From 163aa8f30abea7d25cce8ce6e03570546688fec8 Mon Sep 17 00:00:00 2001
From: Gin <>
Date: Thu, 19 Dec 2024 17:27:20 -0800
Subject: [PATCH] fix neon bug

---
 ...naryImage_BasicFilters_Core_64x8_arm64_NEON.cpp |  3 +++
 .../Kernels_BinaryImage_BasicFilters_arm64_NEON.h  | 14 +++++++-------
 .../Waterfill/Kernels_Waterfill_Session.tpp        |  6 +++---
 3 files changed, 13 insertions(+), 10 deletions(-)

diff --git a/SerialPrograms/Source/Kernels/BinaryImageFilters/Kernels_BinaryImage_BasicFilters_Core_64x8_arm64_NEON.cpp b/SerialPrograms/Source/Kernels/BinaryImageFilters/Kernels_BinaryImage_BasicFilters_Core_64x8_arm64_NEON.cpp
index 716259ebe..52d718340 100644
--- a/SerialPrograms/Source/Kernels/BinaryImageFilters/Kernels_BinaryImage_BasicFilters_Core_64x8_arm64_NEON.cpp
+++ b/SerialPrograms/Source/Kernels/BinaryImageFilters/Kernels_BinaryImage_BasicFilters_Core_64x8_arm64_NEON.cpp
@@ -10,6 +10,7 @@
 #include "Kernels_BinaryImage_BasicFilters_Routines.h"
 #include "Kernels_BinaryImage_BasicFilters_arm64_NEON.h"
 
+
 namespace PokemonAutomation{
 namespace Kernels{
 
@@ -24,6 +25,8 @@ void filter_by_mask_64x8_arm64_NEON(
 }
 
 
+
+
 void compress_rgb32_to_binary_range_64x8_arm64_NEON(
     const uint32_t* image, size_t bytes_per_row,
     PackedBinaryMatrix_IB& matrix0, uint32_t mins0, uint32_t maxs0
diff --git a/SerialPrograms/Source/Kernels/BinaryImageFilters/Kernels_BinaryImage_BasicFilters_arm64_NEON.h b/SerialPrograms/Source/Kernels/BinaryImageFilters/Kernels_BinaryImage_BasicFilters_arm64_NEON.h
index 848e1d49d..514908197 100644
--- a/SerialPrograms/Source/Kernels/BinaryImageFilters/Kernels_BinaryImage_BasicFilters_arm64_NEON.h
+++ b/SerialPrograms/Source/Kernels/BinaryImageFilters/Kernels_BinaryImage_BasicFilters_arm64_NEON.h
@@ -9,9 +9,9 @@
 
 #include "Kernels/PartialWordAccess/Kernels_PartialWordAccess_arm64_NEON.h"
 
-#include <iostream>
-using std::cout;
-using std::endl;
+// #include <iostream>
+// using std::cout;
+// using std::endl;
 
 namespace PokemonAutomation{
 namespace Kernels{
@@ -149,8 +149,6 @@ class FilterByMask_arm64_NEON{
     const uint32x4_t m_zeros;
 };
 
-
-
 // Compress given pixels buffer (of up to 64-pixel long) into bit map and store in one uint64_t.
 class Compressor_RgbRange_arm64_NEON{
 public:
@@ -178,13 +176,14 @@ class Compressor_RgbRange_arm64_NEON{
             bits |= convert16(pixels + c) << c;
             c += 16;
         }
-        for(; c < count; c += 4){
+
+        count %= 16;
+        for(size_t i = 0; i < count / 4; i++, c+=4){
             const uint8x16_t pixel = vld1q_u8((const uint8_t*)(pixels + c));
             bits |= convert4(pixel) << c;
         }
         count %= 4;
         if (count){
-            c -= 4;
             PartialWordAccess_arm64_NEON loader(count * sizeof(uint32_t));
             const uint8x16_t pixel = loader.load(pixels + c);
             const uint64_t mask = ((uint64_t)1 << count) - 1;
@@ -273,6 +272,7 @@ class Compressor_RgbEuclidean_arm64_NEON{
         for(size_t i = 0; i < count / 4; i++, c+=4){
             bits |= convert4(pixels + c) << c;
         }
+        count %= 4;
         if (count){
             PartialWordAccess_arm64_NEON loader(count * sizeof(uint32_t));
             const uint8x16_t pixel = loader.load(pixels + c);
diff --git a/SerialPrograms/Source/Kernels/Waterfill/Kernels_Waterfill_Session.tpp b/SerialPrograms/Source/Kernels/Waterfill/Kernels_Waterfill_Session.tpp
index b6d7db850..880128c9a 100644
--- a/SerialPrograms/Source/Kernels/Waterfill/Kernels_Waterfill_Session.tpp
+++ b/SerialPrograms/Source/Kernels/Waterfill/Kernels_Waterfill_Session.tpp
@@ -16,9 +16,9 @@
 #include "Kernels/BinaryMatrix/Kernels_SparseBinaryMatrixCore.h"
 #include "Kernels_Waterfill_Session.h"
 
-//#include <iostream>
-//using std::cout;
-//using std::endl;
+// #include <iostream>
+// using std::cout;
+// using std::endl;
 
 namespace PokemonAutomation{
 namespace Kernels{