grid sample 耗时

已解决

tsheng2023-05-10

j5 上 grid sample 输入的 quantize 很耗时，如何剔除或者优化？

"latency (ms) by segments": [
4.053
],
"latency (us)": 4052.9,
"layer details": [
[
"layer",
"original ops",
"original cost",
"computing cost (no DDR)",
"layer util (no DDR)",
"load/store cost",
"layer util (with DDR)",
"active period of time"
],
[
"arg1_torch_native",
"0",
"34 us (57.1% of model)",
"2204 us (54.3% of model)",
"1.5%",
"122 us (3.0% of model)",
"1.4%",
"373 ~ 2827 us (2454)"
],
[
"_sample_hz_grid_sample_grid_rescale_torch_native",
"0",
"9 us (14.2% of model)",
"9 us (0.2% of model)",
"93.4%",
"1 us (0% of model)",
"84.8%",
"2813 ~ 3798 us (985)"
],
[
"_sample_hz_grid_sample",
"0",
"17 us (28.5% of model)",
"1213 us (29.9% of model)",
"1.4%",
"534 us (13.1% of model)",
"0.9%",
"0 ~ 4053 us (4053)"
]
],

算法工具链

征程5