def bench_file(filename):
"""
Perform comparative benchmark on matrices loaded from Matrix Market files.
:param filename: a Matrix Market file
"""
bs = dict()
for dtype in WANT_DTYPES:
print("# loading matrix from file ", filename)
lt = - rsb.rsb_time()
a = rsb.rsb_matrix(filename,dtype=dtype)
lt = lt + rsb.rsb_time()
printf("# loaded a matrix with %.1e nnz in %.1e s (%.1e nnz/s)\n",a.nnz,lt,a.nnz/lt)
printf("# loaded as type %s/%c (default is %s/%c)\n", a.dtype, DT2TC[a.dtype], rsb.rsb_dtype, DT2TC[rsb.rsb_dtype])
if not a._is_unsymmetric():
if WANT_SYMMETRIZE:
print("# NOTE: loaded RSB matrix is NOT unsymmetric: expanding symmetry to cope with csr.")
(I, J, V) = a.find()
a = rsb.rsb_matrix((np.append(V,V), (np.append(I,J), np.append(J,I))), a.shape, dtype=dtype)
else:
print("# NOTE: loaded RSB matrix is NOT unsymmetric, but scipy will only perform unsymmetric SpMM")
if a is not None:
(I, J, V) = a.find()
c = sp.sparse.csr_matrix((V, (I, J)))
( mtxname, _ ) = os.path.splitext(os.path.basename(filename))
( mtxname, _ ) = os.path.splitext(mtxname)
bd = bench_matrix(a, c, mtxname)
dict_stat_merge(bs, derived_bench_stats(bd))
bsc = bs.copy()
dict_sum_average_all(bsc)
print_perf_record(bsc,beg="pyrsb:speedups:",fields=True)
return bs
def bench(timeout, a, x, y):
iterations = 0
dt = -rsb.rsb_time()
while dt + rsb.rsb_time() < timeout:
iterations = iterations + 1
y = a * x # See __mul__
# a.spmm(x,y) # This form avoids the copy of y.
dt = dt + rsb.rsb_time()
op_dt = dt / iterations
return (op_dt, dt, iterations)
def bench_file(filename):
"""
Perform comparative benchmark on matrices loaded from Matrix Market files.
:param filename: a Matrix Market file
"""
print("# loading from file ", filename)
lt = - rsb.rsb_time()
a = rsb.rsb_file_mtx_load(bytes(filename, encoding="utf-8"))
lt = lt + rsb.rsb_time()
printf("# loaded a matrix with %.1e nnz in %.1e s (%.1e nnz/s)\n",a.nnz(),lt,a.nnz()/lt)
if not a._is_unsymmetric():
print("# NOTE: loaded RSB matrix is NOT unsymmetric, but scipy will only perform unsymmetric SpMM")
if a is not None:
(I, J, V) = a.find()
c = sp.sparse.csr_matrix((V, (I, J)))
bench_matrix(a, c)
def bench(timeout, a, x, y):
"""
Benchmark multiplication operation.
:param timeout: benchmark time
:param a: matrix
:param x: right hand side vector
:param y: result vector
:return: a tuple with operation time, benchmark time, performed iterations
"""
iterations = 0
dt = -rsb.rsb_time()
while dt + rsb.rsb_time() < timeout:
iterations = iterations + 1
y += a * x # See __mul__
# a.spmm(x,y) # This form avoids the copy of y.
dt = dt + rsb.rsb_time()
op_dt = dt / iterations
return (op_dt, dt, iterations)
def bench_random_files():
"""
Perform comparative benchmark on randomly generated matrices.
"""
for nrA in WANT_NRA:
ncA = nrA
dnst = (math.sqrt(1.0 * nrA)) / nrA
# print("# generating ",nrA,"x",ncA," with density ",dnst)
printf("# generating %d x %d with with density %.1e\n", nrA, ncA, dnst)
gt = -rsb.rsb_time()
c = sp.sparse.rand(nrA, ncA, density=dnst, format=WANT_PSF, dtype=sp.double)
gt = gt + rsb.rsb_time()
(I, J, V) = sp.sparse.find(c)
ct = -rsb.rsb_time()
a = rsb.rsb_matrix((V, (I, J)), [nrA, ncA])
ct = ct + rsb.rsb_time()
printf("# generated a matrix with %.1e nnz in %.1e s (%.1e nnz/s), converted to RSB in %.1e s\n",a.nnz(),gt,a.nnz()/gt,ct)
bench_matrix(a, c)
def bench_random_matrices():
"""
Perform comparative benchmark on randomly generated matrices.
"""
for dtype in WANT_DTYPES:
for nrA in WANT_NRA:
ncA = nrA
dnst = (math.sqrt(1.0 * nrA)) / nrA
# print("# generating ",nrA,"x",ncA," with density ",dnst)
printf("# generating %d x %d with with density %.1e\n", nrA, ncA, dnst)
gt = -rsb.rsb_time()
c = sp.sparse.rand(nrA, ncA, density=dnst, format=WANT_PSF, dtype=rsb.rsb_dtype)
gt = gt + rsb.rsb_time()
(I, J, V) = sp.sparse.find(c)
V = dtype(V)
c = sp.sparse.csr_matrix((V, (I, J)), [nrA, ncA])
ct = -rsb.rsb_time()
a = rsb.rsb_matrix((V, (I, J)), [nrA, ncA], dtype=dtype)
ct = ct + rsb.rsb_time()
printf("# generated a matrix with %.1e nnz in %.1e s (%.1e nnz/s), converted to RSB in %.1e s\n",a.nnz,gt,a.nnz/gt,ct)
bd = bench_matrix(a, c, "random")
derived_bench_stats(bd)
def bench(timeout, a, x, y):
"""
Benchmark multiplication operation.
:param timeout: benchmark time
:param a: matrix
:param x: right hand side vector
:param y: result vector
:return: a tuple with min operation time, benchmark time, performed iterations
"""
iterations = 0
if timeout > 0.0:
bench(0.0, a, x, y) # single op to warm-up caches
op_dt = float('+inf')
t0 = rsb.rsb_time()
t1 = t0
if WANT_ZERO_ALLOC:
if (isinstance(a,rsb.rsb_matrix)):
while t1 - t0 < timeout or iterations == 0:
iterations = iterations + 1
t2 = rsb.rsb_time()
a._spmm(x,y) # This form avoids the copy of y
t1 = rsb.rsb_time()
op_dt = min(op_dt,t1-t2)
else:
while t1 - t0 < timeout or iterations == 0:
iterations = iterations + 1
t2 = rsb.rsb_time()
# y += a._mul_multivector(x) # inefficient
sp.sparse._sparsetools.csr_matvecs(a.shape[0], a.shape[1], x.shape[1], a.indptr, a.indices, a.data, x.ravel(), y.ravel())
t1 = rsb.rsb_time()
op_dt = min(op_dt,t1-t2)
else:
while t1 - t0 < timeout or iterations == 0:
iterations = iterations + 1
t2 = rsb.rsb_time()
y += a * x # Inefficient (result created repeatedly) see __mul__
t1 = rsb.rsb_time()
op_dt = min(op_dt,t1-t2)
bt = rsb.rsb_time() - t0
return (op_dt, bt, iterations)
def bench_matrix(a, c, mtxname):
"""
Perform comparative benchmark: rsb vs csr.
:param a: rsb matrix
:param c: csr matrix
"""
brdict = {
'mtxname': mtxname,
'at_time': 0.0,
'nsubm': a.nsubm(),
'bpnz': a._idx_bpnz()
}
#tmax = 2
#tmax = 0.1
#tmax = -3
tmax = 0
bd = dict()
psf = WANT_PSF
if WANT_RENDER:
filename = sprintf("%s-%c.eps",mtxname,DT2TC[a.dtype])
a.render(filename)
for nrhs in WANT_NRHS:
bd[nrhs] = dict()
if WANT_AUTOTUNE == 0:
for nrhs in WANT_NRHS:
for order in WANT_ORDER:
(rsb_dt,psf_dt) = bench_both(a, c, psf, brdict, order, nrhs)
bd[nrhs][order] = bench_record(a, psf, brdict, order, nrhs, rsb_dt, psf_dt)
elif WANT_AUTOTUNE == 1:
o = a.copy()
if WANT_VERBOSE:
print("Will autotune matrix for SpMV ", a)
at_time = rsb.rsb_time()
o.autotune(verbose=WANT_VERBOSE_TUNING,tmax=tmax)
brdict['at_time'] = rsb.rsb_time() - at_time
if WANT_RENDER:
filename = sprintf("%s-%c-tuned.eps",mtxname,DT2TC[o.dtype])
o.render(filename)
for nrhs in WANT_NRHS:
for order in WANT_ORDER:
(rsb_dt,psf_dt) = bench_both(a, c, psf, brdict, order, nrhs)
(rsb_at_dt,psf_at_dt) = bench_both(o, c, psf, brdict, order, nrhs)
bd[nrhs][order] = bench_record(o, psf, brdict, order, nrhs, rsb_dt, psf_dt, rsb_at_dt, psf_at_dt)
del o
elif WANT_AUTOTUNE == 2:
for nrhs in WANT_NRHS:
for order in WANT_ORDER:
if WANT_VERBOSE:
print("Will autotune one matrix instance for different specific SpMM ", a)
(rsb_dt,psf_dt) = bench_both(a, c, psf, brdict, order, nrhs)
at_time = rsb.rsb_time()
a.autotune(verbose=WANT_VERBOSE_TUNING,nrhs=nrhs,order=order,tmax=tmax)
brdict['at_time'] = rsb.rsb_time() - at_time
if WANT_RENDER:
filename = sprintf("%s-%c-tuned-%c-%d.eps",mtxname,DT2TC[a.dtype],order,nrhs)
a.render(filename)
(rsb_at_dt,psf_at_dt) = bench_both(a, c, psf, brdict, order, nrhs)
bd[nrhs][order] = bench_record(a, psf, brdict, order, nrhs, rsb_dt, psf_dt, rsb_at_dt, psf_at_dt)
elif WANT_AUTOTUNE >= 3:
for nrhs in WANT_NRHS:
for order in WANT_ORDER:
(rsb_dt,psf_dt) = bench_both(a, c, psf, brdict, order, nrhs)
o = a.copy()
if WANT_VERBOSE:
print("Will autotune copies of starting matrix for specific SpMM ", a)
at_time = rsb.rsb_time()
for i in range(2,+WANT_AUTOTUNE):
o.autotune(verbose=WANT_VERBOSE_TUNING,nrhs=nrhs,order=order,tmax=tmax)
brdict['at_time'] = rsb.rsb_time() - at_time
if WANT_RENDER:
filename = sprintf("%s-%c-tuned-%c-%d.eps",mtxname,DT2TC[o.dtype],order,nrhs)
o.render(filename)
(rsb_at_dt,psf_at_dt) = bench_both(o, c, psf, brdict, order, nrhs)
bd[nrhs][order] = bench_record(o, psf, brdict, order, nrhs, rsb_dt, psf_dt, rsb_at_dt, psf_at_dt)
del o
del a
del c
return bd