From 425d68dcbdd681ad3157000360521e8f36eb6c4c Mon Sep 17 00:00:00 2001 From: "Steinar H. Gunderson" Date: Sun, 2 Jul 2017 10:18:34 +0200 Subject: [PATCH] Inline the fp16 conversion routines. Helps 18% on ResampleEffect::calculate_texture. --- fp16.cpp | 86 --------------------------------------------------- fp16.h | 75 ++++++++++++++++++++++++++++++++++++++++++-- fp16_test.cpp | 4 --- 3 files changed, 73 insertions(+), 92 deletions(-) delete mode 100644 fp16.cpp diff --git a/fp16.cpp b/fp16.cpp deleted file mode 100644 index cc1f30e..0000000 --- a/fp16.cpp +++ /dev/null @@ -1,86 +0,0 @@ -#include "fp16.h" - -namespace movit { -namespace { - -union fp32 { - float f; - unsigned int u; -}; - -} // namespace - -#ifndef __F16C__ - -float fp16_to_fp32(fp16_int_t h) -{ - fp32 magic; magic.u = 113 << 23; - unsigned int shifted_exp = 0x7c00 << 13; // exponent mask after shift - fp32 o; - - // mantissa+exponent - unsigned int shifted = (h.val & 0x7fff) << 13; - unsigned int exponent = shifted & shifted_exp; - - // exponent cases - o.u = shifted; - if (exponent == 0) { // Zero / Denormal - o.u += magic.u; - o.f -= magic.f; - } else if (exponent == shifted_exp) { // Inf/NaN - o.u += (255 - 31) << 23; - } else { - o.u += (127 - 15) << 23; - } - - o.u |= (h.val & 0x8000) << 16; // copy sign bit - return o.f; -} - -fp16_int_t fp32_to_fp16(float x) -{ - fp32 f; f.f = x; - unsigned int f32infty = 255 << 23; - unsigned int f16max = (127 + 16) << 23; - fp32 denorm_magic; denorm_magic.u = ((127 - 15) + (23 - 10) + 1) << 23; - unsigned int sign_mask = 0x80000000u; - fp16_int_t o = { 0 }; - - unsigned int sign = f.u & sign_mask; - f.u ^= sign; - - // NOTE all the integer compares in this function can be safely - // compiled into signed compares since all operands are below - // 0x80000000. Important if you want fast straight SSE2 code - // (since there's no unsigned PCMPGTD). - - if (f.u >= f16max) { // result is Inf or NaN (all exponent bits set) - o.val = (f.u > f32infty) ? 0x7e00 : 0x7c00; // NaN->qNaN and Inf->Inf - } else { // (De)normalized number or zero - if (f.u < (113 << 23)) { // resulting FP16 is subnormal or zero - // use a magic value to align our 10 mantissa bits at the bottom of - // the float. as long as FP addition is round-to-nearest-even this - // just works. - f.f += denorm_magic.f; - - // and one integer subtract of the bias later, we have our final float! - o.val = f.u - denorm_magic.u; - } else { - unsigned int mant_odd = (f.u >> 13) & 1; // resulting mantissa is odd - - // update exponent, rounding bias part 1 - f.u += ((15 - 127) << 23) + 0xfff; - // rounding bias part 2 - f.u += mant_odd; - // take the bits! - o.val = f.u >> 13; - } - } - - o.val |= sign >> 16; - return o; -} - -#endif - -} // namespace diff --git a/fp16.h b/fp16.h index 5f71fec..56a5607 100644 --- a/fp16.h +++ b/fp16.h @@ -44,8 +44,79 @@ static inline fp16_int_t fp32_to_fp16(float x) #else -float fp16_to_fp32(fp16_int_t x); -fp16_int_t fp32_to_fp16(float x); +union fp32 { + float f; + unsigned int u; +}; + +static inline float fp16_to_fp32(fp16_int_t h) +{ + fp32 magic; magic.u = 113 << 23; + unsigned int shifted_exp = 0x7c00 << 13; // exponent mask after shift + fp32 o; + + // mantissa+exponent + unsigned int shifted = (h.val & 0x7fff) << 13; + unsigned int exponent = shifted & shifted_exp; + + // exponent cases + o.u = shifted; + if (exponent == 0) { // Zero / Denormal + o.u += magic.u; + o.f -= magic.f; + } else if (exponent == shifted_exp) { // Inf/NaN + o.u += (255 - 31) << 23; + } else { + o.u += (127 - 15) << 23; + } + + o.u |= (h.val & 0x8000) << 16; // copy sign bit + return o.f; +} + +static inline fp16_int_t fp32_to_fp16(float x) +{ + fp32 f; f.f = x; + unsigned int f32infty = 255 << 23; + unsigned int f16max = (127 + 16) << 23; + fp32 denorm_magic; denorm_magic.u = ((127 - 15) + (23 - 10) + 1) << 23; + unsigned int sign_mask = 0x80000000u; + fp16_int_t o = { 0 }; + + unsigned int sign = f.u & sign_mask; + f.u ^= sign; + + // NOTE all the integer compares in this function can be safely + // compiled into signed compares since all operands are below + // 0x80000000. Important if you want fast straight SSE2 code + // (since there's no unsigned PCMPGTD). + + if (f.u >= f16max) { // result is Inf or NaN (all exponent bits set) + o.val = (f.u > f32infty) ? 0x7e00 : 0x7c00; // NaN->qNaN and Inf->Inf + } else { // (De)normalized number or zero + if (f.u < (113 << 23)) { // resulting FP16 is subnormal or zero + // use a magic value to align our 10 mantissa bits at the bottom of + // the float. as long as FP addition is round-to-nearest-even this + // just works. + f.f += denorm_magic.f; + + // and one integer subtract of the bias later, we have our final float! + o.val = f.u - denorm_magic.u; + } else { + unsigned int mant_odd = (f.u >> 13) & 1; // resulting mantissa is odd + + // update exponent, rounding bias part 1 + f.u += ((15 - 127) << 23) + 0xfff; + // rounding bias part 2 + f.u += mant_odd; + // take the bits! + o.val = f.u >> 13; + } + } + + o.val |= sign >> 16; + return o; +} #endif diff --git a/fp16_test.cpp b/fp16_test.cpp index d95e5c6..058c912 100644 --- a/fp16_test.cpp +++ b/fp16_test.cpp @@ -45,10 +45,6 @@ union fp64 { double f; unsigned long long ll; }; -union fp32 { - float f; - unsigned int u; -}; TEST(FP16Test, NaN) { // Ignore the sign bit. -- 2.39.2