Pool2D gives bad output for Integer inputs

Troubleshooting
#1 
def manual_test():
    data_shape=(1, 1, 2, 2)
    data_dtype='int32'
    data = relay.var("data", shape=data_shape,
            dtype=data_dtype)
    pool = relay.op.nn.avg_pool2d(data, pool_size=(2, 2))
    func = pool
    func = run_infer_type(func)
    func = relay.Function(relay.analysis.free_vars(func),
                func)
    print(func)
    with relay.build_config(opt_level=0):
        graph, lib, params = relay.build(func, "llvm", params=None)
        mod = graph_runtime.create(graph, lib, ctx=tvm.cpu(0))
        golden_data = np.array([5, 5, 5, 5]).reshape(data_shape).astype(data_dtype)
        mod.set_input("data",golden_data)
        mod.run()
        res = mod.get_output(0).asnumpy()
        print(res)

The above gives output 4 instead of 5.

Setting the golden_data to (1, 1, 1, 1) gives output 0.

I think we are dividing before accumulating. Since, we have only tested on Float32, this never showed up before. But, in int32, it performs int division.

@tqchen @yzhliu

#2

I tried to further debug this and got stuck. I changed the topi computein following manner


-      return tvm::sum(temp(indices) / divide_factor, { dheight, dwidth });
+      auto reduced_hw = tvm::sum(temp(indices), { dheight, dwidth });
+      return topi::divide(reduced_hw, divide_factor);
+      //return tvm::sum(temp(indices) / divide_factor, { dheight, dwidth });
     };

I got the following error

tvm._ffi.base.TVMError: Traceback (most recent call last):
  [bt] (8) /home/ubuntu/workplace/t1/tvm/build/libtvm.so(tvm::IRFunctor<void (tvm::NodeRef const&, tvm::ir::IRVisitor*)>::operator()(tvm::NodeRef const&, tvm::ir::IRVisitor*) const+0x162) [0x7fb3d791c47e]
  [bt] (7) /home/ubuntu/workplace/t1/tvm/build/libtvm.so(std::function<void (tvm::NodeRef const&, tvm::ir::IRVisitor*)>::operator()(tvm::NodeRef const&, tvm::ir::IRVisitor*) const+0x61) [0x7fb3d791ce35]
  [bt] (6) /home/ubuntu/workplace/t1/tvm/build/libtvm.so(std::_Function_handler<void (tvm::NodeRef const&, tvm::ir::IRVisitor*), tvm::IRFunctor<void (tvm::NodeRef const&, tvm::ir::IRVisitor*)>& tvm::IRFunctor<void (tvm::NodeRef const&, tvm::ir::IRVisitor*)>::set_dispatch<tvm::ir::Reduce>(std::function<void (tvm::ir::Reduce const*, tvm::ir::IRVisitor*)>)::{lambda(tvm::NodeRef const&, tvm::ir::IRVisitor*)#1}>::_M_invoke(std::_Any_data const&, tvm::NodeRef const&, tvm::ir::IRVisitor*&&)+0x4f) [0x7fb3d7c90d01]
  [bt] (5) /home/ubuntu/workplace/t1/tvm/build/libtvm.so(tvm::IRFunctor<void (tvm::NodeRef const&, tvm::ir::IRVisitor*)>& tvm::IRFunctor<void (tvm::NodeRef const&, tvm::ir::IRVisitor*)>::set_dispatch<tvm::ir::Reduce>(std::function<void (tvm::ir::Reduce const*, tvm::ir::IRVisitor*)>)::{lambda(tvm::NodeRef const&, tvm::ir::IRVisitor*)#1}::operator()(tvm::NodeRef const&, tvm::ir::IRVisitor*) const+0x54) [0x7fb3d7c7f376]
  [bt] (4) /home/ubuntu/workplace/t1/tvm/build/libtvm.so(std::function<void (tvm::ir::Reduce const*, tvm::ir::IRVisitor*)>::operator()(tvm::ir::Reduce const*, tvm::ir::IRVisitor*) const+0x61) [0x7fb3d7c88785]
  [bt] (3) /home/ubuntu/workplace/t1/tvm/build/libtvm.so(+0x204b3b1) [0x7fb3d7c753b1]
  [bt] (2) /home/ubuntu/workplace/t1/tvm/build/libtvm.so(+0x2046642) [0x7fb3d7c70642]
  [bt] (1) /home/ubuntu/workplace/t1/tvm/build/libtvm.so(+0x1f66570) [0x7fb3d7b90570]
  [bt] (0) /home/ubuntu/workplace/t1/tvm/build/libtvm.so(dmlc::LogMessageFatal::~LogMessageFatal()+0x34) [0x7fb3d78729f6]
  File "/home/ubuntu/workplace/t1/tvm/src/op/compute_op.cc", line 584
TVMError: Check failed: 0 == level_: Reductions are only allowed at the top level of compute. Please create another tensor for further composition.

Can anybody help me with this?

#3

I guess we can create a tvm::compute doing sum first, like pad, then do divide

#4

Should this get a github issue for fixing ?

#5

This is already fixed. I will find the commit and paste it here.

#6

Nice, in which case the test in test_op_qnn_conv2d.py needs to be updated ?

#7

Just saw the test. I fixed the test case but forgot to remove the comment. Nice catch :slight_smile: