update with refactoring based on the PR comments

Author: cj401-amd

xla/backends/profiler/gpu/BUILD

@@ -1,3 +1,4 @@
+
 load("@local_config_cuda//cuda:build_defs.bzl", "cuda_library")
 load("@local_config_rocm//rocm:build_defs.bzl", "if_rocm_is_configured")
 load("//xla:xla.default.bzl", "xla_cc_test")
@@ -37,7 +38,6 @@ cc_library(
     deps = [
         ":cupti_collector",
         ":cupti_tracer",
-        ":cupti_tracer_options_utils",
         "//xla/tsl/platform:errors",
         "//xla/tsl/profiler/utils:time_utils",
         "//xla/tsl/util:env_var",
@@ -156,7 +156,9 @@ xla_test(
         ":cupti_wrapper",
         ":mock_cupti",
         "//xla/tsl/profiler/utils:time_utils",
+        "@com_google_absl//absl/memory",
         "@com_google_googletest//:gtest_main",
+        "@tsl//tsl/platform:test",
     ],
 )
 
@@ -222,8 +224,6 @@ cc_library(
         "//xla/tsl/profiler/backends/cpu:annotation_stack",
         "//xla/tsl/profiler/utils:lock_free_queue",
         "//xla/tsl/profiler/utils:per_thread",
-        "//xla/tsl/profiler/utils:xplane_builder",
-        "//xla/tsl/profiler/utils:xplane_schema",
         "@com_google_absl//absl/base:core_headers",
         "@com_google_absl//absl/cleanup",
         "@com_google_absl//absl/container:flat_hash_map",
@@ -259,7 +259,6 @@ cc_library(
         "cuda-only",
         "gpu",
     ],
-    visibility = ["//visibility:public"],
     deps = [
         ":cupti_collector",
         ":cupti_interface",
@@ -286,9 +285,9 @@ cc_library(
         "cuda-only",
         "gpu",
     ],
-    visibility = ["//visibility:public"],
     deps = [
         ":cupti_collector",
+        ":cupti_interface",
         "@com_google_absl//absl/status",
         "@com_google_absl//absl/time",
     ],
@@ -307,7 +306,6 @@ cc_library(
         "gpu",
         "manual",  # This target requires CUDA 12.6+, therefore we only built it if it was requested via a dependency.
     ],
-    visibility = ["//visibility:public"],
     deps = [
         ":cupti_collector",
         ":cupti_interface",
@@ -359,11 +357,39 @@ cc_library(
     ],
 )
 
+cc_library(
+  name = "rocm_tracer_utils",
+  srcs = ["rocm_tracer_utils.cc"],
+  hdrs = ["rocm_tracer_utils.h"],
+  deps = [
+    "//xla/tsl/profiler/backends/cpu:annotation_stack",
+    "//xla/tsl/profiler/utils:time_utils",
+    "//xla/tsl/profiler/utils:math_utils",
+    "@com_google_absl//absl/strings:string_view",
+    "@com_google_absl//absl/container:flat_hash_map",
+    "@com_google_absl//absl/container:flat_hash_set",
+    "@com_google_absl//absl/container:node_hash_map",
+    "@com_google_absl//absl/container:node_hash_set",
+    "@tsl//tsl/platform:env_time",
+    "@tsl//tsl/platform:env",
+    "@tsl//tsl/platform:errors",
+    "@tsl//tsl/platform:logging",
+    "@tsl//tsl/platform:macros",
+  ],
+  visibility = ["//visibility:public"],
+)
+
 cc_library(
     name = "rocm_collector",
     srcs = ["rocm_collector.cc"],
     hdrs = ["rocm_collector.h"],
     # copybara:uncomment compatible_with = ["//buildenv/target:non_prod"],
+    linkopts = select({
+        "//conditions:default": [
+            "-L/opt/rocm/lib",      # search path for all ROCm shared objects
+            "-lrocprofiler-sdk",    # the library that owns the missing symbols
+        ],
+    }),
     tags = [
         "gpu",
         "rocm-only",
@@ -372,6 +398,7 @@ cc_library(
         "manual",
     ]),
     deps = [
+        ":rocm_tracer_utils",
         "//xla/stream_executor/rocm:roctracer_wrapper",
         "//xla/tsl/profiler/backends/cpu:annotation_stack",
         "//xla/tsl/profiler/utils:parse_annotation",
@@ -404,6 +431,12 @@ cc_library(
     srcs = ["rocm_tracer.cc"],
     hdrs = ["rocm_tracer.h"],
     # copybara:uncomment compatible_with = ["//buildenv/target:non_prod"],
+    linkopts = select({
+        "//conditions:default": [
+            "-L/opt/rocm/lib",      # search path for all ROCm shared objects
+            "-lrocprofiler-sdk",    # the library that owns the missing symbols
+        ],
+    }),
     tags = [
         "gpu",
         "rocm-only",
@@ -412,10 +445,15 @@ cc_library(
         "manual",
     ]),
     deps = [
+        ":rocm_tracer_utils",
         ":rocm_collector",
         "//xla/stream_executor/rocm:roctracer_wrapper",
         "//xla/tsl/profiler/backends/cpu:annotation_stack",
         "//xla/tsl/profiler/utils:time_utils",
+        "//xla/tsl/profiler/utils:xplane_builder",
+        "//xla/tsl/profiler/utils:xplane_schema",
+        "//xla/tsl/profiler/utils:xplane_utils",
+        "//xla/tsl/util:env_var",
         "@com_google_absl//absl/container:fixed_array",
         "@com_google_absl//absl/container:flat_hash_map",
         "@com_google_absl//absl/container:flat_hash_set",
@@ -432,9 +470,64 @@ cc_library(
         "@tsl//tsl/platform:status",
         "@tsl//tsl/platform:thread_annotations",
         "@tsl//tsl/platform:types",
+        "@tsl//tsl/profiler/lib:profiler_factory",
+        "@tsl//tsl/profiler/lib:profiler_interface",
     ],
 )
 
+xla_cc_test(
+    name = "rocm_tracer_test",
+    size = "small",
+    srcs = ["rocm_tracer_test.cc"],
+    tags = [
+        "gpu",
+        "rocm",
+        "rocm-only",
+    ] + if_google([
+        # Optional: only run internally if ROCm config is enabled
+        "manual",
+    ]),
+    deps = [
+        ":rocm_tracer",
+        ":rocm_tracer_utils",
+        "//xla/tsl/profiler/utils:xplane_builder",
+        "@com_google_absl//absl/container:flat_hash_map",
+        "@com_google_googletest//:gtest_main",
+        "@tsl//tsl/platform:status_matchers",
+        "@tsl//tsl/platform:test",
+        "@tsl//tsl/profiler/protobuf:xplane_proto_cc",
+    ],
+)
+
+xla_cc_test(    
+  name = "rocm_collector_test",
+  size = "small",
+  srcs = ["rocm_collector_test.cc"],
+  tags = [
+      "gpu",
+      "rocm",
+      "rocm-only",
+  ] + if_google([
+      "manual",
+  ]),
+  deps = [
+    # ":rocm_tracer",
+    ":rocm_collector",
+    ":rocm_tracer_utils",
+    "//xla/tsl/profiler/utils:xplane_builder",
+    "@com_google_absl//absl/container:flat_hash_map",
+    "@com_google_googletest//:gtest_main",
+    "@tsl//tsl/platform:env_time",
+    "@tsl//tsl/platform:status_matchers",
+    "@tsl//tsl/platform:test",
+    "@tsl//tsl/profiler/protobuf:xplane_proto_cc",
+    "@tsl//tsl/platform:env",
+    "@tsl//tsl/platform:errors",
+    "@tsl//tsl/platform:logging",
+    "@tsl//tsl/platform:macros",
+  ],
+)
+
 cc_library(
     name = "nvtx_utils",
     srcs = ["nvtx_utils.cc"],
@@ -655,7 +748,6 @@ xla_test(
     deps = [
         ":cupti_collector",
         ":cupti_error_manager",
-        ":cupti_pm_sampler_stub",
         ":cupti_tracer",
         ":cupti_utils",
         ":cupti_wrapper",
@@ -665,27 +757,4 @@ xla_test(
         "@com_google_absl//absl/time",
         "@com_google_googletest//:gtest_main",
     ],
-)
-
-cc_library(
-    name = "cupti_tracer_options_utils",
-    srcs = ["cupti_tracer_options_utils.cc"],
-    hdrs = ["cupti_tracer_options_utils.h"],
-    # copybara:uncomment compatible_with = ["//buildenv/target:non_prod"],
-    tags = [
-        "cuda-only",
-        "gpu",
-    ],
-    visibility = ["//visibility:public"],
-    deps = [
-        ":cupti_collector",
-        ":cupti_tracer",
-        "//xla/tsl/platform:errors",
-        "//xla/tsl/profiler/utils:profiler_options_util",
-        "@com_google_absl//absl/container:flat_hash_set",
-        "@com_google_absl//absl/status",
-        "@com_google_absl//absl/strings",
-        "@local_config_cuda//cuda:cuda_headers",
-        "@tsl//tsl/profiler/protobuf:profiler_options_proto_cc",
-    ],
-)
+)