Previously in TXP and TEX instructions, lambda was passed to

fetch_texel as zero, but I believe this is incorrect.  The spec uses a
pseudocode function:

      vec4 TextureSample(float s, float t, float r, float lodBias,
                         int texImageUnit, enum texTarget);

to specify the behaviour of TEX, TXB and TXP instructions.  For TEX
and TXP, lodBias is passed as zero, TXB is passed with texcoord[4].  In our code we have

      static void
      fetch_texel( GLcontext *ctx, const GLfloat texcoord[4], GLfloat lambda,
                   GLuint unit, GLfloat color[4] )

and were passing zero and a biased lambda value respectively.  The
difference is that TextureSample() would add in the lambda term
itself, while in our code the caller is expected to do this.  Thus in
the TEX and TXP cases, it is necessary to pass an unbiased lambda
value for things to work out correctly.

diff --git a/src/mesa/swrast/s_nvfragprog.c b/src/mesa/swrast/s_nvfragprog.c
index bb9c986..88fe989 100644 (file)
--- a/src/mesa/swrast/s_nvfragprog.c
+++ b/src/mesa/swrast/s_nvfragprog.c
@@ -1179,7 +1179,19 @@ execute_program( GLcontext *ctx,
                /* Note: we pass 0 for LOD.  The ARB extension requires it
                 * while the NV extension says it's implementation dependant.
                 */
-               fetch_texel( ctx, texcoord, 0.0F, inst->TexSrcUnit, color );
+               /* KW: Previously lambda was passed as zero, but I
+               * believe this is incorrect, the spec seems to
+               * indicate rather that lambda should not be
+               * changed/biased, unlike TXB where texcoord[3] is
+               * added to the lambda calculations.  The lambda should
+               * still be calculated normally for TEX & TXP though,
+               * not set to zero.  Otherwise it's very difficult to
+               * implement normal GL semantics through the fragment
+               * shader.
+               */
+               fetch_texel( ctx, texcoord, 
+                           span->array->lambda[inst->TexSrcUnit][column],
+                           inst->TexSrcUnit, color );
                if (color[3])
                   printf("color[3] = %f\n", color[3]);
                store_vector4( inst, machine, color );
@@ -1227,8 +1239,17 @@ execute_program( GLcontext *ctx,
                  texcoord[1] /= texcoord[3];
                  texcoord[2] /= texcoord[3];
               }
-               /* Note: LOD=0 */
-               fetch_texel( ctx, texcoord, 0.0F, inst->TexSrcUnit, color );
+               /* KW: Previously lambda was passed as zero, but I
+               * believe this is incorrect, the spec seems to
+               * indicate rather that lambda should not be
+               * changed/biased, unlike TXB where texcoord[3] is
+               * added to the lambda calculations.  The lambda should
+               * still be calculated normally for TEX & TXP though,
+               * not set to zero.
+               */
+               fetch_texel( ctx, texcoord, 
+                           span->array->lambda[inst->TexSrcUnit][column],
+                           inst->TexSrcUnit, color );
                store_vector4( inst, machine, color );
             }
             break;