mesa/src/compiler/glsl/opt_algebraic.cpp

/*
 * Copyright © 2010 Intel Corporation
 *
 * Permission is hereby granted, free of charge, to any person obtaining a
 * copy of this software and associated documentation files (the "Software"),
 * to deal in the Software without restriction, including without limitation
 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
 * and/or sell copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice (including the next
 * paragraph) shall be included in all copies or substantial portions of the
 * Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
 * DEALINGS IN THE SOFTWARE.
 */

/**
 * \file opt_algebraic.cpp
 *
 * Takes advantage of association, commutivity, and other algebraic
 * properties to simplify expressions.
 */

#include "ir.h"
#include "ir_visitor.h"
#include "ir_rvalue_visitor.h"
#include "ir_optimization.h"
#include "ir_builder.h"
#include "compiler/glsl_types.h"
#include "main/consts_exts.h"

using namespace ir_builder;

namespace {

/**
 * Visitor class for replacing expressions with ir_constant values.
 */

class ir_algebraic_visitor : public ir_rvalue_visitor {
public:
   ir_algebraic_visitor(bool native_integers,
                        const struct gl_shader_compiler_options *options)
      : options(options)
   {
      this->progress = false;
      this->mem_ctx = NULL;
      this->native_integers = native_integers;
   }

   virtual ~ir_algebraic_visitor()
   {
   }

   virtual ir_visitor_status visit_enter(ir_assignment *ir);

   ir_rvalue *handle_expression(ir_expression *ir);
   void handle_rvalue(ir_rvalue **rvalue);
   bool reassociate_constant(ir_expression *ir1,
			     int const_index,
			     ir_constant *constant,
			     ir_expression *ir2);
   void reassociate_operands(ir_expression *ir1,
			     int op1,
			     ir_expression *ir2,
			     int op2);
   ir_rvalue *swizzle_if_required(ir_expression *expr,
				  ir_rvalue *operand);

   const struct gl_shader_compiler_options *options;
   void *mem_ctx;

   bool native_integers;
   bool progress;
};

} /* unnamed namespace */

ir_visitor_status
ir_algebraic_visitor::visit_enter(ir_assignment *ir)
{
   ir_variable *var = ir->lhs->variable_referenced();
   if (var->data.invariant || var->data.precise) {
      /* If we're assigning to an invariant or precise variable, just bail.
       * Most of the algebraic optimizations aren't precision-safe.
       *
       * FINISHME: Find out which optimizations are precision-safe and enable
       * then only for invariant or precise trees.
       */
      return visit_continue_with_parent;
   } else {
      return visit_continue;
   }
}

static inline bool
is_vec_zero(ir_constant *ir)
{
   return (ir == NULL) ? false : ir->is_zero();
}

static inline bool
is_vec_one(ir_constant *ir)
{
   return (ir == NULL) ? false : ir->is_one();
}

static inline bool
is_vec_negative_one(ir_constant *ir)
{
   return (ir == NULL) ? false : ir->is_negative_one();
}

static inline bool
is_valid_vec_const(ir_constant *ir)
{
   if (ir == NULL)
      return false;

   if (!ir->type->is_scalar() && !ir->type->is_vector())
      return false;

   return true;
}

static inline bool
is_less_than_one(ir_constant *ir)
{
   assert(ir->type->is_float());

   if (!is_valid_vec_const(ir))
      return false;

   unsigned component = 0;
   for (int c = 0; c < ir->type->vector_elements; c++) {
      if (ir->get_float_component(c) < 1.0f)
         component++;
   }

   return (component == ir->type->vector_elements);
}

static inline bool
is_greater_than_zero(ir_constant *ir)
{
   assert(ir->type->is_float());

   if (!is_valid_vec_const(ir))
      return false;

   unsigned component = 0;
   for (int c = 0; c < ir->type->vector_elements; c++) {
      if (ir->get_float_component(c) > 0.0f)
         component++;
   }

   return (component == ir->type->vector_elements);
}

static void
update_type(ir_expression *ir)
{
   if (ir->operands[0]->type->is_vector())
      ir->type = ir->operands[0]->type;
   else
      ir->type = ir->operands[1]->type;
}

void
ir_algebraic_visitor::reassociate_operands(ir_expression *ir1,
					   int op1,
					   ir_expression *ir2,
					   int op2)
{
   ir_rvalue *temp = ir2->operands[op2];
   ir2->operands[op2] = ir1->operands[op1];
   ir1->operands[op1] = temp;

   /* Update the type of ir2.  The type of ir1 won't have changed --
    * base types matched, and at least one of the operands of the 2
    * binops is still a vector if any of them were.
    */
   update_type(ir2);

   this->progress = true;
}

/**
 * Reassociates a constant down a tree of adds or multiplies.
 *
 * Consider (2 * (a * (b * 0.5))).  We want to end up with a * b.
 */
bool
ir_algebraic_visitor::reassociate_constant(ir_expression *ir1, int const_index,
					   ir_constant *constant,
					   ir_expression *ir2)
{
   if (!ir2 || ir1->operation != ir2->operation)
      return false;

   /* Don't want to even think about matrices. */
   if (ir1->operands[0]->type->is_matrix() ||
       ir1->operands[1]->type->is_matrix() ||
       ir2->operands[0]->type->is_matrix() ||
       ir2->operands[1]->type->is_matrix())
      return false;

   void *mem_ctx = ralloc_parent(ir2);

   ir_constant *ir2_const[2];
   ir2_const[0] = ir2->operands[0]->constant_expression_value(mem_ctx);
   ir2_const[1] = ir2->operands[1]->constant_expression_value(mem_ctx);

   if (ir2_const[0] && ir2_const[1])
      return false;

   if (ir2_const[0]) {
      reassociate_operands(ir1, const_index, ir2, 1);
      return true;
   } else if (ir2_const[1]) {
      reassociate_operands(ir1, const_index, ir2, 0);
      return true;
   }

   if (reassociate_constant(ir1, const_index, constant,
			    ir2->operands[0]->as_expression())) {
      update_type(ir2);
      return true;
   }

   if (reassociate_constant(ir1, const_index, constant,
			    ir2->operands[1]->as_expression())) {
      update_type(ir2);
      return true;
   }

   return false;
}

/* When eliminating an expression and just returning one of its operands,
 * we may need to swizzle that operand out to a vector if the expression was
 * vector type.
 */
ir_rvalue *
ir_algebraic_visitor::swizzle_if_required(ir_expression *expr,
					  ir_rvalue *operand)
{
   if (expr->type->is_vector() && operand->type->is_scalar()) {
      return new(mem_ctx) ir_swizzle(operand, 0, 0, 0, 0,
				     expr->type->vector_elements);
   } else
      return operand;
}

ir_rvalue *
ir_algebraic_visitor::handle_expression(ir_expression *ir)
{
   ir_constant *op_const[4] = {NULL, NULL, NULL, NULL};
   ir_expression *op_expr[4] = {NULL, NULL, NULL, NULL};

   if (ir->operation == ir_binop_mul &&
       ir->operands[0]->type->is_matrix() &&
       ir->operands[1]->type->is_vector()) {
      ir_expression *matrix_mul = ir->operands[0]->as_expression();

      if (matrix_mul && matrix_mul->operation == ir_binop_mul &&
         matrix_mul->operands[0]->type->is_matrix() &&
         matrix_mul->operands[1]->type->is_matrix()) {

         return mul(matrix_mul->operands[0],
                    mul(matrix_mul->operands[1], ir->operands[1]));
      }
   }

   assert(ir->num_operands <= 4);
   for (unsigned i = 0; i < ir->num_operands; i++) {
      if (ir->operands[i]->type->is_matrix())
	 return ir;

      op_const[i] =
         ir->operands[i]->constant_expression_value(ralloc_parent(ir));
      op_expr[i] = ir->operands[i]->as_expression();
   }

   if (this->mem_ctx == NULL)
      this->mem_ctx = ralloc_parent(ir);

   switch (ir->operation) {
   case ir_binop_add:
      /* Reassociate addition of constants so that we can do constant
       * folding.
       */
      if (op_const[0] && !op_const[1])
	 reassociate_constant(ir, 0, op_const[0], op_expr[1]);
      if (op_const[1] && !op_const[0])
	 reassociate_constant(ir, 1, op_const[1], op_expr[0]);

      /* Replace (-x + y) * a + x and commutative variations with lrp(x, y, a).
       *
       * (-x + y) * a + x
       * (x * -a) + (y * a) + x
       * x + (x * -a) + (y * a)
       * x * (1 - a) + y * a
       * lrp(x, y, a)
       */
      for (int mul_pos = 0; mul_pos < 2; mul_pos++) {
         ir_expression *mul = op_expr[mul_pos];

         if (!mul || mul->operation != ir_binop_mul)
            continue;

         /* Multiply found on one of the operands. Now check for an
          * inner addition operation.
          */
         for (int inner_add_pos = 0; inner_add_pos < 2; inner_add_pos++) {
            ir_expression *inner_add =
               mul->operands[inner_add_pos]->as_expression();

            if (!inner_add || inner_add->operation != ir_binop_add)
               continue;

            /* Inner addition found on one of the operands. Now check for
             * one of the operands of the inner addition to be the negative
             * of x_operand.
             */
            for (int neg_pos = 0; neg_pos < 2; neg_pos++) {
               ir_expression *neg =
                  inner_add->operands[neg_pos]->as_expression();

               if (!neg || neg->operation != ir_unop_neg)
                  continue;

               ir_rvalue *x_operand = ir->operands[1 - mul_pos];

               if (!neg->operands[0]->equals(x_operand))
                  continue;

               ir_rvalue *y_operand = inner_add->operands[1 - neg_pos];
               ir_rvalue *a_operand = mul->operands[1 - inner_add_pos];

               if (!x_operand->type->is_float_16_32_64() ||
                   x_operand->type != y_operand->type ||
                   x_operand->type != a_operand->type)
                  continue;

               return lrp(x_operand, y_operand, a_operand);
            }
         }
      }

      break;

   case ir_binop_mul:
      if (is_vec_one(op_const[0]))
	 return ir->operands[1];
      if (is_vec_one(op_const[1]))
	 return ir->operands[0];

      if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1]))
	 return ir_constant::zero(ir, ir->type);

      if (is_vec_negative_one(op_const[0]))
         return neg(ir->operands[1]);
      if (is_vec_negative_one(op_const[1]))
         return neg(ir->operands[0]);

      if (op_expr[0] && op_expr[0]->operation == ir_unop_b2f &&
          op_expr[1] && op_expr[1]->operation == ir_unop_b2f) {
         return b2f(logic_and(op_expr[0]->operands[0], op_expr[1]->operands[0]));
      }

      /* Reassociate multiplication of constants so that we can do
       * constant folding.
       */
      if (op_const[0] && !op_const[1])
	 reassociate_constant(ir, 0, op_const[0], op_expr[1]);
      if (op_const[1] && !op_const[0])
	 reassociate_constant(ir, 1, op_const[1], op_expr[0]);

      /* Optimizes
       *
       *    (mul (floor (add (abs x) 0.5) (sign x)))
       *
       * into
       *
       *    (trunc (add x (mul (sign x) 0.5)))
       */
      for (int i = 0; i < 2; i++) {
         ir_expression *sign_expr = ir->operands[i]->as_expression();
         ir_expression *floor_expr = ir->operands[1 - i]->as_expression();

         if (!sign_expr || sign_expr->operation != ir_unop_sign ||
             !floor_expr || floor_expr->operation != ir_unop_floor)
            continue;

         ir_expression *add_expr = floor_expr->operands[0]->as_expression();
         if (!add_expr || add_expr->operation != ir_binop_add)
            continue;

         for (int j = 0; j < 2; j++) {
            ir_expression *abs_expr = add_expr->operands[j]->as_expression();
            if (!abs_expr || abs_expr->operation != ir_unop_abs)
               continue;

            ir_constant *point_five = add_expr->operands[1 - j]->as_constant();
            if (!point_five || !point_five->is_value(0.5, 0))
               continue;

            if (abs_expr->operands[0]->equals(sign_expr->operands[0])) {
               return trunc(add(abs_expr->operands[0],
                                mul(sign_expr, point_five)));
            }
         }
      }
      break;

   case ir_binop_min:
   case ir_binop_max:
      if (!ir->type->is_float())
         break;

      /* Replace min(max) operations and its commutative combinations with
       * a saturate operation
       */
      for (int op = 0; op < 2; op++) {
         ir_expression *inner_expr = op_expr[op];
         ir_constant *outer_const = op_const[1 - op];
         ir_expression_operation op_cond = (ir->operation == ir_binop_max) ?
            ir_binop_min : ir_binop_max;

         if (!inner_expr || !outer_const || (inner_expr->operation != op_cond))
            continue;

         /* One of these has to be a constant */
         if (!inner_expr->operands[0]->as_constant() &&
             !inner_expr->operands[1]->as_constant())
            break;

         /* Found a min(max) combination. Now try to see if its operands
          * meet our conditions that we can do just a single saturate operation
          */
         for (int minmax_op = 0; minmax_op < 2; minmax_op++) {
            ir_rvalue *x = inner_expr->operands[minmax_op];
            ir_rvalue *y = inner_expr->operands[1 - minmax_op];

            ir_constant *inner_const = y->as_constant();
            if (!inner_const)
               continue;

            /* min(max(x, 0.0), 1.0) is sat(x) */
            if (ir->operation == ir_binop_min &&
                inner_const->is_zero() &&
                outer_const->is_one())
               return saturate(x);

            /* max(min(x, 1.0), 0.0) is sat(x) */
            if (ir->operation == ir_binop_max &&
                inner_const->is_one() &&
                outer_const->is_zero())
               return saturate(x);

            /* min(max(x, 0.0), b) where b < 1.0 is sat(min(x, b)) */
            if (ir->operation == ir_binop_min &&
                inner_const->is_zero() &&
                is_less_than_one(outer_const))
               return saturate(expr(ir_binop_min, x, outer_const));

            /* max(min(x, b), 0.0) where b < 1.0 is sat(min(x, b)) */
            if (ir->operation == ir_binop_max &&
                is_less_than_one(inner_const) &&
                outer_const->is_zero())
               return saturate(expr(ir_binop_min, x, inner_const));

            /* max(min(x, 1.0), b) where b > 0.0 is sat(max(x, b)) */
            if (ir->operation == ir_binop_max &&
                inner_const->is_one() &&
                is_greater_than_zero(outer_const))
               return saturate(expr(ir_binop_max, x, outer_const));

            /* min(max(x, b), 1.0) where b > 0.0 is sat(max(x, b)) */
            if (ir->operation == ir_binop_min &&
                is_greater_than_zero(inner_const) &&
                outer_const->is_one())
               return saturate(expr(ir_binop_max, x, inner_const));
         }
      }

      break;

   case ir_triop_fma:
      /* Operands are op0 * op1 + op2. */
      if (is_vec_zero(op_const[0]) || is_vec_zero(op_const[1])) {
         return ir->operands[2];
      } else if (is_vec_zero(op_const[2])) {
         return mul(ir->operands[0], ir->operands[1]);
      } else if (is_vec_one(op_const[0])) {
         return add(ir->operands[1], ir->operands[2]);
      } else if (is_vec_one(op_const[1])) {
         return add(ir->operands[0], ir->operands[2]);
      }
      break;

   case ir_triop_lrp:
      /* Operands are (x, y, a). */
      if (is_vec_zero(op_const[2])) {
         return ir->operands[0];
      } else if (is_vec_one(op_const[2])) {
         return ir->operands[1];
      } else if (ir->operands[0]->equals(ir->operands[1])) {
         return ir->operands[0];
      } else if (is_vec_zero(op_const[0])) {
         return mul(ir->operands[1], ir->operands[2]);
      } else if (is_vec_zero(op_const[1])) {
         unsigned op2_components = ir->operands[2]->type->vector_elements;
         ir_constant *one;

         switch (ir->type->base_type) {
         case GLSL_TYPE_FLOAT16:
            one = new(mem_ctx) ir_constant(float16_t::one(), op2_components);
            break;
         case GLSL_TYPE_FLOAT:
            one = new(mem_ctx) ir_constant(1.0f, op2_components);
            break;
         case GLSL_TYPE_DOUBLE:
            one = new(mem_ctx) ir_constant(1.0, op2_components);
            break;
         default:
            one = NULL;
            unreachable("unexpected type");
         }

         return mul(ir->operands[0], add(one, neg(ir->operands[2])));
      }
      break;

   /* Remove interpolateAt* instructions for demoted inputs. They are
    * assigned a constant expression to facilitate this.
    */
   case ir_unop_interpolate_at_centroid:
   case ir_binop_interpolate_at_offset:
   case ir_binop_interpolate_at_sample:
      if (op_const[0])
         return ir->operands[0];
      break;

   default:
      break;
   }

   return ir;
}

void
ir_algebraic_visitor::handle_rvalue(ir_rvalue **rvalue)
{
   if (!*rvalue)
      return;

   ir_expression *expr = (*rvalue)->as_expression();
   if (!expr || expr->operation == ir_quadop_vector)
      return;

   ir_rvalue *new_rvalue = handle_expression(expr);
   if (new_rvalue == *rvalue)
      return;

   /* If the expr used to be some vec OP scalar returning a vector, and the
    * optimization gave us back a scalar, we still need to turn it into a
    * vector.
    */
   *rvalue = swizzle_if_required(expr, new_rvalue);

   this->progress = true;
}

bool
do_algebraic(exec_list *instructions, bool native_integers,
             const struct gl_shader_compiler_options *options)
{
   ir_algebraic_visitor v(native_integers, options);

   visit_list_elements(&v, instructions);

   return v.progress;
}