From 72acb66527df6f38c7b8b15fa5062a616d67074b Mon Sep 17 00:00:00 2001 From: Dylan Baker Date: Thu, 6 Sep 2018 15:30:15 -0700 Subject: [PATCH] u_math: add x86 optimized version of ifloor MIME-Version: 1.0 Content-Type: text/plain; charset=utf8 Content-Transfer-Encoding: 8bit This is copied from the one in src/mesa/main/imports.h, which is the same otherwise. Reviewed-by: Marek Olšák Reviewed-by: Kristian H. Kristensen Reviewed-by: Matt Turner Part-of: --- src/util/u_math.h | 18 ++++++++++++++++++ 1 file changed, 18 insertions(+) diff --git a/src/util/u_math.h b/src/util/u_math.h index 6c2cb5437c0..59266c16922 100644 --- a/src/util/u_math.h +++ b/src/util/u_math.h @@ -185,6 +185,23 @@ util_fast_pow(float x, float y) static inline int util_ifloor(float f) { +#if defined(USE_X86_ASM) && defined(__GNUC__) && defined(__i386__) + /* + * IEEE floor for computers that round to nearest or even. + * 'f' must be between -4194304 and 4194303. + * This floor operation is done by "(iround(f + .5) + iround(f - .5)) >> 1", + * but uses some IEEE specific tricks for better speed. + * Contributed by Josh Vanderhoof + */ + int ai, bi; + double af, bf; + af = (3 << 22) + 0.5 + (double)f; + bf = (3 << 22) + 0.5 - (double)f; + /* GCC generates an extra fstp/fld without this. */ + __asm__ ("fstps %0" : "=m" (ai) : "t" (af) : "st"); + __asm__ ("fstps %0" : "=m" (bi) : "t" (bf) : "st"); + return (ai - bi) >> 1; +#else int ai, bi; double af, bf; union fi u; @@ -193,6 +210,7 @@ util_ifloor(float f) u.f = (float) af; ai = u.i; u.f = (float) bf; bi = u.i; return (ai - bi) >> 1; +#endif } -- 2.30.2