From f1d7b15cd23311423faac145febceb8c5ba09d81 Mon Sep 17 00:00:00 2001 From: Hadrien Grasland <hadrien.grasland@gmx.fr> Date: Mon, 8 Jul 2024 13:10:22 +0200 Subject: [PATCH] Move work-group size tuning after update tuning --- handouts/src/25-simulating.md | 10 ---------- handouts/src/26-up-to-date.md | 10 ++++++++++ 2 files changed, 10 insertions(+), 10 deletions(-) diff --git a/handouts/src/25-simulating.md b/handouts/src/25-simulating.md index 4c123a8..19958b8 100644 --- a/handouts/src/25-simulating.md +++ b/handouts/src/25-simulating.md @@ -300,16 +300,6 @@ adjust the microbenchmark to evaluate its performance: Then run the resulting microbenchmark, and compare the performance characteristics of our first GPU version to the optimized CPU version. -Finally, if you are going fast, experiment with different work-group sizes in -order to study the impact of this tuning parameter on performance. In the -current setup, you need to... - -- Change the work-group size declaration in the GLSL compute shader. -- Change the matching CPU-side declaration in the `gpu::pipeline` module. -- Recompile the program. - -Making this tuning process easier will be the job of specialization constants. - --- diff --git a/handouts/src/26-up-to-date.md b/handouts/src/26-up-to-date.md index d40dab8..534c832 100644 --- a/handouts/src/26-up-to-date.md +++ b/handouts/src/26-up-to-date.md @@ -91,3 +91,13 @@ number of compute steps per image becomes sufficiently large. Since the threshold is likely to be hardware-dependent, you will probably want to make it configurable via command-line arguments. + +Finally, if you are going fast, experiment with different work-group sizes in +order to study the impact of this tuning parameter on performance. In the +current setup, you need to... + +- Change the work-group size declaration in the GLSL compute shader. +- Change the matching CPU-side declaration in the `gpu::pipeline` module. +- Recompile the program. + +Making this tuning process easier will be the job of specialization constants. -- GitLab