build: support ppc64le build for make and CMake

Author: bufferoverflow

CMakeLists.txt

@@ -565,8 +565,12 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE
     endif()
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
     message(STATUS "PowerPC detected")
-    add_compile_options(-mcpu=native -mtune=native)
-    #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
+    if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
+        add_compile_options(-mcpu=powerpc64le)
+    else()
+	add_compile_options(-mcpu=native -mtune=native)
+	#TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
+    endif()
 else()
     message(STATUS "Unknown architecture")
 endif()

Makefile

@@ -337,6 +337,12 @@ ifneq ($(filter ppc64%,$(UNAME_M)),)
 	endif
 endif
 
+ifneq ($(filter ppc64le%,$(UNAME_M)),)
+	MK_CFLAGS   += -mcpu=powerpc64le
+	MK_CXXFLAGS += -mcpu=powerpc64le
+	CUDA_POWER_ARCH = 1
+endif
+
 else
 	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
 	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
@@ -387,6 +393,8 @@ else
 endif #LLAMA_CUDA_NVCC
 ifdef CUDA_DOCKER_ARCH
 	NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
+else ifdef CUDA_POWER_ARCH
+	NVCCFLAGS +=
 else
 	NVCCFLAGS += -arch=native
 endif # CUDA_DOCKER_ARCH

ggml-quants.c

@@ -39,7 +39,7 @@ inline static int32_t vaddvq_s32(int32x4_t v) {
 #ifdef __wasm_simd128__
 #include <wasm_simd128.h>
 #else
-#ifdef __POWER9_VECTOR__
+#ifdef __powerpc64__
 #include <altivec.h>
 #undef bool
 #define bool _Bool