Merge pull request #153 from neil-lindquist/optimize-struct-sizes

Optimize class sizes
icl-utk-edu · Dec 13, 2023 · 979f9d1 · 979f9d1
2 parents d514136 + f0b465c
commit 979f9d1
Show file tree

Hide file tree

Showing 9 changed files with 80 additions and 103 deletions.
diff --git a/include/slate/BaseBandMatrix.hh b/include/slate/BaseBandMatrix.hh
@@ -252,7 +252,7 @@ void BaseBandMatrix<scalar_t>::allocateBatchArrays(
     int64_t batch_size, int64_t num_arrays)
 {
     if (batch_size == 0) {
-        for (int device = 0; device < this->num_devices_; ++device)
+        for (int device = 0; device < this->num_devices(); ++device)
             batch_size = std::max(batch_size, getMaxDeviceTiles(device));
     }
     this->storage_->allocateBatchArrays(batch_size, num_arrays);
@@ -264,7 +264,7 @@ template <typename scalar_t>
 void BaseBandMatrix<scalar_t>::reserveDeviceWorkspace()
 {
     int64_t num_tiles = 0;
-    for (int device = 0; device < this->num_devices_; ++device)
+    for (int device = 0; device < this->num_devices(); ++device)
         num_tiles = std::max(num_tiles, getMaxDeviceTiles(device));
     this->storage_->reserveDeviceWorkspace(num_tiles);
 }

diff --git a/include/slate/BaseMatrix.hh b/include/slate/BaseMatrix.hh
@@ -160,7 +160,7 @@ public:
     }
 
     /// Returns number of devices (per MPI process) to distribute matrix to.
-    int num_devices() const { return num_devices_; }
+    int num_devices() const { return MatrixStorage<scalar_t>::num_devices(); }
 
     void gridinfo( GridOrder* order, int* nprow, int* npcol,
                    int* myrow, int* mycol );
@@ -774,9 +774,6 @@ protected:
     /// shared storage of tiles and buffers
     std::shared_ptr< MatrixStorage<scalar_t> > storage_;
 
-    // ----- consider where to put these, here or in MatrixStorage
-    static int num_devices_;
-
     MPI_Comm  mpi_comm_;
     MPI_Group mpi_group_;
     int mpi_rank_;
@@ -888,10 +885,6 @@ BaseMatrix<scalar_t>::BaseMatrix(
         MPI_Comm_rank(mpi_comm_, &mpi_rank_));
     slate_mpi_call(
         MPI_Comm_group(mpi_comm_, &mpi_group_));
-
-    // todo: these are static, but we (re-)initialize with each matrix.
-    // todo: similar code in BaseMatrix(...) and MatrixStorage(...)
-    num_devices_ = storage_->num_devices_;
 }
 
 //------------------------------------------------------------------------------
@@ -955,10 +948,6 @@ BaseMatrix<scalar_t>::BaseMatrix(
         MPI_Comm_rank(mpi_comm_, &mpi_rank_));
     slate_mpi_call(
         MPI_Comm_group(mpi_comm_, &mpi_group_));
-
-    // todo: these are static, but we (re-)initialize with each matrix.
-    // todo: similar code in BaseMatrix(...) and MatrixStorage(...)
-    num_devices_ = storage_->num_devices_;
 }
 
 //------------------------------------------------------------------------------
@@ -3894,7 +3883,7 @@ std::tuple<int64_t, int64_t, int>
     assert(0 <= j && j < nt());
     // Given AnyDevice = -3, AllDevices = -2, HostNum = -1,
     // GPU devices 0, 1, ..., num_devices-1.
-    assert( AnyDevice <= device && device < num_devices_ );
+    assert( AnyDevice <= device && device < num_devices() );
     if (op_ == Op::NoTrans)
         return { ioffset_ + i, joffset_ + j, device };
     else
@@ -4001,10 +3990,6 @@ void BaseMatrix<scalar_t>::releaseRemoteWorkspace(
     }
 }
 
-//------------------------------------------------------------------------------
-template <typename scalar_t>
-int BaseMatrix<scalar_t>::num_devices_ = 0;
-
 //------------------------------------------------------------------------------
 // from ScaLAPACK's indxg2l
 // todo: where to put utilities like this?

diff --git a/include/slate/BaseTrapezoidMatrix.hh b/include/slate/BaseTrapezoidMatrix.hh
@@ -615,7 +615,7 @@ void BaseTrapezoidMatrix<scalar_t>::allocateBatchArrays(
     int64_t batch_size, int64_t num_arrays)
 {
     if (batch_size == 0) {
-        for (int device = 0; device < this->num_devices_; ++device)
+        for (int device = 0; device < this->num_devices(); ++device)
             batch_size = std::max(batch_size, getMaxDeviceTiles(device));
     }
     this->storage_->allocateBatchArrays(batch_size, num_arrays);
@@ -635,7 +635,7 @@ template <typename scalar_t>
 void BaseTrapezoidMatrix<scalar_t>::reserveDeviceWorkspace()
 {
     int64_t num_tiles = 0;
-    for (int device = 0; device < this->num_devices_; ++device)
+    for (int device = 0; device < this->num_devices(); ++device)
         num_tiles = std::max(num_tiles, getMaxDeviceTiles(device));
     this->storage_->reserveDeviceWorkspace(num_tiles);
 }

diff --git a/include/slate/Matrix.hh b/include/slate/Matrix.hh
@@ -726,7 +726,7 @@ void Matrix<scalar_t>::allocateBatchArrays(
     int64_t batch_size, int64_t num_arrays)
 {
     if (batch_size == 0) {
-        for (int device = 0; device < this->num_devices_; ++device)
+        for (int device = 0; device < this->num_devices(); ++device)
             batch_size = std::max(batch_size, getMaxDeviceTiles(device));
     }
     this->storage_->allocateBatchArrays(batch_size, num_arrays);
@@ -746,7 +746,7 @@ template <typename scalar_t>
 void Matrix<scalar_t>::reserveDeviceWorkspace()
 {
     int64_t num_tiles = 0;
-    for (int device = 0; device < this->num_devices_; ++device)
+    for (int device = 0; device < this->num_devices(); ++device)
         num_tiles = std::max(num_tiles, getMaxDeviceTiles(device));
     this->storage_->reserveDeviceWorkspace(num_tiles);
 }

diff --git a/include/slate/Tile.hh b/include/slate/Tile.hh
@@ -117,10 +117,10 @@ MatrixType conjTranspose( MatrixType&& A )
 /// and who owns (allocated, deallocates) the data.
 /// @ingroup enum
 ///
-enum class TileKind {
-    Workspace,   ///< SLATE allocated workspace tile
-    SlateOwned,  ///< SLATE allocated origin tile
-    UserOwned,   ///< User owned origin tile
+enum class TileKind : char {
+    Workspace  = 'w',  ///< SLATE allocated workspace tile
+    SlateOwned = 'o',  ///< SLATE allocated origin tile
+    UserOwned  = 'u',  ///< User owned origin tile
 };
 
 //------------------------------------------------------------------------------
@@ -179,6 +179,7 @@ public:
     int64_t stride() const { return stride_; }
 
     /// Sets column stride of this tile
+    [[deprecated( "Use setLayout to manage the Tile's layout.  Will be removed 2024-12." )]]
     void stride(int64_t in_stride) { stride_ = in_stride; }
 
     /// Returns const pointer to data, i.e., A(0,0), where A is this tile
@@ -224,6 +225,12 @@ public:
     /// e.g., via a fromScaLAPACK constructor.
     bool allocated() const { return kind_ != TileKind::UserOwned; }
 
+    /// Returns the TileKind of this tile
+    TileKind kind()
+    {
+        return kind_;
+    }
+
     /// Returns number of bytes; but NOT consecutive if stride != mb_.
     size_t bytes() const { return sizeof(scalar_t) * size(); }
 
@@ -371,11 +378,6 @@ protected:
     void nb(int64_t in_nb);
     void offset(int64_t i, int64_t j);
 
-    void kind(TileKind kind)
-    {
-        kind_ = kind;
-    }
-
     void state(MOSI_State stateIn)
     {
         switch (stateIn) {
@@ -405,13 +407,13 @@ protected:
     int64_t stride_;
     int64_t user_stride_; // Temporarily store user-provided-memory's stride
 
-    Op op_;
-    Uplo uplo_;
-
     scalar_t* data_;
     scalar_t* user_data_; // Temporarily point to user-provided memory buffer.
     scalar_t* ext_data_; // Points to auxiliary buffer.
 
+    Op op_;
+    Uplo uplo_;
+
     TileKind kind_;
     /// layout_: The physical ordering of elements in the data buffer:
     ///          - ColMajor: elements of a column are 1-strided
@@ -434,11 +436,11 @@ Tile<scalar_t>::Tile()
       nb_(0),
       stride_(0),
       user_stride_(0),
-      op_(Op::NoTrans),
-      uplo_(Uplo::General),
       data_(nullptr),
       user_data_(nullptr),
       ext_data_(nullptr),
+      op_(Op::NoTrans),
+      uplo_(Uplo::General),
       kind_(TileKind::UserOwned),
       layout_(Layout::ColMajor),
       user_layout_(Layout::ColMajor),
@@ -487,11 +489,11 @@ Tile<scalar_t>::Tile(
       nb_(nb),
       stride_(lda),
       user_stride_(lda),
-      op_(Op::NoTrans),
-      uplo_(Uplo::General),
       data_(A),
       user_data_(A),
       ext_data_(nullptr),
+      op_(Op::NoTrans),
+      uplo_(Uplo::General),
       kind_(kind),
       layout_(layout),
       user_layout_(layout),
@@ -530,11 +532,11 @@ Tile<scalar_t>::Tile(
       nb_(src_tile.nb_),
       stride_(lda),
       user_stride_(lda),
-      op_(src_tile.op_),
-      uplo_(src_tile.uplo_),
       data_(A),
       user_data_(A),
       ext_data_(nullptr),
+      op_(src_tile.op_),
+      uplo_(src_tile.uplo_),
       kind_(kind),
       layout_(src_tile.layout_),
       user_layout_(src_tile.user_layout_),