SPIRV-Tools/source/opt/const_folding_rules.cpp
Steven Perron 6669d8163d Fold binary floating point operators.
Adds the floating rules for FAdd, FDiv, FMul, and FSub.

Contributes to #1164.
2018-02-14 15:48:15 -05:00

266 lines
11 KiB
C++

// Copyright (c) 2018 Google LLC
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "const_folding_rules.h"
namespace spvtools {
namespace opt {
namespace {
const uint32_t kExtractCompositeIdInIdx = 0;
// Returns a vector that contains the two 32-bit integers that result from
// splitting |a| in two. The first entry in vector are the low order bit if
// |a|.
inline std::vector<uint32_t> ExtractInts(uint64_t a) {
std::vector<uint32_t> result;
result.push_back(static_cast<uint32_t>(a));
result.push_back(static_cast<uint32_t>(a >> 32));
return result;
}
// Returns true if we are allowed to fold or otherwise manipulate the
// instruction that defines |id| in the given context.
bool CanFoldFloatingPoint(ir::IRContext* context, uint32_t id) {
// TODO: Add the rules for kernels. For now it will be pessimistic.
if (!context->get_feature_mgr()->HasCapability(SpvCapabilityShader)) {
return false;
}
bool is_nocontract = false;
context->get_decoration_mgr()->WhileEachDecoration(
id, SpvDecorationNoContraction, [&is_nocontract](const ir::Instruction&) {
is_nocontract = true;
return false;
});
return !is_nocontract;
}
// Folds an OpcompositeExtract where input is a composite constant.
ConstantFoldingRule FoldExtractWithConstants() {
return [](ir::Instruction* inst,
const std::vector<const analysis::Constant*>& constants)
-> const analysis::Constant* {
const analysis::Constant* c = constants[kExtractCompositeIdInIdx];
if (c == nullptr) {
return nullptr;
}
for (uint32_t i = 1; i < inst->NumInOperands(); ++i) {
uint32_t element_index = inst->GetSingleWordInOperand(i);
if (c->AsNullConstant()) {
// Return Null for the return type.
ir::IRContext* context = inst->context();
analysis::ConstantManager* const_mgr = context->get_constant_mgr();
analysis::TypeManager* type_mgr = context->get_type_mgr();
return const_mgr->GetConstant(type_mgr->GetType(inst->type_id()), {});
}
auto cc = c->AsCompositeConstant();
assert(cc != nullptr);
auto components = cc->GetComponents();
c = components[element_index];
}
return c;
};
}
ConstantFoldingRule FoldCompositeWithConstants() {
// Folds an OpCompositeConstruct where all of the inputs are constants to a
// constant. A new constant is created if necessary.
return [](ir::Instruction* inst,
const std::vector<const analysis::Constant*>& constants)
-> const analysis::Constant* {
ir::IRContext* context = inst->context();
analysis::ConstantManager* const_mgr = context->get_constant_mgr();
analysis::TypeManager* type_mgr = context->get_type_mgr();
const analysis::Type* new_type = type_mgr->GetType(inst->type_id());
std::vector<uint32_t> ids;
for (const analysis::Constant* element_const : constants) {
if (element_const == nullptr) {
return nullptr;
}
uint32_t element_id = const_mgr->FindDeclaredConstant(element_const);
if (element_id == 0) {
return nullptr;
}
ids.push_back(element_id);
}
return const_mgr->GetConstant(new_type, ids);
};
}
// The interface for a function that returns the result of applying a scalar
// floating-point binary operation on |a| and |b|. The type of the return value
// will be |type|. The input constants must also be of type |type|.
using FloatScalarFoldingRule = std::function<const analysis::FloatConstant*(
const analysis::Float* type, const analysis::Constant* a,
const analysis::Constant* b, analysis::ConstantManager*)>;
// Returns an std::vector containing the elements of |constant|. The type of
// |constant| must be |Vector|.
std::vector<const analysis::Constant*> GetVectorComponents(
const analysis::Constant* constant, analysis::ConstantManager* const_mgr) {
std::vector<const analysis::Constant*> components;
const analysis::VectorConstant* a = constant->AsVectorConstant();
const analysis::Vector* vector_type = constant->type()->AsVector();
assert(vector_type != nullptr);
if (a != nullptr) {
for (uint32_t i = 0; i < vector_type->element_count(); ++i) {
components.push_back(a->GetComponents()[i]);
}
} else {
const analysis::Type* element_type = vector_type->element_type();
const analysis::Constant* element_null_const =
const_mgr->GetConstant(element_type, {});
for (uint32_t i = 0; i < vector_type->element_count(); ++i) {
components.push_back(element_null_const);
}
}
return components;
}
// Returns a |ConstantFoldingRule| that folds floating point scalars using
// |scalar_rule| and vectors of floating point by applying |scalar_rule| to the
// elements of the vector. The |ConstantFoldingRule| that is returned assumes
// that |constants| contains 2 entries. If they are not |nullptr|, then their
// type is either |Float| or a |Vector| whose element type is |Float|.
ConstantFoldingRule FoldFloatingPointOp(FloatScalarFoldingRule scalar_rule) {
return [scalar_rule](ir::Instruction* inst,
const std::vector<const analysis::Constant*>& constants)
-> const analysis::Constant* {
ir::IRContext* context = inst->context();
analysis::ConstantManager* const_mgr = context->get_constant_mgr();
analysis::TypeManager* type_mgr = context->get_type_mgr();
const analysis::Type* result_type = type_mgr->GetType(inst->type_id());
const analysis::Vector* vector_type = result_type->AsVector();
const analysis::Float* float_type = nullptr;
if (!CanFoldFloatingPoint(context, inst->result_id())) {
return nullptr;
}
if (constants[0] == nullptr || constants[1] == nullptr) {
return nullptr;
}
if (vector_type != nullptr) {
std::vector<const analysis::Constant*> a_componenets;
std::vector<const analysis::Constant*> b_componenets;
std::vector<const analysis::FloatConstant*> results_componenets;
float_type = vector_type->element_type()->AsFloat();
a_componenets = GetVectorComponents(constants[0], const_mgr);
b_componenets = GetVectorComponents(constants[1], const_mgr);
// Fold each component of the vector.
for (uint32_t i = 0; i < a_componenets.size(); ++i) {
results_componenets.push_back(scalar_rule(float_type, a_componenets[i],
b_componenets[i], const_mgr));
if (results_componenets[i] == nullptr) {
return nullptr;
}
}
// Build the constant object and return it.
std::vector<uint32_t> ids;
for (const analysis::FloatConstant* member : results_componenets) {
ids.push_back(const_mgr->GetDefiningInstruction(member)->result_id());
}
return const_mgr->GetConstant(vector_type, ids);
} else {
float_type = result_type->AsFloat();
return scalar_rule(float_type, constants[0], constants[1], const_mgr);
}
};
}
// Returns the floating point value of |c|. The constant |c| must have type
// |Float|, and width |32|.
float GetFloatFromConst(const analysis::Constant* c) {
assert(c->type()->AsFloat() != nullptr &&
c->type()->AsFloat()->width() == 32);
const analysis::FloatConstant* fc = c->AsFloatConstant();
if (fc) {
return fc->GetFloatValue();
} else {
assert(c->AsNullConstant() && "c must be a float point constant.");
return 0.0f;
}
}
// Returns the double value of |c|. The constant |c| must have type
// |Float|, and width |64|.
double GetDoubleFromConst(const analysis::Constant* c) {
assert(c->type()->AsFloat() != nullptr &&
c->type()->AsFloat()->width() == 64);
const analysis::FloatConstant* fc = c->AsFloatConstant();
if (fc) {
return fc->GetDoubleValue();
} else {
assert(c->AsNullConstant() && "c must be a float point constant.");
return 0.0;
}
}
// This macro defines a |FloatScalarFoldingRule| that applies |op|. The
// operator |op| must work for both float and double, and use syntax "f1 op f2".
#define FOLD_OP(op) \
[](const analysis::Float* type, const analysis::Constant* a, \
const analysis::Constant* b, \
analysis::ConstantManager* const_mgr) -> const analysis::FloatConstant* { \
assert(type != nullptr && a != nullptr && b != nullptr); \
if (type->width() == 32) { \
float fa = GetFloatFromConst(a); \
float fb = GetFloatFromConst(b); \
spvutils::FloatProxy<float> result(fa op fb); \
std::vector<uint32_t> words = {result.data()}; \
return const_mgr->GetConstant(type, words)->AsFloatConstant(); \
} else if (type->width() == 64) { \
double fa = GetDoubleFromConst(a); \
double fb = GetDoubleFromConst(b); \
spvutils::FloatProxy<double> result(fa op fb); \
std::vector<uint32_t> words(ExtractInts(result.data())); \
return const_mgr->GetConstant(type, words)->AsFloatConstant(); \
} \
return nullptr; \
}
// Define the folding rules for subtraction, addition, multiplication, and
// division for floating point values.
ConstantFoldingRule FoldFSub() { return FoldFloatingPointOp(FOLD_OP(-)); }
ConstantFoldingRule FoldFAdd() { return FoldFloatingPointOp(FOLD_OP(+)); }
ConstantFoldingRule FoldFMul() { return FoldFloatingPointOp(FOLD_OP(*)); }
ConstantFoldingRule FoldFDiv() { return FoldFloatingPointOp(FOLD_OP(/)); }
} // namespace
spvtools::opt::ConstantFoldingRules::ConstantFoldingRules() {
// Add all folding rules to the list for the opcodes to which they apply.
// Note that the order in which rules are added to the list matters. If a rule
// applies to the instruction, the rest of the rules will not be attempted.
// Take that into consideration.
rules_[SpvOpCompositeConstruct].push_back(FoldCompositeWithConstants());
rules_[SpvOpCompositeExtract].push_back(FoldExtractWithConstants());
rules_[SpvOpFAdd].push_back(FoldFAdd());
rules_[SpvOpFDiv].push_back(FoldFDiv());
rules_[SpvOpFMul].push_back(FoldFMul());
rules_[SpvOpFSub].push_back(FoldFSub());
}
} // namespace opt
} // namespace spvtools