int read_csr_mat(const char *file_name, struct csr_mat_t *mat)
{
FILE *fp = fopen(file_name, "rb");
if (fp == NULL)
{
return -1;
}
fread(&mat->rows, sizeof(int), 1, fp);
fread(&mat->cols, sizeof(int), 1, fp);
fread(&mat->non_zeros, sizeof(INT64), 1, fp);
mat->row_ptr = (DWORD*)numa_alloc((mat->rows + 1) * sizeof(DWORD));
mat->col_idx = (int*)numa_alloc(mat->non_zeros * sizeof(int));
mat->vals = (FLOAT*)numa_alloc(mat->non_zeros * sizeof(FLOAT));
fread(mat->row_ptr, sizeof(DWORD), mat->rows + 1, fp);
fread(mat->col_idx, sizeof(int), mat->non_zeros, fp);
fread(mat->vals, sizeof(FLOAT), mat->non_zeros, fp);
printf("Row x Column: %d x %d\n", mat->rows, mat->cols);
printf("Non-zero elements number: %ld\n", mat->non_zeros);
return 0;
}
int csr_reorder(struct csr_mat_t *csr, struct csr_mat_t *csr_re, int *reorder_map)
{
int *row_len = (int*)malloc(csr->rows * sizeof(int));
int i, j;
for (i = 0; i < csr->rows; i++)
{
reorder_map[i] = i;
row_len[i] = csr->row_ptr[i + 1] - csr->row_ptr[i];
}
row_sort(row_len, reorder_map, csr->rows);
csr_re->rows = csr->rows;
csr_re->cols = csr->cols;
csr_re->non_zeros = csr->non_zeros;
csr_re->row_ptr = (DWORD*)numa_alloc((csr_re->rows + 1) * sizeof(DWORD));
csr_re->col_idx = (int*)numa_alloc(csr_re->non_zeros * sizeof(int));
csr_re->vals = (FLOAT*)numa_alloc(csr_re->non_zeros * sizeof(FLOAT));
int idx = 0;
csr_re->row_ptr[0] = 0;
for (i = 0; i < csr_re->rows; i++)
{
memcpy(csr_re->col_idx + idx, csr->col_idx + csr->row_ptr[reorder_map[i]], row_len[i] * sizeof(int));
memcpy(csr_re->vals + idx, csr->vals + csr->row_ptr[reorder_map[i]], row_len[i] * sizeof(FLOAT));
idx += row_len[i];
csr_re->row_ptr[i + 1] = idx;
}
free(row_len);
return 0;
}
JNIEXPORT jlong JNICALL Java_xerial_jnuma_NumaNative_allocMemory
(JNIEnv *env, jobject obj, jlong capacity) {
void* mem = numa_alloc((size_t) capacity);
if(mem == NULL)
printf("failed to allocate local memory\n");
return (jlong) mem;
}
/*
* Class: xerial_jnuma_NumaNative
* Method: alloc
* Signature: (I)Ljava/nio/ByteBuffer;
*/
JNIEXPORT jobject JNICALL Java_xerial_jnuma_NumaNative_alloc
(JNIEnv *env, jobject obj, jint capacity) {
void* mem = numa_alloc((size_t) capacity);
//printf("allocate local memory\n");
if(mem == NULL)
printf("failed to allocate local memory\n");
return (*env)->NewDirectByteBuffer(env, mem, (jlong) capacity);
}
/*
* Class: xerial_jnuma_NumaNative
* Method: allocate
* Signature: (J)J
*/
JNIEXPORT jlong JNICALL Java_xerial_jnuma_NumaNative_allocate
(JNIEnv *env, jobject obj, jlong capacity) {
void* mem = numa_alloc((size_t) capacity);
if(mem != NULL) {
return (jlong) mem;
}
throwException(env, obj, 11);
return 0L;
}
int csr_transpose(struct csr_mat_t *csr, struct csr_mat_t *csr_t)
{
csr_t->cols = csr->rows;
csr_t->rows = csr->cols;
csr_t->non_zeros = csr->non_zeros;
csr_t->row_ptr = (DWORD*)numa_alloc((csr_t->rows + 1) * sizeof(DWORD));
csr_t->col_idx = (int*)numa_alloc(csr_t->non_zeros * sizeof(int));
csr_t->vals = (FLOAT*)numa_alloc(csr_t->non_zeros * sizeof(FLOAT));
memset(csr_t->row_ptr, 0, (csr_t->rows + 1) * sizeof(DWORD));
int i, j;
for (i = 0; i < csr->rows; i++)
{
for (j = csr->row_ptr[i]; j < csr->row_ptr[i + 1]; j++)
{
csr_t->row_ptr[csr->col_idx[j] + 1]++;
}
}
for (i = 1; i <= csr_t->rows; i++)
{
csr_t->row_ptr[i] += csr_t->row_ptr[i - 1];
}
int *row_start = (int*)malloc(csr_t->rows * sizeof(int));
memcpy(row_start, csr_t->row_ptr, csr_t->rows * sizeof(int));
for (i = 0; i < csr->rows; i++)
{
for (j = csr->row_ptr[i]; j < csr->row_ptr[i + 1]; j++)
{
int row = row_start[csr->col_idx[j]];
csr_t->col_idx[row] = i;
csr_t->vals[row] = csr->vals[j];
row_start[csr->col_idx[j]]++;
}
}
free(row_start);
return 0;
}
int split_csr_lb_nz(struct csr_mat_t *csr, struct csr_cont_t *csr_cont, int count, split_dir_t dir)
{
int i, j;
csr_cont->dir = dir;
csr_cont->count = count;
csr_cont->split_idx = (int*)numa_alloc((count + 1) * sizeof(int));
csr_cont->csrs = (struct csr_mat_t*)numa_alloc(count * sizeof(struct csr_mat_t));
if (dir == SPLIT_HORIZON)
{
struct csr_mat_t *csrs = csr_cont->csrs;
int *split_idx = csr_cont->split_idx;
split_idx[0] = 0;
int avg_ele = csr->non_zeros / count, split_val;
for (i = 1, j = 1; i < count; i++)
{
split_val = i * avg_ele;
while (csr->row_ptr[j] < split_val)
{
j++;
}
if (csr->row_ptr[j] - split_val > split_val - csr->row_ptr[j - 1])
{
j--;
}
split_idx[i] = j;
}
split_idx[i] = csr->rows;
int item_idx = 0;
for (i = 0; i < count; i++)
{
csrs[i].rows = split_idx[i + 1] - split_idx[i];
printf("csrs[%d].rows = %d\n", i, csrs[i].rows);
csrs[i].cols = csr->cols;
csrs[i].non_zeros = csr->row_ptr[split_idx[i + 1]] - csr->row_ptr[split_idx[i]];
csrs[i].row_ptr = (DWORD*)numa_alloc((csrs[i].rows + 1) * sizeof(DWORD));
csrs[i].col_idx = (int*)numa_alloc(csrs[i].non_zeros * sizeof(int));
csrs[i].vals = (FLOAT*)numa_alloc(csrs[i].non_zeros * sizeof(FLOAT));
memcpy(csrs[i].row_ptr, csr->row_ptr + split_idx[i], (csrs[i].rows + 1) * sizeof(DWORD));
memcpy(csrs[i].col_idx, csr->col_idx + item_idx, csrs[i].non_zeros * sizeof(int));
memcpy(csrs[i].vals, csr->vals + item_idx, csrs[i].non_zeros * sizeof(FLOAT));
for (j = 0; j <= csrs[i].rows; j++)
{
csrs[i].row_ptr[j] -= csr->row_ptr[split_idx[i]];
}
item_idx += csrs[i].non_zeros;
}
}
else if (dir == SPLIT_VERTICAL)
{
struct csr_mat_t *csrs = csr_cont->csrs;
int *split_idx = csr_cont->split_idx;
split_idx[0] = 0;
INT64 *col_cnt = (INT64*)calloc((csr->cols + 1), sizeof(INT64));
for (i = 0; i < csr->rows; i++)
{
for (j = csr->row_ptr[i]; j < csr->row_ptr[i + 1]; j++)
{
col_cnt[csr->col_idx[j] + 1]++;
}
}
int avg_ele = csr->non_zeros / count, split_val;
int cur_col = 0;
for (i = 1, j = 1; i < count; i++)
{
split_val = i * avg_ele;
do
{
cur_col += col_cnt[j++];
}
while (cur_col < split_val);
if (cur_col - split_val > split_val - (cur_col - col_cnt[j]))
{
cur_col -= col_cnt[j--];
}
split_idx[i] = j;
}
split_idx[i] = csr->cols;
for (i = 0; i < csr->cols; i++)
{
col_cnt[i + 1] += col_cnt[i];
}
for (i = 0; i < count; i++)
{
csrs[i].rows = csr->rows;
csrs[i].cols = split_idx[i + 1] - split_idx[i];
csrs[i].non_zeros = col_cnt[split_idx[i + 1]] - col_cnt[split_idx[i]];
csrs[i].row_ptr = (DWORD*)numa_alloc((csrs[i].rows + 1) * sizeof(DWORD));
csrs[i].col_idx = (int*)numa_alloc(csrs[i].non_zeros * sizeof(int));
csrs[i].vals = (FLOAT*)numa_alloc(csrs[i].non_zeros * sizeof(FLOAT));
memset(csrs[i].row_ptr, 0, (csrs[i].rows + 1) * sizeof(DWORD));
}
int col, k;
for (i = 0; i < csr->rows; i++)
{
for (j = 0; j < count; j++)
{
csrs[j].row_ptr[i + 1] = csrs[j].row_ptr[i];
}
for (j = csr->row_ptr[i]; j < csr->row_ptr[i + 1]; j++)
{
col = csr->col_idx[j];
for (k = 0; k < count; k++)
{
if (col < split_idx[k + 1])
{
break;
}
}
csrs[k].col_idx[csrs[k].row_ptr[i + 1]] = csr->col_idx[j] - split_idx[k];
csrs[k].vals[csrs[k].row_ptr[i + 1]] = csr->vals[j];
csrs[k].row_ptr[i + 1]++;
}
}
free(col_cnt);
}
return 0;
}
void INTERNAL *qt_affinity_alloc(size_t bytes)
{ /*{{{ */
return numa_alloc(bytes);
} /*}}} */
int main(int argc, char *argv[])
{
if (argc != 2)
{
fprintf(stderr, "usage: %s csr_matrix_file\n", argv[0]);
exit(0);
}
int i, j, k;
struct timespec start, end;
int num_threads = 1;
#pragma omp parallel
{
#pragma omp master
{
num_threads = omp_get_num_threads();
}
}
printf("Thread number: %d.\n", num_threads);
#pragma omp parallel for
for (i = 0; i < num_threads; i++)
{
int cpu = omp_get_thread_num();
thread_bind(cpu);
}
FILE *fp;
struct csr_mat_t csr, csr_re, csr_t, csr_t_re, csr_t_t;
struct blk_mat_t blk;
struct csr_cont_t csr_h, csr_v;
struct blk_cont_t blk_h, blk_t_h;
read_csr_mat(argv[1], &csr);
int rows = csr.rows;
int cols = csr.cols;
INT64 non_zeros = csr.non_zeros;
csr_transpose(&csr, &csr_t);
release_csr_mat(&csr);
int *reorder_map = (int*)malloc(cols * sizeof(int));
csr_reorder(&csr_t, &csr_re, reorder_map);
release_csr_mat(&csr_t);
csr_transpose(&csr_re, &csr_t_t);
release_csr_mat(&csr_re);
split_csr_lb_nz(&csr_t_t, &csr_h, num_threads, SPLIT_HORIZON);
release_csr_mat(&csr_t_t);
csr_cont_to_blk_cont(&csr_h, &blk_h);
release_csr_cont(&csr_h);
printf("Notify: finished the preprocessing.\n");
FLOAT *x = (FLOAT*)numa_alloc(cols * sizeof(FLOAT));
FLOAT *y = (FLOAT*)numa_alloc(rows * sizeof(FLOAT));
for (i = 0; i < cols; i++)
{
x[i] = 1.0;
}
// warm up
spmv_blks(&blk_h, x, y, NULL);
printf("Notify: begin csr spmv.\n");
clock_gettime(CLOCK_MONOTONIC_RAW, &start);
for (i = 0; i < LOOP_TIME; i++)
{
spmv_blks(&blk_h, x, y, NULL);
}
clock_gettime(CLOCK_MONOTONIC_RAW, &end);
double time = get_sec(&start, &end) / LOOP_TIME;
double gflops = 2.0 * non_zeros / time * 1e-9;
printf("Notify: blk spmv time = %lfs, perf = %lf GFLOPS.\n", time, gflops);
// result_file(y, rows);
return 0;
}
static
inline
T *alloc(
std::size_t i_size ///< size of block
)
{
T *data = nullptr;
#if NUMA_BLOCK_ALLOCATOR_TYPE == 1 || NUMA_BLOCK_ALLOCATOR_TYPE == 2
# if NUMA_BLOCK_ALLOCATOR_TYPE == 1
// dummy call here to initialize this class as part of the singleton out of critical region
getSingletonRef();
# if SWEET_THREADING || SWEET_REXI_THREAD_PARALLEL_SUM
# pragma omp critical
# endif
# endif
{
std::vector<void*>& block_list = getBlocksSameSize(i_size);
if (block_list.size() > 0)
{
data = (T*)block_list.back();
block_list.pop_back();
}
}
if (data != nullptr)
return data;
return (T*)first_touch_init(numa_alloc(i_size), i_size);
#elif NUMA_BLOCK_ALLOCATOR_TYPE == 3
#if SWEET_THREADING || SWEET_REXI_THREAD_PARALLEL_SUM
# pragma omp critical
#endif
{
std::vector<void*>& block_list = getBlocksSameSize(i_size);
if (block_list.size() > 0)
{
data = (T*)block_list.back();
block_list.pop_back();
}
}
if (data != nullptr)
return data;
int retval = posix_memalign((void**)&data, 4096, i_size);
if (retval != 0)
{
std::cerr << "Unable to allocate memory" << std::endl;
assert(false);
exit(-1);
}
first_touch_init(data, i_size);
return data;
#else
// allocate a new element to the list of blocks given in block_list
// posix_memalign is thread safe
// http://www.qnx.com/developers/docs/6.3.0SP3/neutrino/lib_ref/p/posix_memalign.html
int retval = posix_memalign((void**)&data, 4096, i_size);
if (retval != 0)
{
std::cerr << "Unable to allocate memory" << std::endl;
assert(false);
exit(-1);
}
first_touch_init(data, i_size);
return data;
#endif
}