Automated rollback of rollback. Fixed in preceding change.

PiperOrigin-RevId: 265455751

1 files changed, 108 insertions, 62 deletions

diff --git a/BUILD b/BUILD
index 7dd756a..f3dc408 100644
--- a/BUILD
+++ b/BUILD
@@ -2,31 +2,11 @@
 
 # TODO(b/123403203) actually make TFLite use ruy.
 
-load(":build_defs.bzl", "ruy_copts_avx2", "ruy_copts_skylake")
-load(":ruy_visibility.bzl", "ruy_visibility")
+load(":build_defs.bzl", "ruy_copts_avx2", "ruy_copts_base", "ruy_copts_skylake", "ruy_visibility")
 load(":ruy_test_ext.bzl", "ruy_test_ext_defines", "ruy_test_ext_deps")
 load(":ruy_test.bzl", "ruy_benchmark", "ruy_benchmark_opt_sets", "ruy_test")
 load("//third_party/tensorflow/lite:special_rules.bzl", "tflite_portable_test_suite")
 
-# 1. Enable -mfpu=neon unconditionally on ARM32. If it turns out that we need to support
-#    ARM32 without NEON then we'll implement runtime detection and dispatch at that point.
-# 2. Explicitly pass -O3 on mobile configs where just "-c opt" means "optimize for code size".
-#    We would want to only do that when compilation_mode is "opt", but limitations of
-#    the "select" keyword (no nested selects, no AND boolean) seem to make that difficult
-#    at the moment. For debugging purposes, one needs to manually edit this to remove these
-#    -O3. Otherwise, not even `bazel build --copt=-O0` will override that.
-RUY_COPTS = select({
-    "//third_party/tensorflow:android_arm64": [
-        "-O3",
-    ],
-    "//third_party/tensorflow:android_arm": [
-        "-O3",
-        "-mfpu=neon",
-    ],
-    "//conditions:default": [
-    ],
-})
-
 package(
     default_visibility = ["//visibility:private"],
     licenses = ["notice"],  # Apache 2.0
@@ -35,33 +15,33 @@ package(
 cc_library(
     name = "platform",
     hdrs = ["platform.h"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
 )
 
 cc_library(
     name = "check_macros",
     hdrs = ["check_macros.h"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = ["//third_party/tensorflow/lite/kernels/internal:compatibility"],
 )
 
 cc_library(
     name = "opt_set",
     hdrs = ["opt_set.h"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
 )
 
 cc_library(
     name = "time",
     hdrs = ["time.h"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
 )
 
 cc_library(
     name = "wait",
     srcs = ["wait.cc"],
     hdrs = ["wait.h"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [":time"],
 )
 
@@ -78,7 +58,7 @@ cc_test(
 cc_library(
     name = "size_util",
     hdrs = ["size_util.h"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [":check_macros"],
 )
 
@@ -99,7 +79,7 @@ cc_library(
     hdrs = [
         "tune.h",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [
         ":opt_set",
         ":platform",
@@ -132,7 +112,7 @@ cc_library(
     hdrs = [
         "allocator.h",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [
         ":check_macros",
         ":size_util",
@@ -151,7 +131,7 @@ cc_test(
 cc_library(
     name = "side_pair",
     hdrs = ["side_pair.h"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [":check_macros"],
 )
 
@@ -163,7 +143,7 @@ cc_library(
     hdrs = [
         "block_map.h",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [
         ":check_macros",
         ":opt_set",
@@ -181,7 +161,7 @@ cc_library(
     hdrs = [
         "blocking_counter.h",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [
         ":check_macros",
         ":wait",
@@ -196,7 +176,7 @@ cc_library(
     hdrs = [
         "thread_pool.h",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     visibility = ruy_visibility(),
     deps = [
         ":blocking_counter",
@@ -213,7 +193,7 @@ cc_library(
     hdrs = [
         "detect_arm.h",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     visibility = ruy_visibility(),
 )
 
@@ -225,7 +205,7 @@ cc_library(
     hdrs = [
         "detect_x86.h",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     visibility = ruy_visibility(),
     deps = [
         ":platform",
@@ -235,7 +215,7 @@ cc_library(
 cc_library(
     name = "path",
     hdrs = ["path.h"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     visibility = ruy_visibility(),
     deps = [
         ":platform",
@@ -251,7 +231,7 @@ cc_library(
     hdrs = [
         "trace.h",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [
         ":block_map",
         ":check_macros",
@@ -263,7 +243,7 @@ cc_library(
 cc_library(
     name = "matrix",
     hdrs = ["matrix.h"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     visibility = ruy_visibility(),
     deps = [":check_macros"],
 )
@@ -271,7 +251,7 @@ cc_library(
 cc_library(
     name = "spec",
     hdrs = ["spec.h"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     visibility = ruy_visibility(),
     deps = [":matrix"],
 )
@@ -279,7 +259,7 @@ cc_library(
 cc_library(
     name = "internal_matrix",
     hdrs = ["internal_matrix.h"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [
         ":check_macros",
         ":common",
@@ -293,7 +273,7 @@ cc_library(
     hdrs = [
         "common.h",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [
         ":check_macros",
         ":matrix",
@@ -311,7 +291,7 @@ cc_library(
         "kernel_common.h",
         "kernel_x86.h",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [
         ":check_macros",
         ":common",
@@ -337,7 +317,7 @@ cc_library(
         "pack_common.h",
         "pack_x86.h",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [
         ":check_macros",
         ":common",
@@ -357,7 +337,7 @@ cc_library(
         "kernel_arm32.cc",
         "kernel_arm64.cc",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [
         ":common",
         ":kernel_common",
@@ -372,7 +352,7 @@ cc_library(
     srcs = [
         "pack_arm.cc",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [
         ":common",
         ":opt_set",
@@ -382,12 +362,17 @@ cc_library(
     ],
 )
 
+# AVX-512 compilation units.
+#
+# These must use the same compiler options.
+RUY_COPTS_BUILT_FOR_AVX512 = ruy_copts_base() + ruy_copts_skylake()
+
 cc_library(
     name = "kernel_avx512",
     srcs = [
         "kernel_avx512.cc",
     ],
-    copts = RUY_COPTS + ruy_copts_skylake(),
+    copts = RUY_COPTS_BUILT_FOR_AVX512,
     deps = [
         ":check_macros",
         ":kernel_common",
@@ -402,7 +387,7 @@ cc_library(
     srcs = [
         "pack_avx512.cc",
     ],
-    copts = RUY_COPTS + ruy_copts_skylake(),
+    copts = RUY_COPTS_BUILT_FOR_AVX512,
     deps = [
         ":check_macros",
         ":matrix",
@@ -415,11 +400,32 @@ cc_library(
 )
 
 cc_library(
+    name = "have_built_path_for_avx512",
+    srcs = [
+        "have_built_path_for_avx512.cc",
+    ],
+    hdrs = [
+        "have_built_path_for.h",
+    ],
+    copts = RUY_COPTS_BUILT_FOR_AVX512,
+    deps = [
+        ":opt_set",
+        ":platform",
+    ],
+)
+# End: AVX-512 compilation units.
+
+# AVX2 compilation units.
+#
+# These must use the same compiler options.
+RUY_COPTS_BUILT_FOR_AVX2 = ruy_copts_base() + ruy_copts_avx2()
+
+cc_library(
     name = "kernel_avx2",
     srcs = [
         "kernel_avx2.cc",
     ],
-    copts = RUY_COPTS + ruy_copts_avx2(),
+    copts = RUY_COPTS_BUILT_FOR_AVX2,
     deps = [
         ":check_macros",
         ":kernel_common",
@@ -434,7 +440,7 @@ cc_library(
     srcs = [
         "pack_avx2.cc",
     ],
-    copts = RUY_COPTS + ruy_copts_avx2(),
+    copts = RUY_COPTS_BUILT_FOR_AVX2,
     deps = [
         ":check_macros",
         ":matrix",
@@ -447,12 +453,28 @@ cc_library(
 )
 
 cc_library(
+    name = "have_built_path_for_avx2",
+    srcs = [
+        "have_built_path_for_avx2.cc",
+    ],
+    hdrs = [
+        "have_built_path_for.h",
+    ],
+    copts = RUY_COPTS_BUILT_FOR_AVX2,
+    deps = [
+        ":opt_set",
+        ":platform",
+    ],
+)
+# End: AVX2 compilation units.
+
+cc_library(
     name = "kernel",
     hdrs = [
         "kernel.h",
         "kernel_common.h",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [
         ":check_macros",
         ":common",
@@ -480,7 +502,7 @@ cc_library(
         "pack.h",
         "pack_common.h",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [
         ":check_macros",
         ":common",
@@ -499,6 +521,18 @@ cc_library(
 )
 
 cc_library(
+    name = "have_built_path_for",
+    hdrs = [
+        "have_built_path_for.h",
+    ],
+    deps = [
+        ":have_built_path_for_avx2",
+        ":have_built_path_for_avx512",
+        ":platform",
+    ],
+)
+
+cc_library(
     name = "context",
     srcs = [
         "context.cc",
@@ -506,13 +540,14 @@ cc_library(
     hdrs = [
         "context.h",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     visibility = ruy_visibility(),
     deps = [
         ":allocator",
         ":check_macros",
         ":detect_arm",
         ":detect_x86",
+        ":have_built_path_for",
         ":path",
         ":platform",
         ":thread_pool",
@@ -521,10 +556,21 @@ cc_library(
     ],
 )
 
+cc_test(
+    name = "context_test",
+    srcs = ["context_test.cc"],
+    deps = [
+        ":context",
+        ":path",
+        ":platform",
+        "@com_google_googletest//:gtest",
+    ],
+)
+
 cc_library(
     name = "trmul_params",
     hdrs = ["trmul_params.h"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [
         ":internal_matrix",
         ":side_pair",
@@ -536,7 +582,7 @@ cc_library(
     name = "trmul",
     srcs = ["trmul.cc"],
     hdrs = ["trmul.h"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [
         ":allocator",
         ":block_map",
@@ -568,7 +614,7 @@ cc_library(
         "ruy.h",
         "ruy_advanced.h",
     ],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     visibility = ruy_visibility(),
     deps = [
         ":check_macros",
@@ -610,7 +656,7 @@ cc_library(
     testonly = True,
     srcs = ["pmu.cc"],
     hdrs = ["pmu.h"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     deps = [":check_macros"],
 )
 
@@ -619,7 +665,7 @@ cc_library(
     name = "test_lib",
     testonly = True,
     hdrs = ["test.h"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     # need defines, not copts, because it's controlling a header, test.h
     defines = ruy_test_ext_defines(),
     linkopts = select({
@@ -640,7 +686,7 @@ cc_library(
 ruy_benchmark(
     name = "benchmark",
     srcs = ["benchmark.cc"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     lhs_rhs_accum_dst = [
         ("f32", "f32", "f32", "f32"),
         ("u8", "u8", "i32", "u8"),
@@ -654,7 +700,7 @@ ruy_benchmark(
 ruy_test(
     name = "test_fast",
     srcs = ["test_fast.cc"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     lhs_rhs_accum_dst = [
         ("f32", "f32", "f32", "f32"),
         ("f64", "f32", "f64", "f32"),
@@ -670,7 +716,7 @@ ruy_test(
 ruy_test(
     name = "test_slow",
     srcs = ["test_slow.cc"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     lhs_rhs_accum_dst = [
         ("f32", "f32", "f32", "f32"),
         ("u8", "u8", "i32", "u8"),
@@ -684,7 +730,7 @@ ruy_test(
 ruy_test(
     name = "test_special_specs",
     srcs = ["test_special_specs.cc"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     lhs_rhs_accum_dst = [
         ("f32", "f32", "f32", "f32"),
         ("u8", "u8", "i32", "u8"),
@@ -695,7 +741,7 @@ ruy_test(
 ruy_benchmark_opt_sets(
     name = "benchmark_opt_set",
     srcs = ["benchmark.cc"],
-    copts = RUY_COPTS,
+    copts = ruy_copts_base(),
     lhs_rhs_accum_dst = [
         ("f32", "f32", "f32", "f32"),
         ("u8", "u8", "i32", "u8"),