From 669dd229835122143ff49b0c921e69a3a31ed46b Mon Sep 17 00:00:00 2001 From: =?utf8?q?Marek=20Ol=C5=A1=C3=A1k?= Date: Sat, 22 Sep 2018 20:03:27 -0400 Subject: [PATCH] util: document a limitation of util_fast_udiv32 trivial --- src/util/fast_idiv_by_const.h | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/src/util/fast_idiv_by_const.h b/src/util/fast_idiv_by_const.h index 92a3ccdf222..638b52a3ffb 100644 --- a/src/util/fast_idiv_by_const.h +++ b/src/util/fast_idiv_by_const.h @@ -135,7 +135,13 @@ static inline uint32_t util_fast_udiv32(uint32_t n, struct util_fast_udiv_info info) { n = n >> info.pre_shift; - /* For non-power-of-two divisors, use a 32-bit ADD that clamps to UINT_MAX. */ + /* If the divisor is not 1, you can instead use a 32-bit ADD that clamps + * to UINT_MAX. Dividing by 1 needs the full 64-bit ADD. + * + * If you have unsigned 64-bit MAD with 32-bit inputs, you can do: + * increment = increment ? multiplier : 0; // on the CPU + * (n * multiplier + increment) // on the GPU using unsigned 64-bit MAD + */ n = (((uint64_t)n + info.increment) * info.multiplier) >> 32; n = n >> info.post_shift; return n; -- 2.11.0