Skip to content
Snippets Groups Projects
Select Git revision
  • First-development-phase
  • master default
  • v181-experimental
  • lga-intel181
  • ic-sdx182
  • ga-sdx182
  • lga-sdx182
  • ga-sdx174
  • lga-sdx174
  • lga-intel171
  • ga-intel171
  • lga-aoc160
  • ga-aoc160
  • LGA-SDX182-ToBePorted
  • FSP-2018
  • Fourth-development-phase
  • Third-development-phase
  • Second-development-phase
18 results

ocladock-fpga

user avatar
Leonardo Solis authored
Former-commit-id: 4b1a6b88
65d4bce9
History

Version compiled with Intel Altera Tools 16.0 (Quartus upgraded & patched) and targeting harp2 machine.

Source first:

source init_aoc_esa.sh

ofdock_datapar_alt: data-parallel version

ofdock_taskpar_alt: task-parallel version

Update to a new branch

The FPGA program is made of 8 kernels:

`GA` ->  `IC` + `GG` + `LS` -> `Conform` ->  `IE` + `IA` -> `Store`

The access to off-chip memory in order to update population and energy values is made in GA, IC, GG, LS, and Store. The memory access seems to not be synchronized even though mem_fence(CLK_GLOBAL_MEM_FENCE | CLK_CHANNEL_MEM_FENCE) was used.

According to this forum post, it is suggested to access off-chip memory using a single kernel.

A new branch called fusion is created where GA, IC, GG, LS, and Store are merged into a single kernel GA. That way the design doesn't rely anymore on mem_fences.