ggml-qnn: pr to upstream

Author: zhouwg

CMakeLists.txt

@@ -6,7 +6,6 @@ include(CheckIncludeFileCXX)
 set(CMAKE_WARN_UNUSED_CLI YES)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
-set(CMAKE_VERBOSE_MAKEFILE on)
 
 if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
     set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)

cmake/aarch64-w64-mingw32.cmake

@@ -9,8 +9,8 @@ set( CMAKE_CXX_COMPILER  clang++ )
 set( CMAKE_C_COMPILER_TARGET   ${target} )
 set( CMAKE_CXX_COMPILER_TARGET ${target} )
 
-#set( arch_c_flags "-march=armv8.7-a -fvectorize -ffp-model=fast -fno-finite-math-only" )
-#set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function -Wno-gnu-zero-variadic-macro-arguments" )
+set( arch_c_flags "-march=armv8.7-a -fvectorize -ffp-model=fast -fno-finite-math-only" )
+set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function -Wno-gnu-zero-variadic-macro-arguments" )
 
 set( CMAKE_C_FLAGS_INIT   "${arch_c_flags} ${warn_c_flags}" )
 set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags} ${warn_c_flags}" )

cmake/arm64-windows-cygwin.cmake

@@ -7,7 +7,7 @@ if(CMAKE_SYSTEM_NAME STREQUAL "Android")
 elseif(CMAKE_SYSTEM_NAME STREQUAL "Windows")
     set(QNN_DEFAULT_LIB_SEARCH_PATH "C:\\" CACHE STRING "customized library search path for QNN backend")
 elseif(CMAKE_SYSTEM_NAME STREQUAL "CYGWIN")
-    set(QNN_DEFAULT_LIB_SEARCH_PATH "/cygdrive/c/qairt/2.31.0.250130/" CACHE STRING "customized library search path for QNN backend")
+    set(QNN_DEFAULT_LIB_SEARCH_PATH "/cygdrive/c/qairt/2.32.0.250228/" CACHE STRING "customized library search path for QNN backend")
 else()
     message(FATAL_ERROR "QNN now only available on Android and Windows(Windows on ARM)")
 endif()

cmake/arm64-windows-llvm.cmake

@@ -14,7 +14,7 @@ GGUF_MODEL_NAME=/sdcard/qwen1_5-1_8b-chat-q4_0.gguf
 #https://www.qualcomm.com/developer/software/qualcomm-ai-engine-direct-sdk
 #https://developer.qualcomm.com/software/hexagon-dsp-sdk/tools
 QNN_SDK_URL=https://www.qualcomm.com/developer/software/qualcomm-ai-engine-direct-sdk
-QNN_SDK_PATH=/opt/qcom/aistack/qairt/2.31.0.250130/
+QNN_SDK_PATH=/opt/qcom/aistack/qairt/2.32.0.250228/
 
 #default is QNN NPU
 qnnbackend=2
@@ -97,11 +97,14 @@ function check_qnn_libs()
 {
     #reuse the cached qnn libs on Android phone
     adb shell ls ${REMOTE_PATH}/libQnnCpu.so
+    adb shell ls ${REMOTE_PATH}/libQnnGpu.so
+    adb shell ls ${REMOTE_PATH}/libQnnHtp.so
     if [ $? -eq 0 ]; then
         printf "QNN libs already exist on Android phone\n"
     else
         update_qnn_libs
     fi
+    update_qnn_cfg
 }
 
 
@@ -119,6 +122,12 @@ function update_qnn_libs()
 }
 
 
+function update_qnn_cfg()
+{
+    adb push ./scripts/ggml-qnn.cfg ${REMOTE_PATH}/
+}
+
+
 function build_ggml_qnn()
 {
     show_pwd

ggml/src/ggml-qnn/CMakeLists.txt

@@ -0,0 +1,9 @@
+[general]
+# enable/disable QNN's internal log
+print_qnn_internal_log = 0
+# 0: general approach,similar to ggml-sycl or ggml-cann
+# 1: mapping entire ggml cgraph to QNN graph
+inference_approach = 0
+
+[npu]
+npu_inference_datatype = "fp16"

ggml/src/ggml-qnn/ggml-qnn.cpp

@@ -137,7 +137,6 @@ llama_target_and_test(test-chat-template.cpp)
 # llama_target_and_test(test-opt.cpp) # SLOW
 llama_target_and_test(test-gguf.cpp)
 llama_target_and_test(test-backend-ops.cpp)
-llama_target_and_test(ggml-qnn-ut.cpp)
 
 llama_target_and_test(test-model-load-cancel.cpp  LABEL "model")
 llama_target_and_test(test-autorelease.cpp        LABEL "model")