Test benchs

.gitignore

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

CMakeLists.txt

@@ -10,18 +10,25 @@ 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(OPENMP "Build with OpenMP execution support" ON)
 option(COVERAGE "Build with code coverage analysis" OFF)
 set(TACO_FEATURE_CUDA 0)
-set(TACO_FEATURE_OPENMP 0)
+set(TACO_FEATURE_ISPC 0)
+set(TACO_FEATURE_OPENMP 1)
 set(TACO_FEATURE_PYTHON 0)
 if(CUDA)
   message("-- Searching for CUDA Installation")
   find_package(CUDA REQUIRED)
   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)
@@ -88,6 +95,39 @@ if(OPENMP)
   set(C_CXX_FLAGS "-fopenmp ${C_CXX_FLAGS}")
 endif(OPENMP)
 
+set(PAPI_DIR "/home/min/a/kadhitha/workspace/my_taco/papi/src/install/")
+
+find_path(PAPI_DIR
+    NAMES include/papi.h
+)
+
+find_library(PAPI_LIBRARIES
+    # Pick the static library first for easier run-time linking.
+    NAMES libpapi.a papi
+    HINTS ${PAPI_DIR}/lib ${HILTIDEPS}/lib
+)
+
+find_path(PAPI_INCLUDE_DIRS
+    NAMES papi.h
+    HINTS ${PAPI_DIR}/include ${HILTIDEPS}/include
+)
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(PAPI DEFAULT_MSG
+    PAPI_LIBRARIES
+    PAPI_INCLUDE_DIRS
+)
+
+mark_as_advanced(
+    PAPI_PREFIX_DIRS
+    PAPI_LIBRARIES
+    PAPI_INCLUDE_DIRS
+)
+
+include_directories(${PAPI_INCLUDE_DIRS})
+
+# project (ValgrindExample)
+
 if(COVERAGE)
   find_program(PATH_TO_GCOVR gcovr REQUIRED)
   # add coverage tooling to build flags
@@ -97,7 +137,8 @@ if(COVERAGE)
   message("-- Code coverage analysis (gcovr) enabled")
 endif(COVERAGE)
 
-set(C_CXX_FLAGS "${C_CXX_FLAGS}")
+set(C_CXX_FLAGS "${C_CXX_FLAGS} -I/${PAPI_DIR}/include -L/${PAPI_DIR}/lib")
+# set(C_CXX_FLAGS "${C_CXX_FLAGS}")
 set(CMAKE_C_FLAGS "${C_CXX_FLAGS}")
 set(CMAKE_CXX_FLAGS "${C_CXX_FLAGS} -std=c++14")
 
@@ -110,6 +151,9 @@ set(TACO_INCLUDE_DIR ${TACO_PROJECT_DIR}/include)
 
 enable_testing()
 include_directories(${TACO_INCLUDE_DIR})
+# include_directories("/home/min/a/kadhitha/workspace/my_taco/valgrind")
+# project (ValgrindExample)
+# include (CTest)
 
 set(TACO_LIBRARY_DIR ${CMAKE_LIBRARY_OUTPUT_DIRECTORY})
 

include/taco/codegen/module.h

@@ -17,7 +17,7 @@ class Module {
 public:
   /// Create a module for some target
   Module(Target target=getTargetFromEnvironment())
-    : lib_handle(nullptr), moduleFromUserSource(false), target(target) {
+    : lib_handle(nullptr), so_lib_handle(nullptr), moduleFromUserSource(false), target(target) {
     setJITLibname();
     setJITTmpdir();
   }
@@ -44,11 +44,16 @@ class Module {
   /// before calling. If there's no function of this name then a nullptr is
   /// returned.
   void* getFuncPtr(std::string name);
+  void* getFuncPtr(std::string& sofile, std::string name);
 
   /// Call a raw function in this module and return the result
+  int callFuncPackedRaw(std::string name, std::string& sofile, void** args);
   int callFuncPackedRaw(std::string name, void** args);
 
   /// Call a raw function in this module and return the result
+  int callFuncPackedRaw(std::string name, std::string& sofile, std::vector<void*> args) {
+    return callFuncPackedRaw(name, sofile, args.data());
+  }
   int callFuncPackedRaw(std::string name, std::vector<void*> args) {
     return callFuncPackedRaw(name, args.data());
   }
@@ -57,6 +62,10 @@ class Module {
   int callFuncPacked(std::string name, void** args) {
     return callFuncPackedRaw("_shim_"+name, args);
   }
+
+  int callFuncPacked(std::string name, std::string& sofile, void** args) {
+    return callFuncPackedRaw("_shim_"+name, sofile,args);
+  }
 
   /// Call a function using the taco_tensor_t interface and return the result
   int callFuncPacked(std::string name, std::vector<void*> args) {
@@ -68,10 +77,12 @@ class Module {
 
 private:
   std::stringstream source;
+  std::stringstream additional_source;
   std::stringstream header;
   std::string libname;
   std::string tmpdir;
   void* lib_handle;
+  void* so_lib_handle;
   std::vector<Stmt> funcs;
 
   // true iff the module was created from user-provided source

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 +223,9 @@ IndexStmt parallelizeOuterLoop(IndexStmt stmt);
  */
 IndexStmt reorderLoopsTopologically(IndexStmt stmt);
 
+IndexStmt loopFusionOverFission(IndexStmt stmt, Assignment assignment,
+  std::string side, int iters);
+
 /**
  * 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/taco_tensor_t.h

@@ -6,6 +6,7 @@
 #ifndef TACO_TENSOR_T_DEFINED
 #define TACO_TENSOR_T_DEFINED
 
+#include <cstdint>
 #include <stdint.h>
 
 typedef enum { taco_mode_dense, taco_mode_sparse } taco_mode_t;

include/taco/tensor.h

@@ -413,6 +413,8 @@ class TensorBase {
 
   /// Compile the tensor expression.
   void compile();
+  void compute(std::ofstream& statfile);
+  void compute(std::ofstream& statfile, std::string& sofile);
 
   void compile(IndexStmt stmt, bool assembleWhileCompute=false);
 

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 */