f16<->f32 ftz is an optional thing for speed.

src/core/SkHalf.h

 /*
  * Copyright 2014 Google Inc.
  *
  * Use of this source code is governed by a BSD-style license that can be
  * found in the LICENSE file.
  */
 
 #ifndef SkHalf_DEFINED
 #define SkHalf_DEFINED
 
 #include "SkNx.h"
 #include "SkTypes.h"
 
 // 16-bit floating point value
 // format is 1 bit sign, 5 bits exponent, 10 bits mantissa
 // only used for storage
 typedef uint16_t SkHalf;
 
 static constexpr uint16_t SK_HalfMin     = 0x0400; // 2^-24  (minimum positive normal value)
 static constexpr uint16_t SK_HalfMax     = 0x7bff; // 65504
 static constexpr uint16_t SK_HalfEpsilon = 0x1400; // 2^-10
 static constexpr uint16_t SK_Half1       = 0x3C00; // 1
 
 // convert between half and single precision floating point
 float SkHalfToFloat(SkHalf h);
 SkHalf SkFloatToHalf(float f);
 
 // Convert between half and single precision floating point,
-// assuming inputs and outputs are both finite, and
-// flushing values which would be denormal half floats to zero.
+// assuming inputs and outputs are both finite, and may
+// flush values which would be denormal half floats to zero.
 static inline Sk4f SkHalfToFloat_finite_ftz(uint64_t);
 static inline Sk4h SkFloatToHalf_finite_ftz(const Sk4f&);
 
 // ~~~~~~~~~~~ impl ~~~~~~~~~~~~~~ //
 
 // Like the serial versions in SkHalf.cpp, these are based on
 // https://fgiesen.wordpress.com/2012/03/28/half-to-float-done-quic/
 
 // GCC 4.9 lacks the intrinsics to use ARMv8 f16<->f32 instructions, so we use inline assembly.
 
@@ skipping 38 matching lines @@
     // 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;
 
     Sk4i merged = (sign >> 16) | (will_be_norm & norm);
     return SkNx_cast<uint16_t>(merged);
 #endif
 }
 
 #endif