Merge pull request #107698 from BastiaanOlij/fix_gles3_stereo_output

Fix GLES3 stereo output (sRGB + lens distortion)

Author: akien-mga

drivers/gles3/effects/copy_effects.cpp

@@ -115,31 +115,56 @@ CopyEffects::~CopyEffects() {
 	copy.shader.version_free(copy.shader_version);
 }
 
-void CopyEffects::copy_to_rect(const Rect2 &p_rect) {
-	bool success = copy.shader.version_bind_shader(copy.shader_version, CopyShaderGLES3::MODE_COPY_SECTION);
+void CopyEffects::copy_to_rect(const Rect2 &p_rect, bool p_linear_to_srgb) {
+	uint64_t specializations = p_linear_to_srgb ? CopyShaderGLES3::CONVERT_LINEAR_TO_SRGB : 0;
+
+	bool success = copy.shader.version_bind_shader(copy.shader_version, CopyShaderGLES3::MODE_COPY_SECTION, specializations);
 	if (!success) {
 		return;
 	}
 
-	copy.shader.version_set_uniform(CopyShaderGLES3::COPY_SECTION, p_rect.position.x, p_rect.position.y, p_rect.size.x, p_rect.size.y, copy.shader_version, CopyShaderGLES3::MODE_COPY_SECTION);
+	copy.shader.version_set_uniform(CopyShaderGLES3::COPY_SECTION, p_rect.position.x, p_rect.position.y, p_rect.size.x, p_rect.size.y, copy.shader_version, CopyShaderGLES3::MODE_COPY_SECTION, specializations);
 	draw_screen_quad();
 }
 
-void CopyEffects::copy_to_rect_3d(const Rect2 &p_rect, float p_layer, int p_type, float p_lod) {
+void CopyEffects::copy_to_rect_3d(const Rect2 &p_rect, float p_layer, int p_type, float p_lod, bool p_linear_to_srgb) {
 	ERR_FAIL_COND(p_type != Texture::TYPE_LAYERED && p_type != Texture::TYPE_3D);
 
 	CopyShaderGLES3::ShaderVariant variant = p_type == Texture::TYPE_LAYERED
 			? CopyShaderGLES3::MODE_COPY_SECTION_2D_ARRAY
 			: CopyShaderGLES3::MODE_COPY_SECTION_3D;
+	uint64_t specializations = p_linear_to_srgb ? CopyShaderGLES3::CONVERT_LINEAR_TO_SRGB : 0;
 
-	bool success = copy.shader.version_bind_shader(copy.shader_version, variant);
+	bool success = copy.shader.version_bind_shader(copy.shader_version, variant, specializations);
 	if (!success) {
 		return;
 	}
 
-	copy.shader.version_set_uniform(CopyShaderGLES3::COPY_SECTION, p_rect.position.x, p_rect.position.y, p_rect.size.x, p_rect.size.y, copy.shader_version, variant);
-	copy.shader.version_set_uniform(CopyShaderGLES3::LAYER, p_layer, copy.shader_version, variant);
-	copy.shader.version_set_uniform(CopyShaderGLES3::LOD, p_lod, copy.shader_version, variant);
+	copy.shader.version_set_uniform(CopyShaderGLES3::COPY_SECTION, p_rect.position.x, p_rect.position.y, p_rect.size.x, p_rect.size.y, copy.shader_version, variant, specializations);
+	copy.shader.version_set_uniform(CopyShaderGLES3::LAYER, p_layer, copy.shader_version, variant, specializations);
+	copy.shader.version_set_uniform(CopyShaderGLES3::LOD, p_lod, copy.shader_version, variant, specializations);
+	draw_screen_quad();
+}
+
+void CopyEffects::copy_with_lens_distortion(const Rect2 &p_rect, float p_layer, const Vector2 &p_eye_center, float p_k1, float p_k2, float p_upscale, float p_aspect_ration, bool p_linear_to_srgb) {
+	CopyShaderGLES3::ShaderVariant variant = CopyShaderGLES3::MODE_LENS_DISTORTION;
+
+	uint64_t specializations = p_linear_to_srgb ? CopyShaderGLES3::CONVERT_LINEAR_TO_SRGB : 0;
+
+	bool success = copy.shader.version_bind_shader(copy.shader_version, variant, specializations);
+	if (!success) {
+		return;
+	}
+
+	copy.shader.version_set_uniform(CopyShaderGLES3::COPY_SECTION, p_rect.position.x, p_rect.position.y, p_rect.size.x, p_rect.size.y, copy.shader_version, variant, specializations);
+	copy.shader.version_set_uniform(CopyShaderGLES3::LAYER, p_layer, copy.shader_version, variant, specializations);
+	copy.shader.version_set_uniform(CopyShaderGLES3::LOD, 0.0, copy.shader_version, variant, specializations);
+	copy.shader.version_set_uniform(CopyShaderGLES3::EYE_CENTER, p_eye_center.x, p_eye_center.y, copy.shader_version, variant, specializations);
+	copy.shader.version_set_uniform(CopyShaderGLES3::K1, p_k1, copy.shader_version, variant, specializations);
+	copy.shader.version_set_uniform(CopyShaderGLES3::K2, p_k1, copy.shader_version, variant, specializations);
+	copy.shader.version_set_uniform(CopyShaderGLES3::UPSCALE, p_upscale, copy.shader_version, variant, specializations);
+	copy.shader.version_set_uniform(CopyShaderGLES3::ASPECT_RATIO, p_aspect_ration, copy.shader_version, variant, specializations);
+
 	draw_screen_quad();
 }
 

drivers/gles3/effects/copy_effects.h

@@ -60,8 +60,9 @@ class CopyEffects {
 	~CopyEffects();
 
 	// These functions assume that a framebuffer and texture are bound already. They only manage the shader, uniforms, and vertex array.
-	void copy_to_rect(const Rect2 &p_rect);
-	void copy_to_rect_3d(const Rect2 &p_rect, float p_layer, int p_type, float p_lod = 0.0f);
+	void copy_to_rect(const Rect2 &p_rect, bool p_linear_to_srgb = false);
+	void copy_to_rect_3d(const Rect2 &p_rect, float p_layer, int p_type, float p_lod = 0.0f, bool p_linear_to_srgb = false);
+	void copy_with_lens_distortion(const Rect2 &p_rect, float p_layer, const Vector2 &p_eye_center, float p_k1, float p_k2, float p_upscale, float p_aspect_ration, bool p_linear_to_srgb = false);
 	void copy_to_and_from_rect(const Rect2 &p_rect);
 	void copy_screen(float p_multiply = 1.0);
 	void copy_cube_to_rect(const Rect2 &p_rect);

drivers/gles3/rasterizer_gles3.cpp

@@ -387,18 +387,26 @@ RasterizerGLES3::RasterizerGLES3() {
 RasterizerGLES3::~RasterizerGLES3() {
 }
 
-void RasterizerGLES3::_blit_render_target_to_screen(RID p_render_target, DisplayServer::WindowID p_screen, const Rect2 &p_screen_rect, uint32_t p_layer, bool p_first) {
-	GLES3::RenderTarget *rt = GLES3::TextureStorage::get_singleton()->get_render_target(p_render_target);
+void RasterizerGLES3::_blit_render_target_to_screen(DisplayServer::WindowID p_screen, const BlitToScreen &p_blit, bool p_first) {
+	GLES3::RenderTarget *rt = GLES3::TextureStorage::get_singleton()->get_render_target(p_blit.render_target);
 
 	ERR_FAIL_NULL(rt);
 
 	// We normally render to the render target upside down, so flip Y when blitting to the screen.
 	bool flip_y = true;
+	bool linear_to_srgb = false;
 	if (rt->overridden.color.is_valid()) {
 		// If we've overridden the render target's color texture, that means we
 		// didn't render upside down, so we don't need to flip it.
 		// We're probably rendering directly to an XR device.
 		flip_y = false;
+
+		// It is 99% likely our texture uses the GL_SRGB8_ALPHA8 texture format in
+		// which case we have a GPU sRGB to Linear conversion on texture read.
+		// We need to counter this.
+		// Unfortunately we do not have an API to check this as Godot does not
+		// track this.
+		linear_to_srgb = true;
 	}
 
 #ifdef WINDOWS_ENABLED
@@ -410,7 +418,7 @@ void RasterizerGLES3::_blit_render_target_to_screen(RID p_render_target, Display
 	glBindFramebuffer(GL_DRAW_FRAMEBUFFER, GLES3::TextureStorage::system_fbo);
 
 	if (p_first) {
-		if (p_screen_rect.position != Vector2() || p_screen_rect.size != rt->size) {
+		if (p_blit.dst_rect.position != Vector2() || p_blit.dst_rect.size != rt->size) {
 			// Viewport doesn't cover entire window so clear window to black before blitting.
 			// Querying the actual window size from the DisplayServer would deadlock in separate render thread mode,
 			// so let's set the biggest viewport the implementation supports, to be sure the window is fully covered.
@@ -421,7 +429,7 @@ void RasterizerGLES3::_blit_render_target_to_screen(RID p_render_target, Display
 		}
 	}
 
-	Vector2 screen_rect_end = p_screen_rect.get_end();
+	Vector2 screen_rect_end = p_blit.dst_rect.get_end();
 
 	// Adreno (TM) 3xx devices have a bug that create wrong Landscape rotation of 180 degree
 	// Reversing both the X and Y axis is equivalent to rotating 180 degrees
@@ -431,8 +439,8 @@ void RasterizerGLES3::_blit_render_target_to_screen(RID p_render_target, Display
 		flip_x = !flip_x;
 	}
 
-	Vector2 p1 = Vector2(flip_x ? screen_rect_end.x : p_screen_rect.position.x, flip_y ? screen_rect_end.y : p_screen_rect.position.y);
-	Vector2 p2 = Vector2(flip_x ? p_screen_rect.position.x : screen_rect_end.x, flip_y ? p_screen_rect.position.y : screen_rect_end.y);
+	Vector2 p1 = Vector2(flip_x ? screen_rect_end.x : p_blit.dst_rect.position.x, flip_y ? screen_rect_end.y : p_blit.dst_rect.position.y);
+	Vector2 p2 = Vector2(flip_x ? p_blit.dst_rect.position.x : screen_rect_end.x, flip_y ? p_blit.dst_rect.position.y : screen_rect_end.y);
 	Vector2 size = p2 - p1;
 
 	Rect2 screenrect = Rect2(Vector2(flip_x ? 1.0 : 0.0, flip_y ? 1.0 : 0.0), Vector2(flip_x ? -1.0 : 1.0, flip_y ? -1.0 : 1.0));
@@ -450,10 +458,12 @@ void RasterizerGLES3::_blit_render_target_to_screen(RID p_render_target, Display
 	glEnable(GL_BLEND);
 	glBlendFunc(GL_ONE, GL_ZERO);
 
-	if (rt->view_count > 1) {
-		copy_effects->copy_to_rect_3d(screenrect, p_layer, GLES3::Texture::TYPE_LAYERED);
+	if (p_blit.lens_distortion.apply && (p_blit.lens_distortion.k1 != 0.0 || p_blit.lens_distortion.k2)) {
+		copy_effects->copy_with_lens_distortion(screenrect, p_blit.multi_view.use_layer ? p_blit.multi_view.layer : 0, p_blit.lens_distortion.eye_center, p_blit.lens_distortion.k1, p_blit.lens_distortion.k2, p_blit.lens_distortion.upscale, p_blit.lens_distortion.aspect_ratio, linear_to_srgb);
+	} else if (rt->view_count > 1) {
+		copy_effects->copy_to_rect_3d(screenrect, p_blit.multi_view.use_layer ? p_blit.multi_view.layer : 0, GLES3::Texture::TYPE_LAYERED, 0.0, linear_to_srgb);
 	} else {
-		copy_effects->copy_to_rect(screenrect);
+		copy_effects->copy_to_rect(screenrect, linear_to_srgb);
 	}
 
 	glBindTexture(GL_TEXTURE_2D, 0);
@@ -462,12 +472,7 @@ void RasterizerGLES3::_blit_render_target_to_screen(RID p_render_target, Display
 // is this p_screen useless in a multi window environment?
 void RasterizerGLES3::blit_render_targets_to_screen(DisplayServer::WindowID p_screen, const BlitToScreen *p_render_targets, int p_amount) {
 	for (int i = 0; i < p_amount; i++) {
-		const BlitToScreen &blit = p_render_targets[i];
-
-		RID rid_rt = blit.render_target;
-
-		Rect2 dst_rect = blit.dst_rect;
-		_blit_render_target_to_screen(rid_rt, p_screen, dst_rect, blit.multi_view.use_layer ? blit.multi_view.layer : 0, i == 0);
+		_blit_render_target_to_screen(p_screen, p_render_targets[i], i == 0);
 	}
 }
 

drivers/gles3/rasterizer_gles3.h

@@ -83,7 +83,7 @@ class RasterizerGLES3 : public RendererCompositor {
 	RasterizerSceneGLES3 *scene = nullptr;
 	static RasterizerGLES3 *singleton;
 
-	void _blit_render_target_to_screen(RID p_render_target, DisplayServer::WindowID p_screen, const Rect2 &p_screen_rect, uint32_t p_layer, bool p_first = true);
+	void _blit_render_target_to_screen(DisplayServer::WindowID p_screen, const BlitToScreen &p_blit, bool p_first = true);
 
 public:
 	RendererUtilities *get_utilities() { return utilities; }

drivers/gles3/shaders/effects/copy.glsl

@@ -6,6 +6,7 @@ mode_copy_section = #define USE_COPY_SECTION \n#define MODE_SIMPLE_COPY
 mode_copy_section_source = #define USE_COPY_SECTION \n#define MODE_SIMPLE_COPY \n#define MODE_COPY_FROM
 mode_copy_section_3d = #define USE_COPY_SECTION \n#define MODE_SIMPLE_COPY \n#define USE_TEXTURE_3D
 mode_copy_section_2d_array = #define USE_COPY_SECTION \n#define MODE_SIMPLE_COPY \n#define USE_TEXTURE_2D_ARRAY
+mode_lens_distortion = #define USE_COPY_SECTION \n#define MODE_SIMPLE_COPY \n#define USE_TEXTURE_2D_ARRAY \n#define APPLY_LENS_DISTORTION
 mode_screen = #define MODE_SIMPLE_COPY \n#define MODE_MULTIPLY
 mode_gaussian_blur = #define MODE_GAUSSIAN_BLUR
 mode_mipmap = #define MODE_MIPMAP
@@ -15,6 +16,8 @@ mode_cube_to_panorama = #define CUBE_TO_PANORAMA
 
 #[specializations]
 
+CONVERT_LINEAR_TO_SRGB = false
+
 #[vertex]
 
 layout(location = 0) in vec2 vertex_attrib;
@@ -93,6 +96,14 @@ uniform sampler2D source; // texunit:0
 
 #endif // !(defined(CUBE_TO_OCTAHEDRAL) || defined(CUBE_TO_PANORAMA))
 
+#ifdef APPLY_LENS_DISTORTION
+uniform vec2 eye_center;
+uniform float k1;
+uniform float k2;
+uniform float upscale;
+uniform float aspect_ratio;
+#endif // APPLY_LENS_DISTORTION
+
 layout(location = 0) out vec4 frag_color;
 
 // This expects 0-1 range input, outside that range it behaves poorly.
@@ -101,22 +112,69 @@ vec3 srgb_to_linear(vec3 color) {
 	return color * (color * (color * 0.305306011 + 0.682171111) + 0.012522878);
 }
 
+// This expects 0-1 range input.
+vec3 linear_to_srgb(vec3 color) {
+	// Approximation from http://chilliant.blogspot.com/2012/08/srgb-approximations-for-hlsl.html
+	return max(vec3(1.055) * pow(color, vec3(0.416666667)) - vec3(0.055), vec3(0.0));
+}
+
 void main() {
 #ifdef MODE_SIMPLE_COPY
 
+	vec2 uv = uv_interp;
+
+#ifdef APPLY_LENS_DISTORTION
+	uv = uv * 2.0 - 1.0;
+	vec2 offset = uv - eye_center;
+
+	// take aspect ratio into account
+	offset.y /= aspect_ratio;
+
+	// distort
+	vec2 offset_sq = offset * offset;
+	float radius_sq = offset_sq.x + offset_sq.y;
+	float radius_s4 = radius_sq * radius_sq;
+	float distortion_scale = 1.0 + (k1 * radius_sq) + (k2 * radius_s4);
+	offset *= distortion_scale;
+
+	// reapply aspect ratio
+	offset.y *= aspect_ratio;
+
+	// add our eye center back in
+	uv = offset + eye_center;
+	uv /= upscale;
+
+	// and check our color
+	if (uv.x < -1.0 || uv.y < -1.0 || uv.x > 1.0 || uv.y > 1.0) {
+		frag_color = vec4(0.0, 0.0, 0.0, 1.0);
+	} else {
+		uv = uv * 0.5 + 0.5;
+#endif // APPLY_LENS_DISTORTION
+
 #ifdef USE_TEXTURE_3D
-	vec4 color = textureLod(source_3d, vec3(uv_interp, layer), lod);
+		vec4 color = textureLod(source_3d, vec3(uv, layer), lod);
 #elif defined(USE_TEXTURE_2D_ARRAY)
-	vec4 color = textureLod(source_2d_array, vec3(uv_interp, layer), lod);
+	vec4 color = textureLod(source_2d_array, vec3(uv, layer), lod);
 #else
-	vec4 color = texture(source, uv_interp);
+	vec4 color = texture(source, uv);
 #endif // USE_TEXTURE_3D
 
+#ifdef CONVERT_LINEAR_TO_SRGB
+		// Reading from a *_SRGB texture source will have converted data to linear,
+		// but we should output in sRGB!
+		color.rgb = linear_to_srgb(color.rgb);
+#endif
+
 #ifdef MODE_MULTIPLY
-	color *= multiply;
+		color *= multiply;
 #endif // MODE_MULTIPLY
 
-	frag_color = color;
+		frag_color = color;
+
+#ifdef APPLY_LENS_DISTORTION
+	}
+#endif // APPLY_LENS_DISTORTION
+
 #endif // MODE_SIMPLE_COPY
 
 #ifdef MODE_SIMPLE_COLOR