Split the unary unranked kernels into enabled by default and experimental.
For now, this change will be a no-op, but in a followup change we will redefine the meaning of --define=enable_unranked_kernels=1 to enable the experimental unary kernels, while the others (currently abs and tanh) will be enabled by default and can be disabled with --define=tensorflow_enable_mlir_generated_gpu_kernels=0 The ranked unary kernels will be removed. PiperOrigin-RevId: 345419524 Change-Id: Ice00e32f695fd44d1b70b8cf8c377f6a61de3b19
This commit is contained in:
parent
7fa02aac5f
commit
6c0245b27c
@ -37,6 +37,46 @@ config_setting(
|
||||
define_values = {"enable_unranked_kernels": "1"},
|
||||
)
|
||||
|
||||
filegroup(
|
||||
name = "enabled_unary_unranked_kernel_srcs",
|
||||
srcs = [
|
||||
"unranked_op_gpu_abs.cc",
|
||||
"unranked_op_gpu_tanh.cc",
|
||||
],
|
||||
compatible_with = get_compatible_with_cloud(),
|
||||
)
|
||||
|
||||
filegroup(
|
||||
name = "experimental_unary_unranked_kernel_srcs",
|
||||
srcs = [
|
||||
"unranked_op_gpu_ceil.cc",
|
||||
"unranked_op_gpu_conj.cc",
|
||||
"unranked_op_gpu_cos.cc",
|
||||
"unranked_op_gpu_exp.cc",
|
||||
"unranked_op_gpu_floor.cc",
|
||||
"unranked_op_gpu_imag.cc",
|
||||
"unranked_op_gpu_is_inf.cc",
|
||||
"unranked_op_gpu_log.cc",
|
||||
"unranked_op_gpu_logical_not.cc",
|
||||
"unranked_op_gpu_real.cc",
|
||||
"unranked_op_gpu_rsqrt.cc",
|
||||
"unranked_op_gpu_sign.cc",
|
||||
"unranked_op_gpu_sin.cc",
|
||||
"unranked_op_gpu_sqrt.cc",
|
||||
],
|
||||
compatible_with = get_compatible_with_cloud(),
|
||||
)
|
||||
|
||||
filegroup(
|
||||
name = "unary_unranked_kernel_srcs",
|
||||
srcs = [
|
||||
":enabled_unary_unranked_kernel_srcs",
|
||||
] + if_mlir_unranked_kernels_enabled(
|
||||
if_true = [":experimental_unary_unranked_kernel_srcs"],
|
||||
),
|
||||
compatible_with = get_compatible_with_cloud(),
|
||||
)
|
||||
|
||||
filegroup(
|
||||
name = "unary_kernel_srcs",
|
||||
srcs = if_mlir_unranked_kernels_enabled(
|
||||
@ -46,24 +86,7 @@ filegroup(
|
||||
"cwise_op_gpu_base.h",
|
||||
"cwise_op_gpu_tanh.cc",
|
||||
],
|
||||
if_true = [
|
||||
"unranked_op_gpu_abs.cc",
|
||||
"unranked_op_gpu_ceil.cc",
|
||||
"unranked_op_gpu_conj.cc",
|
||||
"unranked_op_gpu_cos.cc",
|
||||
"unranked_op_gpu_exp.cc",
|
||||
"unranked_op_gpu_floor.cc",
|
||||
"unranked_op_gpu_imag.cc",
|
||||
"unranked_op_gpu_is_inf.cc",
|
||||
"unranked_op_gpu_log.cc",
|
||||
"unranked_op_gpu_logical_not.cc",
|
||||
"unranked_op_gpu_real.cc",
|
||||
"unranked_op_gpu_rsqrt.cc",
|
||||
"unranked_op_gpu_sign.cc",
|
||||
"unranked_op_gpu_sin.cc",
|
||||
"unranked_op_gpu_sqrt.cc",
|
||||
"unranked_op_gpu_tanh.cc",
|
||||
],
|
||||
if_true = [":unary_unranked_kernel_srcs"],
|
||||
),
|
||||
compatible_with = get_compatible_with_cloud(),
|
||||
)
|
||||
@ -104,6 +127,11 @@ tf_kernel_library(
|
||||
"//tensorflow/core/platform:stream_executor",
|
||||
],
|
||||
if_true = [
|
||||
# Technically we only need to depend on the kernel libraries for the
|
||||
# unranked kernels which are enabled by default. But this would
|
||||
# make our BUILD target structure uglier. We already need to make
|
||||
# sure that those targets can be built, so it should not hurt to
|
||||
# link them in even if they are currently not needed yet.
|
||||
":abs_unranked_kernels",
|
||||
":ceil_unranked_kernels",
|
||||
":conj_unranked_kernels",
|
||||
|
Loading…
Reference in New Issue
Block a user