diff --git a/sumpy/e2e.py b/sumpy/e2e.py
index 66e5e97567d8675834c54f7291dba7d49bd20b19..bff0f508c05ea378fb5c7cd31ce85f82c0ffdfcd 100644
--- a/sumpy/e2e.py
+++ b/sumpy/e2e.py
@@ -63,6 +63,22 @@ class E2EBase(KernelCacheWrapper):
         if device is None:
             device = ctx.devices[0]
 
+        if src_expansion is tgt_expansion:
+            from sumpy.kernel import TargetDerivativeRemover, SourceDerivativeRemover
+            tgt_expansion = src_expansion = src_expansion.with_kernel(
+                    SourceDerivativeRemover()(
+                        TargetDerivativeRemover()(src_expansion.kernel)))
+
+        else:
+
+            from sumpy.kernel import TargetDerivativeRemover, SourceDerivativeRemover
+            src_expansion = src_expansion.with_kernel(
+                    SourceDerivativeRemover()(
+                        TargetDerivativeRemover()(src_expansion.kernel)))
+            tgt_expansion = tgt_expansion.with_kernel(
+                    SourceDerivativeRemover()(
+                        TargetDerivativeRemover()(tgt_expansion.kernel)))
+
         self.ctx = ctx
         self.src_expansion = src_expansion
         self.tgt_expansion = tgt_expansion
diff --git a/sumpy/e2p.py b/sumpy/e2p.py
index 6a77a5cc416af3c01df9786bac5fd236927e787c..9a2a8bc1ea564e37cf4e2d248de6952cfd386067 100644
--- a/sumpy/e2p.py
+++ b/sumpy/e2p.py
@@ -59,6 +59,15 @@ class E2PBase(KernelCacheWrapper):
         if device is None:
             device = ctx.devices[0]
 
+        from sumpy.kernel import SourceDerivativeRemover, TargetDerivativeRemover
+        sdr = SourceDerivativeRemover()
+        tdr = TargetDerivativeRemover()
+        expansion = expansion.with_kernel(
+                sdr(expansion.kernel))
+
+        for knl in kernels:
+            assert sdr(tdr(knl)) == expansion.kernel
+
         self.ctx = ctx
         self.expansion = expansion
         self.kernels = kernels
@@ -68,11 +77,6 @@ class E2PBase(KernelCacheWrapper):
 
         self.dim = expansion.dim
 
-        from sumpy.kernel import TargetDerivativeRemover
-        tdr = TargetDerivativeRemover()
-        for knl in kernels:
-            assert tdr(knl) == expansion.kernel
-
     def get_loopy_insns_and_result_names(self):
         from sumpy.symbolic import make_sym_vector
         bvec = make_sym_vector("b", self.dim)
diff --git a/sumpy/expansion/__init__.py b/sumpy/expansion/__init__.py
index 08754a10240061ac9858034a561fb301c7db6925..b5096e0c572af3bc9e779976d3368d6c377cf5d0 100644
--- a/sumpy/expansion/__init__.py
+++ b/sumpy/expansion/__init__.py
@@ -83,6 +83,10 @@ class ExpansionBase(object):
 
     # }}}
 
+    def with_kernel(self, kernel):
+        return type(self)(kernel, self.order)
+        # FIXME: add self.use_rscale once the rscale MR is in
+
     def __len__(self):
         return len(self.get_coefficient_identifiers())
 
diff --git a/sumpy/fmm.py b/sumpy/fmm.py
index b20c9827e46e86e98ab81ab2c4c2822ee1163cd6..77226a9a4df94d72f6c95cad39ecd8e103bd486c 100644
--- a/sumpy/fmm.py
+++ b/sumpy/fmm.py
@@ -517,7 +517,7 @@ class SumpyExpansionWrangler(object):
             p2l = self.code.p2l(self.level_orders[lev])
 
             target_level_start_ibox, target_local_exps_view = \
-                    self.multipole_expansions_view(local_exps, lev)
+                    self.local_expansions_view(local_exps, lev)
 
             evt, (result,) = p2l(
                     self.level_queues[lev],
diff --git a/sumpy/kernel.py b/sumpy/kernel.py
index aed3d931d1fa18cb78825de0ad9ab5c9760c9ea7..780cf8eb4dd2eb482ea0492779232bbff92c75c3 100644
--- a/sumpy/kernel.py
+++ b/sumpy/kernel.py
@@ -902,6 +902,11 @@ class TargetDerivativeRemover(AxisTargetDerivativeRemover):
         return self.rec(kernel.inner_kernel)
 
 
+class SourceDerivativeRemover(KernelIdentityMapper):
+    def map_directional_source_derivative(self, kernel):
+        return self.rec(kernel.inner_kernel)
+
+
 class DerivativeCounter(KernelCombineMapper):
     def combine(self, values):
         return max(values)
diff --git a/sumpy/p2e.py b/sumpy/p2e.py
index 6a8fcfa9d2ea8dd7a034d0adb1a7cc152c8db7bc..b8c60a6423bc1dd704cf63cf915831c084480b7b 100644
--- a/sumpy/p2e.py
+++ b/sumpy/p2e.py
@@ -61,6 +61,10 @@ class P2EBase(KernelCacheWrapper):
         if device is None:
             device = ctx.devices[0]
 
+        from sumpy.kernel import TargetDerivativeRemover
+        expansion = expansion.with_kernel(
+                TargetDerivativeRemover()(expansion.kernel))
+
         self.ctx = ctx
         self.expansion = expansion
         self.options = options
diff --git a/test/test_fmm.py b/test/test_fmm.py
index 0d8f36d81979de64103c50f29ac21ecc46625abf..c1df96a6d65ec147eb523d1326b887f8521aeae8 100644
--- a/test/test_fmm.py
+++ b/test/test_fmm.py
@@ -169,7 +169,7 @@ def test_sumpy_fmm(ctx_getter, knl, local_expn_class, mpole_expn_class):
     # }}}
 
     from pyopencl.clrandom import PhiloxGenerator
-    rng = PhiloxGenerator(ctx)
+    rng = PhiloxGenerator(ctx, seed=44)
     weights = rng.uniform(queue, nsources, dtype=np.float64)
 
     logger.info("computing direct (reference) result")