mpi: fix / adapt setbuf_unit_test to recent changes in the lib

src/opt/mpi/include/rarray.cc.h

@@ -237,27 +237,17 @@ namespace CNORXZ
 	{
 	    return mMap;
 	}
-	
+
 	template <typename T>
-	template <class LoopI, class SrcI, class RSrcI, class I>
-	void RCArray<T>::load(const Sptr<LoopI>& lpi, const Sptr<RIndex<SrcI,RSrcI>>& ai,
-			      const Sptr<I>& i, const Sptr<Vector<SizeT>>& imap) const
+	template <class SrcI, class RSrcI, class I>
+	void RCArray<T>::load(const Sptr<RIndex<SrcI,RSrcI>>& ai,
+			      const Sptr<I>& i, const Vector<bool>& required) const
 	{
 	    mA->checkFormatCompatibility(mindex(ai->local()*i));
 	    const SizeT blocks = i->pmax().val();
-	    setupBuffer(lpi, ai, imap, *mA, mBuf, mMap, blocks);
+	    setupBuffer(ai, required, *mA, mBuf, mMap, blocks);
 	}
-
-	template <typename T>
-	template <class Index, class F>
-	Sptr<Vector<SizeT>> RCArray<T>::load(const Sptr<Index>& i, const F& f) const
-	{
-	    Sptr<Vector<SizeT>> imap = std::make_shared<Vector<SizeT>>();
-	    CXZ_ERROR("!!!");
-	    //load(i, /**/, imap);
-	    return imap;
-	}
-
+	
 	/*==============+
 	 |    RArray    |
 	 +==============*/
@@ -427,14 +417,14 @@ namespace CNORXZ
 	 |    non-member functions    |
 	 +============================*/
 
-	template <class LoopI, class SrcI, class RSrcI, typename T>
-	void setupBuffer(const Sptr<LoopI>& lpi, const Sptr<RIndex<SrcI,RSrcI>>& rgj,
-			 const Sptr<Vector<SizeT>>& imap, const CArrayBase<T>& data,
+	template <class SrcI, class RSrcI, typename T>
+	void setupBuffer(const Sptr<RIndex<SrcI,RSrcI>>& rgj,
+			 const Vector<bool>& required, const CArrayBase<T>& data,
 			 Vector<T>& buf, Vector<const T*>& map, const SizeT blocks)
 	{
 	    const SizeT myrank = getRankNumber();
 	    const SizeT Nranks = getNumRanks();
-
+	    const SizeT locsz = rgj->local()->lmax().val();
 	    const SizeT mapsize = rgj->range()->size();
 	    map = Vector<const T*>(mapsize,nullptr);
 	    Vector<Vector<T>> sendbuf(Nranks);
@@ -442,22 +432,7 @@ namespace CNORXZ
 		sb.reserve(data.size());
 	    }
 	    Vector<Vector<SizeT>> request(Nranks);
-	    const SizeT locsz = rgj->local()->lmax().val();
 
-	    // ==== new ====
-
-	    // First loop: Find out what's needed
-	    Vector<bool> required(rgj->lmax().val(),false);
-	    lpi->ifor( mapXpr(rgj, lpi, imap,
-			      operation
-			      ( [&](SizeT j) {
-				  const SizeT r = j / locsz;
-				  if(myrank != r){
-				      required[j] = true;
-				  }
-			      } , posop(rgj) ) ),
-		       NoF {} )();
-	    
 	    // Second loop: setup send buffer
 	    auto mi = mindexPtr(rgj->rankI(), rgj->local());
 	    mi->ifor( operation
@@ -540,8 +515,8 @@ namespace CNORXZ
 			      map[mpidx] = data.data() + l*blocks;
 			  }
 		      } , posop(mi) ), NoF {} )();
-	}
 
+	}
 	
     } // namespace mpi
 } // namespace CNORXZ

src/opt/mpi/include/rarray.h

@@ -122,25 +122,17 @@ namespace CNORXZ
 		The buffer needs to be initialized before according to required range.
 	     */
 	    const Vector<const T*>& buffermap() const;
-	    
+
 	    /** Load all data from other ranks that is accessed by ai in a loop over lpi.
 		imap indicates the global (!) position of ai for a given (local) position of lpi.
-		@param lpi Loop rank index.
 		@param ai Access rank index.
 		@param i Index for non-ranked dimensions.
-		@param imap Index position map.
+		@param required Vector of booleans indicating global positions to load.
 	     */
-	    template <class LoopI, class SrcI, class RSrcI, class I>
-	    void load(const Sptr<LoopI>& lpi, const Sptr<RIndex<SrcI,RSrcI>>& ai,
-		      const Sptr<I>& i, const Sptr<Vector<SizeT>>& imap) const;
+	    template <class SrcI, class RSrcI, class I>
+	    void load(const Sptr<RIndex<SrcI,RSrcI>>& ai,
+		      const Sptr<I>& i, const Vector<bool>& required) const;
 
-	    /** Load all data from other ranks that is accessed by f(i).
-		@param lpi Loop index.
-		@param f Map function.
-		@return Index position map.
-	     */
-	    template <class Index, class F>
-	    Sptr<Vector<SizeT>> load(const Sptr<Index>& lpi, const F& f) const;
 
 	protected:
 	    ObjHandle<CArrayBase<T>> mA;
@@ -255,9 +247,9 @@ namespace CNORXZ
 
 	};
 
-	template <class LoopI, class SrcI, class RSrcI, typename T>
-	void setupBuffer(const Sptr<LoopI>& lpi, const Sptr<RIndex<SrcI,RSrcI>>& rgj,
-			 const Sptr<Vector<SizeT>>& imap, const CArrayBase<T>& data,
+	template <class SrcI, class RSrcI, typename T>
+	void setupBuffer(const Sptr<RIndex<SrcI,RSrcI>>& rgj,
+			 const Vector<bool>& required, const CArrayBase<T>& data,
 			 Vector<T>& buf, Vector<const T*>& map, const SizeT blocks);
 
 	

src/opt/mpi/tests/roperation_unit_test.cc

@@ -129,7 +129,11 @@ namespace
 	imap1 = setupMap(xp, x, [&](const auto& vec) {
 	    return std::make_tuple((std::get<0>(vec)+1)%T, (std::get<1>(vec)+1)%L,
 				   (std::get<2>(vec)+1)%L, (std::get<3>(vec)+1)%L); } );
-	mM1.load(x, xp, AB, imap1);
+	Vector<bool> req(xp->range()->size(), false);
+	for(const auto& r: *imap1){
+	    req[r] = true;
+	}
+	mM1.load(xp, AB, req);
 	res.rop(x*A*B) = mapXpr(xp,x,imap1, mM1(xp*A*B) - mM1(x*A*B) );
 
 	for(SizeT x0 = 0; x0 != T; ++x0) { 

src/opt/mpi/tests/setbuf_unit_test.cc

@@ -110,8 +110,13 @@ namespace
 	    return o;
 	};
 
-	const Sptr<Vector<SizeT>> fmap = setupMap(rgj, rgi, shift);
-	setupBuffer(rgj, rgi, fmap, data, buf, map, mSRange->size());
+	Vector<bool> req(mRRange->size(), false);
+	for(auto ii = *rgi; ii.lex() != ii.lmax().val(); ++ii){
+	    auto jj = ii;
+	    jj.at( shift( ii.meta() ) );
+	    req[jj.pos()] = true;
+	}
+	setupBuffer(rgi, req, data, buf, map, mSRange->size());
 
 	EXPECT_EQ(mRRange->sub(1)->size(), 16*12*12*12/4);
 	const SizeT locsz = rgj->local()->lmax().val();
@@ -128,10 +133,7 @@ namespace
 
 	    if(rgi->rank() == myrank){
 		const SizeT mpidx = (rgj->pos() - myrankoff + mapsize) % mapsize;
-		VCHECK(mpidx);
-		assert(mpidx < map.size());
 		EXPECT_TRUE(map.data()[mpidx] != nullptr);
-		if(map.data()[mpidx] == nullptr) continue;
 		
 		const Double vn = *map[mpidx]/blocks;
 		const SizeT xp = static_cast<SizeT>(vn);