diff --git a/sumpy/codegen.py b/sumpy/codegen.py
index fa13d52696652f17d0be6a404ea9bf67f624d0dc..3b19fcb9ef7c821399cc3cf830cb8fe198625c6d 100644
--- a/sumpy/codegen.py
+++ b/sumpy/codegen.py
@@ -208,6 +208,20 @@ def register_bessel_callables(loopy_knl):
             Hankel1_01("hank1_01"))
     return loopy_knl
 
+
+def _fp_contract_fast_preamble(preamble_info):
+    yield ("fp_contract_fast_pocl", "#pragma clang fp contract(fast)")
+
+
+def register_optimization_preambles(loopy_knl, device):
+    if isinstance(loopy_knl.target, lp.PyOpenCLTarget):
+        import pyopencl as cl
+        if device.platform.name == "Portable Computing Language" and \
+                (device.type & cl.device_type.GPU):
+            loopy_knl = lp.register_preamble_generators(loopy_knl,
+                [_fp_contract_fast_preamble])
+    return loopy_knl
+
 # }}}
 
 
diff --git a/sumpy/p2p.py b/sumpy/p2p.py
index 4be5916da17f041134e50c2015d607128074a0fd..d03d8bd1f80c5e4ca3f9b870c41308f8cd5d3fb4 100644
--- a/sumpy/p2p.py
+++ b/sumpy/p2p.py
@@ -190,6 +190,9 @@ class P2PBase(KernelCacheMixin, KernelComputation):
         knl = lp.set_options(knl,
                 enforce_variable_access_ordered="no_check")
 
+        from sumpy.codegen import register_optimization_preambles
+        knl = register_optimization_preambles(knl, self.device)
+
         return knl
 
 
@@ -714,6 +717,9 @@ class P2PFromCSR(P2PBase):
         knl = lp.set_options(knl,
                 enforce_variable_access_ordered="no_check")
 
+        from sumpy.codegen import register_optimization_preambles
+        knl = register_optimization_preambles(knl, self.device)
+
         return knl
 
     def __call__(self, queue, **kwargs):