Hardware-accelerated implementation for IOTA PearlDiver, which utilizes multi-threaded SIMD and GPU.

dcurl exploits SIMD instructions on CPU and OpenCL on GPU. Both CPU and GPU accelerations can be enabled in multi-threaded execuction fashion, resulting in much faster proof-of-work (PoW) for IOTA Reference Implementation (IRI).

• You need to configure OpenCL platform and device by yourself in src/clcontext.c
• You need to configure paths and flags of OpenCL installation in Makefile
• Check JDK installation and set JAVA_HOME if you wish to specify.
• Only one GPU can be facilitated with dcurl at the moment.

• dcurl allows various combinations of build configurations to fit final use scenarios.
• You can execute make config and then edit file build/local.mk for custom build options.
  • BUILD_AVX: build Intel AVX-accelerated Curl backend.
  • BUILD_GPU: build OpenCL-based GPU accelerations.
  • BUILD_JNI: build a shared library for IRI. The build system would generate JNI header file from downloading from latest JAVA source.
  • BUILD_COMPAT: build extra cCurl compatible interface.
• Alternatively, you can specify conditional build as following:

$ make BUILD_GPU=0 BUILD_JNI=1 BUILD_AVX=1

• Test with GPU

$ make BUILD_GPU=1 check

• Expected Results

*** Validating build/test-trinary ***
        [ Verified ]
*** Validating build/test-curl ***
        [ Verified ]
*** Validating build/test-multi_pow_cpu ***
        [ Verified ]
*** Validating build/test-pow_sse ***
        [ Verified ]
*** Validating build/test-pow_cl ***
        [ Verified ]
*** Validating build/test-multi_pow_gpu ***
        [ Verified ]

• Test with AVX but no GPU

$ make BUILD_AVX=1 check

• Expected Results

*** Validating build/test-trinary ***
        [ Verified ]
*** Validating build/test-curl ***
        [ Verified ]
*** Validating build/test-multi_pow_cpu ***
        [ Verified ]
*** Validating build/test-pow_avx ***
        [ Verified ]

• dcurl_init(2, 1) in jni/iri-pearldiver-exlib.c
  • 2 means 2 pow tasks executed in CPU,
  • 1 means 1 pow tasks executed in GPU at the same time.
• Number of threads to find nonce in CPU
  • $ export DCURL_NUM_CPU=26

After integrating dcurl into IRI, performance of <attachToTangle> is measured as following.

• Each sampling is measured with 30 transaction trytes and total 200 samples are measured.
• mwm = 14, 26 CPU threads to find nonce
• Settings: enable 2 pow tasks in CPU, 1 pow tasks in GPU at the same time

Modified IRI accepting external PoW Library

• $ cd ~/iri && mvn compile && mvn package
• $ cp ~/dcurl/build/libdcurl.so ~/iri
• $ cd ~/iri && java -Djava.library.path=./ -jar target/iri.jar -p <port> --pearldiver-exlib dcurl

• More test program :(
• Pre-compile OpenCL kernel functions and include it in dcurl.
• Automatically configure dcurl after init()

dcurl is freely redistributable under the MIT License. Use of this source code is governed by a MIT-style license that can be found in the LICENSE file.