COMPMID-2674: Buffer overrun in GEMM assembly tranform kernels.

Block transforms could access elements out of bounds in case of input
sizes smaller than the transform blocks. Fixes the edge cases by
short-circuiting the accesses in such cases.

Change-Id: I11d172ecd80b4dde46496e9d4b446de7fb9d5dc7
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1976
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michalis Spyrou <michalis.spyrou@arm.com>

1 files changed, 12 insertions, 2 deletions

diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_interleave_8way_half_to_float.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_interleave_8way_half_to_float.hpp
index de8e95a6d7..bde3274926 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_interleave_8way_half_to_float.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_interleave_8way_half_to_float.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2018 ARM Limited.
+ * Copyright (c) 2017-2019 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -34,6 +34,7 @@ template<>
 inline void TransformImpl<8, 1, false, 4, 2, false>::Transform(float *out, const __fp16 *in, int ldin, int y0, int ymax, int k0, int kmax) {
     float *outptr = out;
     const __fp16 *inptr = in;
+    bool first = true;
 
     __fp16 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, 2, false>::Transform(float *out, const
         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, 2, false>::Transform(float *out, const
                 }
             }
 
+            if (first) {
+                if (x<=7) {
+                    break;
+                }
+
+                first = false;
+            }
+
             __asm __volatile (
                 // Load up 8 elements (2 vectors) from each of 8 sources.
                 "LDR	q0, [%[inptr0]], #16\n"