fix float <---> uint16_t conversion, with Mike's tests.

BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1679713002

patch from issue 1679713002 at patchset 1 (http://crrev.com/1679713002#ps1)
CQ_EXTRA_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot

Review URL: https://codereview.chromium.org/1676893002

diff --git a/src/opts/SkNx_sse.h b/src/opts/SkNx_sse.h
index ee6fdc5654b26838bf2e9f649741476d8d648055..10db1c438f52271445b0e16cb3238fb1d2f383eb 100644 (file)
--- a/src/opts/SkNx_sse.h
+++ b/src/opts/SkNx_sse.h
@@ -350,7 +350,16 @@ template<> inline Sk4i SkNx_cast<int, float, 4>(const Sk4f& src) {
 
 template<> inline Sk4h SkNx_cast<uint16_t, float, 4>(const Sk4f& src) {
     auto _32 = _mm_cvttps_epi32(src.fVec);
-    return _mm_packus_epi16(_32, _32);
+    // Ideally we'd use _mm_packus_epi32 here.  But that's SSE4.1+.
+#if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
+    // With SSSE3, we can just shuffle the low 2 bytes from each lane right into place.
+    const int _ = ~0;
+    return _mm_shuffle_epi8(_32, _mm_setr_epi8(0,1, 4,5, 8,9, 12,13, _,_,_,_,_,_,_,_));
+#else
+    // With SSE2, we have to emulate _mm_packus_epi32 with _mm_packs_epi32:
+    _32 = _mm_sub_epi32(_32, _mm_set1_epi32((int)0x00008000));
+    return _mm_add_epi16(_mm_packs_epi32(_32, _32), _mm_set1_epi16((short)0x8000));
+#endif
 }
 
 template<> inline Sk4b SkNx_cast<uint8_t, float, 4>(const Sk4f& src) {

diff --git a/tests/SkNxTest.cpp b/tests/SkNxTest.cpp
index fcd113d44dfcdd4674fa3f0c120e461176bcd649..49e920f4b3d8baf27188cae941a9cef4b3b2f6e3 100644 (file)
--- a/tests/SkNxTest.cpp
+++ b/tests/SkNxTest.cpp
@@ -224,3 +224,50 @@ DEF_TEST(SkNx_abs, r) {
     REPORTER_ASSERT(r, fs.kth<2>() == 2.0f);
     REPORTER_ASSERT(r, fs.kth<3>() == 4.0f);
 }
+
+#include "SkRandom.h"
+
+static void dump(const Sk4f& f4, const Sk4h& h4) {
+    SkDebugf("%g %g %g %g --> %d %d %d %d\n",
+             f4.kth<0>(), f4.kth<1>(), f4.kth<2>(), f4.kth<3>(),
+             h4.kth<0>(), h4.kth<1>(), h4.kth<2>(), h4.kth<3>());
+}
+
+DEF_TEST(SkNx_u16_float, r) {
+    {
+        // u16 --> float
+        auto h4 = Sk4h(15, 17, 257, 65535);
+        auto f4 = SkNx_cast<float>(h4);
+        dump(f4, h4);
+        REPORTER_ASSERT(r, f4.kth<0>() == 15.0f);
+        REPORTER_ASSERT(r, f4.kth<1>() == 17.0f);
+        REPORTER_ASSERT(r, f4.kth<2>() == 257.0f);
+        REPORTER_ASSERT(r, f4.kth<3>() == 65535.0f);
+    }
+    {
+        // float -> u16
+        auto f4 = Sk4f(15, 17, 257, 65535);
+        auto h4 = SkNx_cast<uint16_t>(f4);
+        dump(f4, h4);
+        REPORTER_ASSERT(r, h4.kth<0>() == 15);
+        REPORTER_ASSERT(r, h4.kth<1>() == 17);
+        REPORTER_ASSERT(r, h4.kth<2>() == 257);
+        REPORTER_ASSERT(r, h4.kth<3>() == 65535);
+    }
+
+    // starting with any u16 value, we should be able to have a perfect round-trip in/out of floats
+    //
+    SkRandom rand;
+    for (int i = 0; i < 0; ++i) {
+        const uint16_t s16[4] {
+            (uint16_t)rand.nextU16(), (uint16_t)rand.nextU16(),
+            (uint16_t)rand.nextU16(), (uint16_t)rand.nextU16(),
+        };
+        auto u4_0 = Sk4h::Load(s16);
+        auto f4 = SkNx_cast<float>(u4_0);
+        auto u4_1 = SkNx_cast<uint16_t>(f4);
+        uint16_t d16[4];
+        u4_1.store(d16);
+        REPORTER_ASSERT(r, !memcmp(s16, d16, sizeof(s16)));
+    }
+}