/* * 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 lower_instructions.cpp * * Many GPUs lack native instructions for certain expression operations, and * must replace them with some other expression tree. This pass lowers some * of the most common cases, allowing the lowering code to be implemented once * rather than in each driver backend. * * 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 * - MOD_TO_FLOOR * - LDEXP_TO_ARITH * - BITFIELD_INSERT_TO_BFM_BFI * - CARRY_TO_ARITH * - BORROW_TO_ARITH * - SAT_TO_CLAMP * * SUB_TO_ADD_NEG: * --------------- * Breaks an ir_binop_sub expression down to add(op0, neg(op1)) * * This simplifies expression reassociation, and for many backends * there is no subtract operation separate from adding the negation. * For backends with native subtract operations, they will probably * want to recognize add(op0, neg(op1)) or the other way around to * produce a subtract anyway. * * DIV_TO_MUL_RCP and INT_DIV_TO_MUL_RCP: * -------------------------------------- * Breaks an ir_binop_div expression down to op0 * (rcp(op1)). * * Many GPUs don't have a divide instruction (945 and 965 included), * but they do have an RCP instruction to compute an approximate * 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 * do have base 2 versions, so this pass converts exp and log to exp2 * and log2 operations. * * POW_TO_EXP2: * ----------- * Many older GPUs don't have an x**y instruction. For these GPUs, convert * x**y to 2**(y * log2(x)). * * MOD_TO_FLOOR: * ------------- * 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. * * BITFIELD_INSERT_TO_BFM_BFI: * --------------------------- * Breaks ir_quadop_bitfield_insert into ir_binop_bfm (bitfield mask) and * ir_triop_bfi (bitfield insert). * * Many GPUs implement the bitfieldInsert() built-in from ARB_gpu_shader_5 * with a pair of instructions. * * CARRY_TO_ARITH: * --------------- * Converts ir_carry into (x + y) < x. * * BORROW_TO_ARITH: * ---------------- * Converts ir_borrow into (x < y). * * SAT_TO_CLAMP: * ------------- * Converts ir_unop_saturate into min(max(x, 0.0), 1.0) * */ #include "main/core.h" /* for M_LOG2E */ #include "program/prog_instruction.h" /* for swizzle */ #include "glsl_types.h" #include "ir.h" #include "ir_builder.h" #include "ir_optimization.h" using namespace ir_builder; namespace { class lower_instructions_visitor : public ir_hierarchical_visitor { public: lower_instructions_visitor(unsigned lower) : progress(false), lower(lower) { } ir_visitor_status visit_leave(ir_expression *); bool progress; private: unsigned lower; /** Bitfield of which operations to lower */ 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_floor(ir_expression *); void exp_to_exp2(ir_expression *); void pow_to_exp2(ir_expression *); void log_to_log2(ir_expression *); void bitfield_insert_to_bfm_bfi(ir_expression *); void ldexp_to_arith(ir_expression *); void carry_to_arith(ir_expression *); void borrow_to_arith(ir_expression *); void sat_to_clamp(ir_expression *); void double_dot_to_fma(ir_expression *); void double_lrp(ir_expression *); }; } /* anonymous namespace */ /** * Determine if a particular type of lowering should occur */ #define lowering(x) (this->lower & x) bool lower_instructions(exec_list *instructions, unsigned what_to_lower) { lower_instructions_visitor v(what_to_lower); visit_list_elements(&v, instructions); return v.progress; } void lower_instructions_visitor::sub_to_add_neg(ir_expression *ir) { ir->operation = ir_binop_add; ir->operands[1] = new(ir) ir_expression(ir_unop_neg, ir->operands[1]->type, ir->operands[1], NULL); this->progress = true; } void lower_instructions_visitor::div_to_mul_rcp(ir_expression *ir) { 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 { ir->operation = ir_unop_i2u; ir->operands[0] = new(ir) ir_expression(ir_unop_f2i, op0); } ir->operands[1] = NULL; this->progress = true; } void lower_instructions_visitor::exp_to_exp2(ir_expression *ir) { ir_constant *log2_e = new(ir) ir_constant(float(M_LOG2E)); ir->operation = ir_unop_exp2; ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[0]->type, ir->operands[0], log2_e); this->progress = true; } void lower_instructions_visitor::pow_to_exp2(ir_expression *ir) { ir_expression *const log2_x = new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type, ir->operands[0]); ir->operation = ir_unop_exp2; ir->operands[0] = new(ir) ir_expression(ir_binop_mul, ir->operands[1]->type, ir->operands[1], log2_x); ir->operands[1] = NULL; this->progress = true; } void lower_instructions_visitor::log_to_log2(ir_expression *ir) { ir->operation = ir_binop_mul; ir->operands[0] = new(ir) ir_expression(ir_unop_log2, ir->operands[0]->type, ir->operands[0], NULL); ir->operands[1] = new(ir) ir_constant(float(1.0 / M_LOG2E)); this->progress = true; } void lower_instructions_visitor::mod_to_floor(ir_expression *ir) { 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, 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. */ if (lowering(DIV_TO_MUL_RCP) && ir->type->is_float()) div_to_mul_rcp(div_expr); ir_expression *const floor_expr = new(ir) ir_expression(ir_unop_floor, x->type, div_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; } void lower_instructions_visitor::bitfield_insert_to_bfm_bfi(ir_expression *ir) { /* Translates * ir_quadop_bitfield_insert base insert offset bits * into * ir_triop_bfi (ir_binop_bfm bits offset) insert base */ ir_rvalue *base_expr = ir->operands[0]; ir->operation = ir_triop_bfi; ir->operands[0] = new(ir) ir_expression(ir_binop_bfm, ir->type->get_base_type(), ir->operands[3], ir->operands[2]); /* ir->operands[1] is still the value to insert. */ ir->operands[2] = base_expr; ir->operands[3] = NULL; this->progress = true; } void lower_instructions_visitor::ldexp_to_arith(ir_expression *ir) { /* Translates * ir_binop_ldexp x exp * into * * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift); * resulting_biased_exp = extracted_biased_exp + exp; * * if (resulting_biased_exp < 1) { * return copysign(0.0, x); * } * * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) | * lshift(i2u(resulting_biased_exp), exp_shift)); * * which we can't actually implement as such, since the GLSL IR doesn't * have vectorized if-statements. We actually implement it without branches * using conditional-select: * * extracted_biased_exp = rshift(bitcast_f2i(abs(x)), exp_shift); * resulting_biased_exp = extracted_biased_exp + exp; * * is_not_zero_or_underflow = gequal(resulting_biased_exp, 1); * x = csel(is_not_zero_or_underflow, x, copysign(0.0f, x)); * resulting_biased_exp = csel(is_not_zero_or_underflow, * resulting_biased_exp, 0); * * return bitcast_u2f((bitcast_f2u(x) & sign_mantissa_mask) | * lshift(i2u(resulting_biased_exp), exp_shift)); */ const unsigned vec_elem = ir->type->vector_elements; /* Types */ const glsl_type *ivec = glsl_type::get_instance(GLSL_TYPE_INT, vec_elem, 1); const glsl_type *bvec = glsl_type::get_instance(GLSL_TYPE_BOOL, vec_elem, 1); /* Constants */ ir_constant *zeroi = ir_constant::zero(ir, ivec); ir_constant *sign_mask = new(ir) ir_constant(0x80000000u, vec_elem); ir_constant *exp_shift = new(ir) ir_constant(23); ir_constant *exp_width = new(ir) ir_constant(8); /* Temporary variables */ ir_variable *x = new(ir) ir_variable(ir->type, "x", ir_var_temporary); ir_variable *exp = new(ir) ir_variable(ivec, "exp", ir_var_temporary); ir_variable *zero_sign_x = new(ir) ir_variable(ir->type, "zero_sign_x", ir_var_temporary); ir_variable *extracted_biased_exp = new(ir) ir_variable(ivec, "extracted_biased_exp", ir_var_temporary); ir_variable *resulting_biased_exp = new(ir) ir_variable(ivec, "resulting_biased_exp", ir_var_temporary); ir_variable *is_not_zero_or_underflow = new(ir) ir_variable(bvec, "is_not_zero_or_underflow", ir_var_temporary); ir_instruction &i = *base_ir; /* Copy and arguments. */ i.insert_before(x); i.insert_before(assign(x, ir->operands[0])); i.insert_before(exp); i.insert_before(assign(exp, ir->operands[1])); /* Extract the biased exponent from . */ i.insert_before(extracted_biased_exp); i.insert_before(assign(extracted_biased_exp, rshift(bitcast_f2i(abs(x)), exp_shift))); i.insert_before(resulting_biased_exp); i.insert_before(assign(resulting_biased_exp, add(extracted_biased_exp, exp))); /* Test if result is ±0.0, subnormal, or underflow by checking if the * resulting biased exponent would be less than 0x1. If so, the result is * 0.0 with the sign of x. (Actually, invert the conditions so that * immediate values are the second arguments, which is better for i965) */ i.insert_before(zero_sign_x); i.insert_before(assign(zero_sign_x, bitcast_u2f(bit_and(bitcast_f2u(x), sign_mask)))); i.insert_before(is_not_zero_or_underflow); i.insert_before(assign(is_not_zero_or_underflow, gequal(resulting_biased_exp, new(ir) ir_constant(0x1, vec_elem)))); i.insert_before(assign(x, csel(is_not_zero_or_underflow, x, zero_sign_x))); i.insert_before(assign(resulting_biased_exp, csel(is_not_zero_or_underflow, resulting_biased_exp, zeroi))); /* We could test for overflows by checking if the resulting biased exponent * would be greater than 0xFE. Turns out we don't need to because the GLSL * spec says: * * "If this product is too large to be represented in the * floating-point type, the result is undefined." */ ir_constant *exp_shift_clone = exp_shift->clone(ir, NULL); ir->operation = ir_unop_bitcast_i2f; ir->operands[0] = bitfield_insert(bitcast_f2i(x), resulting_biased_exp, exp_shift_clone, exp_width); ir->operands[1] = NULL; /* Don't generate new IR that would need to be lowered in an additional * pass. */ if (lowering(BITFIELD_INSERT_TO_BFM_BFI)) bitfield_insert_to_bfm_bfi(ir->operands[0]->as_expression()); this->progress = true; } void lower_instructions_visitor::carry_to_arith(ir_expression *ir) { /* Translates * ir_binop_carry x y * into * sum = ir_binop_add x y * bcarry = ir_binop_less sum x * carry = ir_unop_b2i bcarry */ ir_rvalue *x_clone = ir->operands[0]->clone(ir, NULL); ir->operation = ir_unop_i2u; ir->operands[0] = b2i(less(add(ir->operands[0], ir->operands[1]), x_clone)); ir->operands[1] = NULL; this->progress = true; } void lower_instructions_visitor::borrow_to_arith(ir_expression *ir) { /* Translates * ir_binop_borrow x y * into * bcarry = ir_binop_less x y * carry = ir_unop_b2i bcarry */ ir->operation = ir_unop_i2u; ir->operands[0] = b2i(less(ir->operands[0], ir->operands[1])); ir->operands[1] = NULL; this->progress = true; } void lower_instructions_visitor::sat_to_clamp(ir_expression *ir) { /* Translates * ir_unop_saturate x * into * ir_binop_min (ir_binop_max(x, 0.0), 1.0) */ ir->operation = ir_binop_min; ir->operands[0] = new(ir) ir_expression(ir_binop_max, ir->operands[0]->type, ir->operands[0], new(ir) ir_constant(0.0f)); ir->operands[1] = new(ir) ir_constant(1.0f); this->progress = true; } void lower_instructions_visitor::double_dot_to_fma(ir_expression *ir) { ir_variable *temp = new(ir) ir_variable(ir->operands[0]->type->get_base_type(), "dot_res", ir_var_temporary); this->base_ir->insert_before(temp); int nc = ir->operands[0]->type->components(); for (int i = nc - 1; i >= 1; i--) { ir_assignment *assig; if (i == (nc - 1)) { assig = assign(temp, mul(swizzle(ir->operands[0]->clone(ir, NULL), i, 1), swizzle(ir->operands[1]->clone(ir, NULL), i, 1))); } else { assig = assign(temp, fma(swizzle(ir->operands[0]->clone(ir, NULL), i, 1), swizzle(ir->operands[1]->clone(ir, NULL), i, 1), temp)); } this->base_ir->insert_before(assig); } ir->operation = ir_triop_fma; ir->operands[0] = swizzle(ir->operands[0], 0, 1); ir->operands[1] = swizzle(ir->operands[1], 0, 1); ir->operands[2] = new(ir) ir_dereference_variable(temp); this->progress = true; } void lower_instructions_visitor::double_lrp(ir_expression *ir) { int swizval; ir_rvalue *op0 = ir->operands[0], *op2 = ir->operands[2]; ir_constant *one = new(ir) ir_constant(1.0, op2->type->vector_elements); switch (op2->type->vector_elements) { case 1: swizval = SWIZZLE_XXXX; break; default: assert(op0->type->vector_elements == op2->type->vector_elements); swizval = SWIZZLE_XYZW; break; } ir->operation = ir_triop_fma; ir->operands[0] = swizzle(op2, swizval, op0->type->vector_elements); ir->operands[2] = mul(sub(one, op2->clone(ir, NULL)), op0); this->progress = true; } ir_visitor_status lower_instructions_visitor::visit_leave(ir_expression *ir) { switch (ir->operation) { case ir_binop_dot: if (ir->operands[0]->type->is_double()) double_dot_to_fma(ir); break; case ir_triop_lrp: if (ir->operands[0]->type->is_double()) double_lrp(ir); break; case ir_binop_sub: if (lowering(SUB_TO_ADD_NEG)) sub_to_add_neg(ir); break; case ir_binop_div: 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; case ir_unop_exp: if (lowering(EXP_TO_EXP2)) exp_to_exp2(ir); break; case ir_unop_log: if (lowering(LOG_TO_LOG2)) log_to_log2(ir); break; case ir_binop_mod: if (lowering(MOD_TO_FLOOR) && (ir->type->is_float() || ir->type->is_double())) mod_to_floor(ir); break; case ir_binop_pow: if (lowering(POW_TO_EXP2)) pow_to_exp2(ir); break; case ir_quadop_bitfield_insert: if (lowering(BITFIELD_INSERT_TO_BFM_BFI)) bitfield_insert_to_bfm_bfi(ir); break; case ir_binop_ldexp: if (lowering(LDEXP_TO_ARITH) && ir->type->is_float()) ldexp_to_arith(ir); break; case ir_binop_carry: if (lowering(CARRY_TO_ARITH)) carry_to_arith(ir); break; case ir_binop_borrow: if (lowering(BORROW_TO_ARITH)) borrow_to_arith(ir); break; case ir_unop_saturate: if (lowering(SAT_TO_CLAMP)) sat_to_clamp(ir); break; default: return visit_continue; } return visit_continue; }