From cde1c8e835f7fdfbb3ab8a4c8a01edb3a7e569aa Mon Sep 17 00:00:00 2001
From: Gunhan Gulsoy <gunan@google.com>
Date: Sun, 29 Jul 2018 00:11:42 -0700
Subject: Add a new kernels option to bazel rules to enable adding dynamic
 kernel dependencies.

PiperOrigin-RevId: 206473913
---
 tensorflow/tensorflow.bzl | 53 ++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 46 insertions(+), 7 deletions(-)

(limited to 'tensorflow/tensorflow.bzl')

diff --git a/tensorflow/tensorflow.bzl b/tensorflow/tensorflow.bzl
index f8e55abe79..58282ec1c7 100644
--- a/tensorflow/tensorflow.bzl
+++ b/tensorflow/tensorflow.bzl
@@ -9,6 +9,7 @@ load(
     "tf_additional_grpc_deps_py",
     "tf_additional_xla_deps_py",
     "if_static",
+    "if_dynamic_kernels",
 )
 load(
     "@local_config_tensorrt//:build_defs.bzl",
@@ -318,18 +319,36 @@ def tf_binary_additional_srcs():
           clean_dep("//tensorflow:libtensorflow_framework.so"),
       ])
 
+
+# Helper functions to add kernel dependencies to tf binaries when using dynamic
+# kernel linking.
+def tf_binary_dynamic_kernel_dsos(kernels):
+  return if_dynamic_kernels(
+      extra_deps=["libtfkernel_%s.so" % clean_dep(k) for k in kernels],
+      otherwise=[])
+
+# Helper functions to add kernel dependencies to tf binaries when using static
+# kernel linking.
+def tf_binary_dynamic_kernel_deps(kernels):
+  return if_dynamic_kernels(
+      extra_deps=[],
+      otherwise=kernels)
+
 def tf_cc_shared_object(
     name,
     srcs=[],
     deps=[],
+    data=[],
     linkopts=[],
     framework_so=tf_binary_additional_srcs(),
+    kernels=[],
     **kwargs):
   native.cc_binary(
       name=name,
       srcs=srcs + framework_so,
-      deps=deps,
+      deps=deps + tf_binary_dynamic_kernel_deps(kernels),
       linkshared = 1,
+      data = data + tf_binary_dynamic_kernel_dsos(kernels),
       linkopts=linkopts + _rpath_linkopts(name) + select({
           clean_dep("//tensorflow:darwin"): [
               "-Wl,-install_name,@rpath/" + name.split("/")[-1],
@@ -353,18 +372,21 @@ register_extension_info(
 def tf_cc_binary(name,
                  srcs=[],
                  deps=[],
+                 data=[],
                  linkopts=[],
                  copts=tf_copts(),
+                 kernels=[],
                  **kwargs):
   native.cc_binary(
       name=name,
       copts=copts,
       srcs=srcs + tf_binary_additional_srcs(),
-      deps=deps + if_mkl(
+      deps=deps + tf_binary_dynamic_kernel_deps(kernels) + if_mkl(
           [
               "//third_party/mkl:intel_binary_blob",
           ],
       ),
+      data=data +  tf_binary_dynamic_kernel_dsos(kernels),
       linkopts=linkopts + _rpath_linkopts(name),
       **kwargs)
 
@@ -652,11 +674,13 @@ def tf_gen_op_wrapper_py(name,
 def tf_cc_test(name,
                srcs,
                deps,
+               data=[],
                linkstatic=0,
                extra_copts=[],
                suffix="",
                linkopts=[],
                nocopts=None,
+               kernels=[],
                **kwargs):
   native.cc_test(
       name="%s%s" % (name, suffix),
@@ -676,11 +700,12 @@ def tf_cc_test(name,
             "-lm"
         ],
       }) + linkopts + _rpath_linkopts(name),
-      deps=deps + if_mkl(
+      deps=deps + tf_binary_dynamic_kernel_deps(kernels) + if_mkl(
           [
               "//third_party/mkl:intel_binary_blob",
           ],
       ),
+      data=data + tf_binary_dynamic_kernel_dsos(kernels),
       # Nested select() statements seem not to be supported when passed to
       # linkstatic, and we already have a cuda select() passed in to this
       # function.
@@ -781,6 +806,7 @@ def tf_cuda_only_cc_test(name,
                     size="medium",
                     linkstatic=0,
                     args=[],
+                    kernels=[],
                     linkopts=[]):
   native.cc_test(
       name="%s%s" % (name, "_gpu"),
@@ -788,8 +814,8 @@ def tf_cuda_only_cc_test(name,
       size=size,
       args=args,
       copts= _cuda_copts() + tf_copts(),
-      data=data,
-      deps=deps + if_cuda([
+      data=data + tf_binary_dynamic_kernel_dsos(kernels),
+      deps=deps + tf_binary_dynamic_kernel_deps(kernels) + if_cuda([
           clean_dep("//tensorflow/core:cuda"),
           clean_dep("//tensorflow/core:gpu_lib")]),
       linkopts=if_not_windows(["-lpthread", "-lm"]) + linkopts + _rpath_linkopts(name),
@@ -832,9 +858,11 @@ def tf_cc_tests(srcs,
 def tf_cc_test_mkl(srcs,
                    deps,
                    name="",
+                   data=[],
                    linkstatic=0,
                    tags=[],
                    size="medium",
+                   kernels=[],
                    args=None):
   # -fno-exceptions in nocopts breaks compilation if header modules are enabled.
   disable_header_modules = ["-use_header_modules"]
@@ -855,11 +883,12 @@ def tf_cc_test_mkl(srcs,
             "-lm"
         ],
       }) + _rpath_linkopts(src_to_test_name(src)),
-      deps=deps + if_mkl(
+      deps=deps + tf_binary_dynamic_kernel_deps(kernels) + if_mkl(
           [
               "//third_party/mkl:intel_binary_blob",
           ],
       ),
+      data=data + tf_binary_dynamic_kernel_dsos(kernels),
       linkstatic=linkstatic,
       tags=tags,
       size=size,
@@ -899,12 +928,13 @@ def tf_cuda_cc_tests(srcs,
 def tf_java_test(name,
                  srcs=[],
                  deps=[],
+                 kernels=[],
                  *args,
                  **kwargs):
   native.java_test(
       name=name,
       srcs=srcs,
-      deps=deps + tf_binary_additional_srcs(),
+      deps=deps + tf_binary_additional_srcs() + tf_binary_dynamic_kernel_dsos(kernels) + tf_binary_dynamic_kernel_deps(kernels),
       *args,
       **kwargs)
 
@@ -1078,6 +1108,15 @@ def tf_kernel_library(
       deps=deps,
       **kwargs)
 
+  # TODO(gunan): CUDA dependency not clear here. Fix it.
+  tf_cc_shared_object(
+      name="libtfkernel_%s.so" % name,
+      srcs=srcs + hdrs,
+      copts=copts,
+      deps=deps,
+      tags=["manual", "notap"])
+
+
 register_extension_info(
     extension_name = "tf_kernel_library",
     label_regex_for_dep = "{extension_name}(_gpu)?",
-- 
cgit v1.2.3