|
1 | 1 | include(CheckCXXSourceCompiles)
|
2 |
| -include(CheckIncludeFile) |
3 | 2 |
|
4 | 3 | set(TEMP_CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS})
|
5 | 4 | set(CMAKE_REQUIRED_FLAGS "-march=rv64gcv")
|
6 | 5 |
|
7 |
| -if(NOT C906) |
8 |
| - check_include_file(riscv_vector.h NCNN_RVV_COMPILER_SUPPORT_INTRINSIC) |
9 |
| -endif() |
10 |
| - |
11 |
| -if(NCNN_RVV_COMPILER_SUPPORT_INTRINSIC) |
12 |
| - check_cxx_source_compiles(" |
13 |
| - #include <riscv_vector.h> |
14 |
| - int main(void) |
15 |
| - { |
16 |
| - float in1[4] = {-1.f,0.f,+1.f,2.f}; |
17 |
| - float out1=0; |
18 |
| - word_type vl = vsetvl_e32m8(4); |
19 |
| - vfloat32m8_t _add = vle32_v_f32m8(in1,vl); |
20 |
| - vfloat32m1_t _sum = vfmv_s_f_f32m1(vundefined_f32m1(),out1,vl); |
21 |
| - _sum = vfredsum_vs_f32m8_f32m1(_sum, _add, _sum, vl); |
22 |
| - out1 = vfmv_f_s_f32m1_f32(_sum); |
23 |
| - return 0; |
24 |
| - } |
25 |
| - " NCNN_COMPILER_USE_VFREDSUM) |
26 |
| - check_cxx_source_compiles(" |
27 |
| - #include <riscv_vector.h> |
28 |
| - int main(void) |
29 |
| - { |
30 |
| - float in1[4] = {-1.f,0.f,+1.f,2.f}; |
31 |
| - float out1=0; |
32 |
| - word_type vl = vsetvl_e32m8(4); |
33 |
| - vfloat32m8_t _add = vle32_v_f32m8(in1,vl); |
34 |
| - vfloat32m1_t _sum = vfmv_s_f_f32m1(vundefined_f32m1(),out1,vl); |
35 |
| - _sum = vfredusum_vs_f32m8_f32m1(_sum, _add, _sum, vl); |
36 |
| - out1 = vfmv_f_s_f32m1_f32(_sum); |
37 |
| - return 0; |
38 |
| - }; |
39 |
| - " NCNN_COMPILER_USE_VFREDUSUM) |
| 6 | +check_cxx_source_compiles(" |
| 7 | +#include <riscv_vector.h> |
| 8 | +int main(void) |
| 9 | +{ |
| 10 | + float in1[4] = {-1.f,0.f,+1.f,2.f}; |
| 11 | + float out1=0; |
| 12 | + word_type vl = vsetvl_e32m8(4); |
| 13 | + vfloat32m8_t _add = vle32_v_f32m8(in1,vl); |
| 14 | + vfloat32m1_t _sum = vfmv_s_f_f32m1(vundefined_f32m1(),out1,vl); |
| 15 | + _sum = vfredsum_vs_f32m8_f32m1(_sum, _add, _sum, vl); |
| 16 | + out1 = vfmv_f_s_f32m1_f32(_sum); |
| 17 | + return 0; |
| 18 | +} |
| 19 | +" NCNN_COMPILER_USE_VFREDSUM) |
| 20 | +check_cxx_source_compiles(" |
| 21 | +#include <riscv_vector.h> |
| 22 | +int main(void) |
| 23 | +{ |
| 24 | + float in1[4] = {-1.f,0.f,+1.f,2.f}; |
| 25 | + float out1=0; |
| 26 | + word_type vl = vsetvl_e32m8(4); |
| 27 | + vfloat32m8_t _add = vle32_v_f32m8(in1,vl); |
| 28 | + vfloat32m1_t _sum = vfmv_s_f_f32m1(vundefined_f32m1(),out1,vl); |
| 29 | + _sum = vfredusum_vs_f32m8_f32m1(_sum, _add, _sum, vl); |
| 30 | + out1 = vfmv_f_s_f32m1_f32(_sum); |
| 31 | + return 0; |
| 32 | +}; |
| 33 | +" NCNN_COMPILER_USE_VFREDUSUM) |
40 | 34 |
|
41 |
| - if(NCNN_COMPILER_USE_VFREDSUM AND NOT NCNN_COMPILER_USE_VFREDUSUM) |
42 |
| - message(WARNING "The compiler uses vfredsum. Upgrading your toolchain is strongly recommended.") |
43 |
| - foreach(LMUL 1 2 4 8) |
44 |
| - add_definitions(-Dvfredusum_vs_f32m${LMUL}_f32m1=vfredsum_vs_f32m${LMUL}_f32m1) |
45 |
| - if(NCNN_COMPILER_SUPPORT_RVV_FP16) |
46 |
| - add_definitions(-Dvfredusum_vs_f16m${LMUL}_f16m1=vfredsum_vs_f16m${LMUL}_f16m1) |
47 |
| - endif() |
48 |
| - endforeach() |
49 |
| - endif() |
| 35 | +if(NCNN_COMPILER_USE_VFREDSUM AND NOT NCNN_COMPILER_USE_VFREDUSUM) |
| 36 | + message(WARNING "The compiler uses vfredsum. Upgrading your toolchain is strongly recommended.") |
| 37 | + foreach(LMUL 1 2 4 8) |
| 38 | + add_definitions(-Dvfredusum_vs_f32m${LMUL}_f32m1=vfredsum_vs_f32m${LMUL}_f32m1) |
| 39 | + if(NCNN_COMPILER_SUPPORT_RVV_FP16) |
| 40 | + add_definitions(-Dvfredusum_vs_f16m${LMUL}_f16m1=vfredsum_vs_f16m${LMUL}_f16m1) |
| 41 | + endif() |
| 42 | + endforeach() |
50 | 43 | endif()
|
51 | 44 |
|
52 | 45 | set(CMAKE_REQUIRED_FLAGS ${TEMP_CMAKE_REQUIRED_FLAGS})
|
|
0 commit comments