sw_engine: introduce avx simd instruction

avx is the cutting edge method for intel & amd cpus simd instruction.
We are going to support this feature for the desktop environment (instead of sse)

You can turn on this with configuration something like this:

$meson . build -Dvectors=avx

Current patch supports only for raster solid color

Change-Id: I068ba30a1f63d480415e2762f8021fc8d6d28a39

meson.build

@@ -18,6 +18,10 @@ if get_option('loaders').contains('svg') == true
     config_h.set10('THORVG_SVG_LOADER_SUPPORT', true)
 endif
 
+if get_option('vectors').contains('avx') == true
+    config_h.set10('THORVG_AVX_VECTOR_SUPPORT', true)
+endif
+
 configure_file(
     output: 'config.h',
     configuration: config_h

meson_options.txt

@@ -6,6 +6,12 @@ option('engines',
 
 option('loaders',
    type: 'array',
-   choices: ['svg'],
+   choices: ['', 'svg'],
    value: ['svg'],
    description: 'Enable Vector File Loader in thorvg')
+
+option('vectors',
+   type: 'array',
+   choices: ['', 'avx'],
+   value: [''],
+   description: 'Enable CPU Vectorization(SIMD) in thorvg')

src/lib/meson.build

@@ -2,10 +2,12 @@ engine_dep = []
 
 if get_option('engines').contains('sw') == true
     subdir('sw_engine')
+    message('Enable SW Raster Engine')
 endif
 
 if get_option('engines').contains('gl') == true
     subdir('gl_engine')
+    message('Enable GL Raster Engine')
 endif
 
 source_file = [

src/lib/sw_engine/tvgSwCommon.h

@@ -19,6 +19,10 @@
 
 #include "tvgCommon.h"
 
+#ifdef THORVG_AVX_VECTOR_SUPPORT
+    #include <immintrin.h>
+#endif
+
 #if 0
 #include <sys/time.h>
 static double timeStamp()
@@ -242,12 +246,6 @@ static inline uint32_t COLOR_ARGB_JOIN(uint8_t r, uint8_t g, uint8_t b, uint8_t
 }
 
 
-static inline void COLOR_SET(uint32_t *dst, uint32_t val, uint32_t len)
-{
-    while (len--) *dst++ = val;
-}
-
-
 int64_t mathMultiply(int64_t a, int64_t b);
 int64_t mathDivide(int64_t a, int64_t b);
 int64_t mathMulDiv(int64_t a, int64_t b, int64_t c);
@@ -283,7 +281,7 @@ void strokeFree(SwStroke* stroke);
 bool fillGenColorTable(SwFill* fill, const Fill* fdata, const Matrix* transform, bool ctable);
 void fillReset(SwFill* fill);
 void fillFree(SwFill* fill);
-void fillFetchLinear(const SwFill* fill, uint32_t* dst, uint32_t y, uint32_t x, uint32_t len);
+void fillFetchLinear(const SwFill* fill, uint32_t* dst, uint32_t y, uint32_t x, uint32_t offset, uint32_t len);
 void fillFetchRadial(const SwFill* fill, uint32_t* dst, uint32_t y, uint32_t x, uint32_t len);
 
 SwRleData* rleRender(const SwOutline* outline, const SwBBox& bbox, const SwSize& clip, bool antiAlias);
@@ -294,4 +292,29 @@ bool rasterSolidShape(Surface& surface, SwShape& shape, uint8_t r, uint8_t g, ui
 bool rasterStroke(Surface& surface, SwShape& shape, uint8_t r, uint8_t g, uint8_t b, uint8_t a);
 bool rasterClear(Surface& surface);
 
+inline void rasterARGB32(uint32_t *dst, uint32_t val, uint32_t offset, int32_t len)
+{
+#ifdef THORVG_AVX_VECTOR_SUPPORT
+    int32_t align = (8 - (offset % 8)) % 8;
+    //Vectorization
+    auto avxDst = (__m256i*)(dst + offset + align);
+    int32_t i = (len - align);
+    for (;i > 7; i -= 8, ++avxDst) {
+       *avxDst = _mm256_set1_epi32(val);
+    }
+    //Alignment
+    if (align > 0) {
+        if (align > len) align -= (align - len);
+        auto tmp = dst + offset;
+        for (; align > 0; --align, ++tmp) *tmp = val;
+    }
+    //Pack Leftovers
+    dst += offset + (len - i);
+    while (i-- > 0) *(dst++) = val;
+#else
+    dst += offset;
+    while (len--) *dst++ = val;
+#endif
+}
+
 #endif /* _TVG_SW_COMMON_H_ */

src/lib/sw_engine/tvgSwFill.cpp

@@ -208,7 +208,7 @@ void fillFetchRadial(const SwFill* fill, uint32_t* dst, uint32_t y, uint32_t x,
 }
 
 
-void fillFetchLinear(const SwFill* fill, uint32_t* dst, uint32_t y, uint32_t x, uint32_t len)
+void fillFetchLinear(const SwFill* fill, uint32_t* dst, uint32_t y, uint32_t x, uint32_t offset, uint32_t len)
 {
     if (fill->linear.len < FLT_EPSILON) return;
 
@@ -220,10 +220,12 @@ void fillFetchLinear(const SwFill* fill, uint32_t* dst, uint32_t y, uint32_t x,
 
     if (fabsf(inc) < FLT_EPSILON) {
         auto color = _fixedPixel(fill, static_cast<uint32_t>(t * FIXPT_SIZE));
-        COLOR_SET(dst, color, len);
+        rasterARGB32(dst, color, offset, len);
         return;
     }
 
+    dst += offset;
+
     auto vMax = static_cast<float>(INT32_MAX >> (FIXPT_BITS + 1));
     auto vMin = -vMax;
     auto v = t + (inc * len);

src/lib/sw_engine/tvgSwRaster.cpp

@@ -19,7 +19,6 @@
 
 #include "tvgSwCommon.h"
 
-
 /************************************************************************/
 /* Internal Class Implementation                                        */
 /************************************************************************/
@@ -55,13 +54,12 @@ static bool _rasterTranslucentRect(Surface& surface, const SwBBox& region, uint3
 
 static bool _rasterSolidRect(Surface& surface, const SwBBox& region, uint32_t color)
 {
-    auto buffer = surface.buffer + (region.min.y * surface.stride) + region.min.x;
-    auto h = static_cast<uint32_t>(region.max.y - region.min.y);
+    auto buffer = surface.buffer + (region.min.y * surface.stride);
     auto w = static_cast<uint32_t>(region.max.x - region.min.x);
+    auto h = static_cast<uint32_t>(region.max.y - region.min.y);
 
     for (uint32_t y = 0; y < h; ++y) {
-        auto dst = &buffer[y * surface.stride];
-        COLOR_SET(dst, color, w);
+        rasterARGB32(buffer + y * surface.stride, color, region.min.x, w);
     }
     return true;
 }
@@ -95,10 +93,10 @@ static bool _rasterSolidRle(Surface& surface, SwRleData* rle, uint32_t color)
     auto span = rle->spans;
 
     for (uint32_t i = 0; i < rle->size; ++i) {
-        auto dst = &surface.buffer[span->y * surface.stride + span->x];
         if (span->coverage == 255) {
-            COLOR_SET(dst, color, span->len);
+            rasterARGB32(surface.buffer + span->y * surface.stride, color, span->x, span->len);
         } else {
+            auto dst = &surface.buffer[span->y * surface.stride + span->x];
             auto src = COLOR_ALPHA_BLEND(color, span->coverage);
             auto ialpha = 255 - span->coverage;
             for (uint32_t i = 0; i < span->len; ++i) {
@@ -127,7 +125,7 @@ static bool _rasterLinearGradientRect(Surface& surface, const SwBBox& region, co
 
         for (uint32_t y = 0; y < h; ++y) {
             auto dst = &buffer[y * surface.stride];
-            fillFetchLinear(fill, tmpBuf, region.min.y + y, region.min.x, w);
+            fillFetchLinear(fill, tmpBuf, region.min.y + y, region.min.x, 0, w);
             for (uint32_t x = 0; x < w; ++x) {
                 dst[x] = tmpBuf[x] + COLOR_ALPHA_BLEND(dst[x], 255 - COLOR_ALPHA(tmpBuf[x]));
             }
@@ -135,8 +133,7 @@ static bool _rasterLinearGradientRect(Surface& surface, const SwBBox& region, co
     //Opaque Gradient
     } else {
         for (uint32_t y = 0; y < h; ++y) {
-            auto dst = &buffer[y * surface.stride];
-            fillFetchLinear(fill, dst, region.min.y + y, region.min.x, w);
+            fillFetchLinear(fill, buffer + y * surface.stride, region.min.y + y, region.min.x, 0, w);
         }
     }
     return true;
@@ -188,7 +185,7 @@ static bool _rasterLinearGradientRle(Surface& surface, SwRleData* rle, const SwF
     if (fill->translucent) {
         for (uint32_t i = 0; i < rle->size; ++i) {
             auto dst = &surface.buffer[span->y * surface.stride + span->x];
-            fillFetchLinear(fill, buf, span->y, span->x, span->len);
+            fillFetchLinear(fill, buf, span->y, span->x, 0, span->len);
             if (span->coverage == 255) {
                 for (uint32_t i = 0; i < span->len; ++i) {
                     dst[i] = buf[i] + COLOR_ALPHA_BLEND(dst[i], 255 - COLOR_ALPHA(buf[i]));
@@ -204,11 +201,11 @@ static bool _rasterLinearGradientRle(Surface& surface, SwRleData* rle, const SwF
     //Opaque Gradient
     } else {
         for (uint32_t i = 0; i < rle->size; ++i) {
-            auto dst = &surface.buffer[span->y * surface.stride + span->x];
             if (span->coverage == 255) {
-                fillFetchLinear(fill, dst, span->y, span->x, span->len);
+                fillFetchLinear(fill, surface.buffer + span->y * surface.stride, span->y, span->x, span->x, span->len);
             } else {
-                fillFetchLinear(fill, buf, span->y, span->x, span->len);
+                auto dst = &surface.buffer[span->y * surface.stride + span->x];
+                fillFetchLinear(fill, buf, span->y, span->x, 0, span->len);
                 auto ialpha = 255 - span->coverage;
                 for (uint32_t i = 0; i < span->len; ++i) {
                     dst[i] = COLOR_ALPHA_BLEND(buf[i], span->coverage) + COLOR_ALPHA_BLEND(dst[i], ialpha);
@@ -313,10 +310,10 @@ bool rasterClear(Surface& surface)
     if (!surface.buffer || surface.stride <= 0 || surface.w <= 0 || surface.h <= 0) return false;
 
     if (surface.w == surface.stride) {
-        COLOR_SET(surface.buffer, 0xff000000, surface.w * surface.h);
+        rasterARGB32(surface.buffer, 0x00000000, 0, surface.w * surface.h);
     } else {
         for (uint32_t i = 0; i < surface.h; i++) {
-            COLOR_SET(surface.buffer + surface.stride * i, 0xff000000, surface.w);
+            rasterARGB32(surface.buffer + surface.stride * i, 0x00000000, 0, surface.w);
         }
     }
     return true;

src/loaders/meson.build

@@ -2,6 +2,7 @@ subloader_dep = []
 
 if get_option('loaders').contains('svg') == true
     subdir('svg_loader')
+    message('Enable SVG Loader')
 endif
 
 loader_dep = declare_dependency(

src/meson.build

@@ -1,5 +1,14 @@
 compiler_flags = ['-DTVG_BUILD']
 
+cc = meson.get_compiler('cpp')
+if (cc.get_id() != 'msvc')
+    if get_option('vectors').contains('avx')
+        compiler_flags += ['-mavx']
+        message('Enable Advanced Vector Extension')
+    endif
+endif
+
+
 subdir('lib')
 subdir('loaders')
 subdir('examples')