Rename and expose "generic" GPU memory transfer functions

[alexxy/gromacs.git] / src / gromacs / gpu_utils / cudautils.cuh

diff --git a/src/gromacs/gpu_utils/cudautils.cuh b/src/gromacs/gpu_utils/cudautils.cuh
index b05cf65f997b8f773566ef76094a21a444b4710b..f8d73c96979688d7b56bde47871635e21b1c27f3 100644 (file)
--- a/src/gromacs/gpu_utils/cudautils.cuh
+++ b/src/gromacs/gpu_utils/cudautils.cuh
@@ -46,6 +46,8 @@
 #include "gromacs/math/vectypes.h"
 #include "gromacs/utility/fatalerror.h"
 
+enum class GpuApiCallBehavior;
+
 /* TODO error checking needs to be rewritten. We have 2 types of error checks needed
    based on where they occur in the code:
    - non performance-critical: these errors are unsafe to be ignored and must be
@@ -134,6 +136,11 @@ struct gmx_device_info_t
 #endif                                           /* HAVE_NVML */
 };
 
+/*! Launches synchronous or asynchronous device to host memory copy.
+ *
+ *  The copy is launched in stream s or if not specified, in stream 0.
+ */
+int cu_copy_D2H(void *h_dest, void *d_src, size_t bytes, GpuApiCallBehavior transferKind, cudaStream_t s /*= 0*/);
 
 /*! Launches synchronous host to device memory copy in stream 0. */
 int cu_copy_D2H_sync(void * /*h_dest*/, void * /*d_src*/, size_t /*bytes*/);
@@ -141,6 +148,12 @@ int cu_copy_D2H_sync(void * /*h_dest*/, void * /*d_src*/, size_t /*bytes*/);
 /*! Launches asynchronous host to device memory copy in stream s. */
 int cu_copy_D2H_async(void * /*h_dest*/, void * /*d_src*/, size_t /*bytes*/, cudaStream_t /*s = 0*/);
 
+/*! Launches synchronous or asynchronous host to device memory copy.
+ *
+ *  The copy is launched in stream s or if not specified, in stream 0.
+ */
+int cu_copy_H2D(void *d_dest, void *h_src, size_t bytes, GpuApiCallBehavior transferKind, cudaStream_t /*s = 0*/);
+
 /*! Launches synchronous host to device memory copy. */
 int cu_copy_H2D_sync(void * /*d_dest*/, void * /*h_src*/, size_t /*bytes*/);