made documentation consistent

ROCm · Sep 26, 2018 · 35ce325 · 35ce325
1 parent 12ec1d3
commit 35ce325
Show file tree

Hide file tree

Showing 8 changed files with 32 additions and 24 deletions.
diff --git a/src/rccl.cpp b/src/rccl.cpp
@@ -47,12 +47,12 @@ struct RcclUniqueId {
     ~RcclUniqueId() { delete pool; }
 };
 
-//! Get value of environment variable RCCL_TRACE_RT
+//! @brief Get value of environment variable RCCL_TRACE_RT
 const char *get_env_val = getenv("RCCL_TRACE_RT");
-//! Get debug trace level from environment variable
+//! @brief Get debug trace level from environment variable
 int RCCL_TRACE_RT = get_env_val != nullptr ? atoi(get_env_val) : 0;
 
-//! Implementation of rcclGetErrorString
+//! @brief Implementation of rcclGetErrorString
 const char *rcclGetErrorString(rcclResult_t result) {
     switch (result) {
     case rcclSuccess:
@@ -90,7 +90,7 @@ const char *rcclGetErrorString(rcclResult_t result) {
     }
 }
 
-//! Definition of rcclGetUniqueId
+//! @brief Definition of rcclGetUniqueId
 rcclResult_t rcclGetUniqueId(rcclUniqueId *uniqueId) {
     if ((RCCL_TRACE_RT & krccl_print_api) == krccl_print_api) {
         fprintf(stderr, "%s<<rccl-api: %s uniqueId:%p%s\n", API_COLOR, __func__,
@@ -107,7 +107,7 @@ rcclResult_t rcclGetUniqueId(rcclUniqueId *uniqueId) {
     return rcclSuccess;
 }
 
-//! Definition of rcclCommInitRank
+//! @brief Definition of rcclCommInitRank
 rcclResult_t rcclCommInitRank(rcclComm_t *comm, int ndev, rcclUniqueId commId,
                               int rank) {
     if ((RCCL_TRACE_RT & krccl_print_api) == krccl_print_api) {
@@ -151,7 +151,7 @@ rcclResult_t rcclCommInitRank(rcclComm_t *comm, int ndev, rcclUniqueId commId,
     return rcclSuccess;
 }
 
-//! Definition of rcclCommInitAll
+//! @brief Definition of rcclCommInitAll
 rcclResult_t rcclCommInitAll(rcclComm_t *comm, int ndev, int *devlist) {
     if ((RCCL_TRACE_RT & krccl_print_api) == krccl_print_api) {
         fprintf(stderr, "%s<<rccl-api: %s comm:%p ndev:%d devlist:%p%s\n",
@@ -225,7 +225,7 @@ rcclResult_t rcclCommInitAll(rcclComm_t *comm, int ndev, int *devlist) {
     return rcclSuccess;
 }
 
-//! Declaration of rcclCommCuDevice
+//! @brief Declaration of rcclCommCuDevice
 rcclResult_t rcclCommCuDevice(rcclComm_t comm, int *dev) {
     if ((RCCL_TRACE_RT & krccl_print_api) == krccl_print_api) {
         fprintf(stderr, "%s<<rccl-api: %s comm:%p *dev:%d dev:%p%s\n",
@@ -238,7 +238,7 @@ rcclResult_t rcclCommCuDevice(rcclComm_t comm, int *dev) {
     return rcclSuccess;
 }
 
-//! Declaration of rcclCommUserRank
+//! @brief Declaration of rcclCommUserRank
 rcclResult_t rcclCommUserRank(rcclComm_t comm, int *rank) {
     if ((RCCL_TRACE_RT & krccl_print_api) == krccl_print_api) {
         fprintf(stderr, "%s<<rccl-api: %s comm:%p *rank:%d rank:%p%s\n",
@@ -251,7 +251,7 @@ rcclResult_t rcclCommUserRank(rcclComm_t comm, int *rank) {
     return rcclSuccess;
 }
 
-//! Declaration of rcclCommCount
+//! @brief Declaration of rcclCommCount
 rcclResult_t rcclCommCount(rcclComm_t comm, int *count) {
     if ((RCCL_TRACE_RT & krccl_print_api) == krccl_print_api) {
         fprintf(stderr, "%s<<rccl-api: %s comm:%p *count:%d count:%p%s\n",
@@ -264,7 +264,7 @@ rcclResult_t rcclCommCount(rcclComm_t comm, int *count) {
     return rcclSuccess;
 }
 
-//! Declaration of rcclCommDestroy
+//! @brief Declaration of rcclCommDestroy
 rcclResult_t rcclCommDestroy(rcclComm_t comm) {
     if ((RCCL_TRACE_RT & krccl_print_api) == krccl_print_api) {
         fprintf(stderr, "%s<<rccl-api: %s comm:%p%s\n", API_COLOR, __func__,
@@ -279,14 +279,14 @@ rcclResult_t rcclCommDestroy(rcclComm_t comm) {
     return rcclSuccess;
 }
 
-//! Declaration of PostEnqueueEventRecord
+//! @brief Declaration of PostEnqueueEventRecord
 void PostEnqueueEventRecord(RcclComm_t *pcomm, hipStream_t stream) {
     if (stream != pcomm->stream_) {
         hipEventRecord(pcomm->event_, stream);
     }
 }
 
-//! Declaration of PreEnqueueEventRecord
+//! @brief Declaration of PreEnqueueEventRecord
 void PreEnqueueEventRecord(RcclComm_t *pcomm, hipStream_t stream) {
     if (stream != pcomm->stream_) {
         hipStreamWaitEvent(stream, pcomm->event_, 0);

diff --git a/src/rcclAllReduce.cpp b/src/rcclAllReduce.cpp
@@ -27,7 +27,7 @@ extern std::unordered_map<int, std::string> umap_datatype;
 
 extern int RCCL_TRACE_RT;
 
-//! Definition of rcclAllReduce
+//! @brief Definition of rcclAllReduce
 rcclResult_t rcclAllReduce(const void *sendbuff, void *recvbuff, int count,
                            rcclDataType_t datatype, rcclRedOp_t op,
                            rcclComm_t comm, hipStream_t stream) {

diff --git a/src/rcclBcast.cpp b/src/rcclBcast.cpp
@@ -27,7 +27,7 @@ extern std::unordered_map<int, std::string> umap_datatype;
 
 extern int RCCL_TRACE_RT;
 
-//! Definition of rcclBcast
+//! @brief Definition of rcclBcast
 rcclResult_t rcclBcast(void *buff, int count, rcclDataType_t datatype, int root,
                        rcclComm_t comm, hipStream_t stream) {
     if ((RCCL_TRACE_RT & krccl_print_api) == krccl_print_api) {

diff --git a/src/rcclHelper.h b/src/rcclHelper.h
@@ -16,8 +16,16 @@ All rights reserved.
 #include <hip/hip_runtime_api.h>
 #include "rcclTracker.h"
 
-//! Record event on the stream before launching kernels related to op
-void PreEnqueueEventRecord(RcclComm_t*, hipStream_t);
+//! Synchronize current stream with stream used before with the same
+//! communicator. If previous stream is same as current stream, don't do
+//! anything
+
+//! \param [in] comm Memory location to internal Rccl communicator
+//! \param [in] stream Stream with which the op will be synchronized with
+void PreEnqueueEventRecord(RcclComm_t* comm, hipStream_t stream);
 
 //! Record event on the stream after launching kernels related to op
-void PostEnqueueEventRecord(RcclComm_t*, hipStream_t);
+
+//! \param [in] comm Memory location to internal Rccl communicator
+//! \param [in] stream Stream with which the op will be synchronized with
+void PostEnqueueEventRecord(RcclComm_t* comm, hipStream_t stream);
diff --git a/src/rcclScalarBroadcastKernels.h b/src/rcclScalarBroadcastKernels.h
@@ -4,7 +4,7 @@ All rights reserved.
 */
 
 /**
- * @file rcclScalarBcastKernels.h
+ * @file rcclScalarBroadcastKernels.h
  * @brief Implementation of root copy kernel
  *
  * This file contains a kernel which reads data from root gpu
@@ -13,7 +13,7 @@ All rights reserved.
  */
 #pragma once
 
-//! Definition of RcclKernelScalarCopyFromRoot
+//! @brief Definition of RcclKernelScalarCopyFromRoot
 template <typename DataType_t>
 __global__ void RcclKernelScalarCopyFromRoot(RingNode_t* proot_track,
                                              void* recv_buff, int count) {

diff --git a/src/rcclScalarBroadcastRuntime.h b/src/rcclScalarBroadcastRuntime.h
@@ -4,7 +4,7 @@ All rights reserved.
 */
 
 /**
- * @file rcclScalarBcastRuntime.h
+ * @file rcclScalarBroadcastRuntime.h
  * @brief Implemenation of rcclBcast internally
  *
  * This file contains internal implementation of rcclBcast

diff --git a/src/rcclScalarReduceKernels.h b/src/rcclScalarReduceKernels.h
@@ -15,7 +15,7 @@ All rights reserved.
  */
 
 //! @brief Definition of RcclKernelScalarReduce
-//!
+//! Gather data from non-root gpus and do reduction op on it
 template <typename DataType_t, rcclRedOp_t Op>
 __global__ void RcclKernelScalarReduce(RingNode_t* pcurr_track, void* send_buff,
                                        void* recv_buff, int count) {

diff --git a/src/rcclTracker.h b/src/rcclTracker.h
@@ -86,7 +86,7 @@ struct RingNode_t {
 
 struct RcclComm_t;
 
-//! @brief
+//! @brief Definition of RingNodePool_t
 //! Pool data structure used to store all RingNode_t data structures and track
 //! rcclComm_t accordingly
 class RingNodePool_t {
@@ -126,8 +126,8 @@ class RingNodePool_t {
     RingNode_t* GetPoolByDeviceIndex(int device_index);
 };
 
-//! Internal representation of rcclComm_t structure, which is allocated for each
-//! gpu.
+//! @brief Internal representation of rcclComm_t structure, which is allocated
+//! for each gpu.
 struct RcclComm_t {
   public:
     //! Pool of gpus rcclComm_t is created with