From b39c7374009d03d448e47e08ddabc30abeee5247 Mon Sep 17 00:00:00 2001
From: Fox Caminiti <fox@foxcam.net>
Date: Thu, 25 Aug 2022 18:55:59 -0400
Subject: minor fixes

---
 bitmap_calls.cpp | 33 ++++++++++++++++++++++++++++-----
 1 file changed, 28 insertions(+), 5 deletions(-)

(limited to 'bitmap_calls.cpp')

diff --git a/bitmap_calls.cpp b/bitmap_calls.cpp
index 46f0c31..0e76039 100644
--- a/bitmap_calls.cpp
+++ b/bitmap_calls.cpp
@@ -248,16 +248,25 @@ Bitmap_StencilAlpha(void *Input, void *Output, uint16 BytesPerPixel, uint64 Tota
     uint16 ByteOffset = Bitmap_CalcByteOffset(BytesPerPixel);
     uint64 RemainderBytes = TotalBytes % ByteOffset;
 
+#if ARM
+#else
     __m256i AlphaBytes = _mm256_set1_epi32(0x00FFFFFF);
     __m256i Zeroi = _mm256_set1_epi32(0);
+#endif
 
     while (bytes <= TotalBytes - RemainderBytes) {
         uint8 *Pixel = (uint8 *)Row + bytes;
         uint8 *Pixel2 = (uint8 *)Row2 + bytes;
 #if ARM
         if (InstructionMode == instruction_mode_neon) {
-            uint32x2x4_t OutputPixel = vld4_u32((uint32 *)Pixel);
-            vst4_u32((uint32 *)Pixel2, OutputPixel);
+            // TODO(fox): Optimize and write NEON!
+            uint8 *DestAlpha = Pixel2 + (BytesPerPixel/4)*3;
+            uint8 Alpha = *DestAlpha;
+            uint32 *DestPixel = (uint32 *)Pixel2;
+            uint32 *SrcPixel = (uint32 *)Pixel;
+            *DestPixel = *SrcPixel;
+            *DestAlpha = Alpha;
+            bytes += BytesPerPixel;
 #else
         if (InstructionMode == instruction_mode_avx) {
             __m256i InputPixel = _mm256_loadu_si256((__m256i *)Pixel);
@@ -266,19 +275,33 @@ Bitmap_StencilAlpha(void *Input, void *Output, uint16 BytesPerPixel, uint64 Tota
                 OutputPixel = _mm256_blendv_epi8(OutputPixel, InputPixel, AlphaBytes);
                 _mm256_storeu_si256((__m256i *)Pixel2, OutputPixel);
             }
+            bytes += ByteOffset;
         } else if (InstructionMode == instruction_mode_sse) {
             __m128i OutputPixel = _mm_loadu_si128((__m128i *)Pixel);
             _mm_storeu_si128((__m128i *)Pixel2, OutputPixel);
+            bytes += ByteOffset;
 #endif
         } else {
-            *(uint32 *)Pixel2 = *(uint32 *)Pixel;
+            // TODO(fox): Optimize and write NEON!
+            uint8 *DestAlpha = Pixel2 + (BytesPerPixel/4)*3;
+            uint8 Alpha = *DestAlpha;
+            uint32 *DestPixel = (uint32 *)Pixel2;
+            uint32 *SrcPixel = (uint32 *)Pixel;
+            *DestPixel = *SrcPixel;
+            *DestAlpha = Alpha;
+            bytes += BytesPerPixel;
         }
-        bytes += ByteOffset;
     }
     while (bytes <= TotalBytes) {
         uint8 *Pixel = (uint8 *)Row + bytes;
         uint8 *Pixel2 = (uint8 *)Row2 + bytes;
-        *(uint32 *)Pixel2 = *(uint32 *)Pixel;
+        // TODO(fox): Optimize and write NEON!
+        uint8 *DestAlpha = Pixel2 + (BytesPerPixel/4)*3;
+        uint8 Alpha = *DestAlpha;
+        uint32 *DestPixel = (uint32 *)Pixel2;
+        uint32 *SrcPixel = (uint32 *)Pixel;
+        *DestPixel = *SrcPixel;
+        *DestAlpha = Alpha;
         bytes += BytesPerPixel;
     }
 }
-- 
cgit v1.2.3