Initial commit: RCCL auto-tuning project

rccl-test/src/gather.cu

+/*************************************************************************
+ * Copyright (c) 2016-2022, NVIDIA CORPORATION. All rights reserved.
+ * Modifications Copyright (c) 2019-2022 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * See LICENSE.txt for license information
+ ************************************************************************/
+
+#include "hip/hip_runtime.h"
+#include "common.h"
+
+void GatherGetCollByteCount(size_t *sendcount, size_t *recvcount, size_t *paramcount, size_t *sendInplaceOffset, size_t *recvInplaceOffset, size_t count, int nranks) {
+  *sendcount = count/nranks;
+  *recvcount = (count/nranks)*nranks;
+  *sendInplaceOffset = count/nranks;
+  *recvInplaceOffset = 0;
+  *paramcount = count/nranks;
+}
+
+testResult_t GatherInitData(struct threadArgs* args, ncclDataType_t type, ncclRedOp_t op, int root, int rep, int in_place) {
+  size_t sendcount = args->sendBytes / wordSize(type);
+  size_t recvcount = args->expectedBytes / wordSize(type);
+  int nranks = args->nProcs*args->nThreads*args->nGpus*args->nRanks;
+
+  int k=0;
+  for (int i=0; i<args->nGpus; i++) {
+    HIPCHECK(hipSetDevice(args->gpus[i]));
+
+    for (int l=0; l<args->nRanks; l++) {
+      int rank = ((args->proc*args->nThreads + args->thread)*args->nGpus*args->nRanks + i*args->nRanks + l);
+      HIPCHECK(hipMemset(args->recvbuffs[k], 0, args->expectedBytes));
+      void* data = in_place ? ((char*)args->recvbuffs[k])+rank*args->sendBytes : args->sendbuffs[k];
+      TESTCHECK(InitData(data, sendcount, rank*sendcount, type, ncclSum, rep, 1, 0));
+      HIPCHECK(hipMemcpy(args->expected[k], args->recvbuffs[k], args->expectedBytes, hipMemcpyDefault));
+      if (rank == root) {
+	for (int j=0; j<nranks; j++) {
+	  TESTCHECK(InitData(((char*)args->expected[k]), nranks*sendcount, 0, type, ncclSum, rep, 1, 0));
+	}
+      }
+      k++;
+    }
+    HIPCHECK(hipDeviceSynchronize());
+  }
+  return testSuccess;
+}
+
+void GatherGetBw(size_t count, int typesize, double sec, double* algBw, double* busBw, int nranks) {
+  double baseBw = (double)(count * nranks * typesize) / 1.0E9 / sec;
+
+  *algBw = baseBw;
+  double factor = ((double)(nranks-1))/((double)(nranks));
+  *busBw = baseBw * factor;
+}
+
+testResult_t GatherRunColl(void* sendbuff, void* recvbuff, size_t count, ncclDataType_t type, ncclRedOp_t op, int root, ncclComm_t comm, hipStream_t stream) {
+  int nRanks;
+  NCCLCHECK(ncclCommCount(comm, &nRanks));
+  int rank;
+  NCCLCHECK(ncclCommUserRank(comm, &rank));
+  size_t rankOffset = count * wordSize(type);
+  if (count == 0) return testSuccess;
+
+  NCCLCHECK(ncclGroupStart());
+  NCCLCHECK(ncclSend(sendbuff, count, type, root, comm, stream));
+  if (rank == root) {
+    for (int r=0; r<nRanks; r++) {
+      NCCLCHECK(ncclRecv(((char*)recvbuff)+r*rankOffset, count, type, r, comm, stream));
+    }
+  }
+  NCCLCHECK(ncclGroupEnd());
+
+  return testSuccess;
+}
+
+struct testColl gatherTest = {
+  "Gather",
+  GatherGetCollByteCount,
+  GatherInitData,
+  GatherGetBw,
+  GatherRunColl
+};
+
+void GatherGetBuffSize(size_t *sendcount, size_t *recvcount, size_t count, int nranks) {
+  size_t paramcount, sendInplaceOffset, recvInplaceOffset;
+  GatherGetCollByteCount(sendcount, recvcount, &paramcount, &sendInplaceOffset, &recvInplaceOffset, count, nranks);
+}
+
+testResult_t GatherRunTest(struct threadArgs* args, int root, ncclDataType_t type, const char* typeName, ncclRedOp_t op, const char* opName) {
+  args->collTest = &gatherTest;
+  ncclDataType_t *run_types;
+  const char **run_typenames;
+  int type_count;
+  int begin_root, end_root;
+
+  if ((int)type != -1) {
+    type_count = 1;
+    run_types = &type;
+    run_typenames = &typeName;
+  } else {
+    type_count = test_typenum;
+    run_types = test_types;
+    run_typenames = test_typenames;
+  }
+
+  if (root != -1) {
+    begin_root = end_root = root;
+  } else {
+    begin_root = 0;
+    end_root = args->nProcs*args->nThreads*args->nGpus-1;
+  }
+
+  for (int i=0; i<type_count; i++) {
+    for (int j=begin_root; j<=end_root; j++) {
+      TESTCHECK(TimeTest(args, run_types[i], run_typenames[i], (ncclRedOp_t)0, "none", j));
+    }
+  }
+  return testSuccess;
+}
+
+struct testEngine ncclTestEngine = {
+  GatherGetBuffSize,
+  GatherRunTest
+};

rccl-test/src/hypercube.cu

+/*************************************************************************
+ * Copyright (c) 2016-2022, NVIDIA CORPORATION. All rights reserved.
+ * Modifications Copyright (c) 2022 Advanced Micro Devices, Inc. All rights reserved.
+ *
+ * See LICENSE.txt for license information
+ ************************************************************************/
+
+#include "hip/hip_runtime.h"
+#include "common.h"
+
+#define ALIGN 4
+
+void HyperCubeGetCollByteCount(size_t *sendcount, size_t *recvcount, size_t *paramcount, size_t *sendInplaceOffset, size_t *recvInplaceOffset, size_t count, int nranks) {
+  size_t base = (count/(ALIGN*nranks))*ALIGN;
+  *sendcount = base;
+  *recvcount = base*nranks;
+  *sendInplaceOffset = base;
+  *recvInplaceOffset = 0;
+  *paramcount = base;
+}
+
+testResult_t HyperCubeInitData(struct threadArgs* args, ncclDataType_t type, ncclRedOp_t op, int root, int rep, int in_place) {
+  size_t sendcount = args->sendBytes / wordSize(type);
+  size_t recvcount = args->expectedBytes / wordSize(type);
+  int nranks = args->nProcs*args->nThreads*args->nGpus*args->nRanks;
+
+  int k=0;
+  for (int i=0; i<args->nGpus; i++) {
+    HIPCHECK(hipSetDevice(args->gpus[i]));
+
+    for (int l=0; l<args->nRanks; l++) {
+      int rank = ((args->proc*args->nThreads + args->thread)*args->nGpus*args->nRanks + i*args->nRanks + l);
+      HIPCHECK(hipMemset(args->recvbuffs[k], 0, args->expectedBytes));
+      void* data = in_place ? ((char*)args->recvbuffs[k])+rank*args->sendBytes : args->sendbuffs[k];
+      TESTCHECK(InitData(data, sendcount, 0, type, ncclSum, 33*rep + rank, 1, 0));
+      for (int j=0; j<nranks; j++) {
+	TESTCHECK(InitData(((char*)args->expected[k])+args->sendBytes*j, sendcount, 0, type, ncclSum, 33*rep + j, 1, 0));
+      }
+      k++;
+    }
+    HIPCHECK(hipDeviceSynchronize());
+  }
+  return testSuccess;
+}
+
+void HyperCubeGetBw(size_t count, int typesize, double sec, double* algBw, double* busBw, int nranks) {
+  double baseBw = (double)(count * typesize * (nranks - 1)) / 1.0E9 / sec;
+
+  *algBw = baseBw;
+  double factor = 1;
+  *busBw = baseBw * factor;
+}
+
+testResult_t HyperCubeRunColl(void* sendbuff, void* recvbuff, size_t count, ncclDataType_t type, ncclRedOp_t op, int root, ncclComm_t comm, hipStream_t stream) {
+  char* sbuff = (char*)sendbuff;
+  char* rbuff = (char*)recvbuff;
+  int nRanks;
+  NCCLCHECK(ncclCommCount(comm, &nRanks));
+  int rank;
+  NCCLCHECK(ncclCommUserRank(comm, &rank));
+  size_t rankSize = count * wordSize(type);
+  if (rbuff+rank*rankSize != sbuff) HIPCHECK(hipMemcpyAsync(rbuff+rank*rankSize, sbuff, rankSize, hipMemcpyDeviceToDevice, stream));
+
+  // Hypercube AllGather
+  for (int mask=1; mask<nRanks; mask<<=1) {
+    NCCLCHECK(ncclGroupStart());
+    int s = rank & ~(mask-1);
+    int r = s ^ mask;
+    NCCLCHECK(ncclSend(rbuff+s*rankSize, count*mask, type, rank^mask, comm, stream));
+    NCCLCHECK(ncclRecv(rbuff+r*rankSize, count*mask, type, rank^mask, comm, stream));
+    NCCLCHECK(ncclGroupEnd());
+  }
+  return testSuccess;
+}
+
+struct testColl hyperCubeTest = {
+  "HyperCube",
+  HyperCubeGetCollByteCount,
+  HyperCubeInitData,
+  HyperCubeGetBw,
+  HyperCubeRunColl
+};
+
+void HyperCubeGetBuffSize(size_t *sendcount, size_t *recvcount, size_t count, int nranks) {
+  size_t paramcount, sendInplaceOffset, recvInplaceOffset;
+  HyperCubeGetCollByteCount(sendcount, recvcount, &paramcount, &sendInplaceOffset, &recvInplaceOffset, count, nranks);
+}
+
+testResult_t HyperCubeRunTest(struct threadArgs* args, int root, ncclDataType_t type, const char* typeName, ncclRedOp_t op, const char* opName) {
+  args->collTest = &hyperCubeTest;
+  ncclDataType_t *run_types;
+  const char **run_typenames;
+  int type_count;
+
+  if ((int)type != -1) {
+    type_count = 1;
+    run_types = &type;
+    run_typenames = &typeName;
+  } else {
+    type_count = test_typenum;
+    run_types = test_types;
+    run_typenames = test_typenames;
+  }
+
+  // Check if this is a power of 2
+  int nRanks = args->nProcs*args->nThreads*args->nGpus;
+  if (nRanks && !(nRanks & (nRanks - 1))) {
+    for (int i=0; i<type_count; i++) {
+      TESTCHECK(TimeTest(args, run_types[i], run_typenames[i], (ncclRedOp_t)0, "", -1));
+    }
+  } else {
+    printf("nRanks %d is not a power of 2, skipping\n", nRanks);
+  }
+
+  return testSuccess;
+}
+
+struct testEngine ncclTestEngine = {
+  HyperCubeGetBuffSize,
+  HyperCubeRunTest
+};

rccl-test/src/timer.cc

+#include "timer.h"
+
+// Make sure to compile this translation unit with the host compiler and not
+// nvcc, lest you hit an internal compiler error (ICE) with GCC 10.3.0
+#include <chrono>
+
+namespace {
+  std::uint64_t now() {
+    using clock = std::chrono::steady_clock;
+    return std::chrono::duration_cast<std::chrono::nanoseconds>(clock::now().time_since_epoch()).count();
+  }
+}
+
+timer::timer() {
+  t0 = now();
+}
+
+double timer::elapsed() const {
+  std::uint64_t t1 = now();
+  return 1.e-9*(t1 - t0);
+}
+
+double timer::reset() {
+  std::uint64_t t1 = now();
+  double ans = 1.e-9*(t1 - t0);
+  t0 = t1;
+  return ans;
+}

rccl-test/src/timer.h

+#ifndef _408319ecdd5b47b28bf8f511c4fdf816
+#define _408319ecdd5b47b28bf8f511c4fdf816
+
+#include <cstdint>
+
+// Can't include <chrono> because of bug with gcc 10.3.0
+class timer {
+  std::uint64_t t0;
+public:
+  timer();
+  double elapsed() const;
+  double reset();
+};
+
+#endif