AnalyticRRectRenderStep.cpp | Explore in Territory

/*
 * Copyright 2022 Google LLC
 *
 * Use of this source code is governed by a BSD-style license that can be
 * found in the LICENSE file.
 */

#include "src/gpu/graphite/render/AnalyticRRectRenderStep.h"

#include "src/base/SkVx.h"
#include "src/core/SkRRectPriv.h"
#include "src/gpu/graphite/DrawParams.h"
#include "src/gpu/graphite/DrawWriter.h"
#include "src/gpu/graphite/render/CommonDepthStencilSettings.h"

// This RenderStep is flexible and can draw filled rectangles, filled quadrilaterals with per-edge
// AA, filled rounded rectangles with arbitrary corner radii, stroked rectangles with any join,
// stroked lines with any cap, stroked rounded rectangles with circular corners (each corner can be
// different or square), hairline rectangles, hairline lines, and hairline rounded rectangles with
// arbitrary corners.
//
// We combine all of these together to maximize batching across simple geometric draws and reduce
// the number pipeline specializations. Additionally, these primitives are the most common
// operations and help us avoid triggering MSAA.
//
// Each of these "primitives" is represented by a single instance. The instance attributes are
// flexible enough to describe any of the above shapes without relying on uniforms to define its
// operation. The attributes encode shape as follows:
//
// float4 xRadiiOrFlags - if any components is > 0, the instance represents a filled round rect
//    with elliptical corners and these values specify the X radii in top-left CW order.
//    Otherwise, if .x < -1, the instance represents a stroked or hairline [round] rect or line,
//    where .y differentiates hairline vs. stroke. If .y is negative, then it is a hairline [round]
//    rect and xRadiiOrFlags stores (-2 - X radii); if .y is zero, it is a regular stroked [round]
//    rect; if .y is positive, then it is a stroked *or* hairline line. For .y >= 0, .z holds the
//    stroke radius and .w stores the join limit (matching StrokeStyle's conventions).
//    Lastly, if -1 <= .x <= 0, it's a filled quadrilateral with per-edge AA defined by each by the
//    component: aa != 0.
// float4 radiiOrQuadXs - if in filled round rect or hairline [round] rect mode, these values
//    provide the Y radii in top-left CW order. If in stroked [round] rect mode, these values
//    provide the circular corner radii (same order). Otherwise, when in per-edge quad mode, these
//    values provide the X coordinates of the quadrilateral (same order).
// float4 ltrbOrQuadYs - if in filled round rect mode or stroked [round] rect mode, these values
//    define the LTRB edge coordinates of the rectangle surrounding the round rect (or the
//    rect itself when the radii are 0s). In stroked line mode, LTRB is treated as (x0,y0) and
//    (x1,y1) that defines the line. Otherwise, in per-edge quad mode, these values provide
//    the Y coordinates of the quadrilateral.
//
// From the other direction, shapes produce instance values like:
//  - filled rect:    [-1 -1 -1 -1]            [L R R L]             [T T B B]
//  - stroked rect:   [-2 0 stroke join]       [0 0 0 0]             [L T R B]
//  - hairline rect:  [-2 -2 -2 -2]            [0 0 0 0]             [L T R B]
//  - filled rrect:   [xRadii(tl,tr,br,bl)]    [yRadii(tl,tr,br,bl)] [L T R B]
//  - stroked rrect:  [-2 0 stroke join]       [radii(tl,tr,br,bl)]  [L T R B]
//  - hairline rrect: [-2-xRadii(tl,tr,br,bl)] [radii(tl,tr,br,bl)]  [L T R B]
//  - filled line:    N/A, discarded higher in the stack
//  - stroked line:   [-2 1 stroke cap]        [0 0 0 0]             [x0,y0,x1,y1]
//  - hairline line:  [-2 1 0 1]               [0 0 0 0]             [x0,y0,x1,y1]
//  - per-edge quad:  [aa(t,r,b,l) ? -1 : 0]   [xs(tl,tr,br,bl)]     [ys(tl,tr,br,bl)]
//
// This encoding relies on the fact that a valid SkRRect with all x radii equal to 0 must have
// y radii equal to 0 (so it's a rectangle and we can treat it as a quadrilateral with
// all edges AA'ed). This avoids other encodings' inability to represent a quad with all edges
// anti-aliased (e.g. checking for negatives in xRadiiOrFlags to turn on per-edge mode).
//
// From this encoding, data can be unpacked for each corner, which are equivalent under
// rotational symmetry. A corner can have an outer curve, be mitered, or be beveled. It can
// have an inner curve, an inner miter, or fill the interior. Per-edge quads are always mitered
// and fill the interior, but the vertices are placed such that the edge coverage ramps can
// collapse to 0 area on non-AA edges.
//
// The vertices that describe each corner are placed so that edges, miters, and bevels calculate
// coverage by interpolating a varying and then clamping in the fragment shader. Triangles that
// cover the inner and outer curves calculate distance to the curve within the fragment shader.
//
// See https://docs.google.com/presentation/d/1MCPstNsSlDBhR8CrsJo0r-cZNbu-sEJEvU9W94GOJoY/edit?usp=sharing
// for diagrams and explanation of how the geometry is defined.
//
// AnalyticRRectRenderStep uses the common technique of approximating distance to the level set by
// one expansion of the Taylor's series for the level set's equation. Given a level set function
// C(x,y), this amounts to calculating C(px,py)/|∇C(px,py)|. For the straight edges the level set
// is linear and calculated in the vertex shader and then interpolated exactly over the rectangle.
// This provides distances to all four exterior edges within the fragment shader and allows it to
// reconstruct a relative position per elliptical corner. Unfortunately this requires the fragment
// shader to calculate the length of the gradient for straight edges instead of interpolating
// exact device-space distance.
//
// All four corner radii are potentially evaluated by the fragment shader although each corner's
// coverage is only calculated when the pixel is within the bounding box of its quadrant. For fills
// and simple strokes it's theoretically valid to have each pixel calculate a single corner's
// coverage that was controlled via the vertex shader. However, testing all four corners is
// necessary in order to correctly handle self-intersecting stroke interiors. Similarly, all four
// edges must be evaluated in order to handle extremely thin shapes; whereas often you could get
// away with tracking a single edge distance per pixel.
//
// Analytic derivatives are used so that a single pipeline can be used regardless of HW derivative
// support or for geometry that would prove difficult for forward differencing. The device-space
// gradient for ellipses is calculated per-pixel by transforming a per-pixel local gradient vector
// with the Jacobian of the inverse local-to-device transform:
//
// (px,py) is the projected point of (u,v) transformed by a 3x3 matrix, M:
//                [x(u,v) / w(u,v)]       [x]   [m00 m01 m02] [u]
//      (px,py) = [y(u,v) / w(u,v)] where [y] = [m10 m11 m12]X[v] = M*(u,v,1)
//                                        [w]   [m20 m21 m22] [1]
//
// C(px,py) can be defined in terms of a local Cl(u,v) as C(px,py) = Cl(p^-1(px,py)), where p^-1 =
//
//               [x'(px,py) / w'(px,py)]       [x']   [m00' m01' * m02'] [px]
//      (u,v) =  [y'(px,py) / w'(px,py)] where [y'] = [m10' m11' * m12']X[py] = M^-1*(px,py,0,1)
//                                             [w']   [m20' m21' * m22'] [ 1]
//
// Note that if the 3x3 M was arrived by dropping the 3rd row and column from a 4x4 since we assume
// a local 3rd coordinate of 0, M^-1 is not equal to the 4x4 inverse with dropped rows and columns.
//
// Using the chain rule, then ∇C(px,py)
//   =  ∇Cl(u,v)X[1/w'(px,py)     0       -x'(px,py)/w'(px,py)^2] [m00' m01']
//               [    0       1/w'(px,py) -y'(px,py)/w'(px,py)^2]X[m10' m11']
//                                                                [m20' m21']
//
//   = 1/w'(px,py)*∇Cl(u,v)X[1 0 -x'(px,py)/w'(px,py)] [m00' m01']
//                          [0 1 -y'(px,py)/w'(px,py)]X[m10' m11']
//                                                     [m20' m21']
//
//   = w(u,v)*∇Cl(u,v)X[1 0 0 -u] [m00' m01']
//                     [0 1 0 -v]X[m10' m11']
//                                [m20' m21']
//
//   = w(u,v)*∇Cl(u,v)X[m00'-m20'u m01'-m21'u]
//                     [m10'-m20'v m11'-m21'v]
//
// The vertex shader calculates the rightmost 2x2 matrix and interpolates it across the shape since
// each component is linear in (u,v). ∇Cl(u,v) is evaluated per pixel in the fragment shader and
// depends on which corner and edge being evaluated. w(u,v) is the device-space W coordinate, so
// its reciprocal is provided in sk_FragCoord.w.
namespace skgpu::graphite {

AAFlags;

static skvx::float4 load_x_radii(const SkRRect& rrect) { … }
static skvx::float4 load_y_radii(const SkRRect& rrect) { … }

static bool opposite_insets_intersect(const SkRRect& rrect, float strokeRadius, float aaRadius) { … }

static bool opposite_insets_intersect(const Rect& rect, float strokeRadius, float aaRadius) { … }

static bool opposite_insets_intersect(const Geometry& geometry,
                                      float strokeRadius,
                                      float aaRadius) { … }

static bool is_clockwise(const EdgeAAQuad& quad) { … }

static skvx::float2 quad_center(const EdgeAAQuad& quad) { … }

// Represents the per-vertex attributes used in each instance.
struct Vertex { … };

// Allowed values for the center weight instance value (selected at record time based on style
// and transform), and are defined such that when (insance-weight > vertex-weight) is true, the
// vertex should be snapped to the center instead of its regular calculation.
static constexpr float kSolidInterior = …;
static constexpr float kStrokeInterior = …;
static constexpr float kFilledStrokeInterior = …;

// Special value for local AA radius to signal when the self-intersections of a stroke interior
// need extra calculations in the vertex shader.
static constexpr float kComplexAAInsets = …;

static constexpr int kCornerVertexCount = …; // sk_VertexID is divided by this in SkSL
static constexpr int kVertexCount = …;
static constexpr int kIndexCount = …;

static void write_index_buffer(VertexWriter writer) { … }

static void write_vertex_buffer(VertexWriter writer) { … }

AnalyticRRectRenderStep::AnalyticRRectRenderStep(StaticBufferManager* bufferManager)
        : … { … }

AnalyticRRectRenderStep::~AnalyticRRectRenderStep() { … }

std::string AnalyticRRectRenderStep::vertexSkSL() const { … }

const char* AnalyticRRectRenderStep::fragmentCoverageSkSL() const { … }

void AnalyticRRectRenderStep::writeVertices(DrawWriter* writer,
                                            const DrawParams& params,
                                            skvx::ushort2 ssboIndices) const { … }

void AnalyticRRectRenderStep::writeUniformsAndTextures(const DrawParams&,
                                                       PipelineDataGatherer*) const { … }

}  // namespace skgpu::graphite
chromium/third_party/skia/src/gpu/graphite/render/AnalyticRRectRenderStep.cpp
