OCCCL - OCCAM-Pi with openCL

ABSTRACT

AIM

Investiage the feasibility and the performance of opencl/C acceleration on existing pure occam programs.

METHODOLOGY

All startup is done by initial.c. Two methods are provided. initialise() returns a cl_context and cl_device_id which is used to compile the program from src. Each CL program creates its own cl_program_queue at every run of the CL program.

buildcl() builds from src a program which can then be executed by the cl_program_queue

All CL programs are in one .c file with the _init call and the call seperated. The _init needs to run once before the CL program can be executed. There is only a check if it is enabled at compile time.

HARDWARE
CL_DEVICE_TYPE_GPU:
  ATI RV770 (HD4850) (cl_khr_byte_addressable_store not supported)
    OpenCL 1.1 SDK = AMD APP SDK 2.5
    Extensions = cl_amd_fp64 cl_khr_gl_sharing cl_amd_device_attribute_query
    Num Compute units = 10
  Nvidia GTX560 (GF114)
    Extensions = OpenCL device : GeForce GTX 560
cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll  cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_fp64
  Nvidia 8600GT
  Nvidia GT430
  Nvidia ION
    OpenCL 1.1 SDK = Nvidia CUDA 4.0.1
    Extensions = cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll
CL_DEVICE_TYPE_CPU:
  2x Intel E5420
  Intel Core i7-2600

RESULTS
mandelbrot (fp64, 32workers, 128frames) benchmark:
Occam (Intel Q6600 @ 3.0ghz):
  real    0m17.877s
  user    0m35.928s
  sys     0m0.773s
OpenCL (Intel Q6600 @ 3.0ghz + nvidia GTX560):
  real    0m3.467s
  user    0m0.780s
  sys     0m1.013s
C (Intel Q6600 @ 3.0ghz):
  real    0m4.323s
  user    0m3.230s
  sys     0m0.670s

mandebrot no disp (gcc 4.6.2 with -O2
  (Intel Q6600 @ 3.0ghz + nvidia GTX560)
  OpenCL implementation (frame - with opts -D USE_DOUBLE -cl-fast-relaxed-math -cl-mad-enable)
    438ms
  C implementation (frame)
    2663ms
  Occam implementation (line)
    13413ms

Occoids-single:
  (Intel Q6600 @ 3.0Ghz + nvidia GTX560)
  Occam implementation
Cycle       50; cycle time =    14719 us;     2180 processes
Cycle      100; cycle time =    14730 us;     2180 processes
Cycle      150; cycle time =    14885 us;     2180 processes
Cycle      200; cycle time =    15706 us;     2180 processes
Cycle      250; cycle time =    15967 us;     2180 processes
Cycle      300; cycle time =    17533 us;     2180 processes
Cycle      350; cycle time =    16094 us;     2180 processes
Cycle      400; cycle time =    16547 us;     2180 processes
Cycle      450; cycle time =    16460 us;     2180 processes
Cycle      500; cycle time =    16140 us;     2180 processes
Cycle      550; cycle time =    16998 us;     2180 processes
Cycle      600; cycle time =    17053 us;     2180 processes
Cycle      650; cycle time =    16883 us;     2180 processes
Cycle      700; cycle time =    16191 us;     2180 processes
Cycle      750; cycle time =    16688 us;     2180 processes
Cycle      800; cycle time =    16797 us;     2180 processes
Cycle      850; cycle time =    17053 us;     2180 processes
Cycle      900; cycle time =    17501 us;     2180 processes
Cycle      950; cycle time =    18061 us;     2180 processes


CONCLUSION

READING