[hexagon-mlir] Fixes for v73,v79 compilation.

mabsar · mabsar · commit 56270c3a3346 · 2026-05-11T16:00:14.000-07:00
Some changes necessary to get hexagon-mlir
compile and run tests on different cpu targets.

Signed-off-by: mabsar &lt;mabsar@qti.qualcommm.com&gt;
diff --git a/qcom_hexagon_backend/backend/hexagon_executor.py b/qcom_hexagon_backend/backend/hexagon_executor.py
@@ -22,6 +22,26 @@
 )
 from triton.backends.qcom_hexagon_backend.hexagon_profiler import HexagonProfiler
 
+
+def _sdk_tool_version(q6_version: str) -> str:
+    """Return the SDK prebuilt tool version for the given Hexagon arch version.
+
+    v79+ devices require toolv88; older devices use toolv87.
+    """
+    return "v88" if int(q6_version.lstrip("v")) >= 79 else "v87"
+
+
+def _qhmath_sdk_path(q6_version: str) -> tuple:
+    """Tool/arch version for qhmath_hvx prebuilt libs.
+
+    v79 qhmath_hvx in SDK <= 6.4.0.0 is missing fp16 _ahf symbols;
+    fall back to v75 toolv87 libs until HSDK 6.6.0.
+    """
+    if int(q6_version.lstrip("v")) >= 79:
+        return ("v87", "75")
+    return (_sdk_tool_version(q6_version), q6_version)
+
+
 # This file is part of a small subset of python files that uses some type-annotations
 # and it passes type-verification with mypy (a type checker).
 # To typecheck this set of files, do:
@@ -31,7 +51,7 @@
 class HexagonExecutor:
     def __init__(
         self,
-        kernel_run_id,
+        kernel_run_id: str,
         enable_lwp=False,
         enable_etm=False,
         compile_only=False,
@@ -50,6 +70,11 @@ def __init__(
                                  environment variables, tools paths, and Q6 version, set by
                                  calling the `get_config` method.
         """
+        if not compile_only:
+            if (not isinstance(kernel_run_id, str)) or (len(kernel_run_id) == 0):
+                raise ValueError(
+                    "kernel_run_id must be well-formed, non-empty string for execution via standalone launcher"
+                )
         self.exec_mode = get_exec_mode() if not compile_only else "compile_only"
         self.device_path = f"/data/local/tmp/{kernel_run_id}"
         self.lib_path = f"{self.device_path}/lib"
@@ -229,9 +254,11 @@ def generate_shared_object(
             HEXAGON_SDK_ROOT=self.config.env_vars["HEXAGON_SDK_ROOT"],
             Q6_VERSION=self.config.Q6_VERSION,
         )
-        QHL_LINK_DIR = """{HEXAGON_SDK_ROOT}/libs/qhl_hvx/prebuilt/hexagon_toolv87_v{Q6_VERSION}""".format(
+        qhmath_tool_ver, qhmath_arch_ver = _qhmath_sdk_path(self.config.Q6_VERSION)
+        QHL_LINK_DIR = "{HEXAGON_SDK_ROOT}/libs/qhl_hvx/prebuilt/hexagon_tool{TOOL_VERSION}_v{Q6_VERSION}".format(
             HEXAGON_SDK_ROOT=self.config.env_vars["HEXAGON_SDK_ROOT"],
-            Q6_VERSION=self.config.Q6_VERSION,
+            TOOL_VERSION=qhmath_tool_ver,
+            Q6_VERSION=qhmath_arch_ver,
         )
 
         if not (
@@ -248,9 +275,10 @@ def generate_shared_object(
             QHL_LINK_DIR,
         )
 
-        QHMATH_DIR = """{HEXAGON_SDK_ROOT}/libs/qhl/prebuilt/hexagon_toolv87_v{Q6_VERSION}""".format(
+        QHMATH_DIR = "{HEXAGON_SDK_ROOT}/libs/qhl/prebuilt/hexagon_tool{TOOL_VERSION}_v{Q6_VERSION}".format(
             HEXAGON_SDK_ROOT=self.config.env_vars["HEXAGON_SDK_ROOT"],
-            Q6_VERSION=self.config.Q6_VERSION,
+            TOOL_VERSION=qhmath_tool_ver,
+            Q6_VERSION=qhmath_arch_ver,
         )
         if os.path.exists(QHMATH_DIR) and os.path.exists(
             os.path.join(QHMATH_DIR, "libqhmath.a")
@@ -366,11 +394,10 @@ def run_kernel_on_device(
         )
         librun_main_on_hexagon_skel_path = os.path.join(
             self.config.env_vars["HEXAGON_SDK_ROOT"],
-            "libs/run_main_on_hexagon/ship/hexagon_toolv87_v{}/librun_main_on_hexagon_skel.so".format(
-                self.config.Q6_VERSION
+            "libs/run_main_on_hexagon/ship/hexagon_tool{}_v{}/librun_main_on_hexagon_skel.so".format(
+                _sdk_tool_version(self.config.Q6_VERSION), self.config.Q6_VERSION
             ),
         )
-
         libcpp_path = os.path.join(
             self.config.env_vars["HEXAGON_TOOLS"],
             "target/hexagon/lib/v{}/G0/pic/libc++.so.1".format(self.config.Q6_VERSION),
@@ -602,13 +629,11 @@ def run_kernel_on_device(
 
         try:
             if self.enable_etm:
-                if "mlir_ciface" in principal_lib_without_ext:
-                    kernel_name = ""  # torch-mlir flow -> not just one kernel
-                else:
-                    kernel_name = principal_lib_without_ext[3:]  # remove lib prefix
-
                 hex_prof = HexagonProfiler(
-                    etm_local_dir, profiling_mode="etm", kernel_name=kernel_name
+                    etm_local_dir,
+                    profiling_mode="etm",
+                    device_lib_path=path_to_principal_lib_on_device,
+                    local_lib_filename=f"{principal_lib_without_ext}.so",
                 )
             for command, should_run in commands:
                 if not should_run:
@@ -724,16 +749,18 @@ def run_kernel_on_simulator(self, paths_to_shared_libs_local: list[str]):
                     self.config.env_vars["HEXAGON_TOOLS"], SIM_Q6SS_PATH
                 )
             ),
-            ("{} -mv{} \
+            (
+                "{} -mv{} \
                 --usefs={}/../Tools/target/hexagon/lib/v{}/G0/pic \
                 --simulated_returnval \
                 --cosim_file {} \
                 --l2tcm_base 0xd800 \
                 --rtos {} \
                 {}/rtos/qurt/computev{}/sdksim_bin/runelf.pbn -- \
-                {}/libs/run_main_on_hexagon/ship/hexagon_toolv87_v{}/run_main_on_hexagon_sim \
+                {}/libs/run_main_on_hexagon/ship/hexagon_tool{}_v{}/run_main_on_hexagon_sim \
                 stack_size=0x400000 -- \
-                {}").format(
+                {}"
+            ).format(
                 self.config.HEX_TOOLS["hexagon-sim"],
                 self.config.Q6_VERSION,
                 self.config.env_vars["HEXAGON_TOOLS"],
@@ -743,6 +770,7 @@ def run_kernel_on_simulator(self, paths_to_shared_libs_local: list[str]):
                 self.config.env_vars["HEXAGON_SDK_ROOT"],
                 self.config.Q6_VERSION,
                 self.config.env_vars["HEXAGON_SDK_ROOT"],
+                _sdk_tool_version(self.config.Q6_VERSION),
                 self.config.Q6_VERSION,
                 path_to_principal_lib_local,
             ),
diff --git a/qcom_hexagon_backend/backend/hexagon_options.py b/qcom_hexagon_backend/backend/hexagon_options.py
@@ -18,7 +18,10 @@ class HexagonOptions:
     allow_fp8e4nv: bool = False
     allowed_dot_input_precisions: Tuple[str] = ("ieee",)
     arch_triple: str = "hexagon"
-    arch_features: str = f'+hvxv{os.getenv("HEXAGON_ARCH_VERSION")},+hvx-length128b'
+    arch_features: str = (
+        f'+hvxv{os.getenv("HEXAGON_ARCH_VERSION")},+hvx-length128b'
+        + (",+hvx-ieee-fp" if int(os.getenv("HEXAGON_ARCH_VERSION", "0")) >= 79 else "")
+    )
     device_type: str = "hexagon"
     vectorize: int = 1
     vector_length: int = 32
diff --git a/qcom_hexagon_backend/backend/hexagon_profiler.py b/qcom_hexagon_backend/backend/hexagon_profiler.py
@@ -214,7 +214,7 @@ def write_config_lanai(self, addresses):
                 "0x0",
                 "0x0",
             ],
-            "ver": "v75",
+            "ver": "v79",
             "target": "Hexagon",
         }
 
diff --git a/qcom_hexagon_backend/bin/runtime/CMakeLists.txt b/qcom_hexagon_backend/bin/runtime/CMakeLists.txt
@@ -49,6 +49,7 @@ set(HEXAGON_INCLUDES
     -I${HEXAGON_SDK_ROOT}/incs
     -I${HEXAGON_SDK_ROOT}/incs/stddef
     -I${HEXAGON_SDK_ROOT}/libs/qhl_hvx/inc
+    -I${HEXAGON_SDK_ROOT}/libs/qhl_hvx/inc/internal
     -I${HEXKL_ROOT}/include
     -I${HEXAGON_SDK_ROOT}/rtos/qurt/computev${Q6_VERSION}/include/posix
     -I${HEXAGON_SDK_ROOT}/rtos/qurt/computev${Q6_VERSION}/include/qurt
diff --git a/qcom_hexagon_backend/bin/runtime/test/CMakeLists.txt b/qcom_hexagon_backend/bin/runtime/test/CMakeLists.txt
@@ -53,12 +53,19 @@ set(DEVICE_PATH "/vendor/bin/${BENCHMARK}")
 set(Q6_VERSION "$ENV{HEXAGON_ARCH_VERSION}")
 set(DSP_STACK_SIZE "262144")
 
+# v79+ devices require toolv88; older devices use toolv87.
+if(Q6_VERSION GREATER_EQUAL 79)
+  set(HEXAGON_TOOL_VERSION "v88")
+else()
+  set(HEXAGON_TOOL_VERSION "v87")
+endif()
+
 # Generate the run_test.sh script to run the test
 # TODO: Add option to pass gtest args in commandline
 set(SHELL_SCRIPT_CONTENT "#!/bin/bash"
   "set -x"
   "adb -H \$ANDROID_HOST -s \$ANDROID_SERIAL push ${HEXAGON_SDK_ROOT}/libs/run_main_on_hexagon/ship/android_aarch64/run_main_on_hexagon ${DEVICE_PATH}/run_main_on_hexagon"
-  "adb -H \$ANDROID_HOST -s \$ANDROID_SERIAL push ${HEXAGON_SDK_ROOT}/libs/run_main_on_hexagon/ship/hexagon_toolv87_v${Q6_VERSION}/librun_main_on_hexagon_skel.so /vendor/lib/rfsa/adsp/"
+  "adb -H \$ANDROID_HOST -s \$ANDROID_SERIAL push ${HEXAGON_SDK_ROOT}/libs/run_main_on_hexagon/ship/hexagon_tool${HEXAGON_TOOL_VERSION}_v${Q6_VERSION}/librun_main_on_hexagon_skel.so /vendor/lib/rfsa/adsp/"
   "adb -H \$ANDROID_HOST -s \$ANDROID_SERIAL push ${TEST_PATH}/${BENCHMARK}.so ${DEVICE_PATH}"
   "adb -H \$ANDROID_HOST -s \$ANDROID_SERIAL shell 'cd ${DEVICE_PATH}\; touch /vendor/lib/rfsa/adsp/run_main_on_hexagon.farf\; ./run_main_on_hexagon 3 ${DEVICE_PATH}/${BENCHMARK}.so --gtest-args=\"\" stack_size=262144'"
   "adb -H \$ANDROID_HOST -s \$ANDROID_SERIAL pull ${DEVICE_PATH}/gtest_output.txt ."
diff --git a/qcom_hexagon_backend/lib/Target/HEX_LLVMIR/LLVMIRTranslation.cpp b/qcom_hexagon_backend/lib/Target/HEX_LLVMIR/LLVMIRTranslation.cpp
@@ -273,8 +273,25 @@ std::unique_ptr<llvm::Module> translateHexagonMlirLlvmToLLVMIR(
   return llvmIR;
 }
 
+// Derive the Hexagon CPU name from arch_features string.
+// e.g., "+hvxv79,+hvx-length128b" -> "hexagonv79"
+static std::string deriveCpuFromFeatures(const std::string &features) {
+  auto pos = features.find("+hvxv");
+  if (pos != std::string::npos) {
+    pos += 5; // skip "+hvxv"
+    size_t end = pos;
+    while (end < features.size() && std::isdigit((unsigned char)features[end]))
+      end++;
+    if (end > pos)
+      return "hexagonv" + features.substr(pos, end - pos);
+  }
+  return "";
+}
+
 static std::unique_ptr<llvm::TargetMachine>
-create_target_machine(const llvm::Triple &target_triple) {
+create_target_machine(const llvm::Triple &target_triple,
+                      const std::string &cpu = "",
+                      const std::string &features = "") {
 
   // Hexagon target instance is needed to give a broader description of the
   // target platform
@@ -296,7 +313,7 @@ create_target_machine(const llvm::Triple &target_triple) {
       llvm::CodeGenOptLevel::Aggressive;
 
   llvm::TargetMachine *targetMachine = target_instance->createTargetMachine(
-      target_triple, "", "", target_options, reloc_model, code_model,
+      target_triple, cpu, features, target_options, reloc_model, code_model,
       codegen_optlevel);
 
   if (!targetMachine)
@@ -373,8 +390,23 @@ llvm_module_to_obj_string(std::unique_ptr<llvm::Module> &llvmModule) {
   initializeHexagonTarget();
 
   auto target_triple = llvmModule->getTargetTriple();
+
+  // Extract target features from function attributes so the TargetMachine
+  // is configured with the correct CPU (e.g. hexagonv79) and features
+  // (e.g. +hvx-ieee-fp).  Without this, the base TargetMachine defaults to
+  // a generic Hexagon CPU and the LLVM backend rejects v79-specific
+  // instructions even when they appear in the LLVM IR.
+  std::string features;
+  for (const auto &f : *llvmModule) {
+    if (!f.isDeclaration() && f.hasFnAttribute("target-features")) {
+      features = f.getFnAttribute("target-features").getValueAsString().str();
+      break;
+    }
+  }
+  std::string cpu = deriveCpuFromFeatures(features);
+
   std::unique_ptr<llvm::TargetMachine> targetMachine =
-      create_target_machine(target_triple);
+      create_target_machine(target_triple, cpu, features);
 
   // Set options for debugging
   setupCodegenDumpOptions();
diff --git a/scripts/build_hexagon_mlir.sh b/scripts/build_hexagon_mlir.sh
@@ -51,6 +51,22 @@ export PATH="${HOST_TOOLCHAIN}/bin:${PATH}"
 export CC="${HOST_TOOLCHAIN}/bin/clang"
 export CXX="${HOST_TOOLCHAIN}/bin/clang++"
 
+# Get libstdc++ from GCC 12 (Ubuntu 22.04) to satisfy GLIBCXX_3.4.30 requirement
+cd ${BASE_DIR}
+echo "Fetching libstdc++ (GCC 12)..."
+mkdir -p LIBSTDCXX
+cd LIBSTDCXX
+
+if [[ ! -f libstdc++.so.6 ]]; then
+  wget http://security.ubuntu.com/ubuntu/pool/main/g/gcc-12/libstdc++6_12.3.0-1ubuntu1~22.04.3_amd64.deb
+  ar x libstdc++6_12.3.0-1ubuntu1~22.04.3_amd64.deb
+  tar --use-compress-program=unzstd -xf data.tar.zst --strip-components=4 ./usr/lib/x86_64-linux-gnu/libstdc++.so.6.0.30
+  ln -sf libstdc++.so.6.0.30 libstdc++.so.6
+else
+  echo "libstdc++ already present. Skipping."
+fi
+export LD_PRELOAD="${BASE_DIR}/LIBSTDCXX/libstdc++.so.6:${LD_PRELOAD:-}"
+
 # Get HEXAGON_SDK
 cd ${BASE_DIR}
 echo "extracting HEXAGON_SDK..."
diff --git a/scripts/build_triton.sh b/scripts/build_triton.sh
@@ -28,7 +28,7 @@ PYTHON_VERSION=$(python3 -c "import sys; print(f'{sys.version_info.major}.{sys.v
 # Triton shared path
 export TRITON_SHARED_OPT_PATH=$TRITON_ROOT/build/cmake.linux-x86_64-cpython-${PYTHON_VERSION}/third_party/triton_shared/tools/triton-shared-opt/triton-shared-opt
 
-export HEXAGON_ARCH_VERSION=75
+export HEXAGON_ARCH_VERSION=79
 export TRITON_HOME=$HEXAGON_MLIR_ROOT
 export TRITON_PLUGIN_DIRS="$HEXAGON_MLIR_ROOT/triton_shared;$HEXAGON_MLIR_ROOT/qcom_hexagon_backend"
 export PATH=$TRITON_ROOT/build/cmake.linux-x86_64-cpython-${PYTHON_VERSION}/third_party/qcom_hexagon_backend/bin/:$TRITON_ROOT/build/cmake.linux-x86_64-cpython-${PYTHON_VERSION}/third_party/triton_shared/tools/triton-shared-opt:$PATH
@@ -39,6 +39,7 @@ TRITON_BUILD_WITH_CCACHE=true \
 LLVM_INCLUDE_DIRS="$LLVM_PROJECT_BUILD_DIR/include" \
 LLVM_LIBRARY_DIR="$LLVM_PROJECT_BUILD_DIR/lib" \
 LLVM_SYSPATH="$LLVM_PROJECT_BUILD_DIR" \
+TRITON_APPEND_CMAKE_ARGS="-DCMAKE_BUILD_WITH_INSTALL_RPATH=ON" \
 pip install -e . --no-build-isolation --verbose
 
 echo "🎉 Triton build completed successfully."
diff --git a/scripts/set_local_env.sh b/scripts/set_local_env.sh
@@ -25,10 +25,15 @@ export HEXAGON_SDK_VERSION=6.4.0.2
 export HEXAGON_SDK_ROOT=${BASE_DIR}/HEXAGON_SDK/Hexagon_SDK/$HEXAGON_SDK_VERSION
 export HEXKL_ROOT=${BASE_DIR}/HEXKL_DIR/hexkl_addon
 
-export HEXAGON_ARCH_VERSION=75
+export HEXAGON_ARCH_VERSION=79
 export TRITON_HOME=$HEXAGON_MLIR_ROOT
 export TRITON_PLUGIN_DIRS="$HEXAGON_MLIR_ROOT/triton_shared;$HEXAGON_MLIR_ROOT/qcom_hexagon_backend"
 export PATH=$TRITON_ROOT/build/cmake.linux-x86_64-cpython-${PYTHON_VERSION}/third_party/qcom_hexagon_backend/bin/:$TRITON_ROOT/build/cmake.linux-x86_64-cpython-${PYTHON_VERSION}/third_party/triton_shared/tools/triton-shared-opt:$PATH
 export PYTHONPATH=$TRITON_ROOT/python:${PYTHONPATH:-}
 
+# Use locally downloaded libstdc++ (GCC 12) to satisfy GLIBCXX_3.4.30 requirement.
+# LD_PRELOAD is used instead of LD_LIBRARY_PATH because libtriton.so has an RPATH
+# pointing to miniconda's lib which would otherwise take precedence.
+export LD_PRELOAD="${BASE_DIR}/LIBSTDCXX/libstdc++.so.6:${LD_PRELOAD:-}"
+
 set +euxo pipefail

Original file line number	Diff line number	Diff line change
`@@ -214,7 +214,7 @@ def write_config_lanai(self, addresses):`
`214`	`214`	`"0x0",`
`215`	`215`	`"0x0",`
`216`	`216`	`],`
`217`		`- "ver": "v75",`
	`217`	`+ "ver": "v79",`
`218`	`218`	`"target": "Hexagon",`
`219`	`219`	`}`
`220`	`220`