perf: Do only one iteration of refinement for FP16 inv

FP16 only has 10 explicitly stored bits of mantissa, so one iteration
of newton-raphson refinement should be precise enough. Numerical
results showed both one iteration and two iterations of newton-raphson
refinement had worst ULP being 1.

Signed-off-by: Ben Niu <[email protected]>

Author: mcfi

src/core/NEON/NEMath.inl

@@ -634,15 +634,13 @@ inline float16x4_t vinv_f16(float16x4_t x)
 {
     float16x4_t recip = vrecpe_f16(x);
     recip             = vmul_f16(vrecps_f16(x, recip), recip);
-    recip             = vmul_f16(vrecps_f16(x, recip), recip);
     return recip;
 }
 
 inline float16x8_t vinvq_f16(float16x8_t x)
 {
     float16x8_t recip = vrecpeq_f16(x);
     recip             = vmulq_f16(vrecpsq_f16(x, recip), recip);
-    recip             = vmulq_f16(vrecpsq_f16(x, recip), recip);
     return recip;
 }
 

src/core/NEON/SVEMath.inl

@@ -82,7 +82,6 @@ inline svfloat16_t svinv_f16_z(svbool_t pg, svfloat16_t x)
 {
     auto recip = svrecpe_f16(x);
     recip      = svmul_f16_z(pg, svrecps_f16(x, recip), recip);
-    recip      = svmul_f16_z(pg, svrecps_f16(x, recip), recip);
     return recip;
 }