26k and fix bpu round

.gitignore

@@ -7,6 +7,8 @@ deploy_45dim_rl_gym/bpu_quantization/calibration_data/
 deploy_45dim_rl_gym/bpu_quantization/calibration_data_fast*/
 deploy_45dim_rl_gym/bpu_quantization/calibration_data_*_gym_fast*/
 deploy_45dim_rl_gym/bpu_quantization/calibration_data_*_gym_fast*_metadata.json
+deploy_45dim_rl_gym/bpu_quantization/calibration_data_*_robotlab_fast*/
+deploy_45dim_rl_gym/bpu_quantization/calibration_data_*_robotlab_fast*_metadata.json
 deploy_45dim_rl_gym/bpu_quantization/mapper_output/
 deploy_45dim_rl_gym/bpu_quantization/mapper_output_gemm/
 deploy_45dim_rl_gym/bpu_quantization/mapper_output_*_gemm/
@@ -16,6 +18,7 @@ deploy_45dim_rl_gym/bpu_quantization/policy_[0-9]*_actions.onnx
 deploy_45dim_rl_gym/bpu_quantization/policy_[0-9]*_bpu4d.onnx
 deploy_45dim_rl_gym/bpu_quantization/policy_[0-9]*_bpu4d_gemm.onnx
 deploy_45dim_rl_gym/bpu_quantization/policy_[0-9]*_int16_gemm.yaml
+deploy_45dim_rl_gym/bpu_quantization/policy_[0-9]*_int8_gemm.yaml
 deploy_45dim_rl_gym/bpu_quantization/policy_[0-9]*_opset11.onnx
 deploy_45dim_rl_gym/bpu_quantization/policy_robotlab_15000_bpu4d.onnx
 deploy_45dim_rl_gym/bpu_quantization/policy_robotlab_15000_bpu4d_gemm.onnx

deploy_45dim_rl_gym/bpu_deploy_x5/README.md

@@ -7,7 +7,7 @@
 
 - `bpu_policy.py`：BPU policy wrapper，默认使用 `cpp_dnn_api_x5` C++ DNN API 后端。
 - `cpp/`：C++ DNN API 推理库源码、板端编译脚本和纯 C++ benchmark。
-- `deploy_go1_robotlab_bpu_x5_fastcpp.py`：基于现有 `lab_fastcpp` 状态机的 BPU 版本。
+- `deploy_go1_robotlab_bpu_x5_fastcpp.py`：基于现有 `lab_fastcpp` 状态机的 BPU 版本，默认 26000。
 
 ## 重要限制
 
@@ -52,7 +52,7 @@ bash build_board.sh
 默认模型：
 
 ```text
-deploy_45dim_rl_gym/bpu_quantization/mapper_output_6500_gemm/policy_robotlab_6500_int16_gemm.bin
+deploy_45dim_rl_gym/bpu_quantization/mapper_output_26000_gemm/policy_robotlab_26000_int16_gemm.bin
 ```
 
 输入输出：
@@ -66,12 +66,12 @@ deploy_45dim_rl_gym/bpu_quantization/mapper_output_6500_gemm/policy_robotlab_650
 
 ```bash
 cd /root/go1_pro_deploy
-ls deploy_45dim_rl_gym/bpu_quantization/mapper_output_6500_gemm/policy_robotlab_6500_int16_gemm.bin
+ls deploy_45dim_rl_gym/bpu_quantization/mapper_output_26000_gemm/policy_robotlab_26000_int16_gemm.bin
 ls deploy_45dim_rl_gym/bpu_quantization/calibration_data/00000.bin
 ls deploy_45dim_rl_gym/bpu_deploy_x5/cpp/libbpu_dnn_policy.so
 ```
 
-如果 `mapper_output_6500_gemm/` 没有同步到 git，需要手动把 `.bin` 放到对应路径，
+如果 `mapper_output_26000_gemm/` 没有同步到 git，需要手动把 `.bin` 放到对应路径，
 或用 `--bpu-model` 指定绝对路径。
 
 ## 纯离线 BPU 自检
@@ -82,38 +82,32 @@ ls deploy_45dim_rl_gym/bpu_deploy_x5/cpp/libbpu_dnn_policy.so
 cd /root/go1_pro_deploy
 PYTHONPATH=/root/go1_pro_deploy \
 python3 deploy_45dim_rl_gym/bpu_deploy_x5/test_bpu_policy.py \
-  --check-reference-00000 \
+  --bpu-model deploy_45dim_rl_gym/bpu_quantization/mapper_output_26000_gemm/policy_robotlab_26000_int16_gemm.bin \
+  --input-bin deploy_45dim_rl_gym/bpu_quantization/calibration_data_26000_robotlab_fast64/00000.bin \
   --repeat 1000
 ```
 
-通过时应看到：
-
-```text
-reference_max_abs_diff 0.0
-```
-
 当前板端 `root@192.168.150.167` 已验证：
 
 ```text
-reference_max_abs_diff 4.76837158203125e-07
-reference_mean_abs_diff 3.071812386679085e-07
-repeat=1000 avg_ms=1.019592  # Python ctypes -> C++ DNN API
-duplicate_errors=0
+repeat=1000 avg_ms=...  # Python ctypes -> C++ DNN API
 ```
 
 纯 C++ benchmark：
 
 ```bash
 cd /root/go1_pro_deploy/deploy_45dim_rl_gym/bpu_deploy_x5/cpp
-./bpu_dnn_bench
+./bpu_dnn_bench \
+  /root/go1_pro_deploy/deploy_45dim_rl_gym/bpu_quantization/mapper_output_26000_gemm/policy_robotlab_26000_int16_gemm.bin \
+  /root/go1_pro_deploy/deploy_45dim_rl_gym/bpu_quantization/calibration_data_26000_robotlab_fast64/00000.bin \
+  1000
 ```
 
 当前板端结果：
 
 ```text
-reference_max_abs_diff 0
-repeat=1000 cpp_avg_ms=0.711932
-duplicate_errors=0
+input_floats=450 output_floats=12
+repeat=1000 cpp_avg_ms=...
 ```
 
 ## 离线推理检查
@@ -132,6 +126,18 @@ python3 deploy_45dim_rl_gym/bpu_deploy_x5/deploy_go1_robotlab_bpu_x5_fastcpp.py
 
 这一步仍会连接 MCU 读取状态，但不会发送电机指令。
 
+如果要切回旧轮次：
+
+```bash
+PYTHONPATH=/root/go1_pro_deploy \
+python3 deploy_45dim_rl_gym/bpu_deploy_x5/deploy_go1_robotlab_bpu_x5_fastcpp.py \
+  --bpu-round 6500
+
+PYTHONPATH=/root/go1_pro_deploy \
+python3 deploy_45dim_rl_gym/bpu_deploy_x5/deploy_go1_robotlab_bpu_x5_fastcpp.py \
+  --bpu-round 15000
+```
+
 ## 悬空状态机测试
 
 只有 `--infer-check` 日志确认 action 正常后，再悬空测试状态机。先不加

deploy_45dim_rl_gym/bpu_deploy_x5/deploy_go1_robotlab_bpu_x5_fastcpp.py

@@ -5,7 +5,7 @@ Deploy the RoboGauge Go1 45-dim RobotLab BPU policy on Unitree Go1 PRO.
 This uses go1_pro_sdk direct MCU control, not LCM or the official Unitree SDK.
 
 Policy:
-  - policy_robotlab_6500_int16_gemm.bin
+  - policy_robotlab_26000_int16_gemm.bin by default
   - single-frame obs: 45 dims
   - BPU input: 10-frame history, 1x1x1x450 featuremap, stacked by observation terms
   - command scale: [1.0, 1.0, 1.0]
@@ -66,10 +66,16 @@ except ImportError as exc:
         "PYTHONPATH=/root/go1_pro_sdk python3 setup.py build_ext --inplace"
     ) from exc
 
-DEFAULT_BPU_MODEL = (
-    HERE.parent / "bpu_quantization" / "mapper_output_6500_gemm" /
-    "policy_robotlab_6500_int16_gemm.bin"
-)
+BPU_MODEL_REGISTRY = {
+    "6500": HERE.parent / "bpu_quantization" / "mapper_output_6500_gemm" /
+            "policy_robotlab_6500_int16_gemm.bin",
+    "15000": HERE.parent / "bpu_quantization" / "mapper_output_gemm" /
+             "policy_robotlab_15000_int16_gemm.bin",
+    "26000": HERE.parent / "bpu_quantization" / "mapper_output_26000_gemm" /
+             "policy_robotlab_26000_int16_gemm.bin",
+}
+DEFAULT_BPU_ROUND = "26000"
+DEFAULT_BPU_MODEL = BPU_MODEL_REGISTRY[DEFAULT_BPU_ROUND]
 LOWCMD_BACKEND = "cpp_lowcmd_cpp_lowstate"
 SPORT_KILL_CMD = (
     'ssh pi@192.168.123.161 "sudo pkill -9 -f keep_sport_alive; '
@@ -217,6 +223,38 @@ def validate_joint_order():
         print(f"  [{i:02d}] {sdk_name:4s} -> {policy_name:8s} default={q0:+.3f}")
 
 
+def resolve_bpu_model(args):
+    if args.bpu_model:
+        return Path(args.bpu_model).expanduser().resolve()
+    key = str(args.bpu_round)
+    if key not in BPU_MODEL_REGISTRY:
+        choices = ", ".join(sorted(BPU_MODEL_REGISTRY))
+        raise ValueError(f"Unknown --bpu-round {args.bpu_round!r}; choices: {choices}")
+    return BPU_MODEL_REGISTRY[key].expanduser().resolve()
+
+
+def make_policy(args):
+    model_path = resolve_bpu_model(args)
+    policy = BpuInferLibPolicy(
+        model_path,
+        priority=args.bpu_priority,
+        bpu_cores=args.bpu_cores,
+    )
+    if policy.input_size != POLICY_INPUT_DIM:
+        policy.close()
+        raise ValueError(
+            f"BPU model input has {policy.input_size} values, expected {POLICY_INPUT_DIM}. "
+            "Use a RobotLab 10-frame/450-dim .bin."
+        )
+    if policy.output_size != NUM_ACTIONS:
+        policy.close()
+        raise ValueError(
+            f"BPU model output has {policy.output_size} values, expected exactly {NUM_ACTIONS}. "
+            "Compile an actions-only ONNX for deployment."
+        )
+    return policy
+
+
 def apply_deadzone(value, deadzone):
     if deadzone <= 0.0:
         return float(value)
@@ -346,7 +384,8 @@ class JsonlLogger:
             "num_obs": NUM_OBS,
             "history_len": HISTORY_LEN,
             "policy_backend": BpuInferLibPolicy.backend_name,
-            "bpu_model": str(Path(args.bpu_model).expanduser()),
+            "bpu_model": str(resolve_bpu_model(args)),
+            "bpu_round": args.bpu_round,
             "policy_input_dim": POLICY_INPUT_DIM,
             "bpu_input_shape": BPU_INPUT_SHAPE,
             "bpu_output_shape": BPU_OUTPUT_SHAPE,
@@ -750,7 +789,7 @@ def run_infer_check(args):
     logger = JsonlLogger(args.log_dir, args)
     client = None
     try:
-        policy = BpuInferLibPolicy(args.bpu_model, priority=args.bpu_priority, bpu_cores=args.bpu_cores)
+        policy = make_policy(args)
         client, state = connect_client(args)
         obs_builder = ObsHistoryBuilder()
         cmd_filter = CommandFilter(args)
@@ -845,7 +884,7 @@ def run_deploy(args):
     print(STARTUP_BANNER)
     input("Press Enter when ready...")
 
-    policy = BpuInferLibPolicy(args.bpu_model, priority=args.bpu_priority, bpu_cores=args.bpu_cores)
+    policy = make_policy(args)
     logger = JsonlLogger(args.log_dir, args)
     client = None
     state = None
@@ -1162,8 +1201,11 @@ def run_deploy(args):
 
 def build_arg_parser():
     parser = argparse.ArgumentParser(description="Deploy RoboGauge Go1 RobotLab BPU on Go1 PRO")
-    parser.add_argument("--bpu-model", default=str(DEFAULT_BPU_MODEL),
-                        help="Path to policy_robotlab_6500_int16_gemm.bin")
+    parser.add_argument("--bpu-model", default="",
+                        help="Path to a compiled RobotLab 10-frame BPU .bin; overrides --bpu-round")
+    parser.add_argument("--bpu-round", default=DEFAULT_BPU_ROUND,
+                        choices=sorted(BPU_MODEL_REGISTRY),
+                        help="Quick-select RobotLab BPU model round")
     parser.add_argument("--bpu-priority", type=int, default=0)
     parser.add_argument("--bpu-cores", type=int, nargs="+", default=[0],
                         help="Reserved BPU core ids for future runtime scheduling")

deploy_45dim_rl_gym/bpu_quantization/README.md

@@ -55,6 +55,30 @@ cd /Users/chenyouyuan/cyy_ws/deploy_go1_pro/deploy_45dim_rl_gym/bpu_quantization
 ./quantize_policy_x5.sh --samples 128 --compare-limit 128
 ```
 
+速度优先 int8 对照：
+
+```bash
+./quantize_policy_x5.sh --quant int8
+```
+
+本次 `policy_35k` int8 结果不建议直接实机使用：
+
+- 产物：`mapper_output_35k_int8_gemm/policy_35k_int8_gemm.bin`
+- `actions` cosine：`0.981343`
+- L1/L2/Chebyshev：`0.053474 / 0.018129 / 0.113345`
+- mapper 编译估计延迟：`314.2 us + 49.7 us`
+- 板端 `hrt_model_exec perf` 实测：`1.583870 ms`
+- 板端 C++ bench：`1.56478 ms`
+
+对比当前 int16：
+
+- `actions` cosine：`0.998524`
+- 板端 `hrt_model_exec perf` 实测：`1.401526 ms`
+- 板端 C++ bench：`1.37792 ms`
+
+结论：这个模型上默认 int8 既更慢，输出误差也更大。当前继续用
+`policy_35k_int16_gemm.bin` 更合理。
+
 Gym BPU 部署入口支持快速切换轮次：
 
 ```bash
@@ -105,25 +129,31 @@ cd /root/go1_pro_deploy/deploy_45dim_rl_gym/bpu_deploy_x5/cpp
   1000
 ```
 
-已经完成 `policy_robotlab_15000.onnx` 和 `policy_robotlab_6500.onnx` 的 int16
-量化。RobotLab BPU 部署默认仍使用 6500 版本：
+已经完成 `policy_robotlab_15000.onnx`、`policy_robotlab_6500.onnx` 和
+`policy_robotlab_26000.onnx` 的 int16 量化。RobotLab BPU 部署当前默认使用
+26000 版本：
 
-- 原始模型：`../policy_robotlab_6500.onnx`
+- 原始模型：`../policy_robotlab_26000.onnx`
 - 原始输入：`obs [1, 450]`
 - BPU 编译输入：`obs_4d [1, 1, 1, 450]`
 - BPU 输出：`actions [1, 12, 1, 1]`
 - Docker 镜像：`openexplorer/ai_toolchain_ubuntu_20_x5_cpu:v1.2.8`
 - `hb_mapper`：`1.24.3`
 - `hbdk`：`3.49.15`
-- 当前产物：`mapper_output_6500_gemm/policy_robotlab_6500_int16_gemm.bin`
+- 当前产物：`mapper_output_26000_gemm/policy_robotlab_26000_int16_gemm.bin`
 
 `mapper_output*/`、`.hb_check/`、校准数据、中间 ONNX、`hb_mapper` 日志都被
 `.gitignore` 忽略；需要时按下面步骤重新生成。仓库里只保留脚本和 YAML 配置。
 
-6500 量化使用 `calibration_data_fast64/` 的 64 个真实样本。原因是
+26000 量化使用 `calibration_data_26000_robotlab_fast64/` 的 64 个真实样本。原因是
 `hb_mapper` 会先尝试 calibration batch 8，但当前 4D featuremap 包装会被工具链
 内部改成固定 batch 的 reshape，batch 8 失败后会退回 batch 1；用 64 样本可以把
-校准时间从 512 次 batch1 显著降下来。
+校准时间压到可接受范围。
+
+如果你要回退到旧版：
+
+- `mapper_output_6500_gemm/policy_robotlab_6500_int16_gemm.bin`
+- `mapper_output_gemm/policy_robotlab_15000_int16_gemm.bin`
 
 ## 1. 生成校准数据
 

deploy_45dim_rl_gym/bpu_quantization/policy_robotlab_26000_int16_gemm.yaml

@@ -0,0 +1,27 @@
+model_parameters:
+  onnx_model: "./policy_robotlab_26000_bpu4d_gemm.onnx"
+  march: "bayes-e"
+  layer_out_dump: false
+  working_dir: "mapper_output_26000_gemm"
+  output_model_file_prefix: "policy_robotlab_26000_int16_gemm"
+
+input_parameters:
+  input_name: "obs_4d"
+  input_shape: "1x1x1x450"
+  input_type_rt: "featuremap"
+  input_layout_rt: "NCHW"
+  input_type_train: "featuremap"
+  input_layout_train: "NCHW"
+  norm_type: "no_preprocess"
+
+calibration_parameters:
+  cal_data_dir: "./calibration_data_26000_robotlab_fast64"
+  cal_data_type: "float32"
+  calibration_type: "default"
+  optimization: "set_all_nodes_int16"
+  per_channel: true
+
+compiler_parameters:
+  compile_mode: "latency"
+  debug: false
+  optimize_level: "O3"

deploy_45dim_rl_gym/bpu_quantization/quantize_policy_x5.sh

@@ -12,9 +12,11 @@ FLAT_DIM=""
 SAMPLES=64
 MIN_SAMPLES=32
 LOG_PREFIX="rlgym_go1_deploy"
+CAL_TAG="gym"
 DOCKER_IMAGE="openexplorer/ai_toolchain_ubuntu_20_x5_cpu:v1.2.8"
 COMPARE_LIMIT=64
 RUN_CHECKER=1
+QUANT="int16"
 
 usage() {
   cat <<'EOF'
@@ -32,8 +34,10 @@ Options:
   --samples N            calibration sample count; default 64 for faster mapping
   --min-samples N        minimum valid samples required; default 32
   --log-prefix PREFIX    log dir prefix below logs/, default rlgym_go1_deploy
+  --cal-tag TAG          calibration dir tag, default gym
   --docker-image IMAGE   D-Robotics CPU toolchain image
   --compare-limit N      float ONNX equivalence sample count, default 64
+  --quant int16|int8     int16 keeps current conservative path; int8 omits set_all_nodes_int16
   --skip-checker         skip hb_mapper checker before makertbin
 EOF
 }
@@ -48,8 +52,10 @@ while [[ $# -gt 0 ]]; do
     --samples) SAMPLES="$2"; shift 2 ;;
     --min-samples) MIN_SAMPLES="$2"; shift 2 ;;
     --log-prefix) LOG_PREFIX="$2"; shift 2 ;;
+    --cal-tag) CAL_TAG="$2"; shift 2 ;;
     --docker-image) DOCKER_IMAGE="$2"; shift 2 ;;
     --compare-limit) COMPARE_LIMIT="$2"; shift 2 ;;
+    --quant) QUANT="$2"; shift 2 ;;
     --skip-checker) RUN_CHECKER=0; shift ;;
     -h|--help) usage; exit 0 ;;
     *) echo "Unknown argument: $1" >&2; usage >&2; exit 2 ;;
@@ -60,6 +66,11 @@ if [[ -z "${FLAT_DIM}" ]]; then
   FLAT_DIM=$((45 * HISTORY_LEN))
 fi
 
+case "${QUANT}" in
+  int16|int8) ;;
+  *) echo "--quant must be int16 or int8, got: ${QUANT}" >&2; exit 2 ;;
+esac
+
 if [[ "${POLICY}" = /* ]]; then
   POLICY_ABS="${POLICY}"
 else
@@ -80,9 +91,14 @@ if [[ -z "${NAME}" ]]; then
   NAME="$(basename "${POLICY_ABS}" .onnx)"
 fi
 
-CAL_DIR="calibration_data_${ROUND}_gym_fast${SAMPLES}"
-OUTPUT_DIR="mapper_output_${ROUND}_gemm"
-OUTPUT_PREFIX="${NAME}_int16_gemm"
+CAL_DIR="calibration_data_${ROUND}_${CAL_TAG}_fast${SAMPLES}"
+if [[ "${QUANT}" = "int16" ]]; then
+  OUTPUT_DIR="mapper_output_${ROUND}_gemm"
+  OUTPUT_PREFIX="${NAME}_int16_gemm"
+else
+  OUTPUT_DIR="mapper_output_${ROUND}_int8_gemm"
+  OUTPUT_PREFIX="${NAME}_int8_gemm"
+fi
 YAML_FILE="${OUTPUT_PREFIX}.yaml"
 
 echo "[INFO] repo          : ${REPO_ROOT}"
@@ -91,6 +107,7 @@ echo "[INFO] name/round    : ${NAME} / ${ROUND}"
 echo "[INFO] history/shape : ${HISTORY_LEN} / 1x1x1x${FLAT_DIM}"
 echo "[INFO] calibration   : ${CAL_DIR} (${SAMPLES} samples, prefix ${LOG_PREFIX})"
 echo "[INFO] output        : ${OUTPUT_DIR}/${OUTPUT_PREFIX}.bin"
+echo "[INFO] quant         : ${QUANT}"
 
 docker run --rm --platform linux/amd64 \
   -e POLICY_REL="${POLICY_REL}" \
@@ -106,6 +123,7 @@ docker run --rm --platform linux/amd64 \
   -e YAML_FILE="${YAML_FILE}" \
   -e COMPARE_LIMIT="${COMPARE_LIMIT}" \
   -e RUN_CHECKER="${RUN_CHECKER}" \
+  -e QUANT="${QUANT}" \
   -v "${REPO_ROOT}:/workspace/deploy_go1_pro" \
   "${DOCKER_IMAGE}" \
   bash -lc '
@@ -151,6 +169,12 @@ docker run --rm --platform linux/amd64 \
       --flat-dim "${FLAT_DIM}" \
       --limit "${COMPARE_LIMIT}"
 
+    if [[ "${QUANT}" = "int16" ]]; then
+      OPTIMIZATION_LINE="  optimization: \"set_all_nodes_int16\""
+    else
+      OPTIMIZATION_LINE=""
+    fi
+
     cat > "${YAML_FILE}" <<YAML
 model_parameters:
   onnx_model: "./${GEMM_ONNX}"
@@ -172,7 +196,7 @@ calibration_parameters:
   cal_data_dir: "./${CAL_DIR}"
   cal_data_type: "float32"
   calibration_type: "default"
-  optimization: "set_all_nodes_int16"
+${OPTIMIZATION_LINE}
   per_channel: true
 
 compiler_parameters:

deploy_45dim_rl_gym/deploy_go1_onnx_mujoco_lab.py

@@ -43,7 +43,7 @@ from mujoco import viewer
 
 # ── path setup ──
 SCRIPT_DIR = Path(__file__).resolve().parent
-DEFAULT_ONNX = str(SCRIPT_DIR / "policy_robotlab_15000.onnx")
+DEFAULT_ONNX = str(SCRIPT_DIR / "policy_robotlab_26000.onnx")
 ROBOT_XML = str(SCRIPT_DIR / "go1.xml")
 TERRAINS_DIR = SCRIPT_DIR / "terrains"