1 /*
2  * Copyright 2018 Google Inc.
3  *
4  * Use of this source code is governed by a BSD-style license that can be
5  * found in the LICENSE file.
6  */
7 
8 #ifndef SkBitmapProcState_opts_DEFINED
9 #define SkBitmapProcState_opts_DEFINED
10 
11 #include "SkBitmapProcState.h"
12 
13 // SkBitmapProcState optimized Shader, Sample, or Matrix procs.
14 //
15 // Only S32_alpha_D32_filter_DX exploits instructions beyond
16 // our common baseline SSE2/NEON instruction sets, so that's
17 // all that lives here.
18 //
19 // The rest are scattershot at the moment but I want to get them
20 // all migrated to be normal code inside SkBitmapProcState.cpp.
21 
22 #if SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
23     #include <immintrin.h>
24 #elif defined(SK_ARM_HAS_NEON)
25     #include <arm_neon.h>
26 #endif
27 
28 namespace SK_OPTS_NS {
29 
30 // This same basic packing scheme is used throughout the file.
31 static void decode_packed_coordinates_and_weight(uint32_t packed, int* v0, int* v1, int* w) {
32     // The top 14 bits are the integer coordinate x0 or y0.
33     *v0 = packed >> 18;
34 
35     // The bottom 14 bits are the integer coordinate x1 or y1.
36     *v1 = packed & 0x3fff;
37 
38     // The middle 4 bits are the interpolating factor between the two, i.e. the weight for v1.
39     *w = (packed >> 14) & 0xf;
40 }
41 
42 #if 1 && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSSE3
43 
44     // As above, 4x.
45     static void decode_packed_coordinates_and_weight(__m128i packed,
46                                                      int v0[4], int v1[4], __m128i* w) {
47         _mm_storeu_si128((__m128i*)v0, _mm_srli_epi32(packed, 18));
48         _mm_storeu_si128((__m128i*)v1, _mm_and_si128 (packed, _mm_set1_epi32(0x3fff)));
49         *w = _mm_and_si128(_mm_srli_epi32(packed, 14), _mm_set1_epi32(0xf));
50     }
51 
52     // This is the crux of the SSSE3 implementation,
53     // interpolating in X for up to two output pixels (A and B) using _mm_maddubs_epi16().
54     static inline __m128i interpolate_in_x(uint32_t A0, uint32_t A1,
55                                            uint32_t B0, uint32_t B1,
56                                            const __m128i& interlaced_x_weights) {
57         // _mm_maddubs_epi16() is a little idiosyncratic, but very helpful as the core of a lerp.
58         //
59         // It takes two arguments interlaced byte-wise:
60         //    - first  arg: [ x,y, ... 7 more pairs of 8-bit values ...]
61         //    - second arg: [ z,w, ... 7 more pairs of 8-bit values ...]
62         // and returns 8 16-bit values: [ x*z + y*w, ... 7 more 16-bit values ... ].
63         //
64         // That's why we go to all this trouble to make interlaced_x_weights,
65         // and here we're interlacing A0 with A1, B0 with B1 to match.
66 
67         __m128i interlaced_A = _mm_unpacklo_epi8(_mm_cvtsi32_si128(A0), _mm_cvtsi32_si128(A1)),
68                 interlaced_B = _mm_unpacklo_epi8(_mm_cvtsi32_si128(B0), _mm_cvtsi32_si128(B1));
69 
70         return _mm_maddubs_epi16(_mm_unpacklo_epi64(interlaced_A, interlaced_B),
71                                  interlaced_x_weights);
72     }
73 
74     // Interpolate {A0..A3} --> output pixel A, and {B0..B3} --> output pixel B.
75     // Returns two pixels, with each channel in a 16-bit lane of the __m128i.
76     static inline __m128i interpolate_in_x_and_y(uint32_t A0, uint32_t A1,
77                                                  uint32_t A2, uint32_t A3,
78                                                  uint32_t B0, uint32_t B1,
79                                                  uint32_t B2, uint32_t B3,
80                                                  const __m128i& interlaced_x_weights,
81                                                  int wy) {
82         // The stored Y weight wy is for y1, and y0 gets a weight 16-wy.
83         const __m128i wy1 = _mm_set1_epi16(wy),
84                       wy0 = _mm_sub_epi16(_mm_set1_epi16(16), wy1);
85 
86         // First interpolate in X,
87         // leaving the values in 16-bit lanes scaled up by those [0,16] interlaced_x_weights.
88         __m128i row0 = interpolate_in_x(A0,A1, B0,B1, interlaced_x_weights),
89                 row1 = interpolate_in_x(A2,A3, B2,B3, interlaced_x_weights);
90 
91         // Interpolate in Y across the two rows,
92         // then scale everything down by the maximum total weight 16x16 = 256.
93         return _mm_srli_epi16(_mm_add_epi16(_mm_mullo_epi16(row0, wy0),
94                                             _mm_mullo_epi16(row1, wy1)), 8);
95     }
96 
97     /*not static*/ inline
98     void S32_alpha_D32_filter_DX(const SkBitmapProcState& s,
99                                  const uint32_t* xy, int count, uint32_t* colors) {
100         SkASSERT(count > 0 && colors != nullptr);
101         SkASSERT(s.fFilterQuality != kNone_SkFilterQuality);
102         SkASSERT(kN32_SkColorType == s.fPixmap.colorType());
103 
104         int alpha = s.fAlphaScale;
105 
106         // Return (px * s.fAlphaScale) / 256.   (s.fAlphaScale is in [0,256].)
107         auto scale_by_alpha = [alpha](const __m128i& px) {
108             return alpha == 256 ? px
109                                 : _mm_srli_epi16(_mm_mullo_epi16(px, _mm_set1_epi16(alpha)), 8);
110         };
111 
112         // We're in _DX_ mode here, so we're only varying in X.
113         // That means the first entry of xy is our constant pair of Y coordinates and weight in Y.
114         // All the other entries in xy will be pairs of X coordinates and the X weight.
115         int y0, y1, wy;
116         decode_packed_coordinates_and_weight(*xy++, &y0, &y1, &wy);
117 
118         auto row0 = (const uint32_t*)((const uint8_t*)s.fPixmap.addr() + y0 * s.fPixmap.rowBytes()),
119              row1 = (const uint32_t*)((const uint8_t*)s.fPixmap.addr() + y1 * s.fPixmap.rowBytes());
120 
121         while (count >= 4) {
122             // We can really get going, loading 4 X pairs at a time to produce 4 output pixels.
123             const __m128i xx = _mm_loadu_si128((const __m128i*)xy);
124 
125             int x0[4],
126                 x1[4];
127             __m128i wx;
128             decode_packed_coordinates_and_weight(xx, x0, x1, &wx);
129 
130             // Splat out each x weight wx four times (one for each pixel channel) as wx1,
131             // and sixteen minus that as the weight for x0, wx0.
132             __m128i wx1 = _mm_shuffle_epi8(wx, _mm_setr_epi8(0,0,0,0,4,4,4,4,8,8,8,8,12,12,12,12)),
133                     wx0 = _mm_sub_epi8(_mm_set1_epi8(16), wx1);
134 
135             // We need to interlace wx0 and wx1 for _mm_maddubs_epi16().
136             __m128i interlaced_x_weights_AB = _mm_unpacklo_epi8(wx0,wx1),
137                     interlaced_x_weights_CD = _mm_unpackhi_epi8(wx0,wx1);
138 
139             // interpolate_in_x_and_y() can produce two output pixels (A and B) at a time
140             // from eight input pixels {A0..A3} and {B0..B3}, arranged in a 2x2 grid for each.
141             __m128i AB = interpolate_in_x_and_y(row0[x0[0]], row0[x1[0]],
142                                                 row1[x0[0]], row1[x1[0]],
143                                                 row0[x0[1]], row0[x1[1]],
144                                                 row1[x0[1]], row1[x1[1]],
145                                                 interlaced_x_weights_AB, wy);
146 
147             // Once more with the other half of the x-weights for two more pixels C,D.
148             __m128i CD = interpolate_in_x_and_y(row0[x0[2]], row0[x1[2]],
149                                                 row1[x0[2]], row1[x1[2]],
150                                                 row0[x0[3]], row0[x1[3]],
151                                                 row1[x0[3]], row1[x1[3]],
152                                                 interlaced_x_weights_CD, wy);
153 
154             // Scale by alpha, pack back together to 8-bit lanes, and write out four pixels!
155             _mm_storeu_si128((__m128i*)colors, _mm_packus_epi16(scale_by_alpha(AB),
156                                                                 scale_by_alpha(CD)));
157             xy     += 4;
158             colors += 4;
159             count  -= 4;
160         }
161 
162         while (count --> 0) {
163             // This is exactly the same flow as the count >= 4 loop above, but writing one pixel.
164             int x0, x1, wx;
165             decode_packed_coordinates_and_weight(*xy++, &x0, &x1, &wx);
166 
167             // As above, splat out wx four times as wx1, and sixteen minus that as wx0.
168             __m128i wx1 = _mm_set1_epi8(wx),     // This splats it out 16 times, but that's fine.
169                     wx0 = _mm_sub_epi8(_mm_set1_epi8(16), wx1);
170 
171             __m128i interlaced_x_weights_A = _mm_unpacklo_epi8(wx0, wx1);
172 
173             __m128i A = interpolate_in_x_and_y(row0[x0], row0[x1],
174                                                row1[x0], row1[x1],
175                                                       0,        0,
176                                                       0,        0,
177                                                interlaced_x_weights_A, wy);
178 
179             *colors++ = _mm_cvtsi128_si32(_mm_packus_epi16(scale_by_alpha(A), _mm_setzero_si128()));
180         }
181     }
182 
183 
184 #elif 1 && SK_CPU_SSE_LEVEL >= SK_CPU_SSE_LEVEL_SSE2
185 
186     // TODO(mtklein): clean up this code, use decode_packed_coordinates_and_weight(), etc.
187 
188     /*not static*/ inline
189     void S32_alpha_D32_filter_DX(const SkBitmapProcState& s,
190                                  const uint32_t* xy, int count, uint32_t* colors) {
191         SkASSERT(count > 0 && colors != nullptr);
192         SkASSERT(s.fFilterQuality != kNone_SkFilterQuality);
193         SkASSERT(kN32_SkColorType == s.fPixmap.colorType());
194         SkASSERT(s.fAlphaScale <= 256);
195 
196         int y0, y1, wy;
197         decode_packed_coordinates_and_weight(*xy++, &y0, &y1, &wy);
198 
199         auto row0 = (const uint32_t*)( (const char*)s.fPixmap.addr() + y0 * s.fPixmap.rowBytes() ),
200              row1 = (const uint32_t*)( (const char*)s.fPixmap.addr() + y1 * s.fPixmap.rowBytes() );
201 
202         // We'll put one pixel in the low 4 16-bit lanes to line up with wy,
203         // and another in the upper 4 16-bit lanes to line up with 16 - wy.
204         const __m128i allY = _mm_unpacklo_epi64(_mm_set1_epi16(   wy),
205                                                 _mm_set1_epi16(16-wy));
206 
207         while (count --> 0) {
208             int x0, x1, wx;
209             decode_packed_coordinates_and_weight(*xy++, &x0, &x1, &wx);
210 
211             // Load the 4 pixels we're interpolating.
212             const __m128i a00 = _mm_cvtsi32_si128(row0[x0]),
213                           a01 = _mm_cvtsi32_si128(row0[x1]),
214                           a10 = _mm_cvtsi32_si128(row1[x0]),
215                           a11 = _mm_cvtsi32_si128(row1[x1]);
216 
217             // Line up low-x pixels a00 and a10 with allY.
218             __m128i a00a10 = _mm_unpacklo_epi8(_mm_unpacklo_epi32(a10, a00),
219                                                _mm_setzero_si128());
220 
221             // Scale by allY and 16-wx.
222             a00a10 = _mm_mullo_epi16(a00a10, allY);
223             a00a10 = _mm_mullo_epi16(a00a10, _mm_set1_epi16(16-wx));
224 
225 
226             // Line up high-x pixels a01 and a11 with allY.
227             __m128i a01a11 = _mm_unpacklo_epi8(_mm_unpacklo_epi32(a11, a01),
228                                                _mm_setzero_si128());
229 
230             // Scale by allY and wx.
231             a01a11 = _mm_mullo_epi16(a01a11, allY);
232             a01a11 = _mm_mullo_epi16(a01a11, _mm_set1_epi16(wx));
233 
234 
235             // Add the two intermediates, summing across in one direction.
236             __m128i halves = _mm_add_epi16(a00a10, a01a11);
237 
238             // Add the two halves to each other to sum in the other direction.
239             __m128i sum = _mm_add_epi16(halves, _mm_srli_si128(halves, 8));
240 
241             // Get back to [0,255] by dividing by maximum weight 16x16 = 256.
242             sum = _mm_srli_epi16(sum, 8);
243 
244             if (s.fAlphaScale < 256) {
245                 // Scale by alpha, which is in [0,256].
246                 sum = _mm_mullo_epi16(sum, _mm_set1_epi16(s.fAlphaScale));
247                 sum = _mm_srli_epi16(sum, 8);
248             }
249 
250             // Pack back into 8-bit values and store.
251             *colors++ = _mm_cvtsi128_si32(_mm_packus_epi16(sum, _mm_setzero_si128()));
252         }
253     }
254 
255 #else
256 
257     // The NEON code only actually differs from the portable code in the
258     // filtering step after we've loaded all four pixels we want to bilerp.
259 
260     #if defined(SK_ARM_HAS_NEON)
261         static void filter_and_scale_by_alpha(unsigned x, unsigned y,
262                                               SkPMColor a00, SkPMColor a01,
263                                               SkPMColor a10, SkPMColor a11,
264                                               SkPMColor *dst,
265                                               uint16_t scale) {
266             uint8x8_t vy, vconst16_8, v16_y, vres;
267             uint16x4_t vx, vconst16_16, v16_x, tmp, vscale;
268             uint32x2_t va0, va1;
269             uint16x8_t tmp1, tmp2;
270 
271             vy = vdup_n_u8(y);                // duplicate y into vy
272             vconst16_8 = vmov_n_u8(16);       // set up constant in vconst16_8
273             v16_y = vsub_u8(vconst16_8, vy);  // v16_y = 16-y
274 
275             va0 = vdup_n_u32(a00);            // duplicate a00
276             va1 = vdup_n_u32(a10);            // duplicate a10
277             va0 = vset_lane_u32(a01, va0, 1); // set top to a01
278             va1 = vset_lane_u32(a11, va1, 1); // set top to a11
279 
280             tmp1 = vmull_u8(vreinterpret_u8_u32(va0), v16_y); // tmp1 = [a01|a00] * (16-y)
281             tmp2 = vmull_u8(vreinterpret_u8_u32(va1), vy);    // tmp2 = [a11|a10] * y
282 
283             vx = vdup_n_u16(x);                // duplicate x into vx
284             vconst16_16 = vmov_n_u16(16);      // set up constant in vconst16_16
285             v16_x = vsub_u16(vconst16_16, vx); // v16_x = 16-x
286 
287             tmp = vmul_u16(vget_high_u16(tmp1), vx);        // tmp  = a01 * x
288             tmp = vmla_u16(tmp, vget_high_u16(tmp2), vx);   // tmp += a11 * x
289             tmp = vmla_u16(tmp, vget_low_u16(tmp1), v16_x); // tmp += a00 * (16-x)
290             tmp = vmla_u16(tmp, vget_low_u16(tmp2), v16_x); // tmp += a10 * (16-x)
291 
292             if (scale < 256) {
293                 vscale = vdup_n_u16(scale);        // duplicate scale
294                 tmp = vshr_n_u16(tmp, 8);          // shift down result by 8
295                 tmp = vmul_u16(tmp, vscale);       // multiply result by scale
296             }
297 
298             vres = vshrn_n_u16(vcombine_u16(tmp, vcreate_u16(0)), 8); // shift down result by 8
299             vst1_lane_u32(dst, vreinterpret_u32_u8(vres), 0);         // store result
300         }
301     #else
302         static void filter_and_scale_by_alpha(unsigned x, unsigned y,
303                                               SkPMColor a00, SkPMColor a01,
304                                               SkPMColor a10, SkPMColor a11,
305                                               SkPMColor* dstColor,
306                                               unsigned alphaScale) {
307             SkASSERT((unsigned)x <= 0xF);
308             SkASSERT((unsigned)y <= 0xF);
309             SkASSERT(alphaScale <= 256);
310 
311             int xy = x * y;
312             const uint32_t mask = 0xFF00FF;
313 
314             int scale = 256 - 16*y - 16*x + xy;
315             uint32_t lo = (a00 & mask) * scale;
316             uint32_t hi = ((a00 >> 8) & mask) * scale;
317 
318             scale = 16*x - xy;
319             lo += (a01 & mask) * scale;
320             hi += ((a01 >> 8) & mask) * scale;
321 
322             scale = 16*y - xy;
323             lo += (a10 & mask) * scale;
324             hi += ((a10 >> 8) & mask) * scale;
325 
326             lo += (a11 & mask) * xy;
327             hi += ((a11 >> 8) & mask) * xy;
328 
329             if (alphaScale < 256) {
330                 lo = ((lo >> 8) & mask) * alphaScale;
331                 hi = ((hi >> 8) & mask) * alphaScale;
332             }
333 
334             *dstColor = ((lo >> 8) & mask) | (hi & ~mask);
335         }
336     #endif
337 
338 
339     /*not static*/ inline
340     void S32_alpha_D32_filter_DX(const SkBitmapProcState& s,
341                                  const uint32_t* xy, int count, SkPMColor* colors) {
342         SkASSERT(count > 0 && colors != nullptr);
343         SkASSERT(s.fFilterQuality != kNone_SkFilterQuality);
344         SkASSERT(4 == s.fPixmap.info().bytesPerPixel());
345         SkASSERT(s.fAlphaScale <= 256);
346 
347         int y0, y1, wy;
348         decode_packed_coordinates_and_weight(*xy++, &y0, &y1, &wy);
349 
350         auto row0 = (const uint32_t*)( (const char*)s.fPixmap.addr() + y0 * s.fPixmap.rowBytes() ),
351              row1 = (const uint32_t*)( (const char*)s.fPixmap.addr() + y1 * s.fPixmap.rowBytes() );
352 
353         while (count --> 0) {
354             int x0, x1, wx;
355             decode_packed_coordinates_and_weight(*xy++, &x0, &x1, &wx);
356 
357             filter_and_scale_by_alpha(wx, wy,
358                                       row0[x0], row0[x1],
359                                       row1[x0], row1[x1],
360                                       colors++,
361                                       s.fAlphaScale);
362         }
363     }
364 
365 #endif
366 
367 }  // namespace SK_OPTS_NS
368 
369 #endif
370