Testing OpenGL 4.1 With An AMD Cypress GPU On The Latest Open-Source Driver

[phoronix]

Phoronix: Testing OpenGL 4.1 With An AMD Cypress GPU On The Latest Open-Source Driver

When it comes to OpenGL 4 support on the AMD R600 Gallium3D driver for pre-GCN graphics cards, currently the only R600g-supported cards advertising OpenGL 4.1 right now are the Radeon HD 5800 "Cypress" and Radeon HD 6900 "Cayman" series. Here are some tests done with OpenGL 4.1 on a Radeon HD 5830 compared to Cayman and various GPUs with the RadeonSI Gallium3D driver...

http://www.phoronix.com/scan.php?page=news_item&px=AMD-Cypress-OpenGL-4.1

[MannerMan]

Will cards like 6870 ever get 4.x support by the r600 driver? Seems so strange to me that 5xxx cards can support it but not the majority of 6xxx cards

[bridgman]

Sure, "all" it needs is compiler code to emulate double-precision floating point shader instructions using integer or single-precision hardware.

IIRC the majority of 5xxx cards can't support it either, it just seems a bit different because we had three SKUs from 58xx, not the usual two.

[MannerMan]

Sure, "all" it needs is compiler code to emulate double-precision floating point shader instructions using integer or single-precision hardware.

IIRC the majority of 5xxx cards can't support it either, it just seems a bit different because we had three SKUs from 58xx, not the usual two.

Okey, thanks for clarifying. I'll keep waiting =)

[duby229]

Sure, "all" it needs is compiler code to emulate double-precision floating point shader instructions using integer or single-precision hardware.

IIRC the majority of 5xxx cards can't support it either, it just seems a bit different because we had three SKUs from 58xx, not the usual two.

Northern Islands does have double precision hardware right? At least it was definitely advertised that it did.

EDIT: Nope, It sure doesn't I guess I just remembered something I thought, rather than something I read.

http://www.rage3d.com/reviews/video/amd_hd6870_hd6850_launch_review/index.php?p=2

[bridgman]

Generally the top model of each generation had DP support (or faster DP support starting with SI) while the rest of the generation did not.

I think that applies to NI as well -- Cayman has it, as do the Cayman derivatives (Trinity/Richland), but don't think the rest do (although that's partially based on 68xx owners asking about DP emulation).

[duby229]

Generally the top model of each generation had DP support (or faster DP support starting with SI) while the rest of the generation did not.

I think that applies to NI as well -- Cayman has it, as do the Cayman derivatives, but don't think the rest do (partially based on 68xx owners asking about DP emulation).

Barts was the top model VLIW5 NI.

[bridgman]

True, but Cayman was the top model NI, same "product generation" even if the guts were different. IWRC (If Wikipedia Remembers Correctly) Barts was launched first, then Cayman, then Turks & Caicos.

One of the reasons for moving from VLIW5 to VLIW4 was improving ALU utilization with compute workloads, so it's tough to argue that the top model VLIW5 NI needed DP as well...

[duby229]

True, but Cayman was the top model NI, same "product generation" even if the guts were different. IWRC (If Wikipedia Remembers Correctly) Barts was launched first, then Cayman, then Turks & Caicos.

I've already proven I don't remember the subject well. But Barts is VLIW5 and Cayman is VLIW4. At least in that respect they are very different. Other than the compute units though they probably were very similar.

One of the reasons for moving from VLIW5 to VLIW4 was improving ALU utilization with compute workloads, so it's tough to argue that the top model VLIW5 NI needed DP as well...

Someones shortsightedness. OGL4.3 was already specified by the time Barts got on the design table.

[bridgman]

Yeah, NI was originally going to be on a different fab process so all the parts were going to be larger (more SIMDs) than what we shipped.

When the fab process didn't happen as expected we used the new VLIW4 core on the top model (since compute tends to focus on the top model and VLIW4's largest benefit was with compute) and stayed with VLIW5 for the rest.

When GCN came along we basically added a GCN core to a largely NI uncore, since NI had originally been designed for 28 in the first place.