From 980d3c23c9673637004d92e6b5b55506b83c520f Mon Sep 17 00:00:00 2001 From: jwu773 Date: Wed, 4 Feb 2026 17:12:26 +0800 Subject: [PATCH 1/2] AI compiler assigmnets my first commit --- .idea/.gitignore | 10 ++ .idea/TinyInfiniTensor.iml | 9 + .idea/misc.xml | 6 + .idea/modules.xml | 8 + .idea/vcs.xml | 8 + include/core/allocator.h | 2 +- include/core/graph.h | 1 + src/core/allocator.cc | 50 +++++- src/core/graph.cc | 165 ++++++++++++++++++ src/operators/concat.cc | 5 +- src/operators/matmul.cc | 22 ++- src/operators/transpose.cc | 7 +- src/operators/unary.cc | 14 +- src/utils/operator_utils.cc | 28 ++- .../nativecpu/test_nativecpu_concat.cc | 3 +- 15 files changed, 327 insertions(+), 11 deletions(-) create mode 100644 .idea/.gitignore create mode 100644 .idea/TinyInfiniTensor.iml create mode 100644 .idea/misc.xml create mode 100644 .idea/modules.xml create mode 100644 .idea/vcs.xml diff --git a/.idea/.gitignore b/.idea/.gitignore new file mode 100644 index 0000000..ab1f416 --- /dev/null +++ b/.idea/.gitignore @@ -0,0 +1,10 @@ +# Default ignored files +/shelf/ +/workspace.xml +# Ignored default folder with query files +/queries/ +# Datasource local storage ignored files +/dataSources/ +/dataSources.local.xml +# Editor-based HTTP Client requests +/httpRequests/ diff --git a/.idea/TinyInfiniTensor.iml b/.idea/TinyInfiniTensor.iml new file mode 100644 index 0000000..d6ebd48 --- /dev/null +++ b/.idea/TinyInfiniTensor.iml @@ -0,0 +1,9 @@ + + + + + + + + + \ No newline at end of file diff --git a/.idea/misc.xml b/.idea/misc.xml new file mode 100644 index 0000000..07115cd --- /dev/null +++ b/.idea/misc.xml @@ -0,0 +1,6 @@ + + + + + + \ No newline at end of file diff --git a/.idea/modules.xml b/.idea/modules.xml new file mode 100644 index 0000000..62de32f --- /dev/null +++ b/.idea/modules.xml @@ -0,0 +1,8 @@ + + + + + + + + \ No newline at end of file diff --git a/.idea/vcs.xml b/.idea/vcs.xml new file mode 100644 index 0000000..67a1dcf --- /dev/null +++ b/.idea/vcs.xml @@ -0,0 +1,8 @@ + + + + + + + + \ No newline at end of file diff --git a/include/core/allocator.h b/include/core/allocator.h index 002601d..f3ff75a 100644 --- a/include/core/allocator.h +++ b/include/core/allocator.h @@ -27,7 +27,7 @@ namespace infini { // TODO：可能需要设计一个数据结构来存储free block，以便于管理和合并 // HINT: 可以使用一个 map 来存储 free block，key 为 block 的起始/结尾地址，value 为 block 的大小 // =================================== 作业 =================================== - + std::map freeByStart; public: Allocator(Runtime runtime); diff --git a/include/core/graph.h b/include/core/graph.h index c45580c..4c534c8 100644 --- a/include/core/graph.h +++ b/include/core/graph.h @@ -50,6 +50,7 @@ namespace infini * so the topological sorting fails. */ bool topo_sort(); + void reconstruct(Operator &op1, Operator &op2, Operator &op3); void optimize(); diff --git a/src/core/allocator.cc b/src/core/allocator.cc index ff593ae..47dd942 100644 --- a/src/core/allocator.cc +++ b/src/core/allocator.cc @@ -32,8 +32,32 @@ namespace infini // =================================== 作业 =================================== // TODO: 设计一个算法来分配内存，返回起始地址偏移量 // =================================== 作业 =================================== + //std::cout<<" ******************* map size "<second >= size){ + freeBlkOk = true; + if(it->second > size) + freeByStart.insert({it->first + size, it->second - size}); + break; + } + } + + if(!freeBlkOk){ + it--; + size_t moreMem = size - it->second; + peak += moreMem; + } + freeByStart.erase(it); + return it->first; - return 0; } void Allocator::free(size_t addr, size_t size) @@ -44,6 +68,30 @@ namespace infini // =================================== 作业 =================================== // TODO: 设计一个算法来回收内存 // =================================== 作业 =================================== + auto it = freeByStart.begin(); + int flag = 0; + for(; it != freeByStart.end(); it++){ + if(it->first + it->second == addr){ + flag = -1; + break; + } + else if(it->first == addr + size){ + flag = 1; + break; + } + } + if(flag == -1){ + freeByStart.insert({it->first, it->second + size}); + freeByStart.erase(it); + } + else if(flag == 0){ + freeByStart.insert({addr, size}); + } + else if(flag == 1){ + freeByStart.insert({addr, size + it->second}); + freeByStart.erase(it); + } + used -= size; } void *Allocator::getPtr() diff --git a/src/core/graph.cc b/src/core/graph.cc index 3a90637..a873de4 100644 --- a/src/core/graph.cc +++ b/src/core/graph.cc @@ -2,6 +2,10 @@ #include #include #include +#include "operators/transpose.h" +#include "operators/matmul.h" +#include +#include namespace infini { @@ -98,6 +102,49 @@ namespace infini return this->sorted = true; } + bool transposeOpsCancel(vector a, vector b){ + if(a.size() != b.size()) + return false; + for(int i = 0; i < (int)a.size(); i++){ + if(b[a[i]] != i) + return false; + } + return true; + } + + bool transOpCanIntegrateToMatmul(vector perm){ + int size = perm.size(); + for(int i = 0; i < size - 2; i++){ + if(perm[i] != i) + return false; + } + if(perm[size - 2] != size - 1 || perm[size - 1] != size - 2) + return false; + else + return true; + } + + void GraphObj::reconstruct(Operator &op1, Operator &op2, Operator &op3){ + Tensor input = op1->getInputs(0); + if(input) {input->addTarget(op3);} + if(op2 == nullptr){ + op3->replaceInput(op1->getOutput(), input); + op1->removeSuccessors(op3); + op3->removePredecessors(op1); + op1->getOutput()->removeTarget(op3); + } + else{ + op3->replaceInput(op2->getOutput(), input); + op3->removePredecessors(op2); + op1->removeSuccessors(op2); + op1->getOutput()->removeTarget(op2); + } + for(auto &pred: op1->getPredecessors()){ + op3->addPredecessors(pred); + pred->addSuccessors(op3); + } + } + void GraphObj::optimize() { // =================================== 作业 =================================== @@ -106,6 +153,93 @@ namespace infini // 1. 去除冗余的算子（例如，两个相邻的算子都是 transpose 算子，且做的是相反的操作，可以将其全部删除） // 2. 合并算子（例如，矩阵乘算子中含有属性transA、transB，如果其输入存在transpose，且对最后两个维度做交换，就可以将transpose融入到矩阵乘算子的属性中去） // =================================== 作业 =================================== + std::unordered_set toDelete; + std::shared_ptr emptyPtr; + + for(auto &op: ops){ + if(toDelete.find(op.get()) != toDelete.end()){ + continue; + } + OpType opType = op->getOpType(); + //std::cout<<"..................... .....................OP................ ...................\n"; + //op->print(); + //std::cout<<"...................... ....................Op................. ....................\n"; + if(opType == OpType::Concat){ //case1:Concat + if(op->numInputs() == 1){ + toDelete.insert(op.get()); + for(auto &succ: op->getSuccessors()){ + reconstruct(op, emptyPtr, succ); + } + } + } + else if(opType == OpType::Transpose){ //case2:Transpose + TransposeObj* transOp = dynamic_cast(op.get()); + //std::cout<<"......2....OP is Transpose........"<getInputs(0); + for(auto &succ: op->getSuccessors()){ + if(succ->getOpType() == OpType::Transpose){ //case2.1:Transpose-->Transpose + TransposeObj* transSucc = dynamic_cast(succ.get()); + //std::cout<<".......2.1...successor is Transpose op.......\n"; + //transSucc->print(); + bool cancelOut = false; + if(transOp && transSucc){ + cancelOut = transposeOpsCancel(transOp->getPermute(), transSucc->getPermute()); + } + //std::cout<<"......2.1...successor is Transpose :they cancel? "<getOutput()); + for(auto &succ_succ: succ->getSuccessors()){ + reconstruct(op, succ, succ_succ); + } + if(op->getSuccessors().size() == 0){ + toDelete.insert(op.get()); + removeTensor(op->getOutput()); + if(input) {input->removeTarget(op);} + for(auto &pred: op->getPredecessors()) + pred->removeSuccessors(op); + } + } + } + else if(succ->getOpType() == OpType::MatMul){ //case2.2:Transpose-->MatMul + bool ok = transOpCanIntegrateToMatmul(transOp->getPermute()); + //std::cout<<"------2.2---successor is Matmul op, can integrate? "<(succ.get()); + if (matmulSucc){ + if(succ->getInputs(0)->getGuid() == op->getOutput()->getGuid()) + matmulSucc->setTransA(!(matmulSucc->getTransA())); + else + matmulSucc->setTransB(!(matmulSucc->getTransB())); + } + reconstruct(op, emptyPtr, succ); + if(op->getSuccessors().size() == 0){ + toDelete.insert(op.get()); + removeTensor(op->getOutput()); + if(input) {input->removeTarget(op);} + for(auto &pred: op->getPredecessors()) + pred->removeSuccessors(op); + } + } + } + } + } + } + //delete redundant operators + //std::cout<<" --------ooo--------- before optimize "<print(); + }**/ } Tensor GraphObj::getTensor(int fuid) const @@ -152,7 +286,38 @@ namespace infini // TODO：利用 allocator 给计算图分配内存 // HINT: 获取分配好的内存指针后，可以调用 tensor 的 setDataBlob 函数给 tensor 绑定内存 // =================================== 作业 =================================== + std::unordered_map tensorOffset; + for(auto &tensor: tensors){ + if(!tensor->getSource()) + tensorOffset.insert({tensor->getFuid(), allocator.alloc(tensor->getBytes())}); + } + std::unordered_set graphOutputsFuid; + for(auto &output: this->getOutputs()){ + graphOutputsFuid.insert(output->getFuid()); + } + for(auto &op: ops){ + for(auto &output: op->getOutputs()){ + tensorOffset.insert({output->getFuid(), allocator.alloc(output->getBytes())}); + } + for(auto input: op->getInputs()){ + if(input->getTargets().size() == 0){ + int fuid = input->getFuid(); + if(graphOutputsFuid.find(fuid) == graphOutputsFuid.end()){ + allocator.free(tensorOffset[fuid], input->getBytes()); + } + } + } + } + void *pointer = allocator.getPtr(); + for(auto &tensor: tensors){ + int fuid = tensor->getFuid(); + if(tensorOffset.count(fuid)){ + size_t offset = tensorOffset[fuid]; + auto blob = make_ref(runtime, pointer + offset); + tensor->setDataBlob(blob); + } + } allocator.info(); } diff --git a/src/operators/concat.cc b/src/operators/concat.cc index d196330..7408414 100644 --- a/src/operators/concat.cc +++ b/src/operators/concat.cc @@ -17,7 +17,10 @@ optional> ConcatObj::inferShape(const TensorVec &inputs) { // TODO：修改 dims，返回正确的 concat 后的 shape // REF: https://onnx.ai/onnx/operators/onnx__Concat.html#concat-13 // =================================== 作业 =================================== - + int n = inputs.size(); + for(int i = 1; i < n; i++){ + dims[dim] += inputs[i]->getDims()[dim]; + } return {{dims}}; } diff --git a/src/operators/matmul.cc b/src/operators/matmul.cc index 7a16ca2..b57063a 100644 --- a/src/operators/matmul.cc +++ b/src/operators/matmul.cc @@ -27,7 +27,27 @@ namespace infini // TODO：返回经过 matmul 操作后的 shape // REF: https://github.com/onnx/onnx/blob/main/docs/Operators.md#gemm // =================================== 作业 =================================== - return std::nullopt; + const auto A = inputs[0]; + auto a_dim = A->getDims(); + const auto B = inputs[1]; + auto b_dim = B->getDims(); + int size = a_dim.size(); + Shape res = a_dim; + //multi-dimensional broadcasting except for the last two dimensions + for(int i = 0; i < size - 2; i++){ + if(a_dim[i] < b_dim[i]) + res[i] = b_dim[i]; + } + + if(transA) + res[size - 2] = a_dim[size - 1]; + if(transB) + res[size - 1] = b_dim[size - 2]; + else + res[size - 1] = b_dim[size - 1]; + + std::vector vec {res}; + return vec; } } // namespace infini \ No newline at end of file diff --git a/src/operators/transpose.cc b/src/operators/transpose.cc index faab2b6..8379a9c 100644 --- a/src/operators/transpose.cc +++ b/src/operators/transpose.cc @@ -33,8 +33,13 @@ namespace infini // TODO：修改 output_dim，返回正确的 transpose 后的 shape // REF: https://onnx.ai/onnx/operators/onnx__Transpose.html#transpose-21 // =================================== 作业 =================================== + for (int i = 0; i < rank; i++){ + output_dim[i] = input_dim[transposePermute[i]]; + } + std::vector res {output_dim}; + return res; - return std::nullopt; + //return std::nullopt; } std::string TransposeObj::toString() const diff --git a/src/operators/unary.cc b/src/operators/unary.cc index 3daad36..ceb50bc 100644 --- a/src/operators/unary.cc +++ b/src/operators/unary.cc @@ -39,7 +39,9 @@ namespace infini // TODO：返回经过 clip 操作后的 shape // REF: https://onnx.ai/onnx/operators/onnx__Clip.html#clip-13 // =================================== 作业 =================================== - return std::nullopt; + const auto A = inputs[0]; + return {{A->getDims()}}; + //return std::nullopt; } std::string ClipObj::toString() const @@ -66,7 +68,11 @@ namespace infini // REF_FILE: src/core/operator.cc // REF: https://onnx.ai/onnx/operators/onnx__Cast.html#cast-21 // =================================== 作业 =================================== - return {}; + DataType outputType = getOutputDataType(); + int num = inputs.size(); + std::vector res(num, outputType); + return res; + //return {}; } optional> CastObj::inferShape(const TensorVec &inputs) @@ -75,7 +81,9 @@ namespace infini // TODO：返回经过 cast 操作后的 shape // REF: https://onnx.ai/onnx/operators/onnx__Cast.html#cast-21 // =================================== 作业 =================================== - return std::nullopt; + const auto A = inputs[0]; + return {{A->getDims()}}; + //return std::nullopt; } std::string CastObj::toString() const diff --git a/src/utils/operator_utils.cc b/src/utils/operator_utils.cc index edbd2c8..37eab0d 100644 --- a/src/utils/operator_utils.cc +++ b/src/utils/operator_utils.cc @@ -9,8 +9,32 @@ Shape infer_broadcast(const Shape &A, const Shape &B) { // TODO：对 A 和 B 进行双向广播，返回广播后的形状。 // REF: https://github.com/onnx/onnx/blob/main/docs/Broadcasting.md // =================================== 作业 =================================== - - return {}; + Shape res; + long unsigned int i = 0; + if(A.size() > B.size()){ + long unsigned int diff = A.size() - B.size(); + for(; i < diff; i++) + res.push_back(A[i]); + for(; i < A.size(); i++){ + if(A[i] == 1) + res.push_back(B[i-diff]); + else + res.push_back(A[i]); + } + } + else{ + long unsigned int diff = B.size() - A.size(); + for(; i < diff; i++) + res.push_back(B[i]); + for(; i < B.size(); i++){ + if(B[i] == 1) + res.push_back(A[i-diff]); + else + res.push_back(B[i]); + } + } + + return res; } int get_real_axis(const int &axis, const int &rank) { diff --git a/test/kernels/nativecpu/test_nativecpu_concat.cc b/test/kernels/nativecpu/test_nativecpu_concat.cc index fc87fb1..8e23bc9 100644 --- a/test/kernels/nativecpu/test_nativecpu_concat.cc +++ b/test/kernels/nativecpu/test_nativecpu_concat.cc @@ -18,8 +18,9 @@ TEST(Concat, NativeCpu) { t1->setData(IncrementalGenerator()); t2->setData(OneGenerator()); t3->setData(OneGenerator()); - runtime->run(g); + //std::cout<<" --------------###--------------- concat output\n"; + op->getOutput()->printData(); EXPECT_TRUE(op->getOutput()->equalData( vector{0, 1, 2, 1, 1, 1, 3, 4, 5, 1, 1, 1, 6, 7, 8, 1, 1, 1, 9, 10, 11, 1, 1, 1})); From bf78e135993e3a5d5331670a7e996707bdfbd14f Mon Sep 17 00:00:00 2001 From: jwu773 Date: Wed, 4 Feb 2026 17:22:52 +0800 Subject: [PATCH 2/2] compilerAssignments --- src/core/graph.cc | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/src/core/graph.cc b/src/core/graph.cc index a873de4..dca2eef 100644 --- a/src/core/graph.cc +++ b/src/core/graph.cc @@ -161,9 +161,9 @@ namespace infini continue; } OpType opType = op->getOpType(); - //std::cout<<"..................... .....................OP................ ...................\n"; + //std::cout<<"..................... .....................OP start................ ...................\n"; //op->print(); - //std::cout<<"...................... ....................Op................. ....................\n"; + //std::cout<<"...................... ....................Op end................. ....................\n"; if(opType == OpType::Concat){ //case1:Concat if(op->numInputs() == 1){ toDelete.insert(op.get());