Skip to content

userguide.md

Latest commit

 

History

History
132 lines (104 loc) · 6.53 KB

userguide.md

File metadata and controls

132 lines (104 loc) · 6.53 KB

User Guide

Repository Structure

  • api : host code to setup and execute FPGA bitstreams. Compiled to static library that can be linked to your application
  • kernels : OpenCL kernel code for 1d, 2d and 3d FFT
  • examples: Sample code that makes use of the api
  • cmake : cmake modules used by the build system
  • scripts: convenience slurm scripts
  • docs : describes models regarding performance and resource utilization

CMake Build Setup

External Libraries

These additional libraries are automatically fetched during system configuration:

  • cxxopts for command line argument parsing
  • hlslib for CMake Intel FPGA OpenCL find packages
  • findFFTW for CMake FFTW find package
  • gtest for unit tests

Configuration Options

The following compile options can be set when creating a CMake build directory either using the -D parameter or by using the cmake-gui such as:

cmake -DCMAKE_BUILD_TYPE=Release ..

ccmake ..

** Name ** ** Description ** ** Default Values ** ** Alternate Values **
AOC\_FLAGS* * Intel offline compiler flags used for kernel compilation -g -v -no-interleaving=default
EMU\_FLAGS Compiler flags used for emulation, with fast emulation as default -march=emulator
FPGA\_BOARD\_NAME Name of the target FPGA board p520\_hpc\_sg280l pac\_s10\_usm
LOG\_FFT\_SIZE Currently supported log2 number of points along each FFT dimension 6 5, 7, 8, 9 
BURST\_INTERLEAVING* Toggle to enable burst interleaved global memory accesses
Sets the -no-interleaving= to the AOC\_FLAGS* parameter 		 NO YES
DDR\_BUFFER\_LOCATION Name of the global memory interface found in the board\_spec.xml
DDR :p520\_hpc\_sg280l, device : pac\_s10\_usm board 		 DDR device
SVM\_BUFFER\_LOCATION Name of the SVM global memory interface found in the board\_spec.xml* *
"" : p520\_hpc\_sg280l, host: pac\_s10\_usm		 host
CMAKE\_BUILD\_TYPE Specify the build type Debug Release, RelWithDebInfo

Additional Kernel Builds

Generation of Intel OpenCL Offline Compiler reports

make <kernel_name>_report
make fft1d_report

Using the GNU Debugger (gdb) with the Debug builds

CMake debug builds lets you step through code using gdb.

gdb --args ./fft -n 64 -d 2 -p <path-to-bitstream>

Runtime Input Parameters

Offloading FFT on FPGA
Usage:
  ./fft<..> [OPTION...]

  -n, --num arg    Number of sample points in a dimension (default: 64)
  -d, --dim arg    Number of dimensions (default: 3)
  -b, --back       Toggle Backward FFT
  -i, --iter arg   Number of iterations (default: 1)
  -p, --path arg   Path to FPGA bitstream
  -y, --noverify   Toggle to not verify with FFTW
  -c, --batch arg  Number of batches of FFT calculations in FPGA (default: 1)
  -t, --burst      Toggle to use burst interleaved global memory accesses  in
                   FPGA
  -m, --use_bram   Toggle to use BRAM instead of DDR for 3D Transpose  
  -s, --use_usm    Toggle to use Unified Shared Memory features for data
                   transfers between host and device
  -e, --emulate    Toggle to enable emulation 
  -h, --help       Print usage

Output Interpretation

The examples measure and output relevant performance metrics that are shown below:

------------------------------------------
FFT CONFIGURATION: 
--------------------------------------------
Type               : Complex to Complex
Points             : 64 
Direction          : Forward 
Placement          : In Place    
Batch              : 1 
Iterations         : 1 
Transpose3D        : DDR 
Burst Interleaving : No 
Emulation          : Yes 
USM Feature        : No 
--------------------------------------------
-- Initializing FPGA ...
-- 1 platforms found
	0: intel(r) fpga emulation platform for opencl(tm)
-- 1 devices found
	Choosing first device by default
-- Getting program binary from path: p520_hpc_sg280l/emulation/fft1d_64_nointer/fft1d.aocx
-- Building the program
0: Calculating FFT - 
-- Launching 1D FFT of 1 batches 
Launching FFT transform for 1 batch 
-- Copying data from host to device
-- Executing kernels
-- Transfering results back to host
-- Cleaning up FPGA resources ...
------------------------------------------
Measurements 
--------------------------------------------
PCIe Write          = 0.0000ms
Kernel Execution    = 0.0182ms
Kernel Exec/Batch   = 0.0182ms
PCIe Read           = 0.0000ms
Total               = 0.0182ms
Throughput          = 0.0982GFLOPS/s | 26.8213 GB/s

  • PCIe Write: time taken in milliseconds to transfer data from host memory of the CPU to the global memory of the FPGA.

  • PCIe Read : the time taken in milliseconds to transfer data from global memory of the FPGA to the host memory of the CPU.

  • Kernel Execution : the time taken in milliseconds for the execution of the required kernels, which includes the global memory accesses.

  • Total : PCIe Write + Kernel Execution + PCIe Read

  • Throughput : $$ \frac{dim * 5 * N^{dim} * log_2 N}{runtime}$$