Merge branch '7.8'
[mesa.git] / src / mesa / state_tracker / st_mesa_to_tgsi.c
index 5c9be46a776fffe6fe02f320e73577b8742ba091..e8eb9ec3904157d224c4ecda8d9e78d5f39aa437 100644 (file)
 #include "pipe/p_compiler.h"
 #include "pipe/p_shader_tokens.h"
 #include "pipe/p_state.h"
+#include "pipe/p_context.h"
 #include "tgsi/tgsi_ureg.h"
 #include "st_mesa_to_tgsi.h"
+#include "st_context.h"
 #include "shader/prog_instruction.h"
 #include "shader/prog_parameter.h"
-#include "shader/prog_print.h"
 #include "util/u_debug.h"
 #include "util/u_math.h"
 #include "util/u_memory.h"
@@ -48,6 +49,10 @@ struct label {
    unsigned token;
 };
 
+
+/**
+ * Intermediate state used during shader translation.
+ */
 struct st_translate {
    struct ureg_program *ureg;
 
@@ -57,6 +62,10 @@ struct st_translate {
    struct ureg_src inputs[PIPE_MAX_SHADER_INPUTS];
    struct ureg_dst address[1];
    struct ureg_src samplers[PIPE_MAX_SAMPLERS];
+   struct ureg_dst psizregreal;
+   struct ureg_src pointSizeConst;
+   GLint psizoutindex;
+   GLboolean prevInstWrotePsiz;
 
    const GLuint *inputMapping;
    const GLuint *outputMapping;
@@ -145,6 +154,8 @@ dst_register( struct st_translate *t,
       return t->temps[index];
 
    case PROGRAM_OUTPUT:
+      if (index == t->psizoutindex)
+         t->prevInstWrotePsiz = GL_TRUE;
       return t->outputs[t->outputMapping[index]];
 
    case PROGRAM_ADDRESS:
@@ -160,28 +171,38 @@ dst_register( struct st_translate *t,
 static struct ureg_src
 src_register( struct st_translate *t,
               gl_register_file file,
-              GLuint index )
+              GLint index )
 {
    switch( file ) {
    case PROGRAM_UNDEFINED:
       return ureg_src_undef();
 
    case PROGRAM_TEMPORARY:
+      ASSERT(index >= 0);
       if (ureg_dst_is_undef(t->temps[index]))
          t->temps[index] = ureg_DECL_temporary( t->ureg );
+      assert(index < Elements(t->temps));
       return ureg_src(t->temps[index]);
 
-   case PROGRAM_STATE_VAR:
    case PROGRAM_NAMED_PARAM:
    case PROGRAM_ENV_PARAM:
+   case PROGRAM_LOCAL_PARAM:
    case PROGRAM_UNIFORM:
-   case PROGRAM_CONSTANT:       /* ie, immediate */
+      ASSERT(index >= 0);
       return t->constants[index];
+   case PROGRAM_STATE_VAR:
+   case PROGRAM_CONSTANT:       /* ie, immediate */
+      if (index < 0)
+         return ureg_DECL_constant( t->ureg, 0 );
+      else
+         return t->constants[index];
 
    case PROGRAM_INPUT:
+      assert(t->inputMapping[index] < Elements(t->inputs));
       return t->inputs[t->inputMapping[index]];
 
    case PROGRAM_OUTPUT:
+      assert(t->outputMapping[index] < Elements(t->outputs));
       return ureg_src(t->outputs[t->outputMapping[index]]); /* not needed? */
 
    case PROGRAM_ADDRESS:
@@ -263,9 +284,14 @@ translate_src( struct st_translate *t,
    if (SrcReg->Abs) 
       src = ureg_abs(src);
 
-   if (SrcReg->RelAddr) 
+   if (SrcReg->RelAddr) {
       src = ureg_src_indirect( src, ureg_src(t->address[0]));
-   
+      /* If SrcReg->Index was negative, it was set to zero in
+       * src_register().  Reassign it now.
+       */
+      src.Index = SrcReg->Index;
+   }
+
    return src;
 }
 
@@ -649,6 +675,22 @@ compile_instruction(
    }
 }
 
+/**
+ * Emit the TGSI instructions to adjust the WPOS pixel center convention
+ */
+static void
+emit_adjusted_wpos( struct st_translate *t,
+                    const struct gl_program *program, GLfloat value)
+{
+   struct ureg_program *ureg = t->ureg;
+   struct ureg_dst wpos_temp = ureg_DECL_temporary(ureg);
+   struct ureg_src wpos_input = t->inputs[t->inputMapping[FRAG_ATTRIB_WPOS]];
+
+   ureg_ADD(ureg, ureg_writemask(wpos_temp, TGSI_WRITEMASK_X | TGSI_WRITEMASK_Y),
+                  wpos_input, ureg_imm1f(ureg, value));
+
+   t->inputs[t->inputMapping[FRAG_ATTRIB_WPOS]] = ureg_src(wpos_temp);
+}
 
 /**
  * Emit the TGSI instructions for inverting the WPOS y coordinate.
@@ -674,12 +716,17 @@ emit_inverted_wpos( struct st_translate *t,
                                                        winSizeState);
 
    struct ureg_src winsize = ureg_DECL_constant( ureg, winHeightConst );
-   struct ureg_dst wpos_temp = ureg_DECL_temporary( ureg );
+   struct ureg_dst wpos_temp;
    struct ureg_src wpos_input = t->inputs[t->inputMapping[FRAG_ATTRIB_WPOS]];
 
    /* MOV wpos_temp, input[wpos]
     */
-   ureg_MOV( ureg, wpos_temp, wpos_input );
+   if (wpos_input.File == TGSI_FILE_TEMPORARY)
+      wpos_temp = ureg_dst(wpos_input);
+   else {
+      wpos_temp = ureg_DECL_temporary( ureg );
+      ureg_MOV( ureg, wpos_temp, wpos_input );
+   }
 
    /* SUB wpos_temp.y, winsize_const, wpos_input
     */
@@ -694,6 +741,65 @@ emit_inverted_wpos( struct st_translate *t,
 }
 
 
+/**
+ * Emit fragment position/ooordinate code.
+ */
+static void
+emit_wpos(struct st_context *st,
+          struct st_translate *t,
+          const struct gl_program *program,
+          struct ureg_program *ureg)
+{
+   const struct gl_fragment_program *fp =
+      (const struct gl_fragment_program *) program;
+   struct pipe_screen *pscreen = st->pipe->screen;
+   boolean invert = FALSE;
+
+   if (fp->OriginUpperLeft) {
+      if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT)) {
+      }
+      else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT)) {
+         ureg_property_fs_coord_origin(ureg, TGSI_FS_COORD_ORIGIN_LOWER_LEFT);
+         invert = TRUE;
+      }
+      else
+         assert(0);
+   }
+   else {
+      if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_LOWER_LEFT))
+         ureg_property_fs_coord_origin(ureg, TGSI_FS_COORD_ORIGIN_LOWER_LEFT);
+      else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_ORIGIN_UPPER_LEFT))
+         invert = TRUE;
+      else
+         assert(0);
+   }
+   
+   if (fp->PixelCenterInteger) {
+      if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER))
+         ureg_property_fs_coord_pixel_center(ureg, TGSI_FS_COORD_PIXEL_CENTER_INTEGER);
+      else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER))
+         emit_adjusted_wpos(t, program, invert ? 0.5f : -0.5f);
+      else
+         assert(0);
+   }
+   else {
+      if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_HALF_INTEGER)) {
+      }
+      else if (pscreen->get_param(pscreen, PIPE_CAP_TGSI_FS_COORD_PIXEL_CENTER_INTEGER)) {
+         ureg_property_fs_coord_pixel_center(ureg, TGSI_FS_COORD_PIXEL_CENTER_INTEGER);
+         emit_adjusted_wpos(t, program, invert ? -0.5f : 0.5f);
+      }
+      else
+         assert(0);
+   }
+
+   /* we invert after adjustment so that we avoid the MOV to temporary,
+    * and reuse the adjustment ADD instead */
+   if (invert)
+      emit_inverted_wpos(t, program);
+}
+
+
 /**
  * OpenGL's fragment gl_FrontFace input is 1 for front-facing, 0 for back.
  * TGSI uses +1 for front, -1 for back.
@@ -718,6 +824,7 @@ emit_face_var( struct st_translate *t,
    t->inputs[t->inputMapping[FRAG_ATTRIB_FACE]] = ureg_src(face_temp);
 }
 
+
 static void
 emit_edgeflags( struct st_translate *t,
                  const struct gl_program *program )
@@ -729,6 +836,7 @@ emit_edgeflags( struct st_translate *t,
    ureg_MOV( ureg, edge_dst, edge_src );
 }
 
+
 /**
  * Translate Mesa program to TGSI format.
  * \param program  the program to translate
@@ -746,7 +854,7 @@ emit_edgeflags( struct st_translate *t,
  * \param outputSemanticIndex  the semantic index (ex: which texcoord) for
  *                             each output
  *
- * \return  array of translated tokens, caller's responsibility to free
+ * \return  PIPE_OK or PIPE_ERROR_OUT_OF_MEMORY
  */
 enum pipe_error
 st_translate_mesa_program(
@@ -767,6 +875,7 @@ st_translate_mesa_program(
 {
    struct st_translate translate, *t;
    unsigned i;
+   enum pipe_error ret = PIPE_OK;
 
    t = &translate;
    memset(t, 0, sizeof *t);
@@ -775,6 +884,8 @@ st_translate_mesa_program(
    t->inputMapping = inputMapping;
    t->outputMapping = outputMapping;
    t->ureg = ureg;
+   t->psizoutindex = -1;
+   t->prevInstWrotePsiz = GL_FALSE;
 
    /*_mesa_print_program(program);*/
 
@@ -783,17 +894,26 @@ st_translate_mesa_program(
     */
    if (procType == TGSI_PROCESSOR_FRAGMENT) {
       for (i = 0; i < numInputs; i++) {
-         t->inputs[i] = ureg_DECL_fs_input(ureg,
-                                           inputSemanticName[i],
-                                           inputSemanticIndex[i],
-                                           interpMode[i]);
+         if (program->InputFlags[0] & PROG_PARAM_BIT_CYL_WRAP) {
+            t->inputs[i] = ureg_DECL_fs_input_cyl(ureg,
+                                                  inputSemanticName[i],
+                                                  inputSemanticIndex[i],
+                                                  interpMode[i],
+                                                  TGSI_CYLINDRICAL_WRAP_X);
+         }
+         else {
+            t->inputs[i] = ureg_DECL_fs_input(ureg,
+                                              inputSemanticName[i],
+                                              inputSemanticIndex[i],
+                                              interpMode[i]);
+         }
       }
 
       if (program->InputsRead & FRAG_BIT_WPOS) {
          /* Must do this after setting up t->inputs, and before
           * emitting constant references, below:
           */
-         emit_inverted_wpos( t, program );
+         emit_wpos(st_context(ctx), t, program, ureg);
       }
 
       if (program->InputsRead & FRAG_BIT_FACE) {
@@ -833,6 +953,22 @@ st_translate_mesa_program(
          t->outputs[i] = ureg_DECL_output( ureg,
                                            outputSemanticName[i],
                                            outputSemanticIndex[i] );
+         if ((outputSemanticName[i] == TGSI_SEMANTIC_PSIZE) && program->Id) {
+            static const gl_state_index pointSizeClampState[STATE_LENGTH]
+               = { STATE_INTERNAL, STATE_POINT_SIZE_IMPL_CLAMP, 0, 0, 0 };
+               /* XXX: note we are modifying the incoming shader here!  Need to
+               * do this before emitting the constant decls below, or this
+               * will be missed:
+               */
+            unsigned pointSizeClampConst =
+               _mesa_add_state_reference(program->Parameters,
+                                         pointSizeClampState);
+            struct ureg_dst psizregtemp = ureg_DECL_temporary( ureg );
+            t->pointSizeConst = ureg_DECL_constant( ureg, pointSizeClampConst );
+            t->psizregreal = t->outputs[i];
+            t->psizoutindex = i;
+            t->outputs[i] = psizregtemp;
+         }
       }
       if (passthrough_edgeflags)
          emit_edgeflags( t, program );
@@ -845,20 +981,21 @@ st_translate_mesa_program(
       t->address[0] = ureg_DECL_address( ureg );
    }
 
-
    /* Emit constants and immediates.  Mesa uses a single index space
     * for these, so we put all the translated regs in t->constants.
     */
    if (program->Parameters) {
-      
       t->constants = CALLOC( program->Parameters->NumParameters,
                              sizeof t->constants[0] );
-      if (t->constants == NULL)
+      if (t->constants == NULL) {
+         ret = PIPE_ERROR_OUT_OF_MEMORY;
          goto out;
+      }
       
       for (i = 0; i < program->Parameters->NumParameters; i++) {
          switch (program->Parameters->Parameters[i].Type) {
          case PROGRAM_ENV_PARAM:
+         case PROGRAM_LOCAL_PARAM:
          case PROGRAM_STATE_VAR:
          case PROGRAM_NAMED_PARAM:
          case PROGRAM_UNIFORM:
@@ -897,6 +1034,20 @@ st_translate_mesa_program(
    for (i = 0; i < program->NumInstructions; i++) {
       set_insn_start( t, ureg_get_instruction_number( ureg ));
       compile_instruction( t, &program->Instructions[i] );
+
+      /* note can't do that easily at the end of prog due to
+         possible early return */
+      if (t->prevInstWrotePsiz && program->Id) {
+         set_insn_start( t, ureg_get_instruction_number( ureg ));
+         ureg_MAX( t->ureg,
+                   ureg_writemask(t->outputs[t->psizoutindex], WRITEMASK_X),
+                   ureg_src(t->outputs[t->psizoutindex]),
+                   ureg_swizzle(t->pointSizeConst, 1,1,1,1));
+         ureg_MIN( t->ureg, ureg_writemask(t->psizregreal, WRITEMASK_X),
+                   ureg_src(t->outputs[t->psizoutindex]),
+                   ureg_swizzle(t->pointSizeConst, 2,2,2,2));
+      }
+      t->prevInstWrotePsiz = GL_FALSE;
    }
 
    /* Fix up all emitted labels:
@@ -907,8 +1058,6 @@ st_translate_mesa_program(
                         t->insn[t->labels[i].branch_target] );
    }
 
-   return PIPE_OK;
-
 out:
    FREE(t->insn);
    FREE(t->labels);
@@ -918,12 +1067,12 @@ out:
       debug_printf("%s: translate error flag set\n", __FUNCTION__);
    }
 
-   return PIPE_ERROR_OUT_OF_MEMORY;
+   return ret;
 }
 
 
 /**
- * Tokens cannot be free with _mesa_free otherwise the builtin gallium
+ * Tokens cannot be free with free otherwise the builtin gallium
  * malloc debugging will get confused.
  */
 void