From a8d8d817c0eb9522ebc581524621d68fbb26e3b3 Mon Sep 17 00:00:00 2001 From: Matteo Cicuttin <datafl4sh@toxicnet.eu> Date: Tue, 11 May 2021 15:21:47 +0000 Subject: [PATCH] Updated changelog & doc --- doc/changelog.md | 3 ++- doc/lua_api.md | 2 +- 2 files changed, 3 insertions(+), 2 deletions(-) diff --git a/doc/changelog.md b/doc/changelog.md index 83e6c43..ec7db78 100644 --- a/doc/changelog.md +++ b/doc/changelog.md @@ -13,4 +13,5 @@ Unavailable features: - Repetitive sources on GPU. ## Version v0.2 (NOT RELEASED YET) -- Fixed performance issue in the evaluation of boundary/interface sources (4463681c). \ No newline at end of file +- Fixed performance issue in the evaluation of boundary/interface sources (4463681c). +- Added HIP support via Hipify (e6f5a020) diff --git a/doc/lua_api.md b/doc/lua_api.md index 41ed455..8800676 100644 --- a/doc/lua_api.md +++ b/doc/lua_api.md @@ -113,4 +113,4 @@ None. - `magnetic_initial_condition(x, y, z)`: if defined, the solver calls this function at the beginning of the simulation in order to initialize the magnetic field. The parameters `x`, `y` and `z` represent the point at which the field needs to be evaluated. The function has to return a triple specifying the three field components, i.e. `return Hx, Hy, Hz`. ### Considerations on the evaluation of sources and the computation on GPU -Currently sources for the timestep `t+1` are evaluated asynchronously on the CPU while the GPU is computing the timestep `t`. When the GPU is done, the updated sources are uploaded from the CPU to the GPU and the cycle restarts. This usually works relatively well on big domains with high approximation order, however on small domains at low approximation order does not give any speedup. For this reason, if the sources do not need to be evaluated along the whole simulation, it is suggested to turn them off at the appropriate timestep by using the `on_timestep()` callback and the appropriate functions described below. A future version of the solver will implement repetitive boundary sources, that will be precomputed on the CPU and uploaded on the GPU only once. \ No newline at end of file +Currently sources for the timestep `t+1` are evaluated asynchronously on the CPU while the GPU is computing the timestep `t`. The sources at `t+1` are then uploaded asynchronously from the CPU to the GPU. This usually works relatively well, especially on big domains with high approximation order, however on small domains at low approximation order you can notice slowdowns. For this reason, if the sources do not need to be evaluated along the whole simulation, it is suggested to turn them off at the appropriate timestep by using the `on_timestep()` callback and the appropriate functions described below. Despite of that, I am not sure that implementing repetitive boundary sources will be necessary. -- GitLab