michael@0: /* This Source Code Form is subject to the terms of the Mozilla Public
michael@0:  * License, v. 2.0. If a copy of the MPL was not distributed with this
michael@0:  * file, You can obtain one at http://mozilla.org/MPL/2.0/. */
michael@0: 
michael@0: #include "Blur.h"
michael@0: 
michael@0: #include "SSEHelpers.h"
michael@0: 
michael@0: #include <string.h>
michael@0: 
michael@0: namespace mozilla {
michael@0: namespace gfx {
michael@0: 
michael@0: MOZ_ALWAYS_INLINE
michael@0: __m128i Divide(__m128i aValues, __m128i aDivisor)
michael@0: {
michael@0:   const __m128i mask = _mm_setr_epi32(0x0, 0xffffffff, 0x0, 0xffffffff);
michael@0:   static const union {
michael@0:     int64_t i64[2];
michael@0:     __m128i m;
michael@0:   } roundingAddition = { { int64_t(1) << 31, int64_t(1) << 31 } };
michael@0: 
michael@0:   __m128i multiplied31 = _mm_mul_epu32(aValues, aDivisor);
michael@0:   __m128i multiplied42 = _mm_mul_epu32(_mm_srli_epi64(aValues, 32), aDivisor);
michael@0: 
michael@0:   // Add 1 << 31 before shifting or masking the lower 32 bits away, so that the
michael@0:   // result is rounded.
michael@0:   __m128i p_3_1 = _mm_srli_epi64(_mm_add_epi64(multiplied31, roundingAddition.m), 32);
michael@0:   __m128i p4_2_ = _mm_and_si128(_mm_add_epi64(multiplied42, roundingAddition.m), mask);
michael@0:   __m128i p4321 = _mm_or_si128(p_3_1, p4_2_);
michael@0:   return p4321;
michael@0: }
michael@0: 
michael@0: MOZ_ALWAYS_INLINE
michael@0: __m128i BlurFourPixels(const __m128i& aTopLeft, const __m128i& aTopRight,
michael@0:                        const __m128i& aBottomRight, const __m128i& aBottomLeft,
michael@0:                        const __m128i& aDivisor)
michael@0: {
michael@0:   __m128i values = _mm_add_epi32(_mm_sub_epi32(_mm_sub_epi32(aBottomRight, aTopRight), aBottomLeft), aTopLeft);
michael@0:   return Divide(values, aDivisor);
michael@0: }
michael@0: 
michael@0: MOZ_ALWAYS_INLINE
michael@0: void LoadIntegralRowFromRow(uint32_t *aDest, const uint8_t *aSource,
michael@0:                             int32_t aSourceWidth, int32_t aLeftInflation,
michael@0:                             int32_t aRightInflation)
michael@0: {
michael@0:   int32_t currentRowSum = 0;
michael@0: 
michael@0:   for (int x = 0; x < aLeftInflation; x++) {
michael@0:     currentRowSum += aSource[0];
michael@0:     aDest[x] = currentRowSum;
michael@0:   }
michael@0:   for (int x = aLeftInflation; x < (aSourceWidth + aLeftInflation); x++) {
michael@0:     currentRowSum += aSource[(x - aLeftInflation)];
michael@0:     aDest[x] = currentRowSum;
michael@0:   }
michael@0:   for (int x = (aSourceWidth + aLeftInflation); x < (aSourceWidth + aLeftInflation + aRightInflation); x++) {
michael@0:     currentRowSum += aSource[aSourceWidth - 1];
michael@0:     aDest[x] = currentRowSum;
michael@0:   }
michael@0: }
michael@0: 
michael@0: // This function calculates an integral of four pixels stored in the 4
michael@0: // 32-bit integers on aPixels. i.e. for { 30, 50, 80, 100 } this returns
michael@0: // { 30, 80, 160, 260 }. This seems to be the fastest way to do this after
michael@0: // much testing.
michael@0: MOZ_ALWAYS_INLINE
michael@0: __m128i AccumulatePixelSums(__m128i aPixels)
michael@0: {
michael@0:   __m128i sumPixels = aPixels;
michael@0:   __m128i currentPixels = _mm_slli_si128(aPixels, 4);
michael@0:   sumPixels = _mm_add_epi32(sumPixels, currentPixels);
michael@0:   currentPixels = _mm_unpacklo_epi64(_mm_setzero_si128(), sumPixels);
michael@0: 
michael@0:   return _mm_add_epi32(sumPixels, currentPixels);
michael@0: }
michael@0: 
michael@0: MOZ_ALWAYS_INLINE void
michael@0: GenerateIntegralImage_SSE2(int32_t aLeftInflation, int32_t aRightInflation,
michael@0:                            int32_t aTopInflation, int32_t aBottomInflation,
michael@0:                            uint32_t *aIntegralImage, size_t aIntegralImageStride,
michael@0:                            uint8_t *aSource, int32_t aSourceStride, const IntSize &aSize)
michael@0: {
michael@0:   MOZ_ASSERT(!(aLeftInflation & 3));
michael@0: 
michael@0:   uint32_t stride32bit = aIntegralImageStride / 4;
michael@0: 
michael@0:   IntSize integralImageSize(aSize.width + aLeftInflation + aRightInflation,
michael@0:                             aSize.height + aTopInflation + aBottomInflation);
michael@0: 
michael@0:   LoadIntegralRowFromRow(aIntegralImage, aSource, aSize.width, aLeftInflation, aRightInflation);
michael@0: 
michael@0:   for (int y = 1; y < aTopInflation + 1; y++) {
michael@0:     uint32_t *intRow = aIntegralImage + (y * stride32bit);
michael@0:     uint32_t *intPrevRow = aIntegralImage + (y - 1) * stride32bit;
michael@0:     uint32_t *intFirstRow = aIntegralImage;
michael@0: 
michael@0:     for (int x = 0; x < integralImageSize.width; x += 4) {
michael@0:       __m128i firstRow = _mm_load_si128((__m128i*)(intFirstRow + x));
michael@0:       __m128i previousRow = _mm_load_si128((__m128i*)(intPrevRow + x));
michael@0:       _mm_store_si128((__m128i*)(intRow + x), _mm_add_epi32(firstRow, previousRow));
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   for (int y = aTopInflation + 1; y < (aSize.height + aTopInflation); y++) {
michael@0:     __m128i currentRowSum = _mm_setzero_si128();
michael@0:     uint32_t *intRow = aIntegralImage + (y * stride32bit);
michael@0:     uint32_t *intPrevRow = aIntegralImage + (y - 1) * stride32bit;
michael@0:     uint8_t *sourceRow = aSource + aSourceStride * (y - aTopInflation);
michael@0: 
michael@0:     uint32_t pixel = sourceRow[0];
michael@0:     for (int x = 0; x < aLeftInflation; x += 4) {
michael@0:       __m128i sumPixels = AccumulatePixelSums(_mm_shuffle_epi32(_mm_set1_epi32(pixel), _MM_SHUFFLE(0, 0, 0, 0)));
michael@0: 
michael@0:       sumPixels = _mm_add_epi32(sumPixels, currentRowSum);
michael@0: 
michael@0:       currentRowSum = _mm_shuffle_epi32(sumPixels, _MM_SHUFFLE(3, 3, 3, 3));
michael@0: 
michael@0:       _mm_store_si128((__m128i*)(intRow + x), _mm_add_epi32(sumPixels, _mm_load_si128((__m128i*)(intPrevRow + x))));
michael@0:     }
michael@0:     for (int x = aLeftInflation; x < (aSize.width + aLeftInflation); x += 4) {
michael@0:       uint32_t pixels = *(uint32_t*)(sourceRow + (x - aLeftInflation));
michael@0: 
michael@0:       // It's important to shuffle here. When we exit this loop currentRowSum
michael@0:       // has to be set to sumPixels, so that the following loop can get the
michael@0:       // correct pixel for the currentRowSum. The highest order pixel in
michael@0:       // currentRowSum could've originated from accumulation in the stride.
michael@0:       currentRowSum = _mm_shuffle_epi32(currentRowSum, _MM_SHUFFLE(3, 3, 3, 3));
michael@0: 
michael@0:       __m128i sumPixels = AccumulatePixelSums(_mm_unpacklo_epi16(_mm_unpacklo_epi8( _mm_set1_epi32(pixels), _mm_setzero_si128()), _mm_setzero_si128()));
michael@0:       sumPixels = _mm_add_epi32(sumPixels, currentRowSum);
michael@0: 
michael@0:       currentRowSum = sumPixels;
michael@0: 
michael@0:       _mm_store_si128((__m128i*)(intRow + x), _mm_add_epi32(sumPixels, _mm_load_si128((__m128i*)(intPrevRow + x))));
michael@0:     }
michael@0: 
michael@0:     pixel = sourceRow[aSize.width - 1];
michael@0:     int x = (aSize.width + aLeftInflation);
michael@0:     if ((aSize.width & 3)) {
michael@0:       // Deal with unaligned portion. Get the correct pixel from currentRowSum,
michael@0:       // see explanation above.
michael@0:       uint32_t intCurrentRowSum = ((uint32_t*)&currentRowSum)[(aSize.width % 4) - 1];
michael@0:       for (; x < integralImageSize.width; x++) {
michael@0:         // We could be unaligned here!
michael@0:         if (!(x & 3)) {
michael@0:           // aligned!
michael@0:           currentRowSum = _mm_set1_epi32(intCurrentRowSum);
michael@0:           break;
michael@0:         }
michael@0:         intCurrentRowSum += pixel;
michael@0:         intRow[x] = intPrevRow[x] + intCurrentRowSum;
michael@0:       }
michael@0:     } else {
michael@0:       currentRowSum = _mm_shuffle_epi32(currentRowSum, _MM_SHUFFLE(3, 3, 3, 3));
michael@0:     }
michael@0:     for (; x < integralImageSize.width; x += 4) {
michael@0:       __m128i sumPixels = AccumulatePixelSums(_mm_set1_epi32(pixel));
michael@0: 
michael@0:       sumPixels = _mm_add_epi32(sumPixels, currentRowSum);
michael@0: 
michael@0:       currentRowSum = _mm_shuffle_epi32(sumPixels, _MM_SHUFFLE(3, 3, 3, 3));
michael@0: 
michael@0:       _mm_store_si128((__m128i*)(intRow + x), _mm_add_epi32(sumPixels, _mm_load_si128((__m128i*)(intPrevRow + x))));
michael@0:     }
michael@0:   }
michael@0: 
michael@0:   if (aBottomInflation) {
michael@0:     // Store the last valid row of our source image in the last row of
michael@0:     // our integral image. This will be overwritten with the correct values
michael@0:     // in the upcoming loop.
michael@0:     LoadIntegralRowFromRow(aIntegralImage + (integralImageSize.height - 1) * stride32bit,
michael@0:                            aSource + (aSize.height - 1) * aSourceStride, aSize.width, aLeftInflation, aRightInflation);
michael@0: 
michael@0: 
michael@0:     for (int y = aSize.height + aTopInflation; y < integralImageSize.height; y++) {
michael@0:       __m128i *intRow = (__m128i*)(aIntegralImage + (y * stride32bit));
michael@0:       __m128i *intPrevRow = (__m128i*)(aIntegralImage + (y - 1) * stride32bit);
michael@0:       __m128i *intLastRow = (__m128i*)(aIntegralImage + (integralImageSize.height - 1) * stride32bit);
michael@0: 
michael@0:       for (int x = 0; x < integralImageSize.width; x += 4) {
michael@0:         _mm_store_si128(intRow + (x / 4),
michael@0:                         _mm_add_epi32(_mm_load_si128(intLastRow + (x / 4)),
michael@0:                                       _mm_load_si128(intPrevRow + (x / 4))));
michael@0:       }
michael@0:     }
michael@0:   }
michael@0: }
michael@0: 
michael@0: /**
michael@0:  * Attempt to do an in-place box blur using an integral image.
michael@0:  */
michael@0: void
michael@0: AlphaBoxBlur::BoxBlur_SSE2(uint8_t* aData,
michael@0:                            int32_t aLeftLobe,
michael@0:                            int32_t aRightLobe,
michael@0:                            int32_t aTopLobe,
michael@0:                            int32_t aBottomLobe,
michael@0:                            uint32_t *aIntegralImage,
michael@0:                            size_t aIntegralImageStride)
michael@0: {
michael@0:   IntSize size = GetSize();
michael@0: 
michael@0:   MOZ_ASSERT(size.height > 0);
michael@0: 
michael@0:   // Our 'left' or 'top' lobe will include the current pixel. i.e. when
michael@0:   // looking at an integral image the value of a pixel at 'x,y' is calculated
michael@0:   // using the value of the integral image values above/below that.
michael@0:   aLeftLobe++;
michael@0:   aTopLobe++;
michael@0:   int32_t boxSize = (aLeftLobe + aRightLobe) * (aTopLobe + aBottomLobe);
michael@0: 
michael@0:   MOZ_ASSERT(boxSize > 0);
michael@0: 
michael@0:   if (boxSize == 1) {
michael@0:       return;
michael@0:   }
michael@0: 
michael@0:   uint32_t reciprocal = uint32_t((uint64_t(1) << 32) / boxSize);
michael@0: 
michael@0:   uint32_t stride32bit = aIntegralImageStride / 4;
michael@0:   int32_t leftInflation = RoundUpToMultipleOf4(aLeftLobe).value();
michael@0: 
michael@0:   GenerateIntegralImage_SSE2(leftInflation, aRightLobe, aTopLobe, aBottomLobe,
michael@0:                              aIntegralImage, aIntegralImageStride, aData,
michael@0:                              mStride, size);
michael@0: 
michael@0:   __m128i divisor = _mm_set1_epi32(reciprocal);
michael@0: 
michael@0:   // This points to the start of the rectangle within the IntegralImage that overlaps
michael@0:   // the surface being blurred.
michael@0:   uint32_t *innerIntegral = aIntegralImage + (aTopLobe * stride32bit) + leftInflation;
michael@0: 
michael@0:   IntRect skipRect = mSkipRect;
michael@0:   int32_t stride = mStride;
michael@0:   uint8_t *data = aData;
michael@0:   for (int32_t y = 0; y < size.height; y++) {
michael@0:     bool inSkipRectY = y > skipRect.y && y < skipRect.YMost();
michael@0: 
michael@0:     uint32_t *topLeftBase = innerIntegral + ((y - aTopLobe) * ptrdiff_t(stride32bit) - aLeftLobe);
michael@0:     uint32_t *topRightBase = innerIntegral + ((y - aTopLobe) * ptrdiff_t(stride32bit) + aRightLobe);
michael@0:     uint32_t *bottomRightBase = innerIntegral + ((y + aBottomLobe) * ptrdiff_t(stride32bit) + aRightLobe);
michael@0:     uint32_t *bottomLeftBase = innerIntegral + ((y + aBottomLobe) * ptrdiff_t(stride32bit) - aLeftLobe);
michael@0: 
michael@0:     int32_t x = 0;
michael@0:     // Process 16 pixels at a time for as long as possible.
michael@0:     for (; x <= size.width - 16; x += 16) {
michael@0:       if (inSkipRectY && x > skipRect.x && x < skipRect.XMost()) {
michael@0:         x = skipRect.XMost() - 16;
michael@0:         // Trigger early jump on coming loop iterations, this will be reset
michael@0:         // next line anyway.
michael@0:         inSkipRectY = false;
michael@0:         continue;
michael@0:       }
michael@0: 
michael@0:       __m128i topLeft;
michael@0:       __m128i topRight;
michael@0:       __m128i bottomRight;
michael@0:       __m128i bottomLeft;
michael@0: 
michael@0:       topLeft = loadUnaligned128((__m128i*)(topLeftBase + x));
michael@0:       topRight = loadUnaligned128((__m128i*)(topRightBase + x));
michael@0:       bottomRight = loadUnaligned128((__m128i*)(bottomRightBase + x));
michael@0:       bottomLeft = loadUnaligned128((__m128i*)(bottomLeftBase + x));
michael@0:       __m128i result1 = BlurFourPixels(topLeft, topRight, bottomRight, bottomLeft, divisor);
michael@0: 
michael@0:       topLeft = loadUnaligned128((__m128i*)(topLeftBase + x + 4));
michael@0:       topRight = loadUnaligned128((__m128i*)(topRightBase + x + 4));
michael@0:       bottomRight = loadUnaligned128((__m128i*)(bottomRightBase + x + 4));
michael@0:       bottomLeft = loadUnaligned128((__m128i*)(bottomLeftBase + x + 4));
michael@0:       __m128i result2 = BlurFourPixels(topLeft, topRight, bottomRight, bottomLeft, divisor);
michael@0: 
michael@0:       topLeft = loadUnaligned128((__m128i*)(topLeftBase + x + 8));
michael@0:       topRight = loadUnaligned128((__m128i*)(topRightBase + x + 8));
michael@0:       bottomRight = loadUnaligned128((__m128i*)(bottomRightBase + x + 8));
michael@0:       bottomLeft = loadUnaligned128((__m128i*)(bottomLeftBase + x + 8));
michael@0:       __m128i result3 = BlurFourPixels(topLeft, topRight, bottomRight, bottomLeft, divisor);
michael@0: 
michael@0:       topLeft = loadUnaligned128((__m128i*)(topLeftBase + x + 12));
michael@0:       topRight = loadUnaligned128((__m128i*)(topRightBase + x + 12));
michael@0:       bottomRight = loadUnaligned128((__m128i*)(bottomRightBase + x + 12));
michael@0:       bottomLeft = loadUnaligned128((__m128i*)(bottomLeftBase + x + 12));
michael@0:       __m128i result4 = BlurFourPixels(topLeft, topRight, bottomRight, bottomLeft, divisor);
michael@0: 
michael@0:       __m128i final = _mm_packus_epi16(_mm_packs_epi32(result1, result2), _mm_packs_epi32(result3, result4));
michael@0: 
michael@0:       _mm_storeu_si128((__m128i*)(data + stride * y + x), final);
michael@0:     }
michael@0: 
michael@0:     // Process the remaining pixels 4 bytes at a time.
michael@0:     for (; x < size.width; x += 4) {
michael@0:       if (inSkipRectY && x > skipRect.x && x < skipRect.XMost()) {
michael@0:         x = skipRect.XMost() - 4;
michael@0:         // Trigger early jump on coming loop iterations, this will be reset
michael@0:         // next line anyway.
michael@0:         inSkipRectY = false;
michael@0:         continue;
michael@0:       }
michael@0:       __m128i topLeft = loadUnaligned128((__m128i*)(topLeftBase + x));
michael@0:       __m128i topRight = loadUnaligned128((__m128i*)(topRightBase + x));
michael@0:       __m128i bottomRight = loadUnaligned128((__m128i*)(bottomRightBase + x));
michael@0:       __m128i bottomLeft = loadUnaligned128((__m128i*)(bottomLeftBase + x));
michael@0: 
michael@0:       __m128i result = BlurFourPixels(topLeft, topRight, bottomRight, bottomLeft, divisor);
michael@0:       __m128i final = _mm_packus_epi16(_mm_packs_epi32(result, _mm_setzero_si128()), _mm_setzero_si128());
michael@0: 
michael@0:       *(uint32_t*)(data + stride * y + x) = _mm_cvtsi128_si32(final);
michael@0:     }
michael@0:   }
michael@0: 
michael@0: }
michael@0: 
michael@0: }
michael@0: }