Files
Robot_Arm/docs/experiment_round34_f32_cert.json
T
cbb 296cd17f98 exp: 轮次35——FP16/BF16与流水线复用都被实测否掉,f32认证需先抬eps到2e-4
- 显卡精度档(B=1,K=8,steps=1):推理 FP32 7.68 / FP16 8.31 / BF16 8.58 ms,整周期不变而FP16 P99 46.6→55.6;
  动作偏差 FP16 0.0050°、BF16 0.0377° ⇒ 该批量下不是算力受限,保持 FP32
- float32 认证:48段×9点×15组对的间隙差 max 1.32e-4 mm(有1例高估)→ 现 eps=1e-4 已被越界,要用需 ≥2e-4;
  判据结果两边一致(各24接受、0 仅f32安全、0 符号翻转),暂不放宽 dtype 守卫
- 流水线复用特征间隙:第一版42.1→31.8ms 但等价性检查不过(我取缓存行索引取错,差3.19mm);
  修正后只剩 −3.1ms 且缓存行与新算差 6.3e-5mm(不同 batch 的结合序)⇒ 放弃
- 顺带修掉轮次34日志里重复的章节标题
2026-10-20 21:01:04 +08:00

41 lines
1.1 KiB
JSON

{
"kin_fp": "2ebcbf17",
"segments": 48,
"n_subsegments": 8,
"wide_cap_mm": 2.0,
"gap_abs_err_mm": {
"p50": 0.0,
"p99": 3.661949063912929e-05,
"p999": 7.4360332600075e-05,
"max": 0.00013247109886016517
},
"signed_err_mm": {
"min": -0.00010937313134817828,
"max": 0.00013247109886016517,
"overestimate_count": 1,
"overestimate_gt_1e-3": 0
},
"sign_flips": 0,
"accepted_f64": 24,
"accepted_f32": 24,
"unsafe_direction_f32_only": 0,
"conservative_f64_only": 0,
"per_group_pair_err_p99": [
0.0,
7.6e-05,
0.0001,
7e-05,
7.1e-05,
0.0,
0.0,
0.0,
0.000132,
6.7e-05,
5.1e-05,
0.0,
0.0,
0.0,
0.0
],
"note": "wide_cap 2.0 mm\uff1a\u8ba4\u8bc1\u5b9e\u9645\u7528 bound+2eps \u7684\u81ea\u9002\u5e94 cap\uff0c\u8fd9\u91cc\u53d6\u56fa\u5b9a 2 mm \u4ee5\u8986\u76d6\u88ab\u68c0\u9a8c\u7684\u6bb5\uff1b\u82e5 max \u8bef\u5dee\u8fdc\u5c0f\u4e8e 1e-4 \u7684 eps\uff0c\u5224\u636e\u5728 f32 \u4e0b\u4ecd\u6210\u7acb\u9700\u518d\u786e\u8ba4 f32 \u7684\u95f4\u9699\u672c\u8eab\u8bef\u5dee\u7b26\u53f7\u65b9\u5411\u3002"
}