Add UsageMask to DECLARATION in TGSI.

Interpolate FS attributes in the shader.
Do not copy WPOS in FS.

11 files changed, 276 insertions, 251 deletions

diff --git a/src/mesa/pipe/softpipe/sp_headers.h b/src/mesa/pipe/softpipe/sp_headers.h
index b7f46cb658..cc8294b18e 100644
--- a/src/mesa/pipe/softpipe/sp_headers.h
+++ b/src/mesa/pipe/softpipe/sp_headers.h
@@ -31,6 +31,7 @@
 #ifndef SP_HEADERS_H
 #define SP_HEADERS_H
 
+#include "../tgsi/core/tgsi_core.h"
 
 #define PRIM_POINT 1
 #define PRIM_LINE  2
@@ -44,7 +45,6 @@
 #define QUAD_BOTTOM_RIGHT 1
 #define QUAD_TOP_LEFT     2
 #define QUAD_TOP_RIGHT    3
-#define QUAD_SIZE        (2*2)
 
 #define MASK_BOTTOM_LEFT  0x1
 #define MASK_BOTTOM_RIGHT 0x2
@@ -53,17 +53,6 @@
 #define MASK_ALL          0xf
 
 
-#define NUM_CHANNELS   4	/* avoid confusion between 4 pixels and 4 channels */
-
-
-struct setup_coefficient {
-   float a0[NUM_CHANNELS];	/* in an xyzw layout */
-   float dadx[NUM_CHANNELS];
-   float dady[NUM_CHANNELS];
-};
-
-
-
 /**
  * Encodes everything we need to know about a 2x2 pixel block.  Uses
  * "Channel-Serial" or "SoA" layout.  
@@ -76,17 +65,13 @@ struct quad_header {
    unsigned prim:2;     /**< PRIM_POINT, LINE, TRI */
 
    struct {
-      float color[4][QUAD_SIZE];	/* rrrr, gggg, bbbb, aaaa */
+      float color[NUM_CHANNELS][QUAD_SIZE];	/* rrrr, gggg, bbbb, aaaa */
       float depth[QUAD_SIZE];
    } outputs;
 
    float coverage[QUAD_SIZE];    /** fragment coverage for antialiasing */
 
-   const struct setup_coefficient *coef;
-
-   const enum interp_mode *interp; /* XXX: this information should be
-				    * encoded in fragment program DECL
-				    * statements. */
+   const struct tgsi_interp_coef *coef;
 
    unsigned nr_attrs;
 };
diff --git a/src/mesa/pipe/softpipe/sp_prim_setup.c b/src/mesa/pipe/softpipe/sp_prim_setup.c
index 45d09860c3..71ef798cbb 100644
--- a/src/mesa/pipe/softpipe/sp_prim_setup.c
+++ b/src/mesa/pipe/softpipe/sp_prim_setup.c
@@ -80,7 +80,7 @@ struct setup_stage {
 
    float oneoverarea;
 
-   struct setup_coefficient coef[FRAG_ATTRIB_MAX];
+   struct tgsi_interp_coef coef[FRAG_ATTRIB_MAX];
    struct quad_header quad; 
 
    struct {
diff --git a/src/mesa/pipe/softpipe/sp_quad_fs.c b/src/mesa/pipe/softpipe/sp_quad_fs.c
index bffb03a6bc..0a345bbf11 100644
--- a/src/mesa/pipe/softpipe/sp_quad_fs.c
+++ b/src/mesa/pipe/softpipe/sp_quad_fs.c
@@ -33,15 +33,12 @@
  */
 
 #include "pipe/p_util.h"
-#include "tgsi/core/tgsi_core.h"
 
 #include "sp_context.h"
 #include "sp_headers.h"
 #include "sp_quad.h"
 #include "sp_tex_sample.h"
 
-#include "main/mtypes.h"
-
 
 #if defined __GNUC__
 #define USE_ALIGNED_ATTRIBS   1
@@ -66,157 +63,33 @@ quad_shade_stage(struct quad_stage *qs)
    return (struct quad_shade_stage *) qs;
 }
 
-
-
-struct exec_machine {
-   const struct setup_coefficient *coef; /**< will point to quad->coef */
-   float attr[PIPE_ATTRIB_MAX][NUM_CHANNELS][QUAD_SIZE] ALIGN16_SUFFIX;
-};
-
-
-/**
- * Compute quad's attributes values, as constants (GL_FLAT shading).
- */
-static INLINE void cinterp( struct exec_machine *exec,
-			    unsigned attrib,
-			    unsigned i )
-{
-   unsigned j;
-
-   for (j = 0; j < QUAD_SIZE; j++) {
-      exec->attr[attrib][i][j] = exec->coef[attrib].a0[i];
-   }
-}
-
-
-/**
- * Compute quad's attribute values by linear interpolation.
- *
- * Push into the fp:
- * 
- *   INPUT[attr] = MAD COEF_A0[attr], COEF_DADX[attr], INPUT_WPOS.xxxx
- *   INPUT[attr] = MAD INPUT[attr],   COEF_DADY[attr], INPUT_WPOS.yyyy
- */
-static INLINE void linterp( struct exec_machine *exec,
-			    unsigned attrib,
-			    unsigned i )
-{
-   unsigned j;
-
-   for (j = 0; j < QUAD_SIZE; j++) {
-      const float x = exec->attr[FRAG_ATTRIB_WPOS][0][j];
-      const float y = exec->attr[FRAG_ATTRIB_WPOS][1][j];
-      exec->attr[attrib][i][j] = (exec->coef[attrib].a0[i] +
-				  exec->coef[attrib].dadx[i] * x + 
-				  exec->coef[attrib].dady[i] * y);
-   }
-}
-
-
-/**
- * Compute quad's attribute values by linear interpolation with
- * perspective correction.
- *
- * Push into the fp:
- * 
- *   INPUT[attr] = MAD COEF_DADX[attr], INPUT_WPOS.xxxx, COEF_A0[attr]
- *   INPUT[attr] = MAD COEF_DADY[attr], INPUT_WPOS.yyyy, INPUT[attr]
- *   TMP         = RCP INPUT_WPOS.w
- *   INPUT[attr] = MUL INPUT[attr], TMP.xxxx
- *
- */
-static INLINE void pinterp( struct exec_machine *exec,
-			    unsigned attrib,
-			    unsigned i )
-{
-   unsigned j;
-
-   for (j = 0; j < QUAD_SIZE; j++) {
-      const float x = exec->attr[FRAG_ATTRIB_WPOS][0][j];
-      const float y = exec->attr[FRAG_ATTRIB_WPOS][1][j];
-      /* FRAG_ATTRIB_WPOS.w here is really 1/w */
-      const float w = 1.0 / exec->attr[FRAG_ATTRIB_WPOS][3][j];
-      exec->attr[attrib][i][j] = ((exec->coef[attrib].a0[i] +
-				   exec->coef[attrib].dadx[i] * x + 
-				   exec->coef[attrib].dady[i] * y) * w);
-   }
-}
-
-
 /* This should be done by the fragment shader execution unit (code
  * generated from the decl instructions).  Do it here for now.
  */
 static void
-shade_quad( struct quad_stage *qs, struct quad_header *quad )
+shade_quad(
+   struct quad_stage *qs,
+   struct quad_header *quad )
 {
-   struct quad_shade_stage *qss = quad_shade_stage(qs);
+   struct quad_shade_stage *qss = quad_shade_stage( qs );
    struct softpipe_context *softpipe = qs->softpipe;
-   struct exec_machine exec;
    const float fx = quad->x0;
    const float fy = quad->y0;
    unsigned attr, i;
    struct tgsi_exec_machine machine;
 
 #if USE_ALIGNED_ATTRIBS
-   struct tgsi_exec_vector outputs[FRAG_ATTRIB_MAX] ALIGN16_SUFFIX;
+   struct tgsi_exec_vector inputs[PIPE_ATTRIB_MAX] ALIGN16_SUFFIX;
+   struct tgsi_exec_vector outputs[PIPE_ATTRIB_MAX] ALIGN16_SUFFIX;
 #else
-   struct tgsi_exec_vector inputs[FRAG_ATTRIB_MAX + 1];
-   struct tgsi_exec_vector outputs[FRAG_ATTRIB_MAX + 1];
+   struct tgsi_exec_vector inputs[PIPE_ATTRIB_MAX + 1];
+   struct tgsi_exec_vector outputs[PIPE_ATTRIB_MAX + 1];
 #endif
 
-   exec.coef = quad->coef;
-
-   /* Position:
-    */
-   exec.attr[FRAG_ATTRIB_WPOS][0][0] = fx;
-   exec.attr[FRAG_ATTRIB_WPOS][0][1] = fx + 1.0;
-   exec.attr[FRAG_ATTRIB_WPOS][0][2] = fx;
-   exec.attr[FRAG_ATTRIB_WPOS][0][3] = fx + 1.0;
-
-   exec.attr[FRAG_ATTRIB_WPOS][1][0] = fy;
-   exec.attr[FRAG_ATTRIB_WPOS][1][1] = fy;
-   exec.attr[FRAG_ATTRIB_WPOS][1][2] = fy + 1.0;
-   exec.attr[FRAG_ATTRIB_WPOS][1][3] = fy + 1.0;
-
-   /* Z and W are done by linear interpolation */
-   if (softpipe->need_z) {
-      linterp(&exec, 0, 2);   /* attr[0].z */
-   }
-
-   if (softpipe->need_w) {
-      linterp(&exec, 0, 3);  /* attr[0].w */
-      /*invert(&exec, 0, 3);*/
-   }
-
-   /* Interpolate all the remaining attributes.  This will get pushed
-    * into the fragment program's responsibilities at some point.
-    * Start at 1 to skip fragment position attribute (computed above).
-    */
-   for (attr = 1; attr < quad->nr_attrs; attr++) {
-      switch (softpipe->interp[attr]) {
-      case INTERP_CONSTANT:
-	 for (i = 0; i < NUM_CHANNELS; i++)
-	    cinterp(&exec, attr, i);
-	 break;
-
-      case INTERP_LINEAR:
-	 for (i = 0; i < NUM_CHANNELS; i++)
-	    linterp(&exec, attr, i);
-	 break;
-
-      case INTERP_PERSPECTIVE:
-	 for (i = 0; i < NUM_CHANNELS; i++)
-	    pinterp(&exec, attr, i);
-	 break;
-      }
-   }
-
 #ifdef DEBUG
    memset( &machine, 0, sizeof( machine ) );
 #endif
 
-   assert( sizeof( struct tgsi_exec_vector ) == sizeof( exec.attr[0] ) );
-
    /* init machine state */
    tgsi_exec_machine_init(
       &machine,
@@ -228,33 +101,40 @@ shade_quad( struct quad_stage *qs, struct quad_header *quad )
    machine.Consts = softpipe->fs.constants->constant;
 
 #if USE_ALIGNED_ATTRIBS
-   machine.Inputs = (struct tgsi_exec_vector *) exec.attr;
+   machine.Inputs = inputs;
    machine.Outputs = outputs;
 #else
    machine.Inputs = (struct tgsi_exec_vector *) tgsi_align_128bit( inputs );
    machine.Outputs = (struct tgsi_exec_vector *) tgsi_align_128bit( outputs );
-
-   memcpy(
-      machine.Inputs,
-      exec.attr,
-      softpipe->nr_attrs * sizeof( struct tgsi_exec_vector ) );
 #endif
 
+   machine.InterpCoefs = quad->coef;
+
+   machine.Inputs[0].xyzw[0].f[0] = fx;
+   machine.Inputs[0].xyzw[0].f[1] = fx + 1.0;
+   machine.Inputs[0].xyzw[0].f[2] = fx;
+   machine.Inputs[0].xyzw[0].f[3] = fx + 1.0;
+
+   machine.Inputs[0].xyzw[1].f[0] = fy;
+   machine.Inputs[0].xyzw[1].f[1] = fy;
+   machine.Inputs[0].xyzw[1].f[2] = fy + 1.0;
+   machine.Inputs[0].xyzw[1].f[3] = fy + 1.0;
+
    /* run shader */
    tgsi_exec_machine_run( &machine );
 
    /* store result color */
    memcpy(
       quad->outputs.color,
-      &machine.Outputs[FRAG_ATTRIB_COL0].xyzw[0].f[0],
+      &machine.Outputs[1].xyzw[0].f[0],
       sizeof( quad->outputs.color ) );
 
    if( softpipe->need_z ) {
       /* XXX temporary */
-      quad->outputs.depth[0] = exec.attr[0][2][0];
-      quad->outputs.depth[1] = exec.attr[0][2][1];
-      quad->outputs.depth[2] = exec.attr[0][2][2];
-      quad->outputs.depth[3] = exec.attr[0][2][3];
+      memcpy(
+         quad->outputs.depth,
+         &machine.Outputs[0].xyzw[2],
+         sizeof( quad->outputs.depth ) );
    }
 
    /* shader may cull fragments */
diff --git a/src/mesa/pipe/softpipe/sp_state_derived.c b/src/mesa/pipe/softpipe/sp_state_derived.c
index 0a2cfbb7d1..cd67d1c46f 100644
--- a/src/mesa/pipe/softpipe/sp_state_derived.c
+++ b/src/mesa/pipe/softpipe/sp_state_derived.c
@@ -87,7 +87,7 @@ static void calculate_vertex_layout( struct softpipe_context *softpipe )
     * fragment position (XYZW).
     */
    if (softpipe->depth_test.enabled ||
-       (inputsRead & FRAG_ATTRIB_WPOS))
+       (inputsRead & (1 << FRAG_ATTRIB_WPOS)))
       softpipe->need_z = TRUE;
    else
       softpipe->need_z = FALSE;
@@ -95,7 +95,7 @@ static void calculate_vertex_layout( struct softpipe_context *softpipe )
    /* Need W if we do any perspective-corrected interpolation or the
     * fragment program uses the fragment position.
     */
-   if (inputsRead & FRAG_ATTRIB_WPOS)
+   if (inputsRead & (1 << FRAG_ATTRIB_WPOS))
       softpipe->need_w = TRUE;
    else
       softpipe->need_w = FALSE;
diff --git a/src/mesa/pipe/tgsi/core/tgsi_build.c b/src/mesa/pipe/tgsi/core/tgsi_build.c
index d9c9a45532..d6f8af656a 100644
--- a/src/mesa/pipe/tgsi/core/tgsi_build.c
+++ b/src/mesa/pipe/tgsi/core/tgsi_build.c
@@ -88,6 +88,7 @@ tgsi_default_declaration( void )
    declaration.Size = 1;
    declaration.File = TGSI_FILE_NULL;
    declaration.Declare = TGSI_DECLARE_RANGE;
+   declaration.UsageMask = TGSI_WRITEMASK_XYZW;
    declaration.Interpolate = 0;
    declaration.Semantic = 0;
    declaration.Padding = 0;
@@ -100,6 +101,7 @@ struct tgsi_declaration
 tgsi_build_declaration(
    unsigned file,
    unsigned declare,
+   unsigned usage_mask,
    unsigned interpolate,
    unsigned semantic,
    struct tgsi_header *header )
@@ -112,6 +114,7 @@ tgsi_build_declaration(
    declaration = tgsi_default_declaration();
    declaration.File = file;
    declaration.Declare = declare;
+   declaration.UsageMask = usage_mask;
    declaration.Interpolate = interpolate;
    declaration.Semantic = semantic;
 
@@ -162,6 +165,7 @@ tgsi_build_full_declaration(
    *declaration = tgsi_build_declaration(
       full_decl->Declaration.File,
       full_decl->Declaration.Declare,
+      full_decl->Declaration.UsageMask,
       full_decl->Declaration.Interpolate,
       full_decl->Declaration.Semantic,
       header );
diff --git a/src/mesa/pipe/tgsi/core/tgsi_build.h b/src/mesa/pipe/tgsi/core/tgsi_build.h
index b3eb0715ee..116c78abf3 100644
--- a/src/mesa/pipe/tgsi/core/tgsi_build.h
+++ b/src/mesa/pipe/tgsi/core/tgsi_build.h
@@ -38,6 +38,7 @@ struct tgsi_declaration
 tgsi_build_declaration(
    unsigned file,
    unsigned declare,
+   unsigned usage_mask,
    unsigned interpolate,
    unsigned semantic,
    struct tgsi_header *header );
diff --git a/src/mesa/pipe/tgsi/core/tgsi_dump.c b/src/mesa/pipe/tgsi/core/tgsi_dump.c
index 5a9f92004e..622d617e29 100644
--- a/src/mesa/pipe/tgsi/core/tgsi_dump.c
+++ b/src/mesa/pipe/tgsi/core/tgsi_dump.c
@@ -633,6 +633,22 @@ dump_declaration_short(
       assert( 0 );
    }
 
+   if( decl->Declaration.UsageMask != TGSI_WRITEMASK_XYZW ) {
+      CHR( '.' );
+      if( decl->Declaration.UsageMask & TGSI_WRITEMASK_X ) {
+         CHR( 'x' );
+      }
+      if( decl->Declaration.UsageMask & TGSI_WRITEMASK_Y ) {
+         CHR( 'y' );
+      }
+      if( decl->Declaration.UsageMask & TGSI_WRITEMASK_Z ) {
+         CHR( 'z' );
+      }
+      if( decl->Declaration.UsageMask & TGSI_WRITEMASK_W ) {
+         CHR( 'w' );
+      }
+   }
+
    if( decl->Declaration.Interpolate ) {
       TXT( ", " );
       ENM( decl->Interpolation.Interpolate, TGSI_INTERPOLATES_SHORT );
@@ -659,6 +675,21 @@ dump_declaration_verbose(
    ENM( decl->Declaration.File, TGSI_FILES );
    TXT( "\nDeclare    : " );
    ENM( decl->Declaration.Declare, TGSI_DECLARES );
+   if( deflt || fd->Declaration.UsageMask != decl->Declaration.UsageMask ) {
+      TXT( "\nUsageMask  : " );
+      if( decl->Declaration.UsageMask & TGSI_WRITEMASK_X ) {
+         CHR( 'X' );
+      }
+      if( decl->Declaration.UsageMask & TGSI_WRITEMASK_Y ) {
+         CHR( 'Y' );
+      }
+      if( decl->Declaration.UsageMask & TGSI_WRITEMASK_Z ) {
+         CHR( 'Z' );
+      }
+      if( decl->Declaration.UsageMask & TGSI_WRITEMASK_W ) {
+         CHR( 'W' );
+      }
+   }
    if( deflt || fd->Declaration.Interpolate != decl->Declaration.Interpolate ) {
       TXT( "\nInterpolate: " );
       UID( decl->Declaration.Interpolate );
diff --git a/src/mesa/pipe/tgsi/core/tgsi_exec.c b/src/mesa/pipe/tgsi/core/tgsi_exec.c
index 0a34754c40..0ffed20066 100644
--- a/src/mesa/pipe/tgsi/core/tgsi_exec.c
+++ b/src/mesa/pipe/tgsi/core/tgsi_exec.c
@@ -62,7 +62,7 @@
 void
 tgsi_exec_machine_init(
    struct tgsi_exec_machine *mach,
-   struct tgsi_token *tokens,
+   const struct tgsi_token *tokens,
    GLuint numSamplers,
    struct tgsi_sampler *samplers)
 {
@@ -1063,7 +1063,131 @@ fetch_texel( struct tgsi_sampler *sampler,
    }
 }
 
+static void
+constant_interpolation(
+   struct tgsi_exec_machine *mach,
+   unsigned attrib,
+   unsigned chan )
+{
+   unsigned i;
+
+   for( i = 0; i < QUAD_SIZE; i++ ) {
+      mach->Inputs[attrib].xyzw[chan].f[i] = mach->InterpCoefs[attrib].a0[chan];
+   }
+}
+
+static void
+linear_interpolation(
+   struct tgsi_exec_machine *mach,
+   unsigned attrib,
+   unsigned chan )
+{
+   unsigned i;
+
+   for( i = 0; i < QUAD_SIZE; i++ ) {
+      const float x = mach->Inputs[0].xyzw[0].f[i];
+      const float y = mach->Inputs[0].xyzw[1].f[i];
+
+      mach->Inputs[attrib].xyzw[chan].f[i] =
+         mach->InterpCoefs[attrib].a0[chan] +
+         mach->InterpCoefs[attrib].dadx[chan] * x +
+         mach->InterpCoefs[attrib].dady[chan] * y;
+   }
+}
+
+static void
+perspective_interpolation(
+   struct tgsi_exec_machine *mach,
+   unsigned attrib,
+   unsigned chan )
+{
+   unsigned i;
+
+   for( i = 0; i < QUAD_SIZE; i++ ) {
+      const float x = mach->Inputs[0].xyzw[0].f[i];
+      const float y = mach->Inputs[0].xyzw[1].f[i];
+      // WPOS.w here is really 1/w
+      const float w = 1.0f / mach->Inputs[0].xyzw[3].f[i];
+
+      mach->Inputs[attrib].xyzw[chan].f[i] =
+         (mach->InterpCoefs[attrib].a0[chan] +
+         mach->InterpCoefs[attrib].dadx[chan] * x +
+         mach->InterpCoefs[attrib].dady[chan] * y) * w;
+   }
+}
+
+typedef void (* interpolation_func)(
+   struct tgsi_exec_machine *mach,
+   unsigned attrib,
+   unsigned chan );
+
+static void
+exec_declaration(
+   struct tgsi_exec_machine *mach,
+   const struct tgsi_full_declaration *decl )
+{
+   if( mach->Processor == TGSI_PROCESSOR_FRAGMENT ) {
+      if( decl->Declaration.File == TGSI_FILE_INPUT ) {
+         unsigned first, last, mask, i, j;
+         interpolation_func interp;
+
+         assert( decl->Declaration.Declare == TGSI_DECLARE_RANGE );
+
+         first = decl->u.DeclarationRange.First;
+         last = decl->u.DeclarationRange.Last;
+         mask = decl->Declaration.UsageMask;
+
+         /* Do not touch WPOS.xy */
+         if( first == 0 ) {
+            mask &= ~TGSI_WRITEMASK_XY;
+            if( mask == TGSI_WRITEMASK_NONE ) {
+               first++;
+               if( first > last ) {
+                  return;
+               }
+            }
+         }
+
+         switch( decl->Interpolation.Interpolate ) {
+         case TGSI_INTERPOLATE_CONSTANT:
+            interp = constant_interpolation;
+            break;
+
+         case TGSI_INTERPOLATE_LINEAR:
+            interp = linear_interpolation;
+            break;
 
+         case TGSI_INTERPOLATE_PERSPECTIVE:
+            interp = perspective_interpolation;
+            break;
+
+         default:
+            assert( 0 );
+         }
+
+         if( mask == TGSI_WRITEMASK_XYZW ) {
+            unsigned i, j;
+
+            for( i = first; i <= last; i++ ) {
+               for( j = 0; j < NUM_CHANNELS; j++ ) {
+                  interp( mach, i, j );
+               }
+            }
+         }
+         else {
+            unsigned i, j;
+
+            for( j = 0; j < NUM_CHANNELS; j++ ) {
+               if( mask & (1 << j) ) {
+                  for( i = first; i <= last; i++ ) {
+                     interp( mach, i, j );
+                  }
+               }
+            }
+         }
+      }
+   }
+}
 
 static void
 exec_instruction(
@@ -2161,6 +2285,7 @@ tgsi_exec_machine_run2(
       tgsi_parse_token( &parse );
       switch( parse.FullToken.Token.Type ) {
       case TGSI_TOKEN_TYPE_DECLARATION:
+         exec_declaration( mach, &parse.FullToken.FullDeclaration );
          break;
       case TGSI_TOKEN_TYPE_IMMEDIATE:
          break;
diff --git a/src/mesa/pipe/tgsi/core/tgsi_exec.h b/src/mesa/pipe/tgsi/core/tgsi_exec.h
index eed2207d7d..e5e8c3608e 100644
--- a/src/mesa/pipe/tgsi/core/tgsi_exec.h
+++ b/src/mesa/pipe/tgsi/core/tgsi_exec.h
@@ -11,23 +11,27 @@
 extern "C" {
 #endif // defined __cplusplus
 
+#define NUM_CHANNELS 4  /* R,G,B,A */
+#define QUAD_SIZE    4  /* 4 pixel/quad */
+
 union tgsi_exec_channel
 {
-   float  f[4];
-   int    i[4];
-   unsigned   u[4];
+   float    f[QUAD_SIZE];
+   int      i[QUAD_SIZE];
+   unsigned u[QUAD_SIZE];
 };
 
 struct tgsi_exec_vector
 {
-   union tgsi_exec_channel xyzw[4];
+   union tgsi_exec_channel xyzw[NUM_CHANNELS];
 };
 
-
-#define NUM_CHANNELS 4  /* R,G,B,A */
-#ifndef QUAD_SIZE
-#define QUAD_SIZE 4     /* 4 pixel/quad */
-#endif
+struct tgsi_interp_coef
+{
+   float a0[NUM_CHANNELS];	/* in an xyzw layout */
+   float dadx[NUM_CHANNELS];
+   float dady[NUM_CHANNELS];
+};
 
 #define TEX_CACHE_TILE_SIZE 8
 #define TEX_CACHE_NUM_ENTRIES 8
@@ -55,8 +59,8 @@ struct tgsi_sampler
 
 struct tgsi_exec_labels
 {
-   unsigned   labels[128][2];
-   unsigned   count;
+   unsigned labels[128][2];
+   unsigned count;
 };
 
 #define TGSI_EXEC_TEMP_00000000_I   32
@@ -109,15 +113,15 @@ struct tgsi_exec_cond_state
 {
    struct tgsi_exec_cond_regs IfPortion;
    struct tgsi_exec_cond_regs ElsePortion;
-   unsigned                     Condition;
-   boolean                  WasElse;
+   unsigned                   Condition;
+   boolean                    WasElse;
 };
 
 /* XXX: This is temporary */
 struct tgsi_exec_cond_stack
 {
    struct tgsi_exec_cond_state   States[8];
-   unsigned                        Index;      /* into States[] */
+   unsigned                      Index;      /* into States[] */
 };
 
 struct tgsi_exec_machine
@@ -138,15 +142,19 @@ struct tgsi_exec_machine
 
    struct tgsi_sampler           *Samplers;
 
-   float                       Imms[256][4];
-   unsigned                        ImmLimit;
-   float                       (*Consts)[4];
-   const struct tgsi_exec_vector *Inputs;
+   float                         Imms[256][4];
+   unsigned                      ImmLimit;
+   float                         (*Consts)[4];
+   struct tgsi_exec_vector       *Inputs;
    struct tgsi_exec_vector       *Outputs;
-   struct tgsi_token             *Tokens;
-   unsigned                        Processor;
+   const struct tgsi_token       *Tokens;
+   unsigned                      Processor;
+
+   /* GEOMETRY processor only. */
+   unsigned                      *Primitives;
 
-   unsigned                        *Primitives;
+   /* FRAGMENT processor only. */
+   const struct tgsi_interp_coef *InterpCoefs;
 
    struct tgsi_exec_cond_stack   CondStack;
 #if XXX_SSE
@@ -157,7 +165,7 @@ struct tgsi_exec_machine
 void
 tgsi_exec_machine_init(
    struct tgsi_exec_machine *mach,
-   struct tgsi_token *tokens,
+   const struct tgsi_token *tokens,
    unsigned numSamplers,
    struct tgsi_sampler *samplers);
 
diff --git a/src/mesa/pipe/tgsi/core/tgsi_token.h b/src/mesa/pipe/tgsi/core/tgsi_token.h
index f72c7dc4be..3ed341fb7b 100644
--- a/src/mesa/pipe/tgsi/core/tgsi_token.h
+++ b/src/mesa/pipe/tgsi/core/tgsi_token.h
@@ -52,15 +52,33 @@ struct tgsi_token
 #define TGSI_DECLARE_RANGE    0
 #define TGSI_DECLARE_MASK     1
 
+#define TGSI_WRITEMASK_NONE     0x00
+#define TGSI_WRITEMASK_X        0x01
+#define TGSI_WRITEMASK_Y        0x02
+#define TGSI_WRITEMASK_XY       0x03
+#define TGSI_WRITEMASK_Z        0x04
+#define TGSI_WRITEMASK_XZ       0x05
+#define TGSI_WRITEMASK_YZ       0x06
+#define TGSI_WRITEMASK_XYZ      0x07
+#define TGSI_WRITEMASK_W        0x08
+#define TGSI_WRITEMASK_XW       0x09
+#define TGSI_WRITEMASK_YW       0x0A
+#define TGSI_WRITEMASK_XYW      0x0B
+#define TGSI_WRITEMASK_ZW       0x0C
+#define TGSI_WRITEMASK_XZW      0x0D
+#define TGSI_WRITEMASK_YZW      0x0E
+#define TGSI_WRITEMASK_XYZW     0x0F
+
 struct tgsi_declaration
 {
    unsigned Type        : 4;  /* TGSI_TOKEN_TYPE_DECLARATION */
    unsigned Size        : 8;  /* UINT */
    unsigned File        : 4;  /* TGSI_FILE_ */
    unsigned Declare     : 4;  /* TGSI_DECLARE_ */
+   unsigned UsageMask   : 4;  /* TGSI_WRITEMASK_ */
    unsigned Interpolate : 1;  /* BOOL */
    unsigned Semantic    : 1;  /* BOOL */
-   unsigned Padding     : 9;
+   unsigned Padding     : 5;
    unsigned Extended    : 1;  /* BOOL */
 };
 
@@ -1226,23 +1244,6 @@ struct tgsi_instruction_ext_texture
    unsigned Extended : 1;    /* BOOL */
 };
 
-#define TGSI_WRITEMASK_NONE     0x00
-#define TGSI_WRITEMASK_X        0x01
-#define TGSI_WRITEMASK_Y        0x02
-#define TGSI_WRITEMASK_XY       0x03
-#define TGSI_WRITEMASK_Z        0x04
-#define TGSI_WRITEMASK_XZ       0x05
-#define TGSI_WRITEMASK_YZ       0x06
-#define TGSI_WRITEMASK_XYZ      0x07
-#define TGSI_WRITEMASK_W        0x08
-#define TGSI_WRITEMASK_XW       0x09
-#define TGSI_WRITEMASK_YW       0x0A
-#define TGSI_WRITEMASK_XYW      0x0B
-#define TGSI_WRITEMASK_ZW       0x0C
-#define TGSI_WRITEMASK_XZW      0x0D
-#define TGSI_WRITEMASK_YZW      0x0E
-#define TGSI_WRITEMASK_XYZW     0x0F
-
 struct tgsi_instruction_ext_predicate
 {
    unsigned Type             : 4;    /* TGSI_INSTRUCTION_EXT_TYPE_PREDICATE */
diff --git a/src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.c b/src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.c
index 609d329242..2ac04e8d56 100644
--- a/src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.c
+++ b/src/mesa/pipe/tgsi/mesa/mesa_to_tgsi.c
@@ -467,13 +467,15 @@ static struct tgsi_full_declaration
 make_frag_input_decl(

    GLuint first,

    GLuint last,

-   GLuint interpolate )

+   GLuint interpolate,

+   GLuint usage_mask )

 {

    struct tgsi_full_declaration decl;

 

    decl = tgsi_default_full_declaration();

    decl.Declaration.File = TGSI_FILE_INPUT;

    decl.Declaration.Declare = TGSI_DECLARE_RANGE;

+   decl.Declaration.UsageMask = usage_mask;

    decl.Declaration.Interpolate = 1;

    decl.u.DeclarationRange.First = first;

    decl.u.DeclarationRange.Last = last;

@@ -485,13 +487,15 @@ make_frag_input_decl(
 static struct tgsi_full_declaration

 make_frag_output_decl(

    GLuint index,

-   GLuint semantic_name )

+   GLuint semantic_name,

+   GLuint usage_mask )

 {

    struct tgsi_full_declaration decl;

 

    decl = tgsi_default_full_declaration();

    decl.Declaration.File = TGSI_FILE_OUTPUT;

    decl.Declaration.Declare = TGSI_DECLARE_RANGE;

+   decl.Declaration.UsageMask = usage_mask;

    decl.Declaration.Semantic = 1;

    decl.u.DeclarationRange.First = index;

    decl.u.DeclarationRange.Last = index;

@@ -514,6 +518,7 @@ tgsi_mesa_compile_fp_program(
    struct tgsi_full_dst_register *fulldst;

    struct tgsi_full_src_register *fullsrc;

    GLuint inputs_read;

+   GLboolean reads_wpos;

    GLuint preamble_size = 0;

 

    *(struct tgsi_version *) &tokens[0] = tgsi_build_version();

@@ -523,19 +528,33 @@ tgsi_mesa_compile_fp_program(
 

    ti = 2;

 

-   /*

-    * Input 0 is always read, at least implicitly by the MOV instruction generated

-    * below, so mark it as used.

-    */

-   inputs_read = program->Base.InputsRead | 1;

+   reads_wpos = program->Base.InputsRead & (1 << FRAG_ATTRIB_WPOS);

+   inputs_read = program->Base.InputsRead | (1 << FRAG_ATTRIB_WPOS);

 

    /*

     * Declare input attributes. Note that we do not interpolate fragment position.

     */

+

+   /* Fragment position. */

+   if( reads_wpos ) {

+      fulldecl = make_frag_input_decl(

+         0,

+         0,

+         TGSI_INTERPOLATE_CONSTANT,

+         TGSI_WRITEMASK_XY );

+      ti += tgsi_build_full_declaration(

+         &fulldecl,

+         &tokens[ti],

+         header,

+         maxTokens - ti );

+   }

+

+   /* Fragment zw. */

    fulldecl = make_frag_input_decl(

       0,

       0,

-      TGSI_INTERPOLATE_CONSTANT );

+      TGSI_INTERPOLATE_LINEAR,

+      reads_wpos ? TGSI_WRITEMASK_ZW : TGSI_WRITEMASK_Z );

    ti += tgsi_build_full_declaration(

       &fulldecl,

       &tokens[ti],

@@ -552,7 +571,8 @@ tgsi_mesa_compile_fp_program(
       fulldecl = make_frag_input_decl(

          1,

          1 + count - 1,

-         TGSI_INTERPOLATE_LINEAR );

+         TGSI_INTERPOLATE_LINEAR,

+         TGSI_WRITEMASK_XYZW );

       ti += tgsi_build_full_declaration(

          &fulldecl,

          &tokens[ti],

@@ -569,7 +589,8 @@ tgsi_mesa_compile_fp_program(
 

    fulldecl = make_frag_output_decl(

       0,

-      TGSI_SEMANTIC_DEPTH );

+      TGSI_SEMANTIC_DEPTH,

+      TGSI_WRITEMASK_Z );

    ti += tgsi_build_full_declaration(

       &fulldecl,

       &tokens[ti],

@@ -579,7 +600,8 @@ tgsi_mesa_compile_fp_program(
    if( program->Base.OutputsWritten & (1 << FRAG_RESULT_COLR) ) {

       fulldecl = make_frag_output_decl(

          1,

-         TGSI_SEMANTIC_COLOR );

+         TGSI_SEMANTIC_COLOR,

+         TGSI_WRITEMASK_XYZW );

       ti += tgsi_build_full_declaration(

          &fulldecl,

          &tokens[ti],

@@ -587,38 +609,6 @@ tgsi_mesa_compile_fp_program(
          maxTokens - ti );

    }

 

-   /*

-    * Copy input fragment xyz to output xyz.

-    * If the shader writes depth, do not copy the z component.

-    */

-

-   fullinst = tgsi_default_full_instruction();

-

-   fullinst.Instruction.Opcode = TGSI_OPCODE_MOV;

-   fullinst.Instruction.NumDstRegs = 1;

-   fullinst.Instruction.NumSrcRegs = 1;

-

-   fulldst = &fullinst.FullDstRegisters[0];

-   fulldst->DstRegister.File = TGSI_FILE_OUTPUT;

-   fulldst->DstRegister.Index = 0;

-   if( program->Base.OutputsWritten & (1 << FRAG_RESULT_DEPR) ) {

-      fulldst->DstRegister.WriteMask = TGSI_WRITEMASK_XY;

-   }

-   else {

-      fulldst->DstRegister.WriteMask = TGSI_WRITEMASK_XYZ;

-   }

-

-   fullsrc = &fullinst.FullSrcRegisters[0];

-   fullsrc->SrcRegister.File = TGSI_FILE_INPUT;

-   fullsrc->SrcRegister.Index = 0;

-

-   ti += tgsi_build_full_instruction(

-      &fullinst,

-      &tokens[ti],

-      header,

-      maxTokens - ti );

-   preamble_size++;

-

    for( i = 0; i < program->Base.NumInstructions; i++ ) {

       if( compile_instruction(

             &program->Base.Instructions[i],