ggml-dsp: refine ggml-dsp and make ggml-dsp more clear

Author: zhouwg

ggml/src/ggml-hexagon/CMakeLists.txt

@@ -89,7 +89,7 @@ set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DGGML_USE_HEXAGON ${DEBUG_FLAG} ${OPT_F
 set(CMAKE_C_FLAGS_RELEASE "${CMAKE_C_FLAGS_RELEASE} -DGGML_USE_HEXAGON ${DEBUG_FLAG} ${OPT_FLAG}")
 set(CMAKE_CXX_FLAGS_RELEASE "${CMAKE_CXX_FLAGS_RELEASE} -DGGML_USE_HEXAGON ${DEBUG_FLAG} ${OPT_FLAG}")
 
-file(GLOB HEXAGON_SOURCES "${CMAKE_CURRENT_LIST_DIR}/*.cpp" "${CMAKE_CURRENT_LIST_DIR}/stub.c")
+file(GLOB HEXAGON_SOURCES "${CMAKE_CURRENT_LIST_DIR}/*.cpp" "${CMAKE_CURRENT_LIST_DIR}/kernels/stub.c")
 ggml_add_backend_library(ggml-hexagon ${HEXAGON_SOURCES})
 
 target_include_directories(ggml-hexagon PRIVATE ${QNN_SDK_PATH}/include/QNN ${HEXAGON_SDK_PATH} ${CMAKE_CURRENT_LIST_DIR})
@@ -98,7 +98,7 @@ target_link_libraries(ggml-hexagon PRIVATE ${QNN_LINK_LIBRARIES})
 string(REGEX REPLACE "/$" "" QNN_DEFAULT_LIB_SEARCH_PATH "${QNN_DEFAULT_LIB_SEARCH_PATH}")
 target_compile_definitions(ggml-hexagon PRIVATE QNN_DEFAULT_LIB_SEARCH_PATH="${QNN_DEFAULT_LIB_SEARCH_PATH}/")
 
-#cross compiling hexagon kernels which running on cDSP side
+#cross compiling source codes of hexagon kernels which running on cDSP side
 function(ggml_hexagon_build_kernel KNAME)
     message(STATUS "ggml_hexagon: build hexagon-kernel ${KNAME}")
 
@@ -109,7 +109,8 @@ function(ggml_hexagon_build_kernel KNAME)
         COMMAND echo "${CMAKE_CURRENT_LIST_DIR}/kernels"
         COMMAND make -C ${CMAKE_CURRENT_LIST_DIR}/kernels/ clean
         COMMAND make -C ${CMAKE_CURRENT_LIST_DIR}/kernels/ HEXAGON_SDK_PATH=${HEXAGON_SDK_PATH} HTP_ARCH_VERSION=${HTP_ARCH_VERSION} DEBUG_FLAG=${DEBUG_FLAG}
-        COMMAND ls -l  ../../../bin/libggmlop_skel.so
+        COMMAND echo "current working path:`pwd`\n"
+        COMMAND ls -l  ../../../bin/libggmlop-skel.so
         COMMENT "build hexagon-kernel"
     )
 endfunction()

ggml/src/ggml-hexagon/kernels/Makefile

@@ -7,7 +7,7 @@ HEXAGON_COMPUTE=compute${HTP_ARCH_VERSION}
 HEXAGON_CC=${HEXAGON_SDK_PATH}/tools/HEXAGON_Tools/8.8.06/Tools/bin/hexagon-clang
 HEXAGON_CXX=${HEXAGON_SDK_PATH}/tools/HEXAGON_Tools/8.8.06/Tools/bin/hexagon-clang
 
-TARGET=libggmlop_skel.so
+TARGET=libggmlop-skel.so
 
 $(info HEXAGON_SDK_PATH:${HEXAGON_SDK_PATH})
 $(info HTP_ARCH_VERSION:${HTP_ARCH_VERSION})
@@ -20,19 +20,20 @@ CFLAGS=-m${HTP_ARCH_VERSION} -c -Ofast -Wall -Wstrict-prototypes -fno-zero-initi
 
 LDFLAGS=-m${HTP_ARCH_VERSION} -Wl,--defsym=ISDB_TRUSTED_FLAG=2 -Wl,--defsym=ISDB_SECURE_FLAG=2 -Wl,--no-threads -fpic -shared -Wl,-Bsymbolic -Wl,--wrap=malloc -Wl,--wrap=calloc -Wl,--wrap=free -Wl,--wrap=realloc -Wl,--wrap=memalign -lc -Wl,-soname=${TARGET}
 
-SRCS = $(wildcard *.c)
+#SRCS = $(wildcard *.c)
+SRCS = ggml-dsp.c skel.c add.c  mulmat.c
 OBJS = $(patsubst %.c, %.o, $(SRCS))
 
 ALL:$(OBJS)
 		${HEXAGON_CC} ${LDFLAGS} -o ${TARGET} -Wl,--start-group ${OBJS} -Wl,--end-group
 		@ls -l ${TARGET}
 		/bin/cp -fv ${TARGET} ../../../../out/android/bin/
-		/bin/cp -fv ${TARGET} ../../../../out/android/bin/libggmlop_skel${HTP_ARCH_VERSION}.so
+		/bin/cp -fv ${TARGET} ../../../../out/android/bin/libggmlop-skel${HTP_ARCH_VERSION}.so
 		/bin/rm -f *.so
 
 %.o:%.c
 		@echo "${HEXAGON_CC} ${CFLAGS} ${DEBUG_FLAG} -D__FILENAME__=\"$<\" -o $@ -c $<"
-		${HEXAGON_CC} ${CFLAGS} -D__FILENAME__=\"$<\" -o $@ -c $<
+		${HEXAGON_CC} ${CFLAGS} ${DEBUG_FLAG} -D__FILENAME__=\"$<\" -o $@ -c $<
 		@echo "\n"
 
 clean:

ggml/src/ggml-hexagon/kernels/ggml-dsp.h

@@ -67,7 +67,7 @@ extern "C" {
 #define GGML_FP16_TO_FP32(x)            ggml_lookup_fp16_to_fp32(x)
 
 //NPU performance will be slower when enable GGMLHEXAGON_DEBUG
-#if 1//def NDEBUG
+#ifdef NDEBUG
 #define GGMLHEXAGON_DEBUG                                   0
 #else
 #define GGMLHEXAGON_DEBUG                                   1

ggml/src/ggml-hexagon/kernels/skel.c

@@ -290,7 +290,7 @@ extern int adsp_mmap_fd_getinfo(int, uint32_t *);
 extern "C" {
 #endif
 _ATTRIBUTE_VISIBILITY uint32_t ggmlop_skel_handle_invoke_qaic_version = 10048;
-_ATTRIBUTE_VISIBILITY char ggmlop_skel_handle_invoke_uri[77+1]="file:///libggmlop_skel.so?ggmlop_skel_handle_invoke&_modver=1.0&_idlver=0.0.1";
+_ATTRIBUTE_VISIBILITY char ggmlop_skel_handle_invoke_uri[77+1]="file:///libggmlop-skel.so?ggmlop_skel_handle_invoke&_modver=1.0&_idlver=0.0.1";
 static __inline int _skel_pack(_ATTRIBUTE_UNUSED remote_arg* _praROutPost, _ATTRIBUTE_UNUSED remote_arg* _ppraROutPost[1], _ATTRIBUTE_UNUSED void* _primROut, _ATTRIBUTE_UNUSED uint32_t _rout0[1], _ATTRIBUTE_UNUSED uint32_t _rout1[4], _ATTRIBUTE_UNUSED uint32_t _rout2[4], _ATTRIBUTE_UNUSED uint32_t _rout3[1], _ATTRIBUTE_UNUSED uint32_t _rout4[16], _ATTRIBUTE_UNUSED uint32_t _rout5[1], _ATTRIBUTE_UNUSED char* _rout6[1], _ATTRIBUTE_UNUSED uint32_t _rout6Len[1]) {
    int _nErr = 0;
    remote_arg* _praROutPostStart = _praROutPost;

ggml/src/ggml-hexagon/kernels/skel.h

@@ -279,7 +279,7 @@ __QAIC_HEADER_EXPORT int __QAIC_HEADER(ggmlop_dsp_softmax)(remote_handle64 _h, c
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(ggmlop_dsp_rmsnorm)(remote_handle64 _h, const dsptensor* src0, const dsptensor* src1, dsptensor* dst) __QAIC_HEADER_ATTRIBUTE;
 __QAIC_HEADER_EXPORT int __QAIC_HEADER(ggmlop_dsp_pool2d)(remote_handle64 _h, const dsptensor* src0, const dsptensor* src1, dsptensor* dst) __QAIC_HEADER_ATTRIBUTE;
 #ifndef ggmlop_URI
-#define ggmlop_URI "file:///libggmlop_skel.so?ggmlop_skel_handle_invoke&_modver=1.0&_idlver=0.0.1"
+#define ggmlop_URI "file:///libggmlop-skel.so?ggmlop_skel_handle_invoke&_modver=1.0&_idlver=0.0.1"
 #endif /*ggmlop_URI*/
 #ifdef __cplusplus
 }