[dense] Re-enable SIMD to work with fixed-point

* src/dense/ftdense.c: Use integer SIMD functions for accumulation

* src/dense/ftdense.h: Change types of FT26D6, FT20D12 to better fit
their usage

2 files changed, 45 insertions, 24 deletions

diff --git a/src/dense/ftdense.c b/src/dense/ftdense.c
index c39d43e2a..1e1eeb4e9 100644
--- a/src/dense/ftdense.c
+++ b/src/dense/ftdense.c
@@ -24,7 +24,7 @@
 
 #if FT_SSE4_1
 
-    #include <tmmintrin.h>
+    #include <immintrin.h>
 
 #endif
 
@@ -384,26 +384,47 @@ dense_render_glyph( dense_worker* worker, const FT_Bitmap* target )
   unsigned char* dest     = target->buffer;
   unsigned char* dest_end = target->buffer + worker->m_w * worker->m_h;
 
-//#if FT_SSE4_1
-
-  // __m128 offset = _mm_setzero_ps();
-  // __m128i mask = _mm_set1_epi32(0x0c080400);
-  // __m128 sign_mask = _mm_set1_ps(-0.f);
-  // for (int i = 0; i < worker->m_h*worker->m_w; i += 4) {
-  //   __m128 x = _mm_load_ps(&source[i]);
-  //   x = _mm_add_ps(x, _mm_castsi128_ps(_mm_slli_si128(_mm_castps_si128(x), 4)));
-  //   x = _mm_add_ps(x, _mm_shuffle_ps(_mm_setzero_ps(), x, 0x40));
-  //   x = _mm_add_ps(x, offset);
-  //   __m128 y = _mm_andnot_ps(sign_mask, x);  // fabs(x)
-  //   y = _mm_min_ps(y, _mm_set1_ps(1.0f));
-  //   y = _mm_mul_ps(y, _mm_set1_ps(255.0f));
-  //   __m128i z = _mm_cvtps_epi32(y);
-  //   z = _mm_shuffle_epi8(z, mask);
-  //   _mm_store_ss((float *)&dest[i], (__m128)z);
-  //   offset = _mm_shuffle_ps(x, x, _MM_SHUFFLE(3, 3, 3, 3));
-  // }
-
-//#else /* FT_SSE4_1 */
+#if FT_SSE4_1
+
+__m128i offset = _mm_setzero_si128();
+  __m128i mask   = _mm_set1_epi32( 0x0c080400 );
+
+  for (int i = 0; i < worker->m_h*worker->m_w; i += 4)
+  {
+    // load 4 floats from source
+
+    __m128i x = _mm_load_si128( (__m128i*)&source[i] );
+
+    x = _mm_add_epi32( x, _mm_slli_si128( x, 4 ) );
+
+    x = _mm_add_epi32(
+        x, _mm_castps_si128( _mm_shuffle_ps( _mm_setzero_ps(),
+                                             _mm_castsi128_ps( x ), 0x40 ) ) );
+
+    // add the prefsum of previous 4 floats to all current floats
+    x = _mm_add_epi32( x, offset );
+
+    // take absolute value
+    __m128i y = _mm_abs_epi32( x );  // fabs(x)
+
+    // cap max value to 1
+    y = _mm_min_epi32( y, _mm_set1_epi32( 4080 ) );
+
+    // reduce to 255
+    y = _mm_srli_epi32( y, 4 );
+
+    // shuffle
+    y = _mm_shuffle_epi8( y, mask );
+
+    _mm_store_ss( (float*)&dest[i], (__m128)y );
+
+    // store the current prefix sum in offset
+    offset = _mm_castps_si128( _mm_shuffle_ps( _mm_castsi128_ps( x ),
+                                               _mm_castsi128_ps( x ),
+                                               _MM_SHUFFLE( 3, 3, 3, 3 ) ) );
+  }
+
+#else /* FT_SSE4_1 */
 
   FT20D12 value = 0;
 
@@ -422,7 +443,7 @@ dense_render_glyph( dense_worker* worker, const FT_Bitmap* target )
     dest++;
   }
 
-//#endif /* FT_SSE4_1 */
+#endif /* FT_SSE4_1 */
 
   free(worker->m_a);
   return error;
diff --git a/src/dense/ftdense.h b/src/dense/ftdense.h
index 0af2bb87f..e3f8f1997 100644
--- a/src/dense/ftdense.h
+++ b/src/dense/ftdense.h
@@ -20,8 +20,8 @@ extern "C"
 #endif
 
 
-  typedef signed long long FT26D6;            /* 26.6 fixed-point representation  */
-  typedef signed long long FT20D12;           /* 20.12 fixed-point representation  */
+  typedef signed long FT26D6;            /* 26.6 fixed-point representation  */
+  typedef signed int FT20D12;            /* 20.12 fixed-point representation  */
 
   typedef struct
   {