Fully remove unranked kernels for now.
PiperOrigin-RevId: 342614495 Change-Id: I3a67871d9f88e3d1a03961b7f5d36cc2616b703a
This commit is contained in:
parent
3202f5d14f
commit
516f4a121c
@ -82,10 +82,8 @@ cc_library(
|
||||
tf_kernel_library(
|
||||
name = "cwise_unary_op",
|
||||
srcs = [":unary_kernel_srcs"],
|
||||
# TODO(herhut): Remove no_rocm once kernels build again.
|
||||
tags = [
|
||||
"manual",
|
||||
"no_rocm",
|
||||
],
|
||||
deps = if_mlir_unranked_kernels_enabled(
|
||||
if_false = [
|
||||
@ -116,27 +114,24 @@ tf_kernel_library(
|
||||
),
|
||||
)
|
||||
|
||||
tf_kernel_library(
|
||||
name = "cwise_binary_op",
|
||||
srcs = ["unranked_gpu_add.cc"],
|
||||
# TODO(herhut): Remove no_rocm once kernels build again.
|
||||
tags = [
|
||||
"manual",
|
||||
"no_rocm",
|
||||
],
|
||||
deps = [
|
||||
":addv2_unranked_kernels",
|
||||
":unranked_op_gpu_base",
|
||||
"//third_party/eigen3",
|
||||
],
|
||||
)
|
||||
# TODO(herhut): Uncomment once unranked kernels build again.
|
||||
# tf_kernel_library(
|
||||
# name = "cwise_binary_op",
|
||||
# srcs = ["unranked_gpu_add.cc"],
|
||||
# tags = [
|
||||
# "manual",
|
||||
# ],
|
||||
# deps = [
|
||||
# ":addv2_unranked_kernels",
|
||||
# ":unranked_op_gpu_base",
|
||||
# "//third_party/eigen3",
|
||||
# ],
|
||||
# )
|
||||
|
||||
tf_kernel_library(
|
||||
name = "cwise_op",
|
||||
srcs = [],
|
||||
# TODO(herhut): Remove manual once kernels build again.
|
||||
tags = [
|
||||
"manual",
|
||||
"no_rocm",
|
||||
],
|
||||
# Technically these libraries don't need --config=cuda or --config=rocm,
|
||||
@ -145,7 +140,8 @@ tf_kernel_library(
|
||||
":cwise_unary_op",
|
||||
]) + if_mlir_unranked_kernels_enabled(
|
||||
[
|
||||
":cwise_binary_op",
|
||||
# TODO(herhut): Uncomment once it builds again.
|
||||
# ":cwise_binary_op",
|
||||
],
|
||||
),
|
||||
)
|
||||
@ -192,26 +188,27 @@ tf_cuda_cc_test(
|
||||
],
|
||||
)
|
||||
|
||||
tf_cuda_cc_test(
|
||||
name = "gpu_add_test",
|
||||
size = "small",
|
||||
srcs = if_mlir_generated_gpu_kernels_enabled(["gpu_add_test.cc"]),
|
||||
tags = tf_cuda_tests_tags() + [
|
||||
"no_cuda_asan", # b/173033461
|
||||
],
|
||||
deps = [
|
||||
"//tensorflow/core:framework",
|
||||
"//tensorflow/core:framework_internal",
|
||||
"//tensorflow/core:tensorflow",
|
||||
"//tensorflow/core:test",
|
||||
"//tensorflow/core:test_main",
|
||||
"//tensorflow/core:testlib",
|
||||
"//tensorflow/core/common_runtime:device",
|
||||
"//tensorflow/core/common_runtime:device_factory",
|
||||
"//tensorflow/core/kernels:cwise_op",
|
||||
"//tensorflow/core/kernels:ops_testutil",
|
||||
],
|
||||
)
|
||||
# TODO(herhut): Uncomment once unranked kernels build again.
|
||||
# tf_cuda_cc_test(
|
||||
# name = "gpu_add_test",
|
||||
# size = "small",
|
||||
# srcs = if_mlir_generated_gpu_kernels_enabled(["gpu_add_test.cc"]),
|
||||
# tags = tf_cuda_tests_tags() + [
|
||||
# "no_cuda_asan", # b/173033461
|
||||
# ],
|
||||
# deps = [
|
||||
# "//tensorflow/core:framework",
|
||||
# "//tensorflow/core:framework_internal",
|
||||
# "//tensorflow/core:tensorflow",
|
||||
# "//tensorflow/core:test",
|
||||
# "//tensorflow/core:test_main",
|
||||
# "//tensorflow/core:testlib",
|
||||
# "//tensorflow/core/common_runtime:device",
|
||||
# "//tensorflow/core/common_runtime:device_factory",
|
||||
# "//tensorflow/core/kernels:cwise_op",
|
||||
# "//tensorflow/core/kernels:ops_testutil",
|
||||
# ],
|
||||
# )
|
||||
# TODO(b/160731748): Re-enable when it works again.
|
||||
# gen_kernel_library(
|
||||
# name = "bias_add",
|
||||
@ -314,19 +311,20 @@ gen_kernel_library(
|
||||
unroll_factors = "4",
|
||||
)
|
||||
|
||||
gen_kernel_library(
|
||||
name = "addv2",
|
||||
generate_ranked = False,
|
||||
generate_unranked = True,
|
||||
tile_size = "256",
|
||||
types = [
|
||||
"f16",
|
||||
"f32",
|
||||
"f64",
|
||||
"i64",
|
||||
],
|
||||
unroll_factors = "4",
|
||||
)
|
||||
# TODO(herhut): Uncomment once it builds again.
|
||||
# gen_kernel_library(
|
||||
# name = "addv2",
|
||||
# generate_ranked = False,
|
||||
# generate_unranked = True,
|
||||
# tile_size = "256",
|
||||
# types = [
|
||||
# "f16",
|
||||
# "f32",
|
||||
# "f64",
|
||||
# "i64",
|
||||
# ],
|
||||
# unroll_factors = "4",
|
||||
# )
|
||||
|
||||
[
|
||||
gen_kernel_library(
|
||||
|
Loading…
Reference in New Issue
Block a user