1
0
mirror of https://github.com/TomHarte/CLK.git synced 2024-07-06 01:28:57 +00:00

Merge pull request #96 from TomHarte/PhaseAlignedSampling

Optimises existing composite flow
This commit is contained in:
Thomas Harte 2017-01-25 21:51:11 -05:00 committed by GitHub
commit 92f928ca42
15 changed files with 320 additions and 240 deletions

View File

@ -84,7 +84,7 @@ void AY38910::set_clock_rate(double clock_rate)
void AY38910::get_samples(unsigned int number_of_samples, int16_t *target)
{
int c = 0;
while((master_divider_&15) && c < number_of_samples)
while((master_divider_&7) && c < number_of_samples)
{
target[c] = output_volume_;
master_divider_++;
@ -131,7 +131,7 @@ void AY38910::get_samples(unsigned int number_of_samples, int16_t *target)
evaluate_output_volume();
for(int ic = 0; ic < 16 && c < number_of_samples; ic++)
for(int ic = 0; ic < 8 && c < number_of_samples; ic++)
{
target[c] = output_volume_;
c++;
@ -139,7 +139,7 @@ void AY38910::get_samples(unsigned int number_of_samples, int16_t *target)
}
}
master_divider_ &= 15;
master_divider_ &= 7;
}
void AY38910::evaluate_output_volume()

View File

@ -71,7 +71,7 @@ void Machine::setup_output(float aspect_ratio)
"uint y = c & 14u;"
"uint iPhase = (c >> 4);"
"float phaseOffset = 6.283185308 * float(iPhase - 1u) / 13.0;"
"float phaseOffset = 6.283185308 * float(iPhase) / 13.0 + 5.074880441076923;"
"return mix(float(y) / 14.0, step(1, iPhase) * cos(phase + phaseOffset), amplitude);"
"}");
speaker_->set_input_rate((float)(get_clock_rate() / 38.0));
@ -777,5 +777,6 @@ void Machine::update_audio()
void Machine::synchronise()
{
update_audio();
speaker_->flush();
}

View File

@ -29,9 +29,9 @@ void CRT::set_new_timing(unsigned int cycles_per_line, unsigned int height_of_di
// for horizontal retrace and 500 to 750 µs for vertical retrace in NTSC and PAL TV."
time_multiplier_ = IntermediateBufferWidth / cycles_per_line;
phase_denominator_ = cycles_per_line * colour_cycle_denominator;
phase_denominator_ = cycles_per_line * colour_cycle_denominator * time_multiplier_;
phase_numerator_ = 0;
colour_cycle_numerator_ = colour_cycle_numerator * time_multiplier_;
colour_cycle_numerator_ = colour_cycle_numerator;
phase_alternates_ = should_alternate;
is_alernate_line_ &= phase_alternates_;
unsigned int multiplied_cycles_per_line = cycles_per_line * time_multiplier_;
@ -112,7 +112,6 @@ Flywheel::SyncEvent CRT::get_next_horizontal_sync_event(bool hsync_is_requested,
#define source_output_position_x2() (*(uint16_t *)&next_run[SourceVertexOffsetOfEnds + 2])
#define source_phase() next_run[SourceVertexOffsetOfPhaseTimeAndAmplitude + 0]
#define source_amplitude() next_run[SourceVertexOffsetOfPhaseTimeAndAmplitude + 2]
#define source_phase_time() next_run[SourceVertexOffsetOfPhaseTimeAndAmplitude + 1]
void CRT::advance_cycles(unsigned int number_of_cycles, bool hsync_requested, bool vsync_requested, const bool vsync_charging, const Scan::Type type)
{
@ -149,7 +148,6 @@ void CRT::advance_cycles(unsigned int number_of_cycles, bool hsync_requested, bo
source_output_position_x1() = (uint16_t)horizontal_flywheel_->get_current_output_position();
source_phase() = colour_burst_phase_;
source_amplitude() = colour_burst_amplitude_;
source_phase_time() = (uint8_t)colour_burst_time_; // assumption: burst was within the first 1/16 of the line
}
// decrement the number of cycles left to run for and increment the
@ -242,7 +240,9 @@ void CRT::advance_cycles(unsigned int number_of_cycles, bool hsync_requested, bo
frames_since_last_delegate_call_++;
if(frames_since_last_delegate_call_ == 20)
{
output_lock.unlock();
delegate_->crt_did_end_batch_of_frames(this, frames_since_last_delegate_call_, vertical_flywheel_->get_and_reset_number_of_surprises());
output_lock.lock();
frames_since_last_delegate_call_ = 0;
}
}
@ -286,9 +286,11 @@ void CRT::output_scan(const Scan *const scan)
{
if(horizontal_flywheel_->get_current_time() < (horizontal_flywheel_->get_standard_period() * 12) >> 6)
{
colour_burst_time_ = (uint16_t)horizontal_flywheel_->get_current_time();
colour_burst_phase_ = scan->phase;
unsigned int position_phase = (horizontal_flywheel_->get_current_time() * colour_cycle_numerator_ * 256) / phase_denominator_;
colour_burst_phase_ = (position_phase + scan->phase) & 255;
colour_burst_amplitude_ = scan->amplitude;
colour_burst_phase_ = (colour_burst_phase_ & ~63) + 32;
}
}
@ -344,7 +346,7 @@ void CRT::output_default_colour_burst(unsigned int number_of_cycles)
Scan scan {
.type = Scan::Type::ColourBurst,
.number_of_cycles = number_of_cycles,
.phase = (uint8_t)((phase_numerator_ * 255) / phase_denominator_ + (is_alernate_line_ ? 128 : 0)),
.phase = (uint8_t)((phase_numerator_ * 256) / phase_denominator_ + (is_alernate_line_ ? 128 : 0)),
.amplitude = 32
};
output_scan(&scan);

View File

@ -60,7 +60,6 @@ class CRT {
void output_scan(const Scan *scan);
uint8_t colour_burst_phase_, colour_burst_amplitude_;
uint16_t colour_burst_time_;
bool is_writing_composite_run_;
unsigned int phase_denominator_, phase_numerator_, colour_cycle_numerator_;

View File

@ -36,7 +36,7 @@ const GLsizei InputBufferBuilderWidth = 2048;
const GLsizei InputBufferBuilderHeight = 512;
// This is the size of the intermediate buffers used during composite to RGB conversion
const GLsizei IntermediateBufferWidth = 4096;
const GLsizei IntermediateBufferWidth = 2048;
const GLsizei IntermediateBufferHeight = 512;
// Some internal buffer sizes

View File

@ -6,8 +6,10 @@
//
#include "CRT.hpp"
#include <stdlib.h>
#include <math.h>
#include <cstdlib>
#include <cstring>
#include <cmath>
#include "CRTOpenGL.hpp"
#include "../../../SignalProcessing/FIRFilter.hpp"
@ -16,12 +18,14 @@
using namespace Outputs::CRT;
namespace {
static const GLenum composite_texture_unit = GL_TEXTURE0;
static const GLenum separated_texture_unit = GL_TEXTURE1;
static const GLenum filtered_y_texture_unit = GL_TEXTURE2;
static const GLenum filtered_texture_unit = GL_TEXTURE3;
static const GLenum source_data_texture_unit = GL_TEXTURE4;
static const GLenum pixel_accumulation_texture_unit = GL_TEXTURE5;
static const GLenum source_data_texture_unit = GL_TEXTURE0;
static const GLenum pixel_accumulation_texture_unit = GL_TEXTURE1;
static const GLenum composite_texture_unit = GL_TEXTURE2;
static const GLenum separated_texture_unit = GL_TEXTURE3;
static const GLenum filtered_texture_unit = GL_TEXTURE4;
static const GLenum work_texture_unit = GL_TEXTURE2;
}
OpenGLOutputBuilder::OpenGLOutputBuilder(size_t bytes_per_pixel) :
@ -33,11 +37,7 @@ OpenGLOutputBuilder::OpenGLOutputBuilder(size_t bytes_per_pixel) :
last_output_height_(0),
fence_(nullptr),
texture_builder(bytes_per_pixel, source_data_texture_unit),
array_builder(SourceVertexBufferDataSize, OutputVertexBufferDataSize),
composite_texture_(IntermediateBufferWidth, IntermediateBufferHeight, composite_texture_unit),
separated_texture_(IntermediateBufferWidth, IntermediateBufferHeight, separated_texture_unit),
filtered_y_texture_(IntermediateBufferWidth, IntermediateBufferHeight, filtered_y_texture_unit),
filtered_texture_(IntermediateBufferWidth, IntermediateBufferHeight, filtered_texture_unit)
array_builder(SourceVertexBufferDataSize, OutputVertexBufferDataSize)
{
glBlendFunc(GL_SRC_ALPHA, GL_CONSTANT_COLOR);
glBlendColor(0.6f, 0.6f, 0.6f, 1.0f);
@ -47,6 +47,32 @@ OpenGLOutputBuilder::OpenGLOutputBuilder(size_t bytes_per_pixel) :
// create the source vertex array
glGenVertexArrays(1, &source_vertex_array_);
bool supports_texture_barrier = false;
#ifdef GL_NV_texture_barrier
GLint number_of_extensions;
glGetIntegerv(GL_NUM_EXTENSIONS, &number_of_extensions);
for(GLuint c = 0; c < (GLuint)number_of_extensions; c++)
{
const char *extension_name = (const char *)glGetStringi(GL_EXTENSIONS, c);
if(!strcmp(extension_name, "GL_NV_texture_barrier"))
{
supports_texture_barrier = true;
}
}
#endif
// if(supports_texture_barrier)
// {
// work_texture_.reset(new OpenGL::TextureTarget(IntermediateBufferWidth, IntermediateBufferHeight*2, work_texture_unit));
// }
// else
{
composite_texture_.reset(new OpenGL::TextureTarget(IntermediateBufferWidth, IntermediateBufferHeight, composite_texture_unit, GL_NEAREST));
separated_texture_.reset(new OpenGL::TextureTarget(IntermediateBufferWidth, IntermediateBufferHeight, separated_texture_unit, GL_NEAREST));
filtered_texture_.reset(new OpenGL::TextureTarget(IntermediateBufferWidth, IntermediateBufferHeight, filtered_texture_unit, GL_LINEAR));
}
}
OpenGLOutputBuilder::~OpenGLOutputBuilder()
@ -57,6 +83,11 @@ OpenGLOutputBuilder::~OpenGLOutputBuilder()
free(rgb_shader_);
}
bool OpenGLOutputBuilder::get_is_television_output()
{
return output_device_ == Television || !rgb_input_shader_program_;
}
void OpenGLOutputBuilder::draw_frame(unsigned int output_width, unsigned int output_height, bool only_if_dirty)
{
// lock down any other draw_frames
@ -91,7 +122,7 @@ void OpenGLOutputBuilder::draw_frame(unsigned int output_width, unsigned int out
// make sure there's a target to draw to
if(!framebuffer_ || framebuffer_->get_height() != output_height || framebuffer_->get_width() != output_width)
{
std::unique_ptr<OpenGL::TextureTarget> new_framebuffer(new OpenGL::TextureTarget((GLsizei)output_width, (GLsizei)output_height, pixel_accumulation_texture_unit));
std::unique_ptr<OpenGL::TextureTarget> new_framebuffer(new OpenGL::TextureTarget((GLsizei)output_width, (GLsizei)output_height, pixel_accumulation_texture_unit, GL_LINEAR));
if(framebuffer_)
{
new_framebuffer->bind_framebuffer();
@ -123,30 +154,29 @@ void OpenGLOutputBuilder::draw_frame(unsigned int output_width, unsigned int out
output_mutex_.unlock();
struct RenderStage {
OpenGL::TextureTarget *const target;
OpenGL::Shader *const shader;
OpenGL::TextureTarget *const target;
float clear_colour[3];
};
// for composite video, go through four steps to get to something that can be painted to the output
RenderStage composite_render_stages[] =
{
{&composite_texture_, composite_input_shader_program_.get(), {0.0, 0.0, 0.0}},
{&separated_texture_, composite_separation_filter_program_.get(), {0.0, 0.5, 0.5}},
{&filtered_y_texture_, composite_y_filter_shader_program_.get(), {0.0, 0.5, 0.5}},
{&filtered_texture_, composite_chrominance_filter_shader_program_.get(), {0.0, 0.0, 0.0}},
{composite_input_shader_program_.get(), composite_texture_.get(), {0.0, 0.0, 0.0}},
{composite_separation_filter_program_.get(), separated_texture_.get(), {0.0, 0.5, 0.5}},
{composite_chrominance_filter_shader_program_.get(), filtered_texture_.get(), {0.0, 0.0, 0.0}},
{nullptr}
};
// for RGB video, there's only two steps
RenderStage rgb_render_stages[] =
{
{&composite_texture_, rgb_input_shader_program_.get(), {0.0, 0.0, 0.0}},
{&filtered_texture_, rgb_filter_shader_program_.get(), {0.0, 0.0, 0.0}},
{rgb_input_shader_program_.get(), composite_texture_.get(), {0.0, 0.0, 0.0}},
{rgb_filter_shader_program_.get(), filtered_texture_.get(), {0.0, 0.0, 0.0}},
{nullptr}
};
RenderStage *active_pipeline = (output_device_ == Television || !rgb_input_shader_program_) ? composite_render_stages : rgb_render_stages;
RenderStage *active_pipeline = get_is_television_output() ? composite_render_stages : rgb_render_stages;
if(array_submission.input_size || array_submission.output_size)
{
@ -154,24 +184,39 @@ void OpenGLOutputBuilder::draw_frame(unsigned int output_width, unsigned int out
glBindVertexArray(source_vertex_array_);
glDisable(GL_BLEND);
while(active_pipeline->target)
#ifdef GL_NV_texture_barrier
if(work_texture_)
{
work_texture_->bind_framebuffer();
glClear(GL_COLOR_BUFFER_BIT);
}
#endif
while(active_pipeline->shader)
{
// switch to the framebuffer and shader associated with this stage
active_pipeline->shader->bind();
active_pipeline->target->bind_framebuffer();
// if this is the final stage before painting to the CRT, clear the framebuffer before drawing in order to blank out
// those portions for which no input was provided
if(!active_pipeline[1].target)
if(!work_texture_)
{
glClearColor(active_pipeline->clear_colour[0], active_pipeline->clear_colour[1], active_pipeline->clear_colour[2], 1.0f);
glClear(GL_COLOR_BUFFER_BIT);
active_pipeline->target->bind_framebuffer();
// if this is the final stage before painting to the CRT, clear the framebuffer before drawing in order to blank out
// those portions for which no input was provided
// if(!active_pipeline[1].shader)
// {
glClearColor(active_pipeline->clear_colour[0], active_pipeline->clear_colour[1], active_pipeline->clear_colour[2], 1.0f);
glClear(GL_COLOR_BUFFER_BIT);
// }
}
// draw
glDrawArraysInstanced(GL_LINES, 0, 2, (GLsizei)array_submission.input_size / SourceVertexSize);
glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, (GLsizei)array_submission.input_size / SourceVertexSize);
active_pipeline++;
#ifdef GL_NV_texture_barrier
glTextureBarrierNV();
#endif
}
// prepare to transfer to framebuffer
@ -194,6 +239,10 @@ void OpenGLOutputBuilder::draw_frame(unsigned int output_width, unsigned int out
glDrawArraysInstanced(GL_TRIANGLE_STRIP, 0, 4, (GLsizei)array_submission.output_size / OutputVertexSize);
}
#ifdef GL_NV_texture_barrier
glTextureBarrierNV();
#endif
// copy framebuffer to the intended place
glDisable(GL_BLEND);
glBindFramebuffer(GL_FRAMEBUFFER, 0);
@ -211,7 +260,6 @@ void OpenGLOutputBuilder::reset_all_OpenGL_state()
{
composite_input_shader_program_ = nullptr;
composite_separation_filter_program_ = nullptr;
composite_y_filter_shader_program_ = nullptr;
composite_chrominance_filter_shader_program_ = nullptr;
rgb_input_shader_program_ = nullptr;
rgb_filter_shader_program_ = nullptr;
@ -229,18 +277,16 @@ void OpenGLOutputBuilder::set_openGL_context_will_change(bool should_delete_reso
void OpenGLOutputBuilder::set_composite_sampling_function(const char *shader)
{
output_mutex_.lock();
std::lock_guard<std::mutex> lock_guard(output_mutex_);
composite_shader_ = strdup(shader);
reset_all_OpenGL_state();
output_mutex_.unlock();
}
void OpenGLOutputBuilder::set_rgb_sampling_function(const char *shader)
{
output_mutex_.lock();
std::lock_guard<std::mutex> lock_guard(output_mutex_);
rgb_shader_ = strdup(shader);
reset_all_OpenGL_state();
output_mutex_.unlock();
}
#pragma mark - Program compilation
@ -252,16 +298,25 @@ void OpenGLOutputBuilder::prepare_composite_input_shaders()
composite_input_shader_program_->set_output_size(IntermediateBufferWidth, IntermediateBufferHeight);
composite_separation_filter_program_ = OpenGL::IntermediateShader::make_chroma_luma_separation_shader();
composite_separation_filter_program_->set_source_texture_unit(composite_texture_unit);
composite_separation_filter_program_->set_source_texture_unit(work_texture_ ? work_texture_unit : composite_texture_unit);
composite_separation_filter_program_->set_output_size(IntermediateBufferWidth, IntermediateBufferHeight);
composite_y_filter_shader_program_ = OpenGL::IntermediateShader::make_luma_filter_shader();
composite_y_filter_shader_program_->set_source_texture_unit(separated_texture_unit);
composite_y_filter_shader_program_->set_output_size(IntermediateBufferWidth, IntermediateBufferHeight);
composite_chrominance_filter_shader_program_ = OpenGL::IntermediateShader::make_chroma_filter_shader();
composite_chrominance_filter_shader_program_->set_source_texture_unit(filtered_y_texture_unit);
composite_chrominance_filter_shader_program_->set_source_texture_unit(work_texture_ ? work_texture_unit : separated_texture_unit);
composite_chrominance_filter_shader_program_->set_output_size(IntermediateBufferWidth, IntermediateBufferHeight);
if(work_texture_)
{
composite_input_shader_program_->set_is_double_height(true, 0.0f, 0.0f);
composite_separation_filter_program_->set_is_double_height(true, 0.0f, 0.5f);
composite_chrominance_filter_shader_program_->set_is_double_height(true, 0.5f, 0.0f);
}
else
{
composite_input_shader_program_->set_is_double_height(false);
composite_separation_filter_program_->set_is_double_height(false);
composite_chrominance_filter_shader_program_->set_is_double_height(false);
}
}
void OpenGLOutputBuilder::prepare_rgb_input_shaders()
@ -295,7 +350,9 @@ void OpenGLOutputBuilder::prepare_source_vertex_array()
void OpenGLOutputBuilder::prepare_output_shader()
{
output_shader_program_ = OpenGL::OutputShader::make_shader("", "texture(texID, srcCoordinatesVarying).rgb", false);
output_shader_program_->set_source_texture_unit(filtered_texture_unit);
output_shader_program_->set_source_texture_unit(work_texture_ ? work_texture_unit : filtered_texture_unit);
// output_shader_program_->set_source_texture_unit(composite_texture_unit);
output_shader_program_->set_origin_is_double_height(!!work_texture_);
}
void OpenGLOutputBuilder::prepare_output_vertex_array()
@ -319,6 +376,7 @@ void OpenGLOutputBuilder::set_output_device(OutputDevice output_device)
composite_src_output_y_ = 0;
last_output_width_ = 0;
last_output_height_ = 0;
set_output_shader_width();
}
}
@ -362,30 +420,58 @@ void OpenGLOutputBuilder::set_colour_space_uniforms()
}
if(composite_input_shader_program_) composite_input_shader_program_->set_colour_conversion_matrices(fromRGB, toRGB);
if(composite_separation_filter_program_) composite_separation_filter_program_->set_colour_conversion_matrices(fromRGB, toRGB);
if(composite_chrominance_filter_shader_program_) composite_chrominance_filter_shader_program_->set_colour_conversion_matrices(fromRGB, toRGB);
}
float OpenGLOutputBuilder::get_composite_output_width() const
{
return ((float)colour_cycle_numerator_ * 4.0f) / (float)(colour_cycle_denominator_ * IntermediateBufferWidth);
}
void OpenGLOutputBuilder::set_output_shader_width()
{
if(output_shader_program_)
{
const float width = get_is_television_output() ? get_composite_output_width() : 1.0f;
output_shader_program_->set_input_width_scaler(width);
}
}
void OpenGLOutputBuilder::set_timing_uniforms()
{
OpenGL::IntermediateShader *intermediate_shaders[] = {
composite_input_shader_program_.get(),
composite_separation_filter_program_.get(),
composite_y_filter_shader_program_.get(),
composite_chrominance_filter_shader_program_.get()
};
bool extends = false;
float phaseCyclesPerTick = (float)colour_cycle_numerator_ / (float)(colour_cycle_denominator_ * cycles_per_line_);
for(int c = 0; c < 3; c++)
const float colour_subcarrier_frequency = (float)colour_cycle_numerator_ / (float)colour_cycle_denominator_;
const float output_width = get_composite_output_width();
const float sample_cycles_per_line = cycles_per_line_ / output_width;
if(composite_separation_filter_program_)
{
if(intermediate_shaders[c]) intermediate_shaders[c]->set_phase_cycles_per_sample(phaseCyclesPerTick, extends);
extends = true;
composite_separation_filter_program_->set_width_scalers(output_width, output_width);
composite_separation_filter_program_->set_separation_frequency(sample_cycles_per_line, colour_subcarrier_frequency);
composite_separation_filter_program_->set_extension(6.0f);
}
if(composite_chrominance_filter_shader_program_)
{
composite_chrominance_filter_shader_program_->set_width_scalers(output_width, output_width);
composite_chrominance_filter_shader_program_->set_extension(5.0f);
}
if(rgb_filter_shader_program_)
{
rgb_filter_shader_program_->set_width_scalers(1.0f, 1.0f);
rgb_filter_shader_program_->set_filter_coefficients(sample_cycles_per_line, (float)input_frequency_ * 0.5f);
}
if(output_shader_program_)
{
set_output_shader_width();
output_shader_program_->set_timing(height_of_display_, cycles_per_line_, horizontal_scan_period_, vertical_scan_period_, vertical_period_divider_);
}
if(composite_input_shader_program_)
{
composite_input_shader_program_->set_width_scalers(1.0f, output_width);
composite_input_shader_program_->set_extension(0.0f);
}
if(rgb_input_shader_program_)
{
rgb_input_shader_program_->set_width_scalers(1.0f, 1.0f);
}
if(output_shader_program_) output_shader_program_->set_timing(height_of_display_, cycles_per_line_, horizontal_scan_period_, vertical_scan_period_, vertical_period_divider_);
float colour_subcarrier_frequency = (float)colour_cycle_numerator_ / (float)colour_cycle_denominator_;
if(composite_separation_filter_program_) composite_separation_filter_program_->set_separation_frequency(cycles_per_line_, colour_subcarrier_frequency);
if(composite_y_filter_shader_program_) composite_y_filter_shader_program_->set_filter_coefficients(cycles_per_line_, colour_subcarrier_frequency * 0.25f);
if(composite_chrominance_filter_shader_program_) composite_chrominance_filter_shader_program_->set_filter_coefficients(cycles_per_line_, colour_subcarrier_frequency * 0.5f);
if(rgb_filter_shader_program_) rgb_filter_shader_program_->set_filter_coefficients(cycles_per_line_, (float)input_frequency_ * 0.5f);
}

View File

@ -66,13 +66,19 @@ class OpenGLOutputBuilder {
GLsizei composite_src_output_y_;
std::unique_ptr<OpenGL::OutputShader> output_shader_program_;
std::unique_ptr<OpenGL::IntermediateShader> composite_input_shader_program_, composite_separation_filter_program_, composite_y_filter_shader_program_, composite_chrominance_filter_shader_program_;
std::unique_ptr<OpenGL::IntermediateShader> rgb_input_shader_program_, rgb_filter_shader_program_;
OpenGL::TextureTarget composite_texture_; // receives raw composite levels
OpenGL::TextureTarget separated_texture_; // receives unfiltered Y in the R channel plus unfiltered but demodulated chrominance in G and B
OpenGL::TextureTarget filtered_y_texture_; // receives filtered Y in the R channel plus unfiltered chrominance in G and B
OpenGL::TextureTarget filtered_texture_; // receives filtered YIQ or YUV
std::unique_ptr<OpenGL::IntermediateShader> composite_input_shader_program_;
std::unique_ptr<OpenGL::IntermediateShader> composite_separation_filter_program_;
std::unique_ptr<OpenGL::IntermediateShader> composite_chrominance_filter_shader_program_;
std::unique_ptr<OpenGL::IntermediateShader> rgb_input_shader_program_;
std::unique_ptr<OpenGL::IntermediateShader> rgb_filter_shader_program_;
std::unique_ptr<OpenGL::TextureTarget> composite_texture_; // receives raw composite levels
std::unique_ptr<OpenGL::TextureTarget> separated_texture_; // receives filtered Y in the R channel plus unfiltered but demodulated chrominance in G and B
std::unique_ptr<OpenGL::TextureTarget> filtered_texture_; // receives filtered YIQ or YUV
std::unique_ptr<OpenGL::TextureTarget> work_texture_; // used for all intermediate rendering if texture fences are supported
std::unique_ptr<OpenGL::TextureTarget> framebuffer_; // the current pixel output
@ -88,6 +94,9 @@ class OpenGLOutputBuilder {
void reset_all_OpenGL_state();
GLsync fence_;
float get_composite_output_width() const;
void set_output_shader_width();
bool get_is_television_output();
public:
// These two are protected by output_mutex_.

View File

@ -15,6 +15,7 @@
#else
#include <OpenGL/OpenGL.h>
#include <OpenGL/gl3.h>
#include <OpenGL/gl3ext.h>
#endif
#endif

View File

@ -40,11 +40,14 @@ std::unique_ptr<IntermediateShader> IntermediateShader::make_shader(const char *
"in vec2 ends;"
"in vec3 phaseTimeAndAmplitude;"
"uniform float phaseCyclesPerTick;"
"uniform ivec2 outputTextureSize;"
"uniform float extension;"
"uniform %s texID;"
"uniform float offsets[5];"
"uniform vec2 widthScalers;"
"uniform float inputVerticalOffset;"
"uniform float outputVerticalOffset;"
"uniform float textureHeightDivisor;"
"out vec2 phaseAndAmplitudeVarying;"
"out vec2 inputPositionsVarying[11];"
@ -53,35 +56,51 @@ std::unique_ptr<IntermediateShader> IntermediateShader::make_shader(const char *
"void main(void)"
"{"
// odd vertices are on the left, even on the right
"float extent = float(gl_VertexID & 1);"
"float longitudinal = float((gl_VertexID & 2) >> 1);"
"vec2 inputPosition = vec2(mix(inputStart.x, ends.x, extent), inputStart.y);"
"vec2 outputPosition = vec2(mix(outputStart.x, ends.y, extent), outputStart.y);"
// inputPosition.x is either inputStart.x or ends.x, depending on whether it is on the left or the right;
// outputPosition.x is either outputStart.x or ends.y;
// .ys are inputStart.y and outputStart.y respectively
"vec2 inputPosition = vec2(mix(inputStart.x, ends.x, extent)*widthScalers[0], inputStart.y + inputVerticalOffset);"
"vec2 outputPosition = vec2(mix(outputStart.x, ends.y, extent)*widthScalers[1], outputStart.y + outputVerticalOffset);"
"inputPosition.y += longitudinal;"
"outputPosition.y += longitudinal;"
// extension is the amount to extend both the input and output by to add a full colour cycle at each end
"vec2 extensionVector = vec2(extension, 0.0) * 2.0 * (extent - 0.5);"
// extended[Input/Output]Position are [input/output]Position with the necessary applied extension
"vec2 extendedInputPosition = %s + extensionVector;"
"vec2 extendedOutputPosition = outputPosition + extensionVector;"
// keep iInputPositionVarying in whole source pixels, scale mappedInputPosition to the ordinary normalised range
"vec2 textureSize = vec2(textureSize(texID, 0));"
"iInputPositionVarying = extendedInputPosition;"
"vec2 mappedInputPosition = (extendedInputPosition + vec2(0.0, 0.5)) / textureSize;"
"vec2 mappedInputPosition = extendedInputPosition / textureSize;" // + vec2(0.0, 0.5)
"inputPositionsVarying[0] = mappedInputPosition - (vec2(offsets[0], 0.0) / textureSize);"
"inputPositionsVarying[1] = mappedInputPosition - (vec2(offsets[1], 0.0) / textureSize);"
"inputPositionsVarying[2] = mappedInputPosition - (vec2(offsets[2], 0.0) / textureSize);"
"inputPositionsVarying[3] = mappedInputPosition - (vec2(offsets[3], 0.0) / textureSize);"
"inputPositionsVarying[4] = mappedInputPosition - (vec2(offsets[4], 0.0) / textureSize);"
// setup input positions spaced as per the supplied offsets; these are for filtering where required
"inputPositionsVarying[0] = mappedInputPosition - (vec2(5.0, 0.0) / textureSize);"
"inputPositionsVarying[1] = mappedInputPosition - (vec2(4.0, 0.0) / textureSize);"
"inputPositionsVarying[2] = mappedInputPosition - (vec2(3.0, 0.0) / textureSize);"
"inputPositionsVarying[3] = mappedInputPosition - (vec2(2.0, 0.0) / textureSize);"
"inputPositionsVarying[4] = mappedInputPosition - (vec2(1.0, 0.0) / textureSize);"
"inputPositionsVarying[5] = mappedInputPosition;"
"inputPositionsVarying[6] = mappedInputPosition + (vec2(offsets[4], 0.0) / textureSize);"
"inputPositionsVarying[7] = mappedInputPosition + (vec2(offsets[3], 0.0) / textureSize);"
"inputPositionsVarying[8] = mappedInputPosition + (vec2(offsets[2], 0.0) / textureSize);"
"inputPositionsVarying[9] = mappedInputPosition + (vec2(offsets[1], 0.0) / textureSize);"
"inputPositionsVarying[10] = mappedInputPosition + (vec2(offsets[0], 0.0) / textureSize);"
"inputPositionsVarying[6] = mappedInputPosition + (vec2(1.0, 0.0) / textureSize);"
"inputPositionsVarying[7] = mappedInputPosition + (vec2(2.0, 0.0) / textureSize);"
"inputPositionsVarying[8] = mappedInputPosition + (vec2(3.0, 0.0) / textureSize);"
"inputPositionsVarying[9] = mappedInputPosition + (vec2(4.0, 0.0) / textureSize);"
"inputPositionsVarying[10] = mappedInputPosition + (vec2(5.0, 0.0) / textureSize);"
"delayLinePositionVarying = mappedInputPosition - vec2(0.0, 1.0);"
"phaseAndAmplitudeVarying.x = (phaseCyclesPerTick * (extendedOutputPosition.x - phaseTimeAndAmplitude.y) + (phaseTimeAndAmplitude.x / 256.0)) * 2.0 * 3.141592654;"
// setup phaseAndAmplitudeVarying.x as colour burst subcarrier phase, in radians;
// setup phaseAndAmplitudeVarying.x as colour burst amplitude
"phaseAndAmplitudeVarying.x = (extendedOutputPosition.x + (phaseTimeAndAmplitude.x / 64.0)) * 0.5 * 3.141592654;"
"phaseAndAmplitudeVarying.y = 0.33;" // TODO: reinstate connection with (phaseTimeAndAmplitude.y/256.0)
// determine output position by scaling the output position according to the texture size
"vec2 eyePosition = 2.0*(extendedOutputPosition / outputTextureSize) - vec2(1.0) + vec2(1.0)/outputTextureSize;"
"gl_Position = vec4(eyePosition, 0.0, 1.0);"
"}", sampler_type, input_variable);
@ -172,7 +191,6 @@ std::unique_ptr<IntermediateShader> IntermediateShader::make_chroma_luma_separat
"in vec2 phaseAndAmplitudeVarying;"
"in vec2 inputPositionsVarying[11];"
"uniform vec4 weights[3];"
"out vec3 fragColour;"
@ -180,37 +198,20 @@ std::unique_ptr<IntermediateShader> IntermediateShader::make_chroma_luma_separat
"void main(void)"
"{"
"vec4 samples[3] = vec4[]("
"vec4("
"texture(texID, inputPositionsVarying[0]).r,"
"texture(texID, inputPositionsVarying[1]).r,"
"texture(texID, inputPositionsVarying[2]).r,"
"texture(texID, inputPositionsVarying[3]).r"
"),"
"vec4("
"texture(texID, inputPositionsVarying[4]).r,"
"texture(texID, inputPositionsVarying[5]).r,"
"texture(texID, inputPositionsVarying[6]).r,"
"texture(texID, inputPositionsVarying[7]).r"
"),"
"vec4("
"texture(texID, inputPositionsVarying[8]).r,"
"texture(texID, inputPositionsVarying[9]).r,"
"texture(texID, inputPositionsVarying[10]).r,"
"0.0"
")"
"vec4 samples = vec4("
"texture(texID, inputPositionsVarying[3]).r,"
"texture(texID, inputPositionsVarying[4]).r,"
"texture(texID, inputPositionsVarying[5]).r,"
"texture(texID, inputPositionsVarying[6]).r"
");"
"float luminance = dot(samples, vec4(0.25));"
"float luminance = "
"dot(vec3("
"dot(samples[0], weights[0]),"
"dot(samples[1], weights[1]),"
"dot(samples[2], weights[2])"
"), vec3(1.0));"
"float chrominance = 0.5 * (samples[1].y - luminance) / phaseAndAmplitudeVarying.y;"
// define chroma to be whatever was here, minus luma
"float chrominance = 0.5 * (samples.z - luminance) / phaseAndAmplitudeVarying.y;"
"luminance /= (1.0 - phaseAndAmplitudeVarying.y);"
// split choma colours here, as the most direct place, writing out
// RGB = (luma, chroma.x, chroma.y)
"vec2 quadrature = vec2(cos(phaseAndAmplitudeVarying.x), -sin(phaseAndAmplitudeVarying.x));"
"fragColour = vec3(luminance, vec2(0.5) + (chrominance * quadrature));"
"}",false, false);
@ -232,41 +233,18 @@ std::unique_ptr<IntermediateShader> IntermediateShader::make_chroma_filter_shade
"void main(void)"
"{"
"vec3 samples[] = vec3[]("
"texture(texID, inputPositionsVarying[0]).rgb,"
"texture(texID, inputPositionsVarying[1]).rgb,"
"texture(texID, inputPositionsVarying[2]).rgb,"
"texture(texID, inputPositionsVarying[3]).rgb,"
"texture(texID, inputPositionsVarying[4]).rgb,"
"texture(texID, inputPositionsVarying[5]).rgb,"
"texture(texID, inputPositionsVarying[6]).rgb,"
"texture(texID, inputPositionsVarying[7]).rgb,"
"texture(texID, inputPositionsVarying[8]).rgb,"
"texture(texID, inputPositionsVarying[9]).rgb,"
"texture(texID, inputPositionsVarying[10]).rgb"
"texture(texID, inputPositionsVarying[6]).rgb"
");"
"vec4 chromaChannel1[] = vec4[]("
"vec4(samples[0].g, samples[1].g, samples[2].g, samples[3].g),"
"vec4(samples[4].g, samples[5].g, samples[6].g, samples[7].g),"
"vec4(samples[8].g, samples[9].g, samples[10].g, 0.0)"
");"
"vec4 chromaChannel2[] = vec4[]("
"vec4(samples[0].b, samples[1].b, samples[2].b, samples[3].b),"
"vec4(samples[4].b, samples[5].b, samples[6].b, samples[7].b),"
"vec4(samples[8].b, samples[9].b, samples[10].b, 0.0)"
");"
"vec4 chromaChannel1 = vec4(samples[0].g, samples[1].g, samples[2].g, samples[3].g);"
"vec4 chromaChannel2 = vec4(samples[0].b, samples[1].b, samples[2].b, samples[3].b);"
"vec3 lumaChromaColour = vec3(samples[5].r,"
"dot(vec3("
"dot(chromaChannel1[0], weights[0]),"
"dot(chromaChannel1[1], weights[1]),"
"dot(chromaChannel1[2], weights[2])"
"), vec3(1.0)),"
"dot(vec3("
"dot(chromaChannel2[0], weights[0]),"
"dot(chromaChannel2[1], weights[1]),"
"dot(chromaChannel2[2], weights[2])"
"), vec3(1.0))"
"vec3 lumaChromaColour = vec3(samples[2].r,"
"dot(chromaChannel1, vec4(0.25)),"
"dot(chromaChannel2, vec4(0.25))"
");"
"vec3 lumaChromaColourInRange = (lumaChromaColour - vec3(0.0, 0.5, 0.5)) * vec3(1.0, 2.0, 2.0);"
@ -274,52 +252,6 @@ std::unique_ptr<IntermediateShader> IntermediateShader::make_chroma_filter_shade
"}", false, false);
}
std::unique_ptr<IntermediateShader> IntermediateShader::make_luma_filter_shader()
{
return make_shader(
"#version 150\n"
"in vec2 inputPositionsVarying[11];"
"uniform vec4 weights[3];"
"out vec3 fragColour;"
"uniform sampler2D texID;"
"uniform mat3 lumaChromaToRGB;"
"void main(void)"
"{"
"vec3 samples[] = vec3[]("
"texture(texID, inputPositionsVarying[0]).rgb,"
"texture(texID, inputPositionsVarying[1]).rgb,"
"texture(texID, inputPositionsVarying[2]).rgb,"
"texture(texID, inputPositionsVarying[3]).rgb,"
"texture(texID, inputPositionsVarying[4]).rgb,"
"texture(texID, inputPositionsVarying[5]).rgb,"
"texture(texID, inputPositionsVarying[6]).rgb,"
"texture(texID, inputPositionsVarying[7]).rgb,"
"texture(texID, inputPositionsVarying[8]).rgb,"
"texture(texID, inputPositionsVarying[9]).rgb,"
"texture(texID, inputPositionsVarying[10]).rgb"
");"
"vec4 luminance[] = vec4[]("
"vec4(samples[0].r, samples[1].r, samples[2].r, samples[3].r),"
"vec4(samples[4].r, samples[5].r, samples[6].r, samples[7].r),"
"vec4(samples[8].r, samples[9].r, samples[10].r, 0.0)"
");"
"fragColour = vec3("
"dot(vec3("
"dot(luminance[0], weights[0]),"
"dot(luminance[1], weights[1]),"
"dot(luminance[2], weights[2])"
"), vec3(1.0)),"
"samples[5].gb"
");"
"}", false, false);
}
std::unique_ptr<IntermediateShader> IntermediateShader::make_rgb_filter_shader()
{
return make_shader(
@ -404,44 +336,53 @@ void IntermediateShader::set_filter_coefficients(float sampling_rate, float cuto
// Perform a linear search for the highest number of taps we can use with 11 samples.
GLfloat weights[12];
GLfloat offsets[5];
unsigned int taps = 21;
unsigned int taps = 11;
// unsigned int taps = 21;
while(1)
{
float coefficients[21];
SignalProcessing::FIRFilter luminance_filter(taps, sampling_rate, 0.0f, cutoff_frequency, SignalProcessing::FIRFilter::DefaultAttenuation);
luminance_filter.get_coefficients(coefficients);
int sample = 0;
int c = 0;
// int sample = 0;
// int c = 0;
memset(weights, 0, sizeof(float)*12);
memset(offsets, 0, sizeof(float)*5);
int halfSize = (taps >> 1);
while(c < halfSize && sample < 5)
for(int c = 0; c < taps; c++)
{
offsets[sample] = (float)(halfSize - c);
if((coefficients[c] < 0.0f) == (coefficients[c+1] < 0.0f) && c+1 < (taps >> 1))
{
weights[sample] = coefficients[c] + coefficients[c+1];
offsets[sample] -= (coefficients[c+1] / weights[sample]);
c += 2;
}
else
{
weights[sample] = coefficients[c];
c++;
}
sample ++;
}
if(c == halfSize) // i.e. we finished combining inputs before we ran out of space
{
weights[sample] = coefficients[c];
for(int c = 0; c < sample; c++)
{
weights[sample+c+1] = weights[sample-c-1];
}
break;
if(c < 5) offsets[c] = (halfSize - c);
weights[c] = coefficients[c];
}
break;
// int halfSize = (taps >> 1);
// while(c < halfSize && sample < 5)
// {
// offsets[sample] = (float)(halfSize - c);
// if((coefficients[c] < 0.0f) == (coefficients[c+1] < 0.0f) && c+1 < (taps >> 1))
// {
// weights[sample] = coefficients[c] + coefficients[c+1];
// offsets[sample] -= (coefficients[c+1] / weights[sample]);
// c += 2;
// }
// else
// {
// weights[sample] = coefficients[c];
// c++;
// }
// sample ++;
// }
// if(c == halfSize) // i.e. we finished combining inputs before we ran out of space
// {
// weights[sample] = coefficients[c];
// for(int c = 0; c < sample; c++)
// {
// weights[sample+c+1] = weights[sample-c-1];
// }
// break;
// }
taps -= 2;
}
@ -454,10 +395,9 @@ void IntermediateShader::set_separation_frequency(float sampling_rate, float col
set_filter_coefficients(sampling_rate, colour_burst_frequency);
}
void IntermediateShader::set_phase_cycles_per_sample(float phase_cycles_per_sample, bool extend_runs_to_full_cycle)
void IntermediateShader::set_extension(float extension)
{
set_uniform("phaseCyclesPerTick", (GLfloat)phase_cycles_per_sample);
set_uniform("extension", extend_runs_to_full_cycle ? ceilf(1.0f / phase_cycles_per_sample) : 0.0f);
set_uniform("extension", extension);
}
void IntermediateShader::set_colour_conversion_matrices(float *fromRGB, float *toRGB)
@ -465,3 +405,15 @@ void IntermediateShader::set_colour_conversion_matrices(float *fromRGB, float *t
set_uniform_matrix("lumaChromaToRGB", 3, false, toRGB);
set_uniform_matrix("rgbToLumaChroma", 3, false, fromRGB);
}
void IntermediateShader::set_width_scalers(float input_scaler, float output_scaler)
{
set_uniform("widthScalers", input_scaler, output_scaler);
}
void IntermediateShader::set_is_double_height(bool is_double_height, float input_offset, float output_offset)
{
set_uniform("textureHeightDivisor", is_double_height ? 2.0f : 1.0f);
set_uniform("inputVerticalOffset", input_offset);
set_uniform("outputVerticalOffset", output_offset);
}

View File

@ -44,11 +44,6 @@ public:
*/
static std::unique_ptr<IntermediateShader> make_chroma_filter_shader();
/*!
Constructs and returns an intermediate shader that will filter R while passing through G and B unchanged.
*/
static std::unique_ptr<IntermediateShader> make_luma_filter_shader();
/*!
Constructs and returns an intermediate shader that will filter R, G and B.
*/
@ -81,13 +76,24 @@ public:
geometry should be extended so that a complete colour cycle is included at both the beginning and end,
to occur upon the next `bind`.
*/
void set_phase_cycles_per_sample(float phase_cycles_per_sample, bool extend_runs_to_full_cycle);
void set_extension(float extension);
/*!
Queues setting the matrices that convert between RGB and chrominance/luminance to occur on the next `bind`.
*/
void set_colour_conversion_matrices(float *fromRGB, float *toRGB);
/*!
Sets the proportions of the input and output areas that should be considered the whole width 1.0 means use all available
space, 0.5 means use half, etc.
*/
void set_width_scalers(float input_scaler, float output_scaler);
/*!
Sets source and target vertical offsets.
*/
void set_is_double_height(bool is_double_height, float input_offset = 0.0f, float output_offset = 0.0f);
private:
static std::unique_ptr<IntermediateShader> make_shader(const char *fragment_shader, bool use_usampler, bool input_is_inputPosition);
};

View File

@ -38,6 +38,8 @@ std::unique_ptr<OutputShader> OutputShader::make_shader(const char *fragment_met
"uniform vec2 positionConversion;"
"uniform vec2 scanNormal;"
"uniform %s texID;"
"uniform float inputScaler;"
"uniform int textureHeightDivisor;"
"out float lateralVarying;"
"out vec2 srcCoordinatesVarying;"
@ -52,9 +54,10 @@ std::unique_ptr<OutputShader> OutputShader::make_shader(const char *fragment_met
"lateralVarying = lateral - 0.5;"
"vec2 vSrcCoordinates = vec2(x, vertical.y);"
"ivec2 textureSize = textureSize(texID, 0);"
"ivec2 textureSize = textureSize(texID, 0) * ivec2(1, textureHeightDivisor);"
"iSrcCoordinatesVarying = vSrcCoordinates;"
"srcCoordinatesVarying = vec2(vSrcCoordinates.x / textureSize.x, (vSrcCoordinates.y + 0.5) / textureSize.y);"
"srcCoordinatesVarying = vec2(inputScaler * vSrcCoordinates.x / textureSize.x, (vSrcCoordinates.y + 0.5) / textureSize.y);"
"srcCoordinatesVarying.x = srcCoordinatesVarying.x - mod(srcCoordinatesVarying.x, 1.0 / textureSize.x);"
"vec2 vPosition = vec2(x, vertical.x);"
"vec2 floatingPosition = (vPosition / positionConversion) + lateral * scanNormal;"
@ -117,3 +120,13 @@ void OutputShader::set_timing(unsigned int height_of_display, unsigned int cycle
set_uniform("scanNormal", scan_normal[0], scan_normal[1]);
set_uniform("positionConversion", (GLfloat)horizontal_scan_period, (GLfloat)vertical_scan_period / (GLfloat)vertical_period_divider);
}
void OutputShader::set_input_width_scaler(float input_scaler)
{
set_uniform("inputScaler", input_scaler);
}
void OutputShader::set_origin_is_double_height(bool is_double_height)
{
set_uniform("textureHeightDivisor", is_double_height ? 2 : 1);
}

View File

@ -58,6 +58,16 @@ public:
to occur upon the next `bind`.
*/
void set_timing(unsigned int height_of_display, unsigned int cycles_per_line, unsigned int horizontal_scan_period, unsigned int vertical_scan_period, unsigned int vertical_period_divider);
/*!
*/
void set_origin_is_double_height(bool is_double_height);
/*!
Sets the proportion of the input area that should be considered the whole width 1.0 means use all available
space, 0.5 means use half, etc.
*/
void set_input_width_scaler(float input_scaler);
};
}

View File

@ -156,7 +156,8 @@ void Shader::set_uniform(const std::string &name, GLint value1, GLint value2)
void Shader::set_uniform(const std::string &name, GLfloat value1, GLfloat value2)
{
enqueue_function([name, value1, value2, this] {
glUniform2f(location(), value1, value2);
GLint location = location();
glUniform2f(location, value1, value2);
});
}

View File

@ -12,7 +12,7 @@
using namespace OpenGL;
TextureTarget::TextureTarget(GLsizei width, GLsizei height, GLenum texture_unit) :
TextureTarget::TextureTarget(GLsizei width, GLsizei height, GLenum texture_unit, GLint mag_filter) :
_width(width),
_height(height),
_pixel_shader(nullptr),
@ -33,8 +33,8 @@ TextureTarget::TextureTarget(GLsizei width, GLsizei height, GLenum texture_unit)
uint8_t *blank_buffer = (uint8_t *)calloc((size_t)(_expanded_width * _expanded_height), 4);
glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, (GLsizei)_expanded_width, (GLsizei)_expanded_height, 0, GL_RGBA, GL_UNSIGNED_BYTE, blank_buffer);
free(blank_buffer);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, mag_filter);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, _texture, 0);

View File

@ -30,7 +30,7 @@ class TextureTarget {
@param height The height of target to create.
@param texture_unit A texture unit on which to bind the texture.
*/
TextureTarget(GLsizei width, GLsizei height, GLenum texture_unit);
TextureTarget(GLsizei width, GLsizei height, GLenum texture_unit, GLint mag_filter);
~TextureTarget();
/*!