yangulei opened a new pull request, #12370:
URL: https://github.com/apache/tvm/pull/12370
So far, the `softmax` in the following test
``` python
import tvm
from tvm import relay
data = relay.var("data", shape=(1, 1000), dtype="bfloat16")
expr = relay.nn.softmax(data)
mod = tvm.IRModule.from_expr(expr)
with tvm.transform.PassContext(opt_level=3):
relay.build(mod, target="llvm")
```
will be lowered to
``` swift
@tvmgen_default_fused_nn_softmax = primfn(placeholder_1: handle,
T_softmax_norm_1: handle) -> ()
attr = {"from_legacy_te_schedule": True, "global_symbol":
"tvmgen_default_fused_nn_softmax", "tir.noalias": True}
buffers = {placeholder: Buffer(placeholder_2: Pointer(uint16), uint16,
[1000], []),
T_softmax_norm: Buffer(T_softmax_norm_2: Pointer(uint16),
uint16, [1000], [])}
buffer_map = {placeholder_1: placeholder, T_softmax_norm_1: T_softmax_norm}
preflattened_buffer_map = {placeholder_1: placeholder_3:
Buffer(placeholder_2, uint16, [1, 1000], []), T_softmax_norm_1:
T_softmax_norm_3: Buffer(T_softmax_norm_2, uint16, [1, 1000], [])} {
allocate(T_softmax_maxelem: Pointer(uint16), uint16, [1]), storage_scope =
{
{
T_softmax_maxelem_1: Buffer(T_softmax_maxelem, uint16, [1], [],
align=2)[0] = 65408u16
for (k: int32, 0, 1000) {
T_softmax_maxelem_1[0] = cast(uint16,
@tir.shift_right((@tir.reinterpret(max(@tir.reinterpret(@tir.shift_left(cast(uint32,
T_softmax_maxelem_1[0]), 16u32, dtype=uint32), dtype=float32),
@tir.reinterpret(@tir.shift_left(cast(uint32, placeholder[k]), 16u32,
dtype=uint32), dtype=float32)), dtype=uint32) +
(@tir.bitwise_and(@tir.shift_right(@tir.reinterpret(max(@tir.reinterpret(@tir.shift_left(cast(uint32,
T_softmax_maxelem_1[0]), 16u32, dtype=uint32), dtype=float32),
@tir.reinterpret(@tir.shift_left(cast(uint32, placeholder[k]), 16u32,
dtype=uint32), dtype=float32)), dtype=uint32), 16u32, dtype=uint32), 1u32,
dtype=uint32) + 32767u32)), 16u32, dtype=uint32))
}
}
allocate(T_softmax_exp: Pointer(uint16), uint16, [1000]), storage_scope
= {
for (i1: int32, 0, 1000) {
T_softmax_exp_1: Buffer(T_softmax_exp, uint16, [1000], [])[i1] =
@tir.exp(cast(uint16,
@tir.shift_right((@tir.reinterpret((@tir.reinterpret(@tir.shift_left(cast(uint32,
placeholder[i1]), 16u32, dtype=uint32), dtype=float32) -
@tir.reinterpret(@tir.shift_left(cast(uint32, T_softmax_maxelem_1[0]), 16u32,
dtype=uint32), dtype=float32)), dtype=uint32) +
(@tir.bitwise_and(@tir.shift_right(@tir.reinterpret((@tir.reinterpret(@tir.shift_left(cast(uint32,
placeholder[i1]), 16u32, dtype=uint32), dtype=float32) -
@tir.reinterpret(@tir.shift_left(cast(uint32, T_softmax_maxelem_1[0]), 16u32,
dtype=uint32), dtype=float32)), dtype=uint32), 16u32, dtype=uint32), 1u32,
dtype=uint32) + 32767u32)), 16u32, dtype=uint32)), dtype=bfloat16)
}
allocate(T_softmax_expsum: Pointer(uint16), uint16, [1]),
storage_scope = {
{
T_softmax_expsum_1: Buffer(T_softmax_expsum, uint16, [1], [],
align=2)[0] = 0u16
for (k_1: int32, 0, 1000) {
T_softmax_expsum_1[0] = cast(uint16,
@tir.shift_right((@tir.reinterpret((@tir.reinterpret(@tir.shift_left(cast(uint32,
T_softmax_expsum_1[0]), 16u32, dtype=uint32), dtype=float32) +
@tir.reinterpret(@tir.shift_left(cast(uint32, T_softmax_exp_1[k_1]), 16u32,
dtype=uint32), dtype=float32)), dtype=uint32) +
(@tir.bitwise_and(@tir.shift_right(@tir.reinterpret((@tir.reinterpret(@tir.shift_left(cast(uint32,
T_softmax_expsum_1[0]), 16u32, dtype=uint32), dtype=float32) +
@tir.reinterpret(@tir.shift_left(cast(uint32, T_softmax_exp_1[k_1]), 16u32,
dtype=uint32), dtype=float32)), dtype=uint32), 16u32, dtype=uint32), 1u32,
dtype=uint32) + 32767u32)), 16u32, dtype=uint32))
}
}
for (i1_1: int32, 0, 1000) {
T_softmax_norm[i1_1] = cast(uint16,
@tir.shift_right((@tir.reinterpret((@tir.reinterpret(@tir.shift_left(cast(uint32,
T_softmax_exp_1[i1_1]), 16u32, dtype=uint32), dtype=float32) /
@tir.reinterpret(@tir.shift_left(cast(uint32, T_softmax_expsum_1[0]), 16u32,
dtype=uint32), dtype=float32)), dtype=uint32) +
(@tir.bitwise_and(@tir.shift_right(@tir.reinterpret((@tir.reinterpret(@tir.shift_left(cast(uint32,
T_softmax_exp_1[i1_1]), 16u32, dtype=uint32), dtype=float32) /
@tir.reinterpret(@tir.shift_left(cast(uint32, T_softmax_expsum_1[0]), 16u32,
dtype=uint32), dtype=float32)), dtype=uint32), 16u32, dtype=uint32), 1u32,
dtype=uint32) + 32767u32)), 16u32, dtype=uint32))
}
}
}
}
}
```
Note that `@tir.exp` is not legalized properly, which leads to an invalid
LLVM intrinsic call codegen and rise segmentation fault finally.
This is solved by adding bfloat16 promotion for `CallNode`.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
