• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 
2 /*
3  * Copyright 2012 Google Inc.
4  *
5  * Use of this source code is governed by a BSD-style license that can be
6  * found in the LICENSE file.
7  */
8 #include "SkBitmapProcState.h"
9 #include "SkBitmapProcState_filter.h"
10 #include "SkColorPriv.h"
11 #include "SkFilterProc.h"
12 #include "SkPaint.h"
13 #include "SkShader.h"   // for tilemodes
14 #include "SkUtilsArm.h"
15 
16 // Required to ensure the table is part of the final binary.
17 extern const SkBitmapProcState::SampleProc32 gSkBitmapProcStateSample32_neon[];
18 extern const SkBitmapProcState::SampleProc16 gSkBitmapProcStateSample16_neon[];
19 
20 #define   NAME_WRAP(x)  x ## _neon
21 #include "SkBitmapProcState_filter_neon.h"
22 #include "SkBitmapProcState_procs.h"
23 
24 const SkBitmapProcState::SampleProc32 gSkBitmapProcStateSample32_neon[] = {
25     S32_opaque_D32_nofilter_DXDY_neon,
26     S32_alpha_D32_nofilter_DXDY_neon,
27     S32_opaque_D32_nofilter_DX_neon,
28     S32_alpha_D32_nofilter_DX_neon,
29     S32_opaque_D32_filter_DXDY_neon,
30     S32_alpha_D32_filter_DXDY_neon,
31     S32_opaque_D32_filter_DX_neon,
32     S32_alpha_D32_filter_DX_neon,
33 
34     S16_opaque_D32_nofilter_DXDY_neon,
35     S16_alpha_D32_nofilter_DXDY_neon,
36     S16_opaque_D32_nofilter_DX_neon,
37     S16_alpha_D32_nofilter_DX_neon,
38     S16_opaque_D32_filter_DXDY_neon,
39     S16_alpha_D32_filter_DXDY_neon,
40     S16_opaque_D32_filter_DX_neon,
41     S16_alpha_D32_filter_DX_neon,
42 
43     SI8_opaque_D32_nofilter_DXDY_neon,
44     SI8_alpha_D32_nofilter_DXDY_neon,
45     SI8_opaque_D32_nofilter_DX_neon,
46     SI8_alpha_D32_nofilter_DX_neon,
47     SI8_opaque_D32_filter_DXDY_neon,
48     SI8_alpha_D32_filter_DXDY_neon,
49     SI8_opaque_D32_filter_DX_neon,
50     SI8_alpha_D32_filter_DX_neon,
51 
52     S4444_opaque_D32_nofilter_DXDY_neon,
53     S4444_alpha_D32_nofilter_DXDY_neon,
54     S4444_opaque_D32_nofilter_DX_neon,
55     S4444_alpha_D32_nofilter_DX_neon,
56     S4444_opaque_D32_filter_DXDY_neon,
57     S4444_alpha_D32_filter_DXDY_neon,
58     S4444_opaque_D32_filter_DX_neon,
59     S4444_alpha_D32_filter_DX_neon,
60 
61     // A8 treats alpha/opauqe the same (equally efficient)
62     SA8_alpha_D32_nofilter_DXDY_neon,
63     SA8_alpha_D32_nofilter_DXDY_neon,
64     SA8_alpha_D32_nofilter_DX_neon,
65     SA8_alpha_D32_nofilter_DX_neon,
66     SA8_alpha_D32_filter_DXDY_neon,
67     SA8_alpha_D32_filter_DXDY_neon,
68     SA8_alpha_D32_filter_DX_neon,
69     SA8_alpha_D32_filter_DX_neon,
70 
71     // todo: possibly specialize on opaqueness
72     SG8_alpha_D32_nofilter_DXDY_neon,
73     SG8_alpha_D32_nofilter_DXDY_neon,
74     SG8_alpha_D32_nofilter_DX_neon,
75     SG8_alpha_D32_nofilter_DX_neon,
76     SG8_alpha_D32_filter_DXDY_neon,
77     SG8_alpha_D32_filter_DXDY_neon,
78     SG8_alpha_D32_filter_DX_neon,
79     SG8_alpha_D32_filter_DX_neon,
80 };
81 
82 const SkBitmapProcState::SampleProc16 gSkBitmapProcStateSample16_neon[] = {
83     S32_D16_nofilter_DXDY_neon,
84     S32_D16_nofilter_DX_neon,
85     S32_D16_filter_DXDY_neon,
86     S32_D16_filter_DX_neon,
87 
88     S16_D16_nofilter_DXDY_neon,
89     S16_D16_nofilter_DX_neon,
90     S16_D16_filter_DXDY_neon,
91     S16_D16_filter_DX_neon,
92 
93     SI8_D16_nofilter_DXDY_neon,
94     SI8_D16_nofilter_DX_neon,
95     SI8_D16_filter_DXDY_neon,
96     SI8_D16_filter_DX_neon,
97 
98     // Don't support 4444 -> 565
99     NULL, NULL, NULL, NULL,
100     // Don't support A8 -> 565
101     NULL, NULL, NULL, NULL,
102     // Don't support G8 -> 565 (but we could)
103     NULL, NULL, NULL, NULL,
104 };
105 
106 ///////////////////////////////////////////////////////////////////////////////
107 
108 #include <arm_neon.h>
109 #include "SkConvolver.h"
110 
111 // Convolves horizontally along a single row. The row data is given in
112 // |srcData| and continues for the numValues() of the filter.
convolveHorizontally_neon(const unsigned char * srcData,const SkConvolutionFilter1D & filter,unsigned char * outRow,bool hasAlpha)113 void convolveHorizontally_neon(const unsigned char* srcData,
114                                const SkConvolutionFilter1D& filter,
115                                unsigned char* outRow,
116                                bool hasAlpha) {
117     // Loop over each pixel on this row in the output image.
118     int numValues = filter.numValues();
119     for (int outX = 0; outX < numValues; outX++) {
120         uint8x8_t coeff_mask0 = vcreate_u8(0x0100010001000100);
121         uint8x8_t coeff_mask1 = vcreate_u8(0x0302030203020302);
122         uint8x8_t coeff_mask2 = vcreate_u8(0x0504050405040504);
123         uint8x8_t coeff_mask3 = vcreate_u8(0x0706070607060706);
124         // Get the filter that determines the current output pixel.
125         int filterOffset, filterLength;
126         const SkConvolutionFilter1D::ConvolutionFixed* filterValues =
127             filter.FilterForValue(outX, &filterOffset, &filterLength);
128 
129         // Compute the first pixel in this row that the filter affects. It will
130         // touch |filterLength| pixels (4 bytes each) after this.
131         const unsigned char* rowToFilter = &srcData[filterOffset * 4];
132 
133         // Apply the filter to the row to get the destination pixel in |accum|.
134         int32x4_t accum = vdupq_n_s32(0);
135         for (int filterX = 0; filterX < filterLength >> 2; filterX++) {
136             // Load 4 coefficients
137             int16x4_t coeffs, coeff0, coeff1, coeff2, coeff3;
138             coeffs = vld1_s16(filterValues);
139             coeff0 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask0));
140             coeff1 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask1));
141             coeff2 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask2));
142             coeff3 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask3));
143 
144             // Load pixels and calc
145             uint8x16_t pixels = vld1q_u8(rowToFilter);
146             int16x8_t p01_16 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pixels)));
147             int16x8_t p23_16 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(pixels)));
148 
149             int16x4_t p0_src = vget_low_s16(p01_16);
150             int16x4_t p1_src = vget_high_s16(p01_16);
151             int16x4_t p2_src = vget_low_s16(p23_16);
152             int16x4_t p3_src = vget_high_s16(p23_16);
153 
154             int32x4_t p0 = vmull_s16(p0_src, coeff0);
155             int32x4_t p1 = vmull_s16(p1_src, coeff1);
156             int32x4_t p2 = vmull_s16(p2_src, coeff2);
157             int32x4_t p3 = vmull_s16(p3_src, coeff3);
158 
159             accum += p0;
160             accum += p1;
161             accum += p2;
162             accum += p3;
163 
164             // Advance the pointers
165             rowToFilter += 16;
166             filterValues += 4;
167         }
168         int r = filterLength & 3;
169         if (r) {
170             const uint16_t mask[4][4] = {
171                 {0, 0, 0, 0},
172                 {0xFFFF, 0, 0, 0},
173                 {0xFFFF, 0xFFFF, 0, 0},
174                 {0xFFFF, 0xFFFF, 0xFFFF, 0}
175             };
176             uint16x4_t coeffs;
177             int16x4_t coeff0, coeff1, coeff2;
178             coeffs = vld1_u16(reinterpret_cast<const uint16_t*>(filterValues));
179             coeffs &= vld1_u16(&mask[r][0]);
180             coeff0 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_u16(coeffs), coeff_mask0));
181             coeff1 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_u16(coeffs), coeff_mask1));
182             coeff2 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_u16(coeffs), coeff_mask2));
183 
184             // Load pixels and calc
185             uint8x16_t pixels = vld1q_u8(rowToFilter);
186             int16x8_t p01_16 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pixels)));
187             int16x8_t p23_16 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(pixels)));
188             int32x4_t p0 = vmull_s16(vget_low_s16(p01_16), coeff0);
189             int32x4_t p1 = vmull_s16(vget_high_s16(p01_16), coeff1);
190             int32x4_t p2 = vmull_s16(vget_low_s16(p23_16), coeff2);
191 
192             accum += p0;
193             accum += p1;
194             accum += p2;
195         }
196 
197         // Bring this value back in range. All of the filter scaling factors
198         // are in fixed point with kShiftBits bits of fractional part.
199         accum = vshrq_n_s32(accum, SkConvolutionFilter1D::kShiftBits);
200 
201         // Pack and store the new pixel.
202         int16x4_t accum16 = vqmovn_s32(accum);
203         uint8x8_t accum8 = vqmovun_s16(vcombine_s16(accum16, accum16));
204         vst1_lane_u32(reinterpret_cast<uint32_t*>(outRow), vreinterpret_u32_u8(accum8), 0);
205         outRow += 4;
206     }
207 }
208 
209 // Does vertical convolution to produce one output row. The filter values and
210 // length are given in the first two parameters. These are applied to each
211 // of the rows pointed to in the |sourceDataRows| array, with each row
212 // being |pixelWidth| wide.
213 //
214 // The output must have room for |pixelWidth * 4| bytes.
215 template<bool hasAlpha>
convolveVertically_neon(const SkConvolutionFilter1D::ConvolutionFixed * filterValues,int filterLength,unsigned char * const * sourceDataRows,int pixelWidth,unsigned char * outRow)216 void convolveVertically_neon(const SkConvolutionFilter1D::ConvolutionFixed* filterValues,
217                              int filterLength,
218                              unsigned char* const* sourceDataRows,
219                              int pixelWidth,
220                              unsigned char* outRow) {
221     int width = pixelWidth & ~3;
222 
223     int32x4_t accum0, accum1, accum2, accum3;
224     int16x4_t coeff16;
225 
226     // Output four pixels per iteration (16 bytes).
227     for (int outX = 0; outX < width; outX += 4) {
228 
229         // Accumulated result for each pixel. 32 bits per RGBA channel.
230         accum0 = accum1 = accum2 = accum3 = vdupq_n_s32(0);
231 
232         // Convolve with one filter coefficient per iteration.
233         for (int filterY = 0; filterY < filterLength; filterY++) {
234 
235             // Duplicate the filter coefficient 4 times.
236             // [16] cj cj cj cj
237             coeff16 = vdup_n_s16(filterValues[filterY]);
238 
239             // Load four pixels (16 bytes) together.
240             // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
241             uint8x16_t src8 = vld1q_u8(&sourceDataRows[filterY][outX << 2]);
242 
243             int16x8_t src16_01 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(src8)));
244             int16x8_t src16_23 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(src8)));
245             int16x4_t src16_0 = vget_low_s16(src16_01);
246             int16x4_t src16_1 = vget_high_s16(src16_01);
247             int16x4_t src16_2 = vget_low_s16(src16_23);
248             int16x4_t src16_3 = vget_high_s16(src16_23);
249 
250             accum0 += vmull_s16(src16_0, coeff16);
251             accum1 += vmull_s16(src16_1, coeff16);
252             accum2 += vmull_s16(src16_2, coeff16);
253             accum3 += vmull_s16(src16_3, coeff16);
254         }
255 
256         // Shift right for fixed point implementation.
257         accum0 = vshrq_n_s32(accum0, SkConvolutionFilter1D::kShiftBits);
258         accum1 = vshrq_n_s32(accum1, SkConvolutionFilter1D::kShiftBits);
259         accum2 = vshrq_n_s32(accum2, SkConvolutionFilter1D::kShiftBits);
260         accum3 = vshrq_n_s32(accum3, SkConvolutionFilter1D::kShiftBits);
261 
262         // Packing 32 bits |accum| to 16 bits per channel (signed saturation).
263         // [16] a1 b1 g1 r1 a0 b0 g0 r0
264         int16x8_t accum16_0 = vcombine_s16(vqmovn_s32(accum0), vqmovn_s32(accum1));
265         // [16] a3 b3 g3 r3 a2 b2 g2 r2
266         int16x8_t accum16_1 = vcombine_s16(vqmovn_s32(accum2), vqmovn_s32(accum3));
267 
268         // Packing 16 bits |accum| to 8 bits per channel (unsigned saturation).
269         // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
270         uint8x16_t accum8 = vcombine_u8(vqmovun_s16(accum16_0), vqmovun_s16(accum16_1));
271 
272         if (hasAlpha) {
273             // Compute the max(ri, gi, bi) for each pixel.
274             // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0
275             uint8x16_t a = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(accum8), 8));
276             // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
277             uint8x16_t b = vmaxq_u8(a, accum8); // Max of r and g
278             // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0
279             a = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(accum8), 16));
280             // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
281             b = vmaxq_u8(a, b); // Max of r and g and b.
282             // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00
283             b = vreinterpretq_u8_u32(vshlq_n_u32(vreinterpretq_u32_u8(b), 24));
284 
285             // Make sure the value of alpha channel is always larger than maximum
286             // value of color channels.
287             accum8 = vmaxq_u8(b, accum8);
288         } else {
289             // Set value of alpha channels to 0xFF.
290             accum8 = vreinterpretq_u8_u32(vreinterpretq_u32_u8(accum8) | vdupq_n_u32(0xFF000000));
291         }
292 
293         // Store the convolution result (16 bytes) and advance the pixel pointers.
294         vst1q_u8(outRow, accum8);
295         outRow += 16;
296     }
297 
298     // Process the leftovers when the width of the output is not divisible
299     // by 4, that is at most 3 pixels.
300     int r = pixelWidth & 3;
301     if (r) {
302 
303         accum0 = accum1 = accum2 = vdupq_n_s32(0);
304 
305         for (int filterY = 0; filterY < filterLength; ++filterY) {
306             coeff16 = vdup_n_s16(filterValues[filterY]);
307 
308             // [8] a3 b3 g3 r3 a2 b2 g2 r2 a1 b1 g1 r1 a0 b0 g0 r0
309             uint8x16_t src8 = vld1q_u8(&sourceDataRows[filterY][width << 2]);
310 
311             int16x8_t src16_01 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(src8)));
312             int16x8_t src16_23 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(src8)));
313             int16x4_t src16_0 = vget_low_s16(src16_01);
314             int16x4_t src16_1 = vget_high_s16(src16_01);
315             int16x4_t src16_2 = vget_low_s16(src16_23);
316 
317             accum0 += vmull_s16(src16_0, coeff16);
318             accum1 += vmull_s16(src16_1, coeff16);
319             accum2 += vmull_s16(src16_2, coeff16);
320         }
321 
322         accum0 = vshrq_n_s32(accum0, SkConvolutionFilter1D::kShiftBits);
323         accum1 = vshrq_n_s32(accum1, SkConvolutionFilter1D::kShiftBits);
324         accum2 = vshrq_n_s32(accum2, SkConvolutionFilter1D::kShiftBits);
325 
326         int16x8_t accum16_0 = vcombine_s16(vqmovn_s32(accum0), vqmovn_s32(accum1));
327         int16x8_t accum16_1 = vcombine_s16(vqmovn_s32(accum2), vqmovn_s32(accum2));
328 
329         uint8x16_t accum8 = vcombine_u8(vqmovun_s16(accum16_0), vqmovun_s16(accum16_1));
330 
331         if (hasAlpha) {
332             // Compute the max(ri, gi, bi) for each pixel.
333             // [8] xx a3 b3 g3 xx a2 b2 g2 xx a1 b1 g1 xx a0 b0 g0
334             uint8x16_t a = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(accum8), 8));
335             // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
336             uint8x16_t b = vmaxq_u8(a, accum8); // Max of r and g
337             // [8] xx xx a3 b3 xx xx a2 b2 xx xx a1 b1 xx xx a0 b0
338             a = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(accum8), 16));
339             // [8] xx xx xx max3 xx xx xx max2 xx xx xx max1 xx xx xx max0
340             b = vmaxq_u8(a, b); // Max of r and g and b.
341             // [8] max3 00 00 00 max2 00 00 00 max1 00 00 00 max0 00 00 00
342             b = vreinterpretq_u8_u32(vshlq_n_u32(vreinterpretq_u32_u8(b), 24));
343 
344             // Make sure the value of alpha channel is always larger than maximum
345             // value of color channels.
346             accum8 = vmaxq_u8(b, accum8);
347         } else {
348             // Set value of alpha channels to 0xFF.
349             accum8 = vreinterpretq_u8_u32(vreinterpretq_u32_u8(accum8) | vdupq_n_u32(0xFF000000));
350         }
351 
352         switch(r) {
353         case 1:
354             vst1q_lane_u32(reinterpret_cast<uint32_t*>(outRow), vreinterpretq_u32_u8(accum8), 0);
355             break;
356         case 2:
357             vst1_u32(reinterpret_cast<uint32_t*>(outRow),
358                      vreinterpret_u32_u8(vget_low_u8(accum8)));
359             break;
360         case 3:
361             vst1_u32(reinterpret_cast<uint32_t*>(outRow),
362                      vreinterpret_u32_u8(vget_low_u8(accum8)));
363             vst1q_lane_u32(reinterpret_cast<uint32_t*>(outRow+8), vreinterpretq_u32_u8(accum8), 2);
364             break;
365         }
366     }
367 }
368 
convolveVertically_neon(const SkConvolutionFilter1D::ConvolutionFixed * filterValues,int filterLength,unsigned char * const * sourceDataRows,int pixelWidth,unsigned char * outRow,bool sourceHasAlpha)369 void convolveVertically_neon(const SkConvolutionFilter1D::ConvolutionFixed* filterValues,
370                              int filterLength,
371                              unsigned char* const* sourceDataRows,
372                              int pixelWidth,
373                              unsigned char* outRow,
374                              bool sourceHasAlpha) {
375     if (sourceHasAlpha) {
376         convolveVertically_neon<true>(filterValues, filterLength,
377                                       sourceDataRows, pixelWidth,
378                                       outRow);
379     } else {
380         convolveVertically_neon<false>(filterValues, filterLength,
381                                        sourceDataRows, pixelWidth,
382                                        outRow);
383     }
384 }
385 
386 // Convolves horizontally along four rows. The row data is given in
387 // |src_data| and continues for the num_values() of the filter.
388 // The algorithm is almost same as |ConvolveHorizontally_SSE2|. Please
389 // refer to that function for detailed comments.
convolve4RowsHorizontally_neon(const unsigned char * srcData[4],const SkConvolutionFilter1D & filter,unsigned char * outRow[4])390 void convolve4RowsHorizontally_neon(const unsigned char* srcData[4],
391                                     const SkConvolutionFilter1D& filter,
392                                     unsigned char* outRow[4]) {
393 
394     uint8x8_t coeff_mask0 = vcreate_u8(0x0100010001000100);
395     uint8x8_t coeff_mask1 = vcreate_u8(0x0302030203020302);
396     uint8x8_t coeff_mask2 = vcreate_u8(0x0504050405040504);
397     uint8x8_t coeff_mask3 = vcreate_u8(0x0706070607060706);
398     int num_values = filter.numValues();
399 
400     int filterOffset, filterLength;
401     // |mask| will be used to decimate all extra filter coefficients that are
402     // loaded by SIMD when |filter_length| is not divisible by 4.
403     // mask[0] is not used in following algorithm.
404     const uint16_t mask[4][4] = {
405         {0, 0, 0, 0},
406         {0xFFFF, 0, 0, 0},
407         {0xFFFF, 0xFFFF, 0, 0},
408         {0xFFFF, 0xFFFF, 0xFFFF, 0}
409     };
410 
411     // Output one pixel each iteration, calculating all channels (RGBA) together.
412     for (int outX = 0; outX < num_values; outX++) {
413 
414         const SkConvolutionFilter1D::ConvolutionFixed* filterValues =
415         filter.FilterForValue(outX, &filterOffset, &filterLength);
416 
417         // four pixels in a column per iteration.
418         int32x4_t accum0 = vdupq_n_s32(0);
419         int32x4_t accum1 = vdupq_n_s32(0);
420         int32x4_t accum2 = vdupq_n_s32(0);
421         int32x4_t accum3 = vdupq_n_s32(0);
422 
423         int start = (filterOffset<<2);
424 
425         // We will load and accumulate with four coefficients per iteration.
426         for (int filter_x = 0; filter_x < (filterLength >> 2); filter_x++) {
427             int16x4_t coeffs, coeff0, coeff1, coeff2, coeff3;
428 
429             coeffs = vld1_s16(filterValues);
430             coeff0 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask0));
431             coeff1 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask1));
432             coeff2 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask2));
433             coeff3 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask3));
434 
435             uint8x16_t pixels;
436             int16x8_t p01_16, p23_16;
437             int32x4_t p0, p1, p2, p3;
438 
439 
440 #define ITERATION(src, accum)                                       \
441     pixels = vld1q_u8(src);                                         \
442     p01_16 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(pixels)));  \
443     p23_16 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(pixels))); \
444     p0 = vmull_s16(vget_low_s16(p01_16), coeff0);                   \
445     p1 = vmull_s16(vget_high_s16(p01_16), coeff1);                  \
446     p2 = vmull_s16(vget_low_s16(p23_16), coeff2);                   \
447     p3 = vmull_s16(vget_high_s16(p23_16), coeff3);                  \
448     accum += p0;                                                    \
449     accum += p1;                                                    \
450     accum += p2;                                                    \
451     accum += p3
452 
453             ITERATION(srcData[0] + start, accum0);
454             ITERATION(srcData[1] + start, accum1);
455             ITERATION(srcData[2] + start, accum2);
456             ITERATION(srcData[3] + start, accum3);
457 
458             start += 16;
459             filterValues += 4;
460         }
461 
462         int r = filterLength & 3;
463         if (r) {
464             int16x4_t coeffs, coeff0, coeff1, coeff2, coeff3;
465             coeffs = vld1_s16(filterValues);
466             coeffs &= vreinterpret_s16_u16(vld1_u16(&mask[r][0]));
467             coeff0 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask0));
468             coeff1 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask1));
469             coeff2 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask2));
470             coeff3 = vreinterpret_s16_u8(vtbl1_u8(vreinterpret_u8_s16(coeffs), coeff_mask3));
471 
472             uint8x16_t pixels;
473             int16x8_t p01_16, p23_16;
474             int32x4_t p0, p1, p2, p3;
475 
476             ITERATION(srcData[0] + start, accum0);
477             ITERATION(srcData[1] + start, accum1);
478             ITERATION(srcData[2] + start, accum2);
479             ITERATION(srcData[3] + start, accum3);
480         }
481 
482         int16x4_t accum16;
483         uint8x8_t res0, res1, res2, res3;
484 
485 #define PACK_RESULT(accum, res)                                         \
486         accum = vshrq_n_s32(accum, SkConvolutionFilter1D::kShiftBits);  \
487         accum16 = vqmovn_s32(accum);                                    \
488         res = vqmovun_s16(vcombine_s16(accum16, accum16));
489 
490         PACK_RESULT(accum0, res0);
491         PACK_RESULT(accum1, res1);
492         PACK_RESULT(accum2, res2);
493         PACK_RESULT(accum3, res3);
494 
495         vst1_lane_u32(reinterpret_cast<uint32_t*>(outRow[0]), vreinterpret_u32_u8(res0), 0);
496         vst1_lane_u32(reinterpret_cast<uint32_t*>(outRow[1]), vreinterpret_u32_u8(res1), 0);
497         vst1_lane_u32(reinterpret_cast<uint32_t*>(outRow[2]), vreinterpret_u32_u8(res2), 0);
498         vst1_lane_u32(reinterpret_cast<uint32_t*>(outRow[3]), vreinterpret_u32_u8(res3), 0);
499         outRow[0] += 4;
500         outRow[1] += 4;
501         outRow[2] += 4;
502         outRow[3] += 4;
503     }
504 }
505 
applySIMDPadding_neon(SkConvolutionFilter1D * filter)506 void applySIMDPadding_neon(SkConvolutionFilter1D *filter) {
507     // Padding |paddingCount| of more dummy coefficients after the coefficients
508     // of last filter to prevent SIMD instructions which load 8 or 16 bytes
509     // together to access invalid memory areas. We are not trying to align the
510     // coefficients right now due to the opaqueness of <vector> implementation.
511     // This has to be done after all |AddFilter| calls.
512     for (int i = 0; i < 8; ++i) {
513         filter->addFilterValue(static_cast<SkConvolutionFilter1D::ConvolutionFixed>(0));
514     }
515 }
516 
platformConvolutionProcs_arm_neon(SkConvolutionProcs * procs)517 void platformConvolutionProcs_arm_neon(SkConvolutionProcs* procs) {
518     procs->fExtraHorizontalReads = 3;
519     procs->fConvolveVertically = &convolveVertically_neon;
520     procs->fConvolve4RowsHorizontally = &convolve4RowsHorizontally_neon;
521     procs->fConvolveHorizontally = &convolveHorizontally_neon;
522     procs->fApplySIMDPadding = &applySIMDPadding_neon;
523 }
524