AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

2.5.6-1 (#255)

Browse Source 
Add LL128 Protocol.

Rewrite the topology detection and tree/ring creation (#179). Improve
tree performance by sending/receiving from different GPUs. Add
model-based tuning to switch between the different algorithms and
protocols.

Rework P2P/SHM detection in containers (#155, #248).

Detect duplicated devices and return an error (#231).

Add tuning for GCP

[ROCm/rccl commit: 299c554dcc]
This commit is contained in:
Sylvain Jeaugey
2019-11-19 14:57:39 -08:00
committed by GitHub
parent 221b65bee1
commit 71560fd67b
65 changed files with 4783 additions and 2832 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/include/collectives.h
+73
View File
@@ -0,0 +1,73 @@
/*************************************************************************
* Copyright (c) 2017-2019, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NCCL_COLLECTIVES_H_
#define NCCL_COLLECTIVES_H_
#include "core.h"
#include "info.h"
#define FUNC_INDEX(coll, redop, dtype, al, pr) ((((((coll)*ncclNumOps + (redop))*ncclNumTypes) + (dtype))*NCCL_NUM_ALGORITHMS+(al))*NCCL_NUM_PROTOCOLS+(pr))
#define NCCL_COLL_NAME(coll, op, dtype) \
coll##_##op##_##dtype
#define NCCL_KERN_NAME(coll, op, dtype) \
coll##Kernel_##op##_##dtype
/* Declare all collective operations */
#define DECL_COLL5(coll, op, dtype) \
extern __device__ void NCCL_COLL_NAME(coll, op, dtype)(struct CollectiveArgs* args); \
extern __global__ void NCCL_KERN_NAME(coll, op, dtype)(struct ncclColl c); \
#define DECL_COLL4(coll, op, dtype) \
DECL_COLL5(coll, op, dtype) \
DECL_COLL5(coll##LL, op, dtype) \
DECL_COLL5(coll##LL128, op, dtype)
#define DECL_COLL3(coll, op, dtype) \
DECL_COLL4(coll##Ring, op, dtype) \
DECL_COLL4(coll##Tree, op, dtype)
#define DECL_COLL2(coll, op) \
DECL_COLL3(coll, op, i8) \
DECL_COLL3(coll, op, u8) \
DECL_COLL3(coll, op, i32) \
DECL_COLL3(coll, op, u32) \
DECL_COLL3(coll, op, i64) \
DECL_COLL3(coll, op, u64) \
DECL_COLL3(coll, op, f16) \
DECL_COLL3(coll, op, f32) \
DECL_COLL3(coll, op, f64)
#define DECL_COLL(coll) \
DECL_COLL2(coll, sum) \
DECL_COLL2(coll, prod) \
DECL_COLL2(coll, min) \
DECL_COLL2(coll, max)
#define DECL_ALL_COLLS \
DECL_COLL2(ncclBroadcast, copy) \
DECL_COLL(ncclReduce) \
DECL_COLL2(ncclAllGather, copy) \
DECL_COLL(ncclReduceScatter) \
DECL_COLL(ncclAllReduce) \
DECL_ALL_COLLS
// CHUNKSIZE must be a multiple of SLICESIZE
#define ALLREDUCE_SLICESTEPS (NCCL_STEPS/4)
#define ALLREDUCE_CHUNKSTEPS (NCCL_STEPS/2)
#define ALLGATHER_SLICESTEPS (NCCL_STEPS/4)
#define ALLGATHER_CHUNKSTEPS (NCCL_STEPS/2)
#define REDUCESCATTER_SLICESTEPS (NCCL_STEPS/4)
#define REDUCESCATTER_CHUNKSTEPS (NCCL_STEPS/2)
#define BROADCAST_SLICESTEPS 1
#define BROADCAST_CHUNKSTEPS 1
#define REDUCE_SLICESTEPS 1
#define REDUCE_CHUNKSTEPS 1
#endif