yuzu/src/video_core/shader/track.cpp
Lioncash ff45c39578 General: Make use of std::nullopt where applicable
Allows some implementations to avoid completely zeroing out the internal
buffer of the optional, and instead only set the validity byte within
the structure.

This also makes it consistent how we return empty optionals.
2020-09-22 17:32:33 -04:00

236 lines
9.1 KiB
C++

// Copyright 2018 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include <algorithm>
#include <utility>
#include <variant>
#include "common/common_types.h"
#include "video_core/shader/node.h"
#include "video_core/shader/node_helper.h"
#include "video_core/shader/shader_ir.h"
namespace VideoCommon::Shader {
namespace {
std::pair<Node, s64> FindOperation(const NodeBlock& code, s64 cursor,
OperationCode operation_code) {
for (; cursor >= 0; --cursor) {
Node node = code.at(cursor);
if (const auto operation = std::get_if<OperationNode>(&*node)) {
if (operation->GetCode() == operation_code) {
return {std::move(node), cursor};
}
}
if (const auto conditional = std::get_if<ConditionalNode>(&*node)) {
const auto& conditional_code = conditional->GetCode();
auto result = FindOperation(
conditional_code, static_cast<s64>(conditional_code.size() - 1), operation_code);
auto& found = result.first;
if (found) {
return {std::move(found), cursor};
}
}
}
return {};
}
std::optional<std::pair<Node, Node>> DecoupleIndirectRead(const OperationNode& operation) {
if (operation.GetCode() != OperationCode::UAdd) {
return std::nullopt;
}
Node gpr;
Node offset;
ASSERT(operation.GetOperandsCount() == 2);
for (std::size_t i = 0; i < operation.GetOperandsCount(); i++) {
Node operand = operation[i];
if (std::holds_alternative<ImmediateNode>(*operand)) {
offset = operation[i];
} else if (std::holds_alternative<GprNode>(*operand)) {
gpr = operation[i];
}
}
if (offset && gpr) {
return std::make_pair(gpr, offset);
}
return std::nullopt;
}
bool AmendNodeCv(std::size_t amend_index, Node node) {
if (const auto operation = std::get_if<OperationNode>(&*node)) {
operation->SetAmendIndex(amend_index);
return true;
}
if (const auto conditional = std::get_if<ConditionalNode>(&*node)) {
conditional->SetAmendIndex(amend_index);
return true;
}
return false;
}
} // Anonymous namespace
std::pair<Node, TrackSampler> ShaderIR::TrackBindlessSampler(Node tracked, const NodeBlock& code,
s64 cursor) {
if (const auto cbuf = std::get_if<CbufNode>(&*tracked)) {
const u32 cbuf_index = cbuf->GetIndex();
// Constant buffer found, test if it's an immediate
const auto& offset = cbuf->GetOffset();
if (const auto immediate = std::get_if<ImmediateNode>(&*offset)) {
auto track = MakeTrackSampler<BindlessSamplerNode>(cbuf_index, immediate->GetValue());
return {tracked, track};
}
if (const auto operation = std::get_if<OperationNode>(&*offset)) {
const u32 bound_buffer = registry.GetBoundBuffer();
if (bound_buffer != cbuf_index) {
return {};
}
if (const std::optional pair = DecoupleIndirectRead(*operation)) {
auto [gpr, base_offset] = *pair;
return HandleBindlessIndirectRead(*cbuf, *operation, gpr, base_offset, tracked,
code, cursor);
}
}
return {};
}
if (const auto gpr = std::get_if<GprNode>(&*tracked)) {
if (gpr->GetIndex() == Tegra::Shader::Register::ZeroIndex) {
return {};
}
// Reduce the cursor in one to avoid infinite loops when the instruction sets the same
// register that it uses as operand
const auto [source, new_cursor] = TrackRegister(gpr, code, cursor - 1);
if (!source) {
return {};
}
return TrackBindlessSampler(source, code, new_cursor);
}
if (const auto operation = std::get_if<OperationNode>(&*tracked)) {
const OperationNode& op = *operation;
const OperationCode opcode = operation->GetCode();
if (opcode == OperationCode::IBitwiseOr || opcode == OperationCode::UBitwiseOr) {
ASSERT(op.GetOperandsCount() == 2);
auto [node_a, index_a, offset_a] = TrackCbuf(op[0], code, cursor);
auto [node_b, index_b, offset_b] = TrackCbuf(op[1], code, cursor);
if (node_a && node_b) {
auto track = MakeTrackSampler<SeparateSamplerNode>(std::pair{index_a, index_b},
std::pair{offset_a, offset_b});
return {tracked, std::move(track)};
}
}
std::size_t i = op.GetOperandsCount();
while (i--) {
if (auto found = TrackBindlessSampler(op[i - 1], code, cursor); std::get<0>(found)) {
// Constant buffer found in operand.
return found;
}
}
return {};
}
if (const auto conditional = std::get_if<ConditionalNode>(&*tracked)) {
const auto& conditional_code = conditional->GetCode();
return TrackBindlessSampler(tracked, conditional_code,
static_cast<s64>(conditional_code.size()));
}
return {};
}
std::pair<Node, TrackSampler> ShaderIR::HandleBindlessIndirectRead(
const CbufNode& cbuf, const OperationNode& operation, Node gpr, Node base_offset, Node tracked,
const NodeBlock& code, s64 cursor) {
const auto offset_imm = std::get<ImmediateNode>(*base_offset);
const auto& gpu_driver = registry.AccessGuestDriverProfile();
const u32 bindless_cv = NewCustomVariable();
const u32 texture_handler_size = gpu_driver.GetTextureHandlerSize();
Node op = Operation(OperationCode::UDiv, gpr, Immediate(texture_handler_size));
Node cv_node = GetCustomVariable(bindless_cv);
Node amend_op = Operation(OperationCode::Assign, std::move(cv_node), std::move(op));
const std::size_t amend_index = DeclareAmend(std::move(amend_op));
AmendNodeCv(amend_index, code[cursor]);
// TODO: Implement bindless index custom variable
auto track =
MakeTrackSampler<ArraySamplerNode>(cbuf.GetIndex(), offset_imm.GetValue(), bindless_cv);
return {tracked, track};
}
std::tuple<Node, u32, u32> ShaderIR::TrackCbuf(Node tracked, const NodeBlock& code,
s64 cursor) const {
if (const auto cbuf = std::get_if<CbufNode>(&*tracked)) {
// Constant buffer found, test if it's an immediate
const auto& offset = cbuf->GetOffset();
if (const auto immediate = std::get_if<ImmediateNode>(&*offset)) {
return {tracked, cbuf->GetIndex(), immediate->GetValue()};
}
return {};
}
if (const auto gpr = std::get_if<GprNode>(&*tracked)) {
if (gpr->GetIndex() == Tegra::Shader::Register::ZeroIndex) {
return {};
}
// Reduce the cursor in one to avoid infinite loops when the instruction sets the same
// register that it uses as operand
const auto [source, new_cursor] = TrackRegister(gpr, code, cursor - 1);
if (!source) {
return {};
}
return TrackCbuf(source, code, new_cursor);
}
if (const auto operation = std::get_if<OperationNode>(&*tracked)) {
for (std::size_t i = operation->GetOperandsCount(); i > 0; --i) {
if (auto found = TrackCbuf((*operation)[i - 1], code, cursor); std::get<0>(found)) {
// Cbuf found in operand.
return found;
}
}
return {};
}
if (const auto conditional = std::get_if<ConditionalNode>(&*tracked)) {
const auto& conditional_code = conditional->GetCode();
return TrackCbuf(tracked, conditional_code, static_cast<s64>(conditional_code.size()));
}
return {};
}
std::optional<u32> ShaderIR::TrackImmediate(Node tracked, const NodeBlock& code, s64 cursor) const {
// Reduce the cursor in one to avoid infinite loops when the instruction sets the same register
// that it uses as operand
const auto result = TrackRegister(&std::get<GprNode>(*tracked), code, cursor - 1);
const auto& found = result.first;
if (!found) {
return std::nullopt;
}
if (const auto immediate = std::get_if<ImmediateNode>(&*found)) {
return immediate->GetValue();
}
return std::nullopt;
}
std::pair<Node, s64> ShaderIR::TrackRegister(const GprNode* tracked, const NodeBlock& code,
s64 cursor) const {
for (; cursor >= 0; --cursor) {
const auto [found_node, new_cursor] = FindOperation(code, cursor, OperationCode::Assign);
if (!found_node) {
return {};
}
const auto operation = std::get_if<OperationNode>(&*found_node);
ASSERT(operation);
const auto& target = (*operation)[0];
if (const auto gpr_target = std::get_if<GprNode>(&*target)) {
if (gpr_target->GetIndex() == tracked->GetIndex()) {
return {(*operation)[1], new_cursor};
}
}
}
return {};
}
} // namespace VideoCommon::Shader