|
3 | 3 |
|
4 | 4 | include(CheckCXXCompilerFlag) |
5 | 5 |
|
6 | | -if(CMAKE_HOST_SYSTEM_PROCESSOR MATCHES "(x86_64)|(AMD64|amd64)|(^i.86$)") |
7 | | - # build SSE/AVX* code only on x64 processors. |
8 | | - # Check that the compiler supports instructions flag. |
9 | | - # This will add the relevant flag both the the space selector and the optimization. |
10 | | - CHECK_CXX_COMPILER_FLAG(-mavx512vl CXX_AVX512VL) |
11 | | - CHECK_CXX_COMPILER_FLAG(-mavx512bf16 CXX_AVX512BF16) |
12 | | - CHECK_CXX_COMPILER_FLAG(-mavx512bw CXX_AVX512BW) |
13 | | - CHECK_CXX_COMPILER_FLAG(-mavx512vbmi2 CXX_AVX512VBMI2) |
14 | | - CHECK_CXX_COMPILER_FLAG(-mavx512fp16 CXX_AVX512FP16) |
15 | | - CHECK_CXX_COMPILER_FLAG(-mavx512f CXX_AVX512F) |
16 | | - CHECK_CXX_COMPILER_FLAG(-mavx512vnni CXX_AVX512VNNI) |
17 | | - CHECK_CXX_COMPILER_FLAG(-mavx2 CXX_AVX2) |
18 | | - CHECK_CXX_COMPILER_FLAG(-mavx CXX_AVX) |
19 | | - CHECK_CXX_COMPILER_FLAG(-mf16c CXX_F16C) |
20 | | - CHECK_CXX_COMPILER_FLAG(-mfma CXX_FMA) |
21 | | - CHECK_CXX_COMPILER_FLAG(-msse3 CXX_SSE3) |
22 | | - CHECK_CXX_COMPILER_FLAG(-msse CXX_SSE) |
| 6 | +# build SSE/AVX* code only on x64 processors. |
| 7 | +# Check that the compiler supports instructions flag. |
| 8 | +# This will add the relevant flag both the the space selector and the optimization. |
| 9 | +CHECK_CXX_COMPILER_FLAG(-mavx512vl CXX_AVX512VL) |
| 10 | +CHECK_CXX_COMPILER_FLAG(-mavx512bf16 CXX_AVX512BF16) |
| 11 | +CHECK_CXX_COMPILER_FLAG(-mavx512bw CXX_AVX512BW) |
| 12 | +CHECK_CXX_COMPILER_FLAG(-mavx512vbmi2 CXX_AVX512VBMI2) |
| 13 | +CHECK_CXX_COMPILER_FLAG(-mavx512fp16 CXX_AVX512FP16) |
| 14 | +CHECK_CXX_COMPILER_FLAG(-mavx512f CXX_AVX512F) |
| 15 | +CHECK_CXX_COMPILER_FLAG(-mavx512vnni CXX_AVX512VNNI) |
| 16 | +CHECK_CXX_COMPILER_FLAG(-mavx2 CXX_AVX2) |
| 17 | +CHECK_CXX_COMPILER_FLAG(-mavx CXX_AVX) |
| 18 | +CHECK_CXX_COMPILER_FLAG(-mf16c CXX_F16C) |
| 19 | +CHECK_CXX_COMPILER_FLAG(-mfma CXX_FMA) |
| 20 | +CHECK_CXX_COMPILER_FLAG(-msse3 CXX_SSE3) |
| 21 | +CHECK_CXX_COMPILER_FLAG(-msse CXX_SSE) |
23 | 22 |
|
24 | | - # Turn off AVX512BF16 on Ubuntu 18.04 as it is not supported by its binutils assembler version. |
25 | | - if(${CMAKE_SYSTEM_NAME} STREQUAL "Linux") |
26 | | - execute_process(COMMAND lsb_release -rs |
27 | | - OUTPUT_VARIABLE UBUNTU_VERSION |
28 | | - OUTPUT_STRIP_TRAILING_WHITESPACE) |
| 23 | +# Turn off AVX512BF16 on Ubuntu 18.04 as it is not supported by its binutils assembler version. |
| 24 | +if(${CMAKE_SYSTEM_NAME} STREQUAL "Linux") |
| 25 | + execute_process(COMMAND lsb_release -rs |
| 26 | + OUTPUT_VARIABLE UBUNTU_VERSION |
| 27 | + OUTPUT_STRIP_TRAILING_WHITESPACE) |
29 | 28 |
|
30 | | - if("${UBUNTU_VERSION}" STREQUAL "18.04") |
31 | | - message(STATUS "Compiling on Ubuntu 18.04, turning off CXX_AVX512BF16 flag.") |
32 | | - set(CXX_AVX512BF16 FALSE) |
33 | | - endif() |
| 29 | + if("${UBUNTU_VERSION}" STREQUAL "18.04") |
| 30 | + message(STATUS "Compiling on Ubuntu 18.04, turning off CXX_AVX512BF16 flag.") |
| 31 | + set(CXX_AVX512BF16 FALSE) |
34 | 32 | endif() |
| 33 | +endif() |
35 | 34 |
|
36 | | - if(CXX_AVX512VL AND CXX_AVX512BF16) |
37 | | - add_compile_definitions(OPT_AVX512_BF16_VL) |
38 | | - endif() |
| 35 | +if(CXX_AVX512VL AND CXX_AVX512BF16) |
| 36 | + add_compile_definitions(OPT_AVX512_BF16_VL) |
| 37 | +endif() |
39 | 38 |
|
40 | | - if(CXX_AVX512VL AND CXX_AVX512FP16) |
41 | | - add_compile_definitions(OPT_AVX512_FP16_VL) |
42 | | - endif() |
| 39 | +if(CXX_AVX512VL AND CXX_AVX512FP16) |
| 40 | + add_compile_definitions(OPT_AVX512_FP16_VL) |
| 41 | +endif() |
43 | 42 |
|
44 | | - if(CXX_AVX512F) |
45 | | - add_compile_definitions(OPT_AVX512F) |
46 | | - endif() |
| 43 | +if(CXX_AVX512F) |
| 44 | + add_compile_definitions(OPT_AVX512F) |
| 45 | +endif() |
47 | 46 |
|
48 | | - if(CXX_AVX512BW AND CXX_AVX512VBMI2) |
49 | | - add_compile_definitions(OPT_AVX512_BW_VBMI2) |
50 | | - endif() |
| 47 | +if(CXX_AVX512BW AND CXX_AVX512VBMI2) |
| 48 | + add_compile_definitions(OPT_AVX512_BW_VBMI2) |
| 49 | +endif() |
51 | 50 |
|
52 | | - if(CXX_AVX512F AND CXX_AVX512BW AND CXX_AVX512VL AND CXX_AVX512VNNI) |
53 | | - add_compile_definitions(OPT_AVX512_F_BW_VL_VNNI) |
54 | | - endif() |
| 51 | +if(CXX_AVX512F AND CXX_AVX512BW AND CXX_AVX512VL AND CXX_AVX512VNNI) |
| 52 | + add_compile_definitions(OPT_AVX512_F_BW_VL_VNNI) |
| 53 | +endif() |
55 | 54 |
|
56 | | - if(CXX_F16C AND CXX_FMA AND CXX_AVX) |
57 | | - add_compile_definitions(OPT_F16C) |
58 | | - endif() |
| 55 | +if(CXX_F16C AND CXX_FMA AND CXX_AVX) |
| 56 | + add_compile_definitions(OPT_F16C) |
| 57 | +endif() |
59 | 58 |
|
60 | | - if(CXX_AVX2) |
61 | | - add_compile_definitions(OPT_AVX2) |
62 | | - endif() |
| 59 | +if(CXX_AVX2) |
| 60 | + add_compile_definitions(OPT_AVX2) |
| 61 | +endif() |
63 | 62 |
|
64 | | - if(CXX_AVX) |
65 | | - add_compile_definitions(OPT_AVX) |
66 | | - endif() |
| 63 | +if(CXX_AVX) |
| 64 | + add_compile_definitions(OPT_AVX) |
| 65 | +endif() |
67 | 66 |
|
68 | | - if(CXX_SSE3) |
69 | | - add_compile_definitions(OPT_SSE3) |
70 | | - endif() |
| 67 | +if(CXX_SSE3) |
| 68 | + add_compile_definitions(OPT_SSE3) |
| 69 | +endif() |
71 | 70 |
|
72 | | - if(CXX_SSE) |
73 | | - add_compile_definitions(OPT_SSE) |
74 | | - endif() |
| 71 | +if(CXX_SSE) |
| 72 | + add_compile_definitions(OPT_SSE) |
75 | 73 | endif() |
0 commit comments