Cuda support_eac3

EACirc provides a version of gate circuit that is capable of running on CUDA devices. EACirc must be compiled with special settings to enable CUDA support. The CUDA-enabled binary can run considerably faster. To run the accelerated computation you need s special GPU supporting CUDA (basically any modern nVIDIA GPU). The minimal supported GPU architecture is Fermi. Note that CUDA is capable of running only on x64 machines.

Building

• Prerequisites: Properly installed CUDA version 7 or newer. To download the CUDA SDK visit https://developer.nvidia.com/cuda-downloads
• Enable CUDA support: When CMake detects CUDA (usually done automatically when CUDA is properly installed), a CMake option BUILD_CUDA becomes available. If this option is chosen then the compiled binary will be capable of running on CUDA devices.

Running

Prerequisites

• nVIDIA driver supporting CUDA 7 (note: I think a version 350 or higher should do)
• EACirc binary compiled with CUDA support

Configuration

• Enable CUDA (option EACIRC/CUDA/ENABLED is set to 1).
• Use Gate 2 circuit backend (option EACIRC/MAIN/CIRCUIT_REPRESENTATION is set to 3).
• Disable circuit memory (option EACIRC/GATE_CIRCUIT/USE_MEMORY is set to 0).
• Disable JVMSim functionality (option EACIRC/GA_CIRCUIT/ALLOWED_FUNCTIONS/FNC_JVM is set to 0).
• Parameter EACIRC/TEST_VECTORS/SET_SIZE should be at least 32000 (any other number will do, but a multiple of cuda block size is the most effective).

Advanced Configuration

Note: Changing these options could dramatically increase or decrease the effectiveness of CUDA circuit. (If you don't know what you are doing decrease is most likely.)

• option EACIRC/CUDA/BLOCK_SIZE - size of the block in kernel (default is 512)
• option EACIRC/TEST_VECTORS/SET_SIZE - number of running threads in kernel
• If any of the following options is changed then it is recommended to experimentally find the most effective settings for CUDA circuit (block size and test set size).
  • EACIRC/MAIN/CIRCUIT_SIZE_INPUT
  • EACIRC/MAIN/CIRCUIT_SIZE_OUTPUT
  • EACIRC/GATE_CIRCUIT/NUM_LAYERS
  • EACIRC/GATE_CIRCUIT/SIZE_LAYER
  • EACIRC/GATE_CIRCUIT/NUM_CONNECTORS
• EACIRC/MAIN/SAVE_STATE_FREQ any number higher than 100 will do, preferably 1000

List of all effective settings

Default (Jirka's thesis)

• EACIRC/CUDA/BLOCK_SIZE = 512
• EACIRC/TEST_VECTORS/SET_SIZE = 3200
• EACIRC/MAIN/SAVE_STATE_FREQ = 1000
• EACIRC/MAIN/CIRCUIT_SIZE_INPUT = 16
• EACIRC/MAIN/CIRCUIT_SIZE_OUTPUT = 1
• EACIRC/GATE_CIRCUIT/NUM_LAYERS = 5
• EACIRC/GATE_CIRCUIT/SIZE_LAYER = 8
• EACIRC/GATE_CIRCUIT/NUM_CONNECTORS = 4