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