Merge pull request #231 from intel/develop

Develop

README.md

@@ -2,6 +2,7 @@
 
 * New YASK users may want to start with the [YASK tutorial](http://intel.github.io/yask/YASK-tutorial.pdf).
 * Users with existing YASK-based code may want to jump to the [backward-compatibility notices](#backward-compatibility-notices).
+* All YASK users will also be interested in the [API documentation](http://intel.github.io/yask/api/html/index.html).
 
 ## Overview
 YASK is a framework to rapidly create high-performance stencil code including optimizations and features such as
@@ -11,7 +12,7 @@ YASK is a framework to rapidly create high-performance stencil code including op
 * Scaling to multiple sockets and nodes via MPI with overlapped communication and compute.
 * Spatial tiling with automatically-tuned block sizes.
 * Temporal tiling in multiple dimensions to further increase cache locality.
-* APIs for C++ and Python: [API documentation](http://intel.github.io/yask/api/html/index.html)
+* APIs for C++ and Python.
 
 YASK contains a domain-specific compiler to convert stencil-equation specifications to SIMD-optimized code for Intel(R) Xeon Phi(TM) and Intel(R) Xeon(R) processors.
 

src/common/common_utils.cpp

@@ -43,7 +43,7 @@ namespace yask {
     // for numbers above 9 (at least up to 99).
 
     // Format: "major.minor.patch".
-    const string version = "3.01.00";
+    const string version = "3.01.01";
 
     string yask_get_version_string() {
         return version;

src/common/common_utils.hpp

@@ -179,51 +179,92 @@ namespace yask {
 
     // Set that retains order of things added.
     // Or, vector that allows insertion if element doesn't exist.
-    // TODO: hide vector inside class and provide proper accessor methods.
     template <typename T>
-    class vector_set : public std::vector<T> {
-        std::map<T, size_t> _posn;
-
-    private:
-        virtual void push_front(const T& val) {
-            THROW_YASK_EXCEPTION("push_front() not allowed");
-        }
+    class vector_set final {
+        std::vector<T> _items;     // no duplicates.
+        std::map<T, size_t> _posn; // _posn[_items[i]] = i;
 
     public:
         vector_set() {}
-        virtual ~vector_set() {}
+        ~vector_set() {}
 
-        // Copy ctor.
-        vector_set(const vector_set& src) :
-            std::vector<T>(src), _posn(src._posn) {}
+        // Default assign and copy ctor are okay.
 
-        virtual size_t count(const T& val) const {
+        // STL methods.
+        // Do not provide any non-const iterators or element access to prevent
+        // breaking _items <-> _posn relationship.
+        typename std::vector<T>::const_iterator begin() const {
+            return _items.begin();
+        }
+        typename std::vector<T>::const_iterator end() const {
+            return _items.end();
+        }
+        const T& at(size_t i) const {
+            return _items.at(i);
+        }
+        const T& operator[](size_t i) const {
+            return _items[i];
+        }
+        const T& front() const {
+            return _items.front();
+        }
+        const T& back() const {
+            return _items.back();
+        }
+        size_t size() const {
+            assert(_items.size() == _posn.size());
+            return _items.size();
+        }
+        bool empty() const {
+            return size() == 0 ;
+        }
+        size_t count(const T& val) const {
+            assert(_items.size() == _posn.size());
             return _posn.count(val);
         }
-        virtual void insert(const T& val) {
+        void insert(const T& val) {
+            assert(_items.size() == _posn.size());
             if (_posn.count(val) == 0) {
-                std::vector<T>::push_back(val);
-                _posn[val] = std::vector<T>::size() - 1;
+                _items.push_back(val);
+                _posn[val] = _items.size() - 1;
             }
+            assert(_items.size() == _posn.size());
         }
-        virtual void push_back(const T& val) {
-            insert(val);
+        void push_back(const T& val) {
+            insert(val);        // Does nothing if already exists.
         }
-        virtual void erase(const T& val) {
+        void erase(const T& val) {
             if (_posn.count(val) > 0) {
                 size_t op = _posn.at(val);
-                std::vector<T>::erase(std::vector<T>::begin() + op);
+                _items.erase(_items.begin() + op);
+
+                // Repair positions of items after 'val'.
                 for (auto pi : _posn) {
                     auto& p = pi.second;
                     if (p > op)
                         p--;
                 }
+                _posn.erase(val);
             }
+            assert(_items.size() == _posn.size());
         }
-        virtual void clear() {
-            std::vector<T>::clear();
+        void clear() {
+            _items.clear();
             _posn.clear();
         }
+
+        // New methods.
+        void swap(size_t i, size_t j) {
+            assert(i < _items.size());
+            assert(j < _items.size());
+            if (i == j)
+                return;
+            T tmp = _items[i];
+            _items[i] = _items[j];
+            _items[j] = tmp;
+            _posn[_items[i]] = i;
+            _posn[_items[j]] = j;
+        }
     };
 
 } // namespace.

src/common/tuple.hpp

@@ -125,6 +125,9 @@ namespace yask {
         void setVal(const T& val) { _val = val; }
 
         // Comparison ops.
+        // Compare name pointers and actual names in case there
+        // is more than one pool, which can happen when loading
+        // more than one dynamic lib.
         bool operator==(const Scalar& rhs) const {
             return _val == rhs._val &&
                 (_namep == rhs._namep || *_namep == *rhs._namep);
@@ -133,7 +136,7 @@ namespace yask {
             return (_val < rhs._val) ? true :
                 (_val > rhs._val) ? false :
                 (_namep == rhs._namep) ? false :
-                (*_namep < *rhs._namep) ? true : false;
+                (*_namep < *rhs._namep);
         }
     };
 
@@ -183,6 +186,12 @@ namespace yask {
         const std::vector<Scalar<T>>& getDims() const {
             return _q;
         }
+        typename std::vector<Scalar<T>>::const_iterator begin() const {
+            return _q.begin();
+        }
+        typename std::vector<Scalar<T>>::const_iterator end() const {
+            return _q.end();
+        }
 
         // Clear data.
         void clear() {
@@ -251,10 +260,17 @@ namespace yask {
         // Return dim posn or -1 if it doesn't exist.
         // Lookup by name.
         int lookup_posn(const std::string& dim) const {
+
+            // First check pointers.
+            for (size_t i = 0; i < _q.size(); i++) {
+                auto& s = _q[i];
+                if (s.getNamePtr() == &dim)
+                    return int(i);
+            }
+
+            // Then check full strings.
             for (size_t i = 0; i < _q.size(); i++) {
                 auto& s = _q[i];
-
-                // Check for match of name.
                 if (s.getName() == dim)
                     return int(i);
             }

src/compiler/lib/Cpp.cpp

@@ -125,7 +125,7 @@ namespace yask {
                 // x=>x_elem, y=>(y_elem+2), z=>(z_elem+1) in var-point
                 // index args.
                 VarMap vMap;
-                for (auto& dim : vecPoint.getDims()) {
+                for (auto& dim : vecPoint) {
                     auto& dname = dim.getName();
                     int dofs = dim.getVal();
 
@@ -692,7 +692,7 @@ namespace yask {
         auto& fold = getFold();
         os << "\n // Element indices derived from vector indices.\n";
         int i = 0;
-        for (auto& dim : fold.getDims()) {
+        for (auto& dim : fold) {
             auto& dname = dim.getName();
             string ename = dname + _elemSuffix;
             string cap_dname = PrinterBase::allCaps(dname);

src/compiler/lib/Eqs.cpp

@@ -209,7 +209,7 @@ namespace yask {
                                      " cannot have a step condition");
 
             // LHS must have all domain dims.
-            for (auto& dd : dims._domainDims.getDims()) {
+            for (auto& dd : dims._domainDims) {
                 auto& dname = dd.getName();
                 numExprPtr dexpr = op1->getArg(dname);
                 if (!dexpr)
@@ -580,7 +580,7 @@ namespace yask {
             // dim is a simple offset.  For example, in var dim 'x', the
             // index in the corresponding posn must be 'x', 'x+n', or 'x-n'.
             int fdoffsets = 0;
-            for (auto fdim : _dims._foldGT1.getDims()) {
+            for (auto fdim : _dims._foldGT1) {
                 auto& fdname = fdim.getName();
                 if (gp->getArgOffsets().lookup(fdname))
                     fdoffsets++;

src/compiler/lib/Eqs.hpp

@@ -273,15 +273,16 @@ namespace yask {
             // Want to keep original order as much as possible.
             // Only reorder if dependencies are in conflict.
 
-            // Scan from beginning to end.
-            for (size_t i = 0; i < _all.size(); i++) {
-                auto& oi = _all.at(i);
+            // Scan from beginning to next-to-last.
+            for (size_t i = 0; i < _all.size() - 1; i++) {
 
                 // Repeat until no dependent found.
                 bool done = false;
                 while (!done) {
 
                     // Does obj[i] depend on any obj after it?
+                    auto& oi = _all.at(i);
+                    done = true;
                     for (size_t j = i+1; j < _all.size(); j++) {
                         auto& oj = _all.at(j);
 
@@ -296,16 +297,13 @@ namespace yask {
                             }
 
                             // Swap them.
-                            auto temp = oi;
-                            oi = oj;
-                            oj = temp;
+                            _all.swap(i, j);
 
                             // Start over at index i.
                             done = false;
                             break;
                         }
                     }
-                    done = true;
                 }
             }
         }

src/compiler/lib/Expr.hpp

@@ -865,7 +865,7 @@ namespace yask {
 
         // Set given args to be given offsets.
         virtual void setArgOffsets(const IntTuple& offsets) {
-            for (auto ofs : offsets.getDims())
+            for (auto ofs : offsets)
                 setArgOffset(ofs);
         }
 

src/compiler/lib/Settings.cpp

@@ -127,7 +127,7 @@ namespace yask {
 
         // Extract domain fold lengths based on cmd-line options.
         IntTuple foldOpts;
-        for (auto& dim : _domainDims.getDims()) {
+        for (auto& dim : _domainDims) {
             auto& dname = dim.getName();
 
             // Was folding specified for this dim?
@@ -193,7 +193,7 @@ namespace yask {
                 // sizes.
                 else {
                     IntTuple innerOpts;
-                    for (auto& dim : _domainDims.getDims()) {
+                    for (auto& dim : _domainDims) {
                         auto& dname = dim.getName();
                         if (dname == _innerDim)
                             continue;
@@ -208,7 +208,7 @@ namespace yask {
                 }
 
                 // Put them into the fold.
-                for (auto& dim : _domainDims.getDims()) {
+                for (auto& dim : _domainDims) {
                     auto& dname = dim.getName();
                     if (dname == _innerDim)
                         _fold[dname] = inner_sz;
@@ -227,7 +227,7 @@ namespace yask {
         }
 
         // Set foldGT1.
-        for (auto i : _fold.getDims()) {
+        for (auto i : _fold) {
             auto& dname = i.getName();
             auto& val = i.getVal();
             if (val > 1)
@@ -254,7 +254,7 @@ namespace yask {
         }
 
         // Create final cluster lengths based on cmd-line options.
-        for (auto& dim : settings._clusterOptions.getDims()) {
+        for (auto& dim : settings._clusterOptions) {
             auto& dname = dim.getName();
             int mult = dim.getVal();
 

src/compiler/lib/Var.cpp

@@ -141,7 +141,7 @@ namespace yask {
         }
 
         // Find the number of folded dims used in this var.
-        for (auto fdim : dims._foldGT1.getDims()) {
+        for (auto fdim : dims._foldGT1) {
             auto& fdname = fdim.getName();
 
             // Search for dim in var.
@@ -203,7 +203,7 @@ namespace yask {
                 for (auto& i1 : m1) {
                     auto& step = i1.first;
                     const IntTuple& ohalos = i1.second;
-                    for (auto& dim : ohalos.getDims()) {
+                    for (auto& dim : ohalos) {
                         auto& dname = dim.getName();
                         auto& val = dim.getVal();
 
@@ -245,7 +245,7 @@ namespace yask {
         int l1Dist = 0;
 
         // Update halo vals.
-        for (auto& dim : offsets.getDims()) {
+        for (auto& dim : offsets) {
             auto& dname = dim.getName();
             int val = dim.getVal();
             bool left = val <= 0;
@@ -283,7 +283,7 @@ namespace yask {
     // Update const indices based on 'indices'.
     void Var::updateConstIndices(const IntTuple& indices) {
 
-        for (auto& dim : indices.getDims()) {
+        for (auto& dim : indices) {
             auto& dname = dim.getName();
             int val = dim.getVal();