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

Merge remote-tracking branch 'remotes/nccl/master' into rccl_2.5.6

Browse Source
This commit is contained in:
Wenkai Du
2019-11-21 13:41:10 -08:00
parent b1ed4b7fa8 299c554dcc
commit 6648c81dc6
76 changed files with 5119 additions and 3071 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/include/devcomm.h
+34 -13
View File
@@ -24,8 +24,6 @@
#define NCCL_MAX_OPS 2048
#define NCCL_STEPS 8
typedef enum { ncclCollBroadcast, ncclCollReduce, ncclCollAllGather, ncclCollReduceScatter, ncclCollAllReduce, ncclCollCount } ncclColl_t;
#define DIVUP(x, y) \
(((x)+(y)-1)/(y))
#define ROUNDUP(x, y) \
@@ -49,16 +47,18 @@ union ncclLLFifoLine {
int4 i4;
};
#define MAXTHREADS 256
#define NCCL_LL_MAX_NTHREADS MAXTHREADS
#define NUM_LINES_PER_THREAD 8
#define NCCL_LL_SLICE_LINES (NUM_LINES_PER_THREAD*NCCL_LL_MAX_NTHREADS)
#define WARP_SIZE 64
#define MAXCHANNELS 32
#define NCCL_MAX_NTHREADS 256
#define NCCL_LL_MAX_NTHREADS NCCL_MAX_NTHREADS
#define NCCL_LL_LINES_PER_THREAD 8
#define NCCL_LL_SLICE_LINES (NCCL_LL_LINES_PER_THREAD*NCCL_LL_MAX_NTHREADS)
#define NCCL_LL_BUFF_LINES (NCCL_LL_SLICE_LINES*NCCL_STEPS)
#define NCCL_LL_BUFF_SIZE (NCCL_LL_BUFF_LINES*sizeof(union ncclLLFifoLine))
#ifdef DEBUG_LL
#define NCCL_LL_CLEAN_MASK 0x00000ff8
#define NCCL_LL_FLAG_MAX 0x00001000
#define NCCL_LL_FLAG(a) ((uint32_t)(a % NCCL_LL_FLAG_MAX))
#ifdef TEST_LL_CLEANUP
#define NCCL_LL_CLEAN_MASK 0x078 // Set to 0x100 to disable cleanup
#define NCCL_LL_FLAG_MAX 0x100
#define NCCL_LL_FLAG(a) ((uint32_t)((a) % NCCL_LL_FLAG_MAX))
#else
#define NCCL_LL_CLEAN_MASK 0x7ffffff8
#define NCCL_LL_FLAG(a) ((uint32_t)(a))
@@ -66,6 +66,24 @@ union ncclLLFifoLine {
// Make sure the clean mask will last for at least NCCL_NSTEPS
static_assert(NCCL_LL_CLEAN_MASK % NCCL_STEPS == 0, "Invalid NCCL_LL_CLEAN_MASK value");
#define NCCL_LL128_LINESIZE 64
#define NCCL_LL128_LINEELEMS (NCCL_LL128_LINESIZE/sizeof(uint64_t))
#define NCCL_LL128_DATAELEMS (NCCL_LL128_LINEELEMS-1)
#define NCCL_LL128_MAX_NTHREADS 256
#define NCCL_LL128_ELEMS_PER_THREAD 120
// Receiving from up to 3 sources is more compute intensive than sending
// to 3 dests. Use 70% for reduce and 30% for bcast.
#define NCCL_LL128_SPLIT(nt) (nt/2)
#define NCCL_LL128_SLICE_ELEMS (NCCL_LL128_ELEMS_PER_THREAD*NCCL_LL128_MAX_NTHREADS)
#define NCCL_LL128_BUFF_ELEMS (NCCL_LL128_SLICE_ELEMS*NCCL_STEPS)
#define NCCL_LL128_BUFF_SIZE (NCCL_LL128_BUFF_ELEMS*sizeof(uint64_t))
#define NCCL_LL128_SHMEM_ELEMS_PER_THREAD 2
#define NCCL_LL128_SHMEM_SIZE (NCCL_LL128_SHMEM_ELEMS_PER_THREAD*NCCL_LL128_MAX_NTHREADS)
struct ncclConnInfo {
// Regular comm mechanism
char *buff; // Local for recv, remote for send
@@ -90,6 +108,9 @@ struct ncclConnInfo {
// descriptions in primitives.h.
uint32_t* next_hdp_reg; // Next GPU in ring (for p2p transport use only)
uint32_t* curr_hdp_reg; // Curr GPU in ring (for rdma transport use only)
// High bandwidth, low latency protocol
uint64_t* ll128Buff; // Local for recv, remote for send
};
struct ncclConnector {
@@ -167,7 +188,8 @@ struct ncclChannel {
union {
struct {
struct ncclRing ring;
struct ncclTree tree;
struct ncclTree treeUp;
struct ncclTree treeDn;
int id;
int nthreads;
@@ -186,6 +208,7 @@ struct ncclChannel {
int collFifoTail; // Only used by CPU
uint32_t* abortCount;
uint32_t* sync;
};
int data[0x80];
};
@@ -193,8 +216,6 @@ struct ncclChannel {
static_assert(sizeof(struct ncclChannel) == 0x80*sizeof(int), "ncclChannel must have a pow2 size");
#pragma pack(pop) /* restore original alignment from stack */
#define MAXCHANNELS 16
#ifdef ENABLE_PROFILING
struct ncclProf {
union {