libwebp: update to 0.5.0

third_party/libwebp/dsp/dec_sse2.c

 // Copyright 2011 Google Inc. All Rights Reserved.
 //
 // Use of this source code is governed by a BSD-style license
 // that can be found in the COPYING file in the root of the source
 // tree. An additional intellectual property rights grant can be found
 // in the file PATENTS. All contributing project authors may
 // be found in the AUTHORS file in the root of the source tree.
 // -----------------------------------------------------------------------------
 //
 // SSE2 version of some decoding functions (idct, loop filtering).
@@ skipping 34 matching lines @@
   //      (x * K) >> 16 = (x * (k + (1 << 16))) >> 16 = ((x * k ) >> 16) + x
   const __m128i k1 = _mm_set1_epi16(20091);
   const __m128i k2 = _mm_set1_epi16(-30068);
   __m128i T0, T1, T2, T3;
 
   // Load and concatenate the transform coefficients (we'll do two transforms
   // in parallel). In the case of only one transform, the second half of the
   // vectors will just contain random value we'll never use nor store.
   __m128i in0, in1, in2, in3;
   {
-    in0 = _mm_loadl_epi64((__m128i*)&in[0]);
-    in1 = _mm_loadl_epi64((__m128i*)&in[4]);
-    in2 = _mm_loadl_epi64((__m128i*)&in[8]);
-    in3 = _mm_loadl_epi64((__m128i*)&in[12]);
+    in0 = _mm_loadl_epi64((const __m128i*)&in[0]);
+    in1 = _mm_loadl_epi64((const __m128i*)&in[4]);
+    in2 = _mm_loadl_epi64((const __m128i*)&in[8]);
+    in3 = _mm_loadl_epi64((const __m128i*)&in[12]);
     // a00 a10 a20 a30   x x x x
     // a01 a11 a21 a31   x x x x
     // a02 a12 a22 a32   x x x x
     // a03 a13 a23 a33   x x x x
     if (do_two) {
-      const __m128i inB0 = _mm_loadl_epi64((__m128i*)&in[16]);
-      const __m128i inB1 = _mm_loadl_epi64((__m128i*)&in[20]);
-      const __m128i inB2 = _mm_loadl_epi64((__m128i*)&in[24]);
-      const __m128i inB3 = _mm_loadl_epi64((__m128i*)&in[28]);
+      const __m128i inB0 = _mm_loadl_epi64((const __m128i*)&in[16]);
+      const __m128i inB1 = _mm_loadl_epi64((const __m128i*)&in[20]);
+      const __m128i inB2 = _mm_loadl_epi64((const __m128i*)&in[24]);
+      const __m128i inB3 = _mm_loadl_epi64((const __m128i*)&in[28]);
       in0 = _mm_unpacklo_epi64(in0, inB0);
       in1 = _mm_unpacklo_epi64(in1, inB1);
       in2 = _mm_unpacklo_epi64(in2, inB2);
       in3 = _mm_unpacklo_epi64(in3, inB3);
       // a00 a10 a20 a30   b00 b10 b20 b30
       // a01 a11 a21 a31   b01 b11 b21 b31
       // a02 a12 a22 a32   b02 b12 b22 b32
       // a03 a13 a23 a33   b03 b13 b23 b33
     }
   }
@@ skipping 122 matching lines @@
     // Load the reference(s).
     __m128i dst0, dst1, dst2, dst3;
     if (do_two) {
       // Load eight bytes/pixels per line.
       dst0 = _mm_loadl_epi64((__m128i*)(dst + 0 * BPS));
       dst1 = _mm_loadl_epi64((__m128i*)(dst + 1 * BPS));
       dst2 = _mm_loadl_epi64((__m128i*)(dst + 2 * BPS));
       dst3 = _mm_loadl_epi64((__m128i*)(dst + 3 * BPS));
     } else {
       // Load four bytes/pixels per line.
-      dst0 = _mm_cvtsi32_si128(*(int*)(dst + 0 * BPS));
-      dst1 = _mm_cvtsi32_si128(*(int*)(dst + 1 * BPS));
-      dst2 = _mm_cvtsi32_si128(*(int*)(dst + 2 * BPS));
-      dst3 = _mm_cvtsi32_si128(*(int*)(dst + 3 * BPS));
+      dst0 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 0 * BPS));
+      dst1 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 1 * BPS));
+      dst2 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 2 * BPS));
+      dst3 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 3 * BPS));
     }
     // Convert to 16b.
     dst0 = _mm_unpacklo_epi8(dst0, zero);
     dst1 = _mm_unpacklo_epi8(dst1, zero);
     dst2 = _mm_unpacklo_epi8(dst2, zero);
     dst3 = _mm_unpacklo_epi8(dst3, zero);
     // Add the inverse transform(s).
     dst0 = _mm_add_epi16(dst0, T0);
     dst1 = _mm_add_epi16(dst1, T1);
     dst2 = _mm_add_epi16(dst2, T2);
     dst3 = _mm_add_epi16(dst3, T3);
     // Unsigned saturate to 8b.
     dst0 = _mm_packus_epi16(dst0, dst0);
     dst1 = _mm_packus_epi16(dst1, dst1);
     dst2 = _mm_packus_epi16(dst2, dst2);
     dst3 = _mm_packus_epi16(dst3, dst3);
     // Store the results.
     if (do_two) {
       // Store eight bytes/pixels per line.
       _mm_storel_epi64((__m128i*)(dst + 0 * BPS), dst0);
       _mm_storel_epi64((__m128i*)(dst + 1 * BPS), dst1);
       _mm_storel_epi64((__m128i*)(dst + 2 * BPS), dst2);
       _mm_storel_epi64((__m128i*)(dst + 3 * BPS), dst3);
     } else {
       // Store four bytes/pixels per line.
-      *(int*)(dst + 0 * BPS) = _mm_cvtsi128_si32(dst0);
-      *(int*)(dst + 1 * BPS) = _mm_cvtsi128_si32(dst1);
-      *(int*)(dst + 2 * BPS) = _mm_cvtsi128_si32(dst2);
-      *(int*)(dst + 3 * BPS) = _mm_cvtsi128_si32(dst3);
+      WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(dst0));
+      WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(dst1));
+      WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(dst2));
+      WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(dst3));
     }
   }
 }
 
 #if defined(USE_TRANSFORM_AC3)
 #define MUL(a, b) (((a) * (b)) >> 16)
 static void TransformAC3(const int16_t* in, uint8_t* dst) {
   static const int kC1 = 20091 + (1 << 16);
   static const int kC2 = 35468;
   const __m128i A = _mm_set1_epi16(in[0] + 4);
   const __m128i c4 = _mm_set1_epi16(MUL(in[4], kC2));
   const __m128i d4 = _mm_set1_epi16(MUL(in[4], kC1));
   const int c1 = MUL(in[1], kC2);
   const int d1 = MUL(in[1], kC1);
   const __m128i CD = _mm_set_epi16(0, 0, 0, 0, -d1, -c1, c1, d1);
   const __m128i B = _mm_adds_epi16(A, CD);
   const __m128i m0 = _mm_adds_epi16(B, d4);
   const __m128i m1 = _mm_adds_epi16(B, c4);
   const __m128i m2 = _mm_subs_epi16(B, c4);
   const __m128i m3 = _mm_subs_epi16(B, d4);
   const __m128i zero = _mm_setzero_si128();
   // Load the source pixels.
-  __m128i dst0 = _mm_cvtsi32_si128(*(int*)(dst + 0 * BPS));
-  __m128i dst1 = _mm_cvtsi32_si128(*(int*)(dst + 1 * BPS));
-  __m128i dst2 = _mm_cvtsi32_si128(*(int*)(dst + 2 * BPS));
-  __m128i dst3 = _mm_cvtsi32_si128(*(int*)(dst + 3 * BPS));
+  __m128i dst0 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 0 * BPS));
+  __m128i dst1 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 1 * BPS));
+  __m128i dst2 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 2 * BPS));
+  __m128i dst3 = _mm_cvtsi32_si128(WebPMemToUint32(dst + 3 * BPS));
   // Convert to 16b.
   dst0 = _mm_unpacklo_epi8(dst0, zero);
   dst1 = _mm_unpacklo_epi8(dst1, zero);
   dst2 = _mm_unpacklo_epi8(dst2, zero);
   dst3 = _mm_unpacklo_epi8(dst3, zero);
   // Add the inverse transform.
   dst0 = _mm_adds_epi16(dst0, _mm_srai_epi16(m0, 3));
   dst1 = _mm_adds_epi16(dst1, _mm_srai_epi16(m1, 3));
   dst2 = _mm_adds_epi16(dst2, _mm_srai_epi16(m2, 3));
   dst3 = _mm_adds_epi16(dst3, _mm_srai_epi16(m3, 3));
   // Unsigned saturate to 8b.
   dst0 = _mm_packus_epi16(dst0, dst0);
   dst1 = _mm_packus_epi16(dst1, dst1);
   dst2 = _mm_packus_epi16(dst2, dst2);
   dst3 = _mm_packus_epi16(dst3, dst3);
   // Store the results.
-  *(int*)(dst + 0 * BPS) = _mm_cvtsi128_si32(dst0);
-  *(int*)(dst + 1 * BPS) = _mm_cvtsi128_si32(dst1);
-  *(int*)(dst + 2 * BPS) = _mm_cvtsi128_si32(dst2);
-  *(int*)(dst + 3 * BPS) = _mm_cvtsi128_si32(dst3);
+  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(dst0));
+  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(dst1));
+  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(dst2));
+  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(dst3));
 }
 #undef MUL
 #endif   // USE_TRANSFORM_AC3
 
 //------------------------------------------------------------------------------
 // Loop Filter (Paragraph 15)
 
 // Compute abs(p - q) = subs(p - q) OR subs(q - p)
 #define MM_ABS(p, q)  _mm_or_si128(                                            \
     _mm_subs_epu8((q), (p)),                                                   \
     _mm_subs_epu8((p), (q)))
 
 // Shift each byte of "x" by 3 bits while preserving by the sign bit.
 static WEBP_INLINE void SignedShift8b(__m128i* const x) {
   const __m128i zero = _mm_setzero_si128();
-  const __m128i signs = _mm_cmpgt_epi8(zero, *x);
-  const __m128i lo_0 = _mm_unpacklo_epi8(*x, signs);  // s8 -> s16 sign extend
-  const __m128i hi_0 = _mm_unpackhi_epi8(*x, signs);
-  const __m128i lo_1 = _mm_srai_epi16(lo_0, 3);
-  const __m128i hi_1 = _mm_srai_epi16(hi_0, 3);
+  const __m128i lo_0 = _mm_unpacklo_epi8(zero, *x);
+  const __m128i hi_0 = _mm_unpackhi_epi8(zero, *x);
+  const __m128i lo_1 = _mm_srai_epi16(lo_0, 3 + 8);
+  const __m128i hi_1 = _mm_srai_epi16(hi_0, 3 + 8);
   *x = _mm_packs_epi16(lo_1, hi_1);
 }
 
 #define FLIP_SIGN_BIT2(a, b) {                                                 \
   a = _mm_xor_si128(a, sign_bit);                                              \
   b = _mm_xor_si128(b, sign_bit);                                              \
 }
 
 #define FLIP_SIGN_BIT4(a, b, c, d) {                                           \
   FLIP_SIGN_BIT2(a, b);                                                        \
   FLIP_SIGN_BIT2(c, d);                                                        \
 }
 
 // input/output is uint8_t
 static WEBP_INLINE void GetNotHEV(const __m128i* const p1,
                                   const __m128i* const p0,
                                   const __m128i* const q0,
                                   const __m128i* const q1,
                                   int hev_thresh, __m128i* const not_hev) {
   const __m128i zero = _mm_setzero_si128();
   const __m128i t_1 = MM_ABS(*p1, *p0);
   const __m128i t_2 = MM_ABS(*q1, *q0);
 
   const __m128i h = _mm_set1_epi8(hev_thresh);
-  const __m128i t_3 = _mm_subs_epu8(t_1, h);  // abs(p1 - p0) - hev_tresh
-  const __m128i t_4 = _mm_subs_epu8(t_2, h);  // abs(q1 - q0) - hev_tresh
+  const __m128i t_max = _mm_max_epu8(t_1, t_2);
 
-  *not_hev = _mm_or_si128(t_3, t_4);
-  *not_hev = _mm_cmpeq_epi8(*not_hev, zero);  // not_hev <= t1 && not_hev <= t2
+  const __m128i t_max_h = _mm_subs_epu8(t_max, h);
+  *not_hev = _mm_cmpeq_epi8(t_max_h, zero);  // not_hev <= t1 && not_hev <= t2
 }
 
 // input pixels are int8_t
 static WEBP_INLINE void GetBaseDelta(const __m128i* const p1,
                                      const __m128i* const p0,
                                      const __m128i* const q0,
                                      const __m128i* const q1,
                                      __m128i* const delta) {
   // beware of addition order, for saturation!
   const __m128i p1_q1 = _mm_subs_epi8(*p1, *q1);   // p1 - q1
@@ skipping 73 matching lines @@
   GetBaseDelta(&p1s, p0, q0, &q1s, &a);
   a = _mm_and_si128(a, mask);     // mask filter values we don't care about
   DoSimpleFilter(p0, q0, &a);
   FLIP_SIGN_BIT2(*p0, *q0);
 }
 
 // Applies filter on 4 pixels (p1, p0, q0 and q1)
 static WEBP_INLINE void DoFilter4(__m128i* const p1, __m128i* const p0,
                                   __m128i* const q0, __m128i* const q1,
                                   const __m128i* const mask, int hev_thresh) {
+  const __m128i zero = _mm_setzero_si128();
   const __m128i sign_bit = _mm_set1_epi8(0x80);
-  const __m128i k64 = _mm_set1_epi8(0x40);
-  const __m128i zero = _mm_setzero_si128();
+  const __m128i k64 = _mm_set1_epi8(64);
+  const __m128i k3 = _mm_set1_epi8(3);
+  const __m128i k4 = _mm_set1_epi8(4);
   __m128i not_hev;
   __m128i t1, t2, t3;
 
   // compute hev mask
   GetNotHEV(p1, p0, q0, q1, hev_thresh, &not_hev);
 
   // convert to signed values
   FLIP_SIGN_BIT4(*p1, *p0, *q0, *q1);
 
   t1 = _mm_subs_epi8(*p1, *q1);        // p1 - q1
   t1 = _mm_andnot_si128(not_hev, t1);  // hev(p1 - q1)
   t2 = _mm_subs_epi8(*q0, *p0);        // q0 - p0
   t1 = _mm_adds_epi8(t1, t2);          // hev(p1 - q1) + 1 * (q0 - p0)
   t1 = _mm_adds_epi8(t1, t2);          // hev(p1 - q1) + 2 * (q0 - p0)
   t1 = _mm_adds_epi8(t1, t2);          // hev(p1 - q1) + 3 * (q0 - p0)
   t1 = _mm_and_si128(t1, *mask);       // mask filter values we don't care about
 
-  t2 = _mm_set1_epi8(3);
-  t3 = _mm_set1_epi8(4);
-  t2 = _mm_adds_epi8(t1, t2);        // 3 * (q0 - p0) + (p1 - q1) + 3
-  t3 = _mm_adds_epi8(t1, t3);        // 3 * (q0 - p0) + (p1 - q1) + 4
+  t2 = _mm_adds_epi8(t1, k3);        // 3 * (q0 - p0) + hev(p1 - q1) + 3
+  t3 = _mm_adds_epi8(t1, k4);        // 3 * (q0 - p0) + hev(p1 - q1) + 4
   SignedShift8b(&t2);                // (3 * (q0 - p0) + hev(p1 - q1) + 3) >> 3
   SignedShift8b(&t3);                // (3 * (q0 - p0) + hev(p1 - q1) + 4) >> 3
   *p0 = _mm_adds_epi8(*p0, t2);      // p0 += t2
   *q0 = _mm_subs_epi8(*q0, t3);      // q0 -= t3
   FLIP_SIGN_BIT2(*p0, *q0);
 
   // this is equivalent to signed (a + 1) >> 1 calculation
   t2 = _mm_add_epi8(t3, sign_bit);
   t3 = _mm_avg_epu8(t2, zero);
   t3 = _mm_sub_epi8(t3, k64);
@@ skipping 48 matching lines @@
     const __m128i a0_lo = _mm_add_epi16(a1_lo, f9_lo);  // Filter * 27 + 63
     const __m128i a0_hi = _mm_add_epi16(a1_hi, f9_hi);  // Filter * 27 + 63
 
     Update2Pixels(p2, q2, &a2_lo, &a2_hi);
     Update2Pixels(p1, q1, &a1_lo, &a1_hi);
     Update2Pixels(p0, q0, &a0_lo, &a0_hi);
   }
 }
 
 // reads 8 rows across a vertical edge.
-//
-// TODO(somnath): Investigate _mm_shuffle* also see if it can be broken into
-// two Load4x4() to avoid code duplication.
 static WEBP_INLINE void Load8x4(const uint8_t* const b, int stride,
                                 __m128i* const p, __m128i* const q) {
-  __m128i t1, t2;
+  // A0 = 63 62 61 60 23 22 21 20 43 42 41 40 03 02 01 00
+  // A1 = 73 72 71 70 33 32 31 30 53 52 51 50 13 12 11 10
+  const __m128i A0 = _mm_set_epi32(
+      WebPMemToUint32(&b[6 * stride]), WebPMemToUint32(&b[2 * stride]),
+      WebPMemToUint32(&b[4 * stride]), WebPMemToUint32(&b[0 * stride]));
+  const __m128i A1 = _mm_set_epi32(
+      WebPMemToUint32(&b[7 * stride]), WebPMemToUint32(&b[3 * stride]),
+      WebPMemToUint32(&b[5 * stride]), WebPMemToUint32(&b[1 * stride]));
 
-  // Load 0th, 1st, 4th and 5th rows
-  __m128i r0 =  _mm_cvtsi32_si128(*((int*)&b[0 * stride]));  // 03 02 01 00
-  __m128i r1 =  _mm_cvtsi32_si128(*((int*)&b[1 * stride]));  // 13 12 11 10
-  __m128i r4 =  _mm_cvtsi32_si128(*((int*)&b[4 * stride]));  // 43 42 41 40
-  __m128i r5 =  _mm_cvtsi32_si128(*((int*)&b[5 * stride]));  // 53 52 51 50
+  // B0 = 53 43 52 42 51 41 50 40 13 03 12 02 11 01 10 00
+  // B1 = 73 63 72 62 71 61 70 60 33 23 32 22 31 21 30 20
+  const __m128i B0 = _mm_unpacklo_epi8(A0, A1);
+  const __m128i B1 = _mm_unpackhi_epi8(A0, A1);
 
-  r0 = _mm_unpacklo_epi32(r0, r4);               // 43 42 41 40 03 02 01 00
-  r1 = _mm_unpacklo_epi32(r1, r5);               // 53 52 51 50 13 12 11 10
-
-  // t1 = 53 43 52 42 51 41 50 40 13 03 12 02 11 01 10 00
-  t1 = _mm_unpacklo_epi8(r0, r1);
-
-  // Load 2nd, 3rd, 6th and 7th rows
-  r0 =  _mm_cvtsi32_si128(*((int*)&b[2 * stride]));          // 23 22 21 22
-  r1 =  _mm_cvtsi32_si128(*((int*)&b[3 * stride]));          // 33 32 31 30
-  r4 =  _mm_cvtsi32_si128(*((int*)&b[6 * stride]));          // 63 62 61 60
-  r5 =  _mm_cvtsi32_si128(*((int*)&b[7 * stride]));          // 73 72 71 70
-
-  r0 = _mm_unpacklo_epi32(r0, r4);               // 63 62 61 60 23 22 21 20
-  r1 = _mm_unpacklo_epi32(r1, r5);               // 73 72 71 70 33 32 31 30
-
-  // t2 = 73 63 72 62 71 61 70 60 33 23 32 22 31 21 30 20
-  t2 = _mm_unpacklo_epi8(r0, r1);
-
-  // t1 = 33 23 13 03 32 22 12 02 31 21 11 01 30 20 10 00
-  // t2 = 73 63 53 43 72 62 52 42 71 61 51 41 70 60 50 40
-  r0 = t1;
-  t1 = _mm_unpacklo_epi16(t1, t2);
-  t2 = _mm_unpackhi_epi16(r0, t2);
+  // C0 = 33 23 13 03 32 22 12 02 31 21 11 01 30 20 10 00
+  // C1 = 73 63 53 43 72 62 52 42 71 61 51 41 70 60 50 40
+  const __m128i C0 = _mm_unpacklo_epi16(B0, B1);
+  const __m128i C1 = _mm_unpackhi_epi16(B0, B1);
 
   // *p = 71 61 51 41 31 21 11 01 70 60 50 40 30 20 10 00
   // *q = 73 63 53 43 33 23 13 03 72 62 52 42 32 22 12 02
-  *p = _mm_unpacklo_epi32(t1, t2);
-  *q = _mm_unpackhi_epi32(t1, t2);
+  *p = _mm_unpacklo_epi32(C0, C1);
+  *q = _mm_unpackhi_epi32(C0, C1);
 }
 
 static WEBP_INLINE void Load16x4(const uint8_t* const r0,
                                  const uint8_t* const r8,
                                  int stride,
                                  __m128i* const p1, __m128i* const p0,
                                  __m128i* const q0, __m128i* const q1) {
-  __m128i t1, t2;
   // Assume the pixels around the edge (|) are numbered as follows
   //                00 01 | 02 03
   //                10 11 | 12 13
   //                 ...  |  ...
   //                e0 e1 | e2 e3
   //                f0 f1 | f2 f3
   //
   // r0 is pointing to the 0th row (00)
   // r8 is pointing to the 8th row (80)
 
   // Load
   // p1 = 71 61 51 41 31 21 11 01 70 60 50 40 30 20 10 00
   // q0 = 73 63 53 43 33 23 13 03 72 62 52 42 32 22 12 02
   // p0 = f1 e1 d1 c1 b1 a1 91 81 f0 e0 d0 c0 b0 a0 90 80
   // q1 = f3 e3 d3 c3 b3 a3 93 83 f2 e2 d2 c2 b2 a2 92 82
   Load8x4(r0, stride, p1, q0);
   Load8x4(r8, stride, p0, q1);
 
-  t1 = *p1;
-  t2 = *q0;
-  // p1 = f0 e0 d0 c0 b0 a0 90 80 70 60 50 40 30 20 10 00
-  // p0 = f1 e1 d1 c1 b1 a1 91 81 71 61 51 41 31 21 11 01
-  // q0 = f2 e2 d2 c2 b2 a2 92 82 72 62 52 42 32 22 12 02
-  // q1 = f3 e3 d3 c3 b3 a3 93 83 73 63 53 43 33 23 13 03
-  *p1 = _mm_unpacklo_epi64(t1, *p0);
-  *p0 = _mm_unpackhi_epi64(t1, *p0);
-  *q0 = _mm_unpacklo_epi64(t2, *q1);
-  *q1 = _mm_unpackhi_epi64(t2, *q1);
+  {
+    // p1 = f0 e0 d0 c0 b0 a0 90 80 70 60 50 40 30 20 10 00
+    // p0 = f1 e1 d1 c1 b1 a1 91 81 71 61 51 41 31 21 11 01
+    // q0 = f2 e2 d2 c2 b2 a2 92 82 72 62 52 42 32 22 12 02
+    // q1 = f3 e3 d3 c3 b3 a3 93 83 73 63 53 43 33 23 13 03
+    const __m128i t1 = *p1;
+    const __m128i t2 = *q0;
+    *p1 = _mm_unpacklo_epi64(t1, *p0);
+    *p0 = _mm_unpackhi_epi64(t1, *p0);
+    *q0 = _mm_unpacklo_epi64(t2, *q1);
+    *q1 = _mm_unpackhi_epi64(t2, *q1);
+  }
 }
 
 static WEBP_INLINE void Store4x4(__m128i* const x, uint8_t* dst, int stride) {
   int i;
   for (i = 0; i < 4; ++i, dst += stride) {
-    *((int32_t*)dst) = _mm_cvtsi128_si32(*x);
+    WebPUint32ToMem(dst, _mm_cvtsi128_si32(*x));
     *x = _mm_srli_si128(*x, 4);
   }
 }
 
 // Transpose back and store
 static WEBP_INLINE void Store16x4(const __m128i* const p1,
                                   const __m128i* const p0,
                                   const __m128i* const q0,
                                   const __m128i* const q1,
                                   uint8_t* r0, uint8_t* r8,
@@ skipping 324 matching lines @@
   MAX_DIFF2(t2, t1, q1, q0, mask);
 
   ComplexMask(&p1, &p0, &q0, &q1, thresh, ithresh, &mask);
   DoFilter4(&p1, &p0, &q0, &q1, &mask, hev_thresh);
 
   u -= 2;  // beginning of p1
   v -= 2;
   Store16x4(&p1, &p0, &q0, &q1, u, v, stride);
 }
 
-#endif   // WEBP_USE_SSE2
+//------------------------------------------------------------------------------
+// 4x4 predictions
+
+#define DST(x, y) dst[(x) + (y) * BPS]
+#define AVG3(a, b, c) (((a) + 2 * (b) + (c) + 2) >> 2)
+
+// We use the following 8b-arithmetic tricks:
+//     (a + 2 * b + c + 2) >> 2 = (AC + b + 1) >> 1
+//   where: AC = (a + c) >> 1 = [(a + c + 1) >> 1] - [(a^c) & 1]
+// and:
+//     (a + 2 * b + c + 2) >> 2 = (AB + BC + 1) >> 1 - (ab|bc)&lsb
+//   where: AC = (a + b + 1) >> 1,   BC = (b + c + 1) >> 1
+//   and ab = a ^ b, bc = b ^ c, lsb = (AC^BC)&1
+
+static void VE4(uint8_t* dst) {    // vertical
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(dst - BPS - 1));
+  const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1);
+  const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 2);
+  const __m128i a = _mm_avg_epu8(ABCDEFGH, CDEFGH00);
+  const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGH00), one);
+  const __m128i b = _mm_subs_epu8(a, lsb);
+  const __m128i avg = _mm_avg_epu8(b, BCDEFGH0);
+  const uint32_t vals = _mm_cvtsi128_si32(avg);
+  int i;
+  for (i = 0; i < 4; ++i) {
+    WebPUint32ToMem(dst + i * BPS, vals);
+  }
+}
+
+static void LD4(uint8_t* dst) {   // Down-Left
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(dst - BPS));
+  const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1);
+  const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 2);
+  const __m128i CDEFGHH0 = _mm_insert_epi16(CDEFGH00, dst[-BPS + 7], 3);
+  const __m128i avg1 = _mm_avg_epu8(ABCDEFGH, CDEFGHH0);
+  const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGHH0), one);
+  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
+  const __m128i abcdefg = _mm_avg_epu8(avg2, BCDEFGH0);
+  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               abcdefg    ));
+  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));
+  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));
+  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));
+}
+
+static void VR4(uint8_t* dst) {   // Vertical-Right
+  const __m128i one = _mm_set1_epi8(1);
+  const int I = dst[-1 + 0 * BPS];
+  const int J = dst[-1 + 1 * BPS];
+  const int K = dst[-1 + 2 * BPS];
+  const int X = dst[-1 - BPS];
+  const __m128i XABCD = _mm_loadl_epi64((__m128i*)(dst - BPS - 1));
+  const __m128i ABCD0 = _mm_srli_si128(XABCD, 1);
+  const __m128i abcd = _mm_avg_epu8(XABCD, ABCD0);
+  const __m128i _XABCD = _mm_slli_si128(XABCD, 1);
+  const __m128i IXABCD = _mm_insert_epi16(_XABCD, I | (X << 8), 0);
+  const __m128i avg1 = _mm_avg_epu8(IXABCD, ABCD0);
+  const __m128i lsb = _mm_and_si128(_mm_xor_si128(IXABCD, ABCD0), one);
+  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
+  const __m128i efgh = _mm_avg_epu8(avg2, XABCD);
+  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               abcd    ));
+  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(               efgh    ));
+  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(abcd, 1)));
+  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_slli_si128(efgh, 1)));
+
+  // these two are hard to implement in SSE2, so we keep the C-version:
+  DST(0, 2) = AVG3(J, I, X);
+  DST(0, 3) = AVG3(K, J, I);
+}
+
+static void VL4(uint8_t* dst) {   // Vertical-Left
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(dst - BPS));
+  const __m128i BCDEFGH_ = _mm_srli_si128(ABCDEFGH, 1);
+  const __m128i CDEFGH__ = _mm_srli_si128(ABCDEFGH, 2);
+  const __m128i avg1 = _mm_avg_epu8(ABCDEFGH, BCDEFGH_);
+  const __m128i avg2 = _mm_avg_epu8(CDEFGH__, BCDEFGH_);
+  const __m128i avg3 = _mm_avg_epu8(avg1, avg2);
+  const __m128i lsb1 = _mm_and_si128(_mm_xor_si128(avg1, avg2), one);
+  const __m128i ab = _mm_xor_si128(ABCDEFGH, BCDEFGH_);
+  const __m128i bc = _mm_xor_si128(CDEFGH__, BCDEFGH_);
+  const __m128i abbc = _mm_or_si128(ab, bc);
+  const __m128i lsb2 = _mm_and_si128(abbc, lsb1);
+  const __m128i avg4 = _mm_subs_epu8(avg3, lsb2);
+  const uint32_t extra_out = _mm_cvtsi128_si32(_mm_srli_si128(avg4, 4));
+  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(               avg1    ));
+  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(               avg4    ));
+  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg1, 1)));
+  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(avg4, 1)));
+
+  // these two are hard to get and irregular
+  DST(3, 2) = (extra_out >> 0) & 0xff;
+  DST(3, 3) = (extra_out >> 8) & 0xff;
+}
+
+static void RD4(uint8_t* dst) {   // Down-right
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i XABCD = _mm_loadl_epi64((__m128i*)(dst - BPS - 1));
+  const __m128i ____XABCD = _mm_slli_si128(XABCD, 4);
+  const uint32_t I = dst[-1 + 0 * BPS];
+  const uint32_t J = dst[-1 + 1 * BPS];
+  const uint32_t K = dst[-1 + 2 * BPS];
+  const uint32_t L = dst[-1 + 3 * BPS];
+  const __m128i LKJI_____ =
+      _mm_cvtsi32_si128(L | (K << 8) | (J << 16) | (I << 24));
+  const __m128i LKJIXABCD = _mm_or_si128(LKJI_____, ____XABCD);
+  const __m128i KJIXABCD_ = _mm_srli_si128(LKJIXABCD, 1);
+  const __m128i JIXABCD__ = _mm_srli_si128(LKJIXABCD, 2);
+  const __m128i avg1 = _mm_avg_epu8(JIXABCD__, LKJIXABCD);
+  const __m128i lsb = _mm_and_si128(_mm_xor_si128(JIXABCD__, LKJIXABCD), one);
+  const __m128i avg2 = _mm_subs_epu8(avg1, lsb);
+  const __m128i abcdefg = _mm_avg_epu8(avg2, KJIXABCD_);
+  WebPUint32ToMem(dst + 3 * BPS, _mm_cvtsi128_si32(               abcdefg    ));
+  WebPUint32ToMem(dst + 2 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 1)));
+  WebPUint32ToMem(dst + 1 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 2)));
+  WebPUint32ToMem(dst + 0 * BPS, _mm_cvtsi128_si32(_mm_srli_si128(abcdefg, 3)));
+}
+
+#undef DST
+#undef AVG3
+
+//------------------------------------------------------------------------------
+// Luma 16x16
+
+static WEBP_INLINE void TrueMotion(uint8_t* dst, int size) {
+  const uint8_t* top = dst - BPS;
+  const __m128i zero = _mm_setzero_si128();
+  int y;
+  if (size == 4) {
+    const __m128i top_values = _mm_cvtsi32_si128(WebPMemToUint32(top));
+    const __m128i top_base = _mm_unpacklo_epi8(top_values, zero);
+    for (y = 0; y < 4; ++y, dst += BPS) {
+      const int val = dst[-1] - top[-1];
+      const __m128i base = _mm_set1_epi16(val);
+      const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero);
+      WebPUint32ToMem(dst, _mm_cvtsi128_si32(out));
+    }
+  } else if (size == 8) {
+    const __m128i top_values = _mm_loadl_epi64((const __m128i*)top);
+    const __m128i top_base = _mm_unpacklo_epi8(top_values, zero);
+    for (y = 0; y < 8; ++y, dst += BPS) {
+      const int val = dst[-1] - top[-1];
+      const __m128i base = _mm_set1_epi16(val);
+      const __m128i out = _mm_packus_epi16(_mm_add_epi16(base, top_base), zero);
+      _mm_storel_epi64((__m128i*)dst, out);
+    }
+  } else {
+    const __m128i top_values = _mm_loadu_si128((const __m128i*)top);
+    const __m128i top_base_0 = _mm_unpacklo_epi8(top_values, zero);
+    const __m128i top_base_1 = _mm_unpackhi_epi8(top_values, zero);
+    for (y = 0; y < 16; ++y, dst += BPS) {
+      const int val = dst[-1] - top[-1];
+      const __m128i base = _mm_set1_epi16(val);
+      const __m128i out_0 = _mm_add_epi16(base, top_base_0);
+      const __m128i out_1 = _mm_add_epi16(base, top_base_1);
+      const __m128i out = _mm_packus_epi16(out_0, out_1);
+      _mm_storeu_si128((__m128i*)dst, out);
+    }
+  }
+}
+
+static void TM4(uint8_t* dst)   { TrueMotion(dst, 4); }
+static void TM8uv(uint8_t* dst) { TrueMotion(dst, 8); }
+static void TM16(uint8_t* dst)  { TrueMotion(dst, 16); }
+
+static void VE16(uint8_t* dst) {
+  const __m128i top = _mm_loadu_si128((const __m128i*)(dst - BPS));
+  int j;
+  for (j = 0; j < 16; ++j) {
+    _mm_storeu_si128((__m128i*)(dst + j * BPS), top);
+  }
+}
+
+static void HE16(uint8_t* dst) {     // horizontal
+  int j;
+  for (j = 16; j > 0; --j) {
+    const __m128i values = _mm_set1_epi8(dst[-1]);
+    _mm_storeu_si128((__m128i*)dst, values);
+    dst += BPS;
+  }
+}
+
+static WEBP_INLINE void Put16(uint8_t v, uint8_t* dst) {
+  int j;
+  const __m128i values = _mm_set1_epi8(v);
+  for (j = 0; j < 16; ++j) {
+    _mm_storeu_si128((__m128i*)(dst + j * BPS), values);
+  }
+}
+
+static void DC16(uint8_t* dst) {    // DC
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top = _mm_loadu_si128((const __m128i*)(dst - BPS));
+  const __m128i sad8x2 = _mm_sad_epu8(top, zero);
+  // sum the two sads: sad8x2[0:1] + sad8x2[8:9]
+  const __m128i sum = _mm_add_epi16(sad8x2, _mm_shuffle_epi32(sad8x2, 2));
+  int left = 0;
+  int j;
+  for (j = 0; j < 16; ++j) {
+    left += dst[-1 + j * BPS];
+  }
+  {
+    const int DC = _mm_cvtsi128_si32(sum) + left + 16;
+    Put16(DC >> 5, dst);
+  }
+}
+
+static void DC16NoTop(uint8_t* dst) {   // DC with top samples not available
+  int DC = 8;
+  int j;
+  for (j = 0; j < 16; ++j) {
+    DC += dst[-1 + j * BPS];
+  }
+  Put16(DC >> 4, dst);
+}
+
+static void DC16NoLeft(uint8_t* dst) {  // DC with left samples not available
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top = _mm_loadu_si128((const __m128i*)(dst - BPS));
+  const __m128i sad8x2 = _mm_sad_epu8(top, zero);
+  // sum the two sads: sad8x2[0:1] + sad8x2[8:9]
+  const __m128i sum = _mm_add_epi16(sad8x2, _mm_shuffle_epi32(sad8x2, 2));
+  const int DC = _mm_cvtsi128_si32(sum) + 8;
+  Put16(DC >> 4, dst);
+}
+
+static void DC16NoTopLeft(uint8_t* dst) {  // DC with no top and left samples
+  Put16(0x80, dst);
+}
+
+//------------------------------------------------------------------------------
+// Chroma
+
+static void VE8uv(uint8_t* dst) {    // vertical
+  int j;
+  const __m128i top = _mm_loadl_epi64((const __m128i*)(dst - BPS));
+  for (j = 0; j < 8; ++j) {
+    _mm_storel_epi64((__m128i*)(dst + j * BPS), top);
+  }
+}
+
+static void HE8uv(uint8_t* dst) {    // horizontal
+  int j;
+  for (j = 0; j < 8; ++j) {
+    const __m128i values = _mm_set1_epi8(dst[-1]);
+    _mm_storel_epi64((__m128i*)dst, values);
+    dst += BPS;
+  }
+}
+
+// helper for chroma-DC predictions
+static WEBP_INLINE void Put8x8uv(uint8_t v, uint8_t* dst) {
+  int j;
+  const __m128i values = _mm_set1_epi8(v);
+  for (j = 0; j < 8; ++j) {
+    _mm_storel_epi64((__m128i*)(dst + j * BPS), values);
+  }
+}
+
+static void DC8uv(uint8_t* dst) {     // DC
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top = _mm_loadl_epi64((const __m128i*)(dst - BPS));
+  const __m128i sum = _mm_sad_epu8(top, zero);
+  int left = 0;
+  int j;
+  for (j = 0; j < 8; ++j) {
+    left += dst[-1 + j * BPS];
+  }
+  {
+    const int DC = _mm_cvtsi128_si32(sum) + left + 8;
+    Put8x8uv(DC >> 4, dst);
+  }
+}
+
+static void DC8uvNoLeft(uint8_t* dst) {   // DC with no left samples
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top = _mm_loadl_epi64((const __m128i*)(dst - BPS));
+  const __m128i sum = _mm_sad_epu8(top, zero);
+  const int DC = _mm_cvtsi128_si32(sum) + 4;
+  Put8x8uv(DC >> 3, dst);
+}
+
+static void DC8uvNoTop(uint8_t* dst) {  // DC with no top samples
+  int dc0 = 4;
+  int i;
+  for (i = 0; i < 8; ++i) {
+    dc0 += dst[-1 + i * BPS];
+  }
+  Put8x8uv(dc0 >> 3, dst);
+}
+
+static void DC8uvNoTopLeft(uint8_t* dst) {    // DC with nothing
+  Put8x8uv(0x80, dst);
+}
 
 //------------------------------------------------------------------------------
 // Entry point
 
 extern void VP8DspInitSSE2(void);
 
-void VP8DspInitSSE2(void) {
-#if defined(WEBP_USE_SSE2)
+WEBP_TSAN_IGNORE_FUNCTION void VP8DspInitSSE2(void) {
   VP8Transform = Transform;
 #if defined(USE_TRANSFORM_AC3)
   VP8TransformAC3 = TransformAC3;
 #endif
 
   VP8VFilter16 = VFilter16;
   VP8HFilter16 = HFilter16;
   VP8VFilter8 = VFilter8;
   VP8HFilter8 = HFilter8;
   VP8VFilter16i = VFilter16i;
   VP8HFilter16i = HFilter16i;
   VP8VFilter8i = VFilter8i;
   VP8HFilter8i = HFilter8i;
 
   VP8SimpleVFilter16 = SimpleVFilter16;
   VP8SimpleHFilter16 = SimpleHFilter16;
   VP8SimpleVFilter16i = SimpleVFilter16i;
   VP8SimpleHFilter16i = SimpleHFilter16i;
-#endif   // WEBP_USE_SSE2
+
+  VP8PredLuma4[1] = TM4;
+  VP8PredLuma4[2] = VE4;
+  VP8PredLuma4[4] = RD4;
+  VP8PredLuma4[5] = VR4;
+  VP8PredLuma4[6] = LD4;
+  VP8PredLuma4[7] = VL4;
+
+  VP8PredLuma16[0] = DC16;
+  VP8PredLuma16[1] = TM16;
+  VP8PredLuma16[2] = VE16;
+  VP8PredLuma16[3] = HE16;
+  VP8PredLuma16[4] = DC16NoTop;
+  VP8PredLuma16[5] = DC16NoLeft;
+  VP8PredLuma16[6] = DC16NoTopLeft;
+
+  VP8PredChroma8[0] = DC8uv;
+  VP8PredChroma8[1] = TM8uv;
+  VP8PredChroma8[2] = VE8uv;
+  VP8PredChroma8[3] = HE8uv;
+  VP8PredChroma8[4] = DC8uvNoTop;
+  VP8PredChroma8[5] = DC8uvNoLeft;
+  VP8PredChroma8[6] = DC8uvNoTopLeft;
 }
+
+#else  // !WEBP_USE_SSE2
+
+WEBP_DSP_INIT_STUB(VP8DspInitSSE2)
+
+#endif  // WEBP_USE_SSE2