improve simd

src/base/ftobjs.c

@@ -3154,7 +3154,7 @@ int conic_to2(FT_GlyphSlot* slot, FT_Vector *control, FT_Vector *from, FT_Vector
       face->garray = (FT_GlyphSlot*)malloc(
           face->driver->clazz->slot_object_size * face->num_glyphs );
       //error           = FT_Set_Char_Size( face, 0, 160 * 64, 300, 300 );
-      error           = FT_Set_Pixel_Sizes( face, 0, 100);
+      error           = FT_Set_Pixel_Sizes( face, 0, 200);
       // int glyph_index = FT_Get_Char_Index( face, 'A' );
       // error           = FT_Load_Glyph( face, glyph_index, FT_LOAD_NO_HINTING );
 

src/dense/ftdense.c

@@ -467,6 +467,7 @@ dense_render_glyph( dense_worker* worker, const FT_Bitmap* target, FT_PreLine pl
 
   __m128i offset = _mm_setzero_si128();
   __m128i nzero = _mm_castps_si128(_mm_set1_ps(-0.0));
+  
 
   for (int i = 0; i < worker->m_h*worker->m_w; i += 4)
   {
@@ -482,13 +483,15 @@ dense_render_glyph( dense_worker* worker, const FT_Bitmap* target, FT_PreLine pl
     x = _mm_add_epi32( x, offset );
 
     __m128i y = _mm_srli_epi32( _mm_abs_epi32( x) , 4 );
-     y = _mm_packs_epi32(y, nzero);
-     y = _mm_packus_epi16(y, nzero);
+     y = _mm_packus_epi16(_mm_packs_epi32(y, nzero), nzero);
 
     // int* ptr = (int*)&dest[i];
     _mm_storeu_si32(&dest[i], y);
 
     offset = _mm_shuffle_epi32(x,_MM_SHUFFLE( 3, 3, 3, 3 ) );
+
+
+
   }
 
 #elif FT_NEON