using GPU

[shockjiang]

the current code is quite slow, which takes about 20 seconds to generate a picture. So I speed it up by using GPU. I set the HL_TARGET and other environment parameters like the following:

target=x86-64-linux-opencl
cwd=$(pwd)
export HL_TARGET=$target
export HL_JIT_TARGET=$target
export HL_DEBUG_CODEGEN=1
export HL_TRACE=1
export HL_NUM_THREADS=1024
export HL_TRACE_FILE=$cwd/trace.log
export HL_PROFILE=1

And run the hdrplus. However, this does not work at all. Any hint to fix it?

[Titaniumtown]

Any progress?

[suhaasr]

Hi!

I'd love to help, but Tim and I are both in industry now, so we're unable to actively work on this to avoid conflicts of interest. We're currently working on handing the project over to someone who expressed interest in polishing the pipeline.

Best,
Suhaas

[Titaniumtown]

It was handed off to me.

[brotherofken]

That's not easy and requires some refactoring. Here is the steps:

1. At first, we have to move to Halide's Generators API as described here in this tutorial in order to rid of JIT compilation which takes a lot of runtime.
2. Second we have to add special "gpu_tile" scheduling across the pipeline, like it was described in this tutorial.