Check dialect changes

[KavithaTipturMadhu]

Compile the checks into calls to an assert at runtime only, and not lower the whole operator into a runtime call : This avoids weakly typing the runtime utilities in order to support different datatypes.

[chelini]

Thanks, let me dump what we discussed.
check.expect_almost_eq(%arg0, %arg1, %arg3) : tensor<5xf32>, tensor<5xf32>, f32

Would be nice if we can lower this op as:

for (int i = 0; i < 5; i++) {
  %load1 = memref.load %arg0[i]
  %load2 = memref.load %arg1[i]
  if (abs(%load1 - %load2) > %th)
    @call_assert()
}

[rengolin]

Some considerations:

• We have expect_true, probably should also have expect_false
• We have expect_almost_eq, should also have expect_eq (as in identical), for checking blocking and copies
• We need to decide what to do for different precision (f32 vs bf16)
• Do we also need expect_ne? expect_gt, expect_lt, etc?
• Do we want to print the tensor (at least the region) when they're not eq / almost_eq?
• What is better:
  • check.expect_almost_eq(%a, %b, %delta) or
  • %0 = check.compare_fp((%a, %b, %delta); check.expect_true(%0)? (%delta = 0 for identical)
• If the former, then we may want to expand on the number of expects
• If the latter, we may only have one check (expect_true, perhaps expect_false too), and use arith for everything else.
• I still don't like the name expect_almost_eq. I think the existence of a %delta is enough to convey the meaning that it's an imprecise comparison...