i965g: get basic texturing working again

Revert to fixed-layout surface binding table -- it's probably the best
way to do this.  Pass sampler and texture numbers separately even
though we're always keeping them the same at present.

9 files changed, 117 insertions, 66 deletions

diff --git a/src/gallium/drivers/i965/brw_context.h b/src/gallium/drivers/i965/brw_context.h
index 096c8cf12b..598e747fe0 100644
--- a/src/gallium/drivers/i965/brw_context.h
+++ b/src/gallium/drivers/i965/brw_context.h
@@ -209,9 +209,9 @@ struct brw_fragment_shader {
 
 
 struct brw_sampler {
-   float border_color[4];
    struct brw_ss0 ss0;
    struct brw_ss1 ss1;
+   float border_color[4];
    struct brw_ss3 ss3;
 };
 
@@ -355,20 +355,23 @@ struct brw_vs_ouput_sizes {
 /** Number of texture sampler units */
 #define BRW_MAX_TEX_UNIT 16
 
+/** Max number of render targets in a shader */
+#define BRW_MAX_DRAW_BUFFERS 4
+
 /**
  * Size of our surface binding table for the WM.
  * This contains pointers to the drawing surfaces and current texture
  * objects and shader constant buffers (+2).
  */
-#define BRW_WM_MAX_SURF (PIPE_MAX_COLOR_BUFS + BRW_MAX_TEX_UNIT + 1)
+#define BRW_WM_MAX_SURF (BRW_MAX_DRAW_BUFFERS + BRW_MAX_TEX_UNIT + 1)
 
 /**
  * Helpers to convert drawing buffers, textures and constant buffers
  * to surface binding table indexes, for WM.
  */
-#define SURF_INDEX_DRAW(d)           (d)
-#define SURF_INDEX_FRAG_CONST_BUFFER (PIPE_MAX_COLOR_BUFS) 
-#define SURF_INDEX_TEXTURE(t)        (PIPE_MAX_COLOR_BUFS + 1 + (t))
+#define BTI_COLOR_BUF(d)          (d)
+#define BTI_FRAGMENT_CONSTANTS    (BRW_MAX_DRAW_BUFFERS) 
+#define BTI_TEXTURE(t)            (BRW_MAX_DRAW_BUFFERS + 1 + (t))
 
 /**
  * Size of surface binding table for the VS.
diff --git a/src/gallium/drivers/i965/brw_pipe_fb.c b/src/gallium/drivers/i965/brw_pipe_fb.c
index 1511220447..6b03094f50 100644
--- a/src/gallium/drivers/i965/brw_pipe_fb.c
+++ b/src/gallium/drivers/i965/brw_pipe_fb.c
@@ -31,7 +31,7 @@ static void brw_set_framebuffer_state( struct pipe_context *pipe,
 
    /* Color buffers:
     */
-   for (i = 0; i < MAX2(fb->nr_cbufs, brw->curr.fb.nr_cbufs); i++) {
+   for (i = 0; i < PIPE_MAX_COLOR_BUFS; i++) {
       if (brw->curr.fb.cbufs[i] != fb->cbufs[i]) {
 	 brw->state.dirty.mesa |= PIPE_NEW_COLOR_BUFFERS;
 	 pipe_surface_reference(&brw->curr.fb.cbufs[i], fb->cbufs[i]);
@@ -39,7 +39,7 @@ static void brw_set_framebuffer_state( struct pipe_context *pipe,
    }
    
    if (brw->curr.fb.nr_cbufs != fb->nr_cbufs) {
-      brw->curr.fb.nr_cbufs = fb->nr_cbufs;
+      brw->curr.fb.nr_cbufs = MIN2(BRW_MAX_DRAW_BUFFERS, fb->nr_cbufs);
       brw->state.dirty.mesa |= PIPE_NEW_NR_CBUFS;
    }
 }
diff --git a/src/gallium/drivers/i965/brw_pipe_sampler.c b/src/gallium/drivers/i965/brw_pipe_sampler.c
index f0a765ecf5..5cd38a43a6 100644
--- a/src/gallium/drivers/i965/brw_pipe_sampler.c
+++ b/src/gallium/drivers/i965/brw_pipe_sampler.c
@@ -107,7 +107,7 @@ static void *
 brw_create_sampler_state( struct pipe_context *pipe,
                           const struct pipe_sampler_state *template )
 {
-   struct brw_sampler_state *sampler = CALLOC_STRUCT(brw_sampler_state);
+   struct brw_sampler *sampler = CALLOC_STRUCT(brw_sampler);
 
    sampler->ss0.min_filter = translate_img_filter( template->min_img_filter );
    sampler->ss0.mag_filter = translate_img_filter( template->mag_img_filter );
@@ -214,7 +214,6 @@ void brw_pipe_sampler_init( struct brw_context *brw )
 
    brw->base.set_sampler_textures = brw_set_sampler_textures;
 }
-
 void brw_pipe_sampler_cleanup( struct brw_context *brw )
 {
 }
diff --git a/src/gallium/drivers/i965/brw_sf.c b/src/gallium/drivers/i965/brw_sf.c
index e1986a9dbb..a28fb71589 100644
--- a/src/gallium/drivers/i965/brw_sf.c
+++ b/src/gallium/drivers/i965/brw_sf.c
@@ -153,9 +153,10 @@ static enum pipe_error upload_sf_prog(struct brw_context *brw)
       case TGSI_INTERPOLATE_CONSTANT:
          break;
       case TGSI_INTERPOLATE_LINEAR:
+      case TGSI_INTERPOLATE_PERSPECTIVE:
          key.linear_attrs |= 1 << (i+1);
          break;
-      case TGSI_INTERPOLATE_PERSPECTIVE:
+//      case TGSI_INTERPOLATE_PERSPECTIVE:
          key.persp_attrs |= 1 << (i+1);
          break;
       }
diff --git a/src/gallium/drivers/i965/brw_wm.c b/src/gallium/drivers/i965/brw_wm.c
index 3c5a2dab7a..2c9d3e5e87 100644
--- a/src/gallium/drivers/i965/brw_wm.c
+++ b/src/gallium/drivers/i965/brw_wm.c
@@ -56,6 +56,15 @@ GLuint brw_wm_nr_args( GLuint opcode )
    case WM_FB_WRITE:
    case WM_PINTERP:
       return 3;
+   case TGSI_OPCODE_TEX:
+   case TGSI_OPCODE_TXP:
+   case TGSI_OPCODE_TXB:
+   case TGSI_OPCODE_TXD:
+      /* sampler arg is held as a field in the instruction, not in an
+       * actual register:
+       */
+      return tgsi_get_opcode_info(opcode)->num_src - 1;
+
    default:
       assert(opcode < MAX_OPCODE);
       return tgsi_get_opcode_info(opcode)->num_src;
diff --git a/src/gallium/drivers/i965/brw_wm.h b/src/gallium/drivers/i965/brw_wm.h
index b7d807dcb3..f1ca9f6369 100644
--- a/src/gallium/drivers/i965/brw_wm.h
+++ b/src/gallium/drivers/i965/brw_wm.h
@@ -135,6 +135,7 @@ struct brw_wm_instruction {
    GLuint opcode:8;
    GLuint saturate:1;
    GLuint writemask:4;
+   GLuint sampler:4;
    GLuint tex_unit:4;   /* texture/sampler unit for texture instructions */
    GLuint target:4;     /* TGSI_TEXTURE_x for texture instructions,
                          * target binding table index for FB_WRITE
@@ -201,7 +202,8 @@ struct brw_fp_instruction {
    unsigned opcode:8;
    unsigned target:8; /* XXX: special usage for FB_WRITE */
    unsigned tex_unit:4;
-   unsigned pad:12;
+   unsigned sampler:4;
+   unsigned pad:8;
 };
 
 
diff --git a/src/gallium/drivers/i965/brw_wm_emit.c b/src/gallium/drivers/i965/brw_wm_emit.c
index a14e12f35b..3250db1848 100644
--- a/src/gallium/drivers/i965/brw_wm_emit.c
+++ b/src/gallium/drivers/i965/brw_wm_emit.c
@@ -792,7 +792,8 @@ static void emit_tex( struct brw_wm_compile *c,
 		      const struct brw_wm_instruction *inst,
 		      struct brw_reg *dst,
 		      GLuint dst_flags,
-		      struct brw_reg *arg )
+		      struct brw_reg *coord,
+		      GLuint sampler)
 {
    struct brw_compile *p = &c->func;
    GLuint msgLength, responseLength;
@@ -838,7 +839,7 @@ static void emit_tex( struct brw_wm_compile *c,
    for (i = 0; i < nr; i++) {
       static const GLuint swz[4] = {0,1,2,2};
       if (emit & (1<<i)) 
-	 brw_MOV(p, brw_message_reg(msgLength+1), arg[swz[i]]);
+	 brw_MOV(p, brw_message_reg(msgLength+1), coord[swz[i]]);
       else
 	 brw_MOV(p, brw_message_reg(msgLength+1), brw_imm_f(0));
       msgLength += 2;
@@ -862,8 +863,8 @@ static void emit_tex( struct brw_wm_compile *c,
 	      retype(vec16(dst[0]), BRW_REGISTER_TYPE_UW),
 	      1,
 	      retype(c->payload.depth[0].hw_reg, BRW_REGISTER_TYPE_UW),
-              SURF_INDEX_TEXTURE(inst->tex_unit),
-	      inst->tex_unit,	  /* sampler */
+              BTI_TEXTURE(inst->tex_unit),
+	      sampler,          /* sampler index */
 	      inst->writemask,
 	      msg_type, 
 	      responseLength,
@@ -878,7 +879,8 @@ static void emit_txb( struct brw_wm_compile *c,
 		      const struct brw_wm_instruction *inst,
 		      struct brw_reg *dst,
 		      GLuint dst_flags,
-		      struct brw_reg *arg )
+		      struct brw_reg *coord,
+		      GLuint sampler )
 {
    struct brw_compile *p = &c->func;
    GLuint msgLength;
@@ -888,7 +890,7 @@ static void emit_txb( struct brw_wm_compile *c,
    switch (inst->target) {
    case TGSI_TEXTURE_1D:
    case TGSI_TEXTURE_SHADOW1D:
-      brw_MOV(p, brw_message_reg(2), arg[0]);
+      brw_MOV(p, brw_message_reg(2), coord[0]);
       brw_MOV(p, brw_message_reg(4), brw_imm_f(0));
       brw_MOV(p, brw_message_reg(6), brw_imm_f(0));
       break;
@@ -896,22 +898,22 @@ static void emit_txb( struct brw_wm_compile *c,
    case TGSI_TEXTURE_RECT:
    case TGSI_TEXTURE_SHADOW2D:
    case TGSI_TEXTURE_SHADOWRECT:
-      brw_MOV(p, brw_message_reg(2), arg[0]);
-      brw_MOV(p, brw_message_reg(4), arg[1]);
+      brw_MOV(p, brw_message_reg(2), coord[0]);
+      brw_MOV(p, brw_message_reg(4), coord[1]);
       brw_MOV(p, brw_message_reg(6), brw_imm_f(0));
       break;
    case TGSI_TEXTURE_3D:
    case TGSI_TEXTURE_CUBE:
-      brw_MOV(p, brw_message_reg(2), arg[0]);
-      brw_MOV(p, brw_message_reg(4), arg[1]);
-      brw_MOV(p, brw_message_reg(6), arg[2]);
+      brw_MOV(p, brw_message_reg(2), coord[0]);
+      brw_MOV(p, brw_message_reg(4), coord[1]);
+      brw_MOV(p, brw_message_reg(6), coord[2]);
       break;
    default:
       /* unexpected target */
       abort();
    }
 
-   brw_MOV(p, brw_message_reg(8), arg[3]);
+   brw_MOV(p, brw_message_reg(8), coord[3]);
    msgLength = 9;
 
    if (BRW_IS_IGDNG(p->brw))
@@ -923,8 +925,8 @@ static void emit_txb( struct brw_wm_compile *c,
 	      retype(vec16(dst[0]), BRW_REGISTER_TYPE_UW),
 	      1,
 	      retype(c->payload.depth[0].hw_reg, BRW_REGISTER_TYPE_UW),
-              SURF_INDEX_TEXTURE(inst->tex_unit),
-	      inst->tex_unit,	  /* sampler */
+              BTI_TEXTURE(inst->tex_unit),
+	      sampler,          /* sampler index */
 	      inst->writemask,
 	      msg_type,
 	      8,		/* responseLength */
@@ -1483,11 +1485,11 @@ void brw_wm_emit( struct brw_wm_compile *c )
 	 /* Texturing operations:
 	  */
       case TGSI_OPCODE_TEX:
-	 emit_tex(c, inst, dst, dst_flags, args[0]);
+	 emit_tex(c, inst, dst, dst_flags, args[0], inst->sampler);
 	 break;
 
       case TGSI_OPCODE_TXB:
-	 emit_txb(c, inst, dst, dst_flags, args[0]);
+	 emit_txb(c, inst, dst, dst_flags, args[0], inst->sampler);
 	 break;
 
       case TGSI_OPCODE_KIL:
diff --git a/src/gallium/drivers/i965/brw_wm_fp.c b/src/gallium/drivers/i965/brw_wm_fp.c
index 174486a101..a8b5e15f36 100644
--- a/src/gallium/drivers/i965/brw_wm_fp.c
+++ b/src/gallium/drivers/i965/brw_wm_fp.c
@@ -282,6 +282,7 @@ static struct brw_fp_instruction * emit_tex_op(struct brw_wm_compile *c,
 					     struct brw_fp_dst dest,
 					     GLuint tex_unit,
 					     GLuint target,
+					     GLuint sampler,
 					     struct brw_fp_src src0,
 					     struct brw_fp_src src1,
 					     struct brw_fp_src src2 )
@@ -298,6 +299,7 @@ static struct brw_fp_instruction * emit_tex_op(struct brw_wm_compile *c,
    inst->dst = dest;
    inst->tex_unit = tex_unit;
    inst->target = target;
+   inst->sampler = sampler;
    inst->src[0] = src0;
    inst->src[1] = src1;
    inst->src[2] = src2;
@@ -313,7 +315,7 @@ static INLINE void emit_op3(struct brw_wm_compile *c,
 			    struct brw_fp_src src1,
 			    struct brw_fp_src src2 )
 {
-   emit_tex_op(c, op, dest, 0, 0, src0, src1, src2);
+   emit_tex_op(c, op, dest, 0, 0, 0, src0, src1, src2);
 }
 
 
@@ -323,7 +325,7 @@ static INLINE void emit_op2(struct brw_wm_compile *c,
 			    struct brw_fp_src src0,
 			    struct brw_fp_src src1)
 {
-   emit_tex_op(c, op, dest, 0, 0, src0, src1, src_undef());
+   emit_tex_op(c, op, dest, 0, 0, 0, src0, src1, src_undef());
 }
 
 static INLINE void emit_op1(struct brw_wm_compile *c,
@@ -331,14 +333,14 @@ static INLINE void emit_op1(struct brw_wm_compile *c,
 			    struct brw_fp_dst dest,
 			    struct brw_fp_src src0)
 {
-   emit_tex_op(c, op, dest, 0, 0, src0, src_undef(), src_undef());
+   emit_tex_op(c, op, dest, 0, 0, 0, src0, src_undef(), src_undef());
 }
 
 static INLINE void emit_op0(struct brw_wm_compile *c,
 			   GLuint op,
 			   struct brw_fp_dst dest)
 {
-   emit_tex_op(c, op, dest, 0, 0, src_undef(), src_undef(), src_undef());
+   emit_tex_op(c, op, dest, 0, 0, 0, src_undef(), src_undef(), src_undef());
 }
 
 
@@ -674,7 +676,8 @@ static void precalc_tex( struct brw_wm_compile *c,
 			 struct brw_fp_dst dst,
 			 unsigned target,
 			 unsigned unit,
-			 struct brw_fp_src src0 )
+			 struct brw_fp_src src0,
+			 struct brw_fp_src sampler )
 {
    struct brw_fp_src coord = src_undef();
    struct brw_fp_dst tmp = dst_undef();
@@ -751,6 +754,7 @@ static void precalc_tex( struct brw_wm_compile *c,
                   dst_saturate(tmp, dst.saturate),
                   unit,
                   target,
+                  sampler.index,
                   coord,
                   src_undef(),
                   src_undef());
@@ -802,6 +806,7 @@ static void precalc_tex( struct brw_wm_compile *c,
                   dst,
                   unit,
                   target,
+                  sampler.index,
                   coord,
                   src_undef(),
                   src_undef());
@@ -851,7 +856,8 @@ static void precalc_txp( struct brw_wm_compile *c,
 			 struct brw_fp_dst dst,
 			 unsigned target,
 			 unsigned unit,
-			 struct brw_fp_src src0 )
+			 struct brw_fp_src src0,
+                         struct brw_fp_src sampler )
 {
    if (projtex(c, target, src0)) {
       struct brw_fp_dst tmp = get_temp(c);
@@ -877,7 +883,8 @@ static void precalc_txp( struct brw_wm_compile *c,
 		  dst,
 		  target,
 		  unit,
-		  src_reg_from_dst(tmp));
+		  src_reg_from_dst(tmp),
+                  sampler );
 
       release_temp(c, tmp);
    }
@@ -885,7 +892,7 @@ static void precalc_txp( struct brw_wm_compile *c,
    {
       /* dst = TEX src0
        */
-      precalc_tex(c, dst, target, unit, src0);
+      precalc_tex(c, dst, target, unit, src0, sampler);
    }
 }
 
@@ -936,6 +943,7 @@ static void emit_fb_write( struct brw_wm_compile *c )
 		  dst_undef(),
 		  (i == c->key.nr_cbufs - 1), /* EOT */
 		  i,
+                  0,            /* no sampler */
 		  outcolor,
 		  payload_r0_depth,
 		  outdepth);
@@ -1056,15 +1064,17 @@ static void emit_insn( struct brw_wm_compile *c,
    case TGSI_OPCODE_TEX:
       precalc_tex(c, dst,
 		  inst->InstructionExtTexture.Texture,
-		  src[0].file,	/* sampler unit */
-		  src[1] );
+		  src[1].index,	/* use sampler unit for tex idx */
+		  src[0],       /* coord */
+                  src[1]);      /* sampler */
       break;
 
    case TGSI_OPCODE_TXP:
       precalc_txp(c, dst,
 		  inst->InstructionExtTexture.Texture,
-		  src[0].file,	/* sampler unit */
-		  src[1] );
+		  src[1].index,	/* use sampler unit for tex idx */
+		  src[0],       /* coord */
+                  src[1]);      /* sampler */
       break;
 
    case TGSI_OPCODE_TXB:
@@ -1072,8 +1082,9 @@ static void emit_insn( struct brw_wm_compile *c,
        */
       precalc_tex(c, dst,
 		  inst->InstructionExtTexture.Texture,
-		  src[0].file,	/* sampler unit */
-		  src[1] );
+		  src[1].index,	/* use sampler unit for tex idx*/
+		  src[0],
+                  src[1]);
       break;
 
    case TGSI_OPCODE_XPD: 
diff --git a/src/gallium/drivers/i965/brw_wm_surface_state.c b/src/gallium/drivers/i965/brw_wm_surface_state.c
index f882331433..f92b8198ed 100644
--- a/src/gallium/drivers/i965/brw_wm_surface_state.c
+++ b/src/gallium/drivers/i965/brw_wm_surface_state.c
@@ -149,19 +149,23 @@ brw_wm_get_binding_table(struct brw_context *brw,
    enum pipe_error ret;
    struct brw_winsys_reloc reloc[BRW_WM_MAX_SURF];
    uint32_t data[BRW_WM_MAX_SURF];
+   GLuint nr_relocs = 0;
    GLuint data_size = brw->wm.nr_surfaces * sizeof data[0];
    int i;
 
    assert(brw->wm.nr_surfaces <= BRW_WM_MAX_SURF);
    assert(brw->wm.nr_surfaces > 0);
 
-   /* Emit binding table relocations to surface state */
+   /* Emit binding table relocations to surface state 
+    */
    for (i = 0; i < brw->wm.nr_surfaces; i++) {
-      make_reloc(&reloc[i],
-                 BRW_USAGE_STATE,
-                 0,
-                 i * sizeof(GLuint),
-                 brw->wm.surf_bo[i]);
+      if (brw->wm.surf_bo[i]) {
+         make_reloc(&reloc[nr_relocs++],
+                    BRW_USAGE_STATE,
+                    0,
+                    i * sizeof(GLuint),
+                    brw->wm.surf_bo[i]);
+      }
    }
 
    /* Note there is no key for this search beyond the values in the
@@ -169,7 +173,7 @@ brw_wm_get_binding_table(struct brw_context *brw,
     */
    if (brw_search_cache(&brw->surface_cache, BRW_SS_SURF_BIND,
                         NULL, 0,
-                        reloc, brw->wm.nr_surfaces,
+                        reloc, nr_relocs,
                         NULL,
                         bo_out))
       return PIPE_OK;
@@ -182,7 +186,7 @@ brw_wm_get_binding_table(struct brw_context *brw,
 
    ret = brw_upload_cache( &brw->surface_cache, BRW_SS_SURF_BIND,
                            NULL, 0,
-                           reloc, brw->wm.nr_surfaces,
+                           reloc, nr_relocs,
                            data, data_size,
                            NULL, NULL,
                            bo_out);
@@ -208,40 +212,60 @@ static enum pipe_error prepare_wm_surfaces(struct brw_context *brw )
    for (i = 0; i < brw->curr.fb.nr_cbufs; i++) {
       ret = brw_update_render_surface(brw, 
                                       brw_surface(brw->curr.fb.cbufs[i]), 
-                                      &brw->wm.surf_bo[nr_surfaces++]);
+                                      &brw->wm.surf_bo[BTI_COLOR_BUF(i)]);
       if (ret)
          return ret;
+      
+      nr_surfaces = BTI_COLOR_BUF(i) + 1;
+   }
+
+
+
+   /* PIPE_NEW_FRAGMENT_CONSTANTS
+    */
+#if 0
+   if (brw->curr.fragment_constants) {
+      ret = brw_update_fragment_constant_surface(
+         brw, 
+         brw->curr.fragment_constants, 
+         &brw->wm.surf_bo[BTI_FRAGMENT_CONSTANTS]);
+
+      if (ret)
+         return ret;
+
+      nr_surfaces = BTI_FRAGMENT_CONSTANTS + 1;
    }
+   else {
+      bo_reference(&brw->wm.surf_bo[SURF_FRAG_CONSTANTS], NULL);      
+   }
+#endif
+
 
    /* PIPE_NEW_TEXTURE 
     */
    for (i = 0; i < brw->curr.num_textures; i++) {
       ret = brw_update_texture_surface(brw, 
                                        brw_texture(brw->curr.texture[i]),
-                                       &brw->wm.surf_bo[nr_surfaces++]);
+                                       &brw->wm.surf_bo[BTI_TEXTURE(i)]);
       if (ret)
          return ret;
+
+      nr_surfaces = BTI_TEXTURE(i) + 1;
    }
 
-   /* PIPE_NEW_FRAGMENT_CONSTANTS
+   /* Clear any inactive entries:
     */
-#if 0
-   if (brw->curr.fragment_constants) {
-      ret = brw_update_fragment_constant_surface(brw, 
-                                                 brw->curr.fragment_constants, 
-                                                 &brw->wm.surf_bo[nr_surfaces++]);
-      if (ret)
-         return ret;
-   }
-#endif
+   for (i = brw->curr.fb.nr_cbufs; i < BRW_MAX_DRAW_BUFFERS; i++) 
+      bo_reference(&brw->wm.surf_bo[BTI_COLOR_BUF(i)], NULL);
 
-   if (brw->wm.nr_surfaces != nr_surfaces) {
+   if (!brw->curr.fragment_constants)
+      bo_reference(&brw->wm.surf_bo[BTI_FRAGMENT_CONSTANTS], NULL);      
 
-      /* Unreference any left-over old buffers
-       */
-      for (i = nr_surfaces; i < brw->wm.nr_surfaces; i++)
-         bo_reference(&brw->wm.surf_bo[i], NULL);
+   /* XXX: no pipe_max_textures define?? */
+   for (i = brw->curr.num_textures; i < PIPE_MAX_SAMPLERS; i++)
+      bo_reference(&brw->wm.surf_bo[BTI_TEXTURE(i)], NULL);
 
+   if (brw->wm.nr_surfaces != nr_surfaces) {
       brw->wm.nr_surfaces = nr_surfaces;
       brw->state.dirty.brw |= BRW_NEW_NR_WM_SURFACES;
    }