[Impeller] Use 32 bit for positional stuff in the rrect blur (flutter/engine#42797)

Resolves https://github.com/flutter/flutter/issues/128410.

16 bit floats aren't precise enough for positional computations, like
computing SDFs in the case of the RRect blur.

engine/src/flutter/impeller/compiler/shader_lib/impeller/gaussian.glsl

@@ -52,7 +52,12 @@ f16vec2 IPVec2GaussianIntegral(f16vec2 x, float16_t sigma) {
 }
 
 /// Simpler (but less accurate) approximation of the Gaussian integral.
-f16vec2 IPVec2FastGaussianIntegral(f16vec2 x, float16_t sigma) {
+vec2 IPVec2FastGaussianIntegral(vec2 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));
 }
 

engine/src/flutter/impeller/entity/shaders/rrect_blur.frag

@@ -7,9 +7,9 @@
 
 uniform FragInfo {
   f16vec4 color;
-  f16vec2 rect_size;
-  float16_t blur_sigma;
-  float16_t corner_radius;
+  vec2 rect_size;
+  float blur_sigma;
+  float corner_radius;
 }
 frag_info;
 
@@ -19,50 +19,47 @@ out f16vec4 frag_color;
 
 const int kSampleCount = 4;
 
-float16_t RRectDistance(f16vec2 sample_position, f16vec2 half_size) {
-  f16vec2 space = abs(sample_position) - half_size + frag_info.corner_radius;
-  return length(max(space, float16_t(0.0hf))) +
-         min(max(space.x, space.y), float16_t(0.0hf)) - frag_info.corner_radius;
+float16_t RRectDistance(vec2 sample_position, vec2 half_size) {
+  vec2 space = abs(sample_position) - half_size + frag_info.corner_radius;
+  return float16_t(length(max(space, 0.0)) + min(max(space.x, space.y), 0.0) -
+                   frag_info.corner_radius);
 }
 
 /// Closed form unidirectional rounded rect blur mask solution using the
 /// analytical Gaussian integral (with approximated erf).
-float16_t RRectBlurX(f16vec2 sample_position, f16vec2 half_size) {
+float RRectBlurX(vec2 sample_position, vec2 half_size) {
   // Compute the X direction distance field (not incorporating the Y distance)
   // for the rounded rect.
-  float16_t space =
-      min(float16_t(0.0hf),
-          half_size.y - frag_info.corner_radius - abs(sample_position.y));
-  float16_t rrect_distance =
+  float space =
+      min(0.0, half_size.y - frag_info.corner_radius - abs(sample_position.y));
+  float rrect_distance =
       half_size.x - frag_info.corner_radius +
-      sqrt(max(
-          float16_t(0.0hf),
-          frag_info.corner_radius * frag_info.corner_radius - space * space));
+      sqrt(max(0.0, frag_info.corner_radius * frag_info.corner_radius -
+                        space * space));
 
   // Map the linear distance field to the approximate Gaussian integral.
-  f16vec2 integral = IPVec2FastGaussianIntegral(
-      sample_position.x + f16vec2(-rrect_distance, rrect_distance),
-      frag_info.blur_sigma);
+  vec2 integral = IPVec2FastGaussianIntegral(
+      float(sample_position.x) + vec2(-rrect_distance, rrect_distance),
+      float(frag_info.blur_sigma));
   return integral.y - integral.x;
 }
 
-float16_t RRectBlur(f16vec2 sample_position, f16vec2 half_size) {
+float RRectBlur(vec2 sample_position, vec2 half_size) {
   // Limit the sampling range to 3 standard deviations in the Y direction from
   // the kernel center to incorporate 99.7% of the color contribution.
-  float16_t half_sampling_range = frag_info.blur_sigma * 3.0hf;
+  float half_sampling_range = frag_info.blur_sigma * 3.0;
 
-  float16_t begin_y =
-      max(-half_sampling_range, sample_position.y - half_size.y);
-  float16_t end_y = min(half_sampling_range, sample_position.y + half_size.y);
-  float16_t interval = (end_y - begin_y) / float16_t(kSampleCount);
+  float begin_y = max(-half_sampling_range, sample_position.y - half_size.y);
+  float end_y = min(half_sampling_range, sample_position.y + half_size.y);
+  float interval = (end_y - begin_y) / kSampleCount;
 
   // Sample the X blur kSampleCount times, weighted by the Gaussian function.
-  float16_t result = 0.0hf;
+  float result = 0.0;
   for (int sample_i = 0; sample_i < kSampleCount; sample_i++) {
-    float16_t y = begin_y + interval * (float16_t(sample_i) + 0.5hf);
-    result += RRectBlurX(f16vec2(sample_position.x, sample_position.y - y),
-                         half_size) *
-              IPHalfGaussian(y, frag_info.blur_sigma) * interval;
+    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;
@@ -71,12 +68,12 @@ float16_t RRectBlur(f16vec2 sample_position, f16vec2 half_size) {
 void main() {
   frag_color = frag_info.color;
 
-  f16vec2 half_size = frag_info.rect_size * 0.5hf;
-  f16vec2 sample_position = f16vec2(v_position) - half_size;
+  vec2 half_size = frag_info.rect_size * 0.5;
+  vec2 sample_position = v_position - half_size;
 
-  if (frag_info.blur_sigma > 0.0hf) {
-    frag_color *= RRectBlur(sample_position, half_size);
+  if (frag_info.blur_sigma > 0.0) {
+    frag_color *= float16_t(RRectBlur(sample_position, half_size));
   } else {
-    frag_color *= -RRectDistance(sample_position, half_size);
+    frag_color *= float16_t(-RRectDistance(sample_position, half_size));
   }
 }