sophgo / tpu-mlir

Machine learning compiler based on MLIR for Sophgo TPU.
Other
611 stars 153 forks source link

1684详细的转换流程 #124

Closed ijiami-01 closed 1 year ago

ijiami-01 commented 1 year ago

如题,我在model_deploy.py将chip从1684x改1684,发现转换报错,失败。 阅读过 https://github.com/sophgo/tpu-mlir/issues/102 ,但是说的不是很清晰。 以下是我的报错上下文。

/workspace/model_yolov5s/workspace# model_deploy.py \ --mlir yolov5s.mlir \ --quantize F16 \ --chip bm1684 \ --test_input yolov5s_in_f32.npz \ --test_reference yolov5s_top_outputs.npz \ --model yolov5s_1684x_f16.bmodel SOPHGO Toolchain v1.2.beta.0-20230705 [Running]: tpuc-opt yolov5s.mlir --chip-assign="chip=bm1684" --chip-top-optimize --convert-top-to-tpu="mode=F16 asymmetric=False" --canonicalize -o yolov5s_bm1684_f16_tpu.mlir Not Implemented UNREACHABLE executed at ../include/tpu_mlir/Conversion/TopToTpu/TopLowering.h:172! PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace.

sdu-gyf commented 1 year ago

1684不支持f16,只支持f32和int8