Prevent implicit conversion of float to double to avoid precision error on i686. https://github.com/Maratyszcza/FP16/issues/20 Taken from Debian: https://salsa.debian.org/deeplearning-team/fp16/-/blob/master/debian/patches/ftbfs-i386.patch Index: fp16/include/fp16/fp16.h =================================================================== --- fp16.orig/include/fp16/fp16.h +++ fp16/include/fp16/fp16.h @@ -228,7 +228,8 @@ static inline uint16_t fp16_ieee_from_fp const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000)); const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000)); #endif - float base = (fabsf(f) * scale_to_inf) * scale_to_zero; + const volatile float base_inf = fabsf(f) * scale_to_inf; + float base = base_inf * scale_to_zero; const uint32_t w = fp32_to_bits(f); const uint32_t shl1_w = w + w;
aboutsummaryrefslogtreecommitdiff |