Omit copy to RAM when possible with VAAPI

This commit is contained in:
loki 2021-08-12 21:11:40 +02:00
parent b8bfc13cf9
commit 6721155155
7 changed files with 361 additions and 159 deletions

View File

@ -106,10 +106,13 @@ else()
add_compile_definitions(SUNSHINE_PLATFORM="linux")
list(APPEND SUNSHINE_DEFINITIONS APPS_JSON="apps_linux.json")
if(NOT DEFINED SUNSHINE_DISABLE_X11)
option(SUNSHINE_ENABLE_DRM "Enable KMS grab if available" ON)
option(SUNSHINE_ENABLE_X11 "Enable X11 grab if available" ON)
if(${SUNSHINE_ENABLE_X11})
find_package(X11)
endif()
if(NOT DEFINED SUNSHINE_DISABLE_DRM)
if(${SUNSHINE_ENABLE_DRM})
find_package(LIBDRM)
endif()

View File

@ -1,10 +1,6 @@
#include "graphics.h"
#include "sunshine/video.h"
extern "C" {
#include <libavcodec/avcodec.h>
}
#include <fcntl.h>
// I want to have as little build dependencies as possible
@ -460,7 +456,7 @@ std::optional<nv12_t> import_target(display_t::pointer egl_display, std::array<f
return nv12;
}
void egl_t::set_colorspace(std::uint32_t colorspace, std::uint32_t color_range) {
void sws_t::set_colorspace(std::uint32_t colorspace, std::uint32_t color_range) {
video::color_t *color_p;
switch(colorspace) {
case 5: // SWS_CS_SMPTE170M
@ -491,31 +487,25 @@ void egl_t::set_colorspace(std::uint32_t colorspace, std::uint32_t color_range)
color_matrix.update(members, sizeof(members) / sizeof(decltype(members[0])));
}
int egl_t::init(int in_width, int in_height, file_t &&fd) {
file = std::move(fd);
std::optional<sws_t> sws_t::make(int in_width, int in_height, int out_width, int out_heigth, gl::tex_t &&tex) {
sws_t sws;
if(!gbm::create_device) {
BOOST_LOG(warning) << "libgbm not initialized"sv;
return -1;
}
// Ensure aspect ratio is maintained
auto scalar = std::fminf(out_width / (float)in_width, out_heigth / (float)in_height);
auto out_width_f = in_width * scalar;
auto out_height_f = in_height * scalar;
gbm.reset(gbm::create_device(file.el));
if(!gbm) {
BOOST_LOG(error) << "Couldn't create GBM device: ["sv << util::hex(eglGetError()).to_string_view() << ']';
return -1;
}
// result is always positive
auto offsetX_f = (out_width - out_width_f) / 2;
auto offsetY_f = (out_heigth - out_height_f) / 2;
display = make_display(gbm.get());
if(!display) {
return -1;
}
sws.width = out_width_f;
sws.height = out_height_f;
auto ctx_opt = make_ctx(display.get());
if(!ctx_opt) {
return -1;
}
sws.offsetX = offsetX_f;
sws.offsetY = offsetY_f;
ctx = std::move(*ctx_opt);
auto width_i = 1.0f / sws.width;
{
const char *sources[] {
@ -549,28 +539,37 @@ int egl_t::init(int in_width, int in_height, file_t &&fd) {
}
if(error_flag) {
return -1;
return std::nullopt;
}
auto program = gl::program_t::link(compiled_sources[1].left(), compiled_sources[0].left());
if(program.has_right()) {
BOOST_LOG(error) << "GL linker: "sv << program.right();
return -1;
return std::nullopt;
}
// UV - shader
this->program[1] = std::move(program.left());
sws.program[1] = std::move(program.left());
program = gl::program_t::link(compiled_sources[3].left(), compiled_sources[2].left());
if(program.has_right()) {
BOOST_LOG(error) << "GL linker: "sv << program.right();
return -1;
return std::nullopt;
}
// Y - shader
this->program[0] = std::move(program.left());
sws.program[0] = std::move(program.left());
}
auto loc_width_i = gl::ctx.GetUniformLocation(sws.program[1].handle(), "width_i");
if(loc_width_i < 0) {
BOOST_LOG(error) << "Couldn't find uniform [width_i]"sv;
return std::nullopt;
}
gl::ctx.UseProgram(sws.program[1].handle());
gl::ctx.Uniform1fv(loc_width_i, 1, &width_i);
auto color_p = &video::colors[0];
std::pair<const char *, std::string_view> members[] {
std::make_pair("color_vec_y", util::view(color_p->color_vec_y)),
@ -580,25 +579,46 @@ int egl_t::init(int in_width, int in_height, file_t &&fd) {
std::make_pair("range_uv", util::view(color_p->range_uv)),
};
auto color_matrix = program[0].uniform("ColorMatrix", members, sizeof(members) / sizeof(decltype(members[0])));
auto color_matrix = sws.program[0].uniform("ColorMatrix", members, sizeof(members) / sizeof(decltype(members[0])));
if(!color_matrix) {
return -1;
return std::nullopt;
}
this->color_matrix = std::move(*color_matrix);
sws.color_matrix = std::move(*color_matrix);
tex_in = gl::tex_t::make(1);
sws.tex = std::move(tex);
this->in_width = in_width;
this->in_height = in_height;
return 0;
gl_drain_errors;
return std::move(sws);
}
int egl_t::convert(platf::img_t &img) {
auto tex = tex_in[0];
std::optional<sws_t> sws_t::make(int in_width, int in_height, int out_width, int out_heigth) {
auto tex = gl::tex_t::make(1);
gl::ctx.BindTexture(GL_TEXTURE_2D, tex[0]);
gl::ctx.TexStorage2D(GL_TEXTURE_2D, 1, GL_RGBA8, in_width, in_height);
gl::ctx.BindTexture(GL_TEXTURE_2D, tex);
gl::ctx.TexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, in_width, in_height, GL_BGRA, GL_UNSIGNED_BYTE, img.data);
return make(in_width, in_height, out_width, out_heigth, std::move(tex));
}
void sws_t::load_ram(platf::img_t &img) {
gl::ctx.BindTexture(GL_TEXTURE_2D, tex[0]);
gl::ctx.TexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, img.width, img.height, GL_BGRA, GL_UNSIGNED_BYTE, img.data);
}
void sws_t::load_vram(platf::img_t &img, int offset_x, int offset_y, int framebuffer) {
gl::ctx.BindFramebuffer(GL_FRAMEBUFFER, framebuffer);
gl::ctx.ReadBuffer(GL_COLOR_ATTACHMENT0);
gl::ctx.BindTexture(GL_TEXTURE_2D, tex[0]);
gl::ctx.CopyTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, offset_x, offset_y, img.width, img.height);
gl::ctx.Flush();
}
int sws_t::convert(nv12_t &nv12) {
auto texture = tex[0];
gl::ctx.BindTexture(GL_TEXTURE_2D, texture);
GLenum attachments[] {
GL_COLOR_ATTACHMENT0,
@ -615,61 +635,17 @@ int egl_t::convert(platf::img_t &img) {
return -1;
}
gl::ctx.BindTexture(GL_TEXTURE_2D, tex);
gl::ctx.BindTexture(GL_TEXTURE_2D, texture);
gl::ctx.UseProgram(program[x].handle());
program[x].bind(color_matrix);
gl::ctx.Viewport(offsetX / (x + 1), offsetY / (x + 1), out_width / (x + 1), out_height / (x + 1));
gl::ctx.Viewport(offsetX / (x + 1), offsetY / (x + 1), width / (x + 1), height / (x + 1));
gl::ctx.DrawArrays(GL_TRIANGLES, 0, 3);
}
return 0;
}
int egl_t::_set_frame(AVFrame *frame) {
this->hwframe.reset(frame);
this->frame = frame;
// Ensure aspect ratio is maintained
auto scalar = std::fminf(frame->width / (float)in_width, frame->height / (float)in_height);
auto out_width_f = in_width * scalar;
auto out_height_f = in_height * scalar;
// result is always positive
auto offsetX_f = (frame->width - out_width_f) / 2;
auto offsetY_f = (frame->height - out_height_f) / 2;
out_width = out_width_f;
out_height = out_height_f;
offsetX = offsetX_f;
offsetY = offsetY_f;
auto tex = tex_in[0];
gl::ctx.BindTexture(GL_TEXTURE_2D, tex);
gl::ctx.TexStorage2D(GL_TEXTURE_2D, 1, GL_RGBA8, in_width, in_height);
auto loc_width_i = gl::ctx.GetUniformLocation(program[1].handle(), "width_i");
if(loc_width_i < 0) {
BOOST_LOG(error) << "Couldn't find uniform [width_i]"sv;
return -1;
}
auto width_i = 1.0f / out_width;
gl::ctx.UseProgram(program[1].handle());
gl::ctx.Uniform1fv(loc_width_i, 1, &width_i);
gl_drain_errors;
return 0;
}
egl_t::~egl_t() {
if(gl::ctx.GetError) {
gl_drain_errors;
}
}
} // namespace egl
void free_frame(AVFrame *frame) {

View File

@ -155,12 +155,6 @@ int init();
namespace egl {
using display_t = util::dyn_safe_ptr_v2<void, EGLBoolean, &eglTerminate>;
KITTY_USING_MOVE_T(file_t, int, -1, {
if(el >= 0) {
close(el);
}
});
struct rgb_img_t {
display_t::pointer display;
EGLImage xrgb8;
@ -227,37 +221,24 @@ std::optional<nv12_t> import_target(
std::array<file_t, nv12_img_t::num_fds> &&fds,
const surface_descriptor_t &r8, const surface_descriptor_t &gr88);
class egl_t : public platf::hwdevice_t {
class sws_t {
public:
void set_colorspace(std::uint32_t colorspace, std::uint32_t color_range) override;
static std::optional<sws_t> make(int in_width, int in_height, int out_width, int out_heigth, gl::tex_t &&tex);
static std::optional<sws_t> make(int in_width, int in_height, int out_width, int out_heigth);
int init(int in_width, int in_height, file_t &&fd);
int convert(nv12_t &nv12);
int convert(platf::img_t &img) override;
void load_ram(platf::img_t &img);
void load_vram(platf::img_t &img, int offset_x, int offset_y, int framebuffer);
/**
* Any specialization needs to populate nv12_t nv12
* Then call this function
*/
int _set_frame(AVFrame *frame);
void set_colorspace(std::uint32_t colorspace, std::uint32_t color_range);
~egl_t() override;
int in_width, in_height;
int out_width, out_height;
int offsetX, offsetY;
frame_t hwframe;
file_t file;
gbm::gbm_t gbm;
display_t display;
ctx_t ctx;
gl::tex_t tex_in;
nv12_t nv12;
gl::tex_t tex;
gl::program_t program[2];
gl::buffer_t color_matrix;
int width, height;
int offsetX, offsetY;
};
bool fail();

View File

@ -125,8 +125,8 @@ public:
return drmModeGetCrtc(fd.el, id);
}
egl::file_t handleFD(std::uint32_t handle) {
egl::file_t fb_fd;
file_t handleFD(std::uint32_t handle) {
file_t fb_fd;
auto status = drmPrimeHandleToFD(fd.el, handle, 0 /* flags */, &fb_fd.el);
if(status) {
@ -180,7 +180,7 @@ public:
}
egl::file_t fd;
file_t fd;
plane_res_t plane_res;
};
@ -228,12 +228,11 @@ class display_t : public platf::display_t {
public:
display_t(mem_type_e mem_type) : platf::display_t(), mem_type { mem_type } {}
int init(const std::string &display_name, int framerate) {
if(!gbm::create_device) {
BOOST_LOG(warning) << "libgbm not initialized"sv;
return -1;
}
mem_type_e mem_type;
std::chrono::nanoseconds delay;
int init(const std::string &display_name, int framerate) {
delay = std::chrono::nanoseconds { 1s } / framerate;
constexpr auto path = "/dev/dri/card1";
@ -244,8 +243,6 @@ public:
int monitor_index = util::from_view(display_name);
int monitor = 0;
int pitch;
auto end = std::end(card);
for(auto plane = std::begin(card); plane != end; ++plane) {
if(monitor != monitor_index) {
@ -299,6 +296,30 @@ public:
return -1;
}
return 0;
}
int img_width, img_height;
int pitch;
card_t card;
file_t fb_fd;
};
class display_ram_t : public display_t {
public:
display_ram_t(mem_type_e mem_type) : display_t(mem_type) {}
int init(const std::string &display_name, int framerate) {
if(!gbm::create_device) {
BOOST_LOG(warning) << "libgbm not initialized"sv;
return -1;
}
if(display_t::init(display_name, framerate)) {
return -1;
}
gbm.reset(gbm::create_device(card.fd.el));
if(!gbm) {
BOOST_LOG(error) << "Couldn't create GBM device: ["sv << util::hex(eglGetError()).to_string_view() << ']';
@ -400,24 +421,100 @@ public:
return 0;
}
int img_width, img_height;
mem_type_e mem_type;
std::chrono::nanoseconds delay;
card_t card;
egl::file_t fb_fd;
gbm::gbm_t gbm;
egl::display_t display;
egl::ctx_t ctx;
egl::rgb_t rgb;
};
class display_vram_t : public display_t {
public:
display_vram_t(mem_type_e mem_type) : display_t(mem_type) {}
std::shared_ptr<hwdevice_t> make_hwdevice(pix_fmt_e pix_fmt) override {
if(mem_type == mem_type_e::vaapi) {
return va::make_hwdevice(width, height, dup(card.fd.el), offset_x, offset_y,
{
fb_fd.el,
img_width,
img_height,
0,
pitch,
});
}
BOOST_LOG(error) << "Unsupported pixel format for egl::display_vram_t: "sv << platf::from_pix_fmt(pix_fmt);
return nullptr;
}
std::shared_ptr<img_t> alloc_img() override {
auto img = std::make_shared<img_t>();
img->width = width;
img->height = height;
img->pixel_pitch = 4;
img->row_pitch = img->pixel_pitch * width;
return img;
}
int dummy_img(platf::img_t *img) override {
return 0;
}
capture_e capture(snapshot_cb_t &&snapshot_cb, std::shared_ptr<img_t> img, bool *cursor) {
auto next_frame = std::chrono::steady_clock::now();
while(img) {
auto now = std::chrono::steady_clock::now();
if(next_frame > now) {
std::this_thread::sleep_for((next_frame - now) / 3 * 2);
}
while(next_frame > now) {
now = std::chrono::steady_clock::now();
}
next_frame = now + delay;
auto status = snapshot(img.get(), 1000ms, *cursor);
switch(status) {
case platf::capture_e::reinit:
case platf::capture_e::error:
return status;
case platf::capture_e::timeout:
std::this_thread::sleep_for(1ms);
continue;
case platf::capture_e::ok:
img = snapshot_cb(img);
break;
default:
BOOST_LOG(error) << "Unrecognized capture status ["sv << (int)status << ']';
return status;
}
}
return capture_e::ok;
}
capture_e snapshot(img_t * /*img_out_base */, std::chrono::milliseconds /* timeout */, bool /* cursor */) {
return capture_e::ok;
}
};
} // namespace kms
std::shared_ptr<display_t> kms_display(mem_type_e hwdevice_type, const std::string &display_name, int framerate) {
auto disp = std::make_shared<kms::display_t>(hwdevice_type);
if(hwdevice_type == mem_type_e::vaapi) {
auto disp = std::make_shared<kms::display_vram_t>(hwdevice_type);
if(!disp->init(display_name, framerate)) {
return disp;
}
// In the case of failure, attempt the old method for VAAPI
}
auto disp = std::make_shared<kms::display_ram_t>(hwdevice_type);
if(disp->init(display_name, framerate)) {
return nullptr;

View File

@ -1,7 +1,17 @@
#ifndef SUNSHINE_PLATFORM_MISC_H
#define SUNSHINE_PLATFORM_MISC_H
#include <unistd.h>
#include <vector>
#include "sunshine/utility.h"
KITTY_USING_MOVE_T(file_t, int, -1, {
if(el >= 0) {
close(el);
}
});
namespace dyn {
typedef void (*apiproc)(void);

View File

@ -160,29 +160,47 @@ int init() {
int vaapi_make_hwdevice_ctx(platf::hwdevice_t *base, AVBufferRef **hw_device_buf);
class va_t : public egl::egl_t {
class va_t : public platf::hwdevice_t {
public:
int init(int in_width, int in_height, const char *render_device) {
if(!va::initialize) {
BOOST_LOG(warning) << "libva not initialized"sv;
int init(int in_width, int in_height, file_t &&render_device) {
file = std::move(render_device);
if(!va::initialize || !gbm::create_device) {
if(!va::initialize) BOOST_LOG(warning) << "libva not initialized"sv;
if(!gbm::create_device) BOOST_LOG(warning) << "libgbm not initialized"sv;
return -1;
}
data = (void *)vaapi_make_hwdevice_ctx;
this->data = (void *)vaapi_make_hwdevice_ctx;
egl::file_t fd = open(render_device, O_RDWR);
if(fd.el < 0) {
gbm.reset(gbm::create_device(file.el));
if(!gbm) {
char string[1024];
BOOST_LOG(error) << "Couldn't open "sv << render_device << ": " << strerror_r(errno, string, sizeof(string));
BOOST_LOG(error) << "Couldn't create GBM device: ["sv << strerror_r(errno, string, sizeof(string)) << ']';
return -1;
}
return egl::egl_t::init(in_width, in_height, std::move(fd));
display = egl::make_display(gbm.get());
if(!display) {
return -1;
}
auto ctx_opt = egl::make_ctx(display.get());
if(!ctx_opt) {
return -1;
}
ctx = std::move(*ctx_opt);
width = in_width;
height = in_height;
return 0;
}
int set_frame(AVFrame *frame) override {
// No deallocation necessary
int _set_frame(AVFrame *frame) {
this->hwframe.reset(frame);
this->frame = frame;
if(av_hwframe_get_buffer(frame->hw_frames_ctx, frame, 0)) {
BOOST_LOG(error) << "Couldn't get hwframe for VAAPI"sv;
@ -194,7 +212,7 @@ public:
va::VASurfaceID surface = (std::uintptr_t)frame->data[3];
auto status = va::exportSurfaceHandle(
va_display,
this->va_display,
surface,
va::SURFACE_ATTRIB_MEM_TYPE_DRM_PRIME_2,
va::EXPORT_SURFACE_WRITE_ONLY | va::EXPORT_SURFACE_COMPOSED_LAYERS,
@ -207,7 +225,7 @@ public:
}
// Keep track of file descriptors
std::array<egl::file_t, egl::nv12_img_t::num_fds> fds;
std::array<file_t, egl::nv12_img_t::num_fds> fds;
for(int x = 0; x < prime.num_objects; ++x) {
fds[x] = prime.objects[x].fd;
}
@ -234,12 +252,106 @@ public:
return -1;
}
nv12 = std::move(*nv12_opt);
this->nv12 = std::move(*nv12_opt);
return egl::egl_t::_set_frame(frame);
return 0;
}
void set_colorspace(std::uint32_t colorspace, std::uint32_t color_range) override {
sws.set_colorspace(colorspace, color_range);
}
va::display_t::pointer va_display;
file_t file;
frame_t hwframe;
gbm::gbm_t gbm;
egl::display_t display;
egl::ctx_t ctx;
egl::sws_t sws;
egl::nv12_t nv12;
int width, height;
};
class va_ram_t : public va_t {
public:
int convert(platf::img_t &img) override {
sws.load_ram(img);
sws.convert(nv12);
return 0;
}
int set_frame(AVFrame *frame) override {
if(_set_frame(frame)) {
return -1;
}
auto sws_opt = egl::sws_t::make(width, height, frame->width, frame->height);
if(!sws_opt) {
return -1;
}
this->sws = std::move(*sws_opt);
return 0;
}
};
class va_vram_t : public va_t {
public:
int convert(platf::img_t &img) override {
sws.load_vram(img, offset_x, offset_y, framebuffer[0]);
sws.convert(nv12);
return 0;
}
int init(int in_width, int in_height, file_t &&render_device, int offset_x, int offset_y, const egl::surface_descriptor_t &sd) {
if(va_t::init(in_width, in_height, std::move(render_device))) {
return -1;
}
auto rgb_opt = egl::import_source(display.get(), sd);
if(!rgb_opt) {
return -1;
}
rgb = std::move(*rgb_opt);
framebuffer = gl::frame_buf_t::make(1);
framebuffer.bind(std::begin(rgb->tex), std::end(rgb->tex));
this->offset_x = offset_x;
this->offset_y = offset_y;
return 0;
}
int set_frame(AVFrame *frame) override {
if(_set_frame(frame)) {
return -1;
}
auto sws_opt = egl::sws_t::make(width, height, frame->width, frame->height);
if(!sws_opt) {
return -1;
}
this->sws = std::move(*sws_opt);
return 0;
}
file_t fb_fd;
egl::rgb_t rgb;
gl::frame_buf_t framebuffer;
int offset_x, offset_y;
};
/**
@ -343,10 +455,27 @@ int vaapi_make_hwdevice_ctx(platf::hwdevice_t *base, AVBufferRef **hw_device_buf
}
std::shared_ptr<platf::hwdevice_t> make_hwdevice(int width, int height) {
auto egl = std::make_shared<va::va_t>();
auto render_device = config::video.adapter_name.empty() ? "/dev/dri/renderD128" : config::video.adapter_name.c_str();
if(egl->init(width, height, render_device)) {
file_t file = open(render_device, O_RDWR);
if(file.el < 0) {
char string[1024];
BOOST_LOG(error) << "Couldn't open "sv << render_device << ": " << strerror_r(errno, string, sizeof(string));
return nullptr;
}
auto egl = std::make_shared<va::va_ram_t>();
if(egl->init(width, height, std::move(file))) {
return nullptr;
}
return egl;
}
std::shared_ptr<platf::hwdevice_t> make_hwdevice(int width, int height, file_t &&card, int offset_x, int offset_y, const egl::surface_descriptor_t &sd) {
auto egl = std::make_shared<va::va_vram_t>();
if(egl->init(width, height, std::move(card), offset_x, offset_y, sd)) {
return nullptr;
}

View File

@ -1,9 +1,15 @@
#ifndef SUNSHINE_DISPLAY_H
#define SUNSHINE_DISPLAY_H
#ifndef SUNSHINE_VAAPI_H
#define SUNSHINE_VAAPI_H
#include "misc.h"
#include "sunshine/platform/common.h"
namespace egl {
struct surface_descriptor_t;
}
namespace va {
std::shared_ptr<platf::hwdevice_t> make_hwdevice(int width, int height);
std::shared_ptr<platf::hwdevice_t> make_hwdevice(int width, int height, file_t &&card, int offset_x, int offset_y, const egl::surface_descriptor_t &sd);
int init();
} // namespace va