Marek Takes To RadeonSI Tweaking For Unigine Superposition
It looks like one of the latest test targets for well known AMD open-source developer Marek Olšák is Unigine Superposition.
Unigine Superposition, released earlier this year for Windows and Linux, is a beast on OpenGL drivers and all current generation graphics cards. Superposition is one of the workloads that hasn't been well optimized yet for RadeonSI and is where the open-source AMD stack struggles compared to the competition.
Fortunately, it looks like Marek is now eyeing Superposition for RadeonSI improvements.
Today he posted a new patch to enable sisched for Superposition. Sisched is the LLVM SI machine instruction scheduler that has been around for a while but isn't enabled by default, but requires setting R600_DEBUG=sisched.
Marek's patch would enable sisched by default where the process name is superposition. This application-specific switch would be hard-coded into the RadeonSI code rather than being a drirc tunable.
Marek noted in the patch, "+2.3% better score on Fiji. It might be better without HBM." Curious about the impact of sisched for Unigine Superposition, I ran some benchmarks:
A Radeon RX 470 with Mesa 17.2-dev Git on Linux 4.12:
Indeed, some minor gains, though in some scenarios the sisched results were flat. All of the Superposition benchmarks I ran today via this OpenBenchmarking.org result file.
Unigine Superposition, released earlier this year for Windows and Linux, is a beast on OpenGL drivers and all current generation graphics cards. Superposition is one of the workloads that hasn't been well optimized yet for RadeonSI and is where the open-source AMD stack struggles compared to the competition.
Fortunately, it looks like Marek is now eyeing Superposition for RadeonSI improvements.
Today he posted a new patch to enable sisched for Superposition. Sisched is the LLVM SI machine instruction scheduler that has been around for a while but isn't enabled by default, but requires setting R600_DEBUG=sisched.
Marek's patch would enable sisched by default where the process name is superposition. This application-specific switch would be hard-coded into the RadeonSI code rather than being a drirc tunable.
Marek noted in the patch, "+2.3% better score on Fiji. It might be better without HBM." Curious about the impact of sisched for Unigine Superposition, I ran some benchmarks:
A Radeon RX 470 with Mesa 17.2-dev Git on Linux 4.12:
Indeed, some minor gains, though in some scenarios the sisched results were flat. All of the Superposition benchmarks I ran today via this OpenBenchmarking.org result file.
40 Comments