[CodeGen] Incorrect vectorization for fp16 on X86-64 (AVX512)

Hi,

There seems a bug related to vectorization with type fp16 on skylake machine. Basically, when cast fp16 to uint8 using select for AVX512 machine, it will generate wrong result with vectorization on.

The generated ll code seems correct after checking dumped llvm code, but the generated asm code seems problematic. It might be some configurations when doing the code generation /src/codegen/llvm/llvm_common.cc

import tvm
import numpy
## shape cause wrong result
m=3
n=2
k=2
## shape cause LLVM ERROR
# m=3
# n=1
# k=4

dtype = "float16"
target = 'llvm -mcpu=skylake-avx512'
ctx = tvm.context(target, 0)

input = [0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0]
in_array = numpy.array(input).reshape(m,n,k).astype(dtype)
a = tvm.nd.array(in_array, ctx)

X=tvm.placeholder((m,n,k), name='X', dtype="float16")
zero = tvm.const(0, "uint8")
one = tvm.const(1, "uint8")
z16 = tvm.const(0, "float16")

Y=tvm.compute((m,n,k), lambda i,j,k: tvm.expr.Select(X[i,j,k] == z16, zero, one))
s=tvm.create_schedule(Y.op)
s[Y].vectorize(Y.op.axis[2])
print tvm.lower(s,[X,Y], simple_mode=True)
func = tvm.build(s, [X,Y], target=target, name='bug')
assert func

output = [False, True, True, True, True, True, True, True, True, True, True, True]
out = numpy.array(output).reshape(m,n,k).astype("bool")
b= tvm.nd.array(numpy.zeros((m,n,k), dtype="bool"), ctx)
func(a,b)

tvm.testing.assert_allclose(b.asnumpy(), out)

[topperc]

FYI, I think might have fixed the LLVM bug this was hitting here https://github.com/llvm/llvm-project/commit/7b49e8ac359bc35f95af548fbed4b7afd625caab

[topperc]

I was asked to look at why -mcpu=cascadelake didn’t generate the same code as -mcpu=skylake-avx512 when using tvm with llvm 8.0. After ruling out llvm I started poking around tvm and happened to see this avx512 issue in my search results. Coincidentally we had hit the same underlying llvm bug internally last week with a different frontend.