path: root/nerv/lib/matrix/cumatrix.c

#define NERV_GENERIC_CUMATRIX
#define MATRIX_CONTEXT CuContext
#include <string.h>
#include <time.h>
#include "../common.h"
#include "cumatrix.h"
#include "cuda_helper.h"

void nerv_cuda_context_print_profile(CuContext *context) {
    HashMap *profile = context->profile;
    size_t i;
    float tmp, tot = 0;
    fprintf(stderr, "*** [nerv cumatrix profile] **\n");
    for (i = 0; i < profile->size; i++)
    {
        HashNode *ptr;
        for (ptr = profile->bucket[i]; ptr; ptr = ptr->next)
        {
            tmp = *(float *)ptr->val;
            fprintf(stderr, "%s:\t%.6f\n", ptr->key, tmp);
            tot += tmp;
        }
    }
    fprintf(stderr, "Total time:\t%.6f\n", tot);
}

void nerv_cuda_context_clear_profile(CuContext *context) {
    nerv_hashmap_clear(context->profile);
}

void nerv_cuda_context_accu_profile(CuContext *context,
                                    const char *name, float delta) {
    HashMap *profile = context->profile;
    float *val = nerv_hashmap_getval(profile, name);
    if (!val)
    {
        val = malloc(sizeof(float));
        *val = 0;
        nerv_hashmap_setval(profile, name, val);
    }
    *val += delta;
}

static void new_cuda_handles(CuContext *context, int dev, Status *status) {
    if (context->has_handle) return;
    CUDA_SAFE_SYNC_CALL(cudaSetDevice(dev), status);
    CUBLAS_SAFE_SYNC_CALL(cublasCreate(&(context->cublas_handle)), status);
    CURAND_SAFE_SYNC_CALL(curandCreateGenerator(&(context->curand_gen),
                            CURAND_RNG_PSEUDO_DEFAULT), status);
    CURAND_SAFE_SYNC_CALL(
            curandSetPseudoRandomGeneratorSeed(context->curand_gen, time(NULL)),
            status);
    CUDA_SAFE_SYNC_CALL(cudaEventCreate(&(context->profile_start)), status);
    CUDA_SAFE_SYNC_CALL(cudaEventCreate(&(context->profile_stop)), status);
    NERV_SET_STATUS(status, NERV_NORMAL, 0);
    context->has_handle = 1;
}

static void free_cuda_handles(CuContext *context, Status *status) {
    if (!context->has_handle) return;
    context->has_handle = 0;
    CUBLAS_SAFE_SYNC_CALL(cublasDestroy(context->cublas_handle), status);
    CURAND_SAFE_SYNC_CALL(curandDestroyGenerator(context->curand_gen), status);
    CUDA_SAFE_SYNC_CALL(cudaEventDestroy(context->profile_start), status);
    CUDA_SAFE_SYNC_CALL(cudaEventDestroy(context->profile_stop), status);
    NERV_SET_STATUS(status, NERV_NORMAL, 0);
}

static int choose_best_gpu(Status *status) {
    int i, n, dev = 0;
    float best_ratio = 0;
    fprintf(stderr, "*** select a GPU based on available space\n");
    CUDA_SAFE_CALL_RET(cudaGetDeviceCount(&n), status);
    for (i = 0; i < n; i++)
    {
        size_t avail, total;
        float ratio;
        CUDA_SAFE_SYNC_CALL_RET(cudaSetDevice(i), status);
        CUDA_SAFE_SYNC_CALL_RET(cuMemGetInfo(&avail, &total), status);
        ratio = (float)avail/total * 100;
        fprintf(stderr, "* card %d: %.2f%%\n", i, ratio);
        if (ratio > best_ratio)
        {
            best_ratio = ratio;
            dev = i;
        }
        CUDA_SAFE_SYNC_CALL_RET(cudaDeviceReset(), status);
    }
    fprintf(stderr, "*** final decision: GPU %d\n", dev);
    NERV_SET_STATUS(status, NERV_NORMAL, 0);
    return dev;
}

CuContext *nerv_cuda_context_create(int dev, Status *status) {
    CuContext *context = (CuContext *)malloc(sizeof(CuContext));
    context->has_handle = 0; /* this line must come first */
    if (dev == -1)
    {
        dev = choose_best_gpu(status);
        if (status->err_code != NERV_NORMAL)
            return NULL;
    }
    new_cuda_handles(context, dev, status);
    if (status->err_code != NERV_NORMAL)
        return NULL;
    context->profile = nerv_hashmap_create(PROFILE_HASHMAP_SIZE, bkdr_hash, strcmp);
    NERV_SET_STATUS(status, NERV_NORMAL, 0);
    return context;
}

void nerv_cuda_context_destroy(CuContext *context, Status *status) {
    free_cuda_handles(context, status);
    if (status->err_code != NERV_NORMAL)
        return;
    nerv_hashmap_destroy(context->profile);
    free(context);
    NERV_SET_STATUS(status, NERV_NORMAL, 0);
}

void nerv_cuda_context_select_gpu(CuContext *context,
                                    int dev, Status *status) {
    /* free_cuda_handles(context, status);
    if (status->err_code != NERV_NORMAL)
        return;
    */
    /* because of cudaDeviceReset */
    context->has_handle = 0;
    CUDA_SAFE_SYNC_CALL(cudaDeviceReset(), status);
    new_cuda_handles(context, dev, status);
    if (status->err_code != NERV_NORMAL)
        return;
    NERV_SET_STATUS(status, NERV_NORMAL, 0);
}

#define MATRIX_USE_FLOAT
#define cuda_matrix_(NAME) cuda_matrix_float_##NAME
#define nerv_matrix_(NAME) nerv_matrix_cuda_float_##NAME
#define cudak_(NAME) cudak_float_ ## NAME
#define NERV_CUBLAS_(NAME) cublasS##NAME
#include "generic/cumatrix.c"

#undef NERV_CUBLAS_
#undef cudak_
#undef nerv_matrix_
#undef cuda_matrix_
#undef MATRIX_USE_FLOAT
#undef MATRIX_ELEM
#undef MATRIX_ELEM_PTR
#undef MATRIX_ELEM_PTR_BASE
#undef MATRIX_ELEM_FMT
#undef MATRIX_ELEM_WRITE_FMT

#define MATRIX_USE_DOUBLE
#define cuda_matrix_(NAME) cuda_matrix_double_##NAME
#define nerv_matrix_(NAME) nerv_matrix_cuda_double_##NAME
#define cudak_(NAME) cudak_double_ ## NAME
#define NERV_CUBLAS_(NAME) cublasD##NAME
#include "generic/cumatrix.c"