use _Float16 instead of __fp16

Change-Id: I039fb02bb95270e641bc4442204e658735859d43

Author: alex-spacemit

cmake/riscv64-spacemit-linux-gnu-gcc.cmake

@@ -27,4 +27,3 @@ set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
 set(CMAKE_C_FLAGS "-march=rv64gcv_zfh_zba_zicbop -mabi=lp64d ${CMAKE_C_FLAGS}")
 set(CMAKE_CXX_FLAGS "-march=rv64gcv_zfh_zba_zicbop -mabi=lp64d ${CXX_FLAGS}")
 set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} -latomic")
-add_definitions(-D__fp16=_Float16)

ggml/src/ggml-cpu/spacemit/ime1_kernels.cpp

@@ -1475,7 +1475,7 @@ void SQ4BitGemmM4Kernel_CompInt8_ScaleFp16_Impl(size_t            BlkLen,
             std::byte * QuantBDataPtr = (std::byte *) QuantBData +           //
                                         n * BlockCountK * BlkLen / 2 +       // b data
                                         n * BlockCountK * sizeof(uint8_t) +  // zp
-                                        n * BlockCountK * sizeof(__fp16);    // scale
+                                        n * BlockCountK * sizeof(_Float16);    // scale
             float * CPtr = C + n;
             if (NBLKS < 16) {
                 CPtr = tmp;
@@ -1656,7 +1656,7 @@ void SQ4BitGemmM4Kernel_CompInt8_ScaleFp16_Impl(size_t            BlkLen,
             size_t      NBLKS         = (CountN - n) > 16 ? 16 : CountN - n;
             std::byte * QuantBDataPtr = (std::byte *) QuantBData +         //
                                         n * BlockCountK * BlkLen / 2 +     // b data
-                                        n * BlockCountK * sizeof(__fp16);  // scale
+                                        n * BlockCountK * sizeof(_Float16);  // scale
             float * CPtr = C + n;
             if (NBLKS < 16) {
                 CPtr = tmp;
@@ -2216,7 +2216,7 @@ void SQ4BitGemmM1Kernel_CompInt8_ScaleFp16_Impl(size_t            BlkLen,
             std::byte * QuantBDataPtr = (std::byte *) QuantBData +           //
                                         n * BlockCountK * BlkLen / 2 +       // b data
                                         n * BlockCountK * sizeof(uint8_t) +  // zp
-                                        n * BlockCountK * sizeof(__fp16);    // scale
+                                        n * BlockCountK * sizeof(_Float16);    // scale
             float * CPtr = C + n;
             size_t  cnt  = BlockCountK;
             if (Bias != nullptr) {
@@ -2455,7 +2455,7 @@ void SQ4BitGemmM1Kernel_CompInt8_ScaleFp16_Impl(size_t            BlkLen,
             size_t      nblks         = (CountN - n) > 16 ? 16 : CountN - n;
             std::byte * QuantBDataPtr = (std::byte *) QuantBData +         //
                                         n * BlockCountK * BlkLen / 2 +     // b data
-                                        n * BlockCountK * sizeof(__fp16);  // scale
+                                        n * BlockCountK * sizeof(_Float16);  // scale
             float * CPtr = C + n;
             size_t  cnt  = BlockCountK;
             if (Bias != nullptr) {