Skip memset() when using internal rasterizer

diff --git a/libass/ass_bitmap.c b/libass/ass_bitmap.c
index 6f77a173c09921f46ac216cdada952b8ebe03148..b14336fa60ee4108a79241faaa4e7ad647d7c81d 100644 (file)
--- a/libass/ass_bitmap.c
+++ b/libass/ass_bitmap.c
@@ -319,20 +319,15 @@ Bitmap *outline_to_bitmap(ASS_Renderer *render_priv,
 
     int tile_w = (w + 2 * bord + mask) & ~mask;
     int tile_h = (h + 2 * bord + mask) & ~mask;
-    Bitmap *bm = alloc_bitmap(tile_w, tile_h);
+    Bitmap *bm = alloc_bitmap_raw(tile_w, tile_h);
     if (!bm)
         return NULL;
     bm->left = x_min - bord;
     bm->top =  y_min - bord;
 
-    int offs = bord & ~mask;
-    if (!rasterizer_fill(rst,
-            bm->buffer + offs * (bm->stride + 1),
-            x_min - bord + offs,
-            y_min - bord + offs,
-            ((w + bord + mask) & ~mask) - offs,
-            ((h + bord + mask) & ~mask) - offs,
-            bm->stride)) {
+    if (!rasterizer_fill(rst, bm->buffer,
+                         x_min - bord, y_min - bord,
+                         bm->stride, tile_h, bm->stride)) {
         ass_msg(render_priv->library, MSGL_WARN, "Failed to rasterize glyph!\n");
         ass_free_bitmap(bm);
         return NULL;

diff --git a/libass/ass_rasterizer.c b/libass/ass_rasterizer.c
index ee19ad4df9ad3c6b9a9b2ed7488381b5dbff79f7..bf217a7df201a2fec7f6b36f29e42614153355d5 100644 (file)
--- a/libass/ass_rasterizer.c
+++ b/libass/ass_rasterizer.c
@@ -592,7 +592,8 @@ static int polyline_split_vert(const struct segment *src, size_t n_src,
 
 
 static inline void rasterizer_fill_solid(ASS_Rasterizer *rst,
-                                         uint8_t *buf, int width, int height, ptrdiff_t stride)
+                                         uint8_t *buf, int width, int height, ptrdiff_t stride,
+                                         int set)
 {
     assert(!(width  & ((1 << rst->tile_order) - 1)));
     assert(!(height & ((1 << rst->tile_order) - 1)));
@@ -604,7 +605,7 @@ static inline void rasterizer_fill_solid(ASS_Rasterizer *rst,
     height >>= rst->tile_order;
     for (j = 0; j < height; ++j) {
         for (i = 0; i < width; ++i)
-            rst->fill_solid(buf + i * step, stride);
+            rst->fill_solid(buf + i * step, stride, set);
         buf += tile_stride;
     }
 }
@@ -637,8 +638,9 @@ static inline void rasterizer_fill_halfplane(ASS_Rasterizer *rst,
             int64_t abs_c = offs_c < 0 ? -offs_c : offs_c;
             if (abs_c < size)
                 rst->fill_halfplane(buf + i * step, stride, a, b, cc, scale);
-            else if (((uint32_t)(offs_c >> 32) ^ scale) & 0x80000000)
-                rst->fill_solid(buf + i * step, stride);
+            else
+                rst->fill_solid(buf + i * step, stride,
+                                ((uint32_t)(offs_c >> 32) ^ scale) & 0x80000000);
         }
         buf += tile_stride;
     }
@@ -664,8 +666,7 @@ static int rasterizer_fill_level(ASS_Rasterizer *rst,
     size_t n = rst->size[index] - offs;
     struct segment *line = rst->linebuf[index] + offs;
     if (!n) {
-        if (winding)
-            rasterizer_fill_solid(rst, buf, width, height, stride);
+        rasterizer_fill_solid(rst, buf, width, height, stride, winding);
         return 1;
     }
     if (n == 1) {
@@ -682,8 +683,8 @@ static int rasterizer_fill_level(ASS_Rasterizer *rst,
             rasterizer_fill_halfplane(rst, buf, width, height, stride,
                                       line->a, line->b, line->c,
                                       flag & 2 ? -line->scale : line->scale);
-        else if (flag & 2)
-            rasterizer_fill_solid(rst, buf, width, height, stride);
+        else
+            rasterizer_fill_solid(rst, buf, width, height, stride, flag & 2);
         rst->size[index] = offs;
         return 1;
     }

diff --git a/libass/ass_rasterizer.h b/libass/ass_rasterizer.h
index 66303178aa7a375da90b551a7e6ac8e2605b956d..d20feb3c771ad828c6175ae6e47d1d629731ba41 100644 (file)
--- a/libass/ass_rasterizer.h
+++ b/libass/ass_rasterizer.h
@@ -43,15 +43,15 @@ struct segment {
 };
 
 
-typedef void (*FillSolidTileFunc)(uint8_t *buf, ptrdiff_t stride);
+typedef void (*FillSolidTileFunc)(uint8_t *buf, ptrdiff_t stride, int set);
 typedef void (*FillHalfplaneTileFunc)(uint8_t *buf, ptrdiff_t stride,
                                       int32_t a, int32_t b, int64_t c, int32_t scale);
 typedef void (*FillGenericTileFunc)(uint8_t *buf, ptrdiff_t stride,
                                     const struct segment *line, size_t n_lines,
                                     int winding);
 
-void ass_fill_solid_tile16_c(uint8_t *buf, ptrdiff_t stride);
-void ass_fill_solid_tile32_c(uint8_t *buf, ptrdiff_t stride);
+void ass_fill_solid_tile16_c(uint8_t *buf, ptrdiff_t stride, int set);
+void ass_fill_solid_tile32_c(uint8_t *buf, ptrdiff_t stride, int set);
 void ass_fill_halfplane_tile16_c(uint8_t *buf, ptrdiff_t stride,
                                  int32_t a, int32_t b, int64_t c, int32_t scale);
 void ass_fill_halfplane_tile32_c(uint8_t *buf, ptrdiff_t stride,

diff --git a/libass/ass_rasterizer_c.c b/libass/ass_rasterizer_c.c
index f15f91b03e6011612860374102d91a90f79dca5e..38d40505e17c2569332d039d42c4269721b7b0dc 100644 (file)
--- a/libass/ass_rasterizer_c.c
+++ b/libass/ass_rasterizer_c.c
@@ -22,10 +22,10 @@
 
 
 
-void ass_fill_solid_tile16_c(uint8_t *buf, ptrdiff_t stride)
+void ass_fill_solid_tile16_c(uint8_t *buf, ptrdiff_t stride, int set)
 {
     int i, j;
-    int8_t value = 255;
+    int8_t value = set ? 255 : 0;
     for (j = 0; j < 16; ++j) {
         for (i = 0; i < 16; ++i)
             buf[i] = value;
@@ -33,10 +33,10 @@ void ass_fill_solid_tile16_c(uint8_t *buf, ptrdiff_t stride)
     }
 }
 
-void ass_fill_solid_tile32_c(uint8_t *buf, ptrdiff_t stride)
+void ass_fill_solid_tile32_c(uint8_t *buf, ptrdiff_t stride, int set)
 {
     int i, j;
-    int8_t value = 255;
+    int8_t value = set ? 255 : 0;
     for (j = 0; j < 32; ++j) {
         for (i = 0; i < 32; ++i)
             buf[i] = value;

diff --git a/libass/x86/rasterizer.asm b/libass/x86/rasterizer.asm
index d4941c6d9f6a72fc0f568f623f711cb5997a37a6..bbb2921e8785baadb758fb99f405b0033e681b29 100644 (file)
--- a/libass/x86/rasterizer.asm
+++ b/libass/x86/rasterizer.asm
@@ -116,12 +116,21 @@ SECTION .text
 
 ;------------------------------------------------------------------------------
 ; FILL_SOLID_TILE tile_order, suffix
-; void fill_solid_tile%2(uint8_t *buf, ptrdiff_t stride);
+; void fill_solid_tile%2(uint8_t *buf, ptrdiff_t stride, int set);
 ;------------------------------------------------------------------------------
 
 %macro FILL_SOLID_TILE 2
-cglobal fill_solid_tile%2, 2,2,1
-    pcmpeqd m0, m0
+cglobal fill_solid_tile%2, 3,4,1
+    mov r3d, -1
+    test r2d, r2d
+    cmovnz r2d, r3d
+    movd xm0, r2d
+%if mmsize == 32
+    vpbroadcastd m0, xm0
+%else
+    pshufd m0, m0, q0000
+%endif
+
 %rep (1 << %1) - 1
     FILL_LINE r0, 0, 1 << %1
     add r0, r1

diff --git a/libass/x86/rasterizer.h b/libass/x86/rasterizer.h
index 11ea3d4f4e13ac9a9973a81bb5d0e0bc5ebfed73..47d72d45e7e7c10448b94483bce8585f9843ab46 100644 (file)
--- a/libass/x86/rasterizer.h
+++ b/libass/x86/rasterizer.h
@@ -25,8 +25,8 @@
 
 struct segment;
 
-void ass_fill_solid_tile16_sse2(uint8_t *buf, ptrdiff_t stride);
-void ass_fill_solid_tile32_sse2(uint8_t *buf, ptrdiff_t stride);
+void ass_fill_solid_tile16_sse2(uint8_t *buf, ptrdiff_t stride, int set);
+void ass_fill_solid_tile32_sse2(uint8_t *buf, ptrdiff_t stride, int set);
 void ass_fill_halfplane_tile16_sse2(uint8_t *buf, ptrdiff_t stride,
                                     int32_t a, int32_t b, int64_t c, int32_t scale);
 void ass_fill_halfplane_tile32_sse2(uint8_t *buf, ptrdiff_t stride,
@@ -38,8 +38,8 @@ void ass_fill_generic_tile32_sse2(uint8_t *buf, ptrdiff_t stride,
                                   const struct segment *line, size_t n_lines,
                                   int winding);
 
-void ass_fill_solid_tile16_avx2(uint8_t *buf, ptrdiff_t stride);
-void ass_fill_solid_tile32_avx2(uint8_t *buf, ptrdiff_t stride);
+void ass_fill_solid_tile16_avx2(uint8_t *buf, ptrdiff_t stride, int set);
+void ass_fill_solid_tile32_avx2(uint8_t *buf, ptrdiff_t stride, int set);
 void ass_fill_halfplane_tile16_avx2(uint8_t *buf, ptrdiff_t stride,
                                     int32_t a, int32_t b, int64_t c, int32_t scale);
 void ass_fill_halfplane_tile32_avx2(uint8_t *buf, ptrdiff_t stride,