[XLA] Remove gpu/gpu_backend-specific flags

Move useful flags into debug_options, and leave some less used flags out - they
can be propagated through debug_options if required (for now there's too much
duplication between them and what's already inside)

PiperOrigin-RevId: 159261661

6 files changed, 17 insertions, 288 deletions

diff --git a/tensorflow/compiler/xla/legacy_flags/BUILD b/tensorflow/compiler/xla/legacy_flags/BUILD
index 09101708f2..abaccb81e5 100644
--- a/tensorflow/compiler/xla/legacy_flags/BUILD
+++ b/tensorflow/compiler/xla/legacy_flags/BUILD
@@ -115,30 +115,6 @@ cc_library(
 )
 
 cc_library(
-    name = "gpu_compiler_flags",
-    srcs = ["gpu_compiler_flags.cc"],
-    hdrs = ["gpu_compiler_flags.h"],
-    deps = [
-        ":parse_flags_from_env",
-        "//tensorflow/compiler/xla:types",
-        "//tensorflow/core:framework_internal",
-        "//tensorflow/core:lib",
-    ],
-)
-
-cc_library(
-    name = "gpu_backend_lib_flags",
-    srcs = ["gpu_backend_lib_flags.cc"],
-    hdrs = ["gpu_backend_lib_flags.h"],
-    deps = [
-        ":parse_flags_from_env",
-        "//tensorflow/compiler/xla:types",
-        "//tensorflow/core:framework_internal",
-        "//tensorflow/core:lib",
-    ],
-)
-
-cc_library(
     name = "stream_assignment_flags",
     srcs = ["stream_assignment_flags.cc"],
     hdrs = ["stream_assignment_flags.h"],
diff --git a/tensorflow/compiler/xla/legacy_flags/debug_options_flags.cc b/tensorflow/compiler/xla/legacy_flags/debug_options_flags.cc
index 52b048ec15..5f029a5f53 100644
--- a/tensorflow/compiler/xla/legacy_flags/debug_options_flags.cc
+++ b/tensorflow/compiler/xla/legacy_flags/debug_options_flags.cc
@@ -30,6 +30,10 @@ struct DebugOptionsFlags {
   int32 xla_backend_optimization_level;
   bool xla_embed_ir_in_executable;
   string xla_dump_debug_json_to;
+
+  string xla_gpu_cuda_data_dir;
+  bool xla_gpu_ftz;
+
   string xla_backend_extra_options;
 };
 
@@ -46,9 +50,11 @@ void AllocateFlags() {
   flag_values->xla_generate_hlo_graph = "";
   flag_values->xla_disable_hlo_passes = "";
   flag_values->xla_enable_fast_math = true;
-  flag_values->xla_backend_optimization_level = 2;
+  flag_values->xla_backend_optimization_level = 3;
   flag_values->xla_embed_ir_in_executable = false;
   flag_values->xla_dump_debug_json_to = "";
+  flag_values->xla_gpu_cuda_data_dir = "./cuda_sdk_lib";
+  flag_values->xla_gpu_ftz = false;
   flag_values->xla_backend_extra_options = "";
 
   flag_objects = new std::vector<tensorflow::Flag>(
@@ -72,6 +78,14 @@ void AllocateFlags() {
        tensorflow::Flag("xla_embed_ir_in_executable",
                         &flag_values->xla_embed_ir_in_executable,
                         "Embed the compiler IR as a string in the executable."),
+       tensorflow::Flag("xla_gpu_cuda_data_dir",
+                        &flag_values->xla_gpu_cuda_data_dir,
+                        "If non-empty, speficies a local directory containing "
+                        "ptxas and nvvm libdevice files; otherwise we use "
+                        "those from runfile directories."),
+       tensorflow::Flag("xla_gpu_ftz", &flag_values->xla_gpu_ftz,
+                        "If true, flush-to-zero semantics are enabled in the "
+                        "code generated for GPUs."),
        tensorflow::Flag(
            "xla_dump_debug_json_to", &flag_values->xla_dump_debug_json_to,
            "Dump compilation artifacts as JSON into this directory."),
@@ -110,6 +124,8 @@ xla::DebugOptions GetDebugOptionsFromFlags() {
   options.set_xla_embed_ir_in_executable(
       flag_values->xla_embed_ir_in_executable);
   options.set_xla_dump_debug_json_to(flag_values->xla_dump_debug_json_to);
+  options.set_xla_gpu_cuda_data_dir(flag_values->xla_gpu_cuda_data_dir);
+  options.set_xla_gpu_ftz(flag_values->xla_gpu_ftz);
 
   std::vector<string> extra_options_parts =
       tensorflow::str_util::Split(flag_values->xla_backend_extra_options, ',');
diff --git a/tensorflow/compiler/xla/legacy_flags/gpu_backend_lib_flags.cc b/tensorflow/compiler/xla/legacy_flags/gpu_backend_lib_flags.cc
deleted file mode 100644
index f8f6ea26b1..0000000000
--- a/tensorflow/compiler/xla/legacy_flags/gpu_backend_lib_flags.cc
+++ /dev/null
@@ -1,88 +0,0 @@
-/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-// Legacy flags for XLA's gpu_backend_lib module.
-
-#include <mutex>  // NOLINT(build/c++11): only using std::call_once, not mutex.
-#include <vector>
-
-#include "tensorflow/compiler/xla/legacy_flags/gpu_backend_lib_flags.h"
-#include "tensorflow/compiler/xla/legacy_flags/parse_flags_from_env.h"
-#include "tensorflow/core/platform/types.h"
-#include "tensorflow/core/util/command_line_flags.h"
-
-namespace xla {
-namespace legacy_flags {
-
-// Pointers to the parsed value of the flags and flag descriptors, initialized
-// via flags_init.
-static GpuBackendLibFlags* flags;
-static std::vector<tensorflow::Flag>* flag_list;
-static std::once_flag flags_init;
-
-// Allocate *flags.  Called via call_once(&flags_init,...).
-static void AllocateFlags() {
-  flags = new GpuBackendLibFlags;
-  flags->dump_temp_products_to = "";
-  flags->ftz = false;
-  flags->fma = true;
-  flags->verbose_ptx_asm = false;
-  flags->kernel = "";
-  flags->llvm_dump_passes = false;
-  flags->llvm_cl_opts = "";
-  flags->dump_ir_before_passes = false;
-  flags->opt_level = 3;
-  flag_list = new std::vector<tensorflow::Flag>({
-      tensorflow::Flag("dump_temp_products_to", &flags->dump_temp_products_to,
-                       "dump temporary compilation products to this directory. "
-                       "If empty, no dump is produced"),
-      tensorflow::Flag("ftz", &flags->ftz, "flush to zero semantics"),
-      tensorflow::Flag("fma", &flags->fma, "use FMA synthesis"),
-      tensorflow::Flag("verbose_ptx_asm", &flags->verbose_ptx_asm,
-                       "emit PTX assembly with extra comments"),
-      tensorflow::Flag("kernel", &flags->kernel,
-                       "only emit the IR and PTX for this kernel"),
-      tensorflow::Flag("llvm_dump_passes", &flags->llvm_dump_passes,
-                       "dump the passes LLVM runs to stderr"),
-      tensorflow::Flag(
-          "llvm_cl_opts", &flags->llvm_cl_opts,
-          "comma-separated list of command line options to pass to "
-          "LLVM.  For example, --llvm_cl_opts=--print-before=loop-unroll"),
-      tensorflow::Flag("dump_ir_before_passes", &flags->dump_ir_before_passes,
-                       "dump the IR before each optimization pass in "
-                       "sequentially-named files."),
-      tensorflow::Flag("opt_level", &flags->opt_level,
-                       "optimization level (default to 3)"),
-  });
-  ParseFlagsFromEnv(*flag_list);
-}
-
-// Append to *append_to flag definitions associated with XLA's gpu_backend_lib
-// module.
-void AppendGpuBackendLibFlags(std::vector<tensorflow::Flag>* append_to) {
-  std::call_once(flags_init, &AllocateFlags);
-  append_to->insert(append_to->end(), flag_list->begin(), flag_list->end());
-}
-
-// Return a pointer to the GpuBackendLibFlags struct;
-// repeated calls return the same pointer.
-// This should be called only after Flags::Parse() has returned.
-GpuBackendLibFlags* GetGpuBackendLibFlags() {
-  std::call_once(flags_init, &AllocateFlags);
-  return flags;
-}
-
-}  // namespace legacy_flags
-}  // namespace xla
diff --git a/tensorflow/compiler/xla/legacy_flags/gpu_backend_lib_flags.h b/tensorflow/compiler/xla/legacy_flags/gpu_backend_lib_flags.h
deleted file mode 100644
index 31cb50e9da..0000000000
--- a/tensorflow/compiler/xla/legacy_flags/gpu_backend_lib_flags.h
+++ /dev/null
@@ -1,55 +0,0 @@
-/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-#ifndef TENSORFLOW_COMPILER_XLA_LEGACY_FLAGS_GPU_BACKEND_LIB_FLAGS_H_
-#define TENSORFLOW_COMPILER_XLA_LEGACY_FLAGS_GPU_BACKEND_LIB_FLAGS_H_
-
-// Legacy flags for XLA's gpu_backend_lib module.
-
-#include <vector>
-
-#include "tensorflow/compiler/xla/types.h"
-#include "tensorflow/core/platform/types.h"
-#include "tensorflow/core/util/command_line_flags.h"
-
-namespace xla {
-namespace legacy_flags {
-
-// Append to *flag_list flag definitions associated with XLA's gpu_backend_lib
-// module.
-void AppendGpuBackendLibFlags(std::vector<tensorflow::Flag>* flag_list);
-
-// The values of flags associated with XLA's gpu_backend_lib module.
-typedef struct {
-  string dump_temp_products_to;  // temporary compilation products dir
-  bool ftz;                      // flush to zero semantics
-  bool fma;                      // use FMA synthesis
-  bool verbose_ptx_asm;          // emit PTX assembly with extra comments
-  string kernel;                 // only emit the IR and PTX for this kernel
-  bool llvm_dump_passes;         // dump the passes LLVM runs to stderr
-  string llvm_cl_opts;           // comma-separated list of LLVM options
-  bool dump_ir_before_passes;    // dump IR before each pass
-  int32 opt_level;               // optimization level
-} GpuBackendLibFlags;
-
-// Return a pointer to the GpuBackendLibFlags struct;
-// repeated calls return the same pointer.
-// This should be called only after Flags::Parse() has returned.
-GpuBackendLibFlags* GetGpuBackendLibFlags();
-
-}  // namespace legacy_flags
-}  // namespace xla
-
-#endif  // TENSORFLOW_COMPILER_XLA_LEGACY_FLAGS_GPU_BACKEND_LIB_FLAGS_H_
diff --git a/tensorflow/compiler/xla/legacy_flags/gpu_compiler_flags.cc b/tensorflow/compiler/xla/legacy_flags/gpu_compiler_flags.cc
deleted file mode 100644
index a8af794494..0000000000
--- a/tensorflow/compiler/xla/legacy_flags/gpu_compiler_flags.cc
+++ /dev/null
@@ -1,68 +0,0 @@
-/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-// Legacy flags for XLA's gpu_compiler module.
-
-#include <mutex>  // NOLINT(build/c++11): only using std::call_once, not mutex.
-#include <vector>
-
-#include "tensorflow/compiler/xla/legacy_flags/gpu_compiler_flags.h"
-#include "tensorflow/compiler/xla/legacy_flags/parse_flags_from_env.h"
-#include "tensorflow/compiler/xla/types.h"
-#include "tensorflow/core/platform/types.h"
-#include "tensorflow/core/util/command_line_flags.h"
-
-namespace xla {
-namespace legacy_flags {
-
-// Pointers to the parsed value of the flags and flag descriptors, initialized
-// via flags_init.
-static GpuCompilerFlags* flags;
-static std::vector<tensorflow::Flag>* flag_list;
-static std::once_flag flags_init;
-
-// Allocate *flags.  Called via call_once(&flags_init,...).
-static void AllocateFlags() {
-  flags = new GpuCompilerFlags;
-  flags->xla_cuda_data_dir = "./cuda_sdk_lib";
-  flag_list = new std::vector<tensorflow::Flag>({
-      tensorflow::Flag(
-          "xla_cuda_data_dir", &flags->xla_cuda_data_dir,
-          "If non-empty, specifies a local directory containing ptxas and "
-          "nvvm libdevice files. Otherwise, by default, we use those from "
-          "runfile directories."),
-      tensorflow::Flag("xla_ptxas_path", &flags->xla_ptxas_path,
-                       "The path to ptxas. Required to log stats of the ptx."),
-  });
-  ParseFlagsFromEnv(*flag_list);
-}
-
-// Append to *append_to flag definitions associated with XLA's gpu_compiler
-// module.
-void AppendGpuCompilerFlags(std::vector<tensorflow::Flag>* append_to) {
-  std::call_once(flags_init, &AllocateFlags);
-  append_to->insert(append_to->end(), flag_list->begin(), flag_list->end());
-}
-
-// Return a pointer to the GpuCompilerFlags struct;
-// repeated calls return the same pointer.
-// This should be called only after Flags::Parse() has returned.
-GpuCompilerFlags* GetGpuCompilerFlags() {
-  std::call_once(flags_init, &AllocateFlags);
-  return flags;
-}
-
-}  // namespace legacy_flags
-}  // namespace xla
diff --git a/tensorflow/compiler/xla/legacy_flags/gpu_compiler_flags.h b/tensorflow/compiler/xla/legacy_flags/gpu_compiler_flags.h
deleted file mode 100644
index 0b0f186b25..0000000000
--- a/tensorflow/compiler/xla/legacy_flags/gpu_compiler_flags.h
+++ /dev/null
@@ -1,52 +0,0 @@
-/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-==============================================================================*/
-
-#ifndef TENSORFLOW_COMPILER_XLA_LEGACY_FLAGS_GPU_COMPILER_FLAGS_H_
-#define TENSORFLOW_COMPILER_XLA_LEGACY_FLAGS_GPU_COMPILER_FLAGS_H_
-
-// Legacy flags for XLA's gpu_compiler module.
-
-#include <vector>
-
-#include "tensorflow/compiler/xla/types.h"
-#include "tensorflow/core/platform/types.h"
-#include "tensorflow/core/util/command_line_flags.h"
-
-namespace xla {
-namespace legacy_flags {
-
-// Append to *flag_list flag definitions associated with XLA's gpu_compiler
-// module.
-void AppendGpuCompilerFlags(std::vector<tensorflow::Flag>* flag_list);
-
-// The values of flags associated with XLA's gpu_compiler module.
-typedef struct {
-  string xla_cuda_data_dir;  // If non-empty, specifies a local directory
-                             // containing ptxas and nvvm libdevice files.
-                             // Otherwise, by default, we use those from runfile
-                             // directories.
-  string xla_ptxas_path;     // The path to ptxas.  Required to log stats of
-                             // the ptx.
-} GpuCompilerFlags;
-
-// Return a pointer to the GpuCompilerFlags struct;
-// repeated calls return the same pointer.
-// This should be called only after Flags::Parse() has returned.
-GpuCompilerFlags* GetGpuCompilerFlags();
-
-}  // namespace legacy_flags
-}  // namespace xla
-
-#endif  // TENSORFLOW_COMPILER_XLA_LEGACY_FLAGS_GPU_COMPILER_FLAGS_H_