• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2017 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 #include "SkJumper.h"
9 #include "SkJumper_misc.h"     // SI, unaligned_load(), bit_cast()
10 #include "SkJumper_vectors.h"  // F, I32, U32, U16, U8, cast(), expand()
11 
12 // Our fundamental vector depth is our pixel stride.
13 static const size_t kStride = sizeof(F) / sizeof(float);
14 
15 // A reminder:
16 // Code guarded by defined(JUMPER) can assume that it will be compiled by Clang
17 // and that F, I32, etc. are kStride-deep ext_vector_types of the appropriate type.
18 // Otherwise, F, I32, etc. just alias the basic scalar types (and so kStride == 1).
19 
20 // You can use most constants in this file, but in a few rare exceptions we read from this struct.
21 using K = const SkJumper_constants;
22 
23 // A little wrapper macro to name Stages differently depending on the instruction set.
24 // That lets us link together several options.
25 #if !defined(JUMPER)
26     #define WRAP(name) sk_##name
27 #elif defined(__aarch64__)
28     #define WRAP(name) sk_##name##_aarch64
29 #elif defined(__arm__)
30     #define WRAP(name) sk_##name##_vfp4
31 #elif defined(__AVX2__)
32     #define WRAP(name) sk_##name##_hsw
33 #elif defined(__AVX__)
34     #define WRAP(name) sk_##name##_avx
35 #elif defined(__SSE4_1__)
36     #define WRAP(name) sk_##name##_sse41
37 #elif defined(__SSE2__)
38     #define WRAP(name) sk_##name##_sse2
39 #endif
40 
41 // We're finally going to get to what a Stage function looks like!
42 //    tail == 0 ~~> work on a full kStride pixels
43 //    tail != 0 ~~> work on only the first tail pixels
44 // tail is always < kStride.
45 //
46 // We keep program the second argument, so that it's passed in rsi for load_and_inc().
47 using Stage = void(K* k, void** program, size_t x, size_t y, size_t tail, F,F,F,F, F,F,F,F);
48 
49 #if defined(JUMPER) && defined(__AVX__)
50     // We really want to make sure all paths go through this function's (implicit) vzeroupper.
51     // If they don't, we'll experience severe slowdowns when we first use SSE instructions again.
52     __attribute__((disable_tail_calls))
53 #endif
54 MAYBE_MSABI
WRAP(start_pipeline)55 extern "C" void WRAP(start_pipeline)(size_t x, size_t y, size_t limit, void** program, K* k) {
56 #if defined(JUMPER)
57     F v;
58 #else
59     F v{};
60 #endif
61     auto start = (Stage*)load_and_inc(program);
62     while (x + kStride <= limit) {
63         start(k,program,x,y,0,    v,v,v,v, v,v,v,v);
64         x += kStride;
65     }
66     if (size_t tail = limit - x) {
67         start(k,program,x,y,tail, v,v,v,v, v,v,v,v);
68     }
69 }
70 
71 #if defined(JUMPER) && defined(__AVX__)
72     // We really want to make sure all paths go through this function's (implicit) vzeroupper.
73     // If they don't, we'll experience severe slowdowns when we first use SSE instructions again.
74     __attribute__((disable_tail_calls))
75 #endif
76 #if defined(JUMPER)
77     __attribute__((flatten))  // Force-inline the call to start_pipeline().
78 #endif
79 MAYBE_MSABI
WRAP(start_pipeline_2d)80 extern "C" void WRAP(start_pipeline_2d)(size_t x, size_t y, size_t xlimit, size_t ylimit,
81                                         void** program, K* k) {
82     for (; y < ylimit; y++) {
83         WRAP(start_pipeline)(x,y,xlimit, program, k);
84     }
85 }
86 
87 #define STAGE(name)                                                                   \
88     SI void name##_k(K* k, LazyCtx ctx, size_t x, size_t y, size_t tail,              \
89                      F& r, F& g, F& b, F& a, F& dr, F& dg, F& db, F& da);             \
90     extern "C" void WRAP(name)(K* k, void** program, size_t x, size_t y, size_t tail, \
91                                F r, F g, F b, F a, F dr, F dg, F db, F da) {          \
92         LazyCtx ctx(program);                                                         \
93         name##_k(k,ctx,x,y,tail, r,g,b,a, dr,dg,db,da);                               \
94         auto next = (Stage*)load_and_inc(program);                                    \
95         next(k,program,x,y,tail, r,g,b,a, dr,dg,db,da);                               \
96     }                                                                                 \
97     SI void name##_k(K* k, LazyCtx ctx, size_t x, size_t y, size_t tail,              \
98                      F& r, F& g, F& b, F& a, F& dr, F& dg, F& db, F& da)
99 
100 
101 // just_return() is a simple no-op stage that only exists to end the chain,
102 // returning back up to start_pipeline(), and from there to the caller.
WRAP(just_return)103 extern "C" void WRAP(just_return)(K*, void**, size_t,size_t,size_t, F,F,F,F, F,F,F,F) {}
104 
105 
106 // We could start defining normal Stages now.  But first, some helper functions.
107 
108 // These load() and store() methods are tail-aware,
109 // but focus mainly on keeping the at-stride tail==0 case fast.
110 
111 template <typename V, typename T>
load(const T * src,size_t tail)112 SI V load(const T* src, size_t tail) {
113 #if defined(JUMPER)
114     __builtin_assume(tail < kStride);
115     if (__builtin_expect(tail, 0)) {
116         V v{};  // Any inactive lanes are zeroed.
117         switch (tail) {
118             case 7: v[6] = src[6];
119             case 6: v[5] = src[5];
120             case 5: v[4] = src[4];
121             case 4: memcpy(&v, src, 4*sizeof(T)); break;
122             case 3: v[2] = src[2];
123             case 2: memcpy(&v, src, 2*sizeof(T)); break;
124             case 1: memcpy(&v, src, 1*sizeof(T)); break;
125         }
126         return v;
127     }
128 #endif
129     return unaligned_load<V>(src);
130 }
131 
132 template <typename V, typename T>
store(T * dst,V v,size_t tail)133 SI void store(T* dst, V v, size_t tail) {
134 #if defined(JUMPER)
135     __builtin_assume(tail < kStride);
136     if (__builtin_expect(tail, 0)) {
137         switch (tail) {
138             case 7: dst[6] = v[6];
139             case 6: dst[5] = v[5];
140             case 5: dst[4] = v[4];
141             case 4: memcpy(dst, &v, 4*sizeof(T)); break;
142             case 3: dst[2] = v[2];
143             case 2: memcpy(dst, &v, 2*sizeof(T)); break;
144             case 1: memcpy(dst, &v, 1*sizeof(T)); break;
145         }
146         return;
147     }
148 #endif
149     unaligned_store(dst, v);
150 }
151 
152 // AVX adds some mask loads and stores that make for shorter, faster code.
153 #if defined(JUMPER) && defined(__AVX__)
mask(size_t tail)154     SI U32 mask(size_t tail) {
155         // We go a little out of our way to avoid needing large constant values here.
156 
157         // It's easiest to build the mask as 8 8-bit values, either 0x00 or 0xff.
158         // Start fully on, then shift away lanes from the top until we've got our mask.
159         uint64_t mask = 0xffffffffffffffff >> 8*(kStride-tail);
160 
161         // Sign-extend each mask lane to its full width, 0x00000000 or 0xffffffff.
162         using S8  = int8_t  __attribute__((ext_vector_type(8)));
163         using S32 = int32_t __attribute__((ext_vector_type(8)));
164         return (U32)__builtin_convertvector(unaligned_load<S8>(&mask), S32);
165     }
166 
167     template <>
load(const uint32_t * src,size_t tail)168     inline U32 load(const uint32_t* src, size_t tail) {
169         __builtin_assume(tail < kStride);
170         if (__builtin_expect(tail, 0)) {
171             return (U32)_mm256_maskload_ps((const float*)src, mask(tail));
172         }
173         return unaligned_load<U32>(src);
174     }
175 
176     template <>
store(uint32_t * dst,U32 v,size_t tail)177     inline void store(uint32_t* dst, U32 v, size_t tail) {
178         __builtin_assume(tail < kStride);
179         if (__builtin_expect(tail, 0)) {
180             return _mm256_maskstore_ps((float*)dst, mask(tail), (F)v);
181         }
182         unaligned_store(dst, v);
183     }
184 #endif
185 
from_byte(U8 b)186 SI F from_byte(U8 b) {
187     return cast(expand(b)) * (1/255.0f);
188 }
from_565(U16 _565,F * r,F * g,F * b)189 SI void from_565(U16 _565, F* r, F* g, F* b) {
190     U32 wide = expand(_565);
191     *r = cast(wide & (31<<11)) * (1.0f / (31<<11));
192     *g = cast(wide & (63<< 5)) * (1.0f / (63<< 5));
193     *b = cast(wide & (31<< 0)) * (1.0f / (31<< 0));
194 }
from_4444(U16 _4444,F * r,F * g,F * b,F * a)195 SI void from_4444(U16 _4444, F* r, F* g, F* b, F* a) {
196     U32 wide = expand(_4444);
197     *r = cast(wide & (15<<12)) * (1.0f / (15<<12));
198     *g = cast(wide & (15<< 8)) * (1.0f / (15<< 8));
199     *b = cast(wide & (15<< 4)) * (1.0f / (15<< 4));
200     *a = cast(wide & (15<< 0)) * (1.0f / (15<< 0));
201 }
from_8888(U32 _8888,F * r,F * g,F * b,F * a)202 SI void from_8888(U32 _8888, F* r, F* g, F* b, F* a) {
203     *r = cast((_8888      ) & 0xff) * (1/255.0f);
204     *g = cast((_8888 >>  8) & 0xff) * (1/255.0f);
205     *b = cast((_8888 >> 16) & 0xff) * (1/255.0f);
206     *a = cast((_8888 >> 24)       ) * (1/255.0f);
207 }
208 
209 template <typename T>
ix_and_ptr(T ** ptr,const SkJumper_MemoryCtx * ctx,F x,F y)210 SI U32 ix_and_ptr(T** ptr, const SkJumper_MemoryCtx* ctx, F x, F y) {
211     *ptr = (const T*)ctx->pixels;
212     return trunc_(y)*ctx->stride + trunc_(x);
213 }
214 
215 // Now finally, normal Stages!
216 
STAGE(seed_shader)217 STAGE(seed_shader) {
218     // It's important for speed to explicitly cast(x) and cast(y),
219     // which has the effect of splatting them to vectors before converting to floats.
220     // On Intel this breaks a data dependency on previous loop iterations' registers.
221     r = cast(x) + 0.5f + unaligned_load<F>(k->iota_F);
222     g = cast(y) + 0.5f;
223     b = 1.0f;
224     a = 0;
225     dr = dg = db = da = 0;
226 }
227 
STAGE(dither)228 STAGE(dither) {
229     auto rate = *(const float*)ctx;
230 
231     // Get [(x,y), (x+1,y), (x+2,y), ...] loaded up in integer vectors.
232     U32 X = x + unaligned_load<U32>(k->iota_U32),
233         Y = y;
234 
235     // We're doing 8x8 ordered dithering, see https://en.wikipedia.org/wiki/Ordered_dithering.
236     // In this case n=8 and we're using the matrix that looks like 1/64 x [ 0 48 12 60 ... ].
237 
238     // We only need X and X^Y from here on, so it's easier to just think of that as "Y".
239     Y ^= X;
240 
241     // We'll mix the bottom 3 bits of each of X and Y to make 6 bits,
242     // for 2^6 == 64 == 8x8 matrix values.  If X=abc and Y=def, we make fcebda.
243     U32 M = (Y & 1) << 5 | (X & 1) << 4
244           | (Y & 2) << 2 | (X & 2) << 1
245           | (Y & 4) >> 1 | (X & 4) >> 2;
246 
247     // Scale that dither to [0,1), then (-0.5,+0.5), here using 63/128 = 0.4921875 as 0.5-epsilon.
248     // We want to make sure our dither is less than 0.5 in either direction to keep exact values
249     // like 0 and 1 unchanged after rounding.
250     F dither = cast(M) * (2/128.0f) - (63/128.0f);
251 
252     r += rate*dither;
253     g += rate*dither;
254     b += rate*dither;
255 
256     r = max(0, min(r, a));
257     g = max(0, min(g, a));
258     b = max(0, min(b, a));
259 }
260 
261 // load 4 floats from memory, and splat them into r,g,b,a
STAGE(uniform_color)262 STAGE(uniform_color) {
263     auto rgba = (const float*)ctx;
264     r = rgba[0];
265     g = rgba[1];
266     b = rgba[2];
267     a = rgba[3];
268 }
269 
270 // splats opaque-black into r,g,b,a
STAGE(black_color)271 STAGE(black_color) {
272     r = g = b = 0.0f;
273     a = 1.0f;
274 }
275 
STAGE(white_color)276 STAGE(white_color) {
277     r = g = b = a = 1.0f;
278 }
279 
280 // load registers r,g,b,a from context (mirrors store_rgba)
STAGE(load_rgba)281 STAGE(load_rgba) {
282     auto ptr = (const float*)ctx;
283     r = unaligned_load<F>(ptr + 0*kStride);
284     g = unaligned_load<F>(ptr + 1*kStride);
285     b = unaligned_load<F>(ptr + 2*kStride);
286     a = unaligned_load<F>(ptr + 3*kStride);
287 }
288 
289 // store registers r,g,b,a into context (mirrors load_rgba)
STAGE(store_rgba)290 STAGE(store_rgba) {
291     auto ptr = (float*)ctx;
292     unaligned_store(ptr + 0*kStride, r);
293     unaligned_store(ptr + 1*kStride, g);
294     unaligned_store(ptr + 2*kStride, b);
295     unaligned_store(ptr + 3*kStride, a);
296 }
297 
298 // Most blend modes apply the same logic to each channel.
299 #define BLEND_MODE(name)                       \
300     SI F name##_channel(F s, F d, F sa, F da); \
301     STAGE(name) {                              \
302         r = name##_channel(r,dr,a,da);         \
303         g = name##_channel(g,dg,a,da);         \
304         b = name##_channel(b,db,a,da);         \
305         a = name##_channel(a,da,a,da);         \
306     }                                          \
307     SI F name##_channel(F s, F d, F sa, F da)
308 
inv(F x)309 SI F inv(F x) { return 1.0f - x; }
two(F x)310 SI F two(F x) { return x + x; }
311 
BLEND_MODE(clear)312 BLEND_MODE(clear)    { return 0; }
BLEND_MODE(srcatop)313 BLEND_MODE(srcatop)  { return s*da + d*inv(sa); }
BLEND_MODE(dstatop)314 BLEND_MODE(dstatop)  { return d*sa + s*inv(da); }
BLEND_MODE(srcin)315 BLEND_MODE(srcin)    { return s * da; }
BLEND_MODE(dstin)316 BLEND_MODE(dstin)    { return d * sa; }
BLEND_MODE(srcout)317 BLEND_MODE(srcout)   { return s * inv(da); }
BLEND_MODE(dstout)318 BLEND_MODE(dstout)   { return d * inv(sa); }
BLEND_MODE(srcover)319 BLEND_MODE(srcover)  { return mad(d, inv(sa), s); }
BLEND_MODE(dstover)320 BLEND_MODE(dstover)  { return mad(s, inv(da), d); }
321 
BLEND_MODE(modulate)322 BLEND_MODE(modulate) { return s*d; }
BLEND_MODE(multiply)323 BLEND_MODE(multiply) { return s*inv(da) + d*inv(sa) + s*d; }
BLEND_MODE(plus_)324 BLEND_MODE(plus_)    { return s + d; }
BLEND_MODE(screen)325 BLEND_MODE(screen)   { return s + d - s*d; }
BLEND_MODE(xor_)326 BLEND_MODE(xor_)     { return s*inv(da) + d*inv(sa); }
327 #undef BLEND_MODE
328 
329 // Most other blend modes apply the same logic to colors, and srcover to alpha.
330 #define BLEND_MODE(name)                       \
331     SI F name##_channel(F s, F d, F sa, F da); \
332     STAGE(name) {                              \
333         r = name##_channel(r,dr,a,da);         \
334         g = name##_channel(g,dg,a,da);         \
335         b = name##_channel(b,db,a,da);         \
336         a = mad(da, inv(a), a);                \
337     }                                          \
338     SI F name##_channel(F s, F d, F sa, F da)
339 
BLEND_MODE(darken)340 BLEND_MODE(darken)     { return s + d -     max(s*da, d*sa) ; }
BLEND_MODE(lighten)341 BLEND_MODE(lighten)    { return s + d -     min(s*da, d*sa) ; }
BLEND_MODE(difference)342 BLEND_MODE(difference) { return s + d - two(min(s*da, d*sa)); }
BLEND_MODE(exclusion)343 BLEND_MODE(exclusion)  { return s + d - two(s*d); }
344 
BLEND_MODE(colorburn)345 BLEND_MODE(colorburn) {
346     return if_then_else(d == da, d + s*inv(da),
347            if_then_else(s ==  0, s + d*inv(sa),
348                                  sa*(da - min(da, (da-d)*sa/s)) + s*inv(da) + d*inv(sa)));
349 }
BLEND_MODE(colordodge)350 BLEND_MODE(colordodge) {
351     return if_then_else(d ==  0, d + s*inv(da),
352            if_then_else(s == sa, s + d*inv(sa),
353                                  sa*min(da, (d*sa)/(sa - s)) + s*inv(da) + d*inv(sa)));
354 }
BLEND_MODE(hardlight)355 BLEND_MODE(hardlight) {
356     return s*inv(da) + d*inv(sa)
357          + if_then_else(two(s) <= sa, two(s*d), sa*da - two((da-d)*(sa-s)));
358 }
BLEND_MODE(overlay)359 BLEND_MODE(overlay) {
360     return s*inv(da) + d*inv(sa)
361          + if_then_else(two(d) <= da, two(s*d), sa*da - two((da-d)*(sa-s)));
362 }
363 
BLEND_MODE(softlight)364 BLEND_MODE(softlight) {
365     F m  = if_then_else(da > 0, d / da, 0),
366       s2 = two(s),
367       m4 = two(two(m));
368 
369     // The logic forks three ways:
370     //    1. dark src?
371     //    2. light src, dark dst?
372     //    3. light src, light dst?
373     F darkSrc = d*(sa + (s2 - sa)*(1.0f - m)),     // Used in case 1.
374       darkDst = (m4*m4 + m4)*(m - 1.0f) + 7.0f*m,  // Used in case 2.
375       liteDst = rcp(rsqrt(m)) - m,                 // Used in case 3.
376       liteSrc = d*sa + da*(s2 - sa) * if_then_else(two(two(d)) <= da, darkDst, liteDst); // 2 or 3?
377     return s*inv(da) + d*inv(sa) + if_then_else(s2 <= sa, darkSrc, liteSrc);      // 1 or (2 or 3)?
378 }
379 #undef BLEND_MODE
380 
381 // We're basing our implemenation of non-separable blend modes on
382 //   https://www.w3.org/TR/compositing-1/#blendingnonseparable.
383 // and
384 //   https://www.khronos.org/registry/OpenGL/specs/es/3.2/es_spec_3.2.pdf
385 // They're equivalent, but ES' math has been better simplified.
386 //
387 // Anything extra we add beyond that is to make the math work with premul inputs.
388 
max(F r,F g,F b)389 SI F max(F r, F g, F b) { return max(r, max(g, b)); }
min(F r,F g,F b)390 SI F min(F r, F g, F b) { return min(r, min(g, b)); }
391 
sat(F r,F g,F b)392 SI F sat(F r, F g, F b) { return max(r,g,b) - min(r,g,b); }
lum(F r,F g,F b)393 SI F lum(F r, F g, F b) { return r*0.30f + g*0.59f + b*0.11f; }
394 
set_sat(F * r,F * g,F * b,F s)395 SI void set_sat(F* r, F* g, F* b, F s) {
396     F mn  = min(*r,*g,*b),
397       mx  = max(*r,*g,*b),
398       sat = mx - mn;
399 
400     // Map min channel to 0, max channel to s, and scale the middle proportionally.
401     auto scale = [=](F c) {
402         return if_then_else(sat == 0, 0, (c - mn) * s / sat);
403     };
404     *r = scale(*r);
405     *g = scale(*g);
406     *b = scale(*b);
407 }
set_lum(F * r,F * g,F * b,F l)408 SI void set_lum(F* r, F* g, F* b, F l) {
409     F diff = l - lum(*r, *g, *b);
410     *r += diff;
411     *g += diff;
412     *b += diff;
413 }
clip_color(F * r,F * g,F * b,F a)414 SI void clip_color(F* r, F* g, F* b, F a) {
415     F mn = min(*r, *g, *b),
416       mx = max(*r, *g, *b),
417       l  = lum(*r, *g, *b);
418 
419     auto clip = [=](F c) {
420         c = if_then_else(mn >= 0, c, l + (c - l) * (    l) / (l - mn)   );
421         c = if_then_else(mx >  a,    l + (c - l) * (a - l) / (mx - l), c);
422         c = max(c, 0);  // Sometimes without this we may dip just a little negative.
423         return c;
424     };
425     *r = clip(*r);
426     *g = clip(*g);
427     *b = clip(*b);
428 }
429 
STAGE(hue)430 STAGE(hue) {
431     F R = r*a,
432       G = g*a,
433       B = b*a;
434 
435     set_sat(&R, &G, &B, sat(dr,dg,db)*a);
436     set_lum(&R, &G, &B, lum(dr,dg,db)*a);
437     clip_color(&R,&G,&B, a*da);
438 
439     r = r*inv(da) + dr*inv(a) + R;
440     g = g*inv(da) + dg*inv(a) + G;
441     b = b*inv(da) + db*inv(a) + B;
442     a = a + da - a*da;
443 }
STAGE(saturation)444 STAGE(saturation) {
445     F R = dr*a,
446       G = dg*a,
447       B = db*a;
448 
449     set_sat(&R, &G, &B, sat( r, g, b)*da);
450     set_lum(&R, &G, &B, lum(dr,dg,db)* a);  // (This is not redundant.)
451     clip_color(&R,&G,&B, a*da);
452 
453     r = r*inv(da) + dr*inv(a) + R;
454     g = g*inv(da) + dg*inv(a) + G;
455     b = b*inv(da) + db*inv(a) + B;
456     a = a + da - a*da;
457 }
STAGE(color)458 STAGE(color) {
459     F R = r*da,
460       G = g*da,
461       B = b*da;
462 
463     set_lum(&R, &G, &B, lum(dr,dg,db)*a);
464     clip_color(&R,&G,&B, a*da);
465 
466     r = r*inv(da) + dr*inv(a) + R;
467     g = g*inv(da) + dg*inv(a) + G;
468     b = b*inv(da) + db*inv(a) + B;
469     a = a + da - a*da;
470 }
STAGE(luminosity)471 STAGE(luminosity) {
472     F R = dr*a,
473       G = dg*a,
474       B = db*a;
475 
476     set_lum(&R, &G, &B, lum(r,g,b)*da);
477     clip_color(&R,&G,&B, a*da);
478 
479     r = r*inv(da) + dr*inv(a) + R;
480     g = g*inv(da) + dg*inv(a) + G;
481     b = b*inv(da) + db*inv(a) + B;
482     a = a + da - a*da;
483 }
484 
STAGE(srcover_rgba_8888)485 STAGE(srcover_rgba_8888) {
486     auto ptr = *(uint32_t**)ctx + x;
487 
488     U32 dst = load<U32>(ptr, tail);
489     dr = cast((dst      ) & 0xff);
490     dg = cast((dst >>  8) & 0xff);
491     db = cast((dst >> 16) & 0xff);
492     da = cast((dst >> 24)       );
493     // {dr,dg,db,da} are in [0,255]
494     // { r, g, b, a} are in [0,  1]
495 
496     r = mad(dr, inv(a), r*255.0f);
497     g = mad(dg, inv(a), g*255.0f);
498     b = mad(db, inv(a), b*255.0f);
499     a = mad(da, inv(a), a*255.0f);
500     // { r, g, b, a} are now in [0,255]
501 
502     dst = round(r, 1.0f)
503         | round(g, 1.0f) <<  8
504         | round(b, 1.0f) << 16
505         | round(a, 1.0f) << 24;
506     store(ptr, dst, tail);
507 }
508 
STAGE(clamp_0)509 STAGE(clamp_0) {
510     r = max(r, 0);
511     g = max(g, 0);
512     b = max(b, 0);
513     a = max(a, 0);
514 }
515 
STAGE(clamp_1)516 STAGE(clamp_1) {
517     r = min(r, 1.0f);
518     g = min(g, 1.0f);
519     b = min(b, 1.0f);
520     a = min(a, 1.0f);
521 }
522 
STAGE(clamp_a)523 STAGE(clamp_a) {
524     a = min(a, 1.0f);
525     r = min(r, a);
526     g = min(g, a);
527     b = min(b, a);
528 }
529 
STAGE(clamp_a_dst)530 STAGE(clamp_a_dst) {
531     da = min(da, 1.0f);
532     dr = min(dr, da);
533     dg = min(dg, da);
534     db = min(db, da);
535 }
536 
STAGE(set_rgb)537 STAGE(set_rgb) {
538     auto rgb = (const float*)ctx;
539     r = rgb[0];
540     g = rgb[1];
541     b = rgb[2];
542 }
STAGE(swap_rb)543 STAGE(swap_rb) {
544     auto tmp = r;
545     r = b;
546     b = tmp;
547 }
548 
STAGE(move_src_dst)549 STAGE(move_src_dst) {
550     dr = r;
551     dg = g;
552     db = b;
553     da = a;
554 }
STAGE(move_dst_src)555 STAGE(move_dst_src) {
556     r = dr;
557     g = dg;
558     b = db;
559     a = da;
560 }
561 
STAGE(premul)562 STAGE(premul) {
563     r = r * a;
564     g = g * a;
565     b = b * a;
566 }
STAGE(premul_dst)567 STAGE(premul_dst) {
568     dr = dr * da;
569     dg = dg * da;
570     db = db * da;
571 }
STAGE(unpremul)572 STAGE(unpremul) {
573     auto scale = if_then_else(a == 0, 0, 1.0f / a);
574     r *= scale;
575     g *= scale;
576     b *= scale;
577 }
578 
from_srgb(F s)579 SI F from_srgb(F s) {
580     auto lo = s * (1/12.92f);
581     auto hi = mad(s*s, mad(s, 0.3000f, 0.6975f), 0.0025f);
582     return if_then_else(s < 0.055f, lo, hi);
583 }
584 
STAGE(from_srgb)585 STAGE(from_srgb) {
586     r = from_srgb(r);
587     g = from_srgb(g);
588     b = from_srgb(b);
589 }
STAGE(from_srgb_dst)590 STAGE(from_srgb_dst) {
591     dr = from_srgb(dr);
592     dg = from_srgb(dg);
593     db = from_srgb(db);
594 }
STAGE(to_srgb)595 STAGE(to_srgb) {
596     auto fn = [&](F l) {
597         // We tweak c and d for each instruction set to make sure fn(1) is exactly 1.
598     #if defined(JUMPER) && defined(__SSE2__)
599         const float c = 1.130048394203f,
600                     d = 0.141357362270f;
601     #elif defined(JUMPER) && (defined(__aarch64__) || defined(__arm__))
602         const float c = 1.129999995232f,
603                     d = 0.141381442547f;
604     #else
605         const float c = 1.129999995232f,
606                     d = 0.141377761960f;
607     #endif
608         F t = rsqrt(l);
609         auto lo = l * 12.92f;
610         auto hi = mad(t, mad(t, -0.0024542345f, 0.013832027f), c)
611                 * rcp(d + t);
612         return if_then_else(l < 0.00465985f, lo, hi);
613     };
614     r = fn(r);
615     g = fn(g);
616     b = fn(b);
617 }
618 
STAGE(rgb_to_hsl)619 STAGE(rgb_to_hsl) {
620     F mx = max(max(r,g), b),
621       mn = min(min(r,g), b),
622       d = mx - mn,
623       d_rcp = 1.0f / d;
624 
625     F h = (1/6.0f) *
626           if_then_else(mx == mn, 0,
627           if_then_else(mx ==  r, (g-b)*d_rcp + if_then_else(g < b, 6.0f, 0),
628           if_then_else(mx ==  g, (b-r)*d_rcp + 2.0f,
629                                  (r-g)*d_rcp + 4.0f)));
630 
631     F l = (mx + mn) * 0.5f;
632     F s = if_then_else(mx == mn, 0,
633                        d / if_then_else(l > 0.5f, 2.0f-mx-mn, mx+mn));
634 
635     r = h;
636     g = s;
637     b = l;
638 }
STAGE(hsl_to_rgb)639 STAGE(hsl_to_rgb) {
640     F h = r,
641       s = g,
642       l = b;
643 
644     F q = l + if_then_else(l >= 0.5f, s - l*s, l*s),
645       p = 2.0f*l - q;
646 
647     auto hue_to_rgb = [&](F t) {
648         t = fract(t);
649 
650         F r = p;
651         r = if_then_else(t >= 4/6.0f, r, p + (q-p)*(4.0f - 6.0f*t));
652         r = if_then_else(t >= 3/6.0f, r, q);
653         r = if_then_else(t >= 1/6.0f, r, p + (q-p)*(       6.0f*t));
654         return r;
655     };
656 
657     r = if_then_else(s == 0, l, hue_to_rgb(h + (1/3.0f)));
658     g = if_then_else(s == 0, l, hue_to_rgb(h           ));
659     b = if_then_else(s == 0, l, hue_to_rgb(h - (1/3.0f)));
660 }
661 
STAGE(scale_1_float)662 STAGE(scale_1_float) {
663     auto c = *(const float*)ctx;
664 
665     r = r * c;
666     g = g * c;
667     b = b * c;
668     a = a * c;
669 }
STAGE(scale_u8)670 STAGE(scale_u8) {
671     auto ptr = *(const uint8_t**)ctx + x;
672 
673     auto scales = load<U8>(ptr, tail);
674     auto c = from_byte(scales);
675 
676     r = r * c;
677     g = g * c;
678     b = b * c;
679     a = a * c;
680 }
681 
lerp(F from,F to,F t)682 SI F lerp(F from, F to, F t) {
683     return mad(to-from, t, from);
684 }
685 
STAGE(lerp_1_float)686 STAGE(lerp_1_float) {
687     auto c = *(const float*)ctx;
688 
689     r = lerp(dr, r, c);
690     g = lerp(dg, g, c);
691     b = lerp(db, b, c);
692     a = lerp(da, a, c);
693 }
STAGE(lerp_u8)694 STAGE(lerp_u8) {
695     auto ptr = *(const uint8_t**)ctx + x;
696 
697     auto scales = load<U8>(ptr, tail);
698     auto c = from_byte(scales);
699 
700     r = lerp(dr, r, c);
701     g = lerp(dg, g, c);
702     b = lerp(db, b, c);
703     a = lerp(da, a, c);
704 }
STAGE(lerp_565)705 STAGE(lerp_565) {
706     auto ptr = *(const uint16_t**)ctx + x;
707 
708     F cr,cg,cb;
709     from_565(load<U16>(ptr, tail), &cr, &cg, &cb);
710 
711     r = lerp(dr, r, cr);
712     g = lerp(dg, g, cg);
713     b = lerp(db, b, cb);
714     a = max(lerp(da, a, cr), lerp(da, a, cg), lerp(da, a, cb));
715 }
716 
STAGE(load_tables)717 STAGE(load_tables) {
718     auto c = (const SkJumper_LoadTablesCtx*)ctx;
719 
720     auto px = load<U32>((const uint32_t*)c->src + x, tail);
721     r = gather(c->r, (px      ) & 0xff);
722     g = gather(c->g, (px >>  8) & 0xff);
723     b = gather(c->b, (px >> 16) & 0xff);
724     a = cast(        (px >> 24)) * (1/255.0f);
725 }
STAGE(load_tables_u16_be)726 STAGE(load_tables_u16_be) {
727     auto c = (const SkJumper_LoadTablesCtx*)ctx;
728     auto ptr = (const uint16_t*)c->src + 4*x;
729 
730     U16 R,G,B,A;
731     load4(ptr, tail, &R,&G,&B,&A);
732 
733     // c->src is big-endian, so & 0xff grabs the 8 most signficant bits.
734     r = gather(c->r, expand(R) & 0xff);
735     g = gather(c->g, expand(G) & 0xff);
736     b = gather(c->b, expand(B) & 0xff);
737     a = (1/65535.0f) * cast(expand(bswap(A)));
738 }
STAGE(load_tables_rgb_u16_be)739 STAGE(load_tables_rgb_u16_be) {
740     auto c = (const SkJumper_LoadTablesCtx*)ctx;
741     auto ptr = (const uint16_t*)c->src + 3*x;
742 
743     U16 R,G,B;
744     load3(ptr, tail, &R,&G,&B);
745 
746     // c->src is big-endian, so & 0xff grabs the 8 most signficant bits.
747     r = gather(c->r, expand(R) & 0xff);
748     g = gather(c->g, expand(G) & 0xff);
749     b = gather(c->b, expand(B) & 0xff);
750     a = 1.0f;
751 }
752 
STAGE(byte_tables)753 STAGE(byte_tables) {
754     struct Tables { const uint8_t *r, *g, *b, *a; };
755     auto tables = (const Tables*)ctx;
756 
757     r = from_byte(gather(tables->r, round(r, 255.0f)));
758     g = from_byte(gather(tables->g, round(g, 255.0f)));
759     b = from_byte(gather(tables->b, round(b, 255.0f)));
760     a = from_byte(gather(tables->a, round(a, 255.0f)));
761 }
762 
STAGE(byte_tables_rgb)763 STAGE(byte_tables_rgb) {
764     struct Tables { const uint8_t *r, *g, *b; int n; };
765     auto tables = (const Tables*)ctx;
766 
767     F scale = tables->n - 1;
768     r = from_byte(gather(tables->r, round(r, scale)));
769     g = from_byte(gather(tables->g, round(g, scale)));
770     b = from_byte(gather(tables->b, round(b, scale)));
771 }
772 
table(F v,const SkJumper_TableCtx * ctx)773 SI F table(F v, const SkJumper_TableCtx* ctx) {
774     return gather(ctx->table, round(v, ctx->size - 1));
775 }
STAGE(table_r)776 STAGE(table_r) { r = table(r, ctx); }
STAGE(table_g)777 STAGE(table_g) { g = table(g, ctx); }
STAGE(table_b)778 STAGE(table_b) { b = table(b, ctx); }
STAGE(table_a)779 STAGE(table_a) { a = table(a, ctx); }
780 
parametric(F v,const SkJumper_ParametricTransferFunction * ctx)781 SI F parametric(F v, const SkJumper_ParametricTransferFunction* ctx) {
782     F r = if_then_else(v <= ctx->D, mad(ctx->C, v, ctx->F)
783                                   , approx_powf(mad(ctx->A, v, ctx->B), ctx->G) + ctx->E);
784     return min(max(r, 0), 1.0f);  // Clamp to [0,1], with argument order mattering to handle NaN.
785 }
STAGE(parametric_r)786 STAGE(parametric_r) { r = parametric(r, ctx); }
STAGE(parametric_g)787 STAGE(parametric_g) { g = parametric(g, ctx); }
STAGE(parametric_b)788 STAGE(parametric_b) { b = parametric(b, ctx); }
STAGE(parametric_a)789 STAGE(parametric_a) { a = parametric(a, ctx); }
790 
STAGE(lab_to_xyz)791 STAGE(lab_to_xyz) {
792     F L = r * 100.0f,
793       A = g * 255.0f - 128.0f,
794       B = b * 255.0f - 128.0f;
795 
796     F Y = (L + 16.0f) * (1/116.0f),
797       X = Y + A*(1/500.0f),
798       Z = Y - B*(1/200.0f);
799 
800     X = if_then_else(X*X*X > 0.008856f, X*X*X, (X - (16/116.0f)) * (1/7.787f));
801     Y = if_then_else(Y*Y*Y > 0.008856f, Y*Y*Y, (Y - (16/116.0f)) * (1/7.787f));
802     Z = if_then_else(Z*Z*Z > 0.008856f, Z*Z*Z, (Z - (16/116.0f)) * (1/7.787f));
803 
804     // Adjust to D50 illuminant.
805     r = X * 0.96422f;
806     g = Y           ;
807     b = Z * 0.82521f;
808 }
809 
STAGE(load_a8)810 STAGE(load_a8) {
811     auto ptr = *(const uint8_t**)ctx + x;
812 
813     r = g = b = 0.0f;
814     a = from_byte(load<U8>(ptr, tail));
815 }
STAGE(load_a8_dst)816 STAGE(load_a8_dst) {
817     auto ptr = *(const uint8_t**)ctx + x;
818 
819     dr = dg = db = 0.0f;
820     da = from_byte(load<U8>(ptr, tail));
821 }
STAGE(gather_a8)822 STAGE(gather_a8) {
823     const uint8_t* ptr;
824     U32 ix = ix_and_ptr(&ptr, ctx, r,g);
825     r = g = b = 0.0f;
826     a = from_byte(gather(ptr, ix));
827 }
STAGE(store_a8)828 STAGE(store_a8) {
829     auto ptr = *(uint8_t**)ctx + x;
830 
831     U8 packed = pack(pack(round(a, 255.0f)));
832     store(ptr, packed, tail);
833 }
834 
STAGE(load_g8)835 STAGE(load_g8) {
836     auto ptr = *(const uint8_t**)ctx + x;
837 
838     r = g = b = from_byte(load<U8>(ptr, tail));
839     a = 1.0f;
840 }
STAGE(load_g8_dst)841 STAGE(load_g8_dst) {
842     auto ptr = *(const uint8_t**)ctx + x;
843 
844     dr = dg = db = from_byte(load<U8>(ptr, tail));
845     da = 1.0f;
846 }
STAGE(gather_g8)847 STAGE(gather_g8) {
848     const uint8_t* ptr;
849     U32 ix = ix_and_ptr(&ptr, ctx, r,g);
850     r = g = b = from_byte(gather(ptr, ix));
851     a = 1.0f;
852 }
853 
STAGE(load_565)854 STAGE(load_565) {
855     auto ptr = *(const uint16_t**)ctx + x;
856 
857     from_565(load<U16>(ptr, tail), &r,&g,&b);
858     a = 1.0f;
859 }
STAGE(load_565_dst)860 STAGE(load_565_dst) {
861     auto ptr = *(const uint16_t**)ctx + x;
862 
863     from_565(load<U16>(ptr, tail), &dr,&dg,&db);
864     da = 1.0f;
865 }
STAGE(gather_565)866 STAGE(gather_565) {
867     const uint16_t* ptr;
868     U32 ix = ix_and_ptr(&ptr, ctx, r,g);
869     from_565(gather(ptr, ix), &r,&g,&b);
870     a = 1.0f;
871 }
STAGE(store_565)872 STAGE(store_565) {
873     auto ptr = *(uint16_t**)ctx + x;
874 
875     U16 px = pack( round(r, 31.0f) << 11
876                  | round(g, 63.0f) <<  5
877                  | round(b, 31.0f)      );
878     store(ptr, px, tail);
879 }
880 
STAGE(load_4444)881 STAGE(load_4444) {
882     auto ptr = *(const uint16_t**)ctx + x;
883     from_4444(load<U16>(ptr, tail), &r,&g,&b,&a);
884 }
STAGE(load_4444_dst)885 STAGE(load_4444_dst) {
886     auto ptr = *(const uint16_t**)ctx + x;
887     from_4444(load<U16>(ptr, tail), &dr,&dg,&db,&da);
888 }
STAGE(gather_4444)889 STAGE(gather_4444) {
890     const uint16_t* ptr;
891     U32 ix = ix_and_ptr(&ptr, ctx, r,g);
892     from_4444(gather(ptr, ix), &r,&g,&b,&a);
893 }
STAGE(store_4444)894 STAGE(store_4444) {
895     auto ptr = *(uint16_t**)ctx + x;
896     U16 px = pack( round(r, 15.0f) << 12
897                  | round(g, 15.0f) <<  8
898                  | round(b, 15.0f) <<  4
899                  | round(a, 15.0f)      );
900     store(ptr, px, tail);
901 }
902 
STAGE(load_8888)903 STAGE(load_8888) {
904     auto ptr = *(const uint32_t**)ctx + x;
905     from_8888(load<U32>(ptr, tail), &r,&g,&b,&a);
906 }
STAGE(load_8888_dst)907 STAGE(load_8888_dst) {
908     auto ptr = *(const uint32_t**)ctx + x;
909     from_8888(load<U32>(ptr, tail), &dr,&dg,&db,&da);
910 }
STAGE(gather_8888)911 STAGE(gather_8888) {
912     const uint32_t* ptr;
913     U32 ix = ix_and_ptr(&ptr, ctx, r,g);
914     from_8888(gather(ptr, ix), &r,&g,&b,&a);
915 }
STAGE(store_8888)916 STAGE(store_8888) {
917     auto ptr = *(uint32_t**)ctx + x;
918 
919     U32 px = round(r, 255.0f)
920            | round(g, 255.0f) <<  8
921            | round(b, 255.0f) << 16
922            | round(a, 255.0f) << 24;
923     store(ptr, px, tail);
924 }
925 
STAGE(store_8888_2d)926 STAGE(store_8888_2d) {
927     auto c = (const SkJumper_MemoryCtx*)ctx;
928     auto ptr = (uint32_t*)c->pixels + y*c->stride + x;
929 
930     U32 px = round(r, 255.0f)
931            | round(g, 255.0f) <<  8
932            | round(b, 255.0f) << 16
933            | round(a, 255.0f) << 24;
934     store(ptr, px, tail);
935 }
936 
STAGE(load_bgra)937 STAGE(load_bgra) {
938     auto ptr = *(const uint32_t**)ctx + x;
939     from_8888(load<U32>(ptr, tail), &b,&g,&r,&a);
940 }
STAGE(load_bgra_dst)941 STAGE(load_bgra_dst) {
942     auto ptr = *(const uint32_t**)ctx + x;
943     from_8888(load<U32>(ptr, tail), &db,&dg,&dr,&da);
944 }
STAGE(gather_bgra)945 STAGE(gather_bgra) {
946     const uint32_t* ptr;
947     U32 ix = ix_and_ptr(&ptr, ctx, r,g);
948     from_8888(gather(ptr, ix), &b,&g,&r,&a);
949 }
STAGE(store_bgra)950 STAGE(store_bgra) {
951     auto ptr = *(uint32_t**)ctx + x;
952 
953     U32 px = round(b, 255.0f)
954            | round(g, 255.0f) <<  8
955            | round(r, 255.0f) << 16
956            | round(a, 255.0f) << 24;
957     store(ptr, px, tail);
958 }
959 
STAGE(load_f16)960 STAGE(load_f16) {
961     auto ptr = *(const uint64_t**)ctx + x;
962 
963     U16 R,G,B,A;
964     load4((const uint16_t*)ptr,tail, &R,&G,&B,&A);
965     r = from_half(R);
966     g = from_half(G);
967     b = from_half(B);
968     a = from_half(A);
969 }
STAGE(load_f16_dst)970 STAGE(load_f16_dst) {
971     auto ptr = *(const uint64_t**)ctx + x;
972 
973     U16 R,G,B,A;
974     load4((const uint16_t*)ptr,tail, &R,&G,&B,&A);
975     dr = from_half(R);
976     dg = from_half(G);
977     db = from_half(B);
978     da = from_half(A);
979 }
STAGE(gather_f16)980 STAGE(gather_f16) {
981     const uint64_t* ptr;
982     U32 ix = ix_and_ptr(&ptr, ctx, r,g);
983     auto px = gather(ptr, ix);
984 
985     U16 R,G,B,A;
986     load4((const uint16_t*)&px,0, &R,&G,&B,&A);
987     r = from_half(R);
988     g = from_half(G);
989     b = from_half(B);
990     a = from_half(A);
991 }
STAGE(store_f16)992 STAGE(store_f16) {
993     auto ptr = *(uint64_t**)ctx + x;
994     store4((uint16_t*)ptr,tail, to_half(r)
995                               , to_half(g)
996                               , to_half(b)
997                               , to_half(a));
998 }
999 
STAGE(load_u16_be)1000 STAGE(load_u16_be) {
1001     auto ptr = *(const uint16_t**)ctx + 4*x;
1002 
1003     U16 R,G,B,A;
1004     load4(ptr,tail, &R,&G,&B,&A);
1005 
1006     r = (1/65535.0f) * cast(expand(bswap(R)));
1007     g = (1/65535.0f) * cast(expand(bswap(G)));
1008     b = (1/65535.0f) * cast(expand(bswap(B)));
1009     a = (1/65535.0f) * cast(expand(bswap(A)));
1010 }
STAGE(load_rgb_u16_be)1011 STAGE(load_rgb_u16_be) {
1012     auto ptr = *(const uint16_t**)ctx + 3*x;
1013 
1014     U16 R,G,B;
1015     load3(ptr,tail, &R,&G,&B);
1016 
1017     r = (1/65535.0f) * cast(expand(bswap(R)));
1018     g = (1/65535.0f) * cast(expand(bswap(G)));
1019     b = (1/65535.0f) * cast(expand(bswap(B)));
1020     a = 1.0f;
1021 }
STAGE(store_u16_be)1022 STAGE(store_u16_be) {
1023     auto ptr = *(uint16_t**)ctx + 4*x;
1024 
1025     U16 R = bswap(pack(round(r, 65535.0f))),
1026         G = bswap(pack(round(g, 65535.0f))),
1027         B = bswap(pack(round(b, 65535.0f))),
1028         A = bswap(pack(round(a, 65535.0f)));
1029 
1030     store4(ptr,tail, R,G,B,A);
1031 }
1032 
STAGE(load_f32)1033 STAGE(load_f32) {
1034     auto ptr = *(const float**)ctx + 4*x;
1035     load4(ptr,tail, &r,&g,&b,&a);
1036 }
STAGE(load_f32_dst)1037 STAGE(load_f32_dst) {
1038     auto ptr = *(const float**)ctx + 4*x;
1039     load4(ptr,tail, &dr,&dg,&db,&da);
1040 }
STAGE(store_f32)1041 STAGE(store_f32) {
1042     auto ptr = *(float**)ctx + 4*x;
1043     store4(ptr,tail, r,g,b,a);
1044 }
1045 
ulp_before(F f)1046 SI F ulp_before(F f) {
1047     U32 bits = -1 + unaligned_load<U32>(&f);
1048     return unaligned_load<F>(&bits);
1049 }
1050 
exclusive_clamp(F v,const SkJumper_TileCtx * ctx)1051 SI F exclusive_clamp(F v, const SkJumper_TileCtx* ctx) {
1052     v = max(0,v);
1053     return min(v, ulp_before(ctx->scale));
1054 }
exclusive_repeat(F v,const SkJumper_TileCtx * ctx)1055 SI F exclusive_repeat(F v, const SkJumper_TileCtx* ctx) {
1056     v = v - floor_(v*ctx->invScale)*ctx->scale;
1057     return min(v, ulp_before(ctx->scale));
1058 }
exclusive_mirror(F v,const SkJumper_TileCtx * ctx)1059 SI F exclusive_mirror(F v, const SkJumper_TileCtx* ctx) {
1060     auto limit = ctx->scale;
1061     auto invLimit = ctx->invScale;
1062     v = abs_( (v-limit) - (limit+limit)*floor_((v-limit)*(invLimit*0.5f)) - limit );
1063     return min(v, ulp_before(limit));
1064 }
1065 // Clamp x or y to [0,limit) == [0,limit - 1 ulp] (think, sampling from images).
STAGE(clamp_x)1066 STAGE(clamp_x)  { r = exclusive_clamp (r, (const SkJumper_TileCtx*)ctx); }
STAGE(clamp_y)1067 STAGE(clamp_y)  { g = exclusive_clamp (g, (const SkJumper_TileCtx*)ctx); }
STAGE(repeat_x)1068 STAGE(repeat_x) { r = exclusive_repeat(r, (const SkJumper_TileCtx*)ctx); }
STAGE(repeat_y)1069 STAGE(repeat_y) { g = exclusive_repeat(g, (const SkJumper_TileCtx*)ctx); }
STAGE(mirror_x)1070 STAGE(mirror_x) { r = exclusive_mirror(r, (const SkJumper_TileCtx*)ctx); }
STAGE(mirror_y)1071 STAGE(mirror_y) { g = exclusive_mirror(g, (const SkJumper_TileCtx*)ctx); }
1072 
1073 // Clamp x to [0,1], both sides exclusive (think, gradients).
STAGE(clamp_x_1)1074 STAGE( clamp_x_1) { r = min(max(0, r), 1.0f); }
STAGE(repeat_x_1)1075 STAGE(repeat_x_1) { r = r - floor_(r); }
STAGE(mirror_x_1)1076 STAGE(mirror_x_1) { r = abs_( (r-1.0f) - two(floor_((r-1.0f)*0.5f)) - 1.0f ); }
1077 
STAGE(luminance_to_alpha)1078 STAGE(luminance_to_alpha) {
1079     a = r*0.2126f + g*0.7152f + b*0.0722f;
1080     r = g = b = 0;
1081 }
1082 
STAGE(matrix_translate)1083 STAGE(matrix_translate) {
1084     auto m = (const float*)ctx;
1085 
1086     r += m[0];
1087     g += m[1];
1088 }
STAGE(matrix_scale_translate)1089 STAGE(matrix_scale_translate) {
1090     auto m = (const float*)ctx;
1091 
1092     r = mad(r,m[2], m[0]);
1093     g = mad(g,m[3], m[1]);
1094 }
STAGE(matrix_2x3)1095 STAGE(matrix_2x3) {
1096     auto m = (const float*)ctx;
1097 
1098     auto R = mad(r,m[0], mad(g,m[2], m[4])),
1099          G = mad(r,m[1], mad(g,m[3], m[5]));
1100     r = R;
1101     g = G;
1102 }
STAGE(matrix_3x4)1103 STAGE(matrix_3x4) {
1104     auto m = (const float*)ctx;
1105 
1106     auto R = mad(r,m[0], mad(g,m[3], mad(b,m[6], m[ 9]))),
1107          G = mad(r,m[1], mad(g,m[4], mad(b,m[7], m[10]))),
1108          B = mad(r,m[2], mad(g,m[5], mad(b,m[8], m[11])));
1109     r = R;
1110     g = G;
1111     b = B;
1112 }
STAGE(matrix_4x5)1113 STAGE(matrix_4x5) {
1114     auto m = (const float*)ctx;
1115 
1116     auto R = mad(r,m[0], mad(g,m[4], mad(b,m[ 8], mad(a,m[12], m[16])))),
1117          G = mad(r,m[1], mad(g,m[5], mad(b,m[ 9], mad(a,m[13], m[17])))),
1118          B = mad(r,m[2], mad(g,m[6], mad(b,m[10], mad(a,m[14], m[18])))),
1119          A = mad(r,m[3], mad(g,m[7], mad(b,m[11], mad(a,m[15], m[19]))));
1120     r = R;
1121     g = G;
1122     b = B;
1123     a = A;
1124 }
STAGE(matrix_4x3)1125 STAGE(matrix_4x3) {
1126     auto m = (const float*)ctx;
1127     auto X = r,
1128          Y = g;
1129 
1130     r = mad(X, m[0], mad(Y, m[4], m[ 8]));
1131     g = mad(X, m[1], mad(Y, m[5], m[ 9]));
1132     b = mad(X, m[2], mad(Y, m[6], m[10]));
1133     a = mad(X, m[3], mad(Y, m[7], m[11]));
1134 }
STAGE(matrix_perspective)1135 STAGE(matrix_perspective) {
1136     // N.B. Unlike the other matrix_ stages, this matrix is row-major.
1137     auto m = (const float*)ctx;
1138 
1139     auto R = mad(r,m[0], mad(g,m[1], m[2])),
1140          G = mad(r,m[3], mad(g,m[4], m[5])),
1141          Z = mad(r,m[6], mad(g,m[7], m[8]));
1142     r = R * rcp(Z);
1143     g = G * rcp(Z);
1144 }
1145 
gradient_lookup(const SkJumper_GradientCtx * c,U32 idx,F t,F * r,F * g,F * b,F * a)1146 SI void gradient_lookup(const SkJumper_GradientCtx* c, U32 idx, F t,
1147                         F* r, F* g, F* b, F* a) {
1148     F fr, br, fg, bg, fb, bb, fa, ba;
1149 #if defined(JUMPER) && defined(__AVX2__)
1150     if (c->stopCount <=8) {
1151         fr = _mm256_permutevar8x32_ps(_mm256_loadu_ps(c->fs[0]), idx);
1152         br = _mm256_permutevar8x32_ps(_mm256_loadu_ps(c->bs[0]), idx);
1153         fg = _mm256_permutevar8x32_ps(_mm256_loadu_ps(c->fs[1]), idx);
1154         bg = _mm256_permutevar8x32_ps(_mm256_loadu_ps(c->bs[1]), idx);
1155         fb = _mm256_permutevar8x32_ps(_mm256_loadu_ps(c->fs[2]), idx);
1156         bb = _mm256_permutevar8x32_ps(_mm256_loadu_ps(c->bs[2]), idx);
1157         fa = _mm256_permutevar8x32_ps(_mm256_loadu_ps(c->fs[3]), idx);
1158         ba = _mm256_permutevar8x32_ps(_mm256_loadu_ps(c->bs[3]), idx);
1159     } else
1160 #endif
1161     {
1162         fr = gather(c->fs[0], idx);
1163         br = gather(c->bs[0], idx);
1164         fg = gather(c->fs[1], idx);
1165         bg = gather(c->bs[1], idx);
1166         fb = gather(c->fs[2], idx);
1167         bb = gather(c->bs[2], idx);
1168         fa = gather(c->fs[3], idx);
1169         ba = gather(c->bs[3], idx);
1170     }
1171 
1172     *r = mad(t, fr, br);
1173     *g = mad(t, fg, bg);
1174     *b = mad(t, fb, bb);
1175     *a = mad(t, fa, ba);
1176 }
1177 
STAGE(evenly_spaced_gradient)1178 STAGE(evenly_spaced_gradient) {
1179     auto c = (const SkJumper_GradientCtx*)ctx;
1180     auto t = r;
1181     auto idx = trunc_(t * (c->stopCount-1));
1182     gradient_lookup(c, idx, t, &r, &g, &b, &a);
1183 }
1184 
STAGE(gauss_a_to_rgba)1185 STAGE(gauss_a_to_rgba) {
1186     // x = 1 - x;
1187     // exp(-x * x * 4) - 0.018f;
1188     // ... now approximate with quartic
1189     //
1190     const float c4 = -2.26661229133605957031f;
1191     const float c3 = 2.89795351028442382812f;
1192     const float c2 = 0.21345567703247070312f;
1193     const float c1 = 0.15489584207534790039f;
1194     const float c0 = 0.00030726194381713867f;
1195     a = mad(a, mad(a, mad(a, mad(a, c4, c3), c2), c1), c0);
1196     r = a;
1197     g = a;
1198     b = a;
1199 }
1200 
STAGE(gradient)1201 STAGE(gradient) {
1202     auto c = (const SkJumper_GradientCtx*)ctx;
1203     auto t = r;
1204     U32 idx = 0;
1205 
1206     // N.B. The loop starts at 1 because idx 0 is the color to use before the first stop.
1207     for (size_t i = 1; i < c->stopCount; i++) {
1208         idx += if_then_else(t >= c->ts[i], U32(1), U32(0));
1209     }
1210 
1211     gradient_lookup(c, idx, t, &r, &g, &b, &a);
1212 }
1213 
STAGE(evenly_spaced_2_stop_gradient)1214 STAGE(evenly_spaced_2_stop_gradient) {
1215     struct Ctx { float f[4], b[4]; };
1216     auto c = (const Ctx*)ctx;
1217 
1218     auto t = r;
1219     r = mad(t, c->f[0], c->b[0]);
1220     g = mad(t, c->f[1], c->b[1]);
1221     b = mad(t, c->f[2], c->b[2]);
1222     a = mad(t, c->f[3], c->b[3]);
1223 }
1224 
STAGE(xy_to_unit_angle)1225 STAGE(xy_to_unit_angle) {
1226     F X = r,
1227       Y = g;
1228     F xabs = abs_(X),
1229       yabs = abs_(Y);
1230 
1231     F slope = min(xabs, yabs)/max(xabs, yabs);
1232     F s = slope * slope;
1233 
1234     // Use a 7th degree polynomial to approximate atan.
1235     // This was generated using sollya.gforge.inria.fr.
1236     // A float optimized polynomial was generated using the following command.
1237     // P1 = fpminimax((1/(2*Pi))*atan(x),[|1,3,5,7|],[|24...|],[2^(-40),1],relative);
1238     F phi = slope
1239              * (0.15912117063999176025390625f     + s
1240              * (-5.185396969318389892578125e-2f   + s
1241              * (2.476101927459239959716796875e-2f + s
1242              * (-7.0547382347285747528076171875e-3f))));
1243 
1244     phi = if_then_else(xabs < yabs, 1.0f/4.0f - phi, phi);
1245     phi = if_then_else(X < 0.0f   , 1.0f/2.0f - phi, phi);
1246     phi = if_then_else(Y < 0.0f   , 1.0f - phi     , phi);
1247     phi = if_then_else(phi != phi , 0              , phi);  // Check for NaN.
1248     r = phi;
1249 }
1250 
STAGE(xy_to_radius)1251 STAGE(xy_to_radius) {
1252     F X2 = r * r,
1253       Y2 = g * g;
1254     r = sqrt_(X2 + Y2);
1255 }
1256 
solve_2pt_conical_quadratic(const SkJumper_2PtConicalCtx * c,F x,F y,F (* select)(F,F))1257 SI F solve_2pt_conical_quadratic(const SkJumper_2PtConicalCtx* c, F x, F y, F (*select)(F, F)) {
1258     // At this point, (x, y) is mapped into a synthetic gradient space with
1259     // the start circle centerd on (0, 0), and the end circle centered on (1, 0)
1260     // (see the stage setup).
1261     //
1262     // We're searching along X-axis for x', such that
1263     //
1264     //   1) r(x') is a linear interpolation between r0 and r1
1265     //   2) (x, y) is on the circle C(x', 0, r(x'))
1266     //
1267     // Solving this system boils down to a quadratic equation with coefficients
1268     //
1269     //   a = 1 - (r1 - r0)^2             <- constant, precomputed in ctx->fCoeffA)
1270     //
1271     //   b = -2 * (x + (r1 - r0) * r0)
1272     //
1273     //   c = x^2 + y^2 - r0^2
1274     //
1275     // Since the start/end circle centers are the extremes of the [0, 1] interval
1276     // on the X axis, the solution (x') is exactly the t we are looking for.
1277 
1278     const F coeffA = c->fCoeffA,
1279             coeffB = -2 * (x + c->fDR*c->fR0),
1280             coeffC = x*x + y*y - c->fR0*c->fR0;
1281 
1282     const F disc      = mad(coeffB, coeffB, -4 * coeffA * coeffC);
1283     const F sqrt_disc = sqrt_(disc);
1284 
1285     const F invCoeffA = c->fInvCoeffA;
1286     return select((-coeffB + sqrt_disc) * (invCoeffA * 0.5f),
1287                   (-coeffB - sqrt_disc) * (invCoeffA * 0.5f));
1288 }
1289 
STAGE(xy_to_2pt_conical_quadratic_max)1290 STAGE(xy_to_2pt_conical_quadratic_max) {
1291     r = solve_2pt_conical_quadratic(ctx, r, g, max);
1292 }
1293 
STAGE(xy_to_2pt_conical_quadratic_min)1294 STAGE(xy_to_2pt_conical_quadratic_min) {
1295     r = solve_2pt_conical_quadratic(ctx, r, g, min);
1296 }
1297 
STAGE(xy_to_2pt_conical_linear)1298 STAGE(xy_to_2pt_conical_linear) {
1299     auto* c = (const SkJumper_2PtConicalCtx*)ctx;
1300 
1301     const F coeffB = -2 * (r + c->fDR*c->fR0),
1302             coeffC = r*r + g*g - c->fR0*c->fR0;
1303 
1304     r = -coeffC / coeffB;
1305 }
1306 
STAGE(mask_2pt_conical_degenerates)1307 STAGE(mask_2pt_conical_degenerates) {
1308     auto* c = (SkJumper_2PtConicalCtx*)ctx;
1309 
1310     // Compute and save a mask for degenerate values.
1311     U32 mask = 0xffffffff;
1312 
1313     // TODO: mtklein kindly volunteered to revisit this at some point.
1314 #if defined(JUMPER)
1315     // Vector comparisons set all bits, so we can use something like this.
1316     mask = mask & (mad(r, c->fDR, c->fR0) >= 0);  // R(t) >= 0
1317     mask = mask & (r == r);                       // t != NaN
1318 #else
1319     // The portable version is more involved, 'cause we only get one bit back.
1320     mask = mask & if_then_else(mad(r, c->fDR, c->fR0) >= 0, U32(0xffffffff), U32(0)); // R(t) >= 0
1321     mask = mask & if_then_else(r == r,                      U32(0xffffffff), U32(0)); // t != NaN
1322 #endif
1323 
1324     unaligned_store(&c->fMask, mask);
1325 }
1326 
STAGE(apply_vector_mask)1327 STAGE(apply_vector_mask) {
1328     const U32 mask = unaligned_load<U32>((const uint32_t*)ctx);
1329     r = bit_cast<F>(bit_cast<U32>(r) & mask);
1330     g = bit_cast<F>(bit_cast<U32>(g) & mask);
1331     b = bit_cast<F>(bit_cast<U32>(b) & mask);
1332     a = bit_cast<F>(bit_cast<U32>(a) & mask);
1333 }
1334 
STAGE(save_xy)1335 STAGE(save_xy) {
1336     auto c = (SkJumper_SamplerCtx*)ctx;
1337 
1338     // Whether bilinear or bicubic, all sample points are at the same fractional offset (fx,fy).
1339     // They're either the 4 corners of a logical 1x1 pixel or the 16 corners of a 3x3 grid
1340     // surrounding (x,y) at (0.5,0.5) off-center.
1341     F fx = fract(r + 0.5f),
1342       fy = fract(g + 0.5f);
1343 
1344     // Samplers will need to load x and fx, or y and fy.
1345     unaligned_store(c->x,  r);
1346     unaligned_store(c->y,  g);
1347     unaligned_store(c->fx, fx);
1348     unaligned_store(c->fy, fy);
1349 }
1350 
STAGE(accumulate)1351 STAGE(accumulate) {
1352     auto c = (const SkJumper_SamplerCtx*)ctx;
1353 
1354     // Bilinear and bicubic filters are both separable, so we produce independent contributions
1355     // from x and y, multiplying them together here to get each pixel's total scale factor.
1356     auto scale = unaligned_load<F>(c->scalex)
1357                * unaligned_load<F>(c->scaley);
1358     dr = mad(scale, r, dr);
1359     dg = mad(scale, g, dg);
1360     db = mad(scale, b, db);
1361     da = mad(scale, a, da);
1362 }
1363 
1364 // In bilinear interpolation, the 4 pixels at +/- 0.5 offsets from the sample pixel center
1365 // are combined in direct proportion to their area overlapping that logical query pixel.
1366 // At positive offsets, the x-axis contribution to that rectangle is fx, or (1-fx) at negative x.
1367 // The y-axis is symmetric.
1368 
1369 template <int kScale>
bilinear_x(SkJumper_SamplerCtx * ctx,F * x)1370 SI void bilinear_x(SkJumper_SamplerCtx* ctx, F* x) {
1371     *x = unaligned_load<F>(ctx->x) + (kScale * 0.5f);
1372     F fx = unaligned_load<F>(ctx->fx);
1373 
1374     F scalex;
1375     if (kScale == -1) { scalex = 1.0f - fx; }
1376     if (kScale == +1) { scalex =        fx; }
1377     unaligned_store(ctx->scalex, scalex);
1378 }
1379 template <int kScale>
bilinear_y(SkJumper_SamplerCtx * ctx,F * y)1380 SI void bilinear_y(SkJumper_SamplerCtx* ctx, F* y) {
1381     *y = unaligned_load<F>(ctx->y) + (kScale * 0.5f);
1382     F fy = unaligned_load<F>(ctx->fy);
1383 
1384     F scaley;
1385     if (kScale == -1) { scaley = 1.0f - fy; }
1386     if (kScale == +1) { scaley =        fy; }
1387     unaligned_store(ctx->scaley, scaley);
1388 }
1389 
STAGE(bilinear_nx)1390 STAGE(bilinear_nx) { bilinear_x<-1>(ctx, &r); }
STAGE(bilinear_px)1391 STAGE(bilinear_px) { bilinear_x<+1>(ctx, &r); }
STAGE(bilinear_ny)1392 STAGE(bilinear_ny) { bilinear_y<-1>(ctx, &g); }
STAGE(bilinear_py)1393 STAGE(bilinear_py) { bilinear_y<+1>(ctx, &g); }
1394 
1395 
1396 // In bicubic interpolation, the 16 pixels and +/- 0.5 and +/- 1.5 offsets from the sample
1397 // pixel center are combined with a non-uniform cubic filter, with higher values near the center.
1398 //
1399 // We break this function into two parts, one for near 0.5 offsets and one for far 1.5 offsets.
1400 // See GrCubicEffect for details of this particular filter.
1401 
bicubic_near(F t)1402 SI F bicubic_near(F t) {
1403     // 1/18 + 9/18t + 27/18t^2 - 21/18t^3 == t ( t ( -21/18t + 27/18) + 9/18) + 1/18
1404     return mad(t, mad(t, mad((-21/18.0f), t, (27/18.0f)), (9/18.0f)), (1/18.0f));
1405 }
bicubic_far(F t)1406 SI F bicubic_far(F t) {
1407     // 0/18 + 0/18*t - 6/18t^2 + 7/18t^3 == t^2 (7/18t - 6/18)
1408     return (t*t)*mad((7/18.0f), t, (-6/18.0f));
1409 }
1410 
1411 template <int kScale>
bicubic_x(SkJumper_SamplerCtx * ctx,F * x)1412 SI void bicubic_x(SkJumper_SamplerCtx* ctx, F* x) {
1413     *x = unaligned_load<F>(ctx->x) + (kScale * 0.5f);
1414     F fx = unaligned_load<F>(ctx->fx);
1415 
1416     F scalex;
1417     if (kScale == -3) { scalex = bicubic_far (1.0f - fx); }
1418     if (kScale == -1) { scalex = bicubic_near(1.0f - fx); }
1419     if (kScale == +1) { scalex = bicubic_near(       fx); }
1420     if (kScale == +3) { scalex = bicubic_far (       fx); }
1421     unaligned_store(ctx->scalex, scalex);
1422 }
1423 template <int kScale>
bicubic_y(SkJumper_SamplerCtx * ctx,F * y)1424 SI void bicubic_y(SkJumper_SamplerCtx* ctx, F* y) {
1425     *y = unaligned_load<F>(ctx->y) + (kScale * 0.5f);
1426     F fy = unaligned_load<F>(ctx->fy);
1427 
1428     F scaley;
1429     if (kScale == -3) { scaley = bicubic_far (1.0f - fy); }
1430     if (kScale == -1) { scaley = bicubic_near(1.0f - fy); }
1431     if (kScale == +1) { scaley = bicubic_near(       fy); }
1432     if (kScale == +3) { scaley = bicubic_far (       fy); }
1433     unaligned_store(ctx->scaley, scaley);
1434 }
1435 
STAGE(bicubic_n3x)1436 STAGE(bicubic_n3x) { bicubic_x<-3>(ctx, &r); }
STAGE(bicubic_n1x)1437 STAGE(bicubic_n1x) { bicubic_x<-1>(ctx, &r); }
STAGE(bicubic_p1x)1438 STAGE(bicubic_p1x) { bicubic_x<+1>(ctx, &r); }
STAGE(bicubic_p3x)1439 STAGE(bicubic_p3x) { bicubic_x<+3>(ctx, &r); }
1440 
STAGE(bicubic_n3y)1441 STAGE(bicubic_n3y) { bicubic_y<-3>(ctx, &g); }
STAGE(bicubic_n1y)1442 STAGE(bicubic_n1y) { bicubic_y<-1>(ctx, &g); }
STAGE(bicubic_p1y)1443 STAGE(bicubic_p1y) { bicubic_y<+1>(ctx, &g); }
STAGE(bicubic_p3y)1444 STAGE(bicubic_p3y) { bicubic_y<+3>(ctx, &g); }
1445 
STAGE(callback)1446 STAGE(callback) {
1447     auto c = (SkJumper_CallbackCtx*)ctx;
1448     store4(c->rgba,0, r,g,b,a);
1449     c->fn(c, tail ? tail : kStride);
1450     load4(c->read_from,0, &r,&g,&b,&a);
1451 }
1452