shithub: opus

Info  •  Files  •  Log  •  Branches

Improve accuracy of AVX sigmoid

Reciprocal approximation could cause the sigmoid output to be
greater than 1.0.

--- a/dnn/vec_avx.h

+++ b/dnn/vec_avx.h

@@ -148,7 +148,8 @@

         __m256 X, Y;

         X = _mm256_loadu_ps(&x[i]);

         Y = exp8_approx(X);

-        Y = _mm256_mul_ps(Y,  _mm256_rcp_ps(_mm256_add_ps(Y, one)));

+        /* Compute as 1-1/(1+e^x) to avoid >1 values caused by the reciprocal approximation. */

+        Y = _mm256_sub_ps(one, _mm256_mul_ps(one,  _mm256_rcp_ps(_mm256_add_ps(Y, one))));

         _mm256_storeu_ps(&y[i], Y);

     }

     for (;i<N;i++)

-- 

⑨