Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Unity][BugFix] Fix a bug in relax gelu_tanh computation #298

Closed
wants to merge 1 commit into from
Closed

[Unity][BugFix] Fix a bug in relax gelu_tanh computation #298

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 1 addition & 10 deletions python/tvm/relax/frontend/nn/op.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -827,17 +827,8 @@ def gelu(x: Tensor, approximate: Optional[str] = None, name: str = "gelu") -> Te
----
The input tensor is required to have float dtype
"""
dtype = x._expr.struct_info.dtype
if approximate == "tanh":
tanh_const = rx.const(1 + np.tanh(np.sqrt(2 / np.pi)), dtype=dtype)
gelu_out = (
rx.const(0.5, dtype)
* x._expr
* (
tanh_const
* (x._expr + (rx.const(0.044715, dtype) * _op.power(x._expr, rx.const(3, "int32"))))
)
)
gelu_out = _op.nn.gelu_tanh(x._expr)
else:
gelu_out = _op.nn.gelu(x._expr)
return _wrap_nested(gelu_out, name)
Expand Down
14 changes: 9 additions & 5 deletions python/tvm/relax/transform/legalize_ops/nn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -302,11 +302,15 @@ def te_gelu(x: te.Tensor):
def _nn_gelu_tanh(bb: BlockBuilder, call: Call) -> Expr:
def te_gelu_tanh(x: te.Tensor):
dtype = x.dtype
return tir.const(0.5, dtype) * (
tir.const(1.0, dtype)
+ topi.tanh(
tir.const(math.sqrt(2.0 / math.pi), dtype)
* (x + tir.const(0.044715, dtype) * topi.power(x, 3))
return (
tir.const(0.5, dtype)
* x
* (
tir.const(1.0, dtype)
+ topi.tanh(
tir.const(math.sqrt(2.0 / math.pi), dtype)
* (x + tir.const(0.044715, dtype) * topi.power(x, 3))
)
)
)

Expand Down
75 changes: 44 additions & 31 deletions tests/python/relax/test_transform_legalize_ops_nn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1258,55 +1258,61 @@ def main(x: R.Tensor((2, 3), dtype="float32")) -> R.Tensor((2, 3), dtype="float3
@T.prim_func(private=True)
def gelu_tanh(A: T.Buffer((T.int64(2), T.int64(3)), "float32"), T_multiply: T.Buffer((T.int64(2), T.int64(3)), "float32")):
T.func_attr({"tir.noalias": T.bool(True)})
T_power = T.alloc_buffer((T.int64(2), T.int64(3)))
T_multiply_1 = T.alloc_buffer((T.int64(2), T.int64(3)))
T_add = T.alloc_buffer((T.int64(2), T.int64(3)))
T_power = T.alloc_buffer((T.int64(2), T.int64(3)))
T_multiply_2 = T.alloc_buffer((T.int64(2), T.int64(3)))
T_add = T.alloc_buffer((T.int64(2), T.int64(3)))
T_multiply_3 = T.alloc_buffer((T.int64(2), T.int64(3)))
compute = T.alloc_buffer((T.int64(2), T.int64(3)))
T_add_1 = T.alloc_buffer((T.int64(2), T.int64(3)))
for ax0, ax1 in T.grid(T.int64(2), T.int64(3)):
with T.block("T_multiply"):
v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1])
T.reads(A[v_ax0, v_ax1])
T.writes(T_multiply_1[v_ax0, v_ax1])
T_multiply_1[v_ax0, v_ax1] = T.float32(0.5) * A[v_ax0, v_ax1]
for ax0, ax1 in T.grid(T.int64(2), T.int64(3)):
with T.block("T_power"):
v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1])
T.reads(A[v_ax0, v_ax1])
T.writes(T_power[v_ax0, v_ax1])
T_power[v_ax0, v_ax1] = T.pow(A[v_ax0, v_ax1], T.float32(3))
for ax0, ax1 in T.grid(T.int64(2), T.int64(3)):
with T.block("T_multiply"):
with T.block("T_multiply_1"):
v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1])
T.reads(T_power[v_ax0, v_ax1])
T.writes(T_multiply_1[v_ax0, v_ax1])
T_multiply_1[v_ax0, v_ax1] = T.float32(0.044714999999999998) * T_power[v_ax0, v_ax1]
T.writes(T_multiply_2[v_ax0, v_ax1])
T_multiply_2[v_ax0, v_ax1] = T.float32(0.044714999999999998) * T_power[v_ax0, v_ax1]
for ax0, ax1 in T.grid(T.int64(2), T.int64(3)):
with T.block("T_add"):
v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1])
T.reads(A[v_ax0, v_ax1], T_multiply_1[v_ax0, v_ax1])
T.reads(A[v_ax0, v_ax1], T_multiply_2[v_ax0, v_ax1])
T.writes(T_add[v_ax0, v_ax1])
T_add[v_ax0, v_ax1] = A[v_ax0, v_ax1] + T_multiply_1[v_ax0, v_ax1]
T_add[v_ax0, v_ax1] = A[v_ax0, v_ax1] + T_multiply_2[v_ax0, v_ax1]
for ax0, ax1 in T.grid(T.int64(2), T.int64(3)):
with T.block("T_multiply_1"):
with T.block("T_multiply_2"):
v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1])
T.reads(T_add[v_ax0, v_ax1])
T.writes(T_multiply_2[v_ax0, v_ax1])
T_multiply_2[v_ax0, v_ax1] = T.float32(0.79788456080286541) * T_add[v_ax0, v_ax1]
T.writes(T_multiply_3[v_ax0, v_ax1])
T_multiply_3[v_ax0, v_ax1] = T.float32(0.79788456080286541) * T_add[v_ax0, v_ax1]
for i0, i1 in T.grid(T.int64(2), T.int64(3)):
with T.block("compute"):
v_i0, v_i1 = T.axis.remap("SS", [i0, i1])
T.reads(T_multiply_2[v_i0, v_i1])
T.reads(T_multiply_3[v_i0, v_i1])
T.writes(compute[v_i0, v_i1])
compute[v_i0, v_i1] = T.tanh(T_multiply_2[v_i0, v_i1])
compute[v_i0, v_i1] = T.tanh(T_multiply_3[v_i0, v_i1])
for ax0, ax1 in T.grid(T.int64(2), T.int64(3)):
with T.block("T_add_1"):
v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1])
T.reads(compute[v_ax0, v_ax1])
T.writes(T_add_1[v_ax0, v_ax1])
T_add_1[v_ax0, v_ax1] = T.float32(1) + compute[v_ax0, v_ax1]
for ax0, ax1 in T.grid(T.int64(2), T.int64(3)):
with T.block("T_multiply_2"):
with T.block("T_multiply_3"):
v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1])
T.reads(T_add_1[v_ax0, v_ax1])
T.reads(T_multiply_1[v_ax0, v_ax1], T_add_1[v_ax0, v_ax1])
T.writes(T_multiply[v_ax0, v_ax1])
T_multiply[v_ax0, v_ax1] = T.float32(0.5) * T_add_1[v_ax0, v_ax1]

T_multiply[v_ax0, v_ax1] = T_multiply_1[v_ax0, v_ax1] * T_add_1[v_ax0, v_ax1]

mod = LegalizeOps()(GeluTanh)
tvm.ir.assert_structural_equal(mod, Expected)
Expand Down Expand Up @@ -1338,54 +1344,61 @@ def gelu_tanh(var_A: T.handle, var_T_multiply: T.handle):
m, n = T.int64(), T.int64()
A = T.match_buffer(var_A, (m, n))
T_multiply = T.match_buffer(var_T_multiply, (m, n))
T_power = T.alloc_buffer((m, n))
T_multiply_1 = T.alloc_buffer((m, n))
T_add = T.alloc_buffer((m, n))
T_power = T.alloc_buffer((m, n))
T_multiply_2 = T.alloc_buffer((m, n))
T_add = T.alloc_buffer((m, n))
T_multiply_3 = T.alloc_buffer((m, n))
compute = T.alloc_buffer((m, n))
T_add_1 = T.alloc_buffer((m, n))
for ax0, ax1 in T.grid(m, n):
with T.block("T_multiply"):
v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1])
T.reads(A[v_ax0, v_ax1])
T.writes(T_multiply_1[v_ax0, v_ax1])
T_multiply_1[v_ax0, v_ax1] = T.float32(0.5) * A[v_ax0, v_ax1]
for ax0, ax1 in T.grid(m, n):
with T.block("T_power"):
v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1])
T.reads(A[v_ax0, v_ax1])
T.writes(T_power[v_ax0, v_ax1])
T_power[v_ax0, v_ax1] = T.pow(A[v_ax0, v_ax1], T.float32(3))
for ax0, ax1 in T.grid(m, n):
with T.block("T_multiply"):
with T.block("T_multiply_1"):
v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1])
T.reads(T_power[v_ax0, v_ax1])
T.writes(T_multiply_1[v_ax0, v_ax1])
T_multiply_1[v_ax0, v_ax1] = T.float32(0.044714999999999998) * T_power[v_ax0, v_ax1]
T.writes(T_multiply_2[v_ax0, v_ax1])
T_multiply_2[v_ax0, v_ax1] = T.float32(0.044714999999999998) * T_power[v_ax0, v_ax1]
for ax0, ax1 in T.grid(m, n):
with T.block("T_add"):
v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1])
T.reads(A[v_ax0, v_ax1], T_multiply_1[v_ax0, v_ax1])
T.reads(A[v_ax0, v_ax1], T_multiply_2[v_ax0, v_ax1])
T.writes(T_add[v_ax0, v_ax1])
T_add[v_ax0, v_ax1] = A[v_ax0, v_ax1] + T_multiply_1[v_ax0, v_ax1]
T_add[v_ax0, v_ax1] = A[v_ax0, v_ax1] + T_multiply_2[v_ax0, v_ax1]
for ax0, ax1 in T.grid(m, n):
with T.block("T_multiply_1"):
with T.block("T_multiply_2"):
v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1])
T.reads(T_add[v_ax0, v_ax1])
T.writes(T_multiply_2[v_ax0, v_ax1])
T_multiply_2[v_ax0, v_ax1] = T.float32(0.79788456080286541) * T_add[v_ax0, v_ax1]
T.writes(T_multiply_3[v_ax0, v_ax1])
T_multiply_3[v_ax0, v_ax1] = T.float32(0.79788456080286541) * T_add[v_ax0, v_ax1]
for i0, i1 in T.grid(m, n):
with T.block("compute"):
v_i0, v_i1 = T.axis.remap("SS", [i0, i1])
T.reads(T_multiply_2[v_i0, v_i1])
T.reads(T_multiply_3[v_i0, v_i1])
T.writes(compute[v_i0, v_i1])
compute[v_i0, v_i1] = T.tanh(T_multiply_2[v_i0, v_i1])
compute[v_i0, v_i1] = T.tanh(T_multiply_3[v_i0, v_i1])
for ax0, ax1 in T.grid(m, n):
with T.block("T_add_1"):
v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1])
T.reads(compute[v_ax0, v_ax1])
T.writes(T_add_1[v_ax0, v_ax1])
T_add_1[v_ax0, v_ax1] = T.float32(1) + compute[v_ax0, v_ax1]
for ax0, ax1 in T.grid(m, n):
with T.block("T_multiply_2"):
with T.block("T_multiply_3"):
v_ax0, v_ax1 = T.axis.remap("SS", [ax0, ax1])
T.reads(T_add_1[v_ax0, v_ax1])
T.reads(T_multiply_1[v_ax0, v_ax1], T_add_1[v_ax0, v_ax1])
T.writes(T_multiply[v_ax0, v_ax1])
T_multiply[v_ax0, v_ax1] = T.float32(0.5) * T_add_1[v_ax0, v_ax1]
T_multiply[v_ax0, v_ax1] = T_multiply_1[v_ax0, v_ax1] * T_add_1[v_ax0, v_ax1]


mod = LegalizeOps()(GeluTanh)
Expand Down
Loading