diff --git a/paddle/fluid/framework/paddle2cinn/cinn_cache_key.cc b/paddle/fluid/framework/paddle2cinn/cinn_cache_key.cc index 368fb4a5fd8c96c03a52249b8243212380bc64ab..0e157ae7d79f38cda25a25be7af8f8a051a3e1ac 100644 --- a/paddle/fluid/framework/paddle2cinn/cinn_cache_key.cc +++ b/paddle/fluid/framework/paddle2cinn/cinn_cache_key.cc @@ -29,55 +29,32 @@ namespace paddle { namespace framework { namespace paddle2cinn { +using GraphHashStrategy = CinnCacheKey::GraphHashStrategy; + +CinnCacheKey::CinnCacheKey(GraphHashStrategy graph_hash) + : graph_hash_(graph_hash) {} + CinnCacheKey::CinnCacheKey( const ir::Graph& graph, const std::map& input_tensors, - const std::string& arch_str) { + const std::string& arch_str, GraphHashStrategy graph_hash) + : graph_hash_(graph_hash) { this->SetKey(graph, input_tensors, arch_str); } CinnCacheKey::CinnCacheKey(const ir::Graph& graph, const std::map& input_shapes, - const std::string& arch_str) { + const std::string& arch_str, + GraphHashStrategy graph_hash) + : graph_hash_(graph_hash) { this->SetKey(graph, input_shapes, arch_str); } -size_t CinnCacheKey::HashGraph(const ir::Graph& graph) { - // using Dot to unqiue graph - inference::analysis::Dot dot; - std::unordered_map node2dot; - int id = 0; - // Create nodes - // graph.Nodes() return unordered_set, the same graph may - // return different result? - for (const ir::Node* n : graph.Nodes()) { - std::string node_id = std::to_string(id++); - dot.AddNode(node_id, {}, n->Name(), true); - node2dot[n] = node_id; - } - - // Create edges - for (const ir::Node* n : graph.Nodes()) { - const auto& src_id = node2dot.at(n); - for (auto* out : n->outputs) { - const auto& dest_id = node2dot.at(out); - dot.AddEdge(src_id, dest_id, {}); - } - } - - const std::string& viz_graph = dot.Build(); - VLOG(1) << "The hash graph:\n" << viz_graph; - - size_t hash_val = std::hash()(viz_graph); - VLOG(4) << "The graph's hash value is: " << hash_val; - return hash_val; -} - void CinnCacheKey::SetKey( const ir::Graph& graph, const std::map& input_tensors, const std::string& arch_str) { - graph_serialize_str_ = std::to_string(HashGraph(graph)); + graph_hash_val_ = graph_hash_(graph); for (const auto& name_tensor : input_tensors) { input_shapes_[name_tensor.first] = name_tensor.second->dims(); } @@ -87,7 +64,7 @@ void CinnCacheKey::SetKey( void CinnCacheKey::SetKey(const ir::Graph& graph, const std::map& input_shapes, const std::string& arch_str) { - graph_serialize_str_ = std::to_string(HashGraph(graph)); + graph_hash_val_ = graph_hash_(graph); input_shapes_ = input_shapes; arch_str_ = arch_str; } @@ -97,7 +74,7 @@ bool CinnCacheKey::operator!=(const CinnCacheKey& other) const { } bool CinnCacheKey::operator==(const CinnCacheKey& other) const { - return graph_serialize_str_ == other.graph_serialize_str_ && + return graph_hash_val_ == other.graph_hash_val_ && input_shapes_ == other.input_shapes_ && arch_str_ == other.arch_str_; } @@ -114,11 +91,48 @@ size_t CinnCacheKey::Hash::operator()(const CinnCacheKey& key) const { ret = hash_combine(ret, string_hasher(name_shape.second.to_str())); } - ret = hash_combine(ret, string_hasher(key.graph_serialize_str_)); + ret = hash_combine(ret, key.graph_hash_val_); ret = hash_combine(ret, string_hasher(key.arch_str_)); return ret; } +size_t CinnCacheKeyByStructure::HashGraph(const ir::Graph& graph) { + // sort grad node by name and id. + auto compare = [](ir::Node* n1, ir::Node* n2) { + return (n1->Name() == n2->Name()) ? (n1->id() < n2->id()) + : (n1->Name() < n2->Name()); + }; + + // graph.Nodes() return unordered_set, here using set to avoid the same graph + // may return different result + std::set node_set(compare), + output_set(compare); + node_set.insert(graph.Nodes().begin(), graph.Nodes().end()); + + std::string hash_str; + for (ir::Node* n : node_set) { + hash_str.append(n->Name()); + + output_set.clear(); + output_set.insert(n->outputs.begin(), n->outputs.end()); + for (auto* out : output_set) { + hash_str.append(out->Name()); + } + } + + VLOG(1) << "The hash graph:\n" << hash_str; + + size_t hash_val = std::hash()(hash_str); + VLOG(4) << "The graph's hash value by graph structure is: " << hash_val; + return hash_val; +} + +size_t CinnCacheKeyByAddress::HashGraph(const ir::Graph& graph) { + size_t hash_val = reinterpret_cast(&graph); + VLOG(4) << "The graph's hash value by graph address is: " << hash_val; + return hash_val; +} + } // namespace paddle2cinn } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/paddle2cinn/cinn_cache_key.h b/paddle/fluid/framework/paddle2cinn/cinn_cache_key.h index 941f8e0cdecc188dde30cda12a6cb59ed48c097d..67325297c4772479781477668d130623a1165afe 100644 --- a/paddle/fluid/framework/paddle2cinn/cinn_cache_key.h +++ b/paddle/fluid/framework/paddle2cinn/cinn_cache_key.h @@ -14,6 +14,7 @@ #pragma once +#include #include #include "paddle/fluid/framework/ddim.h" @@ -33,14 +34,18 @@ namespace paddle2cinn { // shapes. class CinnCacheKey { public: + using GraphHashStrategy = std::function; + + explicit CinnCacheKey(GraphHashStrategy graph_hash); + CinnCacheKey(const ir::Graph& graph, const std::map& input_tensors, - const std::string& arch_str); + const std::string& arch_str, GraphHashStrategy graph_hash); CinnCacheKey(const ir::Graph& graph, const std::map& input_shapes, - const std::string& arch_str); + const std::string& arch_str, GraphHashStrategy graph_hash); - ~CinnCacheKey() {} + ~CinnCacheKey() = default; void SetKey(const ir::Graph& graph, const std::map& input_tensors, @@ -58,13 +63,38 @@ class CinnCacheKey { }; private: - size_t HashGraph(const ir::Graph& graph); - - std::string graph_serialize_str_; + GraphHashStrategy graph_hash_; + size_t graph_hash_val_; std::map input_shapes_; std::string arch_str_; }; +#define CINN_CACHE_KEY_CREATE(NAME) \ + class NAME : public CinnCacheKey { \ + public: \ + NAME() : CinnCacheKey(HashGraph) {} \ + \ + NAME(const ir::Graph& graph, \ + const std::map& input_tensors, \ + const std::string& arch_str) \ + : CinnCacheKey(graph, input_tensors, arch_str, HashGraph) {} \ + \ + NAME(const ir::Graph& graph, \ + const std::map& input_shapes, \ + const std::string& arch_str) \ + : CinnCacheKey(graph, input_shapes, arch_str, HashGraph) {} \ + \ + private: \ + static size_t HashGraph(const ir::Graph& graph); \ + }; + +// Class to store the keys by graph address for compiling CINN. +CINN_CACHE_KEY_CREATE(CinnCacheKeyByAddress) +// Class to store the keys by graph structure for compiling CINN. +CINN_CACHE_KEY_CREATE(CinnCacheKeyByStructure) + +#undef CINN_CACHE_KEY_CREATE + } // namespace paddle2cinn } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/paddle2cinn/cinn_cache_key_test.cc b/paddle/fluid/framework/paddle2cinn/cinn_cache_key_test.cc index f13f44998211f4d2a45088766c175e79f09853fa..f9b48ef4b5ec0cd4ca42635845ea49450727d966 100644 --- a/paddle/fluid/framework/paddle2cinn/cinn_cache_key_test.cc +++ b/paddle/fluid/framework/paddle2cinn/cinn_cache_key_test.cc @@ -26,8 +26,8 @@ namespace paddle { namespace framework { namespace paddle2cinn { -TEST(CinnCacheKeyTest, TestAsUnorderedKey) { - std::unordered_set test_set; +TEST(CinnCacheKeyTest, TestAsUnorderedKeyByStructure) { + std::unordered_set test_set; ProgramDesc empty_program; ir::Graph empty_graph(empty_program); @@ -47,19 +47,20 @@ TEST(CinnCacheKeyTest, TestAsUnorderedKey) { DDim ddim = paddle::framework::make_ddim({1, 2, 3}); std::map feed_shapes = {{"X", ddim}}; - CinnCacheKey cache_key0(empty_graph, feed_tensors, "x86"); - CinnCacheKey cache_key1(empty_graph, feed_shapes, "x86"); + CinnCacheKeyByStructure cache_key0(empty_graph, feed_tensors, "x86"); + CinnCacheKeyByStructure cache_key1(empty_graph, feed_shapes, "x86"); EXPECT_EQ(cache_key0, cache_key1); - CinnCacheKey cache_key2(graph, feed_shapes, "x86"); - CinnCacheKey cache_key3(graph, feed_shapes, "nvgpu"); - CinnCacheKey cache_key4(graph, feed_tensors, "nvgpu"); + CinnCacheKeyByStructure cache_key2(graph, feed_shapes, "x86"); + CinnCacheKeyByStructure cache_key3(graph, feed_shapes, "nvgpu"); + CinnCacheKeyByStructure cache_key4(graph, feed_tensors, "nvgpu"); EXPECT_NE(cache_key2, cache_key3); EXPECT_EQ(cache_key3, cache_key4); - CinnCacheKey cache_key5(empty_graph, - std::map(), "unk"); - CinnCacheKey cache_key6(empty_graph, std::map(), "unk"); + CinnCacheKeyByStructure cache_key5( + empty_graph, std::map(), "unk"); + CinnCacheKeyByStructure cache_key6(empty_graph, std::map(), + "unk"); EXPECT_EQ(cache_key5, cache_key6); EXPECT_NE(cache_key1, cache_key3); @@ -98,6 +99,107 @@ TEST(CinnCacheKeyTest, TestAsUnorderedKey) { EXPECT_EQ(test_set.find(cache_key6), test_set.end()); } +TEST(CinnCacheKeyTest, TestAsUnorderedKeyByAddress) { + std::unordered_set test_set; + + ProgramDesc empty_program; + ir::Graph empty_graph(empty_program); + + ProgramDesc program; + auto *global_block = program.MutableBlock(0); + auto *x = global_block->Var("X"); + x->SetType(proto::VarType::LOD_TENSOR); + ir::Graph graph(program); + + LoDTensor tensor; + tensor.Resize({1, 2, 3}); + const LoDTensor *tensor_pointer = &tensor; + std::map feed_tensors = { + {"X", tensor_pointer}}; + + DDim ddim = paddle::framework::make_ddim({1, 2, 3}); + std::map feed_shapes = {{"X", ddim}}; + + CinnCacheKeyByAddress cache_key0(empty_graph, feed_tensors, "x86"); + CinnCacheKeyByAddress cache_key1(empty_graph, feed_shapes, "x86"); + EXPECT_EQ(cache_key0, cache_key1); + + CinnCacheKeyByAddress cache_key2(graph, feed_shapes, "x86"); + CinnCacheKeyByAddress cache_key3(graph, feed_shapes, "nvgpu"); + CinnCacheKeyByAddress cache_key4(graph, feed_tensors, "nvgpu"); + EXPECT_NE(cache_key2, cache_key3); + EXPECT_EQ(cache_key3, cache_key4); + + CinnCacheKeyByAddress cache_key5( + empty_graph, std::map(), "unk"); + CinnCacheKeyByAddress cache_key6(empty_graph, std::map(), + "unk"); + EXPECT_EQ(cache_key5, cache_key6); + + EXPECT_NE(cache_key1, cache_key3); + EXPECT_NE(cache_key4, cache_key2); + + EXPECT_NE(cache_key3, cache_key5); + EXPECT_NE(cache_key6, cache_key4); + + EXPECT_NE(cache_key5, cache_key1); + EXPECT_NE(cache_key2, cache_key6); + + test_set.insert(cache_key0); + test_set.insert(cache_key1); + test_set.insert(cache_key3); + test_set.insert(cache_key4); + test_set.insert(cache_key5); + test_set.insert(cache_key6); + EXPECT_EQ(test_set.size(), 3U); + + auto iter = test_set.find(cache_key0); + EXPECT_NE(iter, test_set.end()); + test_set.erase(iter); + EXPECT_EQ(test_set.size(), 2U); + EXPECT_EQ(test_set.find(cache_key1), test_set.end()); + + iter = test_set.find(cache_key3); + EXPECT_NE(iter, test_set.end()); + test_set.erase(iter); + EXPECT_EQ(test_set.size(), 1U); + EXPECT_EQ(test_set.find(cache_key4), test_set.end()); + + iter = test_set.find(cache_key5); + EXPECT_NE(iter, test_set.end()); + test_set.erase(iter); + EXPECT_EQ(test_set.size(), 0U); + EXPECT_EQ(test_set.find(cache_key6), test_set.end()); +} + +TEST(CinnCacheKeyTest, TestSameGraph) { + ProgramDesc program1; + auto *global_block1 = program1.MutableBlock(0); + auto *x1 = global_block1->Var("X"); + x1->SetType(proto::VarType::LOD_TENSOR); + ir::Graph graph1(program1); + + ProgramDesc program2; + auto *global_block2 = program2.MutableBlock(0); + auto *x2 = global_block2->Var("X"); + x2->SetType(proto::VarType::LOD_TENSOR); + ir::Graph graph2(program2); + + LoDTensor tensor; + tensor.Resize({1, 2, 3}); + const LoDTensor *tensor_pointer = &tensor; + std::map feed_tensors = { + {"X", tensor_pointer}}; + + CinnCacheKeyByAddress cache_key_by_address1(graph1, feed_tensors, "x86"); + CinnCacheKeyByAddress cache_key_by_address2(graph2, feed_tensors, "x86"); + EXPECT_NE(cache_key_by_address1, cache_key_by_address2); + + CinnCacheKeyByStructure cache_key_by_struct1(graph1, feed_tensors, "x86"); + CinnCacheKeyByStructure cache_key_by_struct2(graph2, feed_tensors, "x86"); + EXPECT_EQ(cache_key_by_struct1, cache_key_by_struct2); +} + } // namespace paddle2cinn } // namespace framework } // namespace paddle diff --git a/paddle/fluid/framework/paddle2cinn/cinn_compiler.cc b/paddle/fluid/framework/paddle2cinn/cinn_compiler.cc index 131a6a09e8f7670026fabef158bf5ebee9fe0c60..54167d95899d6fab81a8657b167012b47bf950ea 100644 --- a/paddle/fluid/framework/paddle2cinn/cinn_compiler.cc +++ b/paddle/fluid/framework/paddle2cinn/cinn_compiler.cc @@ -69,23 +69,41 @@ const CinnCompiledObject& CinnCompiler::Compile( const std::map& input_tensors, const Target& target, void* stream) { VLOG(1) << "-- The graph to be compiled is:\n" << VizGraph(graph); - CinnCacheKey cur_key(graph, input_tensors, target.arch_str()); + CinnCacheKeyByAddress cur_key_by_address(graph, input_tensors, + target.arch_str()); + CinnCacheKeyByStructure cur_key_by_struct; + bool exist = false; { AutoRDLock r_guard{&rwlock_}; - exist = cache_.count(cur_key) != 0; + exist = cache_by_address_.count(cur_key_by_address) != 0; + // if cannot find graph by address, checkout whether the graph structure + // have been stored in cache. + if (!exist) { + // generate the structure cache key + cur_key_by_struct.SetKey(graph, input_tensors, target.arch_str()); + + // if the graph structure can be found, storing the graph address in + // cache for next query. + if (cache_by_struct_.count(cur_key_by_struct) != 0) { + exist = true; + cache_by_address_[cur_key_by_address] = + cache_by_struct_.at(cur_key_by_struct).get(); + } + } } if (!exist) { std::int64_t compiled_num = real_compiled_num_.fetch_add(1); auto compiled_res = CompileGraph(graph, input_tensors, target, compiled_num, stream); AutoWRLock w_guard{&rwlock_}; - if (!cache_.count(cur_key)) { - cache_[cur_key] = std::move(compiled_res); + if (!cache_by_struct_.count(cur_key_by_struct)) { + cache_by_address_[cur_key_by_address] = compiled_res.get(); + cache_by_struct_[cur_key_by_struct] = std::move(compiled_res); } } AutoRDLock guard{&rwlock_}; - const auto& cached_boj = *cache_[cur_key]; + const auto& cached_boj = *cache_by_address_[cur_key_by_address]; return cached_boj; } @@ -182,7 +200,8 @@ void CinnCompiler::Clear() { { AutoWRLock guard{&rwlock_}; graphs_.clear(); - cache_.clear(); + cache_by_address_.clear(); + cache_by_struct_.clear(); } real_compiled_num_.store(0); } diff --git a/paddle/fluid/framework/paddle2cinn/cinn_compiler.h b/paddle/fluid/framework/paddle2cinn/cinn_compiler.h index d75279cfe96b869dc143663eebd5c6703d85fafb..3bc60e555570708f2709b0cdeaa4c35abc49a08b 100644 --- a/paddle/fluid/framework/paddle2cinn/cinn_compiler.h +++ b/paddle/fluid/framework/paddle2cinn/cinn_compiler.h @@ -95,9 +95,12 @@ class CinnCompiler { void* stream = nullptr) const; std::unordered_map> graphs_; - std::unordered_map, + std::unordered_map - cache_; + cache_by_address_; + std::unordered_map, CinnCacheKey::Hash> + cache_by_struct_; std::atomic_int64_t real_compiled_num_{0}; mutable RWLock rwlock_;