more strict compiler rvv checks, drop rvv-071 support (#4094)

Tencent · Jul 28, 2022 · b4ba207 · b4ba207 · baibeta · Jan 5, 2023
1 parent 0666143
commit b4ba207
Show file tree

Hide file tree

Showing 41 changed files with 45 additions and 1,981 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -272,6 +272,7 @@ elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(mips)")
     set(NCNN_TARGET_ARCH mips)
 
     include(CheckCXXCompilerFlag)
+
     check_cxx_compiler_flag("-mmsa" NCNN_COMPILER_SUPPORT_MIPS_MSA)
 
     set(CMAKE_REQUIRED_FLAGS "-mloongson-mmi -I${CMAKE_CURRENT_SOURCE_DIR}/src/layer/mips")
@@ -293,8 +294,14 @@ elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "^(riscv)")
     set(NCNN_TARGET_ARCH riscv)
 
     include(CheckCXXCompilerFlag)
-    check_cxx_compiler_flag("-march=rv64gcv" NCNN_COMPILER_SUPPORT_RVV)
-    check_cxx_compiler_flag("-march=rv64gcv_zfh" NCNN_COMPILER_SUPPORT_RVV_FP16)
+
+    set(CMAKE_REQUIRED_FLAGS "-march=rv64gcv")
+    check_cxx_source_compiles("#include <riscv_vector.h>\nint main() { vfloat32m1_t _s, _w; float _v; word_type vl; _s = vfmacc_vf_f32m1(_s, _v, _w, vl); return 0; }" NCNN_COMPILER_SUPPORT_RVV)
+
+    set(CMAKE_REQUIRED_FLAGS "-march=rv64gcv_zfh")
+    check_cxx_source_compiles("#include <riscv_vector.h>\nint main() { vfloat16m1_t _s, _w; __fp16 _v; word_type vl; _s = vfmacc_vf_f16m1(_s, _v, _w, vl); return 0; }" NCNN_COMPILER_SUPPORT_RVV_FP16)
+
+    unset(CMAKE_REQUIRED_FLAGS)
 
     if(NCNN_COMPILER_SUPPORT_RVV)
         option(NCNN_RVV "optimize risc-v platform with v extension" ON)

diff --git a/cmake/ncnn_check_rvv_vfredusum.cmake b/cmake/ncnn_check_rvv_vfredusum.cmake
@@ -1,52 +1,45 @@
 include(CheckCXXSourceCompiles)
-include(CheckIncludeFile)
 
 set(TEMP_CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS})
 set(CMAKE_REQUIRED_FLAGS "-march=rv64gcv")
 
-if(NOT C906)
-    check_include_file(riscv_vector.h NCNN_RVV_COMPILER_SUPPORT_INTRINSIC)
-endif()
-
-if(NCNN_RVV_COMPILER_SUPPORT_INTRINSIC)
-    check_cxx_source_compiles("
-    #include <riscv_vector.h>
-    int main(void)
-    {
-        float in1[4] = {-1.f,0.f,+1.f,2.f};
-        float out1=0;
-        word_type vl = vsetvl_e32m8(4);
-        vfloat32m8_t _add = vle32_v_f32m8(in1,vl);
-        vfloat32m1_t _sum = vfmv_s_f_f32m1(vundefined_f32m1(),out1,vl);
-        _sum = vfredsum_vs_f32m8_f32m1(_sum, _add, _sum, vl);
-        out1 = vfmv_f_s_f32m1_f32(_sum);
-        return 0;
-    }
-    " NCNN_COMPILER_USE_VFREDSUM)
-    check_cxx_source_compiles("
-    #include <riscv_vector.h>
-    int main(void)
-    {
-        float in1[4] = {-1.f,0.f,+1.f,2.f};
-        float out1=0;
-        word_type vl = vsetvl_e32m8(4);
-        vfloat32m8_t _add = vle32_v_f32m8(in1,vl);
-        vfloat32m1_t _sum = vfmv_s_f_f32m1(vundefined_f32m1(),out1,vl);
-        _sum = vfredusum_vs_f32m8_f32m1(_sum, _add, _sum, vl);
-        out1 = vfmv_f_s_f32m1_f32(_sum);
-        return 0;
-    };
-    " NCNN_COMPILER_USE_VFREDUSUM)
+check_cxx_source_compiles("
+#include <riscv_vector.h>
+int main(void)
+{
+    float in1[4] = {-1.f,0.f,+1.f,2.f};
+    float out1=0;
+    word_type vl = vsetvl_e32m8(4);
+    vfloat32m8_t _add = vle32_v_f32m8(in1,vl);
+    vfloat32m1_t _sum = vfmv_s_f_f32m1(vundefined_f32m1(),out1,vl);
+    _sum = vfredsum_vs_f32m8_f32m1(_sum, _add, _sum, vl);
+    out1 = vfmv_f_s_f32m1_f32(_sum);
+    return 0;
+}
+" NCNN_COMPILER_USE_VFREDSUM)
+check_cxx_source_compiles("
+#include <riscv_vector.h>
+int main(void)
+{
+    float in1[4] = {-1.f,0.f,+1.f,2.f};
+    float out1=0;
+    word_type vl = vsetvl_e32m8(4);
+    vfloat32m8_t _add = vle32_v_f32m8(in1,vl);
+    vfloat32m1_t _sum = vfmv_s_f_f32m1(vundefined_f32m1(),out1,vl);
+    _sum = vfredusum_vs_f32m8_f32m1(_sum, _add, _sum, vl);
+    out1 = vfmv_f_s_f32m1_f32(_sum);
+    return 0;
+};
+" NCNN_COMPILER_USE_VFREDUSUM)
 
-    if(NCNN_COMPILER_USE_VFREDSUM AND NOT NCNN_COMPILER_USE_VFREDUSUM)
-        message(WARNING "The compiler uses vfredsum. Upgrading your toolchain is strongly recommended.")
-        foreach(LMUL 1 2 4 8)
-            add_definitions(-Dvfredusum_vs_f32m${LMUL}_f32m1=vfredsum_vs_f32m${LMUL}_f32m1)
-            if(NCNN_COMPILER_SUPPORT_RVV_FP16)
-                add_definitions(-Dvfredusum_vs_f16m${LMUL}_f16m1=vfredsum_vs_f16m${LMUL}_f16m1)
-            endif()
-        endforeach()
-    endif()
+if(NCNN_COMPILER_USE_VFREDSUM AND NOT NCNN_COMPILER_USE_VFREDUSUM)
+    message(WARNING "The compiler uses vfredsum. Upgrading your toolchain is strongly recommended.")
+    foreach(LMUL 1 2 4 8)
+        add_definitions(-Dvfredusum_vs_f32m${LMUL}_f32m1=vfredsum_vs_f32m${LMUL}_f32m1)
+        if(NCNN_COMPILER_SUPPORT_RVV_FP16)
+            add_definitions(-Dvfredusum_vs_f16m${LMUL}_f16m1=vfredsum_vs_f16m${LMUL}_f16m1)
+        endif()
+    endforeach()
 endif()
 
 set(CMAKE_REQUIRED_FLAGS ${TEMP_CMAKE_REQUIRED_FLAGS})

diff --git a/src/layer/riscv/absval_riscv.cpp b/src/layer/riscv/absval_riscv.cpp
@@ -15,11 +15,7 @@
 #include "absval_riscv.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 
 static inline vfloat32m8_t vfabs_v_f32m8_absval(vfloat32m8_t op1, size_t vl)
 {

diff --git a/src/layer/riscv/binaryop_riscv.cpp b/src/layer/riscv/binaryop_riscv.cpp
@@ -20,11 +20,7 @@
 #include <math.h>
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #include "rvv_mathfun.h"
 #include "rvv_mathfun_fp16s.h"
 #endif // __riscv_vector

diff --git a/src/layer/riscv/cast_riscv.cpp b/src/layer/riscv/cast_riscv.cpp
@@ -15,11 +15,7 @@
 #include "cast_riscv.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 namespace ncnn {

diff --git a/src/layer/riscv/clip_riscv.cpp b/src/layer/riscv/clip_riscv.cpp
@@ -15,11 +15,7 @@
 #include "clip_riscv.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #include "rvv_mathfun.h"
 #include "rvv_mathfun_fp16s.h"
 #endif // __riscv_vector

diff --git a/src/layer/riscv/concat_riscv.cpp b/src/layer/riscv/concat_riscv.cpp
@@ -15,11 +15,7 @@
 #include "concat_riscv.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 #include "riscv_usability.h"

diff --git a/src/layer/riscv/convolution1d_riscv.cpp b/src/layer/riscv/convolution1d_riscv.cpp
@@ -15,11 +15,7 @@
 #include "convolution1d_riscv.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 #include "riscv_activation.h"

diff --git a/src/layer/riscv/convolution_riscv.cpp b/src/layer/riscv/convolution_riscv.cpp
@@ -19,11 +19,7 @@
 #include "layer_type.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 #include "riscv_activation.h"

diff --git a/src/layer/riscv/convolutiondepthwise_riscv.cpp b/src/layer/riscv/convolutiondepthwise_riscv.cpp
@@ -18,11 +18,7 @@
 #include "layer_type.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 #include "riscv_activation.h"

diff --git a/src/layer/riscv/crop_riscv.cpp b/src/layer/riscv/crop_riscv.cpp
@@ -15,11 +15,7 @@
 #include "crop_riscv.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 #include "riscv_usability.h"

diff --git a/src/layer/riscv/deconvolution_riscv.cpp b/src/layer/riscv/deconvolution_riscv.cpp
@@ -17,11 +17,7 @@
 #include "layer_type.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 #include "riscv_activation.h"

diff --git a/src/layer/riscv/deconvolutiondepthwise_riscv.cpp b/src/layer/riscv/deconvolutiondepthwise_riscv.cpp
@@ -17,11 +17,7 @@
 #include "layer_type.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 #include "riscv_activation.h"

diff --git a/src/layer/riscv/dropout_riscv.cpp b/src/layer/riscv/dropout_riscv.cpp
@@ -18,11 +18,7 @@
 #include "dropout_riscv.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 namespace ncnn {

diff --git a/src/layer/riscv/flatten_riscv.cpp b/src/layer/riscv/flatten_riscv.cpp
@@ -15,11 +15,7 @@
 #include "flatten_riscv.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 #include "riscv_usability.h"

diff --git a/src/layer/riscv/gelu_riscv.cpp b/src/layer/riscv/gelu_riscv.cpp
@@ -15,11 +15,7 @@
 #include "gelu_riscv.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #include "rvv_mathfun.h"
 #endif // __riscv_vector
 

diff --git a/src/layer/riscv/gru_riscv.cpp b/src/layer/riscv/gru_riscv.cpp
@@ -15,11 +15,7 @@
 #include "gru_riscv.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 namespace ncnn {

diff --git a/src/layer/riscv/hardsigmoid_riscv.cpp b/src/layer/riscv/hardsigmoid_riscv.cpp
@@ -18,11 +18,7 @@
 #include "hardsigmoid_riscv.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 namespace ncnn {

diff --git a/src/layer/riscv/hardswish_riscv.cpp b/src/layer/riscv/hardswish_riscv.cpp
@@ -18,11 +18,7 @@
 #include "hardswish_riscv.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 namespace ncnn {

diff --git a/src/layer/riscv/innerproduct_riscv.cpp b/src/layer/riscv/innerproduct_riscv.cpp
@@ -17,11 +17,7 @@
 #include "layer_type.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 #include "riscv_activation.h"

diff --git a/src/layer/riscv/interp_riscv.cpp b/src/layer/riscv/interp_riscv.cpp
@@ -17,11 +17,7 @@
 #include <math.h>
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #include "riscv_usability.h"
 #endif // __riscv_vector
 

diff --git a/src/layer/riscv/mish_riscv.cpp b/src/layer/riscv/mish_riscv.cpp
@@ -15,11 +15,7 @@
 #include "mish_riscv.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #include "rvv_mathfun.h"
 #include "rvv_mathfun_fp16s.h"
 #endif // __riscv_vector

diff --git a/src/layer/riscv/packing_riscv.cpp b/src/layer/riscv/packing_riscv.cpp
@@ -15,11 +15,7 @@
 #include "packing_riscv.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 namespace ncnn {

diff --git a/src/layer/riscv/padding_riscv.cpp b/src/layer/riscv/padding_riscv.cpp
@@ -15,11 +15,7 @@
 #include "padding_riscv.h"
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 #include "riscv_usability.h"

diff --git a/src/layer/riscv/pooling_riscv.cpp b/src/layer/riscv/pooling_riscv.cpp
@@ -17,11 +17,7 @@
 #include <float.h>
 
 #if __riscv_vector
-#ifdef RVV_SPEC_0_7
-#include "riscv_v_071_fix.h"
-#else
 #include <riscv_vector.h>
-#endif
 #endif // __riscv_vector
 
 #include "riscv_usability.h"