firka/flutter
3
0
Fork 1
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Vectorize rrect_blur (flutter/engine#55576)

Browse Source 
issue: https://github.com/flutter/flutter/issues/148496

28% speed improvement in the rrect_blur shader.

## before
<img width="1728" alt="Screenshot 2024-10-01 at 3 45 46 PM" src="https://github.com/user-attachments/assets/643068a5-ab1e-4fa3-bc03-184b2ee4a6cf">

## after
<img width="1728" alt="Screenshot 2024-10-01 at 3 42 30 PM" src="https://github.com/user-attachments/assets/41445231-ffea-4279-8142-ce126df8187c">

[C++, Objective-C, Java style guides]: https://github.com/flutter/engine/blob/main/CONTRIBUTING.md#style
This commit is contained in:
gaaclarke
2024-10-02 17:07:22 -07:00
committed by GitHub
parent f3746f9001
commit 569ddd287e
3 changed files with 54 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
engine/src/flutter/impeller/compiler/shader_lib/impeller/gaussian.glsl
View File

@@ -14,6 +14,12 @@ float IPGaussian(float x, float sigma) {
return exp(-0.5f * x * x / variance) / (kSqrtTwoPi * sigma);
}
/// Equivalent to `IPGaussian(float x, float sigma)` but executed 4x.
vec4 IPGaussian(vec4 x, float sigma) {
float variance = sigma * sigma;
return exp(-0.5f * x * x / variance) / (kSqrtTwoPi * sigma);
}
/// Gaussian distribution function.
float16_t IPHalfGaussian(float16_t x, float16_t sigma) {
float16_t variance = sigma * sigma;
@@ -56,6 +62,12 @@ vec2 IPVec2FastGaussianIntegral(vec2 x, float sigma) {
return 1.0 / (1.0 + exp(-kSqrtThree / sigma * x));
}
/// Equivalent to `IPVec2FastGaussianIntegral(vec2 x, float sigma)` but operated
/// 4x instead of 2x.
vec4 IPVec4FastGaussianIntegral(vec4 x, float sigma) {
return 1.0 / (1.0 + exp(-kSqrtThree / sigma * x));
}
/// Simpler (but less accurate) approximation of the Gaussian integral.
f16vec2 IPHalfVec2FastGaussianIntegral(f16vec2 x, float16_t sigma) {
return 1.0hf / (1.0hf + exp(float16_t(-kSqrtThree) / sigma * x));

43
engine/src/flutter/impeller/entity/shaders/rrect_blur.frag
View File

@@ -23,7 +23,9 @@ const int kSampleCount = 4;
/// Closed form unidirectional rounded rect blur mask solution using the
/// analytical Gaussian integral (with approximated erf).
float RRectBlurX(vec2 sample_position, vec2 half_size) {
vec4 RRectBlurX(float sample_position_x,
vec4 sample_position_y,
vec2 half_size) {
// The vertical edge of the rrect consists of a flat portion and a curved
// portion, the two of which vary in size depending on the size of the
// corner radii, both adding up to half_size.y.
@@ -33,8 +35,8 @@ float RRectBlurX(vec2 sample_position, vec2 half_size) {
// negative (and then clamped to 0) for positions that are located
// vertically in the flat part of the rrect, and will be the relative
// distance from the center of curvature otherwise.
float space_y =
min(0.0, half_size.y - frag_info.corner_radii.y - abs(sample_position.y));
vec4 space_y = min(vec4(0.0), half_size.y - frag_info.corner_radii.y -
abs(sample_position_y));
// space is now in the range [0.0, corner_radii.y]. If the y sample was
// in the flat portion of the rrect, it will be 0.0
@@ -52,22 +54,31 @@ float RRectBlurX(vec2 sample_position, vec2 half_size) {
// space_y was larger than corner_radii.y.
// The calling function RRectBlur will never provide a Y sample outside
// of that range, though, so the max(0.0) is mostly a precaution.
float unit_space_y = space_y / frag_info.corner_radii.y;
float unit_space_x = sqrt(max(0.0, 1.0 - unit_space_y * unit_space_y));
float rrect_distance =
vec4 unit_space_y = space_y / frag_info.corner_radii.y;
vec4 unit_space_x = sqrt(max(vec4(0.0), 1.0 - unit_space_y * unit_space_y));
vec4 rrect_distance =
half_size.x - frag_info.corner_radii.x * (1.0 - unit_space_x);
vec4 result;
// Now we integrate the Gaussian over the range of the relative positions
// of the left and right sides of the rrect relative to the sampling
// X coordinate.
vec2 integral = IPVec2FastGaussianIntegral(
float(sample_position.x) + vec2(-rrect_distance, rrect_distance),
vec4 integral = IPVec4FastGaussianIntegral(
float(sample_position_x) + vec4(-rrect_distance[0], rrect_distance[0],
-rrect_distance[1], rrect_distance[1]),
float(frag_info.blur_sigma));
// integral.y contains the evaluation of the indefinite gaussian integral
// function at (X + rrect_distance) and integral.x contains the evaluation
// of it at (X - rrect_distance). Subtracting the two produces the
// integral result over the range from one to the other.
return integral.y - integral.x;
result.xy = integral.yw - integral.xz;
integral = IPVec4FastGaussianIntegral(
float(sample_position_x) + vec4(-rrect_distance[2], rrect_distance[2],
-rrect_distance[3], rrect_distance[3]),
float(frag_info.blur_sigma));
result.zw = integral.yw - integral.xz;
return result;
}
float RRectBlur(vec2 sample_position, vec2 half_size) {
@@ -84,15 +95,11 @@ float RRectBlur(vec2 sample_position, vec2 half_size) {
float interval = (end_y - begin_y) / kSampleCount;
// Sample the X blur kSampleCount times, weighted by the Gaussian function.
float result = 0.0;
for (int sample_i = 0; sample_i < kSampleCount; sample_i++) {
float y = begin_y + interval * (float(sample_i) + 0.5);
result +=
RRectBlurX(vec2(sample_position.x, sample_position.y - y), half_size) *
IPGaussian(float(y), float(frag_info.blur_sigma)) * interval;
}
return result;
vec4 ys = vec4(0.5, 1.5, 2.5, 3.5) * interval + begin_y;
vec4 sample_ys = sample_position.y - ys;
vec4 blurx = RRectBlurX(sample_position.x, sample_ys, half_size);
vec4 gaussian_y = IPGaussian(ys, float(frag_info.blur_sigma));
return dot(blurx, gaussian_y * interval);
}
void main() {

34
engine/src/flutter/impeller/tools/malioc.json
View File

@@ -3858,8 +3858,8 @@
"arith_fma"
],
"longest_path_cycles": [
1.65625,
1.65625,
1.59375,
1.59375,
0.453125,
1.5,
0.0,
@@ -3880,8 +3880,8 @@
"arith_fma"
],
"shortest_path_cycles": [
1.65625,
1.65625,
1.59375,
1.59375,
0.421875,
1.5,
0.0,
@@ -3893,8 +3893,8 @@
"arith_fma"
],
"total_cycles": [
1.65625,
1.65625,
1.59375,
1.59375,
0.453125,
1.5,
0.0,
@@ -3922,7 +3922,7 @@
"arithmetic"
],
"longest_path_cycles": [
25.739999771118164,
12.869999885559082,
1.0,
0.0
],
@@ -3935,7 +3935,7 @@
"arithmetic"
],
"shortest_path_cycles": [
25.739999771118164,
12.869999885559082,
1.0,
0.0
],
@@ -3943,14 +3943,14 @@
"arithmetic"
],
"total_cycles": [
8.333333015441895,
13.333333015441895,
1.0,
0.0
]
},
"thread_occupancy": 100,
"thread_occupancy": 50,
"uniform_registers_used": 3,
"work_registers_used": 4
"work_registers_used": 7
}
}
}
@@ -6816,8 +6816,8 @@
"arith_fma"
],
"longest_path_cycles": [
1.65625,
1.65625,
1.59375,
1.59375,
0.421875,
1.5,
0.0,
@@ -6838,8 +6838,8 @@
"arith_fma"
],
"shortest_path_cycles": [
1.65625,
1.65625,
1.59375,
1.59375,
0.421875,
1.5,
0.0,
@@ -6851,8 +6851,8 @@
"arith_fma"
],
"total_cycles": [
1.65625,
1.65625,
1.59375,
1.59375,
0.421875,
1.5,
0.0,