glsl: Use a separate div_to_mul_rcp lowering flag for integers.

Using multiply and reciprocal for integer division involves potentially
lossy floating point conversions.  This is okay for older GPUs that
represent integers as floating point, but undesirable for GPUs with
native integer division instructions.

TGSI, for example, has UDIV/IDIV instructions for integer division,
so it makes sense to handle this directly.  Likewise for i965.

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Signed-off-by: Bryan Cain <bryancain3@gmail.com>
Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>

5 files changed, 80 insertions, 63 deletions

diff --git a/src/glsl/ir_optimization.h b/src/glsl/ir_optimization.h
index f7808bdda9..48448d4a16 100644
--- a/src/glsl/ir_optimization.h
+++ b/src/glsl/ir_optimization.h
@@ -29,12 +29,13 @@
  */
 
 /* Operations for lower_instructions() */
-#define SUB_TO_ADD_NEG 0x01
-#define DIV_TO_MUL_RCP 0x02
-#define EXP_TO_EXP2    0x04
-#define POW_TO_EXP2    0x08
-#define LOG_TO_LOG2    0x10
-#define MOD_TO_FRACT   0x20
+#define SUB_TO_ADD_NEG     0x01
+#define DIV_TO_MUL_RCP     0x02
+#define EXP_TO_EXP2        0x04
+#define POW_TO_EXP2        0x08
+#define LOG_TO_LOG2        0x10
+#define MOD_TO_FRACT       0x20
+#define INT_DIV_TO_MUL_RCP 0x40
 
 bool do_common_optimization(exec_list *ir, bool linked, unsigned max_unroll_iterations);
 
diff --git a/src/glsl/lower_instructions.cpp b/src/glsl/lower_instructions.cpp
index 23aa19bde6..d79eb0a7ff 100644
--- a/src/glsl/lower_instructions.cpp
+++ b/src/glsl/lower_instructions.cpp
@@ -32,6 +32,7 @@
  * Currently supported transformations:
  * - SUB_TO_ADD_NEG
  * - DIV_TO_MUL_RCP
+ * - INT_DIV_TO_MUL_RCP
  * - EXP_TO_EXP2
  * - POW_TO_EXP2
  * - LOG_TO_LOG2
@@ -47,8 +48,8 @@
  * want to recognize add(op0, neg(op1)) or the other way around to
  * produce a subtract anyway.
  *
- * DIV_TO_MUL_RCP:
- * ---------------
+ * DIV_TO_MUL_RCP and INT_DIV_TO_MUL_RCP:
+ * --------------------------------------
  * Breaks an ir_unop_div expression down to op0 * (rcp(op1)).
  *
  * Many GPUs don't have a divide instruction (945 and 965 included),
@@ -56,6 +57,10 @@
  * reciprocal.  By breaking the operation down, constant reciprocals
  * can get constant folded.
  *
+ * DIV_TO_MUL_RCP only lowers floating point division; INT_DIV_TO_MUL_RCP
+ * handles the integer case, converting to and from floating point so that
+ * RCP is possible.
+ *
  * EXP_TO_EXP2 and LOG_TO_LOG2:
  * ----------------------------
  * Many GPUs don't have a base e log or exponent instruction, but they
@@ -95,6 +100,7 @@ private:
 
    void sub_to_add_neg(ir_expression *);
    void div_to_mul_rcp(ir_expression *);
+   void int_div_to_mul_rcp(ir_expression *);
    void mod_to_fract(ir_expression *);
    void exp_to_exp2(ir_expression *);
    void pow_to_exp2(ir_expression *);
@@ -127,60 +133,67 @@ lower_instructions_visitor::sub_to_add_neg(ir_expression *ir)
 void
 lower_instructions_visitor::div_to_mul_rcp(ir_expression *ir)
 {
-   if (!ir->operands[1]->type->is_integer()) {
-      /* New expression for the 1.0 / op1 */
-      ir_rvalue *expr;
-      expr = new(ir) ir_expression(ir_unop_rcp,
-				   ir->operands[1]->type,
-				   ir->operands[1],
-				   NULL);
-
-      /* op0 / op1 -> op0 * (1.0 / op1) */
-      ir->operation = ir_binop_mul;
-      ir->operands[1] = expr;
+   assert(ir->operands[1]->type->is_float());
+
+   /* New expression for the 1.0 / op1 */
+   ir_rvalue *expr;
+   expr = new(ir) ir_expression(ir_unop_rcp,
+				ir->operands[1]->type,
+				ir->operands[1]);
+
+   /* op0 / op1 -> op0 * (1.0 / op1) */
+   ir->operation = ir_binop_mul;
+   ir->operands[1] = expr;
+
+   this->progress = true;
+}
+
+void
+lower_instructions_visitor::int_div_to_mul_rcp(ir_expression *ir)
+{
+   assert(ir->operands[1]->type->is_integer());
+
+   /* Be careful with integer division -- we need to do it as a
+    * float and re-truncate, since rcp(n > 1) of an integer would
+    * just be 0.
+    */
+   ir_rvalue *op0, *op1;
+   const struct glsl_type *vec_type;
+
+   vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
+				      ir->operands[1]->type->vector_elements,
+				      ir->operands[1]->type->matrix_columns);
+
+   if (ir->operands[1]->type->base_type == GLSL_TYPE_INT)
+      op1 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[1], NULL);
+   else
+      op1 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[1], NULL);
+
+   op1 = new(ir) ir_expression(ir_unop_rcp, op1->type, op1, NULL);
+
+   vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
+				      ir->operands[0]->type->vector_elements,
+				      ir->operands[0]->type->matrix_columns);
+
+   if (ir->operands[0]->type->base_type == GLSL_TYPE_INT)
+      op0 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[0], NULL);
+   else
+      op0 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[0], NULL);
+
+   vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
+				      ir->type->vector_elements,
+				      ir->type->matrix_columns);
+
+   op0 = new(ir) ir_expression(ir_binop_mul, vec_type, op0, op1);
+
+   if (ir->operands[1]->type->base_type == GLSL_TYPE_INT) {
+      ir->operation = ir_unop_f2i;
+      ir->operands[0] = op0;
    } else {
-      /* Be careful with integer division -- we need to do it as a
-       * float and re-truncate, since rcp(n > 1) of an integer would
-       * just be 0.
-       */
-      ir_rvalue *op0, *op1;
-      const struct glsl_type *vec_type;
-
-      vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
-					 ir->operands[1]->type->vector_elements,
-					 ir->operands[1]->type->matrix_columns);
-
-      if (ir->operands[1]->type->base_type == GLSL_TYPE_INT)
-	 op1 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[1], NULL);
-      else
-	 op1 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[1], NULL);
-
-      op1 = new(ir) ir_expression(ir_unop_rcp, op1->type, op1, NULL);
-
-      vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
-					 ir->operands[0]->type->vector_elements,
-					 ir->operands[0]->type->matrix_columns);
-
-      if (ir->operands[0]->type->base_type == GLSL_TYPE_INT)
-	 op0 = new(ir) ir_expression(ir_unop_i2f, vec_type, ir->operands[0], NULL);
-      else
-	 op0 = new(ir) ir_expression(ir_unop_u2f, vec_type, ir->operands[0], NULL);
-
-      vec_type = glsl_type::get_instance(GLSL_TYPE_FLOAT,
-					 ir->type->vector_elements,
-					 ir->type->matrix_columns);
-
-      op0 = new(ir) ir_expression(ir_binop_mul, vec_type, op0, op1);
-
-      if (ir->operands[1]->type->base_type == GLSL_TYPE_INT) {
-	 ir->operation = ir_unop_f2i;
-	 ir->operands[0] = op0;
-      } else {
-	 ir->operation = ir_unop_i2u;
-	 ir->operands[0] = new(ir) ir_expression(ir_unop_f2i, op0);
-      }
-      ir->operands[1] = NULL;
+      ir->operation = ir_unop_i2u;
+      ir->operands[0] = new(ir) ir_expression(ir_unop_f2i, op0);
    }
+   ir->operands[1] = NULL;
 
    this->progress = true;
 }
@@ -265,7 +278,9 @@ lower_instructions_visitor::visit_leave(ir_expression *ir)
       break;
 
    case ir_binop_div:
-      if (lowering(DIV_TO_MUL_RCP))
+      if (ir->operands[1]->type->is_integer() && lowering(INT_DIV_TO_MUL_RCP))
+	 int_div_to_mul_rcp(ir);
+      else if (ir->operands[1]->type->is_float() && lowering(DIV_TO_MUL_RCP))
 	 div_to_mul_rcp(ir);
       break;
 
diff --git a/src/mesa/drivers/dri/i965/brw_shader.cpp b/src/mesa/drivers/dri/i965/brw_shader.cpp
index 0a21094708..a6ed810a16 100644
--- a/src/mesa/drivers/dri/i965/brw_shader.cpp
+++ b/src/mesa/drivers/dri/i965/brw_shader.cpp
@@ -100,6 +100,7 @@ brw_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
       lower_instructions(shader->ir,
 			 MOD_TO_FRACT |
 			 DIV_TO_MUL_RCP |
+			 INT_DIV_TO_MUL_RCP |
 			 SUB_TO_ADD_NEG |
 			 EXP_TO_EXP2 |
 			 LOG_TO_LOG2);
diff --git a/src/mesa/program/ir_to_mesa.cpp b/src/mesa/program/ir_to_mesa.cpp
index 6820e4c6ba..dd154db8b0 100644
--- a/src/mesa/program/ir_to_mesa.cpp
+++ b/src/mesa/program/ir_to_mesa.cpp
@@ -3232,7 +3232,7 @@ _mesa_ir_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
 	 /* Lowering */
 	 do_mat_op_to_vec(ir);
 	 lower_instructions(ir, (MOD_TO_FRACT | DIV_TO_MUL_RCP | EXP_TO_EXP2
-				 | LOG_TO_LOG2
+				 | LOG_TO_LOG2 | INT_DIV_TO_MUL_RCP
 				 | ((options->EmitNoPow) ? POW_TO_EXP2 : 0)));
 
 	 progress = do_lower_jumps(ir, true, true, options->EmitNoMainReturn, options->EmitNoCont, options->EmitNoLoops) || progress;
diff --git a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
index 3fbb0cdd24..98bea69b95 100644
--- a/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
+++ b/src/mesa/state_tracker/st_glsl_to_tgsi.cpp
@@ -4982,7 +4982,7 @@ st_link_shader(struct gl_context *ctx, struct gl_shader_program *prog)
          /* Lowering */
          do_mat_op_to_vec(ir);
          lower_instructions(ir, (MOD_TO_FRACT | DIV_TO_MUL_RCP | EXP_TO_EXP2
-        			 | LOG_TO_LOG2
+				 | LOG_TO_LOG2 | INT_DIV_TO_MUL_RCP
         			 | ((options->EmitNoPow) ? POW_TO_EXP2 : 0)));
 
          progress = do_lower_jumps(ir, true, true, options->EmitNoMainReturn, options->EmitNoCont, options->EmitNoLoops) || progress;