| Message ID | 1458577000-6615-19-git-send-email-nouveau@karolherbst.de |
|---|---|
| State | New |
| Headers | show |
| Series |
"Volting/Clocking improvements for Fermi and newer"
( rev:
2
)
in
Nouveau |
diff --git a/drm/nouveau/nvkm/subdev/volt/base.c b/drm/nouveau/nvkm/subdev/volt/base.c index 58738e3..2b3f460 100644 --- a/drm/nouveau/nvkm/subdev/volt/base.c +++ b/drm/nouveau/nvkm/subdev/volt/base.c @@ -117,13 +117,55 @@ nvkm_volt_map(struct nvkm_volt *volt, u8 id, u8 temp) vmap = nvbios_vmap_entry_parse(bios, id, &ver, &len, &info); if (vmap) { - if (info.link != 0xff) { - int ret = nvkm_volt_map(volt, info.link, temp); - if (ret < 0) - return ret; - info.min += ret; + switch (ver) { + case 0x10: + if (info.link != 0xff) { + int ret = nvkm_volt_map(volt, info.link, temp); + if (ret < 0) + return ret; + info.min += ret; + } + return info.min; + case 0x20: { + s32 result; + u16 temp = 40; + switch (info.mode) { + case 0x0: + result = info.arg[0] / 10; + result += info.arg[1] * 168; + result += info.arg[2] * 28; + break; + case 0x1: + result = (info.arg[0] / 1675) * 100; + result += info.arg[1] * 100; + result += (info.arg[2] / 10) * 153 * temp; + result += info.arg[3] * 100 * temp; + result += info.arg[4] * 41; + result += (info.arg[5] * (temp * temp)) / 15; + break; + case 0x3: + result = (info.min + info.max) / 2; + break; + case 0x2: + default: + result = info.min; + break; + } + result = min(max(result, (s32)info.min), + (s32)info.max); + + if (info.link != 0xff) { + int ret = nvkm_volt_map(volt, info.link, temp); + if (ret < 0) + return ret; + result += ret; + } + + return result; + } + default: + return -ENODEV; } - return info.min; } return id ? id * 10000 : -ENODEV;
On 21/03/16 18:16, Karol Herbst wrote: > I am sure that those are a bit different on other GPUs, but while testing > the error range compared to nvidia was around 100%+-3%. > > Without this change we are most of the time around 10% below nvidias voltage, > so this change causes no harm and improves the situation a lot already. > > The remaining task for this is to figure out which of these constants are > chip specific and from where to get the chip specific factors > > Signed-off-by: Karol Herbst <nouveau@karolherbst.de> Instead of landing these coefficients like that, I would really like if you could dump pfuse and check if you find these coefficients there. What could be done also is to re-do the work you did to find out those coefficients on a GPU where your coefficients are not really good and find out what values would have needed and then compare pfuse with the original card. If this is to be considered a long task, then I am OK landing those coefficients, but I would really appreciate if you could at least have a quick look at pfuse :)
I am sure that those are a bit different on other GPUs, but while testing the error range compared to nvidia was around 100%+-3%. Without this change we are most of the time around 10% below nvidias voltage, so this change causes no harm and improves the situation a lot already. The remaining task for this is to figure out which of these constants are chip specific and from where to get the chip specific factors Signed-off-by: Karol Herbst <nouveau@karolherbst.de> --- drm/nouveau/nvkm/subdev/volt/base.c | 54 ++++++++++++++++++++++++++++++++----- 1 file changed, 48 insertions(+), 6 deletions(-)