Re: [apache/incubator-tvm] [DISCUSS][RFC] Apache TVM Graduation (#6299)
+1 from me as well! -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/apache/incubator-tvm/issues/6299#issuecomment-675857185
[TVM Discuss] [Development] OpenCL AutoTVM questions
I'm trying to enable AutoTVM for OpenCL (`intel_graphics` target). So far I managed to have some success in that area, but values are multiple times worst than for generic scheduler. To begin with I am focusing only on conv2d operation (since this is also only one currently present in `intel_graphics` TOPI). I've used `conv2d_direct.py` file from CUDA to use it as a dummy test file (this scheduler seemed to be the easiest) and get some idea what is required to write my own one. There are few things I don't understand and I'd appreciate guidance on how such scheduler should be written/what values should I provide. Two most pressing questions for now are: 1. From where you have all of the splitting and other numerical values in `schedule_direct_cuda`? 1. How you decided about `tvm.thread_axis` threads? --- [Visit Topic](https://discuss.tvm.ai/t/opencl-autotvm-questions/2665/1) to respond. You are receiving this because you enabled mailing list mode. To unsubscribe from these emails, [click here](https://discuss.tvm.ai/email/unsubscribe/b527b1296934833ce41beda0c13583fae846f8fa9bbf7a803e07670fa3433d6b). Tianqi Chen, UW, Seattle, WA, 98105, United States http://tracking.discuss.tvm.ai/tracking/unsubscribe?msgid=FN0H3-unCsVvE4avJnv8jg2
[TVM Discuss] [Development] OpenCL AutoTVM questions
Thank you for answer. I'll check those materials. --- [Visit Topic](https://discuss.tvm.ai/t/opencl-autotvm-questions/2665/6) to respond. You are receiving this because you enabled mailing list mode. To unsubscribe from these emails, [click here](https://discuss.tvm.ai/email/unsubscribe/7c084339c93cb690a70dfebd6c49c2b87767bb0ecaef799f029e83b380ee3f16). Tianqi Chen, UW, Seattle, WA, 98105, United States http://tracking.discuss.tvm.ai/tracking/unsubscribe?msgid=jwL7NIOMWiZuT5-nA4kSvA2
[TVM Discuss] [Development] OpenCL AutoTVM questions
Thank you @Laurawly for info. Would you be able to explain a bit more on the topic? Especially I'd like to better understand which parts of the intel_graphics scheduler are getting main benefits from subgroups? --- [Visit Topic](https://discuss.tvm.ai/t/opencl-autotvm-questions/2665/8) to respond. You are receiving this because you enabled mailing list mode. To unsubscribe from these emails, [click here](https://discuss.tvm.ai/email/unsubscribe/a3d2c9d00c433bffced20dc6da67abfcee40623b29a0ed97f340f88fd2938446). Tianqi Chen, UW, Seattle, WA, 98105, United States http://tracking.discuss.tvm.ai/tracking/unsubscribe?msgid=K6V4nmcgilfkimyxa3U91A2
[TVM Discuss] [Development] Use block data format for whole model
I'd like to ask if there is any plan or any simple way to enable block (NCHWc) data format for whole model instead of plain (NCHW)? After some playing with TVM it seems that the preferred data type is plain. Significant exceptions are for Intel CPU and Intel Graphics scheduler which are capable of converting from plain to block and back, but it's not optimal as overall approach. It appears that Intel HW prefers block data format and it would be profitable for performance to remove those conversions all together. The idea was to add on the model reading stage decision if model should be converted to block or plain, but it requires a lot of other changes down the compilation pipeline. Could anyone comment here what's the plan or solution? --- [Visit Topic](https://discuss.tvm.ai/t/use-block-data-format-for-whole-model/3693/1) to respond. You are receiving this because you enabled mailing list mode. To unsubscribe from these emails, [click here](https://discuss.tvm.ai/email/unsubscribe/7253a99ac35bd9126ce8395a7f34abfc84720c4378b8fb733f941bfc197795d8). Tianqi Chen, UW, Seattle, WA, 98105, United States http://tracking.discuss.tvm.ai/tracking/unsubscribe?msgid=ht_vxQhU7PXcyVoX-puCVQ2
[TVM Discuss] [Development] Use block data format for whole model
Ok, if I understood this example correctly here is a way to take advantage of NCHWc in case of model already being in this format. What I'm looking for is more universal approach, allowing to convert whole model to NCHWc **and not each operation**. --- [Visit Topic](https://discuss.tvm.ai/t/use-block-data-format-for-whole-model/3693/3) to respond. You are receiving this because you enabled mailing list mode. To unsubscribe from these emails, [click here](https://discuss.tvm.ai/email/unsubscribe/6b2d6cc1a5dbb8d55f258e4e335034c7bc36d840d9aefd582ea5eda1e066fb12). Tianqi Chen, UW, Seattle, WA, 98105, United States http://tracking.discuss.tvm.ai/tracking/unsubscribe?msgid=TgxFA7tQMlaj0l-EEGQIEQ2
[Apache TVM Discuss] [Development/pre-RFC] Export TIR to json
We propose to implement an API to TVM that allows users to export the TIR representation of the generated kernels as json. **Motivation** Json is a standard file format and can easily be processed in python, for example. A json interface allows to analyze the generated kernels target agnostic. For example, this works for both llvm and c targets. The goal that we have is to be able to extract the full TIR representation of the host module and all device modules regardless of the specific executor, target or runtime. In the case of the AOT executor pipeline, this would also contain the tvmgen_default___main__ function. This allows to reconstruct the graph topology of the neural network. For Graph executors, the graph can still be accessed via module.graph_json. **Proposed Implementation** We have already conducted some experiments with a JsonScriptPrinter class similar to the TVMScriptPrinter class in src/printer/tvmscript_printer.cc based on the discussion here: https://discuss.tvm.apache.org/t/why-c-unpacked-api-not-work-anymore/11845/19?u=sebastianboblestetas We use it in driver_api.cc in the build function like this:  We make it accessible in python via json = module.lib.get_json() We do not know how to best export the TIR representation of the device modules. We would add a new boolean tir configuration option “tir.tir_export_to_json” to make this export functionality optional. We think we can provide a working prototype within a few weeks but would appreciate feedback already now. @Khoi @MJKlaiber @UlrikHjort @areusch --- [Visit Topic](https://discuss.tvm.apache.org/t/export-tir-to-json/12329/1) to respond. You are receiving this because you enabled mailing list mode. To unsubscribe from these emails, [click here](https://discuss.tvm.apache.org/email/unsubscribe/ad7414c9b68f7bcc27f4bec4ecd3a06bf9c541b3379701916706cd8c6ba59ff0).
[Apache TVM Discuss] [Development/pre-RFC] Export TIR to json
Hi,
thanks for the reference to this function. I was not aware of it.
However I tried it on a Module(c) and got this:
`{
"root": 1,
"nodes": [
{
"type_key": ""
},
{
"type_key": "runtime.Module"
}
],
"b64ndarrays": [],
"attrs": {"tvm_version": "0.9.dev0"}
}`
This looks like only a very small overview. What we would like to have is
really the full representation that the code generators get to write the code.
Do I miss an option for this function to get more detailed output?
---
[Visit Topic](https://discuss.tvm.apache.org/t/export-tir-to-json/12329/4) to
respond.
You are receiving this because you enabled mailing list mode.
To unsubscribe from these emails, [click
here](https://discuss.tvm.apache.org/email/unsubscribe/1a2fe36d8136482a9d7d6bbc57d128f5c2f2e7786a8b1ddbe6274231e3ce942d).
[Apache TVM Discuss] [Development/pre-RFC] Export TIR to json
1. You are right, we actually tried to save the runtime.Module. We actually also added the TIR export to the runtime.Module because the generated source code is also put to it. We will try the different options you proposed as soon as we can find the time. 2. In the meantime we have written an early prototype of our export. I attached a small sample image of what it looks like, compared to what save_json gives for the same IRModule. I do not think that save_json gives us the required level of detail but to be honest, I did not look at it in great detail.   3. We did not yet look into this, sorry 4. As you can probably see, we want to extract TIR just before the code generator is invoked. We would like to get all the details that also go into the final code generation step. 5. We want to make this useful for as many use cases as possible, so if we can make the point in time of the export configurable, we are definitely in favor of doing so. I am not sure however, how much this will affect the implementation of the TIR export printer. It might then possibly need configuration options as well, right? --- [Visit Topic](https://discuss.tvm.apache.org/t/export-tir-to-json/12329/7) to respond. You are receiving this because you enabled mailing list mode. To unsubscribe from these emails, [click here](https://discuss.tvm.apache.org/email/unsubscribe/c34f12263ea2c2c94ded8aebd5c47c9e0db77f4e83ee06fc1443dfe1f2087864).
