diff --git a/inc/thorvg.h b/inc/thorvg.h
index 28bd7602..1ee898ca 100644
--- a/inc/thorvg.h
+++ b/inc/thorvg.h
@@ -216,7 +216,8 @@ enum class BlendMethod : uint8_t
 enum class SceneEffect : uint8_t
 {
     ClearAll = 0,      ///< Reset all previously applied scene effects, restoring the scene to its original state.
-    GaussianBlur       ///< Apply a blur effect with a Gaussian filter. Param(3) = {sigma(float)[> 0], direction(int)[both: 0 / horizontal: 1 / vertical: 2], border(int)[duplicate: 0 / wrap: 1], quality(int)[0 - 100]}
+    GaussianBlur,      ///< Apply a blur effect with a Gaussian filter. Param(3) = {sigma(float)[> 0], direction(int)[both: 0 / horizontal: 1 / vertical: 2], border(int)[duplicate: 0 / wrap: 1], quality(int)[0 - 100]}
+    DropShadow         ///< Apply a drop shadow effect with a Gaussian Blur filter. Param(8) = {color_R(int)[0 - 255], color_G(int)[0 - 255], color_B(int)[0 - 255], opacity(int)[0 - 255], angle(float)[0 - 360], distance(float), blur_sigma(float)[> 0], quality(int)[0 - 100]}
 };
 
 
diff --git a/src/renderer/gl_engine/tvgGlRenderer.cpp b/src/renderer/gl_engine/tvgGlRenderer.cpp
index f7ec6ee3..20984975 100644
--- a/src/renderer/gl_engine/tvgGlRenderer.cpp
+++ b/src/renderer/gl_engine/tvgGlRenderer.cpp
@@ -1079,7 +1079,7 @@ bool GlRenderer::prepare(TVG_UNUSED RenderEffect* effect)
 }
 
 
-bool GlRenderer::effect(TVG_UNUSED RenderCompositor* cmp, TVG_UNUSED const RenderEffect* effect)
+bool GlRenderer::effect(TVG_UNUSED RenderCompositor* cmp, TVG_UNUSED const RenderEffect* effect, TVG_UNUSED bool direct)
 {
     TVGLOG("GL_ENGINE", "SceneEffect(%d) is not supported", (int)effect->type);
     return false;
diff --git a/src/renderer/gl_engine/tvgGlRenderer.h b/src/renderer/gl_engine/tvgGlRenderer.h
index e676c565..8c8368ff 100644
--- a/src/renderer/gl_engine/tvgGlRenderer.h
+++ b/src/renderer/gl_engine/tvgGlRenderer.h
@@ -87,7 +87,7 @@ public:
     bool endComposite(RenderCompositor* cmp) override;
 
     bool prepare(RenderEffect* effect) override;
-    bool effect(RenderCompositor* cmp, const RenderEffect* effect) override;
+    bool effect(RenderCompositor* cmp, const RenderEffect* effect, bool direct) override;
 
     static GlRenderer* gen();
     static int init(TVG_UNUSED uint32_t threads);
diff --git a/src/renderer/sw_engine/tvgSwCommon.h b/src/renderer/sw_engine/tvgSwCommon.h
index 74b38330..b6973076 100644
--- a/src/renderer/sw_engine/tvgSwCommon.h
+++ b/src/renderer/sw_engine/tvgSwCommon.h
@@ -568,13 +568,17 @@ bool rasterStroke(SwSurface* surface, SwShape* shape, uint8_t r, uint8_t g, uint
 bool rasterGradientStroke(SwSurface* surface, SwShape* shape, const Fill* fdata, uint8_t opacity);
 bool rasterClear(SwSurface* surface, uint32_t x, uint32_t y, uint32_t w, uint32_t h, pixel_t val = 0);
 void rasterPixel32(uint32_t *dst, uint32_t val, uint32_t offset, int32_t len);
+void rasterTranslucentPixel32(uint32_t* dst, uint32_t* src, uint32_t len, uint8_t opacity);
+void rasterPixel32(uint32_t* dst, uint32_t* src, uint32_t len, uint8_t opacity);
 void rasterGrayscale8(uint8_t *dst, uint8_t val, uint32_t offset, int32_t len);
 void rasterXYFlip(uint32_t* src, uint32_t* dst, int32_t stride, int32_t w, int32_t h, const SwBBox& bbox, bool flipped);
 void rasterUnpremultiply(RenderSurface* surface);
 void rasterPremultiply(RenderSurface* surface);
 bool rasterConvertCS(RenderSurface* surface, ColorSpace to);
 
-bool effectGaussianBlur(SwImage& image, SwImage& buffer, const SwBBox& bbox, const RenderEffectGaussian* params);
-bool effectGaussianPrepare(RenderEffectGaussian* effect);
+bool effectGaussianBlur(SwCompositor* cmp, SwSurface* surface, const RenderEffectGaussianBlur* params, bool direct);
+bool effectGaussianBlurPrepare(RenderEffectGaussianBlur* effect);
+bool effectDropShadow(SwCompositor* cmp, SwSurface* surfaces[2], const RenderEffectDropShadow* params, bool direct);
+bool effectDropShadowPrepare(RenderEffectDropShadow* effect);
 
 #endif /* _TVG_SW_COMMON_H_ */
diff --git a/src/renderer/sw_engine/tvgSwPostEffect.cpp b/src/renderer/sw_engine/tvgSwPostEffect.cpp
index 4aea7d5c..46b021d9 100644
--- a/src/renderer/sw_engine/tvgSwPostEffect.cpp
+++ b/src/renderer/sw_engine/tvgSwPostEffect.cpp
@@ -20,10 +20,11 @@
  * SOFTWARE.
  */
 
+#include "tvgMath.h"
 #include "tvgSwCommon.h"
 
 /************************************************************************/
-/* Gaussian Filter Implementation                                       */
+/* Gaussian Blur Implementation                                         */
 /************************************************************************/
 
 struct SwGaussianBlur
@@ -48,49 +49,59 @@ static void _gaussianExtendRegion(RenderRegion& region, int extra, int8_t direct
 }
 
 
-static int _gaussianRemap(int end, int idx, int border)
+static int _gaussianEdgeWrap(int end, int idx)
 {
-    //wrap
-    if (border == 1) return idx % end;
+    auto r = idx % end;
+    return (r < 0) ? end + r : r;
+}
 
-    //duplicate
+
+static int _gaussianEdgeExtend(int end, int idx)
+{
     if (idx < 0) return 0;
     else if (idx >= end) return end - 1;
     return idx;
 }
 
 
+static int _gaussianRemap(int end, int idx, int border)
+{
+    if (border == 1) return _gaussianEdgeWrap(end, idx);
+    return _gaussianEdgeExtend(end, idx);
+}
+
+
 //TODO: SIMD OPTIMIZATION?
-static void _gaussianBlur(uint8_t* src, uint8_t* dst, int32_t stride, int32_t w, int32_t h, const SwBBox& bbox, int32_t dimension, int border, bool flipped)
+static void _gaussianFilter(uint8_t* dst, uint8_t* src, int32_t stride, int32_t w, int32_t h, const SwBBox& bbox, int32_t dimension, int border, bool flipped)
 {
     if (flipped) {
-        src += ((bbox.min.x * stride) + bbox.min.y) << 2;
-        dst += ((bbox.min.x * stride) + bbox.min.y) << 2;
+        src += (bbox.min.x * stride + bbox.min.y) << 2;
+        dst += (bbox.min.x * stride + bbox.min.y) << 2;
     } else {
-        src += ((bbox.min.y * stride) + bbox.min.x) << 2;
-        dst += ((bbox.min.y * stride) + bbox.min.x) << 2;
+        src += (bbox.min.y * stride + bbox.min.x) << 2;
+        dst += (bbox.min.y * stride + bbox.min.x) << 2;
     }
 
     auto iarr = 1.0f / (dimension + dimension + 1);
 
     #pragma omp parallel for
-    for (int x = 0; x < h; x++) {
-        auto p = x * stride;
+    for (int y = 0; y < h; ++y) {
+        auto p = y * stride;
         auto i = p * 4;                 //current index
         auto l = -(dimension + 1);      //left index
         auto r = dimension;             //right index
         int acc[4] = {0, 0, 0, 0};      //sliding accumulator
 
-        //initial acucmulation
-        for (int x2 = l; x2 < r; ++x2) {
-            auto id = (_gaussianRemap(w, x2, border) + p) * 4;
+        //initial accumulation
+        for (int x = l; x < r; ++x) {
+            auto id = (_gaussianRemap(w, x, border) + p) * 4;
             acc[0] += src[id++];
             acc[1] += src[id++];
             acc[2] += src[id++];
             acc[3] += src[id];
         }
         //perform filtering
-        for (int x2 = 0; x2 < w; ++x2, ++r, ++l) {
+        for (int x = 0; x < w; ++x, ++r, ++l) {
             auto rid = (_gaussianRemap(w, r, border) + p) * 4;
             auto lid = (_gaussianRemap(w, l, border) + p) * 4;
             acc[0] += src[rid++] - src[lid++];
@@ -106,11 +117,15 @@ static void _gaussianBlur(uint8_t* src, uint8_t* dst, int32_t stride, int32_t w,
 }
 
 
-static int _gaussianInit(int* kernel, float sigma, int level)
+static int _gaussianInit(SwGaussianBlur* data, float sigma, int quality)
 {
     const auto MAX_LEVEL = SwGaussianBlur::MAX_LEVEL;
 
-    //compute the kernel
+    if (tvg::zero(sigma)) return 0;
+
+    data->level = int(SwGaussianBlur::MAX_LEVEL * ((quality - 1) * 0.01f)) + 1;
+
+    //compute box kernel sizes
     auto wl = (int) sqrt((12 * sigma / MAX_LEVEL) + 1);
     if (wl % 2 == 0) --wl;
     auto wu = wl + 2;
@@ -118,92 +133,286 @@ static int _gaussianInit(int* kernel, float sigma, int level)
     auto m = int(mi + 0.5f);
     auto extends = 0;
 
-    for (int i = 0; i < level; i++) {
-        kernel[i] = ((i < m ? wl : wu) - 1) / 2;
-        extends += kernel[i];
+    for (int i = 0; i < data->level; i++) {
+        data->kernel[i] = ((i < m ? wl : wu) - 1) / 2;
+        extends += data->kernel[i];
     }
 
     return extends;
 }
 
 
-bool effectGaussianPrepare(RenderEffectGaussian* params)
+bool effectGaussianBlurPrepare(RenderEffectGaussianBlur* params)
 {
-    auto data = (SwGaussianBlur*)malloc(sizeof(SwGaussianBlur));
+    auto rd = (SwGaussianBlur*)malloc(sizeof(SwGaussianBlur));
 
-    //compute box kernel sizes
-    data->level = int(SwGaussianBlur::MAX_LEVEL * ((params->quality - 1) * 0.01f)) + 1;
-    auto extends = _gaussianInit(data->kernel, params->sigma * params->sigma, data->level);
+    auto extends = _gaussianInit(rd, params->sigma * params->sigma, params->quality);
 
-    //skip, if the parameters are invalid.
+    //invalid
     if (extends == 0) {
         params->invalid = true;
-        free(data);
+        free(rd);
         return false;
     }
 
     _gaussianExtendRegion(params->extend, extends, params->direction);
 
-    params->rd = data;
+    params->rd = rd;
 
     return true;
 }
 
 
-/* It is best to take advantage of the Gaussian blur’s separable property
-   by dividing the process into two passes. horizontal and vertical.
-   We can expect fewer calculations. */
-bool effectGaussianBlur(SwImage& image, SwImage& buffer, const SwBBox& bbox, const RenderEffectGaussian* params)
+bool effectGaussianBlur(SwCompositor* cmp, SwSurface* surface, const RenderEffectGaussianBlur* params, TVG_UNUSED bool direct)
 {
-    if (params->invalid) return false;
-
-    if (image.channelSize != sizeof(uint32_t)) {
+    if (cmp->image.channelSize != sizeof(uint32_t)) {
         TVGERR("SW_ENGINE", "Not supported grayscale Gaussian Blur!");
         return false;
     }
 
+    auto& buffer = surface->compositor->image;
     auto data = static_cast<SwGaussianBlur*>(params->rd);
+    auto& bbox = cmp->bbox;
     auto w = (bbox.max.x - bbox.min.x);
     auto h = (bbox.max.y - bbox.min.y);
-    auto stride = image.stride;
-    auto front = image.buf8;
-    auto back = buffer.buf8;
+    auto stride = cmp->image.stride;
+    auto front = cmp->image.buf32;
+    auto back = buffer.buf32;
     auto swapped = false;
 
-    //fine-tuning for low-quality (experimental)
-    auto threshold = (std::min(w, h) < 300) ? 2 : 1;
-
     TVGLOG("SW_ENGINE", "GaussianFilter region(%ld, %ld, %ld, %ld) params(%f %d %d), level(%d)", bbox.min.x, bbox.min.y, bbox.max.x, bbox.max.y, params->sigma, params->direction, params->border, data->level);
 
+    /* It is best to take advantage of the Gaussian blur’s separable property
+       by dividing the process into two passes. horizontal and vertical.
+       We can expect fewer calculations. */
+
     //horizontal
-    if (params->direction == 0 || params->direction == 1) {
+    if (params->direction != 2) {
         for (int i = 0; i < data->level; ++i) {
-            auto k = data->kernel[i] / threshold;
-            if (k == 0) continue;
-            _gaussianBlur(front, back, stride, w, h, bbox, k, params->border, false);
+            _gaussianFilter(reinterpret_cast<uint8_t*>(back), reinterpret_cast<uint8_t*>(front), stride, w, h, bbox, data->kernel[i], params->border, false);
             std::swap(front, back);
             swapped = !swapped;
         }
     }
 
     //vertical. x/y flipping and horionztal access is pretty compatible with the memory architecture.
-    if (params->direction == 0 || params->direction == 2) {
-        rasterXYFlip(reinterpret_cast<uint32_t*>(front), reinterpret_cast<uint32_t*>(back), stride, w, h, bbox, false);
+    if (params->direction != 1) {
+        rasterXYFlip(front, back, stride, w, h, bbox, false);
         std::swap(front, back);
 
         for (int i = 0; i < data->level; ++i) {
-            auto k = data->kernel[i] / threshold;
-            if (k == 0) continue;
-            _gaussianBlur(front, back, stride, h, w, bbox, k, params->border, true);
+            _gaussianFilter(reinterpret_cast<uint8_t*>(back), reinterpret_cast<uint8_t*>(front), stride, h, w, bbox, data->kernel[i], params->border, true);
             std::swap(front, back);
             swapped = !swapped;
         }
 
-        rasterXYFlip(reinterpret_cast<uint32_t*>(front), reinterpret_cast<uint32_t*>(back), stride, h, w, bbox, true);
+        rasterXYFlip(front, back, stride, h, w, bbox, true);
         std::swap(front, back);
     }
 
-    if (swapped) std::swap(image.buf8, buffer.buf8);
+    if (swapped) std::swap(cmp->image.buf8, buffer.buf8);
+
+    return true;
+}
+
+/************************************************************************/
+/* Drop Shadow Implementation                                           */
+/************************************************************************/
+
+struct SwDropShadow : SwGaussianBlur
+{
+    SwPoint offset;
+};
+
+
+//TODO: SIMD OPTIMIZATION?
+static void _dropShadowFilter(uint32_t* dst, uint32_t* src, int stride, int w, int h, const SwBBox& bbox, int32_t dimension, uint32_t color, bool flipped)
+{
+    if (flipped) {
+        src += (bbox.min.x * stride + bbox.min.y);
+        dst += (bbox.min.x * stride + bbox.min.y);
+    } else {
+        src += (bbox.min.y * stride + bbox.min.x);
+        dst += (bbox.min.y * stride + bbox.min.x);
+    }
+    auto iarr = 1.0f / (dimension + dimension + 1);
+
+    //#pragma omp parallel for
+    for (int y = 0; y < h; ++y) {
+        auto p = y * stride;
+        auto i = p;                     //current index
+        auto l = -(dimension + 1);      //left index
+        auto r = dimension;             //right index
+        int acc = 0;                    //sliding accumulator
+
+        //initial accumulation
+        for (int x = l; x < r; ++x) {
+            auto id = _gaussianEdgeExtend(w, x) + p;
+            acc += A(src[id]);
+        }
+        //perform filtering
+        for (int x = 0; x < w; ++x, ++r, ++l) {
+            auto rid = _gaussianEdgeExtend(w, r) + p;
+            auto lid = _gaussianEdgeExtend(w, l) + p;
+            acc += A(src[rid]) - A(src[lid]);
+            dst[i++] = ALPHA_BLEND(color, static_cast<uint8_t>(acc * iarr + 0.5f));
+        }
+    }
+}
+
+
+static void _dropShadowShift(uint32_t* dst, uint32_t* src, int stride, SwBBox& region, SwPoint& offset, uint8_t opacity, bool direct)
+{
+    src += (region.min.y * stride + region.min.x);
+    dst += (region.min.y * stride + region.min.x);
+
+    auto w = region.max.x - region.min.x;
+    auto h = region.max.y - region.min.y;
+    auto translucent = (direct || opacity < 255);
+
+    //shift offset
+    if (region.min.x + offset.x < 0) {
+        src -= offset.x;
+    } else {
+        dst += offset.x;
+        w -= offset.x;
+    }
+
+    if (region.min.y + offset.y < 0) {
+        src -= (offset.y * stride);
+    } else {
+        dst += (offset.y * stride);
+        h -= offset.y;
+    }
+
+    for (auto y = 0; y < h; ++y) {
+        if (translucent) rasterTranslucentPixel32(dst, src, w, opacity);
+        else rasterPixel32(dst, src, w, opacity);
+        src += stride;
+        dst += stride;
+    }
+}
+
+
+static void _dropShadowExtendRegion(RenderRegion& region, int extra, SwPoint& offset)
+{
+    //bbox region expansion for feathering
+    region.x = -extra;
+    region.w = extra * 2;
+    region.y = -extra;
+    region.h = extra * 2;
+
+    region.x = std::min(region.x + (int32_t)offset.x, region.x);
+    region.y = std::min(region.y + (int32_t)offset.y, region.y);
+    region.w += abs(offset.x);
+    region.h += abs(offset.y);
+}
+
+
+bool effectDropShadowPrepare(RenderEffectDropShadow* params)
+{
+    auto rd = (SwDropShadow*)malloc(sizeof(SwDropShadow));
+
+    //compute box kernel sizes
+    auto extends = _gaussianInit(rd, params->sigma * params->sigma, params->quality);
+
+    //invalid
+    if (extends == 0 || params->color[3] == 0) {
+        params->invalid = true;
+        free(rd);
+        return false;
+    }
+
+    //offset
+    if (params->distance > 0.0f) {
+        auto radian = tvg::deg2rad(90.0f - params->angle);
+        rd->offset = {(SwCoord)(params->distance * cosf(radian)), (SwCoord)(-1.0f * params->distance * sinf(radian))};
+    } else {
+        rd->offset = {0, 0};
+    }
+
+    //bbox region expansion for feathering
+    _dropShadowExtendRegion(params->extend, extends, rd->offset);
+
+    params->rd = rd;
+
+    return true;
+}
+
+
+//A quite same integration with effectGaussianBlur(). See it for detailed comments.
+//surface[0]: the original image, to overlay it into the filtered image.
+//surface[1]: temporary buffer for generating the filtered image.
+bool effectDropShadow(SwCompositor* cmp, SwSurface* surface[2], const RenderEffectDropShadow* params, bool direct)
+{
+    if (cmp->image.channelSize != sizeof(uint32_t)) {
+        TVGERR("SW_ENGINE", "Not supported grayscale Drop Shadow!");
+        return false;
+    }
+
+    //FIXME: if the body is partially visible due to clipping, the shadow also becomes partially visible.
+
+    auto data = static_cast<SwDropShadow*>(params->rd);
+    auto& bbox = cmp->bbox;
+    auto w = (bbox.max.x - bbox.min.x);
+    auto h = (bbox.max.y - bbox.min.y);
+
+    //outside the screen
+    if (abs(data->offset.x) >= w || abs(data->offset.y) >= h) return true;
+
+    SwImage* buffer[] = {&surface[0]->compositor->image, &surface[1]->compositor->image};
+    auto color = cmp->recoverSfc->join(params->color[0], params->color[1], params->color[2], 255);
+    auto stride = cmp->image.stride;
+    auto front = cmp->image.buf32;
+    auto back = buffer[1]->buf32;
+    auto opacity = params->color[3];
+
+    TVGLOG("SW_ENGINE", "DropShadow region(%ld, %ld, %ld, %ld) params(%f %f %f), level(%d)", bbox.min.x, bbox.min.y, bbox.max.x, bbox.max.y, params->angle, params->distance, params->sigma, data->level);
+
+    //saving the original image in order to overlay it into the filtered image.
+    _dropShadowFilter(back, front, stride, w, h, bbox, data->kernel[0], color, false);
+    std::swap(front, buffer[0]->buf32);
+    std::swap(front, back);
+
+    //horizontal
+    for (int i = 1; i < data->level; ++i) {
+        _dropShadowFilter(back, front, stride, w, h, bbox, data->kernel[i], color, false);
+        std::swap(front, back);
+    }
+
+    //vertical
+    rasterXYFlip(front, back, stride, w, h, bbox, false);
+    std::swap(front, back);
+
+    for (int i = 0; i < data->level; ++i) {
+        _dropShadowFilter(back, front, stride, h, w, bbox, data->kernel[i], color, true);
+        std::swap(front, back);
+    }
+
+    rasterXYFlip(front, back, stride, h, w, bbox, true);
+    std::swap(cmp->image.buf32, back);
+
+    //draw to the main surface directly
+    if (direct) {
+        _dropShadowShift(cmp->recoverSfc->buf32, cmp->image.buf32, stride, bbox, data->offset, opacity, direct);
+        std::swap(cmp->image.buf32, buffer[0]->buf32);
+        return true;
+    }
+
+    //draw to the intermediate surface
+    rasterClear(surface[1], bbox.min.x, bbox.min.y, w, h);
+    _dropShadowShift(buffer[1]->buf32, cmp->image.buf32, stride, bbox, data->offset, opacity, direct);
+    std::swap(cmp->image.buf32, buffer[1]->buf32);
+
+    //compositing shadow and body
+    auto s = buffer[0]->buf32 + (bbox.min.y * buffer[0]->stride + bbox.min.x);
+    auto d = cmp->image.buf32 + (bbox.min.y * cmp->image.stride + bbox.min.x);
+
+    for (auto y = 0; y < h; ++y) {
+        rasterTranslucentPixel32(d, s, w, 255);
+        s += buffer[0]->stride;
+        d += cmp->image.stride;
+    }
 
     return true;
 }
\ No newline at end of file
diff --git a/src/renderer/sw_engine/tvgSwRaster.cpp b/src/renderer/sw_engine/tvgSwRaster.cpp
index fd1238be..18ffc18e 100644
--- a/src/renderer/sw_engine/tvgSwRaster.cpp
+++ b/src/renderer/sw_engine/tvgSwRaster.cpp
@@ -869,16 +869,7 @@ static bool _rasterDirectRleImage(SwSurface* surface, const SwImage* image, uint
         auto dst = &surface->buf32[span->y * surface->stride + span->x];
         auto img = image->buf32 + (span->y + image->oy) * image->stride + (span->x + image->ox);
         auto alpha = MULTIPLY(span->coverage, opacity);
-        if (alpha == 255) {
-            for (uint32_t x = 0; x < span->len; ++x, ++dst, ++img) {
-                *dst = *img + ALPHA_BLEND(*dst, IA(*img));
-            }
-        } else {
-            for (uint32_t x = 0; x < span->len; ++x, ++dst, ++img) {
-                auto src = ALPHA_BLEND(*img, alpha);
-                *dst = src + ALPHA_BLEND(*dst, IA(src));
-            }
-        }
+        rasterTranslucentPixel32(dst, img, span->len, alpha);
     }
     return true;
 }
@@ -1144,27 +1135,14 @@ static bool _rasterDirectImage(SwSurface* surface, const SwImage* image, const S
     //32bits channels
     if (surface->channelSize == sizeof(uint32_t)) {
         auto dbuffer = &surface->buf32[region.min.y * surface->stride + region.min.x];
-
         for (auto y = region.min.y; y < region.max.y; ++y) {
-            auto dst = dbuffer;
-            auto src = sbuffer;
-            if (opacity == 255) {
-                for (auto x = region.min.x; x < region.max.x; x++, dst++, src++) {
-                    *dst = *src + ALPHA_BLEND(*dst, IA(*src));
-                }
-            } else {
-                for (auto x = region.min.x; x < region.max.x; ++x, ++dst, ++src) {
-                    auto tmp = ALPHA_BLEND(*src, opacity);
-                    *dst = tmp + ALPHA_BLEND(*dst, IA(tmp));
-                }
-            }
+            rasterTranslucentPixel32(dbuffer, sbuffer, region.max.x - region.min.x, opacity);
             dbuffer += surface->stride;
             sbuffer += image->stride;
         }
     //8bits grayscale
     } else if (surface->channelSize == sizeof(uint8_t)) {
         auto dbuffer = &surface->buf8[region.min.y * surface->stride + region.min.x];
-
         for (auto y = region.min.y; y < region.max.y; ++y, dbuffer += surface->stride, sbuffer += image->stride) {
             auto dst = dbuffer;
             auto src = sbuffer;
@@ -1598,6 +1576,19 @@ static bool _rasterRadialGradientRle(SwSurface* surface, const SwRle* rle, const
 /* External Class Implementation                                        */
 /************************************************************************/
 
+void rasterTranslucentPixel32(uint32_t* dst, uint32_t* src, uint32_t len, uint8_t opacity)
+{
+    //TODO: Support SIMD accelerations
+    cRasterTranslucentPixels(dst, src, len, opacity);
+}
+
+
+void rasterPixel32(uint32_t* dst, uint32_t* src, uint32_t len, uint8_t opacity)
+{
+    //TODO: Support SIMD accelerations
+    cRasterPixels(dst, src, len, opacity);
+}
+
 
 void rasterGrayscale8(uint8_t *dst, uint8_t val, uint32_t offset, int32_t len)
 {
diff --git a/src/renderer/sw_engine/tvgSwRasterC.h b/src/renderer/sw_engine/tvgSwRasterC.h
index 258edf84..8bea1b13 100644
--- a/src/renderer/sw_engine/tvgSwRasterC.h
+++ b/src/renderer/sw_engine/tvgSwRasterC.h
@@ -20,6 +20,38 @@
  * SOFTWARE.
  */
 
+
+template<typename PIXEL_T>
+static void inline cRasterTranslucentPixels(PIXEL_T* dst, PIXEL_T* src, uint32_t len, uint32_t opacity)
+{
+    //TODO: 64bits faster?
+    if (opacity == 255) {
+        for (uint32_t x = 0; x < len; ++x, ++dst, ++src) {
+            *dst = *src + ALPHA_BLEND(*dst, IA(*src));
+        }
+    } else {
+        for (uint32_t x = 0; x < len; ++x, ++dst, ++src) {
+            auto tmp = ALPHA_BLEND(*src, opacity);
+            *dst = tmp + ALPHA_BLEND(*dst, IA(tmp));
+        }
+    }
+}
+
+
+template<typename PIXEL_T>
+static void inline cRasterPixels(PIXEL_T* dst, PIXEL_T* src, uint32_t len, uint32_t opacity)
+{
+    //TODO: 64bits faster?
+    if (opacity == 255) {
+        for (uint32_t x = 0; x < len; ++x, ++dst, ++src) {
+            *dst = *src;
+        }
+    } else {
+        cRasterTranslucentPixels(dst, src, len, opacity);
+    }
+}
+
+
 template<typename PIXEL_T>
 static void inline cRasterPixels(PIXEL_T* dst, PIXEL_T val, uint32_t offset, int32_t len)
 {
diff --git a/src/renderer/sw_engine/tvgSwRenderer.cpp b/src/renderer/sw_engine/tvgSwRenderer.cpp
index d4db78fb..0cdadb2a 100644
--- a/src/renderer/sw_engine/tvgSwRenderer.cpp
+++ b/src/renderer/sw_engine/tvgSwRenderer.cpp
@@ -649,19 +649,32 @@ bool SwRenderer::endComposite(RenderCompositor* cmp)
 bool SwRenderer::prepare(RenderEffect* effect)
 {
     switch (effect->type) {
-        case SceneEffect::GaussianBlur: return effectGaussianPrepare(static_cast<RenderEffectGaussian*>(effect));
+        case SceneEffect::GaussianBlur: return effectGaussianBlurPrepare(static_cast<RenderEffectGaussianBlur*>(effect));
+        case SceneEffect::DropShadow: return effectDropShadowPrepare(static_cast<RenderEffectDropShadow*>(effect));
         default: return false;
     }
 }
 
 
-bool SwRenderer::effect(RenderCompositor* cmp, const RenderEffect* effect)
+bool SwRenderer::effect(RenderCompositor* cmp, const RenderEffect* effect, bool direct)
 {
+    if (effect->invalid) return false;
+
     auto p = static_cast<SwCompositor*>(cmp);
-    auto& buffer = request(surface->channelSize)->compositor->image;
 
     switch (effect->type) {
-        case SceneEffect::GaussianBlur: return effectGaussianBlur(p->image, buffer, p->bbox, static_cast<const RenderEffectGaussian*>(effect));
+        case SceneEffect::GaussianBlur: {
+            return effectGaussianBlur(p, request(surface->channelSize), static_cast<const RenderEffectGaussianBlur*>(effect), direct);
+        }
+        case SceneEffect::DropShadow: {
+            auto cmp1 = request(surface->channelSize);
+            cmp1->compositor->valid = false;
+            auto cmp2 = request(surface->channelSize);
+            SwSurface* surfaces[] = {cmp1, cmp2};
+            auto ret = effectDropShadow(p, surfaces, static_cast<const RenderEffectDropShadow*>(effect), direct);
+            cmp1->compositor->valid = true;
+            return ret;
+        }
         default: return false;
     }
 }
diff --git a/src/renderer/sw_engine/tvgSwRenderer.h b/src/renderer/sw_engine/tvgSwRenderer.h
index a5d63028..f3f7088d 100644
--- a/src/renderer/sw_engine/tvgSwRenderer.h
+++ b/src/renderer/sw_engine/tvgSwRenderer.h
@@ -61,7 +61,7 @@ public:
     void clearCompositors();
 
     bool prepare(RenderEffect* effect) override;
-    bool effect(RenderCompositor* cmp, const RenderEffect* effect) override;
+    bool effect(RenderCompositor* cmp, const RenderEffect* effect, bool direct) override;
 
     static SwRenderer* gen();
     static bool init(uint32_t threads);
diff --git a/src/renderer/tvgRender.h b/src/renderer/tvgRender.h
index 030e7822..003c7d3b 100644
--- a/src/renderer/tvgRender.h
+++ b/src/renderer/tvgRender.h
@@ -276,20 +276,17 @@ struct RenderEffect
     }
 };
 
-struct RenderEffectGaussian : RenderEffect
+struct RenderEffectGaussianBlur : RenderEffect
 {
     float sigma;
     uint8_t direction; //0: both, 1: horizontal, 2: vertical
     uint8_t border;    //0: duplicate, 1: wrap
     uint8_t quality;   //0 ~ 100  (optional)
 
-    static RenderEffectGaussian* gen(va_list& args)
+    static RenderEffectGaussianBlur* gen(va_list& args)
     {
-        auto sigma = (float) va_arg(args, double);
-        if (sigma <= 0) return nullptr;
-
-        auto inst = new RenderEffectGaussian;
-        inst->sigma = sigma;
+        auto inst = new RenderEffectGaussianBlur;
+        inst->sigma = std::max((float) va_arg(args, double), 0.0f);
         inst->direction = std::min(va_arg(args, int), 2);
         inst->border = std::min(va_arg(args, int), 1);
         inst->quality = std::min(va_arg(args, int), 100);
@@ -298,6 +295,30 @@ struct RenderEffectGaussian : RenderEffect
     }
 };
 
+struct RenderEffectDropShadow : RenderEffect
+{
+    uint8_t color[4];  //rgba
+    float angle;
+    float distance;
+    float sigma;
+    uint8_t quality;   //0 ~ 100  (optional)
+
+    static RenderEffectDropShadow* gen(va_list& args)
+    {
+        auto inst = new RenderEffectDropShadow;
+        inst->color[0] = va_arg(args, int);
+        inst->color[1] = va_arg(args, int);
+        inst->color[2] = va_arg(args, int);
+        inst->color[3] = va_arg(args, int);
+        inst->angle = (float) va_arg(args, double);
+        inst->distance = (float) va_arg(args, double);
+        inst->sigma = std::max((float) va_arg(args, double), 0.0f);
+        inst->quality = std::min(va_arg(args, int), 100);
+        inst->type = SceneEffect::DropShadow;
+        return inst;
+    }
+};
+
 class RenderMethod
 {
 private:
@@ -331,7 +352,7 @@ public:
     virtual bool endComposite(RenderCompositor* cmp) = 0;
 
     virtual bool prepare(RenderEffect* effect) = 0;
-    virtual bool effect(RenderCompositor* cmp, const RenderEffect* effect) = 0;
+    virtual bool effect(RenderCompositor* cmp, const RenderEffect* effect, bool direct) = 0;
 };
 
 static inline bool MASK_REGION_MERGING(CompositeMethod method)
diff --git a/src/renderer/tvgScene.cpp b/src/renderer/tvgScene.cpp
index 38a4f058..ce169d33 100644
--- a/src/renderer/tvgScene.cpp
+++ b/src/renderer/tvgScene.cpp
@@ -121,7 +121,11 @@ Result Scene::push(SceneEffect effect, ...) noexcept
 
     switch (effect) {
         case SceneEffect::GaussianBlur: {
-            re = RenderEffectGaussian::gen(args);
+            re = RenderEffectGaussianBlur::gen(args);
+            break;
+        }
+        case SceneEffect::DropShadow: {
+            re = RenderEffectDropShadow::gen(args);
             break;
         }
         default: break;
diff --git a/src/renderer/tvgScene.h b/src/renderer/tvgScene.h
index cb2ef7f6..47d0f5b6 100644
--- a/src/renderer/tvgScene.h
+++ b/src/renderer/tvgScene.h
@@ -143,8 +143,9 @@ struct Scene::Impl
         if (cmp) {
             //Apply post effects if any.
             if (effects) {
+                auto direct = effects->count == 1 ? true : false;
                 for (auto e = effects->begin(); e < effects->end(); ++e) {
-                    renderer->effect(cmp, *e);
+                    renderer->effect(cmp, *e, direct);
                 }
             }
             renderer->endComposite(cmp);
diff --git a/src/renderer/wg_engine/tvgWgRenderer.cpp b/src/renderer/wg_engine/tvgWgRenderer.cpp
index a418b285..867f652a 100755
--- a/src/renderer/wg_engine/tvgWgRenderer.cpp
+++ b/src/renderer/wg_engine/tvgWgRenderer.cpp
@@ -431,7 +431,7 @@ bool WgRenderer::prepare(TVG_UNUSED RenderEffect* effect)
 }
 
 
-bool WgRenderer::effect(TVG_UNUSED RenderCompositor* cmp, TVG_UNUSED const RenderEffect* effect)
+bool WgRenderer::effect(TVG_UNUSED RenderCompositor* cmp, TVG_UNUSED const RenderEffect* effect, TVG_UNUSED bool direct)
 {
     TVGLOG("WG_ENGINE", "SceneEffect(%d) is not supported", (int)effect->type);
     return false;
diff --git a/src/renderer/wg_engine/tvgWgRenderer.h b/src/renderer/wg_engine/tvgWgRenderer.h
index 043f5df2..3b948144 100755
--- a/src/renderer/wg_engine/tvgWgRenderer.h
+++ b/src/renderer/wg_engine/tvgWgRenderer.h
@@ -53,7 +53,7 @@ public:
     bool endComposite(RenderCompositor* cmp) override;
 
     bool prepare(RenderEffect* effect) override;
-    bool effect(RenderCompositor* cmp, const RenderEffect* effect) override;
+    bool effect(RenderCompositor* cmp, const RenderEffect* effect, bool direct) override;
 
     static WgRenderer* gen();
     static bool init(uint32_t threads);