//
// Arquiteturas de Alto Desempenho 2025/2026
//
// DETI Coin Miner - WebAssembly implementation with SIMD support
//

#include <time.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include "aad_data_types.h"

#ifdef __EMSCRIPTEN__
#include <emscripten.h>
#include "aad_sha1_wasm.h"
#else
#include "aad_sha1_cpu.h"
#include "aad_utilities.h"
#include "aad_vault.h"
#endif

// WASM SIMD support
#if defined(__wasm_simd128__)
#include <wasm_simd128.h>
#endif

// Global mining state
static volatile int keep_running = 1;
static volatile int use_simd = 0;
static u64_t total_attempts = 0;
static u32_t coins_found = 0;
static double mining_start_time = 0;
static double pause_time_offset = 0;
static double last_pause_time = 0;
static u32_t found_coins[1024][14];
static u32_t found_coins_count = 0;

//
// Check if a hash starts with aad20250
//
static int is_valid_coin(u32_t *hash)
{
  return hash[0] == 0xAAD20250u;
}

//
// Increment coin variable part (positions 12-53)
//
static int increment_coin(u32_t coin[14])
{
  int pos = 53;
  while(pos >= 12)
  {
    u08_t *byte = &((u08_t *)coin)[pos ^ 3];
    if(*byte == '\n' || *byte == 0x80)
      *byte = 32;

    (*byte)++;

    if(*byte == '\n')
      (*byte)++;

    if(*byte >= 127)
    {
      *byte = 32;
      pos--;
    }
    else
    {
      break;
    }
  }

  return (pos >= 12);
}

//
// Get current time in seconds
//
static double get_time()
{
#ifdef __EMSCRIPTEN__
  return emscripten_get_now() / 1000.0;
#else
  struct timespec ts;
  clock_gettime(CLOCK_MONOTONIC_RAW, &ts);
  return (double)ts.tv_sec + (double)ts.tv_nsec / 1e9;
#endif
}

#if defined(__wasm_simd128__)
//
// Helper macro for Left Rotate (ROTL) using SIMD shifts
// CORRECTION: Used wasm_u32x4_shr instead of incorrect wasm_i32x4_shr_u
//
#define SIMD_ROTL(x, n) wasm_v128_or(wasm_i32x4_shl(x, n), wasm_u32x4_shr(x, 32 - n))

//
// SIMD implementation for WebAssembly (4-way parallel)
//
static void prepare_coins_simd(u32_t base_coin[14], u32_t *interleaved_data)
{
  const int SIMD_WIDTH = 4;
  for(int lane = 0; lane < SIMD_WIDTH; lane++)
  {
    u32_t coin[14];
    memcpy(coin, base_coin, sizeof(coin));

    for(int idx = 0; idx < 14; idx++)
    {
      interleaved_data[idx * SIMD_WIDTH + lane] = coin[idx];
    }

    increment_coin(base_coin);
  }
}

static void extract_hashes_simd(u32_t *interleaved_hash, u32_t hashes[][5])
{
  const int SIMD_WIDTH = 4;
  for(int lane = 0; lane < SIMD_WIDTH; lane++)
  {
    for(int idx = 0; idx < 5; idx++)
    {
      hashes[lane][idx] = interleaved_hash[idx * SIMD_WIDTH + lane];
    }
  }
}

static void extract_coins_simd(u32_t *interleaved_data, u32_t coins[][14])
{
  const int SIMD_WIDTH = 4;
  for(int lane = 0; lane < SIMD_WIDTH; lane++)
  {
    for(int idx = 0; idx < 14; idx++)
    {
      coins[lane][idx] = interleaved_data[idx * SIMD_WIDTH + lane];
    }
  }
}

static void sha1_wasm_simd(u32_t *interleaved_data, u32_t *interleaved_hash)
{
  const int SIMD_WIDTH = 4;

  // SHA1 State vectors
  v128_t a, b, c, d, e;
  v128_t w[80]; // Message schedule

  // Initial SHA1 constants
  a = wasm_i32x4_splat(0x67452301u);
  b = wasm_i32x4_splat(0xEFCDAB89u);
  c = wasm_i32x4_splat(0x98BADCFEu);
  d = wasm_i32x4_splat(0x10325476u);
  e = wasm_i32x4_splat(0xC3D2E1F0u);

  // 1. Prepare Message Schedule (Interleaved loads)
  // Load first 14 words
  for(int i = 0; i < 14; i++)
  {
    w[i] = wasm_v128_load(&interleaved_data[i * SIMD_WIDTH]);
  }
  // Standard padding (assumed handled by caller/init)
  w[14] = wasm_i32x4_splat(0);
  w[15] = wasm_i32x4_splat(440); // Length in bits (55 bytes * 8)

  // Expand message schedule from 16 to 80
  for (int i = 16; i < 80; i++) {
    v128_t temp = wasm_v128_xor(w[i-3], w[i-8]);
    temp = wasm_v128_xor(temp, w[i-14]);
    temp = wasm_v128_xor(temp, w[i-16]);
    w[i] = SIMD_ROTL(temp, 1);
  }

  // 2. Main Loop (80 rounds)
  for (int i = 0; i < 80; i++) {
    v128_t f, k;

    if (i < 20) {
      // F1: (b & c) | (~b & d)
      f = wasm_v128_or(wasm_v128_and(b, c), wasm_v128_and(wasm_v128_not(b), d));
      k = wasm_i32x4_splat(0x5A827999u);
    } else if (i < 40) {
      // F2: b ^ c ^ d
      f = wasm_v128_xor(wasm_v128_xor(b, c), d);
      k = wasm_i32x4_splat(0x6ED9EBA1u);
    } else if (i < 60) {
      // F3: (b & c) | (b & d) | (c & d)
      f = wasm_v128_or(wasm_v128_or(wasm_v128_and(b, c), wasm_v128_and(b, d)), wasm_v128_and(c, d));
      k = wasm_i32x4_splat(0x8F1BBCDCu);
    } else {
      // F4: b ^ c ^ d
      f = wasm_v128_xor(wasm_v128_xor(b, c), d);
      k = wasm_i32x4_splat(0xCA62C1D6u);
    }

    // temp = ROTL(a, 5) + f + e + k + w[i]
    v128_t temp = wasm_i32x4_add(SIMD_ROTL(a, 5), f);
    temp = wasm_i32x4_add(temp, e);
    temp = wasm_i32x4_add(temp, k);
    temp = wasm_i32x4_add(temp, w[i]);

    e = d;
    d = c;
    c = SIMD_ROTL(b, 30);
    b = a;
    a = temp;
  }

  // 3. Add to initial state and store
  v128_t h0 = wasm_i32x4_add(a, wasm_i32x4_splat(0x67452301u));
  v128_t h1 = wasm_i32x4_add(b, wasm_i32x4_splat(0xEFCDAB89u));
  v128_t h2 = wasm_i32x4_add(c, wasm_i32x4_splat(0x98BADCFEu));
  v128_t h3 = wasm_i32x4_add(d, wasm_i32x4_splat(0x10325476u));
  v128_t h4 = wasm_i32x4_add(e, wasm_i32x4_splat(0xC3D2E1F0u));

  // Store results back to interleaved_hash
  wasm_v128_store(&interleaved_hash[0 * SIMD_WIDTH], h0);
  wasm_v128_store(&interleaved_hash[1 * SIMD_WIDTH], h1);
  wasm_v128_store(&interleaved_hash[2 * SIMD_WIDTH], h2);
  wasm_v128_store(&interleaved_hash[3 * SIMD_WIDTH], h3);
  wasm_v128_store(&interleaved_hash[4 * SIMD_WIDTH], h4);
}

static int mine_coins_wasm_simd_internal(u32_t iterations_per_call, u32_t coin[14])
{
  const int SIMD_WIDTH = 4;
  u32_t interleaved_data[14 * SIMD_WIDTH] __attribute__((aligned(16)));
  u32_t interleaved_hash[5 * SIMD_WIDTH] __attribute__((aligned(16)));

  u32_t batches = (iterations_per_call + SIMD_WIDTH - 1) / SIMD_WIDTH;

  for(u32_t batch = 0; batch < batches && keep_running; batch++)
  {
    prepare_coins_simd(coin, interleaved_data);

    // This now uses the REAL vectorized implementation
    sha1_wasm_simd(interleaved_data, interleaved_hash);

    total_attempts += SIMD_WIDTH;

    u32_t hashes[SIMD_WIDTH][5];
    extract_hashes_simd(interleaved_hash, hashes);

    for(int lane = 0; lane < SIMD_WIDTH; lane++)
    {
      if(is_valid_coin(hashes[lane]))
      {
        if(found_coins_count < 1024)
        {
          u32_t coins[SIMD_WIDTH][14];
          extract_coins_simd(interleaved_data, coins);
          memcpy(found_coins[found_coins_count], coins[lane], sizeof(coins[lane]));
          found_coins_count++;
        }
        coins_found++;

#ifndef __EMSCRIPTEN__
        printf("COIN FOUND! (attempt %llu, lane %d)\n",
               (unsigned long long)(total_attempts - SIMD_WIDTH + lane), lane);
#endif
      }
    }
  }

  return keep_running;
}
#endif

//
// Main mining iteration (called from JavaScript)
//
#ifdef __EMSCRIPTEN__
EMSCRIPTEN_KEEPALIVE
#endif
int mine_coins_wasm(u32_t iterations_per_call)
{
  static u32_t coin[14];
  static int initialized = 0;
  static int last_logged_mode = -1; // Track last reported mode

  u32_t hash[5];

  // Initialize coin template on first call
  if(!initialized)
  {
    memset(coin, 0, sizeof(coin));
    ((u08_t *)coin)[0x0 ^ 3] = 'D';
    ((u08_t *)coin)[0x1 ^ 3] = 'E';
    ((u08_t *)coin)[0x2 ^ 3] = 'T';
    ((u08_t *)coin)[0x3 ^ 3] = 'I';
    ((u08_t *)coin)[0x4 ^ 3] = ' ';
    ((u08_t *)coin)[0x5 ^ 3] = 'c';
    ((u08_t *)coin)[0x6 ^ 3] = 'o';
    ((u08_t *)coin)[0x7 ^ 3] = 'i';
    ((u08_t *)coin)[0x8 ^ 3] = 'n';
    ((u08_t *)coin)[0x9 ^ 3] = ' ';
    ((u08_t *)coin)[0xa ^ 3] = '2';
    ((u08_t *)coin)[0xb ^ 3] = ' ';
    ((u08_t *)coin)[0x36 ^ 3] = '\n';
    ((u08_t *)coin)[0x37 ^ 3] = 0x80;

    for(int i = 12; i < 54; i++)
      ((u08_t *)coin)[i ^ 3] = 'A' + (i - 12) % 26;

    mining_start_time = get_time();
    initialized = 1;
  }

  if(!keep_running)
    return 0;

#if defined(__wasm_simd128__)
  if(use_simd)
  {
    if (last_logged_mode != 1) {
        printf("C: Running in SIMD Mode (Vectorized)\n");
        last_logged_mode = 1;
    }
    return mine_coins_wasm_simd_internal(iterations_per_call, coin);
  }
#endif

  if (last_logged_mode != 0) {
      printf("C: Running in Scalar Mode\n");
      last_logged_mode = 0;
  }

  // Scalar implementation
  for(u32_t i = 0; i < iterations_per_call && keep_running; i++)
  {
    sha1(coin, hash);
    total_attempts++;

    if(is_valid_coin(hash))
    {
      if(found_coins_count < 1024)
      {
        memcpy(found_coins[found_coins_count], coin, sizeof(coin));
        found_coins_count++;
      }
      coins_found++;

#ifndef __EMSCRIPTEN__
      printf("COIN FOUND! (attempt %llu)\n", (unsigned long long)total_attempts);
      save_coin(coin);
#endif
    }

    increment_coin(coin);
  }

  return keep_running;
}

//
// Get mining statistics
//
#ifdef __EMSCRIPTEN__
EMSCRIPTEN_KEEPALIVE
#endif
void get_statistics(u64_t *attempts, u32_t *coins, double *hash_rate, double *elapsed_time)
{
  *attempts = total_attempts;
  *coins = coins_found;

  double current_time = get_time();
  double actual_elapsed;

  if(!keep_running && last_pause_time > 0) {
    actual_elapsed = last_pause_time - mining_start_time - pause_time_offset;
  } else {
    actual_elapsed = current_time - mining_start_time - pause_time_offset;
  }

  *elapsed_time = actual_elapsed;
  *hash_rate = (actual_elapsed > 0) ? (total_attempts / actual_elapsed) : 0;
}

//
// Enable/disable SIMD
//
#ifdef __EMSCRIPTEN__
EMSCRIPTEN_KEEPALIVE
#endif
void set_simd_enabled(int enabled)
{
  use_simd = enabled;
  printf("C: SIMD mode set to: %d\n", use_simd);
}

//
// Check if SIMD is available
//
#ifdef __EMSCRIPTEN__
EMSCRIPTEN_KEEPALIVE
#endif
int is_simd_available()
{
#if defined(__wasm_simd128__)
  return 1;
#else
  return 0;
#endif
}

//
// Get current SIMD state
//
#ifdef __EMSCRIPTEN__
EMSCRIPTEN_KEEPALIVE
#endif
int is_simd_enabled()
{
  return use_simd;
}

//
// Stop mining
//
#ifdef __EMSCRIPTEN__
EMSCRIPTEN_KEEPALIVE
#endif
void stop_mining()
{
  if(keep_running) {
    keep_running = 0;
    last_pause_time = get_time();
  }
}

//
// Resume mining
//
#ifdef __EMSCRIPTEN__
EMSCRIPTEN_KEEPALIVE
#endif
void resume_mining()
{
  if(!keep_running && last_pause_time > 0) {
    double pause_duration = get_time() - last_pause_time;
    pause_time_offset += pause_duration;
    keep_running = 1;
    last_pause_time = 0;
  }
}

//
// Get found coin data
//
#ifdef __EMSCRIPTEN__
EMSCRIPTEN_KEEPALIVE
#endif
u32_t* get_found_coin(u32_t index)
{
  if(index < found_coins_count)
    return found_coins[index];
  return NULL;
}

//
// Get number of found coins
//
#ifdef __EMSCRIPTEN__
EMSCRIPTEN_KEEPALIVE
#endif
u32_t get_found_coins_count()
{
  return found_coins_count;
}

//
// Reset mining state
//
#ifdef __EMSCRIPTEN__
EMSCRIPTEN_KEEPALIVE
#endif
void reset_mining()
{
  keep_running = 1;
  total_attempts = 0;
  coins_found = 0;
  found_coins_count = 0;
  pause_time_offset = 0;
  last_pause_time = 0;
  mining_start_time = get_time();
}

//
// Main function (for standalone compilation/testing)
//
#ifndef __EMSCRIPTEN__
int main(int argc, char *argv[])
{
  u64_t max_attempts = 0;

  if(argc > 1)
    max_attempts = strtoull(argv[1], NULL, 10);

  printf("Mining DETI coins using WebAssembly implementation (standalone mode)...\n");
#if defined(__wasm_simd128__)
  printf("SIMD support: available\n");
  use_simd = 1;
#else
  printf("SIMD support: not available\n");
#endif
  printf("Press Ctrl+C to stop\n\n");

  time_measurement();
  time_measurement();
  double start_time = wall_time_delta();
  double last_report = start_time;

  while(keep_running && (max_attempts == 0 || total_attempts < max_attempts))
  {
    mine_coins_wasm(100000);

    time_measurement();
    double current_time = wall_time_delta() - start_time;

    if(current_time - last_report >= 1.0)
    {
      u64_t attempts;
      u32_t coins;
      double hash_rate, elapsed;
      get_statistics(&attempts, &coins, &hash_rate, &elapsed);

      printf("Attempts: %llu, Rate: %.2f MH/s, Coins: %u\n",
             (unsigned long long)attempts, hash_rate / 1e6, coins);

      last_report = current_time;
    }
  }

  time_measurement();
  double total_time = wall_time_delta() - start_time;

  printf("\n=== Mining Statistics ===\n");
  printf("Total attempts: %llu\n", (unsigned long long)total_attempts);
  printf("Total time: %.2f seconds\n", total_time);
  printf("Average rate: %.2f attempts/second\n", total_attempts / total_time);
  printf("Coins found: %u\n", coins_found);

  save_coin(NULL);

  return 0;
}
#endif