Doing advent of code 2015 day 4 with https://github.com/intel/isa-l_crypto to try t oget it as fast as possibe

2015-day4.cpp

#include <string_view>
#include <cstdio>
#include <cstdlib>

#include <fmt/core.h>
#include <fmt/format.h>

#include <md5_mb.h>
#include <endian_helper.h>

#define DIGEST_NWORDS   MD5_DIGEST_NWORDS
#define MB_BUFS         MD5_MAX_LANES
#define HASH_CTX_MGR    MD5_HASH_CTX_MGR
#define HASH_CTX	MD5_HASH_CTX

#define CTX_MGR_INIT		md5_ctx_mgr_init_avx2
#define CTX_MGR_SUBMIT		md5_ctx_mgr_submit_avx2
#define CTX_MGR_FLUSH		md5_ctx_mgr_flush_avx2

#define rounds_buf MD5_MAX_LANES

int main() {
	HASH_CTX *ctx = nullptr;

	HASH_CTX_MGR* mgr = static_cast<HASH_CTX_MGR*>(std::aligned_alloc(16, sizeof *mgr));
	CTX_MGR_INIT(mgr);

	HASH_CTX ctxpool[16];

	for (int i = 0; i < 16; i++) {
		hash_ctx_init(&ctxpool[i]);
		ctxpool[i].user_data =  (void *)((uint64_t) i);
	}

	std::string base = "iwrupvqb";

	// We calculate 16 md5hashes at a time.
	for (int i = 0; i < 16; i++) {
		auto data = base + fmt::to_string(i+1);
		ctx = CTX_MGR_SUBMIT(mgr, &ctxpool[i], data.data(), data.size(), HASH_ENTIRE);
	}
	
	int idx = 16;
	bool found_first = false;
	while (ctx) {
		if (ctx->error != HASH_CTX_ERROR_NONE) {
			fmt::print("WE GOT A FUCKING ERROR\n");
			return -1;
		}

		auto dg = ctx->job.result_digest[0];
		if (!found_first && ((to_be32(dg) & 0xFFFFF000) == 0)) {
			found_first = true;
			fmt::print("Part 1: {:d}\t0x", (uint64_t)ctx->user_data + 1, to_be32(dg));
			for (int i = 0; i < DIGEST_NWORDS; i++) {
				fmt::print("{:08x}", to_be32(ctx->job.result_digest[i]));
			}
			fmt::print("\n");
		}

		if ((dg & 0xFFFFFF) == 0) {
			fmt::print("Part 2: {:d}\t0x", (uint64_t)ctx->user_data + 1);
			for (int i = 0; i < DIGEST_NWORDS; i++) {
				fmt::print("{:08x}", to_be32(ctx->job.result_digest[i]));
			}
			fmt::print("\n");
			break;
		}

		auto data = base + fmt::to_string(idx+1);
		ctx->user_data = (void *)((uint64_t) idx);
		ctx = CTX_MGR_SUBMIT(mgr, ctx, data.data(), data.size(), HASH_ENTIRE);

		idx++;
	}

	std::free(mgr);

	return 0;
}