clk: fractional-divider: switch to rational best approximation
authorAndy Shevchenko <andriy.shevchenko@linux.intel.com>
Tue, 22 Sep 2015 15:54:11 +0000 (18:54 +0300)
committerStephen Boyd <sboyd@codeaurora.org>
Fri, 2 Oct 2015 18:29:48 +0000 (11:29 -0700)
This patch converts the code to use rational best approximation algorithm which
is much more precise.

Suggested-by: Stephen Boyd <sboyd@codeaurora.org>
Reviewed-by: Heikki Krogerus <heikki.krogerus@linux.intel.com>
Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Signed-off-by: Stephen Boyd <sboyd@codeaurora.org>
drivers/clk/Kconfig
drivers/clk/clk-fractional-divider.c

index 42f7120ca9ceaa1b900e737efeeb67f0913ae4e4..bd9f3120370f4ac57d2c01d54700d0d560182fa9 100644 (file)
@@ -14,6 +14,7 @@ config COMMON_CLK
        select HAVE_CLK_PREPARE
        select CLKDEV_LOOKUP
        select SRCU
+       select RATIONAL
        ---help---
          The common clock framework is a single definition of struct
          clk, useful across many platforms, as well as an
index 0282c7647a494345533be433d788c6b9ca8bc8aa..5c4955e33f7a259711b960593941df38d526f2bb 100644 (file)
@@ -7,13 +7,14 @@
  *
  * Adjustable fractional divider clock implementation.
  * Output rate = (m / n) * parent_rate.
+ * Uses rational best approximation algorithm.
  */
 
 #include <linux/clk-provider.h>
 #include <linux/module.h>
 #include <linux/device.h>
 #include <linux/slab.h>
-#include <linux/gcd.h>
+#include <linux/rational.h>
 
 #define to_clk_fd(_hw) container_of(_hw, struct clk_fractional_divider, hw)
 
@@ -22,7 +23,8 @@ static unsigned long clk_fd_recalc_rate(struct clk_hw *hw,
 {
        struct clk_fractional_divider *fd = to_clk_fd(hw);
        unsigned long flags = 0;
-       u32 val, m, n;
+       unsigned long m, n;
+       u32 val;
        u64 ret;
 
        if (fd->lock)
@@ -53,20 +55,30 @@ static long clk_fd_round_rate(struct clk_hw *hw, unsigned long rate,
                              unsigned long *parent_rate)
 {
        struct clk_fractional_divider *fd = to_clk_fd(hw);
-       unsigned maxn = (fd->nmask >> fd->nshift) + 1;
-       unsigned div;
+       unsigned long scale;
+       unsigned long m, n;
+       u64 ret;
 
        if (!rate || rate >= *parent_rate)
                return *parent_rate;
 
-       div = gcd(*parent_rate, rate);
+       /*
+        * Get rate closer to *parent_rate to guarantee there is no overflow
+        * for m and n. In the result it will be the nearest rate left shifted
+        * by (scale - fd->nwidth) bits.
+        */
+       scale = fls_long(*parent_rate / rate - 1);
+       if (scale > fd->nwidth)
+               rate <<= scale - fd->nwidth;
 
-       while ((*parent_rate / div) > maxn) {
-               div <<= 1;
-               rate <<= 1;
-       }
+       rational_best_approximation(rate, *parent_rate,
+                       GENMASK(fd->mwidth - 1, 0), GENMASK(fd->nwidth - 1, 0),
+                       &m, &n);
 
-       return rate;
+       ret = (u64)*parent_rate * m;
+       do_div(ret, n);
+
+       return ret;
 }
 
 static int clk_fd_set_rate(struct clk_hw *hw, unsigned long rate,
@@ -74,13 +86,12 @@ static int clk_fd_set_rate(struct clk_hw *hw, unsigned long rate,
 {
        struct clk_fractional_divider *fd = to_clk_fd(hw);
        unsigned long flags = 0;
-       unsigned long div;
-       unsigned n, m;
+       unsigned long m, n;
        u32 val;
 
-       div = gcd(parent_rate, rate);
-       m = rate / div;
-       n = parent_rate / div;
+       rational_best_approximation(rate, parent_rate,
+                       GENMASK(fd->mwidth - 1, 0), GENMASK(fd->nwidth - 1, 0),
+                       &m, &n);
 
        if (fd->lock)
                spin_lock_irqsave(fd->lock, flags);