style
This commit is contained in:
parent
14aded64cf
commit
447e50a96f
2 changed files with 64 additions and 65 deletions
|
@ -10,7 +10,7 @@ endif()
|
||||||
|
|
||||||
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
|
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
|
||||||
|
|
||||||
if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
|
if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
|
||||||
set(LLAMA_STANDALONE ON)
|
set(LLAMA_STANDALONE ON)
|
||||||
|
|
||||||
# configure project version
|
# configure project version
|
||||||
|
|
|
@ -1,85 +1,84 @@
|
||||||
include(CheckCSourceRuns)
|
include(CheckCSourceRuns)
|
||||||
|
|
||||||
set(AVX_CODE "
|
set(AVX_CODE "
|
||||||
#include <immintrin.h>
|
#include <immintrin.h>
|
||||||
int main()
|
int main()
|
||||||
{
|
{
|
||||||
__m256 a;
|
__m256 a;
|
||||||
a = _mm256_set1_ps(0);
|
a = _mm256_set1_ps(0);
|
||||||
return 0;
|
return 0;
|
||||||
}
|
}
|
||||||
")
|
")
|
||||||
|
|
||||||
set(AVX512_CODE "
|
set(AVX512_CODE "
|
||||||
#include <immintrin.h>
|
#include <immintrin.h>
|
||||||
int main()
|
int main()
|
||||||
{
|
{
|
||||||
__m512i a = _mm512_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
|
__m512i a = _mm512_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
|
||||||
0, 0, 0, 0, 0, 0, 0, 0,
|
0, 0, 0, 0, 0, 0, 0, 0,
|
||||||
0, 0, 0, 0, 0, 0, 0, 0,
|
0, 0, 0, 0, 0, 0, 0, 0,
|
||||||
0, 0, 0, 0, 0, 0, 0, 0,
|
0, 0, 0, 0, 0, 0, 0, 0,
|
||||||
0, 0, 0, 0, 0, 0, 0, 0,
|
0, 0, 0, 0, 0, 0, 0, 0,
|
||||||
0, 0, 0, 0, 0, 0, 0, 0,
|
0, 0, 0, 0, 0, 0, 0, 0,
|
||||||
0, 0, 0, 0, 0, 0, 0, 0,
|
0, 0, 0, 0, 0, 0, 0, 0,
|
||||||
0, 0, 0, 0, 0, 0, 0, 0);
|
0, 0, 0, 0, 0, 0, 0, 0);
|
||||||
__m512i b = a;
|
__m512i b = a;
|
||||||
__mmask64 equality_mask = _mm512_cmp_epi8_mask(a, b, _MM_CMPINT_EQ);
|
__mmask64 equality_mask = _mm512_cmp_epi8_mask(a, b, _MM_CMPINT_EQ);
|
||||||
return 0;
|
return 0;
|
||||||
}
|
}
|
||||||
")
|
")
|
||||||
|
|
||||||
set(AVX2_CODE "
|
set(AVX2_CODE "
|
||||||
#include <immintrin.h>
|
#include <immintrin.h>
|
||||||
int main()
|
int main()
|
||||||
{
|
{
|
||||||
__m256i a = {0};
|
__m256i a = {0};
|
||||||
a = _mm256_abs_epi16(a);
|
a = _mm256_abs_epi16(a);
|
||||||
__m256i x;
|
__m256i x;
|
||||||
_mm256_extract_epi64(x, 0); // we rely on this in our AVX2 code
|
_mm256_extract_epi64(x, 0); // we rely on this in our AVX2 code
|
||||||
return 0;
|
return 0;
|
||||||
}
|
}
|
||||||
")
|
")
|
||||||
|
|
||||||
set(FMA_CODE "
|
set(FMA_CODE "
|
||||||
#include <immintrin.h>
|
#include <immintrin.h>
|
||||||
int main()
|
int main()
|
||||||
{
|
{
|
||||||
__m256 acc = _mm256_setzero_ps();
|
__m256 acc = _mm256_setzero_ps();
|
||||||
const __m256 d = _mm256_setzero_ps();
|
const __m256 d = _mm256_setzero_ps();
|
||||||
const __m256 p = _mm256_setzero_ps();
|
const __m256 p = _mm256_setzero_ps();
|
||||||
acc = _mm256_fmadd_ps( d, p, acc );
|
acc = _mm256_fmadd_ps( d, p, acc );
|
||||||
return 0;
|
return 0;
|
||||||
}
|
}
|
||||||
")
|
")
|
||||||
|
|
||||||
macro(check_sse type flags)
|
macro(check_sse type flags)
|
||||||
set(__FLAG_I 1)
|
set(__FLAG_I 1)
|
||||||
set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
|
set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
|
||||||
foreach(__FLAG ${flags})
|
foreach (__FLAG ${flags})
|
||||||
if(NOT ${type}_FOUND)
|
if (NOT ${type}_FOUND)
|
||||||
set(CMAKE_REQUIRED_FLAGS ${__FLAG})
|
set(CMAKE_REQUIRED_FLAGS ${__FLAG})
|
||||||
check_c_source_runs("${${type}_CODE}" HAS_${type}_${__FLAG_I})
|
check_c_source_runs("${${type}_CODE}" HAS_${type}_${__FLAG_I})
|
||||||
if(HAS_${type}_${__FLAG_I})
|
if (HAS_${type}_${__FLAG_I})
|
||||||
set(${type}_FOUND TRUE CACHE BOOL "${type} support")
|
set(${type}_FOUND TRUE CACHE BOOL "${type} support")
|
||||||
set(${type}_FLAGS "${__FLAG}" CACHE STRING "${type} flags")
|
set(${type}_FLAGS "${__FLAG}" CACHE STRING "${type} flags")
|
||||||
endif()
|
endif()
|
||||||
math(EXPR __FLAG_I "${__FLAG_I}+1")
|
math(EXPR __FLAG_I "${__FLAG_I}+1")
|
||||||
|
endif()
|
||||||
|
endforeach()
|
||||||
|
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
|
||||||
|
|
||||||
|
if (NOT ${type}_FOUND)
|
||||||
|
set(${type}_FOUND FALSE CACHE BOOL "${type} support")
|
||||||
|
set(${type}_FLAGS "" CACHE STRING "${type} flags")
|
||||||
endif()
|
endif()
|
||||||
endforeach()
|
|
||||||
set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
|
mark_as_advanced(${type}_FOUND ${type}_FLAGS)
|
||||||
|
|
||||||
if(NOT ${type}_FOUND)
|
|
||||||
set(${type}_FOUND FALSE CACHE BOOL "${type} support")
|
|
||||||
set(${type}_FLAGS "" CACHE STRING "${type} flags")
|
|
||||||
endif()
|
|
||||||
|
|
||||||
mark_as_advanced(${type}_FOUND ${type}_FLAGS)
|
|
||||||
|
|
||||||
endmacro()
|
endmacro()
|
||||||
|
|
||||||
# flags are for MSVC only!
|
# flags are for MSVC only!
|
||||||
check_sse("AVX" " ;/arch:AVX")
|
check_sse("AVX" " ;/arch:AVX")
|
||||||
if(NOT ${AVX_FOUND})
|
if (NOT ${AVX_FOUND})
|
||||||
set(LLAMA_AVX OFF)
|
set(LLAMA_AVX OFF)
|
||||||
else()
|
else()
|
||||||
set(LLAMA_AVX ON)
|
set(LLAMA_AVX ON)
|
||||||
|
@ -87,14 +86,14 @@ endif()
|
||||||
|
|
||||||
check_sse("AVX2" " ;/arch:AVX2")
|
check_sse("AVX2" " ;/arch:AVX2")
|
||||||
check_sse("FMA" " ;/arch:AVX2")
|
check_sse("FMA" " ;/arch:AVX2")
|
||||||
if((NOT ${AVX2_FOUND}) OR (NOT ${FMA_FOUND}))
|
if ((NOT ${AVX2_FOUND}) OR (NOT ${FMA_FOUND}))
|
||||||
set(LLAMA_AVX2 OFF)
|
set(LLAMA_AVX2 OFF)
|
||||||
else()
|
else()
|
||||||
set(LLAMA_AVX2 ON)
|
set(LLAMA_AVX2 ON)
|
||||||
endif()
|
endif()
|
||||||
|
|
||||||
check_sse("AVX512" " ;/arch:AVX512")
|
check_sse("AVX512" " ;/arch:AVX512")
|
||||||
if(NOT ${AVX512_FOUND})
|
if (NOT ${AVX512_FOUND})
|
||||||
set(LLAMA_AVX512 OFF)
|
set(LLAMA_AVX512 OFF)
|
||||||
else()
|
else()
|
||||||
set(LLAMA_AVX512 ON)
|
set(LLAMA_AVX512 ON)
|
||||||
|
|
Loading…
Add table
Add a link
Reference in a new issue