AI/rocm-systems
2
0
Форкнуть 0
Код Задачи Запросы на слияние Действия Пакеты Проекты Релизы Вики Активность

2.17.1-1

Просмотр исходного кода 
Add new NVLS algorithm for allreduce using NVLink SHARP (intra-node only).
Add new config options: cgaClusterSize, minCTAs, maxCTAs, netName.
Enable LL128 when we use PXN to close rings.
NVTX3 includes update.
Fix crash when one CollNet (SHARP) rail fails to initialize.


[ROCm/rccl commit: 5d3ab08b69]
Этот коммит содержится в:
Sylvain Jeaugey
2023-02-27 02:48:21 -08:00
родитель e6e8f2555c
Коммит 8dcf8e8720
72 изменённых файлов: 4541 добавлений и 2391 удалений
Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы
projects/rccl/src/include/bootstrap.h
+1
Просмотреть файл
@@ -25,6 +25,7 @@ ncclResult_t bootstrapSend(void* commState, int peer, int tag, void* data, int s
ncclResult_t bootstrapRecv(void* commState, int peer, int tag, void* data, int size);
ncclResult_t bootstrapBarrier(void* commState, int *ranks, int rank, int nranks, int tag);
ncclResult_t bootstrapIntraNodeAllGather(void* commState, int *ranks, int rank, int nranks, void* allData, int size);
ncclResult_t bootstrapIntraNodeBroadcast(void* commState, int *ranks, int rank, int nranks, int root, void* bcastData, int size);
ncclResult_t bootstrapClose(void* commState);
ncclResult_t bootstrapAbort(void* commState);
#endif
projects/rccl/src/include/collectives.h
+11 -1
Просмотреть файл
@@ -53,7 +53,8 @@ struct ncclDevRedOpFull {
DECL4(func, RING, devredop, type, undef) \
DECL4(func, TREE, devredop, type, undef) \
DECL4(func, COLLNET_DIRECT, devredop, type, undef) \
DECL4(func, COLLNET_CHAIN, devredop, type, undef)
DECL4(func, COLLNET_CHAIN, devredop, type, undef) \
DECL4(func, NVLS, devredop, type, undef)
#if defined(__CUDA_BF16_TYPES_EXIST__)
#define DECL2(func, devredop, undefForFloat) \
@@ -121,4 +122,13 @@ extern __device__ void NCCL_ONERANK_REDUCE_NAME(PreMulSum, double)();
#define REDUCE_CHUNKSTEPS 1
#define NCCL_MAX_SLICE_PER_CHUNK 2 // max value for CHUNKSTEPS/SLICESTEPS, must accord with above
// We can't use the enum identifiers like ncclSum, ncclFloat, etc since this
// macro will be used in preprocessor conditionals where enums have no meaning.
#define NCCL_NVLS_SUPPORTS(/*ncclDataType_t*/ type, /*ncclDevRedOp_t*/ red) \
(((type==2 || type==3) && (red==0 || red==2 || red==3)) || \
((type==4 || type==5) && (red==0 || red==2 || red==3)) || \
((type==6 || type==9) && (red==0 || red==2 || red==3)) || \
(type==7 && red==0) || \
(type==8 && red==0))
#endif
projects/rccl/src/include/comm.h
+13
Просмотреть файл
@@ -104,6 +104,7 @@ struct ncclChannel {
struct ncclTree tree;
struct ncclTree collnetChain;
struct ncclDirect collnetDirect;
struct ncclNvls nvls;
int id; // index of this channel
uint32_t workFifoSent; // last used work index+1
uint64_t p2pOpCount;
@@ -177,8 +178,10 @@ struct ncclComm {
int nRanks; // number of GPUs in communicator
int cudaDev; // my cuda device index
int compCap; // compute capability of the GPU
int minCompCap; // min compute capability in the communicator
int64_t busId; // my PCI bus ID in int format
cpu_set_t cpuAffinity; // CPU affinity of the GPU
int cudaArch; // matches __CUDA_ARCH__ of device
int node;
int nNodes;
@@ -201,6 +204,7 @@ struct ncclComm {
// Channels for collectives
int nChannels;
int nvlsChannels;
// Channels (per peer) for p2p
int p2pnChannels;
int p2pnChannelsPerPeer;
@@ -257,6 +261,10 @@ struct ncclComm {
int collNetSupport;
int intraHighestTransportType;
// NVLink SHARP (NVLS) support
int nvlsSupport;
void* nvlsResources;
size_t channelSize; // User requested work size (bytes) for channel partitions
// Internal streams
@@ -288,6 +296,11 @@ struct ncclComm {
// communicator mode
int blocking;
// CGA cluster size
int cgaClusterSize;
int minCTAs, maxCTAs;
// network interface name
char *netName;
// initState is to more conveniently reclaim resources when errors happen.
ncclResult_t initState;
// flag to indicate if ncclCommFinalize() is called
projects/rccl/src/include/cudawrap.h
+23 -1
Просмотреть файл
@@ -73,10 +73,32 @@ DECLARE_CUDA_PFN_EXTERN(cuGetErrorName, 6000);
DECLARE_CUDA_PFN_EXTERN(cuMemGetAddressRange, 3020);
DECLARE_CUDA_PFN_EXTERN(cuCtxCreate, 3020);
DECLARE_CUDA_PFN_EXTERN(cuCtxDestroy, 4000);
DECLARE_CUDA_PFN_EXTERN(cuCtxGetCurrent, 4000);
DECLARE_CUDA_PFN_EXTERN(cuCtxSetCurrent, 4000);
DECLARE_CUDA_PFN_EXTERN(cuCtxGetDevice, 2000);
// cuMem API support
DECLARE_CUDA_PFN_EXTERN(cuMemAddressReserve, 10020);
DECLARE_CUDA_PFN_EXTERN(cuMemAddressFree, 10020);
DECLARE_CUDA_PFN_EXTERN(cuMemCreate, 10020);
DECLARE_CUDA_PFN_EXTERN(cuMemGetAllocationGranularity, 10020);
DECLARE_CUDA_PFN_EXTERN(cuMemExportToShareableHandle, 10020);
DECLARE_CUDA_PFN_EXTERN(cuMemImportFromShareableHandle, 10020);
DECLARE_CUDA_PFN_EXTERN(cuMemMap, 10020);
DECLARE_CUDA_PFN_EXTERN(cuMemRelease, 10020);
DECLARE_CUDA_PFN_EXTERN(cuMemSetAccess, 10020);
DECLARE_CUDA_PFN_EXTERN(cuMemUnmap, 10020);
#if CUDA_VERSION >= 11070
DECLARE_CUDA_PFN_EXTERN(cuMemGetHandleForAddressRange, 11070); // DMA-BUF support
#endif
#if CUDA_VERSION >= 12010
/* NVSwitch Multicast support */
DECLARE_CUDA_PFN_EXTERN(cuMulticastAddDevice, 12010);
DECLARE_CUDA_PFN_EXTERN(cuMulticastBindMem, 12010);
DECLARE_CUDA_PFN_EXTERN(cuMulticastBindAddr, 12010);
DECLARE_CUDA_PFN_EXTERN(cuMulticastCreate, 12010);
DECLARE_CUDA_PFN_EXTERN(cuMulticastGetGranularity, 12010);
DECLARE_CUDA_PFN_EXTERN(cuMulticastUnbind, 12010);
#endif
#endif
/* CUDA Driver functions loaded with dlsym() */
@@ -88,6 +110,7 @@ DECLARE_CUDA_PFN_EXTERN(cuGetProcAddress, 11030);
ncclResult_t ncclCudaLibraryInit(void);
extern int ncclCudaDriverVersionCache;
extern bool ncclCudaLaunchBlocking; // initialized by ncclCudaLibraryInit()
inline ncclResult_t ncclCudaDriverVersion(int* driver) {
int version = __atomic_load_n(&ncclCudaDriverVersionCache, __ATOMIC_RELAXED);
@@ -98,5 +121,4 @@ inline ncclResult_t ncclCudaDriverVersion(int* driver) {
*driver = version;
return ncclSuccess;
}
#endif
projects/rccl/src/include/devcomm.h
+78 -5
Просмотреть файл
@@ -15,11 +15,12 @@
typedef enum { ncclFuncBroadcast, ncclFuncReduce, ncclFuncAllGather, ncclFuncReduceScatter, ncclFuncAllReduce, ncclFuncSendRecv, ncclFuncSend, ncclFuncRecv, ncclNumFuncs} ncclFunc_t;
extern const char* ncclFuncStr[NCCL_NUM_FUNCTIONS];
#define NCCL_NUM_ALGORITHMS 4 // Tree/Ring/CollNet*
#define NCCL_NUM_ALGORITHMS 5 // Tree/Ring/CollNet*
#define NCCL_ALGO_TREE 0
#define NCCL_ALGO_RING 1
#define NCCL_ALGO_COLLNET_DIRECT 2
#define NCCL_ALGO_COLLNET_CHAIN 3
#define NCCL_ALGO_NVLS 4
extern const char* ncclAlgoStr[NCCL_NUM_ALGORITHMS];
#define NCCL_NUM_PROTOCOLS 3 // Simple/LL/LL128
@@ -78,6 +79,7 @@ static_assert(NCCL_LL_CLEAN_MASK % NCCL_STEPS == 0, "Invalid NCCL_LL_CLEAN_MASK
#define NCCL_DIRECT_NIC 0x04
#define NCCL_IPC_WRITE 0x08
#define NCCL_IPC_READ 0x10
#define NCCL_NVLS_MIN_POLL 0x20
struct ncclConnInfo {
// Regular comm mechanism
@@ -85,7 +87,7 @@ struct ncclConnInfo {
uint64_t *tail; // Local for recv, remote for send
uint64_t *head; // Local for send, remote for recv
int direct; // Direct communication
int flags; // Direct communication / other flags
int shared; // Buffers are shared
void **ptrExchange; // Pointer exchange for direct communication
uint64_t* redOpArgExchange; // PreOp scaler exchange for direct pull case
@@ -138,13 +140,22 @@ struct ncclTree {
struct ncclDirect {
int depth;
int out;
int nHeads;
int headRank;
int shift;
int nHeads; // Number of parallel N<->1<->net operations we'll do in parallel; size of up/down
int headRank; // Index in 0..nHeads-1 I am the head rank of. -1 if I'm not a head rank (no local NIC)
int shift; // Shuffling of send/recv for scatter/gather operations, basically localRank%nHeads
int up[NCCL_MAX_DIRECT_ARITY];
int down[NCCL_MAX_DIRECT_ARITY];
};
#define NCCL_MAX_NVLS_ARITY 8
struct ncclNvls {
int out;
int nHeads; // Number of parallel N<->1<->net operations we'll do in parallel; size of up/down
int headRank; // Index in 0..nHeads-1 I am the head rank of. -1 if I'm not a head rank (no local NIC)
int up[NCCL_MAX_NVLS_ARITY];
int down;
};
#define NCCL_MAX_CONNS 2
struct ncclChannelPeer {
struct ncclConnector send[NCCL_MAX_CONNS];
@@ -264,6 +275,7 @@ struct alignas(16) ncclDevChannel {
struct ncclTree tree;
struct ncclTree collnetChain;
struct ncclDirect collnetDirect;
struct ncclNvls nvls;
uint32_t* workFifoDone; // Location of done counter, device writes index+1 of last work processed
};
@@ -288,4 +300,65 @@ struct alignas(16) ncclDevCommAndChannels {
struct ncclDevChannel channels[MAXCHANNELS];
};
#ifdef __CUDA_ARCH__
#define NCCL_CUDA_ARCH __CUDA_ARCH__
#else
#define NCCL_CUDA_ARCH 0
#endif
template<typename T>
__host__ __device__ constexpr T min_constexpr(T a) { return a; }
template<typename T, typename ...Ts>
__host__ __device__ constexpr T min_constexpr(T a, T b, Ts ...c) {
return min_constexpr<T>((a < b ? a : b), c...);
}
template<typename T>
__host__ __device__ constexpr T max_constexpr(T a) { return a; }
template<typename T, typename ...Ts>
__host__ __device__ constexpr T max_constexpr(T a, T b, Ts ...c) {
return max_constexpr<T>((a > b ? a : b), c...);
}
// Calculate the unroll factor given:
// * bytePerPack: number of bytes accessed per instruction
// * insns: max permissible unroll value
// * bytes: desired number of in-flight bytes per iteration ( = unroll*bytePerPack)
__host__ __device__ constexpr int ncclCalcUnroll(int bytePerPack, int insns, int bytes) {
return min_constexpr(insns, (bytes + bytePerPack-1)/bytePerPack);
}
// Note that all unroll value logic should depend on a given cudaArch argument
// and not __CUDA_ARCH__ since these need to be host-side executable where the
// arch value is strictly runtime only. By defaulting to NCCL_CUDA_ARCH, device
// side code can elide passing the arch for brevity.
__host__ __device__ constexpr int ncclCollUnroll(int cudaArch = NCCL_CUDA_ARCH) {
// Our collective unroll should move to the same bytes&insns model as NVLS.
return cudaArch >= 800 ? 8 : 4;
}
__host__ __device__ constexpr int ncclNvlsUnrollBytes(int cudaArch = NCCL_CUDA_ARCH) { return 4*16; }
__host__ __device__ constexpr int ncclNvlsUnrollInsns(int cudaArch = NCCL_CUDA_ARCH) { return 16; }
__host__ __device__ constexpr int ncclNvlsUnroll(int bytePerPack, int cudaArch = NCCL_CUDA_ARCH) {
return ncclCalcUnroll(bytePerPack, ncclNvlsUnrollInsns(cudaArch), ncclNvlsUnrollBytes(cudaArch));
}
// The amount of dynamic shmem per warp
__host__ __device__ constexpr int ncclShmemScratchWarpSize(int cudaArch = NCCL_CUDA_ARCH) {
return (max_constexpr<int>(
/*LL */0,
/*LL128 */(NCCL_LL128_SHMEM_ELEMS_PER_THREAD*WARP_SIZE)*sizeof(uint64_t),
/*SIMPLE*/(ncclCollUnroll(cudaArch)*WARP_SIZE + 1)*16,
// NVLS needs an extra 16B to read unaligned data.
/*NVLS */WARP_SIZE*(cudaArch >= 900 ? ncclNvlsUnrollBytes(cudaArch) : 0) + 16
) + 15) & -16; // pad to 16 bytes
}
// The amount of dynamic shmem per block
__host__ __device__ constexpr int ncclShmemDynamicSize(int cudaArch = NCCL_CUDA_ARCH) {
return cudaArch < 700 ? 0 : ncclShmemScratchWarpSize(cudaArch)*(NCCL_MAX_NTHREADS/WARP_SIZE);
}
#endif
projects/rccl/src/include/enqueue.h
+1 -2
Просмотреть файл
@@ -15,8 +15,7 @@
#define NCCL_MIN_CHANNEL_SIZE (NCCL_LL_THREAD_THRESHOLD*64)
#define NCCL_AGG_CHANNEL_SIZE (1LL << 21) /* 2 MiB, ideal per-channel size to fully utilize bandwidth */
size_t ncclKernMaxLocalSize();
ncclResult_t ncclKernSetSharedMemoryCarveout(int carveOut);
ncclResult_t ncclInitKernelsForDevice(int cudaArch, size_t* maxStackSize);
ncclResult_t ncclEnqueueCheck(struct ncclInfo* info);
ncclResult_t ncclLaunchPrepare(struct ncclComm* comm);
ncclResult_t ncclLaunchKernelBefore_NoUncapturedCuda(struct ncclComm* comm, struct ncclKernelPlan* plan);
projects/rccl/src/include/info.h
+1
Просмотреть файл
@@ -24,6 +24,7 @@ typedef enum : uint8_t {
ncclPatternTreeUpDown,
ncclPatternCollnetChain,
ncclPatternCollnetDirect,
ncclPatternNvls,
ncclPatternSend,
ncclPatternRecv
} ncclPattern_t;
projects/rccl/src/include/ipcsocket.h
+37
Просмотреть файл
@@ -0,0 +1,37 @@
/*
* Copyright (c) 2016-2023, NVIDIA CORPORATION. All rights reserved.
*
* See COPYRIGHT for license information
*/
#ifndef NCCL_IPCSOCKET_H
#define NCCL_IPCSOCKET_H
#include "nccl.h"
#include <stdio.h>
#include <fcntl.h>
#include <sys/mman.h>
#include <unistd.h>
#include <errno.h>
#include <sys/wait.h>
#include <sys/types.h>
#include <sys/socket.h>
#include <memory.h>
#include <sys/un.h>
#include <inttypes.h>
#define NCCL_IPC_SOCKNAME_LEN 64
struct ncclIpcSocket {
int fd;
char socketName[NCCL_IPC_SOCKNAME_LEN];
volatile uint32_t* abortFlag;
};
ncclResult_t ncclIpcSocketInit(struct ncclIpcSocket *handle, int rank, uint64_t hash, volatile uint32_t* abortFlag);
ncclResult_t ncclIpcSocketClose(struct ncclIpcSocket *handle);
ncclResult_t ncclIpcSocketRecvFd(struct ncclIpcSocket *handle, int *fd);
ncclResult_t ncclIpcSocketSendFd(struct ncclIpcSocket *handle, const int fd, int rank, uint64_t hash);
#endif /* NCCL_IPCSOCKET_H */
projects/rccl/src/include/nccl_net.h
+1 -1
Просмотреть файл
@@ -20,7 +20,7 @@
#define NCCL_NET_MAX_REQUESTS 8
typedef enum {NCCL_LOG_NONE=0, NCCL_LOG_VERSION=1, NCCL_LOG_WARN=2, NCCL_LOG_INFO=3, NCCL_LOG_ABORT=4, NCCL_LOG_TRACE=5} ncclDebugLogLevel;
typedef enum {NCCL_INIT=1, NCCL_COLL=2, NCCL_P2P=4, NCCL_SHM=8, NCCL_NET=16, NCCL_GRAPH=32, NCCL_TUNING=64, NCCL_ENV=128, NCCL_ALLOC=256, NCCL_CALL=512, NCCL_ALL=~0} ncclDebugLogSubSys;
typedef enum {NCCL_INIT=1, NCCL_COLL=2, NCCL_P2P=4, NCCL_SHM=8, NCCL_NET=16, NCCL_GRAPH=32, NCCL_TUNING=64, NCCL_ENV=128, NCCL_ALLOC=256, NCCL_CALL=512, NCCL_PROXY=1024, NCCL_NVLS=2048, NCCL_ALL=~0} ncclDebugLogSubSys;
typedef void (*ncclDebugLogger_t)(ncclDebugLogLevel level, unsigned long flags, const char *file, int line, const char *fmt, ...);
projects/rccl/src/include/nvtx.h
+8 -8
Просмотреть файл
@@ -7,12 +7,12 @@
#ifndef NCCL_NVTX_H_
#define NCCL_NVTX_H_
#include "nvtx3.hpp"
#include "nvtx3/nvtx3.hpp"
#if __cpp_constexpr >= 201304L && !defined(NVTX3_RELAXED_CONSTEXPR)
#define NVTX3_RELAXED_CONSTEXPR constexpr
#if __cpp_constexpr >= 201304L && !defined(NVTX3_CONSTEXPR_IF_CPP14)
#define NVTX3_CONSTEXPR_IF_CPP14 constexpr
#else
#define NVTX3_RELAXED_CONSTEXPR
#define NVTX3_CONSTEXPR_IF_CPP14
#endif
// Define all NCCL-provided static schema IDs here (avoid duplicates).
@@ -37,7 +37,7 @@ struct nccl_domain{static constexpr char const* name{"NCCL"};};
class payload_schema {
public:
NVTX3_RELAXED_CONSTEXPR explicit payload_schema(const nvtxPayloadSchemaEntry_t entries[], size_t numEntries, const uint64_t schemaId, const char* schemaName = nullptr) noexcept
explicit payload_schema(const nvtxPayloadSchemaEntry_t entries[], size_t numEntries, const uint64_t schemaId, const char* schemaName = nullptr) noexcept
{
schema_attr.name = schemaName;
schema_attr.entries = entries;
@@ -74,11 +74,11 @@ class payload_schema {
#define NVTX3_FUNC_WITH_PARAMS(ID, S, P) \
static const payload_schema schema{S, std::extent<decltype(S)>::value, \
NVTX_PAYLOAD_ENTRY_TYPE_SCHEMA_ID_STATIC_START + NVTX_SID_##ID, #ID}; \
static ::nvtx3::v1::registered_string<nccl_domain> const nvtx3_func_name__{__func__}; \
static ::nvtx3::v1::registered_string_in<nccl_domain> const nvtx3_func_name__{__func__}; \
nvtxPayloadData_t nvtx3_bpl__[] = { \
{NVTX_PAYLOAD_ENTRY_TYPE_SCHEMA_ID_STATIC_START + NVTX_SID_##ID, sizeof(P), &(P)}}; \
::nvtx3::v1::event_attributes nvtx3_func_attr__{nvtx3_func_name__, nvtx3_bpl__}; \
::nvtx3::v1::domain_thread_range<nccl_domain> const nvtx3_range__{nvtx3_func_attr__};
::nvtx3::v1::event_attributes const nvtx3_func_attr__{nvtx3_func_name__, nvtx3_bpl__}; \
::nvtx3::v1::scoped_range_in<nccl_domain> const nvtx3_range__{nvtx3_func_attr__};
extern void initNvtxRegisteredEnums();
projects/rccl/src/include/nvtx3/nvToolsExt.h
+1 -1
Просмотреть файл
@@ -1,5 +1,5 @@
/*
* Copyright 2009-2020 NVIDIA Corporation. All rights reserved.
* Copyright 2009-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
projects/rccl/src/include/nvtx3/nvToolsExtCuda.h
+1 -1
Просмотреть файл
@@ -1,5 +1,5 @@
/*
* Copyright 2009-2020 NVIDIA Corporation. All rights reserved.
* Copyright 2009-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
projects/rccl/src/include/nvtx3/nvToolsExtCudaRt.h
+1 -1
Просмотреть файл
@@ -1,5 +1,5 @@
/*
* Copyright 2009-2020 NVIDIA Corporation. All rights reserved.
* Copyright 2009-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
projects/rccl/src/include/nvtx3/nvToolsExtOpenCL.h
+1 -1
Просмотреть файл
@@ -1,5 +1,5 @@
/*
* Copyright 2009-2020 NVIDIA Corporation. All rights reserved.
* Copyright 2009-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
projects/rccl/src/include/nvtx3/nvToolsExtPayload.h
+2 -2
Просмотреть файл
@@ -1,12 +1,12 @@
/*
* Copyright 2021 NVIDIA Corporation. All rights reserved.
* Copyright 2021-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
* SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
*/
#include "nvtx3/nvToolsExt.h"
#include "nvToolsExt.h"
#ifndef NVTOOLSEXT_PAYLOAD_H
#define NVTOOLSEXT_PAYLOAD_H
projects/rccl/src/include/nvtx3/nvToolsExtSync.h
+1 -1
Просмотреть файл
@@ -1,5 +1,5 @@
/*
* Copyright 2009-2020 NVIDIA Corporation. All rights reserved.
* Copyright 2009-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
projects/rccl/src/include/nvtx3.hpp → projects/rccl/src/include/nvtx3/nvtx3.hpp
+1233 -613
Просмотреть файл
Разница между файлами не показана из-за своего большого размера Загрузить разницу
projects/rccl/src/include/nvtx3/nvtxDetail/nvtxImpl.h
+1 -1
Просмотреть файл
@@ -1,5 +1,5 @@
/*
* Copyright 2009-2020 NVIDIA Corporation. All rights reserved.
* Copyright 2009-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
projects/rccl/src/include/nvtx3/nvtxDetail/nvtxImplCore.h
+1 -1
Просмотреть файл
@@ -1,5 +1,5 @@
/*
* Copyright 2009-2020 NVIDIA Corporation. All rights reserved.
* Copyright 2009-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
projects/rccl/src/include/nvtx3/nvtxDetail/nvtxImplCudaRt_v3.h
+1 -1
Просмотреть файл
@@ -1,5 +1,5 @@
/*
* Copyright 2009-2020 NVIDIA Corporation. All rights reserved.
* Copyright 2009-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
projects/rccl/src/include/nvtx3/nvtxDetail/nvtxImplCuda_v3.h
+1 -1
Просмотреть файл
@@ -1,5 +1,5 @@
/*
* Copyright 2009-2020 NVIDIA Corporation. All rights reserved.
* Copyright 2009-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
projects/rccl/src/include/nvtx3/nvtxDetail/nvtxImplOpenCL_v3.h
+1 -1
Просмотреть файл
@@ -1,5 +1,5 @@
/*
* Copyright 2009-2020 NVIDIA Corporation. All rights reserved.
* Copyright 2009-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
projects/rccl/src/include/nvtx3/nvtxDetail/nvtxImplSync_v3.h
+1 -1
Просмотреть файл
@@ -1,5 +1,5 @@
/*
* Copyright 2009-2020 NVIDIA Corporation. All rights reserved.
* Copyright 2009-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
projects/rccl/src/include/nvtx3/nvtxDetail/nvtxInit.h
+1 -1
Просмотреть файл
@@ -1,5 +1,5 @@
/*
* Copyright 2009-2020 NVIDIA Corporation. All rights reserved.
* Copyright 2009-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
projects/rccl/src/include/nvtx3/nvtxDetail/nvtxInitDecls.h
+1 -1
Просмотреть файл
@@ -1,5 +1,5 @@
/*
* Copyright 2009-2020 NVIDIA Corporation. All rights reserved.
* Copyright 2009-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
projects/rccl/src/include/nvtx3/nvtxDetail/nvtxInitDefs.h
+1 -1
Просмотреть файл
@@ -1,5 +1,5 @@
/*
* Copyright 2009-2020 NVIDIA Corporation. All rights reserved.
* Copyright 2009-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
projects/rccl/src/include/nvtx3/nvtxDetail/nvtxLinkOnce.h
+1 -1
Просмотреть файл
@@ -1,5 +1,5 @@
/*
* Copyright 2009-2020 NVIDIA Corporation. All rights reserved.
* Copyright 2009-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
projects/rccl/src/include/nvtx3/nvtxDetail/nvtxTypes.h
+1 -1
Просмотреть файл
@@ -1,5 +1,5 @@
/*
* Copyright 2009-2020 NVIDIA Corporation. All rights reserved.
* Copyright 2009-2022 NVIDIA Corporation. All rights reserved.
*
* Licensed under the Apache License v2.0 with LLVM Exceptions.
* See https://llvm.org/LICENSE.txt for license information.
projects/rccl/src/include/nvtx3/nvtxExtDetail/nvtxExtImplPayload_v1.h
+2 -1
Просмотреть файл
@@ -35,10 +35,11 @@ NVTX_EXT_PAYLOAD_VERSIONED_ID(nvtxExtPayloadSlots)[NVTX3EXT_CBID_PAYLOAD_FN_NUM
NVTX_LINKONCE_DEFINE_FUNCTION void NVTX_EXT_PAYLOAD_VERSIONED_ID(nvtxExtPayloadInitOnce)()
{
intptr_t* fnSlots = NVTX_EXT_PAYLOAD_VERSIONED_ID(nvtxExtPayloadSlots) + 1;
nvtxExtModuleSegment_t segment = {
0, // unused (only one segment)
NVTX3EXT_CBID_PAYLOAD_FN_NUM,
NVTX_EXT_PAYLOAD_VERSIONED_ID(nvtxExtPayloadSlots) + 1
fnSlots
};
nvtxExtModuleInfo_t module = {
projects/rccl/src/include/proxy.h
+40 -2
Просмотреть файл
@@ -10,6 +10,7 @@
#include "devcomm.h"
#include "info.h"
#include "socket.h"
#include "ipcsocket.h"
#include <pthread.h>
#include "shm.h"
@@ -161,6 +162,31 @@ struct ncclProxyProgressState {
int nextOps;
};
// Expected proxy response fifo
struct ncclExpectedProxyResponse {
void* opId;
int respSize;
bool done;
void* respBuff;
struct ncclExpectedProxyResponse* next;
};
struct ncclProxyAsyncOp {
int type;
struct ncclProxyConnection* connection;
int reqSize, respSize;
char *reqBuff, *respBuff;
void* opId;
ncclProxyAsyncOp* next;
};
struct ncclProxyLocalPeer {
struct ncclSocket sock;
int localRank;
ncclProxyAsyncOp* asyncOps;
int asyncOpCounter;
};
struct ncclProxyState {
// Service thread
pthread_t thread;
@@ -176,6 +202,9 @@ struct ncclProxyState {
// Progress thread
struct ncclProxyProgressState progressState;
// Queue of expected responses from the proxy
struct ncclExpectedProxyResponse* expectedResponses;
};
enum proxyConnectState {
@@ -220,10 +249,19 @@ enum ncclProxyMsgType {
ncclProxyMsgStart = 5,
ncclProxyMsgClose = 6,
ncclProxyMsgAbort = 7,
ncclProxyMsgStop = 8
ncclProxyMsgStop = 8,
ncclProxyMsgConvertFd = 9 // cuMem API support
};
ncclResult_t ncclProxyCall(struct ncclProxyConnector* proxyConn, int type, void* reqBuff, int reqSize, void* respBuff, int respSize);
// This function is called by a client of the proxy that needs to invoke any of the non-progress proxyOp types
// Call this function on the client, supplying a locally unique opId. Then, poll on the return value of
// ncclPollProxyResponse(), supplying the same opId to confirm the operation has completed
ncclResult_t ncclProxyCallAsync(struct ncclProxyConnector* proxyConn, int type, void* reqBuff, int reqSize, int respSize, void* opId);
// This function will internally call ncclProxyCallAsync() and spin until ncclPollProxyResponse() confirms the result is received
ncclResult_t ncclProxyCallBlocking(struct ncclProxyConnector* proxyConn, int type, void* reqBuff, int reqSize, void* respBuff, int respSize);
ncclResult_t ncclPollProxyResponse(struct ncclProxyConnector* proxyConn, void* respBuff, void* opId);
ncclResult_t ncclProxyDestroy(struct ncclComm* comm);
ncclResult_t ncclProxyShmUnlink(struct ncclComm* comm);
#endif
projects/rccl/src/include/socket.h
+1 -1
Просмотреть файл
@@ -92,6 +92,6 @@ ncclResult_t ncclSocketProgress(int op, struct ncclSocket* sock, void* ptr, int
ncclResult_t ncclSocketWait(int op, struct ncclSocket* sock, void* ptr, int size, int* offset);
ncclResult_t ncclSocketSend(struct ncclSocket* sock, void* ptr, int size);
ncclResult_t ncclSocketRecv(struct ncclSocket* sock, void* ptr, int size);
ncclResult_t ncclSocketTryRecv(struct ncclSocket* sock, void* ptr, int size, int* closed);
ncclResult_t ncclSocketTryRecv(struct ncclSocket* sock, void* ptr, int size, int* closed, bool blocking);
ncclResult_t ncclSocketClose(struct ncclSocket* sock);
#endif
projects/rccl/src/include/transport.h
+4 -1
Просмотреть файл
@@ -62,7 +62,7 @@ struct ncclTransportComm {
};
struct ncclTransport {
const char name[4];
const char name[8];
ncclResult_t (*canConnect)(int*, struct ncclTopoSystem* topo, struct ncclTopoGraph* graph, struct ncclPeerInfo*, struct ncclPeerInfo*);
struct ncclTransportComm send;
struct ncclTransportComm recv;
@@ -71,6 +71,9 @@ struct ncclTransport {
ncclResult_t ncclTransportP2pConnect(struct ncclComm* comm, int channelId, int nrecv, int* peerRecv, int nsend, int* peerSend, int connIndex);
ncclResult_t ncclTransportP2pSetup(struct ncclComm* comm, struct ncclTopoGraph* graph, int connIndex, int* highestTransportType=NULL);
ncclResult_t ncclNvlsSetup(struct ncclComm* comm);
ncclResult_t ncclNvlsFree(struct ncclComm* comm);
enum { collNetRecv=0, collNetSend=1 };
int ncclTransportCollNetSetup(struct ncclComm* comm, struct ncclTopoGraph* collNetGraph, struct ncclChannel* channel, int masterRank, int masterPeer, int collNetGraphChannelId, int type);
ncclResult_t ncclTransportCollNetCheck(struct ncclComm* comm, int collNetSetupFail);