#!/usr/bin/env bash
set -euo pipefail

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"

# Multi-GPU orchestration for run_from_csv.py:
# split CSV row indices round-robin across GPUs, one worker process per GPU.
#
# Usage:
#   bash run_csv_all_per_gpu.sh
#
# Useful overrides:
#   GPU_IDS=0,1,2,3 bash run_csv_all_per_gpu.sh
#   START_INDEX=0 END_INDEX=49 bash run_csv_all_per_gpu.sh
#   CSV=liveworld_i2v_cases.csv OUTPUT_DIR=outputs_spatialvid_i2v bash run_csv_all_per_gpu.sh
#   SKIP_EXISTING=0 bash run_csv_all_per_gpu.sh

PYTHON_BIN="${PYTHON_BIN:-/home/tione/notebook/workspace/zhiyuan/miniconda3/envs/xfuser/bin/python}"

CSV="${CSV:-${SCRIPT_DIR}/liveworld_i2v_cases.csv}"
OUTPUT_DIR="${OUTPUT_DIR:-${SCRIPT_DIR}/outputs_spatialvid_i2v}"
LOG_DIR="${LOG_DIR:-${SCRIPT_DIR}/logs_csv_per_gpu}"
CONFIG="${CONFIG:-${SCRIPT_DIR}/configs/csv_mp4_81.yaml}"
GPU_IDS="${GPU_IDS:-0,1,2,3,4,5,6,7}"
SKIP_EXISTING="${SKIP_EXISTING:-1}"

TOTAL_ENTRIES="$("${PYTHON_BIN}" -c '
import csv, sys
from pathlib import Path
p = Path(sys.argv[1])
with p.open("r", encoding="utf-8-sig", newline="") as f:
    rows = [r for r in csv.DictReader(f) if any((v or "").strip() for v in r.values())]
print(len(rows))
' "${CSV}")"

START_INDEX="${START_INDEX:-0}"
END_INDEX="${END_INDEX:-$((TOTAL_ENTRIES - 1))}"

gpu_array=(${GPU_IDS//,/ })
num_gpus="${#gpu_array[@]}"

if [ "${num_gpus}" -eq 0 ]; then
  echo "GPU_IDS is empty" >&2
  exit 1
fi
if [ "${START_INDEX}" -gt "${END_INDEX}" ]; then
  echo "START_INDEX (${START_INDEX}) must be <= END_INDEX (${END_INDEX})" >&2
  exit 1
fi
if [ "${END_INDEX}" -ge "${TOTAL_ENTRIES}" ]; then
  echo "END_INDEX (${END_INDEX}) out of range (csv has ${TOTAL_ENTRIES} rows)" >&2
  exit 1
fi

mkdir -p "${OUTPUT_DIR}" "${LOG_DIR}"

echo "[batch] csv=${CSV}"
echo "[batch] output=${OUTPUT_DIR}"
echo "[batch] logs=${LOG_DIR}"
echo "[batch] config=${CONFIG}"
echo "[batch] indices=${START_INDEX}..${END_INDEX} (total_csv=${TOTAL_ENTRIES})"
echo "[batch] gpu_ids=${GPU_IDS}"
echo "[batch] skip_existing=${SKIP_EXISTING}"

pids=()
for i in "${!gpu_array[@]}"; do
  gpu_id="${gpu_array[i]}"

  indices=""
  for ((index = START_INDEX + i; index <= END_INDEX; index += num_gpus)); do
    indices="${indices:+${indices},}${index}"
  done

  if [ -z "${indices}" ]; then
    continue
  fi

  echo "[GPU ${gpu_id}] indices: ${indices}"

  GPU_ID="${gpu_id}" \
  INDICES="${indices}" \
  PYTHON_BIN="${PYTHON_BIN}" \
  CSV="${CSV}" \
  OUTPUT_DIR="${OUTPUT_DIR}" \
  LOG_DIR="${LOG_DIR}" \
  CONFIG="${CONFIG}" \
  SKIP_EXISTING="${SKIP_EXISTING}" \
  bash "${SCRIPT_DIR}/run_csv_indices_per_gpu.sh" &
  pids+=("$!")
done

failed=0
for pid in "${pids[@]}"; do
  if ! wait "${pid}"; then
    failed=1
  fi
done

if [ "${failed}" -ne 0 ]; then
  echo "Some GPU jobs failed. Check logs in ${LOG_DIR}." >&2
  exit 1
fi

echo "All jobs finished."