[ADD] Neura demo python script

.github/workflows/cmake.yml

@@ -31,6 +31,16 @@ jobs:
         sudo apt-get -y install llvm-12 llvm-12-dev llvm-12-tools clang-12
         sudo apt-get -y install build-essential
 
+    - name: Set up Python
+      uses: actions/setup-python@v4
+      with:
+        python-version: '3.7.16'
+
+    - name: Install Python Dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install eventlet pandas matplotlib
+
     - name: Configure CMake
       # Configure CMake in a 'build' subdirectory. `CMAKE_BUILD_TYPE` is only required if you are using a single-configuration generator such as make.
       # See https://cmake.org/cmake/help/latest/variable/CMAKE_BUILD_TYPE.html?highlight=cmake_build_type
@@ -105,3 +115,8 @@ jobs:
         sh compile.sh
         sh run.sh
         sh verify.sh
+
+    - name: Test expandable automatic script
+      working-directory: ${{github.workspace}}/tools/expandable
+      run: |
+        ./demo.sh --test y

.gitignore

@@ -9,3 +9,5 @@ build*/
 *.ll
 *.dot
 *.bc
+.vscode/
+.venv/

src/DFG.cpp

@@ -33,8 +33,13 @@ DFG::DFG(Function& t_F, list<Loop*>* t_loops, bool t_targetFunction,
   bool needsCycleCalculation = false;
   for (auto strategy : *t_fusionStrategy) {
     if (strategy == "default_heterogeneous") {
-      combineMulAdd("CoT");
-      combinePhiAdd("BrT");
+      combine("phi", "add", "Ctrl");
+      combine("phi", "fadd", "Ctrl");
+      combine("fcmp", "select", "Ctrl");
+      combine("icmp", "select", "Ctrl");
+      combine("icmp", "br", "Ctrl");
+      combine("fcmp", "br", "Ctrl");
+      tuneForPattern();
       needsCycleCalculation = true;
     }
     else if (strategy == "nonlinear") {

src/Mapper.cpp

@@ -663,9 +663,13 @@ void Mapper::showUtilization(CGRA* t_cgra, DFG* t_dfg, int t_II,
     total_active_tiles += 1;
   }
   float avg_tile_overall_utilization = 0.0;
+  float max_tile_overall_utilization = 0.0;
   float avg_tile_fu_utilization = 0.0;
   float avg_tile_xbar_utilization = 0.0;
   for (int tile = 0; tile < t_cgra->getFUCount(); ++tile) {
+    if (max_tile_overall_utilization < tile_overall_utilization[tile]) {
+      max_tile_overall_utilization = tile_overall_utilization[tile];
+    }
     avg_tile_overall_utilization += tile_overall_utilization[tile];
     avg_tile_fu_utilization += tile_fu_utilization[tile];
     avg_tile_xbar_utilization += tile_xbar_utilization[tile];
@@ -674,8 +678,10 @@ void Mapper::showUtilization(CGRA* t_cgra, DFG* t_dfg, int t_II,
   avg_tile_overall_utilization /= total_active_tiles;
   avg_tile_fu_utilization /= total_active_tiles;
   avg_tile_xbar_utilization /= total_active_tiles;
+  //max_tile_overall_utilization /= total_active_tiles;
 
-  cout << "tile avg fu utilization: " << avg_tile_fu_utilization*100 << "%; avg xbar utilization: " << avg_tile_xbar_utilization*100 << "%; avg overall utilization: " << avg_tile_overall_utilization*100 << "%" << endl;
+  cout << "tile avg fu utilization: " << avg_tile_fu_utilization*100 << "%; avg xbar utilization: " << avg_tile_xbar_utilization*100 << "%; avg overall utilization: " << avg_tile_overall_utilization*t_II*100 << "%" << endl;
+  cout << "max overall utilization: " << max_tile_overall_utilization*t_II*100 << "%" << endl;
 
   // Collects the histogram of tiles' utilization.
   // Histogram for the number of tiles that have utilization of 0%.
@@ -1006,6 +1012,7 @@ void Mapper::showSchedule(CGRA* t_cgra, DFG* t_dfg, int t_II,
   cout<<"[Mapping II: "<<t_II<<"]"<<endl;
 
   if (t_parameterizableCGRA) {
+    // TODO: make it clean
     jsonTilesLinks["tiles"] = jsonTiles;
     jsonTilesLinks["links"] = jsonLinks;
     json jsonMap(jsonTilesLinks);

src/mapperPass.cpp

@@ -189,7 +189,7 @@ namespace {
           // Exclusive: Multi-cyce operations occupy tiles exclusively. Other operations can be mapped onto this tile only if the multi-cycle operation finishs its computation.
           // Distributed: Multi-cycle operations are splitted into multiple single-cycle operations and each of which can be mapped onto a tile.
           // Inclusive: Multi-cycle operations' execution can overlap with other operations on the same tile.
-          // Note that 
+          // Note that
           assert(multiCycleStrategy.compare("exclusive") == 0 or
                  multiCycleStrategy.compare("distributed") == 0 or
                  multiCycleStrategy.compare("inclusive") == 0);
@@ -339,8 +339,8 @@ namespace {
       else {
         mapper->showSchedule(cgra, dfg, II, isStaticElasticCGRA, parameterizableCGRA);
         // cout << "==================================\n";
-        cout << "[show opcode count]\n";
-        dfg->showOpcodeDistribution();
+        // cout << "[show opcode count]\n";
+        // dfg->showOpcodeDistribution();
         cout << "[Mapping Success]\n";
         cout << "==================================\n";
         if (enableExpandableMapping) {
@@ -409,11 +409,11 @@ namespace {
      */
      bool canMap(CGRA* t_cgra, DFG* t_dfg) {
       std::set<std::string> missing_fus;
-    
+
       for (auto it = t_dfg->nodes.begin(); it != t_dfg->nodes.end(); ++it) {
         DFGNode* node = *it;
         bool nodeSupported = false;
-    
+
         for (int i = 0; i < t_cgra->getRows() && !nodeSupported; ++i) {
           for (int j = 0; j < t_cgra->getColumns(); ++j) {
             CGRANode* fu = t_cgra->nodes[i][j];
@@ -423,12 +423,12 @@ namespace {
             }
           }
         }
-    
+
         if (!nodeSupported) {
           missing_fus.insert(node->getOpcodeName());
         }
       }
-    
+
       if (!missing_fus.empty()) {
         std::cout << "[canMap] Missing functional units: ";
         for (const auto& op : missing_fus) {
@@ -437,10 +437,10 @@ namespace {
         std::cout << std::endl;
         return false;
       }
-    
+
       return true;
     }
-    
+
   };
 }
 

test/compile.sh

@@ -1,2 +1,3 @@
-clang-12 -emit-llvm -fno-unroll-loops -O3 -o kernel.bc -c kernel.cpp
-#llvm-dis fir.bc -o fir.ll
+clang-12 -emit-llvm -fno-unroll-loops -O0 -o kernel.bc -c kernel.cpp
+llvm-dis-12 kernel.bc -o O0kernel.ll
+#clang-12 -emit-llvm -fno-unroll-loops -mllvm -force-vector-width=4 -O3 -o kernel.bc -c ./_matmul/src/matmul.c

test/dot.sh

@@ -1,2 +1 @@
-dot -Tpng _Z6kernelPfS_S_.dot -o kernel.png
-
+dot -Tpng _Z6kernelPfS_S_.dot -o kernel.png

test/kernels/conv/conv.c

@@ -76,7 +76,7 @@ int kernel(int ni, int nj, int nk,
   for (x = 0; x < total; x++) {
     i = x / NJ;
     j = x % NJ;
-    out += A[i][j] * B[i][j];
+    out += A  [i][j] * B[i][j];
   }
 
   /*

test/kernels/latnrm/latnrm.c

@@ -45,17 +45,17 @@ void kernel(float input, float *output, float coefficient[16],
   top = input;
   q_coef = coefficient[0];
   // #pragma clang loop unroll_count(4)
-  for (i = 0; i < ORDER; i++) {
-    k_coef = coefficient[2*i];
-    left = top;
-    right = internal_state[i];
-    internal_state[i] = bottom;
-    top = q_coef * left - k_coef * right;
-    bottom = q_coef * right + k_coef * left;
-    q_coef = coefficient[2*i+1];
-  }
-  internal_state[i++] = bottom;
-  internal_state[i] = top;
+  // for (i = 0; i < ORDER; i++) {
+  //   k_coef = coefficient[2*i];
+  //   left = top;
+  //   right = internal_state[i];
+  //   internal_state[i] = bottom;
+  //   top = q_coef * left - k_coef * right;
+  //   bottom = q_coef * right + k_coef * left;
+  //   q_coef = coefficient[2*i+1];
+  // }
+  // internal_state[i++] = bottom;
+  // internal_state[i] = top;
 
   sum = internal_state[1] * q_coef;
 

test/kernels/mvt/mvt.c

@@ -78,8 +78,8 @@ void kernel(int n,
 {
   int i, j;
 
-#pragma scop
-  #pragma clang loop unroll_count(1)
+// #pragma scop
+  // #pragma clang loop unroll_count(1)
   for (j = 0; j < N; j++) {
     // #pragma clang loop unroll_count(1) vectorize(disable)
     //#pragma clang loop unroll_count(1) vectorize_width(4)
@@ -88,7 +88,7 @@ void kernel(int n,
       x2[i] = x2[i] + A[j][i] * y_2[j];
     }
   }
-#pragma endscop
+// #pragma endscop
 }
 
 

test/kernels/relu+histogram/compile.sh

@@ -0,0 +1,4 @@
+clang-12 -emit-llvm -O3 -fno-unroll-loops -o kernel.bc -c relu.c
+llvm-dis-12 kernel.bc -o kernel.ll
+opt-12 --loop-unroll --unroll-count=4 kernel.bc -o kernel_unroll.bc
+llvm-dis-12 kernel_unroll.bc -o kernel_unroll.ll

test/kernels/relu+histogram/param.json

@@ -0,0 +1,19 @@
+{
+  "kernel"                : "kernel",
+  "targetFunction"        : false,
+  "targetNested"          : false,
+  "targetLoopsID"         : [1],
+  "doCGRAMapping"         : false,
+  "row"                   : 4,
+  "column"                : 4,
+  "diagonalVectorization" : true,
+  "fusionStrategy"        : [],
+  "isTrimmedDemo"         : true,
+  "heuristicMapping"      : false,
+  "bypassConstraint"      : 4,
+  "isStaticElasticCGRA"   : false,
+  "precisionAware"        : false,
+  "ctrlMemConstraint"     : 200,
+  "regConstraint"         : 8
+}
+