This is an email posted to nesdev by Ki a while back.  I have removed one
line at the end regarding the B flag of the cpu(the information was 
incorrect, which Ki noted in a later email).

--------------------------------------------------------------------------------

 By reading Brad's NESSOUND document, we know that there is a
"frame counter" in the NES/FC APU.  I would like to post
some more on this.

  The frame counter is reset upon any write to $4017.  It is
reset at system power-on as well, but is NOT reset upon
system reset.

  Thanks to Samus Aran, we now know the exact period of the
PPU's single frame.  In another words, we are now sure that
the NMI occurs on every 29780 2/3 CPU cycles.

  However, the APU's single frame is NOT 29780 2/3 CPU cycles.
What I mean by "APU's single frame" here is that it is the
number of CPU cycles taken between the frame IRQs.

  The APU's single frame seems to be:

    1789772.727... / 60 = 29829 6/11  [CPU CYCLE]

  Below is a simple diagram which shows the difference
in periods of the PPU's single frame and the APU's.


    RESET         29780 2/3 CPU CYCLES          NMI
PPU   |------------------------------------------|
      |           29829 6/11 CPU CYCLES          IRQ
APU   |----------|----------|----------|----------|


  Note that if you write $00 to $4017 on every NMI, the frame
IRQ would NEVER go off even if it is enabled.  This is because
the the period of NMI is slightly shorter than the period of
the frame IRQ.  This causes the frame counter to be reset
before the frame IRQ goes off.

When you write zero to bit 7 of $4017, the frame counter will
be reset, and the first sound update will be done after 7457 CPU
cycles (i.e. 29829/4).  2nd update will be done 7457 after that,
same goes for 3rd update and 4th update, but the frame IRQ occurs
on 4th update, resetting the frame counter as well.

When you write 1 to bit 7 of $4017, the frame counter will be
reset, but the first sound update will occur at the same time.
2nd, 3rd, and 4th update will be done after 7457, 14914, 22371
CPU cycles after the first update respectively, but the 5th
update will be 14914 cycles after the 4th update.  This causes
sound output to last 1.25 times longer than that of bit 7 = 0.


$4017W:

o when the MSB of $4017 is 0:

bit7=0
  |---------|---------|---------|---------|---------|---------|----
            1st       2nd       3rd       4th    5th(1st)  6th(2nd)


o when the MSB of $4017 is 1:

bit7=1
  |---------|---------|---------|-------------------|---------|----
  1st       2nd       3rd       4th              5th(1st)  6th(2nd)


On 1st, 3rd, 5th, ... updates, the envelope decay and the
linear counter are updated.

On 2nd, 4th, 6th, ... updates, the envelope decay, the
linear counter, the length counter, and the frequency sweep
are updated.
----

  The original info was provided by goroh, and verified by me.
However, it could still be wrong.  Please tell me if you
find anything wrong.
----

(Correction from my last posting)

  I have checked once again and it turned out that the frame IRQ
was NOT disabled upon system reset.  What actually prevented the
frame IRQ to occur after system reset was, in fact, the I flag.
I checked this flag shortly after system reset (right after stack
pointer was initialized), and the flag was 1, although I never
executed "sei" after reset.  Therefore the I flag of the PR2A03G
is 1 on system reset.

  Thanks Matthew Conte and Samus Aran for pointing out the
inaccuracy.