for (k = offsets[i]; k < offsets[i + 1]; k++) {
tmp = (int)(((int64_t)src[group * 128 + k] * c + \
(int64_t)0x1000000000) >> 37);
for (k = offsets[i]; k < offsets[i + 1]; k++) {
tmp = (int)(((int64_t)src[group * 128 + k] * c + \
(int64_t)0x1000000000) >> 37);