///////////////////////////////////////////////////////////////////////////
//
// Copyright (c) 2012, Autodesk, Inc.
//
// All rights reserved.
//
// Implementation of IIF-specific file format and speed optimizations
// provided by Innobec Technologies inc on behalf of Autodesk.
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are
// met:
// * Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
// * Redistributions in binary form must reproduce the above
// copyright notice, this list of conditions and the following disclaimer
// in the documentation and/or other materials provided with the
// distribution.
// * Neither the name of Industrial Light & Magic nor the names of
// its contributors may be used to endorse or promote products derived
// from this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
//
///////////////////////////////////////////////////////////////////////////
#pragma once
#ifndef INCLUDED_IMF_OPTIMIZED_PIXEL_READING_H
#define INCLUDED_IMF_OPTIMIZED_PIXEL_READING_H
#include "ImfSimd.h"
#include "ImfSystemSpecific.h"
#include <iostream>
#include "ImfChannelList.h"
#include "ImfFrameBuffer.h"
#include "ImfStringVectorAttribute.h"
OPENEXR_IMF_INTERNAL_NAMESPACE_HEADER_ENTER
class OptimizationMode
{
public:
bool _optimizable;
int _ySampling;
OptimizationMode() : _optimizable(false) {}
};
#if IMF_HAVE_SSE2
//------------------------------------------------------------------------
// Test for SSE pointer alignemnt
//------------------------------------------------------------------------
EXR_FORCEINLINE
bool
isPointerSSEAligned (const void* EXR_RESTRICT pPointer)
{
unsigned long trailingBits = ((unsigned long)pPointer) & 15;
return trailingBits == 0;
}
//------------------------------------------------------------------------
// Load SSE from address into register
//------------------------------------------------------------------------
template<bool IS_ALIGNED>
EXR_FORCEINLINE
__m128i loadSSE (__m128i*& loadAddress)
{
// throw exception :: this is not accepted
return _mm_loadu_si128 (loadAddress);
}
template<>
EXR_FORCEINLINE
__m128i loadSSE<false> (__m128i*& loadAddress)
{
return _mm_loadu_si128 (loadAddress);
}
template<>
EXR_FORCEINLINE
__m128i loadSSE<true> (__m128i*& loadAddress)
{
return _mm_load_si128 (loadAddress);
}
//------------------------------------------------------------------------
// Store SSE from register into address
//------------------------------------------------------------------------
template<bool IS_ALIGNED>
EXR_FORCEINLINE
void storeSSE (__m128i*& storeAddress, __m128i& dataToStore)
{
}
template<>
EXR_FORCEINLINE
void
storeSSE<false> (__m128i*& storeAddress, __m128i& dataToStore)
{
_mm_storeu_si128 (storeAddress, dataToStore);
}
template<>
EXR_FORCEINLINE
void
storeSSE<true> (__m128i*& storeAddress, __m128i& dataToStore)
{
_mm_stream_si128 (storeAddress, dataToStore);
}
//------------------------------------------------------------------------
//
// Write to RGBA
//
//------------------------------------------------------------------------
//
// Using SSE intrinsics
//
template<bool READ_PTR_ALIGNED, bool WRITE_PTR_ALIGNED>
EXR_FORCEINLINE
void writeToRGBASSETemplate
(__m128i*& readPtrSSERed,
__m128i*& readPtrSSEGreen,
__m128i*& readPtrSSEBlue,
__m128i*& readPtrSSEAlpha,
__m128i*& writePtrSSE,
const size_t& lPixelsToCopySSE)
{
for (size_t i = 0; i < lPixelsToCopySSE; ++i)
{
__m128i redRegister = loadSSE<READ_PTR_ALIGNED> (readPtrSSERed);
__m128i greenRegister = loadSSE<READ_PTR_ALIGNED> (readPtrSSEGreen);
__m128i blueRegister = loadSSE<READ_PTR_ALIGNED> (readPtrSSEBlue);
__m128i alphaRegister = loadSSE<READ_PTR_ALIGNED> (readPtrSSEAlpha);
__m128i redGreenRegister = _mm_unpacklo_epi16 (redRegister,
greenRegister);
__m128i blueAlphaRegister = _mm_unpacklo_epi16 (blueRegister,
alphaRegister);
__m128i pixel12Register = _mm_unpacklo_epi32 (redGreenRegister,
blueAlphaRegister);
__m128i pixel34Register = _mm_unpackhi_epi32 (redGreenRegister,
blueAlphaRegister);
storeSSE<WRITE_PTR_ALIGNED> (writePtrSSE, pixel12Register);
++writePtrSSE;
storeSSE<WRITE_PTR_ALIGNED> (writePtrSSE, pixel34Register);
++writePtrSSE;
redGreenRegister = _mm_unpackhi_epi16 (redRegister, greenRegister);
blueAlphaRegister = _mm_unpackhi_epi16 (blueRegister, alphaRegister);
pixel12Register = _mm_unpacklo_epi32 (redGreenRegister,
blueAlphaRegister);
pixel34Register = _mm_unpackhi_epi32 (redGreenRegister,
blueAlphaRegister);
storeSSE<WRITE_PTR_ALIGNED> (writePtrSSE, pixel12Register);
++writePtrSSE;
storeSSE<WRITE_PTR_ALIGNED> (writePtrSSE, pixel34Register);
++writePtrSSE;
++readPtrSSEAlpha;
++readPtrSSEBlue;
++readPtrSSEGreen;
++readPtrSSERed;
}
}
//
// Not using SSE intrinsics. This is still faster than the alternative
// because we have multiple read pointers and therefore we are able to
// take advantage of data locality for write operations.
//
EXR_FORCEINLINE
void writeToRGBANormal (unsigned short*& readPtrRed,
unsigned short*& readPtrGreen,
unsigned short*& readPtrBlue,
unsigned short*& readPtrAlpha,
unsigned short*& writePtr,
const size_t& lPixelsToCopy)
{
for (size_t i = 0; i < lPixelsToCopy; ++i)
{
*(writePtr++) = *(readPtrRed++);
*(writePtr++) = *(readPtrGreen++);
*(writePtr++) = *(readPtrBlue++);
*(writePtr++) = *(readPtrAlpha++);
}
}
//
// Determine which (template) version to use by checking whether pointers
// are aligned
//
EXR_FORCEINLINE
void optimizedWriteToRGBA (unsigned short*& readPtrRed,
unsigned short*& readPtrGreen,
unsigned short*& readPtrBlue,
unsigned short*& readPtrAlpha,
unsigned short*& writePtr,
const size_t& pixelsToCopySSE,
const size_t& pixelsToCopyNormal)
{
bool readPtrAreAligned = true;
readPtrAreAligned &= isPointerSSEAligned(readPtrRed);
readPtrAreAligned &= isPointerSSEAligned(readPtrGreen);
readPtrAreAligned &= isPointerSSEAligned(readPtrBlue);
readPtrAreAligned &= isPointerSSEAligned(readPtrAlpha);
bool writePtrIsAligned = isPointerSSEAligned(writePtr);
if (!readPtrAreAligned && !writePtrIsAligned)
{
writeToRGBASSETemplate<false, false> ((__m128i*&)readPtrRed,
(__m128i*&)readPtrGreen,
(__m128i*&)readPtrBlue,
(__m128i*&)readPtrAlpha,
(__m128i*&)writePtr,
pixelsToCopySSE);
}
else if (!readPtrAreAligned && writePtrIsAligned)
{
writeToRGBASSETemplate<false, true> ((__m128i*&)readPtrRed,
(__m128i*&)readPtrGreen,
(__m128i*&)readPtrBlue,
(__m128i*&)readPtrAlpha,
(__m128i*&)writePtr,
pixelsToCopySSE);
}
else if (readPtrAreAligned && !writePtrIsAligned)
{
writeToRGBASSETemplate<true, false> ((__m128i*&)readPtrRed,
(__m128i*&)readPtrGreen,
(__m128i*&)readPtrBlue,
(__m128i*&)readPtrAlpha,
(__m128i*&)writePtr,
pixelsToCopySSE);
}
else if(readPtrAreAligned && writePtrIsAligned)
{
writeToRGBASSETemplate<true, true> ((__m128i*&)readPtrRed,
(__m128i*&)readPtrGreen,
(__m128i*&)readPtrBlue,
(__m128i*&)readPtrAlpha,
(__m128i*&)writePtr,
pixelsToCopySSE);
}
writeToRGBANormal (readPtrRed, readPtrGreen, readPtrBlue, readPtrAlpha,
writePtr, pixelsToCopyNormal);
}
//------------------------------------------------------------------------
//
// Write to RGBA Fill A
//
//------------------------------------------------------------------------
//
// Using SSE intrinsics
//
template<bool READ_PTR_ALIGNED, bool WRITE_PTR_ALIGNED>
EXR_FORCEINLINE
void
writeToRGBAFillASSETemplate (__m128i*& readPtrSSERed,
__m128i*& readPtrSSEGreen,
__m128i*& readPtrSSEBlue,
const unsigned short& alphaFillValue,
__m128i*& writePtrSSE,
const size_t& pixelsToCopySSE)
{
const __m128i dummyAlphaRegister = _mm_set_epi16 (alphaFillValue,
alphaFillValue,
alphaFillValue,
alphaFillValue,
alphaFillValue,
alphaFillValue,
alphaFillValue,
alphaFillValue);
for (size_t pixelCounter = 0; pixelCounter < pixelsToCopySSE; ++pixelCounter)
{
__m128i redRegister = loadSSE<READ_PTR_ALIGNED> (readPtrSSERed);
__m128i greenRegister = loadSSE<READ_PTR_ALIGNED> (readPtrSSEGreen);
__m128i blueRegister = loadSSE<READ_PTR_ALIGNED> (readPtrSSEBlue);
__m128i redGreenRegister = _mm_unpacklo_epi16 (redRegister,
greenRegister);
__m128i blueAlphaRegister = _mm_unpacklo_epi16 (blueRegister,
dummyAlphaRegister);
__m128i pixel12Register = _mm_unpacklo_epi32 (redGreenRegister,
blueAlphaRegister);
__m128i pixel34Register = _mm_unpackhi_epi32 (redGreenRegister,
blueAlphaRegister);
storeSSE<WRITE_PTR_ALIGNED> (writePtrSSE, pixel12Register);
++writePtrSSE;
storeSSE<WRITE_PTR_ALIGNED> (writePtrSSE, pixel34Register);
++writePtrSSE;
redGreenRegister = _mm_unpackhi_epi16 (redRegister,
greenRegister);
blueAlphaRegister = _mm_unpackhi_epi16 (blueRegister,
dummyAlphaRegister);
pixel12Register = _mm_unpacklo_epi32 (redGreenRegister,
blueAlphaRegister);
pixel34Register = _mm_unpackhi_epi32 (redGreenRegister,
blueAlphaRegister);
storeSSE<WRITE_PTR_ALIGNED> (writePtrSSE, pixel12Register);
++writePtrSSE;
storeSSE<WRITE_PTR_ALIGNED> (writePtrSSE, pixel34Register);
++writePtrSSE;
++readPtrSSEBlue;
++readPtrSSEGreen;
++readPtrSSERed;
}
}
//
// Not using SSE intrinsics. This is still faster than the alternative
// because we have multiple read pointers and therefore we are able to
// take advantage of data locality for write operations.
//
EXR_FORCEINLINE
void
writeToRGBAFillANormal (unsigned short*& readPtrRed,
unsigned short*& readPtrGreen,
unsigned short*& readPtrBlue,
const unsigned short& alphaFillValue,
unsigned short*& writePtr,
const size_t& pixelsToCopy)
{
for (size_t i = 0; i < pixelsToCopy; ++i)
{
*(writePtr++) = *(readPtrRed++);
*(writePtr++) = *(readPtrGreen++);
*(writePtr++) = *(readPtrBlue++);
*(writePtr++) = alphaFillValue;
}
}
//
// Determine which (template) version to use by checking whether pointers
// are aligned.
//
EXR_FORCEINLINE
void
optimizedWriteToRGBAFillA (unsigned short*& readPtrRed,
unsigned short*& readPtrGreen,
unsigned short*& readPtrBlue,
const unsigned short& alphaFillValue,
unsigned short*& writePtr,
const size_t& pixelsToCopySSE,
const size_t& pixelsToCopyNormal)
{
bool readPtrAreAligned = true;
readPtrAreAligned &= isPointerSSEAligned (readPtrRed);
readPtrAreAligned &= isPointerSSEAligned (readPtrGreen);
readPtrAreAligned &= isPointerSSEAligned (readPtrBlue);
bool writePtrIsAligned = isPointerSSEAligned (writePtr);
if (!readPtrAreAligned && !writePtrIsAligned)
{
writeToRGBAFillASSETemplate<false, false> ((__m128i*&)readPtrRed,
(__m128i*&)readPtrGreen,
(__m128i*&)readPtrBlue,
alphaFillValue,
(__m128i*&)writePtr,
pixelsToCopySSE);
}
else if (!readPtrAreAligned && writePtrIsAligned)
{
writeToRGBAFillASSETemplate<false, true> ((__m128i*&)readPtrRed,
(__m128i*&)readPtrGreen,
(__m128i*&)readPtrBlue,
alphaFillValue,
(__m128i*&)writePtr,
pixelsToCopySSE);
}
else if (readPtrAreAligned && !writePtrIsAligned)
{
writeToRGBAFillASSETemplate<true, false> ((__m128i*&)readPtrRed,
(__m128i*&)readPtrGreen,
(__m128i*&)readPtrBlue,
alphaFillValue,
(__m128i*&)writePtr,
pixelsToCopySSE);
}
else if (readPtrAreAligned && writePtrIsAligned)
{
writeToRGBAFillASSETemplate<true, true> ((__m128i*&)readPtrRed,
(__m128i*&)readPtrGreen,
(__m128i*&)readPtrBlue,
alphaFillValue,
(__m128i*&)writePtr,
pixelsToCopySSE);
}
writeToRGBAFillANormal (readPtrRed,
readPtrGreen, readPtrBlue, alphaFillValue,
writePtr, pixelsToCopyNormal);
}
//------------------------------------------------------------------------
//
// Write to RGB
//
//------------------------------------------------------------------------
//
// Using SSE intrinsics
//
template<bool READ_PTR_ALIGNED, bool WRITE_PTR_ALIGNED>
EXR_FORCEINLINE
void
writeToRGBSSETemplate (__m128i*& readPtrSSERed,
__m128i*& readPtrSSEGreen,
__m128i*& readPtrSSEBlue,
__m128i*& writePtrSSE,
const size_t& pixelsToCopySSE)
{
for (size_t pixelCounter = 0; pixelCounter < pixelsToCopySSE; ++pixelCounter)
{
//
// Need to shuffle and unpack pointers to obtain my first register
// We must save 8 pixels at a time, so we must have the following three registers at the end:
// 1) R1 G1 B1 R2 G2 B2 R3 G3
// 2) B3 R4 G4 B4 R5 G5 B5 R6
// 3) G6 B6 R7 G7 B7 R8 G8 B8
//
__m128i redRegister = loadSSE<READ_PTR_ALIGNED> (readPtrSSERed);
__m128i greenRegister = loadSSE<READ_PTR_ALIGNED> (readPtrSSEGreen);
__m128i blueRegister = loadSSE<READ_PTR_ALIGNED> (readPtrSSEBlue);
//
// First register: R1 G1 B1 R2 G2 B2 R3 G3
// Construct 2 registers and then unpack them to obtain our final result:
//
__m128i redGreenRegister = _mm_unpacklo_epi16 (redRegister,
greenRegister);
__m128i redBlueRegister = _mm_unpacklo_epi16 (redRegister,
blueRegister);
__m128i greenBlueRegister = _mm_unpacklo_epi16 (greenRegister,
blueRegister);
// Left Part (R1 G1 B1 R2)
__m128i quarterRight = _mm_shufflelo_epi16 (redBlueRegister,
_MM_SHUFFLE(3,0,2,1));
__m128i halfLeft = _mm_unpacklo_epi32 (redGreenRegister,
quarterRight);
// Right Part (G2 B2 R3 G3)
__m128i quarterLeft = _mm_shuffle_epi32 (greenBlueRegister,
_MM_SHUFFLE(3,2,0,1));
quarterRight = _mm_shuffle_epi32 (redGreenRegister,
_MM_SHUFFLE(3,0,1,2));
__m128i halfRight = _mm_unpacklo_epi32 (quarterLeft, quarterRight);
__m128i fullRegister = _mm_unpacklo_epi64 (halfLeft, halfRight);
storeSSE<WRITE_PTR_ALIGNED> (writePtrSSE, fullRegister);
++writePtrSSE;
//
// Second register: B3 R4 G4 B4 R5 G5 B5 R6
//
// Left Part (B3, R4, G4, B4)
quarterLeft = _mm_shufflehi_epi16 (redBlueRegister,
_MM_SHUFFLE(0, 3, 2, 1));
quarterRight = _mm_shufflehi_epi16 (greenBlueRegister,
_MM_SHUFFLE(1, 0, 3, 2));
halfLeft = _mm_unpackhi_epi32 (quarterLeft, quarterRight);
// Update the registers
redGreenRegister = _mm_unpackhi_epi16 (redRegister, greenRegister);
redBlueRegister = _mm_unpackhi_epi16 (redRegister, blueRegister);
greenBlueRegister = _mm_unpackhi_epi16 (greenRegister, blueRegister);
// Right Part (R5 G5 B5 R6)
quarterRight = _mm_shufflelo_epi16 (redBlueRegister,
_MM_SHUFFLE(3,0,2,1));
halfRight = _mm_unpacklo_epi32 (redGreenRegister, quarterRight);
fullRegister = _mm_unpacklo_epi64 (halfLeft, halfRight);
storeSSE<WRITE_PTR_ALIGNED> (writePtrSSE, fullRegister);
++writePtrSSE;
//
// Third register: G6 B6 R7 G7 B7 R8 G8 B8
//
// Left part (G6 B6 R7 G7)
quarterLeft = _mm_shuffle_epi32 (greenBlueRegister,
_MM_SHUFFLE(3,2,0,1));
quarterRight = _mm_shuffle_epi32 (redGreenRegister,
_MM_SHUFFLE(3,0,1,2));
halfLeft = _mm_unpacklo_epi32 (quarterLeft, quarterRight);
// Right part (B7 R8 G8 B8)
quarterLeft = _mm_shufflehi_epi16 (redBlueRegister,
_MM_SHUFFLE(0, 3, 2, 1));
quarterRight = _mm_shufflehi_epi16 (greenBlueRegister,
_MM_SHUFFLE(1, 0, 3, 2));
halfRight = _mm_unpackhi_epi32 (quarterLeft, quarterRight);
fullRegister = _mm_unpacklo_epi64 (halfLeft, halfRight);
storeSSE<WRITE_PTR_ALIGNED> (writePtrSSE, fullRegister);
++writePtrSSE;
//
// Increment read pointers
//
++readPtrSSEBlue;
++readPtrSSEGreen;
++readPtrSSERed;
}
}
//
// Not using SSE intrinsics. This is still faster than the alternative
// because we have multiple read pointers and therefore we are able to
// take advantage of data locality for write operations.
//
EXR_FORCEINLINE
void
writeToRGBNormal (unsigned short*& readPtrRed,
unsigned short*& readPtrGreen,
unsigned short*& readPtrBlue,
unsigned short*& writePtr,
const size_t& pixelsToCopy)
{
for (size_t i = 0; i < pixelsToCopy; ++i)
{
*(writePtr++) = *(readPtrRed++);
*(writePtr++) = *(readPtrGreen++);
*(writePtr++) = *(readPtrBlue++);
}
}
//
// Determine which (template) version to use by checking whether pointers
// are aligned
//
EXR_FORCEINLINE
void optimizedWriteToRGB (unsigned short*& readPtrRed,
unsigned short*& readPtrGreen,
unsigned short*& readPtrBlue,
unsigned short*& writePtr,
const size_t& pixelsToCopySSE,
const size_t& pixelsToCopyNormal)
{
bool readPtrAreAligned = true;
readPtrAreAligned &= isPointerSSEAligned(readPtrRed);
readPtrAreAligned &= isPointerSSEAligned(readPtrGreen);
readPtrAreAligned &= isPointerSSEAligned(readPtrBlue);
bool writePtrIsAligned = isPointerSSEAligned(writePtr);
if (!readPtrAreAligned && !writePtrIsAligned)
{
writeToRGBSSETemplate<false, false> ((__m128i*&)readPtrRed,
(__m128i*&)readPtrGreen,
(__m128i*&)readPtrBlue,
(__m128i*&)writePtr,
pixelsToCopySSE);
}
else if (!readPtrAreAligned && writePtrIsAligned)
{
writeToRGBSSETemplate<false, true> ((__m128i*&)readPtrRed,
(__m128i*&)readPtrGreen,
(__m128i*&)readPtrBlue,
(__m128i*&)writePtr,
pixelsToCopySSE);
}
else if (readPtrAreAligned && !writePtrIsAligned)
{
writeToRGBSSETemplate<true, false> ((__m128i*&)readPtrRed,
(__m128i*&)readPtrGreen,
(__m128i*&)readPtrBlue,
(__m128i*&)writePtr,
pixelsToCopySSE);
}
else if (readPtrAreAligned && writePtrIsAligned)
{
writeToRGBSSETemplate<true, true> ((__m128i*&)readPtrRed,
(__m128i*&)readPtrGreen,
(__m128i*&)readPtrBlue,
(__m128i*&)writePtr,
pixelsToCopySSE);
}
writeToRGBNormal (readPtrRed, readPtrGreen, readPtrBlue,
writePtr, pixelsToCopyNormal);
}
#else // ! defined IMF_HAVE_SSE2
#endif // defined IMF_HAVE_SSE2
OPENEXR_IMF_INTERNAL_NAMESPACE_HEADER_EXIT
#endif