X-Git-Url: http://git.tdb.fi/?p=libs%2Fgl.git;a=blobdiff_plain;f=source%2Fcore%2Fmodule.cpp;h=5475e434ac2636331c846a529bdf99fb508efb04;hp=6d0122c055583d0bafa40f70d6bae08128322514;hb=HEAD;hpb=6cee6c490ded0981c4da3f9f4ec74a41555e1a4e

diff --git a/source/core/module.cpp b/source/core/module.cpp
index 6d0122c0..04a37fcc 100644
--- a/source/core/module.cpp
+++ b/source/core/module.cpp
@@ -1,5 +1,6 @@
 #include <msp/core/algorithm.h>
 #include <msp/io/print.h>
+#include "device.h"
 #include "module.h"
 #include "resources.h"
 
@@ -13,6 +14,7 @@ enum SpirVConstants
 	OP_NAME = 5,
 	OP_MEMBER_NAME = 6,
 	OP_ENTRY_POINT = 15,
+	OP_EXECUTION_MODE = 16,
 	OP_TYPE_VOID = 19,
 	OP_TYPE_BOOL = 20,
 	OP_TYPE_INT = 21,
@@ -31,8 +33,24 @@ enum SpirVConstants
 	OP_SPEC_CONSTANT_FALSE = 49,
 	OP_SPEC_CONSTANT = 50,
 	OP_VARIABLE = 59,
+	OP_LOAD = 61,
+	OP_STORE = 62,
+	OP_ACCESS_CHAIN = 65,
 	OP_DECORATE = 71,
 	OP_MEMBER_DECORATE = 72,
+	OP_COPY_OBJECT = 83,
+	OP_PHI = 245,
+	OP_SELECTION_MERGE = 247,
+	OP_LABEL = 248,
+	OP_BRANCH = 249,
+	OP_BRANCH_CONDITIONAL = 250,
+	OP_SWITCH = 251,
+	OP_KILL = 252,
+	OP_RETURN = 253,
+	OP_RETURN_VALUE = 254,
+	OP_UNREACHABLE = 255,
+
+	EXEC_LOCAL_SIZE = 17,
 
 	DECO_SPEC_ID = 1,
 	DECO_ARRAY_STRIDE = 6,
@@ -49,14 +67,14 @@ namespace GL {
 
 void Module::set_source(const string &src)
 {
-	SL::Compiler compiler;
+	SL::Compiler compiler(create_features());
 	compiler.set_source(src);
 	compile(compiler);
 }
 
 void Module::load_source(IO::Base &io, Resources *res, const string &name)
 {
-	SL::Compiler compiler;
+	SL::Compiler compiler(create_features());
 	compiler.load_source(io, res, name);
 	compile(compiler);
 }
@@ -66,6 +84,20 @@ void Module::load_source(IO::Base &io, const string &name)
 	load_source(io, 0, name);
 }
 
+SL::Features Module::create_features() const
+{
+	const DeviceInfo &dev_info = Device::get_current().get_info();
+	const SL::Features &device_features = dev_info.glsl_features;
+	SL::Features latest_features = SL::Features::latest(dev_info.api);
+	SL::Features features;
+	features.target_api = latest_features.target_api;
+	features.glsl_version = latest_features.glsl_version;
+	features.constant_id_range = device_features.constant_id_range;
+	features.uniform_binding_range = device_features.uniform_binding_range;
+	features.texture_binding_range = device_features.texture_binding_range;
+	return features;
+}
+
 
 void GlslModule::compile(SL::Compiler &compiler)
 {
@@ -81,41 +113,6 @@ void GlslModule::compile(SL::Compiler &compiler)
 }
 
 
-SpirVModule::SpirVModule(const SpirVModule &other):
-	code(other.code),
-	entry_points(other.entry_points),
-	structs(other.structs),
-	variables(other.variables)
-{
-	remap_pointers_from(other);
-}
-
-SpirVModule &SpirVModule::operator=(const SpirVModule &other)
-{
-	code = other.code;
-	entry_points = other.entry_points;
-	structs = other.structs;
-	variables = other.variables;
-	remap_pointers_from(other);
-	return *this;
-}
-
-void SpirVModule::remap_pointers_from(const SpirVModule &other)
-{
-	for(EntryPoint &e: entry_points)
-		for(const Variable *&v: e.globals)
-			v = &variables[v-&other.variables.front()];
-
-	for(Variable &v: variables)
-		if(v.struct_type)
-			v.struct_type = &structs[v.struct_type-&other.structs.front()];
-
-	for(Structure &s: structs)
-		for(StructMember &m: s.members)
-			if(m.struct_type)
-				m.struct_type = &structs[m.struct_type-&other.structs.front()];
-}
-
 void SpirVModule::load_code(IO::Base &io)
 {
 	uint32_t buffer[1024];
@@ -130,6 +127,7 @@ void SpirVModule::load_code(IO::Base &io)
 	}
 
 	reflect();
+	create();
 }
 
 void SpirVModule::compile(SL::Compiler &compiler)
@@ -137,6 +135,7 @@ void SpirVModule::compile(SL::Compiler &compiler)
 	compiler.compile(SL::Compiler::SPIRV);
 	code = compiler.get_combined_spirv();
 	reflect();
+	create();
 }
 
 void SpirVModule::reflect()
@@ -230,6 +229,255 @@ void SpirVModule::reflect()
 			v = (i!=var_indices.end() ? &variables[i->second] : 0);
 		}
 	}
+
+	map<const InstructionBlock *, unsigned> block_indices;
+	blocks.reserve(reflection.blocks.size());
+	for(const auto &kvp: reflection.blocks)
+	{
+		block_indices[&kvp.second] = blocks.size();
+		blocks.push_back(kvp.second);
+	}
+
+	for(InstructionBlock &b: blocks)
+	{
+		auto i = spec_indices.find(b.condition);
+		b.condition = (i!=spec_indices.end() ? &spec_constants[i->second] : 0);
+		if(b.condition)
+			specializable = true;
+
+		for(const Variable *&v: b.accessed_variables)
+		{
+			auto j = var_indices.find(v);
+			v = (j!=var_indices.end() ? &variables[j->second] : 0);
+		}
+
+		for(const InstructionBlock *&s: b.successors)
+		{
+			auto j = block_indices.find(s);
+			s = (j!=block_indices.end() ? &blocks[j->second] : 0);
+		}
+	}
+}
+
+SpirVModule *SpirVModule::specialize(const map<string, int> &spec_values) const
+{
+	vector<uint8_t> flags(code[3], 1);
+
+	std::map<unsigned, int> spec_values_by_id;
+	for(const Constant &c: spec_constants)
+	{
+		auto i = spec_values.find(c.name);
+		if(i!=spec_values.end())
+		{
+			flags[c.id] = (i->second ? 5 : 3);
+			spec_values_by_id[c.constant_id] = i->second;
+		}
+	}
+
+	for(const EntryPoint &e: entry_points)
+		flags[e.id] = 0;
+	for(const Variable &v: variables)
+		flags[v.id] = 0;
+	for(const InstructionBlock &b: blocks)
+		flags[b.id] = 0;
+	for(const InstructionBlock *b: collect_visited_blocks(spec_values_by_id))
+	{
+		flags[b->id] = 1;
+		for(const Variable *v: b->accessed_variables)
+			flags[v->id] = 1;
+	}
+
+	std::vector<uint32_t> new_code;
+	new_code.reserve(code.size());
+
+	auto op = code.begin()+5;
+	new_code.insert(new_code.begin(), code.begin(), op);
+
+	bool skip_block = false;
+	while(op!=code.end())
+	{
+		unsigned word_count = *op>>16;
+		unsigned opcode = *op&0xFFFF;
+
+		bool copy = !skip_block;
+		if(skip_block)
+		{
+			skip_block = (opcode!=OP_BRANCH && opcode!=OP_BRANCH_CONDITIONAL && opcode!=OP_SWITCH &&
+				opcode!=OP_KILL && opcode!=OP_RETURN && opcode!=OP_RETURN_VALUE && opcode!=OP_UNREACHABLE);
+		}
+		else
+		{
+			if(opcode==OP_NAME)
+				copy = flags[*(op+1)];
+			else if(opcode==OP_ENTRY_POINT)
+			{
+				unsigned start = new_code.size();
+				new_code.push_back(opcode);
+				new_code.push_back(*(op+1));
+				unsigned func_id = *(op+2);
+				new_code.push_back(func_id);
+
+				unsigned i = 3;
+				while(i<word_count)
+				{
+					unsigned word = *(op+i++);
+					new_code.push_back(word);
+					// Strings are nul-terminated and nul-padded
+					if(!(word>>24))
+						break;
+				}
+
+				unsigned var_count = 0;
+				for(; i<word_count; ++i)
+				{
+					unsigned id = *(op+i);
+					if(flags[id])
+					{
+						++var_count;
+						new_code.push_back(id);
+					}
+				}
+
+				if(var_count)
+				{
+					flags[func_id] = 1;
+					new_code[start] |= (new_code.size()-start)<<16;
+				}
+				else
+					new_code.resize(start);
+
+				copy = false;
+			}
+			else if(opcode==OP_EXECUTION_MODE)
+				copy = flags[*(op+1)];
+			else if(opcode==OP_SPEC_CONSTANT_TRUE || opcode==OP_SPEC_CONSTANT_FALSE)
+			{
+				unsigned id = *(op+2);
+				if(flags[id]&2)
+				{
+					new_code.push_back(0x30000 | (flags[id]&4 ? OP_CONSTANT_TRUE : OP_CONSTANT_FALSE));
+					new_code.push_back(*(op+1));
+					new_code.push_back(id);
+
+					copy = false;
+				}
+			}
+			else if(opcode==OP_VARIABLE)
+				copy = flags[*(op+2)];
+			else if(opcode==OP_DECORATE)
+			{
+				unsigned id = *(op+1);
+				copy = flags[id];
+				if(copy && *(op+2)==DECO_SPEC_ID)
+					copy = !(flags[id]&2);
+			}
+			else if(opcode==OP_LABEL)
+			{
+				copy = flags[*(op+1)];
+				skip_block = !copy;
+			}
+			else if(opcode==OP_SELECTION_MERGE)
+			{
+				unsigned next_opcode = *(op+word_count)&0xFFFF;
+				if(next_opcode==OP_BRANCH_CONDITIONAL)
+				{
+					unsigned true_id = *(op+word_count+2);
+					unsigned false_id = *(op+word_count+3);
+					if(!flags[true_id] || !flags[false_id])
+					{
+						new_code.push_back(0x20000 | OP_BRANCH);
+						new_code.push_back(flags[true_id] ? true_id : false_id);
+						copy = false;
+
+						/* Skip the branch instruction when it's encountered on the
+						next iteration */
+						skip_block = true;
+					}
+				}
+			}
+			else if(opcode==OP_PHI)
+			{
+				unsigned active_count = 0;
+				unsigned result_id = 0;
+				for(unsigned i=3; i<word_count; i+=2)
+					if(flags[*(op+i+1)])
+					{
+						++active_count;
+						result_id = *(op+i);
+					}
+
+				if(active_count==1)
+				{
+					new_code.push_back(0x40000 | OP_COPY_OBJECT);
+					new_code.push_back(*(op+1));
+					new_code.push_back(*(op+2));
+					new_code.push_back(result_id);
+					copy = false;
+				}
+			}
+		}
+
+		if(copy)
+		{
+			for(unsigned i=0; i<word_count; ++i)
+				new_code.push_back(*(op+i));
+		}
+
+		op += word_count;
+	}
+
+	SpirVModule *spec_mod = new SpirVModule;
+	spec_mod->code = move(new_code);
+	spec_mod->reflect();
+	spec_mod->create();
+
+	return spec_mod;
+}
+
+vector<const SpirVModule::InstructionBlock *> SpirVModule::collect_visited_blocks(const map<unsigned, int> &spec_values) const
+{
+	vector<uint8_t> visited(blocks.size(), 4);
+	for(unsigned i=0; i<blocks.size(); ++i)
+	{
+		const InstructionBlock &b = blocks[i];
+
+		bool cond = true;
+		if(b.condition)
+		{
+			cond = b.condition->i_value;
+			auto j = spec_values.find(b.condition->constant_id);
+			if(j!=spec_values.end())
+				cond = j->second;
+			if(b.negate_condition)
+				cond = !cond;
+		}
+
+		visited[i] |= cond*2;
+		for(const InstructionBlock *s: b.successors)
+			visited[s-blocks.data()] &= 3;
+	}
+
+	for(unsigned i=0; i<blocks.size(); ++i)
+		if(visited[i]&4)
+			collect_visited_blocks(i, visited);
+
+	vector<const SpirVModule::InstructionBlock *> result;
+	for(unsigned i=0; i<blocks.size(); ++i)
+		if(visited[i]&1)
+			result.push_back(&blocks[i]);
+
+	return result;
+}
+
+void SpirVModule::collect_visited_blocks(unsigned i, vector<uint8_t> &visited) const
+{
+	visited[i] |= 1;
+	for(const InstructionBlock *s: blocks[i].successors)
+	{
+		unsigned j = s-blocks.data();
+		if((visited[j]&3)==2)
+			collect_visited_blocks(j, visited);
+	}
 }
 
 
@@ -290,6 +538,7 @@ void SpirVModule::Reflection::reflect_code(const vector<uint32_t> &code)
 		case OP_NAME: reflect_name(op); break;
 		case OP_MEMBER_NAME: reflect_member_name(op); break;
 		case OP_ENTRY_POINT: reflect_entry_point(op); break;
+		case OP_EXECUTION_MODE: reflect_execution_mode(op); break;
 		case OP_TYPE_VOID: reflect_void_type(op); break;
 		case OP_TYPE_BOOL: reflect_bool_type(op); break;
 		case OP_TYPE_INT: reflect_int_type(op); break;
@@ -308,8 +557,14 @@ void SpirVModule::Reflection::reflect_code(const vector<uint32_t> &code)
 		case OP_SPEC_CONSTANT_FALSE:
 		case OP_SPEC_CONSTANT: reflect_constant(op); break;
 		case OP_VARIABLE: reflect_variable(op); break;
+		case OP_LOAD:
+		case OP_STORE: reflect_access(op); break;
+		case OP_ACCESS_CHAIN: reflect_access_chain(op); break;
 		case OP_DECORATE: reflect_decorate(op); break;
 		case OP_MEMBER_DECORATE: reflect_member_decorate(op); break;
+		case OP_LABEL: reflect_label(op); break;
+		case OP_BRANCH: reflect_branch(op); break;
+		case OP_BRANCH_CONDITIONAL: reflect_branch_conditional(op); break;
 		}
 
 		op += word_count;
@@ -338,7 +593,9 @@ void SpirVModule::Reflection::reflect_member_name(CodeIterator op)
 void SpirVModule::Reflection::reflect_entry_point(CodeIterator op)
 {
 	CodeIterator op_end = get_op_end(op);
-	EntryPoint &entry = entry_points[*(op+2)];
+	unsigned id = *(op+2);
+	EntryPoint &entry = entry_points[id];
+	entry.id = id;
 	entry.stage = static_cast<Stage>(*(op+1));  // Execution model in SPIR-V spec
 	op += 3;
 	entry.name = read_string(op, op_end);
@@ -348,6 +605,18 @@ void SpirVModule::Reflection::reflect_entry_point(CodeIterator op)
 		entry.globals.push_back(&variables[*op]);
 }
 
+void SpirVModule::Reflection::reflect_execution_mode(CodeIterator op)
+{
+	EntryPoint &entry = entry_points[*(op+1)];
+	unsigned mode = *(op+2);
+	if(mode==EXEC_LOCAL_SIZE)
+	{
+		entry.compute_local_size.x = *(op+3);
+		entry.compute_local_size.y = *(op+4);
+		entry.compute_local_size.z = *(op+5);
+	}
+}
+
 void SpirVModule::Reflection::reflect_void_type(CodeIterator op)
 {
 	types[*(op+1)].type = VOID;
@@ -392,11 +661,13 @@ void SpirVModule::Reflection::reflect_matrix_type(CodeIterator op)
 void SpirVModule::Reflection::reflect_image_type(CodeIterator op)
 {
 	TypeInfo &type = types[*(op+1)];
-	DataType sample = types[*(op+2)].type;
+	DataType sample_type = types[*(op+2)].type;
 	unsigned dimensions = *(op+3);
 	bool depth = *(op+4)==1;
 	bool array = *(op+5);
-	type.type = static_cast<DataType>((depth*0x200000) | (array*0x80000) | (dimensions+1) | sample);
+	bool sampled = *(op+7)==1;
+	type.type = static_cast<DataType>((depth*0x200000) | (sampled*0x100000) | (array*0x80000) |
+		((dimensions+1)<<16) | sample_type);
 }
 
 void SpirVModule::Reflection::reflect_sampled_image_type(CodeIterator op)
@@ -424,6 +695,7 @@ void SpirVModule::Reflection::reflect_struct_type(CodeIterator op)
 	unsigned id = *(op+1);
 	Structure &strct = structs[id];
 	strct.name = names[id];
+	strct.id = id;
 	types[id].struct_type = &strct;
 	op += 2;
 
@@ -448,13 +720,15 @@ void SpirVModule::Reflection::reflect_pointer_type(CodeIterator op)
 
 void SpirVModule::Reflection::reflect_constant(CodeIterator op)
 {
+	unsigned opcode = get_opcode(*op);
 	unsigned id = *(op+2);
 	Constant &cnst = constants[id];
 	cnst.name = names[id];
+	cnst.id = id;
 	cnst.type = types[*(op+1)].type;
-	if(*op==OP_CONSTANT_TRUE || *op==OP_SPEC_CONSTANT_TRUE)
+	if(opcode==OP_CONSTANT_TRUE || opcode==OP_SPEC_CONSTANT_TRUE)
 		cnst.i_value = true;
-	else if(*op==OP_CONSTANT_FALSE || *op==OP_SPEC_CONSTANT_FALSE)
+	else if(opcode==OP_CONSTANT_FALSE || opcode==OP_SPEC_CONSTANT_FALSE)
 		cnst.i_value = false;
 	else if(cnst.type==INT || cnst.type==UNSIGNED_INT)
 		cnst.i_value = *(op+3);
@@ -467,6 +741,7 @@ void SpirVModule::Reflection::reflect_variable(CodeIterator op)
 	unsigned id = *(op+2);
 	Variable &var = variables[id];
 	var.name = names[id];
+	var.id = id;
 	const TypeInfo &type = types[*(op+1)];
 	var.storage = type.storage;
 	var.type = type.type;
@@ -474,6 +749,26 @@ void SpirVModule::Reflection::reflect_variable(CodeIterator op)
 	var.array_size = type.array_size;
 }
 
+void SpirVModule::Reflection::reflect_access(CodeIterator op)
+{
+	if(current_block)
+	{
+		unsigned id = (get_opcode(*op)==OP_LOAD ? *(op+3) : *(op+1));
+		auto i = access_chain_bases.find(id);
+		if(i!=access_chain_bases.end())
+			id = i->second;
+		Variable &var = variables[id];
+		auto j = find(current_block->accessed_variables, &var);
+		if(j==current_block->accessed_variables.end())
+			current_block->accessed_variables.push_back(&var);
+	}
+}
+
+void SpirVModule::Reflection::reflect_access_chain(CodeIterator op)
+{
+	access_chain_bases[*(op+2)] = *(op+3);
+}
+
 void SpirVModule::Reflection::reflect_decorate(CodeIterator op)
 {
 	unsigned id = *(op+1);
@@ -527,5 +822,40 @@ void SpirVModule::Reflection::reflect_member_decorate(CodeIterator op)
 	}
 }
 
+void SpirVModule::Reflection::reflect_label(CodeIterator op)
+{
+	unsigned id = *(op+1);
+	current_block = &blocks[id];
+	current_block->id = id;
+}
+
+void SpirVModule::Reflection::reflect_branch(CodeIterator op)
+{
+	InstructionBlock &block = blocks[*(op+1)];
+	block.condition = &true_condition;
+	current_block->successors.push_back(&block);
+}
+
+void SpirVModule::Reflection::reflect_branch_conditional(CodeIterator op)
+{
+	InstructionBlock &true_block = blocks[*(op+2)];
+	InstructionBlock &false_block = blocks[*(op+3)];
+
+	auto i = constants.find(*(op+1));
+	if(i!=constants.end() && i->second.constant_id)
+	{
+		if(!true_block.condition)
+			true_block.condition = &i->second;
+		if(!false_block.condition)
+		{
+			false_block.condition = &i->second;
+			false_block.negate_condition = true;
+		}
+	}
+
+	current_block->successors.push_back(&true_block);
+	current_block->successors.push_back(&false_block);
+}
+
 } // namespace GL
 } // namespace Msp