C compiler from the Android NDK. For example,

	CGO_ENABLED=1 \
	GOOS=android \
	GOARCH=arm64 \
	CC_FOR_TARGET=$NDK/toolchains/llvm/prebuilt/linux-x86_64/bin/aarch64-linux-android21-clang \
	./all.bash

To run tests on the Android device, add the bin directory to PATH so the
go tool can find the go_android_$GOARCH_exec wrapper generated by
make.bash. For example, to run the go1 benchmarks

# force all tests to run locally on the Aarch64 host.
test:rbe_cross_compile_base --strategy=TestRunner=local --build_tests_only
test:rbe_cross_compile_base --verbose_failures=true --local_test_jobs=HOST_CPUS --test_output=errors

# START LINUX AARCH64 CROSS-COMPILE CONFIGS
build:cross_compile_linux_arm64 --config=cross_compile_base

# Set the target CPU to Aarch64

    val buildSpecs: List<FlameGraphGenerationBuildSpec>
        get() =
            scenarios.flatMap { scenario ->
                Os.values().flatMap { os ->
                    val arch = if (os == Os.MACOS) Arch.AARCH64 else Arch.AMD64
                    if (os == Os.WINDOWS) {
                        listOf("jprofiler")
                    } else {
                        listOf("async-profiler", "async-profiler-heap")

object AdHocPerformanceScenarioMacOS : AdHocPerformanceScenario(Os.MACOS, Arch.AMD64)

  *AMD64:CYGWIN* | *AMD64:MINGW*) distributionPlatform=windows-amd64 ;;
  :Darwin*x86_64) distributionPlatform=darwin-amd64 ;;
  :Darwin*arm64) distributionPlatform=darwin-aarch64 ;;
  :Linux*x86_64*) distributionPlatform=linux-amd64 ;;
  *)
    echo "Cannot detect native platform for mvnd on $(uname)-$(uname -m), use pure java version" >&2
    distributionPlatform=linux-amd64
    ;;
  esac

       * which will have an efficient native implementation in JDK 9.
       *
       */
      String arch = System.getProperty("os.arch");
      if (Objects.equals(arch, "amd64") || Objects.equals(arch, "aarch64")) {
        return ByteOrder.nativeOrder().equals(ByteOrder.LITTLE_ENDIAN)
            ? UnsafeByteArray.UNSAFE_LITTLE_ENDIAN
            : UnsafeByteArray.UNSAFE_BIG_ENDIAN;
      }
    } catch (Throwable t) {

	default:
		return 0, errors.New("invalid arrangement in ARM64 register list")
	}
	return (int64(curQ) & 1 << 30) | (int64(curSize&3) << 10), nil
}

// ARM64RegisterListOffset generates offset encoding according to AArch64 specification.
func ARM64RegisterListOffset(firstReg, regCnt int, arrangement int64) (int64, error) {
	offset := int64(firstReg)
	switch regCnt {
	case 1:
		offset |= 0x7 << 12
	case 2:
		offset |= 0xa << 12

    *   **aarch64 CPUs:** Experimental performance optimizations from
        [Compute Library for the Arm® Architecture (ACL)](https://github.com/ARM-software/ComputeLibrary)
        are available through oneDNN in the default Linux aarch64 package (`pip
        install tensorflow`).
        *   The optimizations are disabled by default.

pkg debug/elf, const R_AARCH64_ADR_PREL_LO21 = 274
pkg debug/elf, const R_AARCH64_ADR_PREL_LO21 R_AARCH64
pkg debug/elf, const R_AARCH64_ADR_PREL_PG_HI21 = 275
pkg debug/elf, const R_AARCH64_ADR_PREL_PG_HI21 R_AARCH64
pkg debug/elf, const R_AARCH64_ADR_PREL_PG_HI21_NC = 276
pkg debug/elf, const R_AARCH64_ADR_PREL_PG_HI21_NC R_AARCH64
pkg debug/elf, const R_AARCH64_CALL26 = 283
pkg debug/elf, const R_AARCH64_CALL26 R_AARCH64

pkg debug/elf, const R_AARCH64_LD64_GOTOFF_LO15 R_AARCH64
pkg debug/elf, const R_AARCH64_LD64_GOTPAGE_LO15 = 313
pkg debug/elf, const R_AARCH64_LD64_GOTPAGE_LO15 R_AARCH64
pkg debug/elf, const R_AARCH64_TLSGD_ADR_PREL21 = 512
pkg debug/elf, const R_AARCH64_TLSGD_ADR_PREL21 R_AARCH64
pkg debug/elf, const R_AARCH64_TLSGD_MOVW_G0_NC = 516
pkg debug/elf, const R_AARCH64_TLSGD_MOVW_G0_NC R_AARCH64
pkg debug/elf, const R_AARCH64_TLSGD_MOVW_G1 = 515