xref: /aosp_15_r20/external/skia/src/sksl/sksl_gpu.sksl (revision c8dee2aa9b3f27cf6c858bd81872bdeb2c07ed17)
1// Not exposed in shared module
2
3$pure $genIType mix($genIType x, $genIType y, $genBType a);
4$pure $genBType mix($genBType x, $genBType y, $genBType a);
5$pure $genType fma($genType a, $genType b, $genType c);
6$pure $genHType fma($genHType a, $genHType b, $genHType c);
7      $genType frexp($genType x, out $genIType exp);
8      $genHType frexp($genHType x, out $genIType exp);
9$pure $genType ldexp($genType x, in $genIType exp);
10$pure $genHType ldexp($genHType x, in $genIType exp);
11
12$pure uint packSnorm2x16(float2 v);
13$pure uint packUnorm4x8(float4 v);
14$pure uint packSnorm4x8(float4 v);
15$pure float2 unpackSnorm2x16(uint p);
16$pure float4 unpackUnorm4x8(uint p);
17$pure float4 unpackSnorm4x8(uint p);
18$pure uint packHalf2x16(float2 v);
19$pure float2 unpackHalf2x16(uint v);
20
21$pure $genIType bitCount($genIType value);
22$pure $genIType bitCount($genUType value);
23$pure $genIType findLSB($genIType value);
24$pure $genIType findLSB($genUType value);
25$pure $genIType findMSB($genIType value);
26$pure $genIType findMSB($genUType value);
27
28$pure half4 sample(sampler2D s, float2 P);
29$pure half4 sample(sampler2D s, float3 P);
30$pure half4 sample(sampler2D s, float3 P, float bias);
31
32$pure half4 sample(samplerExternalOES s, float2 P);
33$pure half4 sample(samplerExternalOES s, float2 P, float bias);
34
35$pure half4 sample(sampler2DRect s, float2 P);
36$pure half4 sample(sampler2DRect s, float3 P);
37
38$pure half4 sampleLod(sampler2D s, float2 P, float lod);
39$pure half4 sampleLod(sampler2D s, float3 P, float lod);
40
41$pure half4 sampleGrad(sampler2D s, float2, float2 dPdx, float2 dPdy);
42
43// Currently we do not support the generic types of loading subpassInput so we have some explicit
44// versions that we currently use
45$pure half4 subpassLoad(subpassInput subpass);
46$pure half4 subpassLoad(subpassInputMS subpass, int sample);
47
48/** Atomically loads the value from `a` and returns it. */
49$pure uint atomicLoad(atomicUint a);
50
51/** Atomically stores the value of `value` to `a` */
52void atomicStore(atomicUint a, uint value);
53
54/**
55 * Performs an atomic addition of `value` to the contents of `a` and returns the original contents
56 * of `a` from before the addition occurred.
57 */
58uint atomicAdd(atomicUint a, uint value);
59
60// Definitions of functions implementing all of the SkBlendMode blends.
61
62$pure half4 blend_clear(half4 src, half4 dst) { return half4(0); }
63
64$pure half4 blend_src(half4 src, half4 dst) { return src; }
65
66$pure half4 blend_dst(half4 src, half4 dst) { return dst; }
67
68$pure half4 blend_src_over(half4 src, half4 dst) { return src + (1 - src.a)*dst; }
69
70$pure half4 blend_dst_over(half4 src, half4 dst) { return (1 - dst.a)*src + dst; }
71
72$pure half4 blend_src_in(half4 src, half4 dst) { return src*dst.a; }
73
74$pure half4 blend_dst_in(half4 src, half4 dst) { return dst*src.a; }
75
76$pure half4 blend_src_out(half4 src, half4 dst) { return (1 - dst.a)*src; }
77
78$pure half4 blend_dst_out(half4 src, half4 dst) { return (1 - src.a)*dst; }
79
80$pure half4 blend_src_atop(half4 src, half4 dst) { return dst.a*src + (1 - src.a)*dst; }
81
82$pure half4 blend_dst_atop(half4 src, half4 dst)  { return  (1 - dst.a) * src + src.a*dst; }
83
84$pure half4 blend_xor(half4 src, half4 dst) { return (1 - dst.a)*src + (1 - src.a)*dst; }
85
86// This multi-purpose Porter-Duff blend function can perform any of the twelve blends above,
87// when passed one of the following values for BlendOp:
88// - Clear:          0*src +        0*dst = (0 +  0*dstA)*src + (0 +  0*srcA)*dst = (0,  0,  0,  0)
89// - Src:            1*src +        0*dst = (1 +  0*dstA)*src + (0 +  0*srcA)*dst = (1,  0,  0,  0)
90// - Dst:            0*src +        1*dst = (0 +  0*dstA)*src + (1 +  0*srcA)*dst = (0,  1,  0,  0)
91// - SrcOver:        1*src + (1-srcA)*dst = (1 +  0*dstA)*src + (1 + -1*srcA)*dst = (1,  1,  0, -1)
92// - DstOver: (1-dstA)*src +        1*dst = (1 + -1*dstA)*src + (1 +  0*srcA)*dst = (1,  1, -1,  0)
93// - SrcIn:       dstA*src +        0*dst = (0 +  1*dstA)*src + (0 +  0*srcA)*dst = (0,  0,  1,  0)
94// - DstIn:          0*src +     srcA*dst = (0 +  0*dstA)*src + (0 +  1*srcA)*dst = (0,  0,  0,  1)
95// - SrcOut:  (1-dstA)*src +        0*dst = (1 + -1*dstA)*src + (0 +  0*srcA)*dst = (1,  0, -1,  0)
96// - DstOut:         0*src + (1-srcA)*dst = (0 +  0*dstA)*src + (1 + -1*srcA)*dst = (0,  1,  0, -1)
97// - SrcATop:     dstA*src + (1-srcA)*dst = (0 +  1*dstA)*src + (1 + -1*srcA)*dst = (0,  1,  1, -1)
98// - DstATop: (1-dstA)*src +     srcA*dst = (1 + -1*dstA)*src + (0 +  1*srcA)*dst = (1,  0, -1,  1)
99// - Xor:     (1-dstA)*src + (1-srcA)*dst = (1 + -1*dstA)*src + (1 + -1*srcA)*dst = (1,  1, -1, -1)
100$pure half4 blend_porter_duff(half4 blendOp, half4 src, half4 dst) {
101    // The supported blend modes all have coefficients that are of the form (C + S*alpha), where
102    // alpha is the other color's alpha channel. C can be 0 or 1, S can be -1, 0, or 1.
103    half2 coeff = blendOp.xy + blendOp.zw * half2(dst.a, src.a);
104    return src * coeff.x + dst * coeff.y;
105}
106
107$pure half4 blend_plus(half4 src, half4 dst) { return min(src + dst, 1); }
108
109$pure half4 blend_modulate(half4 src, half4 dst) { return src*dst; }
110
111$pure half4 blend_screen(half4 src, half4 dst) { return src + (1 - src)*dst; }
112
113$pure half $blend_overlay_component(half2 s, half2 d) {
114    return (2*d.x <= d.y) ? 2*s.x*d.x
115                          : s.y*d.y - 2*(d.y - d.x)*(s.y - s.x);
116}
117
118$pure half4 blend_overlay(half4 src, half4 dst) {
119    half4 result = half4($blend_overlay_component(src.ra, dst.ra),
120                         $blend_overlay_component(src.ga, dst.ga),
121                         $blend_overlay_component(src.ba, dst.ba),
122                         src.a + (1 - src.a)*dst.a);
123    result.rgb += dst.rgb*(1 - src.a) + src.rgb*(1 - dst.a);
124    return result;
125}
126
127$pure half4 blend_overlay(half flip, half4 a, half4 b) {
128    return blend_overlay(bool(flip) ? b : a, bool(flip) ? a : b);
129}
130
131$pure half4 blend_lighten(half4 src, half4 dst) {
132    half4 result = blend_src_over(src, dst);
133    result.rgb = max(result.rgb, (1 - dst.a)*src.rgb + dst.rgb);
134    return result;
135}
136
137$pure half4 blend_darken(half mode /* darken: 1, lighten: -1 */, half4 src, half4 dst) {
138    half4 a = blend_src_over(src, dst);
139    half3 b = (1 - dst.a) * src.rgb + dst.rgb;  // DstOver.rgb
140    a.rgb = mode * min(a.rgb * mode, b.rgb * mode);
141    return a;
142}
143
144$pure half4 blend_darken(half4 src, half4 dst) {
145   return blend_darken(1, src, dst);
146}
147
148// A useful constant to check against when dividing a half-precision denominator.
149// Denormal half floats (values less than this) will compare not-equal to 0 but can easily cause the
150// division to overflow to infinity. Even regular values can overflow given the low maximum value.
151// For instance, any value x > ~3.998 will overflow when divided by $kMinNormalHalf. This is a
152// reasonable value even for wide gamut colors being input to these blend functions, but the
153// most correct denominator check is to treat anything with `denom < x/F16_MAX` as division by 0.
154const half $kMinNormalHalf = 1.0 / (1 << 14);
155
156const half $kGuardedDivideEpsilon = sk_Caps.mustGuardDivisionEvenAfterExplicitZeroCheck
157                                        ? 0.00000001
158                                        : 0.0;
159
160$pure inline half $guarded_divide(half n, half d) {
161    return n / (d + $kGuardedDivideEpsilon);
162}
163
164$pure inline half3 $guarded_divide(half3 n, half d) {
165    return n / (d + $kGuardedDivideEpsilon);
166}
167
168$pure half $color_dodge_component(half2 s, half2 d) {
169    // The following is a single flow of control implementation of:
170    //     if (d.x == 0) {
171    //         return s.x*(1 - d.y);
172    //     } else {
173    //         half delta = s.y - s.x;
174    //         if (delta == 0) {
175    //             return s.y*d.y + s.x*(1 - d.y) + d.x*(1 - s.y);
176    //         } else {
177    //             delta = min(d.y, $guarded_divide(d.x*s.y, delta));
178    //             return delta*s.y + s.x*(1 - d.y) + d.x*(1 - s.y);
179    //         }
180    //     }
181    //
182    // When d.x == 0, then dxScale forces delta to 0 and simplifying the return value to s.x*(1-d.y)
183    // When s.y-s.x == 0, the mix selects d.y and min(d.y, d.y) leaves delta = d.y
184    // Otherwise the mix selects the delta expression in the final else branch.
185    half dxScale = d.x == 0 ? 0 : 1;
186    half delta = dxScale * min(d.y, abs(s.y-s.x) >= $kMinNormalHalf
187                                            ? $guarded_divide(d.x*s.y, s.y-s.x)
188                                            : d.y);
189    return delta*s.y + s.x*(1 - d.y) + d.x*(1 - s.y);
190}
191
192$pure half4 blend_color_dodge(half4 src, half4 dst) {
193    return half4($color_dodge_component(src.ra, dst.ra),
194                 $color_dodge_component(src.ga, dst.ga),
195                 $color_dodge_component(src.ba, dst.ba),
196                 src.a + (1 - src.a)*dst.a);
197}
198
199$pure half $color_burn_component(half2 s, half2 d) {
200    // The following is a single flow of control implementation of:
201    //     if (d.y == d.x) {
202    //         return s.y*d.y + s.x*(1 - d.y) + d.x*(1 - s.y);
203    //     } else if (s.x == 0) {
204    //         return d.x*(1 - s.y);
205    //     } else {
206    //         half delta = max(0, d.y - $guarded_divide((d.y - d.x)*s.y, s.x));
207    //         return delta*s.y + s.x*(1 - d.y) + d.x*(1 - s.y);
208    //     }
209    //
210    // When d.y == d.x, dyTerm is d.y. If s.x is also 0, the second ternary selects d.y, matching
211    // the first if condition.  If s.x is not 0, then the $guarded_divide() evaluates to 0 and delta
212    // still evaluates to d.y.
213    //
214    // When d.y != d.x but s.x is 0, then dyTerm is 0 and the delta selects 0, matching the second
215    // if condition.
216    //
217    // Lastly, when d.y != d.x and s.x != 0, the delta evaluates to "d.y - min(d.y,
218    // $guarded_divide(...))", which is equivalent to max(0, d.y - $guarded_divide) except that it
219    // has the benefit of not wrapping the d.y evaluation in a max() to preserve the unclamped
220    // behavior when d.y == d.x.
221    half dyTerm = d.y == d.x ? d.y : 0;
222    half delta = abs(s.x) >= $kMinNormalHalf
223                        ? d.y - min(d.y, $guarded_divide((d.y - d.x)*s.y, s.x))
224                        : dyTerm;
225    return delta*s.y + s.x*(1 - d.y) + d.x*(1 - s.y);
226}
227
228$pure half4 blend_color_burn(half4 src, half4 dst) {
229    return half4($color_burn_component(src.ra, dst.ra),
230                 $color_burn_component(src.ga, dst.ga),
231                 $color_burn_component(src.ba, dst.ba),
232                 src.a + (1 - src.a)*dst.a);
233}
234
235$pure half4 blend_hard_light(half4 src, half4 dst) {
236    return blend_overlay(dst, src);
237}
238
239$pure half $soft_light_component(half2 s, half2 d) {
240    if (2*s.x <= s.y) {
241        return $guarded_divide(d.x*d.x*(s.y - 2*s.x), d.y) + (1 - d.y)*s.x + d.x*(-s.y + 2*s.x + 1);
242    } else if (4.0 * d.x <= d.y) {
243        half DSqd = d.x*d.x;
244        half DCub = DSqd*d.x;
245        half DaSqd = d.y*d.y;
246        half DaCub = DaSqd*d.y;
247        return $guarded_divide(DaSqd*(s.x - d.x*(3*s.y - 6*s.x - 1)) + 12*d.y*DSqd*(s.y - 2*s.x)
248                               - 16*DCub * (s.y - 2*s.x) - DaCub*s.x, DaSqd);
249    } else {
250        return d.x*(s.y - 2*s.x + 1) + s.x - sqrt(d.y*d.x)*(s.y - 2*s.x) - d.y*s.x;
251    }
252}
253
254$pure half4 blend_soft_light(half4 src, half4 dst) {
255    return (dst.a == 0) ? src : half4($soft_light_component(src.ra, dst.ra),
256                                      $soft_light_component(src.ga, dst.ga),
257                                      $soft_light_component(src.ba, dst.ba),
258                                      src.a + (1 - src.a)*dst.a);
259}
260
261$pure half4 blend_difference(half4 src, half4 dst) {
262    return half4(src.rgb + dst.rgb - 2*min(src.rgb*dst.a, dst.rgb*src.a),
263                 src.a + (1 - src.a)*dst.a);
264}
265
266$pure half4 blend_exclusion(half4 src, half4 dst) {
267    return half4(dst.rgb + src.rgb - 2*dst.rgb*src.rgb, src.a + (1 - src.a)*dst.a);
268}
269
270$pure half4 blend_multiply(half4 src, half4 dst) {
271    return half4((1 - src.a)*dst.rgb + (1 - dst.a)*src.rgb + src.rgb*dst.rgb,
272                 src.a + (1 - src.a)*dst.a);
273}
274
275$pure half $blend_color_luminance(half3 color) { return dot(half3(0.3, 0.59, 0.11), color); }
276
277$pure half3 $blend_set_color_luminance(half3 hueSatColor, half alpha, half3 lumColor) {
278    half lum = $blend_color_luminance(lumColor);
279    half3 result = lum - $blend_color_luminance(hueSatColor) + hueSatColor;
280    half minComp = min(min(result.r, result.g), result.b);
281    half maxComp = max(max(result.r, result.g), result.b);
282    if (minComp < 0 && lum != minComp) {
283        result = lum + (result - lum) * $guarded_divide(lum, (lum - minComp) + $kMinNormalHalf);
284    }
285    if (maxComp > alpha && maxComp != lum) {
286        result = lum +
287                 $guarded_divide((result - lum) * (alpha - lum), (maxComp - lum) + $kMinNormalHalf);
288    }
289    return result;
290}
291
292$pure half $blend_color_saturation(half3 color) {
293    return max(max(color.r, color.g), color.b) - min(min(color.r, color.g), color.b);
294}
295
296$pure half3 $blend_set_color_saturation(half3 color, half3 satColor) {
297    half mn = min(min(color.r, color.g), color.b);
298    half mx = max(max(color.r, color.g), color.b);
299
300    return (mx > mn) ? ((color - mn) * $blend_color_saturation(satColor)) / (mx - mn)
301                     : half3(0);
302}
303
304$pure half4 blend_hslc(half2 flipSat, half4 src, half4 dst) {
305    half alpha = dst.a * src.a;
306    half3 sda = src.rgb * dst.a;
307    half3 dsa = dst.rgb * src.a;
308    half3 l = bool(flipSat.x) ? dsa : sda;
309    half3 r = bool(flipSat.x) ? sda : dsa;
310    if (bool(flipSat.y)) {
311        l = $blend_set_color_saturation(l, r);
312        r = dsa;
313    }
314    return half4($blend_set_color_luminance(l, alpha, r) + dst.rgb - dsa + src.rgb - sda,
315                 src.a + dst.a - alpha);
316}
317
318$pure half4 blend_hue(half4 src, half4 dst) {
319    return blend_hslc(half2(0, 1), src, dst);
320}
321
322$pure half4 blend_saturation(half4 src, half4 dst) {
323    return blend_hslc(half2(1), src, dst);
324}
325
326$pure half4 blend_color(half4 src, half4 dst)  {
327    return blend_hslc(half2(0), src, dst);
328}
329
330$pure half4 blend_luminosity(half4 src, half4 dst) {
331    return blend_hslc(half2(1, 0), src, dst);
332}
333
334$pure float2 proj(float3 p) { return p.xy / p.z; }
335
336// Implement cross() as a determinant to communicate our intent more clearly to the compiler.
337// NOTE: Due to precision issues, it might be the case that cross(a, a) != 0.
338$pure float cross_length_2d(float2 a, float2 b) {
339    return determinant(float2x2(a, b));
340}
341
342$pure half cross_length_2d(half2 a, half2 b) {
343    return determinant(half2x2(a, b));
344}
345
346$pure float2 perp(float2 v) {
347    return float2(-v.y, v.x);
348}
349
350$pure half2 perp(half2 v) {
351    return half2(-v.y, v.x);
352}
353
354// Returns a bias given a scale factor, such that 'scale * (dist + bias)' converts the distance to
355// a per-pixel coverage value, automatically widening the visible coverage ramp for subpixel
356// dimensions. The 'scale' must already be equal to the narrowest dimension of the shape and clamped
357// to [0, 1.0].
358$pure float coverage_bias(float scale) {
359    return 1.0 - 0.5 * scale;
360}
361