def test_axpy():
print("==== Program start ====")
N.set(24)
print('Scalar-vector multiplication %d' % (N.get()))
# Initialize arrays: Randomize A and X, zero Y
A = dace.float64(np.random.rand())
X = np.random.rand(N.get()).astype(np.float64)
Y = np.random.rand(N.get()).astype(np.float64)
A_regression = np.float64()
X_regression = np.ndarray([N.get()], dtype=np.float64)
Y_regression = np.ndarray([N.get()], dtype=np.float64)
A_regression = A
X_regression[:] = X[:]
Y_regression[:] = Y[:]
sdfg = common.vectorize(axpy)
sdfg(A=A, X=X, Y=Y, N=N)
c_axpy = sp.linalg.blas.get_blas_funcs('axpy',
arrays=(X_regression, Y_regression))
if dace.Config.get_bool('profiling'):
dace.timethis('axpy', 'BLAS', (2 * N.get()), c_axpy, X_regression,
Y_regression, N.get(), A_regression)
else:
c_axpy(X_regression, Y_regression, N.get(), A_regression)
diff = np.linalg.norm(Y_regression - Y) / N.get()
print("Difference:", diff)
print("==== Program end ====")
assert diff <= 1e-5
def test_spmv():
W.set(64)
H.set(64)
nnz.set(640)
print('Sparse Matrix-Vector Multiplication %dx%d (%d non-zero elements)' %
(W.get(), H.get(), nnz.get()))
A_row = dace.ndarray([H + 1], dtype=dace.uint32)
A_col = dace.ndarray([nnz], dtype=dace.uint32)
A_val = dace.ndarray([nnz], dtype=dace.float32)
x = dace.ndarray([W], dace.float32)
b = dace.ndarray([H], dace.float32)
# Assuming uniform sparsity distribution across rows
nnz_per_row = nnz.get() // H.get()
nnz_last_row = nnz_per_row + (nnz.get() % H.get())
if nnz_last_row > W.get():
print('Too many nonzeros per row')
exit(1)
# RANDOMIZE SPARSE MATRIX
A_row[0] = dace.uint32(0)
A_row[1:H.get()] = dace.uint32(nnz_per_row)
A_row[-1] = dace.uint32(nnz_last_row)
A_row = np.cumsum(A_row, dtype=np.uint32)
# Fill column data
for i in range(H.get() - 1):
A_col[nnz_per_row*i:nnz_per_row*(i+1)] = \
np.sort(np.random.choice(W.get(), nnz_per_row, replace=False))
# Fill column data for last row
A_col[nnz_per_row * (H.get() - 1):] = np.sort(
np.random.choice(W.get(), nnz_last_row, replace=False))
A_val[:] = np.random.rand(nnz.get()).astype(dace.float32.type)
#########################
x[:] = np.random.rand(W.get()).astype(dace.float32.type)
b[:] = dace.float32(0)
# Setup regression
A_sparse = scipy.sparse.csr_matrix((A_val, A_col, A_row),
shape=(H.get(), W.get()))
sdfg = spmv.to_sdfg()
vectorize(sdfg, 'j')
sdfg(A_row=A_row, A_col=A_col, A_val=A_val, x=x, b=b, H=H, W=W, nnz=nnz)
if dace.Config.get_bool('profiling'):
dace.timethis('spmv', 'scipy', 0, A_sparse.dot, x)
diff = np.linalg.norm(A_sparse.dot(x) - b) / float(H.get())
print("Difference:", diff)
print("==== Program end ====")
assert diff <= 1e-5
def test_filter():
N.set(64)
ratio = np.float32(0.5)
print('Predicate-Based Filter. size=%d, ratio=%f' % (N.get(), ratio))
A = np.random.rand(N.get()).astype(np.float32)
B = np.zeros_like(A)
outsize = dace.scalar(dace.uint32)
outsize[0] = 0
sdfg = pbf.to_sdfg()
vectorize(sdfg, 'i')
sdfg(A=A, B=B, outsz=outsize, ratio=ratio, N=N)
if dace.Config.get_bool('profiling'):
dace.timethis('filter', 'numpy', 0, regression, A, ratio)
filtered = regression(A, ratio)
if len(filtered) != outsize[0]:
print(
"Difference in number of filtered items: %d (DaCe) vs. %d (numpy)"
% (outsize[0], len(filtered)))
totalitems = min(outsize[0], N.get())
print('DaCe:', B[:totalitems].view(type=np.ndarray))
print('Regression:', filtered.view(type=np.ndarray))
exit(1)
# Sort the outputs
filtered = np.sort(filtered)
B[:outsize[0]] = np.sort(B[:outsize[0]])
if len(filtered) == 0:
print("==== Program end ====")
exit(0)
diff = np.linalg.norm(filtered - B[:outsize[0]]) / float(outsize[0])
print("Difference:", diff)
if diff > 1e-5:
totalitems = min(outsize[0], N.get())
print('DaCe:', B[:totalitems].view(type=np.ndarray))
print('Regression:', filtered.view(type=np.ndarray))
print("==== Program end ====")
assert diff <= 1e-5
#########################
x[:] = np.random.rand(W.get()).astype(dtype.type)
#b[:] = dtype(0)
# Setup regression
A_sparse = scipy.sparse.csr_matrix((A_val, A_col, A_row),
shape=(H.get(), W.get()))
spmv = make_sdfg(args["specialize"])
if args["specialize"]:
spmv.specialize(dict(H=H, W=W, nnz=nnz))
spmv(A_row=A_row, A_col=A_col, A_val=A_val, x=x, b=b, H=H, W=W, nnz=nnz)
if dace.Config.get_bool('profiling'):
dace.timethis('spmv', 'scipy', 0, A_sparse.dot, x)
diff = np.linalg.norm(A_sparse.dot(x) - b) / float(H.get())
print("Difference:", diff)
if diff >= 1e-5:
print("Validation failed.")
print("Result:")
print(b)
print("Reference:")
print(A_sparse.dot(x))
print("Type \"debug\" to enter debugger, "
"or any other string to quit (timeout in 10 seconds)")
read, _, _ = select.select([sys.stdin], [], [], 10)
if len(read) > 0 and sys.stdin.readline().strip().lower() == "debug":
print("Entering debugger...")
pdb.set_trace()
print('Matrix addition %dx%d' % (M.get(), N.get()))
# Initialize arrays: Randomize A and B, zero C
A[:] = np.random.rand(M.get(), N.get()).astype(dace.float64.type)
B[:] = np.random.rand(M.get(), N.get()).astype(dace.float64.type)
C[:] = dace.float64(0)
A_regression = np.ndarray([M.get(), N.get()], dtype=np.float64)
B_regression = np.ndarray([M.get(), N.get()], dtype=np.float64)
C_regression = np.ndarray([M.get(), N.get()], dtype=np.float64)
A_regression[:] = A[:]
B_regression[:] = B[:]
C_regression[:] = C[:]
if args["compile-only"]:
dace.compile(mat_add, A, B, C)
else:
mat_add(A, B, C)
if dace.Config.get_bool('profiling'):
dace.timethis('mat_add', contender, dace.eval(2 * N * N * N),
np.dot, A_regression, B_regression, C_regression)
else:
np.add(A_regression, B_regression, C_regression)
diff = np.linalg.norm(C_regression - C) / float(dace.eval(M * N))
print("Difference:", diff)
print("==== Program end ====")
exit(0 if diff <= 1e-5 else 1)
a << A[i, j]
out >> hist(1, lambda x, y: x + y)[:]
out[min(int(a * BINS), BINS - 1)] = 1
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("W", type=int, nargs="?", default=32)
parser.add_argument("H", type=int, nargs="?", default=32)
args = vars(parser.parse_args())
W.set(args["W"])
H.set(args["H"])
print('Histogram %dx%d' % (W.get(), H.get()))
A = np.random.rand(H.get(), W.get()).astype(dace.float32.type)
hist = np.zeros([BINS], dtype=np.uint32)
histogram(A, hist)
if dace.Config.get_bool('profiling'):
dace.timethis('histogram', 'numpy', (H.get() * W.get()), np.histogram, A, BINS)
diff = np.linalg.norm(np.histogram(A, bins=BINS, range=(0.0, 1.0))[0][1:-1] - hist[1:-1])
print("Difference:", diff)
print("==== Program end ====")
exit(0 if diff <= 1e-5 else 1)
b >> B[i, j]
b = a
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("W", type=int, nargs="?", default=64)
parser.add_argument("H", type=int, nargs="?", default=64)
args = vars(parser.parse_args())
W.set(args["W"])
H.set(args["H"])
print('Transpose %dx%d' % (W.get(), H.get()))
A = np.random.rand(H.get(), W.get()).astype(np.float32)
B = np.zeros([W.get(), H.get()], dtype=np.float32)
transpose(A, B)
if dace.Config.get_bool('profiling'):
dace.timethis('transpose', 'numpy', (H.get() * W.get()), np.transpose,
A)
diff = np.linalg.norm(np.transpose(A) - B) / (H.get() * W.get())
print("Difference:", diff)
print("==== Program end ====")
exit(0 if diff <= 1e-5 else 1)
N.set(args["N"])
print('Scalar-vector multiplication %d' % (N.get()))
# Initialize arrays: Randomize A and X, zero Y
A = dace.float64(np.random.rand())
X = np.random.rand(N.get()).astype(np.float64)
Y = np.random.rand(N.get()).astype(np.float64)
A_regression = np.float64()
X_regression = np.ndarray([N.get()], dtype=np.float64)
Y_regression = np.ndarray([N.get()], dtype=np.float64)
A_regression = A
X_regression[:] = X[:]
Y_regression[:] = Y[:]
axpy(A, X, Y)
c_axpy = sp.linalg.blas.get_blas_funcs('axpy',
arrays=(X_regression, Y_regression))
if dace.Config.get_bool('profiling'):
dace.timethis('axpy', 'BLAS', dace.eval(2 * N), c_axpy, X_regression,
Y_regression, N.get(), A_regression)
else:
c_axpy(X_regression, Y_regression, N.get(), A_regression)
diff = np.linalg.norm(Y_regression - Y) / float(dace.eval(N))
print("Difference:", diff)
print("==== Program end ====")
exit(0 if diff <= 1e-5 else 1)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("W", type=int, nargs="?", default=32)
parser.add_argument("H", type=int, nargs="?", default=32)
args = vars(parser.parse_args())
A = dace.ndarray([H, W], dtype=dace.uint8)
hist = dace.ndarray([BINS], dtype=dace.uint32)
W.set(args["W"])
H.set(args["H"])
print('Histogram (dec) %dx%d' % (W.get(), H.get()))
A[:] = np.random.randint(0, 256,
(H.get(), W.get())).astype(dace.uint8.type)
hist[:] = dace.uint32(0)
histogram(A, hist)
if dace.Config.get_bool('profiling'):
dace.timethis('histogram', 'numpy', dace.eval(H * W), np.histogram, A,
BINS)
diff = np.linalg.norm(np.histogram(A, bins=BINS)[0] - hist)
print("Difference:", diff)
print("==== Program end ====")
exit(0 if diff <= 1e-5 else 1)
histogram.draw_to_file()
A = dace.ndarray([H, W], dtype=dtype)
hist = dace.ndarray([num_bins], dtype=itype)
A[:] = np.random.rand(H.get(), W.get()).astype(dace.float32.type)
hist[:] = itype(0)
if args["specialize"]:
histogram(A=A, hist=hist)
else:
histogram(A=A, H=H, W=W, hist=hist)
if dace.Config.get_bool('profiling'):
dace.timethis('histogram', 'numpy', dace.eval(H * W), np.histogram, A,
num_bins)
ref = np.histogram(A, bins=num_bins.get(), range=(0.0, 1.0))[0]
diff = np.linalg.norm(ref[1:-1] - hist[1:-1])
print("Difference:", diff)
if diff > 1e-5:
print("** Kernel")
print(hist)
print("** Reference")
print(ref)
print("Validation failed.")
print("==== Program end ====")
exit(0 if diff <= 1e-5 else 1)
def run_spmv(size_w, size_h, num_nonzero, specialize):
cols.set(size_w)
rows.set(size_h)
nnz.set(num_nonzero)
print("Sparse Matrix-Vector Multiplication {}x{} "
"({} non-zero elements, {}specialized)".format(
cols.get(), rows.get(), nnz.get(),
"not " if not specialize else ""))
A_row = dace.ndarray([rows + 1], dtype=itype)
A_col = dace.ndarray([nnz], dtype=itype)
A_val = dace.ndarray([nnz], dtype=dtype)
x = dace.ndarray([cols], dtype)
b = dace.ndarray([rows], dtype)
# Assuming uniform sparsity distribution across rows
nnz_per_row = nnz.get() // rows.get()
nnz_last_row = nnz_per_row + (nnz.get() % rows.get())
if nnz_last_row > cols.get():
print("Too many nonzeros per row")
exit(1)
# RANDOMIZE SPARSE MATRIX
A_row[0] = itype(0)
A_row[1:rows.get()] = itype(nnz_per_row)
A_row[-1] = itype(nnz_last_row)
A_row = np.cumsum(A_row, dtype=itype.type)
# Fill column data
for i in range(rows.get() - 1):
A_col[nnz_per_row*i:nnz_per_row*(i+1)] = \
np.sort(np.random.choice(cols.get(), nnz_per_row, replace=False))
# Fill column data for last row
A_col[nnz_per_row * (rows.get() - 1):] = np.sort(
np.random.choice(cols.get(), nnz_last_row, replace=False))
A_val[:] = np.random.rand(nnz.get()).astype(dtype.type)
#########################
x[:] = np.random.rand(cols.get()).astype(dtype.type)
#b[:] = dtype(0)
# Setup regression
A_sparse = scipy.sparse.csr_matrix((A_val, A_col, A_row),
shape=(rows.get(), cols.get()))
spmv = make_sdfg(specialize)
if specialize:
spmv.specialize(dict(rows=rows, cols=cols, nnz=nnz))
spmv(A_row=A_row,
A_col=A_col,
A_val=A_val,
x=x,
b=b,
rows=rows,
cols=cols,
nnz=nnz)
if dace.Config.get_bool("profiling"):
dace.timethis("spmv", "scipy", 0, A_sparse.dot, x)
diff = np.linalg.norm(A_sparse.dot(x) - b) / float(rows.get())
print("Difference:", diff)
if diff >= 1e-5:
print("Validation failed.")
print("Result:")
print(b)
print("Reference:")
print(A_sparse.dot(x))
print("Type \"debug\" to enter debugger, "
"or any other string to quit (timeout in 10 seconds)")
read, _, _ = select.select([sys.stdin], [], [], 10)
if len(read) > 0 and sys.stdin.readline().strip().lower() == "debug":
print("Entering debugger...")
pdb.set_trace()
else:
print("Exiting...")
print("==== Program end ====")
if diff > 1e-5:
raise RuntimeError("Validation failed.")
return spmv
n = r ** k
print('FFT on vector of length %d' % n)
N.set(n)
R.set(r)
K.set(k)
X = np.random.rand(n).astype(np.complex64) + 1j * np.random.rand(n).astype(np.complex64)
Y_own = np.zeros_like(X, dtype=np.complex64)
own_fft(X, Y_own, N=N, K=K, R=R)
Y_np = fft(X)
if dace.Config.get_bool('profiling'):
dace.timethis('FFT', 'numpy_fft', (n**2), fft, X)
print("\n### RESULTS ###")
print("X:", np.array_str(X, precision=2, suppress_small=True))
print("Y_own: ", np.array_str(Y_own, precision=10, suppress_small=True))
print("Y_np: ", np.array_str(Y_np, precision=10, suppress_small=True))
eps = 1e-10
diff = np.linalg.norm(Y_own - Y_np) / n
if diff < eps:
print("Difference:", diff)
else:
print("\033[91mDifference:", diff, "\033[0m")
print("==== Program end ====")
exit(0 if diff <= 1e-3 else 1)
b = a
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("W", type=int, nargs="?", default=64)
parser.add_argument("H", type=int, nargs="?", default=64)
args = vars(parser.parse_args())
A = dace.ndarray([H, W], dtype=dace.float32)
B = dace.ndarray([H, W], dtype=dace.float32)
W.set(args["W"])
H.set(args["H"])
print('Transpose %dx%d' % (W.get(), H.get()))
A[:] = np.random.rand(H.get(), W.get()).astype(dace.float32.type)
B[:] = dace.float32(0)
transpose(A, B)
if dace.Config.get_bool('profiling'):
dace.timethis('transpose', 'numpy', dace.eval(H * W), np.transpose, A)
diff = np.linalg.norm(np.transpose(A) - B) / float(dace.eval(H * W))
print("Difference:", diff)
print("==== Program end ====")
exit(0 if diff <= 1e-5 else 1)
gemm = gemm.to_sdfg()
from dace.transformation.interstate import GPUTransformSDFG
gemm.apply_transformations(GPUTransformSDFG,
options={'sequential_innermaps': False},
validate=True,
validate_all=False,
strict=True)
gemm.validate()
for m in [k for k, v in gemm.node(0).scope_dict().items() if isinstance(k, dace.nodes.MapEntry) and v is None]:
m.map.schedule = dace.ScheduleType.GPU_Persistent
for m in [k for k, v in gemm.node(0).scope_dict().items() if isinstance(k, dace.nodes.MapEntry) and v is not None]:
m.map.schedule = dace.ScheduleType.GPU_Device
gemm(A, B, C)
if dace.Config.get_bool('profiling'):
dace.timethis('gemm', 'numpy', (2 * M.get() * K.get() * N.get()),
np.dot, A, B, C_regression)
else:
np.dot(A, B, C_regression)
diff = np.linalg.norm(C_regression - C) / (M.get() * N.get())
print("Difference:", diff)
print("==== Program end ====")
exit(0 if diff <= 1e-5 else 1)
# Initialize arrays: Randomize A and X, zero Y
A = dace.float64(np.random.rand())
X[:] = np.random.rand(N.get()).astype(dace.float64.type)
Y[:] = np.random.rand(N.get()).astype(dace.float64.type)
A_regression = np.float64()
X_regression = np.ndarray([N.get()], dtype=np.float64)
Y_regression = np.ndarray([N.get()], dtype=np.float64)
A_regression = A
X_regression[:] = X[:]
Y_regression[:] = Y[:]
if args["compile-only"]:
dace.compile(axpy, A, X, Y)
else:
axpy(A, X, Y)
c_axpy = sp.linalg.blas.get_blas_funcs('axpy',
arrays=(X_regression,
Y_regression))
if dace.Config.get_bool('profiling'):
dace.timethis('axpy', contender, dace.eval(2 * N), c_axpy,
X_regression, Y_regression, N.get(), A_regression)
else:
c_axpy(X_regression, Y_regression, N.get(), A_regression)
diff = np.linalg.norm(Y_regression - Y) / float(dace.eval(N))
print("Difference:", diff)
print("==== Program end ====")
exit(0 if diff <= 1e-5 else 1)
parser = argparse.ArgumentParser()
parser.add_argument("M", type=int, nargs="?", default=24)
parser.add_argument("K", type=int, nargs="?", default=24)
parser.add_argument("N", type=int, nargs="?", default=24)
args = vars(parser.parse_args())
M.set(args["M"])
K.set(args["K"])
N.set(args["N"])
print('Matrix multiplication %dx%dx%d' % (M.get(), K.get(), N.get()))
# Initialize arrays: Randomize A and B, zero C
A = np.random.rand(M.get(), K.get()).astype(np.float64)
B = np.random.rand(K.get(), N.get()).astype(np.float64)
C = np.zeros([M.get(), N.get()], dtype=np.float64)
C_regression = np.zeros_like(C)
gemm(A, B, C)
if dace.Config.get_bool('profiling'):
dace.timethis('gemm', 'numpy', dace.eval(2 * M * K * N), np.dot, A, B,
C_regression)
else:
np.dot(A, B, C_regression)
diff = np.linalg.norm(C_regression - C) / float(dace.eval(M * N))
print("Difference:", diff)
print("==== Program end ====")
exit(0 if diff <= 1e-5 else 1)
print("Predicate-Based Filter. size={}, ratio={} ({}specialized)".format(
N.get(), ratio, "" if args["specialize"] else "not "))
A[:] = np.random.rand(N.get()).astype(dace.float32.type)
B[:] = dace.float32(0)
sdfg = make_sdfg(args["specialize"])
sdfg.draw_to_file()
if args["specialize"]:
sdfg.specialize(dict(N=N))
sdfg(A=A, B=B, outsize=outsize, ratio=ratio)
else:
sdfg(A=A, B=B, outsize=outsize, ratio=ratio, N=N)
if dace.Config.get_bool('profiling'):
dace.timethis('filter', 'numpy', 0, regression, A, ratio)
filtered = regression(A, ratio)
if len(filtered) != outsize[0]:
print(
"Difference in number of filtered items: %d (DaCe) vs. %d (numpy)"
% (outsize[0], len(filtered)))
totalitems = min(outsize[0], N.get())
print('DaCe:', B[:totalitems].view(type=np.ndarray))
print('Regression:', filtered.view(type=np.ndarray))
exit(1)
# Sort the outputs
filtered = np.sort(filtered)
B[:outsize[0]] = np.sort(B[:outsize[0]])
W.set(args["W"])
print("Histogram {}x{} ({}specialized)".format(
H.get(), W.get(), "" if args["specialize"] else "not "))
A = dace.ndarray([H, W], dtype=dtype)
hist = dace.ndarray([num_bins], dtype=dace.uint32)
A[:] = np.random.rand(H.get(), W.get()).astype(dace.float32.type)
hist[:] = dace.uint32(0)
if args["specialize"]:
histogram(A=A, hist=hist)
else:
histogram(A=A, H=H, W=W, hist=hist)
if dace.Config.get_bool('profiling'):
dace.timethis('histogram', 'numpy', (H.get() * W.get()), np.histogram,
A, num_bins)
diff = np.linalg.norm(
np.histogram(A, bins=num_bins.get(), range=(0.0, 1.0))[0][1:-1] -
hist[1:-1])
print("Difference:", diff)
if diff > 1e-5:
print("Validation failed.")
print("==== Program end ====")
exit(0 if diff <= 1e-5 else 1)