Provide an SSE2 implementation of CombinedShannonEntropy.

CombinedShannonEntropy takes 30% for lossless compression.
This implementation speeds up the overall process by 2 to 3 %.

Change-Id: I04a71743284c38814fd0726034d51a02b1b6ba8f

src/dsp/lossless.h

@@ -199,9 +199,12 @@ static WEBP_INLINE float VP8LFastSLog2(uint32_t v) {
 typedef double (*VP8LCostFunc)(const uint32_t* population, int length);
 typedef double (*VP8LCostCombinedFunc)(const uint32_t* X, const uint32_t* Y,
                                        int length);
+typedef float (*VP8LCombinedShannonEntropyFunc)(const int X[256],
+                                                const int Y[256]);
 
 extern VP8LCostFunc VP8LExtraCost;
 extern VP8LCostCombinedFunc VP8LExtraCostCombined;
+extern VP8LCombinedShannonEntropyFunc VP8LCombinedShannonEntropy;
 
 typedef struct {        // small struct to hold counters
   int counts[2];        // index: 0=zero steak, 1=non-zero streak

src/dsp/lossless_enc.c

@@ -413,15 +413,15 @@ static float CombinedShannonEntropy(const int X[256], const int Y[256]) {
   int sumX = 0, sumXY = 0;
   for (i = 0; i < 256; ++i) {
     const int x = X[i];
-    const int xy = x + Y[i];
     if (x != 0) {
+      const int xy = x + Y[i];
       sumX += x;
       retval -= VP8LFastSLog2(x);
       sumXY += xy;
       retval -= VP8LFastSLog2(xy);
-    } else if (xy != 0) {
-      sumXY += xy;
-      retval -= VP8LFastSLog2(xy);
+    } else if (Y[i] != 0) {
+      sumXY += Y[i];
+      retval -= VP8LFastSLog2(Y[i]);
     }
   }
   retval += VP8LFastSLog2(sumX) + VP8LFastSLog2(sumXY);
@@ -435,7 +435,7 @@ static float PredictionCostSpatialHistogram(const int accumulated[4][256],
   for (i = 0; i < 4; ++i) {
     const double kExpValue = 0.94;
     retval += PredictionCostSpatial(tile[i], 1, kExpValue);
-    retval += CombinedShannonEntropy(tile[i], accumulated[i]);
+    retval += VP8LCombinedShannonEntropy(tile[i], accumulated[i]);
   }
   return (float)retval;
 }
@@ -894,7 +894,7 @@ static float PredictionCostCrossColor(const int accumulated[256],
   // Favor low entropy, locally and globally.
   // Favor small absolute values for PredictionCostSpatial
   static const double kExpValue = 2.4;
-  return CombinedShannonEntropy(counts, accumulated) +
+  return VP8LCombinedShannonEntropy(counts, accumulated) +
          PredictionCostSpatial(counts, 3, kExpValue);
 }
 
@@ -1269,6 +1269,7 @@ VP8LFastLog2SlowFunc VP8LFastSLog2Slow;
 
 VP8LCostFunc VP8LExtraCost;
 VP8LCostCombinedFunc VP8LExtraCostCombined;
+VP8LCombinedShannonEntropyFunc VP8LCombinedShannonEntropy;
 
 VP8LCostCountFunc VP8LHuffmanCostCount;
 
@@ -1300,6 +1301,7 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInit(void) {
 
   VP8LExtraCost = ExtraCost;
   VP8LExtraCostCombined = ExtraCostCombined;
+  VP8LCombinedShannonEntropy = CombinedShannonEntropy;
 
   VP8LHuffmanCostCount = HuffmanCostCount;
 

src/dsp/lossless_enc_sse2.c

@@ -250,6 +250,80 @@ static void HistogramAdd(const VP8LHistogram* const a,
   }
 }
 
+//------------------------------------------------------------------------------
+// Entropy
+
+// Checks whether the X or Y contribution is worth computing and adding.
+// Used in loop unrolling.
+#define ANALYZE_X_OR_Y(x_or_y, j)                                   \
+  do {                                                              \
+    if (x_or_y[i + j] != 0) retval -= VP8LFastSLog2(x_or_y[i + j]); \
+  } while (0)
+
+// Checks whether the X + Y contribution is worth computing and adding.
+// Used in loop unrolling.
+#define ANALYZE_XY(j)                  \
+  do {                                 \
+    if (tmp[j] != 0) {                 \
+      retval -= VP8LFastSLog2(tmp[j]); \
+      ANALYZE_X_OR_Y(X, j);            \
+    }                                  \
+  } while (0)
+
+static float CombinedShannonEntropy(const int X[256], const int Y[256]) {
+  int i;
+  double retval = 0.;
+  int sumX, sumXY;
+  int32_t tmp[4];
+  __m128i zero = _mm_setzero_si128();
+  // Sums up X + Y, 4 ints at a time (and will merge it at the end for sumXY).
+  __m128i sumXY_128 = zero;
+  __m128i sumX_128 = zero;
+
+  for (i = 0; i < 256; i += 4) {
+    const __m128i x = _mm_loadu_si128((const __m128i*)(X + i));
+    const __m128i y = _mm_loadu_si128((const __m128i*)(Y + i));
+
+    // Check if any X is non-zero: this actually provides a speedup as X is
+    // usually sparse.
+    if (_mm_movemask_epi8(_mm_cmpeq_epi32(x, zero)) != 0xFFFF) {
+      const __m128i xy_128 = _mm_add_epi32(x, y);
+      sumXY_128 = _mm_add_epi32(sumXY_128, xy_128);
+
+      sumX_128 = _mm_add_epi32(sumX_128, x);
+
+      // Analyze the different X + Y.
+      _mm_storeu_si128((__m128i*)tmp, xy_128);
+
+      ANALYZE_XY(0);
+      ANALYZE_XY(1);
+      ANALYZE_XY(2);
+      ANALYZE_XY(3);
+    } else {
+      // X is fully 0, so only deal with Y.
+      sumXY_128 = _mm_add_epi32(sumXY_128, y);
+
+      ANALYZE_X_OR_Y(Y, 0);
+      ANALYZE_X_OR_Y(Y, 1);
+      ANALYZE_X_OR_Y(Y, 2);
+      ANALYZE_X_OR_Y(Y, 3);
+    }
+  }
+
+  // Sum up sumX_128 to get sumX.
+  _mm_storeu_si128((__m128i*)tmp, sumX_128);
+  sumX = tmp[3] + tmp[2] + tmp[1] + tmp[0];
+
+  // Sum up sumXY_128 to get sumXY.
+  _mm_storeu_si128((__m128i*)tmp, sumXY_128);
+  sumXY = tmp[3] + tmp[2] + tmp[1] + tmp[0];
+
+  retval += VP8LFastSLog2(sumX) + VP8LFastSLog2(sumXY);
+  return (float)retval;
+}
+#undef ANALYZE_X_OR_Y
+#undef ANALYZE_XY
+
 //------------------------------------------------------------------------------
 // Entry point
 
@@ -261,6 +335,7 @@ WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitSSE2(void) {
   VP8LCollectColorBlueTransforms = CollectColorBlueTransforms;
   VP8LCollectColorRedTransforms = CollectColorRedTransforms;
   VP8LHistogramAdd = HistogramAdd;
+  VP8LCombinedShannonEntropy = CombinedShannonEntropy;
 }
 
 #else  // !WEBP_USE_SSE2