Xavier Hsinyuan

Results 12 comments of Xavier Hsinyuan

Failed to pass tests when build with [riscv-gcc/riscv-gcc-rvv-next](https://github.com/riscv-collab/riscv-gcc/tree/riscv-gcc-rvv-next)(edffbea), binutils-2.39 : ``` test_convolution test_convolution1d test_convolutiondepthwise test_deconvolution test_deconvolutiondepthwise test_squeezenet ```

> Failed to pass tests when build with [riscv-gcc/riscv-gcc-rvv-next](https://github.com/riscv-collab/riscv-gcc/tree/riscv-gcc-rvv-next)(edffbea), binutils-2.39 : Succeed with clang-14 with following requirement to build: * Append some type definitions in `riscv-vector.h`; * `clang++` complains VLAs;...

> Failed to pass tests when build with [riscv-gcc/riscv-gcc-rvv-next](https://github.com/riscv-collab/riscv-gcc/tree/riscv-gcc-rvv-next)(edffbea), binutils-2.39 Use [riscv-gcc/riscv-gcc-rvv-next](https://github.com/riscv-collab/riscv-gcc/tree/riscv-gcc-rvv-next)(32c7d7c), binutils-2.39 instead.

I built a new gcc from [riscv-collab/riscv-gcc.git:riscv-gcc-rvv-next](https://github.com/riscv-collab/riscv-gcc/tree/riscv-gcc-rvv-next)(edffbea) with upstream binutils, and it seems that it doesn't require a lot of memory with `make linux -j 8`. And it seems that...

> Hello, there already have tuple type version of segment load/store intrinsics on GCC implementation. Thanks for your reply. It seems that clang-14 doesn't have the tuple version of segment...

> The "rvv-next" is the latest actively maintaining branch and will be eventually push GCC FSF upstream. It support tuple types too. So I would suggest you use the new...

> For fp32 pack1, we always embed the simd-free scalar code implementation at the same time, which is convenient for reference and helps us write optimized code for scalar in...

Fix: poor colorizing results by using pre-trained weights if cuda is disabled.

> > > This patch just resolve model loading, "CPU gray" stands there, not change any more. I've tested on both my laptop and Google CoLab by using pretrained weight...

> > > Yes, I do. infer box first and then run fusion. But I did this on my windows and cpu model. Do it matter. > And I as...