commit 1443c6920c4b7aa80811c30ed9cdc81395d5df4f
author mtklein <mtklein@chromium.org> Wed Mar 23 09:52:13 2016 -0700
committer Commit bot <commit-bot@chromium.org> Wed Mar 23 09:52:13 2016 -0700
tree efa88cb70147b55260017171c4ed903b90b6b8bb
parent ec950c4d215c6141b31b98ed6a8182a97baeb8b9

Sk4x4f: Simplify x86 down to SSE2.

  - This drops the minimum requirement for Sk4x4f on x86 to SSE2 by
    removing calls to _mm_shuffle_epi8().  Instead we use good old
    shifting and masking.

  - Performance is very similar to SSSE3, close enough I'm having trouble
    telling which is faster.  I think we should let ourselves circle back
    on whether we need an SSSE3 version later.  When possible it's nice
    to stick to SSE2: it's most available, and performs most uniformly
    across different chips.

This makes Sk4x4f fast on Windows and Linux, and may help mobile x86.

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1817353005

Review URL: https://codereview.chromium.org/1817353005

src/core/Sk4x4f.h

diff --git a/src/core/Sk4x4f.h b/src/core/Sk4x4f.h
index f7c0566..d289334 100644
--- a/src/core/Sk4x4f.h
+++ b/src/core/Sk4x4f.h
@@ -22,9 +22,9 @@
     void transpose(uint8_t[16]) const;
 };
 
-// TODO: SSE2, NEON
+// TODO: NEON
 
-#if 1 && !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
+#if 1 && !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
 
 inline Sk4x4f Sk4x4f::Transpose(const Sk4f& x, const Sk4f& y, const Sk4f& z, const Sk4f& w) {
     auto r = x.fVec,
@@ -41,15 +41,12 @@
 
 inline Sk4x4f Sk4x4f::Transpose(const uint8_t bs[16]) {
     auto b16 = _mm_loadu_si128((const __m128i*)bs);
-    auto _ = ~0;  // Shuffles in a zero byte.
-    auto r = _mm_cvtepi32_ps(
-            _mm_shuffle_epi8(b16, _mm_setr_epi8(0,_,_,_,4,_,_,_, 8,_,_,_,12,_,_,_)));
-    auto g = _mm_cvtepi32_ps(
-            _mm_shuffle_epi8(b16, _mm_setr_epi8(1,_,_,_,5,_,_,_, 9,_,_,_,13,_,_,_)));
-    auto b = _mm_cvtepi32_ps(
-            _mm_shuffle_epi8(b16, _mm_setr_epi8(2,_,_,_,6,_,_,_,10,_,_,_,14,_,_,_)));
-    auto a = _mm_cvtepi32_ps(
-            _mm_shuffle_epi8(b16, _mm_setr_epi8(3,_,_,_,7,_,_,_,11,_,_,_,15,_,_,_)));
+
+    auto mask = _mm_set1_epi32(0xFF);
+    auto r = _mm_cvtepi32_ps(_mm_and_si128(mask,               (b16    ))),
+         g = _mm_cvtepi32_ps(_mm_and_si128(mask, _mm_srli_epi32(b16,  8))),
+         b = _mm_cvtepi32_ps(_mm_and_si128(mask, _mm_srli_epi32(b16, 16))),
+         a = _mm_cvtepi32_ps(                    _mm_srli_epi32(b16, 24));
     return { r,g,b,a };
 }
 
@@ -75,14 +72,11 @@
 }
 
 inline void Sk4x4f::transpose(uint8_t bs[16]) const {
-    auto packed = _mm_packus_epi16(_mm_packus_epi16(_mm_cvttps_epi32(r.fVec),
-                                                    _mm_cvttps_epi32(g.fVec)),
-                                   _mm_packus_epi16(_mm_cvttps_epi32(b.fVec),
-                                                    _mm_cvttps_epi32(a.fVec)));
-    _mm_storeu_si128((__m128i*)bs, _mm_shuffle_epi8(packed, _mm_setr_epi8(0, 4,  8, 12,
-                                                                          1, 5,  9, 13,
-                                                                          2, 6, 10, 14,
-                                                                          3, 7, 11, 15)));
+    auto R =                _mm_cvttps_epi32(r.fVec),
+         G = _mm_slli_epi32(_mm_cvttps_epi32(g.fVec),  8),
+         B = _mm_slli_epi32(_mm_cvttps_epi32(b.fVec), 16),
+         A = _mm_slli_epi32(_mm_cvttps_epi32(a.fVec), 24);
+    _mm_storeu_si128((__m128i*)bs, _mm_or_si128(A, _mm_or_si128(B, _mm_or_si128(G, R))));
 }
 
 #else