Halide-Issues

[jinderek]

Just leave some ideas and issues that may not be very mature now, so I did not mention directly under the official Halide.

• The official documentation does not provide detailed information on how to set up HL_TARGET (We can only see one or two examples, such as x86-64-cuda/host-cuda/host-debug/cuda).
• Tensor Core Support
• Python User Friendly
• Halide deploy. (Dependency requirement is too high. Maybe AOT is an option, but how to scale)
• CUDA 8.0 SUPPORT： https://docs.nvidia.com/cuda/volta-tuning-guide/index.html提到了volta

[jinderek]

AOT - GPU - Tutorial issue: https://github.com/halide/Halide/issues/967 some ref: https://github.com/halide/Halide/blob/master/test/generator/gpu_only_aottest.cpp

Though gpu_only_aottest is a ref for how to do gpu aot, but it still manupilate cpu malloc. We can add a pure gpu version.

    Buffer<float> ib(nullptr, n, k);
    ib.device_wrap_native(interface, (uintptr_t)data);
    ib.set_device_dirty(true);
...

[jinderek]

Shared Memory

Halide似乎只能分配一个shared memory实例。 如果分配多个，似乎是公共一个。

需要在Halide的pass中进行检查，如果多个则报错。

复现：用custom pass复现