Add comm_error impl and remove gpucomm_get_device

tsirif · tsirif · commit dcb0fa5d33e8 · 2016-06-23T02:49:20.000+03:00
diff --git a/src/gpuarray/buffer_collectives.h b/src/gpuarray/buffer_collectives.h
@@ -93,14 +93,6 @@ GPUARRAY_PUBLIC int gpucomm_gen_clique_id(gpucontext* ctx,
  */
 GPUARRAY_PUBLIC int gpucomm_get_count(gpucomm* comm, int* count);
 
-/**
- * \brief TODO
- * \param comm [gpucomm*] TODO
- * \param device [int*] TODO
- * \return int TODO
- */
-GPUARRAY_PUBLIC int gpucomm_get_device(gpucomm* comm, int* device);
-
 /**
  * \brief TODO
  * \param comm [gpucomm*] TODO
diff --git a/src/gpuarray_buffer_collectives.c b/src/gpuarray_buffer_collectives.c
@@ -42,13 +42,6 @@ int gpucomm_get_count(gpucomm* comm, int* count) {
   return ctx->comm_ops->get_count(comm, count);
 }
 
-int gpucomm_get_device(gpucomm* comm, int* device) {
-  gpucontext* ctx = gpucomm_context(comm);
-  if (ctx->comm_ops == NULL)
-    return GA_COMM_ERROR;
-  return ctx->comm_ops->get_device(comm, device);
-}
-
 int gpucomm_get_rank(gpucomm* comm, int* rank) {
   gpucontext* ctx = gpucomm_context(comm);
   if (ctx->comm_ops == NULL)
diff --git a/src/gpuarray_collectives_cuda_nccl.c b/src/gpuarray_collectives_cuda_nccl.c
@@ -44,13 +44,17 @@ static int comm_new(gpucomm** comm_ptr, gpucontext* ctx, gpucommCliqueId comm_id
 }
 
 static void comm_free(gpucomm* comm) {
+  ASSERT_COMM(comm);
   cuda_enter(comm->ctx);
   ncclCommDestroy(comm->c);
   cuda_exit(comm->ctx);
   comm_clear(comm);
 }
 
-static const char* comm_error(gpucontext* ctx) {
+static const char* comm_error(gpucontext* c) {
+  cuda_context* ctx = (cuda_context*) c;
+  // find a way to concatenate a constact "(nccl) " infront
+  return ncclGetErrorString(ctx->nccl_err);
 }
 
 static int generate_clique_id(gpucontext* ctx, gpucommCliqueId* cliqueId) {
@@ -59,9 +63,6 @@ static int generate_clique_id(gpucontext* ctx, gpucommCliqueId* cliqueId) {
 static int get_count(const gpucomm* comm, int* count) {
 }
 
-static int get_device(const gpucomm* comm, int* device) {
-}
-
 static int get_rank(const gpucomm* comm, int* rank) {
 }
 
@@ -99,7 +100,6 @@ GPUARRAY_LOCAL gpuarray_comm_ops nccl_ops = {
   comm_error,
   generate_clique_id,
   get_count,
-  get_device,
   get_rank,
   reduce,
   all_reduce,
diff --git a/src/private.h b/src/private.h
@@ -209,7 +209,6 @@ struct _gpuarray_comm_ops {
   const char* (*comm_error)(gpucontext* ctx);
   int (*generate_clique_id)(gpucontext* ctx, gpucommCliqueId* cliqueId);
   int (*get_count)(const gpucomm* comm, int* count);
-  int (*get_device)(const gpucomm* comm, int* device);
   int (*get_rank)(const gpucomm* comm, int* rank);
   // collective ops
   int (*reduce)(const gpudata* src, size_t offsrc,
