samples/nvlink_bandwidth/nvlink_bandwidth.cu

/*
* Copyright 2015-2016 NVIDIA Corporation. All rights reserved.
*
* Sample to demonstrate use of NVlink CUPTI APIs
*/

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <cuda.h>
#include <cupti.h>

#ifndef EXIT_WAIVED
#define EXIT_WAIVED 2
#endif

#define CUPTI_CALL(call)                                                      \
do {                                                                          \
    CUptiResult _status = call;                                               \
    if (_status != CUPTI_SUCCESS) {                                           \
        const char *errstr;                                                   \
        cuptiGetResultString(_status, &errstr);                               \
        fprintf(stderr, "%s:%d: error: function %s failed with error %s.\n",  \
                __FILE__, __LINE__, #call, errstr);                           \
        exit(EXIT_FAILURE);                                                   \
    }                                                                         \
} while (0)

#define DRIVER_API_CALL(apiFuncCall)                                           \
do {                                                                           \
    CUresult _status = apiFuncCall;                                            \
    if (_status != CUDA_SUCCESS) {                                             \
        fprintf(stderr, "%s:%d: error: function %s failed with error %d.\n",   \
                __FILE__, __LINE__, #apiFuncCall, _status);                    \
        exit(EXIT_FAILURE);                                                    \
    }                                                                          \
} while (0)

#define RUNTIME_API_CALL(apiFuncCall)                                          \
do {                                                                           \
    cudaError_t _status = apiFuncCall;                                         \
    if (_status != cudaSuccess) {                                              \
        fprintf(stderr, "%s:%d: error: function %s failed with error %s.\n",   \
                __FILE__, __LINE__, #apiFuncCall, cudaGetErrorString(_status));\
        exit(EXIT_FAILURE);                                                     \
    }                                                                          \
} while (0)

#define MEMORY_ALLOCATION_CALL(var)                                            \
do {                                                                           \
    if (var == NULL) {                                                         \
        fprintf(stderr, "%s:%d: Error: Memory Allocation Failed \n",           \
                __FILE__, __LINE__);                                           \
        exit(EXIT_FAILURE);                                                     \
    }                                                                          \
} while (0)

#define MAX_DEVICES    (32)
#define BLOCK_SIZE     (1024)
#define GRID_SIZE      (512)
#define DATA           (1024 * 1024)
#define BUF_SIZE       (32 * 1024)
#define ALIGN_SIZE     (8)
#define SUCCESS        (0)
#define NUM_METRIC     (4)
#define NUM_EVENTS     (2)
#define MAX_SIZE       (64*1024*1024)   // 64 MB
#define NUM_STREAMS    (6)   // gp100 has 6 physical copy engines
#define DATA           (1024 * 1024)

CUpti_ActivityNvLink4 *nvlinkRec = NULL;
int cpuToGpu = 0;
int gpuToGpu = 0;
int cpuToGpuAccess = 0;
int gpuToGpuAccess = 0;
bool metricSupport = true;

extern "C" __global__ void test_nvlink_bandwidth(float *src, float *dst)
{
    int idx = blockIdx.x * blockDim.x + threadIdx.x;
    dst[idx] = src[idx] * 2.0f;
}

static void printActivity(CUpti_Activity *record)
{
    uint32_t i = 0;
    if (record->kind == CUPTI_ACTIVITY_KIND_NVLINK) {
        nvlinkRec = (CUpti_ActivityNvLink4 *)record;

        printf("typeDev0 %d, typeDev1 %d, sysmem %d, peer %d, physical links %d, ",
          nvlinkRec->typeDev0,
          nvlinkRec->typeDev1,
          ((nvlinkRec->flag & CUPTI_LINK_FLAG_SYSMEM_ACCESS) ? 1 : 0),
          ((nvlinkRec->flag & CUPTI_LINK_FLAG_PEER_ACCESS) ? 1 : 0),
          nvlinkRec->physicalNvLinkCount);

        printf("portDev0 ");
        for (i = 0 ; i < nvlinkRec->physicalNvLinkCount ; i++ ) {
            printf("%d, ", nvlinkRec->portDev0[i]);
        }

        printf("portDev1 ");
        for (i = 0 ; i < nvlinkRec->physicalNvLinkCount ; i++ ) {
            printf("%d, ", nvlinkRec->portDev1[i]);
        }

        printf("bandwidth %llu\n", (long long unsigned int)nvlinkRec->bandwidth);

        cpuToGpuAccess |= (nvlinkRec->flag & CUPTI_LINK_FLAG_SYSMEM_ACCESS);
        gpuToGpuAccess |= (nvlinkRec->flag & CUPTI_LINK_FLAG_PEER_ACCESS);
    }
    else {
        printf("Error : Unexpected CUPTI activity kind.\nExpected Activity kind : CUPTI_ACTIVITY_KIND_NVLINK\n");
    }
}

static void CUPTIAPI bufferRequested(uint8_t **buffer, size_t *size, size_t *maxNumRecords)

{
    *size = BUF_SIZE + ALIGN_SIZE;
    *buffer = (uint8_t*) calloc(1, *size);
    MEMORY_ALLOCATION_CALL(*buffer);
    *maxNumRecords = 0;
}

static void CUPTIAPI bufferCompleted(CUcontext ctx, uint32_t streamId,
                                uint8_t *buffer, size_t size,
                                size_t validSize)
{
    CUptiResult status;
    CUpti_Activity *record = NULL;
    do {
        status = cuptiActivityGetNextRecord(buffer, validSize, &record);
        if(status == CUPTI_SUCCESS) {
            printActivity(record);
        }
        else if (status == CUPTI_ERROR_MAX_LIMIT_REACHED) {
            break;
        }
        else {
            CUPTI_CALL(status);
        }
    } while (1);

    size_t dropped;
    CUPTI_CALL(cuptiActivityGetNumDroppedRecords(ctx, streamId, &dropped));
    if (dropped != 0) {
        printf("Dropped %u activity records\n", (unsigned int)dropped);
    }
}

#define DIM(x) (sizeof(x)/sizeof(*(x)))

void calculateSize(char *result, uint64_t size)
{
    int i;

    const char *sizes[]   = { "TB", "GB", "MB", "KB", "B" };
    uint64_t  exbibytes = 1024ULL * 1024ULL * 1024ULL * 1024ULL;

    uint64_t  multiplier = exbibytes;

    for (i = 0; (unsigned)i < DIM(sizes); i++, multiplier /= (uint64_t)1024)
    {
        if (size < multiplier)
            continue;
        sprintf(result, "%.1f %s", (float) size / multiplier, sizes[i]);
        return;
    }
    strcpy(result, "0");
    return;
}

void readMetricValue(CUpti_EventGroup eventGroup, uint32_t numEvents,
                    CUdevice dev, CUpti_MetricID *metricId,
                    uint64_t timeDuration,
                    CUpti_MetricValue *metricValue) {

    size_t bufferSizeBytes, numCountersRead;
    uint64_t *eventValueArray = NULL;
    CUpti_EventID *eventIdArray;
    size_t arraySizeBytes = 0;
    size_t numTotalInstancesSize = 0;
    uint64_t numTotalInstances = 0;
    uint64_t *aggrEventValueArray = NULL;
    size_t aggrEventValueArraySize;
    uint32_t i = 0, j = 0;
    CUpti_EventDomainID domainId;
    size_t domainSize;

    domainSize = sizeof(CUpti_EventDomainID);

    CUPTI_CALL(cuptiEventGroupGetAttribute(eventGroup,
                                           CUPTI_EVENT_GROUP_ATTR_EVENT_DOMAIN_ID,
                                           &domainSize,
                                           (void *)&domainId));

    numTotalInstancesSize = sizeof(uint64_t);

    CUPTI_CALL(cuptiDeviceGetEventDomainAttribute(dev,
                                              domainId,
                                              CUPTI_EVENT_DOMAIN_ATTR_TOTAL_INSTANCE_COUNT,
                                              &numTotalInstancesSize,
                                              (void *)&numTotalInstances));

    arraySizeBytes = sizeof(CUpti_EventID) * numEvents;
    bufferSizeBytes = sizeof(uint64_t) * numEvents * numTotalInstances;

    eventValueArray = (uint64_t *) malloc(bufferSizeBytes);
    MEMORY_ALLOCATION_CALL(eventValueArray);

    eventIdArray = (CUpti_EventID *) malloc(arraySizeBytes);
    MEMORY_ALLOCATION_CALL(eventIdArray);

    aggrEventValueArray = (uint64_t *) calloc(numEvents, sizeof(uint64_t));
    MEMORY_ALLOCATION_CALL(aggrEventValueArray);

    aggrEventValueArraySize = sizeof(uint64_t) * numEvents;

    CUPTI_CALL(cuptiEventGroupReadAllEvents(eventGroup,
                                                CUPTI_EVENT_READ_FLAG_NONE,
                                                &bufferSizeBytes,
                                                eventValueArray,
                                                &arraySizeBytes,
                                                eventIdArray,
                                                &numCountersRead));

    for (i = 0; i < numEvents; i++) {
        for (j = 0; j < numTotalInstances; j++) {
            aggrEventValueArray[i] += eventValueArray[i + numEvents * j];
        }
    }

    for (i = 0; i < NUM_METRIC; i++) {
        CUPTI_CALL(cuptiMetricGetValue(dev, metricId[i], arraySizeBytes,
                              eventIdArray, aggrEventValueArraySize,
                              aggrEventValueArray, timeDuration,
                              &metricValue[i]));
    }

    free(eventValueArray);
    free(eventIdArray);
}

  // Print metric value, we format based on the value kind
int printMetricValue(CUpti_MetricID metricId, CUpti_MetricValue metricValue, const char *metricName) {

    CUpti_MetricValueKind valueKind;
    char str[64];
    size_t valueKindSize = sizeof(valueKind);

    CUPTI_CALL(cuptiMetricGetAttribute(metricId, CUPTI_METRIC_ATTR_VALUE_KIND,
                                       &valueKindSize, &valueKind));
    switch (valueKind) {

    case CUPTI_METRIC_VALUE_KIND_DOUBLE:
        printf("%s = ", metricName);
        calculateSize(str, (uint64_t)metricValue.metricValueDouble);
        printf("%s\n", str);
        break;

    case CUPTI_METRIC_VALUE_KIND_UINT64:
        printf("%s = ", metricName);
        calculateSize(str, (uint64_t)metricValue.metricValueUint64);
        printf("%s\n", str);
        break;

    case CUPTI_METRIC_VALUE_KIND_INT64:
        printf("%s = ", metricName);
        calculateSize(str, (uint64_t)metricValue.metricValueInt64);
        printf("%s\n", str);
        break;

    case CUPTI_METRIC_VALUE_KIND_THROUGHPUT:
        printf("%s = ", metricName);
        calculateSize(str, (uint64_t)metricValue.metricValueThroughput);
        printf("%s/Sec\n", str);
        break;

    default:
        fprintf(stderr, "error: unknown value kind\n");
        return -1;
    }
    return 0;
  }

void testCpuToGpu(CUpti_EventGroup *eventGroup, CUdeviceptr *pDevBuffer,
                    float** pHostBuffer, size_t bufferSize,
                    cudaStream_t *cudaStreams,
                    uint64_t *timeDuration, int numEventGroup)
{
    int i;
    uint32_t value = 1;
    uint64_t startTimestamp, endTimestamp;

    for (i = 0; i < numEventGroup; i++) {
            CUPTI_CALL(cuptiEventGroupEnable(eventGroup[i]));
            CUPTI_CALL(cuptiEventGroupSetAttribute(eventGroup[i],
                                CUPTI_EVENT_GROUP_ATTR_PROFILE_ALL_DOMAIN_INSTANCES,
                                sizeof(uint32_t), (void*)&value));
    }

    CUPTI_CALL(cuptiGetTimestamp(&startTimestamp));

    //Unidirectional copy H2D
    for (i = 0; i < NUM_STREAMS; i++)
    {
        RUNTIME_API_CALL(cudaMemcpyAsync((void *)pDevBuffer[i], pHostBuffer[i], bufferSize, cudaMemcpyHostToDevice, cudaStreams[i]));
    }
    RUNTIME_API_CALL(cudaDeviceSynchronize());

    //Unidirectional copy D2H
    for (i = 0; i < NUM_STREAMS; i++)
    {
        RUNTIME_API_CALL(cudaMemcpyAsync(pHostBuffer[i], (void *)pDevBuffer[i], bufferSize, cudaMemcpyDeviceToHost, cudaStreams[i]));
    }
    RUNTIME_API_CALL(cudaDeviceSynchronize());

    //Bidirectional copy
    for (i = 0; i < NUM_STREAMS; i+=2)
    {
        RUNTIME_API_CALL(cudaMemcpyAsync((void *)pDevBuffer[i], pHostBuffer[i], bufferSize, cudaMemcpyHostToDevice, cudaStreams[i]));
        RUNTIME_API_CALL(cudaMemcpyAsync(pHostBuffer[i+1], (void *)pDevBuffer[i+1], bufferSize, cudaMemcpyDeviceToHost, cudaStreams[i+1]));
    }
    RUNTIME_API_CALL(cudaDeviceSynchronize());

    CUPTI_CALL(cuptiGetTimestamp(&endTimestamp));
    *timeDuration = endTimestamp - startTimestamp;
}

void testGpuToGpu(CUpti_EventGroup *eventGroup, CUdeviceptr *pDevBuffer0, CUdeviceptr *pDevBuffer1,
                    float** pHostBuffer, size_t bufferSize,
                    cudaStream_t *cudaStreams,
                    uint64_t *timeDuration, int numEventGroup)
{
    int i;
    uint32_t value = 1;
    uint64_t startTimestamp, endTimestamp;

    for (i = 0; i < numEventGroup; i++) {
        CUPTI_CALL(cuptiEventGroupEnable(eventGroup[i]));
        CUPTI_CALL(cuptiEventGroupSetAttribute(eventGroup[i],
                            CUPTI_EVENT_GROUP_ATTR_PROFILE_ALL_DOMAIN_INSTANCES,
                            sizeof(uint32_t), (void*)&value));
    }

    RUNTIME_API_CALL(cudaSetDevice(0));
    RUNTIME_API_CALL(cudaDeviceEnablePeerAccess(1, 0));
    RUNTIME_API_CALL(cudaSetDevice(1));
    RUNTIME_API_CALL(cudaDeviceEnablePeerAccess(0, 0));

    //Unidirectional copy H2D
    for (i = 0; i < NUM_STREAMS; i++) {
        RUNTIME_API_CALL(cudaMemcpyAsync((void *)pDevBuffer0[i], pHostBuffer[i], bufferSize, cudaMemcpyHostToDevice, cudaStreams[i]));
    }
    RUNTIME_API_CALL(cudaDeviceSynchronize());

    for (i = 0; i < NUM_STREAMS; i++) {
        RUNTIME_API_CALL(cudaMemcpyAsync((void *)pDevBuffer1[i], pHostBuffer[i], bufferSize, cudaMemcpyHostToDevice, cudaStreams[i]));
    }
    RUNTIME_API_CALL(cudaDeviceSynchronize());

    CUPTI_CALL(cuptiGetTimestamp(&startTimestamp));

    for (i = 0; i < NUM_STREAMS; i++) {
        RUNTIME_API_CALL(cudaMemcpyAsync((void *)pDevBuffer0[i], (void *)pDevBuffer1[i], bufferSize, cudaMemcpyDeviceToDevice, cudaStreams[i]));
    }
    RUNTIME_API_CALL(cudaDeviceSynchronize());

    for (i = 0; i < NUM_STREAMS; i++) {
        RUNTIME_API_CALL(cudaMemcpyAsync((void *)pDevBuffer1[i], (void *)pDevBuffer0[i], bufferSize, cudaMemcpyDeviceToDevice, cudaStreams[i]));
    }
    RUNTIME_API_CALL(cudaDeviceSynchronize());

    for (i = 0; i < NUM_STREAMS; i++) {
        test_nvlink_bandwidth<<<GRID_SIZE, BLOCK_SIZE>>>((float*)pDevBuffer1[i], (float*)pDevBuffer0[i]);
    }

    CUPTI_CALL(cuptiGetTimestamp(&endTimestamp));
    *timeDuration = endTimestamp - startTimestamp;
}

static void printUsage() {
    printf("usage: Demonstrate use of NVlink CUPTI APIs\n");
    printf("       -help           : display help message\n");
    printf("       --cpu-to-gpu    : Show results for data transfer between CPU and GPU \n");
    printf("       --gpu-to-gpu    : Show results for data transfer between two GPUs \n");
}

void parseCommandLineArgs(int argc, char *argv[])
{
    if (argc != 2) {
        printf("Invalid number of options\n");
        exit(EXIT_FAILURE);
    }

    if (strcmp(argv[1], "--cpu-to-gpu") == 0) {
        cpuToGpu = 1;
    }
    else if (strcmp(argv[1], "--gpu-to-gpu") == 0) {
        gpuToGpu = 1;
    }
    else if ((strcmp(argv[1], "--help") == 0) ||
             (strcmp(argv[1], "-help") == 0) ||
             (strcmp(argv[1], "-h") == 0)) {
        printUsage();
        exit(EXIT_SUCCESS);
    }
    else {
        printf("Invalid/incomplete option %s\n", argv[1]);
        exit(EXIT_FAILURE);
    }
}

int main(int argc, char *argv[])
{
    int deviceCount = 0, i = 0, j = 0, numEventGroup = 0;
    size_t bufferSize = 0, freeMemory = 0, totalMemory = 0;
    CUpti_EventGroupSets *passes = NULL;
    CUcontext ctx;
    char str[64];

    CUdeviceptr pDevBuffer0[NUM_STREAMS];
    CUdeviceptr pDevBuffer1[NUM_STREAMS];
    float* pHostBuffer[NUM_STREAMS];

    cudaStream_t cudaStreams[NUM_STREAMS] = {0};

    CUpti_EventGroup eventGroup[32];
    CUpti_MetricID metricId[NUM_METRIC];
    uint32_t numEvents[NUM_METRIC];
    CUpti_MetricValue metricValue[NUM_METRIC];
    cudaDeviceProp prop[MAX_DEVICES];
    uint64_t timeDuration;

    // Adding nvlink Metrics.
    const char *metricName[NUM_METRIC] = {"nvlink_total_data_transmitted",
                                    "nvlink_total_data_received",
                                    "nvlink_transmit_throughput",
                                    "nvlink_receive_throughput"};

    // Parse command line arguments
    parseCommandLineArgs(argc, argv);

    CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_NVLINK));
    CUPTI_CALL(cuptiActivityRegisterCallbacks(bufferRequested, bufferCompleted));

    DRIVER_API_CALL(cuInit(0));

    RUNTIME_API_CALL(cudaGetDeviceCount(&deviceCount));

    printf("There are %d devices.\n", deviceCount);

    if (deviceCount == 0) {
        printf("There is no device supporting CUDA.\n");
        exit(EXIT_WAIVED);
    }

    for (i = 0; i < deviceCount; i++) {
        RUNTIME_API_CALL(cudaGetDeviceProperties(&prop[i], i));
        printf("CUDA Device %d Name: %s\n", i, prop[i].name);
        // Check if any device is Turing+
        if (prop[i].major == 7 && prop[i].minor > 0) {
            metricSupport = false;
        } else if (prop[i].major > 7) {
            metricSupport = false;
        }
    }

    // Set memcpy size based on available device memory
    RUNTIME_API_CALL(cudaMemGetInfo(&freeMemory, &totalMemory));
    bufferSize = MAX_SIZE < (freeMemory/4) ? MAX_SIZE : (freeMemory/4);

    printf("Total Device Memory available : ");
    calculateSize(str, (uint64_t)totalMemory);
    printf("%s\n", str);

    printf("Memcpy size is set to %llu B (%llu MB)\n",
    (unsigned long long)bufferSize, (unsigned long long)bufferSize/(1024*1024));

    for(i = 0; i < NUM_STREAMS; i++) {
       RUNTIME_API_CALL(cudaStreamCreate(&cudaStreams[i]));
    }
    RUNTIME_API_CALL(cudaDeviceSynchronize());

    // Nvlink-topology Records are generated even before cudaMemcpy API is called.
    CUPTI_CALL(cuptiActivityFlushAll(0));

    // Transfer Data between Host And Device, if Nvlink is Present
    // Check condition : nvlinkRec->flag & CUPTI_LINK_FLAG_SYSMEM_ACCESS
    // True : Nvlink is present between CPU & GPU
    // False : Nvlink is not present.
    if ((nvlinkRec) && (((cpuToGpu) && (cpuToGpuAccess)) || ((gpuToGpu) && (gpuToGpuAccess)))) {
        if (!metricSupport) {
            printf("Legacy CUPTI metrics not supported from Turing+ devices.\n");
            exit(EXIT_WAIVED);
        }

        for (i = 0; i < NUM_METRIC; i++) {
            CUPTI_CALL(cuptiMetricGetIdFromName(0, metricName[i], &metricId[i]));
            CUPTI_CALL(cuptiMetricGetNumEvents(metricId[i], &numEvents[i]));
        }

        DRIVER_API_CALL(cuCtxCreate(&ctx, 0, 0));

        CUPTI_CALL(cuptiMetricCreateEventGroupSets(ctx, (sizeof metricId) ,metricId, &passes));

        // EventGroups required to profile Nvlink metrics.
        for (i = 0; i < (signed)passes->numSets; i++) {
            for (j = 0; j < (signed)passes->sets[i].numEventGroups; j++) {
                eventGroup[numEventGroup] = passes->sets[i].eventGroups[j];

                if (!eventGroup[numEventGroup]) {
                    printf("\n eventGroup initialization failed \n");
                    exit(EXIT_FAILURE);
                }

                numEventGroup++;
            }
        }

        CUPTI_CALL(cuptiSetEventCollectionMode(ctx, CUPTI_EVENT_COLLECTION_MODE_CONTINUOUS));

        // ===== Allocate Memory =====================================

        for(i = 0; i < NUM_STREAMS; i++) {
            RUNTIME_API_CALL(cudaMalloc((void**)&pDevBuffer0[i], bufferSize));

            pHostBuffer[i] = (float *)malloc(bufferSize);
            MEMORY_ALLOCATION_CALL(pHostBuffer[i]);
        }

        if (cpuToGpu) {
            testCpuToGpu(eventGroup, pDevBuffer0, pHostBuffer, bufferSize, cudaStreams, &timeDuration, numEventGroup);
            printf("Data tranferred between CPU & Device%d : \n", (int)nvlinkRec->typeDev0);
        }
        else if(gpuToGpu) {
            RUNTIME_API_CALL(cudaSetDevice(1));
            for(i = 0; i < NUM_STREAMS; i++) {
                RUNTIME_API_CALL(cudaMalloc((void**)&pDevBuffer1[i], bufferSize));
            }
            testGpuToGpu(eventGroup, pDevBuffer0, pDevBuffer1,pHostBuffer, bufferSize, cudaStreams, &timeDuration, numEventGroup);
            printf("Data tranferred between Device 0 & Device 1 : \n");
        }

        // Collect Nvlink Metric values for the data transfer via Nvlink for all the eventGroups.
        for (i = 0; i < numEventGroup; i++) {
            readMetricValue(eventGroup[i], NUM_EVENTS, 0, metricId, timeDuration, metricValue);

            CUPTI_CALL(cuptiEventGroupDisable(eventGroup[i]));
            CUPTI_CALL(cuptiEventGroupDestroy(eventGroup[i]));

            for (i = 0; i < NUM_METRIC; i++) {
                if (printMetricValue(metricId[i], metricValue[i], metricName[i]) != 0) {
                    printf("\n printMetricValue failed \n");
                    exit(EXIT_FAILURE);
                }
            }
        }
    }
    else {
        printf("No Nvlink supported device found\n");
        exit(EXIT_WAIVED);
    }

    exit(EXIT_SUCCESS);
}