/*****************************************************************************\ This program is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with this program; if not, write to the Free Software Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. \ \*****************************************************************************/ // copied from vob \di_research on 10/31/00 // MODIFICATIONS BY GE: // 0. remove Windows header references // 1. define assert // 2. set iRastersReady, iRastersDelivered in submitrowtofilter // 3. (constructor) allocate (and delete in destructor) buffers for m_row_ptrs // (instead of setting it to input buffers, since we reuse input buffers) // 4. copy data into m_row_ptrs in submitrowtofilter //#define assert ASSERT #include "ErnieFilter.h" #if defined(__APPLE__) || defined(__linux) || defined(__GLIBC__) || defined(__NetBSD__) #include #endif #if kGatherStats == 1 extern int blockStats[]; #endif #if ((kMemWritesOptimize != 1) && (kMemWritesOptimize != 0)) #error "kMemWritesOptimize must be 0 or 1" #endif inline void AverageNRound(bool roundGreenDown, int &rFinal, int &r0, int &r1, int &gFinal, int &g0, int &g1, int &bFinal, int &b0, int &b1); inline void AverageNRound(bool roundGreenDown, int &rFinal, int &r0, int &r1, int &gFinal, int &g0, int &g1, int &bFinal, int &b0, int &b1) { // By rounding G in the other direction than R and B L* variations are minimized while mathematically alternate rounding is accomplished. EGW 2 Dec. 1999. if (roundGreenDown) { rFinal = (r0 + r1 + 1) / 2; gFinal = (g0 + g1) / 2; bFinal = (b0 + b1 + 1) / 2; } else { rFinal = (r0 + r1) / 2; gFinal = (g0 + g1 + 1) / 2; bFinal = (b0 + b1) / 2; } } // Filter1RawRow. To be used to filter an odd row for which we don't have a pair, // found at the bottom of bands that aren't divisable by 2. This routine // filters its row horizontally forming 4x1 and 2x1 blocks. void ErnieFilter::Filter1RawRow(unsigned char *currPtr, int rowWidthInPixels, unsigned int *flagsPtr) { ASSERT(currPtr); ASSERT(rowWidthInPixels > 0); int R0, G0, B0, R1, G1, B1, lastR, lastG, lastB; const unsigned int maxErrorForFourPixels = m_max_error_for_two_pixels / 2; // const unsigned int maxErrorForEightPixels = maxErrorForFourPixels / 2; // int currPixel, lastPixel; uint32_t currPixel, lastPixel; bool lastPairAveraged = false; bool last2by2Averaged = false; for (int pixelNum = 0; pixelNum < rowWidthInPixels; pixelNum++) { if ((pixelNum & 0x03) == 0x00) // 0,4,8... { last2by2Averaged = false; // Reinitialize every four columns; } currPixel = get4Pixel(currPtr); flagsPtr[0] = (e11n|e11s); // Initialize in case nothing is found for this column if (isWhite(currPixel)) { flagsPtr[0] = eDone; #if kGatherStats == 1 blockStats[esWhiteFound]++; #endif } // Currently we bail entirely if there is white. Later we may still do RLE on the non white pixel if one is present. if (flagsPtr[0] == (e11n|e11s)) { R1 = GetRed(currPixel); G1 = GetGreen(currPixel); B1 = GetBlue(currPixel); // Can only horizontally average every other pixel, much like the 2x2 blocks. if (isOdd(pixelNum)) { // do horizontal on current raster lastPixel = get4Pixel(currPtr,-1); lastR = GetRed(lastPixel); lastG = GetGreen(lastPixel); lastB = GetBlue(lastPixel); if ((m_max_error_for_two_pixels >= 3) && (NewDeltaE(lastR, R1, lastG, G1, lastB, B1, m_max_error_for_two_pixels))) { /* - - | | build 2x1 - - */ int didNotBuild4by1 = true; #if kGatherStats == 1 blockStats[es21nw]++; #endif AverageNRound(isOdd(pixelNum), lastR, lastR, R1, lastG, lastG, G1, lastB, lastB, B1); if ((pixelNum >= 3) && (flagsPtr[-3] & e21nw)) // 4,5,6,7,12,13,14,15,20... { // Look for a 4x1 ASSERT(!((flagsPtr[-3] | flagsPtr[-2] | flagsPtr[-1] | flagsPtr[0]) & eTheRest)); // no vertical blocks lastPixel = get4Pixel(currPtr,-3); R0 = GetRed(lastPixel); G0 = GetGreen(lastPixel); B0 = GetBlue(lastPixel); if ((maxErrorForFourPixels >= 3) && (NewDeltaE(lastR, R0, lastG, G0, lastB, B0, maxErrorForFourPixels))) { /* - - - - | | build 4x1 - - - - */ #if kGatherStats == 1 blockStats[es41ni]++; #endif didNotBuild4by1 = false; AverageNRound((pixelNum & 0x04)== 0x04, lastR, lastR, R0, lastG, lastG, G0, lastB, lastB, B0); // 4,5,6,7,12,13,14,15,20... if(m_eEndian == LITTLEENDIAN) currPixel = (lastR<<16) + (lastG<<8) + lastB; else if(m_eEndian == BIGENDIAN) currPixel = (lastR<<24) + (lastG<<16) + (lastB<<8); #if kMemWritesOptimize == 0 put4Pixel(currPtr, -3, currPixel); put4Pixel(currPtr, -2, currPixel); put4Pixel(currPtr, -1, currPixel); put4Pixel(currPtr, 0, currPixel); #else put4Pixel(currPtr, -3, currPixel); #endif flagsPtr[-3] = (flagsPtr[-3] & ~eNorths) | e41ni; flagsPtr[-2] = (flagsPtr[-2] & ~eNorths) | e41n; flagsPtr[-1] = (flagsPtr[-1] & ~eNorths) | e41n; flagsPtr[0] = (flagsPtr[0] & ~eNorths) | e41n; } } if (didNotBuild4by1) // Not a 4x1 so output 2x1. { ASSERT(!((flagsPtr[-1] | flagsPtr[0]) & eTheRest)); // no vertical blocks if(m_eEndian == LITTLEENDIAN) currPixel = (lastR<<16) + (lastG<<8) + lastB; else if(m_eEndian == BIGENDIAN) currPixel = (lastR<<24) + (lastG<<16) + (lastB<<8); #if kMemWritesOptimize == 0 put4Pixel(currPtr, -1, currPixel); put4Pixel(currPtr, 0, currPixel); #else put4Pixel(currPtr, -1, currPixel); #endif flagsPtr[-1] = (flagsPtr[-1] & ~eNorths) | e21nw; flagsPtr[0] = (flagsPtr[0] & ~eNorths) | e21ne; } } // If DeltaE... Looking for two by one } // IsOdd(pixelNum) } else // no flag bits set. { lastPairAveraged = false; // EGW Fixes bug on business graphics. 11/24/97 } currPtr += eBufferedPixelWidthInBytes; flagsPtr++; } // for each pixel... } // Filter2RawRows: Looks filter two raw rows together to form blocks. Vertical // blocks are prefered over horizontal ones. The routine will create 1x2 blocks // before it will create 4x1's. In total this routine will create 1x2, 2x2, 4x2, // 4x1, and 2x1 blocks sizes, with the potential for two seperate 4x1's or 2x1's // in the upper and lower rasters. void ErnieFilter::Filter2RawRows(unsigned char *currPtr, unsigned char *upPtr, int rowWidthInPixels, unsigned int *flagsPtr) { ASSERT(currPtr); ASSERT(upPtr); ASSERT(rowWidthInPixels > 0); int R0, G0, B0, R1, G1, B1, lastR, lastG, lastB; const unsigned int maxErrorForFourPixels = m_max_error_for_two_pixels / 2; const unsigned int maxErrorForEightPixels = maxErrorForFourPixels / 2; // int currPixel, upPixel, lastPixel; uint32_t currPixel, upPixel, lastPixel; bool lastPairAveraged = false; bool last2by2Averaged = false; for (int pixelNum = 0; pixelNum < rowWidthInPixels; pixelNum++) { if ((pixelNum & 0x03) == 0x00) // 0,4,8... { last2by2Averaged = false; // Reinitialize every four columns; } upPixel = get4Pixel(upPtr); currPixel = get4Pixel(currPtr); flagsPtr[0] = (e11n|e11s); // Initialize in case nothing is found for this column if (isWhite(upPixel) && isWhite(currPixel)) // both white? { flagsPtr[0] = eDone; #if kGatherStats == 1 blockStats[esWhiteFound]++; #endif } // Do vertical average on the current 2 pixel high column // Currently we bail entirely if there is white. Later we may still do RLE on the non white pixel if one is present. if (flagsPtr[0] == (e11n|e11s)) { R1 = GetRed(currPixel); G1 = GetGreen(currPixel); B1 = GetBlue(currPixel); R0 = GetRed(upPixel); G0 = GetGreen(upPixel); B0 = GetBlue(upPixel); if ((m_max_error_for_two_pixels >= 3) && (NewDeltaE(R0, R1, G0, G1, B0, B1, m_max_error_for_two_pixels))) { /* _ | | build 1x2 | | - */ ASSERT(flagsPtr[0] == (e11n|e11s)); flagsPtr[0] = e12; #if kGatherStats == 1 blockStats[es12]++; #endif R1 = GetRed(currPixel); G1 = GetGreen(currPixel); B1 = GetBlue(currPixel); R0 = GetRed(upPixel); G0 = GetGreen(upPixel); B0 = GetBlue(upPixel); AverageNRound(isOdd(pixelNum), R1, R1, R0, G1, G1, G0, B1, B1, B0); // look for a 2x2 block average on every other column if (isOdd(pixelNum)) { // It looks like we are at the end of a 2x2 block if (lastPairAveraged) { // Last pair was averaged so it's ok to try to make a 2x2 block if ((maxErrorForFourPixels >= 3) && (NewDeltaE(lastR, R1, lastG, G1,lastB, B1, maxErrorForFourPixels))) { /* - - | | build 2x2 | | - - */ ASSERT(flagsPtr[-1] == e12); int didNotBuild4by2 = true; #if kGatherStats == 1 blockStats[es22w]++; #endif flagsPtr[-1] = e22w; flagsPtr[0] = e22e; AverageNRound((pixelNum & 0x02) == 0x02, R1, R1, lastR, G1, G1, lastG, B1, B1, lastB); // 2,3,6,7... Alternate between rounding up and down for these 2x2 blocks if ((pixelNum & 0x03) == 0x03) // 3,7,11,15... Looking for a 4x2 block to average { if (last2by2Averaged) { /* - - - - | | | | We have two 2x2s. | | | | - - - - */ lastPixel = get4Pixel(upPtr, -3); // Go back to previous 2x2 block and get the pixel lastR = GetRed(lastPixel); lastG = GetGreen(lastPixel); lastB = GetBlue(lastPixel); if ((maxErrorForEightPixels >= 3) && (NewDeltaE(lastR, R1, lastG, G1,lastB, B1, maxErrorForEightPixels))) { /* - - - - | | build 4x2. | | - - - - */ #if kGatherStats == 1 blockStats[es42i]++; #endif didNotBuild4by2 = false; flagsPtr[-3] = e42i; flagsPtr[-2] = flagsPtr[-1] = flagsPtr[0] = e42; AverageNRound((pixelNum & 0x04) == 0x04, R1, R1, lastR, G1, G1, lastG, B1, B1, lastB); // 4,5,6,7,12,13,14,15,20... Alternate between rounding up down for these 4x2 blocks if(m_eEndian == LITTLEENDIAN) currPixel = (R1<<16) + (G1<<8) + B1; else if(m_eEndian == BIGENDIAN) currPixel = (R1<<24) + (G1<<16) + (B1<<8); #if kMemWritesOptimize == 0 put4Pixel(upPtr, -3, currPixel); put4Pixel(upPtr, -2, currPixel); put4Pixel(upPtr, -1, currPixel); put4Pixel(upPtr, 0, currPixel); put4Pixel(currPtr, -3, currPixel); put4Pixel(currPtr, -2, currPixel); put4Pixel(currPtr, -1, currPixel); put4Pixel(currPtr, 0, currPixel); #else put4Pixel(upPtr, -3, currPixel); #endif } } if (didNotBuild4by2) { // The first 2x2 block of this pair of 2x2 blocks wasn't averaged. /* - - - - |X X| | | not averaged block and averaged 2x2. |X X| | | - - - - */ last2by2Averaged = true; if(m_eEndian == LITTLEENDIAN) currPixel = (R1<<16) + (G1<<8) + B1; else if(m_eEndian == BIGENDIAN) currPixel = (R1<<24) + (G1<<16) + (B1<<8); #if kMemWritesOptimize == 0 put4Pixel(upPtr, -1, currPixel); put4Pixel(upPtr, 0, currPixel); put4Pixel(currPtr, -1, currPixel); put4Pixel(currPtr, 0, currPixel); #else put4Pixel(upPtr, -1, currPixel); #endif } } else // Not looking for a 4x2 block yet so just output this 2x2 block for now. { /* - - - - | | |? ?| 1st 2x2 and maybe another later. | | |? ?| - - - - */ last2by2Averaged = true; if(m_eEndian == LITTLEENDIAN) currPixel = (R1<<16) + (G1<<8) + B1; else if(m_eEndian == BIGENDIAN) currPixel = (R1<<24) + (G1<<16) + (B1<<8); #if kMemWritesOptimize == 0 put4Pixel(upPtr, -1, currPixel); put4Pixel(upPtr, 0, currPixel); put4Pixel(currPtr, -1, currPixel); put4Pixel(currPtr, 0, currPixel); #else put4Pixel(upPtr, -1, currPixel); #endif } } else // The two averaged columns are not close enough in Delta E { /* - _ | | | | 2 1x2 blocks | | | | - - */ last2by2Averaged = false; if(m_eEndian == LITTLEENDIAN) currPixel = (R1<<16) + (G1<<8) + B1; else if(m_eEndian == BIGENDIAN) currPixel = (R1<<24) + (G1<<16) + (B1<<8); #if kMemWritesOptimize == 0 put4Pixel(upPtr, 0, currPixel); put4Pixel(currPtr, 0, currPixel); #else put4Pixel(upPtr,0, currPixel); #endif } lastR = R1; lastG = G1; lastB = B1; lastPairAveraged = true; } else // This is the right place for 2x2 averaging but the previous column wasn't averaged { /* - X | | Two non averaged pixels and a 1x2. X | | - */ last2by2Averaged = false; lastPairAveraged = true; lastR = R1; lastG = G1; lastB = B1; if(m_eEndian == LITTLEENDIAN) currPixel = (R1<<16) + (G1<<8) + B1; else if(m_eEndian == BIGENDIAN) currPixel = (R1<<24) + (G1<<16) + (B1<<8); #if kMemWritesOptimize == 0 put4Pixel(upPtr, 0, currPixel); put4Pixel(currPtr, 0, currPixel); #else put4Pixel(upPtr, 0, currPixel); #endif } } else // Not on the boundary for a 2x2 block, so just output current averaged 1x2 column { /* - | | ? 1x2 | | ? - */ lastPairAveraged = true; lastR = R1; lastG = G1; lastB = B1; if(m_eEndian == LITTLEENDIAN) currPixel = (R1<<16) + (G1<<8) + B1; else if(m_eEndian == BIGENDIAN) currPixel = (R1<<24) + (G1<<16) + (B1<<8); #if kMemWritesOptimize == 0 put4Pixel(upPtr, 0, currPixel); put4Pixel(currPtr, 0, currPixel); #else put4Pixel(upPtr, 0, currPixel); #endif } } else if (lastPairAveraged) { // This is the case where we can't average current column and the last column was averaged. // Don't do anything if last pair was averaged and this one can't be /* - | | X 1x2 averaged block and two non averaged pixels. | | X - */ lastPairAveraged = false; } else // can't vertically average current column so look for some horizontal averaging as a fallback // Only do it if the last pair wasn't averaged either because we don't want to mess up a vertical averaging // just to create a possible horizontal averaging. { // Can only horizontally average every other pixel, much like the 2x2 blocks. if (isOdd(pixelNum)) { // do horizontal averaging on previous raster lastPixel = get4Pixel(upPtr,-1); lastR = GetRed(lastPixel); lastG = GetGreen(lastPixel); lastB = GetBlue(lastPixel); if (((m_max_error_for_two_pixels >= 3)) && (NewDeltaE(lastR, R0, lastG, G0,lastB, B0, m_max_error_for_two_pixels))) { /* - - | | build upper 2x1 - - */ #if kGatherStats == 1 blockStats[es21nw]++; #endif int didNotBuild4by1 = true; AverageNRound(isOdd(pixelNum), lastR, lastR, R0, lastG, lastG, G0, lastB, lastB, B0); if ((pixelNum >= 3) && (flagsPtr[-3] & e21nw)) // 4,5,6,7,12,13,14,15,20... { ASSERT(!((flagsPtr[-3] | flagsPtr[-2] | flagsPtr[-1] | flagsPtr[0]) & eTheRest)); // no vertical blocks // Attempt an upper 4x1 lastPixel = get4Pixel(upPtr,-3); R0 = GetRed(lastPixel); G0 = GetGreen(lastPixel); B0 = GetBlue(lastPixel); if ( (maxErrorForFourPixels >= 3) && (NewDeltaE(lastR, R0, lastG, G0,lastB, B0, maxErrorForFourPixels))) { /* - - - - | | build upper 4x1 - - - - */ #if kGatherStats == 1 blockStats[es41ni]++; #endif didNotBuild4by1 = false; AverageNRound((pixelNum & 0x04)== 0x04, lastR, lastR, R0, lastG, lastG, G0, lastB, lastB, B0); // 4,5,6,7,12,13,14,15,20... if(m_eEndian == LITTLEENDIAN) currPixel = (lastR<<16) + (lastG<<8) + lastB; else if(m_eEndian == BIGENDIAN) currPixel = (lastR<<24) + (lastG<<16) + (lastB<<8); #if kMemWritesOptimize == 0 put4Pixel(upPtr, -3, currPixel); put4Pixel(upPtr, -2, currPixel); put4Pixel(upPtr, -1, currPixel); put4Pixel(upPtr, 0, currPixel); #else put4Pixel(upPtr, -3, currPixel); #endif ASSERT(!((flagsPtr[-3] | flagsPtr[-2] | flagsPtr[-1] | flagsPtr[0]) & eTheRest)); // no vertical blocks flagsPtr[-3] = (flagsPtr[-3] & ~eNorths) | e41ni; flagsPtr[-2] = (flagsPtr[-2] & ~eNorths) | e41n; flagsPtr[-1] = (flagsPtr[-1] & ~eNorths) | e41n; flagsPtr[0] = (flagsPtr[0] & ~eNorths) | e41n; } } if (didNotBuild4by1) // Not an upper 4x1 so output upper 2x1. { if(m_eEndian == LITTLEENDIAN) currPixel = (lastR<<16) + (lastG<<8) + lastB; else if(m_eEndian == BIGENDIAN) currPixel = (lastR<<24) + (lastG<<16) + (lastB<<8); #if kMemWritesOptimize == 0 put4Pixel(upPtr, -1, currPixel); put4Pixel(upPtr, 0, currPixel); #else put4Pixel(upPtr, -1, currPixel); #endif ASSERT(!((flagsPtr[-1] | flagsPtr[0]) & eTheRest)); // no vertical blocks flagsPtr[-1] = (flagsPtr[-1] & ~eNorths) | e21nw; flagsPtr[0] = (flagsPtr[0] & ~eNorths) | e21ne; } } // do horizontal on current raster lastPixel = get4Pixel(currPtr,-1); lastR = GetRed(lastPixel); lastG = GetGreen(lastPixel); lastB = GetBlue(lastPixel); if ((m_max_error_for_two_pixels >= 3) && (NewDeltaE(lastR, R1, lastG, G1, lastB, B1, m_max_error_for_two_pixels))) { /* - - | | build lower 2x1 - - */ int didNotBuild4by1 = true; #if kGatherStats == 1 blockStats[es21sw]++; #endif AverageNRound(isOdd(pixelNum), lastR, lastR, R1, lastG, lastG, G1, lastB, lastB, B1); if ((pixelNum >= 3) && (flagsPtr[-3] & e21sw)) // 4,5,6,7,12,13,14,15,20... { // Look for a lower 4x1 ASSERT(!((flagsPtr[-3] | flagsPtr[-2] | flagsPtr[-1] | flagsPtr[0]) & eTheRest)); // no vertical blocks lastPixel = get4Pixel(currPtr,-3); R0 = GetRed(lastPixel); G0 = GetGreen(lastPixel); B0 = GetBlue(lastPixel); if ((maxErrorForFourPixels >= 3) && (NewDeltaE(lastR, R0, lastG, G0, lastB, B0, maxErrorForFourPixels))) { /* - - - - | | build lower 4x1 - - - - */ #if kGatherStats == 1 blockStats[es41si]++; #endif didNotBuild4by1 = false; AverageNRound((pixelNum & 0x04)== 0x04, lastR, lastR, R0, lastG, lastG, G0, lastB, lastB, B0); // 4,5,6,7,12,13,14,15,20... if(m_eEndian == LITTLEENDIAN) currPixel = (lastR<<16) + (lastG<<8) + lastB; else if(m_eEndian == BIGENDIAN) currPixel = (lastR<<24) + (lastG<<16) + (lastB<<8); #if kMemWritesOptimize == 0 put4Pixel(currPtr, -3, currPixel); put4Pixel(currPtr, -2, currPixel); put4Pixel(currPtr, -1, currPixel); put4Pixel(currPtr, 0, currPixel); #else put4Pixel(currPtr, -3, currPixel); #endif flagsPtr[-3] = (flagsPtr[-3] & ~eSouths) | e41si; flagsPtr[-2] = (flagsPtr[-2] & ~eSouths) | e41s; flagsPtr[-1] = (flagsPtr[-1] & ~eSouths) | e41s; flagsPtr[0] = (flagsPtr[0] & ~eSouths) | e41s; } } if (didNotBuild4by1) // Not a lower 4x1 so output lower 2x1. { ASSERT(!((flagsPtr[-1] | flagsPtr[0]) & eTheRest)); // no vertical blocks if(m_eEndian == LITTLEENDIAN) currPixel = (lastR<<16) + (lastG<<8) + lastB; else if(m_eEndian == BIGENDIAN) currPixel = (lastR<<24) + (lastG<<16) + (lastB<<8); #if kMemWritesOptimize == 0 put4Pixel(currPtr, -1, currPixel); put4Pixel(currPtr, 0, currPixel); #else put4Pixel(currPtr, -1, currPixel); #endif flagsPtr[-1] = (flagsPtr[-1] & ~eSouths) | e21sw; flagsPtr[0] = (flagsPtr[0] & ~eSouths) | e21se; } } // If DeltaE... Looking for two by one } // IsOdd(pixelNum) } } else // no flag bits set. { lastPairAveraged = false; // EGW Fixes bug on business graphics. 11/24/97 } upPtr += eBufferedPixelWidthInBytes; currPtr += eBufferedPixelWidthInBytes; flagsPtr++; } // for each pixel... } // Filter2PairsOfFilteredRows. This routine takes 2 pairs of rows that // have been through the Filter2RawRows routine and puts blocks together // to make bigger blocks. It prefers taking 2 high blocks and putting // them together to make four high blocks, but as a last resort it will // take try to take a 1 high blocks from the second and third rasters and // create 2 high blocks. The possible block sizes this routine could // create are 8x4, 4x4, 2x4, and 1x4, and then with the second and third rasters // 4x2, 2x2, and 1x2. void ErnieFilter::Filter2PairsOfFilteredRows(unsigned char *row1Ptr, unsigned char *row2Ptr, unsigned char *row3Ptr, unsigned char *row4Ptr) { const unsigned int maxErrorForFourPixels = m_max_error_for_two_pixels / 2; const unsigned int maxErrorForEightPixels = maxErrorForFourPixels / 2; const unsigned int maxErrorForSixteenPixels = maxErrorForEightPixels / 2; const unsigned int maxErrorForThirtyTwoPixels = maxErrorForSixteenPixels / 2; for (int pixelNum = 0; pixelNum < (m_row_width_in_pixels-3);) // Make sure we have four pixels to work with { int currPixel, upPixel; int R0, G0, B0, R1, G1, B1; if ((m_pixel_filtered_flags[0][pixelNum] & e42i) && (m_pixel_filtered_flags[1][pixelNum] & e42i)) { /* - - - - | | | | - - - - We have two 4x2s. - - - - | | | | - - - - */ ASSERT(m_pixel_filtered_flags[0][pixelNum] == e42i && m_pixel_filtered_flags[0][pixelNum+1] == e42 && m_pixel_filtered_flags[0][pixelNum+2] == e42 && m_pixel_filtered_flags[0][pixelNum+3] == e42); ASSERT(m_pixel_filtered_flags[1][pixelNum] == e42i && m_pixel_filtered_flags[1][pixelNum+1] == e42 && m_pixel_filtered_flags[1][pixelNum+2] == e42 && m_pixel_filtered_flags[1][pixelNum+3] == e42); upPixel = get4Pixel(row1Ptr); currPixel = get4Pixel(row3Ptr); R1 = GetRed(currPixel); G1 = GetGreen(currPixel); B1 = GetBlue(currPixel); R0 = GetRed(upPixel); G0 = GetGreen(upPixel); B0 = GetBlue(upPixel); if((maxErrorForSixteenPixels >= 3) &&(NewDeltaE(R1, R0, G1, G0, B1, B0, maxErrorForSixteenPixels))) { /* - - - - | | | | build 4x4 | | | | - - - - */ #if kGatherStats == 1 blockStats[es44ni]++; #endif AverageNRound((pixelNum & 0x04) == 0x04, R1, R1, R0, G1, G1, G0, B1, B1, B0); // 4,5,6,7,12,13,14,15,20... Alternate between rounding up down if(m_eEndian == LITTLEENDIAN) currPixel = (R1<<16) + (G1<<8) + B1; else if(m_eEndian == BIGENDIAN) currPixel = (R1<<24) + (G1<<16) + (B1<<8); #if kMemWritesOptimize == 0 put4Pixel(row1Ptr, 0, currPixel); put4Pixel(row1Ptr, 1, currPixel); put4Pixel(row1Ptr, 2, currPixel); put4Pixel(row1Ptr, 3, currPixel); put4Pixel(row2Ptr, 0, currPixel); put4Pixel(row2Ptr, 1, currPixel); put4Pixel(row2Ptr, 2, currPixel); put4Pixel(row2Ptr, 3, currPixel); put4Pixel(row3Ptr, 0, currPixel); put4Pixel(row3Ptr, 1, currPixel); put4Pixel(row3Ptr, 2, currPixel); put4Pixel(row3Ptr, 3, currPixel); put4Pixel(row4Ptr, 0, currPixel); put4Pixel(row4Ptr, 1, currPixel); put4Pixel(row4Ptr, 2, currPixel); put4Pixel(row4Ptr, 3, currPixel); #else put4Pixel(row1Ptr, 0, currPixel); #endif row1Ptr += 4*eBufferedPixelWidthInBytes; row2Ptr += 4*eBufferedPixelWidthInBytes; row3Ptr += 4*eBufferedPixelWidthInBytes; row4Ptr += 4*eBufferedPixelWidthInBytes; m_pixel_filtered_flags[0][pixelNum] = e44ni; m_pixel_filtered_flags[0][pixelNum+1] = m_pixel_filtered_flags[0][pixelNum+2] = m_pixel_filtered_flags[0][pixelNum+3] = e44n; m_pixel_filtered_flags[1][pixelNum] = e44si; m_pixel_filtered_flags[1][pixelNum+1] = m_pixel_filtered_flags[1][pixelNum+2] = m_pixel_filtered_flags[1][pixelNum+3] = e44s; if ((pixelNum >= 4) && (m_pixel_filtered_flags[1][pixelNum-4] & e44si)) // 4,5,6,7,12,13,14,15,20... { /* - - - - - - - - | | | | | | | | We have two 4x4s. | | | | | | | | - - - - - - - - */ ASSERT(m_pixel_filtered_flags[0][pixelNum-4] == e44ni && m_pixel_filtered_flags[0][pixelNum-3] == e44n && m_pixel_filtered_flags[0][pixelNum-2] == e44n && m_pixel_filtered_flags[0][pixelNum-1] == e44n); ASSERT(m_pixel_filtered_flags[1][pixelNum-4] == e44si && m_pixel_filtered_flags[1][pixelNum-3] == e44s && m_pixel_filtered_flags[1][pixelNum-2] == e44s && m_pixel_filtered_flags[1][pixelNum-1] == e44s); upPixel = get4Pixel(row1Ptr, -8); R0 = GetRed(upPixel); G0 = GetGreen(upPixel); B0 = GetBlue(upPixel); if( (maxErrorForThirtyTwoPixels >= 3) && (NewDeltaE(R1, R0, G1, G0, B1, B0, maxErrorForThirtyTwoPixels))) { /* - - - - - - - - | | | | build 8x4 | | | | - - - - - - - - */ #if kGatherStats == 1 blockStats[es84ni]++; #endif AverageNRound((pixelNum & 0x08) == 0x08, R1, R1, R0, G1, G1, G0, B1, B1, B0); if(m_eEndian == LITTLEENDIAN) currPixel = (R1<<16) + (G1<<8) + B1; else if(m_eEndian == BIGENDIAN) currPixel = (R1<<24) + (G1<<16) + (B1<<8); #if kMemWritesOptimize == 0 put4Pixel(row1Ptr, -8, currPixel); put4Pixel(row1Ptr, -7, currPixel); put4Pixel(row1Ptr, -6, currPixel); put4Pixel(row1Ptr, -5, currPixel); put4Pixel(row1Ptr, -4, currPixel); put4Pixel(row1Ptr, -3, currPixel); put4Pixel(row1Ptr, -2, currPixel); put4Pixel(row1Ptr, -1, currPixel); put4Pixel(row2Ptr, -8, currPixel); put4Pixel(row2Ptr, -7, currPixel); put4Pixel(row2Ptr, -6, currPixel); put4Pixel(row2Ptr, -5, currPixel); put4Pixel(row2Ptr, -4, currPixel); put4Pixel(row2Ptr, -3, currPixel); put4Pixel(row2Ptr, -2, currPixel); put4Pixel(row2Ptr, -1, currPixel); put4Pixel(row3Ptr, -8, currPixel); put4Pixel(row3Ptr, -7, currPixel); put4Pixel(row3Ptr, -6, currPixel); put4Pixel(row3Ptr, -5, currPixel); put4Pixel(row3Ptr, -4, currPixel); put4Pixel(row3Ptr, -3, currPixel); put4Pixel(row3Ptr, -2, currPixel); put4Pixel(row3Ptr, -1, currPixel); put4Pixel(row4Ptr, -8, currPixel); put4Pixel(row4Ptr, -7, currPixel); put4Pixel(row4Ptr, -6, currPixel); put4Pixel(row4Ptr, -5, currPixel); put4Pixel(row4Ptr, -4, currPixel); put4Pixel(row4Ptr, -3, currPixel); put4Pixel(row4Ptr, -2, currPixel); put4Pixel(row4Ptr, -1, currPixel); #else put4Pixel(row1Ptr, -8, currPixel); #endif m_pixel_filtered_flags[0][pixelNum-4] = e84ni; m_pixel_filtered_flags[0][pixelNum-3] = m_pixel_filtered_flags[0][pixelNum-2] = m_pixel_filtered_flags[0][pixelNum-1] = m_pixel_filtered_flags[0][pixelNum] = m_pixel_filtered_flags[0][pixelNum+1] = m_pixel_filtered_flags[0][pixelNum+2] = m_pixel_filtered_flags[0][pixelNum+3] = e84n; m_pixel_filtered_flags[1][pixelNum-4] = e84si; m_pixel_filtered_flags[1][pixelNum-3] = m_pixel_filtered_flags[1][pixelNum-2] = m_pixel_filtered_flags[1][pixelNum-1] = m_pixel_filtered_flags[1][pixelNum] = m_pixel_filtered_flags[1][pixelNum+1] = m_pixel_filtered_flags[1][pixelNum+2] = m_pixel_filtered_flags[1][pixelNum+3] = e84s; } } } else // could not build 4x4 so move forward past the stacked 4x2s. { row1Ptr += 4*eBufferedPixelWidthInBytes; row2Ptr += 4*eBufferedPixelWidthInBytes; row3Ptr += 4*eBufferedPixelWidthInBytes; row4Ptr += 4*eBufferedPixelWidthInBytes; } pixelNum += 4; } else if ((m_pixel_filtered_flags[0][pixelNum] & e22w) && (m_pixel_filtered_flags[1][pixelNum] & e22w)) { /* - - | | | | - - we have 2 2x2s. - - | | | | - - */ ASSERT(m_pixel_filtered_flags[0][pixelNum] == e22w && m_pixel_filtered_flags[0][pixelNum+1] == e22e); ASSERT(m_pixel_filtered_flags[1][pixelNum] == e22w && m_pixel_filtered_flags[1][pixelNum+1] == e22e); upPixel = get4Pixel(row1Ptr); currPixel = get4Pixel(row3Ptr); R1 = GetRed(currPixel); G1 = GetGreen(currPixel); B1 = GetBlue(currPixel); R0 = GetRed(upPixel); G0 = GetGreen(upPixel); B0 = GetBlue(upPixel); if ((maxErrorForEightPixels >= 3) && (NewDeltaE(R1, R0, G1, G0, B1, B0, maxErrorForEightPixels))) { /* - - | | | | build 2x4 | | | | - - */ #if kGatherStats == 1 blockStats[es24nw]++; #endif AverageNRound((pixelNum & 0x02) == 0x02, R1, R1, R0, G1, G1, G0, B1, B1, B0); if(m_eEndian == LITTLEENDIAN) currPixel = (R1<<16) + (G1<<8) + B1; else if(m_eEndian == BIGENDIAN) currPixel = (R1<<24) + (G1<<16) + (B1<<8); #if kMemWritesOptimize == 0 put4Pixel(row1Ptr, 0, currPixel); put4Pixel(row1Ptr, 1, currPixel); put4Pixel(row2Ptr, 0, currPixel); put4Pixel(row2Ptr, 1, currPixel); put4Pixel(row3Ptr, 0, currPixel); put4Pixel(row3Ptr, 1, currPixel); put4Pixel(row4Ptr, 0, currPixel); put4Pixel(row4Ptr, 1, currPixel); #else put4Pixel(row1Ptr, 0, currPixel); #endif row1Ptr += 2*eBufferedPixelWidthInBytes; row2Ptr += 2*eBufferedPixelWidthInBytes; row3Ptr += 2*eBufferedPixelWidthInBytes; row4Ptr += 2*eBufferedPixelWidthInBytes; m_pixel_filtered_flags[0][pixelNum] = e24nw; m_pixel_filtered_flags[0][pixelNum+1] = e24ne; m_pixel_filtered_flags[1][pixelNum] = e24sw; m_pixel_filtered_flags[1][pixelNum+1] = e24se; } else { row1Ptr += 2*eBufferedPixelWidthInBytes; row2Ptr += 2*eBufferedPixelWidthInBytes; row3Ptr += 2*eBufferedPixelWidthInBytes; row4Ptr += 2*eBufferedPixelWidthInBytes; } pixelNum += 2; } else if ((m_pixel_filtered_flags[0][pixelNum] & e12) && (m_pixel_filtered_flags[1][pixelNum] & e12)) { /* - | | | | - we have two 1x2s. - | | | | - */ ASSERT(m_pixel_filtered_flags[0][pixelNum] == e12); ASSERT(m_pixel_filtered_flags[1][pixelNum] == e12); upPixel = get4Pixel(row1Ptr); currPixel = get4Pixel(row3Ptr); R1 = GetRed(currPixel); G1 = GetGreen(currPixel); B1 = GetBlue(currPixel); R0 = GetRed(upPixel); G0 = GetGreen(upPixel); B0 = GetBlue(upPixel); if ((maxErrorForFourPixels >= 3) && (NewDeltaE(R1, R0, G1, G0, B1, B0, maxErrorForFourPixels))) { /* - | | | | build 1x4 | | | | - */ #if kGatherStats == 1 blockStats[es14n]++; #endif AverageNRound((pixelNum & 0x01) == 0x01, R1, R1, R0, G1, G1, G0, B1, B1, B0); if(m_eEndian == LITTLEENDIAN) currPixel = (R1<<16) + (G1<<8) + B1; else if(m_eEndian == BIGENDIAN) currPixel = (R1<<24) + (G1<<16) + (B1<<8); #if kMemWritesOptimize == 0 put4Pixel(row1Ptr, 0, currPixel); put4Pixel(row2Ptr, 0, currPixel); put4Pixel(row3Ptr, 0, currPixel); put4Pixel(row4Ptr, 0, currPixel); #else put4Pixel(row1Ptr, 0, currPixel); #endif m_pixel_filtered_flags[0][pixelNum] = e14n; m_pixel_filtered_flags[1][pixelNum] = e14s; } row1Ptr += eBufferedPixelWidthInBytes; row2Ptr += eBufferedPixelWidthInBytes; row3Ptr += eBufferedPixelWidthInBytes; row4Ptr += eBufferedPixelWidthInBytes; pixelNum++; } else if ((m_pixel_filtered_flags[0][pixelNum] & e41si) && (m_pixel_filtered_flags[1][pixelNum] & e41ni)) { /* - - - - | | - - - - We have two 4x1s. - - - - | | - - - - */ upPixel = get4Pixel(row2Ptr); currPixel = get4Pixel(row3Ptr); R1 = GetRed(currPixel); G1 = GetGreen(currPixel); B1 = GetBlue(currPixel); R0 = GetRed(upPixel); G0 = GetGreen(upPixel); B0 = GetBlue(upPixel); if ((maxErrorForEightPixels >= 3) && (NewDeltaE(R1, R0, G1, G0, B1, B0, maxErrorForEightPixels))) { /* - - - - | | build 4x2. | | - - - - */ #if kGatherStats == 1 blockStats[es42w]++; #endif AverageNRound((pixelNum & 0x04) == 0x04, R1, R1, R0, G1, G1, G0, B1, B1, B0); if(m_eEndian == LITTLEENDIAN) currPixel = (R1<<16) + (G1<<8) + B1; else if(m_eEndian == BIGENDIAN) currPixel = (R1<<24) + (G1<<16) + (B1<<8); // Note we write this block out now and do not delay the writes for the postprocessing step since we do not track this block. put4Pixel(row2Ptr, 0, currPixel); put4Pixel(row2Ptr, 1, currPixel); put4Pixel(row2Ptr, 2, currPixel); put4Pixel(row2Ptr, 3, currPixel); put4Pixel(row3Ptr, 0, currPixel); put4Pixel(row3Ptr, 1, currPixel); put4Pixel(row3Ptr, 2, currPixel); put4Pixel(row3Ptr, 3, currPixel); m_pixel_filtered_flags[0][pixelNum] = m_pixel_filtered_flags[0][pixelNum] & ~e41si; m_pixel_filtered_flags[0][pixelNum+1] = m_pixel_filtered_flags[0][pixelNum+1] & ~e41s; m_pixel_filtered_flags[0][pixelNum+2] = m_pixel_filtered_flags[0][pixelNum+1] & ~e41s; m_pixel_filtered_flags[0][pixelNum+3] = m_pixel_filtered_flags[0][pixelNum+1] & ~e41s; m_pixel_filtered_flags[1][pixelNum] = m_pixel_filtered_flags[1][pixelNum] & ~e41ni; // Note that we just formed a 2x2 in the middle of the filtered sets of rows (and do not remember it). We then remove the 2x1s that the 2x2 eliminated. m_pixel_filtered_flags[1][pixelNum+1] = m_pixel_filtered_flags[1][pixelNum+1] & ~e41n; m_pixel_filtered_flags[1][pixelNum+2] = m_pixel_filtered_flags[1][pixelNum+1] & ~e41n; m_pixel_filtered_flags[1][pixelNum+3] = m_pixel_filtered_flags[1][pixelNum+1] & ~e41n; } pixelNum += 4; row1Ptr += 4*eBufferedPixelWidthInBytes; row2Ptr += 4*eBufferedPixelWidthInBytes; row3Ptr += 4*eBufferedPixelWidthInBytes; row4Ptr += 4*eBufferedPixelWidthInBytes; } else if ((m_pixel_filtered_flags[0][pixelNum] & e21sw) && (m_pixel_filtered_flags[1][pixelNum] & e21nw)) { /* - - | | - - We have two 2x1s. - - | | - - */ ASSERT(!((m_pixel_filtered_flags[0][pixelNum] & e11s) | (m_pixel_filtered_flags[0][pixelNum+1] & e11s))); ASSERT(!((m_pixel_filtered_flags[1][pixelNum] & e11n) | (m_pixel_filtered_flags[1][pixelNum+1] & e11n))); upPixel = get4Pixel(row2Ptr); currPixel = get4Pixel(row3Ptr); R1 = GetRed(currPixel); G1 = GetGreen(currPixel); B1 = GetBlue(currPixel); R0 = GetRed(upPixel); G0 = GetGreen(upPixel); B0 = GetBlue(upPixel); if ((maxErrorForFourPixels >= 3) && (NewDeltaE(R1, R0, G1, G0, B1, B0, maxErrorForFourPixels))) { /* - - | | build 2x2. | | - - */ #if kGatherStats == 1 blockStats[es22w]++; #endif AverageNRound((pixelNum & 0x02) == 0x02, R1, R1, R0, G1, G1, G0, B1, B1, B0); if(m_eEndian == LITTLEENDIAN) currPixel = (R1<<16) + (G1<<8) + B1; else if(m_eEndian == BIGENDIAN) currPixel = (R1<<24) + (G1<<16) + (B1<<8); // Note we write this block out now and do not delay the writes for the postprocessing step since we do not track this block. put4Pixel(row2Ptr, 0, currPixel); put4Pixel(row2Ptr, 1, currPixel); put4Pixel(row3Ptr, 0, currPixel); put4Pixel(row3Ptr, 1, currPixel); m_pixel_filtered_flags[0][pixelNum] = m_pixel_filtered_flags[0][pixelNum] & ~e21sw; m_pixel_filtered_flags[0][pixelNum+1] = m_pixel_filtered_flags[0][pixelNum+1] & ~e21se; m_pixel_filtered_flags[1][pixelNum] = m_pixel_filtered_flags[1][pixelNum] & ~e21nw; // Note that we just formed a 2x2 in the middle of the filtered sets of rows (and do not remember it). We then remove the 2x1s that the 2x2 eliminated. m_pixel_filtered_flags[1][pixelNum+1] = m_pixel_filtered_flags[1][pixelNum+1] & ~e21ne; } pixelNum += 2; row1Ptr += 2*eBufferedPixelWidthInBytes; row2Ptr += 2*eBufferedPixelWidthInBytes; row3Ptr += 2*eBufferedPixelWidthInBytes; row4Ptr += 2*eBufferedPixelWidthInBytes; } else if ((m_pixel_filtered_flags[0][pixelNum] & e11s) && (m_pixel_filtered_flags[1][pixelNum] & e11n)) { /* - | | - We have two 1x1s. - | | - */ upPixel = get4Pixel(row2Ptr); currPixel = get4Pixel(row3Ptr); R1 = GetRed(currPixel); G1 = GetGreen(currPixel); B1 = GetBlue(currPixel); R0 = GetRed(upPixel); G0 = GetGreen(upPixel); B0 = GetBlue(upPixel); if ((m_max_error_for_two_pixels >= 3) && (NewDeltaE(R1, R0, G1, G0, B1, B0, m_max_error_for_two_pixels))) { /* - | | build 1x2. | | - */ #if kGatherStats == 1 blockStats[es12w]++; #endif AverageNRound(isOdd(pixelNum), R1, R1, R0, G1, G1, G0, B1, B1, B0); if(m_eEndian == LITTLEENDIAN) currPixel = (R1<<16) + (G1<<8) + B1; else if(m_eEndian == BIGENDIAN) currPixel = (R1<<24) + (G1<<16) + (B1<<8); // Note we write this block out now and do not delay the writes for the postprocessing step since we do not track this block. put4Pixel(row2Ptr, 0, currPixel); put4Pixel(row3Ptr, 0, currPixel); m_pixel_filtered_flags[0][pixelNum] = m_pixel_filtered_flags[0][pixelNum] & ~e11s; m_pixel_filtered_flags[1][pixelNum] = m_pixel_filtered_flags[1][pixelNum] & ~e11n; // Note that we just formed a 2x2 in the middle of the filtered sets of rows (and do not remember it). We then remove the 2x1s that the 2x2 eliminated. } pixelNum += 1; row1Ptr += eBufferedPixelWidthInBytes; row2Ptr += eBufferedPixelWidthInBytes; row3Ptr += eBufferedPixelWidthInBytes; row4Ptr += eBufferedPixelWidthInBytes; } else // Do no vertical filtering here. { pixelNum += 1; row1Ptr += eBufferedPixelWidthInBytes; row2Ptr += eBufferedPixelWidthInBytes; row3Ptr += eBufferedPixelWidthInBytes; row4Ptr += eBufferedPixelWidthInBytes; } } } // Filter3FilteredRows. This routine only exists for the case of the odd size band // with three rasters left over. I'm not sure how much extra benifit we really // get from running this, but for now I'm leaving it in. Since Ernie deals with // block sizes that are powers of two its rather difficult to filter 3 rows together, // about all I've been able to do is look for 1 high blocks in the second and third // rasters to put together into 2 high blocks. This routine will create 4x2, 2x2, and // 1x2 blocks from those second and third rasters. void ErnieFilter::Filter3FilteredRows(unsigned char *row1Ptr, unsigned char *row2Ptr, unsigned char *row3Ptr) { const unsigned int maxErrorForFourPixels = m_max_error_for_two_pixels / 2; const unsigned int maxErrorForEightPixels = maxErrorForFourPixels / 2; // const unsigned int maxErrorForSixteenPixels = maxErrorForEightPixels / 2; // const unsigned int maxErrorForThirtyTwoPixels = maxErrorForSixteenPixels / 2; for (int pixelNum = 0; pixelNum < (m_row_width_in_pixels-3);) // Make sure we have four pixels to work with { // int currPixel, upPixel; uint32_t currPixel, upPixel; int R0, G0, B0, R1, G1, B1; if ((m_pixel_filtered_flags[0][pixelNum] & e41si) && (m_pixel_filtered_flags[1][pixelNum] & e41ni)) { /* - - - - | | - - - - We have two 4x1s. - - - - | | - - - - */ ASSERT(!((m_pixel_filtered_flags[0][pixelNum] & e11s) | (m_pixel_filtered_flags[0][pixelNum+1] & e11s))); ASSERT(!((m_pixel_filtered_flags[1][pixelNum] & e11n) | (m_pixel_filtered_flags[1][pixelNum+1] & e11n))); upPixel = get4Pixel(row2Ptr); currPixel = get4Pixel(row3Ptr); R1 = GetRed(currPixel); G1 = GetGreen(currPixel); B1 = GetBlue(currPixel); R0 = GetRed(upPixel); G0 = GetGreen(upPixel); B0 = GetBlue(upPixel); if ((maxErrorForEightPixels >= 3) && (NewDeltaE(R1, R0, G1, G0, B1, B0, maxErrorForEightPixels))) { /* - - - - | | build 4x2. | | - - - - */ #if kGatherStats == 1 blockStats[es42w]++; #endif AverageNRound((pixelNum & 0x04) == 0x04, R1, R1, R0, G1, G1, G0, B1, B1, B0); if(m_eEndian == LITTLEENDIAN) currPixel = (R1<<16) + (G1<<8) + B1; else if(m_eEndian == BIGENDIAN) currPixel = (R1<<24) + (G1<<16) + (B1<<8); // Note we write this block out now and do not delay the writes for the postprocessing step since we do not track this block. put4Pixel(row2Ptr, 0, currPixel); put4Pixel(row2Ptr, 1, currPixel); put4Pixel(row2Ptr, 2, currPixel); put4Pixel(row2Ptr, 3, currPixel); put4Pixel(row3Ptr, 0, currPixel); put4Pixel(row3Ptr, 1, currPixel); put4Pixel(row3Ptr, 2, currPixel); put4Pixel(row3Ptr, 3, currPixel); m_pixel_filtered_flags[0][pixelNum] = m_pixel_filtered_flags[0][pixelNum] & ~e41si; m_pixel_filtered_flags[0][pixelNum+1] = m_pixel_filtered_flags[0][pixelNum+1] & ~e41s; m_pixel_filtered_flags[0][pixelNum+2] = m_pixel_filtered_flags[0][pixelNum+1] & ~e41s; m_pixel_filtered_flags[0][pixelNum+3] = m_pixel_filtered_flags[0][pixelNum+1] & ~e41s; m_pixel_filtered_flags[1][pixelNum] = m_pixel_filtered_flags[1][pixelNum] & ~e41ni; // Note that we just formed a 2x2 in the middle of the filtered sets of rows (and do not remember it). We then remove the 2x1s that the 2x2 eliminated. m_pixel_filtered_flags[1][pixelNum+1] = m_pixel_filtered_flags[1][pixelNum+1] & ~e41n; m_pixel_filtered_flags[1][pixelNum+2] = m_pixel_filtered_flags[1][pixelNum+1] & ~e41n; m_pixel_filtered_flags[1][pixelNum+3] = m_pixel_filtered_flags[1][pixelNum+1] & ~e41n; } pixelNum += 4; row1Ptr += 4*eBufferedPixelWidthInBytes; row2Ptr += 4*eBufferedPixelWidthInBytes; row3Ptr += 4*eBufferedPixelWidthInBytes; } else if ((m_pixel_filtered_flags[0][pixelNum] & e21sw) && (m_pixel_filtered_flags[1][pixelNum] & e21nw)) { /* - - | | - - We have two 2x1s. - - | | - - */ ASSERT(!((m_pixel_filtered_flags[0][pixelNum] & e11s) | (m_pixel_filtered_flags[0][pixelNum+1] & e11s))); ASSERT(!((m_pixel_filtered_flags[1][pixelNum] & e11n) | (m_pixel_filtered_flags[1][pixelNum+1] & e11n))); upPixel = get4Pixel(row2Ptr); currPixel = get4Pixel(row3Ptr); R1 = GetRed(currPixel); G1 = GetGreen(currPixel); B1 = GetBlue(currPixel); R0 = GetRed(upPixel); G0 = GetGreen(upPixel); B0 = GetBlue(upPixel); if ((maxErrorForFourPixels >= 3) && (NewDeltaE(R1, R0, G1, G0, B1, B0, maxErrorForFourPixels))) { /* - - | | build 2x2. | | - - */ #if kGatherStats == 1 blockStats[es22w]++; #endif AverageNRound((pixelNum & 0x02) == 0x02, R1, R1, R0, G1, G1, G0, B1, B1, B0); if(m_eEndian == LITTLEENDIAN) currPixel = (R1<<16) + (G1<<8) + B1; else if(m_eEndian == BIGENDIAN) currPixel = (R1<<24) + (G1<<16) + (B1<<8); // Note we write this block out now and do not delay the writes for the postprocessing step since we do not track this block. put4Pixel(row2Ptr, 0, currPixel); put4Pixel(row2Ptr, 1, currPixel); put4Pixel(row3Ptr, 0, currPixel); put4Pixel(row3Ptr, 1, currPixel); m_pixel_filtered_flags[0][pixelNum] = m_pixel_filtered_flags[0][pixelNum] & ~e21sw; m_pixel_filtered_flags[0][pixelNum+1] = m_pixel_filtered_flags[0][pixelNum+1] & ~e21se; m_pixel_filtered_flags[1][pixelNum] = m_pixel_filtered_flags[1][pixelNum] & ~e21nw; // Note that we just formed a 2x2 in the middle of the filtered sets of rows (and do not remember it). We then remove the 2x1s that the 2x2 eliminated. m_pixel_filtered_flags[1][pixelNum+1] = m_pixel_filtered_flags[1][pixelNum+1] & ~e21ne; } pixelNum += 2; row1Ptr += 2*eBufferedPixelWidthInBytes; row2Ptr += 2*eBufferedPixelWidthInBytes; row3Ptr += 2*eBufferedPixelWidthInBytes; } else if ((m_pixel_filtered_flags[0][pixelNum] & e11s) && (m_pixel_filtered_flags[1][pixelNum] & e11n)) { /* - | | - We have two 1x1s. - | | - */ upPixel = get4Pixel(row2Ptr); currPixel = get4Pixel(row3Ptr); R1 = GetRed(currPixel); G1 = GetGreen(currPixel); B1 = GetBlue(currPixel); R0 = GetRed(upPixel); G0 = GetGreen(upPixel); B0 = GetBlue(upPixel); if ((m_max_error_for_two_pixels >= 3) && (NewDeltaE(R1, R0, G1, G0, B1, B0, m_max_error_for_two_pixels))) { /* - | | build 1x2. | | - */ #if kGatherStats == 1 blockStats[es12w]++; #endif AverageNRound(isOdd(pixelNum), R1, R1, R0, G1, G1, G0, B1, B1, B0); if(m_eEndian == LITTLEENDIAN) currPixel = (R1<<16) + (G1<<8) + B1; else if(m_eEndian == BIGENDIAN) currPixel = (R1<<24) + (G1<<16) + (B1<<8); // Note we write this block out now and do not delay the writes for the postprocessing step since we do not track this block. put4Pixel(row2Ptr, 0, currPixel); put4Pixel(row3Ptr, 0, currPixel); m_pixel_filtered_flags[0][pixelNum] = m_pixel_filtered_flags[0][pixelNum] & ~e11s; m_pixel_filtered_flags[1][pixelNum] = m_pixel_filtered_flags[1][pixelNum] & ~e11n; // Note that we just formed a 2x2 in the middle of the filtered sets of rows (and do not remember it). We then remove the 2x1s that the 2x2 eliminated. } pixelNum += 1; row1Ptr += eBufferedPixelWidthInBytes; row2Ptr += eBufferedPixelWidthInBytes; row3Ptr += eBufferedPixelWidthInBytes; } else // Do no vertical filtering here. { pixelNum += 1; row1Ptr += eBufferedPixelWidthInBytes; row2Ptr += eBufferedPixelWidthInBytes; row3Ptr += eBufferedPixelWidthInBytes; } } } #define NEWTEST true inline bool ErnieFilter::NewDeltaE(int dr0, int dr1, int dg0, int dg1, int db0, int db1, int tolerance) { int Y0, Y1, dY, Cr0, Cr1, Cb0, Cb1, dCr, dCb; // new Delta E stuff from Jay Y0 = 5*dr0 + 9*dg0 + 2*db0; Y1 = 5*dr1 + 9*dg1 + 2*db1; dY = ABS(Y0 - Y1) >> 4; if(dY > tolerance) { return false; } else { Cr0 = (dr0 << 4) - Y0; Cr1 = (dr1 << 4) - Y1; dCr = ABS(Cr0 - Cr1) >> 5; if(dCr > tolerance) { return false; } else { Cb0 = (db0 << 4) - Y0; Cb1 = (db1 << 4) - Y1; dCb = ABS(Cb0 - Cb1) >> 6; if(dCb > tolerance) { return false; } } } return true; } ErnieFilter::ErnieFilter(int rowWidthInPixels, pixelTypes pixelType, unsigned int maxErrorForTwoPixels) { int index; m_input_bytes_per_pixel = 3; ASSERT(rowWidthInPixels > 0); ASSERT(pixelType == eBGRPixelData); union { short s; char c[2]; } uEndian; uEndian.s = 0x0A0B; m_eEndian = LITTLEENDIAN; if (uEndian.c[0] == 0x0A) m_eEndian = BIGENDIAN; m_internal_bytes_per_pixel = 4; m_pixel_offsets_index = 0; m_row_width_in_pixels = rowWidthInPixels; m_row_width_in_bytes = m_row_width_in_pixels*m_internal_bytes_per_pixel; m_max_error_for_two_pixels = maxErrorForTwoPixels; for (index = 0; index < 4; index++) { m_row_bufs[index] = new uint32_t[rowWidthInPixels]; ASSERT(m_row_bufs[index]); m_row_ptrs[index] = new unsigned char[rowWidthInPixels*m_input_bytes_per_pixel]; ASSERT(m_row_ptrs[index]); m_black_row_ptrs[index] = new BYTE[rowWidthInPixels*m_input_bytes_per_pixel]; ASSERT(m_black_row_ptrs[index]); m_black_raster_sizes[index] = 0; } for (index = 0; index < 2; index++) { m_pixel_filtered_flags[index] = new unsigned int[rowWidthInPixels]; ASSERT(m_pixel_filtered_flags[index]); } // The least compressible image will be all raw pixels. Maximum compressed size is: // full size + a bloat of Cmd byte + 1 VLI byte per 255 pixels rounded up to nearest integer. int maxCompressionBufSize = m_row_width_in_bytes + 1 + ((int)ceil((double) MAX((rowWidthInPixels-2)/255, 0))); m_compression_out_buf = new unsigned char[maxCompressionBufSize]; ASSERT(m_compression_out_buf); m_buffered_row_count = 0; m_pixel_offsets[0] = 0; m_pixel_offsets[1] = 5; m_pixel_offsets[2] = 2; m_pixel_offsets[3] = 7; m_pixel_offsets[4] = 1; m_pixel_offsets[5] = 4; m_pixel_offsets[6] = 6; m_pixel_offsets[7] = 3; m_row_index = 0; } ErnieFilter::~ErnieFilter() { // Deallocate memory next. int index; for (index = 0; index < 4; index++) { delete [] m_row_bufs[index]; delete [] m_row_ptrs[index]; delete [] m_black_row_ptrs[index]; } for (index = 0; index < 2; index++) { delete [] m_pixel_filtered_flags[index]; } delete [] m_compression_out_buf; } void ErnieFilter::writeBufferedRows() { int pixelIndex = 0; // We just have one lonely raster left. Nothing // we can do but filter it horizontally. if( 1 == m_buffered_row_count) { int offset2 = m_pixel_offsets[m_pixel_offsets_index]; Filter1RawRow( (unsigned char*)(m_row_bufs[0] + offset2), m_row_width_in_pixels - m_pixel_offsets[m_pixel_offsets_index], m_pixel_filtered_flags[0] + m_pixel_offsets[m_pixel_offsets_index]); unsigned char *rowPtr = m_row_ptrs[0]; ASSERT(rowPtr); pixelIndex = 0; do { memcpy(rowPtr, &m_row_bufs[0][pixelIndex], 3); rowPtr += 3; } while (++pixelIndex < m_row_width_in_pixels); } // If we've got a pair of rasters in the buffer, that pair // has already been filtered somewhat. So lets just write them // out, some filtering is better than none. else if (2 == m_buffered_row_count) { // Write the two rows back out. int k; for (k = 0; k < 2; k++) { unsigned char *rowPtr = m_row_ptrs[k]; ASSERT(rowPtr); pixelIndex = 0; do { memcpy(rowPtr, &m_row_bufs[k][pixelIndex], 3); rowPtr += 3; } while (++pixelIndex < m_row_width_in_pixels); } } // Okay, if we had three rasters in the buffer, the pair // should have already been written out above, so lets // just run the odd raster through Ernie with to // get the horizontal filtering. [Need to look to see // if there's something more we can do with filtering // all three together.] else if (3 == m_buffered_row_count) { int offset2 = m_pixel_offsets[m_pixel_offsets_index]; Filter1RawRow( (unsigned char*)(m_row_bufs[2] + offset2), m_row_width_in_pixels - m_pixel_offsets[m_pixel_offsets_index], m_pixel_filtered_flags[1] + m_pixel_offsets[m_pixel_offsets_index]); Filter3FilteredRows( (unsigned char*)m_row_bufs[0], (unsigned char*)m_row_bufs[1], (unsigned char*)m_row_bufs[2]); int k; for (k = 0; k < 3; k++) { unsigned char *rowPtr = m_row_ptrs[k]; ASSERT(rowPtr); pixelIndex = 0; do { memcpy(rowPtr, &m_row_bufs[k][pixelIndex], 3); rowPtr += 3; } while (++pixelIndex < m_row_width_in_pixels); } } } void ErnieFilter::submitRowToFilter(unsigned char *rowPtr) { memcpy(m_row_ptrs[m_buffered_row_count], rowPtr, m_row_width_in_pixels*m_input_bytes_per_pixel); int pixelIndex = 0; uint32_t *RowPtrDest = m_row_bufs[m_buffered_row_count]; BYTE byte1 = 0; BYTE byte2 = 0; BYTE byte3 = 0; do { byte1 = *rowPtr++; byte2 = *rowPtr++; byte3 = *rowPtr++; if(m_eEndian == LITTLEENDIAN) RowPtrDest[pixelIndex] = ((byte3 << 16) | (byte2 << 8) | (byte1)) & 0x00FFFFFF; else if(m_eEndian == BIGENDIAN) RowPtrDest[pixelIndex] = ((byte1 << 24) | (byte2 << 16) | (byte3 << 8)) & 0xFFFFFF00; } while (++pixelIndex < m_row_width_in_pixels); m_buffered_row_count++; iRastersReady=0; iRastersDelivered=0; // Next see about filtering & compression. // NOTE 1: as an optimization only do subsections of the raster at a time to stay in cache. // NOTE 2: Could filter the pixels left of the offset. if (2 == m_buffered_row_count) { int offset2 = m_pixel_offsets[m_pixel_offsets_index]; Filter2RawRows( (unsigned char*)(m_row_bufs[1] + offset2), (unsigned char*)(m_row_bufs[0] + offset2), m_row_width_in_pixels - m_pixel_offsets[m_pixel_offsets_index], m_pixel_filtered_flags[0] + m_pixel_offsets[m_pixel_offsets_index]); } if (4 == m_buffered_row_count) { int offset4 = m_pixel_offsets[m_pixel_offsets_index]; Filter2RawRows( (unsigned char*)(m_row_bufs[3] + offset4), (unsigned char*)(m_row_bufs[2] + offset4), m_row_width_in_pixels - m_pixel_offsets[m_pixel_offsets_index], m_pixel_filtered_flags[1] + m_pixel_offsets[m_pixel_offsets_index]); Filter2PairsOfFilteredRows( (unsigned char*)m_row_bufs[0], (unsigned char*)m_row_bufs[1], (unsigned char*)m_row_bufs[2], (unsigned char*)m_row_bufs[3]); #if kMemWritesOptimize == 1 // Writing the blocks out on a post processing step in this manner could leave the last 3 rows // unfiltered. This is a trade off we make for simplicity. The resulting loss in compression is small. WriteBlockPixels(); #endif m_pixel_offsets_index = (m_pixel_offsets_index + 1) % 8; // cycle the offset index. int k; for (k = 0; k < m_pixel_offsets[m_pixel_offsets_index]; k++) // Clear out the flags that we're offsetting past for this next iteration. { m_pixel_filtered_flags[0][k] = eDone; m_pixel_filtered_flags[1][k] = eDone; } // Write the four rows back out. for (k = 0; k < 4; k++) { unsigned char *rowPtr = m_row_ptrs[k]; ASSERT(rowPtr); pixelIndex = 0; do { memcpy(rowPtr, &m_row_bufs[k][pixelIndex], m_input_bytes_per_pixel); rowPtr += m_input_bytes_per_pixel; } while (++pixelIndex < m_row_width_in_pixels); } m_buffered_row_count = 0; iRastersReady = 4; } } #if kMemWritesOptimize == 1 /* At this point the color for the entire block is stored in the top left corner of the block. This routine takes that pixel and smears it into the rest of the block. */ void ErnieFilter::WriteBlockPixels(void) { unsigned char *row1Ptr = (unsigned char*)m_row_bufs[0]; unsigned char *row2Ptr = (unsigned char*)m_row_bufs[1]; unsigned char *row3Ptr = (unsigned char*)m_row_bufs[2]; unsigned char *row4Ptr = (unsigned char*)m_row_bufs[3]; for (int flagSet = 0; flagSet <= 1; flagSet++) { unsigned int *flagsPtr = m_pixel_filtered_flags[0]; unsigned char *rowA = (unsigned char*)m_row_bufs[0]; unsigned char *rowB = (unsigned char*)m_row_bufs[1]; if (flagSet == 1) { flagsPtr = m_pixel_filtered_flags[1]; rowA = (unsigned char*)m_row_bufs[2]; rowB = (unsigned char*)m_row_bufs[3]; } for (int rowIndex = 0; rowIndex < m_row_width_in_pixels;) { unsigned int currentFlags = flagsPtr[rowIndex]; #ifndef NDEBUG /* only done for debug builds */ int numberOfBitsSet = 0; unsigned int currentFlagsCopy = currentFlags & eTopLeftOfBlocks; while (currentFlagsCopy) { if (currentFlagsCopy & 1) numberOfBitsSet++; currentFlagsCopy >>= 1; } ASSERT( (numberOfBitsSet <= 1) || ((numberOfBitsSet == 2) && (((currentFlags & eTopLeftOfBlocks) & ~(e21nw|e21sw|e41ni|e41si))==0))); #endif if (currentFlags & eTopLeftOfBlocks) // Avoids doing a lot of checks if nothing is set. { // unsigned int pixel; uint32_t pixel; // The three possible scenerios are: // 1: No top left of block bits are set. // 2: 1 top left block bit is set. // 3: 2 top left block bits are set. They are 21nw and 21sw. // Note: Due to possibly having two groups tracked by this flag we require the north checks to occur before the south checks. if (currentFlags & e22w) { pixel = get4Pixel(rowA, rowIndex); put4Pixel(rowB, rowIndex, pixel); rowIndex += 1; put4Pixel(rowA, rowIndex, pixel); put4Pixel(rowB, rowIndex, pixel); rowIndex += 1; continue; } if (currentFlags & e12) { put4Pixel(rowB, rowIndex, get4Pixel(rowA, rowIndex)); rowIndex += 1; continue; } if (currentFlags & e42i) { pixel = get4Pixel(rowA, rowIndex); put4Pixel(rowB, rowIndex, pixel); rowIndex += 1; put4Pixel(rowA, rowIndex, pixel); put4Pixel(rowB, rowIndex, pixel); rowIndex += 1; put4Pixel(rowB, rowIndex, pixel); put4Pixel(rowA, rowIndex, pixel); rowIndex += 1; put4Pixel(rowA, rowIndex, pixel); put4Pixel(rowB, rowIndex, pixel); rowIndex += 1; continue; } if (currentFlags & e84ni) { pixel = get4Pixel(rowA, rowIndex); put4Pixel(row2Ptr, rowIndex, pixel); put4Pixel(row3Ptr, rowIndex, pixel); put4Pixel(row4Ptr, rowIndex, pixel); rowIndex += 1; put4Pixel(row1Ptr, rowIndex, pixel); put4Pixel(row2Ptr, rowIndex, pixel); put4Pixel(row3Ptr, rowIndex, pixel); put4Pixel(row4Ptr, rowIndex, pixel); rowIndex += 1; put4Pixel(row1Ptr, rowIndex, pixel); put4Pixel(row2Ptr, rowIndex, pixel); put4Pixel(row3Ptr, rowIndex, pixel); put4Pixel(row4Ptr, rowIndex, pixel); rowIndex += 1; put4Pixel(row1Ptr, rowIndex, pixel); put4Pixel(row2Ptr, rowIndex, pixel); put4Pixel(row3Ptr, rowIndex, pixel); put4Pixel(row4Ptr, rowIndex, pixel); rowIndex += 1; put4Pixel(row1Ptr, rowIndex, pixel); put4Pixel(row2Ptr, rowIndex, pixel); put4Pixel(row3Ptr, rowIndex, pixel); put4Pixel(row4Ptr, rowIndex, pixel); rowIndex += 1; put4Pixel(row1Ptr, rowIndex, pixel); put4Pixel(row2Ptr, rowIndex, pixel); put4Pixel(row3Ptr, rowIndex, pixel); put4Pixel(row4Ptr, rowIndex, pixel); rowIndex += 1; put4Pixel(row1Ptr, rowIndex, pixel); put4Pixel(row2Ptr, rowIndex, pixel); put4Pixel(row3Ptr, rowIndex, pixel); put4Pixel(row4Ptr, rowIndex, pixel); rowIndex += 1; put4Pixel(row1Ptr, rowIndex, pixel); put4Pixel(row2Ptr, rowIndex, pixel); put4Pixel(row3Ptr, rowIndex, pixel); put4Pixel(row4Ptr, rowIndex, pixel); rowIndex += 1; continue; } if (currentFlags & e24nw) { pixel = get4Pixel(row1Ptr, rowIndex); put4Pixel(row2Ptr, rowIndex, pixel); put4Pixel(row3Ptr, rowIndex, pixel); put4Pixel(row4Ptr, rowIndex, pixel); rowIndex += 1; put4Pixel(row1Ptr, rowIndex, pixel); put4Pixel(row2Ptr, rowIndex, pixel); put4Pixel(row3Ptr, rowIndex, pixel); put4Pixel(row4Ptr, rowIndex, pixel); rowIndex += 1; continue; } if (currentFlags & e44ni) { pixel = get4Pixel(row1Ptr, rowIndex); put4Pixel(row2Ptr, rowIndex, pixel); put4Pixel(row3Ptr, rowIndex, pixel); put4Pixel(row4Ptr, rowIndex, pixel); rowIndex += 1; put4Pixel(row1Ptr, rowIndex, pixel); put4Pixel(row2Ptr, rowIndex, pixel); put4Pixel(row3Ptr, rowIndex, pixel); put4Pixel(row4Ptr, rowIndex, pixel); rowIndex += 1; put4Pixel(row1Ptr, rowIndex, pixel); put4Pixel(row2Ptr, rowIndex, pixel); put4Pixel(row3Ptr, rowIndex, pixel); put4Pixel(row4Ptr, rowIndex, pixel); rowIndex += 1; put4Pixel(row1Ptr, rowIndex, pixel); put4Pixel(row2Ptr, rowIndex, pixel); put4Pixel(row3Ptr, rowIndex, pixel); put4Pixel(row4Ptr, rowIndex, pixel); rowIndex += 1; continue; } if (currentFlags & e14n) { pixel = get4Pixel(row1Ptr, rowIndex); put4Pixel(row2Ptr, rowIndex, pixel); put4Pixel(row3Ptr, rowIndex, pixel); put4Pixel(row4Ptr, rowIndex, pixel); rowIndex += 1; continue; } if (currentFlags & e21nw) { put4Pixel(rowA, rowIndex+1, get4Pixel(rowA, rowIndex)); if (!(currentFlags & (e21sw|e41si))) // if no south groups { rowIndex += 2; continue; } } if (currentFlags & e41ni) { pixel = get4Pixel(rowA, rowIndex); put4Pixel(rowA, rowIndex+1, pixel); put4Pixel(rowA, rowIndex+2, pixel); put4Pixel(rowA, rowIndex+3, pixel); if (!(currentFlags & (e21sw|e41si))) // if no south groups. { rowIndex += 2; continue; } } if (currentFlags & e21sw) { put4Pixel(rowB, rowIndex+1, get4Pixel(rowB, rowIndex)); rowIndex += 2; continue; } if (currentFlags & e41si) { pixel = get4Pixel(rowB, rowIndex); put4Pixel(rowB, rowIndex+1, pixel); put4Pixel(rowB, rowIndex+2, pixel); put4Pixel(rowB, rowIndex+3, pixel); rowIndex += 2; continue; } } rowIndex += 1; } } } #endif // kMemWritesOptimize bool ErnieFilter::Process (RASTERDATA* ImageData) { if ( ImageData == NULL || (ImageData->rasterdata[COLORTYPE_COLOR] == NULL && ImageData->rasterdata[COLORTYPE_BLACK] == NULL)) { return false; } if (ImageData->rasterdata[COLORTYPE_BLACK]) { memcpy(m_black_row_ptrs[m_row_index], ImageData->rasterdata[COLORTYPE_BLACK], (ImageData->rastersize[COLORTYPE_BLACK] + 7) / 8); } m_black_raster_sizes[m_row_index++] = ImageData->rastersize[COLORTYPE_BLACK]; if (m_row_index == 4) m_row_index = 0; if (ImageData->rasterdata[COLORTYPE_COLOR]) { submitRowToFilter(ImageData->rasterdata[COLORTYPE_COLOR]); // something ready after 4th time only return (m_buffered_row_count == 0); } iRastersReady = 1; return true; } //Process bool ErnieFilter::NextOutputRaster(RASTERDATA& next_raster) { if (iRastersReady == 0){ return false; } next_raster.rastersize[COLORTYPE_COLOR] = m_row_width_in_pixels * m_input_bytes_per_pixel; next_raster.rasterdata[COLORTYPE_COLOR] = m_row_ptrs[iRastersDelivered]; next_raster.rastersize[COLORTYPE_BLACK] = m_black_raster_sizes[iRastersDelivered]; if ( m_black_raster_sizes[iRastersDelivered] > 0 ){ next_raster.rasterdata[COLORTYPE_BLACK] = m_black_row_ptrs[iRastersDelivered]; } else { next_raster.rasterdata[COLORTYPE_BLACK] = NULL; } iRastersReady--; iRastersDelivered++; if (iRastersDelivered == 4) iRastersDelivered = 0; return true; } //NextOutputRaster unsigned int ErnieFilter::GetMaxOutputWidth() { return m_row_width_in_pixels * m_input_bytes_per_pixel; } //GetMaxOutputWidth void ErnieFilter::Flush() { writeBufferedRows(); iRastersDelivered=0; m_pixel_offsets_index = 0; iRastersReady = m_buffered_row_count; m_buffered_row_count = 0; m_row_index = 0; } //Flush