Default rpc_mem to false for bazel

docs/user_guide/advanced_usage.rst

@@ -598,3 +598,16 @@ These weights can be quantized to 8bit to reduce the size to a quarter, whereas
 For example, the top-1 accuracy of MobileNetV1 after quantization of weights is 68.2% on the ImageNet validation set.
 ``quantize_large_weights`` can be specified as 1 in the deployment file to save these weights in 8bit and actual inference in float.
 It can be used for both CPU and GPU.
+
+Build with rpcmem
+-------------------
+For Qualcomm-based devices, it's possible to build the models and the ``mace_run`` executable with ```rpcmem`` support:
+
+.. code-block:: sh
+
+    python tools/converter.py convert --config=/path/to/model_deployment_file.yml --enable_rpcmem
+    python tools/converter.py run --config=/path/to/model_deployment_file.yml --enable_rpcmem
+
+For deployment, make sure the use of ``rpcmem`` is consisten between ``libmace.a`` and built models. Both have to be built
+with or without ``rpcmem`` support. For details about how to built ``libmace.a`` with ``rpcmem`` support see 
+`here <https://mace.readthedocs.io/en/latest/user_guide/basic_usage.html#build-mace-into-a-library>`.

docs/user_guide/advanced_usage_cmake.rst

@@ -175,7 +175,7 @@ After that you can rebuild the engine.
 
     .. code-block:: bash
 
-        RUNTIME=GPU RUNMODE=code QUANTIZE=OFF bash tools/cmake/cmake-build-armeabi-v7a.sh
+        RUNTIME=GPU RUNMODE=code QUANTIZE=OFF RPCMEM=OFF bash tools/cmake/cmake-build-armeabi-v7a.sh
 
 ``RUNMODE=code`` means you compile and link model library with MACE engine.
 

docs/user_guide/basic_usage.rst

@@ -45,14 +45,15 @@ Here we use the mobilenet-v2 model as an example.
         cd path/to/mace
         # Build library
         # output lib path: build/lib
-        bash tools/bazel_build_standalone_lib.sh [-abi=abi][-runtimes=rt1,rt2,...][-quantize][-static]
+        bash tools/bazel_build_standalone_lib.sh [-abi=abi][-runtimes=rt1,rt2,...][-quantize][-static][-rpcmem]
 
     .. note::
 
         - This step can be skipped if you just want to run a model using ``tools/converter.py``, such as commands in step 5.
         - Use the `-abi` parameter to specify the ABI. Supported ABIs are armeabi-v7a, arm64-v8a, arm_linux_gnueabihf, aarch64_linux_gnu and host (for host machine, linux-x86-64). The default ABI is arm64-v8a.
         - For each ABI, several runtimes can be chosen by specifying the `-runtimes` parameter. Supported runtimes are CPU, GPU, DSP and APU. By default, the library is built to run on CPU.
         - Omit the `-static` option if a shared library is desired instead of a static one. By default, a shared library is built.
+        - Omit the `-rpcmem` option if your target device chipset is not manufactured by Qualcomm.
         - See 'bash tools/bazel_build_standalone_lib.sh -help' for detailed information.
         - DO respect the hyphens ('-') and the underscores ('_') in the ABI.
 
@@ -189,7 +190,7 @@ Or use bazel to build MACE source code into a library.
         cd path/to/mace
         # Build library
         # output lib path: build/lib
-        bash tools/bazel_build_standalone_lib.sh [-abi=abi][-runtimes=rt1,rt2,...][-static]
+        bash tools/bazel_build_standalone_lib.sh [-abi=abi][-runtimes=rt1,rt2,...][-static][-rpcmem]
 
 The above command will generate static library ``build/lib/libmace.a`` dynamic library ``build/lib/libmace.so``.
 

docs/user_guide/basic_usage_cmake.rst

@@ -20,13 +20,15 @@ Please make sure you have CMake installed.
 
     .. code-block:: sh
 
-        RUNTIME=GPU QUANTIZE=OFF bash tools/cmake/cmake-build-armeabi-v7a.sh
+        RUNTIME=GPU QUANTIZE=OFF RPCMEM=OFF bash tools/cmake/cmake-build-armeabi-v7a.sh
 
 which generate libraries in ``build/cmake-build/armeabi-v7a``, you can use either static libraries or the ``libmace.so`` shared library.
 
 You can also build for other target abis: ``arm64-v8a``, ``arm-linux-gnueabihf``, ``aarch64-linux-gnu``, ``host``;
 and runtime: ``GPU``, ``HEXAGON``, ``HTA``, ``APU``.
 
+For Qualcomm devices, it is possible to set ``RPCMEM=ON`` to enable the use of specific memory optimizations.
+
 
 Model Conversion
 -------------------------------

tools/bazel_adb_run.py

@@ -108,7 +108,7 @@ def parse_args():
     parser.add_argument(
         "--enable_rpcmem",
         type=str2bool,
-        default=True,
+        default=False,
         help="Whether to use rpcmem")
     parser.add_argument(
         "--enable_hta",

tools/bazel_build_standalone_lib.sh

@@ -42,7 +42,7 @@ enable_dsp=false
 enable_apu=false
 enable_quantize=false
 enable_bfloat16=false
-enable_rpcmem=true
+enable_rpcmem=false
 static_lib=false
 symbol_hidden=
 runtime_label="cpu"
@@ -103,6 +103,9 @@ for opt in "${@}";do
     bfloat16|-bfloat16|--bfloat16)
       enable_bfloat16=true
       ;;
+    rpcmem|-rpcmem|--rpcmem)
+      enable_rpcmem=true
+      ;;
     help|-help|--help)
       helper
       ;;
@@ -115,10 +118,6 @@ $(echo "$1" | cut -d '=' -f -1)"
   esac
 done
 
-if [[ "${enable_apu}" == true || ("${abi}" != armeabi-v7a && "${abi}" != arm64-v8a) ]];then
-  enable_rpcmem=false
-fi
-
 if [[ "${static_lib}" == true ]];then
   lib_type=static
   lib_label=static

tools/cmake/cmake-build-arm64-v8a.sh

@@ -37,6 +37,11 @@ if [[ "$BFLOAT16" == "ON" ]]; then
     DMACE_ENABLE_BFLOAT16=ON
 fi
 
+MACE_ENABLE_RPCMEM=OFF
+if [[ "$RPCMEM" == "ON" ]]; then
+    MACE_ENABLE_RPCMEM=ON
+fi
+
 mkdir -p ${BUILD_DIR} && cd ${BUILD_DIR}
 cmake -DANDROID_ABI="arm64-v8a" \
       -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK_HOME}/build/cmake/android.toolchain.cmake \
@@ -56,7 +61,7 @@ cmake -DANDROID_ABI="arm64-v8a" \
       -DMACE_ENABLE_TESTS=ON              \
       -DMACE_ENABLE_BENCHMARKS=ON         \
       -DMACE_ENABLE_CODE_MODE=${MACE_ENABLE_CODE_MODE}        \
-      -DMACE_ENABLE_RPCMEM=ON                                 \
+      -DMACE_ENABLE_RPCMEM=${MACE_ENABLE_RPCMEM}              \
       -DCMAKE_INSTALL_PREFIX=install      \
       ../../..
 make -j$(nproc) VERBOSE=1 && make install

tools/cmake/cmake-build-armeabi-v7a.sh

@@ -37,6 +37,11 @@ if [[ "$BFLOAT16" == "ON" ]]; then
     DMACE_ENABLE_BFLOAT16=ON
 fi
 
+MACE_ENABLE_RPCMEM=OFF
+if [[ "$RPCMEM" == "ON" ]]; then
+    MACE_ENABLE_RPCMEM=ON
+fi
+
 mkdir -p ${BUILD_DIR} && cd ${BUILD_DIR}
 cmake -DANDROID_ABI="armeabi-v7a" \
       -DANDROID_ARM_NEON=ON \
@@ -56,7 +61,7 @@ cmake -DANDROID_ABI="armeabi-v7a" \
       -DMACE_ENABLE_TESTS=ON                                 \
       -DMACE_ENABLE_BENCHMARKS=ON                            \
       -DMACE_ENABLE_CODE_MODE=${MACE_ENABLE_CODE_MODE}       \
-      -DMACE_ENABLE_RPCMEM=ON                                \
+      -DMACE_ENABLE_RPCMEM=${MACE_ENABLE_RPCMEM}             \
       -DCMAKE_INSTALL_PREFIX=install                         \
       ../../..
 make -j$(nproc) VERBOSE=1 && make install

tools/converter.py

@@ -753,7 +753,7 @@ def print_configuration(configs):
     MaceLogger.summary(StringFormatter.table(header, data, title))
 
 
-def build_model_lib(configs, address_sanitizer, debug_mode):
+def build_model_lib(configs, address_sanitizer, use_rpc_mem, debug_mode):
     MaceLogger.header(StringFormatter.block("Building model library"))
 
     # create model library dir
@@ -776,6 +776,7 @@ def build_model_lib(configs, address_sanitizer, debug_mode):
             enable_quantize=quantize_enabled(configs),
             enable_bfloat16=bfloat16_enabled(configs),
             enable_fp16=fp16_enabled(configs),
+            enable_rpcmem=use_rpc_mem,
             address_sanitizer=address_sanitizer,
             symbol_hidden=get_symbol_hidden_mode(debug_mode),
             debug_mode=debug_mode
@@ -901,7 +902,10 @@ def convert_func(flags):
             StringFormatter.block("Model %s converted" % model_name))
 
     if model_graph_format == ModelFormat.code:
-        build_model_lib(configs, flags.address_sanitizer, flags.debug_mode)
+        build_model_lib(configs,
+                        flags.address_sanitizer,
+                        flags.use_rpc_mem,
+                        flags.debug_mode)
 
     print_library_summary(configs)
 
@@ -910,7 +914,7 @@ def convert_func(flags):
 # run
 ################################
 def build_mace_run(configs, target_abi, toolchain, address_sanitizer,
-                   mace_lib_type, debug_mode, device):
+                   mace_lib_type, use_rpc_mem, debug_mode, device):
     library_name = configs[YAMLKeyword.library_name]
 
     build_tmp_binary_dir = get_build_binary_dir(library_name, target_abi)
@@ -941,6 +945,7 @@ def build_mace_run(configs, target_abi, toolchain, address_sanitizer,
         enable_quantize=quantize_enabled(configs),
         enable_bfloat16=bfloat16_enabled(configs),
         enable_fp16=fp16_enabled(configs),
+        enable_rpcmem=use_rpc_mem,
         address_sanitizer=address_sanitizer,
         symbol_hidden=get_symbol_hidden_mode(debug_mode, mace_lib_type),
         debug_mode=debug_mode,
@@ -987,6 +992,7 @@ def run_mace(flags):
                                toolchain,
                                flags.address_sanitizer,
                                flags.mace_lib_type,
+                               flags.use_rpc_mem,
                                flags.debug_mode,
                                device)
                 # run
@@ -1065,6 +1071,10 @@ def parse_args():
         type=str,
         default="",
         help="Target SOCs, comma seperated list.")
+    all_type_parent_parser.add_argument(
+        "--use_rpc_mem",
+        action="store_true",
+        help="Enable rpc_mem optimizations.")
     all_type_parent_parser.add_argument(
         "--debug_mode",
         action="store_true",

tools/sh_commands.py

@@ -283,7 +283,7 @@ def bazel_build(target,
                 enable_quantize=True,
                 enable_bfloat16=False,
                 enable_fp16=False,
-                enable_rpcmem=True,
+                enable_rpcmem=False,
                 address_sanitizer=False,
                 symbol_hidden=True,
                 debug_mode=False,