liz/rawTherapee
liz/rawTherapee
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge pull request #4657 from Beep6581/dng_halffloat_load_speedup

Browse Source 
If available, use intrinsics to convert half precision float values to single precision float values
This commit is contained in:
Ingo Weyrich 2018-07-04 12:29:43 +02:00 committed by GitHub
commit 370e5d0222
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 26 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

29
rtengine/dcraw.cc
View File

@ -9937,6 +9937,7 @@ static void decodeFPDeltaRow(Bytef * src, Bytef * dst, size_t tileWidth, size_t
}
#ifndef __F16C__
// From DNG SDK dng_utils.h
static inline uint32_t DNG_HalfToFloat(uint16_t halfValue) {
int32_t sign = (halfValue >> 15) & 0x00000001;
@ -9970,6 +9971,7 @@ static inline uint32_t DNG_HalfToFloat(uint16_t halfValue) {
// Assemble sign, exponent and mantissa.
return (uint32_t) ((sign << 31) | (exponent << 23) | mantissa);
}
#endif
static inline uint32_t DNG_FP24ToFloat(const uint8_t * input) {
int32_t sign = (input [0] >> 7) & 0x01;
@ -10006,11 +10008,32 @@ static inline uint32_t DNG_FP24ToFloat(const uint8_t * input) {
static void expandFloats(Bytef * dst, int tileWidth, int bytesps) {
if (bytesps == 2) {
uint16_t * dst16 = (uint16_t *) dst;
uint32_t * dst32 = (uint32_t *) dst;
uint16_t* const dst16 = reinterpret_cast<uint16_t*>(dst);
#ifndef __F16C__
uint32_t* const dst32 = reinterpret_cast<uint32_t*>(dst);
for (int index = tileWidth - 1; index >= 0; --index) {
dst32[index] = DNG_HalfToFloat(dst16[index]);
dst32[index] = DNG_HalfToFloat(dst16[index]);
}
#else
float* const dst32 = reinterpret_cast<float*>(dst);
int index = tileWidth - 8;
for (; index >= 0; index -= 8) {
__m128i halfFloatv = _mm_loadu_si128((__m128i*)&dst16[index]);
STVFU(dst32[index], _mm_cvtph_ps(halfFloatv));
STVFU(dst32[index + 4], _mm_cvtph_ps(_mm_shuffle_epi32(halfFloatv, _MM_SHUFFLE(0,0,3,2))));
}
index += 4;
if(index >= 0) {
__m128i halfFloatv = _mm_loadu_si128((__m128i*)&dst16[index]);
STVFU(dst32[index], _mm_cvtph_ps(halfFloatv));
index--;
} else {
index += 3;
}
for (; index >= 0; --index) {
dst32[index] = _cvtsh_ss(dst16[index]);
}
#endif
} else if (bytesps == 3) {
uint8_t * dst8 = ((uint8_t *) dst) + (tileWidth - 1) * 3;
uint32_t * dst32 = (uint32_t *) dst;