[BUG] float16 and bfloat16 missing critical operators

[kian-mahmoodi]

Describe the bug

float16 and bfloat16 is missing exp
bfloat16 is missing multiple comparator operators (eg >, >=, etc)

To Reproduce EXP

def testExpT -> T:
return allo.exp(a)

def top_vitis_csyn():
sch_top = allo.customize(testExp, instantiate=[float16])
mod = sch_top.build(
target="vitis_hls",
mode="csyn",
project="top.prj",
)
mod()
print("Vitis HLS synthesis complete!")

top_vitis_csyn()

Buggy output

Traceback (most recent call last):
File "/scratch/km777/hls_final/allo/allo/ir/builder.py", line 3262, in build_stmts
results.append(build_stmt(ctx, stmt))
^^^^^^^^^^^^^^^^^^^^^
File "/scratch/km777/hls_final/allo/allo/ir/builder.py", line 101, in call
res = method(ctx, node, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/scratch/km777/hls_final/allo/allo/ir/builder.py", line 3125, in build_Return
ret = build_stmt(ctx, node.value)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/scratch/km777/hls_final/allo/allo/ir/builder.py", line 101, in call
res = method(ctx, node, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/scratch/km777/hls_final/allo/allo/ir/builder.py", line 2702, in build_Call
raise RuntimeError(f"Unsupported function {fn_name} with type {arg_types}")
RuntimeError: Unsupported function exp with type [F16Type(f16)]

Error: Unsupported function exp with type [F16Type(f16)]

To Reproduce Comparator

Note this compiles without issue for float32 but not b/float16

def testGTT -> T:
c : T = 0
if a < 0:
c = 1.0
else:
c = 2.0
return 2.0

def top_vitis_csyn():
sch_top = allo.customize(testGT, instantiate=[float16])
mod = sch_top.build(
target="vitis_hls",
mode="csyn",
project="top.prj",
)
mod()
print("Vitis HLS synthesis complete!")

top_vitis_csyn()

Buggy output (NOTE DIFFERENT FOR float16 and bfloat16!

Output for float16

File "/scratch/km777/hls_final/baseline/HLS_final_project/src/allo/test.py", line 36, in
top_vitis_csyn()
File "/scratch/km777/hls_final/baseline/HLS_final_project/src/allo/test.py", line 28, in top_vitis_csyn
mod = sch_top.build(
^^^^^^^^^^^^^^
File "/scratch/km777/hls_final/allo/allo/customize.py", line 1236, in build
return HLSModule(
^^^^^^^^^^
File "/scratch/km777/hls_final/allo/allo/backend/hls.py", line 271, in init
header, self.args = separate_header(self.hls_code, self.top_func_name)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/scratch/km777/hls_final/allo/allo/backend/hls.py", line 158, in separate_header
raise ValueError(f"Fail to resolve ctype {ele_type}")
ValueError: Fail to resolve ctype half

Output for bfloat16

Traceback (most recent call last):
File "/scratch/km777/hls_final/allo/allo/ir/builder.py", line 3262, in build_stmts
results.append(build_stmt(ctx, stmt))
^^^^^^^^^^^^^^^^^^^^^
File "/scratch/km777/hls_final/allo/allo/ir/builder.py", line 101, in call
res = method(ctx, node, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/scratch/km777/hls_final/allo/allo/ir/builder.py", line 2019, in build_If
cond = build_stmt(ctx, node.test)
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/scratch/km777/hls_final/allo/allo/ir/builder.py", line 101, in call
res = method(ctx, node, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/scratch/km777/hls_final/allo/allo/ir/builder.py", line 1967, in build_Compare
raise RuntimeError(f"Unsupported types for binary op: {dtype}")
RuntimeError: Unsupported types for binary op: bf16

Error: Unsupported types for binary op: bf16

[chhzh123]

I believe the HLS backend cannot support bfloat16. For float16, could you help fix the issue? @Fangtangtang

[Fangtangtang]

Hi @kian83m, thanks for reporting this! Could you please attach the full buggy code here? You can either upload the file directly, or wrap the entire snippet in a code block.

[kian-mahmoodi]

@Fangtangtang The "To Reproduce EXP" and "To Reproduce Comparator" portions of the report have the code to reproduce the bug. Is there something else you are looking for?

[Fangtangtang]

It's still confusing to me. the code aren't formatted, at least we need indentation in allo (just like python).
I tried adding indentation manually based on my understanding, but the resulting code is semantically incorrect.

e.g., in this one

def testExpT -> T:
    return allo.exp(a)

def top_vitis_csyn():
    sch_top = allo.customize(testExp, instantiate=[float16])
    mod = sch_top.build(
    target="vitis_hls",
    mode="csyn",
    project="top.prj",
    )
    mod()
    print("Vitis HLS synthesis complete!")

top_vitis_csyn()

we need to import allo right? def testExpT -> T seems semantically incorrect. and what is a in allo.exp(a)?

[Fangtangtang]

@chhzh123 One problem with float16 exp is that here is no overload of exp that accepts half data type.
As a result, the following hls code leads to an error call to 'exp' is ambiguous (kernel.cpp:40:16).

  for (int v8 = 0; v8 < 32; v8++) {	// L4
    half v9 = buf0[v8];	// L4
    half v10 = exp(v9);	//
    B[v8] = v10;	// L4
  }

The missing piece here is an implicit type conversion rule for float16 exp.

since this is hls backend specific, we shouldn't apply this conversion unconditionally in IR builder. A quick fix is to temporarily introduce a hls backend-specific flag in the IR builder.

---

For long-term solution, I think it's reasonable to allow registering different sets of implicit conversion rules for different backend. This may also relate to the previous draft #462 on optional implicit conversion.

---

A quick fix is to temporarily introduce a hls backend-specific flag in the ir builder.

please let me know if such kind of ad-hoc fix is desired