add benchmarks for different spmv kernels

vbjax/sparse.py

@@ -7,7 +7,7 @@
 _to_jax = lambda x: jax.dlpack.from_dlpack(x.__dlpack__())
 
 
-def make_spmv(A, is_symmetric=False):
+def make_spmv(A, is_symmetric=False, use_scipy=False):
     """
     Make a closure for a general sparse matrix-vector multiplication.
 
@@ -17,6 +17,8 @@ def make_spmv(A, is_symmetric=False):
         Constant sparse matrix.
     is_symmetric : bool, optional, default False
         Whether matrix is symmetric.
+    use_scipy: bool, optional, default False
+        Use scipy.
 
     Returns
     -------

vbjax/tests/test_sparse.py

@@ -7,6 +7,7 @@
 import jax.experimental.sparse as jsp
 import scipy.sparse
 import vbjax
+import pytest
 
 
 def _test_spmv(spmv, A, n):
@@ -68,7 +69,49 @@ def bench_csr_to_bcoo():
             jb2 = f(jx)
         t1 = time.time()
         print(f'{name}: {t1 - t0:.3f} s')
-
 
-if __name__ == '__main__':
-    bench_csr_to_bcoo()
+
+def create_sample_data(n=1000, density_pct=10):
+    A = scipy.sparse.random(n, n, density=density_pct/100).tocsr()
+    jx = jax.numpy.r_[:n].astype('f')
+    return A, jx
+
+
+# some performance testing values
+_perf_args = 'n,density_pct,grad,impl,jit'
+_perf_values = [(1000, 10), (10_000, 0.02)]
+
+# we want to test each of the above values with grad on and off
+_perf_values = [vals + (flag, impl, jit)
+                for flag in (True, False)
+                for impl in 'scipy jaxbcoo'.split(' ')
+                for jit in (True, False)
+                for vals in _perf_values]
+
+@pytest.mark.parametrize(_perf_args, _perf_values)
+def test_perf_jbcoo(benchmark, n, density_pct, grad, impl, jit):
+    A, x = create_sample_data(n=n, density_pct=density_pct)
+
+    if impl == 'scipy': # TODO enum
+        spmv1 = vbjax.make_spmv(A)
+    elif impl == 'jaxbcoo':
+        jA = _csr_to_jax_bcoo(A)
+        spmv1 = lambda x: jA @ x
+    else:
+        raise ValueError(impl)
+    assert callable(spmv1)
+
+    if grad:
+        spmv2 = jax.grad(lambda x: jnp.sum(spmv1(x)))
+    else:
+        spmv2 = spmv1
+    assert callable(spmv2)
+
+    if jit and impl not in ('scipy', ):
+        spmv3 = jax.jit(spmv2)
+        spmv3(x)
+    else:
+        spmv3 = spmv2
+    assert callable(spmv3)
+
+    benchmark(lambda : spmv3(x))