ggml-hexagon: fix a minior issue in dev ops

Author: zhouwg

ggml/src/ggml-hexagon/CMakeLists.txt

@@ -115,4 +115,16 @@ function(ggml_hexagon_build_kernel KNAME)
     )
 endfunction()
 
+function(ggml_hexagon_setup_cfg KNAME)
+    message(STATUS "ggml_hexagon: setup runtime configuration file ${KNAME}")
+    add_custom_command(
+        TARGET ${PROJECT_NAME}
+        POST_BUILD
+        COMMAND echo "current working path:`pwd`\n"
+        COMMAND /bin/cp -fv ../../../../../scripts/${KNAME}  ../../../bin/
+        COMMENT "setup runtime configuration file"
+    )
+endfunction()
+
 ggml_hexagon_build_kernel("cdsp")
+ggml_hexagon_setup_cfg("ggml-hexagon.cfg")

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -1830,7 +1830,7 @@ static void ggmlhexagon_load_cfg() {
     });
     std::string precision_mode;
     std::string ggml_hexagon_version;
-    hexagoncfg_instance.get_stringvalue("general", "ggml_hexagon_version", ggml_hexagon_version, "1.03");
+    hexagoncfg_instance.get_stringvalue("general", "version", ggml_hexagon_version, "1.00");
     hexagoncfg_instance.get_intvalue("general", "enable_perf", g_hexagon_appcfg.enable_perf, 1);
     hexagoncfg_instance.get_intvalue("general", "print_tensors_info", g_hexagon_appcfg.print_tensors_info, 0);
     hexagoncfg_instance.get_intvalue("general", "dump_op_info", g_hexagon_appcfg.dump_op_info, 0);

scripts/build-run-android.sh

@@ -12,6 +12,7 @@ PWD=`pwd`
 REMOTE_PATH=/data/local/tmp/
 #LLM model file on Android phone
 GGUF_MODEL_NAME=/sdcard/gemma-3-4b-it-Q8_0.gguf
+#https://huggingface.co/Qwen/Qwen1.5-1.8B-Chat-GGUF/blob/main/qwen1_5-1_8b-chat-q4_0.gguf
 GGUF_MODEL_NAME=/sdcard/qwen1_5-1_8b-chat-q4_0.gguf
 
 #Android NDK can be found at:
@@ -241,6 +242,10 @@ function prepare_run_on_phone()
         adb push ./out/android/bin/*.so ${REMOTE_PATH}/
     fi
     adb push ./out/android/bin/${program} ${REMOTE_PATH}/
+    #for non developers: deploy dev ops once time with build outputs in ./out/android/bin/
+    #adb push ./out/android/bin/ggml-hexagon.cfg ${REMOTE_PATH}/
+    #for developers: modify ./scritps/ggml-hexagon.cfg before run
+    adb push ./scripts/ggml-hexagon.cfg ${REMOTE_PATH}/
     adb shell chmod +x ${REMOTE_PATH}/${program}
 }
 

scripts/ggml-hexagon.cfg

@@ -40,14 +40,16 @@ hwaccel_approach = 2
 #attention:
 #          a. HWACCEL_QNN_SINGLEGRAPH not supported at the moment;
 #          b. following combinations are valid:
-#             1: hwaccel_approach = 2 AND hexagon_backend = 2(this is the default setting)
-#             2: hwaccel_approach = 0 AND hexagon_backend = 2(QNNNPU)
+#             1: hwaccel_approach = 2 AND hexagon_backend = 2(HWACCEL_CDSP, this is the default setting)
+#             2: hwaccel_approach = 0 AND hexagon_backend = 2(HWACCEL_QNN, QNNNPU)
 #             3: hwaccel_approach = 0 AND hexagon_backend = 1(QNNGPU)
 #             4: hwaccel_approach = 0 AND hexagon_backend = 0(QNNCPU)
-#             5: hwaccel_approach = 2 AND hexagon_backend = 3
-#             6: hwaccel_approach = 0 AND hexagon_backend = 3
+#             5: hwaccel_approach = 2 AND hexagon_backend = 3(fall back to the default ggml backend)
+#             6: hwaccel_approach = 0 AND hexagon_backend = 3(fall back to the default ggml backend)
 #
-#generally speaking, we only need to focus on b-1 and b-2 in this PR.
+#generally speaking,
+#          a. we only need to focus on b-1 and b-2 in this PR.
+#          b. we can compare Hexagon NPU performance between HWACCEL_CDSP/HWACCEL_QNN(QNNNPU)/the default ggml backend accordingly
 
 
 #enable/disable offload quantized type mulmat