diff options
author | Paul Walmsley <paul@pwsan.com> | 2009-01-28 14:08:41 -0500 |
---|---|---|
committer | Russell King <rmk+kernel@arm.linux.org.uk> | 2009-02-08 12:50:34 -0500 |
commit | 85a5f78d2b15a2e73b6486a24b77bb3ab67d5bbc (patch) | |
tree | 6ea454fb3e0a26ede82ea2c3e0d881758b41c0fc /arch/arm/mach-omap2/clock.c | |
parent | b32450409847dddf060a468707212d3493df4f63 (diff) |
[ARM] OMAP3 clock: optimize DPLL rate rounding algorithm
The previous DPLL rate rounding algorithm counted the divider (N) down
from the maximum to 1. Since we currently use a broad DPLL rate
tolerance, and lower N values are more power-efficient, we can often
bypass several iterations through the loop by counting N upwards from
1.
Peter de Schrijver <peter.de-schrijver@nokia.com> put up with several
test cycles of this patch - thanks Peter.
linux-omap source commit is 6f6d82bb2f80fa20a841ac3e95a6f44a5a156188.
Signed-off-by: Paul Walmsley <paul@pwsan.com>
Cc: Peter de Schrijver <peter.de-schrijver@nokia.com>
Signed-off-by: Tony Lindgren <tony@atomide.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
Diffstat (limited to 'arch/arm/mach-omap2/clock.c')
-rw-r--r-- | arch/arm/mach-omap2/clock.c | 35 |
1 files changed, 18 insertions, 17 deletions
diff --git a/arch/arm/mach-omap2/clock.c b/arch/arm/mach-omap2/clock.c index 2f0eaaad4819..76e20bcc4e8a 100644 --- a/arch/arm/mach-omap2/clock.c +++ b/arch/arm/mach-omap2/clock.c | |||
@@ -45,7 +45,7 @@ | |||
45 | #define DPLL_MIN_DIVIDER 1 | 45 | #define DPLL_MIN_DIVIDER 1 |
46 | 46 | ||
47 | /* Possible error results from _dpll_test_mult */ | 47 | /* Possible error results from _dpll_test_mult */ |
48 | #define DPLL_MULT_UNDERFLOW (1 << 0) | 48 | #define DPLL_MULT_UNDERFLOW -1 |
49 | 49 | ||
50 | /* | 50 | /* |
51 | * Scale factor to mitigate roundoff errors in DPLL rate rounding. | 51 | * Scale factor to mitigate roundoff errors in DPLL rate rounding. |
@@ -826,7 +826,7 @@ static int _dpll_test_mult(int *m, int n, unsigned long *new_rate, | |||
826 | unsigned long target_rate, | 826 | unsigned long target_rate, |
827 | unsigned long parent_rate) | 827 | unsigned long parent_rate) |
828 | { | 828 | { |
829 | int flags = 0, carry = 0; | 829 | int r = 0, carry = 0; |
830 | 830 | ||
831 | /* Unscale m and round if necessary */ | 831 | /* Unscale m and round if necessary */ |
832 | if (*m % DPLL_SCALE_FACTOR >= DPLL_ROUNDING_VAL) | 832 | if (*m % DPLL_SCALE_FACTOR >= DPLL_ROUNDING_VAL) |
@@ -847,13 +847,13 @@ static int _dpll_test_mult(int *m, int n, unsigned long *new_rate, | |||
847 | if (*m < DPLL_MIN_MULTIPLIER) { | 847 | if (*m < DPLL_MIN_MULTIPLIER) { |
848 | *m = DPLL_MIN_MULTIPLIER; | 848 | *m = DPLL_MIN_MULTIPLIER; |
849 | *new_rate = 0; | 849 | *new_rate = 0; |
850 | flags = DPLL_MULT_UNDERFLOW; | 850 | r = DPLL_MULT_UNDERFLOW; |
851 | } | 851 | } |
852 | 852 | ||
853 | if (*new_rate == 0) | 853 | if (*new_rate == 0) |
854 | *new_rate = _dpll_compute_new_rate(parent_rate, *m, n); | 854 | *new_rate = _dpll_compute_new_rate(parent_rate, *m, n); |
855 | 855 | ||
856 | return flags; | 856 | return r; |
857 | } | 857 | } |
858 | 858 | ||
859 | /** | 859 | /** |
@@ -892,21 +892,27 @@ long omap2_dpll_round_rate(struct clk *clk, unsigned long target_rate) | |||
892 | 892 | ||
893 | dd->last_rounded_rate = 0; | 893 | dd->last_rounded_rate = 0; |
894 | 894 | ||
895 | for (n = dd->max_divider; n >= DPLL_MIN_DIVIDER; n--) { | 895 | for (n = DPLL_MIN_DIVIDER; n <= dd->max_divider; n++) { |
896 | 896 | ||
897 | /* Compute the scaled DPLL multiplier, based on the divider */ | 897 | /* Compute the scaled DPLL multiplier, based on the divider */ |
898 | m = scaled_rt_rp * n; | 898 | m = scaled_rt_rp * n; |
899 | 899 | ||
900 | /* | 900 | /* |
901 | * Since we're counting n down, a m overflow means we can | 901 | * Since we're counting n up, a m overflow means we |
902 | * can immediately skip to the next n | 902 | * can bail out completely (since as n increases in |
903 | * the next iteration, there's no way that m can | ||
904 | * increase beyond the current m) | ||
903 | */ | 905 | */ |
904 | if (m > scaled_max_m) | 906 | if (m > scaled_max_m) |
905 | continue; | 907 | break; |
906 | 908 | ||
907 | r = _dpll_test_mult(&m, n, &new_rate, target_rate, | 909 | r = _dpll_test_mult(&m, n, &new_rate, target_rate, |
908 | clk->parent->rate); | 910 | clk->parent->rate); |
909 | 911 | ||
912 | /* m can't be set low enough for this n - try with a larger n */ | ||
913 | if (r == DPLL_MULT_UNDERFLOW) | ||
914 | continue; | ||
915 | |||
910 | e = target_rate - new_rate; | 916 | e = target_rate - new_rate; |
911 | pr_debug("clock: n = %d: m = %d: rate error is %d " | 917 | pr_debug("clock: n = %d: m = %d: rate error is %d " |
912 | "(new_rate = %ld)\n", n, m, e, new_rate); | 918 | "(new_rate = %ld)\n", n, m, e, new_rate); |
@@ -918,16 +924,11 @@ long omap2_dpll_round_rate(struct clk *clk, unsigned long target_rate) | |||
918 | min_e_n = n; | 924 | min_e_n = n; |
919 | 925 | ||
920 | pr_debug("clock: found new least error %d\n", min_e); | 926 | pr_debug("clock: found new least error %d\n", min_e); |
921 | } | ||
922 | 927 | ||
923 | /* | 928 | /* We found good settings -- bail out now */ |
924 | * Since we're counting n down, a m underflow means we | 929 | if (min_e <= clk->dpll_data->rate_tolerance) |
925 | * can bail out completely (since as n decreases in | 930 | break; |
926 | * the next iteration, there's no way that m can | 931 | } |
927 | * increase beyond the current m) | ||
928 | */ | ||
929 | if (r & DPLL_MULT_UNDERFLOW) | ||
930 | break; | ||
931 | } | 932 | } |
932 | 933 | ||
933 | if (min_e < 0) { | 934 | if (min_e < 0) { |