diff options
author | Cástor Muñoz <cmvidal@gmail.com> | 2016-05-15 02:32:40 +0200 |
---|---|---|
committer | Cástor Muñoz <cmvidal@gmail.com> | 2016-05-15 22:57:12 +0200 |
commit | 52a6db3a363e915b1af7eb37778baaee81306c48 (patch) | |
tree | 0987c86a4f429f32b0ec236eba23d5b88ac65a3b /firmware/export | |
parent | 201f57171a67530a288f2a29a7a7060293e93bf2 (diff) | |
download | rockbox-52a6db3a363e915b1af7eb37778baaee81306c48.tar.gz rockbox-52a6db3a363e915b1af7eb37778baaee81306c48.zip |
iPod Classic: optimize UDMA timming for ATA drives
This patch optimizes UDMA timings to increase write transfer rate on
ATA bus, these transfers are clocked by HCLK, tDVS+tDVH is modified to
decrease Tcyctyp (typical write cycle period). This is not overclocking,
we meet the ATA standar, the settings used by OF are not well optimized
for each UDMA mode, we will never know but probably this was due some
documentation issue.
ATA_UDMA_TIME register is documented on s3c6400 datasheet, information
included in s5l8700 datasheet is wrong or not valid for s5l8702.
From ATA specs, (Minimum, Maximum) values in nanoseconds:
UDMA 0 UDMA 1 UDMA 2 UDMA 3 UDMA 4
tACKENV (20, 70) (20, 70) (20, 70) (20, 55) (20, 55)
tRP (160, --) (125, --) (100, --) (100, --) (100, --)
tSS (50, --) (50, --) (50, --) (50, --) (50, --)
tDVS (70, --) (48, --) (31, --) (20, --) (6.7, --)
tDVH (6.2, --) (6.2, --) (6.2, --) (6.2, --) (6.2, --)
tDVS+tDVH (120, --) (80, --) (60, --) (45, --) (30, --)
Tcyc = tDVS+tDVH
WR[bytes/s] = 1/Tcyc[s] * 2[bytes]
On Classic (boosted):
HClk = 108 MHz. -> T = ~9.26 ns.
Old values (used by OF):
UDMA ATA_UDMA_TIME tACK tRP tSS tDVS tDVH Tcyc WR(MB/s)
0 0x5071152 27.8 166.7 55.6 74.1 55.6 129.7 15.4
1 0x3050a52 27.8 101.8 55.6 55.6 37 92.6 21.6
2 0x3030a52 27.8 101.8 55.6 37 37 74 27
3 0x2020a52 27.8 101.8 55.6 27.8 27.8 55.6 36
4 0x2010a52 27.8 101.8 55.6 18.5 27.8 46.3 43.2
New values:
UDMA ATA_UDMA_TIME tACK tRP tSS tDVS tDVH Tcyc WR(MB/s)
0 0x4071152 27.8 166.7 55.6 74.1 46.3 120.4 16.6
1 0x2050d52 27.8 129.6 55.6 55.6 27.8 83.4 24
2 0x2030a52 27.8 101.8 55.6 37 27.8 64.8 30.9
3 0x1020a52 27.8 101.8 55.6 27.8 18.5 46.3 43.2
4 0x1010a52 27.8 101.8 55.6 18.5 18.5 37 54
To verify that the settings are correct, a write-to-cache test was
performed using emCORE, the measured transfer rate (WRm) is compared
against the theoric transfer rate (WR) at 108 Mhz for the old and
the new UDMA4 settings (iPod 160, HDD Toshiba MK1634GAL):
UDMA ATA_UDMA_TIME Tcyc(ns) WR(MB/s) WRm(MB/s) RDm(MB/s)
4 0x2010a52 46.3 43.2 42.9 59.8
4 0x1010a52 37 54 53.5 59.8
Notes:
- The new UDMA4 settings increases ~25% the ATA transfer rate for
cached-writes. The real HDD write speed is limited by the internal
transfer rate (depends on cilinder, for the MK1634GAL it is 276 to
573 Mbits/s). Sequential write benchmark using diskdump on USB are
~8% faster.
- Read transfers are clocked by the device, it depends on UDMA mode
selected and are not affected by HClk or ATA_UDMA_TIME settings.
Read-from-cache tests results (RDm) using HClk=108 and HClk=54 for
UDMA4 are 59.8 MB/s on MK1634GAL.
- Minimum HClk is limited by tACKENV specs, using current settings
it is 54 MHz for UDMA4,UDMA3 and 43 MHz for UDMA2,UDMA1,UDMA0.
Change-Id: I61d67060410752518a59e1ff08072b21747ca997
Diffstat (limited to 'firmware/export')
0 files changed, 0 insertions, 0 deletions