git.sesse.net Git - nageru/blob - nageru/pbo_frame_allocator.cpp

   1 #include "pbo_frame_allocator.h"
   2
   3 #include <bmusb/bmusb.h>
   4 #include <movit/util.h>
   5 #include <stdbool.h>
   6 #include <stdint.h>
   7 #include <stdio.h>
   8 #include <cstddef>
   9
  10 #include "flags.h"
  11 #include "mjpeg_encoder.h"
  12 #include "v210_converter.h"
  13 #include "va_display_with_cleanup.h"
  14
  15 using namespace std;
  16
  17 namespace {
  18
  19 void set_clamp_to_edge()
  20 {
  21         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
  22         check_error();
  23         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
  24         check_error();
  25         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
  26         check_error();
  27 }
  28
  29 }  // namespace
  30
  31 PBOFrameAllocator::PBOFrameAllocator(bmusb::PixelFormat pixel_format, size_t frame_size, GLuint width, GLuint height, unsigned card_index, MJPEGEncoder *mjpeg_encoder, size_t num_queued_frames, GLenum buffer, GLenum permissions, GLenum map_bits)
  32         : card_index(card_index), mjpeg_encoder(mjpeg_encoder), pixel_format(pixel_format), buffer(buffer)
  33 {
  34         userdata.reset(new Userdata[num_queued_frames]);
  35         for (size_t i = 0; i < num_queued_frames; ++i) {
  36                 init_frame(i, frame_size, width, height, permissions, map_bits);
  37         }
  38         glBindBuffer(buffer, 0);
  39         check_error();
  40         glBindTexture(GL_TEXTURE_2D, 0);
  41         check_error();
  42 }
  43
  44 void PBOFrameAllocator::init_frame(size_t frame_idx, size_t frame_size, GLuint width, GLuint height, GLenum permissions, GLenum map_bits)
  45 {
  46         GLuint pbo;
  47         glGenBuffers(1, &pbo);
  48         check_error();
  49         glBindBuffer(buffer, pbo);
  50         check_error();
  51         glBufferStorage(buffer, frame_size, nullptr, permissions | GL_MAP_PERSISTENT_BIT);
  52         check_error();
  53
  54         Frame frame;
  55         frame.data = (uint8_t *)glMapBufferRange(buffer, 0, frame_size, permissions | map_bits | GL_MAP_PERSISTENT_BIT);
  56         frame.data2 = frame.data + frame_size / 2;
  57         check_error();
  58         frame.size = frame_size;
  59         Userdata *ud = &userdata[frame_idx];
  60         frame.userdata = ud;
  61         ud->pbo = pbo;
  62         ud->pixel_format = pixel_format;
  63         ud->data_copy_malloc = new uint8_t[frame_size];
  64         frame.owner = this;
  65
  66         // For 8-bit non-planar Y'CbCr, we ask the driver to split Y' and Cb/Cr
  67         // into separate textures. For 10-bit, the input format (v210)
  68         // is complicated enough that we need to interpolate up to 4:4:4,
  69         // which we do in a compute shader ourselves. For BGRA, the data
  70         // is already 4:4:4:4.
  71         frame.interleaved = (pixel_format == bmusb::PixelFormat_8BitYCbCr);
  72
  73         // Create textures. We don't allocate any data for the second field at this point
  74         // (just create the texture state with the samplers), since our default assumed
  75         // resolution is progressive.
  76         switch (pixel_format) {
  77         case bmusb::PixelFormat_8BitYCbCr:
  78                 glGenTextures(2, ud->tex_y);
  79                 check_error();
  80                 glGenTextures(2, ud->tex_cbcr);
  81                 check_error();
  82                 break;
  83         case bmusb::PixelFormat_10BitYCbCr:
  84                 glGenTextures(2, ud->tex_v210);
  85                 check_error();
  86                 glGenTextures(2, ud->tex_444);
  87                 check_error();
  88                 break;
  89         case bmusb::PixelFormat_8BitBGRA:
  90                 glGenTextures(2, ud->tex_rgba);
  91                 check_error();
  92                 break;
  93         case bmusb::PixelFormat_8BitYCbCrPlanar:
  94                 glGenTextures(2, ud->tex_y);
  95                 check_error();
  96                 glGenTextures(2, ud->tex_cb);
  97                 check_error();
  98                 glGenTextures(2, ud->tex_cr);
  99                 check_error();
 100                 break;
 101         default:
 102                 assert(false);
 103         }
 104
 105         ud->last_width[0] = width;
 106         ud->last_height[0] = height;
 107         ud->last_cbcr_width[0] = width / 2;
 108         ud->last_cbcr_height[0] = height;
 109         ud->last_v210_width[0] = 0;
 110
 111         ud->last_width[1] = 0;
 112         ud->last_height[1] = 0;
 113         ud->last_cbcr_width[1] = 0;
 114         ud->last_cbcr_height[1] = 0;
 115         ud->last_v210_width[1] = 0;
 116
 117         ud->last_interlaced = false;
 118         ud->last_has_signal = false;
 119         ud->last_is_connected = false;
 120         for (unsigned field = 0; field < 2; ++field) {
 121                 switch (pixel_format) {
 122                 case bmusb::PixelFormat_10BitYCbCr: {
 123                         const size_t v210_width = v210Converter::get_minimum_v210_texture_width(width);
 124
 125                         // Seemingly we need to set the minification filter even though
 126                         // shader image loads don't use them, or NVIDIA will just give us
 127                         // zero back.
 128                         glBindTexture(GL_TEXTURE_2D, ud->tex_v210[field]);
 129                         check_error();
 130                         glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 131                         check_error();
 132                         if (field == 0) {
 133                                 ud->last_v210_width[0] = v210_width;
 134                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, v210_width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 135                                 check_error();
 136                         }
 137
 138                         glBindTexture(GL_TEXTURE_2D, ud->tex_444[field]);
 139                         check_error();
 140                         set_clamp_to_edge();
 141                         if (field == 0) {
 142                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB10_A2, width, height, 0, GL_RGBA, GL_UNSIGNED_INT_2_10_10_10_REV, nullptr);
 143                                 check_error();
 144                         }
 145                         break;
 146                 }
 147                 case bmusb::PixelFormat_8BitYCbCr:
 148                         glBindTexture(GL_TEXTURE_2D, ud->tex_y[field]);
 149                         check_error();
 150                         set_clamp_to_edge();
 151                         if (field == 0) {
 152                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 153                                 check_error();
 154                         }
 155
 156                         glBindTexture(GL_TEXTURE_2D, ud->tex_cbcr[field]);
 157                         check_error();
 158                         set_clamp_to_edge();
 159                         if (field == 0) {
 160                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, width / 2, height, 0, GL_RG, GL_UNSIGNED_BYTE, nullptr);
 161                                 check_error();
 162                         }
 163                         break;
 164                 case bmusb::PixelFormat_8BitBGRA:
 165                         glBindTexture(GL_TEXTURE_2D, ud->tex_rgba[field]);
 166                         check_error();
 167                         set_clamp_to_edge();
 168                         if (field == 0) {
 169                                 if (global_flags.can_disable_srgb_decoder) {  // See the comments in tweaked_inputs.h.
 170                                         glTexImage2D(GL_TEXTURE_2D, 0, GL_SRGB8_ALPHA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 171                                 } else {
 172                                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, width, height, 0, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, nullptr);
 173                                 }
 174                                 check_error();
 175                         }
 176                         break;
 177                 case bmusb::PixelFormat_8BitYCbCrPlanar:
 178                         glBindTexture(GL_TEXTURE_2D, ud->tex_y[field]);
 179                         check_error();
 180                         set_clamp_to_edge();
 181                         if (field == 0) {
 182                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 183                                 check_error();
 184                         }
 185
 186                         glBindTexture(GL_TEXTURE_2D, ud->tex_cb[field]);
 187                         check_error();
 188                         set_clamp_to_edge();
 189                         if (field == 0) {
 190                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width / 2, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 191                                 check_error();
 192                         }
 193
 194                         glBindTexture(GL_TEXTURE_2D, ud->tex_cr[field]);
 195                         check_error();
 196                         set_clamp_to_edge();
 197                         if (field == 0) {
 198                                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, width / 2, height, 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 199                                 check_error();
 200                         }
 201                         break;
 202                 default:
 203                         assert(false);
 204                 }
 205         }
 206
 207         freelist.push(frame);
 208 }
 209
 210 PBOFrameAllocator::~PBOFrameAllocator()
 211 {
 212         while (!freelist.empty()) {
 213                 Frame frame = freelist.front();
 214                 freelist.pop();
 215                 destroy_frame(&frame);
 216         }
 217 }
 218
 219 void PBOFrameAllocator::destroy_frame(Frame *frame)
 220 {
 221         Userdata *ud = (Userdata *)frame->userdata;
 222         delete[] ud->data_copy_malloc;
 223
 224         GLuint pbo = ud->pbo;
 225         glBindBuffer(buffer, pbo);
 226         check_error();
 227         glUnmapBuffer(buffer);
 228         check_error();
 229         glBindBuffer(buffer, 0);
 230         check_error();
 231         glDeleteBuffers(1, &pbo);
 232         check_error();
 233         switch (pixel_format) {
 234         case bmusb::PixelFormat_10BitYCbCr:
 235                 glDeleteTextures(2, ud->tex_v210);
 236                 check_error();
 237                 glDeleteTextures(2, ud->tex_444);
 238                 check_error();
 239                 break;
 240         case bmusb::PixelFormat_8BitYCbCr:
 241                 glDeleteTextures(2, ud->tex_y);
 242                 check_error();
 243                 glDeleteTextures(2, ud->tex_cbcr);
 244                 check_error();
 245                 break;
 246         case bmusb::PixelFormat_8BitBGRA:
 247                 glDeleteTextures(2, ud->tex_rgba);
 248                 check_error();
 249                 break;
 250         case bmusb::PixelFormat_8BitYCbCrPlanar:
 251                 glDeleteTextures(2, ud->tex_y);
 252                 check_error();
 253                 glDeleteTextures(2, ud->tex_cb);
 254                 check_error();
 255                 glDeleteTextures(2, ud->tex_cr);
 256                 check_error();
 257                 break;
 258         default:
 259                 assert(false);
 260         }
 261 }
 262 //static int sumsum = 0;
 263
 264 bmusb::FrameAllocator::Frame PBOFrameAllocator::alloc_frame()
 265 {
 266         Frame vf;
 267
 268         lock_guard<mutex> lock(freelist_mutex);  // Meh.
 269         if (freelist.empty()) {
 270                 printf("Frame overrun (no more spare PBO frames), dropping frame!\n");
 271         } else {
 272                 //fprintf(stderr, "freelist has %d allocated\n", ++sumsum);
 273                 vf = freelist.front();
 274                 freelist.pop();  // Meh.
 275         }
 276         vf.len = 0;
 277         vf.overflow = 0;
 278
 279         if (mjpeg_encoder != nullptr && mjpeg_encoder->using_vaapi() &&
 280             mjpeg_encoder->get_mjpeg_stream_for_card(card_index) != -1) {
 281                 Userdata *ud = (Userdata *)vf.userdata;
 282                 vf.data_copy = ud->data_copy_malloc;
 283                 ud->data_copy_current_src = Userdata::FROM_MALLOC;
 284         } else {
 285                 vf.data_copy = nullptr;
 286         }
 287
 288         return vf;
 289 }
 290
 291 bmusb::FrameAllocator::Frame PBOFrameAllocator::create_frame(size_t width, size_t height, size_t stride)
 292 {
 293         Frame vf;
 294
 295         {
 296                 lock_guard<mutex> lock(freelist_mutex);
 297                 if (freelist.empty()) {
 298                         printf("Frame overrun (no more spare PBO frames), dropping frame!\n");
 299                         vf.len = 0;
 300                         vf.overflow = 0;
 301                         return vf;
 302                 } else {
 303                         vf = freelist.front();
 304                         freelist.pop();
 305                 }
 306         }
 307         vf.len = 0;
 308         vf.overflow = 0;
 309
 310         Userdata *userdata = (Userdata *)vf.userdata;
 311
 312         if (mjpeg_encoder != nullptr && mjpeg_encoder->using_vaapi() &&
 313             mjpeg_encoder->get_mjpeg_stream_for_card(card_index) != -1) {
 314                 VADisplay va_dpy = mjpeg_encoder->va_dpy->va_dpy;
 315                 MJPEGEncoder::VAResources resources = mjpeg_encoder->get_va_resources(width, height);
 316                 MJPEGEncoder::ReleaseVAResources release(mjpeg_encoder, resources);
 317
 318                 VAImage image;
 319                 VAStatus va_status = vaDeriveImage(va_dpy, resources.surface, &image);
 320                 CHECK_VASTATUS(va_status, "vaDeriveImage");
 321
 322                 if (image.pitches[0] == stride) {
 323                         userdata->va_resources = move(resources);
 324                         userdata->va_resources_release = move(release);
 325                         userdata->va_image = move(image);
 326
 327                         va_status = vaMapBuffer(va_dpy, image.buf, (void **)&vf.data_copy);
 328                         CHECK_VASTATUS(va_status, "vaMapBuffer");
 329                         vf.data_copy += image.offsets[0];
 330                         userdata->data_copy_current_src = Userdata::FROM_VA_API;
 331                 } else {
 332                         printf("WARNING: Could not copy directly into VA-API MJPEG buffer for %zu x %zu, since producer and consumer disagreed on stride (%zu != %d).\n", width, height, stride, image.pitches[0]);
 333                         vf.data_copy = userdata->data_copy_malloc;
 334                         userdata->data_copy_current_src = Userdata::FROM_MALLOC;
 335
 336                         va_status = vaDestroyImage(va_dpy, image.image_id);
 337                         CHECK_VASTATUS(va_status, "vaDestroyImage");
 338                 }
 339         } else {
 340                 vf.data_copy = nullptr;
 341         }
 342
 343         return vf;
 344 }
 345
 346 void PBOFrameAllocator::release_frame(Frame frame)
 347 {
 348         if (frame.overflow > 0) {
 349                 printf("%d bytes overflow after last (PBO) frame\n", int(frame.overflow));
 350         }
 351
 352 #if 0
 353         // Poison the page. (Note that this might be bogus if you don't have an OpenGL context.)
 354         memset(frame.data, 0, frame.size);
 355         Userdata *userdata = (Userdata *)frame.userdata;
 356         for (unsigned field = 0; field < 2; ++field) {
 357                 glBindTexture(GL_TEXTURE_2D, userdata->tex_y[field]);
 358                 check_error();
 359                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 360                 check_error();
 361                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 362                 check_error();
 363                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 364                 check_error();
 365                 glTexImage2D(GL_TEXTURE_2D, 0, GL_R8, userdata->last_width[field], userdata->last_height[field], 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 366                 check_error();
 367
 368                 glBindTexture(GL_TEXTURE_2D, userdata->tex_cbcr[field]);
 369                 check_error();
 370                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
 371                 check_error();
 372                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
 373                 check_error();
 374                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
 375                 check_error();
 376                 glTexImage2D(GL_TEXTURE_2D, 0, GL_RG8, userdata->last_width[field] / 2, userdata->last_height[field], 0, GL_RED, GL_UNSIGNED_BYTE, nullptr);
 377                 check_error();
 378         }
 379 #endif
 380
 381         lock_guard<mutex> lock(freelist_mutex);
 382         freelist.push(frame);
 383         //--sumsum;
 384 }