1 /*
2 * Copyright 2015 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 SkBlitMask_opts_DEFINED
9 #define SkBlitMask_opts_DEFINED
10
11 #include "include/private/base/SkFeatures.h"
12 #include "src/core/Sk4px.h"
13
14 #if defined(SK_ARM_HAS_NEON)
15 #include <arm_neon.h>
16 #endif
17
18 namespace SK_OPTS_NS {
19
20 #if defined(SK_ARM_HAS_NEON)
21 // The Sk4px versions below will work fine with NEON, but we have had many indications
22 // that it doesn't perform as well as this NEON-specific code. TODO(mtklein): why?
23
24 #define NEON_A (SK_A32_SHIFT / 8)
25 #define NEON_R (SK_R32_SHIFT / 8)
26 #define NEON_G (SK_G32_SHIFT / 8)
27 #define NEON_B (SK_B32_SHIFT / 8)
28
SkAlpha255To256_neon8(uint8x8_t alpha)29 static inline uint16x8_t SkAlpha255To256_neon8(uint8x8_t alpha) {
30 return vaddw_u8(vdupq_n_u16(1), alpha);
31 }
32
SkAlphaMul_neon8(uint8x8_t color,uint16x8_t scale)33 static inline uint8x8_t SkAlphaMul_neon8(uint8x8_t color, uint16x8_t scale) {
34 return vshrn_n_u16(vmovl_u8(color) * scale, 8);
35 }
36
SkAlphaMulQ_neon8(uint8x8x4_t color,uint16x8_t scale)37 static inline uint8x8x4_t SkAlphaMulQ_neon8(uint8x8x4_t color, uint16x8_t scale) {
38 uint8x8x4_t ret;
39
40 ret.val[0] = SkAlphaMul_neon8(color.val[0], scale);
41 ret.val[1] = SkAlphaMul_neon8(color.val[1], scale);
42 ret.val[2] = SkAlphaMul_neon8(color.val[2], scale);
43 ret.val[3] = SkAlphaMul_neon8(color.val[3], scale);
44
45 return ret;
46 }
47
48
49 template <bool isColor>
D32_A8_Opaque_Color_neon(void * SK_RESTRICT dst,size_t dstRB,const void * SK_RESTRICT maskPtr,size_t maskRB,SkColor color,int width,int height)50 static void D32_A8_Opaque_Color_neon(void* SK_RESTRICT dst, size_t dstRB,
51 const void* SK_RESTRICT maskPtr, size_t maskRB,
52 SkColor color, int width, int height) {
53 SkPMColor pmc = SkPreMultiplyColor(color);
54 SkPMColor* SK_RESTRICT device = (SkPMColor*)dst;
55 const uint8_t* SK_RESTRICT mask = (const uint8_t*)maskPtr;
56 uint8x8x4_t vpmc;
57
58 // Nine patch may set maskRB to 0 to blit the same row repeatedly.
59 ptrdiff_t mask_adjust = (ptrdiff_t)maskRB - width;
60 dstRB -= (width << 2);
61
62 if (width >= 8) {
63 vpmc.val[NEON_A] = vdup_n_u8(SkGetPackedA32(pmc));
64 vpmc.val[NEON_R] = vdup_n_u8(SkGetPackedR32(pmc));
65 vpmc.val[NEON_G] = vdup_n_u8(SkGetPackedG32(pmc));
66 vpmc.val[NEON_B] = vdup_n_u8(SkGetPackedB32(pmc));
67 }
68 do {
69 int w = width;
70 while (w >= 8) {
71 uint8x8_t vmask = vld1_u8(mask);
72 uint16x8_t vscale, vmask256 = SkAlpha255To256_neon8(vmask);
73 if (isColor) {
74 vscale = vsubw_u8(vdupq_n_u16(256),
75 SkAlphaMul_neon8(vpmc.val[NEON_A], vmask256));
76 } else {
77 vscale = vsubw_u8(vdupq_n_u16(256), vmask);
78 }
79 uint8x8x4_t vdev = vld4_u8((uint8_t*)device);
80
81 vdev.val[NEON_A] = SkAlphaMul_neon8(vpmc.val[NEON_A], vmask256)
82 + SkAlphaMul_neon8(vdev.val[NEON_A], vscale);
83 vdev.val[NEON_R] = SkAlphaMul_neon8(vpmc.val[NEON_R], vmask256)
84 + SkAlphaMul_neon8(vdev.val[NEON_R], vscale);
85 vdev.val[NEON_G] = SkAlphaMul_neon8(vpmc.val[NEON_G], vmask256)
86 + SkAlphaMul_neon8(vdev.val[NEON_G], vscale);
87 vdev.val[NEON_B] = SkAlphaMul_neon8(vpmc.val[NEON_B], vmask256)
88 + SkAlphaMul_neon8(vdev.val[NEON_B], vscale);
89
90 vst4_u8((uint8_t*)device, vdev);
91
92 mask += 8;
93 device += 8;
94 w -= 8;
95 }
96
97 while (w--) {
98 unsigned aa = *mask++;
99 if (isColor) {
100 *device = SkBlendARGB32(pmc, *device, aa);
101 } else {
102 *device = SkAlphaMulQ(pmc, SkAlpha255To256(aa))
103 + SkAlphaMulQ(*device, SkAlpha255To256(255 - aa));
104 }
105 device += 1;
106 }
107
108 device = (uint32_t*)((char*)device + dstRB);
109 mask += mask_adjust;
110
111 } while (--height != 0);
112 }
113
blit_mask_d32_a8_general(SkPMColor * dst,size_t dstRB,const SkAlpha * mask,size_t maskRB,SkColor color,int w,int h)114 static void blit_mask_d32_a8_general(SkPMColor* dst, size_t dstRB,
115 const SkAlpha* mask, size_t maskRB,
116 SkColor color, int w, int h) {
117 D32_A8_Opaque_Color_neon<true>(dst, dstRB, mask, maskRB, color, w, h);
118 }
119
120 // As above, but made slightly simpler by requiring that color is opaque.
blit_mask_d32_a8_opaque(SkPMColor * dst,size_t dstRB,const SkAlpha * mask,size_t maskRB,SkColor color,int w,int h)121 static void blit_mask_d32_a8_opaque(SkPMColor* dst, size_t dstRB,
122 const SkAlpha* mask, size_t maskRB,
123 SkColor color, int w, int h) {
124 D32_A8_Opaque_Color_neon<false>(dst, dstRB, mask, maskRB, color, w, h);
125 }
126
127 // Same as _opaque, but assumes color == SK_ColorBLACK, a very common and even simpler case.
blit_mask_d32_a8_black(SkPMColor * dst,size_t dstRB,const SkAlpha * maskPtr,size_t maskRB,int width,int height)128 static void blit_mask_d32_a8_black(SkPMColor* dst, size_t dstRB,
129 const SkAlpha* maskPtr, size_t maskRB,
130 int width, int height) {
131 SkPMColor* SK_RESTRICT device = (SkPMColor*)dst;
132 const uint8_t* SK_RESTRICT mask = (const uint8_t*)maskPtr;
133
134 // Nine patch may set maskRB to 0 to blit the same row repeatedly.
135 ptrdiff_t mask_adjust = (ptrdiff_t)maskRB - width;
136 dstRB -= (width << 2);
137 do {
138 int w = width;
139 while (w >= 8) {
140 uint8x8_t vmask = vld1_u8(mask);
141 uint16x8_t vscale = vsubw_u8(vdupq_n_u16(256), vmask);
142 uint8x8x4_t vdevice = vld4_u8((uint8_t*)device);
143
144 vdevice = SkAlphaMulQ_neon8(vdevice, vscale);
145 vdevice.val[NEON_A] += vmask;
146
147 vst4_u8((uint8_t*)device, vdevice);
148
149 mask += 8;
150 device += 8;
151 w -= 8;
152 }
153 while (w-- > 0) {
154 unsigned aa = *mask++;
155 *device = (aa << SK_A32_SHIFT)
156 + SkAlphaMulQ(*device, SkAlpha255To256(255 - aa));
157 device += 1;
158 }
159 device = (uint32_t*)((char*)device + dstRB);
160 mask += mask_adjust;
161 } while (--height != 0);
162 }
163
164 #elif SK_CPU_LSX_LEVEL >= SK_CPU_LSX_LEVEL_LSX
165 #include <lsxintrin.h>
166
167 static __m128i SkAlphaMul_lsx(__m128i x, __m128i y) {
168 __m128i tmp = __lsx_vmul_h(x, y);
169 __m128i mask = __lsx_vreplgr2vr_h(0xff00);
170 return __lsx_vsrlri_h(__lsx_vand_v(tmp, mask), 8);
171 }
172
173 template <bool isColor>
174 static void D32_A8_Opaque_Color_lsx(void* SK_RESTRICT dst, size_t dstRB,
175 const void* SK_RESTRICT maskPtr, size_t maskRB,
176 SkColor color, int width, int height) {
177 SkPMColor pmc = SkPreMultiplyColor(color);
178 SkPMColor* SK_RESTRICT device = (SkPMColor*)dst;
179 const uint8_t* SK_RESTRICT mask = (const uint8_t*)maskPtr;
180 __m128i vpmc_b = __lsx_vldi(0);
181 __m128i vpmc_g = __lsx_vldi(0);
182 __m128i vpmc_r = __lsx_vldi(0);
183 __m128i vpmc_a = __lsx_vldi(0);
184
185 // Nine patch may set maskRB to 0 to blit the same row repeatedly.
186 ptrdiff_t mask_adjust = (ptrdiff_t)maskRB - width;
187 dstRB -= (width << 2);
188
189 if (width >= 8) {
190 vpmc_b = __lsx_vreplgr2vr_h(SkGetPackedB32(pmc));
191 vpmc_g = __lsx_vreplgr2vr_h(SkGetPackedG32(pmc));
192 vpmc_r = __lsx_vreplgr2vr_h(SkGetPackedR32(pmc));
193 vpmc_a = __lsx_vreplgr2vr_h(SkGetPackedA32(pmc));
194 }
195
196 const __m128i zeros = __lsx_vldi(0);
197 __m128i planar = __lsx_vldi(0);
198 planar = __lsx_vinsgr2vr_d(planar, 0x0d0905010c080400, 0);
199 planar = __lsx_vinsgr2vr_d(planar, 0x0f0b07030e0a0602, 1);
200
201 do{
202 int w = width;
203 while(w >= 8){
204 __m128i lo = __lsx_vld(device, 0); // bgra bgra bgra bgra
205 __m128i hi = __lsx_vld(device, 16); // BGRA BGRA BGRA BGRA
206 lo = __lsx_vshuf_b(zeros, lo, planar); // bbbb gggg rrrr aaaa
207 hi = __lsx_vshuf_b(zeros, hi, planar); // BBBB GGGG RRRR AAAA
208 __m128i bg = __lsx_vilvl_w(hi, lo), // bbbb BBBB gggg GGGG
209 ra = __lsx_vilvh_w(hi, lo); // rrrr RRRR aaaa AAAA
210
211 __m128i b = __lsx_vilvl_b(zeros, bg), // _b_b _b_b _B_B _B_B
212 g = __lsx_vilvh_b(zeros, bg), // _g_g _g_g _G_G _G_G
213 r = __lsx_vilvl_b(zeros, ra), // _r_r _r_r _R_R _R_R
214 a = __lsx_vilvh_b(zeros, ra); // _a_a _a_a _A_A _A_A
215
216 __m128i vmask = __lsx_vld(mask, 0);
217 vmask = __lsx_vilvl_b(zeros, vmask);
218 __m128i vscale, vmask256 = __lsx_vadd_h(vmask, __lsx_vreplgr2vr_h(1));
219
220 if (isColor) {
221 __m128i tmp = SkAlphaMul_lsx(vpmc_a, vmask256);
222 vscale = __lsx_vsub_h(__lsx_vreplgr2vr_h(256), tmp);
223 } else {
224 vscale = __lsx_vsub_h(__lsx_vreplgr2vr_h(256), vmask);
225 }
226
227 b = SkAlphaMul_lsx(vpmc_b, vmask256) + SkAlphaMul_lsx(b, vscale);
228 g = SkAlphaMul_lsx(vpmc_g, vmask256) + SkAlphaMul_lsx(g, vscale);
229 r = SkAlphaMul_lsx(vpmc_r, vmask256) + SkAlphaMul_lsx(r, vscale);
230 a = SkAlphaMul_lsx(vpmc_a, vmask256) + SkAlphaMul_lsx(a, vscale);
231
232 bg = __lsx_vor_v(b, __lsx_vslli_h(g, 8)); // bgbg bgbg BGBG BGBG
233 ra = __lsx_vor_v(r, __lsx_vslli_h(a, 8)); // rara rara RARA RARA
234 lo = __lsx_vilvl_h(ra, bg); // bgra bgra bgra bgra
235 hi = __lsx_vilvh_h(ra, bg); // BGRA BGRA BGRA BGRA
236
237 __lsx_vst(lo, device, 0);
238 __lsx_vst(hi, device, 16);
239
240 mask += 8;
241 device += 8;
242 w -= 8;
243 }
244
245 while (w--) {
246 unsigned aa = *mask++;
247 if (isColor) {
248 *device = SkBlendARGB32(pmc, *device, aa);
249 } else {
250 *device = SkAlphaMulQ(pmc, SkAlpha255To256(aa))
251 + SkAlphaMulQ(*device, SkAlpha255To256(255 - aa));
252 }
253 device += 1;
254 }
255
256 device = (uint32_t *)((char*)device + dstRB);
257 mask += mask_adjust;
258
259 } while (--height != 0);
260 }
261
262 static void blit_mask_d32_a8_general(SkPMColor* dst, size_t dstRB,
263 const SkAlpha* mask, size_t maskRB,
264 SkColor color, int w, int h) {
265 D32_A8_Opaque_Color_lsx<true>(dst, dstRB, mask, maskRB, color, w, h);
266 }
267
268 static void blit_mask_d32_a8_opaque(SkPMColor* dst, size_t dstRB,
269 const SkAlpha* mask, size_t maskRB,
270 SkColor color, int w, int h) {
271 D32_A8_Opaque_Color_lsx<false>(dst, dstRB, mask, maskRB, color, w, h);
272 }
273
274 // Same as _opaque, but assumes color == SK_ColorBLACK, a very common and even simpler case.
275 static void blit_mask_d32_a8_black(SkPMColor* dst, size_t dstRB,
276 const SkAlpha* maskPtr, size_t maskRB,
277 int width, int height) {
278 SkPMColor* SK_RESTRICT device = (SkPMColor*)dst;
279 const uint8_t* SK_RESTRICT mask = (const uint8_t*)maskPtr;
280
281 // Nine patch may set maskRB to 0 to blit the same row repeatedly.
282 ptrdiff_t mask_adjust = (ptrdiff_t)maskRB - width;
283 dstRB -= (width << 2);
284 const __m128i zeros = __lsx_vldi(0);
285 __m128i planar = __lsx_vldi(0);
286 planar = __lsx_vinsgr2vr_d(planar, 0x0d0905010c080400, 0);
287 planar = __lsx_vinsgr2vr_d(planar, 0x0f0b07030e0a0602, 1);
288
289 do {
290 int w = width;
291 while (w >= 8) {
292 __m128i vmask = __lsx_vld(mask, 0);
293 vmask = __lsx_vilvl_b(zeros, vmask);
294 __m128i vscale = __lsx_vsub_h(__lsx_vreplgr2vr_h(256), vmask);
295 __m128i lo = __lsx_vld(device, 0); // bgra bgra bgra bgra
296 __m128i hi = __lsx_vld(device, 16); // BGRA BGRA BGRA BGRA
297 lo = __lsx_vshuf_b(zeros, lo, planar); // bbbb gggg rrrr aaaa
298 hi = __lsx_vshuf_b(zeros, hi, planar); // BBBB GGGG RRRR AAAA
299 __m128i bg = __lsx_vilvl_w(hi, lo), // bbbb BBBB gggg GGGG
300 ra = __lsx_vilvh_w(hi, lo); // rrrr RRRR aaaa AAAA
301
302 __m128i b = __lsx_vilvl_b(zeros, bg), // _b_b _b_b _B_B _B_B
303 g = __lsx_vilvh_b(zeros, bg), // _g_g _g_g _G_G _G_G
304 r = __lsx_vilvl_b(zeros, ra), // _r_r _r_r _R_R _R_R
305 a = __lsx_vilvh_b(zeros, ra); // _a_a _a_a _A_A _A_A
306
307 b = SkAlphaMul_lsx(b, vscale);
308 g = SkAlphaMul_lsx(g, vscale);
309 r = SkAlphaMul_lsx(r, vscale);
310 a = SkAlphaMul_lsx(a, vscale);
311
312 a += vmask;
313
314 bg = __lsx_vor_v(b, __lsx_vslli_h(g, 8)); // bgbg bgbg BGBG BGBG
315 ra = __lsx_vor_v(r, __lsx_vslli_h(a, 8)); // rara rara RARA RARA
316 lo = __lsx_vilvl_h(ra, bg); // bgra bgra bgra bgra
317 hi = __lsx_vilvh_h(ra, bg); // BGRA BGRA BGRA BGRA
318
319 __lsx_vst(lo, device, 0);
320 __lsx_vst(hi, device, 16);
321
322 mask += 8;
323 device += 8;
324 w -= 8;
325 }
326
327 while (w-- > 0) {
328 unsigned aa = *mask++;
329 *device = (aa << SK_A32_SHIFT)
330 + SkAlphaMulQ(*device, SkAlpha255To256(255 - aa));
331 device += 1;
332 }
333
334 device = (uint32_t*)((char*)device + dstRB);
335 mask += mask_adjust;
336
337 } while (--height != 0);
338 }
339
340 #else
341 static void blit_mask_d32_a8_general(SkPMColor* dst, size_t dstRB,
342 const SkAlpha* mask, size_t maskRB,
343 SkColor color, int w, int h) {
344 auto s = Sk4px::DupPMColor(SkPreMultiplyColor(color));
345 auto fn = [&](const Sk4px& d, const Sk4px& aa) {
346 // = (s + d(1-sa))aa + d(1-aa)
347 // = s*aa + d(1-sa*aa)
348 auto left = s.approxMulDiv255(aa),
349 right = d.approxMulDiv255(left.alphas().inv());
350 return left + right; // This does not overflow (exhaustively checked).
351 };
352 while (h --> 0) {
353 Sk4px::MapDstAlpha(w, dst, mask, fn);
354 dst += dstRB / sizeof(*dst);
355 mask += maskRB / sizeof(*mask);
356 }
357 }
358
359 // As above, but made slightly simpler by requiring that color is opaque.
360 static void blit_mask_d32_a8_opaque(SkPMColor* dst, size_t dstRB,
361 const SkAlpha* mask, size_t maskRB,
362 SkColor color, int w, int h) {
363 SkASSERT(SkColorGetA(color) == 0xFF);
364 auto s = Sk4px::DupPMColor(SkPreMultiplyColor(color));
365 auto fn = [&](const Sk4px& d, const Sk4px& aa) {
366 // = (s + d(1-sa))aa + d(1-aa)
367 // = s*aa + d(1-sa*aa)
368 // ~~~>
369 // = s*aa + d(1-aa)
370 return s.approxMulDiv255(aa) + d.approxMulDiv255(aa.inv());
371 };
372 while (h --> 0) {
373 Sk4px::MapDstAlpha(w, dst, mask, fn);
374 dst += dstRB / sizeof(*dst);
375 mask += maskRB / sizeof(*mask);
376 }
377 }
378
379 // Same as _opaque, but assumes color == SK_ColorBLACK, a very common and even simpler case.
380 static void blit_mask_d32_a8_black(SkPMColor* dst, size_t dstRB,
381 const SkAlpha* mask, size_t maskRB,
382 int w, int h) {
383 auto fn = [](const Sk4px& d, const Sk4px& aa) {
384 // = (s + d(1-sa))aa + d(1-aa)
385 // = s*aa + d(1-sa*aa)
386 // ~~~>
387 // a = 1*aa + d(1-1*aa) = aa + d(1-aa)
388 // c = 0*aa + d(1-1*aa) = d(1-aa)
389 return (aa & Sk4px(skvx::byte16{0,0,0,255, 0,0,0,255, 0,0,0,255, 0,0,0,255}))
390 + d.approxMulDiv255(aa.inv());
391 };
392 while (h --> 0) {
393 Sk4px::MapDstAlpha(w, dst, mask, fn);
394 dst += dstRB / sizeof(*dst);
395 mask += maskRB / sizeof(*mask);
396 }
397 }
398 #endif
399
blit_mask_d32_a8(SkPMColor * dst,size_t dstRB,const SkAlpha * mask,size_t maskRB,SkColor color,int w,int h)400 /*not static*/ inline void blit_mask_d32_a8(SkPMColor* dst, size_t dstRB,
401 const SkAlpha* mask, size_t maskRB,
402 SkColor color, int w, int h) {
403 if (color == SK_ColorBLACK) {
404 blit_mask_d32_a8_black(dst, dstRB, mask, maskRB, w, h);
405 } else if (SkColorGetA(color) == 0xFF) {
406 blit_mask_d32_a8_opaque(dst, dstRB, mask, maskRB, color, w, h);
407 } else {
408 blit_mask_d32_a8_general(dst, dstRB, mask, maskRB, color, w, h);
409 }
410 }
411
412 } // namespace SK_OPTS_NS
413
414 #endif//SkBlitMask_opts_DEFINED
415