fixing no output vals (#9)

don't pass output buffer if no return value; adding some debug helpers

Author: fschlimb

src/Creator.cpp

@@ -18,8 +18,10 @@
 #include <mlir/Dialect/Tensor/IR/Tensor.h>
 #include <mlir/IR/Builders.h>
 
+static const char *FORCE_DIST = getenv("DDPT_FORCE_DIST");
+
 inline uint64_t mkTeam(uint64_t team) {
-  if (team && getTransceiver()->nranks() > 1) {
+  if (team && (FORCE_DIST || getTransceiver()->nranks() > 1)) {
     return 1;
   }
   return 0;

src/Deferred.cpp

@@ -99,6 +99,15 @@ void process_promises() {
     // Create the jit func
     // create dummy type, we'll replace it with the actual type later
     auto dummyFuncType = builder.getFunctionType({}, {});
+    if (false) {
+      ::mlir::OpBuilder::InsertionGuard guard(builder);
+      // Insert before module terminator.
+      builder.setInsertionPoint(module.getBody(),
+                                std::prev(module.getBody()->end()));
+      auto func = builder.create<::mlir::func::FuncOp>(loc, "_debugFunc",
+                                                       dummyFuncType);
+      func.setPrivate();
+    }
     std::string fname("ddpt_jit");
     auto function =
         builder.create<::mlir::func::FuncOp>(loc, fname, dummyFuncType);

src/idtr.cpp

@@ -422,4 +422,5 @@ void _idtr_extractslice(int64_t *slcOffs, int64_t *slcSizes,
               << std::endl;
 }
 
+void _debugFunc() { std::cerr << "_debugfunc\n"; }
 } // extern "C"

src/include/ddptensor/tensor_i.hpp

@@ -43,7 +43,7 @@ class TensorMeta {
   // @ return team, 0 means non-distributed
   int team() const { return _team; }
 
-  /// @return rank (number of dims) of future tensor
+  /// @return if future tensor will have balanced partitions
   int balanced() const { return _balanced; }
 
   void set_guid(id_type guid) { _guid = guid; }

src/jit/mlir.cpp

@@ -375,7 +375,9 @@ std::vector<intptr_t> JIT::run(::mlir::ModuleOp &module,
   std::vector<intptr_t> out(osz);
   auto tmp = out.data();
   // first arg must be the result ptr
-  args.push_back(&tmp);
+  if (osz) {
+    args.push_back(&tmp);
+  }
   // we need a void*& for every input tensor
   // we refer directly to the storage in inp
   for (auto &arg : inp) {

test/stencil-2d.py

@@ -144,7 +144,6 @@ def main():
             W[r + j, r + j] = +1.0 / (4 * j * r)
             W[r - j, r - j] = -1.0 / (4 * j * r)
 
-    # A = np.empty((n, n), dtype=np.float64)
     A = np.numpy.fromfunction(lambda i, j: i + j, (n, n), dtype=np.float64)
 
     for k in range(iterations + 1):