4141 } \
4242 } while (0)
4343
44- extern const gpuarray_buffer_ops
45- cuda_ops ; //!< Link wrapped cuda core operations
44+ //!< Link wrapped cuda core operations
45+ extern const gpuarray_buffer_ops cuda_ops ;
4646
4747/**
4848 * Definition of struct _gpucomm
@@ -73,19 +73,20 @@ static void comm_clear(gpucomm* comm) {
7373 */
7474static int comm_new (gpucomm * * comm_ptr , gpucontext * ctx ,
7575 gpucommCliqueId comm_id , int ndev , int rank ) {
76- ASSERT_CTX (ctx );
7776 gpucomm * comm ;
77+ ncclResult_t nccl_err ;
78+
79+ ASSERT_CTX (ctx );
7880 comm = calloc (1 , sizeof (* comm )); // Allocate memory
7981 if (comm == NULL ) {
8082 * comm_ptr = NULL ; // Set to NULL if failed
8183 return GA_MEMORY_ERROR ;
8284 }
8385 comm -> ctx = (cuda_context * )ctx ; // convert to underlying cuda context
84- comm -> ctx
85- -> refcnt ++ ; // So that context would not be destroyed before communicator
86+ // So that context would not be destroyed before communicator
87+ comm -> ctx -> refcnt ++ ;
8688 cuda_enter (comm -> ctx ); // Use device
87- ncclResult_t nccl_err =
88- ncclCommInitRank (& comm -> c , ndev , * ((ncclUniqueId * )& comm_id ), rank );
89+ nccl_err = ncclCommInitRank (& comm -> c , ndev , * ((ncclUniqueId * )& comm_id ), rank );
8990 cuda_exit (comm -> ctx );
9091 TAG_COMM (comm );
9192 if (nccl_err != ncclSuccess ) {
@@ -181,6 +182,7 @@ static inline int check_restrictions(gpudata* src, size_t offsrc, gpudata* dest,
181182 int opcode , gpucomm * comm ,
182183 ncclDataType_t * datatype ,
183184 ncclRedOp_t * op ) {
185+ size_t op_size ;
184186 // Check if count is larger than INT_MAX
185187 // TODO remove whenif nccl adapts to size_t
186188 if (count > INT_MAX )
@@ -207,7 +209,7 @@ static inline int check_restrictions(gpudata* src, size_t offsrc, gpudata* dest,
207209 assert (!(offsrc > src -> sz ));
208210 assert (!(dest != NULL && offdest > dest -> sz ));
209211 // size to operate upon must be able to fit inside the gpudata (incl offsets)
210- size_t op_size = count * gpuarray_get_elsize (typecode );
212+ op_size = count * gpuarray_get_elsize (typecode );
211213 if ((src -> sz - offsrc ) < op_size )
212214 return GA_VALUE_ERROR ;
213215 if (dest != NULL && (dest -> sz - offdest ) < op_size )
@@ -223,10 +225,12 @@ static int reduce(gpudata* src, size_t offsrc, gpudata* dest, size_t offdest,
223225 gpucomm * comm ) {
224226 ncclRedOp_t op ;
225227 ncclDataType_t datatype ;
226- ASSERT_BUF (src );
227- ASSERT_COMM (comm );
228228 gpudata * dst = NULL ;
229229 int rank = 0 ;
230+ cuda_context * ctx ;
231+
232+ ASSERT_BUF (src );
233+ ASSERT_COMM (comm );
230234 GA_CHECK (get_rank (comm , & rank ));
231235 if (rank == root ) {
232236 dst = dest ;
@@ -235,7 +239,7 @@ static int reduce(gpudata* src, size_t offsrc, gpudata* dest, size_t offdest,
235239 GA_CHECK (check_restrictions (src , offsrc , dst , offdest , count , typecode ,
236240 opcode , comm , & datatype , & op ));
237241
238- cuda_context * ctx = comm -> ctx ;
242+ ctx = comm -> ctx ;
239243 cuda_enter (ctx );
240244
241245 // sync: wait till a write has finished (out of concurrent kernels)
@@ -270,13 +274,15 @@ static int all_reduce(gpudata* src, size_t offsrc, gpudata* dest,
270274 gpucomm * comm ) {
271275 ncclRedOp_t op ;
272276 ncclDataType_t datatype ;
277+ cuda_context * ctx ;
278+
273279 ASSERT_BUF (src );
274280 ASSERT_COMM (comm );
275281 ASSERT_BUF (dest );
276282 GA_CHECK (check_restrictions (src , offsrc , dest , offdest , count , typecode ,
277283 opcode , comm , & datatype , & op ));
278284
279- cuda_context * ctx = comm -> ctx ;
285+ ctx = comm -> ctx ;
280286 cuda_enter (ctx );
281287
282288 // sync: wait till a write has finished (out of concurrent kernels)
@@ -303,23 +309,26 @@ static int all_reduce(gpudata* src, size_t offsrc, gpudata* dest,
303309static int reduce_scatter (gpudata * src , size_t offsrc , gpudata * dest ,
304310 size_t offdest , size_t count , int typecode ,
305311 int opcode , gpucomm * comm ) {
306- ASSERT_BUF (src );
307- ASSERT_COMM (comm );
308- ASSERT_BUF (dest );
309312 ncclRedOp_t op ;
310313 ncclDataType_t datatype ;
311314 int ndev = 0 ;
315+ size_t resc_size ;
316+ cuda_context * ctx ;
317+
318+ ASSERT_BUF (src );
319+ ASSERT_COMM (comm );
320+ ASSERT_BUF (dest );
312321 GA_CHECK (get_count (comm , & ndev ));
313322 GA_CHECK (check_restrictions (src , offsrc , NULL , 0 , count * ndev , typecode ,
314323 opcode , comm , & datatype , & op ));
315324 if (dest -> ctx != comm -> ctx )
316325 return GA_VALUE_ERROR ;
317- size_t resc_size = count * gpuarray_get_elsize (typecode );
326+ resc_size = count * gpuarray_get_elsize (typecode );
318327 if ((dest -> sz - offdest ) < resc_size )
319328 return GA_VALUE_ERROR ;
320329 assert (!(offdest > dest -> sz ));
321330
322- cuda_context * ctx = comm -> ctx ;
331+ ctx = comm -> ctx ;
323332 cuda_enter (ctx );
324333
325334 // sync: wait till a write has finished (out of concurrent kernels)
@@ -345,15 +354,17 @@ static int reduce_scatter(gpudata* src, size_t offsrc, gpudata* dest,
345354 */
346355static int broadcast (gpudata * array , size_t offset , size_t count , int typecode ,
347356 int root , gpucomm * comm ) {
357+ ncclDataType_t datatype ;
358+ int rank = 0 ;
359+ cuda_context * ctx ;
360+
348361 ASSERT_BUF (array );
349362 ASSERT_COMM (comm );
350- ncclDataType_t datatype ;
351363 GA_CHECK (check_restrictions (array , offset , NULL , 0 , count , typecode , 0 , comm ,
352364 & datatype , NULL ));
353- int rank = 0 ;
354365 GA_CHECK (get_rank (comm , & rank ));
355366
356- cuda_context * ctx = comm -> ctx ;
367+ ctx = comm -> ctx ;
357368 cuda_enter (ctx );
358369
359370 // sync: wait till a write has finished (out of concurrent kernels)
@@ -382,22 +393,25 @@ static int broadcast(gpudata* array, size_t offset, size_t count, int typecode,
382393static int all_gather (gpudata * src , size_t offsrc , gpudata * dest ,
383394 size_t offdest , size_t count , int typecode ,
384395 gpucomm * comm ) {
396+ ncclDataType_t datatype ;
397+ int ndev = 0 ;
398+ size_t resc_size ;
399+ cuda_context * ctx ;
400+
385401 ASSERT_BUF (src );
386402 ASSERT_COMM (comm );
387403 ASSERT_BUF (dest );
388- ncclDataType_t datatype ;
389404 GA_CHECK (check_restrictions (src , offsrc , NULL , 0 , count , typecode , 0 , comm ,
390405 & datatype , NULL ));
391406 if (dest -> ctx != comm -> ctx )
392407 return GA_VALUE_ERROR ;
393- int ndev = 0 ;
394408 GA_CHECK (get_count (comm , & ndev ));
395- size_t resc_size = ndev * count * gpuarray_get_elsize (typecode );
409+ resc_size = ndev * count * gpuarray_get_elsize (typecode );
396410 if ((dest -> sz - offdest ) < resc_size )
397411 return GA_VALUE_ERROR ;
398412 assert (!(offdest > dest -> sz ));
399413
400- cuda_context * ctx = comm -> ctx ;
414+ ctx = comm -> ctx ;
401415 cuda_enter (ctx );
402416
403417 // sync: wait till a write has finished (out of concurrent kernels)
@@ -420,8 +434,7 @@ static int all_gather(gpudata* src, size_t offsrc, gpudata* dest,
420434
421435/**
422436 * Instance of `gpuarray_comm_ops` which contains NCCL implementations. To be
423- * linked
424- * in \ref gpuarray_buffer_cuda.c, in order to fill a /ref gpucontext's
437+ * linked in \ref gpuarray_buffer_cuda.c, in order to fill a /ref gpucontext's
425438 * comm_ops.
426439 */
427440GPUARRAY_LOCAL gpuarray_comm_ops nccl_ops = {
0 commit comments