From 04b7ccfd87d802e6b9a22b86c8d098979164b8ba Mon Sep 17 00:00:00 2001
From: Fox Caminiti <fox@foxcam.net>
Date: Mon, 15 Aug 2022 23:03:30 -0400
Subject: undo started

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

(limited to 'bitmap_calls.cpp')

diff --git a/bitmap_calls.cpp b/bitmap_calls.cpp
index dd5c793..6425e6d 100644
--- a/bitmap_calls.cpp
+++ b/bitmap_calls.cpp
@@ -31,10 +31,17 @@ void Bitmap_ConvertPacking(void *Buffer, void *DestBuffer, uint16 Width, uint16
                 DPixel = Temp + Y*Width*4 + X*BytesPerPixel;
             }
 
+#if ARM
+            if (InstructionMode == instruction_mode_neon) {
+                uint32x2x2_t Row = vld2_u32((uint32 *)Pixel);
+                vst2_u32((uint32 *)DPixel, Row);
+                X += 4;
+#else
             if (InstructionMode == instruction_mode_sse || InstructionMode == instruction_mode_avx) {
                 __m128i Row = _mm_loadu_si128((__m128i *)Pixel);
                 _mm_storeu_si128((__m128i *)DPixel, Row);
                 X+=4;
+#endif
             } else {
                 *(uint32 *)DPixel = *(uint32 *)Pixel;
                 X++;
@@ -101,8 +108,13 @@ static void
 Bitmap_Clear(void *Buffer, uint16 Width, uint16 Height, uint16 BytesPerPixel)
 {
     uint8 *Row = (uint8 *)Buffer;
+#if ARM
+    uint32 Zero[4] = {0};
+    uint32x2x4_t Zero8 = vld4_dup_u32(Zero);
+#else
     __m256i Zero8 = _mm256_setzero_si256();
     __m128i Zero = _mm_setzero_si128();
+#endif
     uint64 bytes = 0;
 
     uint16 ByteOffset = Bitmap_CalcByteOffset(BytesPerPixel);
@@ -110,10 +122,15 @@ Bitmap_Clear(void *Buffer, uint16 Width, uint16 Height, uint16 BytesPerPixel)
 
     while (bytes < TotalBytes) {
         uint8 *Pixel = Row + bytes;
+#if ARM
+        if (InstructionMode == instruction_mode_neon) {
+            vst4_u32((uint32 *)Pixel, Zero8);
+#else
         if (InstructionMode == instruction_mode_avx) {
             _mm256_storeu_si256((__m256i *)Pixel, Zero8);
         } else if (InstructionMode == instruction_mode_sse) {
             _mm_storeu_si128((__m128i *)Pixel, Zero);
+#endif
         } else {
             *(uint32 *)Pixel = 0x00000000;
         }
@@ -136,10 +153,15 @@ Bitmap_CalcPackedDimensions(uint16 Width, uint16 Height, uint16 *WidthP, uint16
 static uint16
 Bitmap_CalcByteOffset(uint16 BytesPerPixel) {
     uint16 ByteOffset = BytesPerPixel;
+#if ARM
+    if (InstructionMode == instruction_mode_neon)
+        ByteOffset = 8*BytesPerPixel;
+#else
     if (InstructionMode == instruction_mode_avx)
         ByteOffset = 8*BytesPerPixel;
     if (InstructionMode == instruction_mode_sse)
         ByteOffset = 4*BytesPerPixel;
+#endif
     return ByteOffset;
 }
 
@@ -174,12 +196,18 @@ Bitmap_CopyToPointer(void *Input, void *Output, uint16 BytesPerPixel, uint64 Tot
             int pp = 0;
         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);
+#else
         if (InstructionMode == instruction_mode_avx) {
             __m256i OutputPixel = _mm256_loadu_si256((__m256i *)Pixel);
             _mm256_storeu_si256((__m256i *)Pixel2, OutputPixel);
         } else if (InstructionMode == instruction_mode_sse) {
             __m128i OutputPixel = _mm_loadu_si128((__m128i *)Pixel);
             _mm_storeu_si128((__m128i *)Pixel2, OutputPixel);
+#endif
         } else {
             *(uint32 *)Pixel2 = *(uint32 *)Pixel;
         }
-- 
cgit v1.2.3