[BenchmarkGen] support generating the last dense layer, and impl corresponding affine lowering (#15)

README.md

@@ -34,10 +34,14 @@ After the installation and test successfully completed, you should be able to pl
 ```sh
 $ export PATH=$SCALEHLS_DIR/build/bin:$PATH
 $ cd $SCALEHLS_DIR
+$
+$ benchmark-gen -type "cnn" -config "$SCALEHLS_DIR/config/cnn-config.ini" -number 1
+$ scalehls-opt -hlskernel-to-affine test/Conversion/HLSKernelToAffine/test_*.mlir
+$
 $ scalehls-opt -convert-to-hlscpp test/Conversion/ConvertToHLSCpp/test_*.mlir
 $ scalehls-opt -convert-to-hlscpp test/EmitHLSCpp/test_*.mlir | scalehls-translate -emit-hlscpp
+$
 $ scalehls-opt -qor-estimation test/Analysis/QoREstimation/test_for.mlir
-$ scalehls-opt -hlskernel-to-affine test/Conversion/HLSKernelToAffine/test_*.mlir
 ```
 
 ## References

config/cnn-config.ini

@@ -1,13 +1,13 @@
 [config]
 # input and output information
 inputChannel=3
-inputHeight=224
-inputWidth=224
-outputChannel=1000
+inputHeight=32
+inputWidth=32
+outputChannel=10
 
 # Generation rules
 batchSize=1
-minChannel=64
-maxChannel=512
-poolingNumber=5
+minChannel=8
+maxChannel=64
+poolingNumber=3
 bypassNumber=0

include/Dialect/HLSKernel/CNNOps.td

@@ -10,8 +10,8 @@ def DenseOp : HLSKernelOp<"dense", [HLSKernelOpInterface]> {
   let description = [{
     Dense layer (fully-connected layer).
 
-    I: (N, C)
-    K: (F, C)
+    I: (N, C) or (N, C, H, W)
+    K: (F, C) or (F, C, H, W)
     B: (F)
     O: (N, F)
   }];

lib/Conversion/HLSKernelToAffine/HLSKernelToAffine.cpp

@@ -135,7 +135,7 @@ bool HLSKernelVisitor::visitOp(DenseOp op) {
   auto B = op.getOperand(2);
   auto O = op.getOperand(3);
 
-  auto KShape = K.getType().cast<MemRefType>().getShape();
+  auto IShape = I.getType().cast<MemRefType>().getShape();
   auto OShape = O.getType().cast<MemRefType>().getShape();
 
   // Set insertion point of builder.
@@ -145,24 +145,42 @@ bool HLSKernelVisitor::visitOp(DenseOp op) {
   auto n = createLoop(0, OShape[0]);
 
   // Create output channel loop.
-  auto f = createLoop(0, KShape[0]);
+  auto f = createLoop(0, OShape[1]);
 
   // Load bias into O array.
   auto bias = createLoad(B, {f});
   createStore(bias, O, {n, f});
 
   // Create input channel loop.
-  auto c = createLoop(0, KShape[1]);
+  auto c = createLoop(0, IShape[1]);
 
-  // Fetch feature map, kernel and carry out multiplication.
-  auto fmap = createLoad(I, {n, c});
-  auto kernel = createLoad(K, {f, c});
-  auto mult = createBinaryOp<mlir::MulFOp>(fmap, kernel);
+  if (IShape.size() == 2) {
+    // Fetch feature map, kernel and carry out multiplication.
+    auto fmap = createLoad(I, {n, c});
+    auto kernel = createLoad(K, {f, c});
+    auto mult = createBinaryOp<mlir::MulFOp>(fmap, kernel);
 
-  // Fetch partial result and carry out accumulation.
-  auto partial = createLoad(O, {n, f});
-  auto accum = createBinaryOp<mlir::AddFOp>(partial, mult);
-  createStore(accum, O, {n, f});
+    // Fetch partial result and carry out accumulation.
+    auto partial = createLoad(O, {n, f});
+    auto accum = createBinaryOp<mlir::AddFOp>(partial, mult);
+    createStore(accum, O, {n, f});
+  } else {
+    // Create kernel height loop.
+    auto r = createLoop(0, IShape[2]);
+
+    // Create kernel width loop.
+    auto s = createLoop(0, IShape[3]);
+
+    // Fetch feature map, kernel and carry out multiplication.
+    auto fmap = createLoad(I, {n, c, r, s});
+    auto kernel = createLoad(K, {f, c, r, s});
+    auto mult = createBinaryOp<mlir::MulFOp>(fmap, kernel);
+
+    // Fetch partial result and carry out accumulation.
+    auto partial = createLoad(O, {n, f});
+    auto accum = createBinaryOp<mlir::AddFOp>(partial, mult);
+    createStore(accum, O, {n, f});
+  }
 
   return true;
 }

tools/benchmark-gen/benchmark-gen.cpp

@@ -105,7 +105,7 @@ LogicalResult BenchmarkGenerator::genCNN(INIReader config) {
   SmallVector<mlir::Type, 2> inputTypes;
   inputTypes.push_back(
       getMemType({batchSize, inputChannel, inputHeight, inputWidth}));
-  inputTypes.push_back(getMemType({outputChannel}));
+  inputTypes.push_back(getMemType({batchSize, outputChannel}));
   SmallVector<mlir::Type, 2> outputTypes;
 
   auto func = builder.create<FuncOp>(
@@ -178,10 +178,17 @@ LogicalResult BenchmarkGenerator::genCNN(INIReader config) {
     btmWidth = poolingFlag ? topWidth / 2 : topWidth;
   }
 
-  // TODO: Create the last dense layer.
+  // Create the last dense layer.
+  fmaps.push_back(func.getArgument(1));
+  kernels.push_back(builder.create<mlir::AllocOp>(
+      loc, getMemType({outputChannel, topChannel, topHeight, topWidth})));
+  biases.push_back(
+      builder.create<mlir::AllocOp>(loc, getMemType({outputChannel})));
+
+  builder.create<DenseOp>(loc, *std::prev(fmaps.end(), 2), kernels.back(),
+                          biases.back(), fmaps.back());
 
   builder.create<mlir::ReturnOp>(loc);
-
   os << module << "\n";
   return success();
 }