CLK/Outputs/OpenGL/ScanTarget.cpp

//
//  ScanTarget.cpp
//  Clock Signal
//
//  Created by Thomas Harte on 05/11/2018.
//  Copyright © 2018 Thomas Harte. All rights reserved.
//

#include "ScanTarget.hpp"

#include "OpenGL.hpp"

#include "Outputs/ScanTargets/FilterGenerator.hpp"
#include "Outputs/OpenGL/Shaders/CompositionShader.hpp"
#include "Outputs/OpenGL/Shaders/CopyShader.hpp"

#include <algorithm>
#include <cassert>
#include <cstring>
#include <limits>

using namespace Outputs::Display::OpenGL;

#ifndef NDEBUG
//#define LOG_LINES
//#define LOG_SCANS
#endif

namespace {

/// The texture unit from which to source input data.
constexpr GLenum SourceDataTextureUnit = GL_TEXTURE0;

/// The texture unit which contains raw line-by-line composite, S-Video or RGB data.
constexpr GLenum UnprocessedLineBufferTextureUnit = GL_TEXTURE1;

/// The texture unit that contains a pre-lowpass-filtered but fixed-resolution version of the chroma signal;
/// this is used when processing composite video only, and for chroma information only. Luminance is calculated
/// at the fidelity permitted by the output target, but my efforts to separate, demodulate and filter
/// chrominance during output without either massively sampling or else incurring significant high-frequency
/// noise that sampling reduces into a Moire, have proven to be unsuccessful for the time being.
constexpr GLenum QAMChromaTextureUnit = GL_TEXTURE2;

/// The texture unit that contains the current display.
constexpr GLenum AccumulationTextureUnit = GL_TEXTURE3;

using Logger = Log::Logger<Log::Source::OpenGL>;

constexpr GLint internalFormatForDepth(const std::size_t depth) {
	switch(depth) {
		default: return GL_FALSE;
		case 1: return GL_R8UI;
		case 2: return GL_RG8UI;
		case 3: return GL_RGB8UI;
		case 4: return GL_RGBA8UI;
	}
}

constexpr GLenum formatForDepth(const std::size_t depth) {
	switch(depth) {
		default: return GL_FALSE;
		case 1: return GL_RED_INTEGER;
		case 2: return GL_RG_INTEGER;
		case 3: return GL_RGB_INTEGER;
		case 4: return GL_RGBA_INTEGER;
	}
}

template <typename T> void allocate_buffer(
	const T &array,
	GLuint &buffer_name,
	GLuint &vertex_array_name
) {
	const auto buffer_size = array.size() * sizeof(array[0]);
	test_gl([&]{ glGenBuffers(1, &buffer_name); });
	test_gl([&]{ glBindBuffer(GL_ARRAY_BUFFER, buffer_name); });
	test_gl([&]{ glBufferData(GL_ARRAY_BUFFER, GLsizeiptr(buffer_size), NULL, GL_STREAM_DRAW); });

	test_gl([&]{ glGenVertexArrays(1, &vertex_array_name); });
	test_gl([&]{ glBindVertexArray(vertex_array_name); });
	test_gl([&]{ glBindBuffer(GL_ARRAY_BUFFER, buffer_name); });
}
}

ScanTarget::ScanTarget(const API api, const GLuint target_framebuffer, const float output_gamma) :
	api_(api),
	target_framebuffer_(target_framebuffer),
	output_gamma_(output_gamma),
	unprocessed_line_texture_(api, LineBufferWidth, LineBufferHeight, UnprocessedLineBufferTextureUnit, GL_NEAREST, false),
	full_display_rectangle_(api, -1.0f, -1.0f, 2.0f, 2.0f),
	scans_(scan_buffer_) {

	set_scan_buffer(scan_buffer_.data(), scan_buffer_.size());
	set_line_buffer(line_buffer_.data(), line_metadata_buffer_.data(), line_buffer_.size());

	// Allocate space for the scans and lines.
	allocate_buffer(scan_buffer_, scan_buffer_name_, scan_vertex_array_);
	allocate_buffer(line_buffer_, line_buffer_name_, line_vertex_array_);

	// TODO: if this is OpenGL 4.4 or newer, use glBufferStorage rather than glBufferData
	// and specify GL_MAP_PERSISTENT_BIT. Then map the buffer now, and let the client
	// write straight into it.

	test_gl([&]{ glBlendFunc(GL_SRC_ALPHA, GL_CONSTANT_COLOR); });
	test_gl([&]{ glBlendColor(0.4f, 0.4f, 0.4f, 1.0f); });

	// Establish initial state for is_drawing_to_accumulation_buffer_.
	is_drawing_to_accumulation_buffer_.clear();


	// TEST CODE. NOCOMMIT.
//	const auto buffer_width = FilterGenerator::SuggestedBufferWidth;
//	const float sample_multiplier =
//		FilterGenerator::suggested_sample_multiplier(227.5f, 1320);
//
//	VertexArray va(scan_buffer_);
//	for(auto &pair: {
//		std::make_pair(InputDataType::Luminance1, DisplayType::CompositeColour),
//		std::make_pair(InputDataType::Luminance8, DisplayType::CompositeColour),
//		std::make_pair(InputDataType::PhaseLinkedLuminance8, DisplayType::CompositeColour),
//
//		std::make_pair(InputDataType::Luminance8Phase8, DisplayType::SVideo),
//		std::make_pair(InputDataType::Luminance8Phase8, DisplayType::CompositeColour),
//
//		std::make_pair(InputDataType::Red1Green1Blue1, DisplayType::RGB),
//		std::make_pair(InputDataType::Red1Green1Blue1, DisplayType::SVideo),
//		std::make_pair(InputDataType::Red1Green1Blue1, DisplayType::CompositeColour),
//
//		std::make_pair(InputDataType::Red2Green2Blue2, DisplayType::RGB),
//		std::make_pair(InputDataType::Red2Green2Blue2, DisplayType::SVideo),
//		std::make_pair(InputDataType::Red2Green2Blue2, DisplayType::CompositeColour),
//
//		std::make_pair(InputDataType::Red4Green4Blue4, DisplayType::RGB),
//		std::make_pair(InputDataType::Red4Green4Blue4, DisplayType::SVideo),
//		std::make_pair(InputDataType::Red4Green4Blue4, DisplayType::CompositeColour),
//
//		std::make_pair(InputDataType::Red8Green8Blue8, DisplayType::RGB),
//		std::make_pair(InputDataType::Red8Green8Blue8, DisplayType::SVideo),
//		std::make_pair(InputDataType::Red8Green8Blue8, DisplayType::CompositeColour),
//	}) {
//		OpenGL::composition_shader(
//			api,
//			pair.first,
//			pair.second,
//			ColourSpace::YIQ,
//			sample_multiplier,
//			BufferingScanTarget::WriteAreaWidth, BufferingScanTarget::WriteAreaHeight,
//			buffer_width, 2048,	// TODO: substitute real composition buffer sizes.
//			va,
//			GL_TEXTURE0
//		).bind();
//	}
}

ScanTarget::~ScanTarget() {
	perform([&] {
		glDeleteBuffers(1, &scan_buffer_name_);
		glDeleteVertexArrays(1, &scan_vertex_array_);
	});
}

void ScanTarget::set_target_framebuffer(GLuint target_framebuffer) {
	perform([&] {
		target_framebuffer_ = target_framebuffer;
	});
}

void ScanTarget::setup_pipeline() {
	const auto modals = BufferingScanTarget::modals();
	const auto data_type_size = Outputs::Display::size_for_data_type(modals.input_data_type);

	// Possibly create a new source texture.
	if(source_texture_.empty() || source_texture_.channels() != data_type_size) {
		source_texture_ = Texture(
			data_type_size,
			SourceDataTextureUnit,
			GL_NEAREST,
			GL_NEAREST,
			WriteAreaWidth,
			WriteAreaHeight
		);
	}

	// Resize the texture only if required.
	const size_t required_size = WriteAreaWidth*WriteAreaHeight*data_type_size;
	if(required_size != write_area_texture_.size()) {
		write_area_texture_.resize(required_size);
		set_write_area(write_area_texture_.data());
	}

	// Prepare to bind line shaders.
	test_gl([&]{ glBindVertexArray(line_vertex_array_); });
	test_gl([&]{ glBindBuffer(GL_ARRAY_BUFFER, line_buffer_name_); });

	// Destroy or create a QAM buffer and shader, if appropriate.
	if(!existing_modals_ || existing_modals_->display_type != modals.display_type) {
		const bool needs_qam_buffer =
			modals.display_type == DisplayType::CompositeColour ||
			modals.display_type == DisplayType::SVideo;
		if(needs_qam_buffer) {
			if(!qam_chroma_texture_) {
				qam_chroma_texture_ = std::make_unique<TextureTarget>(
					api_,
					LineBufferWidth,
					LineBufferHeight,
					QAMChromaTextureUnit,
					GL_NEAREST,
					false
				);
			}

			qam_separation_shader_ = qam_separation_shader();
			enable_vertex_attributes(ShaderType::QAMSeparation, *qam_separation_shader_);
			qam_separation_shader_->set_uniform("textureName", GLint(UnprocessedLineBufferTextureUnit - GL_TEXTURE0));
		} else {
			qam_chroma_texture_.reset();
			qam_separation_shader_.reset();
		}

		// Establish an output shader.
		output_shader_ = conversion_shader();
		enable_vertex_attributes(ShaderType::Conversion, *output_shader_);
		set_uniforms(ShaderType::Conversion, *output_shader_);

		output_shader_->set_uniform("textureName", GLint(UnprocessedLineBufferTextureUnit - GL_TEXTURE0));
		output_shader_->set_uniform("qamTextureName", GLint(QAMChromaTextureUnit - GL_TEXTURE0));
	}

	if(qam_separation_shader_) {
		set_uniforms(ShaderType::QAMSeparation, *qam_separation_shader_);
	}

	// Select whichever of a letterbox or pillarbox avoids cropping.
	constexpr float output_ratio = 4.0f / 3.0f;
	const float aspect_ratio_stretch = modals.aspect_ratio / output_ratio;

	auto adjusted_rect = modals.visible_area;
	const float letterbox_scale = adjusted_rect.size.height / (adjusted_rect.size.width * aspect_ratio_stretch);
	if(letterbox_scale > 1.0f) {
		adjusted_rect.scale(letterbox_scale, 1.0f);
	} else {
		adjusted_rect.scale(1.0f, 1.0f / letterbox_scale);
	}

	// Provide to shader.
	output_shader_->set_uniform("origin", adjusted_rect.origin.x, adjusted_rect.origin.y);
	output_shader_->set_uniform("size", 1.0f / adjusted_rect.size.width, 1.0f / adjusted_rect.size.height);

	// Establish an input shader.
	if(!existing_modals_ || existing_modals_->input_data_type != modals.input_data_type) {
		input_shader_ = composition_shader();
		test_gl([&]{ glBindVertexArray(scan_vertex_array_); });
		test_gl([&]{ glBindBuffer(GL_ARRAY_BUFFER, scan_buffer_name_); });
		enable_vertex_attributes(ShaderType::Composition, *input_shader_);
		set_uniforms(ShaderType::Composition, *input_shader_);
		input_shader_->set_uniform("textureName", GLint(SourceDataTextureUnit - GL_TEXTURE0));
	}

	//
	// New pipeline starts here!
	//
	const auto buffer_width = FilterGenerator::SuggestedBufferWidth;
	const auto subcarrier_frequency = [](const Modals &modals) {
		return float(modals.colour_cycle_numerator) / float(modals.colour_cycle_denominator);
	};
	const float sample_multiplier =
		FilterGenerator::suggested_sample_multiplier(
			subcarrier_frequency(modals),
			modals.cycles_per_line
		);

//	if(copy_shader_.empty()) {
//		copy_shader_ = copy_shader(api_, GL_TEXTURE4, {}, {});
//	}

	if(
		!existing_modals_ ||
		existing_modals_->input_data_type != modals.input_data_type ||
		existing_modals_->display_type != modals.display_type ||
		existing_modals_->composite_colour_space != modals.composite_colour_space ||
		subcarrier_frequency(*existing_modals_) != subcarrier_frequency(modals)
	) {
		composition_shader_ = OpenGL::composition_shader(
			api_,
			modals.input_data_type,
			modals.display_type,
			modals.composite_colour_space,
			sample_multiplier,
			2048, 2048,
			buffer_width, 2048,
			scans_,
			GL_TEXTURE0
		);
	}

	if(composition_buffer_.empty()) {
		composition_buffer_ = TextureTarget(api_, buffer_width, 2048, GL_TEXTURE4, GL_NEAREST, false);
		composition_buffer_.bind_texture();

		// The below establishes that the composition buffer texture is ultimately being drawn.
		std::vector<uint8_t> image(buffer_width * 2048 * 4);
		for(auto &c : image) {
			c = rand();
		}
		test_gl([&]{ 
			glTexImage2D(
				GL_TEXTURE_2D,
				0,
				GL_RGBA,
				buffer_width,
				2048,
				0,
				GL_RGBA,
				GL_UNSIGNED_BYTE,
				image.data()
			);
		});
	}

	existing_modals_ = modals;
}

bool ScanTarget::is_soft_display_type() {
	const auto display_type = modals().display_type;
	return display_type == DisplayType::CompositeColour || display_type == DisplayType::CompositeMonochrome;
}

void ScanTarget::update(int, int output_height) {
	// If the GPU is still busy, don't wait; we'll catch it next time.
	if(fence_ != nullptr) {
		if(glClientWaitSync(fence_, GL_SYNC_FLUSH_COMMANDS_BIT, 0) == GL_TIMEOUT_EXPIRED) {
			display_metrics_.announce_draw_status(
				lines_submitted_,
				std::chrono::high_resolution_clock::now() - line_submission_begin_time_,
				false);
			return;
		}
		fence_ = nullptr;
	}

	// Update the display metrics.
	display_metrics_.announce_draw_status(
		lines_submitted_,
		std::chrono::high_resolution_clock::now() - line_submission_begin_time_,
		true);

	// Grab the new output list.
	perform([&] {
		const OutputArea area = get_output_area();

		// Establish the pipeline if necessary.
		const auto new_modals = BufferingScanTarget::new_modals();
		const bool did_setup_pipeline = [&] {
			if(bool(new_modals)) {
				setup_pipeline();
				return true;
			}
			return false;
		} ();

		// Determine the start time of this submission group and the number of lines it will contain.
		line_submission_begin_time_ = std::chrono::high_resolution_clock::now();
		lines_submitted_ = (area.end.line - area.start.line + line_buffer_.size()) % line_buffer_.size();

		// Submit texture.
		if(area.start.write_area_x != area.end.write_area_x || area.start.write_area_y != area.end.write_area_y) {
			source_texture_.bind();

			const auto submit = [&](const GLint y_begin, const GLint y_end) {
				test_gl([&]{
					glTexSubImage2D(
						GL_TEXTURE_2D, 0,
						0, y_begin,
						WriteAreaWidth,
						y_end - y_begin,
						formatForDepth(write_area_data_size()),
						GL_UNSIGNED_BYTE,
						&write_area_texture_[size_t(y_begin * WriteAreaWidth) * source_texture_.channels()]
					);
				});
			};

			// Both of the following upload to area.end.write_area_y + 1 to include whatever line the write area
			// is currently on. It may have partial source areas along it, despite being incomplete.
			if(area.end.write_area_y >= area.start.write_area_y) {
				// Submit the direct region from the submit pointer to the read pointer.
				submit(area.start.write_area_y, area.end.write_area_y + 1);
			} else {
				// The circular buffer wrapped around; submit the data from the read pointer to the end of
				// the buffer and from the start of the buffer to the submit pointer.
				submit(area.start.write_area_y, WriteAreaHeight);
				submit(0, area.end.write_area_y + 1);
			}
		}

		// Submit scans; only the new ones need to be communicated.
		if(area.end.scan != area.start.scan) {
			const size_t new_scans = (area.end.scan - area.start.scan + scan_buffer_.size()) % scan_buffer_.size();

			//
			// OLD PIPELINE: submit new scans.
			//
			test_gl([&]{ glBindBuffer(GL_ARRAY_BUFFER, scan_buffer_name_); });

			// Map only the required portion of the buffer.
			const size_t new_scans_size = new_scans * sizeof(Scan);
			const auto destination = static_cast<Scan *>(
				glMapBufferRange(
					GL_ARRAY_BUFFER,
					0,
					GLsizeiptr(new_scans_size),
					GL_MAP_WRITE_BIT | GL_MAP_FLUSH_EXPLICIT_BIT
				)
			);
			test_gl_error();

			// Copy as a single chunk if possible; otherwise copy in two parts.
			if(area.start.scan < area.end.scan) {
				std::copy_n(scan_buffer_.begin() + area.start.scan, new_scans, destination);
			} else {
				const size_t first_portion_count = scan_buffer_.size() - area.start.scan;
				std::copy_n(scan_buffer_.begin() + area.start.scan, first_portion_count, destination);
				std::copy_n(scan_buffer_.begin(), new_scans - first_portion_count, destination + first_portion_count);
			}

			// Flush and unmap the buffer.
			test_gl([&]{ glFlushMappedBufferRange(GL_ARRAY_BUFFER, 0, GLsizeiptr(new_scans_size)); });
			test_gl([&]{ glUnmapBuffer(GL_ARRAY_BUFFER); });

			//
			// OLD PIPELINE: draw new scans.
			//

			// Push new input to the unprocessed line buffer.
			unprocessed_line_texture_.bind_framebuffer();

			// Clear newly-touched lines; that is everything from (read+1) to submit.
			const auto first_line_to_clear = GLsizei((area.start.line+1)%line_buffer_.size());
			const auto final_line_to_clear = GLsizei(area.end.line);
			if(first_line_to_clear != final_line_to_clear) {
				test_gl([&]{ glEnable(GL_SCISSOR_TEST); });

				// Determine the proper clear colour — this needs to be anything that describes black
				// in the input colour encoding at use.
				if(modals().input_data_type == InputDataType::Luminance8Phase8) {
					// Supply both a zero luminance and a colour-subcarrier-disengaging phase.
					test_gl([&]{ glClearColor(0.0f, 1.0f, 0.0f, 0.0f); });
				} else {
					test_gl([&]{ glClearColor(0.0f, 0.0f, 0.0f, 0.0f); });
				}

				if(first_line_to_clear < final_line_to_clear) {
					test_gl([&]{ glScissor(GLint(0), GLint(first_line_to_clear), unprocessed_line_texture_.width(), final_line_to_clear - first_line_to_clear); });
					test_gl([&]{ glClear(GL_COLOR_BUFFER_BIT); });
				} else {
					test_gl([&]{ glScissor(GLint(0), GLint(0), unprocessed_line_texture_.width(), final_line_to_clear); });
					test_gl([&]{ glClear(GL_COLOR_BUFFER_BIT); });
					test_gl([&]{ glScissor(GLint(0), GLint(first_line_to_clear), unprocessed_line_texture_.width(), unprocessed_line_texture_.height() - first_line_to_clear); });
					test_gl([&]{ glClear(GL_COLOR_BUFFER_BIT); });
				}

				test_gl([&]{ glDisable(GL_SCISSOR_TEST); });
			}

			// Apply new spans. They definitely always go to the first buffer.
			test_gl([&]{ glBindVertexArray(scan_vertex_array_); });
			input_shader_->bind();
			test_gl([&]{ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, GLsizei(new_scans)); });


			//
			// NEW PIPELINE.
			//

			// Submit new scans.
			// First implementation: put all new scans at the start of the buffer, for a simple
			// glDrawArraysInstanced call below.
			scans_.bind_buffer();
			size_t buffer_destination = 0;
			const auto submit = [&](const size_t begin, const size_t end) {
				test_gl([&]{ 
					glBufferSubData(
						GL_ARRAY_BUFFER,
						buffer_destination,
						(end - begin) * sizeof(Scan),
						&scan_buffer_[begin]
					);
				});
				buffer_destination += (end - begin) * sizeof(Scan);
			};
			if(area.start.scan < area.end.scan) {
				submit(area.start.scan, area.end.scan);
			} else {
				submit(area.start.scan, scan_buffer_.size());
				submit(0, area.end.scan);
			}

			// Populate composition buffer.
			composition_buffer_.bind_framebuffer();
			scans_.bind();
			composition_shader_.bind();
			test_gl([&]{ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, GLsizei(new_scans)); });
		}

		// Logic for reducing resolution: start doing so if the metrics object reports that
		// it's a good idea. Go up to a quarter of the requested resolution, subject to
		// clamping at each stage. If the output resolution changes, or anything else about
		// the output pipeline, just start trying the highest size again.
		if(display_metrics_.should_lower_resolution() && is_soft_display_type()) {
			resolution_reduction_level_ = std::min(resolution_reduction_level_+1, 4);
		}
		if(output_height_ != output_height || did_setup_pipeline) {
			resolution_reduction_level_ = 1;
			output_height_ = output_height;
		}

		// Ensure the accumulation buffer is properly sized, allowing for the metrics object's
		// feelings about whether too high a resolution is being used.
		const int framebuffer_height = std::max(output_height / resolution_reduction_level_, std::min(540, output_height));
		const int proportional_width = (framebuffer_height * 4) / 3;
		const bool did_create_accumulation_texture =
			!accumulation_texture_ ||
			(
				accumulation_texture_->width() != proportional_width ||
				accumulation_texture_->height() != framebuffer_height
			);

		// Work with the accumulation_buffer_ potentially starts from here onwards; set its flag.
		while(is_drawing_to_accumulation_buffer_.test_and_set());
		if(did_create_accumulation_texture) {
			Logger::info().append("Changed output resolution to %d by %d", proportional_width, framebuffer_height);
			display_metrics_.announce_did_resize();
			auto new_framebuffer = std::make_unique<TextureTarget>(
				api_,
				GLsizei(proportional_width),
				GLsizei(framebuffer_height),
				AccumulationTextureUnit,
				GL_NEAREST,
				true
			);
			if(accumulation_texture_) {
				new_framebuffer->bind_framebuffer();
				test_gl([&]{ glClear(GL_COLOR_BUFFER_BIT | GL_STENCIL_BUFFER_BIT); });

				test_gl([&]{ glActiveTexture(AccumulationTextureUnit); });
				accumulation_texture_->bind_texture();
				accumulation_texture_->draw(4.0f / 3.0f);

				test_gl([&]{ glClear(GL_STENCIL_BUFFER_BIT); });

				new_framebuffer->bind_texture();
			}
			accumulation_texture_ = std::move(new_framebuffer);

			// In the absence of a way to resize a stencil buffer, just mark
			// what's currently present as invalid to avoid an improper clear
			// for this frame.
			stencil_is_valid_ = false;
		}

		if(did_setup_pipeline || did_create_accumulation_texture) {
			set_sampling_window(proportional_width, framebuffer_height, *output_shader_);
		}

		// Figure out how many new lines are ready.
		auto new_lines = (area.end.line - area.start.line + LineBufferHeight) % LineBufferHeight;
		if(new_lines) {
			// Prepare to output lines.
			test_gl([&]{ glBindVertexArray(line_vertex_array_); });

			// Bind the accumulation framebuffer, unless there's going to be QAM work first.
			if(!qam_separation_shader_ || line_metadata_buffer_[area.start.line].is_first_in_frame) {
				accumulation_texture_->bind_framebuffer();
				output_shader_->bind();

				// Enable blending and stenciling.
				test_gl([&]{ glEnable(GL_BLEND); });
				test_gl([&]{ glEnable(GL_STENCIL_TEST); });
			}

			// Set the proper stencil function regardless.
			test_gl([&]{ glStencilFunc(GL_EQUAL, 0, GLuint(~0)); });
			test_gl([&]{ glStencilOp(GL_KEEP, GL_KEEP, GL_INCR); });

			// Prepare to upload data that will consitute lines.
			test_gl([&]{ glBindBuffer(GL_ARRAY_BUFFER, line_buffer_name_); });

			// Divide spans by which frame they're in.
			auto start_line = area.start.line;
			while(new_lines) {
				uint16_t end_line = (start_line + 1) % LineBufferHeight;

				// Find the limit of spans to draw in this cycle.
				size_t lines = 1;
				while(end_line != area.end.line && !line_metadata_buffer_[end_line].is_first_in_frame) {
					end_line = (end_line + 1) % LineBufferHeight;
					++lines;
				}

				// If this is start-of-frame, clear any untouched pixels and flush the stencil buffer
				if(line_metadata_buffer_[start_line].is_first_in_frame) {
					if(stencil_is_valid_ && line_metadata_buffer_[start_line].previous_frame_was_complete) {
						full_display_rectangle_.draw(0.0f, 0.0f, 0.0f);
					}
					stencil_is_valid_ = true;
					test_gl([&]{ glClear(GL_STENCIL_BUFFER_BIT); });

					// Rebind the program for span output.
					test_gl([&]{ glBindVertexArray(line_vertex_array_); });
					if(!qam_separation_shader_) {
						output_shader_->bind();
					}
				}

				// Upload.
				const auto buffer_size = lines * sizeof(Line);
				if(!end_line || end_line > start_line) {
					test_gl([&]{ glBufferSubData(GL_ARRAY_BUFFER, 0, GLsizeiptr(buffer_size), &line_buffer_[start_line]); });
				} else {
					auto destination = static_cast<Line *>(
						glMapBufferRange(
							GL_ARRAY_BUFFER,
							0,
							GLsizeiptr(buffer_size),
							GL_MAP_WRITE_BIT | GL_MAP_FLUSH_EXPLICIT_BIT
						)
					);
					assert(destination);
					test_gl_error();

					std::copy_n(line_buffer_.begin(), end_line, destination + line_buffer_.size() - start_line);
					std::copy_n(line_buffer_.begin() + start_line, line_buffer_.size() - start_line, destination);

					test_gl([&]{ glFlushMappedBufferRange(GL_ARRAY_BUFFER, 0, GLsizeiptr(buffer_size)); });
					test_gl([&]{ glUnmapBuffer(GL_ARRAY_BUFFER); });
				}

				// Produce colour information, if required.
				if(qam_separation_shader_) {
					qam_separation_shader_->bind();
					qam_chroma_texture_->bind_framebuffer();
					test_gl([&]{ glClear(GL_COLOR_BUFFER_BIT); });	// TODO: this is here as a hint that the old framebuffer doesn't need reloading;
															// test whether that's a valid optimisation on desktop OpenGL.

					test_gl([&]{ glDisable(GL_BLEND); });
					test_gl([&]{ glDisable(GL_STENCIL_TEST); });
					test_gl([&]{ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, GLsizei(lines)); });

					accumulation_texture_->bind_framebuffer();
					output_shader_->bind();
					test_gl([&]{ glEnable(GL_BLEND); });
					test_gl([&]{ glEnable(GL_STENCIL_TEST); });
				}

				// Render to the output.
				test_gl([&]{ glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, GLsizei(lines)); });

				start_line = end_line;
				new_lines -= lines;
			}

			// Disable blending and the stencil test again.
			test_gl([&]{ glDisable(GL_STENCIL_TEST); });
			test_gl([&]{ glDisable(GL_BLEND); });
		}

		// That's it for operations affecting the accumulation buffer.
		is_drawing_to_accumulation_buffer_.clear();

		// Grab a fence sync object to avoid busy waiting upon the next extry into this
		// function, and reset the is_updating_ flag.
		fence_ = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
		complete_output_area(area);
	});
}

void ScanTarget::draw(int output_width, int output_height) {
	while(is_drawing_to_accumulation_buffer_.test_and_set(std::memory_order_acquire));

	if(accumulation_texture_) {
		// Copy the accumulation texture to the target.
		test_gl([&]{ glBindFramebuffer(GL_FRAMEBUFFER, target_framebuffer_); });
		test_gl([&]{ glViewport(0, 0, (GLsizei)output_width, (GLsizei)output_height); });

		test_gl([&]{ glClearColor(0.0f, 0.0f, 0.0f, 0.0f); });
		test_gl([&]{ glClear(GL_COLOR_BUFFER_BIT); });
		accumulation_texture_->bind_texture();
		accumulation_texture_->draw(float(output_width) / float(output_height), 4.0f / 255.0f);
	}

	if(!composition_buffer_.empty()) {
		// Copy the accumulation texture to the target.
		test_gl([&]{ glBindFramebuffer(GL_FRAMEBUFFER, target_framebuffer_); });
		test_gl([&]{ glViewport(0, 0, (GLsizei)output_width, (GLsizei)output_height); });
		composition_buffer_.draw(float(output_width) / float(output_height), 4.0f / 255.0f);
	}

	is_drawing_to_accumulation_buffer_.clear(std::memory_order_release);
}