fix memcpy_neon

1 files changed, 25 insertions, 8 deletions

diff --git a/src/mem_utils.h b/src/mem_utils.h
index 8601f78..cb5f179 100644
--- a/src/mem_utils.h
+++ b/src/mem_utils.h
@@ -1,25 +1,42 @@
 #pragma once
 
 #include <cstddef>
+#include <iostream>
 
 #include <arm_neon.h>
 
+// using ARRAY_TYPE = uint16_t;
+// using ARRAY_TYPE = uint32_t;
+using ARRAY_TYPE = uint64_t;
+// using ARRAY_TYPE = uint64_t;
+
 template<std::size_t S>
-void memcpy_1by1(std::byte *dst, const std::byte *src)
+void memcpy_1by1(ARRAY_TYPE *dst, const ARRAY_TYPE *src)
 {
     for (std::size_t i{0}; i < S; ++i) {
         dst[i] = src[i];
     }
 }
-
-using ARRAY_TYPE = uint32_t;
 template<std::size_t S>
-void memcpy_neon(ARRAY_TYPE *dst, const ARRAY_TYPE *src)
+inline void memcpy_neon(ARRAY_TYPE *dst, const ARRAY_TYPE *src)
 {
-    uint32x4_t tmp;
+    // for (std::size_t i{0}; i < S; i += 4) {
+    //     vst1q_u32(&dst[i], vld1q_u32(src + i));
+    // }
+    // if (uint64_t(src) % 16 != 0) {
+    //     std::cerr << "src misaligned" << std::endl;
+    // }
+    // if (uint64_t(dst) % 16 != 0) {
+    //     std::cerr << "dst misaligned" << std::endl;
+    // }
 
-    for (std::size_t i{0}; i < (S / 4); i += 4) {
-        tmp = vld1q_u32(src + i);
-        vst1q_u32(&dst[i], tmp);
+    for (std::size_t i{0}; i < S; i += 2) {
+        vst1q_u64(&dst[i], vld1q_u64(src + i));
     }
+    // for (std::size_t i{0}; i < S; i += 8) {
+    //     vst1q_u16(&dst[i], vld1q_u16(src + i));
+    // }
+    // for (std::size_t i{0}; i < S; i += 4) {
+    //     vst1q_u64(&dst[i], vld1q_u64(src + i));
+    // }
 }