Revert of Expand _01 half<->float limitation to _finite. Simplify. (patchset #7 id:120001 of https://codereview.chromium.org/2145663003/ )

Reason for revert:
Unit tests fail on Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-Fast

Original issue's description:
> Expand _01 half<->float limitation to _finite.  Simplify.
>
>     It's become clear we need to sometimes deal with values <0 or >1.
>     I'm not yet convinced we care about NaN or +-inf.
>
>     We had some fairly clever tricks and optimizations here for NEON
>     and SSE.  I've thrown them out in favor of a single implementation.
>     If we find the specializations mattered, we can certainly figure out
>     how to extend them to this new range/domain.
>
>     This happens to add a vectorized float -> half for ARMv7, which was
>     missing from the _01 version.  (The SSE strategy was not portable to
>     platforms that flush denorm floats to zero.)
>
>     I've tested the full float range for FloatToHalf on my desktop and a 5x.
>
> BUG=skia:
> GOLD_TRYBOT_URL= https://gold.skia.org/search?issue=2145663003
> CQ_INCLUDE_TRYBOTS=client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot;master.client.skia:Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-SKNX_NO_SIMD-Trybot,Test-Ubuntu-GCC-GCE-CPU-AVX2-x86_64-Release-Fast-Trybot
>
> Committed: https://skia.googlesource.com/skia/+/3296bee70d074bb8094b3229dbe12fa016657e90

TBR=msarett@google.com,mtklein@chromium.org
# Skipping CQ checks because original CL landed less than 1 days ago.
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=skia:

Review-Url: https://codereview.chromium.org/2151023003

1 files changed, 66 insertions, 41 deletions

diff --git a/src/core/SkHalf.h b/src/core/SkHalf.h
index 2f2ed66c6a..5f5575ae1a 100644
--- a/src/core/SkHalf.h
+++ b/src/core/SkHalf.h
@@ -24,10 +24,10 @@ typedef uint16_t SkHalf;
 float SkHalfToFloat(SkHalf h);
 SkHalf SkFloatToHalf(float f);
 
-// Convert between half and single precision floating point,
-// assuming inputs and outputs are both finite.
-static inline     Sk4f SkHalfToFloat_finite(uint64_t);
-static inline uint64_t SkFloatToHalf_finite(const Sk4f&);
+// Convert between half and single precision floating point, but pull any dirty
+// trick we can to make it faster as long as it's correct enough for values in [0,1].
+static inline     Sk4f SkHalfToFloat_01(uint64_t);
+static inline uint64_t SkFloatToHalf_01(const Sk4f&);
 
 // ~~~~~~~~~~~ impl ~~~~~~~~~~~~~~ //
 
@@ -36,7 +36,7 @@ static inline uint64_t SkFloatToHalf_finite(const Sk4f&);
 
 // GCC 4.9 lacks the intrinsics to use ARMv8 f16<->f32 instructions, so we use inline assembly.
 
-static inline Sk4f SkHalfToFloat_finite(uint64_t hs) {
+static inline Sk4f SkHalfToFloat_01(uint64_t hs) {
 #if !defined(SKNX_NO_SIMD) && defined(SK_CPU_ARM64)
     float32x4_t fs;
     asm ("fmov  %d[fs], %[hs]        \n"   // vcreate_f16(hs)
@@ -44,28 +44,53 @@ static inline Sk4f SkHalfToFloat_finite(uint64_t hs) {
         : [fs] "=w" (fs)                   // =w: write-only NEON register
         : [hs] "r" (hs));                  //  r: read-only 64-bit general register
     return fs;
+
+#elif !defined(SKNX_NO_SIMD) && defined(SK_ARM_HAS_NEON)
+    // NEON makes this pretty easy:
+    //   - denormals are 10-bit * 2^-14 == 24-bit fixed point;
+    //   - handle normals the same way as in SSE: align mantissa, then rebias exponent.
+    uint32x4_t h = vmovl_u16(vcreate_u16(hs)),
+               is_denorm = vcltq_u32(h, vdupq_n_u32(1<<10));
+    float32x4_t denorm = vcvtq_n_f32_u32(h, 24),
+                  norm = vreinterpretq_f32_u32(vaddq_u32(vshlq_n_u32(h, 13),
+                                                         vdupq_n_u32((127-15) << 23)));
+    return vbslq_f32(is_denorm, denorm, norm);
+
+#elif !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
+    // If our input is a normal 16-bit float, things are pretty easy:
+    //   - shift left by 13 to put the mantissa in the right place;
+    //   - the exponent is wrong, but it just needs to be rebiased;
+    //   - re-bias the exponent from 15-bias to 127-bias by adding (127-15).
+
+    // If our input is denormalized, we're going to do the same steps, plus a few more fix ups:
+    //   - the input is h = K*2^-14, for some 10-bit fixed point K in [0,1);
+    //   - by shifting left 13 and adding (127-15) to the exponent, we constructed the float value
+    //     2^-15*(1+K);
+    //   - we'd need to subtract 2^-15 and multiply by 2 to get back to K*2^-14, or equivallently
+    //     multiply by 2 then subtract 2^-14.
+    //
+    //   - We'll work that multiply by 2 into the rebias, by adding 1 more to the exponent.
+    //   - Conveniently, this leaves that rebias constant 2^-14, exactly what we want to subtract.
+
+    __m128i h = _mm_unpacklo_epi16(_mm_loadl_epi64((const __m128i*)&hs), _mm_setzero_si128());
+    const __m128i is_denorm = _mm_cmplt_epi32(h, _mm_set1_epi32(1<<10));
+
+    __m128i rebias = _mm_set1_epi32((127-15) << 23);
+    rebias = _mm_add_epi32(rebias, _mm_and_si128(is_denorm, _mm_set1_epi32(1<<23)));
+
+    __m128i f = _mm_add_epi32(_mm_slli_epi32(h, 13), rebias);
+    return _mm_sub_ps(_mm_castsi128_ps(f),
+                      _mm_castsi128_ps(_mm_and_si128(is_denorm, rebias)));
 #else
-    Sk4i bits      = SkNx_cast<int>(Sk4h::Load(&hs)),   // Expand to 32 bit.
-         sign      = bits & 0x00008000,                 // Save the sign bit for later...
-         positive  = bits ^ sign,                       // ...but strip it off for now.
-         is_denorm = positive < (1<<10);                // Exponent == 0?
-
-    // For normal half floats, extend the mantissa by 13 zero bits,
-    // then adjust the exponent from 15 bias to 127 bias.
-    Sk4i norm = (positive << 13) + ((127 - 15) << 23);
-
-    // For denorm half floats, mask in the exponent-only float K that turns our
-    // denorm value V*2^-14 into a normalized float K + V*2^-14.  Then subtract off K.
-    const Sk4i K = ((127-15) + (23-10) + 1) << 23;
-    Sk4i mask_K = positive | K;
-    Sk4f denorm = Sk4f::Load(&mask_K) - Sk4f::Load(&K);
-
-    Sk4i merged = (sign << 16) | is_denorm.thenElse(Sk4i::Load(&denorm), norm);
-    return Sk4f::Load(&merged);
+    float fs[4];
+    for (int i = 0; i < 4; i++) {
+        fs[i] = SkHalfToFloat(hs >> (i*16));
+    }
+    return Sk4f::Load(fs);
 #endif
 }
 
-static inline uint64_t SkFloatToHalf_finite(const Sk4f& fs) {
+static inline uint64_t SkFloatToHalf_01(const Sk4f& fs) {
     uint64_t r;
 #if !defined(SKNX_NO_SIMD) && defined(SK_CPU_ARM64)
     float32x4_t vec = fs.fVec;
@@ -73,25 +98,25 @@ static inline uint64_t SkFloatToHalf_finite(const Sk4f& fs) {
          "fmov  %[r], %d[vec]         \n"   // vst1_f16(&r, ...)
         : [r] "=r" (r)                      // =r: write-only 64-bit general register
         , [vec] "+w" (vec));                // +w: read-write NEON register
+
+// TODO: ARMv7 NEON float->half?
+
+#elif !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
+    // Scale down from 127-bias to 15-bias, then cut off bottom 13 mantissa bits.
+    // This doesn't round, so it can be 1 bit too small.
+    const __m128 rebias = _mm_castsi128_ps(_mm_set1_epi32((127 - (127-15)) << 23));
+    __m128i h = _mm_srli_epi32(_mm_castps_si128(_mm_mul_ps(fs.fVec, rebias)), 13);
+    _mm_storel_epi64((__m128i*)&r, _mm_packs_epi32(h,h));
+
 #else
-    Sk4i bits           = Sk4i::Load(&fs),
-         sign           = bits & 0x80000000,              // Save the sign bit for later...
-         positive       = bits ^ sign,                    // ...but strip it off for now.
-         will_be_denorm = positive < ((127-15+1) << 23);  // positve < smallest normal half?
-
-    // For normal half floats, adjust the exponent from 127 bias to 15 bias,
-    // then drop the bottom 13 mantissa bits.
-    Sk4i norm = (positive - ((127 - 15) << 23)) >> 13;
-
-    // This mechanically inverts the denorm half -> normal float conversion above.
-    // Knowning that and reading its explanation will leave you feeling more confident
-    // than reading my best attempt at explaining this directly.
-    const Sk4i K = ((127-15) + (23-10) + 1) << 23;
-    Sk4f plus_K = Sk4f::Load(&positive) + Sk4f::Load(&K);
-    Sk4i denorm = Sk4i::Load(&plus_K) ^ K;
-
-    Sk4i merged = (sign >> 16) | will_be_denorm.thenElse(denorm, norm);
-    SkNx_cast<uint16_t>(merged).store(&r);
+    SkHalf hs[4];
+    for (int i = 0; i < 4; i++) {
+        hs[i] = SkFloatToHalf(fs[i]);
+    }
+    r = (uint64_t)hs[3] << 48
+      | (uint64_t)hs[2] << 32
+      | (uint64_t)hs[1] << 16
+      | (uint64_t)hs[0] <<  0;
 #endif
     return r;
 }