3 files changed, 21 insertions, 2 deletions

diff --git a/src/solvers/h48/solve.h b/src/solvers/h48/solve.h
index c51fa29..6a29b82 100644
--- a/src/solvers/h48/solve.h
+++ b/src/solvers/h48/solve.h
@@ -138,6 +138,7 @@ h48_prune_pipeline(
         bool normal
 )
 {
+        uint64_t i;
         uint32_t cdata;
         uint8_t m, p;
 
@@ -184,7 +185,8 @@ h48_prune_pipeline(
                 if (prune[m].pi == 0) {
                         prune[m].coord = coord_h48_edges(prune[m].inverse,
                             COCLASS(cdata), TTREP(cdata), arg->h);
-                        // TODO prefetch
+                        i = H48_INDEX(H48_LINE_EXT(prune[m].coord));
+                        prefetch(arg->h48data, i);
                 }
         }
 
@@ -210,7 +212,8 @@ h48_prune_pipeline(
                 }
                 prune[m].coord = coord_h48_edges(
                     prune[m].cube, COCLASS(cdata), TTREP(cdata), arg->h);
-                // TODO prefetch
+                i = H48_INDEX(H48_LINE_EXT(prune[m].coord));
+                prefetch(arg->h48data, i);
         }
 
         /* Stage 3: get pval from normal */
diff --git a/src/utils/prefetch.h b/src/utils/prefetch.h
new file mode 100644
index 0000000..89b14e8
--- /dev/null
+++ b/src/utils/prefetch.h
@@ -0,0 +1,15 @@
+#if defined(AVX2)
+
+#define prefetch(a, i) _mm_prefetch(a+i, _MM_HINT_T0)
+
+#else
+#if defined(__GNUC__) || defined(__clang__)
+
+#define prefetch(a, i) __builtin_prefetch(a+i, 0, 0)
+
+#else
+
+#define prefetch(a, i) (void)i
+
+#endif
+#endif
diff --git a/src/utils/utils.h b/src/utils/utils.h
index 383781a..e322f20 100644
--- a/src/utils/utils.h
+++ b/src/utils/utils.h
@@ -4,3 +4,4 @@
 #include "math.h"
 #include "sleep.h"
 #include "wrapthread.h"
+#include "prefetch.h"