CPU feature logging

Signed-off-by: Benoit Jacob <jacob.benoit.1@gmail.com>

build_tools/bazel/iree_check_test.bzl

@@ -24,7 +24,6 @@ def iree_check_test(
         input_type = None,
         runner_args = [],
         tags = [],
-        target_cpu_features = None,
         timeout = None,
         **kwargs):
     """Creates an iree-check-module test for the specified source file.
@@ -43,19 +42,16 @@ def iree_check_test(
           driver and input file are passed automatically.
       tags: additional tags to apply to the generated test. Tag
           "driver=DRIVER" and "target=TARGET" are added automatically.
-      target_cpu_features: currently unimplemented (must be empty), will
-          eventually allow specifying target CPU features.
       timeout: timeout for the generated tests.
       **kwargs: any additional attributes to pass to the underlying native_test.
     """
 
-    if target_cpu_features:
-        fail("target_cpu_features must currently be empty")
     input_type_flags = []
     if input_type:
         input_type_flags = ["--iree-input-type=%s" % input_type]
     flags = [
         "--iree-hal-target-backends=%s" % target_backend,
+        "--iree-hal-target-cpu=generic",
     ] + compiler_flags + input_type_flags
     bytecode_module_name = name + "_bytecode_module"
     iree_bytecode_module(
@@ -91,7 +87,6 @@ def iree_check_single_backend_test_suite(
         input_type = None,
         runner_args = [],
         tags = [],
-        target_cpu_features = None,
         timeout = None,
         **kwargs):
     """Creates a test suite of iree-check-module tests for a single backend/driver pair.
@@ -112,8 +107,6 @@ def iree_check_single_backend_test_suite(
           iree-check-module tests. The driver and input file are passed
           automatically. To use different runner_args per test, create a
           separate suite or iree_check_test.
-      target_cpu_features: currently unimplemented (must be empty), will
-          eventually allow specifying target CPU features.
       tags: tags to apply to the generated tests. Note that as in standard test
           suites, manual is treated specially and will also apply to the test
           suite itself.
@@ -130,13 +123,6 @@ def iree_check_single_backend_test_suite(
     if target_backend == "rocm" or driver == "hip":
         return
 
-    # We haven't implemented this so far because we have been using target_cpu_features so far only
-    # for aarch64 targets, for which we use the CMake build. To future people implementing this:
-    # target_cpu_features should be a list, and here it should be joined into a comma-separated
-    # string to be passed to --iree-llvmcpu-target-cpu-features
-    if target_cpu_features:
-        fail("target_cpu_features must currently be empty")
-
     tests = []
     for src in srcs:
         test_name = "_".join([name, src])
@@ -200,10 +186,10 @@ def iree_check_test_suite(
           Currently unimplemented in Bazel due to difficulty of specializing
           to target architecture in Bazel. The following describes the
           semantics that this should have if implemented. Each
-          entry is either "default" for the architecture defaults, or a colon-
-          separated triple "arch:name:cpu_features" where "arch" filters
-          for a target CPU architecture (in IREE_ARCH format), "name" is a
-          short name for the CPU features set (used to generate target names)
+          entry is either "generic" for the architecture defaults, or "host"
+          for the host CPU,or a colon-separated triple "arch:name:cpu_features"
+          where "arch" filters for a target CPU architecture (in IREE_ARCH format),
+          "name" is a short name for the CPU features set (used to generate target names)
           and cpu_features is a comma-separated list of LLVM target attributes
           to enable. Example:
             x86_64:avx2_fma:+avx,+avx2,+fma

build_tools/bazel/iree_e2e_generated_runner_test.bzl

@@ -22,7 +22,6 @@ def iree_e2e_runner_test(
         compiler_flags = [],
         runner_args = [],
         tags = [],
-        target_cpu_features = None,
         timeout = None,
         **kwargs):
     """Creates a test using a specified test runner program.
@@ -44,23 +43,18 @@ def iree_e2e_runner_test(
             added automatically.
         test_runner: test runner program to run.
         timeout: timeout for the generated tests.
-        target_cpu_features: target CPU features. Only for llvm-cpu backend.
         **kwargs: any additional attributes to pass to the underlying tests and
             test suite.
     """
 
-    if target_cpu_features:
-        fail("target_cpu_features must currently be empty")
-
     iree_bytecode_module(
         name = name + "_%s_module" % test_type,
         module_name = tests_vmfb,
         src = tests_src,
         flags = [
             "--iree-hal-target-backends=%s" % target_backend,
-        ] + ([
-            "--iree-llvmcpu-target-cpu-features=%s" % target_cpu_features,
-        ] if target_cpu_features else []) + compiler_flags,
+            "--iree-hal-target-cpu=generic",
+        ] + compiler_flags,
         visibility = ["//visibility:private"],
         testonly = True,
         **kwargs
@@ -106,7 +100,6 @@ def iree_single_backend_e2e_runner_test(
         compiler_flags = [],
         runner_args = [],
         tags = [],
-        target_cpu_features = None,
         timeout = None,
         **kwargs):
     """Generates an iree_e2e_runner_test using a custom python generator script.
@@ -133,7 +126,6 @@ def iree_single_backend_e2e_runner_test(
             added automatically.
         test_runner: test runner program to run.
         timeout: timeout for the generated tests.
-        target_cpu_features: target CPU features. Only for llvm-cpu backend.
         **kwargs: any additional attributes to pass to the underlying tests and
             test suite.
     """
@@ -171,7 +163,6 @@ def iree_single_backend_e2e_runner_test(
         runner_args = runner_args,
         tags = tags,
         timeout = timeout,
-        target_cpu_features = target_cpu_features,
         **kwargs
     )
 
@@ -213,13 +204,13 @@ def iree_generated_e2e_runner_test(
             Currently unimplemented in Bazel due to difficulty of specializing
             to target architecture in Bazel. The following describes the
             semantics that this should have if implemented. Each
-            entry is either "default" for the architecture defaults, or a colon-
-            separated triple "arch:name:cpu_features" where "arch" filters
-            for a target CPU architecture (in IREE_ARCH format), "name" is a
-            short name for the CPU features set (used to generate target names)
+            entry is either "generic" for the architecture defaults, or "host"
+            for the host CPU,or a colon-separated triple "arch:name:cpu_features"
+            where "arch" filters for a target CPU architecture (in IREE_ARCH format),
+            "name" is a short name for the CPU features set (used to generate target names)
             and cpu_features is a comma-separated list of LLVM target attributes
             to enable. Example:
-              x86_64:avx2_fma:+avx,+avx2,+fma
+                x86_64:avx2_fma:+avx,+avx2,+fma
         **kwargs: any additional attributes to pass to the underlying tests and test suite.
     """
 

build_tools/bazel_to_cmake/bazel_to_cmake_converter.py

@@ -769,7 +769,6 @@ def iree_check_single_backend_test_suite(
         target_backends_and_drivers=None,
         runner_args=None,
         tags=None,
-        target_cpu_features=None,
         timeout=None,
         **kwargs,
     ):
@@ -787,9 +786,6 @@ def iree_check_single_backend_test_suite(
         input_type_block = self._convert_string_arg_block("INPUT_TYPE", input_type)
         runner_args_block = self._convert_string_list_block("RUNNER_ARGS", runner_args)
         labels_block = self._convert_string_list_block("LABELS", tags)
-        target_cpu_features_block = self._convert_string_arg_block(
-            "TARGET_CPU_FEATURES", target_cpu_features
-        )
         timeout_block = self._convert_timeout_arg_block("TIMEOUT", timeout)
 
         self._converter.body += (
@@ -802,7 +798,6 @@ def iree_check_single_backend_test_suite(
             f"{input_type_block}"
             f"{runner_args_block}"
             f"{labels_block}"
-            f"{target_cpu_features_block}"
             f"{timeout_block}"
             f")\n\n"
         )

build_tools/cmake/iree_check_test.cmake

@@ -39,8 +39,6 @@ endfunction()
 #       "driver=${DRIVER}" are added automatically.
 #   MODULE_FILE_NAME: Optional, specifies the absolute path to the filename
 #       to use for the generated IREE module (.vmfb).
-#   TARGET_CPU_FEATURES: If specified, a string passed as argument to
-#       --iree-llvmcpu-target-cpu-features.
 #   DEPENDS: Optional. Additional dependencies beyond SRC and the tools.
 #   INPUT_TYPE: The value for the --iree-input-type= flag. Also disables tests
 #       if no compiled support for that configuration.
@@ -53,7 +51,7 @@ function(iree_check_test)
     _RULE
     ""
     "NAME;SRC;TARGET_BACKEND;DRIVER;MODULE_FILE_NAME;INPUT_TYPE"
-    "COMPILER_FLAGS;RUNNER_ARGS;LABELS;TARGET_CPU_FEATURES;DEPENDS;TIMEOUT"
+    "COMPILER_FLAGS;RUNNER_ARGS;LABELS;DEPENDS;TIMEOUT"
     ${ARGN}
   )
 
@@ -162,9 +160,6 @@ function(iree_check_test)
   if(_RULE_INPUT_TYPE)
     list(APPEND _BASE_COMPILER_FLAGS "--iree-input-type=${_RULE_INPUT_TYPE}")
   endif()
-  if(_RULE_TARGET_CPU_FEATURES)
-    list(APPEND _BASE_COMPILER_FLAGS "--iree-llvmcpu-target-cpu-features=${_RULE_TARGET_CPU_FEATURES}")
-  endif()
   if(_NORMALIZED_TARGET_BACKEND STREQUAL "ROCM")
     list(APPEND _BASE_COMPILER_FLAGS "--iree-hip-target=${IREE_HIP_TEST_TARGET_CHIP}")
   endif()
@@ -242,8 +237,6 @@ endfunction()
 #       different args per test, create a separate suite or iree_check_test.
 #   LABELS: Additional labels to apply to the generated tests. The package path
 #       is added automatically.
-#   TARGET_CPU_FEATURES: If specified, a string passed as argument to
-#       --iree-llvmcpu-target-cpu-features.
 #   DEPENDS: Optional. Additional dependencies beyond SRC and the tools.
 #   INPUT_TYPE: The value for the --iree-input-type= flag. Also disables tests
 #       if no compiled support for that configuration.
@@ -256,7 +249,7 @@ function(iree_check_single_backend_test_suite)
     _RULE
     ""
     "NAME;TARGET_BACKEND;DRIVER;INPUT_TYPE"
-    "SRCS;COMPILER_FLAGS;RUNNER_ARGS;LABELS;TARGET_CPU_FEATURES;DEPENDS;TIMEOUT"
+    "SRCS;COMPILER_FLAGS;RUNNER_ARGS;LABELS;DEPENDS;TIMEOUT"
     ${ARGN}
   )
 
@@ -284,7 +277,6 @@ function(iree_check_single_backend_test_suite)
         INPUT_TYPE ${_RULE_INPUT_TYPE}
         RUNNER_ARGS ${_RULE_RUNNER_ARGS}
         LABELS ${_RULE_LABELS}
-        TARGET_CPU_FEATURES ${_RULE_TARGET_CPU_FEATURES}
         DEPENDS ${_RULE_DEPENDS}
         TIMEOUT ${_RULE_TIMEOUT}
       )
@@ -309,7 +301,6 @@ function(iree_check_single_backend_test_suite)
           INPUT_TYPE ${_RULE_INPUT_TYPE}
           RUNNER_ARGS ${_RULE_RUNNER_ARGS}
           LABELS ${_RULE_LABELS}
-          TARGET_CPU_FEATURES ${_RULE_TARGET_CPU_FEATURES}
           DEPENDS ${_RULE_DEPENDS}
           TIMEOUT ${_RULE_TIMEOUT}
         )
@@ -328,7 +319,6 @@ function(iree_check_single_backend_test_suite)
           INPUT_TYPE ${_RULE_INPUT_TYPE}
           RUNNER_ARGS ${_RULE_RUNNER_ARGS}
           LABELS ${_RULE_LABELS}
-          TARGET_CPU_FEATURES ${_RULE_TARGET_CPU_FEATURES}
           DEPENDS ${_RULE_DEPENDS}
           TIMEOUT ${_RULE_TIMEOUT}
         )
@@ -342,9 +332,7 @@ endfunction()
 # This function has 3 output-params: variables that it sets with PARENT_SCOPE:
 # _ENABLED, _FEATURES_NAME, _FEATURES.
 #
-# "default" is handled specially. _ENABLED is always set to "TRUE" and
-# _FEATURES_NAME and _FEATURES are set to
-# the empty string.
+# "generic" and "host" are handled specially, as CPU names.
 #
 # Other values are parsed as "arch:features_name:features". The `arch`
 # component is  matched with `IREE_ARCH`, `_ENABLED` is set to "TRUE" if and
@@ -354,23 +342,28 @@ endfunction()
 #
 # Examples:
 #
-# default:
-#    _ENABLED="TRUE" unconditionally,
-#        other output strings are "".
+# generic:
+#    _ENABLED="TRUE" unconditionally, _FEATURES_NAME="generic", _FEATURES="".
+#
+# host:
+#    _ENABLED="TRUE" unconditionally, _FEATURES_NAME="host", _FEATURES="".
 #
 # aarch64:dotprod:+dotprod:
 #    _ENABLED="TRUE" if the target architecture is aarch64, and in that case:
 #        _FEATURES_NAME="dotprod".
 #        _FEATURES="+dotprod".
 function(parse_target_cpu_features_variant _VARIANT_STRING _ENABLED_VAR
-             _FEATURES_NAME_VAR _FEATURES_VAR)
+             _FEATURES_NAME_VAR _COMPILER_FLAGS_VAR)
   set("${_ENABLED_VAR}" FALSE PARENT_SCOPE)
   set("${_FEATURES_NAME_VAR}" "" PARENT_SCOPE)
-  set("${_FEATURES_VAR}" "" PARENT_SCOPE)
-  if("${_VARIANT_STRING}" STREQUAL "default")
+  set("${_COMPILER_FLAGS_VAR}" "" PARENT_SCOPE)
+  if("${_VARIANT_STRING}" STREQUAL "generic" OR "${_VARIANT_STRING}" STREQUAL "host")
     set("${_ENABLED_VAR}" TRUE PARENT_SCOPE)
+    set("${_FEATURES_NAME_VAR}" "${_VARIANT_STRING}" PARENT_SCOPE)
+    set("${_COMPILER_FLAGS_VAR}" "--iree-llvmcpu-target-cpu=${_VARIANT_STRING}" PARENT_SCOPE)
     return()
   endif()
+
   # Interpret _VARIANT_STRING as a CMake list (;-separated).
   string(REPLACE ":" ";" _COMPONENTS "${_VARIANT_STRING}")
   list(LENGTH _COMPONENTS _NUM_COMPONENTS)
@@ -385,7 +378,7 @@ function(parse_target_cpu_features_variant _VARIANT_STRING _ENABLED_VAR
   if(_FILTER_ARCH STREQUAL IREE_ARCH)
     set("${_ENABLED_VAR}" TRUE PARENT_SCOPE)
     set("${_FEATURES_NAME_VAR}" "${_FEATURES_NAME}" PARENT_SCOPE)
-    set("${_FEATURES_VAR}" "${_FEATURES}" PARENT_SCOPE)
+    set("${_COMPILER_FLAGS_VAR}" "--iree-llvmcpu-target-cpu-features=${_FEATURES}" PARENT_SCOPE)
   endif()
 endfunction()
 
@@ -413,8 +406,8 @@ endfunction()
 #   LABELS: Additional labels to apply to the generated tests. The package path is
 #       added automatically.
 #   TARGET_CPU_FEATURES_VARIANTS: list of target cpu features variants. Each
-#       entry is either "default" for the architecture defaults, or a colon-
-#       separated triple "arch:name:cpu_features" where "arch" filters
+#       entry is either "generic" for the architecture defaults, or "host" for
+#       the host CPU, or a colon-separated triple "arch:name:cpu_features" where "arch" filters
 #       for a target CPU architecture (in IREE_ARCH format), "name" is a
 #       short name for the CPU features set (used to generate target names)
 #       and cpu_features is a comma-separated list of LLVM target attributes
@@ -443,7 +436,7 @@ function(iree_check_test_suite)
   if(_RULE_TARGET_CPU_FEATURES_VARIANTS)
     set(_TARGET_CPU_FEATURES_VARIANTS "${_RULE_TARGET_CPU_FEATURES_VARIANTS}")
   else()
-    set(_TARGET_CPU_FEATURES_VARIANTS "default")
+    set(_TARGET_CPU_FEATURES_VARIANTS "generic")
   endif()
 
   if(NOT DEFINED _RULE_TARGET_BACKENDS AND NOT DEFINED _RULE_DRIVERS)
@@ -466,7 +459,7 @@ function(iree_check_test_suite)
     list(GET _RULE_DRIVERS ${_INDEX} _DRIVER)
     foreach(_VARIANT_STRING IN LISTS _TARGET_CPU_FEATURES_VARIANTS)
       parse_target_cpu_features_variant("${_VARIANT_STRING}"
-        _ENABLED _TARGET_CPU_FEATURES_NAME _TARGET_CPU_FEATURES)
+        _ENABLED _TARGET_CPU_FEATURES_NAME _VARIANT_COMPILER_FLAGS)
       if(NOT _ENABLED)
         # The current entry is disabled on the target CPU architecture.
         continue()
@@ -488,12 +481,11 @@ function(iree_check_test_suite)
           ${_DRIVER}
         COMPILER_FLAGS
           ${_RULE_COMPILER_FLAGS}
+          ${_VARIANT_COMPILER_FLAGS}
         RUNNER_ARGS
           ${_RULE_RUNNER_ARGS}
         LABELS
           ${_LABELS}
-        TARGET_CPU_FEATURES
-          ${_TARGET_CPU_FEATURES}
         TIMEOUT
           ${_RULE_TIMEOUT}
         INPUT_TYPE