From 5cdcd561e261fa4be5ac8c057f36a4146bf63048 Mon Sep 17 00:00:00 2001
From: James Zern <jzern@google.com>
Date: Tue, 4 Aug 2015 23:09:53 -0700
Subject: [PATCH] lossless_enc_neon: add VP8LTransformColor

based on SSE2, ~32% faster

Change-Id: I76f4d6fe456baceba46ffebf2f699e98691eefdf
---
 src/dsp/lossless_enc_neon.c | 57 +++++++++++++++++++++++++++++++++++++
 1 file changed, 57 insertions(+)

diff --git a/src/dsp/lossless_enc_neon.c b/src/dsp/lossless_enc_neon.c
index ee634b93..4c56f259 100644
--- a/src/dsp/lossless_enc_neon.c
+++ b/src/dsp/lossless_enc_neon.c
@@ -68,6 +68,62 @@ static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) {
   VP8LSubtractGreenFromBlueAndRed_C(argb_data, num_pixels & 3);
 }
 
+//------------------------------------------------------------------------------
+// Color Transform
+
+static void TransformColor(const VP8LMultipliers* const m,
+                           uint32_t* argb_data, int num_pixels) {
+  // sign-extended multiplying constants, pre-shifted by 6.
+#define CST(X)  (((int16_t)(m->X << 8)) >> 6)
+  const int16_t rb[8] = {
+    CST(green_to_blue_), CST(green_to_red_),
+    CST(green_to_blue_), CST(green_to_red_),
+    CST(green_to_blue_), CST(green_to_red_),
+    CST(green_to_blue_), CST(green_to_red_)
+  };
+  const int16x8_t mults_rb = vld1q_s16(rb);
+  const int16_t b2[8] = {
+    0, CST(red_to_blue_), 0, CST(red_to_blue_),
+    0, CST(red_to_blue_), 0, CST(red_to_blue_),
+  };
+  const int16x8_t mults_b2 = vld1q_s16(b2);
+#undef CST
+#ifdef USE_VTBLQ
+  static const uint8_t kg0g0[16] = {
+    255, 1, 255, 1, 255, 5, 255, 5, 255, 9, 255, 9, 255, 13, 255, 13
+  };
+  const uint8x16_t shuffle = vld1q_u8(kg0g0);
+#else
+  static const uint8_t k0g0g[8] = { 255, 1, 255, 1, 255, 5, 255, 5 };
+  const uint8x8_t shuffle = vld1_u8(k0g0g);
+#endif
+  const uint32x4_t mask_rb = vdupq_n_u32(0x00ff00ffu);  // red-blue masks
+  int i;
+  for (i = 0; i + 4 <= num_pixels; i += 4) {
+    const uint8x16_t in = vld1q_u8((uint8_t*)(argb_data + i));
+    // 0 g 0 g
+    const uint8x16_t greens = DoGreenShuffle(in, shuffle);
+    // x dr  x db1
+    const int16x8_t A = vqdmulhq_s16(vreinterpretq_s16_u8(greens), mults_rb);
+    // r 0   b   0
+    const int16x8_t B = vshlq_n_s16(vreinterpretq_s16_u8(in), 8);
+    // x db2 0   0
+    const int16x8_t C = vqdmulhq_s16(B, mults_b2);
+    // 0 0   x db2
+    const uint32x4_t D = vshrq_n_u32(vreinterpretq_u32_s16(C), 16);
+    // x dr  x  db
+    const int8x16_t E = vaddq_s8(vreinterpretq_s8_u32(D),
+                                 vreinterpretq_s8_s16(A));
+    // 0 dr  0  db
+    const uint32x4_t F = vandq_u32(vreinterpretq_u32_s8(E), mask_rb);
+    const int8x16_t out = vsubq_s8(vreinterpretq_s8_u8(in),
+                                   vreinterpretq_s8_u32(F));
+    vst1q_s8((int8_t*)(argb_data + i), out);
+  }
+  // fallthrough and finish off with plain-C
+  VP8LTransformColor_C(m, argb_data + i, num_pixels - i);
+}
+
 #undef USE_VTBLQ
 
 //------------------------------------------------------------------------------
@@ -77,6 +133,7 @@ extern void VP8LEncDspInitNEON(void);
 
 WEBP_TSAN_IGNORE_FUNCTION void VP8LEncDspInitNEON(void) {
   VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed;
+  VP8LTransformColor = TransformColor;
 }
 
 #else  // !WEBP_USE_NEON