diff --git a/ruy/kernel_common.h b/ruy/kernel_common.h
index 69e819b..d4cf293 100644
--- a/ruy/kernel_common.h
+++ b/ruy/kernel_common.h
@@ -177,7 +177,13 @@
   params->rhs_zero_point = rhs.zero_point;
   params->dst_zero_point = dst->zero_point;
   params->depth = depth;
-  params->prod_zp_depth = lhs.zero_point * rhs.zero_point * depth;
+  // prod_zp_depth intentionally relies on two's-complement wraparound to match
+  // the wrapping arithmetic performed by the asm kernels. Compute it in unsigned
+  // so the overflow is well-defined.
+  params->prod_zp_depth = static_cast<std::int32_t>(
+      static_cast<std::uint32_t>(lhs.zero_point) *
+      static_cast<std::uint32_t>(rhs.zero_point) *
+      static_cast<std::uint32_t>(depth));
   params->flags |= RUY_ASM_FLAG_NEEDS_LEFT_SHIFT;
   if (mul_params.multiplier_fixedpoint_perchannel()) {
     // Temporary release-assert to debug some crashes in an application.