Optimize VP8SetResidualCoeffs.

Brings down WebP lossy encoding timings by 5%

Change-Id: Ia4a2fab0a887aaaf7841ce6d9ee16270d3e15489

src/dsp/enc_sse2.c

@@ -17,7 +17,9 @@
 #include <stdlib.h>  // for abs()
 #include <emmintrin.h>
 
+#include "../enc/cost.h"
 #include "../enc/vp8enci.h"
+#include "../utils/utils.h"
 
 //------------------------------------------------------------------------------
 // Quite useful macro for debugging. Left here for convenience.
@@ -929,6 +931,33 @@ static int QuantizeBlockWHT(int16_t in[16], int16_t out[16],
   return DoQuantizeBlock(in, out, 0, &mtx->sharpen_[0], mtx);
 }
 
+// Forward declaration.
+void VP8SetResidualCoeffsSSE2(const int16_t* const coeffs,
+                              VP8Residual* const res);
+
+void VP8SetResidualCoeffsSSE2(const int16_t* const coeffs,
+                              VP8Residual* const res) {
+  const __m128i c0 = _mm_loadu_si128((const __m128i*)coeffs);
+  const __m128i c1 = _mm_loadu_si128((const __m128i*)(coeffs + 8));
+  // Use SSE to compare 8 values with a single instruction.
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i m0 = _mm_cmpeq_epi16(c0, zero);
+  const __m128i m1 = _mm_cmpeq_epi16(c1, zero);
+  // Get the comparison results as a bitmask, consisting of two times 16 bits:
+  // two identical bits for each result. Concatenate both bitmasks to get a
+  // single 32 bit value. Negate the mask to get the position of entries that
+  // are not equal to zero. Finally, mask out least significant bits according
+  // to res->first.
+  const uint32_t mask =
+      ~((_mm_movemask_epi8(m1) << 16) | _mm_movemask_epi8(m0)) &
+      -(1U << (res->first << 1));
+  // The position of the most significant non-zero bit indicates the position of
+  // the last non-zero value. Divide the result by two because __movemask_epi8
+  // operates on 8 bit values instead of 16 bit values.
+  res->last = mask ? (BitsLog2Floor(mask) >> 1) : -1;
+  res->coeffs = coeffs;
+}
+
 #endif   // WEBP_USE_SSE2
 
 //------------------------------------------------------------------------------