[Do not merge] Overlap benchmark: AG+GEMM distributed matmul with HostIr and ParallelType::Stream

bench/process_outputs

@@ -0,0 +1,7 @@
+#!/bin/bash
+
+FILE="/opt/pytorch/Fuser/bench/logs/${1}/info"
+
+cat $FILE | grep "rank 0: "  #| awk '{print $4}'
+
+# | grep -E 'Streams32\b'

bench/test

@@ -0,0 +1,87 @@
+#!/bin/bash
+EXPERIMENT=StreamParallelType_tests
+DATE=$(date +%Y%m%d-%H%M)
+LOG_BASE="/opt/pytorch/Fuser/bench/logs"
+
+NP=8
+BACKEND=UCC
+M=32768
+K=32768
+N=1024
+
+S=8
+Streams=3
+Pgs=1
+
+# M=131072 #32768
+# K=131072
+# N=32768 #1024
+# L=1048576  #268435456 #67108864 #131072
+# PRE_COMM="_pre_comm"
+# POST_COMM="_post_comm"
+# UNFUSE="_unfused"
+# GRAPH="_WithCudaGraph"
+# cuStreamWrite=WithcuStreamWriteValue32_
+# GTEST_PREFIX="OverlapBenchmark.PipelinedAGMatmulBenchmark/"
+# GTEST_PREFIX="DummyOverlapBenchmark.PipelinedAGMatmulBenchmark/"
+GTEST_PREFIX="OverlapBenchmark.PipelinedAGMatmulBenchmarkStreamParallelType/"
+GTEST_POSTFIX="${BACKEND}_S${S}_M${M}_K${K}_N${N}_Streams${Streams}_${cuStreamWrite}Pgs${Pgs}${UNFUSE}${GRAPH}"
+# GTEST_POSTFIX="${BACKEND}_M${M}_K${K}_N${N}_L${L}${PRE_COMM}${POST_COMM}"
+export GTEST_FILTER="${GTEST_PREFIX}${GTEST_POSTFIX}"
+echo "gtest filter: $GTEST_FILTER" | tee -a $LOG_FILE_INFO
+
+MPIFLAGS=" -np $NP"
+
+# MPIFLAGS+=" -x NCCL_P2P_NET_CHUNKSIZE=2MB"
+# MPIFLAGS+=" -x NCCL_DEBUG=TRACE" #INFO
+# MPIFLAGS+=" -x NCCL_MAX_NCHANNELS=1"
+
+MPIFLAGS+=" -x UCC_CL_BASIC_TLS=nccl"
+# MPIFLAGS+=" -x UCC_TL_NCCL_SYNC=event"
+
+# MPIFLAGS+=" -x UCC_CL_BASIC_TLS=cuda"
+# MPIFLAGS+=" -x UCC_TL_CUDA_SCRATCH_SIZE=32mb"
+# MPIFLAGS+=" -x UCC_TL_CUDA_ALLGATHER_RING_MAX_RINGS=32"
+# MPIFLAGS+=" -x UCC_TL_CUDA_ALLGATHER_RING_NUM_CHUNKS=32"
+
+# MPIFLAGS+=" -x UCC_EC_CUDA_EXEC_NUM_WORKERS=8"
+# MPIFLAGS+=" -x UCC_EC_CUDA_USE_COOPERATIVE_LAUNCH=0"
+# MPIFLAGS+=" -x UCC_EC_CUDA_STREAM_TASK_MODE=driver"
+# MPIFLAGS+=" -x UCC_EC_CUDA_STREAM_TASK_MODE=kernel"
+# MPIFLAGS+=" -x UCC_EC_CUDA_EXEC_COPY_LARGE_THRESH=1M"
+# MPIFLAGS+=" -x UCC_EC_CUDA_EXEC_NUM_THREADS=512"
+
+# MPIFLAGS+=" -x UCC_CL_BASIC_TLS=ucp"
+# MPIFLAGS+=" -x UCX_RNDV_THRESH=0 -x UCX_TLS=ib,cuda_copy"
+# MPIFLAGS+=" -x UCX_RNDV_SCHEME=put_zcopy"
+# MPIFLAGS+=" -x UCX_RNDV_SCHEME=get_zcopy"
+
+
+MPIFLAGS+=" -x UCX_NET_DEVICES=mlx5_0:1"
+# MPIFLAGS+=" -x UCC_CL_BASIC_TLS=^sharp,mlx5"
+MPIFLAGS+=" -x UCC_COLL_TRACE=info"
+# MPIFLAGS+=" -x UCC_LOG_LEVEL=debug"
+# MPIFLAGS+=" -x TORCH_NCCL_AVOID_RECORD_STREAMS=1"
+# MPIFLAGS+=" -x CUDA_DEVICE_MAX_CONNECTIONS=2"
+
+
+export LOGS="${LOG_BASE}/${EXPERIMENT}_${BACKEND}_${DATE}"
+mkdir -p $LOGS
+export LOG_FILE_INFO="${LOGS}/info.txt"
+echo "Writing to $LOG_FILE_INFO" | tee -a $LOG_FILE_INFO
+
+echo "mpi flags: $MPIFLAGS" | tee -a $LOG_FILE_INFO
+
+TEST_CMD="$BUILD_DIRECTORY/test_multidevice --gtest_filter=${GTEST_FILTER}"
+echo "test cmd: $TEST_CMD" | tee -a $LOG_FILE_INFO
+
+MPICMD="mpirun $MPIFLAGS $TEST_CMD"
+echo $MPICMD | tee -a $LOG_FILE_INFO
+
+NSYSCMD="nsys profile --stats=false -w true -t cublas,cuda,nvtx,osrt,mpi,ucx -o ${LOGS}/${GTEST_POSTFIX} --capture-range-end stop --capture-range=cudaProfilerApi --cudabacktrace=memory,sync,kernel,other"
+
+CMD="${NSYSCMD} ${MPICMD}"
+sudo /bin/sh -c "echo '1' > /proc/sys/kernel/perf_event_paranoid"
+echo $CMD | tee -a ${LOG_FILE_INFO}
+$CMD | tee -a ${LOG_FILE_INFO}
+

csrc/driver_api.h

@@ -32,6 +32,7 @@ namespace nvfuser {
   fn(cuModuleGetFunction);                \
   fn(cuModuleLoadDataEx);                 \
   fn(cuModuleUnload);                     \
+  fn(cuStreamWriteValue32);               \
   fn(cuOccupancyMaxActiveBlocksPerMultiprocessor)
 
 #if (CUDA_VERSION >= 12000)