fix: expf overflow on avx2.

b4rtaz · b4rtaz · commit cba7e3716bd0 · 2025-08-08T12:11:27.000+02:00
diff --git a/src/nn/nn-cpu-ops.cpp b/src/nn/nn-cpu-ops.cpp
@@ -84,6 +84,8 @@ static inline float horizontalMax_avx2(__m256 v) {
 }
 
 static inline __m256 expf_avx2(__m256 x) {
+    x = _mm256_max_ps(x, _mm256_set1_ps(-88.0f));
+
     const __m256 log2e = _mm256_set1_ps(1.4426950408889634f);
     const __m256 c0 = _mm256_set1_ps(1.0f);
     const __m256 c1 = _mm256_set1_ps(0.6931471805599453f);

Original file line number	Diff line number	Diff line change
`@@ -84,6 +84,8 @@ static inline float horizontalMax_avx2(__m256 v) {`
`84`	`84`	`}`
`85`	`85`
`86`	`86`	`static inline __m256 expf_avx2(__m256 x) {`
	`87`	`+ x = _mm256_max_ps(x, _mm256_set1_ps(-88.0f));`
	`88`	`+`
`87`	`89`	`const __m256 log2e = _mm256_set1_ps(1.4426950408889634f);`
`88`	`90`	`const __m256 c0 = _mm256_set1_ps(1.0f);`
`89`	`91`	`const __m256 c1 = _mm256_set1_ps(0.6931471805599453f);`