1 files changed, 11 insertions, 1 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_interleave_8way_32bit.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_interleave_8way_32bit.hpp
index 758c084a46..9dfc1346e6 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_interleave_8way_32bit.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_interleave_8way_32bit.hpp
@@ -34,6 +34,7 @@ template<typename T>
 inline void TransformImpl<8, 1, false, 4, 4, false>::Transform(T *out, const T *in, int ldin, int y0, int ymax, int k0, int kmax) {
     uint32_t *outptr = (uint32_t *)out;
     const uint32_t *inptr = (uint32_t *)in;
+    bool first = true;
 
     uint32_t zerobuff[16] = { 0 }; // 8 for asm loop plus up to 7 for overflow loop
 
@@ -57,8 +58,9 @@ inline void TransformImpl<8, 1, false, 4, 4, false>::Transform(T *out, const T *
         prefetch_2x(inptr7);
 
         int x=(kmax-k0);
-        for (;x>7;x-=8) {
+        for (;(x>7) || first;x-=8) {
             /* Cope with ragged cases by copying from a buffer of zeroes instead */
+            /* 'first' forces this to always run at least once, needed if the total size is <=7. */
             if ((y + 7) >= ymax) {
                 switch ((y + 7) - ymax) {
                     /* Everything falls through in here */
@@ -89,6 +91,14 @@ inline void TransformImpl<8, 1, false, 4, 4, false>::Transform(T *out, const T *
                 }
             }
 
+            if (first) {
+                if (x<=7) {
+                    break;
+                }
+
+                first = false;
+            }
+
             __asm __volatile (
                 // Load up 8 elements (2 vectors) from each of 8 sources.
                 "LDP        q0, q1, [%[inptr0]], #32\n" // q0=A0A1A2A3