From 8362ddda746fc73b8e71982755d851f5c1125aea Mon Sep 17 00:00:00 2001
From: Mira Grudzinska <m.grudzinska@samsung.com>
Date: Sun, 14 Nov 2021 00:30:18 +0100
Subject: [PATCH] sw_engine: avxRasterTranslucentRle implemented

---
 src/lib/sw_engine/tvgSwRaster.cpp  |  6 ++--
 src/lib/sw_engine/tvgSwRasterAvx.h | 52 ++++++++++++++++++++++++++++++
 2 files changed, 56 insertions(+), 2 deletions(-)

diff --git a/src/lib/sw_engine/tvgSwRaster.cpp b/src/lib/sw_engine/tvgSwRaster.cpp
index b5866b29..28c62350 100644
--- a/src/lib/sw_engine/tvgSwRaster.cpp
+++ b/src/lib/sw_engine/tvgSwRaster.cpp
@@ -200,7 +200,9 @@ static bool _rasterTranslucentRle(SwSurface* surface, SwRleData* rle, uint32_t c
         }
     }
 
-#if defined(THORVG_NEON_VECTOR_SUPPORT)
+#if defined(THORVG_AVX_VECTOR_SUPPORT)
+    return avxRasterTranslucentRle(surface, rle, color);
+#elif defined(THORVG_NEON_VECTOR_SUPPORT)
     return neonRasterTranslucentRle(surface, rle, color);
 #else
     return cRasterTranslucentRle(surface, rle, color);
@@ -1586,4 +1588,4 @@ bool rasterImage(SwSurface* surface, SwImage* image, const Matrix* transform, co
             return _rasterImage(surface, image, bbox);
         }
     }
-}
\ No newline at end of file
+}
diff --git a/src/lib/sw_engine/tvgSwRasterAvx.h b/src/lib/sw_engine/tvgSwRasterAvx.h
index 48d34f6b..c14a2138 100644
--- a/src/lib/sw_engine/tvgSwRasterAvx.h
+++ b/src/lib/sw_engine/tvgSwRasterAvx.h
@@ -124,4 +124,56 @@ static inline bool avxRasterTranslucentRect(SwSurface* surface, const SwBBox& re
     return true;
 }
 
+
+static inline bool avxRasterTranslucentRle(SwSurface* surface, const SwRleData* rle, uint32_t color)
+{
+    auto span = rle->spans;
+    uint32_t src;
+
+    for (uint32_t i = 0; i < rle->size; ++i) {
+        auto dst = &surface->buffer[span->y * surface->stride + span->x];
+
+        if (span->coverage < 255) src = ALPHA_BLEND(color, span->coverage);
+        else src = color;
+
+        auto ialpha = 255 - static_cast<uint8_t>(surface->blender.alpha(src));
+
+        //1. fill the not aligned memory (for 128-bit registers a 16-bytes alignment is required)
+        auto notAligned = ((uintptr_t)dst & 0xf) / 4;
+        if (notAligned) {
+            notAligned = (N_32BITS_IN_128REG - notAligned > span->len ? span->len : N_32BITS_IN_128REG - notAligned);
+            for (uint32_t x = 0; x < notAligned; ++x, ++dst) {
+                *dst = src + ALPHA_BLEND(*dst, ialpha);
+            }
+        }
+
+        //2. fill the aligned memory using avx - N_32BITS_IN_128REG pixels processed at once
+        //In order to avoid unneccessary avx variables declarations a check is made whether there are any iterations at all
+        uint32_t iterations = (span->len - notAligned) / N_32BITS_IN_128REG;
+        uint32_t avxFilled = 0;
+        if (iterations > 0) {
+            auto avxSrc = _mm_set1_epi32(src);
+            auto avxIalpha = _mm_set1_epi8(ialpha);
+
+            avxFilled = iterations * N_32BITS_IN_128REG;
+            auto avxDst = (__m128i*)dst;
+            for (uint32_t x = 0; x < iterations; ++x, ++avxDst) {
+                *avxDst = _mm_add_epi32(avxSrc, ALPHA_BLEND(*avxDst, avxIalpha));
+            }
+        }
+
+        //3. fill the remaining pixels
+        int32_t leftovers = span->len - notAligned - avxFilled;
+        dst += avxFilled;
+        while (leftovers--) {
+            *dst = src + ALPHA_BLEND(*dst, ialpha);
+            dst++;
+        }
+
+        ++span;
+    }
+    return true;
+}
+
+
 #endif