/* * Copyright (C) 2014 Patrick Mours * SPDX-License-Identifier: BSD-3-Clause */ #include "effect_parser.hpp" #include "effect_codegen.hpp" #include #include // memcmp #include // std::find_if, std::max #include // Use the C++ variant of the SPIR-V headers #include namespace spv { #include } using namespace reshadefx; /// /// A single instruction in a SPIR-V module /// struct spirv_instruction { spv::Op op; spv::Id type; spv::Id result; std::vector operands; explicit spirv_instruction(spv::Op op = spv::OpNop) : op(op), type(0), result(0) {} spirv_instruction(spv::Op op, spv::Id result) : op(op), type(result), result(0) {} spirv_instruction(spv::Op op, spv::Id type, spv::Id result) : op(op), type(type), result(result) {} /// /// Add a single operand to the instruction. /// spirv_instruction &add(spv::Id operand) { operands.push_back(operand); return *this; } /// /// Add a range of operands to the instruction. /// template spirv_instruction &add(It begin, It end) { operands.insert(operands.end(), begin, end); return *this; } /// /// Add a null-terminated literal UTF-8 string to the instruction. /// spirv_instruction &add_string(const char *string) { uint32_t word; do { word = 0; for (uint32_t i = 0; i < 4 && *string; ++i) reinterpret_cast(&word)[i] = *string++; add(word); } while (*string || (word & 0xFF000000)); return *this; } /// /// Write this instruction to a SPIR-V module. /// /// The output stream to append this instruction to. void write(std::vector &output) const { // See https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html // 0 | Opcode: The 16 high-order bits are the WordCount of the instruction. The 16 low-order bits are the opcode enumerant. // 1 | Optional instruction type // . | Optional instruction Result // . | Operand 1 (if needed) // . | Operand 2 (if needed) // ... | ... // WordCount - 1 | Operand N (N is determined by WordCount minus the 1 to 3 words used for the opcode, instruction type , and instruction Result ). const uint32_t num_words = 1 + (type != 0) + (result != 0) + static_cast(operands.size()); output.push_back((num_words << spv::WordCountShift) | op); // Optional instruction type ID if (type != 0) output.push_back(type); // Optional instruction result ID if (result != 0) output.push_back(result); // Write out the operands output.insert(output.end(), operands.begin(), operands.end()); } }; /// /// A list of instructions forming a basic block in the SPIR-V module /// struct spirv_basic_block { std::vector instructions; /// /// Append another basic block the end of this one. /// void append(const spirv_basic_block &block) { instructions.insert(instructions.end(), block.instructions.begin(), block.instructions.end()); } }; class codegen_spirv final : public codegen { public: codegen_spirv(bool vulkan_semantics, bool debug_info, bool uniforms_to_spec_constants, bool enable_16bit_types, bool flip_vert_y) : _debug_info(debug_info), _vulkan_semantics(vulkan_semantics), _uniforms_to_spec_constants(uniforms_to_spec_constants), _enable_16bit_types(enable_16bit_types), _flip_vert_y(flip_vert_y) { _glsl_ext = make_id(); } private: struct type_lookup { reshadefx::type type; bool is_ptr; uint32_t array_stride; std::pair storage; friend bool operator==(const type_lookup &lhs, const type_lookup &rhs) { return lhs.type == rhs.type && lhs.is_ptr == rhs.is_ptr && lhs.array_stride == rhs.array_stride && lhs.storage == rhs.storage; } }; struct function_blocks { spirv_basic_block declaration; spirv_basic_block variables; spirv_basic_block definition; type return_type; std::vector param_types; bool is_entry_point = false; friend bool operator==(const function_blocks &lhs, const function_blocks &rhs) { if (lhs.param_types.size() != rhs.param_types.size()) return false; for (size_t i = 0; i < lhs.param_types.size(); ++i) if (!(lhs.param_types[i] == rhs.param_types[i])) return false; return lhs.return_type == rhs.return_type; } }; spirv_basic_block _entries; spirv_basic_block _execution_modes; spirv_basic_block _debug_a; spirv_basic_block _debug_b; spirv_basic_block _annotations; spirv_basic_block _types_and_constants; spirv_basic_block _variables; std::unordered_set _spec_constants; std::unordered_set _capabilities; std::vector> _type_lookup; std::vector> _constant_lookup; std::vector> _function_type_lookup; std::unordered_map _string_lookup; std::unordered_map> _storage_lookup; std::unordered_map _semantic_to_location; std::vector _functions_blocks; std::unordered_map _block_data; spirv_basic_block *_current_block_data = nullptr; bool _debug_info = false; bool _vulkan_semantics = false; bool _uniforms_to_spec_constants = false; bool _enable_16bit_types = false; bool _flip_vert_y = false; id _glsl_ext = 0; id _global_ubo_type = 0; id _global_ubo_variable = 0; std::vector _global_ubo_types; function_blocks *_current_function = nullptr; inline void add_location(const location &loc, spirv_basic_block &block) { if (loc.source.empty() || !_debug_info) return; spv::Id file; if (const auto it = _string_lookup.find(loc.source); it != _string_lookup.end()) file = it->second; else { add_instruction(spv::OpString, 0, _debug_a, file) .add_string(loc.source.c_str()); _string_lookup.emplace(loc.source, file); } // https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#OpLine add_instruction_without_result(spv::OpLine, block) .add(file) .add(loc.line) .add(loc.column); } inline spirv_instruction &add_instruction(spv::Op op, spv::Id type = 0) { assert(is_in_function() && is_in_block()); return add_instruction(op, type, *_current_block_data); } inline spirv_instruction &add_instruction(spv::Op op, spv::Id type, spirv_basic_block &block) { spirv_instruction &instruction = add_instruction_without_result(op, block); instruction.type = type; instruction.result = make_id(); return instruction; } inline spirv_instruction &add_instruction(spv::Op op, spv::Id type, spirv_basic_block &block, spv::Id &result) { spirv_instruction &instruction = add_instruction_without_result(op, block); instruction.type = type; instruction.result = result = make_id(); return instruction; } inline spirv_instruction &add_instruction_without_result(spv::Op op) { assert(is_in_function() && is_in_block()); return add_instruction_without_result(op, *_current_block_data); } inline spirv_instruction &add_instruction_without_result(spv::Op op, spirv_basic_block &block) { return block.instructions.emplace_back(op); } void write_result(module &module) override { // First initialize the UBO type now that all member types are known if (_global_ubo_type != 0) { spirv_instruction &type_inst = add_instruction_without_result(spv::OpTypeStruct, _types_and_constants); type_inst.add(_global_ubo_types.begin(), _global_ubo_types.end()); type_inst.result = _global_ubo_type; spirv_instruction &variable_inst = add_instruction_without_result(spv::OpVariable, _variables); variable_inst.add(spv::StorageClassUniform); variable_inst.type = convert_type({ type::t_struct, 0, 0, type::q_uniform, 0, _global_ubo_type }, true, spv::StorageClassUniform); variable_inst.result = _global_ubo_variable; add_name(variable_inst.result, "$Globals"); } module = std::move(_module); std::vector spirv; // Write SPIRV header info spirv.push_back(spv::MagicNumber); spirv.push_back(0x10300); // Force SPIR-V 1.3 spirv.push_back(0u); // Generator magic number, see https://www.khronos.org/registry/spir-v/api/spir-v.xml spirv.push_back(_next_id); // Maximum ID spirv.push_back(0u); // Reserved for instruction schema // All capabilities spirv_instruction(spv::OpCapability) .add(spv::CapabilityShader) // Implicitly declares the Matrix capability too .write(spirv); for (spv::Capability capability : _capabilities) spirv_instruction(spv::OpCapability) .add(capability) .write(spirv); // Optional extension instructions spirv_instruction(spv::OpExtInstImport, _glsl_ext) .add_string("GLSL.std.450") // Import GLSL extension .write(spirv); // Single required memory model instruction spirv_instruction(spv::OpMemoryModel) .add(spv::AddressingModelLogical) .add(spv::MemoryModelGLSL450) .write(spirv); // All entry point declarations for (const auto &node : _entries.instructions) node.write(spirv); // All execution mode declarations for (const auto &node : _execution_modes.instructions) node.write(spirv); spirv_instruction(spv::OpSource) .add(spv::SourceLanguageUnknown) // ReShade FX is not a reserved token at the moment .add(0) // Language version, TODO: Maybe fill in ReShade version here? .write(spirv); if (_debug_info) { // All debug instructions for (const auto &node : _debug_a.instructions) node.write(spirv); for (const auto &node : _debug_b.instructions) node.write(spirv); } // All annotation instructions for (const auto &node : _annotations.instructions) node.write(spirv); // All type declarations for (const auto &node : _types_and_constants.instructions) node.write(spirv); for (const auto &node : _variables.instructions) node.write(spirv); // All function definitions for (const auto &function : _functions_blocks) { if (function.definition.instructions.empty()) continue; for (const auto &node : function.declaration.instructions) node.write(spirv); // Grab first label and move it in front of variable declarations function.definition.instructions.front().write(spirv); assert(function.definition.instructions.front().op == spv::OpLabel); for (const auto &node : function.variables.instructions) node.write(spirv); for (auto it = function.definition.instructions.begin() + 1; it != function.definition.instructions.end(); ++it) it->write(spirv); } module.code.assign(reinterpret_cast(spirv.data()), reinterpret_cast(spirv.data() + spirv.size())); } spv::Id convert_type(type info, bool is_ptr = false, spv::StorageClass storage = spv::StorageClassFunction, spv::ImageFormat format = spv::ImageFormatUnknown, uint32_t array_stride = 0) { assert(array_stride == 0 || info.is_array()); // The storage class is only relevant for pointers, so ignore it for other types during lookup if (is_ptr == false) storage = spv::StorageClassFunction; // There cannot be sampler variables that are local to a function, so always assume uniform storage for them if (info.is_object()) storage = spv::StorageClassUniformConstant; else assert(format == spv::ImageFormatUnknown); if (info.is_sampler() || info.is_storage()) info.rows = info.cols = 1; // Fall back to 32-bit types and use relaxed precision decoration instead if 16-bit types are not enabled if (!_enable_16bit_types && info.is_numeric() && info.precision() < 32) info.base = static_cast(info.base + 1); // min16int -> int, min16uint -> uint, min16float -> float const type_lookup lookup { info, is_ptr, array_stride, { storage, format } }; if (const auto it = std::find_if(_type_lookup.begin(), _type_lookup.end(), [&lookup](const auto &lookup_it) { return lookup_it.first == lookup; }); it != _type_lookup.end()) return it->second; spv::Id type, elem_type; if (is_ptr) { elem_type = convert_type(info, false, storage, format, array_stride); add_instruction(spv::OpTypePointer, 0, _types_and_constants, type) .add(storage) .add(elem_type); } else if (info.is_array()) { auto elem_info = info; elem_info.array_length = 0; // Make sure we don't get any dynamic arrays here assert(info.array_length > 0); elem_type = convert_type(elem_info, false, storage, format); const spv::Id array_length = emit_constant(info.array_length); add_instruction(spv::OpTypeArray, 0, _types_and_constants, type) .add(elem_type) .add(array_length); if (array_stride != 0) add_decoration(type, spv::DecorationArrayStride, { array_stride }); } else if (info.is_matrix()) { // Convert MxN matrix to a SPIR-V matrix with M vectors with N elements auto elem_info = info; elem_info.rows = info.cols; elem_info.cols = 1; elem_type = convert_type(elem_info, false, storage, format); // Matrix types with just one row are interpreted as if they were a vector type if (info.rows == 1) return elem_type; add_instruction(spv::OpTypeMatrix, 0, _types_and_constants, type) .add(elem_type) .add(info.rows); } else if (info.is_vector()) { auto elem_info = info; elem_info.rows = 1; elem_info.cols = 1; elem_type = convert_type(elem_info, false, storage, format); add_instruction(spv::OpTypeVector, 0, _types_and_constants, type) .add(elem_type) .add(info.rows); } else { switch (info.base) { case type::t_void: assert(info.rows == 0 && info.cols == 0); add_instruction(spv::OpTypeVoid, 0, _types_and_constants, type); break; case type::t_bool: assert(info.rows == 1 && info.cols == 1); add_instruction(spv::OpTypeBool, 0, _types_and_constants, type); break; case type::t_min16int: assert(_enable_16bit_types && info.rows == 1 && info.cols == 1); add_capability(spv::CapabilityInt16); if (storage == spv::StorageClassInput || storage == spv::StorageClassOutput) add_capability(spv::CapabilityStorageInputOutput16); add_instruction(spv::OpTypeInt, 0, _types_and_constants, type) .add(16) // Width .add(1); // Signedness break; case type::t_int: assert(info.rows == 1 && info.cols == 1); add_instruction(spv::OpTypeInt, 0, _types_and_constants, type) .add(32) // Width .add(1); // Signedness break; case type::t_min16uint: assert(_enable_16bit_types && info.rows == 1 && info.cols == 1); add_capability(spv::CapabilityInt16); if (storage == spv::StorageClassInput || storage == spv::StorageClassOutput) add_capability(spv::CapabilityStorageInputOutput16); add_instruction(spv::OpTypeInt, 0, _types_and_constants, type) .add(16) // Width .add(0); // Signedness break; case type::t_uint: assert(info.rows == 1 && info.cols == 1); add_instruction(spv::OpTypeInt, 0, _types_and_constants, type) .add(32) // Width .add(0); // Signedness break; case type::t_min16float: assert(_enable_16bit_types && info.rows == 1 && info.cols == 1); add_capability(spv::CapabilityFloat16); if (storage == spv::StorageClassInput || storage == spv::StorageClassOutput) add_capability(spv::CapabilityStorageInputOutput16); add_instruction(spv::OpTypeFloat, 0, _types_and_constants, type) .add(16); // Width break; case type::t_float: assert(info.rows == 1 && info.cols == 1); add_instruction(spv::OpTypeFloat, 0, _types_and_constants, type) .add(32); // Width break; case type::t_struct: assert(info.rows == 0 && info.cols == 0 && info.definition != 0); type = info.definition; break; case type::t_sampler1d_int: case type::t_sampler1d_uint: case type::t_sampler1d_float: add_capability(spv::CapabilitySampled1D); [[fallthrough]]; case type::t_sampler2d_int: case type::t_sampler2d_uint: case type::t_sampler2d_float: case type::t_sampler3d_int: case type::t_sampler3d_uint: case type::t_sampler3d_float: elem_type = convert_image_type(info, format); add_instruction(spv::OpTypeSampledImage, 0, _types_and_constants, type) .add(elem_type); break; case type::t_storage1d_int: case type::t_storage1d_uint: case type::t_storage1d_float: add_capability(spv::CapabilityImage1D); [[fallthrough]]; case type::t_storage2d_int: case type::t_storage2d_uint: case type::t_storage2d_float: case type::t_storage3d_int: case type::t_storage3d_uint: case type::t_storage3d_float: // No format specified for the storage image if (format == spv::ImageFormatUnknown) add_capability(spv::CapabilityStorageImageWriteWithoutFormat); return convert_image_type(info, format); default: return assert(false), 0; } } _type_lookup.push_back({ lookup, type }); return type; } spv::Id convert_type(const function_blocks &info) { if (const auto it = std::find_if(_function_type_lookup.begin(), _function_type_lookup.end(), [&lookup = info](const auto &lookup_it) { return lookup_it.first == lookup; }); it != _function_type_lookup.end()) return it->second; auto return_type = convert_type(info.return_type); assert(return_type != 0); std::vector param_type_ids; param_type_ids.reserve(info.param_types.size()); for (const type ¶m_type : info.param_types) param_type_ids.push_back(convert_type(param_type, true)); spirv_instruction &inst = add_instruction(spv::OpTypeFunction, 0, _types_and_constants); inst.add(return_type); inst.add(param_type_ids.begin(), param_type_ids.end()); _function_type_lookup.push_back({ info, inst.result });; return inst.result; } spv::Id convert_image_type(type info, spv::ImageFormat format = spv::ImageFormatUnknown) { type_lookup lookup { info, false, 0u, { spv::StorageClassUniformConstant, format } }; auto elem_info = info; elem_info.rows = 1; elem_info.cols = 1; if (!info.is_numeric()) { if ((info.is_integral() && info.is_signed()) || (format >= spv::ImageFormatRgba32i && format <= spv::ImageFormatR8i)) elem_info.base = type::t_int; else if ((info.is_integral() && info.is_unsigned()) || (format >= spv::ImageFormatRgba32ui && format <= spv::ImageFormatR8ui)) elem_info.base = type::t_uint; else elem_info.base = type::t_float; } if (!info.is_storage()) { lookup.type = elem_info; lookup.type.base = static_cast(type::t_texture1d + info.texture_dimension() - 1); lookup.type.definition = static_cast(elem_info.base); } if (const auto it = std::find_if(_type_lookup.begin(), _type_lookup.end(), [&lookup](const auto &lookup_it) { return lookup_it.first == lookup; }); it != _type_lookup.end()) return it->second; spv::Id type, elem_type = convert_type(elem_info, false, spv::StorageClassUniformConstant); add_instruction(spv::OpTypeImage, 0, _types_and_constants, type) .add(elem_type) // Sampled Type (always a scalar type) .add(spv::Dim1D + info.texture_dimension() - 1) .add(0) // Not a depth image .add(0) // Not an array .add(0) // Not multi-sampled .add(info.is_storage() ? 2 : 1) // Used with a sampler or as storage .add(format); _type_lookup.push_back({ lookup, type }); return type; } uint32_t semantic_to_location(const std::string &semantic, uint32_t max_array_length = 1) { if (semantic.compare(0, 5, "COLOR") == 0) return std::strtoul(semantic.c_str() + 5, nullptr, 10); if (semantic.compare(0, 9, "SV_TARGET") == 0) return std::strtoul(semantic.c_str() + 9, nullptr, 10); if (const auto it = _semantic_to_location.find(semantic); it != _semantic_to_location.end()) return it->second; // Extract the semantic index from the semantic name (e.g. 2 for "TEXCOORD2") size_t digit_index = semantic.size() - 1; while (digit_index != 0 && semantic[digit_index] >= '0' && semantic[digit_index] <= '9') digit_index--; digit_index++; const uint32_t semantic_digit = std::strtoul(semantic.c_str() + digit_index, nullptr, 10); const std::string semantic_base = semantic.substr(0, digit_index); uint32_t location = static_cast(_semantic_to_location.size()); // Now create adjoining location indices for all possible semantic indices belonging to this semantic name for (uint32_t a = 0; a < semantic_digit + max_array_length; ++a) { const auto insert = _semantic_to_location.emplace(semantic_base + std::to_string(a), location + a); if (!insert.second) { assert(a == 0 || (insert.first->second - a) == location); // Semantic was already created with a different location index, so need to remap to that location = insert.first->second - a; } } return location + semantic_digit; } const spv::BuiltIn semantic_to_builtin(const std::string &semantic, shader_type stype) const { if (semantic == "SV_POSITION") return stype == shader_type::ps ? spv::BuiltInFragCoord : spv::BuiltInPosition; if (semantic == "SV_POINTSIZE") return spv::BuiltInPointSize; if (semantic == "SV_DEPTH") return spv::BuiltInFragDepth; if (semantic == "SV_VERTEXID") return _vulkan_semantics ? spv::BuiltInVertexIndex : spv::BuiltInVertexId; if (semantic == "SV_ISFRONTFACE") return spv::BuiltInFrontFacing; if (semantic == "SV_GROUPID") return spv::BuiltInWorkgroupId; if (semantic == "SV_GROUPINDEX") return spv::BuiltInLocalInvocationIndex; if (semantic == "SV_GROUPTHREADID") return spv::BuiltInLocalInvocationId; if (semantic == "SV_DISPATCHTHREADID") return spv::BuiltInGlobalInvocationId; return spv::BuiltInMax; } const spv::ImageFormat format_to_image_format(texture_format format) { switch (format) { default: assert(false); [[fallthrough]]; case texture_format::unknown: return spv::ImageFormatUnknown; case texture_format::r8: add_capability(spv::CapabilityStorageImageExtendedFormats); return spv::ImageFormatR8; case texture_format::r16: add_capability(spv::CapabilityStorageImageExtendedFormats); return spv::ImageFormatR16; case texture_format::r16f: add_capability(spv::CapabilityStorageImageExtendedFormats); return spv::ImageFormatR16f; case texture_format::r32i: return spv::ImageFormatR32i; case texture_format::r32u: return spv::ImageFormatR32ui; case texture_format::r32f: return spv::ImageFormatR32f; case texture_format::rg8: add_capability(spv::CapabilityStorageImageExtendedFormats); return spv::ImageFormatRg8; case texture_format::rg16: add_capability(spv::CapabilityStorageImageExtendedFormats); return spv::ImageFormatRg16; case texture_format::rg16f: add_capability(spv::CapabilityStorageImageExtendedFormats); return spv::ImageFormatRg16f; case texture_format::rg32f: add_capability(spv::CapabilityStorageImageExtendedFormats); return spv::ImageFormatRg32f; case texture_format::rgba8: return spv::ImageFormatRgba8; case texture_format::rgba16: add_capability(spv::CapabilityStorageImageExtendedFormats); return spv::ImageFormatRgba16; case texture_format::rgba16f: return spv::ImageFormatRgba16f; case texture_format::rgba32f: return spv::ImageFormatRgba32f; case texture_format::rgb10a2: add_capability(spv::CapabilityStorageImageExtendedFormats); return spv::ImageFormatRgb10A2; } } inline void add_name(id id, const char *name) { if (!_debug_info) return; assert(name != nullptr); // https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#OpName add_instruction_without_result(spv::OpName, _debug_b) .add(id) .add_string(name); } inline void add_builtin(id id, spv::BuiltIn builtin) { add_instruction_without_result(spv::OpDecorate, _annotations) .add(id) .add(spv::DecorationBuiltIn) .add(builtin); } inline void add_decoration(id id, spv::Decoration decoration, std::initializer_list values = {}) { // https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#OpDecorate add_instruction_without_result(spv::OpDecorate, _annotations) .add(id) .add(decoration) .add(values.begin(), values.end()); } inline void add_member_name(id id, uint32_t member_index, const char *name) { if (!_debug_info) return; assert(name != nullptr); // https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#OpMemberName add_instruction_without_result(spv::OpMemberName, _debug_b) .add(id) .add(member_index) .add_string(name); } inline void add_member_builtin(id id, uint32_t member_index, spv::BuiltIn builtin) { add_instruction_without_result(spv::OpMemberDecorate, _annotations) .add(id) .add(member_index) .add(spv::DecorationBuiltIn) .add(builtin); } inline void add_member_decoration(id id, uint32_t member_index, spv::Decoration decoration, std::initializer_list values = {}) { // https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#OpMemberDecorate add_instruction_without_result(spv::OpMemberDecorate, _annotations) .add(id) .add(member_index) .add(decoration) .add(values.begin(), values.end()); } inline void add_capability(spv::Capability capability) { _capabilities.insert(capability); } id define_struct(const location &loc, struct_info &info) override { // First define all member types to make sure they are declared before the struct type references them std::vector member_types; member_types.reserve(info.member_list.size()); for (const struct_member_info &member : info.member_list) member_types.push_back(convert_type(member.type)); // Afterwards define the actual struct type add_location(loc, _types_and_constants); add_instruction(spv::OpTypeStruct, 0, _types_and_constants, info.definition) .add(member_types.begin(), member_types.end()); if (!info.unique_name.empty()) add_name(info.definition, info.unique_name.c_str()); for (uint32_t index = 0; index < info.member_list.size(); ++index) { const struct_member_info &member = info.member_list[index]; add_member_name(info.definition, index, member.name.c_str()); if (!_enable_16bit_types && member.type.is_numeric() && member.type.precision() < 32) add_member_decoration(info.definition, index, spv::DecorationRelaxedPrecision); } _structs.push_back(info); return info.definition; } id define_texture(const location &, texture_info &info) override { info.id = make_id(); // Need to create an unique ID here too, so that the symbol lookup for textures works info.binding = ~0u; _module.textures.push_back(info); return info.id; } id define_sampler(const location &loc, const texture_info &, sampler_info &info) override { info.id = define_variable(loc, info.type, info.unique_name.c_str(), spv::StorageClassUniformConstant); info.binding = _module.num_sampler_bindings++; info.texture_binding = ~0u; add_decoration(info.id, spv::DecorationBinding, { info.binding }); add_decoration(info.id, spv::DecorationDescriptorSet, { 1 }); _module.samplers.push_back(info); return info.id; } id define_storage(const location &loc, const texture_info &tex_info, storage_info &info) override { info.id = define_variable(loc, info.type, info.unique_name.c_str(), spv::StorageClassUniformConstant, format_to_image_format(tex_info.format)); info.binding = _module.num_storage_bindings++; add_decoration(info.id, spv::DecorationBinding, { info.binding }); add_decoration(info.id, spv::DecorationDescriptorSet, { 2 }); _module.storages.push_back(info); return info.id; } id define_uniform(const location &, uniform_info &info) override { if (_uniforms_to_spec_constants && info.has_initializer_value) { const id res = emit_constant(info.type, info.initializer_value, true); add_name(res, info.name.c_str()); const auto add_spec_constant = [this](const spirv_instruction &inst, const uniform_info &info, const constant &initializer_value, size_t initializer_offset) { assert(inst.op == spv::OpSpecConstant || inst.op == spv::OpSpecConstantTrue || inst.op == spv::OpSpecConstantFalse); const uint32_t spec_id = static_cast(_module.spec_constants.size()); add_decoration(inst.result, spv::DecorationSpecId, { spec_id }); uniform_info scalar_info = info; scalar_info.type.rows = 1; scalar_info.type.cols = 1; scalar_info.size = 4; scalar_info.offset = static_cast(initializer_offset); scalar_info.initializer_value = {}; scalar_info.initializer_value.as_uint[0] = initializer_value.as_uint[initializer_offset]; _module.spec_constants.push_back(scalar_info); }; const spirv_instruction &base_inst = _types_and_constants.instructions.back(); assert(base_inst.result == res); // External specialization constants need to be scalars if (info.type.is_scalar()) { add_spec_constant(base_inst, info, info.initializer_value, 0); } else { assert(base_inst.op == spv::OpSpecConstantComposite); // Add each individual scalar component of the constant as a separate external specialization constant for (size_t i = 0; i < (info.type.is_array() ? base_inst.operands.size() : 1); ++i) { constant initializer_value = info.initializer_value; spirv_instruction elem_inst = base_inst; if (info.type.is_array()) { elem_inst = *std::find_if(_types_and_constants.instructions.rbegin(), _types_and_constants.instructions.rend(), [elem = base_inst.operands[i]](const auto &it) { return it.result == elem; }); assert(initializer_value.array_data.size() == base_inst.operands.size()); initializer_value = initializer_value.array_data[i]; } for (size_t row = 0; row < elem_inst.operands.size(); ++row) { const spirv_instruction &row_inst = *std::find_if(_types_and_constants.instructions.rbegin(), _types_and_constants.instructions.rend(), [elem = elem_inst.operands[row]](const auto &it) { return it.result == elem; }); if (row_inst.op != spv::OpSpecConstantComposite) { add_spec_constant(row_inst, info, initializer_value, row); continue; } for (size_t col = 0; col < row_inst.operands.size(); ++col) { const spirv_instruction &col_inst = *std::find_if(_types_and_constants.instructions.rbegin(), _types_and_constants.instructions.rend(), [elem = row_inst.operands[col]](const auto &it) { return it.result == elem; }); add_spec_constant(col_inst, info, initializer_value, row * info.type.cols + col); } } } } return res; } else { // Create global uniform buffer variable on demand if (_global_ubo_type == 0) { _global_ubo_type = make_id(); add_decoration(_global_ubo_type, spv::DecorationBlock); } if (_global_ubo_variable == 0) { _global_ubo_variable = make_id(); add_decoration(_global_ubo_variable, spv::DecorationDescriptorSet, { 0 }); add_decoration(_global_ubo_variable, spv::DecorationBinding, { 0 }); } uint32_t alignment = (info.type.rows == 3 ? 4 : info.type.rows) * 4; info.size = info.type.rows * 4; uint32_t array_stride = 16; const uint32_t matrix_stride = 16; if (info.type.is_matrix()) { alignment = matrix_stride; info.size = info.type.rows * matrix_stride; } if (info.type.is_array()) { alignment = array_stride; array_stride = align_up(info.size, array_stride); // Uniform block rules do not permit anything in the padding of an array info.size = array_stride * info.type.array_length; } info.offset = _module.total_uniform_size; info.offset = align_up(info.offset, alignment); _module.total_uniform_size = info.offset + info.size; type ubo_type = info.type; // Convert boolean uniform variables to integer type so that they have a defined size if (info.type.is_boolean()) ubo_type.base = type::t_uint; const uint32_t member_index = static_cast(_global_ubo_types.size()); // Composite objects in the uniform storage class must be explicitly laid out, which includes array types requiring a stride decoration _global_ubo_types.push_back( convert_type(ubo_type, false, spv::StorageClassUniform, spv::ImageFormatUnknown, info.type.is_array() ? array_stride : 0u)); add_member_name(_global_ubo_type, member_index, info.name.c_str()); add_member_decoration(_global_ubo_type, member_index, spv::DecorationOffset, { info.offset }); if (info.type.is_matrix()) { // Read matrices in column major layout, even though they are actually row major, to avoid transposing them on every access (since SPIR-V uses column matrices) // TODO: This technically only works with square matrices add_member_decoration(_global_ubo_type, member_index, spv::DecorationColMajor); add_member_decoration(_global_ubo_type, member_index, spv::DecorationMatrixStride, { matrix_stride }); } _module.uniforms.push_back(info); return 0xF0000000 | member_index; } } id define_variable(const location &loc, const type &type, std::string name, bool global, id initializer_value) override { spv::StorageClass storage = spv::StorageClassFunction; if (type.has(type::q_groupshared)) storage = spv::StorageClassWorkgroup; else if (global) storage = spv::StorageClassPrivate; return define_variable(loc, type, name.c_str(), storage, spv::ImageFormatUnknown, initializer_value); } id define_variable(const location &loc, const type &type, const char *name, spv::StorageClass storage, spv::ImageFormat format = spv::ImageFormatUnknown, spv::Id initializer_value = 0) { assert(storage != spv::StorageClassFunction || _current_function != nullptr); spirv_basic_block &block = (storage != spv::StorageClassFunction) ? _variables : _current_function->variables; add_location(loc, block); spv::Id res; // https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#OpVariable spirv_instruction &inst = add_instruction(spv::OpVariable, convert_type(type, true, storage, format), block, res) .add(storage); if (initializer_value != 0) { if (storage != spv::StorageClassFunction || _current_function->is_entry_point) { // The initializer for variables must be a constant inst.add(initializer_value); } else { // Only use the variable initializer on global variables, since local variables for e.g. "for" statements need to be assigned in their respective scope and not their declaration expression variable; variable.reset_to_lvalue(loc, res, type); emit_store(variable, initializer_value); } } if (name != nullptr && *name != '\0') add_name(res, name); if (!_enable_16bit_types && type.is_numeric() && type.precision() < 32) add_decoration(res, spv::DecorationRelaxedPrecision); _storage_lookup[res] = { storage, format }; return res; } id define_function(const location &loc, function_info &info) override { assert(!is_in_function()); auto &function = _functions_blocks.emplace_back(); function.return_type = info.return_type; _current_function = &function; for (auto ¶m : info.parameter_list) function.param_types.push_back(param.type); add_location(loc, function.declaration); // https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#OpFunction add_instruction(spv::OpFunction, convert_type(info.return_type), function.declaration, info.definition) .add(spv::FunctionControlMaskNone) .add(convert_type(function)); if (!info.name.empty()) add_name(info.definition, info.name.c_str()); for (auto ¶m : info.parameter_list) { add_location(param.location, function.declaration); param.definition = add_instruction(spv::OpFunctionParameter, convert_type(param.type, true), function.declaration).result; add_name(param.definition, param.name.c_str()); } _functions.push_back(std::make_unique(info)); return info.definition; } void define_entry_point(function_info &func, shader_type stype, int num_threads[3]) override { // Modify entry point name so each thread configuration is made separate if (stype == shader_type::cs) func.unique_name = 'E' + func.unique_name + '_' + std::to_string(num_threads[0]) + '_' + std::to_string(num_threads[1]) + '_' + std::to_string(num_threads[2]); if (const auto it = std::find_if(_module.entry_points.begin(), _module.entry_points.end(), [&func](const auto &ep) { return ep.name == func.unique_name; }); it != _module.entry_points.end()) return; _module.entry_points.push_back({ func.unique_name, stype }); spv::Id position_variable = 0, point_size_variable = 0; std::vector inputs_and_outputs; std::vector call_params; // Generate the glue entry point function function_info entry_point; entry_point.return_type = { type::t_void }; define_function({}, entry_point); enter_block(create_block()); _current_function->is_entry_point = true; const auto create_varying_param = [this, &call_params](const struct_member_info ¶m) { // Initialize all output variables with zero const spv::Id variable = define_variable({}, param.type, nullptr, spv::StorageClassFunction, spv::ImageFormatUnknown, emit_constant(param.type, 0u)); expression &call_param = call_params.emplace_back(); call_param.reset_to_lvalue({}, variable, param.type); return variable; }; const auto create_varying_variable = [this, &inputs_and_outputs, &position_variable, &point_size_variable, stype](const type ¶m_type, const std::string &semantic, spv::StorageClass storage, int a = 0) { const spv::Id variable = define_variable({}, param_type, nullptr, storage); if (const spv::BuiltIn builtin = semantic_to_builtin(semantic, stype); builtin != spv::BuiltInMax) { assert(a == 0); // Built-in variables cannot be arrays add_builtin(variable, builtin); if (builtin == spv::BuiltInPosition && storage == spv::StorageClassOutput) position_variable = variable; if (builtin == spv::BuiltInPointSize && storage == spv::StorageClassOutput) point_size_variable = variable; } else { assert(stype != shader_type::cs); // Compute shaders cannot have custom inputs or outputs const uint32_t location = semantic_to_location(semantic, std::max(1, param_type.array_length)); add_decoration(variable, spv::DecorationLocation, { location + a }); } if (param_type.has(type::q_noperspective)) add_decoration(variable, spv::DecorationNoPerspective); if (param_type.has(type::q_centroid)) add_decoration(variable, spv::DecorationCentroid); if (param_type.has(type::q_nointerpolation)) add_decoration(variable, spv::DecorationFlat); inputs_and_outputs.push_back(variable); return variable; }; // Translate function parameters to input/output variables for (const struct_member_info ¶m : func.parameter_list) { spv::Id param_var = create_varying_param(param); // Create separate input/output variables for "inout" parameters if (param.type.has(type::q_in)) { spv::Id param_value = 0; // Flatten structure parameters if (param.type.is_struct()) { const struct_info &definition = get_struct(param.type.definition); type struct_type = param.type; const int array_length = std::max(1, param.type.array_length); struct_type.array_length = 0; // Struct arrays need to be flattened into individual elements as well std::vector array_elements; array_elements.reserve(array_length); for (int a = 0; a < array_length; a++) { std::vector struct_elements; struct_elements.reserve(definition.member_list.size()); for (const struct_member_info &member : definition.member_list) { spv::Id input_var = create_varying_variable(member.type, member.semantic, spv::StorageClassInput, a); param_value = add_instruction(spv::OpLoad, convert_type(member.type)) .add(input_var).result; struct_elements.push_back(param_value); } param_value = add_instruction(spv::OpCompositeConstruct, convert_type(struct_type)) .add(struct_elements.begin(), struct_elements.end()).result; array_elements.push_back(param_value); } if (param.type.is_array()) { // Build the array from all constructed struct elements param_value = add_instruction(spv::OpCompositeConstruct, convert_type(param.type)) .add(array_elements.begin(), array_elements.end()).result; } } else { spv::Id input_var = create_varying_variable(param.type, param.semantic, spv::StorageClassInput); param_value = add_instruction(spv::OpLoad, convert_type(param.type)) .add(input_var).result; } add_instruction_without_result(spv::OpStore) .add(param_var) .add(param_value); } if (param.type.has(type::q_out)) { if (param.type.is_struct()) { const struct_info &definition = get_struct(param.type.definition); for (int a = 0, array_length = std::max(1, param.type.array_length); a < array_length; a++) { for (const struct_member_info &member : definition.member_list) { create_varying_variable(member.type, member.semantic, spv::StorageClassOutput, a); } } } else { create_varying_variable(param.type, param.semantic, spv::StorageClassOutput); } } } const auto call_result = emit_call({}, func.definition, func.return_type, call_params); for (size_t i = 0, inputs_and_outputs_index = 0; i < func.parameter_list.size(); ++i) { const struct_member_info ¶m = func.parameter_list[i]; if (param.type.has(type::q_out)) { const spv::Id value = add_instruction(spv::OpLoad, convert_type(param.type)) .add(call_params[i].base).result; if (param.type.is_struct()) { const struct_info &definition = get_struct(param.type.definition); type struct_type = param.type; const int array_length = std::max(1, param.type.array_length); struct_type.array_length = 0; // Skip input variables if this is an "inout" parameter if (param.type.has(type::q_in)) inputs_and_outputs_index += definition.member_list.size() * array_length; // Split up struct array into individual struct elements again for (int a = 0; a < array_length; a++) { spv::Id element_value = value; if (param.type.is_array()) { element_value = add_instruction(spv::OpCompositeExtract, convert_type(struct_type)) .add(value) .add(a).result; } // Split out struct fields into separate output variables again for (uint32_t member_index = 0; member_index < definition.member_list.size(); ++member_index) { const struct_member_info &member = definition.member_list[member_index]; const spv::Id member_value = add_instruction(spv::OpCompositeExtract, convert_type(member.type)) .add(element_value) .add(member_index).result; add_instruction_without_result(spv::OpStore) .add(inputs_and_outputs[inputs_and_outputs_index++]) .add(member_value); } } } else { // Skip input variable if this is an "inout" parameter (see loop above) if (param.type.has(type::q_in)) inputs_and_outputs_index += 1; add_instruction_without_result(spv::OpStore) .add(inputs_and_outputs[inputs_and_outputs_index++]) .add(value); } } else { // Input parameters do not need to store anything, but increase the input/output variable index if (param.type.is_struct()) { const struct_info &definition = get_struct(param.type.definition); inputs_and_outputs_index += definition.member_list.size() * std::max(1, param.type.array_length); } else { inputs_and_outputs_index += 1; } } } if (func.return_type.is_struct()) { const struct_info &definition = get_struct(func.return_type.definition); for (uint32_t member_index = 0; member_index < definition.member_list.size(); ++member_index) { const struct_member_info &member = definition.member_list[member_index]; const spv::Id result = create_varying_variable(member.type, member.semantic, spv::StorageClassOutput); const spv::Id member_result = add_instruction(spv::OpCompositeExtract, convert_type(member.type)) .add(call_result) .add(member_index).result; add_instruction_without_result(spv::OpStore) .add(result) .add(member_result); } } else if (!func.return_type.is_void()) { const spv::Id result = create_varying_variable(func.return_type, func.return_semantic, spv::StorageClassOutput); add_instruction_without_result(spv::OpStore) .add(result) .add(call_result); } // Add code to flip the output vertically if (_flip_vert_y && position_variable != 0 && stype == shader_type::vs) { expression position; position.reset_to_lvalue({}, position_variable, { type::t_float, 4, 1 }); position.add_constant_index_access(1); // Y component // gl_Position.y = -gl_Position.y emit_store(position, emit_unary_op({}, tokenid::minus, { type::t_float, 1, 1 }, emit_load(position, false))); } // Add code that sets the point size to a default value (in case this vertex shader is used with point primitives) if (point_size_variable == 0 && stype == shader_type::vs) { create_varying_variable({ type::t_float, 1, 1 }, "SV_POINTSIZE", spv::StorageClassOutput); expression point_size; point_size.reset_to_lvalue({}, point_size_variable, { type::t_float, 1, 1 }); // gl_PointSize = 1.0 emit_store(point_size, emit_constant({ type::t_float, 1, 1 }, 1)); } leave_block_and_return(0); leave_function(); spv::ExecutionModel model; switch (stype) { case shader_type::vs: model = spv::ExecutionModelVertex; break; case shader_type::ps: model = spv::ExecutionModelFragment; add_instruction_without_result(spv::OpExecutionMode, _execution_modes) .add(entry_point.definition) .add(_vulkan_semantics ? spv::ExecutionModeOriginUpperLeft : spv::ExecutionModeOriginLowerLeft); break; case shader_type::cs: model = spv::ExecutionModelGLCompute; add_instruction_without_result(spv::OpExecutionMode, _execution_modes) .add(entry_point.definition) .add(spv::ExecutionModeLocalSize) .add(num_threads[0]) .add(num_threads[1]) .add(num_threads[2]); break; default: assert(false); return; } assert(!func.unique_name.empty()); add_instruction_without_result(spv::OpEntryPoint, _entries) .add(model) .add(entry_point.definition) .add_string(func.unique_name.c_str()) .add(inputs_and_outputs.begin(), inputs_and_outputs.end()); } id emit_load(const expression &exp, bool) override { if (exp.is_constant) // Constant expressions do not have a complex access chain return emit_constant(exp.type, exp.constant); size_t i = 0; spv::Id result = exp.base; auto base_type = exp.type; bool is_uniform_bool = false; if (exp.is_lvalue || !exp.chain.empty()) add_location(exp.location, *_current_block_data); // If a variable is referenced, load the value first if (exp.is_lvalue && _spec_constants.find(exp.base) == _spec_constants.end()) { if (!exp.chain.empty()) base_type = exp.chain[0].from; std::pair storage = { spv::StorageClassFunction, spv::ImageFormatUnknown }; if (const auto it = _storage_lookup.find(exp.base); it != _storage_lookup.end()) storage = it->second; spirv_instruction *access_chain = nullptr; // Check if this is a uniform variable (see 'define_uniform' function above) and dereference it if (result & 0xF0000000) { const uint32_t member_index = result ^ 0xF0000000; storage.first = spv::StorageClassUniform; is_uniform_bool = base_type.is_boolean(); if (is_uniform_bool) base_type.base = type::t_uint; access_chain = &add_instruction(spv::OpAccessChain) .add(_global_ubo_variable) .add(emit_constant(member_index)); } // Any indexing expressions can be resolved during load with an 'OpAccessChain' already if (!exp.chain.empty() && ( exp.chain[0].op == expression::operation::op_member || exp.chain[0].op == expression::operation::op_dynamic_index || exp.chain[0].op == expression::operation::op_constant_index)) { // Ensure that 'access_chain' cannot get invalidated by calls to 'emit_constant' or 'convert_type' assert(_current_block_data != &_types_and_constants); // Use access chain from uniform if possible, otherwise create new one if (access_chain == nullptr) access_chain = &add_instruction(spv::OpAccessChain).add(result); // Base // Ignore first index into 1xN matrices, since they were translated to a vector type in SPIR-V if (exp.chain[0].from.rows == 1 && exp.chain[0].from.cols > 1) i = 1; for (; i < exp.chain.size() && ( exp.chain[i].op == expression::operation::op_member || exp.chain[i].op == expression::operation::op_dynamic_index || exp.chain[i].op == expression::operation::op_constant_index); ++i) access_chain->add(exp.chain[i].op == expression::operation::op_dynamic_index ? exp.chain[i].index : emit_constant(exp.chain[i].index)); // Indexes base_type = exp.chain[i - 1].to; access_chain->type = convert_type(base_type, true, storage.first, storage.second); // Last type is the result result = access_chain->result; } else if (access_chain != nullptr) { access_chain->type = convert_type(base_type, true, storage.first, storage.second, base_type.is_array() ? 16u : 0u); result = access_chain->result; } result = add_instruction(spv::OpLoad, convert_type(base_type, false, spv::StorageClassFunction, storage.second)) .add(result) // Pointer .result; } // Need to convert boolean uniforms which are actually integers in SPIR-V if (is_uniform_bool) { base_type.base = type::t_bool; result = add_instruction(spv::OpINotEqual, convert_type(base_type)) .add(result) .add(emit_constant(0)) .result; } // Work through all remaining operations in the access chain and apply them to the value for (; i < exp.chain.size(); ++i) { assert(result != 0); const auto &op = exp.chain[i]; switch (op.op) { case expression::operation::op_cast: if (op.from.is_scalar() && !op.to.is_scalar()) { type cast_type = op.to; cast_type.base = op.from.base; std::vector args; for (unsigned int c = 0; c < op.to.components(); ++c) args.emplace_back().reset_to_rvalue(exp.location, result, op.from); result = emit_construct(exp.location, cast_type, args); } if (op.from.is_boolean()) { const spv::Id true_constant = emit_constant(op.to, 1); const spv::Id false_constant = emit_constant(op.to, 0); result = add_instruction(spv::OpSelect, convert_type(op.to)) .add(result) // Condition .add(true_constant) .add(false_constant) .result; } else { spv::Op spv_op = spv::OpNop; switch (op.to.base) { case type::t_bool: if (op.from.is_floating_point()) spv_op = spv::OpFOrdNotEqual; else spv_op = spv::OpINotEqual; // Add instruction to compare value against zero instead of casting result = add_instruction(spv_op, convert_type(op.to)) .add(result) .add(emit_constant(op.from, 0)) .result; continue; case type::t_min16int: case type::t_int: if (op.from.is_floating_point()) spv_op = spv::OpConvertFToS; else if (op.from.precision() == op.to.precision()) spv_op = spv::OpBitcast; else if (_enable_16bit_types) spv_op = spv::OpSConvert; else continue; // Do not have to add conversion instruction between min16int/int if 16-bit types are not enabled break; case type::t_min16uint: case type::t_uint: if (op.from.is_floating_point()) spv_op = spv::OpConvertFToU; else if (op.from.precision() == op.to.precision()) spv_op = spv::OpBitcast; else if (_enable_16bit_types) spv_op = spv::OpUConvert; else continue; break; case type::t_min16float: case type::t_float: if (op.from.is_floating_point() && !_enable_16bit_types) continue; // Do not have to add conversion instruction between min16float/float if 16-bit types are not enabled else if (op.from.is_floating_point()) spv_op = spv::OpFConvert; else if (op.from.is_signed()) spv_op = spv::OpConvertSToF; else spv_op = spv::OpConvertUToF; break; default: assert(false); } result = add_instruction(spv_op, convert_type(op.to)) .add(result) .result; } break; case expression::operation::op_dynamic_index: assert(op.from.is_vector() && op.to.is_scalar()); result = add_instruction(spv::OpVectorExtractDynamic, convert_type(op.to)) .add(result) // Vector .add(op.index) // Index .result; break; case expression::operation::op_member: // In case of struct return values, which are r-values case expression::operation::op_constant_index: assert(op.from.is_vector() || op.from.is_matrix() || op.from.is_struct()); result = add_instruction(spv::OpCompositeExtract, convert_type(op.to)) .add(result) .add(op.index) // Literal Index .result; break; case expression::operation::op_swizzle: if (op.to.is_vector()) { if (op.from.is_matrix()) { spv::Id components[4]; for (unsigned int c = 0; c < 4 && op.swizzle[c] >= 0; ++c) { const unsigned int row = op.swizzle[c] / 4; const unsigned int column = op.swizzle[c] - row * 4; type scalar_type = op.to; scalar_type.rows = 1; scalar_type.cols = 1; spirv_instruction &node = add_instruction(spv::OpCompositeExtract, convert_type(scalar_type)) .add(result); if (op.from.rows > 1) // Matrix types with a single row are actually vectors, so they don't need the extra index node.add(row); node.add(column); components[c] = node.result; } spirv_instruction &node = add_instruction(spv::OpCompositeConstruct, convert_type(op.to)); for (unsigned int c = 0; c < 4 && op.swizzle[c] >= 0; ++c) node.add(components[c]); result = node.result; break; } else if (op.from.is_vector()) { spirv_instruction &node = add_instruction(spv::OpVectorShuffle, convert_type(op.to)) .add(result) // Vector 1 .add(result); // Vector 2 for (unsigned int c = 0; c < 4 && op.swizzle[c] >= 0; ++c) node.add(op.swizzle[c]); result = node.result; break; } else { spirv_instruction &node = add_instruction(spv::OpCompositeConstruct, convert_type(op.to)); for (unsigned int c = 0; c < op.to.rows; ++c) node.add(result); result = node.result; break; } } else if (op.from.is_matrix() && op.to.is_scalar()) { assert(op.swizzle[1] < 0); spirv_instruction &node = add_instruction(spv::OpCompositeExtract, convert_type(op.to)) .add(result); // Composite if (op.from.rows > 1) { const unsigned int row = op.swizzle[0] / 4; const unsigned int column = op.swizzle[0] - row * 4; node.add(row); node.add(column); } else { node.add(op.swizzle[0]); } result = node.result; // Result ID break; } assert(false); break; } } return result; } void emit_store(const expression &exp, id value) override { assert(value != 0 && exp.is_lvalue && !exp.is_constant && !exp.type.is_sampler()); add_location(exp.location, *_current_block_data); size_t i = 0; // Any indexing expressions can be resolved with an 'OpAccessChain' already spv::Id target = emit_access_chain(exp, i); auto base_type = exp.chain.empty() ? exp.type : i == 0 ? exp.chain[0].from : exp.chain[i - 1].to; // TODO: Complex access chains like float4x4[0].m00m10[0] = 0; // Work through all remaining operations in the access chain and apply them to the value for (; i < exp.chain.size(); ++i) { const auto &op = exp.chain[i]; switch (op.op) { case expression::operation::op_cast: case expression::operation::op_member: // These should have been handled above already (and casting does not make sense for a store operation) break; case expression::operation::op_dynamic_index: case expression::operation::op_constant_index: assert(false); break; case expression::operation::op_swizzle: { spv::Id result = add_instruction(spv::OpLoad, convert_type(base_type)) .add(target) // Pointer .result; // Result ID if (base_type.is_vector()) { spirv_instruction &node = add_instruction(spv::OpVectorShuffle, convert_type(base_type)) .add(result) // Vector 1 .add(value); // Vector 2 unsigned int shuffle[4] = { 0, 1, 2, 3 }; for (unsigned int c = 0; c < base_type.rows; ++c) if (op.swizzle[c] >= 0) shuffle[op.swizzle[c]] = base_type.rows + c; for (unsigned int c = 0; c < base_type.rows; ++c) node.add(shuffle[c]); value = node.result; } else if (op.to.is_scalar()) { assert(op.swizzle[1] < 0); spirv_instruction &node = add_instruction(spv::OpCompositeInsert, convert_type(base_type)) .add(value) // Object .add(result); // Composite if (op.from.is_matrix() && op.from.rows > 1) { const unsigned int row = op.swizzle[0] / 4; const unsigned int column = op.swizzle[0] - row * 4; node.add(row); node.add(column); } else { node.add(op.swizzle[0]); } value = node.result; // Result ID } else { // TODO: Implement matrix to vector swizzles assert(false); } break; } } } add_instruction_without_result(spv::OpStore) .add(target) .add(value); } id emit_access_chain(const expression &exp, size_t &i) override { // This function cannot create access chains for uniform variables assert((exp.base & 0xF0000000) == 0); i = 0; if (exp.chain.empty() || ( exp.chain[0].op != expression::operation::op_member && exp.chain[0].op != expression::operation::op_dynamic_index && exp.chain[0].op != expression::operation::op_constant_index)) return exp.base; std::pair storage = { spv::StorageClassFunction, spv::ImageFormatUnknown }; if (const auto it = _storage_lookup.find(exp.base); it != _storage_lookup.end()) storage = it->second; // Ensure that 'access_chain' cannot get invalidated by calls to 'emit_constant' or 'convert_type' assert(_current_block_data != &_types_and_constants); spirv_instruction *access_chain = &add_instruction(spv::OpAccessChain).add(exp.base); // Base // Ignore first index into 1xN matrices, since they were translated to a vector type in SPIR-V if (exp.chain[0].from.rows == 1 && exp.chain[0].from.cols > 1) i = 1; for (; i < exp.chain.size() && ( exp.chain[i].op == expression::operation::op_member || exp.chain[i].op == expression::operation::op_dynamic_index || exp.chain[i].op == expression::operation::op_constant_index); ++i) access_chain->add(exp.chain[i].op == expression::operation::op_dynamic_index ? exp.chain[i].index : emit_constant(exp.chain[i].index)); // Indexes access_chain->type = convert_type(exp.chain[i - 1].to, true, storage.first, storage.second); // Last type is the result return access_chain->result; } id emit_constant(uint32_t value) { return emit_constant({ type::t_uint, 1, 1 }, value); } id emit_constant(const type &type, uint32_t value) { // Create a constant value of the specified type constant data = {}; // Initialize to zero, so that components not set below still have a defined value for the lookup via std::memcmp for (unsigned int i = 0; i < type.components(); ++i) if (type.is_integral()) data.as_uint[i] = value; else data.as_float[i] = static_cast(value); return emit_constant(type, data, false); } id emit_constant(const type &type, const constant &data) override { return emit_constant(type, data, false); } id emit_constant(const type &type, const constant &data, bool spec_constant) { if (!spec_constant) // Specialization constants cannot reuse other constants { if (const auto it = std::find_if(_constant_lookup.begin(), _constant_lookup.end(), [&type, &data](auto &x) { if (!(std::get<0>(x) == type && std::memcmp(&std::get<1>(x).as_uint[0], &data.as_uint[0], sizeof(uint32_t) * 16) == 0 && std::get<1>(x).array_data.size() == data.array_data.size())) return false; for (size_t i = 0; i < data.array_data.size(); ++i) if (std::memcmp(&std::get<1>(x).array_data[i].as_uint[0], &data.array_data[i].as_uint[0], sizeof(uint32_t) * 16) != 0) return false; return true; }); it != _constant_lookup.end()) return std::get<2>(*it); // Re-use existing constant instead of duplicating the definition } spv::Id result; if (type.is_array()) { assert(type.array_length > 0); // Unsized arrays cannot be constants auto elem_type = type; elem_type.array_length = 0; std::vector elements; elements.reserve(type.array_length); // Fill up elements with constant array data for (const constant &elem : data.array_data) elements.push_back(emit_constant(elem_type, elem, spec_constant)); // Fill up any remaining elements with a default value (when the array data did not specify them) for (size_t i = elements.size(); i < static_cast(type.array_length); ++i) elements.push_back(emit_constant(elem_type, {}, spec_constant)); result = add_instruction(spec_constant ? spv::OpSpecConstantComposite : spv::OpConstantComposite, convert_type(type), _types_and_constants) .add(elements.begin(), elements.end()) .result; } else if (type.is_struct()) { assert(!spec_constant); // Structures cannot be specialization constants result = add_instruction(spv::OpConstantNull, convert_type(type), _types_and_constants) .result; } else if (type.is_vector() || type.is_matrix()) { auto elem_type = type; elem_type.rows = type.cols; elem_type.cols = 1; spv::Id rows[4] = {}; // Construct matrix constant out of row vector constants // Construct vector constant out of scalar constants for each element for (unsigned int i = 0; i < type.rows; ++i) { constant row_data = {}; for (unsigned int k = 0; k < type.cols; ++k) row_data.as_uint[k] = data.as_uint[i * type.cols + k]; rows[i] = emit_constant(elem_type, row_data, spec_constant); } if (type.rows == 1) { result = rows[0]; } else { spirv_instruction &node = add_instruction(spec_constant ? spv::OpSpecConstantComposite : spv::OpConstantComposite, convert_type(type), _types_and_constants); for (unsigned int i = 0; i < type.rows; ++i) node.add(rows[i]); result = node.result; } } else if (type.is_boolean()) { result = add_instruction(data.as_uint[0] ? (spec_constant ? spv::OpSpecConstantTrue : spv::OpConstantTrue) : (spec_constant ? spv::OpSpecConstantFalse : spv::OpConstantFalse), convert_type(type), _types_and_constants) .result; } else { assert(type.is_scalar()); result = add_instruction(spec_constant ? spv::OpSpecConstant : spv::OpConstant, convert_type(type), _types_and_constants) .add(data.as_uint[0]) .result; } if (spec_constant) // Keep track of all specialization constants _spec_constants.insert(result); else _constant_lookup.push_back({ type, data, result }); return result; } id emit_unary_op(const location &loc, tokenid op, const type &type, id val) override { spv::Op spv_op = spv::OpNop; switch (op) { case tokenid::minus: spv_op = type.is_floating_point() ? spv::OpFNegate : spv::OpSNegate; break; case tokenid::tilde: spv_op = spv::OpNot; break; case tokenid::exclaim: spv_op = spv::OpLogicalNot; break; default: return assert(false), 0; } add_location(loc, *_current_block_data); spirv_instruction &inst = add_instruction(spv_op, convert_type(type)); inst.add(val); // Operand return inst.result; } id emit_binary_op(const location &loc, tokenid op, const type &res_type, const type &type, id lhs, id rhs) override { spv::Op spv_op = spv::OpNop; switch (op) { case tokenid::plus: case tokenid::plus_plus: case tokenid::plus_equal: spv_op = type.is_floating_point() ? spv::OpFAdd : spv::OpIAdd; break; case tokenid::minus: case tokenid::minus_minus: case tokenid::minus_equal: spv_op = type.is_floating_point() ? spv::OpFSub : spv::OpISub; break; case tokenid::star: case tokenid::star_equal: spv_op = type.is_floating_point() ? spv::OpFMul : spv::OpIMul; break; case tokenid::slash: case tokenid::slash_equal: spv_op = type.is_floating_point() ? spv::OpFDiv : type.is_signed() ? spv::OpSDiv : spv::OpUDiv; break; case tokenid::percent: case tokenid::percent_equal: spv_op = type.is_floating_point() ? spv::OpFRem : type.is_signed() ? spv::OpSRem : spv::OpUMod; break; case tokenid::caret: case tokenid::caret_equal: spv_op = spv::OpBitwiseXor; break; case tokenid::pipe: case tokenid::pipe_equal: spv_op = spv::OpBitwiseOr; break; case tokenid::ampersand: case tokenid::ampersand_equal: spv_op = spv::OpBitwiseAnd; break; case tokenid::less_less: case tokenid::less_less_equal: spv_op = spv::OpShiftLeftLogical; break; case tokenid::greater_greater: case tokenid::greater_greater_equal: spv_op = type.is_signed() ? spv::OpShiftRightArithmetic : spv::OpShiftRightLogical; break; case tokenid::pipe_pipe: spv_op = spv::OpLogicalOr; break; case tokenid::ampersand_ampersand: spv_op = spv::OpLogicalAnd; break; case tokenid::less: spv_op = type.is_floating_point() ? spv::OpFOrdLessThan : type.is_signed() ? spv::OpSLessThan : spv::OpULessThan; break; case tokenid::less_equal: spv_op = type.is_floating_point() ? spv::OpFOrdLessThanEqual : type.is_signed() ? spv::OpSLessThanEqual : spv::OpULessThanEqual; break; case tokenid::greater: spv_op = type.is_floating_point() ? spv::OpFOrdGreaterThan : type.is_signed() ? spv::OpSGreaterThan : spv::OpUGreaterThan; break; case tokenid::greater_equal: spv_op = type.is_floating_point() ? spv::OpFOrdGreaterThanEqual : type.is_signed() ? spv::OpSGreaterThanEqual : spv::OpUGreaterThanEqual; break; case tokenid::equal_equal: spv_op = type.is_floating_point() ? spv::OpFOrdEqual : type.is_boolean() ? spv::OpLogicalEqual : spv::OpIEqual; break; case tokenid::exclaim_equal: spv_op = type.is_floating_point() ? spv::OpFOrdNotEqual : type.is_boolean() ? spv::OpLogicalNotEqual : spv::OpINotEqual; break; default: return assert(false), 0; } add_location(loc, *_current_block_data); // Binary operators generally only work on scalars and vectors in SPIR-V, so need to apply them to matrices component-wise if (type.is_matrix() && type.rows != 1) { std::vector ids; ids.reserve(type.cols); auto vector_type = type; vector_type.rows = type.cols; vector_type.cols = 1; for (unsigned int row = 0; row < type.rows; ++row) { const spv::Id lhs_elem = add_instruction(spv::OpCompositeExtract, convert_type(vector_type)) .add(lhs) .add(row) .result; const spv::Id rhs_elem = add_instruction(spv::OpCompositeExtract, convert_type(vector_type)) .add(rhs) .add(row) .result; spirv_instruction &inst = add_instruction(spv_op, convert_type(vector_type)); inst.add(lhs_elem); // Operand 1 inst.add(rhs_elem); // Operand 2 if (res_type.has(type::q_precise)) add_decoration(inst.result, spv::DecorationNoContraction); if (!_enable_16bit_types && res_type.precision() < 32) add_decoration(inst.result, spv::DecorationRelaxedPrecision); ids.push_back(inst.result); } spirv_instruction &inst = add_instruction(spv::OpCompositeConstruct, convert_type(res_type)); inst.add(ids.begin(), ids.end()); return inst.result; } else { spirv_instruction &inst = add_instruction(spv_op, convert_type(res_type)); inst.add(lhs); // Operand 1 inst.add(rhs); // Operand 2 if (res_type.has(type::q_precise)) add_decoration(inst.result, spv::DecorationNoContraction); if (!_enable_16bit_types && res_type.precision() < 32) add_decoration(inst.result, spv::DecorationRelaxedPrecision); return inst.result; } } id emit_ternary_op(const location &loc, tokenid op, const type &type, id condition, id true_value, id false_value) override { if (op != tokenid::question) return assert(false), 0; add_location(loc, *_current_block_data); spirv_instruction &inst = add_instruction(spv::OpSelect, convert_type(type)); inst.add(condition); // Condition inst.add(true_value); // Object 1 inst.add(false_value); // Object 2 return inst.result; } id emit_call(const location &loc, id function, const type &res_type, const std::vector &args) override { #ifndef NDEBUG for (const expression &arg : args) assert(arg.chain.empty() && arg.base != 0); #endif add_location(loc, *_current_block_data); // https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#OpFunctionCall spirv_instruction &inst = add_instruction(spv::OpFunctionCall, convert_type(res_type)); inst.add(function); // Function for (const expression &arg : args) inst.add(arg.base); // Arguments return inst.result; } id emit_call_intrinsic(const location &loc, id intrinsic, const type &res_type, const std::vector &args) override { #ifndef NDEBUG for (const expression &arg : args) assert(arg.chain.empty() && arg.base != 0); #endif add_location(loc, *_current_block_data); enum { #define IMPLEMENT_INTRINSIC_SPIRV(name, i, code) name##i, #include "effect_symbol_table_intrinsics.inl" }; switch (intrinsic) { #define IMPLEMENT_INTRINSIC_SPIRV(name, i, code) case name##i: code #include "effect_symbol_table_intrinsics.inl" default: return assert(false), 0; } } id emit_construct(const location &loc, const type &type, const std::vector &args) override { #ifndef NDEBUG for (const expression &arg : args) assert((arg.type.is_scalar() || type.is_array()) && arg.chain.empty() && arg.base != 0); #endif add_location(loc, *_current_block_data); std::vector ids; ids.reserve(args.size()); // There must be exactly one constituent for each top-level component of the result if (type.is_matrix()) { auto vector_type = type; vector_type.rows = type.cols; vector_type.cols = 1; // Turn the list of scalar arguments into a list of column vectors for (size_t arg = 0; arg < args.size(); arg += vector_type.rows) { spirv_instruction &inst = add_instruction(spv::OpCompositeConstruct, convert_type(vector_type)); for (unsigned row = 0; row < vector_type.rows; ++row) inst.add(args[arg + row].base); ids.push_back(inst.result); } } else { assert(type.is_vector() || type.is_array()); // The exception is that for constructing a vector, a contiguous subset of the scalars consumed can be represented by a vector operand instead for (const expression &arg : args) ids.push_back(arg.base); } spirv_instruction &inst = add_instruction(spv::OpCompositeConstruct, convert_type(type)); inst.add(ids.begin(), ids.end()); return inst.result; } void emit_if(const location &loc, id, id condition_block, id true_statement_block, id false_statement_block, unsigned int selection_control) override { spirv_instruction merge_label = _current_block_data->instructions.back(); assert(merge_label.op == spv::OpLabel); _current_block_data->instructions.pop_back(); // Add previous block containing the condition value first _current_block_data->append(_block_data[condition_block]); spirv_instruction branch_inst = _current_block_data->instructions.back(); assert(branch_inst.op == spv::OpBranchConditional); _current_block_data->instructions.pop_back(); // Add structured control flow instruction add_location(loc, *_current_block_data); add_instruction_without_result(spv::OpSelectionMerge) .add(merge_label.result) .add(selection_control & 0x3); // 'SelectionControl' happens to match the flags produced by the parser // Append all blocks belonging to the branch _current_block_data->instructions.push_back(branch_inst); _current_block_data->append(_block_data[true_statement_block]); _current_block_data->append(_block_data[false_statement_block]); _current_block_data->instructions.push_back(merge_label); } id emit_phi(const location &loc, id, id condition_block, id true_value, id true_statement_block, id false_value, id false_statement_block, const type &type) override { spirv_instruction merge_label = _current_block_data->instructions.back(); assert(merge_label.op == spv::OpLabel); _current_block_data->instructions.pop_back(); // Add previous block containing the condition value first _current_block_data->append(_block_data[condition_block]); if (true_statement_block != condition_block) _current_block_data->append(_block_data[true_statement_block]); if (false_statement_block != condition_block) _current_block_data->append(_block_data[false_statement_block]); _current_block_data->instructions.push_back(merge_label); add_location(loc, *_current_block_data); // https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#OpPhi spirv_instruction &inst = add_instruction(spv::OpPhi, convert_type(type)) .add(true_value) // Variable 0 .add(true_statement_block) // Parent 0 .add(false_value) // Variable 1 .add(false_statement_block); // Parent 1 return inst.result; } void emit_loop(const location &loc, id, id prev_block, id header_block, id condition_block, id loop_block, id continue_block, unsigned int loop_control) override { spirv_instruction merge_label = _current_block_data->instructions.back(); assert(merge_label.op == spv::OpLabel); _current_block_data->instructions.pop_back(); // Add previous block first _current_block_data->append(_block_data[prev_block]); // Fill header block assert(_block_data[header_block].instructions.size() == 2); _current_block_data->instructions.push_back(_block_data[header_block].instructions[0]); assert(_current_block_data->instructions.back().op == spv::OpLabel); // Add structured control flow instruction add_location(loc, *_current_block_data); add_instruction_without_result(spv::OpLoopMerge) .add(merge_label.result) .add(continue_block) .add(loop_control & 0x3); // 'LoopControl' happens to match the flags produced by the parser _current_block_data->instructions.push_back(_block_data[header_block].instructions[1]); assert(_current_block_data->instructions.back().op == spv::OpBranch); // Add condition block if it exists if (condition_block != 0) _current_block_data->append(_block_data[condition_block]); // Append loop body block before continue block _current_block_data->append(_block_data[loop_block]); _current_block_data->append(_block_data[continue_block]); _current_block_data->instructions.push_back(merge_label); } void emit_switch(const location &loc, id, id selector_block, id default_label, id default_block, const std::vector &case_literal_and_labels, const std::vector &case_blocks, unsigned int selection_control) override { assert(case_blocks.size() == case_literal_and_labels.size() / 2); spirv_instruction merge_label = _current_block_data->instructions.back(); assert(merge_label.op == spv::OpLabel); _current_block_data->instructions.pop_back(); // Add previous block containing the selector value first _current_block_data->append(_block_data[selector_block]); spirv_instruction switch_inst = _current_block_data->instructions.back(); assert(switch_inst.op == spv::OpSwitch); _current_block_data->instructions.pop_back(); // Add structured control flow instruction add_location(loc, *_current_block_data); add_instruction_without_result(spv::OpSelectionMerge) .add(merge_label.result) .add(selection_control & 0x3); // 'SelectionControl' happens to match the flags produced by the parser // Update switch instruction to contain all case labels switch_inst.operands[1] = default_label; switch_inst.add(case_literal_and_labels.begin(), case_literal_and_labels.end()); // Append all blocks belonging to the switch _current_block_data->instructions.push_back(switch_inst); std::vector blocks = case_blocks; if (default_label != merge_label.result) blocks.push_back(default_block); // Eliminate duplicates (because of multiple case labels pointing to the same block) std::sort(blocks.begin(), blocks.end()); blocks.erase(std::unique(blocks.begin(), blocks.end()), blocks.end()); for (const id case_block : blocks) _current_block_data->append(_block_data[case_block]); _current_block_data->instructions.push_back(merge_label); } bool is_in_function() const override { return _current_function != nullptr; } id set_block(id id) override { _last_block = _current_block; _current_block = id; _current_block_data = &_block_data[id]; return _last_block; } void enter_block(id id) override { assert(id != 0); // Can only use labels inside functions and should never be in another basic block if creating a new one assert(is_in_function() && !is_in_block()); set_block(id); add_instruction_without_result(spv::OpLabel) .result = id; } id leave_block_and_kill() override { assert(is_in_function()); // Can only discard inside functions if (!is_in_block()) return 0; add_instruction_without_result(spv::OpKill); return set_block(0); } id leave_block_and_return(id value) override { assert(is_in_function()); // Can only return from inside functions if (!is_in_block()) // Might already have left the last block in which case this has to be ignored return 0; if (_current_function->return_type.is_void()) { add_instruction_without_result(spv::OpReturn); } else { if (0 == value) // The implicit return statement needs this value = add_instruction(spv::OpUndef, convert_type(_current_function->return_type), _types_and_constants).result; add_instruction_without_result(spv::OpReturnValue) .add(value); } return set_block(0); } id leave_block_and_switch(id value, id default_target) override { assert(value != 0 && default_target != 0); assert(is_in_function()); // Can only switch inside functions if (!is_in_block()) return _last_block; add_instruction_without_result(spv::OpSwitch) .add(value) .add(default_target); return set_block(0); } id leave_block_and_branch(id target, unsigned int) override { assert(target != 0); assert(is_in_function()); // Can only branch inside functions if (!is_in_block()) return _last_block; add_instruction_without_result(spv::OpBranch) .add(target); return set_block(0); } id leave_block_and_branch_conditional(id condition, id true_target, id false_target) override { assert(condition != 0 && true_target != 0 && false_target != 0); assert(is_in_function()); // Can only branch inside functions if (!is_in_block()) return _last_block; add_instruction_without_result(spv::OpBranchConditional) .add(condition) .add(true_target) .add(false_target); return set_block(0); } void leave_function() override { assert(is_in_function()); // Can only leave if there was a function to begin with _current_function->definition = _block_data[_last_block]; // Append function end instruction add_instruction_without_result(spv::OpFunctionEnd, _current_function->definition); _current_function = nullptr; } }; codegen *reshadefx::create_codegen_spirv(bool vulkan_semantics, bool debug_info, bool uniforms_to_spec_constants, bool enable_16bit_types, bool flip_vert_y) { return new codegen_spirv(vulkan_semantics, debug_info, uniforms_to_spec_constants, enable_16bit_types, flip_vert_y); }