Some Early Bits Of The "Soft FP64" Infrastructure Will Be Mainlined Soon In Mesa

[phoronix]

Phoronix: Some Early Bits Of The "Soft FP64" Infrastructure Will Be Mainlined Soon In Mesa

David Airlie has announced his plans to begin mainlining some early infrastructure work on the "soft" FP64 code into Mesa Git. This doesn't yet allow for soft FP64 on older GPUs lacking the hardware capability to do this otherwise, but will help in another area and can make for easier mainlining of the actual soft FP64 support in the future...

http://www.phoronix.com/scan.php?page=news_item&px=Early-Soft-FP64-Work-Going-Git

[gsedej]

How come that gcn and intel supports fp64? Is all implemented on hardware like hd6900?
If now, there is some kind soft fp64? Why this code isnt working on r600?
Edit: i was talking about opengl fp64

[bridgman]

How come that gcn and intel supports fp64? Is all implemented on hardware like hd6900?

Correct - all GCN parts include HW support for fp64, although the ratio between fp32 and fp64 varies.

Before GCN, only specific high-end parts included fp64 - HD 58xx and HD69xx IIRC.

[gsedej]

Thank you for the reply!

[Dukenukemx]

Correct - all GCN parts include HW support for fp64, although the ratio between fp32 and fp64 varies.

Before GCN, only specific high-end parts included fp64 - HD 58xx and HD69xx IIRC.

Find it odd that a HD 58xx can do it while a HD 68xx can't.

[AJenbo]

Find it odd that a HD 58xx can do it while a HD 68xx can't.

Branding versus actual hardware generation probably

[chithanh]

Before GCN, only specific high-end parts included fp64 - HD 58xx and HD69xx IIRC.

And the Trinity/Richland APUs I believe.

Find it odd that a HD 58xx can do it while a HD 68xx can't.

From what I have read, 68xx were originally supposed to be called 67xx but ended up performing so close to 58xx (except for FP64) that the name was chosen.

[bridgman]

Find it odd that a HD 58xx can do it while a HD 68xx can't.

58xx was the top-end 5xxx part, and 69xx was the top-end 6xxx part. IIRC we had a 59xx board but it was dual 58xx, so 58xx was the top GPU.

The other reason fp64 was limited to 69xx is that we were moving from 5-wide to 4-wide ALUs to improve compute efficiency, and 69xx was the first 4-wide part. The rest of the 6xxx line were all 5-wide IIRC.

[chithanh]

5970 was mostly a dual 5870, the dual 6970 was named 6990.

5870 was VLIW5, yet supported FP64 in hardware, as do all APUs since Kaveri. Interestingly, Carrizo/Bristol Ridge could even do FP64 at 1:2 the FP32 rate. Unfortunately it seems that Raven Ridge will be much slower at FP64 than Bristol Ridge.

[starshipeleven]

How come that gcn and intel supports fp64? Is all implemented on hardware like hd6900?
If now, there is some kind soft fp64? Why this code isnt working on r600?
Edit: i was talking about opengl fp64

Last time I asked this, I think the Intel guy told me that from HD4000 onwards the Intel iGPUs actually have hardware support for fp64.