Resync all tests with test code from NCCL 2.4
Major rework to merge most of the changes from the NCCL internal tests into the public ones Added "-m <agg_iters>" operation aggregation option. Data integrity checking is now much more performant at scale. Startup times at scale are improved. Test latency units are now displayed in usec.
This commit is contained in:
David Addison
+104
-26
@@ -1,8 +1,10 @@
|
||||
/*************************************************************************
|
||||
* Copyright (c) 2016-2017, NVIDIA CORPORATION. All rights reserved.
|
||||
* Copyright (c) 2016-2019, NVIDIA CORPORATION. All rights reserved.
|
||||
*
|
||||
* See LICENCE.txt for license information
|
||||
* See LICENSE.txt for license information
|
||||
************************************************************************/
|
||||
#ifndef __COMMON_H__
|
||||
#define __COMMON_H__
|
||||
|
||||
#include "nccl.h"
|
||||
#include <stdio.h>
|
||||
@@ -17,23 +19,75 @@
|
||||
#define CUDACHECK(cmd) do { \
|
||||
cudaError_t e = cmd; \
|
||||
if( e != cudaSuccess ) { \
|
||||
printf("Cuda failure %s:%d '%s'\n", \
|
||||
char hostname[1024]; \
|
||||
getHostName(hostname, 1024); \
|
||||
printf("%s: Test CUDA failure %s:%d '%s'\n", \
|
||||
hostname, \
|
||||
__FILE__,__LINE__,cudaGetErrorString(e)); \
|
||||
exit(EXIT_FAILURE); \
|
||||
return testCudaError; \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
#define NCCLCHECK(cmd) do { \
|
||||
ncclResult_t r = cmd; \
|
||||
if (r!= ncclSuccess) { \
|
||||
printf("NCCL failure %s:%d '%s'\n", \
|
||||
char hostname[1024]; \
|
||||
getHostName(hostname, 1024); \
|
||||
printf("%s: Test NCCL failure %s:%d '%s'\n", \
|
||||
hostname, \
|
||||
__FILE__,__LINE__,ncclGetErrorString(r)); \
|
||||
exit(EXIT_FAILURE); \
|
||||
return testNcclError; \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
struct threadArgs_t {
|
||||
void *proc_args;
|
||||
typedef enum {
|
||||
testSuccess = 0,
|
||||
testInternalError = 1,
|
||||
testCudaError = 2,
|
||||
testNcclError = 3,
|
||||
testCuRandError = 4
|
||||
} testResult_t;
|
||||
|
||||
// Relay errors up and trace
|
||||
#define TESTCHECK(cmd) do { \
|
||||
testResult_t r = cmd; \
|
||||
if (r!= testSuccess) { \
|
||||
char hostname[1024]; \
|
||||
getHostName(hostname, 1024); \
|
||||
printf(" .. %s: Test failure %s:%d\n", \
|
||||
hostname, \
|
||||
__FILE__,__LINE__); \
|
||||
return r; \
|
||||
} \
|
||||
} while(0)
|
||||
|
||||
struct testColl {
|
||||
const char name[20];
|
||||
void (*getCollByteCount)(
|
||||
size_t *sendcount, size_t *recvcount, size_t *paramcount,
|
||||
size_t *sendInplaceOffset, size_t *recvInplaceOffset,
|
||||
size_t count, int nranks);
|
||||
testResult_t (*initData)(struct threadArgs* args, ncclDataType_t type,
|
||||
ncclRedOp_t op, int root, int rep, int in_place);
|
||||
void (*getBw)(size_t count, int typesize, double sec, double* algBw, double* busBw, int nranks);
|
||||
testResult_t (*runColl)(void* sendbuff, void* recvbuff, size_t count, ncclDataType_t type,
|
||||
ncclRedOp_t op, int root, ncclComm_t comm, cudaStream_t stream);
|
||||
};
|
||||
extern struct testColl allReduceTest;
|
||||
extern struct testColl allGatherTest;
|
||||
extern struct testColl reduceScatterTest;
|
||||
extern struct testColl broadcastTest;
|
||||
extern struct testColl reduceTest;
|
||||
|
||||
struct testEngine {
|
||||
void (*getBuffSize)(size_t *sendcount, size_t *recvcount, size_t count, int nranks);
|
||||
testResult_t (*runTest)(struct threadArgs* args, int root, ncclDataType_t type,
|
||||
const char* typeName, ncclRedOp_t op, const char* opName);
|
||||
};
|
||||
|
||||
extern struct testEngine ncclTestEngine;
|
||||
|
||||
struct threadArgs {
|
||||
size_t nbytes;
|
||||
size_t minbytes;
|
||||
size_t maxbytes;
|
||||
@@ -55,11 +109,8 @@ struct threadArgs_t {
|
||||
ncclComm_t* comms;
|
||||
cudaStream_t* streams;
|
||||
|
||||
void** expectedHost;
|
||||
void** expected;
|
||||
size_t expectedBytes;
|
||||
void* procSharedHost;
|
||||
void* procShared;
|
||||
volatile int* sync;
|
||||
int sync_idx;
|
||||
volatile int* barrier;
|
||||
@@ -72,27 +123,28 @@ struct threadArgs_t {
|
||||
int* errors;
|
||||
double* bw;
|
||||
int* bw_count;
|
||||
|
||||
struct testColl* collTest;
|
||||
};
|
||||
|
||||
typedef testResult_t (*threadFunc_t)(struct threadArgs* args);
|
||||
struct testThread {
|
||||
pthread_t thread;
|
||||
threadFunc_t func;
|
||||
struct threadArgs args;
|
||||
testResult_t ret;
|
||||
};
|
||||
|
||||
#include <chrono>
|
||||
|
||||
// Provided by common.cu
|
||||
extern void Barrier(struct threadArgs_t* args);
|
||||
extern void TimeTest(struct threadArgs_t* args, ncclDataType_t type, const char* typeName, ncclRedOp_t op, const char* opName, int root, int inPlace);
|
||||
extern void Randomize(void* ptr, size_t count, ncclDataType_t type, int seed);
|
||||
extern void Accumulate(void* out, void* in, size_t n, ncclDataType_t type, ncclRedOp_t op);
|
||||
extern void CheckDelta(void* expected, void* results, size_t count, ncclDataType_t type, double* devmax);
|
||||
extern double DeltaMaxValue(ncclDataType_t type);
|
||||
extern void Barrier(struct threadArgs* args);
|
||||
extern testResult_t TimeTest(struct threadArgs* args, ncclDataType_t type, const char* typeName, ncclRedOp_t op, const char* opName, int root);
|
||||
extern testResult_t InitDataReduce(void* data, const size_t count, const size_t offset, ncclDataType_t type, ncclRedOp_t op, const int rep, const int nranks);
|
||||
extern testResult_t InitData(void* data, const size_t count, ncclDataType_t type, const int rep, const int rank);
|
||||
extern void AllocateBuffs(void **sendbuff, void **recvbuff, void **expected, void **expectedHost, size_t nbytes, int nranks);
|
||||
|
||||
// Provided by each coll
|
||||
void RunTest(struct threadArgs_t* args, int root, ncclDataType_t type, const char* typeName, ncclRedOp_t op, const char* opName);
|
||||
extern void GetBw(size_t count, int typeSize, double sec, double* algBw, double* busBw, int nranks);
|
||||
extern void RunColl(void* sendbuf, void* recvbuff, size_t count, ncclDataType_t type, ncclRedOp_t op, int root, ncclComm_t comm, cudaStream_t stream);
|
||||
extern void InitData(struct threadArgs_t* args, ncclDataType_t type, ncclRedOp_t op, int in_place, int is_first);
|
||||
extern double CheckData(struct threadArgs_t* args, ncclDataType_t type, ncclRedOp_t op);
|
||||
extern void AllocateBuffs(void **sendbuff, void **recvbuff, void **expected, void **expectedHost, size_t nbytes, int nranks);
|
||||
extern void InitRecvResult(struct threadArgs_t* args, ncclDataType_t type, ncclRedOp_t op, int root, int in_place, int is_first);
|
||||
extern void getCollByteCount(size_t *sendbytes, size_t *recvbytes, size_t *parambytes, size_t *sendInlineOffset, size_t *recvInlineOffset, size_t *procSharedBytes, int *sameexpected, size_t nbytes, int nranks);
|
||||
extern void print_line_header (size_t size, size_t count, const char *typeName, const char *opName, int root);
|
||||
extern void print_header();
|
||||
|
||||
@@ -152,7 +204,33 @@ extern const char *test_typenames[ncclNumTypes];
|
||||
extern ncclRedOp_t test_ops[ncclNumOps];
|
||||
extern const char *test_opnames[ncclNumOps];
|
||||
|
||||
static int ncclstringtotype(char *str) {
|
||||
for (int t=0; t<ncclNumTypes; t++) {
|
||||
if (strcmp(str, test_typenames[t]) == 0) {
|
||||
return t;
|
||||
}
|
||||
}
|
||||
if (strcmp(str, "all") == 0) {
|
||||
return -1;
|
||||
}
|
||||
printf("invalid type %s, defaulting to %s .. \n", str, test_typenames[ncclFloat]);
|
||||
return ncclFloat;
|
||||
}
|
||||
|
||||
static int ncclstringtoop (char *str) {
|
||||
for (int o=0; o<ncclNumOps; o++) {
|
||||
if (strcmp(str, test_opnames[o]) == 0) {
|
||||
return o;
|
||||
}
|
||||
}
|
||||
if (strcmp(str, "all") == 0) {
|
||||
return -1;
|
||||
}
|
||||
printf("invalid op %s, defaulting to %s .. \n", str, test_opnames[ncclSum]);
|
||||
return ncclSum;
|
||||
}
|
||||
|
||||
extern thread_local int is_main_thread;
|
||||
#define PRINT if (is_main_thread) printf
|
||||
|
||||
|
||||
#endif
|
||||
|
||||
Reference in New Issue
Block a user