Add auto test script.
Add experimental FP8 1.4.3/1.5.2 model, run simple mnist model successed.
Rename arch_O0 to arch_cpu, Move O1,O2 to TM_OPT_LEVEL setting, start write tm_layer_O1.c
Add Support for RV64V's FP16 model, use RVV1.0 inntrinsic
Optimize reduce sum with RVV's vfredosum_vs_f32m1_f32m1
speed up TM_ARCH_RV64V infer mbnet v1 0.25:
| model type | before | after | speedup |
|---|---|---|---|
| INT8 | 212ms | 178ms | 19% |
| FP32 | 190ms | 182ms | 4% |
| FP16 | 178ms | 161ms | 11% |
tested infer time duty: sbuf:dot:sum ~ 9:5:2
need opt sbuf time
First commit to github, support INT8/FP32 model, support arch:
arch_O0,
arch_ARM_SIMD, arch_ARM_NEON,arch_ARM_MVEI,
arch_RV32P, arch_RV64V
support OPS:
Conv2D,DWConv2D,(with relu/relu6,stride)
FC,GAP,SOFTMAX
support demos:
layer_test
mnist
mnist port for Arduino ATmega328
mbnet(mobilenet v1)