Merge branch 'master' into rearhart/plaidml-opt-m

.ci/onnx/onnxruntime/Jenkinsfile

+// ******************************************************************************
+// Copyright 2018-2019 Intel Corporation
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+// ******************************************************************************
+
+try{ if(LABEL.trim() == "") {throw new Exception();} }catch(Exception e){LABEL="onnx && ci"}; echo "${LABEL}"
+
+NGRPAH_REPOSITORY = "https://github.com/NervanaSystems/ngraph.git"
+NGRAPH_COMMIT_HASH = "${ghprbActualCommit}" // particular nGraph PR commit hash 
+
+ONNX_REPOSITORY = "https://github.com/NervanaSystems/onnxruntime.git"
+ONNX_RUNTIME_BRANCH = "release"
+
+def main(){
+    timeout(activity: true, time: 15) {
+        try{
+            stage("CloneRepos"){
+                CloneRepos()
+            }
+            stage("Apply Patch"){
+                ApplyPatch()
+            }
+            stage("Onnx Models"){
+                BuildAndTest()
+            }            
+        }
+        catch(e) {
+            // Set result to ABORTED if exception contains exit code of a process interrupted by SIGTERM
+            if ("$e".contains("143")) {
+                currentBuild.result = "ABORTED"
+            } else {
+                currentBuild.result = "FAILURE"
+            }
+        }
+        stage("Clean"){
+            Clean()
+        }
+    }
+}
+
+
+def CloneRepos() {
+    dir("ngraph"){
+        checkout([
+            $class: 'GitSCM',
+            branches: [[name: "${NGRAPH_COMMIT_HASH}"]],
+            doGenerateSubmoduleConfigurations: false,
+            extensions: [[
+                $class: 'SubmoduleOption',
+                disableSubmodules: false,
+                parentCredentials: true,
+                recursiveSubmodules: true,
+                reference: '',
+                trackingSubmodules: false,
+                timeout: 15
+            ]],
+            submoduleCfg: [],
+            userRemoteConfigs: [[
+                refspec: '+refs/pull/*:refs/remotes/origin/pr/*',
+                url: "${NGRPAH_REPOSITORY}"
+            ]]
+        ])
+    }
+    dir("onnxruntime") {
+        checkout([
+            $class: 'GitSCM',
+            branches: [[name: "${ONNX_RUNTIME_BRANCH}"]],
+            doGenerateSubmoduleConfigurations: false,
+            extensions: [[
+                $class: 'SubmoduleOption',
+                disableSubmodules: false,
+                parentCredentials: true,
+                recursiveSubmodules: true,
+                reference: '',
+                trackingSubmodules: false,
+                timeout: 15
+            ]],
+            submoduleCfg: [],
+            userRemoteConfigs: [[                    
+                url: "${ONNX_REPOSITORY}"
+            ]]
+        ])
+    }
+}
+
+def ApplyPatch(){
+    dir("onnxruntime"){
+        echo "Update cmake/external/ngraph.cmake with ${NGRAPH_COMMIT_HASH}"
+        sh """
+            sed -i 's/set(ngraph_TAG ".*")/set(ngraph_TAG "${NGRAPH_COMMIT_HASH}")/g' cmake/external/ngraph.cmake 
+            grep -q "${NGRAPH_COMMIT_HASH}" cmake/external/ngraph.cmake 
+        """
+        echo "Add proxy to tools/ci_build/github/linux/docker/Dockerfile.ubuntu"
+        sh """
+            sed -i 's|{HTTP_PROXY}|${env.http_proxy}|g' ../ngraph/.ci/onnx/onnxruntime/proxy.patch
+            sed -i 's|{SOCKS_PROXY}|${env.socks_proxy}|g' ../ngraph/.ci/onnx/onnxruntime/proxy.patch
+            grep -q "${env.http_proxy}" ../ngraph/.ci/onnx/onnxruntime/proxy.patch
+            git apply ../ngraph/.ci/onnx/onnxruntime/proxy.patch
+        """
+    }
+}
+
+def BuildAndTest(){
+    dir("onnxruntime"){
+        sh "mkdir -p `pwd`/build/models && chmod 777 `pwd`/build/models"
+        sh """
+            //!/bin/bash
+            ./tools/ci_build/github/linux/run_dockerbuild.sh \
+                -o ubuntu16.04 \
+                -d ngraph \
+                -r `pwd`/build -x '--use_ngraph --use_full_protobuf --test_data_url https://onnxruntimetestdata.blob.core.windows.net/models/20190327.zip --test_data_checksum 45166d81c021c8aae212b53c92101792'
+        """
+    }
+}
+
+def Clean(){
+    deleteDir()
+}
+
+node(LABEL) {
+    main()
+}

.ci/onnx/onnxruntime/proxy.patch

+diff --git a/tools/ci_build/github/linux/docker/Dockerfile.ubuntu b/tools/ci_build/github/linux/docker/Dockerfile.ubuntu
+index bdff95e1..cd9c0008 100644
+--- a/tools/ci_build/github/linux/docker/Dockerfile.ubuntu
++++ b/tools/ci_build/github/linux/docker/Dockerfile.ubuntu
+@@ -3,6 +3,18 @@ FROM ubuntu:${OS_VERSION}
+ 
+ ARG PYTHON_VERSION=3.5
+ 
++ENV http_proxy={HTTP_PROXY}
++ENV socks_proxy={SOCKS_PROXY}
++ENV https_proxy={HTTP_PROXY}
++ENV ftp_proxy={HTTP_PROXY}
++ENV rsync_proxy={HTTP_PROXY}
++ENV no_proxy=intel.com,.intel.com,localhost
++ENV HTTP_PROXY={HTTP_PROXY}
++ENV HTTPS_PROXY={HTTP_PROXY}
++ENV FTP_PROXY={HTTP_PROXY}
++ENV SOCKS_PROXY={SOCKS_PROXY}
++ENV NO_PROXY=intel.com,.intel.com,localhost
++
+ ADD scripts /tmp/scripts
+ RUN /tmp/scripts/install_ubuntu.sh -p ${PYTHON_VERSION} && /tmp/scripts/install_deps.sh && rm -rf /tmp/scripts
+ 

CONTRIB.md

 Contributor Guidelines
 ======================
 
-https://ngraph.nervanasys.com/docs/latest/project/code-contributor-README.html
+The latest version of this file can be found at: 
+
+https://ngraph.nervanasys.com/docs/latest/project/contribution-guide.html
 
 
 License

cmake/external_mlir.cmake

@@ -20,8 +20,8 @@ set(MLIR_LLVM_REPO_URL https://github.com/llvm/llvm-project.git)
 set(MLIR_REPO_URL https://github.com/tensorflow/mlir.git)
 
 # Change these commit IDs to move to latest stable versions
-set(MLIR_LLVM_COMMIT_ID bb2b527)
-set(MLIR_COMMIT_ID 49f7efc)
+set(MLIR_LLVM_COMMIT_ID c0cad98)
+set(MLIR_COMMIT_ID 82d5084)
 set(MLIR_PROJECT_ROOT ${CMAKE_CURRENT_BINARY_DIR}/mlir_project)
 set(MLIR_LLVM_ROOT ${MLIR_PROJECT_ROOT}/llvm-projects)
 set(MLIR_SOURCE_DIR ${MLIR_LLVM_ROOT}/llvm/projects/mlir)

src/contrib/mlir/CMakeLists.txt

@@ -56,6 +56,7 @@ if (NGRAPH_MLIR_ENABLE)
         MLIRExecutionEngine
         MLIRIR
         MLIRLLVMIR
+        MLIRStandardToLLVM
         MLIRParser
         MLIRPass
         MLIRTargetLLVMIR

src/contrib/mlir/compiler.cpp

@@ -34,11 +34,12 @@
 #include <llvm/Support/MemoryBuffer.h>
 #include <llvm/Support/SourceMgr.h>
 #include <llvm/Support/TargetSelect.h>
+#include <mlir/Conversion/StandardToLLVM/ConvertStandardToLLVM.h>
+#include <mlir/Conversion/StandardToLLVM/ConvertStandardToLLVMPass.h>
 #include <mlir/ExecutionEngine/ExecutionEngine.h>
 #include <mlir/ExecutionEngine/MemRefUtils.h>
 #include <mlir/ExecutionEngine/OptUtils.h>
 #include <mlir/LLVMIR/LLVMDialect.h>
-#include <mlir/LLVMIR/Transforms.h>
 #include <mlir/Pass/PassManager.h>
 #include <mlir/Target/LLVMIR.h>
 #include <mlir/Transforms/DialectConversion.h>
@@ -50,6 +51,7 @@
 using llvm::SmallVector;
 using llvm::StringRef;
 using llvm::make_unique;
+
 using namespace ngraph::runtime::ngmlir;
 
 #define COMPILE_OP_DECL(op_name)                                                                   \
@@ -75,7 +77,7 @@ void MLIRCompiler::init_mlir()
 
     if (!initialized)
     {
-        mlir::registerDialect<mlir::NGDialect>();
+        mlir::registerDialect<mlir::NGraphOpsDialect>();
         // Register any LLVM command line options
         llvm::cl::ParseEnvironmentOptions("ngraph", "MLIR_LLVM_OPTIONS", "");
         initialized = true;
@@ -133,7 +135,7 @@ void MLIRCompiler::build_ng_dialect_module()
     }
 
     // create builder
-    m_builder = llvm::make_unique<mlir::FuncBuilder>(function.get());
+    m_builder = llvm::make_unique<mlir::OpBuilder>(function->getBody());
     build_ng_dialect();
     m_module->getFunctions().push_back(function.release());
     if (failed(m_module->verify()))
@@ -359,10 +361,14 @@ void MLIRCompiler::execute()
     NGRAPH_CHECK(m_module, "MLIR module is not ready.");
 
     // Lower Standard dialect to LLVM dialect.
-    auto converter = mlir::createStdToLLVMConverter();
-    auto r = converter->convert(m_module.get());
-    (void)r;
-    NGRAPH_CHECK(succeeded(r), "second conversion failed");
+    mlir::LLVMTypeConverter llvm_converter(&m_context);
+    OwningRewritePatternList patterns;
+    mlir::populateStdToLLVMConversionPatterns(llvm_converter, patterns);
+
+    mlir::ConversionTarget target(m_context);
+    target.addLegalDialect<mlir::LLVM::LLVMDialect>();
+    auto result = applyConversionPatterns(*m_module, target, llvm_converter, std::move(patterns));
+    NGRAPH_CHECK(succeeded(result), "Standard to LLVM dialect conversion failed");
 
     dump_mlir_module("LLVM-IR Dialect Dump:");
 

src/contrib/mlir/compiler.hpp

@@ -132,7 +132,7 @@ namespace ngraph
                 mlir::MLIRContext m_context;
 
                 std::unique_ptr<mlir::Module> m_module;
-                std::unique_ptr<mlir::FuncBuilder> m_builder;
+                std::unique_ptr<mlir::OpBuilder> m_builder;
                 std::unique_ptr<mlir::ExecutionEngine> m_engine;
 
                 using TensorToInfo = std::pair<descriptor::Tensor*, TensorInfo>;

src/contrib/mlir/dialect/dialect.cpp

@@ -21,8 +21,8 @@
 
 using namespace mlir;
 
-NGDialect::NGDialect(mlir::MLIRContext* ctx)
-    : mlir::Dialect("ng", ctx)
+NGraphOpsDialect::NGraphOpsDialect(mlir::MLIRContext* ctx)
+    : mlir::Dialect(getDialectNamespace(), ctx)
 {
     addTypes<NGTensorType>();
     addTypes<NGIntegerType>();
@@ -34,7 +34,7 @@ NGDialect::NGDialect(mlir::MLIRContext* ctx)
         >();
 }
 
-void NGDialect::printType(mlir::Type type, raw_ostream& os) const
+void NGraphOpsDialect::printType(mlir::Type type, raw_ostream& os) const
 {
     switch (type.getKind())
     {

src/contrib/mlir/dialect/dialect.hpp

@@ -25,15 +25,17 @@
 #include "ngraph/check.hpp"
 namespace mlir
 {
-    class NGDialect : public mlir::Dialect
+    class NGraphOpsDialect : public mlir::Dialect
     {
     public:
-        explicit NGDialect(mlir::MLIRContext* ctx);
+        explicit NGraphOpsDialect(mlir::MLIRContext* ctx);
         mlir::Type parseType(llvm::StringRef tyData, mlir::Location loc) const override
         {
             NGRAPH_CHECK(false, "Unsupported type parsing.");
             return mlir::Type();
         }
         void printType(mlir::Type type, llvm::raw_ostream& os) const override;
+
+        static StringRef getDialectNamespace() { return "ng"; }
     };
 }

src/contrib/mlir/lowerer.cpp

@@ -41,31 +41,34 @@ namespace
 
     class DialectLoweringPass;
 
+    /// Base class for nGraph operation conversions to affine/standard dialect. Provides
+    /// conversion patterns with an access to the DialectLoweringPass which holds the state of the
+    /// conversion.
+    class NGraphOpLowering : public ConversionPattern
+    {
+    public:
+        NGraphOpLowering(StringRef rootOpName, MLIRContext* context, DialectLoweringPass& pass)
+            : ConversionPattern(rootOpName, /*benefit=*/1, context)
+            , m_pass(pass){};
+
+    protected:
+        // Back-reference to the lowering pass which contains the lowering state, including the
+        // nGraph type converter.
+        DialectLoweringPass& m_pass;
+    };
+
 #include "op_lowerers.inc"
 
-    /// Use Dialect Converson Framework
-    class DialectLowerer : public DialectConversion
+    /// Conversion from types in the nGraph dialect to the Standard dialect.
+    class NGraphTypeConverter : public TypeConverter
     {
     public:
-        DialectLowerer(DialectLoweringPass& pass)
-            : DialectConversion()
-            , m_pass(pass)
+        NGraphTypeConverter()
+            : TypeConverter()
         {
         }
 
         Type convertType(Type t) override;
-
-    protected:
-        // Initialize the list of converters.
-        void initConverters(OwningRewritePatternList& patterns, MLIRContext* mlirContext) override
-        {
-            RewriteListBuilder<NGAddOpConversion, NGDotOpConversion, NGReturnOpConversion>::build(
-                patterns, mlirContext, m_pass);
-        }
-
-    private:
-        DialectLoweringPass& m_pass;
-        llvm::BumpPtrAllocator allocator;
     };
 
     /// Dialect Lowering Pass to affine ops
@@ -73,14 +76,17 @@ namespace
     {
     public:
         DialectLoweringPass(ngmlir::MLIRCompiler& compiler)
-            : m_dialectLowerer(*this)
-            , m_compiler(compiler)
+            : m_compiler(compiler)
         {
         }
+
         void runOnModule() override;
         SmallVector<Value*, 4> buildOutputDefs(Operation* op, PatternRewriter& rewriter);
 
     private:
+        /// Collect a set of patterns to convert from the nGraph dialect to Affine dialect.
+        void populateNGraphToAffineConversionPatterns(OwningRewritePatternList& patterns);
+
         mlir::Function* getCallDecl(StringRef name,
                                     ArrayRef<Type> args,
                                     ArrayRef<Type> output,
@@ -90,7 +96,7 @@ namespace
         Value* insertMemMgrDef(PatternRewriter* rewriter = nullptr);
 
     private:
-        DialectLowerer m_dialectLowerer;
+        NGraphTypeConverter m_typeConverter;
         // Value holding mem manager passed pointer
         SmallVector<Value*, 4> m_memMgrDefs;
 
@@ -101,21 +107,39 @@ namespace
 
     void DialectLoweringPass::runOnModule()
     {
+        // Create type converter and initialize conversion patterns.
+        NGraphTypeConverter converter;
+        OwningRewritePatternList patterns;
+        populateNGraphToAffineConversionPatterns(patterns);
+
+        // Create target that defines legal ops for nGraph dialect to be lowered to.
+        ConversionTarget target(getContext());
+        // TODO: Remove NGFakeInputOp. We need to set NGFakeInputOp as legal op because we generate
+        // it as part of the lowering to affine/standard.
+        target.addLegalDialect<AffineOpsDialect, StandardOpsDialect>();
+        target.addLegalOp<NGFakeInputOp>();
+
         // capture output values by looking for the Return and grabbing the values
         // the order of the returned values matches the order of the lowered func signature for
         // results. This is used to find the arg_id that a defined value maps to if it is an output
         findOutputValues();
 
-        if (failed(m_dialectLowerer.convert(&getModule())))
+        if (failed(applyConversionPatterns(getModule(), target, converter, std::move(patterns))))
         {
-            getModule().getContext()->emitError(mlir::UnknownLoc::get(getModule().getContext()),
-                                                "Error lowering dialect\n");
+            emitError(mlir::UnknownLoc::get(&getContext()), "Error lowering nGraph dialect\n");
             signalPassFailure();
         }
 
         processFakeInstrs();
     }
 
+    void DialectLoweringPass::populateNGraphToAffineConversionPatterns(
+        OwningRewritePatternList& patterns)
+    {
+        RewriteListBuilder<NGAddOpConversion, NGDotOpConversion, NGReturnOpConversion>::build(
+            patterns, &getContext(), *this);
+    }
+
     void DialectLoweringPass::findOutputValues()
     {
         // get original function
@@ -138,6 +162,9 @@ namespace
             outputCount = ret.getNumOperands();
         });
         // will be populated with lowered output values later
+        // TODO: This resize is making debugging obscure. When the container is not populated due
+        // to a bug, null pointers are used by the consumer leading to a crash more difficult to
+        // root-cause. We should try to change the current approach or introduce verification code.
         m_loweredOutputValues.resize(outputCount, nullptr);
     }
 
@@ -146,10 +173,11 @@ namespace
     {
         // it would be nice to insert one fake def at the start of the new func
         // however, due to how DialectConversion framework works, new func is only
-        // materialized after conversion is done (rewriter->getFunction, or even rewriter->getInsertionBlock()->getFunction()
-        // will give you the original func). This makes it very convoluted to insert instructions at entry block.
+        // materialized after conversion is done (rewriter->getFunction, or even
+        // rewriter->getInsertionBlock()->getFunction() will give you the original func). This
+        // makes it very convoluted to insert instructions at entry block.
         auto op = rewriter->create<NGFakeInputOp>(rewriter->getUnknownLoc(),
-                                                  IndexType::get(getModule().getContext()));
+                                                  IndexType::get(&getContext()));
         // will be fixed later to read passed arg instead.
         m_memMgrDefs.push_back(op.getResult());
         return op.getResult();
@@ -167,8 +195,7 @@ namespace
                 unsigned argId = (int)attr.getInt();
                 auto fakeOp = rewriter.create<NGFakeInputOp>(
                     op->getLoc(),
-                    m_dialectLowerer.convertType(
-                        origResult->getType()) /* convert to lowered type */
+                    m_typeConverter.convertType(origResult->getType()) /* convert to lowered type */
                     );
                 // Fake instrution is short-lived. Verify here.
                 fakeOp.verify();
@@ -181,7 +208,7 @@ namespace
                 auto tensorType = origResult->getType().cast<NGTensorType>();
                 auto callBackFunc = getCallDecl("__mlir_allocate",
                                                 {rewriter.getIndexType(), rewriter.getIndexType()},
-                                                {m_dialectLowerer.convertType(tensorType)},
+                                                {m_typeConverter.convertType(tensorType)},
                                                 rewriter);
 
                 auto size = tensorType.getSizeInBytes();
@@ -261,10 +288,10 @@ namespace
         return callBackFuncPtr;
     }
     // NGDialect converters
-    Type DialectLowerer::convertType(Type type)
+    Type NGraphTypeConverter::convertType(Type type)
     {
         // We may need to refactor this code to a external utility if type conversion is needed
-        // outside of the lowering context since DialectLowerer is private.
+        // outside of the lowering context since NGraphTypeConverter is private.
 
         if (auto tensor_type = type.dyn_cast<NGTensorType>())
         {
@@ -294,7 +321,7 @@ namespace
     }
 
 #define REWRITER(OP)                                                                               \
-    void OP##Conversion::rewrite(                                                                  \
+    PatternMatchResult OP##Conversion::matchAndRewrite(                                            \
         Operation* op, ArrayRef<Value*> operands, PatternRewriter& rewriter) const
 
     // ADD
@@ -334,6 +361,8 @@ namespace
                 });
         // clang-format on
         rewriter.replaceOp(op, {result});
+
+        return matchSuccess();
     }
 
     REWRITER(NGDotOp)
@@ -396,9 +425,16 @@ namespace
         });
 
         rewriter.replaceOp(op, {result});
+
+        return matchSuccess();
+    }
+
+    REWRITER(NGReturnOp)
+    {
+        rewriter.replaceOpWithNewOp<ReturnOp>(op);
+        return matchSuccess();
     }
 
-    REWRITER(NGReturnOp) { rewriter.replaceOpWithNewOp<ReturnOp>(op); }
 #undef REWRITER
 }
 

src/contrib/mlir/lowerer.hpp

@@ -27,6 +27,8 @@ namespace ngraph
         namespace ngmlir
         {
             class MLIRCompiler;
+
+            using OwningRewritePatternList = std::vector<std::unique_ptr<mlir::RewritePattern>>;
         }
     }
 }

src/contrib/mlir/op_lowerers.inc

@@ -18,16 +18,18 @@
 // Add new dialect ops lowerers to this file
 
 #define DECL_OP_CONV(OP)                                                                           \
-class OP##Conversion : public mlir::DialectConversionPattern \
-{\
-public:\
-    explicit OP##Conversion(mlir::MLIRContext *context, DialectLoweringPass& pass)\
-    : mlir::DialectConversionPattern(mlir::OP::getOperationName(), 1, context),\
-      m_pass(pass)\
-    {} \
-    void rewrite(Operation *op, ArrayRef<Value *> operands, PatternRewriter &rewriter) const override; \
-    DialectLoweringPass& m_pass;\
-};
+    class OP##Conversion : public NGraphOpLowering                                                 \
+    {                                                                                              \
+    public:                                                                                        \
+        explicit OP##Conversion(mlir::MLIRContext* context, DialectLoweringPass& pass)             \
+            : NGraphOpLowering(mlir::OP::getOperationName(), context, pass)                        \
+        {                                                                                          \
+        }                                                                                          \
+                                                                                                   \
+        PatternMatchResult matchAndRewrite(Operation* op,                                          \
+                                           ArrayRef<Value*> operands,                              \
+                                           PatternRewriter& rewriter) const override;              \
+    };
 
 DECL_OP_CONV(NGAddOp)
 DECL_OP_CONV(NGDotOp)

src/ngraph/CMakeLists.txt

@@ -142,6 +142,8 @@ set (SRC
     op/experimental/dyn_broadcast.hpp
     op/experimental/dyn_pad.cpp
     op/experimental/dyn_pad.hpp
+    op/experimental/dyn_replace_slice.cpp
+    op/experimental/dyn_replace_slice.hpp
     op/experimental/dyn_reshape.cpp
     op/experimental/dyn_reshape.hpp
     op/experimental/dyn_slice.cpp

src/ngraph/descriptor/tensor.cpp

@@ -46,6 +46,9 @@ descriptor::Tensor::Tensor(const element::Type& element_type,
 void descriptor::Tensor::set_tensor_type(const element::Type& element_type,
                                          const PartialShape& pshape)
 {
+    NGRAPH_CHECK(pshape.all_non_negative(),
+                 "set_tensor_type called on a PartialShape containing negative dimensions: ",
+                 pshape);
     if (pshape.is_static())
     {
         m_shape = pshape.to_shape();

src/ngraph/dimension.cpp

@@ -23,7 +23,7 @@
 
 using namespace ngraph;
 
-Dimension::Dimension(size_t dimension)
+Dimension::Dimension(int64_t dimension)
     : m_dimension(dimension)
 {
     if (dimension == s_dynamic_val)
@@ -40,7 +40,7 @@ std::ostream& ngraph::operator<<(std::ostream& str, const Dimension& dimension)
 {
     if (dimension.is_static())
     {
-        return (str << size_t(dimension));
+        return (str << int64_t(dimension));
     }
     else
     {
@@ -50,36 +50,36 @@ std::ostream& ngraph::operator<<(std::ostream& str, const Dimension& dimension)
 
 Dimension Dimension::operator+(const Dimension& dim) const
 {
-    return (is_static() && dim.is_static() ? m_dimension + size_t(dim) : Dimension::dynamic());
+    return (is_static() && dim.is_static() ? m_dimension + int64_t(dim) : Dimension::dynamic());
 }
 
 Dimension Dimension::operator-(const Dimension& dim) const
 {
-    return (is_static() && dim.is_static() ? m_dimension - size_t(dim) : Dimension::dynamic());
+    return (is_static() && dim.is_static() ? m_dimension - int64_t(dim) : Dimension::dynamic());
 }
 
 Dimension Dimension::operator*(const Dimension& dim) const
 {
     return ((is_static() && dim.is_static())
-                ? m_dimension * size_t(dim)
+                ? m_dimension * int64_t(dim)
                 : (is_static() && m_dimension == 0)
                       ? 0
-                      : (dim.is_static() && size_t(dim) == 0) ? 0 : Dimension::dynamic());
+                      : (dim.is_static() && int64_t(dim) == 0) ? 0 : Dimension::dynamic());
 }
 
 bool Dimension::compatible(const Dimension& d) const
 {
-    return (is_dynamic() || d.is_dynamic() || m_dimension == size_t(d));
+    return (is_dynamic() || d.is_dynamic() || m_dimension == int64_t(d));
 }
 
 bool Dimension::relaxes(const Dimension& d) const
 {
-    return (is_dynamic() || (d.is_static() && size_t(*this) == size_t(d)));
+    return (is_dynamic() || (d.is_static() && int64_t(*this) == int64_t(d)));
 }
 
 bool Dimension::refines(const Dimension& d) const
 {
-    return (d.is_dynamic() || (is_static() && size_t(d) == size_t(*this)));
+    return (d.is_dynamic() || (is_static() && int64_t(d) == int64_t(*this)));
 }
 
 bool Dimension::merge(Dimension& dst, const Dimension d1, const Dimension d2)
@@ -94,7 +94,7 @@ bool Dimension::merge(Dimension& dst, const Dimension d1, const Dimension d2)
         dst = d1;
         return true;
     }
-    else if (size_t(d1) != size_t(d2))
+    else if (int64_t(d1) != int64_t(d2))
     {
         return false;
     }
@@ -115,16 +115,16 @@ bool Dimension::broadcast_merge(Dimension& dst, const Dimension d1, const Dimens
     else if (d1.is_dynamic() || d2.is_dynamic())
     {
         // One static. Set dst to static size if >1
-        auto ds = d1.is_dynamic() ? size_t(d2) : size_t(d1);
+        auto ds = d1.is_dynamic() ? int64_t(d2) : int64_t(d1);
         dst = (ds > 1) ? ds : Dimension::dynamic();
         return true;
     }
     else
     {
         // Static sizes. Both match or one of them is 1.
-        if (size_t(d1) == size_t(d2) || size_t(d1) == 1 || size_t(d2) == 1)
+        if (int64_t(d1) == int64_t(d2) || int64_t(d1) == 1 || int64_t(d2) == 1)
         {
-            dst = std::max(size_t(d1), size_t(d2));
+            dst = std::max(int64_t(d1), int64_t(d2));
             return true;
         }
         else

src/ngraph/dimension.hpp

@@ -25,7 +25,7 @@ namespace ngraph
     /// \brief Class representing a dimension, which may be dynamic (undetermined until runtime),
     ///        in a shape or shape-like object.
     ///
-    /// Static dimensions may be implicitly converted from size_t. A dynamic dimension is
+    /// Static dimensions may be implicitly converted from int64_t. A dynamic dimension is
     /// constructed with Dimension() or Dimension::dynamic().
     ///
     /// XXX: THIS CLASS IS NOT IN USE YET AND THE ENTIRE DESIGN IS SUBJECT TO CHANGE.
@@ -36,7 +36,7 @@ namespace ngraph
         /// \param dimension Value of the dimension. Must not be equal to
         ///                  Dimension::s_dynamic_val.
         /// \throws std::invalid_argument If `dimension` == Dimension::s_dynamic_val.
-        Dimension(size_t dimension);
+        Dimension(int64_t dimension);
 
         /// \brief Construct a dynamic dimension.
         Dimension() { m_dimension = s_dynamic_val; }
@@ -46,25 +46,30 @@ namespace ngraph
         /// \brief Check whether this dimension is dynamic.
         /// \return `false` if the dimension is static, else `true`.
         bool is_dynamic() const { return !is_static(); }
-        /// \brief Convert this dimension to `size_t`. This dimension must be static.
+        /// \brief Convert this dimension to `int64_t`. This dimension must be static.
         /// \throws std::invalid_argument If this dimension is dynamic.
-        explicit operator size_t() const
+        explicit operator int64_t() const
         {
             if (is_dynamic())
             {
-                throw std::invalid_argument("Cannot convert dynamic dimension to size_t");
+                throw std::invalid_argument("Cannot convert dynamic dimension to int64_t");
             }
             return m_dimension;
         }
-        /// \brief Convert this dimension to `ptrdiff_t`. This dimension must be static.
-        /// \throws std::invalid_argument If this dimension is dynamic.
-        explicit operator ptrdiff_t() const
+        /// \brief Convert this dimension to `size_t`. This dimension must be static and
+        ///        non-negative.
+        /// \throws std::invalid_argument If this dimension is dynamic or negative.
+        explicit operator size_t() const
         {
             if (is_dynamic())
             {
-                throw std::invalid_argument("Cannot convert dynamic dimension to ptrdiff_t");
+                throw std::invalid_argument("Cannot convert dynamic dimension to size_t");
+            }
+            if (m_dimension < 0)
+            {
+                throw std::invalid_argument("Cannot convert negative dimension to size_t");
             }
-            return static_cast<ptrdiff_t>(m_dimension);
+            return m_dimension;
         }
 
         /// \brief Check whether this dimension represents the same scheme as the argument (both
@@ -75,7 +80,7 @@ namespace ngraph
         bool same_scheme(const Dimension& dim) const
         {
             return (is_dynamic() && dim.is_dynamic()) ||
-                   (is_static() && dim.is_static() && m_dimension == size_t(dim));
+                   (is_static() && dim.is_static() && m_dimension == int64_t(dim));
         }
 
         /// \brief Try to merge two Dimension objects together.
@@ -128,25 +133,25 @@ namespace ngraph
         /// \return A dynamic dimension.
         static Dimension dynamic() { return Dimension(); }
         /// \brief Constant for the value used internally to represent a dynamic dimension.
-        static const size_t s_dynamic_val{(std::numeric_limits<size_t>::max())};
+        static const int64_t s_dynamic_val{(std::numeric_limits<int64_t>::max())};
 
         /// \brief Addition operator for Dimension.
         /// \param dim Right operand for addition.
         /// \return Dimension::dynamic() if either of `*this` or `dim` is dynamic; else, a static
-        ///         dimension with value `size_t(*this)+size_t(dim)`.
+        ///         dimension with value `int64_t(*this)+in64_t(dim)`.
         Dimension operator+(const Dimension& dim) const;
 
         /// \brief Subtraction operator for Dimension.
         /// \param dim Right operand for subtraction.
         /// \return Dimension::dynamic() if either of `*this` or `dim` is dynamic; else, a static
-        ///         dimension with value `size_t(*this)-size_t(dim)`.
+        ///         dimension with value `int64_t(*this)-int64_t(dim)`.
         Dimension operator-(const Dimension& dim) const;
 
         /// \brief Multiplication operator for Dimension.
         /// \param dim Right operand for multiplicaiton.
         /// \return 0 if either of `*this` or `dim` is static and 0; else, Dimension::dynamic() if
         ///         either of `*this` or `dim` is dynamic; else, a static dimension with value
-        ///         `size_t(*this)*size_t(dim)`.
+        ///         `int64_t(*this)*int64_t(dim)`.
         Dimension operator*(const Dimension& dim) const;
 
         /// \brief Add-into operator for Dimension.
@@ -160,7 +165,7 @@ namespace ngraph
     private:
         // The actual numerical value of the dimension. s_dynamic_val is a special case,
         // representing a dynamic dimension.
-        size_t m_dimension;
+        int64_t m_dimension;
     };
 
     /// \brief Insert a human-readable representation of a dimension into an output stream.
@@ -168,6 +173,6 @@ namespace ngraph
     /// \param dimension The dimension to be inserted into `str`.
     /// \return A reference to `str` after insertion.
     ///
-    /// Inserts the string `?` if `dimension` is dynamic; else inserts `size_t(dimension)`.
+    /// Inserts the string `?` if `dimension` is dynamic; else inserts `int64_t(dimension)`.
     std::ostream& operator<<(std::ostream& str, const Dimension& dimension);
 }

src/ngraph/ngraph.hpp

@@ -89,6 +89,7 @@
 #include "ngraph/op/experimental/batch_mat_mul.hpp"
 #include "ngraph/op/experimental/dyn_broadcast.hpp"
 #include "ngraph/op/experimental/dyn_pad.hpp"
+#include "ngraph/op/experimental/dyn_replace_slice.hpp"
 #include "ngraph/op/experimental/dyn_reshape.hpp"
 #include "ngraph/op/experimental/dyn_slice.hpp"
 #include "ngraph/op/experimental/range.hpp"

src/ngraph/op/experimental/dyn_replace_slice.cpp

+//*****************************************************************************
+// Copyright 2017-2019 Intel Corporation
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//*****************************************************************************
+
+#include "ngraph/op/experimental/dyn_replace_slice.hpp"
+
+#include "ngraph/op/constant.hpp"
+#include "ngraph/validation_util.hpp"
+
+#include <memory>
+
+using namespace std;
+using namespace ngraph;
+
+op::DynReplaceSlice::DynReplaceSlice(const shared_ptr<Node>& arg,
+                                     const shared_ptr<Node>& replacement,
+                                     const shared_ptr<Node>& lower_bounds,
+                                     const shared_ptr<Node>& upper_bounds,
+                                     const shared_ptr<Node>& strides,
+                                     const AxisSet& lower_bounds_mask,
+                                     const AxisSet& upper_bounds_mask,
+                                     const AxisSet& new_axis,
+                                     const AxisSet& shrink_axis,
+                                     const AxisSet& ellipsis_mask)
+    : Op("DynReplaceSlice",
+         check_single_output_args({arg, replacement, lower_bounds, upper_bounds, strides}))
+    , m_lower_bounds_mask(lower_bounds_mask)
+    , m_upper_bounds_mask(upper_bounds_mask)
+    , m_new_axis(new_axis)
+    , m_shrink_axis(shrink_axis)
+    , m_ellipsis_mask(ellipsis_mask)
+{
+    constructor_validate_and_infer_types();
+}
+
+void op::DynReplaceSlice::validate_and_infer_types()
+{
+    auto arg_et = get_input_element_type(0);
+    auto replacement_et = get_input_element_type(1);
+    auto lower_bounds_et = get_input_element_type(2);
+    auto upper_bounds_et = get_input_element_type(3);
+    auto strides_et = get_input_element_type(4);
+
+    element::Type result_et;
+
+    // check data types
+    NODE_VALIDATION_CHECK(this,
+                          element::Type::merge(result_et, arg_et, replacement_et),
+                          "Argument element type is not compatible with replacement element type");
+
+    NODE_VALIDATION_CHECK(this,
+                          lower_bounds_et.compatible(element::Type_t::i64),
+                          "Lower bounds must have element type i64.");
+    NODE_VALIDATION_CHECK(this,
+                          upper_bounds_et.compatible(element::Type_t::i64),
+                          "Upper bounds must have element type i64.");
+    NODE_VALIDATION_CHECK(
+        this, strides_et.compatible(element::Type_t::i64), "Strides must have element type i64");
+
+    // check shapes
+    auto arg_shape = get_input_partial_shape(0);
+    auto replacement_shape = get_input_partial_shape(1);
+    auto lower_bounds_shape = get_input_partial_shape(2);
+    auto upper_bounds_shape = get_input_partial_shape(3);
+    auto strides_shape = get_input_partial_shape(4);
+    NODE_VALIDATION_CHECK(this,
+                          lower_bounds_shape.rank().compatible(1),
+                          "Lower bounds shape must have rank 1, got ",
+                          lower_bounds_shape.rank(),
+                          ".");
+    NODE_VALIDATION_CHECK(this,
+                          upper_bounds_shape.rank().compatible(1),
+                          "Upper bounds shape must have rank 1, got ",
+                          upper_bounds_shape.rank(),
+                          ".");
+    NODE_VALIDATION_CHECK(this,
+                          strides_shape.rank().compatible(1),
+                          "Strides shape must have rank 1, got ",
+                          strides_shape.rank(),
+                          ".");
+
+    PartialShape attrs_shape{PartialShape::dynamic()};
+
+    NODE_VALIDATION_CHECK(this,
+                          (lower_bounds_shape.same_scheme(PartialShape{0}) ||
+                           PartialShape::merge_into(attrs_shape, lower_bounds_shape)) &&
+                              (upper_bounds_shape.same_scheme(PartialShape{0}) ||
+                               PartialShape::merge_into(attrs_shape, upper_bounds_shape)) &&
+                              (strides_shape.same_scheme(PartialShape{0}) ||
+                               PartialShape::merge_into(attrs_shape, strides_shape)),
+                          "Shapes for lower bounds, upper bounds, and strides do not match");
+
+    set_input_is_relevant_to_shape(2);
+    set_input_is_relevant_to_shape(3);
+    set_input_is_relevant_to_shape(4);
+
+    auto lower_bounds = dynamic_pointer_cast<op::Constant>(get_argument(2));
+    auto upper_bounds = dynamic_pointer_cast<op::Constant>(get_argument(3));
+    auto strides = dynamic_pointer_cast<op::Constant>(get_argument(4));
+
+    // TODO(amprocte): We can get a bit more information here about the ranks of arg and
+    // replacement by inspecting the attributes.
+    auto slice_shape = PartialShape::dynamic();
+
+    if (lower_bounds && upper_bounds && strides)
+    {
+        slice_shape = infer_slice_shape(this,
+                                        get_input_partial_shape(0),
+                                        lower_bounds->get_vector<int64_t>(),
+                                        upper_bounds->get_vector<int64_t>(),
+                                        strides->get_vector<int64_t>(),
+                                        m_lower_bounds_mask,
+                                        m_upper_bounds_mask,
+                                        m_new_axis,
+                                        m_shrink_axis,
+                                        m_ellipsis_mask);
+    }
+
+    NODE_VALIDATION_CHECK(this,
+                          slice_shape.compatible(replacement_shape),
+                          "Shape of the replacement is not compatible with the shape of the "
+                          "slice (shape of slice: ",
+                          slice_shape,
+                          ")");
+
+    set_output_type(0, result_et, arg_shape);
+}
+
+shared_ptr<Node> op::DynReplaceSlice::copy_with_new_args(const NodeVector& new_args) const
+{
+    check_new_args_count(this, new_args);
+    return make_shared<DynReplaceSlice>(new_args.at(0),
+                                        new_args.at(1),
+                                        new_args.at(2),
+                                        new_args.at(3),
+                                        new_args.at(4),
+                                        m_lower_bounds_mask,
+                                        m_upper_bounds_mask,
+                                        m_new_axis,
+                                        m_shrink_axis,
+                                        m_ellipsis_mask);
+}
+
+void op::DynReplaceSlice::generate_adjoints(autodiff::Adjoints& adjoints, const NodeVector& deltas)
+{
+    throw ngraph_error("generate_adjoints not implemented for DynReplaceSlice");
+}

src/ngraph/op/experimental/dyn_replace_slice.hpp

+//*****************************************************************************
+// Copyright 2017-2019 Intel Corporation
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//*****************************************************************************
+
+#pragma once
+
+#include "ngraph/node.hpp"
+#include "ngraph/op/op.hpp"
+
+namespace ngraph
+{
+    namespace op
+    {
+        /// \brief Takes a slice of an input tensor, i.e., the sub-tensor that resides within a bounding box, optionally with stride.
+        class DynReplaceSlice : public Op
+        {
+        public:
+            /// \brief Constructs a dynamic tensor replace-slice operation.
+            ///
+            /// \param arg The tensor in which to replace the slice.
+            /// \param replacement Data to copy to the slice for replacement.
+            /// \param lower_bounds The axiswise lower bounds of the slice (inclusive).
+            /// \param upper_bounds The axiswise upper bounds of the slice (exclusive).
+            /// \param strides The slicing strides; for example, strides of `{n,m}` means to take
+            ///                every nth row and every mth column of the input matrix.
+            /// \param lower_bounds_mask Ignores lower_bounds for axis with the mask set
+            /// \param upper_bounds_mask Ignores upper_bounds for axis with the mask set
+            /// \param new_axis          Add dimension one axis at the set positions
+            /// \param shrink_axis       Delete dimensions at the set positions
+            /// \param ellipsis_mask     Inserts missing dimensions on the set position
+            DynReplaceSlice(const std::shared_ptr<Node>& arg,
+                            const std::shared_ptr<Node>& replacement,
+                            const std::shared_ptr<Node>& lower_bounds,
+                            const std::shared_ptr<Node>& upper_bounds,
+                            const std::shared_ptr<Node>& strides,
+                            const AxisSet& lower_bounds_mask = AxisSet{},
+                            const AxisSet& upper_bounds_mask = AxisSet{},
+                            const AxisSet& new_axis = AxisSet{},
+                            const AxisSet& shrink_axis = AxisSet{},
+                            const AxisSet& ellipsis_mask = AxisSet{});
+
+            const AxisSet& get_lower_bounds_mask() const { return m_lower_bounds_mask; }
+            const AxisSet& get_upper_bounds_mask() const { return m_upper_bounds_mask; }
+            const AxisSet& get_new_axis() const { return m_new_axis; }
+            const AxisSet& get_shrink_axis() const { return m_shrink_axis; }
+            const AxisSet& get_ellipsis_mask() const { return m_ellipsis_mask; }
+            virtual std::shared_ptr<Node>
+                copy_with_new_args(const NodeVector& new_args) const override;
+
+        protected:
+            virtual void generate_adjoints(autodiff::Adjoints& adjoints,
+                                           const NodeVector& deltas) override;
+            void validate_and_infer_types() override;
+
+        private:
+            /// Helper method to compute output shape
+            Shape compute_output_shape() const;
+
+            AxisSet m_lower_bounds_mask;
+            AxisSet m_upper_bounds_mask;
+            AxisSet m_new_axis;
+            AxisSet m_shrink_axis;
+            AxisSet m_ellipsis_mask;
+        };
+    }
+}

src/ngraph/op/experimental/dyn_slice.cpp

@@ -17,6 +17,7 @@
 #include "ngraph/op/experimental/dyn_slice.hpp"
 
 #include "ngraph/op/constant.hpp"
+#include "ngraph/validation_util.hpp"
 
 #include <memory>
 
@@ -42,142 +43,6 @@ op::DynSlice::DynSlice(const shared_ptr<Node>& arg,
     constructor_validate_and_infer_types();
 }
 
-Shape op::DynSlice::compute_output_shape() const
-{
-    auto input_shape = get_input_partial_shape(0).to_shape();
-    auto lower_bounds = dynamic_pointer_cast<op::Constant>(get_argument(1));
-    auto upper_bounds = dynamic_pointer_cast<op::Constant>(get_argument(2));
-    auto strides = dynamic_pointer_cast<op::Constant>(get_argument(3));
-
-    if (lower_bounds && upper_bounds && strides)
-    {
-        auto lb = lower_bounds->get_vector<int64_t>();
-        auto ub = upper_bounds->get_vector<int64_t>();
-        auto str = strides->get_vector<int64_t>();
-
-        int max_dims = input_shape.size() + m_new_axis.size();
-        if (lb.size() && ub.size())
-        {
-            NODE_VALIDATION_CHECK(
-                this,
-                lb.size() == ub.size(),
-                "Lower bounds and Upper bounds needs to have same number of values");
-        }
-        if (lb.size() && str.size())
-        {
-            NODE_VALIDATION_CHECK(this,
-                                  lb.size() == str.size(),
-                                  "Lower bounds and strides needs to have same number of values");
-        }
-        if (ub.size() && str.size())
-        {
-            NODE_VALIDATION_CHECK(this,
-                                  ub.size() == str.size(),
-                                  "Upper bounds and strides needs to have same number of values");
-        }
-
-        int bounds_size =
-            lb.size() ? lb.size() : (ub.size() ? ub.size() : (str.size() ? str.size() : 0));
-
-        NODE_VALIDATION_CHECK(
-            this, m_ellipsis_mask.size() <= 1, "Ellipsis mask cannot specify more than one axis");
-
-        int ellipsis_pos1 = m_ellipsis_mask.size() ? *m_ellipsis_mask.begin() : max_dims;
-
-        int ellipsis_pos2 = max_dims;
-        bounds_size -= ellipsis_pos1;
-        if (bounds_size > 0 && (max_dims - bounds_size) > ellipsis_pos1)
-        {
-            ellipsis_pos2 = max_dims - bounds_size;
-        }
-
-        std::vector<int> begin_dms(max_dims, 0);
-        std::vector<int> end_dms(max_dims, -1);
-        std::vector<int> stride_dms(max_dims, 1);
-
-        int i, j, k, bj, ej, sj;
-        Shape out_dims;
-        for (i = 0, j = 0, k = 0, bj = 0, ej = 0, sj = 0; i < max_dims; i++)
-        {
-            if (i >= ellipsis_pos1 && i < ellipsis_pos2)
-            {
-                if (m_new_axis.find(i) == m_new_axis.end())
-                {
-                    end_dms[i] = end_dms[i] >= 0 ? end_dms[i] : input_shape[j++] + end_dms[i];
-                }
-                else
-                {
-                    end_dms[i] = begin_dms[i];
-                }
-                out_dims.push_back(
-                    static_cast<int>(ceil(static_cast<float>(abs(end_dms[i] - begin_dms[i]) + 1) /
-                                          static_cast<float>(abs(stride_dms[i])))));
-                k = ellipsis_pos1;
-                continue;
-            }
-            stride_dms[i] = (str.size() > sj && str[sj] != 0) ? str[sj++] : 1;
-
-            // Use lower_bounds if mask is not set
-            if (m_lower_bounds_mask.find(j) == m_lower_bounds_mask.end())
-            {
-                begin_dms[i] = lb.size() > bj ? lb[bj] : (stride_dms[i] > 0 ? 0 : -1);
-            }
-            else
-            {
-                begin_dms[i] = stride_dms[i] > 0 ? 0 : -1;
-            }
-            bj++;
-
-            begin_dms[i] = begin_dms[i] >= 0 ? begin_dms[i] : input_shape[j] + begin_dms[i];
-            //  Clipping 'begin'
-            begin_dms[i] =
-                (begin_dms[i] < 0) ? 0 : (begin_dms[i] >= input_shape[j] ? input_shape[j] - 1
-                                                                         : begin_dms[i]);
-
-            // Use upper_bounds if mask is not set
-            if (m_upper_bounds_mask.find(j) == m_upper_bounds_mask.end())
-            {
-                int end_dms_tmp =
-                    ub.size() > ej ? (stride_dms[i] > 0 ? ub[ej] - 1 : ub[ej] + 1) : end_dms[i];
-                end_dms[i] = ub.size() > ej ? end_dms_tmp : (stride_dms[i] > 0 ? -1 : 0);
-            }
-            else
-            {
-                end_dms[i] = stride_dms[i] > 0 ? -1 : 0;
-            }
-            ej++;
-            end_dms[i] = end_dms[i] >= 0 ? end_dms[i] : input_shape[j] + end_dms[i];
-            //  Clipping 'end'
-            end_dms[i] = (end_dms[i] < 0) ? 0 : (end_dms[i] >= input_shape[j] ? input_shape[j] - 1
-                                                                              : end_dms[i]);
-
-            if (m_new_axis.find(i) == m_new_axis.end())
-            {
-                j++;
-            }
-            else
-            {
-                end_dms[i] = 0;
-            }
-
-            if (m_shrink_axis.find(k) != m_shrink_axis.end())
-            {
-                end_dms[i] = begin_dms[i];
-            }
-            else
-            {
-                out_dims.push_back(
-                    static_cast<int>(ceil(static_cast<float>(abs(end_dms[i] - begin_dms[i]) + 1) /
-                                          static_cast<float>(abs(stride_dms[i])))));
-            }
-
-            k++;
-        }
-        return out_dims;
-    }
-    return Shape{};
-}
-
 void op::DynSlice::validate_and_infer_types()
 {
     auto lower_bounds_et = get_input_element_type(1);
@@ -219,17 +84,24 @@ void op::DynSlice::validate_and_infer_types()
     set_input_is_relevant_to_shape(2);
     set_input_is_relevant_to_shape(3);
 
-    if (get_input_partial_shape(0).is_static())
-    {
-        auto shape = compute_output_shape();
-        if (shape != Shape{})
-        {
-            set_output_type(0, get_input_element_type(0), shape);
-        }
-        else
+    auto lower_bounds = dynamic_pointer_cast<op::Constant>(get_argument(1));
+    auto upper_bounds = dynamic_pointer_cast<op::Constant>(get_argument(2));
+    auto strides = dynamic_pointer_cast<op::Constant>(get_argument(3));
+
+    if (lower_bounds && upper_bounds && strides)
     {
-            set_output_type(0, get_input_element_type(0), PartialShape::dynamic(arg_shape.rank()));
-        }
+        set_output_type(0,
+                        get_input_element_type(0),
+                        infer_slice_shape(this,
+                                          get_input_partial_shape(0),
+                                          lower_bounds->get_vector<int64_t>(),
+                                          upper_bounds->get_vector<int64_t>(),
+                                          strides->get_vector<int64_t>(),
+                                          m_lower_bounds_mask,
+                                          m_upper_bounds_mask,
+                                          m_new_axis,
+                                          m_shrink_axis,
+                                          m_ellipsis_mask));
     }
     else
     {

src/ngraph/op/op_tbl.hpp

@@ -84,6 +84,7 @@ NGRAPH_OP(Divide, ngraph::op)
 NGRAPH_OP(Dot, ngraph::op)
 NGRAPH_OP(DynBroadcast, ngraph::op)
 NGRAPH_OP(DynPad, ngraph::op)
+NGRAPH_OP(DynReplaceSlice, ngraph::op)
 NGRAPH_OP(DynReshape, ngraph::op)
 NGRAPH_OP(DynSlice, ngraph::op)
 NGRAPH_OP(EmbeddingLookup, ngraph::op)

src/ngraph/op/pad.cpp

@@ -84,7 +84,7 @@ void op::Pad::validate_and_infer_types()
             if (arg_shape[i].is_static())
             {
                 ptrdiff_t result_dim =
-                    m_padding_below[i] + static_cast<ptrdiff_t>(arg_shape[i]) + m_padding_above[i];
+                    m_padding_below[i] + static_cast<int64_t>(arg_shape[i]) + m_padding_above[i];
                 NODE_VALIDATION_CHECK(this,
                                       result_dim >= 0,
                                       "Inferred result dimension at axis ",

src/ngraph/partial_shape.cpp

@@ -275,3 +275,16 @@ bool PartialShape::broadcast_merge_into(PartialShape& dst,
         return success;
     }
 }
+
+bool PartialShape::all_non_negative() const
+{
+    for (auto& d : m_dimensions)
+    {
+        if (d.is_static() && int64_t(d) < 0)
+        {
+            return false;
+        }
+    }
+
+    return true;
+}

src/ngraph/partial_shape.hpp

@@ -164,6 +164,10 @@ namespace ngraph
         /// \throws std::invalid_argument If this PartialShape is dynamic.
         Shape to_shape() const;
 
+        /// \brief Returns `true` if all static dimensions of the tensor are non-negative, else
+        ///        `false`.
+        bool all_non_negative() const;
+
         /// \brief Index operator for PartialShape.
         /// \param i The index of the dimension being selected.
         /// \return A reference to the `i`th Dimension of this shape.

src/ngraph/pass/dyn_elimination.cpp

@@ -19,10 +19,12 @@
 #include "dyn_elimination.hpp"
 #include "ngraph/op/broadcast.hpp"
 #include "ngraph/op/experimental/dyn_broadcast.hpp"
+#include "ngraph/op/experimental/dyn_replace_slice.hpp"
 #include "ngraph/op/experimental/dyn_reshape.hpp"
 #include "ngraph/op/experimental/dyn_slice.hpp"
 #include "ngraph/op/experimental/range.hpp"
 #include "ngraph/op/experimental/transpose.hpp"
+#include "ngraph/op/replace_slice.hpp"
 #include "ngraph/op/reshape.hpp"
 #include "ngraph/op/reverse.hpp"
 #include "ngraph/op/slice.hpp"
@@ -36,7 +38,8 @@ pass::DynElimination::DynElimination()
     : GraphRewrite()
 {
     construct_transpose();
-    construct_broadcast();
+    construct_dyn_broadcast();
+    construct_dyn_replace_slice();
     construct_dyn_slice();
     construct_dyn_reshape();
     construct_range();
@@ -89,7 +92,7 @@ void pass::DynElimination::construct_transpose()
     add_matcher(transpose_matcher, transpose_callback, all_pass_property_off);
 }
 
-void pass::DynElimination::construct_broadcast()
+void pass::DynElimination::construct_dyn_broadcast()
 {
     auto data_arg_label = make_shared<pattern::op::Label>(element::f32, Shape{1, 2, 3});
     auto shape_arg_label =
@@ -444,6 +447,92 @@ void pass::DynElimination::construct_dyn_slice()
     add_matcher(dyn_slice_matcher, dyn_slice_callback, all_pass_property_off);
 }
 
+void pass::DynElimination::construct_dyn_replace_slice()
+{
+    auto data_arg_label = make_shared<pattern::op::Label>(element::f32, Shape{1, 2, 3});
+    auto replacement_arg_label = make_shared<pattern::op::Label>(element::f32, Shape{1, 2, 3});
+    auto begins_arg_label =
+        make_shared<pattern::op::Label>(element::i64, Shape{3}, pattern::has_class<op::Constant>());
+    auto ends_arg_label =
+        make_shared<pattern::op::Label>(element::i64, Shape{3}, pattern::has_class<op::Constant>());
+    auto strides_arg_label =
+        make_shared<pattern::op::Label>(element::i64, Shape{3}, pattern::has_class<op::Constant>());
+
+    auto dyn_replace_slice_pat = make_shared<op::DynReplaceSlice>(data_arg_label,
+                                                                  replacement_arg_label,
+                                                                  begins_arg_label,
+                                                                  ends_arg_label,
+                                                                  strides_arg_label,
+                                                                  AxisSet{},
+                                                                  AxisSet{},
+                                                                  AxisSet{},
+                                                                  AxisSet{},
+                                                                  AxisSet{});
+    auto dyn_replace_slice_callback = [data_arg_label,
+                                       replacement_arg_label,
+                                       begins_arg_label,
+                                       ends_arg_label,
+                                       strides_arg_label](pattern::Matcher& m) {
+        auto pattern_map = m.get_pattern_map();
+
+        auto data_arg = pattern_map[data_arg_label];
+        auto replacement_arg = pattern_map[replacement_arg_label];
+        auto begins_arg = static_pointer_cast<op::Constant>(pattern_map[begins_arg_label]);
+        auto ends_arg = static_pointer_cast<op::Constant>(pattern_map[ends_arg_label]);
+        auto strides_arg = static_pointer_cast<op::Constant>(pattern_map[strides_arg_label]);
+        auto dyn_replace_slice = static_pointer_cast<op::DynReplaceSlice>(m.get_match_root());
+
+        if (data_arg->get_output_partial_shape(0).is_dynamic() ||
+            replacement_arg->get_output_partial_shape(0).is_dynamic() ||
+            begins_arg->get_element_type() != element::i64 ||
+            ends_arg->get_element_type() != element::i64 ||
+            strides_arg->get_element_type() != element::i64)
+        {
+            return false;
+        }
+
+        SlicePlan p = make_plan(data_arg->get_output_shape(0),
+                                begins_arg->get_vector<int64_t>(),
+                                ends_arg->get_vector<int64_t>(),
+                                strides_arg->get_vector<int64_t>(),
+                                dyn_replace_slice->get_lower_bounds_mask(),
+                                dyn_replace_slice->get_upper_bounds_mask(),
+                                dyn_replace_slice->get_new_axis(),
+                                dyn_replace_slice->get_shrink_axis(),
+                                dyn_replace_slice->get_ellipsis_mask());
+
+        shared_ptr<Node> substitute_replacement_arg = replacement_arg;
+
+        if (!p.reverse_axes.empty())
+        {
+            substitute_replacement_arg =
+                make_shared<op::Reverse>(substitute_replacement_arg, p.reverse_axes);
+        }
+
+        if (p.reshape_in_shape != p.reshape_out_shape)
+        {
+            substitute_replacement_arg =
+                make_shared<op::Reshape>(substitute_replacement_arg,
+                                         ngraph::get_default_order(p.reshape_out_shape),
+                                         p.reshape_in_shape);
+        }
+
+        auto substitute_rsl =
+            make_shared<op::ReplaceSlice>(data_arg,
+                                          substitute_replacement_arg,
+                                          Coordinate(p.begins.begin(), p.begins.end()),
+                                          Coordinate(p.ends.begin(), p.ends.end()),
+                                          Strides(p.strides.begin(), p.strides.end()));
+
+        replace_node(m.get_match_root(), substitute_rsl);
+        return true;
+    };
+
+    auto dyn_replace_slice_matcher =
+        make_shared<pattern::Matcher>(dyn_replace_slice_pat, "DynElimination.DynReplaceShape");
+    add_matcher(dyn_replace_slice_matcher, dyn_replace_slice_callback, all_pass_property_off);
+}
+
 void pass::DynElimination::construct_dyn_reshape()
 {
     auto data_arg_label = make_shared<pattern::op::Label>(element::f32, Shape{1, 2, 3});

src/ngraph/pass/dyn_elimination.hpp

@@ -30,7 +30,8 @@ namespace ngraph
 
         private:
             void construct_transpose();
-            void construct_broadcast();
+            void construct_dyn_broadcast();
+            void construct_dyn_replace_slice();
             void construct_dyn_slice();
             void construct_dyn_reshape();
             void construct_range();

src/ngraph/runtime/cpu/cpu_call_frame.cpp

@@ -216,8 +216,6 @@ void runtime::cpu::CPU_CallFrame::setup_runtime_context()
         {
             // single thread for codegen
             NGRAPH_CHECK(m_num_ctx == 1);
-            ctx->mkldnn_primitives.swap(mkldnn_emitter->get_mkldnn_primitives());
-            ctx->mkldnn_workspaces = mkldnn_emitter->get_mkldnn_workspaces();
         }
 
         ctx->states = m_external_function->m_states.data();

src/ngraph/runtime/cpu/cpu_external_function.cpp

@@ -33,7 +33,10 @@
 #include "ngraph/codegen/execution_engine.hpp"
 #endif
 
+#ifdef NGRAPH_MLIR_ENABLE
 #include "contrib/mlir/pass/mlir_subgraph_extraction.hpp"
+#endif
+
 #include "ngraph/descriptor/input.hpp"
 #include "ngraph/descriptor/output.hpp"
 #include "ngraph/file_util.hpp"

src/ngraph/runtime/gpu/gpu_backend.cpp

@@ -218,6 +218,7 @@ bool runtime::gpu::GPU_Backend::is_supported(const Node& op) const
 {
     set<string> unsupported_ops = {"Quantize",
                                    "Dequantize",
+                                   "DynReplaceSlice",
                                    "DynReshape",
                                    "DynSlice",
                                    "ShapeOf",

src/ngraph/runtime/gpu/gpu_compiled_function.cpp

@@ -173,9 +173,6 @@ void runtime::gpu::GPUCompiledFunction::compile()
     pass_manager.register_pass<runtime::gpu::pass::BatchNormCache>();
     pass_manager.register_pass<ngraph::pass::LikeReplacement>();
     pass_manager.register_pass<ngraph::pass::FusedOpDecomposition>();
-    // Run this pass for the second time since, some fused operators like LSTMCell may use
-    // other fused operators inside.
-    pass_manager.register_pass<ngraph::pass::FusedOpDecomposition>();
     pass_manager.register_pass<ngraph::pass::ImplicitBroadcastElimination>();
     pass_manager.register_pass<runtime::gpu::pass::GPULayout>(this);
     pass_manager.register_pass<ngraph::pass::AssignLayout<descriptor::layout::DenseTensorLayout>>();

src/ngraph/runtime/gpu/gpu_emitter.cpp

@@ -62,6 +62,7 @@
 #include "ngraph/op/experimental/batch_mat_mul.hpp"
 #include "ngraph/op/experimental/dyn_broadcast.hpp"
 #include "ngraph/op/experimental/dyn_pad.hpp"
+#include "ngraph/op/experimental/dyn_replace_slice.hpp"
 #include "ngraph/op/experimental/dyn_reshape.hpp"
 #include "ngraph/op/experimental/dyn_slice.hpp"
 #include "ngraph/op/experimental/generate_mask.hpp"
@@ -612,6 +613,11 @@ std::string runtime::gpu::GPU_Emitter::emit_Dot(EMIT_ARGS)
     return compiled_function->add_to_runtime(index, function_name, args, out);
 }
 
+std::string runtime::gpu::GPU_Emitter::emit_DynReplaceSlice(EMIT_ARGS)
+{
+    throw unsupported_op("Unsupported op '" + node->description() + "'");
+}
+
 std::string runtime::gpu::GPU_Emitter::emit_DynReshape(EMIT_ARGS)
 {
     throw unsupported_op("Unsupported op '" + node->description() + "'");

src/ngraph/runtime/intelgpu/intelgpu_backend.cpp

@@ -430,10 +430,6 @@ shared_ptr<runtime::Executable>
 
     if (m_disable_backend_optimizations < 2)
     {
-        pass_manager.register_pass<ngraph::pass::FusedOpDecomposition>(
-            IntelGPUBackend::is_supported_impl);
-        // Run this pass for the second time since, some fused operators like LSTMCell may use
-        // other fused operators inside.
         pass_manager.register_pass<ngraph::pass::FusedOpDecomposition>(
             IntelGPUBackend::is_supported_impl);
         pass_manager.register_pass<ngraph::pass::ImplicitBroadcastElimination>();
@@ -2061,6 +2057,7 @@ shared_ptr<runtime::Executable>
         case OP_TYPEID::DepthToSpace:
         case OP_TYPEID::DynBroadcast:
         case OP_TYPEID::DynPad:
+        case OP_TYPEID::DynReplaceSlice:
         case OP_TYPEID::DynReshape:
         case OP_TYPEID::DynSlice:
         case OP_TYPEID::Elu:

src/ngraph/runtime/intelgpu/unit_test.manifest

@@ -18,6 +18,7 @@ replace_slice_matrix
 replace_slice_matrix_inplace
 replace_slice_scalar
 replace_slice_vector
+dyn_replace_slice
 shape_of_5d
 shape_of_matrix
 shape_of_scalar

src/ngraph/runtime/interpreter/int_executable.cpp

@@ -47,9 +47,6 @@ runtime::interpreter::INTExecutable::INTExecutable(const shared_ptr<Function>& f
     pass::Manager pass_manager;
     pass_manager.register_pass<pass::LikeReplacement>();
     pass_manager.register_pass<pass::FusedOpDecomposition>();
-    // Run this pass for the second time since, some fused operators like LSTMCell may use
-    // other fused operators inside.
-    pass_manager.register_pass<pass::FusedOpDecomposition>();
     pass_manager.register_pass<pass::ImplicitBroadcastElimination>();
     pass_manager.register_pass<pass::AssignLayout<DenseTensorLayout>>();
     pass_manager.register_pass<pass::Liveness>();

src/ngraph/runtime/interpreter/int_executable.hpp

@@ -1503,7 +1503,8 @@ private:
         case OP_TYPEID::Transpose:
         case OP_TYPEID::DynPad:
         case OP_TYPEID::Tile:
-        default: throw unsupported_op("Unsupported op '" + node.description() + "'");
+        case OP_TYPEID::DynReplaceSlice:
+            throw unsupported_op("Unsupported op '" + node.description() + "'");
 #if !(defined(__GNUC__) && (__GNUC__ == 4 && __GNUC_MINOR__ == 8))
 #pragma GCC diagnostic pop
 #endif

src/ngraph/runtime/plaidml/CMakeLists.txt

@@ -41,6 +41,7 @@ set(SRC
     plaidml_ops_one_hot.cpp
     plaidml_ops_passthrough.cpp
     plaidml_ops_pool.cpp
+    plaidml_ops_quantize.cpp
     plaidml_ops_reduce.cpp
     plaidml_ops_replace_slice.cpp
     plaidml_ops_replicate.cpp

src/ngraph/runtime/plaidml/plaidml_builder.hpp

@@ -188,7 +188,8 @@ class ngraph::runtime::plaidml::builder::Elementwise final : public Statement
 {
 public:
     Elementwise(std::string lhs, std::string rhs);
-
+    void set_lhs(const std::string& lhs) { m_lhs = lhs; }
+    void set_rhs(const std::string& rhs) { m_rhs = rhs; }
 private:
     friend class Function;
 

src/ngraph/runtime/plaidml/plaidml_compiler.cpp

@@ -20,6 +20,7 @@
 #include "ngraph/pass/algebraic_simplification.hpp"
 #include "ngraph/pass/core_fusion.hpp"
 #include "ngraph/pass/cse.hpp"
+#include "ngraph/pass/fused_op_decomposition.hpp"
 #include "ngraph/pass/get_output_element_elimination.hpp"
 #include "ngraph/pass/like_replacement.hpp"
 #include "ngraph/pass/liveness.hpp"
@@ -87,6 +88,7 @@ std::shared_ptr<ngraph::runtime::plaidml::PlaidML_Executable>
     ngraph::pass::Manager pass_manager;
 
     // We apply the same general-purposes passes as the CPU backend.
+    pass_manager.register_pass<ngraph::pass::FusedOpDecomposition>();
     pass_manager.register_pass<ngraph::pass::LikeReplacement>();
     pass_manager.register_pass<ngraph::pass::NopElimination>();
     pass_manager.register_pass<ngraph::pass::ZeroDimTensorElimination>();

src/ngraph/runtime/plaidml/unit_test.manifest

@@ -60,37 +60,7 @@ generate_mask
 generate_mask2
 avg_pool_3d
 avg_pool_3d_uneven_strided_padded_include_in_computation
-quantize_dynamic_offset                 # Quantization/Dequantization is unimplemented
-dequantize_dynamic_offset               # Quantization/Dequantization is unimplemented
-dequantize_int8_zero_offset             # Quantization/Dequantization is unimplemented
-dequantize_int32                        # Quantization/Dequantization is unimplemented
-dequantize_int32_zero_offset            # Quantization/Dequantization is unimplemented
-dequantize_zero_offset                  # Quantization/Dequantization is unimplemented
-quantize_ROUND_NEAREST_TOWARD_ZERO      # Quantization/Dequantization is unimplemented
-quantize_ROUND_NEAREST_UPWARD           # Quantization/Dequantization is unimplemented
-quantize_ROUND_NEAREST_DOWNWARD         # Quantization/Dequantization is unimplemented
-quantize_ROUND_NEAREST_TOWARD_EVEN      # Quantization/Dequantization is unimplemented
-quantize_ROUND_NEAREST_TOWARD_INFINITY  # Quantization/Dequantization is unimplemented
-quantize_ROUND_TOWARD_INFINITY          # Quantization/Dequantization is unimplemented
-quantize_ROUND_TOWARD_ZERO              # Quantization/Dequantization is unimplemented
-quantize_ROUND_UP                       # Quantization/Dequantization is unimplemented
-quantize_ROUND_DOWN                     # Quantization/Dequantization is unimplemented
-quantize                                # Quantization/Dequantization is unimplemented
-quantize_zero_offset                    # Quantization/Dequantization is unimplemented
-quantize_axes                           # Quantization/Dequantization is unimplemented
-quantize_dynamic_offset                 # Quantization/Dequantization is unimplemented
-quantize_int8                           # Quantization/Dequantization is unimplemented
-quantize_int8_zero_offset               # Quantization/Dequantization is unimplemented
-quantize_int32                          # Quantization/Dequantization is unimplemented
-quantize_int32_zero_offset              # Quantization/Dequantization is unimplemented
-quantize_clamp                          # Quantization/Dequantization is unimplemented
-quantize_clamp_int8                     # Quantization/Dequantization is unimplemented
-quantize_clamp_int32                    # Quantization/Dequantization is unimplemented
-quantize_clamp_int32_zero_offset        # Quantization/Dequantization is unimplemented
-quantize_clamp_uint8                    # Quantization/Dequantization is unimplemented
-dequantize                              # Quantization/Dequantization is unimplemented
-dequantize_axes                         # Quantization/Dequantization is unimplemented
-dequantize_int8                         # Quantization/Dequantization is unimplemented
+quantize_clamp_int32                    # Requires fp64 inputs, which won't work on GPUs
 numeric_float_nan
 numeric_double_nan
 shape_of_scalar
@@ -259,12 +229,6 @@ backwards_softmax_underflow
 backwards_softmax_3d
 batch_mat_mul_forward
 dot_matrix_2x0_0x2
-rnn_cell_no_bias
-rnn_cell_bias_clip
-rnn_cell_activation_function
-gru_cell_bias_clip
-gru_cell_linear_before_reset
-gru_cell_activation_function
 
 # dgkutnic ww24.5: these tests are to be triaged by the PlaidML team
 # ww25.2: re-scrubbed this list of tests after fixing check_inputs
@@ -289,3 +253,29 @@ group_conv_transpose
 group_conv_transpose_output_shape
 divide_python_rounding_int32
 backwards_batchmatmul_tensor2_tensor2
+
+# unsupported ops: 'QuantizedConvolution', 'QuantizedDot', 'TopK', 'Erf', 'EmbeddingLookup'
+model_quant_conv_linear
+model_conv_integer_no_zero_point
+model_matmul_integer_no_zero_point
+model_matmul_integer_4d_no_zero_point
+model_top_k
+model_erf
+model_erf_int32
+model_hardmax
+
+# node validation error: "Argument shapes are inconsistent."
+model_lstm_fwd_with_clip
+model_lstm_fwd_mixed_seq
+model_lstm_fwd_hardsigmoid_activation
+model_reduce_log_sum
+model_reduce_log_sum_exp
+model_reduce_mean
+
+# result mismatch
+model_dequantize_linear_scalar_zero_scale_int8
+model_softmax
+avg_pool_3d_uneven_strided_padded
+rnn_cell_activation_function
+gru_cell_bias_clip
+gru_cell_linear_before_reset

src/ngraph/serializer.cpp

@@ -55,6 +55,7 @@
 #include "ngraph/op/experimental/compiled_kernel.hpp"
 #include "ngraph/op/experimental/dyn_broadcast.hpp"
 #include "ngraph/op/experimental/dyn_pad.hpp"
+#include "ngraph/op/experimental/dyn_replace_slice.hpp"
 #include "ngraph/op/experimental/dyn_reshape.hpp"
 #include "ngraph/op/experimental/dyn_slice.hpp"
 #include "ngraph/op/experimental/generate_mask.hpp"
@@ -1130,6 +1131,25 @@ shared_ptr<Node> JSONDeserializer::deserialize_node(json node_js)
             node = make_shared<op::DynPad>(args[0], args[1], args[2], args[3]);
             break;
         }
+        case OP_TYPEID::DynReplaceSlice:
+        {
+            auto lower_bounds_mask = node_js.at("lower_bounds_mask").get<set<size_t>>();
+            auto upper_bounds_mask = node_js.at("upper_bounds_mask").get<set<size_t>>();
+            auto new_axis = node_js.at("new_axis").get<set<size_t>>();
+            auto shrink_axis = node_js.at("shrink_axis").get<set<size_t>>();
+            auto ellipsis_mask = node_js.at("ellipsis_mask").get<set<size_t>>();
+            node = make_shared<op::DynReplaceSlice>(args[0],
+                                                    args[1],
+                                                    args[2],
+                                                    args[3],
+                                                    args[4],
+                                                    lower_bounds_mask,
+                                                    upper_bounds_mask,
+                                                    new_axis,
+                                                    shrink_axis,
+                                                    ellipsis_mask);
+            break;
+        }
         case OP_TYPEID::DynReshape:
         {
             node = make_shared<op::DynReshape>(args[0], args[1]);
@@ -1137,7 +1157,20 @@ shared_ptr<Node> JSONDeserializer::deserialize_node(json node_js)
         }
         case OP_TYPEID::DynSlice:
         {
-            node = make_shared<op::DynSlice>(args[0], args[1], args[2], args[3]);
+            auto lower_bounds_mask = node_js.at("lower_bounds_mask").get<set<size_t>>();
+            auto upper_bounds_mask = node_js.at("upper_bounds_mask").get<set<size_t>>();
+            auto new_axis = node_js.at("new_axis").get<set<size_t>>();
+            auto shrink_axis = node_js.at("shrink_axis").get<set<size_t>>();
+            auto ellipsis_mask = node_js.at("ellipsis_mask").get<set<size_t>>();
+            node = make_shared<op::DynSlice>(args[0],
+                                             args[1],
+                                             args[2],
+                                             args[3],
+                                             lower_bounds_mask,
+                                             upper_bounds_mask,
+                                             new_axis,
+                                             shrink_axis,
+                                             ellipsis_mask);
             break;
         }
         case OP_TYPEID::Elu:
@@ -2297,9 +2330,27 @@ json JSONSerializer::serialize_node(const Node& n)
     }
     case OP_TYPEID::DynPad: { break;
     }
+    case OP_TYPEID::DynReplaceSlice:
+    {
+        auto tmp = dynamic_cast<const op::DynReplaceSlice*>(&n);
+        node["lower_bounds_mask"] = tmp->get_lower_bounds_mask();
+        node["upper_bounds_mask"] = tmp->get_upper_bounds_mask();
+        node["new_axis"] = tmp->get_new_axis();
+        node["shrink_axis"] = tmp->get_shrink_axis();
+        node["ellipsis_mask"] = tmp->get_ellipsis_mask();
+        break;
+    }
     case OP_TYPEID::DynReshape: { break;
     }
-    case OP_TYPEID::DynSlice: { break;
+    case OP_TYPEID::DynSlice:
+    {
+        auto tmp = dynamic_cast<const op::DynSlice*>(&n);
+        node["lower_bounds_mask"] = tmp->get_lower_bounds_mask();
+        node["upper_bounds_mask"] = tmp->get_upper_bounds_mask();
+        node["new_axis"] = tmp->get_new_axis();
+        node["shrink_axis"] = tmp->get_shrink_axis();
+        node["ellipsis_mask"] = tmp->get_ellipsis_mask();
+        break;
     }
     case OP_TYPEID::Elu: { break;
     }

src/ngraph/validation_util.cpp

@@ -142,8 +142,8 @@ PartialShape ngraph::infer_windowed_reduction_output_shape(const Node* node,
             ptrdiff_t data_padded_dilated_dim = -1;
             if (data_dim_static)
             {
-                data_padded_dilated_dim = (static_cast<ptrdiff_t>(data_dilation[i]) *
-                                           (static_cast<ptrdiff_t>(data_shape[i]) - 1)) +
+                data_padded_dilated_dim = (static_cast<int64_t>(data_dilation[i]) *
+                                           (static_cast<int64_t>(data_shape[i]) - 1)) +
                                           1 + data_padding_below[i] + data_padding_above[i];
                 NODE_VALIDATION_CHECK(
                     node,
@@ -158,8 +158,8 @@ PartialShape ngraph::infer_windowed_reduction_output_shape(const Node* node,
             ptrdiff_t window_dilated_dim = -1;
             if (window_dim_static)
             {
-                window_dilated_dim = static_cast<ptrdiff_t>(window_dilation[i]) *
-                                         (static_cast<ptrdiff_t>(window_shape[i]) - 1) +
+                window_dilated_dim = static_cast<int64_t>(window_dilation[i]) *
+                                         (static_cast<int64_t>(window_shape[i]) - 1) +
                                      1;
 
                 NODE_VALIDATION_CHECK(node,
@@ -628,3 +628,257 @@ void ngraph::infer_auto_padding(const Shape& image_shape,
         padding_above.push_back(pad_type == op::PadType::SAME_UPPER ? padding_rhs : padding_lhs);
     }
 }
+
+PartialShape ngraph::infer_slice_shape(const Node* node,
+                                       const PartialShape& input_shape,
+                                       const std::vector<int64_t>& lb,
+                                       const std::vector<int64_t>& ub,
+                                       const std::vector<int64_t>& str,
+                                       const AxisSet& lb_mask,
+                                       const AxisSet& ub_mask,
+                                       const AxisSet& new_axis,
+                                       const AxisSet& shrink_axis,
+                                       const AxisSet& ellipsis_mask)
+{
+    if (lb.size() && ub.size())
+    {
+        NODE_VALIDATION_CHECK(node,
+                              lb.size() == ub.size(),
+                              "Lower bounds and Upper bounds needs to have same number of values");
+    }
+    if (lb.size() && str.size())
+    {
+        NODE_VALIDATION_CHECK(node,
+                              lb.size() == str.size(),
+                              "Lower bounds and strides needs to have same number of values");
+    }
+    if (ub.size() && str.size())
+    {
+        NODE_VALIDATION_CHECK(node,
+                              ub.size() == str.size(),
+                              "Upper bounds and strides needs to have same number of values");
+    }
+
+    if (input_shape.rank().is_dynamic())
+    {
+        return PartialShape::dynamic();
+    }
+
+    int max_dims = size_t(input_shape.rank()) + new_axis.size();
+
+    int bounds_size =
+        lb.size() ? lb.size() : (ub.size() ? ub.size() : (str.size() ? str.size() : 0));
+
+    int ellipsis_pos1 = ellipsis_mask.size() ? *ellipsis_mask.begin() : max_dims;
+
+    int ellipsis_pos2 = max_dims;
+    bounds_size -= ellipsis_pos1;
+    if (bounds_size > 0 && (max_dims - bounds_size) > ellipsis_pos1)
+    {
+        ellipsis_pos2 = max_dims - bounds_size;
+    }
+
+    std::vector<Dimension> begin_dms(max_dims, 0);
+    std::vector<Dimension> end_dms(max_dims, -1);
+    std::vector<Dimension> stride_dms(max_dims, 1);
+
+    std::vector<Dimension> out_dims;
+
+    int j = 0;
+    int k = 0;
+    int bj = 0;
+    int ej = 0;
+    int sj = 0;
+
+    for (int i = 0; i < max_dims; i++)
+    {
+        if (i >= ellipsis_pos1 && i < ellipsis_pos2)
+        {
+            if (new_axis.find(i) == new_axis.end())
+            {
+                if (end_dms[i].is_static() && int64_t(end_dms[i]) < 0)
+                {
+                    end_dms[i] = input_shape[j++] + end_dms[i];
+                }
+            }
+            else
+            {
+                end_dms[i] = begin_dms[i];
+            }
+
+            if (end_dms[i].is_dynamic() || begin_dms[i].is_dynamic() || stride_dms[i].is_dynamic())
+            {
+                out_dims.push_back(Dimension::dynamic());
+            }
+            else
+            {
+                out_dims.push_back(static_cast<int64_t>(
+                    ceil(static_cast<float>(abs(int64_t(end_dms[i]) - int64_t(begin_dms[i])) + 1) /
+                         static_cast<float>(abs(int64_t(stride_dms[i]))))));
+            }
+            k = ellipsis_pos1;
+            continue;
+        }
+        stride_dms[i] = (str.size() > sj && str[sj] != 0) ? str[sj++] : 1;
+
+        // Use lower_bounds if mask is not set
+        if (lb_mask.find(j) == lb_mask.end())
+        {
+            if (lb.size() > bj)
+            {
+                begin_dms[i] = lb[bj];
+            }
+            else if (stride_dms[i].is_dynamic())
+            {
+                begin_dms[i] = Dimension::dynamic();
+            }
+            else if (int64_t(stride_dms[i]) > 0)
+            {
+                begin_dms[i] = 0;
+            }
+            else
+            {
+                begin_dms[i] = -1;
+            }
+        }
+        else if (stride_dms[i].is_dynamic())
+        {
+            begin_dms[i] = Dimension::dynamic();
+        }
+        else if (int64_t(stride_dms[i]) > 0)
+        {
+            begin_dms[i] = 0;
+        }
+        else
+        {
+            begin_dms[i] = -1;
+        }
+
+        bj++;
+
+        if (begin_dms[i].is_static() && int64_t(begin_dms[i]) < 0)
+        {
+            begin_dms[i] = input_shape[j] + begin_dms[i];
+        }
+        // Clipping 'begin'
+        if (begin_dms[i].is_static())
+        {
+            if (int64_t(begin_dms[i]) < 0)
+            {
+                begin_dms[i] = 0;
+            }
+            else if (input_shape[j].is_dynamic())
+            {
+                begin_dms[i] = Dimension::dynamic();
+            }
+            else if (int64_t(begin_dms[i]) >= int64_t(input_shape[j]))
+            {
+                begin_dms[i] = input_shape[j] - 1;
+            }
+        }
+
+        // Use upper_bounds if mask is not set
+        if (ub_mask.find(j) == ub_mask.end())
+        {
+            Dimension end_dms_tmp;
+
+            if (ub.size() <= ej)
+            {
+                end_dms_tmp = end_dms[i];
+            }
+            else if (stride_dms[i].is_dynamic())
+            {
+                end_dms_tmp = Dimension::dynamic();
+            }
+            else if (int64_t(stride_dms[i]) > 0)
+            {
+                end_dms_tmp = ub[ej] - 1;
+            }
+            else
+            {
+                end_dms_tmp = ub[ej] + 1;
+            }
+
+            if (ub.size() > ej)
+            {
+                end_dms[i] = end_dms_tmp;
+            }
+            else if (stride_dms[i].is_dynamic())
+            {
+                end_dms[i] = Dimension::dynamic();
+            }
+            else if (int64_t(stride_dms[i]) > 0)
+            {
+                end_dms[i] = -1;
+            }
+            else
+            {
+                end_dms[i] = 0;
+            }
+        }
+        else
+        {
+            if (stride_dms[i].is_dynamic())
+            {
+                end_dms[i] = Dimension::dynamic();
+            }
+            else if (int64_t(stride_dms[i]) > 0)
+            {
+                end_dms[i] = -1;
+            }
+            else
+            {
+                end_dms[i] = 0;
+            }
+        }
+        ej++;
+
+        if (end_dms[i].is_static() && int64_t(end_dms[i]) < 0)
+        {
+            end_dms[i] = input_shape[j] + end_dms[i];
+        }
+        // Clipping 'end'
+        if (end_dms[i].is_static())
+        {
+            if (int64_t(end_dms[i]) < 0)
+            {
+                end_dms[i] = 0;
+            }
+            else if (input_shape[j].is_dynamic())
+            {
+                end_dms[i] = Dimension::dynamic();
+            }
+            else if (int64_t(end_dms[i]) >= int64_t(input_shape[j]))
+            {
+                end_dms[i] = input_shape[j] - 1;
+            }
+        }
+
+        if (new_axis.find(i) == new_axis.end())
+        {
+            j++;
+        }
+        else
+        {
+            end_dms[i] = 0;
+        }
+
+        if (shrink_axis.find(k) != shrink_axis.end())
+        {
+            end_dms[i] = begin_dms[i];
+        }
+        else if (end_dms[i].is_dynamic() || begin_dms[i].is_dynamic() || stride_dms[i].is_dynamic())
+        {
+            out_dims.push_back(Dimension::dynamic());
+        }
+        else
+        {
+            out_dims.push_back(static_cast<int64_t>(
+                ceil(static_cast<float>(abs(int64_t(end_dms[i]) - int64_t(begin_dms[i])) + 1) /
+                     static_cast<float>(abs(int64_t(stride_dms[i]))))));
+        }
+
+        k++;
+    }
+    return out_dims;
+}

src/ngraph/validation_util.hpp

@@ -94,4 +94,15 @@ namespace ngraph
                             const op::PadType pad_type,
                             CoordinateDiff& padding_above,
                             CoordinateDiff& padding_below);
+
+    PartialShape infer_slice_shape(const Node* node,
+                                   const PartialShape& input_shape,
+                                   const std::vector<int64_t>& lb,
+                                   const std::vector<int64_t>& ub,
+                                   const std::vector<int64_t>& str,
+                                   const AxisSet& lb_mask,
+                                   const AxisSet& ub_mask,
+                                   const AxisSet& new_axis,
+                                   const AxisSet& shrink_mask,
+                                   const AxisSet& ellipsis_mask);
 }

test/CMakeLists.txt

@@ -167,6 +167,7 @@ set(MULTI_TEST_SRC
     backend_test.in.cpp
     backend_unary_elementwise.in.cpp
     convolution_test.in.cpp
+    dyn_replace_slice_test.in.cpp
     dyn_slice_test.in.cpp
     dynamic.in.cpp
 )

test/dyn_elimination.cpp

@@ -132,6 +132,56 @@ TEST(dyn_elimination, slice)
     ASSERT_EQ(f->get_results().at(0)->get_shape(), (Shape{2, 4, 2, 2, 1, 2, 2}));
 }
 
+TEST(dyn_elimination, replace_slice)
+{
+    // input has shape [2,4,6,8,2,2,2]
+    // slice in numpy syntax is [0:,:4,2:6:2,7:3:-2,np.newaxis,...,1]
+    // slice shape should be [2,4,2,2,1,2,2] (so sayeth numpy!)
+    Shape shape_in{2, 4, 6, 8, 2, 2, 2};
+    Shape shape_slice{2, 4, 2, 2, 1, 2, 2};
+    auto input = make_shared<op::Parameter>(element::f32, shape_in);
+    auto replacement = make_shared<op::Parameter>(element::f32, shape_slice);
+    auto constant_lb =
+        make_shared<op::Constant>(element::i64, Shape{7}, vector<int64_t>{0, 3, 2, 7, 0, 0, 1});
+    auto constant_ub =
+        make_shared<op::Constant>(element::i64, Shape{7}, vector<int64_t>{0, 4, 6, 3, 0, 0, 0});
+    auto constant_strides =
+        make_shared<op::Constant>(element::i64, Shape{7}, vector<int64_t>{1, 1, 2, -2, 0, 0, 0});
+    AxisSet lower_bounds_mask{1};
+    AxisSet upper_bounds_mask{0};
+    AxisSet new_axis_mask{4};
+    AxisSet shrink_mask{6};
+    AxisSet ellipsis_mask{5};
+
+    auto rsl = make_shared<op::DynReplaceSlice>(input,
+                                                replacement,
+                                                constant_lb,
+                                                constant_ub,
+                                                constant_strides,
+                                                lower_bounds_mask,
+                                                upper_bounds_mask,
+                                                new_axis_mask,
+                                                shrink_mask,
+                                                ellipsis_mask);
+
+    ASSERT_EQ(rsl->get_element_type(), element::f32);
+    ASSERT_EQ(rsl->get_shape(), (Shape{2, 4, 6, 8, 2, 2, 2}));
+
+    auto f = make_shared<Function>(rsl, ParameterVector{input, replacement});
+
+    pass::Manager pass_manager;
+    pass_manager.register_pass<pass::DynElimination>();
+    pass_manager.run_passes(f);
+
+    ASSERT_EQ(count_ops_of_type<op::DynReplaceSlice>(f), 0);
+    ASSERT_EQ(count_ops_of_type<op::ReplaceSlice>(f), 1);
+    ASSERT_EQ(count_ops_of_type<op::Reshape>(f), 1);
+    ASSERT_EQ(count_ops_of_type<op::Reverse>(f), 1);
+
+    ASSERT_EQ(f->get_results().at(0)->get_element_type(), element::f32);
+    ASSERT_EQ(f->get_results().at(0)->get_shape(), (Shape{2, 4, 6, 8, 2, 2, 2}));
+}
+
 TEST(dyn_elimination, reshape)
 {
     auto input_arg = make_shared<op::Parameter>(element::f32, Shape{2, 4, 6, 8});

test/models/onnx/lstm_fwd_large_batch_no_clip.prototxt

+ir_version: 4
+producer_name: "nGraph ONNX Importer"
+graph {
+  node {
+    input: "X"
+    input: "W"
+    input: "R"
+    output: ""
+    output: "Y_h"
+    op_type: "LSTM"
+    attribute {
+      name: "clip"
+      f: 9999.0
+      type: FLOAT
+    }
+    attribute {
+      name: "direction"
+      s: "forward"
+      type: STRING
+    }
+    attribute {
+      name: "hidden_size"
+      i: 3
+      type: INT
+    }
+  }
+  name: "compute_graph"
+  input {
+    name: "X"
+    type {
+      tensor_type {
+        elem_type: 1
+        shape {
+          dim {
+            dim_value: 2
+          }
+          dim {
+            dim_value: 32
+          }
+          dim {
+            dim_value: 1
+          }
+        }
+      }
+    }
+  }
+  input {
+    name: "W"
+    type {
+      tensor_type {
+        elem_type: 1
+        shape {
+          dim {
+            dim_value: 1
+          }
+          dim {
+            dim_value: 12
+          }
+          dim {
+            dim_value: 1
+          }
+        }
+      }
+    }
+  }
+  input {
+    name: "R"
+    type {
+      tensor_type {
+        elem_type: 1
+        shape {
+          dim {
+            dim_value: 1
+          }
+          dim {
+            dim_value: 12
+          }
+          dim {
+            dim_value: 3
+          }
+        }
+      }
+    }
+  }
+  output {
+    name: "Y_h"
+    type {
+      tensor_type {
+        elem_type: 1
+        shape {
+          dim {
+            dim_value: 1
+          }
+          dim {
+            dim_value: 32
+          }
+          dim {
+            dim_value: 3
+          }
+        }
+      }
+    }
+  }
+}
+opset_import {
+  version: 7
+}

test/onnx/onnx_import_rnn.in.cpp

@@ -20,6 +20,7 @@
 #include <fstream>
 #include <iterator>
 #include <limits>
+#include <numeric>
 #include <sstream>
 #include <stdexcept>
 #include <vector>
@@ -203,3 +204,48 @@ NGRAPH_TEST(onnx_${BACKEND_NAME}, model_lstm_fwd_hardsigmoid_activation)
     test_case.set_tolerance(6);
     test_case.run();
 }
+
+NGRAPH_TEST(onnx_${BACKEND_NAME}, model_lstm_fwd_large_batch_no_clip)
+{
+    auto function = onnx_import::import_onnx_model(
+        file_util::path_join(SERIALIZED_ZOO, "onnx/lstm_fwd_large_batch_no_clip.prototxt"));
+
+    auto test_case = ngraph::test::NgraphTestCase(function, "${BACKEND_NAME}");
+
+    std::size_t seq_length = 2;
+    std::size_t batch_size = 32;
+    std::size_t input_size = 1;
+    std::size_t hidden_size = 3;
+
+    std::vector<float> in_X(seq_length * batch_size * input_size);
+    std::iota(std::begin(in_X), std::end(in_X), 1.f);
+    std::vector<float> in_R(4 * hidden_size * hidden_size, 0.1f);
+
+    // X
+    test_case.add_input<float>(in_X);
+    // W
+    test_case.add_input<float>(
+        {0.1f, 0.2f, 0.3f, 0.4f, 1.f, 2.f, 3.f, 4.f, 10.f, 11.f, 12.f, 13.f});
+    // R
+    test_case.add_input<float>(in_R);
+
+    // Y_h_data
+    test_case.add_expected_output<float>(
+        Shape{1, batch_size, hidden_size},
+        {0.90387899f, 0.9135572f,  0.91772245f, 0.90897038f, 0.92132433f, 0.92825467f, 0.91365823f,
+         0.92815113f, 0.93676105f, 0.91799162f, 0.93406357f, 0.94344562f, 0.92199681f, 0.93912057f,
+         0.94859476f, 0.92569357f, 0.94340185f, 0.95250664f, 0.92909964f, 0.94699686f, 0.95545127f,
+         0.93223207f, 0.94999634f, 0.95765468f, 0.93510761f, 0.9524867f,  0.95929726f, 0.93774272f,
+         0.9545467f,  0.96051891f, 0.9401536f,  0.95624603f, 0.96142619f, 0.94235605f, 0.95764499f,
+         0.96209939f, 0.94436539f, 0.95879495f, 0.96259862f, 0.94619635f, 0.95973921f, 0.96296872f,
+         0.94786299f, 0.96051397f, 0.96324302f, 0.94937864f, 0.96114929f, 0.96344629f, 0.95075587f,
+         0.96167006f, 0.96359692f, 0.95200645f, 0.96209679f, 0.96370852f, 0.95314133f, 0.9624464f,
+         0.9637912f,  0.95417069f, 0.96273278f, 0.96385246f, 0.95510395f, 0.96296733f, 0.96389785f,
+         0.95594975f, 0.96315942f, 0.96393147f, 0.95671607f, 0.96331673f, 0.96395638f, 0.9574102f,
+         0.96344554f, 0.96397483f, 0.9580388f,  0.96355102f, 0.9639885f,  0.95860795f, 0.96363739f,
+         0.96399863f, 0.95912322f, 0.96370811f, 0.96400613f, 0.95958963f, 0.96376601f, 0.96401169f,
+         0.96001179f, 0.96381342f, 0.96401581f, 0.96039386f, 0.96385224f, 0.96401886f, 0.96073964f,
+         0.96388402f, 0.96402112f, 0.96105254f, 0.96391004f, 0.96402279f});
+
+    test_case.run();
+}

test/update_dyn_replace_slice_reference.sh

+#!/bin/bash
+# ******************************************************************************
+# Copyright 2017-2019 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ******************************************************************************
+declare THIS_SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
+python ${THIS_SCRIPT_DIR}/ref_generators/generate_dyn_replace_slice_ref.py ${THIS_SCRIPT_DIR}/dyn_replace_slice_test.in.cpp