brw_vs_emit.c

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

	      tmp_d, brw_imm_d(23));

      release_tmp(c, tmp);
   }

   if (dst.dw1.bits.writemask & WRITEMASK_Y) {
      /* result[1] = arg0.x - floor(arg0.x) */
      brw_FRC(p, brw_writemask(dst, WRITEMASK_Y), brw_swizzle1(arg0, 0));
   }
   
   if (dst.dw1.bits.writemask & WRITEMASK_Z) {
      /* As with the LOG instruction, we might be better off just
       * doing a taylor expansion here, seeing as we have to do all
       * the prep work.
       *
       * If mathbox partial precision is too low, consider also:
       * result[3] = result[0] * EXP(result[1])
       */
      emit_math1(c, 
		 BRW_MATH_FUNCTION_EXP, 
		 brw_writemask(dst, WRITEMASK_Z),
		 brw_swizzle1(arg0, 0), 
		 BRW_MATH_PRECISION_PARTIAL);
   }  

   if (dst.dw1.bits.writemask & WRITEMASK_W) {
      /* result[3] = 1.0; */
      brw_MOV(p, brw_writemask(dst, WRITEMASK_W), brw_imm_f(1));
   }
}


static void emit_log_noalias( struct brw_vs_compile *c,
			      struct brw_reg dst,
			      struct brw_reg arg0 )
{
   struct brw_compile *p = &c->func;
   struct brw_reg tmp = dst;
   struct brw_reg tmp_ud = retype(tmp, BRW_REGISTER_TYPE_UD);
   struct brw_reg arg0_ud = retype(arg0, BRW_REGISTER_TYPE_UD);
   GLboolean need_tmp = (dst.dw1.bits.writemask != 0xf ||
			 dst.file != BRW_GENERAL_REGISTER_FILE);

   if (need_tmp) {
      tmp = get_tmp(c);
      tmp_ud = retype(tmp, BRW_REGISTER_TYPE_UD);
   }
   
   /* Perform mant = frexpf(fabsf(x), &exp), adjust exp and mnt
    * according to spec:
    *
    * These almost look likey they could be joined up, but not really
    * practical:
    *
    * result[0].f = (x.i & ((1<<31)-1) >> 23) - 127
    * result[1].i = (x.i & ((1<<23)-1)        + (127<<23)
    */
   if (dst.dw1.bits.writemask & WRITEMASK_XZ) {
      brw_AND(p, 
	      brw_writemask(tmp_ud, WRITEMASK_X),
	      brw_swizzle1(arg0_ud, 0),
	      brw_imm_ud((1U<<31)-1));

      brw_SHR(p, 
	      brw_writemask(tmp_ud, WRITEMASK_X), 
	      tmp_ud,
	      brw_imm_ud(23));

      brw_ADD(p, 
	      brw_writemask(tmp, WRITEMASK_X), 
	      retype(tmp_ud, BRW_REGISTER_TYPE_D),	/* does it matter? */
	      brw_imm_d(-127));
   }

   if (dst.dw1.bits.writemask & WRITEMASK_YZ) {
      brw_AND(p, 
	      brw_writemask(tmp_ud, WRITEMASK_Y),
	      brw_swizzle1(arg0_ud, 0),
	      brw_imm_ud((1<<23)-1));

      brw_OR(p, 
	     brw_writemask(tmp_ud, WRITEMASK_Y), 
	     tmp_ud,
	     brw_imm_ud(127<<23));
   }
   
   if (dst.dw1.bits.writemask & WRITEMASK_Z) {
      /* result[2] = result[0] + LOG2(result[1]); */

      /* Why bother?  The above is just a hint how to do this with a
       * taylor series.  Maybe we *should* use a taylor series as by
       * the time all the above has been done it's almost certainly
       * quicker than calling the mathbox, even with low precision.
       * 
       * Options are:
       *    - result[0] + mathbox.LOG2(result[1])
       *    - mathbox.LOG2(arg0.x)
       *    - result[0] + inline_taylor_approx(result[1])
       */
      emit_math1(c, 
		 BRW_MATH_FUNCTION_LOG, 
		 brw_writemask(tmp, WRITEMASK_Z), 
		 brw_swizzle1(tmp, 1), 
		 BRW_MATH_PRECISION_FULL);
      
      brw_ADD(p, 
	      brw_writemask(tmp, WRITEMASK_Z), 
	      brw_swizzle1(tmp, 2), 
	      brw_swizzle1(tmp, 0));
   }  

   if (dst.dw1.bits.writemask & WRITEMASK_W) {
      /* result[3] = 1.0; */
      brw_MOV(p, brw_writemask(tmp, WRITEMASK_W), brw_imm_f(1));
   }

   if (need_tmp) {
      brw_MOV(p, dst, tmp);
      release_tmp(c, tmp);
   }
}


      

/* Need to unalias - consider swizzles:   r0 = DST r0.xxxx r1
 */
static void emit_dst_noalias( struct brw_vs_compile *c, 
			      struct brw_reg dst,
			      struct brw_reg arg0,
			      struct brw_reg arg1)
{
   struct brw_compile *p = &c->func;

   /* There must be a better way to do this: 
    */
   if (dst.dw1.bits.writemask & WRITEMASK_X)
      brw_MOV(p, brw_writemask(dst, WRITEMASK_X), brw_imm_f(1.0));
   if (dst.dw1.bits.writemask & WRITEMASK_Y)
      brw_MUL(p, brw_writemask(dst, WRITEMASK_Y), arg0, arg1);
   if (dst.dw1.bits.writemask & WRITEMASK_Z)
      brw_MOV(p, brw_writemask(dst, WRITEMASK_Z), arg0);
   if (dst.dw1.bits.writemask & WRITEMASK_W)
      brw_MOV(p, brw_writemask(dst, WRITEMASK_W), arg1);
}

static void emit_xpd( struct brw_compile *p,
		      struct brw_reg dst,
		      struct brw_reg t,
		      struct brw_reg u)
{
   brw_MUL(p, brw_null_reg(), brw_swizzle(t, 1,2,0,3),  brw_swizzle(u,2,0,1,3));
   brw_MAC(p, dst,     negate(brw_swizzle(t, 2,0,1,3)), brw_swizzle(u,1,2,0,3));
}



static void emit_lit_noalias( struct brw_vs_compile *c, 
			      struct brw_reg dst,
			      struct brw_reg arg0 )
{
   struct brw_compile *p = &c->func;
   struct brw_instruction *if_insn;
   struct brw_reg tmp = dst;
   GLboolean need_tmp = (dst.file != BRW_GENERAL_REGISTER_FILE);

   if (need_tmp) 
      tmp = get_tmp(c);
   
   brw_MOV(p, brw_writemask(dst, WRITEMASK_YZ), brw_imm_f(0)); 
   brw_MOV(p, brw_writemask(dst, WRITEMASK_XW), brw_imm_f(1)); 

   /* Need to use BRW_EXECUTE_8 and also do an 8-wide compare in order
    * to get all channels active inside the IF.  In the clipping code
    * we run with NoMask, so it's not an option and we can use
    * BRW_EXECUTE_1 for all comparisions.
    */
   brw_CMP(p, brw_null_reg(), BRW_CONDITIONAL_G, brw_swizzle1(arg0,0), brw_imm_f(0));
   if_insn = brw_IF(p, BRW_EXECUTE_8);
   {
      brw_MOV(p, brw_writemask(dst, WRITEMASK_Y), brw_swizzle1(arg0,0));

      brw_CMP(p, brw_null_reg(), BRW_CONDITIONAL_G, brw_swizzle1(arg0,1), brw_imm_f(0));
      brw_MOV(p, brw_writemask(tmp, WRITEMASK_Z),  brw_swizzle1(arg0,1));
      brw_set_predicate_control(p, BRW_PREDICATE_NONE);

      emit_math2(c, 
		 BRW_MATH_FUNCTION_POW, 
		 brw_writemask(dst, WRITEMASK_Z),
		 brw_swizzle1(tmp, 2),
		 brw_swizzle1(arg0, 3),
		 BRW_MATH_PRECISION_PARTIAL);      
   }

   brw_ENDIF(p, if_insn);
}





/* TODO: relative addressing!
 */
static struct brw_reg get_reg( struct brw_vs_compile *c,
			       GLuint file,
			       GLuint index )
{

   switch (file) {
   case PROGRAM_TEMPORARY:
   case PROGRAM_INPUT:
   case PROGRAM_OUTPUT:
      assert(c->regs[file][index].nr != 0);
      return c->regs[file][index];
   case PROGRAM_STATE_VAR:
   case PROGRAM_CONSTANT:
   case PROGRAM_UNIFORM:
      assert(c->regs[PROGRAM_STATE_VAR][index].nr != 0);
      return c->regs[PROGRAM_STATE_VAR][index];
   case PROGRAM_ADDRESS:
      assert(index == 0);
      return c->regs[file][index];

   case PROGRAM_UNDEFINED:			/* undef values */
      return brw_null_reg();

   case PROGRAM_LOCAL_PARAM: 
   case PROGRAM_ENV_PARAM: 
   case PROGRAM_WRITE_ONLY:
   default:
      assert(0);
      return brw_null_reg();
   }
}



static struct brw_reg deref( struct brw_vs_compile *c,
			     struct brw_reg arg,
			     GLint offset)
{
   struct brw_compile *p = &c->func;
   struct brw_reg tmp = vec4(get_tmp(c));
   struct brw_reg vp_address = retype(vec1(get_reg(c, PROGRAM_ADDRESS, 0)), BRW_REGISTER_TYPE_UW);
   GLuint byte_offset = arg.nr * 32 + arg.subnr + offset * 16;
   struct brw_reg indirect = brw_vec4_indirect(0,0);

   {
      brw_push_insn_state(p);
      brw_set_access_mode(p, BRW_ALIGN_1);

      /* This is pretty clunky - load the address register twice and
       * fetch each 4-dword value in turn.  There must be a way to do
       * this in a single pass, but I couldn't get it to work.
       */
      brw_ADD(p, brw_address_reg(0), vp_address, brw_imm_d(byte_offset));
      brw_MOV(p, tmp, indirect);

      brw_ADD(p, brw_address_reg(0), suboffset(vp_address, 8), brw_imm_d(byte_offset));
      brw_MOV(p, suboffset(tmp, 4), indirect);

      brw_pop_insn_state(p);
   }
   
   return vec8(tmp);
}


static void emit_arl( struct brw_vs_compile *c,
		      struct brw_reg dst,
		      struct brw_reg arg0 )
{
   struct brw_compile *p = &c->func;
   struct brw_reg tmp = dst;
   GLboolean need_tmp = (dst.file != BRW_GENERAL_REGISTER_FILE);
   
   if (need_tmp) 
      tmp = get_tmp(c);

   brw_RNDD(p, tmp, arg0);
   brw_MUL(p, dst, tmp, brw_imm_d(16));

   if (need_tmp)
      release_tmp(c, tmp);
}


/* Will return mangled results for SWZ op.  The emit_swz() function
 * ignores this result and recalculates taking extended swizzles into
 * account.
 */
static struct brw_reg get_arg( struct brw_vs_compile *c,
			       struct prog_src_register *src )
{
   struct brw_reg reg;

   if (src->File == PROGRAM_UNDEFINED)
      return brw_null_reg();

   if (src->RelAddr) 
      reg = deref(c, c->regs[PROGRAM_STATE_VAR][0], src->Index);
   else
      reg = get_reg(c, src->File, src->Index);

   /* Convert 3-bit swizzle to 2-bit.  
    */
   reg.dw1.bits.swizzle = BRW_SWIZZLE4(GET_SWZ(src->Swizzle, 0),
				       GET_SWZ(src->Swizzle, 1),
				       GET_SWZ(src->Swizzle, 2),
				       GET_SWZ(src->Swizzle, 3));

   /* Note this is ok for non-swizzle instructions: 
    */
   reg.negate = src->NegateBase ? 1 : 0;   

   return reg;
}


static struct brw_reg get_dst( struct brw_vs_compile *c,
			       struct prog_dst_register dst )
{
   struct brw_reg reg = get_reg(c, dst.File, dst.Index);

   reg.dw1.bits.writemask = dst.WriteMask;

   return reg;
}




static void emit_swz( struct brw_vs_compile *c, 
		      struct brw_reg dst,
		      struct prog_src_register src )
{
   struct brw_compile *p = &c->func;
   GLuint zeros_mask = 0;
   GLuint ones_mask = 0;
   GLuint src_mask = 0;
   GLubyte src_swz[4];
   GLboolean need_tmp = (src.NegateBase &&
			 dst.file != BRW_GENERAL_REGISTER_FILE);
   struct brw_reg tmp = dst;
   GLuint i;

   if (need_tmp)
      tmp = get_tmp(c);

   for (i = 0; i < 4; i++) {
      if (dst.dw1.bits.writemask & (1<<i)) {
	 GLubyte s = GET_SWZ(src.Swizzle, i);
	 switch (s) {
	 case SWIZZLE_X:
	 case SWIZZLE_Y:
	 case SWIZZLE_Z:
	 case SWIZZLE_W:
	    src_mask |= 1<<i;
	    src_swz[i] = s;
	    break;
	 case SWIZZLE_ZERO:
	    zeros_mask |= 1<<i;
	    break;
	 case SWIZZLE_ONE:
	    ones_mask |= 1<<i;
	    break;
	 }
      }
   }
   
   /* Do src first, in case dst aliases src:
    */
   if (src_mask) {
      struct brw_reg arg0;

      if (src.RelAddr) 
	 arg0 = deref(c, c->regs[PROGRAM_STATE_VAR][0], src.Index);
      else
	 arg0 = get_reg(c, src.File, src.Index);

      arg0 = brw_swizzle(arg0, 
			 src_swz[0], src_swz[1], 
			 src_swz[2], src_swz[3]);

      brw_MOV(p, brw_writemask(tmp, src_mask), arg0);
   } 
   
   if (zeros_mask) 
      brw_MOV(p, brw_writemask(tmp, zeros_mask), brw_imm_f(0));

   if (ones_mask) 
      brw_MOV(p, brw_writemask(tmp, ones_mask), brw_imm_f(1));

   if (src.NegateBase)
      brw_MOV(p, brw_writemask(tmp, src.NegateBase), negate(tmp));
   
   if (need_tmp) {
      brw_MOV(p, dst, tmp);
      release_tmp(c, tmp);
   }