[DO_NOT_MERGE] Ispc dev temp

.gitignore

@@ -12,3 +12,7 @@ CMakeCache.txt
 doc
 
 apps/tensor_times_vector/tensor_times_vector
+
+.cache
+.vscode
+compile_commands.json

CMakeLists.txt

@@ -10,10 +10,12 @@ project(taco
   LANGUAGES C CXX
 )
 option(CUDA "Build for NVIDIA GPU (CUDA must be preinstalled)" OFF)
+option(ISPC "Build for Intel ISPC Compiler (ISPC Compiler must be preinstalled)" OFF)
 option(PYTHON "Build TACO for python environment" OFF)
 option(OPENMP "Build with OpenMP execution support" OFF)
 option(COVERAGE "Build with code coverage analysis" OFF)
 set(TACO_FEATURE_CUDA 0)
+set(TACO_FEATURE_ISPC 0)
 set(TACO_FEATURE_OPENMP 0)
 set(TACO_FEATURE_PYTHON 0)
 if(CUDA)
@@ -22,6 +24,11 @@ if(CUDA)
   add_definitions(-DCUDA_BUILT)
   set(TACO_FEATURE_CUDA 1)
 endif(CUDA)
+if(ISPC)
+  message("-- Searching for ISPC Installation")
+  add_definitions(-DISPC_BUILT)
+  set(TACO_FEATURE_ISPC 1)
+endif(ISPC) 
 if(OPENMP)
   message("-- Will use OpenMP for parallel execution")
   add_definitions(-DUSE_OPENMP)

include/taco/codegen/module.h

@@ -68,6 +68,7 @@ class Module {
 
 private:
   std::stringstream source;
+  std::stringstream additional_source;
   std::stringstream header;
   std::string libname;
   std::string tmpdir;

include/taco/cuda.h

@@ -9,7 +9,19 @@
   #define CUDA_BUILT false
 #endif
 
+#ifndef ISPC_BUILT
+  #define ISPC_BUILT false
+#endif
+
 namespace taco {
+
+/// Functions used by taco to interface with ISPC
+bool should_use_ISPC_codegen();
+void set_ISPC_codegen_enabled(bool enabled);
+bool is_ISPC_code_stream_enabled();
+void set_ISPC_code_stream_enabled(bool enabled);
+
+
 /// Functions used by taco to interface with CUDA (especially unified memory)
 /// Check if should use CUDA codegen
 bool should_use_CUDA_codegen();

include/taco/index_notation/transformations.h

@@ -223,6 +223,8 @@ IndexStmt parallelizeOuterLoop(IndexStmt stmt);
  */
 IndexStmt reorderLoopsTopologically(IndexStmt stmt);
 
+IndexStmt justTraverseThroughTheIndexStmt(IndexStmt stmt);
+
 /**
  * Performs scalar promotion so that reductions are done by accumulating into 
  * scalar temporaries whenever possible.

include/taco/ir/ir.h

@@ -591,7 +591,7 @@ struct Switch : public StmtNode<Switch> {
   static const IRNodeType _type_info = IRNodeType::Switch;
 };
 
-enum class LoopKind {Serial, Static, Dynamic, Runtime, Vectorized, Static_Chunked};
+enum class LoopKind {Serial, Static, Dynamic, Runtime, Vectorized, Static_Chunked, Foreach, Mul_Thread, Init};
 
 /** A for loop from start to end by increment.
  * A vectorized loop will require the increment to be 1 and the

include/taco/ir/ir_printer.h

@@ -16,6 +16,7 @@ class IRPrinter : public IRVisitorStrict {
 public:
   IRPrinter(std::ostream& stream);
   IRPrinter(std::ostream& stream, bool color, bool simplify);
+  IRPrinter(std::ostream& stream, std::ostream& stream2, bool color, bool simplify);
   virtual ~IRPrinter();
 
   void setColor(bool color);
@@ -72,6 +73,7 @@ class IRPrinter : public IRVisitorStrict {
   virtual void visit(const Break*);
 
   std::ostream &stream;
+  std::ostream &stream2;
   int indent;
   bool color;
   bool simplify;
@@ -109,6 +111,7 @@ class IRPrinter : public IRVisitorStrict {
   void doIndent();
   void printBinOp(Expr a, Expr b, std::string op, Precedence precedence);
   bool needsParentheses(Precedence precedence);
+  void sendToStream(std::stringstream &stream);
 
   std::string keywordString(std::string);
   std::string commentString(std::string);

include/taco/ir_tags.h

@@ -9,7 +9,7 @@ namespace taco {
 /// ParallelUnit::GPUWarp can be optionally used to allow for GPU warp-level primitives
 /// ParallelUnit::GPUThread causes for every iteration to be executed on a separate GPU thread
 enum class ParallelUnit {
-  NotParallel, DefaultUnit, GPUBlock, GPUWarp, GPUThread, CPUThread, CPUVector, CPUThreadGroupReduction, GPUBlockReduction, GPUWarpReduction
+  NotParallel, DefaultUnit, GPUBlock, GPUWarp, GPUThread, CPUThread, CPUVector, CPUThreadGroupReduction, GPUBlockReduction, GPUWarpReduction, CPUSimd, CPUSpmd
 };
 extern const char *ParallelUnit_NAMES[];
 

include/taco/lower/lowerer_impl_imperative.h

@@ -499,10 +499,13 @@ class LowererImplImperative : public LowererImpl {
 
   bool emitUnderivedGuards = true;
 
+  int loopDepth = 0;
   int inParallelLoopDepth = 0;
 
   std::map<ParallelUnit, ir::Expr> parallelUnitSizes;
   std::map<ParallelUnit, IndexVar> parallelUnitIndexVars;
+  std::map<int, ParallelUnit> forUnits; // <loopdepth, ParallelUnit>
+  std::map<TensorVar,int> whereTempsWithLoopDepth;
 
   /// Keep track of what IndexVars have already been defined
   std::set<IndexVar> definedIndexVars;

include/taco/util/strings.h

@@ -1,20 +1,42 @@
 #ifndef TACO_UTIL_STRINGS_H
 #define TACO_UTIL_STRINGS_H
 
+#include "taco/cuda.h"
 #include <string>
 #include <sstream>
 #include <vector>
 #include <map>
 #include <iomanip>
 #include <limits>
 
+#include "taco/type.h"
+
 // To get the value of a compiler macro variable
 #define STRINGIFY(x) #x
 #define TO_STRING(x) STRINGIFY(x)
 
 namespace taco {
 namespace util {
 
+// /// Turn anything except floating points that can be written to a stream
+// /// into a string.
+// template <class T>
+// typename std::enable_if<!std::is_floating_point<T>::value, std::string>::type
+// toStringISPC(const T &val) {
+
+//   std::stringstream sstream;
+//   if (val == Int32) {
+//     sstream << "int32";
+//   }
+//   else if (val == Int64) {
+//     sstream << "int64";
+//   }
+//   else {
+//     sstream << val;
+//   }
+//   return sstream.str();
+// }
+
 /// Turn anything except floating points that can be written to a stream
 /// into a string.
 template <class T>

include/taco/version.h.in

@@ -20,5 +20,6 @@
 #define TACO_FEATURE_OPENMP @TACO_FEATURE_OPENMP@
 #define TACO_FEATURE_PYTHON @TACO_FEATURE_PYTHON@
 #define TACO_FEATURE_CUDA   @TACO_FEATURE_CUDA@
+#define TACO_FEATURE_ISPC   @TACO_FEATURE_ISPC@
 
 #endif /* TACO_VERSION_H */