enable fp16

Tencent · Dec 3, 2024 · bc4ec08 · bc4ec08
1 parent 37be473
commit bc4ec08
Show file tree

Hide file tree

Showing 7 changed files with 17 additions and 17 deletions.
diff --git a/.github/workflows/linux-riscv64.yml b/.github/workflows/linux-riscv64.yml
@@ -92,7 +92,7 @@ jobs:
       run: |
         export RISCV_ROOT_PATH=/data/action/osd/Xuantie-900-gcc-linux-6.6.0-glibc-x86_64-V3.0.1
         mkdir build && cd build
-        cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/c906-v301.toolchain.cmake -DCMAKE_BUILD_TYPE=release -DNCNN_OPENMP=OFF -DNCNN_THREADS=OFF -DNCNN_RUNTIME_CPU=OFF -DNCNN_XTHEADVECTOR=ON -DNCNN_RVV=OFF -DNCNN_ZFH=OFF -DNCNN_SIMPLEOCV=ON -DNCNN_BUILD_EXAMPLES=ON -DNCNN_BUILD_TESTS=ON ..
+        cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/c906-v301.toolchain.cmake -DCMAKE_BUILD_TYPE=release -DNCNN_OPENMP=OFF -DNCNN_THREADS=OFF -DNCNN_RUNTIME_CPU=OFF -DNCNN_RVV=OFF -DNCNN_XTHEADVECTOR=ON -DNCNN_ZFH=ON -DNCNN_SIMPLEOCV=ON -DNCNN_BUILD_EXAMPLES=ON -DNCNN_BUILD_TESTS=ON ..
     - name: build
       continue-on-error: true
       run: cmake --build build -j 8
@@ -114,7 +114,7 @@ jobs:
       run: |
         export RISCV_ROOT_PATH=/data/action/osd/Xuantie-900-gcc-linux-6.6.0-glibc-x86_64-V3.0.1
         mkdir build && cd build
-        cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/c910-v301.toolchain.cmake -DCMAKE_BUILD_TYPE=release -DNCNN_OPENMP=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_XTHEADVECTOR=ON -DNCNN_RVV=OFF -DNCNN_ZFH=OFF -DNCNN_SIMPLEOCV=ON -DNCNN_BUILD_EXAMPLES=ON -DNCNN_BUILD_TESTS=ON ..
+        cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/c910-v301.toolchain.cmake -DCMAKE_BUILD_TYPE=release -DNCNN_OPENMP=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_RVV=OFF -DNCNN_XTHEADVECTOR=ON -DNCNN_ZFH=ON -DNCNN_SIMPLEOCV=ON -DNCNN_BUILD_EXAMPLES=ON -DNCNN_BUILD_TESTS=ON ..
     - name: build
       continue-on-error: true
       run: cmake --build build -j 8

diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -420,9 +420,9 @@ elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(riscv)")
 
     if(CMAKE_SIZEOF_VOID_P EQUAL 8)
         check_cxx_source_compiles("#if __riscv_v == 7000\nint main() { return 0; }\n#endif\n" NCNN_COMPILER_IS_XUANTIE_GCC)
-        # if(NCNN_COMPILER_IS_XUANTIE_GCC)
-        #     message(FATAL_ERROR "Since XuanTie GCC Toolchain is not yet ready for RVV1.0, the risc-v optimized code of ncnn has been migrated to the RVV1.0 standard. Please use the old version of ncnn, which can be found at https://github.com/Tencent/ncnn/releases/download/20240820/ncnn-20240820-full-source.zip")
-        # endif()
+        if(NCNN_COMPILER_IS_XUANTIE_GCC)
+            message(WARNING "Since XuanTie GCC Toolchain is not yet ready for RVV1.0, the risc-v optimized code of ncnn has been migrated to the RVV1.0 standard. Please use the old version of ncnn, which can be found at https://github.com/Tencent/ncnn/releases/download/20240820/ncnn-20240820-full-source.zip")
+        endif()
 
         set(CMAKE_REQUIRED_FLAGS "-march=rv64gcv")
         check_cxx_source_compiles("#include <riscv_vector.h>\nint main() { vfloat32m8_t _s, _w; float _v; size_t vl; _s = __riscv_vfmacc_vf_f32m8(_s, _v, _w, vl); vfloat32m1_t _x; vfloat32m1x2_t _xx = __riscv_vcreate_v_f32m1x2(_x, _x); return 0; }" NCNN_COMPILER_SUPPORT_RISCV_V)

diff --git a/cmake/ncnn_add_layer.cmake b/cmake/ncnn_add_layer.cmake
@@ -420,7 +420,7 @@ macro(ncnn_add_layer class)
             if(NOT NCNN_RUNTIME_CPU AND NCNN_ZVFH)
                 ncnn_add_arch_opt_source(${class} zfh "-march=rv64gcv_zfh_zvfh -D__fp16=_Float16")
             elseif(NOT NCNN_RUNTIME_CPU AND NCNN_XTHEADVECTOR)
-                ncnn_add_arch_opt_source(${class} zfh "-march=rv64gc_zfh_xtheadvector -D__fp16=_Float16")
+                ncnn_add_arch_opt_source(${class} zfh "-march=rv64gc_zfh_xtheadvector -D__riscv_zvfh=1 -D__fp16=_Float16")
             else()
                 ncnn_add_arch_opt_source(${class} zfh "-march=rv64gc_zfh -D__fp16=_Float16")
             endif()

diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
@@ -680,7 +680,7 @@ if(NCNN_TARGET_ARCH STREQUAL "riscv" AND CMAKE_SIZEOF_VOID_P EQUAL 8 AND NOT C90
         set(RISCV_MARCH_FLAG "-march=rv64gc_xtheadvector")
         if(NCNN_ZFH)
             set(RISCV_MARCH_FLAG "${RISCV_MARCH_FLAG}_zfh")
-            target_compile_options(ncnn PRIVATE -D__fp16=_Float16)
+            target_compile_options(ncnn PRIVATE -D__riscv_zvfh=1 -D__fp16=_Float16)
         endif()
     endif()
     target_compile_options(ncnn PRIVATE ${RISCV_MARCH_FLAG})

diff --git a/src/layer/riscv/instancenorm_riscv_zfh.cpp b/src/layer/riscv/instancenorm_riscv_zfh.cpp
@@ -45,7 +45,7 @@ int InstanceNorm_riscv::forward_inplace_fp16s(Mat& bottom_top_blob, const Option
             // mean and var
             float sum = 0.f;
             float sqsum = 0.f;
-#if __riscv_zvfh
+#if __riscv_zvfh && !defined(C906)
             vfloat32m1_t _sum = __riscv_vfmv_s_f_f32m1(0.f, __riscv_vsetvlmax_e32m1());
             vfloat32m1_t _sqsum = __riscv_vfmv_s_f_f32m1(0.f, __riscv_vsetvlmax_e32m1());
             {
@@ -70,7 +70,7 @@ int InstanceNorm_riscv::forward_inplace_fp16s(Mat& bottom_top_blob, const Option
             }
 #endif // __riscv_zvfh
             float mean = sum / size;
-#if __riscv_zvfh
+#if __riscv_zvfh && !defined(C906)
             {
                 int n = size;
                 __fp16* ptr_sqsum = ptr;
@@ -215,7 +215,7 @@ int InstanceNorm_riscv::forward_inplace_fp16sa(Mat& bottom_top_blob, const Optio
             // mean and var
             __fp16 sum = 0.f;
             __fp16 sqsum = 0.f;
-#if __riscv_zvfh
+#if __riscv_zvfh && !defined(C906)
             vfloat16m1_t _sum = __riscv_vfmv_s_f_f16m1(0.f, __riscv_vsetvlmax_e32m1());
             vfloat16m1_t _sqsum = __riscv_vfmv_s_f_f16m1(0.f, __riscv_vsetvlmax_e32m1());
             {
@@ -240,7 +240,7 @@ int InstanceNorm_riscv::forward_inplace_fp16sa(Mat& bottom_top_blob, const Optio
             }
 #endif // __riscv_zvfh
             __fp16 mean = sum / size;
-#if __riscv_zvfh
+#if __riscv_zvfh && !defined(C906)
             {
                 int n = size;
                 __fp16* ptr_sqsum = ptr;

diff --git a/toolchains/c906-v301.toolchain.cmake b/toolchains/c906-v301.toolchain.cmake
@@ -30,12 +30,12 @@ if(NOT CMAKE_FIND_ROOT_PATH_MODE_PACKAGE)
     set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
 endif()
 
-set(CMAKE_C_FLAGS "-march=rv64gc_zfh_xtheadvector_xtheadc -mabi=lp64d -mtune=c906 -DC906=1 -static")
-set(CMAKE_CXX_FLAGS "-march=rv64gc_zfh_xtheadvector_xtheadc -mabi=lp64d -mtune=c906 -DC906=1 -static")
+set(CMAKE_C_FLAGS "-march=rv64gc_zfh_xtheadvector_xtheadc -mabi=lp64d -mtune=c906 -DC906=1 -D__riscv_zvfh=1 -static")
+set(CMAKE_CXX_FLAGS "-march=rv64gc_zfh_xtheadvector_xtheadc -mabi=lp64d -mtune=c906 -DC906=1 -D__riscv_zvfh=1 -static")
 
 # cache flags
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}" CACHE STRING "c flags")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}" CACHE STRING "c++ flags")
 
 # export RISCV_ROOT_PATH=/home/nihui/osd/Xuantie-900-gcc-linux-6.6.0-glibc-x86_64-V3.0.1
-# cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/c906-v301.toolchain.cmake -DCMAKE_BUILD_TYPE=release -DNCNN_BUILD_TESTS=ON -DNCNN_OPENMP=OFF -DNCNN_THREADS=OFF -DNCNN_RUNTIME_CPU=OFF -DNCNN_ZFH=ON -DNCNN_XTHEADVECTOR=ON -DNCNN_SIMPLEOCV=ON -DNCNN_BUILD_EXAMPLES=ON ..
+# cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/c906-v301.toolchain.cmake -DCMAKE_BUILD_TYPE=release -DNCNN_BUILD_TESTS=ON -DNCNN_OPENMP=OFF -DNCNN_THREADS=OFF -DNCNN_RUNTIME_CPU=OFF -DNCNN_RVV=OFF -DNCNN_XTHEADVECTOR=ON -DNCNN_ZFH=ON -DNCNN_SIMPLEOCV=ON -DNCNN_BUILD_EXAMPLES=ON ..
diff --git a/toolchains/c910-v301.toolchain.cmake b/toolchains/c910-v301.toolchain.cmake
@@ -30,12 +30,12 @@ if(NOT CMAKE_FIND_ROOT_PATH_MODE_PACKAGE)
     set(CMAKE_FIND_ROOT_PATH_MODE_PACKAGE ONLY)
 endif()
 
-set(CMAKE_C_FLAGS "-march=rv64gc_zfh_xtheadvector_xtheadc -mabi=lp64d -mtune=c910 -DC906=1 -static")
-set(CMAKE_CXX_FLAGS "-march=rv64gc_zfh_xtheadvector_xtheadc -mabi=lp64d -mtune=c910 -DC906=1 -static")
+set(CMAKE_C_FLAGS "-march=rv64gc_zfh_xtheadvector_xtheadc -mabi=lp64d -mtune=c910 -DC906=1 -D__riscv_zvfh=1 -static")
+set(CMAKE_CXX_FLAGS "-march=rv64gc_zfh_xtheadvector_xtheadc -mabi=lp64d -mtune=c910 -DC906=1 -D__riscv_zvfh=1 -static")
 
 # cache flags
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}" CACHE STRING "c flags")
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}" CACHE STRING "c++ flags")
 
 # export RISCV_ROOT_PATH=/home/nihui/osd/Xuantie-900-gcc-linux-6.6.0-glibc-x86_64-V3.0.1
-# cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/c910-v301.toolchain.cmake -DCMAKE_BUILD_TYPE=release -DNCNN_BUILD_TESTS=ON -DNCNN_OPENMP=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_ZFH=ON -DNCNN_XTHEADVECTOR=ON -DNCNN_SIMPLEOCV=ON -DNCNN_BUILD_EXAMPLES=ON ..
+# cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/c910-v301.toolchain.cmake -DCMAKE_BUILD_TYPE=release -DNCNN_BUILD_TESTS=ON -DNCNN_OPENMP=ON -DNCNN_RUNTIME_CPU=OFF -DNCNN_RVV=OFF -DNCNN_XTHEADVECTOR=ON -DNCNN_ZFH=ON -DNCNN_SIMPLEOCV=ON -DNCNN_BUILD_EXAMPLES=ON ..