIN2(r, 1.0F);
-                  rgba[i][GCOMP] = MIN2(g, 1.0F);
-                  rgba[i][BCOMP] = MIN2(b, 1.0F);
-                  /* Av = Af */
-               }
-               break;
-            case GL_LUMINANCE_ALPHA:
-               for (i=0;i<n;i++) {
-                  GLfloat Lt = texel[i][RCOMP];
-                  GLfloat r = rgba[i][RCOMP] + Lt;
-                  GLfloat g = rgba[i][GCOMP] + Lt;
-                  GLfloat b = rgba[i][BCOMP] + Lt;
-                  rgba[i][RCOMP] = MIN2(r, 1.0F);
-                  rgba[i][GCOMP] = MIN2(g, 1.0F);
-                  rgba[i][BCOMP] = MIN2(b, 1.0F);
-                  rgba[i][ACOMP] = rgba[i][ACOMP] * texel[i][ACOMP];
-               }
-               break;
-            case GL_INTENSITY:
-               for (i=0;i<n;i++) {
-                  GLfloat It = texel[i][RCOMP];
-                  GLfloat r = rgba[i][RCOMP] + It;
-                  GLfloat g = rgba[i][GCOMP] + It;
-                  GLfloat b = rgba[i][BCOMP] + It;
-                  GLfloat a = rgba[i][ACOMP] + It;
-                  rgba[i][RCOMP] = MIN2(r, 1.0F);
-                  rgba[i][GCOMP] = MIN2(g, 1.0F);
-                  rgba[i][BCOMP] = MIN2(b, 1.0F);
-                  rgba[i][ACOMP] = MIN2(a, 1.0F);
-               }
-               break;
-	    case GL_RGB:
-	       for (i=0;i<n;i++) {
-                  GLfloat r = rgba[i][RCOMP] + texel[i][RCOMP];
-                  GLfloat g = rgba[i][GCOMP] + texel[i][GCOMP];
-                  GLfloat b = rgba[i][BCOMP] + texel[i][BCOMP];
-		  rgba[i][RCOMP] = MIN2(r, 1.0F);
-		  rgba[i][GCOMP] = MIN2(g, 1.0F);
-		  rgba[i][BCOMP] = MIN2(b, 1.0F);
-		  /* Av = Af */
-	       }
-	       break;
-	    case GL_RGBA:
-	       for (i=0;i<n;i++) {
-                  GLfloat r = rgba[i][RCOMP] + texel[i][RCOMP];
-                  GLfloat g = rgba[i][GCOMP] + texel[i][GCOMP];
-                  GLfloat b = rgba[i][BCOMP] + texel[i][BCOMP];
-		  rgba[i][RCOMP] = MIN2(r, 1.0F);
-		  rgba[i][GCOMP] = MIN2(g, 1.0F);
-		  rgba[i][BCOMP] = MIN2(b, 1.0F);
-                  rgba[i][ACOMP] = rgba[i][ACOMP] * texel[i][ACOMP];
-               }
-               break;
-            default:
-               _mesa_problem(ctx, "Bad format (GL_ADD) in texture_apply");
-               return;
-	 }
-	 break;
-
-      default:
-         _mesa_problem(ctx, "Bad env mode in texture_apply");
-         return;
-   }
-
-   /* convert GLfloat colors to GLchan */
-   for (i = 0; i < n; i++) {
-      CLAMPED_FLOAT_TO_CHAN(rgbaChan[i][RCOMP], rgba[i][RCOMP]);
-      CLAMPED_FLOAT_TO_CHAN(rgbaChan[i][GCOMP], rgba[i][GCOMP]);
-      CLAMPED_FLOAT_TO_CHAN(rgbaChan[i][BCOMP], rgba[i][BCOMP]);
-      CLAMPED_FLOAT_TO_CHAN(rgbaChan[i][ACOMP], rgba[i][ACOMP]);
-   }
-}
-
-
-
 /**
  * Apply texture mapping to a span of fragments.
  */
@@ -944,12 +563,12 @@ _swrast_texture_span( GLcontext *ctx, SWspan *span )
    GLfloat primary_rgba[MAX_WIDTH][4];
    GLuint unit;
 
-   ASSERT(span->end < MAX_WIDTH);
+   ASSERT(span->end <= MAX_WIDTH);
 
    /*
     * Save copy of the incoming fragment colors (the GL_PRIMARY_COLOR)
     */
-   if (swrast->_AnyTextureCombine) {
+   if (swrast->_TextureCombinePrimary) {
       GLuint i;
       for (i = 0; i < span->end; i++) {
          primary_rgba[i][RCOMP] = CHAN_TO_FLOAT(span->array->rgba[i][RCOMP]);
@@ -961,21 +580,19 @@ _swrast_texture_span( GLcontext *ctx, SWspan *span )
 
    /* First must sample all bump maps */
    for (unit = 0; unit < ctx->Const.MaxTextureUnits; unit++) {
-      if (ctx->Texture.Unit[unit]._ReallyEnabled &&
-         ctx->Texture.Unit[unit]._CurrentCombine->ModeRGB == GL_BUMP_ENVMAP_ATI) {
-         const GLfloat (*texcoords)[4]
-            = (const GLfloat (*)[4])
+      const struct gl_texture_unit *texUnit = &ctx->Texture.Unit[unit];
+
+      if (texUnit->_ReallyEnabled &&
+         texUnit->_CurrentCombine->ModeRGB == GL_BUMP_ENVMAP_ATI) {
+         const GLfloat (*texcoords)[4] = (const GLfloat (*)[4])
             span->array->attribs[FRAG_ATTRIB_TEX0 + unit];
-         GLfloat (*targetcoords)[4]
-            = (GLfloat (*)[4])
+         float4_array targetcoords =
             span->array->attribs[FRAG_ATTRIB_TEX0 +
                ctx->Texture.Unit[unit].BumpTarget - GL_TEXTURE0];
 
-         const struct gl_texture_unit *texUnit = &ctx->Texture.Unit[unit];
          const struct gl_texture_object *curObj = texUnit->_Current;
          GLfloat *lambda = span->array->lambda[unit];
-         GLchan (*texels)[4] = (GLchan (*)[4])
-            (swrast->TexelBuffer + unit * (span->end * 4 * sizeof(GLchan)));
+         float4_array texels = get_texel_array(swrast, unit);
          GLuint i;
          GLfloat rotMatrix00 = ctx->Texture.Unit[unit].RotMatrix[0];
          GLfloat rotMatrix01 = ctx->Texture.Unit[unit].RotMatrix[1];
@@ -1015,19 +632,10 @@ _swrast_texture_span( GLcontext *ctx, SWspan *span )
             not sure this can work correctly even ignoring
             the problem that channel is unsigned */
          for (i = 0; i < span->end; i++) {
-#if CHAN_TYPE == GL_FLOAT
             targetcoords[i][0] += (texels[i][0] * rotMatrix00 + texels[i][1] *
                                   rotMatrix01) / targetcoords[i][3];
             targetcoords[i][1] += (texels[i][0] * rotMatrix10 + texels[i][1] *
                                   rotMatrix11) / targetcoords[i][3];
-#else
-            targetcoords[i][0] += (CHAN_TO_FLOAT(texels[i][1]) * rotMatrix00 +
-                                  CHAN_TO_FLOAT(texels[i][1]) * rotMatrix01) /
-                                  targetcoords[i][3];
-            targetcoords[i][1] += (CHAN_TO_FLOAT(texels[i][0]) * rotMatrix10 + 
-                                  CHAN_TO_FLOAT(texels[i][1]) * rotMatrix11) /
-                                  targetcoords[i][3];
-#endif
          }
       }
    }
@@ -1037,16 +645,14 @@ _swrast_texture_span( GLcontext *ctx, SWspan *span )
     * accomodate GL_ARB_texture_env_crossbar.
     */
    for (unit = 0; unit < ctx->Const.MaxTextureUnits; unit++) {
-      if (ctx->Texture.Unit[unit]._ReallyEnabled &&
-         ctx->Texture.Unit[unit]._CurrentCombine->ModeRGB != GL_BUMP_ENVMAP_ATI) {
-         const GLfloat (*texcoords)[4]
-            = (const GLfloat (*)[4])
+      const struct gl_texture_unit *texUnit = &ctx->Texture.Unit[unit];
+      if (texUnit->_ReallyEnabled &&
+          texUnit->_CurrentCombine->ModeRGB != GL_BUMP_ENVMAP_ATI) {
+         const GLfloat (*texcoords)[4] = (const GLfloat (*)[4])
             span->array->attribs[FRAG_ATTRIB_TEX0 + unit];
-         const struct gl_texture_unit *texUnit = &ctx->Texture.Unit[unit];
          const struct gl_texture_object *curObj = texUnit->_Current;
          GLfloat *lambda = span->array->lambda[unit];
-         GLfloat (*texels)[4] = (GLfloat (*)[4])
-            (swrast->TexelBuffer + unit * (span->end * 4 * sizeof(GLfloat)));
+         float4_array texels = get_texel_array(swrast, unit);
 
          /* adjust texture lod (lambda) */
          if (span->arrayMask & SPAN_LAMBDA) {
@@ -1089,31 +695,16 @@ _swrast_texture_span( GLcontext *ctx, SWspan *span )
       }
    }
 
-
    /*
     * OK, now apply the texture (aka texture combine/blend).
     * We modify the span->color.rgba values.
     */
    for (unit = 0; unit < ctx->Const.MaxTextureUnits; unit++) {
       if (ctx->Texture.Unit[unit]._ReallyEnabled) {
-         const struct gl_texture_unit *texUnit = &ctx->Texture.Unit[unit];
-         if (texUnit->_CurrentCombine != &texUnit->_EnvMode ) {
-            texture_combine( ctx, unit, span->end,
-                             (CONST GLfloat (*)[4]) primary_rgba,
-                             swrast->TexelBuffer,
-                             span->array->rgba );
-         }
-         else {
-            /* conventional texture blend */
-            const GLfloat (*texels)[4] = (const GLfloat (*)[4])
-               (swrast->TexelBuffer + unit *
-                (span->end * 4 * sizeof(GLfloat)));
-
-
-            texture_apply( ctx, texUnit, span->end,
-                           (CONST GLfloat (*)[4]) primary_rgba, texels,
-                           span->array->rgba );
-         }
+         texture_combine( ctx, unit, span->end,
+                          primary_rgba,
+                          swrast->TexelBuffer,
+                          span->array->rgba );
       }
    }
 }