i915_program.c

Mesa is an open-source implementation of the OpenGL specification - a system for rendering interacti

      return swizzle(UREG(REG_TYPE_R, 0), ZERO, ZERO, ZERO, ZERO);
   if (c0 == 1.0)
      return swizzle(UREG(REG_TYPE_R, 0), ONE, ONE, ONE, ONE);

   for (reg = 0; reg < I915_MAX_CONSTANT; reg++) {
      if (p->constant_flags[reg] == I915_CONSTFLAG_PARAM)
         continue;
      for (idx = 0; idx < 4; idx++) {
         if (!(p->constant_flags[reg] & (1 << idx)) ||
             p->constant[reg][idx] == c0) {
            p->constant[reg][idx] = c0;
            p->constant_flags[reg] |= 1 << idx;
            if (reg + 1 > p->nr_constants)
               p->nr_constants = reg + 1;
            return swizzle(UREG(REG_TYPE_CONST, reg), idx, ZERO, ZERO, ONE);
         }
      }
   }

   fprintf(stderr, "%s: out of constants\n", __FUNCTION__);
   p->error = 1;
   return 0;
}

GLuint
i915_emit_const2f(struct i915_fragment_program * p, GLfloat c0, GLfloat c1)
{
   GLint reg, idx;

   if (c0 == 0.0)
      return swizzle(i915_emit_const1f(p, c1), ZERO, X, Z, W);
   if (c0 == 1.0)
      return swizzle(i915_emit_const1f(p, c1), ONE, X, Z, W);

   if (c1 == 0.0)
      return swizzle(i915_emit_const1f(p, c0), X, ZERO, Z, W);
   if (c1 == 1.0)
      return swizzle(i915_emit_const1f(p, c0), X, ONE, Z, W);

   for (reg = 0; reg < I915_MAX_CONSTANT; reg++) {
      if (p->constant_flags[reg] == 0xf ||
          p->constant_flags[reg] == I915_CONSTFLAG_PARAM)
         continue;
      for (idx = 0; idx < 3; idx++) {
         if (!(p->constant_flags[reg] & (3 << idx))) {
            p->constant[reg][idx] = c0;
            p->constant[reg][idx + 1] = c1;
            p->constant_flags[reg] |= 3 << idx;
            if (reg + 1 > p->nr_constants)
               p->nr_constants = reg + 1;
            return swizzle(UREG(REG_TYPE_CONST, reg), idx, idx + 1, ZERO,
                           ONE);
         }
      }
   }

   fprintf(stderr, "%s: out of constants\n", __FUNCTION__);
   p->error = 1;
   return 0;
}



GLuint
i915_emit_const4f(struct i915_fragment_program * p,
                  GLfloat c0, GLfloat c1, GLfloat c2, GLfloat c3)
{
   GLint reg;

   for (reg = 0; reg < I915_MAX_CONSTANT; reg++) {
      if (p->constant_flags[reg] == 0xf &&
          p->constant[reg][0] == c0 &&
          p->constant[reg][1] == c1 &&
          p->constant[reg][2] == c2 && p->constant[reg][3] == c3) {
         return UREG(REG_TYPE_CONST, reg);
      }
      else if (p->constant_flags[reg] == 0) {
         p->constant[reg][0] = c0;
         p->constant[reg][1] = c1;
         p->constant[reg][2] = c2;
         p->constant[reg][3] = c3;
         p->constant_flags[reg] = 0xf;
         if (reg + 1 > p->nr_constants)
            p->nr_constants = reg + 1;
         return UREG(REG_TYPE_CONST, reg);
      }
   }

   fprintf(stderr, "%s: out of constants\n", __FUNCTION__);
   p->error = 1;
   return 0;
}


GLuint
i915_emit_const4fv(struct i915_fragment_program * p, const GLfloat * c)
{
   return i915_emit_const4f(p, c[0], c[1], c[2], c[3]);
}


GLuint
i915_emit_param4fv(struct i915_fragment_program * p, const GLfloat * values)
{
   GLint reg, i;

   for (i = 0; i < p->nr_params; i++) {
      if (p->param[i].values == values)
         return UREG(REG_TYPE_CONST, p->param[i].reg);
   }


   for (reg = 0; reg < I915_MAX_CONSTANT; reg++) {
      if (p->constant_flags[reg] == 0) {
         p->constant_flags[reg] = I915_CONSTFLAG_PARAM;
         i = p->nr_params++;

         p->param[i].values = values;
         p->param[i].reg = reg;
         p->params_uptodate = 0;

         if (reg + 1 > p->nr_constants)
            p->nr_constants = reg + 1;
         return UREG(REG_TYPE_CONST, reg);
      }
   }

   fprintf(stderr, "%s: out of constants\n", __FUNCTION__);
   p->error = 1;
   return 0;
}



void
i915_program_error(struct i915_fragment_program *p, const char *msg)
{
   _mesa_problem(NULL, "i915_program_error: %s", msg);
   p->error = 1;
}


void
i915_init_program(struct i915_context *i915, struct i915_fragment_program *p)
{
   GLcontext *ctx = &i915->intel.ctx;

   p->translated = 0;
   p->params_uptodate = 0;
   p->on_hardware = 0;
   p->error = 0;

   p->nr_tex_indirect = 1;      /* correct? */
   p->nr_tex_insn = 0;
   p->nr_alu_insn = 0;
   p->nr_decl_insn = 0;

   p->ctx = ctx;
   memset(p->constant_flags, 0, sizeof(p->constant_flags));

   p->nr_constants = 0;
   p->csr = p->program;
   p->decl = p->declarations;
   p->decl_s = 0;
   p->decl_t = 0;
   p->temp_flag = 0xffff000;
   p->utemp_flag = ~0x7;
   p->wpos_tex = -1;
   p->depth_written = 0;
   p->nr_params = 0;

   *(p->decl++) = _3DSTATE_PIXEL_SHADER_PROGRAM;
}


void
i915_fini_program(struct i915_fragment_program *p)
{
   GLuint program_size = p->csr - p->program;
   GLuint decl_size = p->decl - p->declarations;

   if (p->nr_tex_indirect > I915_MAX_TEX_INDIRECT)
      i915_program_error(p, "Exceeded max nr indirect texture lookups");

   if (p->nr_tex_insn > I915_MAX_TEX_INSN)
      i915_program_error(p, "Exceeded max TEX instructions");

   if (p->nr_alu_insn > I915_MAX_ALU_INSN)
      i915_program_error(p, "Exceeded max ALU instructions");

   if (p->nr_decl_insn > I915_MAX_DECL_INSN)
      i915_program_error(p, "Exceeded max DECL instructions");

   if (p->error) {
      p->FragProg.Base.NumNativeInstructions = 0;
      p->FragProg.Base.NumNativeAluInstructions = 0;
      p->FragProg.Base.NumNativeTexInstructions = 0;
      p->FragProg.Base.NumNativeTexIndirections = 0;
   }
   else {
      p->FragProg.Base.NumNativeInstructions = (p->nr_alu_insn +
                                                p->nr_tex_insn +
                                                p->nr_decl_insn);
      p->FragProg.Base.NumNativeAluInstructions = p->nr_alu_insn;
      p->FragProg.Base.NumNativeTexInstructions = p->nr_tex_insn;
      p->FragProg.Base.NumNativeTexIndirections = p->nr_tex_indirect;
   }

   p->declarations[0] |= program_size + decl_size - 2;
}

void
i915_upload_program(struct i915_context *i915,
                    struct i915_fragment_program *p)
{
   GLuint program_size = p->csr - p->program;
   GLuint decl_size = p->decl - p->declarations;

   FALLBACK(&i915->intel, I915_FALLBACK_PROGRAM, p->error);

   /* Could just go straight to the batchbuffer from here:
    */
   if (i915->state.ProgramSize != (program_size + decl_size) ||
       memcmp(i915->state.Program + decl_size, p->program,
              program_size * sizeof(int)) != 0) {
      I915_STATECHANGE(i915, I915_UPLOAD_PROGRAM);
      memcpy(i915->state.Program, p->declarations, decl_size * sizeof(int));
      memcpy(i915->state.Program + decl_size, p->program,
             program_size * sizeof(int));
      i915->state.ProgramSize = decl_size + program_size;
   }

   /* Always seemed to get a failure if I used memcmp() to
    * shortcircuit this state upload.  Needs further investigation?
    */
   if (p->nr_constants) {
      GLuint nr = p->nr_constants;

      I915_ACTIVESTATE(i915, I915_UPLOAD_CONSTANTS, 1);
      I915_STATECHANGE(i915, I915_UPLOAD_CONSTANTS);

      i915->state.Constant[0] = _3DSTATE_PIXEL_SHADER_CONSTANTS | ((nr) * 4);
      i915->state.Constant[1] = (1 << (nr - 1)) | ((1 << (nr - 1)) - 1);

      memcpy(&i915->state.Constant[2], p->constant, 4 * sizeof(int) * (nr));
      i915->state.ConstantSize = 2 + (nr) * 4;

      if (0) {
         GLuint i;
         for (i = 0; i < nr; i++) {
            fprintf(stderr, "const[%d]: %f %f %f %f\n", i,
                    p->constant[i][0],
                    p->constant[i][1], p->constant[i][2], p->constant[i][3]);
         }
      }
   }
   else {
      I915_ACTIVESTATE(i915, I915_UPLOAD_CONSTANTS, 0);
   }

   p->on_hardware = 1;
}