Merge remote-tracking branch 'nccl/master' into develop
Этот коммит содержится в:
Wenkai Du
@@ -29,7 +29,7 @@ all: $(STATICLIB)
|
||||
all_deps: $(DEPENDFILES)
|
||||
|
||||
# Auto-generating the rules per op/reduction/datatype/algorithm
|
||||
$(RULESFILE) :
|
||||
$(RULESFILE) : gen_rules.sh
|
||||
@printf "Generating %-35s > %s\n" rules $@
|
||||
@mkdir -p $(OBJDIR)
|
||||
@CUDA_MAJOR=${CUDA_MAJOR} CUDA_MINOR=${CUDA_MINOR} ./gen_rules.sh $(OBJDIR) > $@
|
||||
|
||||
@@ -584,22 +584,22 @@ struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_TREE, NCCL_PROTO_SI
|
||||
};
|
||||
|
||||
template<typename T, typename RedOp>
|
||||
struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_COLLNET, NCCL_PROTO_SIMPLE> {
|
||||
__device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
|
||||
struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_COLLNET_DIRECT, NCCL_PROTO_SIMPLE> {
|
||||
__device__ __forceinline__ void run(ncclWorkElem *args) {
|
||||
static constexpr int COLLNET_COPY_THREADS = 64;
|
||||
const int tid = threadIdx.x;
|
||||
const int bid = args->bid;
|
||||
const int nChannels = args->nChannels;
|
||||
struct ncclDirect* tree = &ncclShmem.channel.collTree;
|
||||
struct ncclDirect* direct = &ncclShmem.channel.collnetDirect;
|
||||
const ssize_t chunkSize = int(args->lastChunkSize);
|
||||
const ssize_t size = args->count;
|
||||
const ssize_t loopSize = nChannels*tree->nHeads*chunkSize;
|
||||
const ssize_t loopSize = nChannels*direct->nHeads*chunkSize;
|
||||
|
||||
const int hasUp = (tree->up[0] >= 0) ? 1 : 0;
|
||||
const int hasDn = (tree->down[0] >= 0) ? 1 : 0;
|
||||
const int nThreadsScatter = ((hasUp && hasDn) ? COLLNET_COPY_THREADS : hasUp ? 2*COLLNET_COPY_THREADS : 0);
|
||||
const int hasUp = (direct->up[0] >= 0) ? 1 : 0;
|
||||
const int hasDn = (direct->down[0] >= 0) ? 1 : 0;
|
||||
const int nThreadsScatter = WARP_SIZE + ((hasUp && hasDn) ? COLLNET_COPY_THREADS : hasUp ? 2*COLLNET_COPY_THREADS : 0);
|
||||
const int nThreadsGather = ((hasUp && hasDn) ? COLLNET_COPY_THREADS : hasUp ? 1*COLLNET_COPY_THREADS : 0);
|
||||
const int nThreadsBcast = ((hasUp && hasDn) ? COLLNET_COPY_THREADS : hasUp ? 0 : 1*COLLNET_COPY_THREADS);
|
||||
const int nThreadsBcast = WARP_SIZE + ((hasUp && hasDn) ? COLLNET_COPY_THREADS : hasUp ? 0 : 1*COLLNET_COPY_THREADS);
|
||||
const int nThreadsReduce = args->nWarps*WARP_SIZE - nThreadsScatter - nThreadsGather - nThreadsBcast;
|
||||
const int tidStartBcast = nThreadsGather;
|
||||
const int tidStartScatter = tidStartBcast + nThreadsBcast;
|
||||
@@ -611,24 +611,24 @@ struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_COLLNET, NCCL_PROTO
|
||||
// Scatter
|
||||
int group = (2*Proto::MaxGroupWidth) | (1<<16);
|
||||
Primitives<T, RedOp, FanAsymmetric<0, NCCL_MAX_DIRECT_ARITY>, /*Direct=*/0, Proto, 0>
|
||||
prims(tid-tidStartScatter, nThreadsScatter, NULL, tree->up, args->sendbuff, args->recvbuff, args->redOpArg, group, args);
|
||||
prims(tid-tidStartScatter, nThreadsScatter, NULL, direct->up, args->sendbuff, args->recvbuff, args->redOpArg, group, args);
|
||||
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
|
||||
ssize_t offset = gridOffset + bid*tree->nHeads*chunkSize;
|
||||
int nelem = min(tree->nHeads*chunkSize, size-offset);
|
||||
ssize_t offset = gridOffset + bid*direct->nHeads*chunkSize;
|
||||
int nelem = min(direct->nHeads*chunkSize, size-offset);
|
||||
if (args->regUsed) {
|
||||
prims.directScatter(offset, nelem, chunkSize, tree->headRank, tree->shift);
|
||||
prims.directScatter(offset, nelem, chunkSize, direct->headRank, direct->shift);
|
||||
} else {
|
||||
prims.scatter(offset, nelem, chunkSize, tree->headRank, tree->shift);
|
||||
prims.scatter(offset, nelem, chunkSize, direct->headRank, direct->shift);
|
||||
}
|
||||
}
|
||||
} else if (tid >= tidStartReduce && tree->out != -1) {
|
||||
} else if (tid >= tidStartReduce && direct->out != -1) {
|
||||
int group = (3*Proto::MaxGroupWidth) | (1<<16);
|
||||
if (hasDn) {
|
||||
// Reduce, send to network
|
||||
Primitives<T, RedOp, FanAsymmetric<NCCL_MAX_DIRECT_ARITY, 1>, /*Direct=*/0, Proto, 0>
|
||||
prims(tid-tidStartReduce, nThreadsReduce, tree->down, &tree->out, args->sendbuff, args->recvbuff, args->redOpArg, group, args);
|
||||
prims(tid-tidStartReduce, nThreadsReduce, direct->down, &direct->out, args->sendbuff, args->recvbuff, args->redOpArg, group, args);
|
||||
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
|
||||
ssize_t offset = gridOffset + (bid*tree->nHeads+tree->headRank)*chunkSize;
|
||||
ssize_t offset = gridOffset + (bid*direct->nHeads+direct->headRank)*chunkSize;
|
||||
int nelem = min(chunkSize, size-offset);
|
||||
if (args->regUsed) {
|
||||
prims.directRecvReduceSend(offset, offset, nelem);
|
||||
@@ -639,9 +639,9 @@ struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_COLLNET, NCCL_PROTO
|
||||
} else {
|
||||
// Directly send to network
|
||||
Primitives<T, RedOp, FanAsymmetric<0, 1>, /*Direct=*/0, Proto, 0>
|
||||
prims(tid-tidStartReduce, nThreadsReduce, nullptr, &tree->out, args->sendbuff, args->recvbuff, args->redOpArg, group);
|
||||
prims(tid-tidStartReduce, nThreadsReduce, nullptr, &direct->out, args->sendbuff, args->recvbuff, args->redOpArg, group);
|
||||
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
|
||||
ssize_t offset = gridOffset + (bid*tree->nHeads+tree->headRank)*chunkSize;
|
||||
ssize_t offset = gridOffset + (bid*direct->nHeads+direct->headRank)*chunkSize;
|
||||
int nelem = min(chunkSize, size-offset);
|
||||
prims.send(offset, nelem);
|
||||
}
|
||||
@@ -650,29 +650,29 @@ struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_COLLNET, NCCL_PROTO
|
||||
// Gather
|
||||
int group = (0*Proto::MaxGroupWidth) | (0<<16);
|
||||
Primitives<T, RedOp, FanAsymmetric<NCCL_MAX_DIRECT_ARITY, 0>, /*Direct=*/0, Proto, 0>
|
||||
prims(tid, nThreadsGather, tree->up, NULL, args->sendbuff, args->recvbuff, args->redOpArg, group, args);
|
||||
prims(tid, nThreadsGather, direct->up, NULL, args->sendbuff, args->recvbuff, args->redOpArg, group, args);
|
||||
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
|
||||
ssize_t offset = gridOffset + bid*tree->nHeads*chunkSize;
|
||||
int nelem = min(tree->nHeads*chunkSize, size-offset);
|
||||
prims.directGather(offset, nelem, chunkSize, tree->headRank, tree->shift);
|
||||
ssize_t offset = gridOffset + bid*direct->nHeads*chunkSize;
|
||||
int nelem = min(direct->nHeads*chunkSize, size-offset);
|
||||
prims.directGather(offset, nelem, chunkSize, direct->headRank, direct->shift);
|
||||
}
|
||||
} else if (tid >= tidStartBcast && tid < tidStartScatter && tree->out != -1) {
|
||||
} else if (tid >= tidStartBcast && tid < tidStartScatter && direct->out != -1) {
|
||||
int group = (1*Proto::MaxGroupWidth) | (0<<16);
|
||||
if (hasDn) {
|
||||
// Recv from network, broadcast
|
||||
Primitives<T, RedOp, FanAsymmetric<1, NCCL_MAX_DIRECT_ARITY>, /*Direct=*/0, Proto, 0>
|
||||
prims(tid-tidStartBcast, nThreadsBcast, &tree->out, tree->down, args->sendbuff, args->recvbuff, args->redOpArg, group, args);
|
||||
prims(tid-tidStartBcast, nThreadsBcast, &direct->out, direct->down, args->sendbuff, args->recvbuff, args->redOpArg, group, args);
|
||||
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
|
||||
ssize_t offset = gridOffset + (bid*tree->nHeads+tree->headRank)*chunkSize;
|
||||
ssize_t offset = gridOffset + (bid*direct->nHeads+direct->headRank)*chunkSize;
|
||||
int nelem = min(chunkSize, size-offset);
|
||||
prims.recvCopyDirectSend(offset, offset, nelem, /*postOp=*/true);
|
||||
}
|
||||
} else {
|
||||
// Recv from network (no post thread needed)
|
||||
Primitives<T, RedOp, FanAsymmetric<1, 0>, /*Direct=*/0, Proto, 0>
|
||||
prims(tid-tidStartBcast, nThreadsBcast, &tree->out, nullptr, args->sendbuff, args->recvbuff, args->redOpArg, group);
|
||||
prims(tid-tidStartBcast, nThreadsBcast, &direct->out, nullptr, args->sendbuff, args->recvbuff, args->redOpArg, group);
|
||||
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
|
||||
ssize_t offset = gridOffset + (bid*tree->nHeads+tree->headRank)*chunkSize;
|
||||
ssize_t offset = gridOffset + (bid*direct->nHeads+direct->headRank)*chunkSize;
|
||||
int nelem = min(chunkSize, size-offset);
|
||||
prims.recv(offset, nelem, /*postOp=*/true);
|
||||
}
|
||||
@@ -681,6 +681,73 @@ struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_COLLNET, NCCL_PROTO
|
||||
}
|
||||
};
|
||||
|
||||
template<typename T, typename RedOp>
|
||||
struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_COLLNET_CHAIN, NCCL_PROTO_SIMPLE> {
|
||||
__device__ __forceinline__ void run(ncclWorkElem *args) {
|
||||
const int tid = threadIdx.x;
|
||||
const int nthreads = args->nWarps*WARP_SIZE;
|
||||
const int bid = args->bid;
|
||||
const int nChannels = args->nChannels;
|
||||
ncclTree *tree = &ncclShmem.channel.collnetChain;
|
||||
ssize_t chunkSize = int(args->lastChunkSize);
|
||||
const ssize_t loopSize = int(nChannels*chunkSize);
|
||||
const ssize_t size = args->count;
|
||||
|
||||
int nthreadsSplit = nthreads/2;
|
||||
if (nthreadsSplit >= 256) nthreadsSplit += 64;
|
||||
|
||||
int group, send, recv, groupTid, groupNthreads;
|
||||
using Proto = ProtoSimple<1, 1>;
|
||||
if (tid < nthreadsSplit) {
|
||||
group = (0*Proto::MaxGroupWidth) | (1<<16);
|
||||
recv = tree->down[0];
|
||||
send = tree->up;
|
||||
groupTid = tid;
|
||||
groupNthreads = nthreadsSplit;
|
||||
} else {
|
||||
group = (1*Proto::MaxGroupWidth);
|
||||
recv = tree->up;
|
||||
send = tree->down[0];
|
||||
groupTid = tid - nthreadsSplit;
|
||||
groupNthreads = nthreads-nthreadsSplit;
|
||||
}
|
||||
|
||||
Primitives<T, RedOp, FanSymmetric<1>, /*Direct=*/1, Proto, 0>
|
||||
prims(groupTid, groupNthreads, &recv, &send, args->sendbuff, args->recvbuff, args->redOpArg, group);
|
||||
|
||||
if (tid < nthreadsSplit) {
|
||||
if (recv == -1) {
|
||||
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
|
||||
ssize_t offset = gridOffset + bid*int(chunkSize);
|
||||
int nelem = min(chunkSize, size-offset);
|
||||
prims.send(offset, nelem);
|
||||
}
|
||||
} else {
|
||||
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
|
||||
ssize_t offset = gridOffset + bid*int(chunkSize);
|
||||
int nelem = min(chunkSize, size-offset);
|
||||
prims.recvReduceSend(offset, nelem);
|
||||
}
|
||||
}
|
||||
}
|
||||
else {
|
||||
if (send == -1) {
|
||||
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
|
||||
ssize_t offset = gridOffset + bid*int(chunkSize);
|
||||
int nelem = min(chunkSize, size-offset);
|
||||
prims.directRecv(offset, nelem);
|
||||
}
|
||||
} else {
|
||||
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
|
||||
ssize_t offset = gridOffset + bid*int(chunkSize);
|
||||
int nelem = min(chunkSize, size-offset);
|
||||
prims.directRecvCopySend(offset, offset, nelem);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
template<typename T, typename RedOp>
|
||||
struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_LL> {
|
||||
__device__ __forceinline__ void run(ncclWorkElem *args) {
|
||||
|
||||
@@ -35,7 +35,8 @@
|
||||
#define NCCL_FUNC4(func, devredop, type, nullify) \
|
||||
NCCL_FUNC5(func, TREE, devredop, type, nullify), \
|
||||
NCCL_FUNC5(func, RING, devredop, type, nullify), \
|
||||
NCCL_FUNC5(func, COLLNET, devredop, type, nullify)
|
||||
NCCL_FUNC5(func, COLLNET_DIRECT, devredop, type, nullify), \
|
||||
NCCL_FUNC5(func, COLLNET_CHAIN, devredop, type, nullify)
|
||||
|
||||
// Must be consistent with ncclDataType_t
|
||||
#define NCCL_FUNCS3A(func, devredop, nullForFloat) \
|
||||
@@ -120,7 +121,8 @@ static const __device__ constexpr ncclKernelFunc_t ncclFuncs[]{
|
||||
#define NCCL_FUNC4_LL128(func, devredop, type, nullify) \
|
||||
NCCL_FUNC5_LL128(func, TREE, devredop, type, nullify), \
|
||||
NCCL_FUNC5_LL128(func, RING, devredop, type, nullify), \
|
||||
NCCL_FUNC5_LL128(func, COLLNET, devredop, type, nullify)
|
||||
NCCL_FUNC5_LL128(func, COLLNET_DIRECT, devredop, type, nullify), \
|
||||
NCCL_FUNC5_LL128(func, COLLNET_CHAIN, devredop, type, nullify)
|
||||
|
||||
// Must be consistent with ncclDataType_t
|
||||
#define NCCL_FUNCS3A_LL128(func, devredop, nullForFloat) \
|
||||
@@ -214,8 +216,8 @@ struct Caller<f, f + 1, u>{
|
||||
void call(unsigned short funcIndex) noexcept { if (u) ncclFuncs_ll128[f](); else ncclFuncs[f](); }
|
||||
};
|
||||
|
||||
static_assert(FUNC_INDEX_P2P == 2710, "Wrong P2P function index");
|
||||
static_assert(FUNC_INDEX_ALLTOALL_PIVOT == 2711, "Wrong AllToAllPivot function index");
|
||||
static_assert(FUNC_INDEX_P2P == 3610, "Wrong P2P function index");
|
||||
static_assert(FUNC_INDEX_ALLTOALL_PIVOT == 3611, "Wrong AllToAllPivot function index");
|
||||
|
||||
template<bool USING_LL128>
|
||||
__forceinline__
|
||||
@@ -238,49 +240,65 @@ void NCCL_CALL_FUNCTIONS(unsigned short funcIndex) noexcept {
|
||||
ncclFunction_AllReduce_TREE_LL128_Sum_float();
|
||||
else if (!USING_LL128 && funcIndex == FUNC_INDEX(ncclFuncAllReduce, ncclSum, ncclFloat32, NCCL_ALGO_TREE, NCCL_PROTO_LL128))
|
||||
ncclFunction_AllReduce_TREE_LL_Sum_float();
|
||||
else if (funcIndex == FUNC_INDEX(ncclFuncAllReduce, ncclSum, ncclFloat32, NCCL_ALGO_COLLNET, NCCL_PROTO_SIMPLE))
|
||||
ncclFunction_AllReduce_COLLNET_SIMPLE_Sum_float();
|
||||
else if (funcIndex == FUNC_INDEX(ncclFuncAllReduce, ncclSum, ncclFloat32, NCCL_ALGO_COLLNET, NCCL_PROTO_LL))
|
||||
ncclFunction_AllReduce_COLLNET_LL_Sum_float();
|
||||
else if (USING_LL128 && funcIndex == FUNC_INDEX(ncclFuncAllReduce, ncclSum, ncclFloat32, NCCL_ALGO_COLLNET, NCCL_PROTO_LL128))
|
||||
ncclFunction_AllReduce_COLLNET_LL128_Sum_float();
|
||||
else if (!USING_LL128 && funcIndex == FUNC_INDEX(ncclFuncAllReduce, ncclSum, ncclFloat32, NCCL_ALGO_COLLNET, NCCL_PROTO_LL128))
|
||||
ncclFunction_AllReduce_COLLNET_LL_Sum_float();
|
||||
else if (funcIndex == FUNC_INDEX(ncclFuncAllReduce, ncclSum, ncclFloat32, NCCL_ALGO_COLLNET_DIRECT, NCCL_PROTO_SIMPLE))
|
||||
ncclFunction_AllReduce_COLLNET_DIRECT_SIMPLE_Sum_float();
|
||||
else if (funcIndex == FUNC_INDEX(ncclFuncAllReduce, ncclSum, ncclFloat32, NCCL_ALGO_COLLNET_DIRECT, NCCL_PROTO_LL))
|
||||
ncclFunction_AllReduce_COLLNET_DIRECT_LL_Sum_float();
|
||||
else if (USING_LL128 && funcIndex == FUNC_INDEX(ncclFuncAllReduce, ncclSum, ncclFloat32, NCCL_ALGO_COLLNET_DIRECT, NCCL_PROTO_LL128))
|
||||
ncclFunction_AllReduce_COLLNET_DIRECT_LL128_Sum_float();
|
||||
else if (!USING_LL128 && funcIndex == FUNC_INDEX(ncclFuncAllReduce, ncclSum, ncclFloat32, NCCL_ALGO_COLLNET_DIRECT, NCCL_PROTO_LL128))
|
||||
ncclFunction_AllReduce_COLLNET_DIRECT_LL_Sum_float();
|
||||
else if (funcIndex == FUNC_INDEX(ncclFuncAllReduce, ncclSum, ncclFloat32, NCCL_ALGO_COLLNET_CHAIN, NCCL_PROTO_SIMPLE))
|
||||
ncclFunction_AllReduce_COLLNET_CHAIN_SIMPLE_Sum_float();
|
||||
else if (funcIndex == FUNC_INDEX(ncclFuncAllReduce, ncclSum, ncclFloat32, NCCL_ALGO_COLLNET_CHAIN, NCCL_PROTO_LL))
|
||||
ncclFunction_AllReduce_COLLNET_CHAIN_LL_Sum_float();
|
||||
else if (USING_LL128 && funcIndex == FUNC_INDEX(ncclFuncAllReduce, ncclSum, ncclFloat32, NCCL_ALGO_COLLNET_CHAIN, NCCL_PROTO_LL128))
|
||||
ncclFunction_AllReduce_COLLNET_CHAIN_LL128_Sum_float();
|
||||
else if (!USING_LL128 && funcIndex == FUNC_INDEX(ncclFuncAllReduce, ncclSum, ncclFloat32, NCCL_ALGO_COLLNET_CHAIN, NCCL_PROTO_LL128))
|
||||
ncclFunction_AllReduce_COLLNET_CHAIN_LL_Sum_float();
|
||||
else
|
||||
assert("Unsupported function index");
|
||||
#else
|
||||
if (funcIndex < 540) {
|
||||
if (funcIndex % 9 == 0) ncclFunction_Broadcast_TREE_LL_Sum_int8_t();
|
||||
else if (USING_LL128 && funcIndex % 9 == 1) ncclFunction_Broadcast_TREE_LL128_Sum_int8_t();
|
||||
else if (!USING_LL128 && funcIndex % 9 == 1) ncclFunction_Broadcast_TREE_LL_Sum_int8_t();
|
||||
else if (funcIndex % 9 == 2) ncclFunction_Broadcast_TREE_SIMPLE_Sum_int8_t();
|
||||
else if (funcIndex % 9 == 3) ncclFunction_Broadcast_RING_LL_Sum_int8_t();
|
||||
else if (USING_LL128 && funcIndex % 9 == 4) ncclFunction_Broadcast_RING_LL128_Sum_int8_t();
|
||||
else if (!USING_LL128 && funcIndex % 9 == 4) ncclFunction_Broadcast_RING_LL_Sum_int8_t();
|
||||
else if (funcIndex % 9 == 5) ncclFunction_Broadcast_RING_SIMPLE_Sum_int8_t();
|
||||
else if (funcIndex % 9 == 6) ncclFunction_Broadcast_COLLNET_LL_Sum_int8_t();
|
||||
else if (USING_LL128 && funcIndex % 9 == 7) ncclFunction_Broadcast_COLLNET_LL128_Sum_int8_t();
|
||||
else if (!USING_LL128 && funcIndex % 9 == 7) ncclFunction_Broadcast_COLLNET_LL_Sum_int8_t();
|
||||
else ncclFunction_Broadcast_COLLNET_SIMPLE_Sum_int8_t();
|
||||
if (funcIndex < 720) {
|
||||
if (funcIndex % 12 == 0) ncclFunction_Broadcast_TREE_LL_Sum_int8_t();
|
||||
else if (USING_LL128 && funcIndex % 12 == 1) ncclFunction_Broadcast_TREE_LL128_Sum_int8_t();
|
||||
else if (!USING_LL128 && funcIndex % 12 == 1) ncclFunction_Broadcast_TREE_LL_Sum_int8_t();
|
||||
else if (funcIndex % 12 == 2) ncclFunction_Broadcast_TREE_SIMPLE_Sum_int8_t();
|
||||
else if (funcIndex % 12 == 3) ncclFunction_Broadcast_RING_LL_Sum_int8_t();
|
||||
else if (USING_LL128 && funcIndex % 12 == 4) ncclFunction_Broadcast_RING_LL128_Sum_int8_t();
|
||||
else if (!USING_LL128 && funcIndex % 12 == 4) ncclFunction_Broadcast_RING_LL_Sum_int8_t();
|
||||
else if (funcIndex % 12 == 5) ncclFunction_Broadcast_RING_SIMPLE_Sum_int8_t();
|
||||
else if (funcIndex % 12 == 6) ncclFunction_Broadcast_COLLNET_DIRECT_LL_Sum_int8_t();
|
||||
else if (USING_LL128 && funcIndex % 12 == 7) ncclFunction_Broadcast_COLLNET_DIRECT_LL128_Sum_int8_t();
|
||||
else if (!USING_LL128 && funcIndex % 12 == 7) ncclFunction_Broadcast_COLLNET_DIRECT_LL_Sum_int8_t();
|
||||
else if (funcIndex % 12 == 8) ncclFunction_Broadcast_COLLNET_DIRECT_SIMPLE_Sum_int8_t();
|
||||
else if (funcIndex % 12 == 9) ncclFunction_Broadcast_COLLNET_CHAIN_LL_Sum_int8_t();
|
||||
else if (USING_LL128 && funcIndex % 12 == 10) ncclFunction_Broadcast_COLLNET_CHAIN_LL128_Sum_int8_t();
|
||||
else if (!USING_LL128 && funcIndex % 12 == 10) ncclFunction_Broadcast_COLLNET_CHAIN_LL_Sum_int8_t();
|
||||
else ncclFunction_Broadcast_COLLNET_CHAIN_SIMPLE_Sum_int8_t();
|
||||
}
|
||||
else if (funcIndex < 1080) Caller<540, 1080, USING_LL128>::call(funcIndex);
|
||||
else if (funcIndex < 1620) {
|
||||
if (funcIndex % 9 == 0) ncclFunction_AllGather_TREE_LL_Sum_int8_t();
|
||||
else if (USING_LL128 && funcIndex % 9 == 1) ncclFunction_AllGather_TREE_LL128_Sum_int8_t();
|
||||
else if (!USING_LL128 && funcIndex % 9 == 1) ncclFunction_AllGather_TREE_LL_Sum_int8_t();
|
||||
else if (funcIndex % 9 == 2) ncclFunction_AllGather_TREE_SIMPLE_Sum_int8_t();
|
||||
else if (funcIndex % 9 == 3) ncclFunction_AllGather_RING_LL_Sum_int8_t();
|
||||
else if (USING_LL128 && funcIndex % 9 == 4) ncclFunction_AllGather_RING_LL128_Sum_int8_t();
|
||||
else if (!USING_LL128 && funcIndex % 9 == 4) ncclFunction_AllGather_RING_LL_Sum_int8_t();
|
||||
else if (funcIndex % 9 == 5) ncclFunction_AllGather_RING_SIMPLE_Sum_int8_t();
|
||||
else if (funcIndex % 9 == 6) ncclFunction_AllGather_COLLNET_LL_Sum_int8_t();
|
||||
else if (USING_LL128 && funcIndex % 9 == 7) ncclFunction_AllGather_COLLNET_LL128_Sum_int8_t();
|
||||
else if (!USING_LL128 && funcIndex % 9 == 7) ncclFunction_AllGather_COLLNET_LL_Sum_int8_t();
|
||||
else ncclFunction_AllGather_COLLNET_SIMPLE_Sum_int8_t();
|
||||
else if (funcIndex < 1440) Caller<720, 1440, USING_LL128>::call(funcIndex);
|
||||
else if (funcIndex < 2160) {
|
||||
if (funcIndex % 12 == 0) ncclFunction_AllGather_TREE_LL_Sum_int8_t();
|
||||
else if (USING_LL128 && funcIndex % 12 == 1) ncclFunction_AllGather_TREE_LL128_Sum_int8_t();
|
||||
else if (!USING_LL128 && funcIndex % 12 == 1) ncclFunction_AllGather_TREE_LL_Sum_int8_t();
|
||||
else if (funcIndex % 12 == 2) ncclFunction_AllGather_TREE_SIMPLE_Sum_int8_t();
|
||||
else if (funcIndex % 12 == 3) ncclFunction_AllGather_RING_LL_Sum_int8_t();
|
||||
else if (USING_LL128 && funcIndex % 12 == 4) ncclFunction_AllGather_RING_LL128_Sum_int8_t();
|
||||
else if (!USING_LL128 && funcIndex % 12 == 4) ncclFunction_AllGather_RING_LL_Sum_int8_t();
|
||||
else if (funcIndex % 12 == 5) ncclFunction_AllGather_RING_SIMPLE_Sum_int8_t();
|
||||
else if (funcIndex % 12 == 6) ncclFunction_AllGather_COLLNET_DIRECT_LL_Sum_int8_t();
|
||||
else if (USING_LL128 && funcIndex % 12 == 7) ncclFunction_AllGather_COLLNET_DIRECT_LL128_Sum_int8_t();
|
||||
else if (!USING_LL128 && funcIndex % 12 == 7) ncclFunction_AllGather_COLLNET_DIRECT_LL_Sum_int8_t();
|
||||
else if (funcIndex % 12 == 8) ncclFunction_AllGather_COLLNET_DIRECT_SIMPLE_Sum_int8_t();
|
||||
else if (funcIndex % 12 == 9) ncclFunction_AllGather_COLLNET_CHAIN_LL_Sum_int8_t();
|
||||
else if (USING_LL128 && funcIndex % 12 == 10) ncclFunction_AllGather_COLLNET_CHAIN_LL128_Sum_int8_t();
|
||||
else if (!USING_LL128 && funcIndex % 12 == 10) ncclFunction_AllGather_COLLNET_CHAIN_LL_Sum_int8_t();
|
||||
else ncclFunction_AllGather_COLLNET_CHAIN_SIMPLE_Sum_int8_t();
|
||||
}
|
||||
else if (funcIndex < 2700) Caller<1620, 2700, USING_LL128>::call(funcIndex);
|
||||
else if (funcIndex < 3600) Caller<2160, 3600, USING_LL128>::call(funcIndex);
|
||||
else {
|
||||
switch (funcIndex - 2700) {
|
||||
switch (funcIndex - 3600) {
|
||||
case 0:
|
||||
ncclFunction_OneRankReduce_PreMulSum_int8_t();
|
||||
break;
|
||||
@@ -707,7 +725,8 @@ __device__ __attribute__((noinline)) void NCCL_FUNC_NAME(func, algo, proto, dev
|
||||
#define IMPL_COLL3(func, devredop, type, ncclType) \
|
||||
IMPL_COLL4(func, TREE, devredop, type, ncclType) \
|
||||
IMPL_COLL4(func, RING, devredop, type, ncclType) \
|
||||
IMPL_COLL4(func, COLLNET, devredop, type, ncclType)
|
||||
IMPL_COLL4(func, COLLNET_DIRECT, devredop, type, ncclType) \
|
||||
IMPL_COLL4(func, COLLNET_CHAIN, devredop, type, ncclType)
|
||||
|
||||
#define IMPL_COLL2(func, devredop) \
|
||||
IMPL_COLL3(func, devredop, int8_t, ncclInt8) \
|
||||
|
||||
@@ -21,7 +21,8 @@ __shared__ ncclShmemData ncclShmem;
|
||||
#define NCCL_FUNC4(func, devredop, type, nullify) \
|
||||
NCCL_FUNC5(func, TREE, devredop, type, nullify), \
|
||||
NCCL_FUNC5(func, RING, devredop, type, nullify), \
|
||||
NCCL_FUNC5(func, COLLNET, devredop, type, nullify)
|
||||
NCCL_FUNC5(func, COLLNET_DIRECT, devredop, type, nullify), \
|
||||
NCCL_FUNC5(func, COLLNET_CHAIN, devredop, type, nullify)
|
||||
|
||||
#if defined(__CUDA_BF16_TYPES_EXIST__)
|
||||
// Must be consistent with ncclDataType_t
|
||||
|
||||
@@ -21,10 +21,18 @@ for base in sendrecv all_reduce all_gather broadcast reduce reduce_scatter; do
|
||||
dtn=0
|
||||
# Order must match that of the ncclDataType_t enum
|
||||
for dt in ${datatypes}; do
|
||||
echo "${dir}/${base}_${op}_${dt}.o : ${base}.cu ${dir}/${base}.dep"
|
||||
# Generate a unique filename for each compilation unit,
|
||||
# otherwise the __nv_module_id may conflict at link time
|
||||
echo "${dir}/${base}_${op}_${dt}.cu : ${base}.cu"
|
||||
echo " @printf \"Copying %-35s > %s\\\\n\" \$< \$@"
|
||||
echo " cp \$< \$@"
|
||||
echo ""
|
||||
# Compile the file
|
||||
echo "${dir}/${base}_${op}_${dt}.o : ${dir}/${base}_${op}_${dt}.cu ${base}.cu ${dir}/${base}.dep"
|
||||
|
||||
echo " @printf \"Compiling %-35s > %s\\\\n\" ${base}.cu ${dir}/${base}_${op}_${dt}.o"
|
||||
echo " mkdir -p ${dir}"
|
||||
echo " \${NVCC} -DNCCL_OP=${opn} -DNCCL_TYPE=${dtn} \${NVCUFLAGS} -dc ${base}.cu -o ${dir}/${base}_${op}_${dt}.o"
|
||||
echo " \${NVCC} -DNCCL_OP=${opn} -DNCCL_TYPE=${dtn} \${NVCUFLAGS} -dc \$< -o \$@"
|
||||
echo ""
|
||||
targets="$targets\t${dir}/${base}_${op}_${dt}.o \\\\\n"
|
||||
dtn=$(($dtn + 1))
|
||||
|
||||
@@ -13,7 +13,10 @@ template<typename T, typename RedOp, typename Fan, int Direct, int P2p>
|
||||
class Primitives<T, RedOp, Fan, Direct, ProtoLL, P2p>:
|
||||
public PrimitivesWithoutDirect<Primitives<T, RedOp, Fan, Direct, ProtoLL, P2p>> {
|
||||
|
||||
static constexpr int MaxRecv = Fan::MaxRecv, MaxSend = Fan::MaxSend;
|
||||
// In the case of Fan::MaxRecv == 0, we need to force MaxRecv to 1 for this to compile
|
||||
// This is because of a recv buffer which is allocated to MaxRecv length in send-only cases
|
||||
static constexpr int MaxRecv = Fan::MaxRecv > 1 ? Fan::MaxRecv : 1;
|
||||
static constexpr int MaxSend = Fan::MaxSend;
|
||||
static constexpr int Input=0, Output=1;
|
||||
RedOp redOp;
|
||||
const int tid;
|
||||
@@ -495,18 +498,19 @@ private:
|
||||
redOp(redOpArg),
|
||||
tid(tid), nthreads(nthreads), wid(tid%WARP_SIZE), group(group&(uint16_t)0xFFFF),
|
||||
stepLines(ncclShmem.comm.buffSizes[NCCL_PROTO_LL]/NCCL_STEPS/sizeof(ncclLLFifoLine)) {
|
||||
int connIndex = group >> 16;
|
||||
auto *channel = &ncclShmem.channel;
|
||||
barriers = &ncclShmem.groups[this->group].barrier;
|
||||
barrier_next = ncclShmem.groups[this->group].barrier_next;
|
||||
|
||||
auto *channel = &ncclShmem.channel;
|
||||
// If we are going to support oneshot collNet + LL, then we would need to add connector index here
|
||||
int nrecv=0, nsend=0;
|
||||
while (nrecv < MaxRecv && recvPeers[nrecv] >= 0) {
|
||||
loadRecvConn(&channel->peers[recvPeers[nrecv]].recv[0], nrecv);
|
||||
// We compare with Fan::MaxRecv here because this->MaxRecv is always at least 1
|
||||
while (nrecv < Fan::MaxRecv && recvPeers[nrecv] >= 0) {
|
||||
loadRecvConn(&channel->peers[recvPeers[nrecv]].recv[connIndex], nrecv);
|
||||
nrecv++;
|
||||
}
|
||||
while (nsend < MaxSend && sendPeers[nsend] >= 0) {
|
||||
loadSendConn(&channel->peers[sendPeers[nsend]].send[0], nsend);
|
||||
loadSendConn(&channel->peers[sendPeers[nsend]].send[connIndex], nsend);
|
||||
nsend++;
|
||||
}
|
||||
this->fan = Fan(nrecv, nsend);
|
||||
|
||||
@@ -236,7 +236,8 @@ private:
|
||||
vr[u+1] = __builtin_nontemporal_load(ptr+u*WARP_SIZE+1);
|
||||
needReload |= flagThread && (vr[u+1] != flag);
|
||||
}
|
||||
} while (__any(needReload) && checkAbort(spins, 0, 0) == 0);
|
||||
needReload &= (0 == checkAbort(spins, 0, 0));
|
||||
} while (__any(needReload));
|
||||
}
|
||||
|
||||
/************* Finish register load **************/
|
||||
@@ -278,7 +279,8 @@ private:
|
||||
vr[u+1] = __builtin_nontemporal_load(ptr+u*WARP_SIZE+1);
|
||||
needReload |= flagThread && (vr[u+1] != flag);
|
||||
}
|
||||
} while (__any(needReload) && checkAbort(spins, i, 0) == 0);
|
||||
needReload &= (0 == checkAbort(spins, i, 0));
|
||||
} while (__any(needReload));
|
||||
|
||||
#pragma unroll
|
||||
for (int u=0; u<ELEMS_PER_THREAD; u+=2) {
|
||||
@@ -409,17 +411,17 @@ public:
|
||||
tid(tid), nthreads(nthreads), wid(tid%WARP_SIZE), warp(tid/WARP_SIZE),
|
||||
flagThread((tid%4)==3), group(group&(uint16_t)0xFFFF),
|
||||
stepSize(ncclShmem.comm.buffSizes[NCCL_PROTO_LL128]/NCCL_STEPS/sizeof(uint64_t)) {
|
||||
int connIndex = group >> 16;
|
||||
auto *channel = &ncclShmem.channel;
|
||||
barriers = &ncclShmem.groups[this->group].barrier;
|
||||
barrier_next = ncclShmem.groups[this->group].barrier_next;
|
||||
|
||||
auto *channel = &ncclShmem.channel;
|
||||
int nrecv=0, nsend=0;
|
||||
while (nrecv < MaxRecv && recvPeers[nrecv] >= 0) {
|
||||
loadRecvConn(&channel->peers[recvPeers[nrecv]].recv[0], nrecv);
|
||||
loadRecvConn(&channel->peers[recvPeers[nrecv]].recv[connIndex], nrecv);
|
||||
nrecv++;
|
||||
}
|
||||
while (nsend < MaxSend && sendPeers[nsend] >= 0) {
|
||||
loadSendConn(&channel->peers[sendPeers[nsend]].send[0], nsend);
|
||||
loadSendConn(&channel->peers[sendPeers[nsend]].send[connIndex], nsend);
|
||||
nsend++;
|
||||
}
|
||||
this->fan = Fan(nrecv, nsend);
|
||||
|
||||
@@ -14,7 +14,8 @@
|
||||
|
||||
template<typename T, typename RedOp>
|
||||
struct RunWork<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE> {
|
||||
__device__ __forceinline__ void runSend(const int tid, const int nthreads, const int group, struct ncclWorkElemP2p* args) {
|
||||
template<typename Proto>
|
||||
__device__ void runSend(const int tid, const int nthreads, const int group, struct ncclWorkElemP2p* args) {
|
||||
void* buff = reinterpret_cast<void*>(uintptr_t(args->buffHi32)<<32 | args->buffLo32);
|
||||
size_t count = reinterpret_cast<size_t>(size_t(args->countHi32)<<32 | args->countLo32);
|
||||
|
||||
@@ -75,8 +76,8 @@ struct RunWork<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE> {
|
||||
|
||||
}
|
||||
} else {
|
||||
using Proto = ProtoSimple<1, 1>;
|
||||
int const chunkSize = args->chunkSize/sizeof(T);
|
||||
int chunkSize = args->chunkSize/sizeof(T);
|
||||
if (args->proto == NCCL_PROTO_LL) chunkSize /= 2;
|
||||
int const peer = args->peer;
|
||||
Primitives<T, RedOp, FanAsymmetric<0, 1>, 0, Proto, 1> prims
|
||||
(tid, nthreads, nullptr, &peer, buff, nullptr, /*redOpArg(ignored)=*/0, group);
|
||||
@@ -112,7 +113,8 @@ struct RunWork<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE> {
|
||||
}
|
||||
}
|
||||
|
||||
__device__ __forceinline__ void runRecv(const int tid, const int nthreads, const int group, struct ncclWorkElemP2p* args) {
|
||||
template<typename Proto>
|
||||
__device__ void runRecv(const int tid, const int nthreads, const int group, struct ncclWorkElemP2p* args) {
|
||||
#if defined(ENABLE_NPKIT)
|
||||
bool isNpKitThread = (tid == 0);
|
||||
int npKitCtxIdx = blockIdx.x * NCCL_MAX_WORK_ELEMENTS_P2P + 1;
|
||||
@@ -134,10 +136,10 @@ struct RunWork<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE> {
|
||||
#endif
|
||||
|
||||
if (args->peer != ncclShmem.comm.rank) {
|
||||
using Proto = ProtoSimple<1, 1>;
|
||||
void* buff = reinterpret_cast<void*>(uintptr_t(args->buffHi32)<<32 | args->buffLo32);
|
||||
ssize_t count = reinterpret_cast<size_t>(size_t(args->countHi32)<<32 | args->countLo32);
|
||||
int const chunkSize = args->chunkSize/sizeof(T);
|
||||
int chunkSize = args->chunkSize/sizeof(T);
|
||||
if (args->proto == NCCL_PROTO_LL) chunkSize /= 2; // This is to account for chunkEffectiveSize
|
||||
int const peer = args->peer;
|
||||
Primitives<T, RedOp, FanAsymmetric<1, 0>, 0, Proto, 1> prims
|
||||
(tid, nthreads, &peer, nullptr, nullptr, buff, /*redOpArg(ignored)=*/0, group);
|
||||
@@ -191,10 +193,21 @@ struct RunWork<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE> {
|
||||
|
||||
if (args->p2pType == ncclWorkP2pTypeUnused) return;
|
||||
if (tid >= nthreads || args->peer == -1) return;
|
||||
|
||||
// Select Proto here
|
||||
// This is to allow the same kernel to run multiple primitives on different warps (thread groups)
|
||||
if ((group%2) == 0) {
|
||||
runRecv(tid, nthreads, group, args);
|
||||
if (args->proto == NCCL_PROTO_LL) {
|
||||
runRecv<ProtoLL>(tid, nthreads, group, args);
|
||||
} else {
|
||||
runRecv<ProtoSimple<1,1>>(tid, nthreads, group, args);
|
||||
}
|
||||
} else {
|
||||
runSend(tid, nthreads, group, args);
|
||||
if (args->proto == NCCL_PROTO_LL) {
|
||||
runSend<ProtoLL>(tid, nthreads, group, args);
|
||||
} else {
|
||||
runSend<ProtoSimple<1,1>>(tid, nthreads, group, args);
|
||||
}
|
||||
}
|
||||
}
|
||||
};
|
||||
|
||||
Ссылка в новой задаче
Block a user