https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77776
--- Comment #17 from Matthias Kretz (Vir) <mkretz at gcc dot gnu.org> ---
hypotf(a, b) is implemented using double precision and hypot(a, b) uses 80-bit
long double on i386 and x86_64 hypot does what you describe, right?
std::experimental::simd benchmarks of hypot(a, b), where simd_abi::scalar uses
the <cmath> implementation (i.e. glibc):
-march=skylake-avx512 -ffast-math -O3 -lmvec:
TYPE Latency Speedup Throughput
Speedup
[cycles/call] [per value] [cycles/call] [per
value]
float, simd_abi::scalar 37.5 1 11.5
1
float, 37.6 0.999 10.2
1.13
float, simd_abi::__sse 34 4.42 6.46
7.15
float, simd_abi::__avx 34.1 8.79 6.56
14.1
float, simd_abi::_Avx512<32> 34.3 8.76 6.01
15.4
float, simd_abi::_Avx512<64> 44.1 13.6 12
15.4
float, [[gnu::vector_size(16)]] 58.3 2.57 47.5
0.974
float, [[gnu::vector_size(32)]] 132 2.27 104
0.892
float, [[gnu::vector_size(64)]] 240 2.5 222
0.832
--------------------------------------------------------------------------------------
TYPE Latency Speedup Throughput
Speedup
[cycles/call] [per value] [cycles/call] [per
value]
double, simd_abi::scalar 81 1 21.5
1
double, 80.1 1.01 21.3
1.01
double, simd_abi::__sse 39.9 4.06 6.47
6.64
double, simd_abi::__avx 40.2 8.05 12
7.14
double, simd_abi::_Avx512<32> 40.3 8.04 12
7.14
double, simd_abi::_Avx512<64> 56.2 11.5 24
7.14
double, [[gnu::vector_size(16)]] 89.3 1.81 42.5
1.01
double, [[gnu::vector_size(32)]] 150 2.16 110
0.777
double, [[gnu::vector_size(64)]] 297 2.18 242
0.71
--------------------------------------------------------------------------------------
-march=skylake-avx512 -O3 -lmvec:
TYPE Latency Speedup Throughput
Speedup
[cycles/call] [per value] [cycles/call] [per
value]
float, simd_abi::scalar 37.6 1 10.4
1
float, 37.7 0.998 10.2
1.02
float, simd_abi::__sse 37.6 4 8.83
4.71
float, simd_abi::__avx 37.5 8.01 9.42
8.82
float, simd_abi::_Avx512<64> 47.8 12.6 12
13.8
float, [[gnu::vector_size(16)]] 98.7 1.52 57.2
0.727
float, [[gnu::vector_size(32)]] 151 2 114
0.728
float, [[gnu::vector_size(64)]] 260 2.31 230
0.722
--------------------------------------------------------------------------------------
TYPE Latency Speedup Throughput
Speedup
[cycles/call] [per value] [cycles/call] [per
value]
double, simd_abi::scalar 79.7 1 21.7
1
double, 80.1 0.995 21.6
1
double, simd_abi::__sse 44.2 3.6 9.99
4.33
double, simd_abi::__avx 43.6 7.32 12
7.21
double, simd_abi::_Avx512<64> 59.9 10.6 24
7.21
double, [[gnu::vector_size(16)]] 88.3 1.8 44.2
0.98
double, [[gnu::vector_size(32)]] 163 1.96 115
0.75
double, [[gnu::vector_size(64)]] 302 2.11 233
0.742
--------------------------------------------------------------------------------------
I have never ported my SIMD implementation back to scalar and benchmarked it
against glibc.
