diff options
| author | Shawn Nematbakhsh <shawnn@chromium.org> | 2017-06-15 15:07:35 -0700 |
|---|---|---|
| committer | chrome-bot <chrome-bot@chromium.org> | 2017-06-19 15:33:09 -0700 |
| commit | 90167c1764ab7c532b121bd08990b81631a3c896 (patch) | |
| tree | cb2b1dadc7052437d071b872b6aff04df221d0d8 /common/math_util.c | |
| parent | 0696dd1e68692ca8dae17b737e0d895561f65cd4 (diff) | |
| download | chrome-ec-90167c1764ab7c532b121bd08990b81631a3c896.tar.gz | |
math_util: Use HW instruction for int_sqrtf when available
Cortex-M4 supports a floating point square root function that takes 14
cycles to execute, which is a speed improvement over the existing binary
search, and saves flash space.
BUG=chromium:687624
BRANCH=None
TEST=On kevin, verify that both sqrtf methods (binary search vs HW
instruction) have identical results for fractional input (eg.
sqrt(15.999999) = 3), except when floating point representation of
square root rounds up to an integer. Verify identical results for all
integers [-100, 16793602). Note that 16793602 is the first integer for
which the floating point representation of sqrt rounds up to an integer.
Also verify basic motion sense functions on kevin.
Signed-off-by: Shawn Nematbakhsh <shawnn@chromium.org>
Change-Id: I8521c9a28c958b340ca83c37342253e424df0c91
Reviewed-on: https://chromium-review.googlesource.com/537734
Commit-Ready: Shawn N <shawnn@chromium.org>
Tested-by: Shawn N <shawnn@chromium.org>
Reviewed-by: Shawn N <shawnn@chromium.org>
Diffstat (limited to 'common/math_util.c')
| -rw-r--r-- | common/math_util.c | 17 |
1 files changed, 16 insertions, 1 deletions
diff --git a/common/math_util.c b/common/math_util.c index ad98aecdf5..6a49dc31c3 100644 --- a/common/math_util.c +++ b/common/math_util.c @@ -6,6 +6,7 @@ /* Common math functions. */ #include "common.h" +#include "math.h" #include "math_util.h" #include "util.h" @@ -73,7 +74,20 @@ fp_t arc_cos(fp_t x) /** * Integer square root. */ -int int_sqrtf(fp_inter_t x) +#ifdef CONFIG_FPU +/* + * Use library sqrtf instruction, if available, since it's usually much faster + * and smaller. On Cortex-M4, this becomes a single instruction which takes + * 14 cycles to execute. This produces identical results to binary search, + * except when the floating point representation of the square root rounds up + * to an integer. + */ +static inline int int_sqrtf(fp_inter_t x) +{ + return sqrtf(x); +} +#else +static int int_sqrtf(fp_inter_t x) { int rmax = 0x7fffffff; int rmin = 0; @@ -110,6 +124,7 @@ int int_sqrtf(fp_inter_t x) } } } +#endif int vector_magnitude(const vector_3_t v) { |