Fix intertype gpu copy (elemental#92)

* Fix copying between types on GPU * Update version to 1.3.2
benson31 · Dec 14, 2019 · 752d139 · 752d139
1 parent aced0a7
commit 752d139
Show file tree

Hide file tree

Showing 4 changed files with 55 additions and 5 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -54,7 +54,7 @@ endif (__GIT_EXECUTABLE)
 # This must be set because version tags
 set(HYDROGEN_VERSION_MAJOR 1)
 set(HYDROGEN_VERSION_MINOR 3)
-set(HYDROGEN_VERSION_PATCH 1)
+set(HYDROGEN_VERSION_PATCH 2)
 set(HYDROGEN_VERSION_MAJOR_MINOR
     "${HYDROGEN_VERSION_MAJOR}.${HYDROGEN_VERSION_MINOR}")
 set(HYDROGEN_VERSION

diff --git a/include/El/blas_like/level1/Copy.hpp b/include/El/blas_like/level1/Copy.hpp
@@ -17,6 +17,10 @@
 #include <El/blas_like/level1/Copy/GeneralPurpose.hpp>
 #include <El/blas_like/level1/Copy/util.hpp>
 
+#ifdef HYDROGEN_HAVE_GPU
+#include <hydrogen/blas/GPU_BLAS.hpp>
+#endif
+
 namespace El {
 namespace details {
 
@@ -248,6 +252,29 @@ void Copy( const Matrix<T>& A, Matrix<T>& B )
 }
 
 #ifdef HYDROGEN_HAVE_CUDA
+template <typename T, typename U>
+void Copy(Matrix<T, Device::GPU> const& A, Matrix<U, Device::GPU>& B)
+{
+    EL_DEBUG_CSE;
+    Int const height = A.Height();
+    Int const width = A.Width();
+    B.Resize(height, width);
+    Int const ldA = A.LDim();
+    Int const ldB = B.LDim();
+    T const* ABuf = A.LockedBuffer();
+    U* BBuf = B.Buffer();
+
+    SyncInfo<Device::GPU> syncInfoA = SyncInfoFromMatrix(A),
+        syncInfoB = SyncInfoFromMatrix(B);
+    auto syncHelper = MakeMultiSync(syncInfoB, syncInfoA);
+
+    gpu_blas::Copy(TransposeMode::NORMAL,
+                   height, width,
+                   ABuf, ldA,
+                   BBuf, ldB,
+                   syncInfoB);
+}
+
 template<typename T>
 void Copy(const Matrix<T,Device::GPU>& A, Matrix<T,Device::GPU>& B)
 {

diff --git a/include/hydrogen/blas/GPU_BLAS_decl.hpp b/include/hydrogen/blas/GPU_BLAS_decl.hpp
@@ -216,11 +216,11 @@ void Copy(SizeT size,
  *
  *  @ingroup device_blas
  */
-template <typename T, typename SizeT>
+template <typename T, typename U, typename SizeT>
 void Copy(TransposeMode transpA,
           SizeT num_rows, SizeT num_cols,
           T const* A, SizeT lda,
-          T* B, SizeT ldb,
+          U* B, SizeT ldb,
           SyncInfo<Device::GPU> const& syncinfo);
 
 /** @brief 2-D Copy operation in GPU memory with 2 strides.

diff --git a/include/hydrogen/blas/GPU_BLAS_impl.hpp b/include/hydrogen/blas/GPU_BLAS_impl.hpp
@@ -394,6 +394,29 @@ void Copy2DImpl(SizeT nrows, SizeT ncols,
     }
 }
 
+template <typename T, typename U, typename SizeT>
+void Copy2DImpl(SizeT nrows, SizeT ncols,
+                TransposeMode transA,
+                T const* A, SizeT lda,
+                U* B, SizeT ldb,
+                SyncInfo<Device::GPU> const& si)
+{
+    switch (transA)
+    {
+    case TransposeMode::NORMAL:
+        Copy_GPU_impl(nrows, ncols,
+                      A, SizeT(1), lda,
+                      B, SizeT(1), ldb, si.stream_);
+        break;
+    case TransposeMode::TRANSPOSE:
+        throw std::logic_error(
+            "Copy2DImpl: Need to implement multitype transpose");
+        break;
+    default:
+        throw std::logic_error("Copy2DImpl: TransposeMode not supported!");
+    }
+}
+
 template <typename T, typename SizeT,
           typename=EnableUnless<IsSupportedType<T,BLAS_Op::COPY>>,
           typename=void>
@@ -555,11 +578,11 @@ void Copy(SizeT size,
     details::CopyImpl(size, X, incx, Y, incy, si);
 }
 
-template <typename T, typename SizeT>
+template <typename T, typename U, typename SizeT>
 void Copy(TransposeMode transA,
           SizeT num_rows, SizeT num_cols,
           T const* A, SizeT lda,
-          T* B, SizeT ldb,
+          U* B, SizeT ldb,
           SyncInfo<Device::GPU> const& si)
 {
     details::Copy2DImpl(num_rows, num_cols, transA,