1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102
|
function gbtest_perf2
%GBTEST_PERF2 test A'*x performance
% SuiteSparse:GraphBLAS, Timothy A. Davis, (c) 2017-2022, All Rights Reserved.
% SPDX-License-Identifier: Apache-2.0
max_nthreads = GrB.threads ;
threads = [1 2 4 8 16 20 32 40 64] ;
desc = struct ('in0', 'transpose') ;
rng ('default') ;
n = 1e6 ; nz = 20e6 ;
% n = 1e5 ; nz = 1e6 ;
d = nz / n^2 ;
% same as A = sprand (n,n,d), but faster:
G = GrB.random (n,n,d) ;
A = double (G) ;
% warmup to make sure the GrB library is loaded
y = GrB (rand (2)) * GrB (rand (2)) ;
degree = sum (spones (G)) ;
nempty = length (find (degree == 0)) ;
fprintf ('matrix: n: %d nnz: %d # empty cols: %d\n', n, nnz (A), nempty) ;
ntrials = 1 ;
for test = 1:4
if (test == 1)
X = 'sparse (rand (n,1))' ;
x = sparse (rand (n,1)) ;
elseif (test == 2)
X = 'rand (n,1)' ;
x = rand (n,1) ;
elseif (test == 3)
X = 'sprand (n,1,0.5)' ;
x = sprand (n,1,0.5) ;
else
X = 'sprand (n,1,0.05)' ;
x = sprand (n,1,0.05) ;
end
fprintf ('\n\n========================\n') ;
fprintf ('built-in: y = A''*x where x = %s\n', X) ;
tic
for trial = 1:ntrials
y = A'*x ;
end
tbuiltin = toc ;
fprintf ('built-in time: %8.4f sec\n', tbuiltin) ;
ybuiltin = y ;
fprintf ('\nGrB: y = A''*x where x = %s\n', X) ;
for nthreads = threads
if (nthreads > max_nthreads)
break ;
end
GrB.threads (nthreads) ;
tic
for trial = 1:ntrials
% y = G'*x ;
y = GrB.mxm (G, '+.*', x, desc) ;
end
t = toc ;
if (nthreads == 1)
t1 = t ;
end
fprintf (...
'threads: %2d GrB time: %8.4f speedup vs built-in: %8.2f vs: GrB(1 thread) %8.2f\n', ...
nthreads, t, tbuiltin / t, t1 / t) ;
assert (norm (y-ybuiltin, 1) / norm (ybuiltin,1) < 1e-12)
end
fprintf ('\nGrB: y = zeros(n,1) + A''*x where x = %s\n', X) ;
for nthreads = threads
if (nthreads > max_nthreads)
break ;
end
GrB.threads (nthreads) ;
tic
for trial = 1:ntrials
y = zeros (n,1) ;
% y = y + G'*x
y = GrB.mxm (y, '+', G, '+.*', x, desc) ;
end
t = toc ;
if (nthreads == 1)
t1 = t ;
end
fprintf (...
'threads: %2d GrB time: %8.4f speedup vs built-in: %8.2f vs: GrB(1 thread) %8.2f\n', ...
nthreads, t, tbuiltin / t, t1 / t) ;
assert (norm (y-ybuiltin, 1) / norm (ybuiltin,1) < 1e-12)
end
end
GrB.burble (0) ;
|