void blake_large_lastBlock(blake_large_ctx_t* ctx, const void* msg, uint16_t length_b){
while(length_b>=BLAKE_LARGE_BLOCKSIZE){
blake_large_nextBlock(ctx, msg);
msg = (uint8_t*)msg + BLAKE_LARGE_BLOCKSIZE_B;
length_b -= BLAKE_LARGE_BLOCKSIZE;
}
uint8_t buffer[128];
uint64_t v[16];
uint64_t ctr;
ctr = ctx->counter*1024+length_b;
memset(buffer, 0, 128);
memcpy(buffer, msg, (length_b+7)/8);
buffer[length_b/8] |= 0x80 >> (length_b&0x7);
blake_large_changeendian(buffer, buffer);
blake_large_expand(v, ctx);
if(length_b>1024-128-2){
v[12] ^= ctr;
v[13] ^= ctr;
blake_large_compress(v, buffer);
blake_large_collapse(ctx, v);
memset(buffer, 0, 128-8);
blake_large_expand(v, ctx);
} else {
if(length_b){
v[12] ^= ctr;
v[13] ^= ctr;
}
}
if(ctx->appendone)
buffer[128-16-8] |= 0x01;
*((uint64_t*)(&(buffer[128-8]))) = ctr;
blake_large_compress(v, buffer);
blake_large_collapse(ctx, v);
}
void blake_large_nextBlock(blake_large_ctx_t *ctx, const void *msg){
uint64_t v[16];
uint64_t m[16];
union {
uint64_t v64;
uint32_t v32[2];
}ctr;
blake_large_expand(v,ctx);
ctx->counter++;
ctr.v64 = ctx->counter*1024;
v[12] ^= ctr.v64;
v[13] ^= ctr.v64;
blake_large_changeendian(m, msg);
blake_large_compress(v, m);
blake_large_collapse(ctx, v);
}