diff options
Diffstat (limited to 'src/glsl/lower_instructions.cpp')
| -rw-r--r-- | src/glsl/lower_instructions.cpp | 65 |
1 files changed, 39 insertions, 26 deletions
diff --git a/src/glsl/lower_instructions.cpp b/src/glsl/lower_instructions.cpp index 6842853..09afe55 100644 --- a/src/glsl/lower_instructions.cpp +++ b/src/glsl/lower_instructions.cpp @@ -36,7 +36,7 @@ * - EXP_TO_EXP2 * - POW_TO_EXP2 * - LOG_TO_LOG2 - * - MOD_TO_FRACT + * - MOD_TO_FLOOR * - LDEXP_TO_ARITH * - BITFIELD_INSERT_TO_BFM_BFI * - CARRY_TO_ARITH @@ -77,14 +77,17 @@ * Many older GPUs don't have an x**y instruction. For these GPUs, convert * x**y to 2**(y * log2(x)). * - * MOD_TO_FRACT: + * MOD_TO_FLOOR: * ------------- - * Breaks an ir_binop_mod expression down to (op1 * fract(op0 / op1)) + * Breaks an ir_binop_mod expression down to (op0 - op1 * floor(op0 / op1)) * * Many GPUs don't have a MOD instruction (945 and 965 included), and * if we have to break it down like this anyway, it gives an * opportunity to do things like constant fold the (1.0 / op1) easily. * + * Note: before we used to implement this as op1 * fract(op / op1) but this + * implementation had significant precision errors. + * * LDEXP_TO_ARITH: * ------------- * Converts ir_binop_ldexp to arithmetic and bit operations. @@ -136,7 +139,7 @@ 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 mod_to_floor(ir_expression *); void exp_to_exp2(ir_expression *); void pow_to_exp2(ir_expression *); void log_to_log2(ir_expression *); @@ -276,22 +279,29 @@ lower_instructions_visitor::log_to_log2(ir_expression *ir) } void -lower_instructions_visitor::mod_to_fract(ir_expression *ir) +lower_instructions_visitor::mod_to_floor(ir_expression *ir) { - ir_variable *temp = new(ir) ir_variable(ir->operands[1]->type, "mod_b", - ir_var_temporary); - this->base_ir->insert_before(temp); - - ir_assignment *const assign = - new(ir) ir_assignment(new(ir) ir_dereference_variable(temp), - ir->operands[1], NULL); - - this->base_ir->insert_before(assign); + ir_variable *x = new(ir) ir_variable(ir->operands[0]->type, "mod_x", + ir_var_temporary); + ir_variable *y = new(ir) ir_variable(ir->operands[1]->type, "mod_y", + ir_var_temporary); + this->base_ir->insert_before(x); + this->base_ir->insert_before(y); + + ir_assignment *const assign_x = + new(ir) ir_assignment(new(ir) ir_dereference_variable(x), + ir->operands[0], NULL); + ir_assignment *const assign_y = + new(ir) ir_assignment(new(ir) ir_dereference_variable(y), + ir->operands[1], NULL); + + this->base_ir->insert_before(assign_x); + this->base_ir->insert_before(assign_y); ir_expression *const div_expr = - new(ir) ir_expression(ir_binop_div, ir->operands[0]->type, - ir->operands[0], - new(ir) ir_dereference_variable(temp)); + new(ir) ir_expression(ir_binop_div, x->type, + new(ir) ir_dereference_variable(x), + new(ir) ir_dereference_variable(y)); /* Don't generate new IR that would need to be lowered in an additional * pass. @@ -299,14 +309,17 @@ lower_instructions_visitor::mod_to_fract(ir_expression *ir) if (lowering(DIV_TO_MUL_RCP)) div_to_mul_rcp(div_expr); - ir_rvalue *expr = new(ir) ir_expression(ir_unop_fract, - ir->operands[0]->type, - div_expr, - NULL); + ir_expression *const floor_expr = + new(ir) ir_expression(ir_unop_floor, x->type, div_expr); - ir->operation = ir_binop_mul; - ir->operands[0] = new(ir) ir_dereference_variable(temp); - ir->operands[1] = expr; + ir_expression *const mul_expr = + new(ir) ir_expression(ir_binop_mul, + new(ir) ir_dereference_variable(y), + floor_expr); + + ir->operation = ir_binop_sub; + ir->operands[0] = new(ir) ir_dereference_variable(x); + ir->operands[1] = mul_expr; this->progress = true; } @@ -535,8 +548,8 @@ lower_instructions_visitor::visit_leave(ir_expression *ir) break; case ir_binop_mod: - if (lowering(MOD_TO_FRACT) && ir->type->is_float()) - mod_to_fract(ir); + if (lowering(MOD_TO_FLOOR) && ir->type->is_float()) + mod_to_floor(ir); break; case ir_binop_pow: |