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

NCCL 2.26.2-1

Browse Source 
Profiler improvements
 * Add events for CUDA kernel start and end.
 * Allow network plugins to generate profiling events
 * Enable profiling on a per-operation basis, rather than per-communicator.
 * Add support for graph capturing.

Add implicit launch order
 * Allow to prevent deadlocks when using multiple NCCL communicators per
   device by implicitly ordering NCCL operations using the host program
   order. Disabled by default, set NCCL_LAUNCH_ORDER_IMPLICIT=1 to enable.
 * Add a complementary mechanism to detect host threads racing to launch
   to the same device. Enabled by default, set NCCL_LAUNCH_RACE_FATAL=0 to
   disable.

Optimize the PAT algorithm
 * Separate the computation and execution of PAT steps on different warps,
   allowing to run up to 16 PAT steps in parallel to significantly
   accelerate PAT and reduce its linear part.

Add support for setting QoS per communicator
 * Add a new trafficClass field to the communicator configuration, to
   allow the application to select a particular traffic class for a
   given communicator. The meaning of the traffic class is
   network-specific and should be set in accordance with the network
   configuration.
 * For the IB/RoCE plugin, existing config variables such as NCCL_IB_SL
   and NCCL_IB_TC take precedence.

Allow to enable GPU Direct RDMA specifically on C2C platforms
 * Disabled by default, set NCCL_NET_GDR_C2C=1 to enable.

Do not disable user buffer registration unless PXN is really used
 * Only disable UB when a communicator has more than one rank per
   node on any node.

RAS subsystem improvements
 * Report operation counts separately for each collective operation type.
 * Provide details about missing communicator ranks and reliably
   distinguish ranks that are no longer a given communicator's members
   (now reported as NOCOMM) from those that failed to respond.

Add support for timestamps to NCCL diagnostic messages
 * On by default for WARN messages; NCCL_DEBUG_TIMESTAMP_LEVELS can be
   used to enable them for other debug levels as well.
 * The format can be changed using the NCCL_DEBUG_TIMESTAMP_FORMAT config
   variable.

Reduce the memory usage with NVLink SHARP (NVLS)
 * Potentially save hundreds of MBs of device memory, considering the
   multicast buffer size granularity separately from the address alignment.

Update performance tuning for recent Intel CPUs
 * Improve algorithm/protocol selection on recent CPUs such as Emerald
   Rapids and Sapphire Rapids.

Improve channel scheduling when mixing LL and Simple operations.
 * Make LL operations account for 4x more traffic to ensure LL and simple
   operations complete at the same time.

Refactor the plugin code
 * Clean up and harmonize the support code across the network, tuner,
   and profiler plugins.

Add support for comment lines (starting with #) in the nccl.conf file
* Issue #1540.

Make user buffer registration problems print an INFO instead of a WARN.

Drop support for network plugin interface version 5.

Fix a race condition with split-shared communicators
 * NCCL could hang during connection setup if multiple communicators
   were grouped together that share resources.

Fix a performance regression when using NCCL_CROSS_NIC=1
 * NCCL would unnecessarily alternate rings, breaking the GPU-NIC
   associations.

Make GID index detection code more resilient
 * Dynamic GID detection code was giving up too soon if the
   detected index was not available (e.g., wasn't mapped to the
   container's sysfs).
 * Issues #1538, #1573.

Fix a race condition with non-blocking operation
 * Fix issue when creating a non-blocking communicator after a non-
   blocking collective operation on another communicator.

Fix shared memory usage on recent Blackwell GPUs.
 * Issues NVIDIA/nccl-tests#287, NVIDIA/nccl-tests#291, #1637.

Fix an error with NIC fusion and IB SHARP when recreating communicators
 * Disable the unloading of network plugins

Make the auto-merge failures in the NIC fusion non-fatal
 * This could happen when trying to merge IB and RoCE devices.

Fixes to ncclCommAbort
 * Fix hangs due to the progress thread spinning indefinitely on the
   network progress.
 * Reduce the abort time by up to two orders of magnitude.

Fix a crash when libnccl.so was dynamically unloaded
 * The RAS subsystem was missing a clean-up handler.

Fix a hang if the network plugin's test() call returns an error.

Fix a hang on heterogeneous architectures
 * Ensure we harmonize the tuning to avoid different tuning choices,
   causing a hang.

Fix double-free on failed ncclCommInitRank and ncclCommFinalize.

Fix a potential list traversal bug during a group launch of multiple
communicators
 * Issue #1599.

Unify the handling of NCCL configuration variables
 * Under rare circumstances, some variables specified in the config file
   could be ignored.
This commit is contained in:
Kamil Iskra
2025-03-12 13:46:21 -07:00
parent 80f6bda437
commit f44ac759fe
116 changed files with 7522 additions and 5278 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/include/bitops.h
+47 -6
View File
@@ -8,6 +8,7 @@
#define NCCL_BITOPS_H_
#include <stdint.h>
#include <string.h>
#if !__NVCC__
#ifndef __host__
@@ -276,13 +277,53 @@ inline __host__ __device__ uint32_t u32fp8Decode(uint8_t x) {
return u32fpDecode(x, 3);
}
inline __host__ __device__ uint64_t getHash(const char* string, int n) {
// Based on DJB2a, result = result * 33 ^ char
uint64_t result = 5381;
for (int c = 0; c < n; c++) {
result = ((result << 5) + result) ^ string[c];
// The hash isn't just a function of the bytes but also where the bytes are split
// into different calls to eatHash().
inline __host__ __device__ void eatHash(uint64_t acc[2], const void* bytes, size_t size) {
char const* ptr = (char const*)bytes;
acc[0] ^= size;
while (size != 0) {
// Mix the accumulator bits.
acc[0] += acc[1];
acc[1] ^= acc[0];
acc[0] ^= acc[0] >> 31;
acc[0] *= 0x9de62bbc8cef3ce3;
acc[1] ^= acc[1] >> 32;
acc[1] *= 0x485cd6311b599e79;
// Read in a chunk of input.
size_t chunkSize = size < sizeof(uint64_t) ? size : sizeof(uint64_t);
uint64_t x = 0;
memcpy(&x, ptr, chunkSize);
ptr += chunkSize;
size -= chunkSize;
// Add to accumulator.
acc[0] += x;
}
return result;
}
template<typename T>
inline __host__ __device__ void eatHash(uint64_t acc[2], const T* bytes) {
eatHash(acc, (const void*)bytes, sizeof(T));
}
inline __host__ __device__ uint64_t digestHash(uint64_t const acc[2]) {
uint64_t h = acc[0];
h ^= h >> 31;
h *= 0xbac3bd562846de6b;
h += acc[1];
h ^= h >> 32;
h *= 0x995a187a14e7b445;
return h;
}
inline __host__ __device__ uint64_t getHash(const void* bytes, size_t size) {
uint64_t acc[2] = {1, 1};
eatHash(acc, bytes, size);
return digestHash(acc);
}
template<typename T>
inline __host__ __device__ uint64_t getHash(const T* bytes) {
return getHash((const void*)bytes, sizeof(T));
}
#endif
src/include/collectives.h
+239 -211
View File
@@ -10,6 +10,7 @@
#include "nccl.h"
#include "nccl_common.h"
#include "device.h"
#define NCCL_MAX_NET_SIZE (1024*1024*1024L) // Rather than send INT_MAX which is 2G-1, send a power of two.
// CHUNKSIZE must be a multiple of SLICESIZE
@@ -382,6 +383,42 @@ public:
~RingBCAlgorithm() {}
};
#if !defined (__CUDA_ARCH__) || __CUDA_ARCH__ >= 600
#include <cuda/atomic>
#endif
// Need a power of two to ensure it divides by parallelFactor (which is also a power of two)
#define NCCL_PAT_NWORKERS 512
static constexpr int PatUsed = 0x1,
PatSkipped = 0x2;
struct ncclPatStep {
int recvDim, sendDim, recvOffset, sendOffset, stepOffset, postRecv, postSend, nelem, last, flags;
size_t inpIx, outIx;
};
struct ncclPatPeer {
uint64_t step;
struct ncclConnInfo* conn;
struct ncclConnFifo* connFifo;
void* buff;
uint64_t *headPtr;
uint64_t *tailPtr;
uint64_t stepCache;
long long int accSize;
int connStepSize;
};
#define NCCL_SHMEM_PAT_STEPS 32
struct ncclPatShmem {
struct ncclPatStep patSteps[NCCL_SHMEM_PAT_STEPS];
int parallelFactor;
long long int localAccSize;
struct ncclPatPeer sendDims[32]; // Should cover 2^32 ranks
struct ncclPatPeer recvDims[32];
};
template<typename T>
class PatRSAlgorithm{
size_t offset;
@@ -394,18 +431,17 @@ class PatRSAlgorithm{
int nrPow2;
int postFreq;
int lastA;
int parallelFactor;
int aggFactor;
int as; // aggregated steps
int a; // step inside aggregated step
int sendSkipped; // number of skipped steps during aggregation
int recvSkipped; // number of skipped steps during aggregation
int phase2recv; // receive offset for phase 2
int stepOffset;
int aggDelta;
int scale;
int phase;
__device__ __host__ int min(int a, int b) {
__device__ __host__ ssize_t min(ssize_t a, ssize_t b) {
return (a<b)?a:b;
}
@@ -433,16 +469,16 @@ class PatRSAlgorithm{
__device__ __host__ void resetA() {
a = 0;
sendSkipped = recvSkipped = 0;
sendSkipped = stepOffset = 0;
lastA = aggFactor;
if (phase >= 2) lastA /= 2*scale;
if (phase == 4) lastA = 1;
}
__device__ __host__ void reset() {
nelem = getNelem();
phase = 0;
scale = 1;
phase2recv = 0;
as = aggDelta - 1;
resetA();
}
@@ -465,8 +501,9 @@ class PatRSAlgorithm{
}
public:
__device__ __host__ PatRSAlgorithm(int stepSize, int stepDepth, size_t offset, size_t end, size_t count, int chunkCount, int rank, int nranks):
__device__ __host__ PatRSAlgorithm(int stepSize, int stepDepth, int maxParallelFactor, size_t offset, size_t end, size_t count, int chunkCount, int rank, int nranks):
offset(offset), end(end), count(count), chunkCount(chunkCount), rank(rank), nranks(nranks) {
parallelFactor = maxParallelFactor;
aggDelta = nrPow2 = (1<<log2Up(nranks));
aggFactor = 1;
@@ -476,6 +513,7 @@ public:
aggDelta /= 2;
}
postFreq = aggFactor;
if (postFreq < parallelFactor) parallelFactor = postFreq;
int d = stepDepth;
while (d > 1 && aggFactor < nranks/2) {
d /= 2;
@@ -486,160 +524,151 @@ public:
reset();
}
__device__ __host__ void getNextOp(int &recvDim, int &sendDim, size_t &inpIx, size_t &outIx, int &recvOffset, int &sendOffset, int &sendStepOffset, int &nelemOut, int &postRecv, int &postSend, int &last) {
restart:
last = 0;
nelemOut = nelem;
outIx = offset;
__device__ __host__ int getParallelFactor() {
return parallelFactor;
}
__device__ __host__ void getNextOp(struct ncclPatStep* ps) {
ps->last = 0;
ps->nelem = nelem;
ps->outIx = offset;
ps->stepOffset = stepOffset;
int skip = 0;
//printf("Phase %d as %d/%d a %d/%d scale %d\n", phase, as, aggDelta, a, lastA, scale);
if (phase == 0) {
if (a >= lastA) {
skip = 1;
} else if (phase == 0) {
int s = mirrorInvert(a, lastA)*aggDelta + as;
if (s >= nranks) skip = 1;
int sendDataRank = (rank + s) % nranks;
inpIx = sendDataRank * count + offset;
recvDim = -1;
sendDim = 0;
outIx = 0;
recvOffset = -1;
sendOffset = ((a - sendSkipped)%postFreq) * nelem;
sendStepOffset = 0;
if ((((a - sendSkipped)%postFreq) + 1 >= postFreq) || (a == lastA-1)) {
postSend = 1;
ps->inpIx = sendDataRank * count + offset;
ps->recvDim = -1;
ps->sendDim = 0;
ps->outIx = 0;
ps->recvOffset = -1;
ps->sendOffset = (a%postFreq) * nelem;
if (((a%postFreq) + 1 >= postFreq) || (a == lastA-1)) {
ps->postSend = 1;
} else {
postSend = 0;
ps->postSend = 0;
}
postRecv = 0;
if (skip) sendSkipped++;
if (++a == lastA) {
phase = as == 1 ? (aggFactor > 1 ? 2 : 4) : 1; // If as == 1, switch to phase 2
resetA();
}
if (skip == 0) return;
ps->postRecv = 0;
} else if (phase == 1) {
int s = mirrorInvert(a, lastA)*aggDelta + as;
if (s >= nranks) skip = 1;
recvDim = firstBitSet(s, nrPow2);
sendOffset = ((a - sendSkipped)%postFreq)*nelem;
recvOffset = ((a - recvSkipped)%postFreq)*nelem;
postSend = 0;
if (recvDim == 0) {
if ((((a - sendSkipped)%postFreq) + 1 >= postFreq) || (a == lastA-1)) postSend = 1;
sendStepOffset = 0;
ps->recvDim = firstBitSet(s, nrPow2);
ps->sendOffset = (a%postFreq)*nelem;
ps->recvOffset = (a%postFreq)*nelem;
ps->postSend = 0;
if (ps->recvDim == 0 && (((a%postFreq) + 1 >= postFreq) || (a == lastA-1))) ps->postSend = 1;
if (((a%postFreq) + 1 >= postFreq) || (a == lastA-1)) {
ps->postRecv = 1;
} else {
sendStepOffset = (a - sendSkipped)/postFreq;
ps->postRecv = 0;
}
if ((((a - recvSkipped)%postFreq) + 1 >= postFreq) || (a == lastA-1)) {
postRecv = 1;
} else {
postRecv = 0;
}
s -= (1<<recvDim);
s -= (1<<ps->recvDim);
int recvDataRank = (rank + nranks + s) % nranks;
inpIx = recvDataRank * count + offset;
sendDim = s ? firstBitSet(s, nrPow2) : -1;
if (sendDim == -1) {
sendOffset = -1;
sendStepOffset = 0;
} else if (as - (1<<recvDim) == 0) {
if (newPeer(a, aggFactor)) sendSkipped = a;
ps->inpIx = recvDataRank * count + offset;
ps->sendDim = s ? firstBitSet(s, nrPow2) : -1;
if (ps->sendDim == -1) {
ps->sendOffset = -1;
} else if (as - (1<<ps->recvDim) == 0) {
if (newPeer(a, aggFactor)) { sendSkipped = a; ps->stepOffset = stepOffset = 0; }
int foffset = a - sendSkipped;
sendStepOffset = recvDim == 0 ? 0 : foffset/postFreq;
sendOffset = (foffset%postFreq)*nelem;
ps->sendOffset = (foffset%postFreq)*nelem;
}
int recvDim = ps->recvDim;
if (s < nranks && skip) {
recvDim = -1;
recvOffset = -1;
postRecv = 0;
ps->recvDim = -1;
ps->recvOffset = -1;
ps->postRecv = 0;
skip = 0;
}
if (skip || recvDim == -1) recvSkipped++;
if (skip) sendSkipped++;
if (++a == lastA) {
as--;
phase = as % 2 == 1 ? 0 : 1;
resetA();
}
if (skip == 0) return;
if (recvDim > 0 && (((a-sendSkipped)%postFreq) + 1 >= postFreq) && skip == 0) stepOffset++;
} else if (phase == 2) {
int s = (2*mirrorInvert(a, lastA)+1)*scale*aggDelta + 1;
postRecv = 0;
ps->postRecv = 0;
if (s >= nranks) skip = 1;
recvDim = 0;
postSend = a == lastA-1 ? 1 : 0;
ps->recvDim = 0;
ps->postSend = a == lastA-1 ? 1 : 0;
s -= 1;
if (s < nranks && skip) {
recvDim = -1;
recvOffset = -1;
ps->recvDim = -1;
ps->recvOffset = -1;
skip = 0;
} else if (!skip) {
int foffset = phase2recv;
phase2recv++;
postRecv |= ((foffset+1)%postFreq) == 0 ? 1 : 0;
recvOffset = (foffset%postFreq) * nelem;
int foffset = a + aggFactor - aggFactor/scale;
ps->postRecv |= ((foffset+1)%postFreq) == 0 ? 1 : 0;
ps->recvOffset = (foffset%postFreq) * nelem;
}
int recvDataRank = (rank + nranks + s) % nranks;
inpIx = recvDataRank * count + offset;
sendDim = s ? firstBitSet(s, nrPow2) : -1;
int foffset = a - sendSkipped;
postSend |= ((foffset+1)%postFreq) == 0 ? 1 : 0;
sendStepOffset = 0;
sendOffset = (foffset%postFreq) * nelem;
if (skip || sendDim == -1) sendSkipped++;
if (++a == lastA) {
phase = 3;
resetA();
}
if (skip == 0) return;
ps->inpIx = recvDataRank * count + offset;
ps->sendDim = s ? firstBitSet(s, nrPow2) : -1;
int foffset = a;
ps->postSend |= ((foffset+1)%postFreq) == 0 ? 1 : 0;
ps->sendOffset = (foffset%postFreq) * nelem;
} else if (phase == 3) {
int s = (2*mirrorInvert(a, lastA)+1)*scale*aggDelta;
postRecv = a == lastA-1 ? 1 : 0;
ps->postRecv = a == lastA-1 ? 1 : 0;
if (s >= nranks) skip = 1;
recvDim = firstBitSet(s, nrPow2);
postSend = 0;
s -= (1<<recvDim);
int foffset = a - recvSkipped;
postRecv |= (foffset+1)%postFreq == 0 ? 1 : 0;
recvOffset = (foffset%postFreq) * nelem;
ps->recvDim = firstBitSet(s, nrPow2);
ps->postSend = 0;
s -= (1<<ps->recvDim);
int foffset = a;
ps->postRecv |= (foffset+1)%postFreq == 0 ? 1 : 0;
ps->recvOffset = (foffset%postFreq) * nelem;
int recvDataRank = (rank + nranks + s) % nranks;
inpIx = recvDataRank * count + offset;
sendDim = s ? firstBitSet(s, nrPow2) : -1;
ps->inpIx = recvDataRank * count + offset;
ps->sendDim = s ? firstBitSet(s, nrPow2) : -1;
if (s < nranks && skip) {
recvDim = -1;
recvOffset = -1;
postRecv = 0;
ps->recvDim = -1;
ps->recvOffset = -1;
ps->postRecv = 0;
skip = 0;
}
if (newPeer(a, aggFactor/(2*scale))) sendSkipped = a;
if (newPeer(a, aggFactor/(2*scale))) { sendSkipped = a; ps->stepOffset = stepOffset = 0; }
foffset = a - sendSkipped;
sendStepOffset = foffset / postFreq; // Accumulate on next steps
sendOffset = sendDim >= 0 ? (foffset%postFreq) * nelem : -1;
if (skip || recvDim == -1) recvSkipped++;
if (skip) sendSkipped++;
if (++a == lastA) {
scale *= 2;
phase = scale < aggFactor ? 2 : 4;
if ((foffset%postFreq) + 1 >= postFreq && skip == 0) stepOffset++;
ps->sendOffset = ps->sendDim >= 0 ? (foffset%postFreq) * nelem : -1;
} else if (phase == 4) {
ps->recvDim = 0;
ps->sendDim = -1;
ps->inpIx = rank * count + offset;
ps->recvOffset = ((aggFactor-1)%postFreq) * nelem;
ps->sendOffset = -1;
ps->postRecv = 1;
ps->postSend = 0;
offset += chunkCount;
}
a++;
if (a >= lastA && a >= parallelFactor) {
int p = phase;
if (p == 1) as--;
if (p == 3) scale *= 2;
phase =
p == 0 ? as == 1 ? (aggFactor > 1 ? 2 : 4) : 1 :
p == 1 ? as % 2 == 1 ? 0 : 1 :
p == 2 ? 3 :
p == 3 ? scale < aggFactor ? 2 : 4 :
5;
if (p == 4) {
if (offset >= end) {
ps->last = 2;
} else {
reset();
}
} else {
resetA();
}
if (skip == 0) return;
} else if (phase == 4) {
recvDim = 0;
sendDim = -1;
inpIx = rank * count + offset;
recvOffset = (phase2recv%postFreq) * nelem;
sendStepOffset = 0;
sendOffset = -1;
postRecv = 1;
postSend = 0;
offset += chunkCount;
if (offset >= end) {
last = 1;
} else {
reset();
}
return;
} else if (phase == 4 && offset >= end) {
ps->last = 1;
}
goto restart;
int flags = PatUsed | (skip ? PatSkipped : 0);
#if __CUDA_ARCH__ >= 600
cuda::atomic_ref<int, cuda::thread_scope_block> a(ps->flags);
a.store(flags, cuda::memory_order_release);
#else
ps->flags = flags;
#endif
}
};
@@ -655,14 +684,12 @@ class PatAGAlgorithm{
int nrPow2;
int postFreq;
int lastA;
int parallelFactor;
int aggFactor;
int as; // aggregated steps
int a; // step inside aggregated step
int aggDelta;
int scale;
int phase;
// AS computation
@@ -671,7 +698,7 @@ class PatAGAlgorithm{
int bitCount[32];
int bitZeroStep[32];
__device__ __host__ int min(int a, int b) {
__device__ __host__ ssize_t min(ssize_t a, ssize_t b) {
return (a<b)?a:b;
}
@@ -738,8 +765,9 @@ class PatAGAlgorithm{
public:
__device__ __host__ PatAGAlgorithm(int stepSize, int stepDepth, size_t offset, size_t end, size_t count, int chunkCount, int rank, int nranks):
__device__ __host__ PatAGAlgorithm(int stepSize, int stepDepth, int maxParallelFactor, size_t offset, size_t end, size_t count, int chunkCount, int rank, int nranks):
offset(offset), end(end), count(count), chunkCount(chunkCount), rank(rank), nranks(nranks) {
parallelFactor = maxParallelFactor;
aggDelta = nrPow2 = (1<<log2Up(nranks));
aggFactor = 1;
@@ -749,120 +777,120 @@ public:
aggDelta /= 2;
}
postFreq = aggFactor;
if (postFreq < parallelFactor) parallelFactor = postFreq;
int d = stepDepth;
while (d > 1 && aggFactor < nranks/2) {
d /= 2;
aggFactor *= 2;
aggDelta /= 2;
}
//printf("AggFactor %d PostFreq %d AggDelta %d\n", aggFactor, postFreq, aggDelta);
asDim = log2Up(aggDelta);
reset();
}
__device__ __host__ void getNextOp(int &recvDim, int &sendDim, size_t &inpIx, size_t &outIx, int &recvOffset, int &sendOffset, int &recvStepOffset, int &nelemOut, int &postRecv, int &postSend, int &last) {
restart:
//printf("Phase %d as %d/%d a %d/%d scale %d\n", phase, as, aggDelta, a, lastA, scale);
last = 0;
nelemOut = nelem;
inpIx = offset;
__device__ __host__ int getParallelFactor() {
return parallelFactor;
}
__device__ __host__ void getNextOp(struct ncclPatStep* ps) {
ps->last = 0;
ps->nelem = nelem;
ps->inpIx = offset;
int skip = 0;
if (phase == 0) {
if (a >= lastA) {
skip = 1;
} else if (phase == 0) {
int s = a*aggDelta + as;
if (s >= nranks) skip = 1;
int nextSkip = (a+1)*aggDelta + as >= nranks ? 1 : 0;
int recvDataRank = (rank + s) % nranks;
outIx = recvDataRank * count + offset;
sendDim = -1;
recvDim = 0;
inpIx = 0;
sendOffset = -1;
recvOffset = (a % postFreq) * nelem;
recvStepOffset = 0;
postRecv = (a % postFreq == postFreq-1) || ((a+1)*aggDelta+as >= nranks) ? 1 : 0;
postSend = 0;
a++;
if (nextSkip) {
as = nextAs();
if (as == aggDelta/2) {
offset += chunkCount;
if (offset >= end) {
last = 1;
} else {
reset();
}
return;
}
phase = 1;
resetA();
}
if (skip == 0) return;
ps->outIx = recvDataRank * count + offset;
ps->sendDim = -1;
ps->recvDim = 0;
ps->inpIx = 0;
ps->sendOffset = -1;
ps->recvOffset = (a % postFreq) * nelem;
ps->stepOffset = 0;
ps->postRecv = (a % postFreq == postFreq-1) || ((a+1)*aggDelta+as >= nranks) ? 1 : 0;
ps->postSend = 0;
} else if (phase == 1) {
int s = a*aggDelta + as;
if (s >= nranks) skip = 1;
sendDim = firstBitSet(s, nrPow2);
s -= (1<<sendDim);
ps->sendDim = firstBitSet(s, nrPow2);
s -= (1<<ps->sendDim);
int sendDataRank = (rank + nranks + s) % nranks;
outIx = sendDataRank * count + offset;
recvDim = s ? firstBitSet(s, nrPow2) : -1;
sendOffset = recvOffset = (a % postFreq) * nelem;
postSend = (a % postFreq == postFreq-1) || ((a+1)*aggDelta+as >= nranks) ? 1 : 0;
postRecv = (sendDim == 0) && ((a % postFreq == postFreq-1) || ((a+1)*aggDelta+as-1 >= nranks)) ? 1 : 0;
recvStepOffset = (sendDim == 0) ? 0 : a/postFreq;
if (recvDim == -1) {
recvOffset = -1;
postRecv = 0;
} else if (as - (1<<sendDim) == 0) {
int foffset = (a*aggDelta) >> (recvDim+1);
recvOffset = (foffset%postFreq)*nelem;
postRecv = (sendDim == 0) && ((foffset % postFreq == postFreq-1) || ((((foffset+1)*2)+1)<<recvDim) >= nranks) ? 1 : 0;
recvStepOffset = (sendDim == 0) ? 0 : foffset/postFreq;
ps->outIx = sendDataRank * count + offset;
ps->recvDim = s ? firstBitSet(s, nrPow2) : -1;
ps->sendOffset = ps->recvOffset = (a % postFreq) * nelem;
ps->postSend = (a % postFreq == postFreq-1) || ((a+1)*aggDelta+as >= nranks) ? 1 : 0;
ps->postRecv = (ps->sendDim == 0) && ((a % postFreq == postFreq-1) || ((a+1)*aggDelta+as-1 >= nranks)) ? 1 : 0;
ps->stepOffset = (ps->sendDim == 0) ? 0 : a/postFreq;
if (ps->recvDim == -1) {
ps->recvOffset = -1;
ps->postRecv = 0;
} else if (as - (1<<ps->sendDim) == 0) {
int foffset = (a*aggDelta) >> (ps->recvDim+1);
ps->recvOffset = (foffset%postFreq)*nelem;
ps->postRecv = (ps->sendDim == 0) && ((foffset % postFreq == postFreq-1) || ((((foffset+1)*2)+1)<<ps->recvDim) >= nranks) ? 1 : 0;
ps->stepOffset = (ps->sendDim == 0) ? 0 : foffset/postFreq;
}
if (s < nranks && sendDim == 0 && skip) {
if (s < nranks && ps->sendDim == 0 && skip) {
// Don't forget to receive at least once even if we don't send afterwards
sendDim = -1;
sendOffset = -1;
postSend = 0;
ps->sendDim = -1;
ps->sendOffset = -1;
ps->postSend = 0;
skip = 0;
}
if (++a == lastA) {
if (as % 2 == 1) {
phase = 0;
} else {
as = nextAs();
}
resetA();
}
if (skip == 0) return;
} else if (phase == 2) {
int s = (2*a+1)*scale*aggDelta;
postSend = (a % postFreq == postFreq-1) || ((2*(a+1)+1)*scale*aggDelta >= nranks) ? 1 : 0;
postRecv = 0;
ps->postSend = (a % postFreq == postFreq-1) || ((2*(a+1)+1)*scale*aggDelta >= nranks) ? 1 : 0;
ps->postRecv = 0;
if (s >= nranks) skip = 1;
sendDim = firstBitSet(s, nrPow2);
s -= (1<<sendDim);
sendOffset = (a%postFreq) * nelem;
recvStepOffset = a / postFreq;
ps->sendDim = firstBitSet(s, nrPow2);
s -= (1<<ps->sendDim);
ps->sendOffset = (a%postFreq) * nelem;
ps->stepOffset = a / postFreq;
int sendDataRank = (rank + nranks + s) % nranks;
outIx = sendDataRank * count + offset;
recvDim = s ? firstBitSet(s, nrPow2) : -1;
if (recvDim == -1) {
recvOffset = -1;
ps->outIx = sendDataRank * count + offset;
ps->recvDim = s ? firstBitSet(s, nrPow2) : -1;
if (ps->recvDim == -1) {
ps->recvOffset = -1;
} else {
s -= (1<<recvDim);
int foffset = (a*2*scale*aggDelta) >> (recvDim+1);
recvOffset = (foffset%postFreq)*nelem;
recvStepOffset = foffset / postFreq;
s -= (1<<ps->recvDim);
int foffset = (a*2*scale*aggDelta) >> (ps->recvDim+1);
ps->recvOffset = (foffset%postFreq)*nelem;
ps->stepOffset = foffset / postFreq;
}
if (++a == lastA) {
scale /= 2;
phase = scale ? 2 : 1;
}
a++;
if (a >= lastA && a >= parallelFactor) {
int p = phase;
if (p == 2) scale /= 2;
phase =
p == 2 ? scale ? 2 : 1 :
p == 1 ? as % 2 == 1 ? 0 : 1 :
1;
if (p == 0 || (p == 1 && as % 2 == 0)) as = nextAs();
if (p == 0 && as == aggDelta/2) {
offset += chunkCount;
if (offset >= end) {
ps->last = 2;
} else {
reset();
}
} else {
resetA();
}
if (skip == 0) return;
} else if (phase == 0 && as == 1 && offset + chunkCount >= end && a-1 >= ((lastA-1) / parallelFactor) * parallelFactor) {
ps->last = 1;
}
goto restart;
int flags = PatUsed | (skip ? PatSkipped : 0);
#if __CUDA_ARCH__ >= 600
cuda::atomic_ref<int, cuda::thread_scope_block> a(ps->flags);
a.store(flags, cuda::memory_order_release);
#else
ps->flags = flags;
#endif
}
};
#endif
src/include/comm.h
+8 -2
View File
@@ -131,6 +131,9 @@ struct ncclSharedResources {
int* tpRankToLocalRank;
// Internal streams
struct ncclStrongStream deviceStream, hostStream;
int noncapturedRefs; // number of non-captured hostStreamPlanCallback on the stream
int persistentRefs;
cudaEvent_t launchEvent, scratchEvent;
/* proxy related shared res */
struct ncclProxyState* proxyState;
@@ -407,6 +410,7 @@ struct ncclComm {
// List of destructors to run when comm is destructed
struct ncclDestructor* destructorHead;
struct ncclCudaContext* context;
struct ncclSharedResources* sharedRes;
/* map to top parent ranks. */
int* topParentRanks;
@@ -419,6 +423,7 @@ struct ncclComm {
int netPluginLoaded;
ncclNet_t* ncclNet;
int ncclNetVer;
ncclNetDeviceType netDeviceType;
ncclCollNet_t* ncclCollNet;
void* bootstrap;
@@ -426,6 +431,7 @@ struct ncclComm {
uint64_t* connectSend;
uint64_t* connectRecv;
struct ncclTopoGraph graphs[NCCL_NUM_ALGORITHMS];
int maxTreePattern;
bool initAlgoChannels[NCCL_NUM_ALGORITHMS];
bool runtimeConn; // if dynamic connection is supported
bool directMode;
@@ -565,8 +571,7 @@ struct ncclComm {
struct ncclComm* groupNext;
// Subset of those in groupNext list. Holds 0x1 if not needing preconnect.
struct ncclComm* preconnectNext;
int persistentRefs; // number of persistent plan-lists capturing this comm
int noncapturedRefs; // number of non-captured hostStreamPlanCallback on the stream
int localPersistentRefs; // number of persistent plan-lists capturing this comm
struct P2pSchedulePair { int sendRank; int recvRank; } *p2pSchedule;
struct ncclKernelPlanner planner;
@@ -603,6 +608,7 @@ struct ncclComm {
// Profiler plugin
void* profilerContext;
uint64_t seqNumber[NCCL_NUM_FUNCTIONS];
struct ncclProfilerProxy profiler;
// buffer registration cache
struct ncclRegCache regCache;
src/include/device.h
+7 -1
View File
@@ -133,6 +133,7 @@ struct ncclProxyConnector {
struct ncclConnector {
int connected;
int hasSeen;
struct ncclProxyConnector proxyConn;
struct ncclTransportComm* transportComm;
void* transportResources;
@@ -374,6 +375,7 @@ struct alignas(16) ncclDevChannel {
struct ncclDirect collnetDirect;
struct ncclNvls nvls;
uint32_t* workFifoDone; // Location of done counter, device writes index+1 of last work processed
uint64_t workCounter;
};
struct ncclDevComm {
@@ -396,6 +398,10 @@ struct ncclDevComm {
// Channels, device side
struct ncclDevChannel* channels/*[MAXCHANNELS]*/;
int* rankToLocalRank;
// Profiler counters
uint64_t* workStarted/*[MAXCHANNELS]*/;
uint64_t* workCompleted/*[MAXCHANNELS]*/;
};
struct alignas(16) ncclDevCommAndChannels {
@@ -468,7 +474,7 @@ __host__ __device__ constexpr int ncclCalcUnroll(int bytePerPack, int insns, int
__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;
return cudaArch >= 800 ? (cudaArch == 1200 ? 6 : 8) : 4;
}
__host__ __device__ constexpr int ncclNvlsUnrollBytes(int cudaArch = NCCL_CUDA_ARCH) { return 4*16; }
src/include/graph.h
+12 -4
View File
@@ -36,7 +36,13 @@ ncclResult_t ncclTopoComputeCommCPU(struct ncclComm* comm);
ncclResult_t ncclTopoGetNetDev(struct ncclComm* comm, int rank, struct ncclTopoGraph* graph, int channelId, int peerRank, int64_t* id, int* dev, int* proxyRank);
ncclResult_t ncclTopoCheckP2p(struct ncclComm* comm, struct ncclTopoSystem* system, int rank1, int rank2, int* p2p, int *read, int* intermediateRank);
ncclResult_t ncclTopoCheckMNNVL(struct ncclTopoSystem* system, struct ncclPeerInfo* info1, struct ncclPeerInfo* info2, int* ret);
ncclResult_t ncclTopoCheckGdr(struct ncclTopoSystem* topo, int rank, int64_t netId, int read, int* useGdr);
enum ncclTopoGdrMode {
ncclTopoGdrModeDisable = 0,
ncclTopoGdrModeDefault = 1,
ncclTopoGdrModePci = 2,
ncclTopoGdrModeNum = 3
};
ncclResult_t ncclTopoCheckGdr(struct ncclTopoSystem* topo, int rank, int64_t netId, int read, enum ncclTopoGdrMode* gdrMode);
ncclResult_t ncclTopoNeedFlush(struct ncclComm* comm, int netDev, int rank, int* flush);
ncclResult_t ncclTopoIsGdrAvail(struct ncclTopoSystem* system, int rank, bool *avail);
ncclResult_t ncclTopoCheckNet(struct ncclTopoSystem* system, int rank1, int rank2, int* net);
@@ -55,9 +61,11 @@ ncclResult_t ncclTopoGetCpuAffinity(struct ncclTopoSystem* system, int rank, cpu
#define NCCL_TOPO_CPU_VENDOR_AMD 2
#define NCCL_TOPO_CPU_VENDOR_ZHAOXIN 3
#define NCCL_TOPO_CPU_VENDOR_MIXED 4
#define NCCL_TOPO_CPU_TYPE_BDW 1
#define NCCL_TOPO_CPU_TYPE_SKL 2
#define NCCL_TOPO_CPU_TYPE_YONGFENG 1
#define NCCL_TOPO_CPU_MODEL_INTEL_BDW 1
#define NCCL_TOPO_CPU_MODEL_INTEL_SKL 2
#define NCCL_TOPO_CPU_MODEL_INTEL_SRP 3
#define NCCL_TOPO_CPU_MODEL_INTEL_ERP 4
#define NCCL_TOPO_CPU_MODEL_YONGFENG 1
ncclResult_t ncclTopoCpuType(struct ncclTopoSystem* system, int* arch, int* vendor, int* model);
ncclResult_t ncclTopoGetGpuCount(struct ncclTopoSystem* system, int* count);
ncclResult_t ncclTopoGetNetCount(struct ncclTopoSystem* system, int* count);
src/include/group.h
+6
View File
@@ -112,6 +112,12 @@ inline void ncclGroupCommJoin(struct ncclComm* comm) {
struct ncclComm** pp = &ncclGroupCommHead;
while (*pp != nullptr && comm->intraComm0 != (*pp)->intraComm0)
pp = &(*pp)->groupNext;
// didn't find its clique, we need to insert it with ascending order based on commHash
if (*pp == nullptr) {
pp = &ncclGroupCommHead;
while (*pp != nullptr && (*pp)->commHash < comm->commHash) pp = &(*pp)->groupNext;
}
comm->groupNext = *pp;
*pp = comm;
// Comms gets a new memory stack scope upon joining. Each task batched for
src/include/nccl_net.h
-604
View File
@@ -1,604 +0,0 @@
/*************************************************************************
* Copyright (c) 2017-2022, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NCCL_NET_H_
#define NCCL_NET_H_
#include "nccl.h"
#include "nccl_common.h"
#include "net_device.h"
#include <stdint.h>
#define NCCL_NET_HANDLE_MAXSIZE 128
//Maximum value NCCL can accept for maxP2pBytes and maxCollBytes net properties
#define NCCL_MAX_NET_SIZE_BYTES (1*1024*1024*1024*1024L)
#define NCCL_NET_OPTIONAL_RECV_COMPLETION 0x1
#define NCCL_PTR_HOST 0x1
#define NCCL_PTR_CUDA 0x2
#define NCCL_PTR_DMABUF 0x4
// Maximum number of requests per comm object
#define NCCL_NET_MAX_REQUESTS 32
// Max number of ncclNet objects which can live in the same process
#define NCCL_NET_MAX_PLUGINS 3
#define NCCL_NET_MAX_DEVS_PER_NIC_V9 4
#define NCCL_NET_MAX_DEVS_PER_NIC NCCL_NET_MAX_DEVS_PER_NIC_V9
typedef struct {
int ndevs;
int devs[NCCL_NET_MAX_DEVS_PER_NIC_V9];
} ncclNetVDeviceProps_v9_t;
typedef ncclNetVDeviceProps_v9_t ncclNetVDeviceProps_t;
typedef struct {
char* name; // Used mostly for logging.
char* pciPath; // Path to the PCI device in /sys.
uint64_t guid; // Unique identifier for the NIC chip. Important for
// cards with multiple PCI functions (Physical or virtual).
int ptrSupport; // [NCCL_PTR_HOST|NCCL_PTR_CUDA|NCCL_PTR_DMABUF]
int regIsGlobal; // regMr is not tied to a particular comm
int forceFlush; // Force a flush on receives
int speed; // Port speed in Mbps.
int port; // Port number.
float latency; // Network latency
int maxComms; // Maximum number of comms we can create
int maxRecvs; // Maximum number of grouped receives.
ncclNetDeviceType netDeviceType; // Network offload type
int netDeviceVersion; // Version number for network offload
ncclNetVDeviceProps_v9_t vProps;
size_t maxP2pBytes; // Max transfer size for point-to-point operations
size_t maxCollBytes; // Max transfer size for collective operations
} ncclNetProperties_v9_t;
typedef ncclNetProperties_v9_t ncclNetProperties_t;
typedef struct {
// Name of the network (mainly for logs)
const char* name;
// Initialize the network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v9_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create a connection.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Connect to a handle and return a sending comm object for that peer.
// This call must not block for the connection to be established, and instead
// should return successfully with sendComm == NULL with the expectation that
// it will be called again until sendComm != NULL.
// If *sendDevComm points to a valid object, then NCCL is requesting device offload for this connection
ncclResult_t (*connect)(int dev, void* handle, void** sendComm, ncclNetDeviceHandle_v8_t** sendDevComm);
// Finalize connection establishment after remote peer has called connect.
// This call must not block for the connection to be established, and instead
// should return successfully with recvComm == NULL with the expectation that
// it will be called again until recvComm != NULL.
// If *recvDevComm points to a valid object, then NCCL is requesting device offload for this connection
ncclResult_t (*accept)(void* listenComm, void** recvComm, ncclNetDeviceHandle_v8_t** recvDevComm);
// Register/Deregister memory. Comm can be either a sendComm or a recvComm.
// Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* comm, void* data, size_t size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* comm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* comm, void* mhandle);
// Asynchronous send to a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*isend)(void* sendComm, void* data, size_t size, int tag, void* mhandle, void** request);
// Asynchronous recv from a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*irecv)(void* recvComm, int n, void** data, size_t* sizes, int* tags, void** mhandles, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* recvComm, int n, void** data, int* sizes, void** mhandles, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* sizes);
// Close and free send/recv comm objects
ncclResult_t (*closeSend)(void* sendComm);
ncclResult_t (*closeRecv)(void* recvComm);
ncclResult_t (*closeListen)(void* listenComm);
// Copy the given mhandle to a dptr in a format usable by this plugin's device code
ncclResult_t (*getDeviceMr)(void* comm, void* mhandle, void** dptr_mhandle);
// Notify the plugin that a recv has completed by the device
ncclResult_t (*irecvConsumed)(void* recvComm, int n, void* request);
// Create a virtual NIC given the specified properties, which can be accessed at device index d
ncclResult_t (*makeVDevice)(int* d, ncclNetVDeviceProps_t* props);
} ncclNet_v9_t;
typedef ncclNet_v9_t ncclNet_t;
#define NCCL_NET_PLUGIN_SYMBOL ncclNetPlugin_v9
typedef struct {
void* mhandle;
void* address;
size_t size;
} ncclNetSGE_v9_t;
typedef struct {
// Name of the collective network (mainly for logs)
const char* name;
// Initialize the collective network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters capable of doing collective operations.
// If ndev returns 0, all other functions might be set to NULL.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v9_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create connections.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Create a group for collective operations. handles have been created
// using listen() above. rank indicates caller's rank in the collective network.
ncclResult_t (*connect)(void* handles[], int nranks, int rank, void* listenComm, void** collComm);
// Returns whether a reduction operation on a data type is supported.
// 1 for supported, 0 otherwise.
ncclResult_t (*reduceSupport)(ncclDataType_t dataType, ncclRedOp_t redOp, int* supported);
// Register/Deregister memory. Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* collComm, void* data, size_t size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* collComm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* collComm, void* mhandle);
// Performs an asynchronous allreduce operation on the collective group.
// May return request == NULL if the call cannot be performed (or would block).
ncclResult_t (*iallreduce)(void* collComm, void* sendData, void* recvData, size_t count,
ncclDataType_t dataType, ncclRedOp_t redOp, void* sendMhandle, void* recvMhandle, void** request);
ncclResult_t (*iallgather)(void* collComm, void* sendData, int nRecvParts, ncclNetSGE_v9_t* recvParts,
size_t bytesPerRank, size_t windowOffset, size_t windowBytes,
void* sendMhandle, void** request);
ncclResult_t (*ireducescatter)(void* collComm, int nSendParts, ncclNetSGE_v9_t* sendParts, void* recvData,
size_t bytesPerRank, size_t windowOffset, size_t windowBytes,
ncclDataType_t dataType, ncclRedOp_t redOp,
void* recvMhandle, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* collComm, void* data, int size, void* mhandle, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* size);
// Close and free collective comm objects
ncclResult_t (*closeColl)(void* collComm);
ncclResult_t (*closeListen)(void* listenComm);
// Create a virtual NIC given the specified properties, which can be accessed at device index d
ncclResult_t (*makeVDevice)(int* d, ncclNetVDeviceProps_t* props);
} ncclCollNet_v9_t;
typedef ncclCollNet_v9_t ncclCollNet_t;
#define NCCL_COLLNET_PLUGIN_SYMBOL ncclCollNetPlugin_v9
typedef struct {
char* name; // Used mostly for logging.
char* pciPath; // Path to the PCI device in /sys.
uint64_t guid; // Unique identifier for the NIC chip. Important for
// cards with multiple PCI functions (Physical or virtual).
int ptrSupport; // [NCCL_PTR_HOST|NCCL_PTR_CUDA|NCCL_PTR_DMABUF]
int regIsGlobal; // regMr is not tied to a particular comm
int speed; // Port speed in Mbps.
int port; // Port number.
float latency; // Network latency
int maxComms; // Maximum number of comms we can create
int maxRecvs; // Maximum number of grouped receives.
ncclNetDeviceType netDeviceType; // Network offload type
int netDeviceVersion; // Version number for network offload
} ncclNetProperties_v8_t;
typedef struct {
// Name of the network (mainly for logs)
const char* name;
// Initialize the network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v8_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create a connection.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Connect to a handle and return a sending comm object for that peer.
// This call must not block for the connection to be established, and instead
// should return successfully with sendComm == NULL with the expectation that
// it will be called again until sendComm != NULL.
// If *sendDevComm points to a valid object, then NCCL is requesting device offload for this connection
ncclResult_t (*connect)(int dev, void* handle, void** sendComm, ncclNetDeviceHandle_v8_t** sendDevComm);
// Finalize connection establishment after remote peer has called connect.
// This call must not block for the connection to be established, and instead
// should return successfully with recvComm == NULL with the expectation that
// it will be called again until recvComm != NULL.
// If *recvDevComm points to a valid object, then NCCL is requesting device offload for this connection
ncclResult_t (*accept)(void* listenComm, void** recvComm, ncclNetDeviceHandle_v8_t** recvDevComm);
// Register/Deregister memory. Comm can be either a sendComm or a recvComm.
// Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* comm, void* data, size_t size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* comm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* comm, void* mhandle);
// Asynchronous send to a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*isend)(void* sendComm, void* data, int size, int tag, void* mhandle, void** request);
// Asynchronous recv from a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*irecv)(void* recvComm, int n, void** data, int* sizes, int* tags, void** mhandles, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* recvComm, int n, void** data, int* sizes, void** mhandles, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* sizes);
// Close and free send/recv comm objects
ncclResult_t (*closeSend)(void* sendComm);
ncclResult_t (*closeRecv)(void* recvComm);
ncclResult_t (*closeListen)(void* listenComm);
// Copy the given mhandle to a dptr in a format usable by this plugin's device code
ncclResult_t (*getDeviceMr)(void* comm, void* mhandle, void** dptr_mhandle);
// Notify the plugin that a recv has completed by the device
ncclResult_t (*irecvConsumed)(void* recvComm, int n, void* request);
} ncclNet_v8_t;
typedef struct {
void* mhandle;
void* address;
uint32_t size;
} ncclNetSGE_v8_t;
typedef struct {
// Name of the collective network (mainly for logs)
const char* name;
// Initialize the collective network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters capable of doing collective operations.
// If ndev returns 0, all other functions might be set to NULL.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v8_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create connections.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Create a group for collective operations. handles have been created
// using listen() above. rank indicates caller's rank in the collective network.
ncclResult_t (*connect)(void* handles[], int nranks, int rank, void* listenComm, void** collComm);
// Returns whether a reduction operation on a data type is supported.
// 1 for supported, 0 otherwise.
ncclResult_t (*reduceSupport)(ncclDataType_t dataType, ncclRedOp_t redOp, int* supported);
// Register/Deregister memory. Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* collComm, void* data, size_t size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* collComm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* collComm, void* mhandle);
// Performs an asynchronous allreduce operation on the collective group.
// May return request == NULL if the call cannot be performed (or would block).
ncclResult_t (*iallreduce)(void* collComm, void* sendData, void* recvData, int count,
ncclDataType_t dataType, ncclRedOp_t redOp, void* sendMhandle, void* recvMhandle, void** request);
ncclResult_t (*iallgather)(void* collComm, void* sendData, int nRecvParts, ncclNetSGE_v8_t* recvParts,
size_t bytesPerRank, size_t windowOffset, size_t windowBytes,
void* sendMhandle, void** request);
ncclResult_t (*ireducescatter)(void* collComm, int nSendParts, ncclNetSGE_v8_t* sendParts, void* recvData,
size_t bytesPerRank, size_t windowOffset, size_t windowBytes,
ncclDataType_t dataType, ncclRedOp_t redOp,
void* recvMhandle, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* collComm, void* data, int size, void* mhandle, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* size);
// Close and free collective comm objects
ncclResult_t (*closeColl)(void* collComm);
ncclResult_t (*closeListen)(void* listenComm);
} ncclCollNet_v8_t;
typedef struct {
char* name; // Used mostly for logging.
char* pciPath; // Path to the PCI device in /sys.
uint64_t guid; // Unique identifier for the NIC chip. Important for
// cards with multiple PCI functions (Physical or virtual).
int ptrSupport; // [NCCL_PTR_HOST|NCCL_PTR_CUDA|NCCL_PTR_DMABUF]
int speed; // Port speed in Mbps.
int port; // Port number.
float latency; // Network latency
int maxComms; // Maximum number of comms we can create
int maxRecvs; // Maximum number of grouped receives.
ncclNetDeviceType netDeviceType; // Network offload type
int netDeviceVersion; // Version number for network offload
} ncclNetProperties_v7_t;
typedef struct {
// Name of the network (mainly for logs)
const char* name;
// Initialize the network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v7_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create a connection.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Connect to a handle and return a sending comm object for that peer.
// This call must not block for the connection to be established, and instead
// should return successfully with sendComm == NULL with the expectation that
// it will be called again until sendComm != NULL.
// If *sendDevComm points to a valid object, then NCCL is requesting device offload for this connection
ncclResult_t (*connect)(int dev, void* handle, void** sendComm, ncclNetDeviceHandle_v7_t** sendDevComm);
// Finalize connection establishment after remote peer has called connect.
// This call must not block for the connection to be established, and instead
// should return successfully with recvComm == NULL with the expectation that
// it will be called again until recvComm != NULL.
// If *recvDevComm points to a valid object, then NCCL is requesting device offload for this connection
ncclResult_t (*accept)(void* listenComm, void** recvComm, ncclNetDeviceHandle_v7_t** recvDevComm);
// Register/Deregister memory. Comm can be either a sendComm or a recvComm.
// Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* comm, void* data, int size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* comm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* comm, void* mhandle);
// Asynchronous send to a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*isend)(void* sendComm, void* data, int size, int tag, void* mhandle, void** request);
// Asynchronous recv from a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*irecv)(void* recvComm, int n, void** data, int* sizes, int* tags, void** mhandles, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* recvComm, int n, void** data, int* sizes, void** mhandles, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* sizes);
// Close and free send/recv comm objects
ncclResult_t (*closeSend)(void* sendComm);
ncclResult_t (*closeRecv)(void* recvComm);
ncclResult_t (*closeListen)(void* listenComm);
// Copy the given mhandle to a dptr in a format usable by this plugin's device code
ncclResult_t (*getDeviceMr)(void* comm, void* mhandle, void** dptr_mhandle);
// Notify the plugin that a recv has completed by the device
ncclResult_t (*irecvConsumed)(void* recvComm, int n, void* request);
} ncclNet_v7_t;
typedef struct {
// Name of the collective network (mainly for logs)
const char* name;
// Initialize the collective network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters capable of doing collective operations.
// If ndev returns 0, all other functions might be set to NULL.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v7_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create connections.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Create a group for collective operations. handles have been created
// using listen() above. rank indicates caller's rank in the collective network.
ncclResult_t (*connect)(void* handles[], int nranks, int rank, void* listenComm, void** collComm);
// Returns whether a reduction operation on a data type is supported.
// 1 for supported, 0 otherwise.
ncclResult_t (*reduceSupport)(ncclDataType_t dataType, ncclRedOp_t redOp, int* supported);
// Register/Deregister memory. Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* collComm, void* data, int size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* collComm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* collComm, void* mhandle);
// Performs an asynchronous allreduce operation on the collective group.
// May return request == NULL if the call cannot be performed (or would block).
ncclResult_t (*iallreduce)(void* collComm, void* sendData, void* recvData, int count,
ncclDataType_t dataType, ncclRedOp_t redOp, void* sendMhandle, void* recvMhandle, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* collComm, void* data, int size, void* mhandle, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* size);
// Close and free collective comm objects
ncclResult_t (*closeColl)(void* collComm);
ncclResult_t (*closeListen)(void* listenComm);
} ncclCollNet_v7_t;
#define NCCL_NET_MAX_REQUESTS_V6 8
// v6 struct for backwards compatibility
typedef struct {
char* name; // Used mostly for logging.
char* pciPath; // Path to the PCI device in /sys.
uint64_t guid; // Unique identifier for the NIC chip. Important for
// cards with multiple PCI functions (Physical or virtual).
int ptrSupport; // [NCCL_PTR_HOST|NCCL_PTR_CUDA|NCCL_PTR_DMABUF]
int speed; // Port speed in Mbps.
int port; // Port number.
float latency; // Network latency
int maxComms; // Maximum number of comms we can create
int maxRecvs; // Maximum number of grouped receives.
} ncclNetProperties_v6_t;
typedef struct {
// Name of the network (mainly for logs)
const char* name;
// Initialize the network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v6_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create a connection.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Connect to a handle and return a sending comm object for that peer.
// This call must not block for the connection to be established, and instead
// should return successfully with sendComm == NULL with the expectation that
// it will be called again until sendComm != NULL.
ncclResult_t (*connect)(int dev, void* handle, void** sendComm);
// Finalize connection establishment after remote peer has called connect.
// This call must not block for the connection to be established, and instead
// should return successfully with recvComm == NULL with the expectation that
// it will be called again until recvComm != NULL.
ncclResult_t (*accept)(void* listenComm, void** recvComm);
// Register/Deregister memory. Comm can be either a sendComm or a recvComm.
// Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* comm, void* data, int size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* comm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* comm, void* mhandle);
// Asynchronous send to a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*isend)(void* sendComm, void* data, int size, int tag, void* mhandle, void** request);
// Asynchronous recv from a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*irecv)(void* recvComm, int n, void** data, int* sizes, int* tags, void** mhandles, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* recvComm, int n, void** data, int* sizes, void** mhandles, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* sizes);
// Close and free send/recv comm objects
ncclResult_t (*closeSend)(void* sendComm);
ncclResult_t (*closeRecv)(void* recvComm);
ncclResult_t (*closeListen)(void* listenComm);
} ncclNet_v6_t;
typedef struct {
// Name of the collective network (mainly for logs)
const char* name;
// Initialize the collective network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters capable of doing collective operations.
// If ndev returns 0, all other functions might be set to NULL.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v6_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create connections.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Create a group for collective operations. handles have been created
// using listen() above. rank indicates caller's rank in the collective network.
ncclResult_t (*connect)(void* handles[], int nranks, int rank, void* listenComm, void** collComm);
// Returns whether a reduction operation on a data type is supported.
// 1 for supported, 0 otherwise.
ncclResult_t (*reduceSupport)(ncclDataType_t dataType, ncclRedOp_t redOp, int* supported);
// Register/Deregister memory. Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* collComm, void* data, int size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* collComm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* collComm, void* mhandle);
// Performs an asynchronous allreduce operation on the collective group.
// May return request == NULL if the call cannot be performed (or would block).
ncclResult_t (*iallreduce)(void* collComm, void* sendData, void* recvData, int count,
ncclDataType_t dataType, ncclRedOp_t redOp, void* sendMhandle, void* recvMhandle, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* collComm, void* data, int size, void* mhandle, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* size);
// Close and free collective comm objects
ncclResult_t (*closeColl)(void* collComm);
ncclResult_t (*closeListen)(void* listenComm);
} ncclCollNet_v6_t;
// v5 struct for backwards compatibility
typedef struct {
// Name of the network (mainly for logs)
const char* name;
// Initialize the network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v6_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create a connection.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Connect to a handle and return a sending comm object for that peer.
// This call must not block for the connection to be established, and instead
// should return successfully with sendComm == NULL with the expectation that
// it will be called again until sendComm != NULL.
ncclResult_t (*connect)(int dev, void* handle, void** sendComm);
// Finalize connection establishment after remote peer has called connect.
// This call must not block for the connection to be established, and instead
// should return successfully with recvComm == NULL with the expectation that
// it will be called again until recvComm != NULL.
ncclResult_t (*accept)(void* listenComm, void** recvComm);
// Register/Deregister memory. Comm can be either a sendComm or a recvComm.
// Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* comm, void* data, int size, int type, void** mhandle);
ncclResult_t (*deregMr)(void* comm, void* mhandle);
// Asynchronous send to a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*isend)(void* sendComm, void* data, int size, int tag, void* mhandle, void** request);
// Asynchronous recv from a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*irecv)(void* recvComm, int n, void** data, int* sizes, int* tags, void** mhandles, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* recvComm, int n, void** data, int* sizes, void** mhandles, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* sizes);
// Close and free send/recv comm objects
ncclResult_t (*closeSend)(void* sendComm);
ncclResult_t (*closeRecv)(void* recvComm);
ncclResult_t (*closeListen)(void* listenComm);
} ncclNet_v5_t;
// v5 struct for backwards compatibility
typedef struct {
// Name of the collective network (mainly for logs)
const char* name;
// Initialize the collective network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters capable of doing collective operations.
// If ndev returns 0, all other functions might be set to NULL.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v6_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create connections.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Create a group for collective operations. handles have been created
// using listen() above. rank indicates caller's rank in the collective network.
ncclResult_t (*connect)(void* handles[], int nranks, int rank, void* listenComm, void** collComm);
// Returns whether a reduction operation on a data type is supported.
// 1 for supported, 0 otherwise.
ncclResult_t (*reduceSupport)(ncclDataType_t dataType, ncclRedOp_t redOp, int* supported);
// Register/Deregister memory. Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* collComm, void* data, int size, int type, void** mhandle);
ncclResult_t (*deregMr)(void* collComm, void* mhandle);
// Performs an asynchronous allreduce operation on the collective group.
// May return request == NULL if the call cannot be performed (or would block).
ncclResult_t (*iallreduce)(void* collComm, void* sendData, void* recvData, int count,
ncclDataType_t dataType, ncclRedOp_t redOp, void* sendMhandle, void* recvMhandle, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* collComm, void* data, int size, void* mhandle, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* size);
// Close and free collective comm objects
ncclResult_t (*closeColl)(void* collComm);
ncclResult_t (*closeListen)(void* listenComm);
} ncclCollNet_v5_t;
#endif // end include guard
src/include/nccl_profiler.h
-235
View File
@@ -1,235 +0,0 @@
/*************************************************************************
* Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NCCL_PROFILER_H_
#define NCCL_PROFILER_H_
#include <cstdint>
enum {
ncclProfileGroup = (1 << 0), // group event type
ncclProfileColl = (1 << 1), // host collective call event type
ncclProfileP2p = (1 << 2), // host point-to-point call event type
ncclProfileProxyOp = (1 << 3), // proxy operation event type
ncclProfileProxyStep = (1 << 4), // proxy step event type
ncclProfileProxyCtrl = (1 << 5), // proxy control event type
};
typedef struct {
uint8_t type; // event type descriptor: ncclProfileColl, ...
void* parentObj; // pointer to the profiler parent object (for coll is the group)
int rank; // originating rank
union {
struct {
const char* name;
uint64_t commHash;
uint64_t seqNumber;
const char* func;
void const* sendBuff;
void* recvBuff;
size_t count;
int root;
const char* datatype;
size_t trafficBytes;
uint8_t nMaxChannels;
uint8_t nWarps;
const char* algo;
const char* proto;
} coll;
struct {
const char* name;
uint64_t commHash;
const char* func;
void* buff;
const char* datatype;
size_t count;
int peer;
} p2p;
struct {
pid_t pid; // pid of the originating process
uint8_t channelId; // channel id for this proxy operation
int peer; // remote rank for send/recv
int nSteps; // number of steps for this proxy operation
int chunkSize; // amount of data transferred by this proxy operation
int isSend;
} proxyOp;
struct {
int step;
} proxyStep;
};
} ncclProfilerEventDescr_v2_t;
typedef enum {
ncclProfilerProxyOpSendPosted,
ncclProfilerProxyOpSendRemFifoWait,
ncclProfilerProxyOpSendTransmitted,
ncclProfilerProxyOpSendDone,
ncclProfilerProxyOpRecvPosted,
ncclProfilerProxyOpRecvReceived,
ncclProfilerProxyOpRecvTransmitted,
ncclProfilerProxyOpRecvDone,
/* Legacy proxy profiler states */
ncclProfilerProxyStepSendGPUWait,
ncclProfilerProxyStepSendWait,
ncclProfilerProxyStepRecvWait,
ncclProfilerProxyStepRecvFlushWait,
ncclProfilerProxyStepRecvGPUWait,
/* Legacy proxy control states */
ncclProfilerProxyCtrlIdle,
ncclProfilerProxyCtrlActive,
ncclProfilerProxyCtrlSleep,
ncclProfilerProxyCtrlWakeup,
ncclProfilerProxyCtrlAppend,
ncclProfilerProxyCtrlAppendEnd,
} ncclProfilerEventState_v2_t;
typedef union {
struct {
size_t transSize;
int steps;
} proxyOp;
struct {
int appendedProxyOps;
} proxyCtrl;
} ncclProfilerEventStateArgs_v2_t;
typedef struct {
const char* name;
// init - initialize the profiler plugin
// Input
// - context : opaque profiler context object for separating profiler behavior across comms
// Output
// - eActivationMask: bitmask of active events set by the plugin
ncclResult_t (*init)(void** context, int* eActivationMask);
// startEvent - initialize and start a new event for the supplied event descriptor inside the eventset
// Input
// - context: opaque profiler context object
// - eDescr : pointer to ncclProfilerEventDescr_t object
// Output
// - eHandle: return event handle for supplied event descriptor object
ncclResult_t (*startEvent)(void* context, void** eHandle, ncclProfilerEventDescr_v2_t* eDescr);
// stopEvent - stop/finalize an event inside and event set
// Input
// - eHandle: handle to event object
ncclResult_t (*stopEvent)(void* eHandle);
// recordEventState - record event state transitions and event attribute updates
// Input
// - eHandle : handle to event object created through startEvent
// - eStateArgs: optional argument used to capture event attribute updates associated with the state transition
// - eState : event state transition
ncclResult_t (*recordEventState)(void* eHandle, ncclProfilerEventState_v2_t eState, ncclProfilerEventStateArgs_v2_t* eStateArgs);
// finalize - finalize the profiler plugin
// Input
// - context: opaque profiler context object
ncclResult_t (*finalize)(void* context);
} ncclProfiler_v2_t;
typedef ncclProfilerEventDescr_v2_t ncclProfilerEventDescr_t;
typedef ncclProfilerEventState_v2_t ncclProfilerEventState_t;
typedef ncclProfilerEventStateArgs_v2_t ncclProfilerEventStateArgs_t;
typedef ncclProfiler_v2_t ncclProfiler_t;
typedef struct {
uint8_t type; // event type descriptor: ncclProfileColl, ...
void* parentObj; // pointer to the profiler parent object (for coll is the group)
int rank; // originating rank
union {
struct {
const char* name;
uint64_t commHash;
uint64_t seqNumber;
uint8_t func;
void const* sendBuff;
void* recvBuff;
size_t count;
int root;
uint8_t datatype;
uint32_t op;
size_t trafficBytes;
uint8_t nMaxChannels;
uint8_t nWarps;
uint8_t algo;
uint8_t proto;
int isCollnet;
int isNvls;
} coll;
struct {
const char* name;
uint64_t commHash;
uint8_t func;
void* buff;
uint8_t datatype;
size_t count;
int peer;
} p2p;
struct {
pid_t pid; // pid of the originating process
uint8_t channelId; // channel id for this proxy operation
int peer; // remote rank for send/recv
int nSteps; // number of steps for this proxy operation
int chunkSize; // amount of data transferred by this proxy operation
int isSend;
} proxyOp;
struct {
int step;
} proxyStep;
};
} ncclProfilerEventDescr_v1_t;
typedef ncclProfilerEventState_v2_t ncclProfilerEventState_v1_t;
typedef ncclProfilerEventStateArgs_v2_t ncclProfilerEventStateArgs_v1_t;
typedef struct {
const char* name;
// init - initialize the profiler plugin
// Input
// - context : opaque profiler context object for separating profiler behavior across comms
// Output
// - eActivationMask: bitmask of active events set by the plugin
ncclResult_t (*init)(void** context, int* eActivationMask);
// startEvent - initialize and start a new event for the supplied event descriptor inside the eventset
// Input
// - context: opaque profiler context object
// - eDescr : pointer to ncclProfilerEventDescr_t object
// Output
// - eHandle: return event handle for supplied event descriptor object
ncclResult_t (*startEvent)(void* context, void** eHandle, ncclProfilerEventDescr_v1_t* eDescr);
// stopEvent - stop/finalize an event inside and event set
// Input
// - eHandle: handle to event object
ncclResult_t (*stopEvent)(void* eHandle);
// recordEventState - record event state transitions and event attribute updates
// Input
// - eHandle : handle to event object created through startEvent
// - eStateArgs: optional argument used to capture event attribute updates associated with the state transition
// - eState : event state transition
ncclResult_t (*recordEventState)(void* eHandle, ncclProfilerEventState_v1_t eState, ncclProfilerEventStateArgs_v1_t* eStateArgs);
// finalize - finalize the profiler plugin
// Input
// - context: opaque profiler context object
ncclResult_t (*finalize)(void* context);
} ncclProfiler_v1_t;
#endif
src/include/nccl_tuner.h
-149
View File
@@ -1,149 +0,0 @@
/*************************************************************************
* Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2023, Meta Platforms, Inc. and affiliates.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NCCL_TUNER_H_
#define NCCL_TUNER_H_
#include "nccl.h"
#include "nccl_common.h"
// API to be implemented by external tuner
typedef struct {
// Name of the tuner
const char* name;
// Initializes tuner states.
// Inputs:
// - nRanks: number of ranks in current communicator. Each communicator initialize its own tuner.
// - nNodes: number of nodes in current communicator.
// - logFunction: a logFunction can be useful to integrate logging together with NCCL core.
// Outputs:
// - context: tuner context object
ncclResult_t (*init)(size_t nRanks, size_t nNodes, ncclDebugLogger_t logFunction, void **context);
// Gets info (algo, protocol, number of ctas and threads) for a given collective.
// Inputs:
// - context: tuner context object
// - collType: collective type , e.g., allreduce, allgather…
// - nBytes: collective size in bytes
// - numPipeOps: number of operations in the group
// - numAlgo: number of algorithms in collCostTable
// - numProto: number of protocols in collCostTable
// - regBuff: can register user buffer
//
// Outputs:
// - nChannels: number of channels (hence SMs) to be used.
//
// InOut:
// - collCostTable: collective cost table, generated by NCCL core, containing algo|proto|time entries for collType.
// NCCL core sets ignored algo/proto cost table entries to -1.0 (NCCL_ALGO_PROTO_IGNORE).
//
// If getCollInfo() does not return ncclSuccess, NCCL will fall back to the
// default tuning for the given collective.
// Also, the plugin is allowed to not set any output, or set only the
// algorithm and protocol, but not only the algorithm or only the protocol.
// Unset fields will be set automatically by NCCL.
ncclResult_t (*getCollInfo)(void* context, ncclFunc_t collType, size_t nBytes,
int numPipeOps, float** collCostTable, int numAlgo, int numProto,
int regBuff, int* nChannels);
// Terminates the plugin and cleans up any resources that the plugin allocated.
// context: tuner context object
ncclResult_t (*destroy)(void* context);
} ncclTuner_v4_t;
typedef ncclTuner_v4_t ncclTuner_t;
#define NCCL_TUNER_PLUGIN_SYMBOL "ncclTunerPlugin_v4"
// API to be implemented by external tuner
typedef struct {
// Name of the tuner
const char* name;
// Initializes tuner states.
// Inputs:
// - nRanks: number of ranks in current communicator. Each communicator initialize its own tuner.
// - nNodes: number of nodes in current communicator.
// - logFunction: a logFunction can be useful to integrate logging together with NCCL core.
// Outputs:
// - context: tuner context object
ncclResult_t (*init)(size_t nRanks, size_t nNodes, ncclDebugLogger_t logFunction, void **context);
// Gets info (algo, protocol, number of ctas and threads) for a given collective.
// Inputs:
// - context: tuner context object
// - collType: collective type , e.g., allreduce, allgather…
// - nBytes: collective size in bytes
// - numPipeOps: number of operations in the group
// - numAlgo: number of algorithms in collCostTable
// - numProto: number of protocols in collCostTable
//
// Outputs:
// - nChannels: number of channels (hence SMs) to be used.
//
// InOut:
// - collCostTable: collective cost table, generated by NCCL core, containing algo|proto|time entries for collType.
// NCCL core sets ignored algo/proto cost table entries to -1.0 (NCCL_ALGO_PROTO_IGNORE).
//
// If getCollInfo() does not return ncclSuccess, NCCL will fall back to the
// default tuning for the given collective.
// Also, the plugin is allowed to not set any output, or set only the
// algorithm and protocol, but not only the algorithm or only the protocol.
// Unset fields will be set automatically by NCCL.
ncclResult_t (*getCollInfo)(void* context, ncclFunc_t collType, size_t nBytes,
int numPipeOps, float** collCostTable, int numAlgo, int numProto,
int* nChannels);
// Terminates the plugin and cleans up any resources that the plugin allocated.
// context: tuner context object
ncclResult_t (*destroy)(void* context);
} ncclTuner_v3_t;
// API to be implemented by external tuner
typedef struct {
// Name of the tuner
const char* name;
// Initializes tuner states.
// Inputs:
// - nRanks: number of ranks in current communicator. Each communicator initialize its own tuner.
// - nNodes: number of nodes in current communicator.
// - logFunction: a logFunction can be useful to integrate logging together with NCCL core.
// Outputs:
// - context: tuner context object
ncclResult_t (*init)(size_t nRanks, size_t nNodes, ncclDebugLogger_t logFunction, void **context);
// Gets info (algo, protocol, number of ctas and threads) for a given collective.
// Inputs:
// - context: tuner context object
// - collType: collective type , e.g., allreduce, allgather…
// - nBytes: collective size in bytes
// - collNetTypeSupport: whether collnet supports this type
// - nvlsTypeSupport: whether nvlink sharp supports this time
// - numPipeOps: number of operations in the group
//
// Outputs:
// - algorithm: selected algorithm to be used for the given collective
// - protocol: selected protocol to be used for the give collective
// - nChannels: number of channels (hence SMs) to be used.
//
// If getCollInfo() does not return ncclSuccess, NCCL will fall back to the
// default tuning for the given collective.
// Also, the plugin is allowed to not set any output, or set only the
// algorithm and protocol, but not only the algorithm or only the protocol.
// Unset fields will be set automatically by NCCL.
ncclResult_t (*getCollInfo)(void* context, ncclFunc_t collType, size_t nBytes,
int collNetSupport, int nvlsSupport, int numPipeOps,
int* algorithm, int* protocol, int* nChannels);
// Terminates the plugin and cleans up any resources that the plugin allocated.
// context: tuner context object
ncclResult_t (*destroy)(void* context);
} ncclTuner_v2_t;
#endif
src/include/net.h
-1
View File
@@ -18,7 +18,6 @@ ncclResult_t ncclNetPluginLoad(struct ncclComm* comm);
ncclResult_t ncclNetPluginUnload(struct ncclComm* comm);
ncclResult_t ncclNetInit(struct ncclComm* comm);
ncclResult_t ncclNetFinalize(struct ncclComm* comm);
int ncclNetVersion(struct ncclComm* comm);
// Test whether the current GPU support GPU Direct RDMA.
ncclResult_t ncclGpuGdrSupport(struct ncclComm* comm, int* gdrSupport);
src/include/net_device.h
+2 -1
View File
@@ -26,6 +26,7 @@ typedef struct {
typedef ncclNetDeviceHandle_v7_t ncclNetDeviceHandle_v8_t;
typedef ncclNetDeviceHandle_v8_t ncclNetDeviceHandle_v9_t;
typedef ncclNetDeviceHandle_v9_t ncclNetDeviceHandle_t;
typedef ncclNetDeviceHandle_v9_t ncclNetDeviceHandle_v10_t;
typedef ncclNetDeviceHandle_v10_t ncclNetDeviceHandle_t;
#endif
src/include/nvtx.h
+2 -1
View File
@@ -31,9 +31,10 @@
#define NVTX_SID_CommInitRankScalable 12 // same schema as NVTX_SID_CommInitRank
#define NVTX_SID_CommSplit 13
#define NVTX_SID_CommFinalize 14
// When adding new schema IDs, DO NOT re-use/overlap with the enum schema ID below!
// Define static schema ID for the reduction operation.
#define NVTX_PAYLOAD_ENTRY_NCCL_REDOP 14 + NVTX_PAYLOAD_ENTRY_TYPE_SCHEMA_ID_STATIC_START
#define NVTX_PAYLOAD_ENTRY_NCCL_REDOP 15 + NVTX_PAYLOAD_ENTRY_TYPE_SCHEMA_ID_STATIC_START
extern const nvtxDomainHandle_t ncclNvtxDomainHandle;
src/include/plugin/nccl_net.h
+54
View File
@@ -0,0 +1,54 @@
/*************************************************************************
* Copyright (c) 2017-2022, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NCCL_NET_H_
#define NCCL_NET_H_
#include "nccl.h"
#include "nccl_common.h"
#include "net_device.h"
#include <stdint.h>
#define NCCL_NET_HANDLE_MAXSIZE 128
//Maximum value NCCL can accept for maxP2pBytes and maxCollBytes net properties
#define NCCL_MAX_NET_SIZE_BYTES (1*1024*1024*1024*1024L)
#define NCCL_NET_OPTIONAL_RECV_COMPLETION 0x1
#define MAX_NET_SIZE (1024*1024*1024L) // Rather than send INT_MAX which is 2G-1, send a power of two.
#define MAX_COLLNET_SIZE (512*1024*1024L) //Set for initial collent plugins when size was not dynamically queried
#define NCCL_PTR_HOST 0x1
#define NCCL_PTR_CUDA 0x2
#define NCCL_PTR_DMABUF 0x4
// Maximum number of requests per comm object
#define NCCL_NET_MAX_REQUESTS 32
// Max number of ncclNet objects which can live in the same process
#define NCCL_NET_MAX_PLUGINS 3
// NCCL core profiler callback for network defined events instrumentation
typedef ncclResult_t (*ncclProfilerCallback_t)(void** eHandle, int type, void* pHandle, int64_t pluginId, void* extData);
#include "net/net_v10.h"
#include "net/net_v9.h"
#include "net/net_v8.h"
#include "net/net_v7.h"
#include "net/net_v6.h"
typedef ncclNet_v10_t ncclNet_t;
typedef ncclCollNet_v10_t ncclCollNet_t;
typedef ncclNetSGE_v10_t ncclNetSGE_t;
typedef ncclNetProperties_v10_t ncclNetProperties_t;
typedef ncclNetVDeviceProps_v10_t ncclNetVDeviceProps_t;
typedef ncclNetCommConfig_v10_t ncclNetCommConfig_t;
#define NCCL_NET_MAX_DEVS_PER_NIC NCCL_NET_MAX_DEVS_PER_NIC_V10
#define NCCL_NET_PLUGIN_SYMBOL ncclNetPlugin_v10
#define NCCL_COLLNET_PLUGIN_SYMBOL ncclCollNetPlugin_v10
#endif // end include guard
src/include/plugin/nccl_profiler.h
+69
View File
@@ -0,0 +1,69 @@
/*************************************************************************
* Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NCCL_PROFILER_H_
#define NCCL_PROFILER_H_
enum {
ncclProfileGroup = (1 << 0), // group event type
ncclProfileColl = (1 << 1), // host collective call event type
ncclProfileP2p = (1 << 2), // host point-to-point call event type
ncclProfileProxyOp = (1 << 3), // proxy operation event type
ncclProfileProxyStep = (1 << 4), // proxy step event type
ncclProfileProxyCtrl = (1 << 5), // proxy control event type
ncclProfileKernelCh = (1 << 6), // kernel channel event type
ncclProfileNetPlugin = (1 << 7), // network plugin-defined, events
};
typedef enum {
ncclProfilerProxyOpSendPosted,
ncclProfilerProxyOpSendRemFifoWait,
ncclProfilerProxyOpSendTransmitted,
ncclProfilerProxyOpSendDone,
ncclProfilerProxyOpRecvPosted,
ncclProfilerProxyOpRecvReceived,
ncclProfilerProxyOpRecvTransmitted,
ncclProfilerProxyOpRecvDone,
/* Legacy proxy profiler states */
ncclProfilerProxyStepSendGPUWait,
ncclProfilerProxyStepSendWait,
ncclProfilerProxyStepRecvWait,
ncclProfilerProxyStepRecvFlushWait,
ncclProfilerProxyStepRecvGPUWait,
/* Legacy proxy control states */
ncclProfilerProxyCtrlIdle,
ncclProfilerProxyCtrlActive,
ncclProfilerProxyCtrlSleep,
ncclProfilerProxyCtrlWakeup,
ncclProfilerProxyCtrlAppend,
ncclProfilerProxyCtrlAppendEnd,
} ncclProfilerEventState_t;
typedef ncclProfilerEventState_t ncclProfilerEventState_v1_t;
typedef ncclProfilerEventState_t ncclProfilerEventState_v2_t;
typedef ncclProfilerEventState_t ncclProfilerEventState_v3_t;
#include <cstdint>
#include "profiler/profiler_v3.h"
#include "profiler/profiler_v2.h"
#include "profiler/profiler_v1.h"
typedef ncclProfiler_v3_t ncclProfiler_t;
typedef ncclProfilerEventDescr_v3_t ncclProfilerEventDescr_t;
typedef ncclProfilerEventStateArgs_v3_t ncclProfilerEventStateArgs_t;
#define NCCL_PROFILER_NET_VER_BITS (16)
#define NCCL_PROFILER_NET_VER_MASK (~0U >> NCCL_PROFILER_NET_VER_BITS)
#define NCCL_PROFILER_NET_TYPE_MASK (~0U << NCCL_PROFILER_NET_VER_BITS)
typedef enum {
NCCL_PROFILER_NET_TYPE_IB = (1U << NCCL_PROFILER_NET_VER_BITS),
NCCL_PROFILER_NET_TYPE_SOCK = (2U << NCCL_PROFILER_NET_VER_BITS),
} ncclProfilerNetType;
#endif
src/include/plugin/nccl_tuner.h
+22
View File
@@ -0,0 +1,22 @@
/*************************************************************************
* Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2023, Meta Platforms, Inc. and affiliates.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NCCL_TUNER_H_
#define NCCL_TUNER_H_
#include "nccl.h"
#include "nccl_common.h"
#include "tuner/tuner_v4.h"
#include "tuner/tuner_v3.h"
#include "tuner/tuner_v2.h"
typedef ncclTuner_v4_t ncclTuner_t;
#define NCCL_TUNER_PLUGIN_SYMBOL "ncclTunerPlugin_v4"
#endif
src/include/plugin/net/net_v10.h
+158
View File
@@ -0,0 +1,158 @@
/*
* Copyright (c) 2017-2022, NVIDIA CORPORATION. All rights reserved.
*/
#ifndef NET_V10_H_
#define NET_V10_H_
#define NCCL_NET_MAX_DEVS_PER_NIC_V10 4
typedef struct {
int ndevs;
int devs[NCCL_NET_MAX_DEVS_PER_NIC_V10];
} ncclNetVDeviceProps_v10_t;
#define NCCL_NET_TRAFFIC_CLASS_UNDEF -1
typedef struct {
// Plugin-specific TC value
int trafficClass;
} ncclNetCommConfig_v10_t;
typedef struct {
char* name; // Used mostly for logging.
char* pciPath; // Path to the PCI device in /sys.
uint64_t guid; // Unique identifier for the NIC chip. Important for
// cards with multiple PCI functions (Physical or virtual).
int ptrSupport; // [NCCL_PTR_HOST|NCCL_PTR_CUDA|NCCL_PTR_DMABUF]
int regIsGlobal; // regMr is not tied to a particular comm
int forceFlush; // Force a flush on receives
int speed; // Port speed in Mbps.
int port; // Port number.
float latency; // Network latency
int maxComms; // Maximum number of comms we can create
int maxRecvs; // Maximum number of grouped receives.
ncclNetDeviceType netDeviceType; // Network offload type
int netDeviceVersion; // Version number for network offload
ncclNetVDeviceProps_v10_t vProps;
size_t maxP2pBytes; // Max transfer size for point-to-point operations
size_t maxCollBytes; // Max transfer size for collective operations
} ncclNetProperties_v10_t;
typedef struct {
// Name of the network (mainly for logs)
const char* name;
// Initialize the network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction, ncclProfilerCallback_t profFunction);
// Return the number of adapters.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v10_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create a connection.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Connect to a handle and return a sending comm object for that peer.
// This call must not block for the connection to be established, and instead
// should return successfully with sendComm == NULL with the expectation that
// it will be called again until sendComm != NULL.
// If *sendDevComm points to a valid object, then NCCL is requesting device offload for this connection
ncclResult_t (*connect)(int dev, ncclNetCommConfig_v10_t* config, void* handle, void** sendComm, ncclNetDeviceHandle_v10_t** sendDevComm);
// Finalize connection establishment after remote peer has called connect.
// This call must not block for the connection to be established, and instead
// should return successfully with recvComm == NULL with the expectation that
// it will be called again until recvComm != NULL.
// If *recvDevComm points to a valid object, then NCCL is requesting device offload for this connection
ncclResult_t (*accept)(void* listenComm, void** recvComm, ncclNetDeviceHandle_v10_t** recvDevComm);
// Register/Deregister memory. Comm can be either a sendComm or a recvComm.
// Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* comm, void* data, size_t size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* comm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* comm, void* mhandle);
// Asynchronous send to a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*isend)(void* sendComm, void* data, size_t size, int tag, void* mhandle, void* phandle, void** request);
// Asynchronous recv from a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*irecv)(void* recvComm, int n, void** data, size_t* sizes, int* tags, void** mhandles, void** phandles, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* recvComm, int n, void** data, int* sizes, void** mhandles, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* sizes);
// Close and free send/recv comm objects
ncclResult_t (*closeSend)(void* sendComm);
ncclResult_t (*closeRecv)(void* recvComm);
ncclResult_t (*closeListen)(void* listenComm);
// Copy the given mhandle to a dptr in a format usable by this plugin's device code
ncclResult_t (*getDeviceMr)(void* comm, void* mhandle, void** dptr_mhandle);
// Notify the plugin that a recv has completed by the device
ncclResult_t (*irecvConsumed)(void* recvComm, int n, void* request);
// Virtual NIC APIs. makeVDevice will create a virtual NIC given the specified properties, and tell the caller
// what index this new vNIC exists at
ncclResult_t (*makeVDevice)(int* d, ncclNetVDeviceProps_v10_t* props);
} ncclNet_v10_t;
typedef struct {
void* mhandle;
void* address;
size_t size;
} ncclNetSGE_v10_t;
typedef struct {
// Name of the collective network (mainly for logs)
const char* name;
// Initialize the collective network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters capable of doing collective operations.
// If ndev returns 0, all other functions might be set to NULL.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v10_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create connections.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Create a group for collective operations. handles have been created
// using listen() above. rank indicates caller's rank in the collective network.
ncclResult_t (*connect)(void* handles[], int nranks, int rank, void* listenComm, void** collComm);
// Returns whether a reduction operation on a data type is supported.
// 1 for supported, 0 otherwise.
ncclResult_t (*reduceSupport)(ncclDataType_t dataType, ncclRedOp_t redOp, int* supported);
// Register/Deregister memory. Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* collComm, void* data, size_t size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* collComm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* collComm, void* mhandle);
// Performs an asynchronous allreduce operation on the collective group.
// May return request == NULL if the call cannot be performed (or would block).
ncclResult_t (*iallreduce)(void* collComm, void* sendData, void* recvData, size_t count,
ncclDataType_t dataType, ncclRedOp_t redOp, void* sendMhandle, void* recvMhandle, void** request);
ncclResult_t (*iallgather)(void* collComm, void* sendData, int nRecvParts, ncclNetSGE_v10_t* recvParts,
size_t bytesPerRank, size_t windowOffset, size_t windowBytes,
void* sendMhandle, void** request);
ncclResult_t (*ireducescatter)(void* collComm, int nSendParts, ncclNetSGE_v10_t* sendParts, void* recvData,
size_t bytesPerRank, size_t windowOffset, size_t windowBytes,
ncclDataType_t dataType, ncclRedOp_t redOp,
void* recvMhandle, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* collComm, void* data, int size, void* mhandle, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* size);
// Close and free collective comm objects
ncclResult_t (*closeColl)(void* collComm);
ncclResult_t (*closeListen)(void* listenComm);
// Create a virtual NIC given the specified properties, which can be accessed at device index d
ncclResult_t (*makeVDevice)(int* d, ncclNetVDeviceProps_v10_t* props);
} ncclCollNet_v10_t;
#endif // end include guard
src/include/plugin/net/net_v6.h
+113
View File
@@ -0,0 +1,113 @@
/*************************************************************************
* Copyright (c) 2017-2022, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NET_V6_H_
#define NET_V6_H_
#define NCCL_NET_MAX_REQUESTS_V6 8
// v6 struct for backwards compatibility
typedef struct {
char* name; // Used mostly for logging.
char* pciPath; // Path to the PCI device in /sys.
uint64_t guid; // Unique identifier for the NIC chip. Important for
// cards with multiple PCI functions (Physical or virtual).
int ptrSupport; // [NCCL_PTR_HOST|NCCL_PTR_CUDA|NCCL_PTR_DMABUF]
int speed; // Port speed in Mbps.
int port; // Port number.
float latency; // Network latency
int maxComms; // Maximum number of comms we can create
int maxRecvs; // Maximum number of grouped receives.
} ncclNetProperties_v6_t;
typedef struct {
// Name of the network (mainly for logs)
const char* name;
// Initialize the network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v6_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create a connection.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Connect to a handle and return a sending comm object for that peer.
// This call must not block for the connection to be established, and instead
// should return successfully with sendComm == NULL with the expectation that
// it will be called again until sendComm != NULL.
ncclResult_t (*connect)(int dev, void* handle, void** sendComm);
// Finalize connection establishment after remote peer has called connect.
// This call must not block for the connection to be established, and instead
// should return successfully with recvComm == NULL with the expectation that
// it will be called again until recvComm != NULL.
ncclResult_t (*accept)(void* listenComm, void** recvComm);
// Register/Deregister memory. Comm can be either a sendComm or a recvComm.
// Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* comm, void* data, int size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* comm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* comm, void* mhandle);
// Asynchronous send to a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*isend)(void* sendComm, void* data, int size, int tag, void* mhandle, void** request);
// Asynchronous recv from a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*irecv)(void* recvComm, int n, void** data, int* sizes, int* tags, void** mhandles, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* recvComm, int n, void** data, int* sizes, void** mhandles, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* sizes);
// Close and free send/recv comm objects
ncclResult_t (*closeSend)(void* sendComm);
ncclResult_t (*closeRecv)(void* recvComm);
ncclResult_t (*closeListen)(void* listenComm);
} ncclNet_v6_t;
typedef struct {
// Name of the collective network (mainly for logs)
const char* name;
// Initialize the collective network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters capable of doing collective operations.
// If ndev returns 0, all other functions might be set to NULL.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v6_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create connections.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Create a group for collective operations. handles have been created
// using listen() above. rank indicates caller's rank in the collective network.
ncclResult_t (*connect)(void* handles[], int nranks, int rank, void* listenComm, void** collComm);
// Returns whether a reduction operation on a data type is supported.
// 1 for supported, 0 otherwise.
ncclResult_t (*reduceSupport)(ncclDataType_t dataType, ncclRedOp_t redOp, int* supported);
// Register/Deregister memory. Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* collComm, void* data, int size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* collComm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* collComm, void* mhandle);
// Performs an asynchronous allreduce operation on the collective group.
// May return request == NULL if the call cannot be performed (or would block).
ncclResult_t (*iallreduce)(void* collComm, void* sendData, void* recvData, int count,
ncclDataType_t dataType, ncclRedOp_t redOp, void* sendMhandle, void* recvMhandle, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* collComm, void* data, int size, void* mhandle, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* size);
// Close and free collective comm objects
ncclResult_t (*closeColl)(void* collComm);
ncclResult_t (*closeListen)(void* listenComm);
} ncclCollNet_v6_t;
#endif
src/include/plugin/net/net_v7.h
+120
View File
@@ -0,0 +1,120 @@
/*************************************************************************
* Copyright (c) 2017-2022, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NET_V7_H_
#define NET_V7_H_
typedef struct {
char* name; // Used mostly for logging.
char* pciPath; // Path to the PCI device in /sys.
uint64_t guid; // Unique identifier for the NIC chip. Important for
// cards with multiple PCI functions (Physical or virtual).
int ptrSupport; // [NCCL_PTR_HOST|NCCL_PTR_CUDA|NCCL_PTR_DMABUF]
int speed; // Port speed in Mbps.
int port; // Port number.
float latency; // Network latency
int maxComms; // Maximum number of comms we can create
int maxRecvs; // Maximum number of grouped receives.
ncclNetDeviceType netDeviceType; // Network offload type
int netDeviceVersion; // Version number for network offload
} ncclNetProperties_v7_t;
typedef struct {
// Name of the network (mainly for logs)
const char* name;
// Initialize the network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v7_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create a connection.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Connect to a handle and return a sending comm object for that peer.
// This call must not block for the connection to be established, and instead
// should return successfully with sendComm == NULL with the expectation that
// it will be called again until sendComm != NULL.
// If *sendDevComm points to a valid object, then NCCL is requesting device offload for this connection
ncclResult_t (*connect)(int dev, void* handle, void** sendComm, ncclNetDeviceHandle_v7_t** sendDevComm);
// Finalize connection establishment after remote peer has called connect.
// This call must not block for the connection to be established, and instead
// should return successfully with recvComm == NULL with the expectation that
// it will be called again until recvComm != NULL.
// If *recvDevComm points to a valid object, then NCCL is requesting device offload for this connection
ncclResult_t (*accept)(void* listenComm, void** recvComm, ncclNetDeviceHandle_v7_t** recvDevComm);
// Register/Deregister memory. Comm can be either a sendComm or a recvComm.
// Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* comm, void* data, int size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* comm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* comm, void* mhandle);
// Asynchronous send to a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*isend)(void* sendComm, void* data, int size, int tag, void* mhandle, void** request);
// Asynchronous recv from a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*irecv)(void* recvComm, int n, void** data, int* sizes, int* tags, void** mhandles, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* recvComm, int n, void** data, int* sizes, void** mhandles, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* sizes);
// Close and free send/recv comm objects
ncclResult_t (*closeSend)(void* sendComm);
ncclResult_t (*closeRecv)(void* recvComm);
ncclResult_t (*closeListen)(void* listenComm);
// Copy the given mhandle to a dptr in a format usable by this plugin's device code
ncclResult_t (*getDeviceMr)(void* comm, void* mhandle, void** dptr_mhandle);
// Notify the plugin that a recv has completed by the device
ncclResult_t (*irecvConsumed)(void* recvComm, int n, void* request);
} ncclNet_v7_t;
typedef struct {
// Name of the collective network (mainly for logs)
const char* name;
// Initialize the collective network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters capable of doing collective operations.
// If ndev returns 0, all other functions might be set to NULL.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v7_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create connections.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Create a group for collective operations. handles have been created
// using listen() above. rank indicates caller's rank in the collective network.
ncclResult_t (*connect)(void* handles[], int nranks, int rank, void* listenComm, void** collComm);
// Returns whether a reduction operation on a data type is supported.
// 1 for supported, 0 otherwise.
ncclResult_t (*reduceSupport)(ncclDataType_t dataType, ncclRedOp_t redOp, int* supported);
// Register/Deregister memory. Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* collComm, void* data, int size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* collComm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* collComm, void* mhandle);
// Performs an asynchronous allreduce operation on the collective group.
// May return request == NULL if the call cannot be performed (or would block).
ncclResult_t (*iallreduce)(void* collComm, void* sendData, void* recvData, int count,
ncclDataType_t dataType, ncclRedOp_t redOp, void* sendMhandle, void* recvMhandle, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* collComm, void* data, int size, void* mhandle, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* size);
// Close and free collective comm objects
ncclResult_t (*closeColl)(void* collComm);
ncclResult_t (*closeListen)(void* listenComm);
} ncclCollNet_v7_t;
#endif
src/include/plugin/net/net_v8.h
+134
View File
@@ -0,0 +1,134 @@
/*************************************************************************
* Copyright (c) 2017-2022, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NET_V8_H_
#define NET_V8_H_
typedef struct {
char* name; // Used mostly for logging.
char* pciPath; // Path to the PCI device in /sys.
uint64_t guid; // Unique identifier for the NIC chip. Important for
// cards with multiple PCI functions (Physical or virtual).
int ptrSupport; // [NCCL_PTR_HOST|NCCL_PTR_CUDA|NCCL_PTR_DMABUF]
int regIsGlobal; // regMr is not tied to a particular comm
int speed; // Port speed in Mbps.
int port; // Port number.
float latency; // Network latency
int maxComms; // Maximum number of comms we can create
int maxRecvs; // Maximum number of grouped receives.
ncclNetDeviceType netDeviceType; // Network offload type
int netDeviceVersion; // Version number for network offload
} ncclNetProperties_v8_t;
typedef struct {
// Name of the network (mainly for logs)
const char* name;
// Initialize the network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v8_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create a connection.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Connect to a handle and return a sending comm object for that peer.
// This call must not block for the connection to be established, and instead
// should return successfully with sendComm == NULL with the expectation that
// it will be called again until sendComm != NULL.
// If *sendDevComm points to a valid object, then NCCL is requesting device offload for this connection
ncclResult_t (*connect)(int dev, void* handle, void** sendComm, ncclNetDeviceHandle_v8_t** sendDevComm);
// Finalize connection establishment after remote peer has called connect.
// This call must not block for the connection to be established, and instead
// should return successfully with recvComm == NULL with the expectation that
// it will be called again until recvComm != NULL.
// If *recvDevComm points to a valid object, then NCCL is requesting device offload for this connection
ncclResult_t (*accept)(void* listenComm, void** recvComm, ncclNetDeviceHandle_v8_t** recvDevComm);
// Register/Deregister memory. Comm can be either a sendComm or a recvComm.
// Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* comm, void* data, size_t size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* comm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* comm, void* mhandle);
// Asynchronous send to a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*isend)(void* sendComm, void* data, int size, int tag, void* mhandle, void** request);
// Asynchronous recv from a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*irecv)(void* recvComm, int n, void** data, int* sizes, int* tags, void** mhandles, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* recvComm, int n, void** data, int* sizes, void** mhandles, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* sizes);
// Close and free send/recv comm objects
ncclResult_t (*closeSend)(void* sendComm);
ncclResult_t (*closeRecv)(void* recvComm);
ncclResult_t (*closeListen)(void* listenComm);
// Copy the given mhandle to a dptr in a format usable by this plugin's device code
ncclResult_t (*getDeviceMr)(void* comm, void* mhandle, void** dptr_mhandle);
// Notify the plugin that a recv has completed by the device
ncclResult_t (*irecvConsumed)(void* recvComm, int n, void* request);
} ncclNet_v8_t;
typedef struct {
void* mhandle;
void* address;
uint32_t size;
} ncclNetSGE_v8_t;
typedef struct {
// Name of the collective network (mainly for logs)
const char* name;
// Initialize the collective network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters capable of doing collective operations.
// If ndev returns 0, all other functions might be set to NULL.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v8_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create connections.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Create a group for collective operations. handles have been created
// using listen() above. rank indicates caller's rank in the collective network.
ncclResult_t (*connect)(void* handles[], int nranks, int rank, void* listenComm, void** collComm);
// Returns whether a reduction operation on a data type is supported.
// 1 for supported, 0 otherwise.
ncclResult_t (*reduceSupport)(ncclDataType_t dataType, ncclRedOp_t redOp, int* supported);
// Register/Deregister memory. Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* collComm, void* data, size_t size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* collComm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* collComm, void* mhandle);
// Performs an asynchronous allreduce operation on the collective group.
// May return request == NULL if the call cannot be performed (or would block).
ncclResult_t (*iallreduce)(void* collComm, void* sendData, void* recvData, int count,
ncclDataType_t dataType, ncclRedOp_t redOp, void* sendMhandle, void* recvMhandle, void** request);
ncclResult_t (*iallgather)(void* collComm, void* sendData, int nRecvParts, ncclNetSGE_v8_t* recvParts,
size_t bytesPerRank, size_t windowOffset, size_t windowBytes,
void* sendMhandle, void** request);
ncclResult_t (*ireducescatter)(void* collComm, int nSendParts, ncclNetSGE_v8_t* sendParts, void* recvData,
size_t bytesPerRank, size_t windowOffset, size_t windowBytes,
ncclDataType_t dataType, ncclRedOp_t redOp,
void* recvMhandle, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* collComm, void* data, int size, void* mhandle, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* size);
// Close and free collective comm objects
ncclResult_t (*closeColl)(void* collComm);
ncclResult_t (*closeListen)(void* listenComm);
} ncclCollNet_v8_t;
#endif
src/include/plugin/net/net_v9.h
+152
View File
@@ -0,0 +1,152 @@
/*************************************************************************
* Copyright (c) 2017-2022, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NET_V9_H_
#define NET_V9_H_
#define NCCL_NET_MAX_DEVS_PER_NIC_V9 4
typedef struct {
int ndevs;
int devs[NCCL_NET_MAX_DEVS_PER_NIC_V9];
} ncclNetVDeviceProps_v9_t;
typedef struct {
char* name; // Used mostly for logging.
char* pciPath; // Path to the PCI device in /sys.
uint64_t guid; // Unique identifier for the NIC chip. Important for
// cards with multiple PCI functions (Physical or virtual).
int ptrSupport; // [NCCL_PTR_HOST|NCCL_PTR_CUDA|NCCL_PTR_DMABUF]
int regIsGlobal; // regMr is not tied to a particular comm
int forceFlush; // Force a flush on receives
int speed; // Port speed in Mbps.
int port; // Port number.
float latency; // Network latency
int maxComms; // Maximum number of comms we can create
int maxRecvs; // Maximum number of grouped receives.
ncclNetDeviceType netDeviceType; // Network offload type
int netDeviceVersion; // Version number for network offload
ncclNetVDeviceProps_v9_t vProps;
size_t maxP2pBytes; // Max transfer size for point-to-point operations
size_t maxCollBytes; // Max transfer size for collective operations
} ncclNetProperties_v9_t;
typedef struct {
// Name of the network (mainly for logs)
const char* name;
// Initialize the network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v9_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create a connection.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Connect to a handle and return a sending comm object for that peer.
// This call must not block for the connection to be established, and instead
// should return successfully with sendComm == NULL with the expectation that
// it will be called again until sendComm != NULL.
// If *sendDevComm points to a valid object, then NCCL is requesting device offload for this connection
ncclResult_t (*connect)(int dev, void* handle, void** sendComm, ncclNetDeviceHandle_v9_t** sendDevComm);
// Finalize connection establishment after remote peer has called connect.
// This call must not block for the connection to be established, and instead
// should return successfully with recvComm == NULL with the expectation that
// it will be called again until recvComm != NULL.
// If *recvDevComm points to a valid object, then NCCL is requesting device offload for this connection
ncclResult_t (*accept)(void* listenComm, void** recvComm, ncclNetDeviceHandle_v9_t** recvDevComm);
// Register/Deregister memory. Comm can be either a sendComm or a recvComm.
// Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* comm, void* data, size_t size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* comm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* comm, void* mhandle);
// Asynchronous send to a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*isend)(void* sendComm, void* data, size_t size, int tag, void* mhandle, void** request);
// Asynchronous recv from a peer.
// May return request == NULL if the call cannot be performed (or would block)
ncclResult_t (*irecv)(void* recvComm, int n, void** data, size_t* sizes, int* tags, void** mhandles, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* recvComm, int n, void** data, int* sizes, void** mhandles, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* sizes);
// Close and free send/recv comm objects
ncclResult_t (*closeSend)(void* sendComm);
ncclResult_t (*closeRecv)(void* recvComm);
ncclResult_t (*closeListen)(void* listenComm);
// Copy the given mhandle to a dptr in a format usable by this plugin's device code
ncclResult_t (*getDeviceMr)(void* comm, void* mhandle, void** dptr_mhandle);
// Notify the plugin that a recv has completed by the device
ncclResult_t (*irecvConsumed)(void* recvComm, int n, void* request);
// Virtual NIC APIs. makeVDevice will create a virtual NIC given the specified properties, and tell the caller
// what index this new vNIC exists at
ncclResult_t (*makeVDevice)(int* d, ncclNetVDeviceProps_v9_t* props);
} ncclNet_v9_t;
typedef struct {
void* mhandle;
void* address;
size_t size;
} ncclNetSGE_v9_t;
typedef struct {
// Name of the collective network (mainly for logs)
const char* name;
// Initialize the collective network.
ncclResult_t (*init)(ncclDebugLogger_t logFunction);
// Return the number of adapters capable of doing collective operations.
// If ndev returns 0, all other functions might be set to NULL.
ncclResult_t (*devices)(int* ndev);
// Get various device properties.
ncclResult_t (*getProperties)(int dev, ncclNetProperties_v9_t* props);
// Create a receiving object and provide a handle to connect to it. The
// handle can be up to NCCL_NET_HANDLE_MAXSIZE bytes and will be exchanged
// between ranks to create connections.
ncclResult_t (*listen)(int dev, void* handle, void** listenComm);
// Create a group for collective operations. handles have been created
// using listen() above. rank indicates caller's rank in the collective network.
ncclResult_t (*connect)(void* handles[], int nranks, int rank, void* listenComm, void** collComm);
// Returns whether a reduction operation on a data type is supported.
// 1 for supported, 0 otherwise.
ncclResult_t (*reduceSupport)(ncclDataType_t dataType, ncclRedOp_t redOp, int* supported);
// Register/Deregister memory. Type is either NCCL_PTR_HOST or NCCL_PTR_CUDA.
ncclResult_t (*regMr)(void* collComm, void* data, size_t size, int type, void** mhandle);
/* DMA-BUF support */
ncclResult_t (*regMrDmaBuf)(void* collComm, void* data, size_t size, int type, uint64_t offset, int fd, void** mhandle);
ncclResult_t (*deregMr)(void* collComm, void* mhandle);
// Performs an asynchronous allreduce operation on the collective group.
// May return request == NULL if the call cannot be performed (or would block).
ncclResult_t (*iallreduce)(void* collComm, void* sendData, void* recvData, size_t count,
ncclDataType_t dataType, ncclRedOp_t redOp, void* sendMhandle, void* recvMhandle, void** request);
ncclResult_t (*iallgather)(void* collComm, void* sendData, int nRecvParts, ncclNetSGE_v9_t* recvParts,
size_t bytesPerRank, size_t windowOffset, size_t windowBytes,
void* sendMhandle, void** request);
ncclResult_t (*ireducescatter)(void* collComm, int nSendParts, ncclNetSGE_v9_t* sendParts, void* recvData,
size_t bytesPerRank, size_t windowOffset, size_t windowBytes,
ncclDataType_t dataType, ncclRedOp_t redOp,
void* recvMhandle, void** request);
// Perform a flush/fence to make sure all data received with NCCL_PTR_CUDA is
// visible to the GPU
ncclResult_t (*iflush)(void* collComm, void* data, int size, void* mhandle, void** request);
// Test whether a request is complete. If size is not NULL, it returns the
// number of bytes sent/received.
ncclResult_t (*test)(void* request, int* done, int* size);
// Close and free collective comm objects
ncclResult_t (*closeColl)(void* collComm);
ncclResult_t (*closeListen)(void* listenComm);
// Create a virtual NIC given the specified properties, which can be accessed at device index d
ncclResult_t (*makeVDevice)(int* d, ncclNetVDeviceProps_v9_t* props);
} ncclCollNet_v9_t;
#endif // end include guard
src/include/plugin/plugin.h
+18
View File
@@ -0,0 +1,18 @@
/*************************************************************************
* Copyright (c) 2024-2025, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NCCL_PLUGIN_H_
#define NCCL_PLUGIN_H_
#include "nccl.h"
void* ncclOpenNetPluginLib(const char* name);
void* ncclOpenTunerPluginLib(const char* name);
void* ncclOpenProfilerPluginLib(const char* name);
void* ncclGetNetPluginLib(void);
ncclResult_t ncclClosePluginLib(void* handle);
#endif
src/include/plugin/profiler/net_ib.h
+13
View File
@@ -0,0 +1,13 @@
/*************************************************************************
* Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NET_IB_H_
#define NET_IB_H_
#include "nccl_profiler.h"
#include "net_ib_v1.h"
#endif
src/include/plugin/profiler/net_ib_v1.h
+34
View File
@@ -0,0 +1,34 @@
/*************************************************************************
* Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NET_IB_V1_H_
#define NET_IB_V1_H_
#define NCCL_PROFILER_NET_IB_VER 1
enum {
ncclProfileQp = (1 << 0),
};
// The data structure version is encoded in the plugin identifier bitmask and
// passed to NCCL core through the profiler callback. NCCL copies the plugin
// identifier in the event descriptor before calling the profiler startEvent
// function. The profiler should inspect the plugin id to find out the source
// plugin as well as the version of the event struct
typedef struct {
uint8_t type; // event type (plugin defined)
union {
struct {
int device; // network device id
uint64_t wr_id; // work request id
int opcode; // ibv opcode
int qpNum; // QP number
size_t length; // work request data length
} qp;
};
} ncclProfilerNetIbDescr_v1_t;
#endif
src/include/plugin/profiler/net_socket.h
+13
View File
@@ -0,0 +1,13 @@
/*************************************************************************
* Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NET_SOCKET_H_
#define NET_SOCKET_H_
#include "nccl_profiler.h"
#include "net_socket_v1.h"
#endif
src/include/plugin/profiler/net_socket_v1.h
+32
View File
@@ -0,0 +1,32 @@
/*************************************************************************
* Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef NET_SOCKET_V1_H_
#define NET_SOCKET_V1_H_
#define NCCL_PROFILER_NET_SOCKET_VER 1
enum {
ncclProfileSocket = (1 << 0),
};
// The data structure version is encoded in the plugin identifier bitmask and
// passed to NCCL core through the profiler callback. NCCL copies the plugin
// identifier in the event descriptor before calling the profiler startEvent
// function. The profiler should inspect the plugin id to find out the source
// plugin as well as the version of the event struct
typedef struct {
uint8_t type; // event type (plugin defined)
union {
struct {
int fd;
int op;
size_t length;
} sock;
};
} ncclProfilerNetSockDescr_v1_t;
#endif
src/include/plugin/profiler/profiler_v1.h
+107
View File
@@ -0,0 +1,107 @@
/*************************************************************************
* Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef PROFILER_V1_H_
#define PROFILER_V1_H_
typedef struct {
uint8_t type; // event type descriptor: ncclProfileColl, ...
void* parentObj; // pointer to the profiler parent object (for coll is the group)
int rank; // originating rank
union {
struct {
const char* name;
uint64_t commHash;
uint64_t seqNumber;
uint8_t func;
void const* sendBuff;
void* recvBuff;
size_t count;
int root;
uint8_t datatype;
uint32_t op;
size_t trafficBytes;
uint8_t nMaxChannels;
uint8_t nWarps;
uint8_t algo;
uint8_t proto;
int isCollnet;
int isNvls;
} coll;
struct {
const char* name;
uint64_t commHash;
uint8_t func;
void* buff;
uint8_t datatype;
size_t count;
int peer;
} p2p;
struct {
pid_t pid; // pid of the originating process
uint8_t channelId; // channel id for this proxy operation
int peer; // remote rank for send/recv
int nSteps; // number of steps for this proxy operation
int chunkSize; // amount of data transferred by this proxy operation
int isSend;
} proxyOp;
struct {
int step;
} proxyStep;
};
} ncclProfilerEventDescr_v1_t;
typedef union {
struct {
size_t transSize;
int steps;
} proxyOp;
struct {
int appendedProxyOps;
} proxyCtrl;
} ncclProfilerEventStateArgs_v1_t;
typedef struct {
const char* name;
// init - initialize the profiler plugin
// Input
// - context : opaque profiler context object for separating profiler behavior across comms
// Output
// - eActivationMask: bitmask of active events set by the plugin
ncclResult_t (*init)(void** context, int* eActivationMask);
// startEvent - initialize and start a new event for the supplied event descriptor inside the eventset
// Input
// - context: opaque profiler context object
// - eDescr : pointer to ncclProfilerEventDescr_t object
// Output
// - eHandle: return event handle for supplied event descriptor object
ncclResult_t (*startEvent)(void* context, void** eHandle, ncclProfilerEventDescr_v1_t* eDescr);
// stopEvent - stop/finalize an event inside and event set
// Input
// - eHandle: handle to event object
ncclResult_t (*stopEvent)(void* eHandle);
// recordEventState - record event state transitions and event attribute updates
// Input
// - eHandle : handle to event object created through startEvent
// - eStateArgs: optional argument used to capture event attribute updates associated with the state transition
// - eState : event state transition
ncclResult_t (*recordEventState)(void* eHandle, ncclProfilerEventState_v1_t eState, ncclProfilerEventStateArgs_v1_t* eStateArgs);
// finalize - finalize the profiler plugin
// Input
// - context: opaque profiler context object
ncclResult_t (*finalize)(void* context);
} ncclProfiler_v1_t;
#endif
src/include/plugin/profiler/profiler_v2.h
+104
View File
@@ -0,0 +1,104 @@
/*************************************************************************
* Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef PROFILER_V2_H_
#define PROFILER_V2_H_
typedef struct {
uint8_t type; // event type descriptor: ncclProfileColl, ...
void* parentObj; // pointer to the profiler parent object (for coll is the group)
int rank; // originating rank
union {
struct {
const char* name;
uint64_t commHash;
uint64_t seqNumber;
const char* func;
void const* sendBuff;
void* recvBuff;
size_t count;
int root;
const char* datatype;
size_t trafficBytes;
uint8_t nMaxChannels;
uint8_t nWarps;
const char* algo;
const char* proto;
} coll;
struct {
const char* name;
uint64_t commHash;
const char* func;
void* buff;
const char* datatype;
size_t count;
int peer;
} p2p;
struct {
pid_t pid; // pid of the originating process
uint8_t channelId; // channel id for this proxy operation
int peer; // remote rank for send/recv
int nSteps; // number of steps for this proxy operation
int chunkSize; // amount of data transferred by this proxy operation
int isSend;
} proxyOp;
struct {
int step;
} proxyStep;
};
} ncclProfilerEventDescr_v2_t;
typedef union {
struct {
size_t transSize;
int steps;
} proxyOp;
struct {
int appendedProxyOps;
} proxyCtrl;
} ncclProfilerEventStateArgs_v2_t;
typedef struct {
const char* name;
// init - initialize the profiler plugin
// Input
// - context : opaque profiler context object for separating profiler behavior across comms
// Output
// - eActivationMask: bitmask of active events set by the plugin
ncclResult_t (*init)(void** context, int* eActivationMask);
// startEvent - initialize and start a new event for the supplied event descriptor inside the eventset
// Input
// - context: opaque profiler context object
// - eDescr : pointer to ncclProfilerEventDescr_t object
// Output
// - eHandle: return event handle for supplied event descriptor object
ncclResult_t (*startEvent)(void* context, void** eHandle, ncclProfilerEventDescr_v2_t* eDescr);
// stopEvent - stop/finalize an event inside and event set
// Input
// - eHandle: handle to event object
ncclResult_t (*stopEvent)(void* eHandle);
// recordEventState - record event state transitions and event attribute updates
// Input
// - eHandle : handle to event object created through startEvent
// - eStateArgs: optional argument used to capture event attribute updates associated with the state transition
// - eState : event state transition
ncclResult_t (*recordEventState)(void* eHandle, ncclProfilerEventState_v2_t eState, ncclProfilerEventStateArgs_v2_t* eStateArgs);
// finalize - finalize the profiler plugin
// Input
// - context: opaque profiler context object
ncclResult_t (*finalize)(void* context);
} ncclProfiler_v2_t;
#endif
src/include/plugin/profiler/profiler_v3.h
+112
View File
@@ -0,0 +1,112 @@
/*************************************************************************
* Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef PROFILER_V3_H_
#define PROFILER_V3_H_
typedef struct {
uint8_t type; // event type descriptor: ncclProfileColl, ...
void* parentObj; // pointer to the profiler parent object (for coll is the group)
int rank; // originating rank
union {
struct {
const char* name;
uint64_t commHash;
uint64_t seqNumber;
const char* func;
void const* sendBuff;
void* recvBuff;
size_t count;
int root;
const char* datatype;
uint8_t nMaxChannels;
uint8_t nWarps;
const char* algo;
const char* proto;
} coll;
struct {
const char* name;
uint64_t commHash;
const char* func;
void* buff;
const char* datatype;
size_t count;
int peer;
} p2p;
struct {
pid_t pid; // pid of the originating process
uint8_t channelId; // channel id for this proxy operation
int peer; // remote rank for send/recv
int nSteps; // number of steps for this proxy operation
int chunkSize; // amount of data transferred by this proxy operation
int isSend;
} proxyOp;
struct {
int step;
} proxyStep;
struct {
uint8_t channelId;
} kernelCh;
struct {
int64_t id;
void* data;
} netPlugin;
};
} ncclProfilerEventDescr_v3_t;
typedef union {
struct {
size_t transSize;
int steps;
} proxyOp;
struct {
int appendedProxyOps;
} proxyCtrl;
} ncclProfilerEventStateArgs_v3_t;
typedef struct {
const char* name;
// init - initialize the profiler plugin
// Input
// - context : opaque profiler context object for separating profiler behavior across comms
// Output
// - eActivationMask: bitmask of active events set by the plugin
ncclResult_t (*init)(void** context, int* eActivationMask);
// startEvent - initialize and start a new event for the supplied event descriptor inside the eventset
// Input
// - context: opaque profiler context object
// - eDescr : pointer to ncclProfilerEventDescr_t object
// Output
// - eHandle: return event handle for supplied event descriptor object
ncclResult_t (*startEvent)(void* context, void** eHandle, ncclProfilerEventDescr_v3_t* eDescr);
// stopEvent - stop/finalize an event inside and event set
// Input
// - eHandle: handle to event object
ncclResult_t (*stopEvent)(void* eHandle);
// recordEventState - record event state transitions and event attribute updates
// Input
// - eHandle : handle to event object created through startEvent
// - eStateArgs: optional argument used to capture event attribute updates associated with the state transition
// - eState : event state transition
ncclResult_t (*recordEventState)(void* eHandle, ncclProfilerEventState_v3_t eState, ncclProfilerEventStateArgs_v3_t* eStateArgs);
// finalize - finalize the profiler plugin
// Input
// - context: opaque profiler context object
ncclResult_t (*finalize)(void* context);
} ncclProfiler_v3_t;
#endif
src/include/plugin/tuner/tuner_v2.h
+53
View File
@@ -0,0 +1,53 @@
/*************************************************************************
* Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2023, Meta Platforms, Inc. and affiliates.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef TUNER_V2_H_
#define TUNER_V2_H_
// API to be implemented by external tuner
typedef struct {
// Name of the tuner
const char* name;
// Initializes tuner states.
// Inputs:
// - nRanks: number of ranks in current communicator. Each communicator initialize its own tuner.
// - nNodes: number of nodes in current communicator.
// - logFunction: a logFunction can be useful to integrate logging together with NCCL core.
// Outputs:
// - context: tuner context object
ncclResult_t (*init)(size_t nRanks, size_t nNodes, ncclDebugLogger_t logFunction, void **context);
// Gets info (algo, protocol, number of ctas and threads) for a given collective.
// Inputs:
// - context: tuner context object
// - collType: collective type , e.g., allreduce, allgather…
// - nBytes: collective size in bytes
// - collNetTypeSupport: whether collnet supports this type
// - nvlsTypeSupport: whether nvlink sharp supports this time
// - numPipeOps: number of operations in the group
//
// Outputs:
// - algorithm: selected algorithm to be used for the given collective
// - protocol: selected protocol to be used for the give collective
// - nChannels: number of channels (hence SMs) to be used.
//
// If getCollInfo() does not return ncclSuccess, NCCL will fall back to the
// default tuning for the given collective.
// Also, the plugin is allowed to not set any output, or set only the
// algorithm and protocol, but not only the algorithm or only the protocol.
// Unset fields will be set automatically by NCCL.
ncclResult_t (*getCollInfo)(void* context, ncclFunc_t collType, size_t nBytes,
int collNetSupport, int nvlsSupport, int numPipeOps,
int* algorithm, int* protocol, int* nChannels);
// Terminates the plugin and cleans up any resources that the plugin allocated.
// context: tuner context object
ncclResult_t (*destroy)(void* context);
} ncclTuner_v2_t;
#endif
src/include/plugin/tuner/tuner_v3.h
+55
View File
@@ -0,0 +1,55 @@
/*************************************************************************
* Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2023, Meta Platforms, Inc. and affiliates.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef TUNER_V3_H_
#define TUNER_V3_H_
// API to be implemented by external tuner
typedef struct {
// Name of the tuner
const char* name;
// Initializes tuner states.
// Inputs:
// - nRanks: number of ranks in current communicator. Each communicator initialize its own tuner.
// - nNodes: number of nodes in current communicator.
// - logFunction: a logFunction can be useful to integrate logging together with NCCL core.
// Outputs:
// - context: tuner context object
ncclResult_t (*init)(size_t nRanks, size_t nNodes, ncclDebugLogger_t logFunction, void **context);
// Gets info (algo, protocol, number of ctas and threads) for a given collective.
// Inputs:
// - context: tuner context object
// - collType: collective type , e.g., allreduce, allgather…
// - nBytes: collective size in bytes
// - numPipeOps: number of operations in the group
// - numAlgo: number of algorithms in collCostTable
// - numProto: number of protocols in collCostTable
//
// Outputs:
// - nChannels: number of channels (hence SMs) to be used.
//
// InOut:
// - collCostTable: collective cost table, generated by NCCL core, containing algo|proto|time entries for collType.
// NCCL core sets ignored algo/proto cost table entries to -1.0 (NCCL_ALGO_PROTO_IGNORE).
//
// If getCollInfo() does not return ncclSuccess, NCCL will fall back to the
// default tuning for the given collective.
// Also, the plugin is allowed to not set any output, or set only the
// algorithm and protocol, but not only the algorithm or only the protocol.
// Unset fields will be set automatically by NCCL.
ncclResult_t (*getCollInfo)(void* context, ncclFunc_t collType, size_t nBytes,
int numPipeOps, float** collCostTable, int numAlgo, int numProto,
int* nChannels);
// Terminates the plugin and cleans up any resources that the plugin allocated.
// context: tuner context object
ncclResult_t (*destroy)(void* context);
} ncclTuner_v3_t;
#endif
src/include/plugin/tuner/tuner_v4.h
+56
View File
@@ -0,0 +1,56 @@
/*************************************************************************
* Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved.
* Copyright (c) 2023, Meta Platforms, Inc. and affiliates.
*
* See LICENSE.txt for license information
************************************************************************/
#ifndef TUNER_V4_H_
#define TUNER_V4_H_
// API to be implemented by external tuner
typedef struct {
// Name of the tuner
const char* name;
// Initializes tuner states.
// Inputs:
// - nRanks: number of ranks in current communicator. Each communicator initialize its own tuner.
// - nNodes: number of nodes in current communicator.
// - logFunction: a logFunction can be useful to integrate logging together with NCCL core.
// Outputs:
// - context: tuner context object
ncclResult_t (*init)(size_t nRanks, size_t nNodes, ncclDebugLogger_t logFunction, void **context);
// Gets info (algo, protocol, number of ctas and threads) for a given collective.
// Inputs:
// - context: tuner context object
// - collType: collective type , e.g., allreduce, allgather…
// - nBytes: collective size in bytes
// - numPipeOps: number of operations in the group
// - numAlgo: number of algorithms in collCostTable
// - numProto: number of protocols in collCostTable
// - regBuff: can register user buffer
//
// Outputs:
// - nChannels: number of channels (hence SMs) to be used.
//
// InOut:
// - collCostTable: collective cost table, generated by NCCL core, containing algo|proto|time entries for collType.
// NCCL core sets ignored algo/proto cost table entries to -1.0 (NCCL_ALGO_PROTO_IGNORE).
//
// If getCollInfo() does not return ncclSuccess, NCCL will fall back to the
// default tuning for the given collective.
// Also, the plugin is allowed to not set any output, or set only the
// algorithm and protocol, but not only the algorithm or only the protocol.
// Unset fields will be set automatically by NCCL.
ncclResult_t (*getCollInfo)(void* context, ncclFunc_t collType, size_t nBytes,
int numPipeOps, float** collCostTable, int numAlgo, int numProto,
int regBuff, int* nChannels);
// Terminates the plugin and cleans up any resources that the plugin allocated.
// context: tuner context object
ncclResult_t (*destroy)(void* context);
} ncclTuner_v4_t;
#endif
src/include/profiler.h
+20
View File
@@ -17,6 +17,18 @@ struct ncclTaskP2p;
struct ncclInfo;
struct ncclComm;
struct ncclProxyOp;
struct ncclProxyConnector;
struct ncclProfilerProxy {
bool initialized;
uint64_t* workStarted/*[MAXCHANNELS]*/;
uint64_t* workCompleted/*[MAXCHANNELS]*/;
uint64_t workCounter[MAXCHANNELS]; // host work counter
struct ncclProxyConnector sendProxyConn[MAXCHANNELS];
struct ncclProxyConnector recvProxyConn[MAXCHANNELS];
};
extern int ncclProfilerEventMask;
// Plugin Init/Finalize Wrappers
ncclResult_t ncclProfilerPluginInit(struct ncclComm* comm);
@@ -44,6 +56,10 @@ ncclResult_t ncclProfilerStopProxyStepEvent(int sub, struct ncclProxyArgs* args,
ncclResult_t ncclProfilerStartProxyCtrlEvent(void* profilerContext, void** eHandle);
ncclResult_t ncclProfilerStopProxyCtrlEvent(void* eHandle);
// Kernel Channel Start/Stop Event Wrappers
ncclResult_t ncclProfilerStartKernelChEvent(struct ncclProxyArgs* args, int s);
ncclResult_t ncclProfilerStopKernelChEvent(struct ncclProxyArgs* args, int s);
// Record Event Wrappers
ncclResult_t ncclProfilerRecordProxyOpEventState(int sub, struct ncclProxyArgs* args, int steps, size_t transSize, ncclProfilerEventState_t eState);
ncclResult_t ncclProfilerRecordProxyStepEventState(int sub, struct ncclProxyArgs* args, int stepId, ncclProfilerEventState_t eState);
@@ -51,5 +67,9 @@ ncclResult_t ncclProfilerRecordProxyCtrlEventState(void*eHandle, int appended, n
// Profiler utility functions
ncclResult_t ncclProfilerAddPidToProxyOp(struct ncclProxyOp* op);
bool ncclProfilerNeedsProxy(struct ncclComm* comm, struct ncclProxyOp* op);
// Profiler callback for network plugin
ncclResult_t ncclProfilerCallback(void** eHandle, int type, void* pHandle, int64_t pluginId, void* extData);
#endif
src/include/proxy.h
+6 -1
View File
@@ -32,7 +32,8 @@ typedef enum : uint8_t {
ncclPatternPatUp,
ncclPatternPatDown,
ncclPatternSend,
ncclPatternRecv
ncclPatternRecv,
ncclPatternProfiler,
} ncclPattern_t;
enum ncclProxyOpState { ncclProxyOpNone, ncclProxyOpReady, ncclProxyOpProgress };
@@ -93,6 +94,7 @@ struct ncclProxyOp {
int peer;
pid_t pid;
void* profilerContext;
uint64_t workCounter;
struct ncclProxyOp *enqNext;
};
@@ -129,12 +131,15 @@ struct ncclProxySubArgs {
// Profiler plugin
int eActivationMask;
int rank;
uint64_t profilerSteps;
pid_t pid;
void* profilerContext;
void* taskEventHandle;
void* opEventHandle;
void* kernelEventHandle;
void* stepEventHandles[NCCL_STEPS];
size_t transSize;
uint64_t workCounter;
void* recvRequestsCache[NCCL_STEPS];
int recvRequestsSubCount;
src/include/ras.h
+2
View File
@@ -15,6 +15,8 @@ struct rasRankInit {
pid_t pid;
int cudaDev;
int nvmlDev;
uint64_t hostHash;
uint64_t pidHash;
};
ncclResult_t ncclRasCommInit(struct ncclComm* comm, struct rasRankInit* myRank);
src/include/register.h
+1 -1
View File
@@ -42,7 +42,7 @@ struct ncclReg {
uintptr_t baseAddr;
size_t baseSize;
CUdeviceptr regAddr;
size_t regSize;
size_t regUCSize, regMCSize;
int dev;
CUmemGenericAllocationHandle mcHandle;
uintptr_t caddrs[NCCL_MAX_LOCAL_RANKS]; /* use to check if NVLS buffers match among intra-node ranks */
src/include/shm.h
+2 -3
View File
@@ -14,7 +14,6 @@ struct shmCuIpc {
CUmemFabricHandle handle;
CUmemGenericAllocationHandle data;
};
int tpProxyRank;
void *ptr;
size_t size;
};
@@ -30,8 +29,8 @@ struct shmIpcDesc {
typedef struct shmIpcDesc ncclShmIpcDesc_t;
ncclResult_t ncclShmAllocateShareableBuffer(int tpProxyRank, size_t size, bool legacy, ncclShmIpcDesc_t *descOut, void **hptr, void **dptr);
ncclResult_t ncclShmImportShareableBuffer(struct ncclComm *comm, ncclShmIpcDesc_t *desc, void **hptr, void **dptr, ncclShmIpcDesc_t *descOut);
ncclResult_t ncclShmAllocateShareableBuffer(size_t size, bool legacy, ncclShmIpcDesc_t *descOut, void **hptr, void **dptr);
ncclResult_t ncclShmImportShareableBuffer(struct ncclComm *comm, int proxyRank, ncclShmIpcDesc_t *desc, void **hptr, void **dptr, ncclShmIpcDesc_t *descOut);
ncclResult_t ncclShmIpcClose(ncclShmIpcDesc_t *desc);
#endif
src/include/socket.h
+1 -1
View File
@@ -96,5 +96,5 @@ ncclResult_t ncclSocketRecv(struct ncclSocket* sock, void* ptr, int size);
ncclResult_t ncclSocketSendRecv(struct ncclSocket* sendSock, void* sendPtr, int sendSize, struct ncclSocket* recvSock, void* recvPtr, int recvSize);
ncclResult_t ncclSocketTryRecv(struct ncclSocket* sock, void* ptr, int size, int* closed, bool blocking);
ncclResult_t ncclSocketShutdown(struct ncclSocket* sock, int how);
ncclResult_t ncclSocketClose(struct ncclSocket* sock);
ncclResult_t ncclSocketClose(struct ncclSocket* sock, bool wait = false);
#endif
src/include/strongstream.h
+46 -52
View File
@@ -10,13 +10,24 @@
#include "nccl.h"
#include "checks.h"
#include <cuda.h>
#include <cuda_runtime.h>
#include <stdint.h>
// ncclCudaContext: wraps a CUDA context with per-context state.
struct ncclCudaContext;
// Get a ncclCudaContext to track the currently active CUDA context.
ncclResult_t ncclCudaContextTrack(struct ncclCudaContext** out);
// Drop reference.
void ncclCudaContextDrop(struct ncclCudaContext* cxt);
/* ncclCudaGraph: Wraps a cudaGraph_t so that we can support pre-graph CUDA runtimes
* easily.
*/
struct ncclCudaGraph {
#if CUDART_VERSION >= 11030
cudaStream_t origin;
cudaGraph_t graph;
unsigned long long graphId;
#endif
@@ -25,6 +36,7 @@ struct ncclCudaGraph {
inline struct ncclCudaGraph ncclCudaGraphNone() {
struct ncclCudaGraph tmp;
#if CUDART_VERSION >= 11030
tmp.origin = nullptr;
tmp.graph = nullptr;
tmp.graphId = ULLONG_MAX;
#endif
@@ -33,7 +45,7 @@ inline struct ncclCudaGraph ncclCudaGraphNone() {
inline bool ncclCudaGraphValid(struct ncclCudaGraph graph) {
#if CUDART_VERSION >= 11030
return graph.graph != nullptr;
return graph.graphId != ULLONG_MAX;
#else
return false;
#endif
@@ -57,60 +69,37 @@ ncclResult_t ncclCudaGraphAddDestructor(struct ncclCudaGraph graph, cudaHostFn_t
* streams unfit for the use of serializing access to a persistent resource.
* Strong streams have been introduced to address this need.
*
* - All updates to a strong stream must be enclosed by a Acquire/Release pair.
* All updates to a strong stream must be enclosed by a Acquire/Release pair.
*
* - The Acquire, Release, and all updates take a ncclCudaGraph parameter
* indicating the currently capturing graph (or none). This parameter must be
* the same for the entire sequence of {Acquire; ...; Release}.
* Acquire retrieves a "work" stream (cudaStream_t) which may be used to add
* work.
*
* - An {Acquire; ...; Release} sequence must not be concurrent with any
* other operations against the strong stream including graph launches which
* reference this stream.
* Release publishes the work streams work into the strong stream. The Release
* must be issued by the same thread that did the Acquire.
*/
struct ncclStrongStream;
ncclResult_t ncclStrongStreamConstruct(struct ncclStrongStream* ss);
ncclResult_t ncclStrongStreamDestruct(struct ncclStrongStream* ss);
// Acquire-fence the strong stream.
// Acquire the strong stream. Upon return `*workStream` will be usable to add work.
// `concurrent` indicates if other threads may be using the strong stream.
ncclResult_t ncclStrongStreamAcquire(
struct ncclCudaGraph graph, struct ncclStrongStream* ss
struct ncclCudaGraph graph, struct ncclStrongStream* ss, bool concurrent, cudaStream_t* workStream
);
// Acquire-fence the strong stream assuming no graph is capturing. This permits
// the caller to enqueue directly to the `ss->cudaStream` member using native CUDA
// calls. Strong stream still must be released via:
// ncclStrongStreamRelease(ncclCudaGraphNone(), ss);
ncclResult_t ncclStrongStreamAcquireUncaptured(struct ncclStrongStream* ss);
// Release-fence of the strong stream.
ncclResult_t ncclStrongStreamRelease(struct ncclCudaGraph graph, struct ncclStrongStream* ss);
// Add a host launch to the stream.
ncclResult_t ncclStrongStreamLaunchHost(
struct ncclCudaGraph graph, struct ncclStrongStream* ss,
cudaHostFn_t fn, void* arg
);
// Add a kernel launch to the stream.
ncclResult_t ncclStrongStreamLaunchKernel(
struct ncclCudaGraph graph, struct ncclStrongStream* ss,
void* fn, dim3 grid, dim3 block, void** args, size_t sharedMemBytes
// Get the workStream for an already acquired strong stream.
// `concurrent` indicates if other threads may be using the strong stream.
ncclResult_t ncclStrongStreamAcquiredWorkStream(
struct ncclCudaGraph graph, struct ncclStrongStream* ss, bool concurrent, cudaStream_t* workStream
);
// Cause `a` to wait for the current state `b`. Both `a` and `b` must be acquired.
// `b_subsumes_a` indicates that all work in `a` is already present in `b`, thus
// we want to fast-forward `a` to be a clone of `b`. Knowing this permits the
// implementation to induce few graph dependencies.
ncclResult_t ncclStrongStreamWaitStream(
struct ncclCudaGraph graph, struct ncclStrongStream* a, struct ncclStrongStream* b, bool b_subsumes_a=false
);
// `b` must be capturing within `graph`.
ncclResult_t ncclStrongStreamWaitStream(
struct ncclCudaGraph graph, struct ncclStrongStream* a, cudaStream_t b, bool b_subsumes_a=false
);
// `a` must be capturing within `graph`.
ncclResult_t ncclStrongStreamWaitStream(
struct ncclCudaGraph graph, cudaStream_t a, struct ncclStrongStream* b, bool b_subsumes_a=false
// Release of the strong stream.
// `concurrent` indicates if other threads may be using the strong stream.
ncclResult_t ncclStrongStreamRelease(struct ncclCudaGraph graph, struct ncclStrongStream* ss, bool concurrent);
ncclResult_t ncclStreamWaitStream(
cudaStream_t a, cudaStream_t b, cudaEvent_t scratchEvent
);
// Synchrnoization does not need the strong stream to be acquired.
@@ -118,23 +107,28 @@ ncclResult_t ncclStrongStreamSynchronize(struct ncclStrongStream* ss);
////////////////////////////////////////////////////////////////////////////////
struct ncclStrongStreamGraph; // internal to ncclStrongStream
struct ncclStrongStreamCapture; // internal to ncclStrongStream
struct ncclStrongStream {
// Used when not graph capturing.
cudaStream_t cudaStream;
// The stream to use for non-captured work.
cudaStream_t liveStream;
void* liveAcquiredBy;
#if CUDART_VERSION >= 11030
// This stream ever appeared in a graph capture.
bool everCaptured;
pthread_mutex_t lock;
struct ncclStrongStreamCapture* captureHead;
// The event used to establish order between graphs and streams. During acquire
// this event is waited on, during release it is recorded to.
cudaEvent_t serialEvent;
// This stream ever appeared in a graph capture.
bool everCaptured;
// Tracks whether serialEvent needs to be recorded to upon Release().
bool serialEventNeedsRecord;
struct ncclStrongStreamGraph* graphHead;
#else
cudaEvent_t scratchEvent;
#endif
};
struct ncclCudaContext {
struct ncclCudaContext* next;
CUcontext hcontext;
int refCount;
struct ncclStrongStream launchOrder;
};
#endif
src/include/transport.h
+7 -3
View File
@@ -18,6 +18,7 @@
#define TRANSPORT_SHM 1
#define TRANSPORT_NET 2
#define TRANSPORT_COLLNET 3
#define TRANSPORT_PROFILER 4
#include "proxy.h"
#include "comm.h"
@@ -26,6 +27,7 @@ extern struct ncclTransport p2pTransport;
extern struct ncclTransport shmTransport;
extern struct ncclTransport netTransport;
extern struct ncclTransport collNetTransport;
extern struct ncclTransport profilerTransport;
extern struct ncclTransport* ncclTransports[];
// Forward declarations
@@ -65,8 +67,10 @@ struct ncclNvlsSharedRes {
CUmulticastObjectProp signalProp;
CUmemAccessDesc accessDesc;
int dev;
size_t buffSize;
size_t creditSize;
size_t creditUCSize;
size_t creditMCSize;
size_t buffUCSize;
size_t buffMCSize;
CUmemGenericAllocationHandle mcBuffHandle; // Multicast handle for NVLS buffer
CUmemGenericAllocationHandle mcCreditHandle; // Multicast handle for NVLS credit buffer
char* mcBuff; // Multicast NVLS buffer address
@@ -123,7 +127,7 @@ ncclResult_t ncclNvlsBufferSetup(struct ncclComm* comm);
ncclResult_t ncclNvlsTreeConnect(struct ncclComm* comm);
ncclResult_t ncclNvlsGraphRegisterBuffer(struct ncclComm *comm, const void *sendbuff, void *recvbuff, size_t sendbuffSize, size_t recvbuffSize, int *outRegBufUsed, void **outRegBufSend, void **outRegBufRecv, struct ncclIntruQueue<struct ncclCommCallback, &ncclCommCallback::next>* cleanupQueue, int* nCleanupQueueElts);
ncclResult_t ncclNvlsLocalRegisterBuffer(struct ncclComm *comm, const void *sendbuff, void *recvbuff, size_t sendbuffSize, size_t recvbuffSize, int *outRegBufUsed, void **outRegBufSend, void **outRegBufRecv);
ncclResult_t ncclNvlsDeregBuffer(struct ncclComm* comm, CUmemGenericAllocationHandle *mcHandler, CUdeviceptr ptr, int dev, size_t size);
ncclResult_t ncclNvlsDeregBuffer(struct ncclComm* comm, CUmemGenericAllocationHandle *mcHandler, CUdeviceptr ptr, int dev, size_t ucsize, size_t mcsize);
ncclResult_t ncclNvlsFree(struct ncclComm* comm);
enum { collNetRecv=0, collNetSend=1 };