libgomp: cuda.h and omp_target_memcpy_rect cleanup

Fixes for commit r14-2792-g25072a477a56a727b369bf9b20f4d18198ff5894
"OpenMP: Call cuMemcpy2D/cuMemcpy3D for nvptx for omp_target_memcpy_rect",
namely:

In that commit, the code was changed to handle shared-memory devices;
however, as pointed out, omp_target_memcpy_check already set the pointer
to NULL in that case.  Hence, this commit reverts to the prior version.

In cuda.h, it adds cuMemcpyPeer{,Async} for symmetry for cuMemcpy3DPeer
(all currently unused) and in three structs, fixes reserved-member names
and remove a bogus 'const' in three structs.

And it changes a DLSYM to DLSYM_OPT as not all plugins support the new
functions, yet.

include/ChangeLog:

	* cuda/cuda.h (CUDA_MEMCPY2D, CUDA_MEMCPY3D, CUDA_MEMCPY3D_PEER):
	Remove bogus 'const' from 'const void *dst' and fix reserved-name
	name in those structs.
	(cuMemcpyPeer, cuMemcpyPeerAsync): Add.

libgomp/ChangeLog:

	* target.c (omp_target_memcpy_rect_worker): Undo dim=1 change for
	GOMP_OFFLOAD_CAP_SHARED_MEM.
	(omp_target_memcpy_rect_copy): Likewise for lock condition.
	(gomp_load_plugin_for_device): Use DLSYM_OPT not DLSYM for
	memcpy3d/memcpy2d.
	* plugin/plugin-nvptx.c (GOMP_OFFLOAD_memcpy2d,
	GOMP_OFFLOAD_memcpy3d): Use memset 0 to nullify reserved and
	unused src/dst fields for that mem type; remove '{src,dst}LOD = 0'.

3 files changed, 28 insertions, 42 deletions

diff --git a/include/cuda/cuda.h b/include/cuda/cuda.h
index 09c3c2b..94fc64a 100644
--- a/include/cuda/cuda.h
+++ b/include/cuda/cuda.h
@@ -147,7 +147,7 @@ typedef struct {
 
   size_t dstXInBytes, dstY;
   CUmemorytype dstMemoryType;
-  const void *dstHost;
+  void *dstHost;
   CUdeviceptr dstDevice;
   CUarray dstArray;
   size_t dstPitch;
@@ -162,16 +162,16 @@ typedef struct {
   const void *srcHost;
   CUdeviceptr srcDevice;
   CUarray srcArray;
-  void *dummy;
+  void *reserved0;
   size_t srcPitch, srcHeight;
 
   size_t dstXInBytes, dstY, dstZ;
   size_t dstLOD;
   CUmemorytype dstMemoryType;
-  const void *dstHost;
+  void *dstHost;
   CUdeviceptr dstDevice;
   CUarray dstArray;
-  void *dummy2;
+  void *reserved1;
   size_t dstPitch, dstHeight;
 
   size_t WidthInBytes, Height, Depth;
@@ -190,7 +190,7 @@ typedef struct {
   size_t dstXInBytes, dstY, dstZ;
   size_t dstLOD;
   CUmemorytype dstMemoryType;
-  const void *dstHost;
+  void *dstHost;
   CUdeviceptr dstDevice;
   CUarray dstArray;
   CUcontext dstContext;
@@ -246,6 +246,8 @@ CUresult cuMemAlloc (CUdeviceptr *, size_t);
 CUresult cuMemAllocHost (void **, size_t);
 CUresult cuMemHostAlloc (void **, size_t, unsigned int);
 CUresult cuMemcpy (CUdeviceptr, CUdeviceptr, size_t);
+CUresult cuMemcpyPeer (CUdeviceptr, CUcontext, CUdeviceptr, CUcontext, size_t);
+CUresult cuMemcpyPeerAsync (CUdeviceptr, CUcontext, CUdeviceptr, CUcontext, size_t, CUstream);
 #define cuMemcpyDtoDAsync cuMemcpyDtoDAsync_v2
 CUresult cuMemcpyDtoDAsync (CUdeviceptr, CUdeviceptr, size_t, CUstream);
 #define cuMemcpyDtoH cuMemcpyDtoH_v2
diff --git a/libgomp/plugin/plugin-nvptx.c b/libgomp/plugin/plugin-nvptx.c
index 9cdc55c..00d4241 100644
--- a/libgomp/plugin/plugin-nvptx.c
+++ b/libgomp/plugin/plugin-nvptx.c
@@ -1794,6 +1794,8 @@ GOMP_OFFLOAD_memcpy2d (int dst_ord, int src_ord, size_t dim1_size,
   /* TODO: Consider using CU_MEMORYTYPE_UNIFIED if supported.  */
 
   CUDA_MEMCPY2D data;
+
+  memset (&data, 0, sizeof (data));
   data.WidthInBytes = dim1_size;
   data.Height = dim0_len;
 
@@ -1855,6 +1857,8 @@ GOMP_OFFLOAD_memcpy3d (int dst_ord, int src_ord, size_t dim2_size,
   /* TODO: Consider using CU_MEMORYTYPE_UNIFIED if supported.  */
 
   CUDA_MEMCPY3D data;
+
+  memset (&data, 0, sizeof (data));
   data.WidthInBytes = dim2_size;
   data.Height = dim1_len;
   data.Depth = dim0_len;
@@ -1874,7 +1878,6 @@ GOMP_OFFLOAD_memcpy3d (int dst_ord, int src_ord, size_t dim2_size,
   data.dstXInBytes = dst_offset2_size;
   data.dstY = dst_offset1_len;
   data.dstZ = dst_offset0_len;
-  data.dstLOD = 0;
 
   if (src_ord == -1)
     {
@@ -1891,7 +1894,6 @@ GOMP_OFFLOAD_memcpy3d (int dst_ord, int src_ord, size_t dim2_size,
   data.srcXInBytes = src_offset2_size;
   data.srcY = src_offset1_len;
   data.srcZ = src_offset0_len;
-  data.srcLOD = 0;
 
   CUDA_CALL (cuMemcpy3D, &data);
   return true;
diff --git a/libgomp/target.c b/libgomp/target.c
index 5cf2e8d..cd4cc1b 100644
--- a/libgomp/target.c
+++ b/libgomp/target.c
@@ -4540,33 +4540,22 @@ omp_target_memcpy_rect_worker (void *dst, const void *src, size_t element_size,
 	  || __builtin_mul_overflow (element_size, dst_offsets[0], &dst_off)
 	  || __builtin_mul_overflow (element_size, src_offsets[0], &src_off))
 	return EINVAL;
-      if (src_devicep != NULL && src_devicep == dst_devicep)
-	ret = src_devicep->dev2dev_func (src_devicep->target_id,
-					 (char *) dst + dst_off,
-					 (const char *) src + src_off,
-					 length);
-      else if (src_devicep != NULL
-	       && (dst_devicep == NULL
-		   || (dst_devicep->capabilities
-		       & GOMP_OFFLOAD_CAP_SHARED_MEM)))
-	ret = src_devicep->dev2host_func (src_devicep->target_id,
+      if (dst_devicep == NULL && src_devicep == NULL)
+	{
+	  memcpy ((char *) dst + dst_off, (const char *) src + src_off,
+		  length);
+	  ret = 1;
+	}
+      else if (src_devicep == NULL)
+	ret = dst_devicep->host2dev_func (dst_devicep->target_id,
 					  (char *) dst + dst_off,
 					  (const char *) src + src_off,
 					  length);
-      else if (dst_devicep != NULL
-	       && (src_devicep == NULL
-		   || (src_devicep->capabilities
-			& GOMP_OFFLOAD_CAP_SHARED_MEM)))
-	ret = dst_devicep->host2dev_func (dst_devicep->target_id,
+      else if (dst_devicep == NULL)
+	ret = src_devicep->dev2host_func (src_devicep->target_id,
 					  (char *) dst + dst_off,
 					  (const char *) src + src_off,
 					  length);
-      else if (dst_devicep == NULL && src_devicep == NULL)
-	{
-	  memcpy ((char *) dst + dst_off, (const char *) src + src_off,
-		  length);
-	  ret = 1;
-	}
       else if (src_devicep == dst_devicep)
 	ret = src_devicep->dev2dev_func (src_devicep->target_id,
 					 (char *) dst + dst_off,
@@ -4584,7 +4573,8 @@ omp_target_memcpy_rect_worker (void *dst, const void *src, size_t element_size,
 	  else if (*tmp_size < length)
 	    {
 	      *tmp_size = length;
-	      *tmp = realloc (*tmp, length);
+	      free (*tmp);
+	      *tmp = malloc (length);
 	      if (*tmp == NULL)
 		return ENOMEM;
 	    }
@@ -4599,7 +4589,7 @@ omp_target_memcpy_rect_worker (void *dst, const void *src, size_t element_size,
       return ret ? 0 : EINVAL;
     }
 
-  /* host->device, device->host and same-device device->device.  */
+  /* host->device, device->host and intra device.  */
   if (num_dims == 2
       && ((src_devicep
 	   && src_devicep == dst_devicep
@@ -4711,16 +4701,8 @@ omp_target_memcpy_rect_copy (void *dst, const void *src,
   bool lock_src;
   bool lock_dst;
 
-  lock_src = (src_devicep
-	      && (!dst_devicep
-		  || src_devicep == dst_devicep
-		  || !(src_devicep->capabilities
-		       & GOMP_OFFLOAD_CAP_SHARED_MEM)));
-  lock_dst = (dst_devicep
-	      && (!lock_src
-		  || (src_devicep != dst_devicep
-		      && !(dst_devicep->capabilities
-			   & GOMP_OFFLOAD_CAP_SHARED_MEM))));
+  lock_src = src_devicep != NULL;
+  lock_dst = dst_devicep != NULL && src_devicep != dst_devicep;
   if (lock_src)
     gomp_mutex_lock (&src_devicep->lock);
   if (lock_dst)
@@ -5076,8 +5058,8 @@ gomp_load_plugin_for_device (struct gomp_device_descr *device,
   DLSYM (free);
   DLSYM (dev2host);
   DLSYM (host2dev);
-  DLSYM (memcpy2d);
-  DLSYM (memcpy3d);
+  DLSYM_OPT (memcpy2d, memcpy2d);
+  DLSYM_OPT (memcpy3d, memcpy3d);
   device->capabilities = device->get_caps_func ();
   if (device->capabilities & GOMP_OFFLOAD_CAP_OPENMP_400)
     {