/* Bitshuffle a single block. */
int64_t bshuf_bitshuffle_block(ioc_chain *C_ptr,
const size_t size, const size_t elem_size) {
size_t this_iter;
void *in = ioc_get_in(C_ptr, &this_iter);
ioc_set_next_in(C_ptr, &this_iter,
(void*) ((char*) in + size * elem_size));
void *out = ioc_get_out(C_ptr, &this_iter);
ioc_set_next_out(C_ptr, &this_iter,
(void *) ((char *) out + size * elem_size));
int64_t count = bshuf_trans_bit_elem(in, out, size, elem_size);
return count;
}
/* Wrap a function for processing a single block to process an entire buffer in
* parallel. */
int64_t bshuf_blocked_wrap_fun(bshufBlockFunDef fun, void* in, void* out,
const size_t size, const size_t elem_size, size_t block_size) {
size_t ii;
ioc_chain C;
ioc_init(&C, in, out);
int64_t err = 0, count, cum_count = 0;
size_t last_block_size;
if (block_size == 0) {
block_size = bshuf_default_block_size(elem_size);
}
if (block_size < 0 || block_size % BSHUF_BLOCKED_MULT) return -81;
#if defined(_OPENMP)
#pragma omp parallel for schedule(dynamic, 1) \
private(count) reduction(+ : cum_count)
#endif
for (ii = 0; ii < size / block_size; ii ++) {
count = fun(&C, block_size, elem_size);
if (count < 0) err = count;
cum_count += count;
}
last_block_size = size % block_size;
last_block_size = last_block_size - last_block_size % BSHUF_BLOCKED_MULT;
if (last_block_size) {
count = fun(&C, last_block_size, elem_size);
if (count < 0) err = count;
cum_count += count;
}
if (err < 0) return err;
size_t leftover_bytes = size % BSHUF_BLOCKED_MULT * elem_size;
size_t this_iter;
char *last_in = (char *) ioc_get_in(&C, &this_iter);
ioc_set_next_in(&C, &this_iter, (void *) (last_in + leftover_bytes));
char *last_out = (char *) ioc_get_out(&C, &this_iter);
ioc_set_next_out(&C, &this_iter, (void *) (last_out + leftover_bytes));
memcpy(last_out, last_in, leftover_bytes);
ioc_destroy(&C);
return cum_count + leftover_bytes;
}
/* Decompress and bitunshuffle a single block. */
int64_t bshuf_decompress_lz4_block(ioc_chain *C_ptr,
const size_t size, const size_t elem_size) {
int64_t nbytes, count;
size_t this_iter;
void *in = ioc_get_in(C_ptr, &this_iter);
int32_t nbytes_from_header = bshuf_read_uint32_BE(in);
ioc_set_next_in(C_ptr, &this_iter,
(void*) ((char*) in + nbytes_from_header + 4));
void *out = ioc_get_out(C_ptr, &this_iter);
ioc_set_next_out(C_ptr, &this_iter,
(void *) ((char *) out + size * elem_size));
void* tmp_buf = malloc(size * elem_size);
if (tmp_buf == NULL) return -1;
#ifdef BSHUF_LZ4_DECOMPRESS_FAST
nbytes = LZ4_decompress_fast((const char*) in + 4, (char*) tmp_buf, size * elem_size);
CHECK_ERR_FREE_LZ(nbytes, tmp_buf);
if (nbytes != nbytes_from_header) {
free(tmp_buf);
return -91;
}
#else
nbytes = LZ4_decompress_safe((const char*) in + 4, (char *) tmp_buf, nbytes_from_header,
size * elem_size);
CHECK_ERR_FREE_LZ(nbytes, tmp_buf);
if (nbytes != size * elem_size) {
free(tmp_buf);
return -91;
}
nbytes = nbytes_from_header;
#endif
count = bshuf_untrans_bit_elem(tmp_buf, out, size, elem_size);
CHECK_ERR_FREE(count, tmp_buf);
nbytes += 4;
free(tmp_buf);
return nbytes;
}
/* Bitshuffle and compress a single block. */
int64_t bshuf_compress_lz4_block(ioc_chain *C_ptr,
const size_t size, const size_t elem_size) {
int64_t nbytes, count;
void* tmp_buf_bshuf = malloc(size * elem_size);
if (tmp_buf_bshuf == NULL) return -1;
void* tmp_buf_lz4 = malloc(LZ4_compressBound(size * elem_size));
if (tmp_buf_lz4 == NULL){
free(tmp_buf_bshuf);
return -1;
}
size_t this_iter;
void *in = ioc_get_in(C_ptr, &this_iter);
ioc_set_next_in(C_ptr, &this_iter, (void*) ((char*) in + size * elem_size));
count = bshuf_trans_bit_elem(in, tmp_buf_bshuf, size, elem_size);
if (count < 0) {
free(tmp_buf_lz4);
free(tmp_buf_bshuf);
return count;
}
nbytes = LZ4_compress((const char*) tmp_buf_bshuf, (char*) tmp_buf_lz4, size * elem_size);
free(tmp_buf_bshuf);
CHECK_ERR_FREE_LZ(nbytes, tmp_buf_lz4);
void *out = ioc_get_out(C_ptr, &this_iter);
ioc_set_next_out(C_ptr, &this_iter, (void *) ((char *) out + nbytes + 4));
bshuf_write_uint32_BE(out, nbytes);
memcpy((char *) out + 4, tmp_buf_lz4, nbytes);
free(tmp_buf_lz4);
return nbytes + 4;
}
