 |
StencilStream v3.0.0
SYCL-based Stencil Simulation Framework Targeting FPGAs
|
Loading...
Searching...
No Matches
|
StencilStream v3.0.0
SYCL-based Stencil Simulation Framework Targeting FPGAs
|
A grid updater that applies an iterative stencil code to a grid. More...
#include <StencilUpdate.hpp>
Classes | |
| struct | Params |
| Parameters for the stencil updater. More... | |
Public Types | |
| using | GridImpl = Grid< Cell, word_size > |
| Shorthand for the used and supported grid type. | |
Public Member Functions | |
| StencilUpdate (Params params) | |
| Create a new stencil updater object. | |
| Params & | get_params () |
| Return a reference to the parameters. | |
| GridImpl | operator() (GridImpl &source_grid) |
| Compute a new grid based on the source grid, using the configured transition function. | |
| uindex_t | get_n_processed_cells () const |
| Return the accumulated total number of cells processed by this updater. | |
| double | get_kernel_runtime () const |
| Return the accumulated total runtime of the execution kernel. | |
| double | get_walltime () const |
| Return the accumulated runtime of the updater, measured from the host side. | |
A grid updater that applies an iterative stencil code to a grid.
This updater applies an iterative stencil code, defined by the template parameter F, to the grid; As often as requested. Since the underlying FPGA design follows the Monotile architecture (See monotile), an instance of this updater template can only process grids up to the defined max_grid_width and max_grid_height.
| F | The transition function to apply to input grids. |
| n_processing_elements | (Optimization parameter) The number of processing elements (PEs) to implement. Increasing the number of PEs leads to a higher performance since more iterations are computed in parallel. However, it will also increase the resource and space usage of the design. Too many PEs might also decrease the clock frequency. |
| max_grid_width | (Optimization parameter) The maximally supported grid width. For best hardware utilization, this should be a power of two. Increase this parameter to the maximum your application is expected to handle. However, higher maximal grid width might lead to increased logic and space usage as well as decreased clock frequencies. |
| max_grid_height | (Optimization parameter) The maximally supported grid height. Increase this parameter to the maximum your application is expected to handle. However, increasing the maximal grid height will increase the BRAM usage of each PE. |
| TDVStrategy | (Optimization parameter) The precomputation strategy for the time-dependent value system. |
| word_size | (Optimization parameter) The width of the global memory channel, in bytes. For DDR-based systems, this should be 512 bits, or 64 bytes. |
| using stencil::monotile::StencilUpdate< F, n_processing_elements, max_grid_width, max_grid_height, TDVStrategy, word_size >::GridImpl = Grid<Cell, word_size> |
Shorthand for the used and supported grid type.
|
inline |
Create a new stencil updater object.
|
inline |
Return the accumulated total runtime of the execution kernel.
This runtime is accumulated across multiple calls to operator()(). However, this is only possible if Params::profiling is set to true.
|
inline |
Return the accumulated total number of cells processed by this updater.
For each call of to operator()(), this is the width times the height of the grid, times the number of computed iterations. This will also be accumulated across multiple calls to operator()().
|
inline |
Return a reference to the parameters.
Modifications to the parameters struct will be used in the next call to operator()().
|
inline |
Return the accumulated runtime of the updater, measured from the host side.
For each call to operator()(), the time it took to submit all kernels and, if Params::blocking is true, to finish the computation is recorded and accumulated.
|
inline |
Compute a new grid based on the source grid, using the configured transition function.
The computation does not work in-place. Instead, it will allocate two additional grids with the same size as the source grid and use them for a double buffering scheme. Therefore, you are free to reuse the source grid as it will not be altered.
If Params::blocking is set to true, this method will block until the computation is complete. Otherwise, it will return as soon as all kernels are submitted.
1.9.8