[cleanup] Remove depth parameter from ConvolutionNode.

artemrakhov-glow · artemrakhov-glow · commit 28bac255ebd8 · 2018-04-12T16:53:49.000-07:00
diff --git a/include/glow/Graph/Graph.h b/include/glow/Graph/Graph.h
@@ -164,8 +164,8 @@ class Function final : public Named {
 
   ConvolutionNode *createConv(llvm::StringRef name, NodeValue input,
                               NodeValue filter, NodeValue bias, TypeRef outTy,
-                              size_t depth, size_t kernel, size_t stride,
-                              size_t pad, size_t group);
+                              size_t kernel, size_t stride, size_t pad,
+                              size_t group);
 
   PoolMaxNode *createPoolMax(llvm::StringRef name, NodeValue input,
                              size_t kernel, size_t stride, size_t pad);
diff --git a/lib/Backends/CPU/Transforms.cpp b/lib/Backends/CPU/Transforms.cpp
@@ -18,7 +18,7 @@ using llvm::isa;
 /// pre-swizzle the data in the weights to make the access pattern more
 /// efficient.
 static Node *optimizeCPUConv(ConvolutionNode *CN, Function *F) {
-  auto depth = CN->getDepth();
+  auto depth = CN->getFilter().dims()[0];
   auto *M = F->getParent();
 
   // The depth dimension must be a multiple of 64 to perform the
@@ -62,7 +62,7 @@ static Node *optimizeCPUConv(ConvolutionNode *CN, Function *F) {
 
   return F->addNode(new CPUConvDKKC8Node(
       CN->getName(), CN->getType(), CN->getInput(), filter8, CN->getBias(),
-      CN->getKernel(), CN->getStride(), CN->getPad(), CN->getDepth()));
+      CN->getKernel(), CN->getStride(), CN->getPad()));
 }
 
 bool CPUBackend::transformPostLowering(Function *F) {
diff --git a/lib/Backends/OpenCL/OpenCL.cpp b/lib/Backends/OpenCL/OpenCL.cpp
@@ -431,11 +431,9 @@ void OCLBackend::doForwardPass() {
       setKernelArg(kernel, 9, idim);
       setKernelArg(kernel, 10, ShapeNHWC(CC->getFilter()->getType()->dims()));
 
-      auto depth = CC->getDepth();
-
       // Use a 3D grid where the first dimension is the depth and the second
       // dimension is the slice index in the batch.
-      enqueueKernel(commands_, kernel, deviceId_, {odim.h, odim.w, depth},
+      enqueueKernel(commands_, kernel, deviceId_, {odim.h, odim.w, odim.c},
                     kernelLaunches);
       continue;
     }
diff --git a/lib/Graph/Graph.cpp b/lib/Graph/Graph.cpp
@@ -372,38 +372,37 @@ ConvolutionNode *Function::createConv(llvm::StringRef name, NodeValue input,
   auto OT = getParent()->uniqueType(ElemKind::FloatTy, outDims);
 
   return addNode(new ConvolutionNode(name, OT, input, filter, bias, kernel,
-                                     stride, pad, depth, /*group = */ 1));
+                                     stride, pad, /*group = */ 1));
 }
 
 /// Check that the dimensions that are passed in when the convolution is
 /// constructed are correct.
 static void assertConvDims(NodeValue input, NodeValue filter, NodeValue bias,
-                           size_t depth, size_t kernel, size_t stride,
-                           size_t pad, size_t group) {
+                           size_t kernel, size_t stride, size_t pad,
+                           size_t group) {
   ShapeNHWC idim = ShapeNHWC(input.dims());
   assert(idim.w >= kernel && idim.h >= kernel &&
          "buffer too small for selected stride");
   assert(idim.c % group == 0 && "channels number must be divisible by groups");
   (void)idim;
 
   auto filterDims = filter->dims();
-  assert(filterDims[0] == depth * group && filterDims[1] == kernel &&
+  assert(filterDims[0] % group == 0 && filterDims[1] == kernel &&
          filterDims[2] == kernel && filterDims[3] == idim.c / group &&
          "Invalid filter dims");
   (void)filterDims;
 
-  assert(bias->getType()->size() == depth * group && "Invalid bias size");
+  assert(bias->getType()->size() == filterDims[0] && "Invalid bias size");
 }
 
 ConvolutionNode *Function::createConv(llvm::StringRef name, NodeValue input,
                                       NodeValue filter, NodeValue bias,
-                                      TypeRef outTy, size_t depth,
-                                      size_t kernel, size_t stride, size_t pad,
-                                      size_t group) {
-  assertConvDims(input, filter, bias, depth, kernel, stride, pad, group);
+                                      TypeRef outTy, size_t kernel,
+                                      size_t stride, size_t pad, size_t group) {
+  assertConvDims(input, filter, bias, kernel, stride, pad, group);
   auto OT = getParent()->uniqueType(*outTy);
   return addNode(new ConvolutionNode(name, OT, input, filter, bias, kernel,
-                                     stride, pad, depth, group));
+                                     stride, pad, group));
 }
 
 PoolMaxNode *Function::createPoolMax(llvm::StringRef name, NodeValue input,
diff --git a/lib/Graph/Nodes.cpp b/lib/Graph/Nodes.cpp
@@ -488,7 +488,7 @@ static void checkType(NodeValue A, ElemKind expectedType) {
 
 static void verifyConvolution(NodeValue src, NodeValue dest, NodeValue filter,
                               NodeValue bias, size_t kernel, size_t stride,
-                              size_t pad, size_t depth, size_t group) {
+                              size_t pad, size_t group) {
   assert(src.getElementType() == dest.getElementType() && "Invalid Type");
   assert(src.getElementType() == filter.getElementType() && "Invalid Type");
   assert(src.getElementType() == bias.getElementType() && "Invalid Type");
@@ -501,15 +501,14 @@ static void verifyConvolution(NodeValue src, NodeValue dest, NodeValue filter,
   assert(idim.c % group == 0 && "channels number must be divisible by groups");
 
   auto outSz = calculateConvOutputDims(idim.h, idim.w, kernel, stride, pad);
-  ShapeNHWC exp(idim.n, outSz.first, outSz.second, depth * group);
-  (void)exp;
-  assert(exp == odim && "Invalid output dimensions");
+  assert(odim.n == idim.n && odim.h == outSz.first && odim.w == outSz.second &&
+         odim.c % group == 0 && "Invalid output dimensions");
 
-  auto filterDims = {depth * group, kernel, kernel, idim.c / group};
+  auto filterDims = {odim.c, kernel, kernel, idim.c / group};
   assert(filter.getType()->dims().equals(filterDims) && "Invalid filter dims");
   (void)filterDims;
 
-  auto biasDims = {depth * group};
+  auto biasDims = {odim.c};
   assert(bias.getType()->dims().equals(biasDims) && "Invalid bias dims");
   (void)biasDims;
 }
@@ -596,14 +595,14 @@ static void verifyRegression(NodeValue src, NodeValue dest,
 
 void ConvolutionNode::verify() const {
   verifyConvolution(getInput(), getResult(), getFilter(), getBias(), Kernel_,
-                    Stride_, Pad_, Depth_, Group_);
+                    Stride_, Pad_, Group_);
 }
 
 void ConvolutionGradNode::verify() const {
   verifyConvolution(getGradOfInputNamedInput(),
                     getGradOfOriginalOutputNamedResult(),
                     getGradOfInputNamedFilter(), getGradOfInputNamedBias(),
-                    Kernel_, Stride_, Pad_, Depth_, Group_);
+                    Kernel_, Stride_, Pad_, Group_);
 }
 
 void PoolMaxNode::verify() const {
diff --git a/lib/IR/IRGen.cpp b/lib/IR/IRGen.cpp
@@ -115,10 +115,9 @@ struct IRGenVisitor : NodeWalker {
       auto *filterG = builder_.createAllocActivationInst("conv.filter.G",
                                                          filter->getType());
 
-      builder_.createConvolutionGradInst(N->getName(), input, filter, outGrad,
-                                         inG, filterG, biasG, CG->getKernel(),
-                                         CG->getStride(), CG->getPad(),
-                                         CG->getDepth(), CG->getGroup());
+      builder_.createConvolutionGradInst(
+          N->getName(), input, filter, outGrad, inG, filterG, biasG,
+          CG->getKernel(), CG->getStride(), CG->getPad(), CG->getGroup());
 
       registerIR(CG->getGradOfInputNamedInput(), inG);
       registerIR(CG->getGradOfInputNamedFilter(), filterG);
diff --git a/lib/Importer/Caffe2.cpp b/lib/Importer/Caffe2.cpp
@@ -218,8 +218,8 @@ void caffe2ModelLoader::loadOperator(const caffe2::OperatorDef &op) {
         {idim.n, outSz.first, outSz.second, depth}};
     auto outTy = G_.getParent()->uniqueType(ElemKind::FloatTy, outDims);
 
-    auto *node = G_.createConv(opName, tr, filter, bias, outTy, depth / group,
-                               kernel, stride, pad, group);
+    auto *node = G_.createConv(opName, tr, filter, bias, outTy, kernel, stride,
+                               pad, group);
 
     // Transpose the output back.
     auto *N = G_.createTranspose(opName, node, NHWC2NCHW);
diff --git a/lib/Optimizer/GraphOptimizer.cpp b/lib/Optimizer/GraphOptimizer.cpp
@@ -762,10 +762,10 @@ static void optimizeQuantization(Function *F) {
       if (auto *CN = dyn_cast<ConvolutionNode>(RS->getInput())) {
         // Create the exact same convolution but with a different scaling
         // return type.
-        auto *newCN = F->createConv(
-            CN->getName(), CN->getInput(), CN->getFilter(), CN->getBias(),
-            RS->getType(), CN->getDepth(), CN->getKernel(), CN->getStride(),
-            CN->getPad(), CN->getGroup());
+        auto *newCN =
+            F->createConv(CN->getName(), CN->getInput(), CN->getFilter(),
+                          CN->getBias(), RS->getType(), CN->getKernel(),
+                          CN->getStride(), CN->getPad(), CN->getGroup());
         RS->getResult().replaceAllUsesOfWith(newCN);
         continue;
       }
diff --git a/lib/Optimizer/Lower.cpp b/lib/Optimizer/Lower.cpp
@@ -504,9 +504,8 @@ void lowerGroupConvolutionNode(Function *F, ConvolutionNode &BNG) {
                        {(groupId + 1) * outCperG, kernel, kernel, inCperG});
     auto *bias_slice = F->createSlice(BNG.getName(), bias, {groupId * outCperG},
                                       {(groupId + 1) * outCperG});
-    convs[groupId] =
-        F->createConv(BNG.getName(), in_slice, filter_slice, bias_slice, outTy,
-                      outCperG, kernel, stride, pad, 1);
+    convs[groupId] = F->createConv(BNG.getName(), in_slice, filter_slice,
+                                   bias_slice, outTy, kernel, stride, pad, 1);
   }
   auto result = F->createConcat(BNG.getName(), convs, 3);
   BNG.getResult().replaceAllUsesOfWith(result);
diff --git a/lib/Quantization/Quantization.cpp b/lib/Quantization/Quantization.cpp
@@ -322,10 +322,10 @@ void generateQuantizedGraph(
         auto QT = F->getParent()->uniqueType(ElemKind::Int8QTy,
                                              CV->getResult()->dims(),
                                              TQP.scale_, TQP.offset_);
-        quantizedNode = F->createConv(
-            CV->getName(), quantizedInputs[0], quantizedInputs[1],
-            quantizedInputs[2], QT, CV->getDepth(), CV->getKernel(),
-            CV->getStride(), CV->getPad(), CV->getGroup());
+        quantizedNode =
+            F->createConv(CV->getName(), quantizedInputs[0], quantizedInputs[1],
+                          quantizedInputs[2], QT, CV->getKernel(),
+                          CV->getStride(), CV->getPad(), CV->getGroup());
         break;
       }
       case Kinded::Kind::SliceNodeKind: {
diff --git a/tests/unittests/BackendTestUtils.cpp b/tests/unittests/BackendTestUtils.cpp
@@ -104,7 +104,7 @@ void inferConvNet(Tensor *inputs, Tensor *filter, Tensor *bias, Tensor *out,
     OT = F->getParent()->uniqueType(out->getElementType(), out->dims());
   }
   auto *conv =
-      F->createConv("conv", inputVar, filterVar, biasVar, OT, 10, 5, 3, 4, 1);
+      F->createConv("conv", inputVar, filterVar, biasVar, OT, 5, 3, 4, 1);
   auto result = F->createSave("ret", conv, outVar);
   EE.compile(CompilationMode::Infer, F);
   EE.run({inputVar, filterVar, biasVar}, {inputs, filter, bias});
diff --git a/tests/unittests/OperatorTest.cpp b/tests/unittests/OperatorTest.cpp
@@ -502,8 +502,8 @@ void checkIntConvolution(ExecutionEngine &EE, unsigned convDepth) {
   auto *biasq = F->createQuantize("bias.q", bias, biasTy);
 
   auto *convq =
-      F->createConv("convq", inputq, filterq, biasq, resTy, conv->getDepth(),
-                    conv->getKernel(), conv->getStride(), conv->getPad(), 1);
+      F->createConv("convq", inputq, filterq, biasq, resTy, conv->getKernel(),
+                    conv->getStride(), conv->getPad(), 1);
   auto *dequantRes = F->createDequantize("dequant", convq);
 
   // Subtract the results of the convolution from the quantized convolution.
diff --git a/tests/unittests/basicIRTest.cpp b/tests/unittests/basicIRTest.cpp
@@ -115,7 +115,7 @@ TEST(IR, allInstrs) {
     XY->setName("srcXY");
 
     builder.createCopyInst("", I1, I0);
-    builder.createConvolutionInst("", I3, I1, F0, B0, 7, 2, 3, 64, 1);
+    builder.createConvolutionInst("", I3, I1, F0, B0, 7, 2, 3, 1);
     builder.createPoolMaxInst("", I4, I0, 7, 2, 3);
     builder.createSigmoidInst("", I1, I0);
     builder.createTanhInst("", I1, I0);
diff --git a/tests/unittests/graphTest.cpp b/tests/unittests/graphTest.cpp
@@ -123,8 +123,8 @@ TEST(Graph, simpleQuant) {
   std::array<size_t, 4> outDims = {{1, outSz.first, outSz.second, 16}};
   auto t = F->getParent()->uniqueType(glow::ElemKind::Int8QTy, outDims, 1.5, 6);
 
-  auto *conv = F->createConv("conv", input, filter, bias, t, depth, kernel,
-                             step, pad, 1);
+  auto *conv =
+      F->createConv("conv", input, filter, bias, t, kernel, step, pad, 1);
 
   auto s = conv->getType()->size();
   auto *fcFilter = MD.createVariable(ElemKind::Int8QTy, {s, 6}, 0.4, 2, "F");
diff --git a/tools/ClassGen/Backends/CPU/CPUSpecificInstrs.h b/tools/ClassGen/Backends/CPU/CPUSpecificInstrs.h
@@ -16,7 +16,6 @@ BB.newBackendSpecificInstr("CPUConvDKKC8")
     .addMember(MemberType::SizeT, "Kernel")
     .addMember(MemberType::SizeT, "Stride")
     .addMember(MemberType::SizeT, "Pad")
-    .addMember(MemberType::SizeT, "Depth")
     .autoIRGen()
     .autoVerify(VerifyKind::SameElementType, {"Dest", "Src", "Filter", "Bias"});
 
diff --git a/tools/ClassGen/Backends/CPU/CPUSpecificNodes.h b/tools/ClassGen/Backends/CPU/CPUSpecificNodes.h
@@ -13,7 +13,6 @@ BB.newNode("CPUConvDKKC8")
     .addMember(MemberType::SizeT, "Kernel")
     .addMember(MemberType::SizeT, "Stride")
     .addMember(MemberType::SizeT, "Pad")
-    .addMember(MemberType::SizeT, "Depth")
     .addResultFromCtorArg()
     .setDocstring("This is a cpu-specific convolution implementation where the "
                   "filter is transposed to the shape [D/8, K, K, C, 8]");
diff --git a/tools/ClassGen/Backends/CPU/CPUSpecificNodesVerification.h b/tools/ClassGen/Backends/CPU/CPUSpecificNodesVerification.h
@@ -10,7 +10,7 @@ void CPUConvDKKC8Node::verify() const {
   ShapeNHWC odim(getResult().getType()->dims());
   auto outSz = calculateConvOutputDims(idim.h, idim.w, getKernel(), getStride(),
                                        getPad());
-  ShapeNHWC exp(idim.n, outSz.first, outSz.second, getDepth());
+  ShapeNHWC exp(idim.n, outSz.first, outSz.second, getBias().dims()[0]);
   (void)exp;
   assert(exp == odim && "Invalid output dimensions");
 }
diff --git a/tools/ClassGen/InstrGen.cpp b/tools/ClassGen/InstrGen.cpp
@@ -69,7 +69,6 @@ int main(int argc, char **argv) {
       .addMember(MemberType::SizeT, "Kernel")
       .addMember(MemberType::SizeT, "Stride")
       .addMember(MemberType::SizeT, "Pad")
-      .addMember(MemberType::SizeT, "Depth")
       .addMember(MemberType::SizeT, "Group")
       .autoIRGen()
       .autoVerify(VerifyKind::SameElementType,
diff --git a/tools/ClassGen/NodeGen.cpp b/tools/ClassGen/NodeGen.cpp
@@ -49,13 +49,12 @@ int main(int argc, char **argv) {
       .addMember(MemberType::SizeT, "Kernel")
       .addMember(MemberType::SizeT, "Stride")
       .addMember(MemberType::SizeT, "Pad")
-      .addMember(MemberType::SizeT, "Depth")
       .addMember(MemberType::SizeT, "Group")
       .addResultFromCtorArg()
       .addGradient()
       .setDocstring("Performs Convolution using a given Input, Filter, and "
                     "Bias tensors, as well as provided Kernel, Stride, Pad, "
-                    "Depth and Group.");
+                    "and Group.");
 
   BB.newNode("PoolMax")
       .addInput("Input")
