diff options
| author | tgrogers <[email protected]> | 2018-06-29 17:34:22 -0400 |
|---|---|---|
| committer | tgrogers <[email protected]> | 2018-06-29 17:34:22 -0400 |
| commit | 112a3a2e3c9d17af420d2389f68a35407c692744 (patch) | |
| tree | 728a9157fae6d77097fd19799e80a0c0511c8098 /configs/QuadroFX5600 | |
| parent | 8334828738e3dec7221bd411d3c541bd940fde59 (diff) | |
Allowing modern GPU configrations to properly run PTXPLUS.
There are several problems with this:
1) To get proper occupancy, based on register file usage, you must querry ptxas with
the sm version that corresponds to the register usage in the config file. To enable this,
a new config option has been added that determines what sm version you pass to ptxas to compute reg-usage.
This configuration option is always required in the gpgpusim.config file
2) If you are running PTXPLUS with a modern card (i.e. volta/pascal), you need ptxas from CUDA 9.1.
However since PTXPLUS only supports sm_13 - you need a version of CUDA where cudaobjectdump
supports sm_1x. This ended at CUDA 5 - and PTXPLUS requires CUDA 4.2. Therefore, to run PTXPLUS
on a modern card, you need CUDA 4.2 + modern CUDA installed. To fascilitate this, a new envronment
varaible is added and the setup envrionment script prints an appropraite warning if you are using a newer CUDA.
We have tried to make this as fail-proof as possible - and die appropraitely when something is wrong.
Diffstat (limited to 'configs/QuadroFX5600')
| -rw-r--r-- | configs/QuadroFX5600/gpgpusim.config | 3 |
1 files changed, 3 insertions, 0 deletions
diff --git a/configs/QuadroFX5600/gpgpusim.config b/configs/QuadroFX5600/gpgpusim.config index cb87b65..6f836ee 100644 --- a/configs/QuadroFX5600/gpgpusim.config +++ b/configs/QuadroFX5600/gpgpusim.config @@ -11,6 +11,9 @@ # shader core pipeline config -gpgpu_shader_registers 16384 +-gpgpu_occupancy_sm_number 12 + +-gpgpu_occupancy_sm_number 12 #8192 (registers per block as written by device Query and which used in this option in our other configurations but this break some benchmarks execution! it does not affect performance modeling though) -gpgpu_shader_core_pipeline 768:32 -gpgpu_shader_cta 8 |