summaryrefslogtreecommitdiffstats
path: root/gfx/2d/BlurNEON.cpp
blob: 978b3cdc0d7552c80008d0c9032a75735abe7801 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
/* This Source Code Form is subject to the terms of the Mozilla Public
 * License, v. 2.0. If a copy of the MPL was not distributed with this
 * file, You can obtain one at http://mozilla.org/MPL/2.0/. */

#include "Blur.h"
#include <arm_neon.h>

namespace mozilla {
namespace gfx {

MOZ_ALWAYS_INLINE
uint16x4_t Divide(uint32x4_t aValues, uint32x2_t aDivisor)
{
  uint64x2_t roundingAddition = vdupq_n_u64(int64_t(1) << 31);
  uint64x2_t multiplied21 = vmull_u32(vget_low_u32(aValues), aDivisor);
  uint64x2_t multiplied43 = vmull_u32(vget_high_u32(aValues), aDivisor);
  return vqmovn_u32(vcombine_u32(vshrn_n_u64(vaddq_u64(multiplied21, roundingAddition), 32),
                                 vshrn_n_u64(vaddq_u64(multiplied43, roundingAddition), 32)));
}

MOZ_ALWAYS_INLINE
uint16x4_t BlurFourPixels(const uint32x4_t& aTopLeft, const uint32x4_t& aTopRight,
                          const uint32x4_t& aBottomRight, const uint32x4_t& aBottomLeft,
                          const uint32x2_t& aDivisor)
{
  uint32x4_t values = vaddq_u32(vsubq_u32(vsubq_u32(aBottomRight, aTopRight), aBottomLeft), aTopLeft);
  return Divide(values, aDivisor);
}

MOZ_ALWAYS_INLINE
void LoadIntegralRowFromRow(uint32_t *aDest, const uint8_t *aSource,
                            int32_t aSourceWidth, int32_t aLeftInflation,
                            int32_t aRightInflation)
{
  int32_t currentRowSum = 0;

  for (int x = 0; x < aLeftInflation; x++) {
    currentRowSum += aSource[0];
    aDest[x] = currentRowSum;
  }
  for (int x = aLeftInflation; x < (aSourceWidth + aLeftInflation); x++) {
    currentRowSum += aSource[(x - aLeftInflation)];
    aDest[x] = currentRowSum;
  }
  for (int x = (aSourceWidth + aLeftInflation); x < (aSourceWidth + aLeftInflation + aRightInflation); x++) {
    currentRowSum += aSource[aSourceWidth - 1];
    aDest[x] = currentRowSum;
  }
}

MOZ_ALWAYS_INLINE void
GenerateIntegralImage_NEON(int32_t aLeftInflation, int32_t aRightInflation,
                           int32_t aTopInflation, int32_t aBottomInflation,
                           uint32_t *aIntegralImage, size_t aIntegralImageStride,
                           uint8_t *aSource, int32_t aSourceStride, const IntSize &aSize)
{
  MOZ_ASSERT(!(aLeftInflation & 3));

  uint32_t stride32bit = aIntegralImageStride / 4;
  IntSize integralImageSize(aSize.width + aLeftInflation + aRightInflation,
                            aSize.height + aTopInflation + aBottomInflation);

  LoadIntegralRowFromRow(aIntegralImage, aSource, aSize.width, aLeftInflation, aRightInflation);

  for (int y = 1; y < aTopInflation + 1; y++) {
    uint32_t *intRow = aIntegralImage + (y * stride32bit);
    uint32_t *intPrevRow = aIntegralImage + (y - 1) * stride32bit;
    uint32_t *intFirstRow = aIntegralImage;

    for (int x = 0; x < integralImageSize.width; x += 4) {
      uint32x4_t firstRow = vld1q_u32(intFirstRow + x);
      uint32x4_t previousRow = vld1q_u32(intPrevRow + x);
      vst1q_u32(intRow + x, vaddq_u32(firstRow, previousRow));
    }
  }

  for (int y = aTopInflation + 1; y < (aSize.height + aTopInflation); y++) {
    uint32x4_t currentRowSum = vdupq_n_u32(0);
    uint32_t *intRow = aIntegralImage + (y * stride32bit);
    uint32_t *intPrevRow = aIntegralImage + (y - 1) * stride32bit;
    uint8_t *sourceRow = aSource + aSourceStride * (y - aTopInflation);

    uint32_t pixel = sourceRow[0];
    for (int x = 0; x < aLeftInflation; x += 4) {
      uint32_t temp[4];
      temp[0] = pixel;
      temp[1] = temp[0] + pixel;
      temp[2] = temp[1] + pixel;
      temp[3] = temp[2] + pixel;
      uint32x4_t sumPixels = vld1q_u32(temp);
      sumPixels = vaddq_u32(sumPixels, currentRowSum);
      currentRowSum = vdupq_n_u32(vgetq_lane_u32(sumPixels, 3));
      vst1q_u32(intRow + x, vaddq_u32(sumPixels, vld1q_u32(intPrevRow + x)));
    }

    for (int x = aLeftInflation; x < (aSize.width + aLeftInflation); x += 4) {
      // It's important to shuffle here. When we exit this loop currentRowSum
      // has to be set to sumPixels, so that the following loop can get the
      // correct pixel for the currentRowSum. The highest order pixel in
      // currentRowSum could've originated from accumulation in the stride.
      currentRowSum = vdupq_n_u32(vgetq_lane_u32(currentRowSum, 3));

      uint32_t temp[4];
      temp[0] = *(sourceRow + (x - aLeftInflation));
      temp[1] = temp[0] + *(sourceRow + (x - aLeftInflation) + 1);
      temp[2] = temp[1] + *(sourceRow + (x - aLeftInflation) + 2);
      temp[3] = temp[2] + *(sourceRow + (x - aLeftInflation) + 3);
      uint32x4_t sumPixels = vld1q_u32(temp);
      sumPixels = vaddq_u32(sumPixels, currentRowSum);
      currentRowSum = sumPixels;
      vst1q_u32(intRow + x, vaddq_u32(sumPixels, vld1q_u32(intPrevRow + x)));
    }

    pixel = sourceRow[aSize.width - 1];
    int x = (aSize.width + aLeftInflation);
    if ((aSize.width & 3)) {
      // Deal with unaligned portion. Get the correct pixel from currentRowSum,
      // see explanation above.
      uint32_t intCurrentRowSum = ((uint32_t*)&currentRowSum)[(aSize.width % 4) - 1];
      for (; x < integralImageSize.width; x++) {
        // We could be unaligned here!
        if (!(x & 3)) {
          // aligned!
          currentRowSum = vdupq_n_u32(intCurrentRowSum);
          break;
        }
        intCurrentRowSum += pixel;
        intRow[x] = intPrevRow[x] + intCurrentRowSum;
      }
    } else {
      currentRowSum = vdupq_n_u32(vgetq_lane_u32(currentRowSum, 3));
    }

    for (; x < integralImageSize.width; x += 4) {
      uint32_t temp[4];
      temp[0] = pixel;
      temp[1] = temp[0] + pixel;
      temp[2] = temp[1] + pixel;
      temp[3] = temp[2] + pixel;
      uint32x4_t sumPixels = vld1q_u32(temp);
      sumPixels = vaddq_u32(sumPixels, currentRowSum);
      currentRowSum = vdupq_n_u32(vgetq_lane_u32(sumPixels, 3));
      vst1q_u32(intRow + x, vaddq_u32(sumPixels, vld1q_u32(intPrevRow + x)));
    }
  }

  if (aBottomInflation) {
    // Store the last valid row of our source image in the last row of
    // our integral image. This will be overwritten with the correct values
    // in the upcoming loop.
    LoadIntegralRowFromRow(aIntegralImage + (integralImageSize.height - 1) * stride32bit,
                           aSource + (aSize.height - 1) * aSourceStride, aSize.width, aLeftInflation, aRightInflation);

    for (int y = aSize.height + aTopInflation; y < integralImageSize.height; y++) {
      uint32_t *intRow = aIntegralImage + (y * stride32bit);
      uint32_t *intPrevRow = aIntegralImage + (y - 1) * stride32bit;
      uint32_t *intLastRow = aIntegralImage + (integralImageSize.height - 1) * stride32bit;
      for (int x = 0; x < integralImageSize.width; x += 4) {
        vst1q_u32(intRow + x,
                        vaddq_u32(vld1q_u32(intLastRow + x),
                                  vld1q_u32(intPrevRow + x)));
      }
    }
  }
}

/**
 * Attempt to do an in-place box blur using an integral image.
 */
void
AlphaBoxBlur::BoxBlur_NEON(uint8_t* aData,
                           int32_t aLeftLobe,
                           int32_t aRightLobe,
                           int32_t aTopLobe,
                           int32_t aBottomLobe,
                           uint32_t *aIntegralImage,
                           size_t aIntegralImageStride)
{
  IntSize size = GetSize();

  MOZ_ASSERT(size.height > 0);

  // Our 'left' or 'top' lobe will include the current pixel. i.e. when
  // looking at an integral image the value of a pixel at 'x,y' is calculated
  // using the value of the integral image values above/below that.
  aLeftLobe++;
  aTopLobe++;
  int32_t boxSize = (aLeftLobe + aRightLobe) * (aTopLobe + aBottomLobe);

  MOZ_ASSERT(boxSize > 0);

  if (boxSize == 1) {
      return;
  }

  uint32_t reciprocal = uint32_t((uint64_t(1) << 32) / boxSize);
  uint32_t stride32bit = aIntegralImageStride / 4;
  int32_t leftInflation = RoundUpToMultipleOf4(aLeftLobe).value();

  GenerateIntegralImage_NEON(leftInflation, aRightLobe, aTopLobe, aBottomLobe,
                             aIntegralImage, aIntegralImageStride, aData,
                             mStride, size);

  uint32x2_t divisor = vdup_n_u32(reciprocal);

  // This points to the start of the rectangle within the IntegralImage that overlaps
  // the surface being blurred.
  uint32_t *innerIntegral = aIntegralImage + (aTopLobe * stride32bit) + leftInflation;
  IntRect skipRect = mSkipRect;
  int32_t stride = mStride;
  uint8_t *data = aData;

  for (int32_t y = 0; y < size.height; y++) {
    bool inSkipRectY = y > skipRect.y && y < skipRect.YMost();
    uint32_t *topLeftBase = innerIntegral + ((y - aTopLobe) * ptrdiff_t(stride32bit) - aLeftLobe);
    uint32_t *topRightBase = innerIntegral + ((y - aTopLobe) * ptrdiff_t(stride32bit) + aRightLobe);
    uint32_t *bottomRightBase = innerIntegral + ((y + aBottomLobe) * ptrdiff_t(stride32bit) + aRightLobe);
    uint32_t *bottomLeftBase = innerIntegral + ((y + aBottomLobe) * ptrdiff_t(stride32bit) - aLeftLobe);

    int32_t x = 0;
    // Process 16 pixels at a time for as long as possible.
    for (; x <= size.width - 16; x += 16) {
      if (inSkipRectY && x > skipRect.x && x < skipRect.XMost()) {
        x = skipRect.XMost() - 16;
        // Trigger early jump on coming loop iterations, this will be reset
        // next line anyway.
        inSkipRectY = false;
        continue;
      }

      uint32x4_t topLeft;
      uint32x4_t topRight;
      uint32x4_t bottomRight;
      uint32x4_t bottomLeft;
      topLeft = vld1q_u32(topLeftBase + x);
      topRight = vld1q_u32(topRightBase + x);
      bottomRight = vld1q_u32(bottomRightBase + x);
      bottomLeft = vld1q_u32(bottomLeftBase + x);
      uint16x4_t result1 = BlurFourPixels(topLeft, topRight, bottomRight, bottomLeft, divisor);

      topLeft = vld1q_u32(topLeftBase + x + 4);
      topRight = vld1q_u32(topRightBase + x + 4);
      bottomRight = vld1q_u32(bottomRightBase + x + 4);
      bottomLeft = vld1q_u32(bottomLeftBase + x + 4);
      uint16x4_t result2 = BlurFourPixels(topLeft, topRight, bottomRight, bottomLeft, divisor);

      topLeft = vld1q_u32(topLeftBase + x + 8);
      topRight = vld1q_u32(topRightBase + x + 8);
      bottomRight = vld1q_u32(bottomRightBase + x + 8);
      bottomLeft = vld1q_u32(bottomLeftBase + x + 8);
      uint16x4_t result3 = BlurFourPixels(topLeft, topRight, bottomRight, bottomLeft, divisor);

      topLeft = vld1q_u32(topLeftBase + x + 12);
      topRight = vld1q_u32(topRightBase + x + 12);
      bottomRight = vld1q_u32(bottomRightBase + x + 12);
      bottomLeft = vld1q_u32(bottomLeftBase + x + 12);
      uint16x4_t result4 = BlurFourPixels(topLeft, topRight, bottomRight, bottomLeft, divisor);

      uint8x8_t combine1 = vqmovn_u16(vcombine_u16(result1, result2));
      uint8x8_t combine2 = vqmovn_u16(vcombine_u16(result3, result4));
      uint8x16_t final = vcombine_u8(combine1, combine2);
      vst1q_u8(data + stride * y + x, final);
    }

    // Process the remaining pixels 4 bytes at a time.
    for (; x < size.width; x += 4) {
      if (inSkipRectY && x > skipRect.x && x < skipRect.XMost()) {
        x = skipRect.XMost() - 4;
        // Trigger early jump on coming loop iterations, this will be reset
        // next line anyway.
        inSkipRectY = false;
        continue;
      }

      uint32x4_t topLeft = vld1q_u32(topLeftBase + x);
      uint32x4_t topRight = vld1q_u32(topRightBase + x);
      uint32x4_t bottomRight = vld1q_u32(bottomRightBase + x);
      uint32x4_t bottomLeft = vld1q_u32(bottomLeftBase + x);
      uint16x4_t result = BlurFourPixels(topLeft, topRight, bottomRight, bottomLeft, divisor);
      uint32x2_t final = vreinterpret_u32_u8(vmovn_u16(vcombine_u16(result, vdup_n_u16(0))));
      *(uint32_t*)(data + stride * y + x) = vget_lane_u32(final, 0);
    }
  }
}

}
}