• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 // Intentionally NO #pragma once... included multiple times.
9 
10 // This file is included from skcms.cc in a namespace with some pre-defines:
11 //    - N:    depth of all vectors, 1,4,8, or 16 (preprocessor define)
12 //    - V<T>: a template to create a vector of N T's.
13 
14 using F   = V<Color>;   // Called F for historic reasons... maybe rename C?
15 using I32 = V<int32_t>;
16 using U64 = V<uint64_t>;
17 using U32 = V<uint32_t>;
18 using U16 = V<uint16_t>;
19 using U8  = V<uint8_t>;
20 
21 
22 #if defined(__GNUC__) && !defined(__clang__)
23     // Once again, GCC is kind of weird, not allowing vector = scalar directly.
24     static constexpr F F0 = F() + 0.0f,
25                        F1 = F() + 1.0f;
26 #else
27     static constexpr F F0 = 0.0f,
28                        F1 = 1.0f;
29 #endif
30 
31 // Instead of checking __AVX__ below, we'll check USING_AVX.
32 // This lets skcms.cc set USING_AVX to force us in even if the compiler's not set that way.
33 // Same deal for __F16C__ and __AVX2__ ~~~> USING_AVX_F16C, USING_AVX2.
34 
35 #if !defined(USING_AVX)      && N == 8 && defined(__AVX__)
36     #define  USING_AVX
37 #endif
38 #if !defined(USING_AVX_F16C) && defined(USING_AVX) && defined(__F16C__)
39     #define  USING AVX_F16C
40 #endif
41 #if !defined(USING_AVX2)     && defined(USING_AVX) && defined(__AVX2__)
42     #define  USING_AVX2
43 #endif
44 #if !defined(USING_AVX512F)  && N == 16 && defined(__AVX512F__)
45     #define  USING_AVX512F
46 #endif
47 
48 // Similar to the AVX+ features, we define USING_NEON and USING_NEON_F16C.
49 // This is more for organizational clarity... skcms.cc doesn't force these.
50 #if N > 1 && defined(__ARM_NEON)
51     #define USING_NEON
52     #if __ARM_FP & 2
53         #define USING_NEON_F16C
54     #endif
55     #if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(SKCMS_OPT_INTO_NEON_FP16)
56         #define USING_NEON_FP16
57     #endif
58 #endif
59 
60 // These -Wvector-conversion warnings seem to trigger in very bogus situations,
61 // like vst3q_f32() expecting a 16x char rather than a 4x float vector.  :/
62 #if defined(USING_NEON) && defined(__clang__)
63     #pragma clang diagnostic ignored "-Wvector-conversion"
64 #endif
65 
66 // GCC & Clang (but not clang-cl) warn returning U64 on x86 is larger than a register.
67 // You'd see warnings like, "using AVX even though AVX is not enabled".
68 // We stifle these warnings; our helpers that return U64 are always inlined.
69 #if defined(__SSE__) && defined(__GNUC__)
70     #pragma GCC diagnostic ignored "-Wpsabi"
71 #endif
72 
73 #if defined(__clang__)
74     #define FALLTHROUGH [[clang::fallthrough]]
75 #else
76     #define FALLTHROUGH
77 #endif
78 
79 // We tag most helper functions as SI, to enforce good code generation
80 // but also work around what we think is a bug in GCC: when targeting 32-bit
81 // x86, GCC tends to pass U16 (4x uint16_t vector) function arguments in the
82 // MMX mm0 register, which seems to mess with unrelated code that later uses
83 // x87 FP instructions (MMX's mm0 is an alias for x87's st0 register).
84 //
85 // It helps codegen to call __builtin_memcpy() when we know the byte count at compile time.
86 #if defined(__clang__) || defined(__GNUC__)
87     #define SI static inline __attribute__((always_inline))
88 #else
89     #define SI static inline
90 #endif
91 
92 template <typename T, typename P>
load(const P * ptr)93 SI T load(const P* ptr) {
94     T val;
95     small_memcpy(&val, ptr, sizeof(val));
96     return val;
97 }
98 template <typename T, typename P>
store(P * ptr,const T & val)99 SI void store(P* ptr, const T& val) {
100     small_memcpy(ptr, &val, sizeof(val));
101 }
102 
103 // (T)v is a cast when N == 1 and a bit-pun when N>1,
104 // so we use cast<T>(v) to actually cast or bit_pun<T>(v) to bit-pun.
105 template <typename D, typename S>
cast(const S & v)106 SI D cast(const S& v) {
107 #if N == 1
108     return (D)v;
109 #elif defined(__clang__)
110     return __builtin_convertvector(v, D);
111 #else
112     D d;
113     for (int i = 0; i < N; i++) {
114         d[i] = v[i];
115     }
116     return d;
117 #endif
118 }
119 
120 template <typename D, typename S>
bit_pun(const S & v)121 SI D bit_pun(const S& v) {
122     static_assert(sizeof(D) == sizeof(v), "");
123     return load<D>(&v);
124 }
125 
126 // When we convert from float to fixed point, it's very common to want to round,
127 // and for some reason compilers generate better code when converting to int32_t.
128 // To serve both those ends, we use this function to_fixed() instead of direct cast().
129 #if defined(USING_NEON_FP16)
130     // NEON's got a F16 -> U16 instruction, so this should be fine without going via I16.
to_fixed(F f)131     SI U16 to_fixed(F f) {  return cast<U16>(f + 0.5f); }
132 #else
to_fixed(F f)133     SI U32 to_fixed(F f) {  return (U32)cast<I32>(f + 0.5f); }
134 #endif
135 
136 
137 // Sometimes we do something crazy on one branch of a conditonal,
138 // like divide by zero or convert a huge float to an integer,
139 // but then harmlessly select the other side.  That trips up N==1
140 // sanitizer builds, so we make if_then_else() a macro to avoid
141 // evaluating the unused side.
142 
143 #if N == 1
144     #define if_then_else(cond, t, e) ((cond) ? (t) : (e))
145 #else
146     template <typename C, typename T>
if_then_else(C cond,T t,T e)147     SI T if_then_else(C cond, T t, T e) {
148         return bit_pun<T>( ( cond & bit_pun<C>(t)) |
149                            (~cond & bit_pun<C>(e)) );
150     }
151 #endif
152 
153 
F_from_Half(U16 half)154 SI F F_from_Half(U16 half) {
155 #if defined(USING_NEON_FP16)
156     return bit_pun<F>(half);
157 #elif defined(USING_NEON_F16C)
158     return vcvt_f32_f16((float16x4_t)half);
159 #elif defined(USING_AVX512F)
160     return (F)_mm512_cvtph_ps((__m256i)half);
161 #elif defined(USING_AVX_F16C)
162     typedef int16_t __attribute__((vector_size(16))) I16;
163     return __builtin_ia32_vcvtph2ps256((I16)half);
164 #else
165     U32 wide = cast<U32>(half);
166     // A half is 1-5-10 sign-exponent-mantissa, with 15 exponent bias.
167     U32 s  = wide & 0x8000,
168         em = wide ^ s;
169 
170     // Constructing the float is easy if the half is not denormalized.
171     F norm = bit_pun<F>( (s<<16) + (em<<13) + ((127-15)<<23) );
172 
173     // Simply flush all denorm half floats to zero.
174     return if_then_else(em < 0x0400, F0, norm);
175 #endif
176 }
177 
178 #if defined(__clang__)
179     // The -((127-15)<<10) underflows that side of the math when
180     // we pass a denorm half float.  It's harmless... we'll take the 0 side anyway.
181     __attribute__((no_sanitize("unsigned-integer-overflow")))
182 #endif
Half_from_F(F f)183 SI U16 Half_from_F(F f) {
184 #if defined(USING_NEON_FP16)
185     return bit_pun<U16>(f);
186 #elif defined(USING_NEON_F16C)
187     return (U16)vcvt_f16_f32(f);
188 #elif defined(USING_AVX512F)
189     return (U16)_mm512_cvtps_ph((__m512 )f, _MM_FROUND_CUR_DIRECTION );
190 #elif defined(USING_AVX_F16C)
191     return (U16)__builtin_ia32_vcvtps2ph256(f, 0x04/*_MM_FROUND_CUR_DIRECTION*/);
192 #else
193     // A float is 1-8-23 sign-exponent-mantissa, with 127 exponent bias.
194     U32 sem = bit_pun<U32>(f),
195         s   = sem & 0x80000000,
196          em = sem ^ s;
197 
198     // For simplicity we flush denorm half floats (including all denorm floats) to zero.
199     return cast<U16>(if_then_else(em < 0x38800000, (U32)F0
200                                                  , (s>>16) + (em>>13) - ((127-15)<<10)));
201 #endif
202 }
203 
204 // Swap high and low bytes of 16-bit lanes, converting between big-endian and little-endian.
205 #if defined(USING_NEON_FP16)
swap_endian_16(U16 v)206     SI U16 swap_endian_16(U16 v) {
207         return (U16)vrev16q_u8((uint8x16_t) v);
208     }
209 #elif defined(USING_NEON)
swap_endian_16(U16 v)210     SI U16 swap_endian_16(U16 v) {
211         return (U16)vrev16_u8((uint8x8_t) v);
212     }
213 #endif
214 
swap_endian_16x4(const U64 & rgba)215 SI U64 swap_endian_16x4(const U64& rgba) {
216     return (rgba & 0x00ff00ff00ff00ff) << 8
217          | (rgba & 0xff00ff00ff00ff00) >> 8;
218 }
219 
220 #if defined(USING_NEON_FP16)
min_(F x,F y)221     SI F min_(F x, F y) { return (F)vminq_f16((float16x8_t)x, (float16x8_t)y); }
max_(F x,F y)222     SI F max_(F x, F y) { return (F)vmaxq_f16((float16x8_t)x, (float16x8_t)y); }
223 #elif defined(USING_NEON)
min_(F x,F y)224     SI F min_(F x, F y) { return (F)vminq_f32((float32x4_t)x, (float32x4_t)y); }
max_(F x,F y)225     SI F max_(F x, F y) { return (F)vmaxq_f32((float32x4_t)x, (float32x4_t)y); }
226 #else
min_(F x,F y)227     SI F min_(F x, F y) { return if_then_else(x > y, y, x); }
max_(F x,F y)228     SI F max_(F x, F y) { return if_then_else(x < y, y, x); }
229 #endif
230 
floor_(F x)231 SI F floor_(F x) {
232 #if N == 1
233     return floorf_(x);
234 #elif defined(USING_NEON_FP16)
235     return vrndmq_f16(x);
236 #elif defined(__aarch64__)
237     return vrndmq_f32(x);
238 #elif defined(USING_AVX512F)
239     // Clang's _mm512_floor_ps() passes its mask as -1, not (__mmask16)-1,
240     // and integer santizer catches that this implicit cast changes the
241     // value from -1 to 65535.  We'll cast manually to work around it.
242     // Read this as `return _mm512_floor_ps(x)`.
243     return _mm512_mask_floor_ps(x, (__mmask16)-1, x);
244 #elif defined(USING_AVX)
245     return __builtin_ia32_roundps256(x, 0x01/*_MM_FROUND_FLOOR*/);
246 #elif defined(__SSE4_1__)
247     return _mm_floor_ps(x);
248 #else
249     // Round trip through integers with a truncating cast.
250     F roundtrip = cast<F>(cast<I32>(x));
251     // If x is negative, truncating gives the ceiling instead of the floor.
252     return roundtrip - if_then_else(roundtrip > x, F1, F0);
253 
254     // This implementation fails for values of x that are outside
255     // the range an integer can represent.  We expect most x to be small.
256 #endif
257 }
258 
approx_log2(F x)259 SI F approx_log2(F x) {
260 #if defined(USING_NEON_FP16)
261     // TODO(mtklein)
262     return x;
263 #else
264     // The first approximation of log2(x) is its exponent 'e', minus 127.
265     I32 bits = bit_pun<I32>(x);
266 
267     F e = cast<F>(bits) * (1.0f / (1<<23));
268 
269     // If we use the mantissa too we can refine the error signficantly.
270     F m = bit_pun<F>( (bits & 0x007fffff) | 0x3f000000 );
271 
272     return e - 124.225514990f
273              -   1.498030302f*m
274              -   1.725879990f/(0.3520887068f + m);
275 #endif
276 }
277 
approx_log(F x)278 SI F approx_log(F x) {
279     const float ln2 = 0.69314718f;
280     return ln2 * approx_log2(x);
281 }
282 
approx_exp2(F x)283 SI F approx_exp2(F x) {
284 #if defined(USING_NEON_FP16)
285     // TODO(mtklein)
286     return x;
287 #else
288     F fract = x - floor_(x);
289 
290     F fbits = (1.0f * (1<<23)) * (x + 121.274057500f
291                                     -   1.490129070f*fract
292                                     +  27.728023300f/(4.84252568f - fract));
293     I32 bits = cast<I32>(max_(fbits, F0));
294 
295     return bit_pun<F>(bits);
296 #endif
297 }
298 
approx_pow(F x,float y)299 SI F approx_pow(F x, float y) {
300     return if_then_else((x == F0) | (x == F1), x
301                                              , approx_exp2(approx_log2(x) * y));
302 }
303 
approx_exp(F x)304 SI F approx_exp(F x) {
305     const float log2_e = 1.4426950408889634074f;
306     return approx_exp2(log2_e * x);
307 }
308 
309 // Return tf(x).
apply_tf(const skcms_TransferFunction * tf,F x)310 SI F apply_tf(const skcms_TransferFunction* tf, F x) {
311 #if defined(USING_NEON_FP16)
312     // TODO(mtklein)
313     (void)tf;
314     return x;
315 #else
316     // Peel off the sign bit and set x = |x|.
317     U32 bits = bit_pun<U32>(x),
318         sign = bits & 0x80000000;
319     x = bit_pun<F>(bits ^ sign);
320 
321     // The transfer function has a linear part up to d, exponential at d and after.
322     F v = if_then_else(x < tf->d,            tf->c*x + tf->f
323                                 , approx_pow(tf->a*x + tf->b, tf->g) + tf->e);
324 
325     // Tack the sign bit back on.
326     return bit_pun<F>(sign | bit_pun<U32>(v));
327 #endif
328 }
329 
apply_pq(const skcms_TransferFunction * tf,F x)330 SI F apply_pq(const skcms_TransferFunction* tf, F x) {
331 #if defined(USING_NEON_FP16)
332     // TODO(mtklein)
333     (void)tf;
334     return x;
335 #else
336     U32 bits = bit_pun<U32>(x),
337         sign = bits & 0x80000000;
338     x = bit_pun<F>(bits ^ sign);
339 
340     F v = approx_pow(max_(tf->a + tf->b * approx_pow(x, tf->c), F0)
341                        / (tf->d + tf->e * approx_pow(x, tf->c)),
342                      tf->f);
343 
344     return bit_pun<F>(sign | bit_pun<U32>(v));
345 #endif
346 }
347 
apply_hlg(const skcms_TransferFunction * tf,F x)348 SI F apply_hlg(const skcms_TransferFunction* tf, F x) {
349 #if defined(USING_NEON_FP16)
350     // TODO(mtklein)
351     (void)tf;
352     return x;
353 #else
354     const float R = tf->a, G = tf->b,
355                 a = tf->c, b = tf->d, c = tf->e,
356                 K = tf->f + 1;
357     U32 bits = bit_pun<U32>(x),
358         sign = bits & 0x80000000;
359     x = bit_pun<F>(bits ^ sign);
360 
361     F v = if_then_else(x*R <= 1, approx_pow(x*R, G)
362                                , approx_exp((x-c)*a) + b);
363 
364     return K*bit_pun<F>(sign | bit_pun<U32>(v));
365 #endif
366 }
367 
apply_hlginv(const skcms_TransferFunction * tf,F x)368 SI F apply_hlginv(const skcms_TransferFunction* tf, F x) {
369 #if defined(USING_NEON_FP16)
370     // TODO(mtklein)
371     (void)tf;
372     return x;
373 #else
374     const float R = tf->a, G = tf->b,
375                 a = tf->c, b = tf->d, c = tf->e,
376                 K = tf->f + 1;
377     U32 bits = bit_pun<U32>(x),
378         sign = bits & 0x80000000;
379     x = bit_pun<F>(bits ^ sign);
380     x /= K;
381 
382     F v = if_then_else(x <= 1, R * approx_pow(x, G)
383                              , a * approx_log(x - b) + c);
384 
385     return bit_pun<F>(sign | bit_pun<U32>(v));
386 #endif
387 }
388 
389 
390 // Strided loads and stores of N values, starting from p.
391 template <typename T, typename P>
load_3(const P * p)392 SI T load_3(const P* p) {
393 #if N == 1
394     return (T)p[0];
395 #elif N == 4
396     return T{p[ 0],p[ 3],p[ 6],p[ 9]};
397 #elif N == 8
398     return T{p[ 0],p[ 3],p[ 6],p[ 9], p[12],p[15],p[18],p[21]};
399 #elif N == 16
400     return T{p[ 0],p[ 3],p[ 6],p[ 9], p[12],p[15],p[18],p[21],
401              p[24],p[27],p[30],p[33], p[36],p[39],p[42],p[45]};
402 #endif
403 }
404 
405 template <typename T, typename P>
load_4(const P * p)406 SI T load_4(const P* p) {
407 #if N == 1
408     return (T)p[0];
409 #elif N == 4
410     return T{p[ 0],p[ 4],p[ 8],p[12]};
411 #elif N == 8
412     return T{p[ 0],p[ 4],p[ 8],p[12], p[16],p[20],p[24],p[28]};
413 #elif N == 16
414     return T{p[ 0],p[ 4],p[ 8],p[12], p[16],p[20],p[24],p[28],
415              p[32],p[36],p[40],p[44], p[48],p[52],p[56],p[60]};
416 #endif
417 }
418 
419 template <typename T, typename P>
store_3(P * p,const T & v)420 SI void store_3(P* p, const T& v) {
421 #if N == 1
422     p[0] = v;
423 #elif N == 4
424     p[ 0] = v[ 0]; p[ 3] = v[ 1]; p[ 6] = v[ 2]; p[ 9] = v[ 3];
425 #elif N == 8
426     p[ 0] = v[ 0]; p[ 3] = v[ 1]; p[ 6] = v[ 2]; p[ 9] = v[ 3];
427     p[12] = v[ 4]; p[15] = v[ 5]; p[18] = v[ 6]; p[21] = v[ 7];
428 #elif N == 16
429     p[ 0] = v[ 0]; p[ 3] = v[ 1]; p[ 6] = v[ 2]; p[ 9] = v[ 3];
430     p[12] = v[ 4]; p[15] = v[ 5]; p[18] = v[ 6]; p[21] = v[ 7];
431     p[24] = v[ 8]; p[27] = v[ 9]; p[30] = v[10]; p[33] = v[11];
432     p[36] = v[12]; p[39] = v[13]; p[42] = v[14]; p[45] = v[15];
433 #endif
434 }
435 
436 template <typename T, typename P>
store_4(P * p,const T & v)437 SI void store_4(P* p, const T& v) {
438 #if N == 1
439     p[0] = v;
440 #elif N == 4
441     p[ 0] = v[ 0]; p[ 4] = v[ 1]; p[ 8] = v[ 2]; p[12] = v[ 3];
442 #elif N == 8
443     p[ 0] = v[ 0]; p[ 4] = v[ 1]; p[ 8] = v[ 2]; p[12] = v[ 3];
444     p[16] = v[ 4]; p[20] = v[ 5]; p[24] = v[ 6]; p[28] = v[ 7];
445 #elif N == 16
446     p[ 0] = v[ 0]; p[ 4] = v[ 1]; p[ 8] = v[ 2]; p[12] = v[ 3];
447     p[16] = v[ 4]; p[20] = v[ 5]; p[24] = v[ 6]; p[28] = v[ 7];
448     p[32] = v[ 8]; p[36] = v[ 9]; p[40] = v[10]; p[44] = v[11];
449     p[48] = v[12]; p[52] = v[13]; p[56] = v[14]; p[60] = v[15];
450 #endif
451 }
452 
453 
gather_8(const uint8_t * p,I32 ix)454 SI U8 gather_8(const uint8_t* p, I32 ix) {
455 #if N == 1
456     U8 v = p[ix];
457 #elif N == 4
458     U8 v = { p[ix[0]], p[ix[1]], p[ix[2]], p[ix[3]] };
459 #elif N == 8
460     U8 v = { p[ix[0]], p[ix[1]], p[ix[2]], p[ix[3]],
461              p[ix[4]], p[ix[5]], p[ix[6]], p[ix[7]] };
462 #elif N == 16
463     U8 v = { p[ix[ 0]], p[ix[ 1]], p[ix[ 2]], p[ix[ 3]],
464              p[ix[ 4]], p[ix[ 5]], p[ix[ 6]], p[ix[ 7]],
465              p[ix[ 8]], p[ix[ 9]], p[ix[10]], p[ix[11]],
466              p[ix[12]], p[ix[13]], p[ix[14]], p[ix[15]] };
467 #endif
468     return v;
469 }
470 
gather_16(const uint8_t * p,I32 ix)471 SI U16 gather_16(const uint8_t* p, I32 ix) {
472     // Load the i'th 16-bit value from p.
473     auto load_16 = [p](int i) {
474         return load<uint16_t>(p + 2*i);
475     };
476 #if N == 1
477     U16 v = load_16(ix);
478 #elif N == 4
479     U16 v = { load_16(ix[0]), load_16(ix[1]), load_16(ix[2]), load_16(ix[3]) };
480 #elif N == 8
481     U16 v = { load_16(ix[0]), load_16(ix[1]), load_16(ix[2]), load_16(ix[3]),
482               load_16(ix[4]), load_16(ix[5]), load_16(ix[6]), load_16(ix[7]) };
483 #elif N == 16
484     U16 v = { load_16(ix[ 0]), load_16(ix[ 1]), load_16(ix[ 2]), load_16(ix[ 3]),
485               load_16(ix[ 4]), load_16(ix[ 5]), load_16(ix[ 6]), load_16(ix[ 7]),
486               load_16(ix[ 8]), load_16(ix[ 9]), load_16(ix[10]), load_16(ix[11]),
487               load_16(ix[12]), load_16(ix[13]), load_16(ix[14]), load_16(ix[15]) };
488 #endif
489     return v;
490 }
491 
gather_32(const uint8_t * p,I32 ix)492 SI U32 gather_32(const uint8_t* p, I32 ix) {
493     // Load the i'th 32-bit value from p.
494     auto load_32 = [p](int i) {
495         return load<uint32_t>(p + 4*i);
496     };
497 #if N == 1
498     U32 v = load_32(ix);
499 #elif N == 4
500     U32 v = { load_32(ix[0]), load_32(ix[1]), load_32(ix[2]), load_32(ix[3]) };
501 #elif N == 8
502     U32 v = { load_32(ix[0]), load_32(ix[1]), load_32(ix[2]), load_32(ix[3]),
503               load_32(ix[4]), load_32(ix[5]), load_32(ix[6]), load_32(ix[7]) };
504 #elif N == 16
505     U32 v = { load_32(ix[ 0]), load_32(ix[ 1]), load_32(ix[ 2]), load_32(ix[ 3]),
506               load_32(ix[ 4]), load_32(ix[ 5]), load_32(ix[ 6]), load_32(ix[ 7]),
507               load_32(ix[ 8]), load_32(ix[ 9]), load_32(ix[10]), load_32(ix[11]),
508               load_32(ix[12]), load_32(ix[13]), load_32(ix[14]), load_32(ix[15]) };
509 #endif
510     // TODO: AVX2 and AVX-512 gathers (c.f. gather_24).
511     return v;
512 }
513 
gather_24(const uint8_t * p,I32 ix)514 SI U32 gather_24(const uint8_t* p, I32 ix) {
515     // First, back up a byte.  Any place we're gathering from has a safe junk byte to read
516     // in front of it, either a previous table value, or some tag metadata.
517     p -= 1;
518 
519     // Load the i'th 24-bit value from p, and 1 extra byte.
520     auto load_24_32 = [p](int i) {
521         return load<uint32_t>(p + 3*i);
522     };
523 
524     // Now load multiples of 4 bytes (a junk byte, then r,g,b).
525 #if N == 1
526     U32 v = load_24_32(ix);
527 #elif N == 4
528     U32 v = { load_24_32(ix[0]), load_24_32(ix[1]), load_24_32(ix[2]), load_24_32(ix[3]) };
529 #elif N == 8 && !defined(USING_AVX2)
530     U32 v = { load_24_32(ix[0]), load_24_32(ix[1]), load_24_32(ix[2]), load_24_32(ix[3]),
531               load_24_32(ix[4]), load_24_32(ix[5]), load_24_32(ix[6]), load_24_32(ix[7]) };
532 #elif N == 8
533     (void)load_24_32;
534     // The gather instruction here doesn't need any particular alignment,
535     // but the intrinsic takes a const int*.
536     const int* p4 = bit_pun<const int*>(p);
537     I32 zero = { 0, 0, 0, 0,  0, 0, 0, 0},
538         mask = {-1,-1,-1,-1, -1,-1,-1,-1};
539     #if defined(__clang__)
540         U32 v = (U32)__builtin_ia32_gatherd_d256(zero, p4, 3*ix, mask, 1);
541     #elif defined(__GNUC__)
542         U32 v = (U32)__builtin_ia32_gathersiv8si(zero, p4, 3*ix, mask, 1);
543     #endif
544 #elif N == 16
545     (void)load_24_32;
546     // The intrinsic is supposed to take const void* now, but it takes const int*, just like AVX2.
547     // And AVX-512 swapped the order of arguments.  :/
548     const int* p4 = bit_pun<const int*>(p);
549     U32 v = (U32)_mm512_i32gather_epi32((__m512i)(3*ix), p4, 1);
550 #endif
551 
552     // Shift off the junk byte, leaving r,g,b in low 24 bits (and zero in the top 8).
553     return v >> 8;
554 }
555 
556 #if !defined(__arm__)
gather_48(const uint8_t * p,I32 ix,U64 * v)557     SI void gather_48(const uint8_t* p, I32 ix, U64* v) {
558         // As in gather_24(), with everything doubled.
559         p -= 2;
560 
561         // Load the i'th 48-bit value from p, and 2 extra bytes.
562         auto load_48_64 = [p](int i) {
563             return load<uint64_t>(p + 6*i);
564         };
565 
566     #if N == 1
567         *v = load_48_64(ix);
568     #elif N == 4
569         *v = U64{
570             load_48_64(ix[0]), load_48_64(ix[1]), load_48_64(ix[2]), load_48_64(ix[3]),
571         };
572     #elif N == 8 && !defined(USING_AVX2)
573         *v = U64{
574             load_48_64(ix[0]), load_48_64(ix[1]), load_48_64(ix[2]), load_48_64(ix[3]),
575             load_48_64(ix[4]), load_48_64(ix[5]), load_48_64(ix[6]), load_48_64(ix[7]),
576         };
577     #elif N == 8
578         (void)load_48_64;
579         typedef int32_t   __attribute__((vector_size(16))) Half_I32;
580         typedef long long __attribute__((vector_size(32))) Half_I64;
581 
582         // The gather instruction here doesn't need any particular alignment,
583         // but the intrinsic takes a const long long*.
584         const long long int* p8 = bit_pun<const long long int*>(p);
585 
586         Half_I64 zero = { 0, 0, 0, 0},
587                  mask = {-1,-1,-1,-1};
588 
589         ix *= 6;
590         Half_I32 ix_lo = { ix[0], ix[1], ix[2], ix[3] },
591                  ix_hi = { ix[4], ix[5], ix[6], ix[7] };
592 
593         #if defined(__clang__)
594             Half_I64 lo = (Half_I64)__builtin_ia32_gatherd_q256(zero, p8, ix_lo, mask, 1),
595                      hi = (Half_I64)__builtin_ia32_gatherd_q256(zero, p8, ix_hi, mask, 1);
596         #elif defined(__GNUC__)
597             Half_I64 lo = (Half_I64)__builtin_ia32_gathersiv4di(zero, p8, ix_lo, mask, 1),
598                      hi = (Half_I64)__builtin_ia32_gathersiv4di(zero, p8, ix_hi, mask, 1);
599         #endif
600         store((char*)v +  0, lo);
601         store((char*)v + 32, hi);
602     #elif N == 16
603         (void)load_48_64;
604         const long long int* p8 = bit_pun<const long long int*>(p);
605         __m512i lo = _mm512_i32gather_epi64(_mm512_extracti32x8_epi32((__m512i)(6*ix), 0), p8, 1),
606                 hi = _mm512_i32gather_epi64(_mm512_extracti32x8_epi32((__m512i)(6*ix), 1), p8, 1);
607         store((char*)v +  0, lo);
608         store((char*)v + 64, hi);
609     #endif
610 
611         *v >>= 16;
612     }
613 #endif
614 
F_from_U8(U8 v)615 SI F F_from_U8(U8 v) {
616     return cast<F>(v) * (1/255.0f);
617 }
618 
F_from_U16_BE(U16 v)619 SI F F_from_U16_BE(U16 v) {
620     // All 16-bit ICC values are big-endian, so we byte swap before converting to float.
621     // MSVC catches the "loss" of data here in the portable path, so we also make sure to mask.
622     U16 lo = (v >> 8),
623         hi = (v << 8) & 0xffff;
624     return cast<F>(lo|hi) * (1/65535.0f);
625 }
626 
U16_from_F(F v)627 SI U16 U16_from_F(F v) {
628     // 65535 == inf in FP16, so promote to FP32 before converting.
629     return cast<U16>(cast<V<float>>(v) * 65535 + 0.5f);
630 }
631 
minus_1_ulp(F v)632 SI F minus_1_ulp(F v) {
633 #if defined(USING_NEON_FP16)
634     return bit_pun<F>( bit_pun<U16>(v) - 1 );
635 #else
636     return bit_pun<F>( bit_pun<U32>(v) - 1 );
637 #endif
638 }
639 
table(const skcms_Curve * curve,F v)640 SI F table(const skcms_Curve* curve, F v) {
641     // Clamp the input to [0,1], then scale to a table index.
642     F ix = max_(F0, min_(v, F1)) * (float)(curve->table_entries - 1);
643 
644     // We'll look up (equal or adjacent) entries at lo and hi, then lerp by t between the two.
645     I32 lo = cast<I32>(            ix      ),
646         hi = cast<I32>(minus_1_ulp(ix+1.0f));
647     F t = ix - cast<F>(lo);  // i.e. the fractional part of ix.
648 
649     // TODO: can we load l and h simultaneously?  Each entry in 'h' is either
650     // the same as in 'l' or adjacent.  We have a rough idea that's it'd always be safe
651     // to read adjacent entries and perhaps underflow the table by a byte or two
652     // (it'd be junk, but always safe to read).  Not sure how to lerp yet.
653     F l,h;
654     if (curve->table_8) {
655         l = F_from_U8(gather_8(curve->table_8, lo));
656         h = F_from_U8(gather_8(curve->table_8, hi));
657     } else {
658         l = F_from_U16_BE(gather_16(curve->table_16, lo));
659         h = F_from_U16_BE(gather_16(curve->table_16, hi));
660     }
661     return l + (h-l)*t;
662 }
663 
sample_clut_8(const uint8_t * grid_8,I32 ix,F * r,F * g,F * b)664 SI void sample_clut_8(const uint8_t* grid_8, I32 ix, F* r, F* g, F* b) {
665     U32 rgb = gather_24(grid_8, ix);
666 
667     *r = cast<F>((rgb >>  0) & 0xff) * (1/255.0f);
668     *g = cast<F>((rgb >>  8) & 0xff) * (1/255.0f);
669     *b = cast<F>((rgb >> 16) & 0xff) * (1/255.0f);
670 }
671 
sample_clut_8(const uint8_t * grid_8,I32 ix,F * r,F * g,F * b,F * a)672 SI void sample_clut_8(const uint8_t* grid_8, I32 ix, F* r, F* g, F* b, F* a) {
673     // TODO: don't forget to optimize gather_32().
674     U32 rgba = gather_32(grid_8, ix);
675 
676     *r = cast<F>((rgba >>  0) & 0xff) * (1/255.0f);
677     *g = cast<F>((rgba >>  8) & 0xff) * (1/255.0f);
678     *b = cast<F>((rgba >> 16) & 0xff) * (1/255.0f);
679     *a = cast<F>((rgba >> 24) & 0xff) * (1/255.0f);
680 }
681 
sample_clut_16(const uint8_t * grid_16,I32 ix,F * r,F * g,F * b)682 SI void sample_clut_16(const uint8_t* grid_16, I32 ix, F* r, F* g, F* b) {
683 #if defined(__arm__)
684     // This is up to 2x faster on 32-bit ARM than the #else-case fast path.
685     *r = F_from_U16_BE(gather_16(grid_16, 3*ix+0));
686     *g = F_from_U16_BE(gather_16(grid_16, 3*ix+1));
687     *b = F_from_U16_BE(gather_16(grid_16, 3*ix+2));
688 #else
689     // This strategy is much faster for 64-bit builds, and fine for 32-bit x86 too.
690     U64 rgb;
691     gather_48(grid_16, ix, &rgb);
692     rgb = swap_endian_16x4(rgb);
693 
694     *r = cast<F>((rgb >>  0) & 0xffff) * (1/65535.0f);
695     *g = cast<F>((rgb >> 16) & 0xffff) * (1/65535.0f);
696     *b = cast<F>((rgb >> 32) & 0xffff) * (1/65535.0f);
697 #endif
698 }
699 
sample_clut_16(const uint8_t * grid_16,I32 ix,F * r,F * g,F * b,F * a)700 SI void sample_clut_16(const uint8_t* grid_16, I32 ix, F* r, F* g, F* b, F* a) {
701     // TODO: gather_64()-based fast path?
702     *r = F_from_U16_BE(gather_16(grid_16, 4*ix+0));
703     *g = F_from_U16_BE(gather_16(grid_16, 4*ix+1));
704     *b = F_from_U16_BE(gather_16(grid_16, 4*ix+2));
705     *a = F_from_U16_BE(gather_16(grid_16, 4*ix+3));
706 }
707 
clut(uint32_t input_channels,uint32_t output_channels,const uint8_t grid_points[4],const uint8_t * grid_8,const uint8_t * grid_16,F * r,F * g,F * b,F * a)708 static void clut(uint32_t input_channels, uint32_t output_channels,
709                  const uint8_t grid_points[4], const uint8_t* grid_8, const uint8_t* grid_16,
710                  F* r, F* g, F* b, F* a) {
711 
712     const int dim = (int)input_channels;
713     assert (0 < dim && dim <= 4);
714     assert (output_channels == 3 ||
715             output_channels == 4);
716 
717     // For each of these arrays, think foo[2*dim], but we use foo[8] since we know dim <= 4.
718     I32 index [8];  // Index contribution by dimension, first low from 0, then high from 4.
719     F   weight[8];  // Weight for each contribution, again first low, then high.
720 
721     // O(dim) work first: calculate index,weight from r,g,b,a.
722     const F inputs[] = { *r,*g,*b,*a };
723     for (int i = dim-1, stride = 1; i >= 0; i--) {
724         // x is where we logically want to sample the grid in the i-th dimension.
725         F x = inputs[i] * (float)(grid_points[i] - 1);
726 
727         // But we can't index at floats.  lo and hi are the two integer grid points surrounding x.
728         I32 lo = cast<I32>(            x      ),   // i.e. trunc(x) == floor(x) here.
729             hi = cast<I32>(minus_1_ulp(x+1.0f));
730         // Notice how we fold in the accumulated stride across previous dimensions here.
731         index[i+0] = lo * stride;
732         index[i+4] = hi * stride;
733         stride *= grid_points[i];
734 
735         // We'll interpolate between those two integer grid points by t.
736         F t = x - cast<F>(lo);  // i.e. fract(x)
737         weight[i+0] = 1-t;
738         weight[i+4] = t;
739     }
740 
741     *r = *g = *b = F0;
742     if (output_channels == 4) {
743         *a = F0;
744     }
745 
746     // We'll sample 2^dim == 1<<dim table entries per pixel,
747     // in all combinations of low and high in each dimension.
748     for (int combo = 0; combo < (1<<dim); combo++) {  // This loop can be done in any order.
749 
750         // Each of these upcoming (combo&N)*K expressions here evaluates to 0 or 4,
751         // where 0 selects the low index contribution and its weight 1-t,
752         // or 4 the high index contribution and its weight t.
753 
754         // Since 0<dim≤4, we can always just start off with the 0-th channel,
755         // then handle the others conditionally.
756         I32 ix = index [0 + (combo&1)*4];
757         F    w = weight[0 + (combo&1)*4];
758 
759         switch ((dim-1)&3) {  // This lets the compiler know there are no other cases to handle.
760             case 3: ix += index [3 + (combo&8)/2];
761                     w  *= weight[3 + (combo&8)/2];
762                     FALLTHROUGH;
763                     // fall through
764 
765             case 2: ix += index [2 + (combo&4)*1];
766                     w  *= weight[2 + (combo&4)*1];
767                     FALLTHROUGH;
768                     // fall through
769 
770             case 1: ix += index [1 + (combo&2)*2];
771                     w  *= weight[1 + (combo&2)*2];
772         }
773 
774         F R,G,B,A=F0;
775         if (output_channels == 3) {
776             if (grid_8) { sample_clut_8 (grid_8 ,ix, &R,&G,&B); }
777             else        { sample_clut_16(grid_16,ix, &R,&G,&B); }
778         } else {
779             if (grid_8) { sample_clut_8 (grid_8 ,ix, &R,&G,&B,&A); }
780             else        { sample_clut_16(grid_16,ix, &R,&G,&B,&A); }
781         }
782         *r += w*R;
783         *g += w*G;
784         *b += w*B;
785         *a += w*A;
786     }
787 }
788 
clut(const skcms_A2B * a2b,F * r,F * g,F * b,F a)789 static void clut(const skcms_A2B* a2b, F* r, F* g, F* b, F a) {
790     clut(a2b->input_channels, a2b->output_channels,
791          a2b->grid_points, a2b->grid_8, a2b->grid_16,
792          r,g,b,&a);
793 }
clut(const skcms_B2A * b2a,F * r,F * g,F * b,F * a)794 static void clut(const skcms_B2A* b2a, F* r, F* g, F* b, F* a) {
795     clut(b2a->input_channels, b2a->output_channels,
796          b2a->grid_points, b2a->grid_8, b2a->grid_16,
797          r,g,b,a);
798 }
799 
exec_ops(const Op * ops,const void ** args,const char * src,char * dst,int i)800 static void exec_ops(const Op* ops, const void** args,
801                      const char* src, char* dst, int i) {
802     F r = F0, g = F0, b = F0, a = F1;
803     while (true) {
804         switch (*ops++) {
805             case Op_load_a8:{
806                 a = F_from_U8(load<U8>(src + 1*i));
807             } break;
808 
809             case Op_load_g8:{
810                 r = g = b = F_from_U8(load<U8>(src + 1*i));
811             } break;
812 
813             case Op_load_4444:{
814                 U16 abgr = load<U16>(src + 2*i);
815 
816                 r = cast<F>((abgr >> 12) & 0xf) * (1/15.0f);
817                 g = cast<F>((abgr >>  8) & 0xf) * (1/15.0f);
818                 b = cast<F>((abgr >>  4) & 0xf) * (1/15.0f);
819                 a = cast<F>((abgr >>  0) & 0xf) * (1/15.0f);
820             } break;
821 
822             case Op_load_565:{
823                 U16 rgb = load<U16>(src + 2*i);
824 
825                 r = cast<F>(rgb & (uint16_t)(31<< 0)) * (1.0f / (31<< 0));
826                 g = cast<F>(rgb & (uint16_t)(63<< 5)) * (1.0f / (63<< 5));
827                 b = cast<F>(rgb & (uint16_t)(31<<11)) * (1.0f / (31<<11));
828             } break;
829 
830             case Op_load_888:{
831                 const uint8_t* rgb = (const uint8_t*)(src + 3*i);
832             #if defined(USING_NEON_FP16)
833                 // See the explanation under USING_NEON below.  This is that doubled up.
834                 uint8x16x3_t v = {{ vdupq_n_u8(0), vdupq_n_u8(0), vdupq_n_u8(0) }};
835                 v = vld3q_lane_u8(rgb+ 0, v,  0);
836                 v = vld3q_lane_u8(rgb+ 3, v,  2);
837                 v = vld3q_lane_u8(rgb+ 6, v,  4);
838                 v = vld3q_lane_u8(rgb+ 9, v,  6);
839 
840                 v = vld3q_lane_u8(rgb+12, v,  8);
841                 v = vld3q_lane_u8(rgb+15, v, 10);
842                 v = vld3q_lane_u8(rgb+18, v, 12);
843                 v = vld3q_lane_u8(rgb+21, v, 14);
844 
845                 r = cast<F>((U16)v.val[0]) * (1/255.0f);
846                 g = cast<F>((U16)v.val[1]) * (1/255.0f);
847                 b = cast<F>((U16)v.val[2]) * (1/255.0f);
848             #elif defined(USING_NEON)
849                 // There's no uint8x4x3_t or vld3 load for it, so we'll load each rgb pixel one at
850                 // a time.  Since we're doing that, we might as well load them into 16-bit lanes.
851                 // (We'd even load into 32-bit lanes, but that's not possible on ARMv7.)
852                 uint8x8x3_t v = {{ vdup_n_u8(0), vdup_n_u8(0), vdup_n_u8(0) }};
853                 v = vld3_lane_u8(rgb+0, v, 0);
854                 v = vld3_lane_u8(rgb+3, v, 2);
855                 v = vld3_lane_u8(rgb+6, v, 4);
856                 v = vld3_lane_u8(rgb+9, v, 6);
857 
858                 // Now if we squint, those 3 uint8x8_t we constructed are really U16s, easy to
859                 // convert to F.  (Again, U32 would be even better here if drop ARMv7 or split
860                 // ARMv7 and ARMv8 impls.)
861                 r = cast<F>((U16)v.val[0]) * (1/255.0f);
862                 g = cast<F>((U16)v.val[1]) * (1/255.0f);
863                 b = cast<F>((U16)v.val[2]) * (1/255.0f);
864             #else
865                 r = cast<F>(load_3<U32>(rgb+0) ) * (1/255.0f);
866                 g = cast<F>(load_3<U32>(rgb+1) ) * (1/255.0f);
867                 b = cast<F>(load_3<U32>(rgb+2) ) * (1/255.0f);
868             #endif
869             } break;
870 
871             case Op_load_8888:{
872                 U32 rgba = load<U32>(src + 4*i);
873 
874                 r = cast<F>((rgba >>  0) & 0xff) * (1/255.0f);
875                 g = cast<F>((rgba >>  8) & 0xff) * (1/255.0f);
876                 b = cast<F>((rgba >> 16) & 0xff) * (1/255.0f);
877                 a = cast<F>((rgba >> 24) & 0xff) * (1/255.0f);
878             } break;
879 
880             case Op_load_8888_palette8:{
881                 const uint8_t* palette = (const uint8_t*) *args++;
882                 I32 ix = cast<I32>(load<U8>(src + 1*i));
883                 U32 rgba = gather_32(palette, ix);
884 
885                 r = cast<F>((rgba >>  0) & 0xff) * (1/255.0f);
886                 g = cast<F>((rgba >>  8) & 0xff) * (1/255.0f);
887                 b = cast<F>((rgba >> 16) & 0xff) * (1/255.0f);
888                 a = cast<F>((rgba >> 24) & 0xff) * (1/255.0f);
889             } break;
890 
891             case Op_load_1010102:{
892                 U32 rgba = load<U32>(src + 4*i);
893 
894                 r = cast<F>((rgba >>  0) & 0x3ff) * (1/1023.0f);
895                 g = cast<F>((rgba >> 10) & 0x3ff) * (1/1023.0f);
896                 b = cast<F>((rgba >> 20) & 0x3ff) * (1/1023.0f);
897                 a = cast<F>((rgba >> 30) & 0x3  ) * (1/   3.0f);
898             } break;
899 
900             case Op_load_161616LE:{
901                 uintptr_t ptr = (uintptr_t)(src + 6*i);
902                 assert( (ptr & 1) == 0 );                   // src must be 2-byte aligned for this
903                 const uint16_t* rgb = (const uint16_t*)ptr; // cast to const uint16_t* to be safe.
904             #if defined(USING_NEON_FP16)
905                 uint16x8x3_t v = vld3q_u16(rgb);
906                 r = cast<F>((U16)v.val[0]) * (1/65535.0f);
907                 g = cast<F>((U16)v.val[1]) * (1/65535.0f);
908                 b = cast<F>((U16)v.val[2]) * (1/65535.0f);
909             #elif defined(USING_NEON)
910                 uint16x4x3_t v = vld3_u16(rgb);
911                 r = cast<F>((U16)v.val[0]) * (1/65535.0f);
912                 g = cast<F>((U16)v.val[1]) * (1/65535.0f);
913                 b = cast<F>((U16)v.val[2]) * (1/65535.0f);
914             #else
915                 r = cast<F>(load_3<U32>(rgb+0)) * (1/65535.0f);
916                 g = cast<F>(load_3<U32>(rgb+1)) * (1/65535.0f);
917                 b = cast<F>(load_3<U32>(rgb+2)) * (1/65535.0f);
918             #endif
919             } break;
920 
921             case Op_load_16161616LE:{
922                 uintptr_t ptr = (uintptr_t)(src + 8*i);
923                 assert( (ptr & 1) == 0 );                    // src must be 2-byte aligned for this
924                 const uint16_t* rgba = (const uint16_t*)ptr; // cast to const uint16_t* to be safe.
925             #if defined(USING_NEON_FP16)
926                 uint16x8x4_t v = vld4q_u16(rgba);
927                 r = cast<F>((U16)v.val[0]) * (1/65535.0f);
928                 g = cast<F>((U16)v.val[1]) * (1/65535.0f);
929                 b = cast<F>((U16)v.val[2]) * (1/65535.0f);
930                 a = cast<F>((U16)v.val[3]) * (1/65535.0f);
931             #elif defined(USING_NEON)
932                 uint16x4x4_t v = vld4_u16(rgba);
933                 r = cast<F>((U16)v.val[0]) * (1/65535.0f);
934                 g = cast<F>((U16)v.val[1]) * (1/65535.0f);
935                 b = cast<F>((U16)v.val[2]) * (1/65535.0f);
936                 a = cast<F>((U16)v.val[3]) * (1/65535.0f);
937             #else
938                 U64 px = load<U64>(rgba);
939 
940                 r = cast<F>((px >>  0) & 0xffff) * (1/65535.0f);
941                 g = cast<F>((px >> 16) & 0xffff) * (1/65535.0f);
942                 b = cast<F>((px >> 32) & 0xffff) * (1/65535.0f);
943                 a = cast<F>((px >> 48) & 0xffff) * (1/65535.0f);
944             #endif
945             } break;
946 
947             case Op_load_161616BE:{
948                 uintptr_t ptr = (uintptr_t)(src + 6*i);
949                 assert( (ptr & 1) == 0 );                   // src must be 2-byte aligned for this
950                 const uint16_t* rgb = (const uint16_t*)ptr; // cast to const uint16_t* to be safe.
951             #if defined(USING_NEON_FP16)
952                 uint16x8x3_t v = vld3q_u16(rgb);
953                 r = cast<F>(swap_endian_16((U16)v.val[0])) * (1/65535.0f);
954                 g = cast<F>(swap_endian_16((U16)v.val[1])) * (1/65535.0f);
955                 b = cast<F>(swap_endian_16((U16)v.val[2])) * (1/65535.0f);
956             #elif defined(USING_NEON)
957                 uint16x4x3_t v = vld3_u16(rgb);
958                 r = cast<F>(swap_endian_16((U16)v.val[0])) * (1/65535.0f);
959                 g = cast<F>(swap_endian_16((U16)v.val[1])) * (1/65535.0f);
960                 b = cast<F>(swap_endian_16((U16)v.val[2])) * (1/65535.0f);
961             #else
962                 U32 R = load_3<U32>(rgb+0),
963                     G = load_3<U32>(rgb+1),
964                     B = load_3<U32>(rgb+2);
965                 // R,G,B are big-endian 16-bit, so byte swap them before converting to float.
966                 r = cast<F>((R & 0x00ff)<<8 | (R & 0xff00)>>8) * (1/65535.0f);
967                 g = cast<F>((G & 0x00ff)<<8 | (G & 0xff00)>>8) * (1/65535.0f);
968                 b = cast<F>((B & 0x00ff)<<8 | (B & 0xff00)>>8) * (1/65535.0f);
969             #endif
970             } break;
971 
972             case Op_load_16161616BE:{
973                 uintptr_t ptr = (uintptr_t)(src + 8*i);
974                 assert( (ptr & 1) == 0 );                    // src must be 2-byte aligned for this
975                 const uint16_t* rgba = (const uint16_t*)ptr; // cast to const uint16_t* to be safe.
976             #if defined(USING_NEON_FP16)
977                 uint16x8x4_t v = vld4q_u16(rgba);
978                 r = cast<F>(swap_endian_16((U16)v.val[0])) * (1/65535.0f);
979                 g = cast<F>(swap_endian_16((U16)v.val[1])) * (1/65535.0f);
980                 b = cast<F>(swap_endian_16((U16)v.val[2])) * (1/65535.0f);
981                 a = cast<F>(swap_endian_16((U16)v.val[3])) * (1/65535.0f);
982             #elif defined(USING_NEON)
983                 uint16x4x4_t v = vld4_u16(rgba);
984                 r = cast<F>(swap_endian_16((U16)v.val[0])) * (1/65535.0f);
985                 g = cast<F>(swap_endian_16((U16)v.val[1])) * (1/65535.0f);
986                 b = cast<F>(swap_endian_16((U16)v.val[2])) * (1/65535.0f);
987                 a = cast<F>(swap_endian_16((U16)v.val[3])) * (1/65535.0f);
988             #else
989                 U64 px = swap_endian_16x4(load<U64>(rgba));
990 
991                 r = cast<F>((px >>  0) & 0xffff) * (1/65535.0f);
992                 g = cast<F>((px >> 16) & 0xffff) * (1/65535.0f);
993                 b = cast<F>((px >> 32) & 0xffff) * (1/65535.0f);
994                 a = cast<F>((px >> 48) & 0xffff) * (1/65535.0f);
995             #endif
996             } break;
997 
998             case Op_load_hhh:{
999                 uintptr_t ptr = (uintptr_t)(src + 6*i);
1000                 assert( (ptr & 1) == 0 );                   // src must be 2-byte aligned for this
1001                 const uint16_t* rgb = (const uint16_t*)ptr; // cast to const uint16_t* to be safe.
1002             #if defined(USING_NEON_FP16)
1003                 uint16x8x3_t v = vld3q_u16(rgb);
1004                 U16 R = (U16)v.val[0],
1005                     G = (U16)v.val[1],
1006                     B = (U16)v.val[2];
1007             #elif defined(USING_NEON)
1008                 uint16x4x3_t v = vld3_u16(rgb);
1009                 U16 R = (U16)v.val[0],
1010                     G = (U16)v.val[1],
1011                     B = (U16)v.val[2];
1012             #else
1013                 U16 R = load_3<U16>(rgb+0),
1014                     G = load_3<U16>(rgb+1),
1015                     B = load_3<U16>(rgb+2);
1016             #endif
1017                 r = F_from_Half(R);
1018                 g = F_from_Half(G);
1019                 b = F_from_Half(B);
1020             } break;
1021 
1022             case Op_load_hhhh:{
1023                 uintptr_t ptr = (uintptr_t)(src + 8*i);
1024                 assert( (ptr & 1) == 0 );                    // src must be 2-byte aligned for this
1025                 const uint16_t* rgba = (const uint16_t*)ptr; // cast to const uint16_t* to be safe.
1026             #if defined(USING_NEON_FP16)
1027                 uint16x8x4_t v = vld4q_u16(rgba);
1028                 U16 R = (U16)v.val[0],
1029                     G = (U16)v.val[1],
1030                     B = (U16)v.val[2],
1031                     A = (U16)v.val[3];
1032             #elif defined(USING_NEON)
1033                 uint16x4x4_t v = vld4_u16(rgba);
1034                 U16 R = (U16)v.val[0],
1035                     G = (U16)v.val[1],
1036                     B = (U16)v.val[2],
1037                     A = (U16)v.val[3];
1038             #else
1039                 U64 px = load<U64>(rgba);
1040                 U16 R = cast<U16>((px >>  0) & 0xffff),
1041                     G = cast<U16>((px >> 16) & 0xffff),
1042                     B = cast<U16>((px >> 32) & 0xffff),
1043                     A = cast<U16>((px >> 48) & 0xffff);
1044             #endif
1045                 r = F_from_Half(R);
1046                 g = F_from_Half(G);
1047                 b = F_from_Half(B);
1048                 a = F_from_Half(A);
1049             } break;
1050 
1051             case Op_load_fff:{
1052                 uintptr_t ptr = (uintptr_t)(src + 12*i);
1053                 assert( (ptr & 3) == 0 );                   // src must be 4-byte aligned for this
1054                 const float* rgb = (const float*)ptr;       // cast to const float* to be safe.
1055             #if defined(USING_NEON_FP16)
1056                 float32x4x3_t lo = vld3q_f32(rgb +  0),
1057                               hi = vld3q_f32(rgb + 12);
1058                 r = (F)vcombine_f16(vcvt_f16_f32(lo.val[0]), vcvt_f16_f32(hi.val[0]));
1059                 g = (F)vcombine_f16(vcvt_f16_f32(lo.val[1]), vcvt_f16_f32(hi.val[1]));
1060                 b = (F)vcombine_f16(vcvt_f16_f32(lo.val[2]), vcvt_f16_f32(hi.val[2]));
1061             #elif defined(USING_NEON)
1062                 float32x4x3_t v = vld3q_f32(rgb);
1063                 r = (F)v.val[0];
1064                 g = (F)v.val[1];
1065                 b = (F)v.val[2];
1066             #else
1067                 r = load_3<F>(rgb+0);
1068                 g = load_3<F>(rgb+1);
1069                 b = load_3<F>(rgb+2);
1070             #endif
1071             } break;
1072 
1073             case Op_load_ffff:{
1074                 uintptr_t ptr = (uintptr_t)(src + 16*i);
1075                 assert( (ptr & 3) == 0 );                   // src must be 4-byte aligned for this
1076                 const float* rgba = (const float*)ptr;      // cast to const float* to be safe.
1077             #if defined(USING_NEON_FP16)
1078                 float32x4x4_t lo = vld4q_f32(rgba +  0),
1079                               hi = vld4q_f32(rgba + 16);
1080                 r = (F)vcombine_f16(vcvt_f16_f32(lo.val[0]), vcvt_f16_f32(hi.val[0]));
1081                 g = (F)vcombine_f16(vcvt_f16_f32(lo.val[1]), vcvt_f16_f32(hi.val[1]));
1082                 b = (F)vcombine_f16(vcvt_f16_f32(lo.val[2]), vcvt_f16_f32(hi.val[2]));
1083                 a = (F)vcombine_f16(vcvt_f16_f32(lo.val[3]), vcvt_f16_f32(hi.val[3]));
1084             #elif defined(USING_NEON)
1085                 float32x4x4_t v = vld4q_f32(rgba);
1086                 r = (F)v.val[0];
1087                 g = (F)v.val[1];
1088                 b = (F)v.val[2];
1089                 a = (F)v.val[3];
1090             #else
1091                 r = load_4<F>(rgba+0);
1092                 g = load_4<F>(rgba+1);
1093                 b = load_4<F>(rgba+2);
1094                 a = load_4<F>(rgba+3);
1095             #endif
1096             } break;
1097 
1098             case Op_swap_rb:{
1099                 F t = r;
1100                 r = b;
1101                 b = t;
1102             } break;
1103 
1104             case Op_clamp:{
1105                 r = max_(F0, min_(r, F1));
1106                 g = max_(F0, min_(g, F1));
1107                 b = max_(F0, min_(b, F1));
1108                 a = max_(F0, min_(a, F1));
1109             } break;
1110 
1111             case Op_invert:{
1112                 r = F1 - r;
1113                 g = F1 - g;
1114                 b = F1 - b;
1115                 a = F1 - a;
1116             } break;
1117 
1118             case Op_force_opaque:{
1119                 a = F1;
1120             } break;
1121 
1122             case Op_premul:{
1123                 r *= a;
1124                 g *= a;
1125                 b *= a;
1126             } break;
1127 
1128             case Op_unpremul:{
1129                 F scale = if_then_else(F1 / a < INFINITY_, F1 / a, F0);
1130                 r *= scale;
1131                 g *= scale;
1132                 b *= scale;
1133             } break;
1134 
1135             case Op_matrix_3x3:{
1136                 const skcms_Matrix3x3* matrix = (const skcms_Matrix3x3*) *args++;
1137                 const float* m = &matrix->vals[0][0];
1138 
1139                 F R = m[0]*r + m[1]*g + m[2]*b,
1140                   G = m[3]*r + m[4]*g + m[5]*b,
1141                   B = m[6]*r + m[7]*g + m[8]*b;
1142 
1143                 r = R;
1144                 g = G;
1145                 b = B;
1146             } break;
1147 
1148             case Op_matrix_3x4:{
1149                 const skcms_Matrix3x4* matrix = (const skcms_Matrix3x4*) *args++;
1150                 const float* m = &matrix->vals[0][0];
1151 
1152                 F R = m[0]*r + m[1]*g + m[ 2]*b + m[ 3],
1153                   G = m[4]*r + m[5]*g + m[ 6]*b + m[ 7],
1154                   B = m[8]*r + m[9]*g + m[10]*b + m[11];
1155 
1156                 r = R;
1157                 g = G;
1158                 b = B;
1159             } break;
1160 
1161             case Op_lab_to_xyz:{
1162                 // The L*a*b values are in r,g,b, but normalized to [0,1].  Reconstruct them:
1163                 F L = r * 100.0f,
1164                   A = g * 255.0f - 128.0f,
1165                   B = b * 255.0f - 128.0f;
1166 
1167                 // Convert to CIE XYZ.
1168                 F Y = (L + 16.0f) * (1/116.0f),
1169                   X = Y + A*(1/500.0f),
1170                   Z = Y - B*(1/200.0f);
1171 
1172                 X = if_then_else(X*X*X > 0.008856f, X*X*X, (X - (16/116.0f)) * (1/7.787f));
1173                 Y = if_then_else(Y*Y*Y > 0.008856f, Y*Y*Y, (Y - (16/116.0f)) * (1/7.787f));
1174                 Z = if_then_else(Z*Z*Z > 0.008856f, Z*Z*Z, (Z - (16/116.0f)) * (1/7.787f));
1175 
1176                 // Adjust to XYZD50 illuminant, and stuff back into r,g,b for the next op.
1177                 r = X * 0.9642f;
1178                 g = Y          ;
1179                 b = Z * 0.8249f;
1180             } break;
1181 
1182             // As above, in reverse.
1183             case Op_xyz_to_lab:{
1184                 F X = r * (1/0.9642f),
1185                   Y = g,
1186                   Z = b * (1/0.8249f);
1187 
1188                 X = if_then_else(X > 0.008856f, approx_pow(X, 1/3.0f), X*7.787f + (16/116.0f));
1189                 Y = if_then_else(Y > 0.008856f, approx_pow(Y, 1/3.0f), Y*7.787f + (16/116.0f));
1190                 Z = if_then_else(Z > 0.008856f, approx_pow(Z, 1/3.0f), Z*7.787f + (16/116.0f));
1191 
1192                 F L = Y*116.0f - 16.0f,
1193                   A = (X-Y)*500.0f,
1194                   B = (Y-Z)*200.0f;
1195 
1196                 r = L * (1/100.f);
1197                 g = (A + 128.0f) * (1/255.0f);
1198                 b = (B + 128.0f) * (1/255.0f);
1199             } break;
1200 
1201             case Op_tf_r:{ r = apply_tf((const skcms_TransferFunction*)*args++, r); } break;
1202             case Op_tf_g:{ g = apply_tf((const skcms_TransferFunction*)*args++, g); } break;
1203             case Op_tf_b:{ b = apply_tf((const skcms_TransferFunction*)*args++, b); } break;
1204             case Op_tf_a:{ a = apply_tf((const skcms_TransferFunction*)*args++, a); } break;
1205 
1206             case Op_pq_r:{ r = apply_pq((const skcms_TransferFunction*)*args++, r); } break;
1207             case Op_pq_g:{ g = apply_pq((const skcms_TransferFunction*)*args++, g); } break;
1208             case Op_pq_b:{ b = apply_pq((const skcms_TransferFunction*)*args++, b); } break;
1209             case Op_pq_a:{ a = apply_pq((const skcms_TransferFunction*)*args++, a); } break;
1210 
1211             case Op_hlg_r:{ r = apply_hlg((const skcms_TransferFunction*)*args++, r); } break;
1212             case Op_hlg_g:{ g = apply_hlg((const skcms_TransferFunction*)*args++, g); } break;
1213             case Op_hlg_b:{ b = apply_hlg((const skcms_TransferFunction*)*args++, b); } break;
1214             case Op_hlg_a:{ a = apply_hlg((const skcms_TransferFunction*)*args++, a); } break;
1215 
1216             case Op_hlginv_r:{ r = apply_hlginv((const skcms_TransferFunction*)*args++, r); } break;
1217             case Op_hlginv_g:{ g = apply_hlginv((const skcms_TransferFunction*)*args++, g); } break;
1218             case Op_hlginv_b:{ b = apply_hlginv((const skcms_TransferFunction*)*args++, b); } break;
1219             case Op_hlginv_a:{ a = apply_hlginv((const skcms_TransferFunction*)*args++, a); } break;
1220 
1221             case Op_table_r: { r = table((const skcms_Curve*)*args++, r); } break;
1222             case Op_table_g: { g = table((const skcms_Curve*)*args++, g); } break;
1223             case Op_table_b: { b = table((const skcms_Curve*)*args++, b); } break;
1224             case Op_table_a: { a = table((const skcms_Curve*)*args++, a); } break;
1225 
1226             case Op_clut_A2B: {
1227                 const skcms_A2B* a2b = (const skcms_A2B*) *args++;
1228                 clut(a2b, &r,&g,&b,a);
1229 
1230                 if (a2b->input_channels == 4) {
1231                     // CMYK is opaque.
1232                     a = F1;
1233                 }
1234             } break;
1235 
1236             case Op_clut_B2A: {
1237                 const skcms_B2A* b2a = (const skcms_B2A*) *args++;
1238                 clut(b2a, &r,&g,&b,&a);
1239             } break;
1240 
1241     // Notice, from here on down the store_ ops all return, ending the loop.
1242 
1243             case Op_store_a8: {
1244                 store(dst + 1*i, cast<U8>(to_fixed(a * 255)));
1245             } return;
1246 
1247             case Op_store_g8: {
1248                 // g should be holding luminance (Y) (r,g,b ~~~> X,Y,Z)
1249                 store(dst + 1*i, cast<U8>(to_fixed(g * 255)));
1250             } return;
1251 
1252             case Op_store_4444: {
1253                 store<U16>(dst + 2*i, cast<U16>(to_fixed(r * 15) << 12)
1254                                     | cast<U16>(to_fixed(g * 15) <<  8)
1255                                     | cast<U16>(to_fixed(b * 15) <<  4)
1256                                     | cast<U16>(to_fixed(a * 15) <<  0));
1257             } return;
1258 
1259             case Op_store_565: {
1260                 store<U16>(dst + 2*i, cast<U16>(to_fixed(r * 31) <<  0 )
1261                                     | cast<U16>(to_fixed(g * 63) <<  5 )
1262                                     | cast<U16>(to_fixed(b * 31) << 11 ));
1263             } return;
1264 
1265             case Op_store_888: {
1266                 uint8_t* rgb = (uint8_t*)dst + 3*i;
1267             #if defined(USING_NEON_FP16)
1268                 // See the explanation under USING_NEON below.  This is that doubled up.
1269                 U16 R = to_fixed(r * 255),
1270                     G = to_fixed(g * 255),
1271                     B = to_fixed(b * 255);
1272 
1273                 uint8x16x3_t v = {{ (uint8x16_t)R, (uint8x16_t)G, (uint8x16_t)B }};
1274                 vst3q_lane_u8(rgb+ 0, v,  0);
1275                 vst3q_lane_u8(rgb+ 3, v,  2);
1276                 vst3q_lane_u8(rgb+ 6, v,  4);
1277                 vst3q_lane_u8(rgb+ 9, v,  6);
1278 
1279                 vst3q_lane_u8(rgb+12, v,  8);
1280                 vst3q_lane_u8(rgb+15, v, 10);
1281                 vst3q_lane_u8(rgb+18, v, 12);
1282                 vst3q_lane_u8(rgb+21, v, 14);
1283             #elif defined(USING_NEON)
1284                 // Same deal as load_888 but in reverse... we'll store using uint8x8x3_t, but
1285                 // get there via U16 to save some instructions converting to float.  And just
1286                 // like load_888, we'd prefer to go via U32 but for ARMv7 support.
1287                 U16 R = cast<U16>(to_fixed(r * 255)),
1288                     G = cast<U16>(to_fixed(g * 255)),
1289                     B = cast<U16>(to_fixed(b * 255));
1290 
1291                 uint8x8x3_t v = {{ (uint8x8_t)R, (uint8x8_t)G, (uint8x8_t)B }};
1292                 vst3_lane_u8(rgb+0, v, 0);
1293                 vst3_lane_u8(rgb+3, v, 2);
1294                 vst3_lane_u8(rgb+6, v, 4);
1295                 vst3_lane_u8(rgb+9, v, 6);
1296             #else
1297                 store_3(rgb+0, cast<U8>(to_fixed(r * 255)) );
1298                 store_3(rgb+1, cast<U8>(to_fixed(g * 255)) );
1299                 store_3(rgb+2, cast<U8>(to_fixed(b * 255)) );
1300             #endif
1301             } return;
1302 
1303             case Op_store_8888: {
1304                 store(dst + 4*i, cast<U32>(to_fixed(r * 255)) <<  0
1305                                | cast<U32>(to_fixed(g * 255)) <<  8
1306                                | cast<U32>(to_fixed(b * 255)) << 16
1307                                | cast<U32>(to_fixed(a * 255)) << 24);
1308             } return;
1309 
1310             case Op_store_1010102: {
1311                 store(dst + 4*i, cast<U32>(to_fixed(r * 1023)) <<  0
1312                                | cast<U32>(to_fixed(g * 1023)) << 10
1313                                | cast<U32>(to_fixed(b * 1023)) << 20
1314                                | cast<U32>(to_fixed(a *    3)) << 30);
1315             } return;
1316 
1317             case Op_store_161616LE: {
1318                 uintptr_t ptr = (uintptr_t)(dst + 6*i);
1319                 assert( (ptr & 1) == 0 );                // The dst pointer must be 2-byte aligned
1320                 uint16_t* rgb = (uint16_t*)ptr;          // for this cast to uint16_t* to be safe.
1321             #if defined(USING_NEON_FP16)
1322                 uint16x8x3_t v = {{
1323                     (uint16x8_t)U16_from_F(r),
1324                     (uint16x8_t)U16_from_F(g),
1325                     (uint16x8_t)U16_from_F(b),
1326                 }};
1327                 vst3q_u16(rgb, v);
1328             #elif defined(USING_NEON)
1329                 uint16x4x3_t v = {{
1330                     (uint16x4_t)U16_from_F(r),
1331                     (uint16x4_t)U16_from_F(g),
1332                     (uint16x4_t)U16_from_F(b),
1333                 }};
1334                 vst3_u16(rgb, v);
1335             #else
1336                 store_3(rgb+0, U16_from_F(r));
1337                 store_3(rgb+1, U16_from_F(g));
1338                 store_3(rgb+2, U16_from_F(b));
1339             #endif
1340 
1341             } return;
1342 
1343             case Op_store_16161616LE: {
1344                 uintptr_t ptr = (uintptr_t)(dst + 8*i);
1345                 assert( (ptr & 1) == 0 );               // The dst pointer must be 2-byte aligned
1346                 uint16_t* rgba = (uint16_t*)ptr;        // for this cast to uint16_t* to be safe.
1347             #if defined(USING_NEON_FP16)
1348                 uint16x8x4_t v = {{
1349                     (uint16x8_t)U16_from_F(r),
1350                     (uint16x8_t)U16_from_F(g),
1351                     (uint16x8_t)U16_from_F(b),
1352                     (uint16x8_t)U16_from_F(a),
1353                 }};
1354                 vst4q_u16(rgba, v);
1355             #elif defined(USING_NEON)
1356                 uint16x4x4_t v = {{
1357                     (uint16x4_t)U16_from_F(r),
1358                     (uint16x4_t)U16_from_F(g),
1359                     (uint16x4_t)U16_from_F(b),
1360                     (uint16x4_t)U16_from_F(a),
1361                 }};
1362                 vst4_u16(rgba, v);
1363             #else
1364                 U64 px = cast<U64>(to_fixed(r * 65535)) <<  0
1365                        | cast<U64>(to_fixed(g * 65535)) << 16
1366                        | cast<U64>(to_fixed(b * 65535)) << 32
1367                        | cast<U64>(to_fixed(a * 65535)) << 48;
1368                 store(rgba, px);
1369             #endif
1370             } return;
1371 
1372             case Op_store_161616BE: {
1373                 uintptr_t ptr = (uintptr_t)(dst + 6*i);
1374                 assert( (ptr & 1) == 0 );                // The dst pointer must be 2-byte aligned
1375                 uint16_t* rgb = (uint16_t*)ptr;          // for this cast to uint16_t* to be safe.
1376             #if defined(USING_NEON_FP16)
1377                 uint16x8x3_t v = {{
1378                     (uint16x8_t)swap_endian_16(U16_from_F(r)),
1379                     (uint16x8_t)swap_endian_16(U16_from_F(g)),
1380                     (uint16x8_t)swap_endian_16(U16_from_F(b)),
1381                 }};
1382                 vst3q_u16(rgb, v);
1383             #elif defined(USING_NEON)
1384                 uint16x4x3_t v = {{
1385                     (uint16x4_t)swap_endian_16(cast<U16>(U16_from_F(r))),
1386                     (uint16x4_t)swap_endian_16(cast<U16>(U16_from_F(g))),
1387                     (uint16x4_t)swap_endian_16(cast<U16>(U16_from_F(b))),
1388                 }};
1389                 vst3_u16(rgb, v);
1390             #else
1391                 U32 R = to_fixed(r * 65535),
1392                     G = to_fixed(g * 65535),
1393                     B = to_fixed(b * 65535);
1394                 store_3(rgb+0, cast<U16>((R & 0x00ff) << 8 | (R & 0xff00) >> 8) );
1395                 store_3(rgb+1, cast<U16>((G & 0x00ff) << 8 | (G & 0xff00) >> 8) );
1396                 store_3(rgb+2, cast<U16>((B & 0x00ff) << 8 | (B & 0xff00) >> 8) );
1397             #endif
1398 
1399             } return;
1400 
1401             case Op_store_16161616BE: {
1402                 uintptr_t ptr = (uintptr_t)(dst + 8*i);
1403                 assert( (ptr & 1) == 0 );               // The dst pointer must be 2-byte aligned
1404                 uint16_t* rgba = (uint16_t*)ptr;        // for this cast to uint16_t* to be safe.
1405             #if defined(USING_NEON_FP16)
1406                 uint16x8x4_t v = {{
1407                     (uint16x8_t)swap_endian_16(U16_from_F(r)),
1408                     (uint16x8_t)swap_endian_16(U16_from_F(g)),
1409                     (uint16x8_t)swap_endian_16(U16_from_F(b)),
1410                     (uint16x8_t)swap_endian_16(U16_from_F(a)),
1411                 }};
1412                 vst4q_u16(rgba, v);
1413             #elif defined(USING_NEON)
1414                 uint16x4x4_t v = {{
1415                     (uint16x4_t)swap_endian_16(cast<U16>(U16_from_F(r))),
1416                     (uint16x4_t)swap_endian_16(cast<U16>(U16_from_F(g))),
1417                     (uint16x4_t)swap_endian_16(cast<U16>(U16_from_F(b))),
1418                     (uint16x4_t)swap_endian_16(cast<U16>(U16_from_F(a))),
1419                 }};
1420                 vst4_u16(rgba, v);
1421             #else
1422                 U64 px = cast<U64>(to_fixed(r * 65535)) <<  0
1423                        | cast<U64>(to_fixed(g * 65535)) << 16
1424                        | cast<U64>(to_fixed(b * 65535)) << 32
1425                        | cast<U64>(to_fixed(a * 65535)) << 48;
1426                 store(rgba, swap_endian_16x4(px));
1427             #endif
1428             } return;
1429 
1430             case Op_store_hhh: {
1431                 uintptr_t ptr = (uintptr_t)(dst + 6*i);
1432                 assert( (ptr & 1) == 0 );                // The dst pointer must be 2-byte aligned
1433                 uint16_t* rgb = (uint16_t*)ptr;          // for this cast to uint16_t* to be safe.
1434 
1435                 U16 R = Half_from_F(r),
1436                     G = Half_from_F(g),
1437                     B = Half_from_F(b);
1438             #if defined(USING_NEON_FP16)
1439                 uint16x8x3_t v = {{
1440                     (uint16x8_t)R,
1441                     (uint16x8_t)G,
1442                     (uint16x8_t)B,
1443                 }};
1444                 vst3q_u16(rgb, v);
1445             #elif defined(USING_NEON)
1446                 uint16x4x3_t v = {{
1447                     (uint16x4_t)R,
1448                     (uint16x4_t)G,
1449                     (uint16x4_t)B,
1450                 }};
1451                 vst3_u16(rgb, v);
1452             #else
1453                 store_3(rgb+0, R);
1454                 store_3(rgb+1, G);
1455                 store_3(rgb+2, B);
1456             #endif
1457             } return;
1458 
1459             case Op_store_hhhh: {
1460                 uintptr_t ptr = (uintptr_t)(dst + 8*i);
1461                 assert( (ptr & 1) == 0 );                // The dst pointer must be 2-byte aligned
1462                 uint16_t* rgba = (uint16_t*)ptr;         // for this cast to uint16_t* to be safe.
1463 
1464                 U16 R = Half_from_F(r),
1465                     G = Half_from_F(g),
1466                     B = Half_from_F(b),
1467                     A = Half_from_F(a);
1468             #if defined(USING_NEON_FP16)
1469                 uint16x8x4_t v = {{
1470                     (uint16x8_t)R,
1471                     (uint16x8_t)G,
1472                     (uint16x8_t)B,
1473                     (uint16x8_t)A,
1474                 }};
1475                 vst4q_u16(rgba, v);
1476             #elif defined(USING_NEON)
1477                 uint16x4x4_t v = {{
1478                     (uint16x4_t)R,
1479                     (uint16x4_t)G,
1480                     (uint16x4_t)B,
1481                     (uint16x4_t)A,
1482                 }};
1483                 vst4_u16(rgba, v);
1484             #else
1485                 store(rgba, cast<U64>(R) <<  0
1486                           | cast<U64>(G) << 16
1487                           | cast<U64>(B) << 32
1488                           | cast<U64>(A) << 48);
1489             #endif
1490 
1491             } return;
1492 
1493             case Op_store_fff: {
1494                 uintptr_t ptr = (uintptr_t)(dst + 12*i);
1495                 assert( (ptr & 3) == 0 );                // The dst pointer must be 4-byte aligned
1496                 float* rgb = (float*)ptr;                // for this cast to float* to be safe.
1497             #if defined(USING_NEON_FP16)
1498                 float32x4x3_t lo = {{
1499                     vcvt_f32_f16(vget_low_f16(r)),
1500                     vcvt_f32_f16(vget_low_f16(g)),
1501                     vcvt_f32_f16(vget_low_f16(b)),
1502                 }}, hi = {{
1503                     vcvt_f32_f16(vget_high_f16(r)),
1504                     vcvt_f32_f16(vget_high_f16(g)),
1505                     vcvt_f32_f16(vget_high_f16(b)),
1506                 }};
1507                 vst3q_f32(rgb +  0, lo);
1508                 vst3q_f32(rgb + 12, hi);
1509             #elif defined(USING_NEON)
1510                 float32x4x3_t v = {{
1511                     (float32x4_t)r,
1512                     (float32x4_t)g,
1513                     (float32x4_t)b,
1514                 }};
1515                 vst3q_f32(rgb, v);
1516             #else
1517                 store_3(rgb+0, r);
1518                 store_3(rgb+1, g);
1519                 store_3(rgb+2, b);
1520             #endif
1521             } return;
1522 
1523             case Op_store_ffff: {
1524                 uintptr_t ptr = (uintptr_t)(dst + 16*i);
1525                 assert( (ptr & 3) == 0 );                // The dst pointer must be 4-byte aligned
1526                 float* rgba = (float*)ptr;               // for this cast to float* to be safe.
1527             #if defined(USING_NEON_FP16)
1528                 float32x4x4_t lo = {{
1529                     vcvt_f32_f16(vget_low_f16(r)),
1530                     vcvt_f32_f16(vget_low_f16(g)),
1531                     vcvt_f32_f16(vget_low_f16(b)),
1532                     vcvt_f32_f16(vget_low_f16(a)),
1533                 }}, hi = {{
1534                     vcvt_f32_f16(vget_high_f16(r)),
1535                     vcvt_f32_f16(vget_high_f16(g)),
1536                     vcvt_f32_f16(vget_high_f16(b)),
1537                     vcvt_f32_f16(vget_high_f16(a)),
1538                 }};
1539                 vst4q_f32(rgba +  0, lo);
1540                 vst4q_f32(rgba + 16, hi);
1541             #elif defined(USING_NEON)
1542                 float32x4x4_t v = {{
1543                     (float32x4_t)r,
1544                     (float32x4_t)g,
1545                     (float32x4_t)b,
1546                     (float32x4_t)a,
1547                 }};
1548                 vst4q_f32(rgba, v);
1549             #else
1550                 store_4(rgba+0, r);
1551                 store_4(rgba+1, g);
1552                 store_4(rgba+2, b);
1553                 store_4(rgba+3, a);
1554             #endif
1555             } return;
1556         }
1557     }
1558 }
1559 
1560 
run_program(const Op * program,const void ** arguments,const char * src,char * dst,int n,const size_t src_bpp,const size_t dst_bpp)1561 static void run_program(const Op* program, const void** arguments,
1562                         const char* src, char* dst, int n,
1563                         const size_t src_bpp, const size_t dst_bpp) {
1564     int i = 0;
1565     while (n >= N) {
1566         exec_ops(program, arguments, src, dst, i);
1567         i += N;
1568         n -= N;
1569     }
1570     if (n > 0) {
1571         char tmp[4*4*N] = {0};
1572 
1573         memcpy(tmp, (const char*)src + (size_t)i*src_bpp, (size_t)n*src_bpp);
1574         exec_ops(program, arguments, tmp, tmp, 0);
1575         memcpy((char*)dst + (size_t)i*dst_bpp, tmp, (size_t)n*dst_bpp);
1576     }
1577 }
1578 
1579 // Clean up any #defines we may have set so that we can be #included again.
1580 #if defined(USING_AVX)
1581     #undef  USING_AVX
1582 #endif
1583 #if defined(USING_AVX_F16C)
1584     #undef  USING_AVX_F16C
1585 #endif
1586 #if defined(USING_AVX2)
1587     #undef  USING_AVX2
1588 #endif
1589 #if defined(USING_AVX512F)
1590     #undef  USING_AVX512F
1591 #endif
1592 
1593 #if defined(USING_NEON)
1594     #undef  USING_NEON
1595 #endif
1596 #if defined(USING_NEON_F16C)
1597     #undef  USING_NEON_F16C
1598 #endif
1599 #if defined(USING_NEON_FP16)
1600     #undef  USING_NEON_FP16
1601 #endif
1602 
1603 #undef FALLTHROUGH
1604