Merge pull request #101515 from allenwp/agx-negative-optimizations

Optimize AgX tonemapper's handling of negative values

drivers/gles3/shaders/tonemap_inc.glsl

@@ -86,7 +86,7 @@ vec3 tonemap_aces(vec3 color, float p_white) {
 
 // Polynomial approximation of EaryChow's AgX sigmoid curve.
 // x must be within the range [0.0, 1.0]
-vec3 agx_default_contrast_approx(vec3 x) {
+vec3 agx_contrast_approx(vec3 x) {
 	// Generated with Excel trendline
 	// Input data: Generated using python sigmoid with EaryChow's configuration and 57 steps
 	// Additional padding values were added to give correct intersections at 0.0 and 1.0
@@ -96,25 +96,21 @@ vec3 agx_default_contrast_approx(vec3 x) {
 	return 0.021 * x + 4.0111 * x2 - 25.682 * x2 * x + 70.359 * x4 - 74.778 * x4 * x + 27.069 * x4 * x2;
 }
 
-const mat3 LINEAR_SRGB_TO_LINEAR_REC2020 = mat3(
-		vec3(0.6274, 0.0691, 0.0164),
-		vec3(0.3293, 0.9195, 0.0880),
-		vec3(0.0433, 0.0113, 0.8956));
-
 // This is an approximation and simplification of EaryChow's AgX implementation that is used by Blender.
 // This code is based off of the script that generates the AgX_Base_sRGB.cube LUT that Blender uses.
 // Source: https://github.com/EaryChow/AgX_LUT_Gen/blob/main/AgXBasesRGB.py
 vec3 tonemap_agx(vec3 color) {
-	const mat3 agx_inset_matrix = mat3(
-			0.856627153315983, 0.137318972929847, 0.11189821299995,
-			0.0951212405381588, 0.761241990602591, 0.0767994186031903,
-			0.0482516061458583, 0.101439036467562, 0.811302368396859);
+	// Combined linear sRGB to linear Rec 2020 and Blender AgX inset matrices:
+	const mat3 srgb_to_rec2020_agx_inset_matrix = mat3(
+			0.54490813676363087053, 0.14044005884001287035, 0.088827411851915368603,
+			0.37377945959812267119, 0.75410959864013760045, 0.17887712465043811023,
+			0.081384976686407536266, 0.10543358536857773485, 0.73224999956948382528);
 
 	// Combined inverse AgX outset matrix and linear Rec 2020 to linear sRGB matrices.
 	const mat3 agx_outset_rec2020_to_srgb_matrix = mat3(
-			1.9648846919172409596, -0.29937618452442253746, -0.16440106280678278299,
-			-0.85594737466675834968, 1.3263980951083531115, -0.23819967517076844919,
-			-0.10883731725048386702, -0.02702191058393112346, 1.4025007379775505276);
+			1.9645509602733325934, -0.29932243390911083839, -0.16436833806080403409,
+			-0.85585845117807513559, 1.3264510741502356555, -0.23822464068860595117,
+			-0.10886710826831608324, -0.027084020983874825605, 1.402665347143271889);
 
 	// LOG2_MIN      = -10.0
 	// LOG2_MAX      =  +6.5
@@ -122,26 +118,32 @@ vec3 tonemap_agx(vec3 color) {
 	const float min_ev = -12.4739311883324; // log2(pow(2, LOG2_MIN) * MIDDLE_GRAY)
 	const float max_ev = 4.02606881166759; // log2(pow(2, LOG2_MAX) * MIDDLE_GRAY)
 
-	// Do AGX in rec2020 to match Blender.
-	color = LINEAR_SRGB_TO_LINEAR_REC2020 * color;
+	// Large negative values in one channel and large positive values in other
+	// channels can result in a colour that appears darker and more saturated than
+	// desired after passing it through the inset matrix. For this reason, it is
+	// best to prevent negative input values.
+	// This is done before the Rec. 2020 transform to allow the Rec. 2020
+	// transform to be combined with the AgX inset matrix. This results in a loss
+	// of color information that could be correctly interpreted within the
+	// Rec. 2020 color space as positive RGB values, but it is less common for Godot
+	// to provide this function with negative sRGB values and therefore not worth
+	// the performance cost of an additional matrix multiplication.
+	// A value of 2e-10 intentionally introduces insignificant error to prevent
+	// log2(0.0) after the inset matrix is applied; color will be >= 1e-10 after
+	// the matrix transform.
+	color = max(color, 2e-10);
 
-	// Preventing negative values is required for the AgX inset matrix to behave correctly.
-	// This could also be done before the Rec. 2020 transform, allowing the transform to
-	// be combined with the AgX inset matrix, but doing this causes a loss of color information
-	// that could be correctly interpreted within the Rec. 2020 color space.
-	color = max(color, vec3(0.0));
-
-	color = agx_inset_matrix * color;
+	// Do AGX in rec2020 to match Blender and then apply inset matrix.
+	color = srgb_to_rec2020_agx_inset_matrix * color;
 
 	// Log2 space encoding.
-	color = max(color, 1e-10); // Prevent log2(0.0). Possibly unnecessary.
-	// Must be clamped because agx_blender_default_contrast_approx may not work
+	// Must be clamped because agx_contrast_approx may not work
 	// well with values outside of the range [0.0, 1.0]
 	color = clamp(log2(color), min_ev, max_ev);
 	color = (color - min_ev) / (max_ev - min_ev);
 
 	// Apply sigmoid function approximation.
-	color = agx_default_contrast_approx(color);
+	color = agx_contrast_approx(color);
 
 	// Convert back to linear before applying outset matrix.
 	color = pow(color, vec3(2.4));
@@ -149,9 +151,9 @@ vec3 tonemap_agx(vec3 color) {
 	// Apply outset to make the result more chroma-laden and then go back to linear sRGB.
 	color = agx_outset_rec2020_to_srgb_matrix * color;
 
-	// Simply hard clip instead of Blender's complex lusRGB.compensate_low_side.
-	color = max(color, vec3(0.0));
-
+	// Blender's lusRGB.compensate_low_side is too complex for this shader, so
+	// simply return the color, even if it has negative components. These negative
+	// components may be useful for subsequent color adjustments.
 	return color;
 }
 

servers/rendering/renderer_rd/shaders/effects/tonemap.glsl

@@ -266,7 +266,7 @@ vec3 tonemap_aces(vec3 color, float white) {
 
 // Polynomial approximation of EaryChow's AgX sigmoid curve.
 // x must be within the range [0.0, 1.0]
-vec3 agx_default_contrast_approx(vec3 x) {
+vec3 agx_contrast_approx(vec3 x) {
 	// Generated with Excel trendline
 	// Input data: Generated using python sigmoid with EaryChow's configuration and 57 steps
 	// Additional padding values were added to give correct intersections at 0.0 and 1.0
@@ -276,25 +276,21 @@ vec3 agx_default_contrast_approx(vec3 x) {
 	return 0.021 * x + 4.0111 * x2 - 25.682 * x2 * x + 70.359 * x4 - 74.778 * x4 * x + 27.069 * x4 * x2;
 }
 
-const mat3 LINEAR_SRGB_TO_LINEAR_REC2020 = mat3(
-		vec3(0.6274, 0.0691, 0.0164),
-		vec3(0.3293, 0.9195, 0.0880),
-		vec3(0.0433, 0.0113, 0.8956));
-
 // This is an approximation and simplification of EaryChow's AgX implementation that is used by Blender.
 // This code is based off of the script that generates the AgX_Base_sRGB.cube LUT that Blender uses.
 // Source: https://github.com/EaryChow/AgX_LUT_Gen/blob/main/AgXBasesRGB.py
 vec3 tonemap_agx(vec3 color) {
-	const mat3 agx_inset_matrix = mat3(
-			0.856627153315983, 0.137318972929847, 0.11189821299995,
-			0.0951212405381588, 0.761241990602591, 0.0767994186031903,
-			0.0482516061458583, 0.101439036467562, 0.811302368396859);
+	// Combined linear sRGB to linear Rec 2020 and Blender AgX inset matrices:
+	const mat3 srgb_to_rec2020_agx_inset_matrix = mat3(
+			0.54490813676363087053, 0.14044005884001287035, 0.088827411851915368603,
+			0.37377945959812267119, 0.75410959864013760045, 0.17887712465043811023,
+			0.081384976686407536266, 0.10543358536857773485, 0.73224999956948382528);
 
 	// Combined inverse AgX outset matrix and linear Rec 2020 to linear sRGB matrices.
 	const mat3 agx_outset_rec2020_to_srgb_matrix = mat3(
-			1.9648846919172409596, -0.29937618452442253746, -0.16440106280678278299,
-			-0.85594737466675834968, 1.3263980951083531115, -0.23819967517076844919,
-			-0.10883731725048386702, -0.02702191058393112346, 1.4025007379775505276);
+			1.9645509602733325934, -0.29932243390911083839, -0.16436833806080403409,
+			-0.85585845117807513559, 1.3264510741502356555, -0.23822464068860595117,
+			-0.10886710826831608324, -0.027084020983874825605, 1.402665347143271889);
 
 	// LOG2_MIN      = -10.0
 	// LOG2_MAX      =  +6.5
@@ -302,26 +298,32 @@ vec3 tonemap_agx(vec3 color) {
 	const float min_ev = -12.4739311883324; // log2(pow(2, LOG2_MIN) * MIDDLE_GRAY)
 	const float max_ev = 4.02606881166759; // log2(pow(2, LOG2_MAX) * MIDDLE_GRAY)
 
-	// Do AGX in rec2020 to match Blender.
-	color = LINEAR_SRGB_TO_LINEAR_REC2020 * color;
+	// Large negative values in one channel and large positive values in other
+	// channels can result in a colour that appears darker and more saturated than
+	// desired after passing it through the inset matrix. For this reason, it is
+	// best to prevent negative input values.
+	// This is done before the Rec. 2020 transform to allow the Rec. 2020
+	// transform to be combined with the AgX inset matrix. This results in a loss
+	// of color information that could be correctly interpreted within the
+	// Rec. 2020 color space as positive RGB values, but it is less common for Godot
+	// to provide this function with negative sRGB values and therefore not worth
+	// the performance cost of an additional matrix multiplication.
+	// A value of 2e-10 intentionally introduces insignificant error to prevent
+	// log2(0.0) after the inset matrix is applied; color will be >= 1e-10 after
+	// the matrix transform.
+	color = max(color, 2e-10);
 
-	// Preventing negative values is required for the AgX inset matrix to behave correctly.
-	// This could also be done before the Rec. 2020 transform, allowing the transform to
-	// be combined with the AgX inset matrix, but doing this causes a loss of color information
-	// that could be correctly interpreted within the Rec. 2020 color space.
-	color = max(color, vec3(0.0));
-
-	color = agx_inset_matrix * color;
+	// Do AGX in rec2020 to match Blender and then apply inset matrix.
+	color = srgb_to_rec2020_agx_inset_matrix * color;
 
 	// Log2 space encoding.
-	color = max(color, 1e-10); // Prevent log2(0.0). Possibly unnecessary.
-	// Must be clamped because agx_blender_default_contrast_approx may not work
+	// Must be clamped because agx_contrast_approx may not work
 	// well with values outside of the range [0.0, 1.0]
 	color = clamp(log2(color), min_ev, max_ev);
 	color = (color - min_ev) / (max_ev - min_ev);
 
 	// Apply sigmoid function approximation.
-	color = agx_default_contrast_approx(color);
+	color = agx_contrast_approx(color);
 
 	// Convert back to linear before applying outset matrix.
 	color = pow(color, vec3(2.4));
@@ -329,9 +331,9 @@ vec3 tonemap_agx(vec3 color) {
 	// Apply outset to make the result more chroma-laden and then go back to linear sRGB.
 	color = agx_outset_rec2020_to_srgb_matrix * color;
 
-	// Simply hard clip instead of Blender's complex lusRGB.compensate_low_side.
-	color = max(color, vec3(0.0));
-
+	// Blender's lusRGB.compensate_low_side is too complex for this shader, so
+	// simply return the color, even if it has negative components. These negative
+	// components may be useful for subsequent color adjustments.
 	return color;
 }