diff options
Diffstat (limited to 'sc/source/core/tool')
-rw-r--r-- | sc/source/core/tool/arraysumAVX.cxx | 21 | ||||
-rw-r--r-- | sc/source/core/tool/arraysumAVX512.cxx | 41 | ||||
-rw-r--r-- | sc/source/core/tool/arraysumSSE2.cxx | 21 |
3 files changed, 30 insertions, 53 deletions
diff --git a/sc/source/core/tool/arraysumAVX.cxx b/sc/source/core/tool/arraysumAVX.cxx index c55d71f22983..e256248047d0 100644 --- a/sc/source/core/tool/arraysumAVX.cxx +++ b/sc/source/core/tool/arraysumAVX.cxx @@ -20,7 +20,9 @@ namespace sc::op { -#ifdef LO_AVX_AVAILABLE // Old processors +#ifdef LO_AVX_AVAILABLE + +bool hasAVXCode() { return true; } using namespace AVX; @@ -48,13 +50,10 @@ static inline void sumAVX(__m256d& sum, __m256d& err, const __m256d& value) sum = t; } -#endif - /** Execute Kahan sum with AVX. */ KahanSumSimple executeAVX(size_t& i, size_t nSize, const double* pCurrent) { -#ifdef LO_AVX_AVAILABLE // Make sure we don't fall out of bounds. // This works by sums of 8 terms. // So the 8'th term is i+7 @@ -107,14 +106,16 @@ KahanSumSimple executeAVX(size_t& i, size_t nSize, const double* pCurrent) return { sums[0], errs[0] }; } return { 0.0, 0.0 }; -#else - (void)i; - (void)nSize; - (void)pCurrent; - abort(); -#endif } +#else // LO_AVX_AVAILABLE + +bool hasAVXCode() { return false; } + +KahanSumSimple executeAVX(size_t&, size_t, const double*) { abort(); } + +#endif + } // end namespace sc::op /* vim:set shiftwidth=4 softtabstop=4 expandtab: */ diff --git a/sc/source/core/tool/arraysumAVX512.cxx b/sc/source/core/tool/arraysumAVX512.cxx index 987e5a3e6ff6..6a3235a58e2e 100644 --- a/sc/source/core/tool/arraysumAVX512.cxx +++ b/sc/source/core/tool/arraysumAVX512.cxx @@ -18,25 +18,11 @@ #include <stdlib.h> -/* TODO Remove this once GCC updated and AVX512 can work. */ -#ifdef __GNUC__ -#if __GNUC__ < 9 -#ifdef LO_AVX512F_AVAILABLE -#define HAS_LO_AVX512F_AVAILABLE -#undef LO_AVX512F_AVAILABLE -#endif -#endif -#endif - namespace sc::op { #ifdef LO_AVX512F_AVAILABLE -const bool hasAVX512F = cpuid::hasAVX512F(); -#else -const bool hasAVX512F = false; -#endif -#ifdef LO_AVX512F_AVAILABLE // New processors +bool hasAVX512FCode() { return true; } using namespace AVX512; @@ -62,13 +48,10 @@ static inline void sumAVX512(__m512d& sum, __m512d& err, const __m512d& value) sum = t; } -#endif - /** Execute Kahan sum with AVX512. */ KahanSumSimple executeAVX512F(size_t& i, size_t nSize, const double* pCurrent) { -#ifdef LO_AVX512F_AVAILABLE // New processors // Make sure we don't fall out of bounds. // This works by sums of 8 terms. // So the 8'th term is i+7 @@ -122,24 +105,16 @@ KahanSumSimple executeAVX512F(size_t& i, size_t nSize, const double* pCurrent) return { sums[0], errs[0] }; } return { 0.0, 0.0 }; -#else - (void)i; - (void)nSize; - (void)pCurrent; - abort(); -#endif } -} // end namespace sc::op +#else // LO_AVX512F_AVAILABLE + +bool hasAVX512FCode() { return false; } + +KahanSumSimple executeAVX512F(size_t&, size_t, const double*) { abort(); } -/* TODO Remove this once GCC updated and AVX512 can work. */ -#ifdef __GNUC__ -#if __GNUC__ < 9 -#ifdef HAS_LO_AVX512F_AVAILABLE -#define LO_AVX512F_AVAILABLE -#undef HAS_LO_AVX512F_AVAILABLE -#endif -#endif #endif +} // end namespace sc::op + /* vim:set shiftwidth=4 softtabstop=4 expandtab: */ diff --git a/sc/source/core/tool/arraysumSSE2.cxx b/sc/source/core/tool/arraysumSSE2.cxx index b4edb98286f9..1a5cc2f00dfe 100644 --- a/sc/source/core/tool/arraysumSSE2.cxx +++ b/sc/source/core/tool/arraysumSSE2.cxx @@ -20,7 +20,9 @@ namespace sc::op { -#ifdef LO_SSE2_AVAILABLE // Old processors +#ifdef LO_SSE2_AVAILABLE + +bool hasSSE2Code() { return true; } using namespace SSE2; @@ -47,13 +49,10 @@ static inline void sumSSE2(__m128d& sum, __m128d& err, const __m128d& value) sum = t; } -#endif - /** Execute Kahan sum with SSE2. */ KahanSumSimple executeSSE2(size_t& i, size_t nSize, const double* pCurrent) { -#ifdef LO_SSE2_AVAILABLE // Make sure we don't fall out of bounds. // This works by sums of 8 terms. // So the 8'th term is i+7 @@ -120,13 +119,15 @@ KahanSumSimple executeSSE2(size_t& i, size_t nSize, const double* pCurrent) return { sums[0], errs[0] }; } return { 0.0, 0.0 }; -#else - (void)i; - (void)nSize; - (void)pCurrent; - abort(); -#endif } + +#else // LO_SSE2_AVAILABLE + +bool hasSSE2Code() { return false; } + +KahanSumSimple executeSSE2(size_t&, size_t, const double*) { abort(); } + +#endif } /* vim:set shiftwidth=4 softtabstop=4 expandtab: */ |