src/mesa/state_tracker/st_cb_readpixels.c

   1 /**************************************************************************
   2  *
   3  * Copyright 2007 VMware, Inc.
   4  * All Rights Reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the
   8  * "Software"), to deal in the Software without restriction, including
   9  * without limitation the rights to use, copy, modify, merge, publish,
  10  * distribute, sub license, and/or sell copies of the Software, and to
  11  * permit persons to whom the Software is furnished to do so, subject to
  12  * the following conditions:
  13  *
  14  * The above copyright notice and this permission notice (including the
  15  * next paragraph) shall be included in all copies or substantial portions
  16  * of the Software.
  17  *
  18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
  19  * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  20  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
  21  * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
  22  * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
  23  * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
  24  * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  25  *
  26  **************************************************************************/
  27
  28 #include "main/bufferobj.h"
  29 #include "main/image.h"
  30 #include "main/pbo.h"
  31 #include "main/imports.h"
  32 #include "main/readpix.h"
  33 #include "main/enums.h"
  34 #include "main/framebuffer.h"
  35 #include "util/u_inlines.h"
  36 #include "util/u_format.h"
  37 #include "cso_cache/cso_context.h"
  38
  39 #include "st_cb_fbo.h"
  40 #include "st_atom.h"
  41 #include "st_context.h"
  42 #include "st_cb_bitmap.h"
  43 #include "st_cb_readpixels.h"
  44 #include "st_debug.h"
  45 #include "state_tracker/st_cb_texture.h"
  46 #include "state_tracker/st_format.h"
  47 #include "state_tracker/st_pbo.h"
  48 #include "state_tracker/st_texture.h"
  49
  50 /* The readpixels cache caches a blitted staging texture so that back-to-back
  51  * calls to glReadPixels with user pointers require less CPU-GPU synchronization.
  52  *
  53  * Assumptions:
  54  *
  55  * (1) Blits have high synchronization overheads, and it is beneficial to
  56  *     use a single blit of the entire framebuffer instead of many smaller
  57  *     blits (because the smaller blits cannot be batched, and we have to wait
  58  *     for the GPU after each one).
  59  *
  60  * (2) transfer_map implicitly involves a blit as well (for de-tiling, copy
  61  *     from VRAM, etc.), so that it is beneficial to replace the
  62  *     _mesa_readpixels path as well when possible.
  63  *
  64  * Change this #define to true to fill and use the cache whenever possible
  65  * (this is inefficient and only meant for testing / debugging).
  66  */
  67 #define ALWAYS_READPIXELS_CACHE false
  68
  69 static boolean
  70 needs_integer_signed_unsigned_conversion(const struct gl_context *ctx,
  71                                          GLenum format, GLenum type)
  72 {
  73    struct gl_renderbuffer *rb =
  74       _mesa_get_read_renderbuffer_for_format(ctx, format);
  75
  76    assert(rb);
  77
  78    GLenum srcType = _mesa_get_format_datatype(rb->Format);
  79
  80     if ((srcType == GL_INT &&
  81         (type == GL_UNSIGNED_INT ||
  82          type == GL_UNSIGNED_SHORT ||
  83          type == GL_UNSIGNED_BYTE)) ||
  84        (srcType == GL_UNSIGNED_INT &&
  85         (type == GL_INT ||
  86          type == GL_SHORT ||
  87          type == GL_BYTE))) {
  88       return TRUE;
  89    }
  90
  91    return FALSE;
  92 }
  93
  94 static bool
  95 try_pbo_readpixels(struct st_context *st, struct st_renderbuffer *strb,
  96                    bool invert_y,
  97                    GLint x, GLint y, GLsizei width, GLsizei height,
  98                    enum pipe_format src_format, enum pipe_format dst_format,
  99                    const struct gl_pixelstore_attrib *pack, void *pixels)
 100 {
 101    struct pipe_context *pipe = st->pipe;
 102    struct pipe_screen *screen = pipe->screen;
 103    struct cso_context *cso = st->cso_context;
 104    struct pipe_surface *surface = strb->surface;
 105    struct pipe_resource *texture = strb->texture;
 106    const struct util_format_description *desc;
 107    struct st_pbo_addresses addr;
 108    struct pipe_framebuffer_state fb;
 109    enum pipe_texture_target view_target;
 110    bool success = false;
 111
 112    if (texture->nr_samples > 1)
 113       return false;
 114
 115    if (!screen->is_format_supported(screen, dst_format, PIPE_BUFFER, 0,
 116                                     PIPE_BIND_SHADER_IMAGE))
 117       return false;
 118
 119    desc = util_format_description(dst_format);
 120
 121    /* Compute PBO addresses */
 122    addr.bytes_per_pixel = desc->block.bits / 8;
 123    addr.xoffset = x;
 124    addr.yoffset = y;
 125    addr.width = width;
 126    addr.height = height;
 127    addr.depth = 1;
 128    if (!st_pbo_addresses_pixelstore(st, GL_TEXTURE_2D, false, pack, pixels, &addr))
 129       return false;
 130
 131    cso_save_state(cso, (CSO_BIT_FRAGMENT_SAMPLER_VIEWS |
 132                         CSO_BIT_FRAGMENT_SAMPLERS |
 133                         CSO_BIT_FRAGMENT_IMAGE0 |
 134                         CSO_BIT_VERTEX_ELEMENTS |
 135                         CSO_BIT_AUX_VERTEX_BUFFER_SLOT |
 136                         CSO_BIT_FRAMEBUFFER |
 137                         CSO_BIT_VIEWPORT |
 138                         CSO_BIT_RASTERIZER |
 139                         CSO_BIT_DEPTH_STENCIL_ALPHA |
 140                         CSO_BIT_STREAM_OUTPUTS |
 141                         CSO_BIT_PAUSE_QUERIES |
 142                         CSO_BITS_ALL_SHADERS));
 143    cso_save_constant_buffer_slot0(cso, PIPE_SHADER_FRAGMENT);
 144
 145    /* Set up the sampler_view */
 146    {
 147       struct pipe_sampler_view templ;
 148       struct pipe_sampler_view *sampler_view;
 149       struct pipe_sampler_state sampler = {0};
 150       const struct pipe_sampler_state *samplers[1] = {&sampler};
 151
 152       u_sampler_view_default_template(&templ, texture, src_format);
 153
 154       switch (texture->target) {
 155       case PIPE_TEXTURE_CUBE:
 156       case PIPE_TEXTURE_CUBE_ARRAY:
 157          view_target = PIPE_TEXTURE_2D_ARRAY;
 158          break;
 159       default:
 160          view_target = texture->target;
 161          break;
 162       }
 163
 164       templ.target = view_target;
 165       templ.u.tex.first_level = surface->u.tex.level;
 166       templ.u.tex.last_level = templ.u.tex.first_level;
 167
 168       if (view_target != PIPE_TEXTURE_3D) {
 169          templ.u.tex.first_layer = surface->u.tex.first_layer;
 170          templ.u.tex.last_layer = templ.u.tex.last_layer;
 171       } else {
 172          addr.constants.layer_offset = surface->u.tex.first_layer;
 173       }
 174
 175       sampler_view = pipe->create_sampler_view(pipe, texture, &templ);
 176       if (sampler_view == NULL)
 177          goto fail;
 178
 179       cso_set_sampler_views(cso, PIPE_SHADER_FRAGMENT, 1, &sampler_view);
 180
 181       pipe_sampler_view_reference(&sampler_view, NULL);
 182
 183       cso_set_samplers(cso, PIPE_SHADER_FRAGMENT, 1, samplers);
 184    }
 185
 186    /* Set up destination image */
 187    {
 188       struct pipe_image_view image;
 189
 190       memset(&image, 0, sizeof(image));
 191       image.resource = addr.buffer;
 192       image.format = dst_format;
 193       image.access = PIPE_IMAGE_ACCESS_WRITE;
 194       image.u.buf.offset = addr.first_element * addr.bytes_per_pixel;
 195       image.u.buf.size = (addr.last_element - addr.first_element + 1) *
 196                          addr.bytes_per_pixel;
 197
 198       cso_set_shader_images(cso, PIPE_SHADER_FRAGMENT, 0, 1, &image);
 199    }
 200
 201    /* Set up no-attachment framebuffer */
 202    memset(&fb, 0, sizeof(fb));
 203    fb.width = surface->width;
 204    fb.height = surface->height;
 205    fb.samples = 1;
 206    fb.layers = 1;
 207    cso_set_framebuffer(cso, &fb);
 208
 209    cso_set_viewport_dims(cso, fb.width, fb.height, invert_y);
 210
 211    if (invert_y)
 212       st_pbo_addresses_invert_y(&addr, fb.height);
 213
 214    {
 215       struct pipe_depth_stencil_alpha_state dsa;
 216       memset(&dsa, 0, sizeof(dsa));
 217       cso_set_depth_stencil_alpha(cso, &dsa);
 218    }
 219
 220    /* Set up the fragment shader */
 221    {
 222       void *fs = st_pbo_get_download_fs(st, view_target, src_format, dst_format);
 223       if (!fs)
 224          goto fail;
 225
 226       cso_set_fragment_shader_handle(cso, fs);
 227    }
 228
 229    success = st_pbo_draw(st, &addr, fb.width, fb.height);
 230
 231    /* Buffer written via shader images needs explicit synchronization. */
 232    pipe->memory_barrier(pipe, PIPE_BARRIER_ALL);
 233
 234 fail:
 235    cso_restore_state(cso);
 236    cso_restore_constant_buffer_slot0(cso, PIPE_SHADER_FRAGMENT);
 237
 238    return success;
 239 }
 240
 241 /* Invalidate the readpixels cache to ensure we don't read stale data.
 242  */
 243 void st_invalidate_readpix_cache(struct st_context *st)
 244 {
 245    pipe_resource_reference(&st->readpix_cache.src, NULL);
 246    pipe_resource_reference(&st->readpix_cache.cache, NULL);
 247 }
 248
 249 /**
 250  * Create a staging texture and blit the requested region to it.
 251  */
 252 static struct pipe_resource *
 253 blit_to_staging(struct st_context *st, struct st_renderbuffer *strb,
 254                    bool invert_y,
 255                    GLint x, GLint y, GLsizei width, GLsizei height,
 256                    GLenum format,
 257                    enum pipe_format src_format, enum pipe_format dst_format)
 258 {
 259    struct pipe_context *pipe = st->pipe;
 260    struct pipe_screen *screen = pipe->screen;
 261    struct pipe_resource dst_templ;
 262    struct pipe_resource *dst;
 263    struct pipe_blit_info blit;
 264
 265    /* We are creating a texture of the size of the region being read back.
 266     * Need to check for NPOT texture support. */
 267    if (!screen->get_param(screen, PIPE_CAP_NPOT_TEXTURES) &&
 268        (!util_is_power_of_two(width) ||
 269         !util_is_power_of_two(height)))
 270       return NULL;
 271
 272    /* create the destination texture */
 273    memset(&dst_templ, 0, sizeof(dst_templ));
 274    dst_templ.target = PIPE_TEXTURE_2D;
 275    dst_templ.format = dst_format;
 276    if (util_format_is_depth_or_stencil(dst_format))
 277       dst_templ.bind |= PIPE_BIND_DEPTH_STENCIL;
 278    else
 279       dst_templ.bind |= PIPE_BIND_RENDER_TARGET;
 280    dst_templ.usage = PIPE_USAGE_STAGING;
 281
 282    st_gl_texture_dims_to_pipe_dims(GL_TEXTURE_2D, width, height, 1,
 283                                    &dst_templ.width0, &dst_templ.height0,
 284                                    &dst_templ.depth0, &dst_templ.array_size);
 285
 286    dst = screen->resource_create(screen, &dst_templ);
 287    if (!dst)
 288       return NULL;
 289
 290    memset(&blit, 0, sizeof(blit));
 291    blit.src.resource = strb->texture;
 292    blit.src.level = strb->surface->u.tex.level;
 293    blit.src.format = src_format;
 294    blit.dst.resource = dst;
 295    blit.dst.level = 0;
 296    blit.dst.format = dst->format;
 297    blit.src.box.x = x;
 298    blit.dst.box.x = 0;
 299    blit.src.box.y = y;
 300    blit.dst.box.y = 0;
 301    blit.src.box.z = strb->surface->u.tex.first_layer;
 302    blit.dst.box.z = 0;
 303    blit.src.box.width = blit.dst.box.width = width;
 304    blit.src.box.height = blit.dst.box.height = height;
 305    blit.src.box.depth = blit.dst.box.depth = 1;
 306    blit.mask = st_get_blit_mask(strb->Base._BaseFormat, format);
 307    blit.filter = PIPE_TEX_FILTER_NEAREST;
 308    blit.scissor_enable = FALSE;
 309
 310    if (invert_y) {
 311       blit.src.box.y = strb->Base.Height - blit.src.box.y;
 312       blit.src.box.height = -blit.src.box.height;
 313    }
 314
 315    /* blit */
 316    st->pipe->blit(st->pipe, &blit);
 317
 318    return dst;
 319 }
 320
 321 static struct pipe_resource *
 322 try_cached_readpixels(struct st_context *st, struct st_renderbuffer *strb,
 323                       bool invert_y,
 324                       GLsizei width, GLsizei height,
 325                       GLenum format,
 326                       enum pipe_format src_format, enum pipe_format dst_format)
 327 {
 328    struct pipe_resource *src = strb->texture;
 329    struct pipe_resource *dst = NULL;
 330
 331    if (ST_DEBUG & DEBUG_NOREADPIXCACHE)
 332       return NULL;
 333
 334    /* Reset cache after invalidation or switch of parameters. */
 335    if (st->readpix_cache.src != src ||
 336        st->readpix_cache.dst_format != dst_format ||
 337        st->readpix_cache.level != strb->surface->u.tex.level ||
 338        st->readpix_cache.layer != strb->surface->u.tex.first_layer) {
 339       pipe_resource_reference(&st->readpix_cache.src, src);
 340       pipe_resource_reference(&st->readpix_cache.cache, NULL);
 341       st->readpix_cache.dst_format = dst_format;
 342       st->readpix_cache.level = strb->surface->u.tex.level;
 343       st->readpix_cache.layer = strb->surface->u.tex.first_layer;
 344       st->readpix_cache.hits = 0;
 345    }
 346
 347    /* Decide whether to trigger the cache. */
 348    if (!st->readpix_cache.cache) {
 349       if (!strb->use_readpix_cache && !ALWAYS_READPIXELS_CACHE) {
 350          /* Heuristic: If previous successive calls read at least a fraction
 351           * of the surface _and_ we read again, trigger the cache.
 352           */
 353          unsigned threshold = MAX2(1, strb->Base.Width * strb->Base.Height / 8);
 354
 355          if (st->readpix_cache.hits < threshold) {
 356             st->readpix_cache.hits += width * height;
 357             return NULL;
 358          }
 359
 360          strb->use_readpix_cache = true;
 361       }
 362
 363       /* Fill the cache */
 364       st->readpix_cache.cache = blit_to_staging(st, strb, invert_y,
 365                                                 0, 0,
 366                                                 strb->Base.Width,
 367                                                 strb->Base.Height, format,
 368                                                 src_format, dst_format);
 369    }
 370
 371    /* Return an owning reference to stay consistent with the non-cached path */
 372    pipe_resource_reference(&dst, st->readpix_cache.cache);
 373
 374    return dst;
 375 }
 376
 377 /**
 378  * This uses a blit to copy the read buffer to a texture format which matches
 379  * the format and type combo and then a fast read-back is done using memcpy.
 380  * We can do arbitrary X/Y/Z/W/0/1 swizzling here as long as there is
 381  * a format which matches the swizzling.
 382  *
 383  * If such a format isn't available, we fall back to _mesa_readpixels.
 384  *
 385  * NOTE: Some drivers use a blit to convert between tiled and linear
 386  *       texture layouts during texture uploads/downloads, so the blit
 387  *       we do here should be free in such cases.
 388  */
 389 static void
 390 st_ReadPixels(struct gl_context *ctx, GLint x, GLint y,
 391               GLsizei width, GLsizei height,
 392               GLenum format, GLenum type,
 393               const struct gl_pixelstore_attrib *pack,
 394               void *pixels)
 395 {
 396    struct st_context *st = st_context(ctx);
 397    struct gl_renderbuffer *rb =
 398          _mesa_get_read_renderbuffer_for_format(ctx, format);
 399    struct st_renderbuffer *strb = st_renderbuffer(rb);
 400    struct pipe_context *pipe = st->pipe;
 401    struct pipe_screen *screen = pipe->screen;
 402    struct pipe_resource *src;
 403    struct pipe_resource *dst = NULL;
 404    enum pipe_format dst_format, src_format;
 405    unsigned bind;
 406    struct pipe_transfer *tex_xfer;
 407    ubyte *map = NULL;
 408    int dst_x, dst_y;
 409
 410    /* Validate state (to be sure we have up-to-date framebuffer surfaces)
 411     * and flush the bitmap cache prior to reading. */
 412    st_validate_state(st, ST_PIPELINE_RENDER);
 413    st_flush_bitmap_cache(st);
 414
 415    if (!st->prefer_blit_based_texture_transfer) {
 416       goto fallback;
 417    }
 418
 419    /* This must be done after state validation. */
 420    src = strb->texture;
 421
 422    /* XXX Fallback for depth-stencil formats due to an incomplete
 423     * stencil blit implementation in some drivers. */
 424    if (format == GL_DEPTH_STENCIL) {
 425       goto fallback;
 426    }
 427
 428    /* If the base internal format and the texture format don't match, we have
 429     * to use the slow path. */
 430    if (rb->_BaseFormat !=
 431        _mesa_get_format_base_format(rb->Format)) {
 432       goto fallback;
 433    }
 434
 435    if (_mesa_readpixels_needs_slow_path(ctx, format, type, GL_TRUE)) {
 436       goto fallback;
 437    }
 438
 439    /* Convert the source format to what is expected by ReadPixels
 440     * and see if it's supported. */
 441    src_format = util_format_linear(src->format);
 442    src_format = util_format_luminance_to_red(src_format);
 443    src_format = util_format_intensity_to_red(src_format);
 444
 445    if (!src_format ||
 446        !screen->is_format_supported(screen, src_format, src->target,
 447                                     src->nr_samples,
 448                                     PIPE_BIND_SAMPLER_VIEW)) {
 449       goto fallback;
 450    }
 451
 452    if (format == GL_DEPTH_COMPONENT || format == GL_DEPTH_STENCIL)
 453       bind = PIPE_BIND_DEPTH_STENCIL;
 454    else
 455       bind = PIPE_BIND_RENDER_TARGET;
 456
 457    /* Choose the destination format by finding the best match
 458     * for the format+type combo. */
 459    dst_format = st_choose_matching_format(st, bind, format, type,
 460                                           pack->SwapBytes);
 461    if (dst_format == PIPE_FORMAT_NONE) {
 462       goto fallback;
 463    }
 464
 465    if (st->pbo.download_enabled && _mesa_is_bufferobj(pack->BufferObj)) {
 466       if (try_pbo_readpixels(st, strb,
 467                              st_fb_orientation(ctx->ReadBuffer) == Y_0_TOP,
 468                              x, y, width, height,
 469                              src_format, dst_format,
 470                              pack, pixels))
 471          return;
 472    }
 473
 474    if (needs_integer_signed_unsigned_conversion(ctx, format, type)) {
 475       goto fallback;
 476    }
 477
 478    /* Cache a staging texture for back-to-back ReadPixels, to avoid CPU-GPU
 479     * synchronization overhead.
 480     */
 481    dst = try_cached_readpixels(st, strb,
 482                                st_fb_orientation(ctx->ReadBuffer) == Y_0_TOP,
 483                                width, height, format, src_format, dst_format);
 484    if (dst) {
 485       dst_x = x;
 486       dst_y = y;
 487    } else {
 488       /* See if the texture format already matches the format and type,
 489        * in which case the memcpy-based fast path will likely be used and
 490        * we don't have to blit. */
 491       if (_mesa_format_matches_format_and_type(rb->Format, format,
 492                                                type, pack->SwapBytes, NULL)) {
 493          goto fallback;
 494       }
 495
 496       dst = blit_to_staging(st, strb,
 497                             st_fb_orientation(ctx->ReadBuffer) == Y_0_TOP,
 498                             x, y, width, height, format,
 499                             src_format, dst_format);
 500       if (!dst)
 501          goto fallback;
 502
 503       dst_x = 0;
 504       dst_y = 0;
 505    }
 506
 507    /* map resources */
 508    pixels = _mesa_map_pbo_dest(ctx, pack, pixels);
 509
 510    map = pipe_transfer_map_3d(pipe, dst, 0, PIPE_TRANSFER_READ,
 511                               dst_x, dst_y, 0, width, height, 1, &tex_xfer);
 512    if (!map) {
 513       _mesa_unmap_pbo_dest(ctx, pack);
 514       pipe_resource_reference(&dst, NULL);
 515       goto fallback;
 516    }
 517
 518    /* memcpy data into a user buffer */
 519    {
 520       const uint bytesPerRow = width * util_format_get_blocksize(dst_format);
 521       const int destStride = _mesa_image_row_stride(pack, width, format, type);
 522       char *dest = _mesa_image_address2d(pack, pixels,
 523                                          width, height, format,
 524                                          type, 0, 0);
 525
 526       if (tex_xfer->stride == bytesPerRow && destStride == bytesPerRow) {
 527          memcpy(dest, map, bytesPerRow * height);
 528       } else {
 529          GLuint row;
 530
 531          for (row = 0; row < (unsigned) height; row++) {
 532             memcpy(dest, map, bytesPerRow);
 533             map += tex_xfer->stride;
 534             dest += destStride;
 535          }
 536       }
 537    }
 538
 539    pipe_transfer_unmap(pipe, tex_xfer);
 540    _mesa_unmap_pbo_dest(ctx, pack);
 541    pipe_resource_reference(&dst, NULL);
 542    return;
 543
 544 fallback:
 545    _mesa_readpixels(ctx, x, y, width, height, format, type, pack, pixels);
 546 }
 547
 548 void st_init_readpixels_functions(struct dd_function_table *functions)
 549 {
 550    functions->ReadPixels = st_ReadPixels;
 551 }