COM_SMAAOperation.cc

Go to the documentation of this file.

/* SPDX-FileCopyrightText: 2013 Jorge Jimenez <jorge@iryoku.com>
 * SPDX-FileCopyrightText: 2013 Jose I. Echevarria <joseignacioechevarria@gmail.com>
 * SPDX-FileCopyrightText: 2013 Belen Masia <bmasia@unizar.es>
 * SPDX-FileCopyrightText: 2013 Fernando Navarro <fernandn@microsoft.com>
 * SPDX-FileCopyrightText: 2013 Diego Gutierrez <diegog@unizar.es>
 * SPDX-FileCopyrightText: 2019-2023 Blender Authors
 *
 * SPDX-License-Identifier: MIT AND GPL-2.0-or-later */
 
#include "BLI_math_vector.h"
#include "BLI_math_vector.hh"
#include "BLI_smaa_textures.h"
#include "BLI_span.hh"
#include "BLI_task.hh"
 
#include "IMB_colormanagement.hh"
 
#include "COM_MemoryBuffer.h"
#include "COM_SMAAOperation.h"
 
/* ----------------------------------------------------------------------------
 * Blender's Defines */
 
#define SMAA_CUSTOM_SL
#define SMAA_AREATEX_SELECT(sample) sample.xy()
#define SMAA_SEARCHTEX_SELECT(sample) sample.x
#define SMAATexture2D(tex) const MemoryBuffer *tex
#define SMAATexturePass2D(tex) tex
#define SMAASampleLevelZero(tex, coord) tex->texture_bilinear_extend(coord)
#define SMAASampleLevelZeroPoint(tex, coord) tex->texture_bilinear_extend(coord)
#define SMAASampleLevelZeroOffset(tex, coord, offset, size) \
  tex->texture_bilinear_extend(coord + float2(offset) / float2(size))
#define SMAASample(tex, coord) tex->texture_bilinear_extend(coord)
#define SMAASamplePoint(tex, coord) tex->texture_nearest_extend(coord)
#define SMAASamplePointOffset(tex, coord, offset, size) \
  tex->texture_nearest_extend(coord + float2(offset) / float2(size))
#define SMAASampleOffset(tex, coord, offset, size) \
  tex->texture_bilinear_extend(coord + float2(offset) / float2(size))
#define SMAA_FLATTEN
#define SMAA_BRANCH
#define lerp(a, b, t) math::interpolate(a, b, t)
#define saturate(a) math::clamp(a, 0.0f, 1.0f)
#define mad(a, b, c) (a * b + c)
 
/* ----------------------------------------------------------------------------
 * SMAA Presets */
 
#if defined(SMAA_PRESET_LOW)
#  define SMAA_THRESHOLD 0.15f
#  define SMAA_MAX_SEARCH_STEPS 4
#  define SMAA_DISABLE_DIAG_DETECTION
#  define SMAA_DISABLE_CORNER_DETECTION
#elif defined(SMAA_PRESET_MEDIUM)
#  define SMAA_THRESHOLD 0.1f
#  define SMAA_MAX_SEARCH_STEPS 8
#  define SMAA_DISABLE_DIAG_DETECTION
#  define SMAA_DISABLE_CORNER_DETECTION
#elif defined(SMAA_PRESET_HIGH)
#  define SMAA_THRESHOLD 0.1f
#  define SMAA_MAX_SEARCH_STEPS 16
#  define SMAA_MAX_SEARCH_STEPS_DIAG 8
#  define SMAA_CORNER_ROUNDING 25
#elif defined(SMAA_PRESET_ULTRA)
#  define SMAA_THRESHOLD 0.05f
#  define SMAA_MAX_SEARCH_STEPS 32
#  define SMAA_MAX_SEARCH_STEPS_DIAG 16
#  define SMAA_CORNER_ROUNDING 25
#endif
 
/* ----------------------------------------------------------------------------
 * Configurable Defines */
 
#ifndef SMAA_THRESHOLD
#  define SMAA_THRESHOLD 0.1f
#endif
 
#ifndef SMAA_DEPTH_THRESHOLD
#  define SMAA_DEPTH_THRESHOLD (0.1f * SMAA_THRESHOLD)
#endif
 
#ifndef SMAA_MAX_SEARCH_STEPS
#  define SMAA_MAX_SEARCH_STEPS 16
#endif
 
#ifndef SMAA_MAX_SEARCH_STEPS_DIAG
#  define SMAA_MAX_SEARCH_STEPS_DIAG 8
#endif
 
#ifndef SMAA_CORNER_ROUNDING
#  define SMAA_CORNER_ROUNDING 25
#endif
 
#ifndef SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR
#  define SMAA_LOCAL_CONTRAST_ADAPTATION_FACTOR 2.0f
#endif
 
#ifndef SMAA_PREDICATION
#  define SMAA_PREDICATION 0
#endif
 
#ifndef SMAA_PREDICATION_THRESHOLD
#  define SMAA_PREDICATION_THRESHOLD 0.01f
#endif
 
#ifndef SMAA_PREDICATION_SCALE
#  define SMAA_PREDICATION_SCALE 2.0f
#endif
 
#ifndef SMAA_PREDICATION_STRENGTH
#  define SMAA_PREDICATION_STRENGTH 0.4f
#endif
 
#ifndef SMAA_REPROJECTION
#  define SMAA_REPROJECTION 0
#endif
 
#ifndef SMAA_REPROJECTION_WEIGHT_SCALE
#  define SMAA_REPROJECTION_WEIGHT_SCALE 30.0f
#endif
 
#ifndef SMAA_INCLUDE_VS
#  define SMAA_INCLUDE_VS 1
#endif
#ifndef SMAA_INCLUDE_PS
#  define SMAA_INCLUDE_PS 1
#endif
 
/* ----------------------------------------------------------------------------
 * Texture Access Defines */
 
#ifndef SMAA_AREATEX_SELECT
#  if defined(SMAA_HLSL_3)
#    define SMAA_AREATEX_SELECT(sample) sample.ra
#  else
#    define SMAA_AREATEX_SELECT(sample) sample.rg
#  endif
#endif
 
#ifndef SMAA_SEARCHTEX_SELECT
#  define SMAA_SEARCHTEX_SELECT(sample) sample.r
#endif
 
#ifndef SMAA_DECODE_VELOCITY
#  define SMAA_DECODE_VELOCITY(sample) sample.rg
#endif
 
/* ----------------------------------------------------------------------------
 * Non-Configurable Defines */
 
#define SMAA_AREATEX_MAX_DISTANCE 16
#define SMAA_AREATEX_MAX_DISTANCE_DIAG 20
#define SMAA_AREATEX_PIXEL_SIZE (1.0f / float2(160.0f, 560.0f))
#define SMAA_AREATEX_SUBTEX_SIZE (1.0f / 7.0f)
#define SMAA_SEARCHTEX_SIZE float2(66.0f, 33.0f)
#define SMAA_SEARCHTEX_PACKED_SIZE float2(64.0f, 16.0f)
#define SMAA_CORNER_ROUNDING_NORM (float(SMAA_CORNER_ROUNDING) / 100.0f)
 
/* ----------------------------------------------------------------------------
 * Porting Functions */
 
#if defined(SMAA_HLSL_3)
#  define SMAATexture2D(tex) sampler2D tex
#  define SMAATexturePass2D(tex) tex
#  define SMAASampleLevelZero(tex, coord) tex2Dlod(tex, float4(coord, 0.0, 0.0))
#  define SMAASampleLevelZeroPoint(tex, coord) tex2Dlod(tex, float4(coord, 0.0, 0.0))
/* clang-format off */
#  define SMAASampleLevelZeroOffset(tex, coord, offset) tex2Dlod(tex, float4(coord + offset * SMAA_RT_METRICS.xy, 0.0, 0.0))
/* clang-format on */
#  define SMAASample(tex, coord) tex2D(tex, coord)
#  define SMAASamplePoint(tex, coord) tex2D(tex, coord)
#  define SMAASampleOffset(tex, coord, offset) tex2D(tex, coord + offset * SMAA_RT_METRICS.xy)
#  define SMAA_FLATTEN [flatten]
#  define SMAA_BRANCH [branch]
#endif
#if defined(SMAA_HLSL_4) || defined(SMAA_HLSL_4_1)
SamplerState LinearSampler
{
  Filter = MIN_MAG_LINEAR_MIP_POINT;
  AddressU = Clamp;
  AddressV = Clamp;
};
SamplerState PointSampler
{
  Filter = MIN_MAG_MIP_POINT;
  AddressU = Clamp;
  AddressV = Clamp;
};
#  define SMAATexture2D(tex) Texture2D tex
#  define SMAATexturePass2D(tex) tex
#  define SMAASampleLevelZero(tex, coord) tex.SampleLevel(LinearSampler, coord, 0)
#  define SMAASampleLevelZeroPoint(tex, coord) tex.SampleLevel(PointSampler, coord, 0)
/* clang-format off */
#  define SMAASampleLevelZeroOffset(tex, coord, offset) tex.SampleLevel(LinearSampler, coord, 0, offset)
/* clang-format on */
#  define SMAASample(tex, coord) tex.Sample(LinearSampler, coord)
#  define SMAASamplePoint(tex, coord) tex.Sample(PointSampler, coord)
#  define SMAASampleOffset(tex, coord, offset) tex.Sample(LinearSampler, coord, offset)
#  define SMAA_FLATTEN [flatten]
#  define SMAA_BRANCH [branch]
#  define SMAATexture2DMS2(tex) Texture2DMS<float4, 2> tex
#  define SMAALoad(tex, pos, sample) tex.Load(pos, sample)
#  if defined(SMAA_HLSL_4_1)
#    define SMAAGather(tex, coord) tex.Gather(LinearSampler, coord, 0)
#  endif
#endif
#if defined(SMAA_GLSL_3) || defined(SMAA_GLSL_4) || defined(GPU_METAL) || defined(GPU_VULKAN)
#  define SMAATexture2D(tex) sampler2D tex
#  define SMAATexturePass2D(tex) tex
#  define SMAASampleLevelZero(tex, coord) textureLod(tex, coord, 0.0)
#  define SMAASampleLevelZeroPoint(tex, coord) textureLod(tex, coord, 0.0)
#  define SMAASampleLevelZeroOffset(tex, coord, offset) textureLodOffset(tex, coord, 0.0, offset)
#  define SMAASample(tex, coord) texture(tex, coord)
#  define SMAASamplePoint(tex, coord) texture(tex, coord)
#  define SMAASampleOffset(tex, coord, offset) texture(tex, coord, offset)
#  define SMAA_FLATTEN
#  define SMAA_BRANCH
#  define lerp(a, b, t) mix(a, b, t)
#  define saturate(a) clamp(a, 0.0, 1.0)
#  if defined(SMAA_GLSL_4)
#    define SMAAGather(tex, coord) textureGather(tex, coord)
#  endif
#  if defined(SMAA_GLSL_4)
#    define mad(a, b, c) fma(a, b, c)
#  elif defined(GPU_VULKAN)
/* NOTE(Vulkan) mad macro doesn't work, define each override as work-around. */
vec4 mad(vec4 a, vec4 b, vec4 c)
{
  return fma(a, b, c);
}
vec3 mad(vec3 a, vec3 b, vec3 c)
{
  return fma(a, b, c);
}
vec2 mad(vec2 a, vec2 b, vec2 c)
{
  return fma(a, b, c);
}
float mad(float a, float b, float c)
{
  return fma(a, b, c);
}
#  else
#    define mad(a, b, c) (a * b + c)
#  endif
/* NOTE(Metal): Types already natively declared in MSL. */
#  ifndef GPU_METAL
#    define float2 vec2
#    define float3 vec3
#    define float4 vec4
#    define int2 ivec2
#    define int3 ivec3
#    define int4 ivec4
#    define bool2 bvec2
#    define bool3 bvec3
#    define bool4 bvec4
#  endif
#endif
 
/* clang-format off */
#if !defined(SMAA_HLSL_3) && !defined(SMAA_HLSL_4) && !defined(SMAA_HLSL_4_1) && !defined(SMAA_GLSL_3) && !defined(SMAA_GLSL_4) && !defined(SMAA_CUSTOM_SL)
#  error you must define the shading language: SMAA_HLSL_*, SMAA_GLSL_* or SMAA_CUSTOM_SL
#endif
/* clang-format on */
 
namespace blender::compositor {
 
/* ----------------------------------------------------------------------------
 * Misc functions */
 
static void SMAAMovc(float2 cond, float2 &variable, float2 value)
{
  /* Use select function (select(genType A, genType B, genBType cond)). */
  variable = math::interpolate(variable, value, cond);
}
 
static void SMAAMovc(float4 cond, float4 &variable, float4 value)
{
  /* Use select function (select(genType A, genType B, genBType cond)). */
  variable = math::interpolate(variable, value, cond);
}
 
#if SMAA_INCLUDE_VS
/* ----------------------------------------------------------------------------
 * Vertex Shaders */
 
static void SMAAEdgeDetectionVS(float2 texcoord, int2 size, float4 offset[3])
{
  offset[0] = float4(texcoord.xy(), texcoord.xy()) +
              float4(-1.0f, 0.0f, 0.0f, -1.0f) / float4(size, size);
  offset[1] = float4(texcoord.xy(), texcoord.xy()) +
              float4(1.0f, 0.0f, 0.0f, 1.0f) / float4(size, size);
  offset[2] = float4(texcoord.xy(), texcoord.xy()) +
              float4(-2.0f, 0.0f, 0.0f, -2.0f) / float4(size, size);
}
 
static void SMAABlendingWeightCalculationVS(float2 texcoord,
                                            int2 size,
                                            float2 &pixcoord,
                                            float4 offset[3])
{
  pixcoord = texcoord * float2(size);
 
  /* We will use these offsets for the searches later on (see @PSEUDO_GATHER4): */
  offset[0] = float4(texcoord.xy(), texcoord.xy()) +
              float4(-0.25f, -0.125f, 1.25f, -0.125f) / float4(size, size);
  offset[1] = float4(texcoord.xy(), texcoord.xy()) +
              float4(-0.125f, -0.25f, -0.125f, 1.25f) / float4(size, size);
 
  /* And these for the searches, they indicate the ends of the loops: */
  offset[2] = float4(offset[0].x, offset[0].z, offset[1].y, offset[1].w) +
              (float4(-2.0f, 2.0f, -2.0f, 2.0f) * float(SMAA_MAX_SEARCH_STEPS)) /
                  float4(float2(size.x), float2(size.y));
}
 
static void SMAANeighborhoodBlendingVS(float2 texcoord, int2 size, float4 &offset)
{
  offset = float4(texcoord, texcoord) + float4(1.0f, 0.0f, 0.0f, 1.0f) / float4(size, size);
}
#endif /* SMAA_INCLUDE_VS */
 
static float2 SMAALumaEdgeDetectionPS(float2 texcoord,
                                      float4 offset[3],
                                      SMAATexture2D(colorTex),
#if SMAA_PREDICATION
                                      SMAATexture2D(predicationTex),
#endif
                                      float edge_threshold,
                                      float3 luminance_coefficients,
                                      float local_contrast_adaptation_factor)
{
#if SMAA_PREDICATION
  float2 threshold = SMAACalculatePredicatedThreshold(
      texcoord, offset, SMAATexturePass2D(predicationTex));
#else
  /* Calculate the threshold: */
  float2 threshold = float2(edge_threshold, edge_threshold);
#endif
 
  /* Calculate lumas: */
  // float4 weights = float4(0.2126, 0.7152, 0.0722, 0.0);
  float4 weights = float4(luminance_coefficients, 0.0f);
  float L = math::dot(SMAASamplePoint(colorTex, texcoord), weights);
 
  float Lleft = math::dot(SMAASamplePoint(colorTex, offset[0].xy()), weights);
  float Ltop = math::dot(SMAASamplePoint(colorTex, offset[0].zw()), weights);
 
  /* We do the usual threshold: */
  float4 delta;
  float2 delta_left_top = math::abs(L - float2(Lleft, Ltop));
  delta.x = delta_left_top.x;
  delta.y = delta_left_top.y;
  float2 edges = math::step(threshold, delta.xy());
 
  /* Then return early if there is no edge: */
  if (math::dot(edges, float2(1.0f, 1.0f)) == 0.0f) {
    return float2(0.0f);
  }
 
  /* Calculate right and bottom deltas: */
  float Lright = math::dot(SMAASamplePoint(colorTex, offset[1].xy()), weights);
  float Lbottom = math::dot(SMAASamplePoint(colorTex, offset[1].zw()), weights);
  float2 delta_right_bottom = math::abs(L - float2(Lright, Lbottom));
  delta.z = delta_right_bottom.x;
  delta.w = delta_right_bottom.y;
 
  /* Calculate the maximum delta in the direct neighborhood: */
  float2 maxDelta = math::max(delta.xy(), delta.zw());
 
  /* Calculate left-left and top-top deltas: */
  float Lleftleft = math::dot(SMAASamplePoint(colorTex, offset[2].xy()), weights);
  float Ltoptop = math::dot(SMAASamplePoint(colorTex, offset[2].zw()), weights);
  float2 delta_left_left_top_top = math::abs(float2(Lleft, Ltop) - float2(Lleftleft, Ltoptop));
  delta.z = delta_left_left_top_top.x;
  delta.w = delta_left_left_top_top.y;
 
  /* Calculate the final maximum delta: */
  maxDelta = math::max(maxDelta.xy(), delta.zw());
  float finalDelta = math::max(maxDelta.x, maxDelta.y);
 
  /* Local contrast adaptation: */
  edges *= math::step(finalDelta, local_contrast_adaptation_factor * delta.xy());
 
  return edges;
}
 
/* ----------------------------------------------------------------------------
 * Diagonal Search Functions */
 
#if !defined(SMAA_DISABLE_DIAG_DETECTION)
 
static float2 SMAADecodeDiagBilinearAccess(float2 e)
{
  /* Bilinear access for fetching 'e' have a 0.25 offset, and we are
   * interested in the R and G edges:
   *
   * +---G---+-------+
   * |   x o R   x   |
   * +-------+-------+
   *
   * Then, if one of these edge is enabled:
   *   Red:   `(0.75 * X + 0.25 * 1) => 0.25 or 1.0`
   *   Green: `(0.75 * 1 + 0.25 * X) => 0.75 or 1.0`
   *
   * This function will unpack the values `(mad + mul + round)`:
   * wolframalpha.com: `round(x * abs(5 * x - 5 * 0.75))` plot 0 to 1
   */
  e.x = e.x * math::abs(5.0f * e.x - 5.0f * 0.75f);
  return math::round(e);
}
 
static float4 SMAADecodeDiagBilinearAccess(float4 e)
{
  e.x = e.x * math::abs(5.0f * e.x - 5.0f * 0.75f);
  e.z = e.z * math::abs(5.0f * e.z - 5.0f * 0.75f);
  return math::round(e);
}
 
static float2 SMAASearchDiag1(
    SMAATexture2D(edgesTex), float2 texcoord, float2 dir, int2 size, float2 &e)
{
  float4 coord = float4(texcoord, -1.0f, 1.0f);
  float3 t = float3(1.0f / float2(size), 1.0f);
  while (coord.z < float(SMAA_MAX_SEARCH_STEPS_DIAG - 1) && coord.w > 0.9f) {
    float3 increment = mad(t, float3(dir, 1.0f), coord.xyz());
    coord.x = increment.x;
    coord.y = increment.y;
    coord.z = increment.z;
    e = SMAASamplePoint(edgesTex, coord.xy()).xy();
    coord.w = math::dot(e, float2(0.5f, 0.5f));
  }
  return coord.zw();
}
 
static float2 SMAASearchDiag2(
    SMAATexture2D(edgesTex), float2 texcoord, float2 dir, int2 size, float2 &e)
{
  float4 coord = float4(texcoord, -1.0f, 1.0f);
  coord.x += 0.25f / size.x; /* See @SearchDiag2Optimization */
  float3 t = float3(1.0f / float2(size), 1.0f);
  while (coord.z < float(SMAA_MAX_SEARCH_STEPS_DIAG - 1) && coord.w > 0.9f) {
    float3 increment = mad(t, float3(dir, 1.0f), coord.xyz());
    coord.x = increment.x;
    coord.y = increment.y;
    coord.z = increment.z;
 
    /* @SearchDiag2Optimization */
    /* Fetch both edges at once using bilinear filtering: */
    e = SMAASampleLevelZero(edgesTex, coord.xy()).xy();
    e = SMAADecodeDiagBilinearAccess(e);
 
    /* Non-optimized version: */
    // e.g = SMAASampleLevelZero(edgesTex, coord.xy).g;
    // e.r = SMAASampleLevelZeroOffset(edgesTex, coord.xy, int2(1, 0), size).r;
 
    coord.w = math::dot(e, float2(0.5f, 0.5f));
  }
  return coord.zw();
}
 
static float2 SMAAAreaDiag(SMAATexture2D(areaTex), float2 dist, float2 e, float offset)
{
  float2 texcoord = mad(
      float2(SMAA_AREATEX_MAX_DISTANCE_DIAG, SMAA_AREATEX_MAX_DISTANCE_DIAG), e, dist);
 
  /* We do a scale and bias for mapping to texel space: */
  texcoord = mad(SMAA_AREATEX_PIXEL_SIZE, texcoord, 0.5f * SMAA_AREATEX_PIXEL_SIZE);
 
  /* Diagonal areas are on the second half of the texture: */
  texcoord.x += 0.5f;
 
  /* Move to proper place, according to the sub-pixel offset: */
  texcoord.y += SMAA_AREATEX_SUBTEX_SIZE * offset;
 
  /* Do it! */
  return SMAA_AREATEX_SELECT(SMAASampleLevelZero(areaTex, texcoord));
}
 
static float2 SMAACalculateDiagWeights(SMAATexture2D(edgesTex),
                                       SMAATexture2D(areaTex),
                                       float2 texcoord,
                                       float2 e,
                                       float4 subsampleIndices,
                                       int2 size)
{
  float2 weights = float2(0.0f, 0.0f);
 
  /* Search for the line ends: */
  float4 d;
  float2 end;
  if (e.x > 0.0f) {
    float2 negative_diagonal = SMAASearchDiag1(
        SMAATexturePass2D(edgesTex), texcoord, float2(-1.0f, 1.0f), size, end);
    d.x = negative_diagonal.x;
    d.z = negative_diagonal.y;
    d.x += float(end.y > 0.9f);
  }
  else {
    d.x = 0.0f;
    d.z = 0.0f;
  }
  float2 positive_diagonal = SMAASearchDiag1(
      SMAATexturePass2D(edgesTex), texcoord, float2(1.0, -1.0), size, end);
  d.y = positive_diagonal.x;
  d.w = positive_diagonal.y;
 
  SMAA_BRANCH
  if (d.x + d.y > 2.0f) { /* `d.x + d.y + 1 > 3`. */
    /* Fetch the crossing edges: */
    float4 coords = float4(texcoord, texcoord) +
                    float4(-d.x + 0.25f, d.x, d.y, -d.y - 0.25f) / float4(size, size);
    float4 c;
    float2 left_edge = SMAASampleLevelZeroOffset(edgesTex, coords.xy(), int2(-1, 0), size).xy();
    float2 right_edge = SMAASampleLevelZeroOffset(edgesTex, coords.zw(), int2(1, 0), size).xy();
    c.x = left_edge.x;
    c.y = left_edge.y;
    c.z = right_edge.x;
    c.w = right_edge.y;
    float4 decoded_access = SMAADecodeDiagBilinearAccess(c);
    c.y = decoded_access.x;
    c.x = decoded_access.y;
    c.w = decoded_access.z;
    c.z = decoded_access.w;
 
    /* Non-optimized version: */
    // float4 coords = mad(float4(-d.x, d.x, d.y, -d.y), SMAA_RT_METRICS.xyxy, texcoord.xyxy);
    // float4 c;
    // c.x = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2(-1,  0), size).g;
    // c.y = SMAASampleLevelZeroOffset(edgesTex, coords.xy, int2( 0,  0), size).r;
    // c.z = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1,  0), size).g;
    // c.w = SMAASampleLevelZeroOffset(edgesTex, coords.zw, int2( 1, -1), size).r;
 
    /* Merge crossing edges at each side into a single value: */
    float2 cc = mad(float2(2.0f, 2.0f), float2(c.x, c.z), float2(c.y, c.w));
 
    /* Remove the crossing edge if we didn't found the end of the line: */
    SMAAMovc(math::step(0.9f, d.zw()), cc, float2(0.0f, 0.0f));
 
    /* Fetch the areas for this line: */
    weights += SMAAAreaDiag(SMAATexturePass2D(areaTex), d.xy(), cc, subsampleIndices.z);
  }
 
  /* Search for the line ends: */
  float2 negative_diagonal = SMAASearchDiag2(
      SMAATexturePass2D(edgesTex), texcoord, float2(-1.0f, -1.0f), size, end);
  d.x = negative_diagonal.x;
  d.z = negative_diagonal.y;
  if (SMAASamplePointOffset(edgesTex, texcoord, int2(1, 0), size).x > 0.0f) {
    float2 positive_diagonal = SMAASearchDiag2(
        SMAATexturePass2D(edgesTex), texcoord, float2(1.0f, 1.0f), size, end);
    d.y = positive_diagonal.x;
    d.w = positive_diagonal.y;
    d.y += float(end.y > 0.9f);
  }
  else {
    d.y = 0.0f;
    d.w = 0.0f;
  }
 
  SMAA_BRANCH
  if (d.x + d.y > 2.0f) { /* `d.x + d.y + 1 > 3` */
    /* Fetch the crossing edges: */
    float4 coords = float4(texcoord, texcoord) + float4(-d.x, -d.x, d.y, d.y) / float4(size, size);
    float4 c;
    c.x = SMAASampleLevelZeroOffset(edgesTex, coords.xy(), int2(-1, 0), size).y;
    c.y = SMAASampleLevelZeroOffset(edgesTex, coords.xy(), int2(0, -1), size).x;
    float2 left_edge = SMAASampleLevelZeroOffset(edgesTex, coords.zw(), int2(1, 0), size).xy();
    c.z = left_edge.y;
    c.w = left_edge.x;
    float2 cc = mad(float2(2.0f, 2.0f), float2(c.x, c.z), float2(c.y, c.w));
 
    /* Remove the crossing edge if we didn't found the end of the line: */
    SMAAMovc(math::step(0.9f, d.zw()), cc, float2(0.0f, 0.0f));
 
    /* Fetch the areas for this line: */
    float2 area = SMAAAreaDiag(SMAATexturePass2D(areaTex), d.xy(), cc, subsampleIndices.w).xy();
    weights.x += area.y;
    weights.y += area.x;
  }
 
  return weights;
}
#endif
 
/* ----------------------------------------------------------------------------
 * Horizontal/Vertical Search Functions */
 
static float SMAASearchLength(SMAATexture2D(searchTex), float2 e, float offset)
{
  /* The texture is flipped vertically, with left and right cases taking half
   * of the space horizontally: */
  float2 scale = SMAA_SEARCHTEX_SIZE * float2(0.5f, -1.0f);
  float2 bias = SMAA_SEARCHTEX_SIZE * float2(offset, 1.0f);
 
  /* Scale and bias to access texel centers: */
  scale += float2(-1.0f, 1.0f);
  bias += float2(0.5f, -0.5f);
 
  /* Convert from pixel coordinates to texcoords:
   * (We use SMAA_SEARCHTEX_PACKED_SIZE because the texture is cropped). */
  scale *= 1.0f / SMAA_SEARCHTEX_PACKED_SIZE;
  bias *= 1.0f / SMAA_SEARCHTEX_PACKED_SIZE;
 
  /* Lookup the search texture: */
  return SMAA_SEARCHTEX_SELECT(SMAASampleLevelZero(searchTex, mad(scale, e, bias)));
}
 
static float SMAASearchXLeft(
    SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end, int2 size)
{
  float2 e = float2(0.0f, 1.0f);
  while (texcoord.x > end && e.y > 0.8281f && /* Is there some edge not activated? */
         e.x == 0.0f) /* Or is there a crossing edge that breaks the line? */
  {
    e = SMAASampleLevelZero(edgesTex, texcoord).xy();
    texcoord = texcoord - float2(2.0f, 0.0f) / float2(size);
  }
 
  float offset = mad(
      -(255.0f / 127.0f), SMAASearchLength(SMAATexturePass2D(searchTex), e, 0.0f), 3.25f);
  return texcoord.x + offset / size.x;
 
  /* Non-optimized version:
   * We correct the previous (-0.25, -0.125) offset we applied: */
  // texcoord.x += 0.25 * SMAA_RT_METRICS.x;
 
  /* The searches are bias by 1, so adjust the coords accordingly: */
  // texcoord.x += SMAA_RT_METRICS.x;
 
  /* Disambiguate the length added by the last step: */
  // texcoord.x += 2.0 * SMAA_RT_METRICS.x; /* Undo last step. */
  // texcoord.x -= SMAA_RT_METRICS.x * (255.0 / 127.0) *
  //               SMAASearchLength(SMAATexturePass2D(searchTex), e, 0.0);
  // return mad(SMAA_RT_METRICS.x, offset, texcoord.x);
}
 
static float SMAASearchXRight(
    SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end, int2 size)
{
  float2 e = float2(0.0f, 1.0f);
  while (texcoord.x < end && e.y > 0.8281f && /* Is there some edge not activated? */
         e.x == 0.0f) /* Or is there a crossing edge that breaks the line? */
  {
    e = SMAASampleLevelZero(edgesTex, texcoord).xy();
    texcoord = texcoord + float2(2.0f, 0.0f) / float2(size);
  }
  float offset = mad(
      -(255.0f / 127.0f), SMAASearchLength(SMAATexturePass2D(searchTex), e, 0.5f), 3.25f);
  return texcoord.x - offset / size.x;
}
 
static float SMAASearchYUp(
    SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end, int2 size)
{
  float2 e = float2(1.0f, 0.0f);
  while (texcoord.y > end && e.x > 0.8281f && /* Is there some edge not activated? */
         e.y == 0.0f) /* Or is there a crossing edge that breaks the line? */
  {
    e = SMAASampleLevelZero(edgesTex, texcoord).xy();
    texcoord = texcoord - float2(0.0f, 2.0f) / float2(size);
  }
  float2 flipped_edge = float2(e.y, e.x);
  float offset = mad(-(255.0f / 127.0f),
                     SMAASearchLength(SMAATexturePass2D(searchTex), flipped_edge, 0.0f),
                     3.25f);
  return texcoord.y + offset / size.y;
}
 
static float SMAASearchYDown(
    SMAATexture2D(edgesTex), SMAATexture2D(searchTex), float2 texcoord, float end, int2 size)
{
  float2 e = float2(1.0f, 0.0f);
  while (texcoord.y < end && e.x > 0.8281f && /* Is there some edge not activated? */
         e.y == 0.0f) /* Or is there a crossing edge that breaks the line? */
  {
    e = SMAASampleLevelZero(edgesTex, texcoord).xy();
    texcoord = texcoord + float2(0.0f, 2.0f) / float2(size);
  }
  float2 flipped_edge = float2(e.y, e.x);
  float offset = mad(-(255.0f / 127.0f),
                     SMAASearchLength(SMAATexturePass2D(searchTex), flipped_edge, 0.5f),
                     3.25f);
  return texcoord.y - offset / size.y;
}
 
static float2 SMAAArea(SMAATexture2D(areaTex), float2 dist, float e1, float e2, float offset)
{
  /* Rounding prevents precision errors of bilinear filtering: */
  float2 texcoord = mad(float2(SMAA_AREATEX_MAX_DISTANCE, SMAA_AREATEX_MAX_DISTANCE),
                        math::round(4.0f * float2(e1, e2)),
                        dist);
 
  /* We do a scale and bias for mapping to texel space: */
  texcoord = mad(SMAA_AREATEX_PIXEL_SIZE, texcoord, 0.5f * SMAA_AREATEX_PIXEL_SIZE);
 
  /* Move to proper place, according to the sub-pixel offset: */
  texcoord.y = mad(SMAA_AREATEX_SUBTEX_SIZE, offset, texcoord.y);
 
  /* Do it! */
  return SMAA_AREATEX_SELECT(SMAASampleLevelZero(areaTex, texcoord));
}
 
/* ----------------------------------------------------------------------------
 * Corner Detection Functions */
 
static void SMAADetectHorizontalCornerPattern(SMAATexture2D(edgesTex),
                                              float2 &weights,
                                              float4 texcoord,
                                              float2 d,
                                              int2 size,
                                              int corner_rounding)
{
#if !defined(SMAA_DISABLE_CORNER_DETECTION)
  float2 leftRight = math::step(d, float2(d.y, d.x));
  float2 rounding = (1.0f - corner_rounding / 100.0f) * leftRight;
 
  rounding /= leftRight.x + leftRight.y; /* Reduce blending for pixels in the center of a line. */
 
  float2 factor = float2(1.0f, 1.0f);
  factor.x -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy(), int2(0, 1), size).x;
  factor.x -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw(), int2(1, 1), size).x;
  factor.y -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy(), int2(0, -2), size).x;
  factor.y -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw(), int2(1, -2), size).x;
 
  weights *= saturate(factor);
#endif
}
 
static void SMAADetectVerticalCornerPattern(SMAATexture2D(edgesTex),
                                            float2 &weights,
                                            float4 texcoord,
                                            float2 d,
                                            int2 size,
                                            int corner_rounding)
{
#if !defined(SMAA_DISABLE_CORNER_DETECTION)
  float2 leftRight = math::step(d, float2(d.y, d.x));
  float2 rounding = (1.0f - corner_rounding / 100.0f) * leftRight;
 
  rounding /= leftRight.x + leftRight.y;
 
  float2 factor = float2(1.0f, 1.0f);
  factor.x -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy(), int2(1, 0), size).y;
  factor.x -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw(), int2(1, 1), size).y;
  factor.y -= rounding.x * SMAASampleLevelZeroOffset(edgesTex, texcoord.xy(), int2(-2, 0), size).y;
  factor.y -= rounding.y * SMAASampleLevelZeroOffset(edgesTex, texcoord.zw(), int2(-2, 1), size).y;
 
  weights *= saturate(factor);
#endif
}
 
/* ----------------------------------------------------------------------------
 * Blending Weight Calculation Pixel Shader (Second Pass) */
 
static float4 SMAABlendingWeightCalculationPS(float2 texcoord,
                                              float2 pixcoord,
                                              float4 offset[3],
                                              MemoryBuffer *edgesTex,
                                              MemoryBuffer *areaTex,
                                              MemoryBuffer *searchTex,
                                              float4 subsampleIndices,
                                              int2 size,
                                              int corner_rounding)
{
  /* Just pass zero for SMAA 1x, see @SUBSAMPLE_INDICES. */
  float4 weights = float4(0.0f, 0.0f, 0.0f, 0.0f);
 
  float2 e = SMAASamplePoint(edgesTex, texcoord).xy();
 
  SMAA_BRANCH
  if (e.y > 0.0f) { /* Edge at north. */
#if !defined(SMAA_DISABLE_DIAG_DETECTION)
    /* Diagonals have both north and west edges, so searching for them in
     * one of the boundaries is enough. */
    float2 diagonal_weights = SMAACalculateDiagWeights(SMAATexturePass2D(edgesTex),
                                                       SMAATexturePass2D(areaTex),
                                                       texcoord,
                                                       e,
                                                       subsampleIndices,
                                                       size);
 
    weights.x = diagonal_weights.x;
    weights.y = diagonal_weights.y;
 
    /* We give priority to diagonals, so if we find a diagonal we skip
     * horizontal/vertical processing. */
    SMAA_BRANCH
    if (weights.x == -weights.y) { /* `weights.x + weights.y == 0.0` */
#endif
 
      float2 d;
 
      /* Find the distance to the left: */
      float3 coords;
      coords.x = SMAASearchXLeft(SMAATexturePass2D(edgesTex),
                                 SMAATexturePass2D(searchTex),
                                 offset[0].xy(),
                                 offset[2].x,
                                 size);
      coords.y =
          offset[1].y;  // offset[1].y = texcoord.y - 0.25 * SMAA_RT_METRICS.y (@CROSSING_OFFSET)
      d.x = coords.x;
 
      /* Now fetch the left crossing edges, two at a time using bilinear
       * filtering. Sampling at -0.25 (see @CROSSING_OFFSET) enables to
       * discern what value each edge has: */
      float e1 = SMAASampleLevelZero(edgesTex, coords.xy()).x;
 
      /* Find the distance to the right: */
      coords.z = SMAASearchXRight(SMAATexturePass2D(edgesTex),
                                  SMAATexturePass2D(searchTex),
                                  offset[0].zw(),
                                  offset[2].y,
                                  size);
      d.y = coords.z;
 
      /* We want the distances to be in pixel units (doing this here allows
       * better interleaving of arithmetic and memory accesses): */
      d = math::abs(math::round(mad(float2(size.x), d, -float2(pixcoord.x))));
 
      /* SMAAArea below needs a sqrt, as the areas texture is compressed quadratically: */
      float2 sqrt_d = math::sqrt(d);
 
      /* Fetch the right crossing edges: */
      float e2 =
          SMAASampleLevelZeroOffset(edgesTex, float2(coords.z, coords.y), int2(1, 0), size).x;
 
      /* Ok, we know how this pattern looks like, now it is time for getting the actual area: */
      float2 area = SMAAArea(SMAATexturePass2D(areaTex), sqrt_d, e1, e2, subsampleIndices.y);
      weights.x = area.x;
      weights.y = area.y;
 
      /* Fix corners: */
      coords.y = texcoord.y;
 
      float2 corner_weight = weights.xy();
      SMAADetectHorizontalCornerPattern(SMAATexturePass2D(edgesTex),
                                        corner_weight,
                                        float4(coords.xy(), coords.z, coords.y),
                                        d,
                                        size,
                                        corner_rounding);
      weights.x = corner_weight.x;
      weights.y = corner_weight.y;
 
#if !defined(SMAA_DISABLE_DIAG_DETECTION)
    }
    else
      e.x = 0.0f; /* Skip vertical processing. */
#endif
  }
 
  SMAA_BRANCH
  if (e.x > 0.0f) { /* Edge at west. */
    float2 d;
 
    /* Find the distance to the top: */
    float3 coords;
    coords.y = SMAASearchYUp(SMAATexturePass2D(edgesTex),
                             SMAATexturePass2D(searchTex),
                             offset[1].xy(),
                             offset[2].z,
                             size);
    coords.x = offset[0].x;  // offset[1].x = texcoord.x - 0.25 * SMAA_RT_METRICS.x;
    d.x = coords.y;
 
    /* Fetch the top crossing edges: */
    float e1 = SMAASampleLevelZero(edgesTex, coords.xy()).y;
 
    /* Find the distance to the bottom: */
    coords.z = SMAASearchYDown(SMAATexturePass2D(edgesTex),
                               SMAATexturePass2D(searchTex),
                               offset[1].zw(),
                               offset[2].w,
                               size);
    d.y = coords.z;
 
    /* We want the distances to be in pixel units: */
    d = math::abs(math::round(mad(float2(size.y), d, -float2(pixcoord.y))));
 
    /* SMAAArea below needs a sqrt, as the areas texture is compressed quadratically: */
    float2 sqrt_d = math::sqrt(d);
 
    /* Fetch the bottom crossing edges: */
    float e2 = SMAASampleLevelZeroOffset(edgesTex, float2(coords.x, coords.z), int2(0, 1), size).y;
 
    /* Get the area for this direction: */
    float2 area = SMAAArea(SMAATexturePass2D(areaTex), sqrt_d, e1, e2, subsampleIndices.x);
    weights.z = area.x;
    weights.w = area.y;
 
    /* Fix corners: */
    coords.x = texcoord.x;
 
    float2 corner_weight = weights.zw();
    SMAADetectVerticalCornerPattern(SMAATexturePass2D(edgesTex),
                                    corner_weight,
                                    float4(coords.xy(), coords.x, coords.z),
                                    d,
                                    size,
                                    corner_rounding);
    weights.z = corner_weight.x;
    weights.w = corner_weight.y;
  }
 
  return weights;
}
 
/* ----------------------------------------------------------------------------
 * Neighborhood Blending Pixel Shader (Third Pass) */
 
static float4 SMAANeighborhoodBlendingPS(float2 texcoord,
                                         float4 offset,
                                         SMAATexture2D(colorTex),
                                         SMAATexture2D(blendTex),
#if SMAA_REPROJECTION
                                         SMAATexture2D(velocityTex),
#endif
                                         int2 size)
{
  /* Fetch the blending weights for current pixel: */
  float4 a;
  a.x = SMAASample(blendTex, offset.xy()).w;  // Right
  a.y = SMAASample(blendTex, offset.zw()).y;  // Top
  a.z = SMAASample(blendTex, texcoord).z;     // Left
  a.w = SMAASample(blendTex, texcoord).x;     // Bottom
 
  /* Is there any blending weight with a value greater than 0.0? */
  SMAA_BRANCH
  if (math::dot(a, float4(1.0f, 1.0f, 1.0f, 1.0f)) < 1e-5f) {
    float4 color = SMAASampleLevelZero(colorTex, texcoord);
 
#if SMAA_REPROJECTION
    float2 velocity = SMAA_DECODE_VELOCITY(SMAASampleLevelZero(velocityTex, texcoord));
 
    /* Pack velocity into the alpha channel: */
    color.a = math::sqrt(5.0f * math::length(velocity));
#endif
 
    return color;
  }
  else {
    bool h = math::max(a.x, a.z) > math::max(a.y, a.w); /* `max(horizontal) > max(vertical)`. */
 
    /* Calculate the blending offsets: */
    float4 blendingOffset = float4(0.0f, a.y, 0.0f, a.w);
    float2 blendingWeight = float2(a.y, a.w);
    SMAAMovc(float4(h), blendingOffset, float4(a.x, 0.0f, a.z, 0.0f));
    SMAAMovc(float2(h), blendingWeight, float2(a.x, a.z));
    blendingWeight /= math::dot(blendingWeight, float2(1.0f, 1.0f));
 
    /* Calculate the texture coordinates: */
    float4 blendingCoord = float4(texcoord, texcoord) + blendingOffset / float4(size, -size);
 
    /* We exploit bilinear filtering to mix current pixel with the chosen neighbor: */
    float4 color = blendingWeight.x * SMAASampleLevelZero(colorTex, blendingCoord.xy());
    color += blendingWeight.y * SMAASampleLevelZero(colorTex, blendingCoord.zw());
 
#if SMAA_REPROJECTION
    /* Anti-alias velocity for proper reprojection in a later stage: */
    float2 velocity = blendingWeight.x *
                      SMAA_DECODE_VELOCITY(SMAASampleLevelZero(velocityTex, blendingCoord.xy()));
    velocity += blendingWeight.y *
                SMAA_DECODE_VELOCITY(SMAASampleLevelZero(velocityTex, blendingCoord.zw()));
 
    /* Pack velocity into the alpha channel: */
    color.a = math::sqrt(5.0f * math::length(velocity));
#endif
 
    return color;
  }
}
 
SMAAOperation::SMAAOperation()
{
  this->add_input_socket(DataType::Color);
  this->add_output_socket(DataType::Color);
  flags_.can_be_constant = true;
}
 
void SMAAOperation::get_area_of_interest(const int /*input_idx*/,
                                         const rcti &output_area,
                                         rcti &r_input_area)
{
  r_input_area.xmax = output_area.xmax +
                      math::max(SMAA_MAX_SEARCH_STEPS, SMAA_MAX_SEARCH_STEPS_DIAG + 1);
  r_input_area.xmin = output_area.xmin - math::max(math::max(SMAA_MAX_SEARCH_STEPS - 1, 1),
                                                   SMAA_MAX_SEARCH_STEPS_DIAG + 1);
  r_input_area.ymax = output_area.ymax +
                      math::max(SMAA_MAX_SEARCH_STEPS, SMAA_MAX_SEARCH_STEPS_DIAG);
  r_input_area.ymin = output_area.ymin - math::max(math::max(SMAA_MAX_SEARCH_STEPS - 1, 1),
                                                   SMAA_MAX_SEARCH_STEPS_DIAG);
}
 
void SMAAOperation::update_memory_buffer(MemoryBuffer *output,
                                         const rcti & /*area*/,
                                         Span<MemoryBuffer *> inputs)
{
  const MemoryBuffer *image = inputs[0];
  if (image->is_a_single_elem()) {
    copy_v4_v4(output->get_elem(0, 0), image->get_elem(0, 0));
    return;
  }
 
  const int2 size = int2(image->get_width(), image->get_height());
  MemoryBuffer edges(DataType::Float2, size.x, size.y);
 
  float3 luminance_coefficients;
  IMB_colormanagement_get_luminance_coefficients(luminance_coefficients);
 
  threading::parallel_for(IndexRange(size.y), 1, [&](const IndexRange sub_y_range) {
    for (const int64_t y : sub_y_range) {
      for (const int64_t x : IndexRange(size.x)) {
        int2 texel = int2(x, y);
        float2 coordinates = (float2(texel) + float2(0.5f)) / float2(size);
 
        float4 offset[3];
        SMAAEdgeDetectionVS(coordinates, size, offset);
 
        float2 edge = SMAALumaEdgeDetectionPS(coordinates,
                                              offset,
                                              image,
                                              threshold_,
                                              luminance_coefficients,
                                              local_contrast_adaptation_factor_);
        copy_v2_v2(edges.get_elem(texel.x, texel.y), edge);
      }
    }
  });
 
  MemoryBuffer blending_weights(DataType::Color, size.x, size.y);
 
  MemoryBuffer area_texture(DataType::Float2, AREATEX_WIDTH, AREATEX_HEIGHT);
  area_texture.copy_from(areaTexBytes, area_texture.get_rect());
 
  MemoryBuffer search_texture(DataType::Value, SEARCHTEX_WIDTH, SEARCHTEX_HEIGHT);
  search_texture.copy_from(searchTexBytes, search_texture.get_rect());
 
  threading::parallel_for(IndexRange(size.y), 1, [&](const IndexRange sub_y_range) {
    for (const int64_t y : sub_y_range) {
      for (const int64_t x : IndexRange(size.x)) {
        int2 texel = int2(x, y);
        float2 coordinates = (float2(texel) + float2(0.5f)) / float2(size);
 
        float4 offset[3];
        float2 pixel_coordinates;
        SMAABlendingWeightCalculationVS(coordinates, size, pixel_coordinates, offset);
 
        float4 weights = SMAABlendingWeightCalculationPS(coordinates,
                                                         pixel_coordinates,
                                                         offset,
                                                         &edges,
                                                         &area_texture,
                                                         &search_texture,
                                                         float4(0.0f),
                                                         size,
                                                         corner_rounding_);
        copy_v4_v4(blending_weights.get_elem(texel.x, texel.y), weights);
      }
    }
  });
 
  threading::parallel_for(IndexRange(size.y), 1, [&](const IndexRange sub_y_range) {
    for (const int64_t y : sub_y_range) {
      for (const int64_t x : IndexRange(size.x)) {
        int2 texel = int2(x, y);
        float2 coordinates = (float2(texel) + float2(0.5f)) / float2(size);
 
        float4 offset;
        SMAANeighborhoodBlendingVS(coordinates, size, offset);
 
        float4 result = SMAANeighborhoodBlendingPS(
            coordinates, offset, image, &blending_weights, size);
        copy_v4_v4(output->get_elem(texel.x, texel.y), result);
      }
    }
  });
}
 
}  // namespace blender::compositor