new version of SkHalfToFloat_01
This is a little faster than the previous version, and much better explained.
BUG=skia:
GOLD_TRYBOT_URL= https://gold.skia.org/search2?unt=true&query=source_type%3Dgm&master=false&issue=1688233002
Review URL: https://codereview.chromium.org/1688233002
diff --git a/src/core/SkHalf.h b/src/core/SkHalf.h
index 3937343..a42f57d 100644
--- a/src/core/SkHalf.h
+++ b/src/core/SkHalf.h
@@ -37,23 +37,30 @@
// TODO: NEON versions
static inline Sk4f SkHalfToFloat_01(uint64_t hs) {
#if !defined(SKNX_NO_SIMD) && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
- // Load our 16-bit floats into the bottom 16 bits of each 32-bit lane, with zeroes on top.
+ // 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));
- // Fork into two paths, depending on whether the 16-bit float is denormalized.
- __m128 is_denorm = _mm_castsi128_ps(_mm_cmplt_epi32(h, _mm_set1_epi32(0x0400)));
+ __m128i rebias = _mm_set1_epi32((127-15) << 23);
+ rebias = _mm_add_epi32(rebias, _mm_and_si128(is_denorm, _mm_set1_epi32(1<<23)));
- // TODO: figure out, explain
- const __m128 half = _mm_set1_ps(0.5f);
- __m128 denorm = _mm_sub_ps(_mm_or_ps(_mm_castsi128_ps(h), half), half);
-
- // If we're normalized, just shift ourselves so the exponent/mantissa dividing line
- // is correct, then re-bias the exponent from 15 to 127.
- __m128 norm = _mm_castsi128_ps(_mm_add_epi32(_mm_slli_epi32(h, 13),
- _mm_set1_epi32((127-15) << 23)));
-
- return _mm_or_ps(_mm_and_ps (is_denorm, denorm),
- _mm_andnot_ps(is_denorm, norm));
+ __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
float fs[4];
for (int i = 0; i < 4; i++) {
@@ -68,8 +75,8 @@
// Scale our floats down by a tiny power of 2 to pull up our mantissa bits,
// then shift back down to 16-bit float layout. This doesn't round, so can be 1 bit small.
// TODO: understand better. Why this scale factor?
- const __m128 scale = _mm_castsi128_ps(_mm_set1_epi32(15 << 23));
- __m128i h = _mm_srli_epi32(_mm_castps_si128(_mm_mul_ps(fs.fVec, scale)), 13);
+ 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);
uint64_t r;
_mm_storel_epi64((__m128i*)&r, _mm_packs_epi32(h,h));
diff --git a/tests/Float16Test.cpp b/tests/Float16Test.cpp
index f96d904..3b13533 100644
--- a/tests/Float16Test.cpp
+++ b/tests/Float16Test.cpp
@@ -66,10 +66,23 @@
REPORTER_ASSERT(reporter, 0 == memcmp(fscratch, fs, sizeof(fs)));
}
+static uint32_t u(float f) {
+ uint32_t x;
+ memcpy(&x, &f, 4);
+ return x;
+}
+
DEF_TEST(HalfToFloat_01, r) {
for (uint16_t h = 0; h < 0x8000; h++) {
float f = SkHalfToFloat(h);
if (f >= 0 && f <= 1) {
+ float got = SkHalfToFloat_01(h)[0];
+ if (got != f) {
+ SkDebugf("0x%04x -> 0x%08x (%g), want 0x%08x (%g)\n",
+ h,
+ u(got), got,
+ u(f), f);
+ }
REPORTER_ASSERT(r, SkHalfToFloat_01(h)[0] == f);
REPORTER_ASSERT(r, SkFloatToHalf_01(SkHalfToFloat_01(h)) == h);
}