pentschev

Description

All benchmarks are composed of 10 batches of 2-dimensional matrices, with sizes varying from 128x128 to 4096x4096 with single-precision.

CUDA Results

NVIDIA Tesla K20

Matrix dimensions: 128x128 In-place C2C FFT time for 10 runs: 0.538662 ms

pentschev

pentschev

pentschev

(lldb)
There is a running process, detach from it and attach?: [Y/n]
Process 6843 detached
_bt.cpython-37m-darwin.so was compiled with optimization - stepping may behave oddly; variables may not be available.
Process 6843 stopped
* thread #1, queue = 'com.apple.main-thread', stop reason = signal SIGSTOP
    frame #0: 0x0000000114b54ab0 _bt.cpython-37m-darwin.so`backtrace_thread [inlined] _wait_and_reset_signal at bt.c:182:35 [opt]
   179 	static
   180 	void _wait_and_reset_signal(struct sigaction *old_sa) {
   181 	  // spin and wait.

pentschev

FROM python:3

ENV PYTHON 3.7
ENV NUMPY 1.16.2
ENV UPSTREAM_DEV 1
ENV TEST true
ENV LINT true
ENV COVERAGE false
ENV PARALLEL true
ENV XTRATESTARGS ''

pentschev

pentschev

pentschev

from dask.distributed import Client
from dask_cuda import LocalCUDACluster
from dask.array.utils import assert_eq

import dask.array as da
import cupy as cp

add_broadcast_kernel = cp.RawKernel(
    r'''
    extern "C" __global__

pentschev

pentschev