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