Some questions wanna disscuss

[xqdan]

Hi,

Nice work! after went thru the code, some questions want to ask,

• How to lower ops like batchnorm to meta ops, tf xla has tf2xla phase to lower big op into a group of meta ops, I don't see the related code in jittor for this, do I miss anything here?

• How many ops is using meta ops, how many is supported by extern library? take restnet50 as example.

• How to do auto schedule for ops like conv, gemm, could you elaborate with a specfic case?

Thanks!

[Gword]

Thanks for your feedback.

• Jittor's op are all composed of meta op, you can see their implementation method in nn.py. The implementation of batchnorm can refer to the class BatchNorm in nn.py.
• In resnet50, only conv op uses cudnn for calculation.
• Conv, gemm and other algorithm hardware libraries have been optimized very well, we will forward the calculation process. And unsupported op will use some general optimization strategies, such as remove atomic.

[xqdan]

Thanks, make sense. more questions,

• notice that for registered ops, only infer shape is provided, do we need infer type? do you consider employ a type system to save infer and optimize graph optimizations?

• How to do auto schedule for unsupported op, I mean auto schedule is none trival, I wanna know the solution jittor is using.

• about vectorization, seems the capability of traditional compiler is limited, is it enough to only emit pragma for LLVM?

Thanks

[xmyqsh]

@xqdan shape could be inferred like this y = create_output(nullptr, x->dtype()); .

@Gword Where has been set resnet50's conv op to use cudnn? I can only refer to the nn.py implementation. How does the extern implementation to be selected?

[jackmsye]

@xmyqsh same confused as me. i saw some example and tests in python/jittor /test/test_cudnn_op.py, but I can not find how they triggle cudnn, the only flag in use_cuda=1

[xmyqsh]

@jackmsye I have got it!

267         string relay_conv_name = fop->flags.get(NodeFlags::_cpu) ?
268             "mkl_conv" : "cudnn_conv";

[jackmsye]

@xmyqsh thx, they just put on conv_tunner

[xmyqsh]

Some tuners use TunerManager, some tuners use PassManager, while others use both. Do you know why?

[jackmsye]

first they use jit compiler, when it is ops, it executes tunner_manager, you can see the code in src/ops_compiler.cc, then in TunnerManager, the function run_tunner's member is PassManager. see code in tuner_manager.cc

[xmyqsh]

@jackmsye Exactly! Can you give a brief summary of TunnerManager and PassManager?

    run_tuner<ReorderTuner>(&pm);
    run_tuner<BroadcastTuner>(&pm);
    run_tuner<ReduceTuner>(&pm);
    run_tuner<MatmulTuner>(&pm);
    run_tuner<ConvTuner>(&pm);

    run_pass<MarkRawPass>();
    run_pass<ReplaceForNumPass>();
    run_pass<LoopVarAnalyzePass>();
    run_pass<RemoveLoopPass>();
    run_pass<RenameLoopIndexPass>();
    run_pass<CompileShapesPass>();
....
    run_pass<SplitLoopPass>();
    run_pass<ReorderLoopPass>();
    run_pass<MergeLoopPass>();
    run_pass<ExpandEmptyBlockPass>();
    run_pass<SolveConflictDefinePass>();

    run_pass<RemoveIntermediatePass>();
....
    run_pass<SolveConflictDefinePass>();

    run_pass<RestridePass>();
....
    if (cc_type == "icc") {
        // only icc supports pragma
        run_pass<VectorizePass>();
        run_pass<UnrollPass>();
        run_pass<UnrollPass>();
    }
    run_pass<UseMovntPass>();
    run_pass<CheckCachePass>();
    run_pass<LoopToFuncPass>();
    run_pass<AssumeAlignedPass>();
    run_pass<ParallelPass>();
    run_pass<AtomicTunerPass>();
    run_pass<FloatAtomicFixPass>();
....
    run_pass<InsertProfileLoopPass>();
....
    run_pass<SolveConflictDefinePass>();
....
    run_pass<FakeMainPass>();

[Jittor]

Thanks @xmyqsh , we are polishing our backend documents and will be released soon~