neon: GetResidualCost

Direct copy of sse2. Slight improvement because neon has
abs().

flower.ppm had minimal improvement. Somewhat expected because
GetResidualCost_C is only ~3.6%

mug.ppm had a better improvement because GetResidualCost_C is
almost 9%.

C    2.150
NEON 2.130

BUG=b/118740850

Change-Id: Ibc0dd97a81596635f5599cf568205974b4fd2597

src/dsp/cost_neon.c

@@ -48,6 +48,63 @@ static void SetResidualCoeffs_NEON(const int16_t* const coeffs,
   res->coeffs = coeffs;
 }
 
+static int GetResidualCost_NEON(int ctx0, const VP8Residual* const res) {
+  uint8_t levels[16], ctxs[16];
+  uint16_t abs_levels[16];
+  int n = res->first;
+  // should be prob[VP8EncBands[n]], but it's equivalent for n=0 or 1
+  const int p0 = res->prob[n][ctx0][0];
+  CostArrayPtr const costs = res->costs;
+  const uint16_t* t = costs[n][ctx0];
+  // bit_cost(1, p0) is already incorporated in t[] tables, but only if ctx != 0
+  // (as required by the syntax). For ctx0 == 0, we need to add it here or it'll
+  // be missing during the loop.
+  int cost = (ctx0 == 0) ? VP8BitCost(1, p0) : 0;
+
+  if (res->last < 0) {
+    return VP8BitCost(0, p0);
+  }
+
+  {   // precompute clamped levels and contexts, packed to 8b.
+    const uint8x16_t kCst2 = vdupq_n_u8(2);
+    const uint8x16_t kCst67 = vdupq_n_u8(MAX_VARIABLE_LEVEL);
+    const int16x8_t c0 = vld1q_s16(res->coeffs);
+    const int16x8_t c1 = vld1q_s16(res->coeffs + 8);
+    const uint16x8_t E0 = vreinterpretq_u16_s16(vabsq_s16(c0));
+    const uint16x8_t E1 = vreinterpretq_u16_s16(vabsq_s16(c1));
+    const uint8x16_t F = vcombine_u8(vqmovn_u16(E0), vqmovn_u16(E1));
+    const uint8x16_t G = vminq_u8(F, kCst2);   // context = 0,1,2
+    const uint8x16_t H = vminq_u8(F, kCst67);  // clamp_level in [0..67]
+
+    vst1q_u8(ctxs, G);
+    vst1q_u8(levels, H);
+
+    vst1q_u16(abs_levels, E0);
+    vst1q_u16(abs_levels + 8, E1);
+  }
+  for (; n < res->last; ++n) {
+    const int ctx = ctxs[n];
+    const int level = levels[n];
+    const int flevel = abs_levels[n];   // full level
+    cost += VP8LevelFixedCosts[flevel] + t[level];  // simplified VP8LevelCost()
+    t = costs[n + 1][ctx];
+  }
+  // Last coefficient is always non-zero
+  {
+    const int level = levels[n];
+    const int flevel = abs_levels[n];
+    assert(flevel != 0);
+    cost += VP8LevelFixedCosts[flevel] + t[level];
+    if (n < 15) {
+      const int b = VP8EncBands[n + 1];
+      const int ctx = ctxs[n];
+      const int last_p0 = res->prob[b][ctx][0];
+      cost += VP8BitCost(0, last_p0);
+    }
+  }
+  return cost;
+}
+
 //------------------------------------------------------------------------------
 // Entry point
 
@@ -55,6 +112,7 @@ extern void VP8EncDspCostInitNEON(void);
 
 WEBP_TSAN_IGNORE_FUNCTION void VP8EncDspCostInitNEON(void) {
   VP8SetResidualCoeffs = SetResidualCoeffs_NEON;
+  VP8GetResidualCost = GetResidualCost_NEON;
 }
 
 #else  // !WEBP_USE_NEON