1/* 2 * Copyright 2010 Jerome Glisse <glisse@freedesktop.org> 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * on the rights to use, copy, modify, merge, publish, distribute, sub 8 * license, and/or sell copies of the Software, and to permit persons to whom 9 * the Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM, 19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 21 * USE OR OTHER DEALINGS IN THE SOFTWARE. 22 */ 23#include "r600_pipe.h" 24#include "compute_memory_pool.h" 25#include "evergreen_compute.h" 26#include "util/u_surface.h" 27#include "util/format/u_format.h" 28#include "evergreend.h" 29 30enum r600_blitter_op /* bitmask */ 31{ 32 R600_SAVE_FRAGMENT_STATE = 1, 33 R600_SAVE_TEXTURES = 2, 34 R600_SAVE_FRAMEBUFFER = 4, 35 R600_DISABLE_RENDER_COND = 8, 36 37 R600_CLEAR = R600_SAVE_FRAGMENT_STATE, 38 39 R600_CLEAR_SURFACE = R600_SAVE_FRAGMENT_STATE | R600_SAVE_FRAMEBUFFER, 40 41 R600_COPY_BUFFER = R600_DISABLE_RENDER_COND, 42 43 R600_COPY_TEXTURE = R600_SAVE_FRAGMENT_STATE | R600_SAVE_FRAMEBUFFER | R600_SAVE_TEXTURES | 44 R600_DISABLE_RENDER_COND, 45 46 R600_BLIT = R600_SAVE_FRAGMENT_STATE | R600_SAVE_FRAMEBUFFER | R600_SAVE_TEXTURES, 47 48 R600_DECOMPRESS = R600_SAVE_FRAGMENT_STATE | R600_SAVE_FRAMEBUFFER | R600_DISABLE_RENDER_COND, 49 50 R600_COLOR_RESOLVE = R600_SAVE_FRAGMENT_STATE | R600_SAVE_FRAMEBUFFER 51}; 52 53static void r600_blitter_begin(struct pipe_context *ctx, enum r600_blitter_op op) 54{ 55 struct r600_context *rctx = (struct r600_context *)ctx; 56 57 if (rctx->cmd_buf_is_compute) { 58 rctx->b.gfx.flush(rctx, PIPE_FLUSH_ASYNC, NULL); 59 rctx->cmd_buf_is_compute = false; 60 } 61 62 util_blitter_save_vertex_buffer_slot(rctx->blitter, rctx->vertex_buffer_state.vb); 63 util_blitter_save_vertex_elements(rctx->blitter, rctx->vertex_fetch_shader.cso); 64 util_blitter_save_vertex_shader(rctx->blitter, rctx->vs_shader); 65 util_blitter_save_geometry_shader(rctx->blitter, rctx->gs_shader); 66 util_blitter_save_tessctrl_shader(rctx->blitter, rctx->tcs_shader); 67 util_blitter_save_tesseval_shader(rctx->blitter, rctx->tes_shader); 68 util_blitter_save_so_targets(rctx->blitter, rctx->b.streamout.num_targets, 69 (struct pipe_stream_output_target**)rctx->b.streamout.targets); 70 util_blitter_save_rasterizer(rctx->blitter, rctx->rasterizer_state.cso); 71 72 if (op & R600_SAVE_FRAGMENT_STATE) { 73 util_blitter_save_viewport(rctx->blitter, &rctx->b.viewports.states[0]); 74 util_blitter_save_scissor(rctx->blitter, &rctx->b.scissors.states[0]); 75 util_blitter_save_fragment_shader(rctx->blitter, rctx->ps_shader); 76 util_blitter_save_blend(rctx->blitter, rctx->blend_state.cso); 77 util_blitter_save_depth_stencil_alpha(rctx->blitter, rctx->dsa_state.cso); 78 util_blitter_save_stencil_ref(rctx->blitter, &rctx->stencil_ref.pipe_state); 79 util_blitter_save_sample_mask(rctx->blitter, rctx->sample_mask.sample_mask, rctx->ps_iter_samples); 80 util_blitter_save_fragment_constant_buffer_slot(rctx->blitter, 81 &rctx->constbuf_state[PIPE_SHADER_FRAGMENT].cb[0]); 82 } 83 84 if (op & R600_SAVE_FRAMEBUFFER) 85 util_blitter_save_framebuffer(rctx->blitter, &rctx->framebuffer.state); 86 87 if (op & R600_SAVE_TEXTURES) { 88 util_blitter_save_fragment_sampler_states( 89 rctx->blitter, util_last_bit(rctx->samplers[PIPE_SHADER_FRAGMENT].states.enabled_mask), 90 (void**)rctx->samplers[PIPE_SHADER_FRAGMENT].states.states); 91 92 util_blitter_save_fragment_sampler_views( 93 rctx->blitter, util_last_bit(rctx->samplers[PIPE_SHADER_FRAGMENT].views.enabled_mask), 94 (struct pipe_sampler_view**)rctx->samplers[PIPE_SHADER_FRAGMENT].views.views); 95 } 96 97 if (op & R600_DISABLE_RENDER_COND) 98 rctx->b.render_cond_force_off = true; 99} 100 101static void r600_blitter_end(struct pipe_context *ctx) 102{ 103 struct r600_context *rctx = (struct r600_context *)ctx; 104 105 rctx->b.render_cond_force_off = false; 106} 107 108static unsigned u_max_sample(struct pipe_resource *r) 109{ 110 return r->nr_samples ? r->nr_samples - 1 : 0; 111} 112 113static void r600_blit_decompress_depth(struct pipe_context *ctx, 114 struct r600_texture *texture, 115 struct r600_texture *staging, 116 unsigned first_level, unsigned last_level, 117 unsigned first_layer, unsigned last_layer, 118 unsigned first_sample, unsigned last_sample) 119{ 120 struct r600_context *rctx = (struct r600_context *)ctx; 121 unsigned layer, level, sample, checked_last_layer, max_layer, max_sample; 122 struct r600_texture *flushed_depth_texture = staging ? 123 staging : texture->flushed_depth_texture; 124 const struct util_format_description *desc = 125 util_format_description(texture->resource.b.b.format); 126 float depth; 127 128 if (!staging && !texture->dirty_level_mask) 129 return; 130 131 max_sample = u_max_sample(&texture->resource.b.b); 132 133 /* XXX Decompressing MSAA depth textures is broken on R6xx. 134 * There is also a hardlock if CMASK and FMASK are not present. 135 * Just skip this until we find out how to fix it. */ 136 if (rctx->b.gfx_level == R600 && max_sample > 0) { 137 texture->dirty_level_mask = 0; 138 return; 139 } 140 141 if (rctx->b.family == CHIP_RV610 || rctx->b.family == CHIP_RV630 || 142 rctx->b.family == CHIP_RV620 || rctx->b.family == CHIP_RV635) 143 depth = 0.0f; 144 else 145 depth = 1.0f; 146 147 /* Enable decompression in DB_RENDER_CONTROL */ 148 rctx->db_misc_state.flush_depthstencil_through_cb = true; 149 rctx->db_misc_state.copy_depth = util_format_has_depth(desc); 150 rctx->db_misc_state.copy_stencil = util_format_has_stencil(desc); 151 rctx->db_misc_state.copy_sample = first_sample; 152 r600_mark_atom_dirty(rctx, &rctx->db_misc_state.atom); 153 154 for (level = first_level; level <= last_level; level++) { 155 if (!staging && !(texture->dirty_level_mask & (1 << level))) 156 continue; 157 158 /* The smaller the mipmap level, the less layers there are 159 * as far as 3D textures are concerned. */ 160 max_layer = util_max_layer(&texture->resource.b.b, level); 161 checked_last_layer = last_layer < max_layer ? last_layer : max_layer; 162 163 for (layer = first_layer; layer <= checked_last_layer; layer++) { 164 for (sample = first_sample; sample <= last_sample; sample++) { 165 struct pipe_surface *zsurf, *cbsurf, surf_tmpl; 166 167 if (sample != rctx->db_misc_state.copy_sample) { 168 rctx->db_misc_state.copy_sample = sample; 169 r600_mark_atom_dirty(rctx, &rctx->db_misc_state.atom); 170 } 171 172 surf_tmpl.format = texture->resource.b.b.format; 173 surf_tmpl.u.tex.level = level; 174 surf_tmpl.u.tex.first_layer = layer; 175 surf_tmpl.u.tex.last_layer = layer; 176 177 zsurf = ctx->create_surface(ctx, &texture->resource.b.b, &surf_tmpl); 178 179 surf_tmpl.format = flushed_depth_texture->resource.b.b.format; 180 cbsurf = ctx->create_surface(ctx, 181 &flushed_depth_texture->resource.b.b, &surf_tmpl); 182 183 r600_blitter_begin(ctx, R600_DECOMPRESS); 184 util_blitter_custom_depth_stencil(rctx->blitter, zsurf, cbsurf, 1 << sample, 185 rctx->custom_dsa_flush, depth); 186 r600_blitter_end(ctx); 187 188 pipe_surface_reference(&zsurf, NULL); 189 pipe_surface_reference(&cbsurf, NULL); 190 } 191 } 192 193 /* The texture will always be dirty if some layers or samples aren't flushed. 194 * I don't think this case occurs often though. */ 195 if (!staging && 196 first_layer == 0 && last_layer == max_layer && 197 first_sample == 0 && last_sample == max_sample) { 198 texture->dirty_level_mask &= ~(1 << level); 199 } 200 } 201 202 /* reenable compression in DB_RENDER_CONTROL */ 203 rctx->db_misc_state.flush_depthstencil_through_cb = false; 204 r600_mark_atom_dirty(rctx, &rctx->db_misc_state.atom); 205} 206 207static void r600_blit_decompress_depth_in_place(struct r600_context *rctx, 208 struct r600_texture *texture, 209 bool is_stencil_sampler, 210 unsigned first_level, unsigned last_level, 211 unsigned first_layer, unsigned last_layer) 212{ 213 struct pipe_surface *zsurf, surf_tmpl = {{0}}; 214 unsigned layer, max_layer, checked_last_layer, level; 215 unsigned *dirty_level_mask; 216 217 /* Enable decompression in DB_RENDER_CONTROL */ 218 if (is_stencil_sampler) { 219 rctx->db_misc_state.flush_stencil_inplace = true; 220 dirty_level_mask = &texture->stencil_dirty_level_mask; 221 } else { 222 rctx->db_misc_state.flush_depth_inplace = true; 223 dirty_level_mask = &texture->dirty_level_mask; 224 } 225 r600_mark_atom_dirty(rctx, &rctx->db_misc_state.atom); 226 227 surf_tmpl.format = texture->resource.b.b.format; 228 229 for (level = first_level; level <= last_level; level++) { 230 if (!(*dirty_level_mask & (1 << level))) 231 continue; 232 233 surf_tmpl.u.tex.level = level; 234 235 /* The smaller the mipmap level, the less layers there are 236 * as far as 3D textures are concerned. */ 237 max_layer = util_max_layer(&texture->resource.b.b, level); 238 checked_last_layer = last_layer < max_layer ? last_layer : max_layer; 239 240 for (layer = first_layer; layer <= checked_last_layer; layer++) { 241 surf_tmpl.u.tex.first_layer = layer; 242 surf_tmpl.u.tex.last_layer = layer; 243 244 zsurf = rctx->b.b.create_surface(&rctx->b.b, &texture->resource.b.b, &surf_tmpl); 245 246 r600_blitter_begin(&rctx->b.b, R600_DECOMPRESS); 247 util_blitter_custom_depth_stencil(rctx->blitter, zsurf, NULL, ~0, 248 rctx->custom_dsa_flush, 1.0f); 249 r600_blitter_end(&rctx->b.b); 250 251 pipe_surface_reference(&zsurf, NULL); 252 } 253 254 /* The texture will always be dirty if some layers or samples aren't flushed. 255 * I don't think this case occurs often though. */ 256 if (first_layer == 0 && last_layer == max_layer) { 257 *dirty_level_mask &= ~(1 << level); 258 } 259 } 260 261 /* Disable decompression in DB_RENDER_CONTROL */ 262 rctx->db_misc_state.flush_depth_inplace = false; 263 rctx->db_misc_state.flush_stencil_inplace = false; 264 r600_mark_atom_dirty(rctx, &rctx->db_misc_state.atom); 265} 266 267void r600_decompress_depth_textures(struct r600_context *rctx, 268 struct r600_samplerview_state *textures) 269{ 270 unsigned i; 271 unsigned depth_texture_mask = textures->compressed_depthtex_mask; 272 273 while (depth_texture_mask) { 274 struct pipe_sampler_view *view; 275 struct r600_pipe_sampler_view *rview; 276 struct r600_texture *tex; 277 278 i = u_bit_scan(&depth_texture_mask); 279 280 view = &textures->views[i]->base; 281 assert(view); 282 rview = (struct r600_pipe_sampler_view*)view; 283 284 tex = (struct r600_texture *)view->texture; 285 assert(tex->db_compatible); 286 287 if (r600_can_sample_zs(tex, rview->is_stencil_sampler)) { 288 r600_blit_decompress_depth_in_place(rctx, tex, 289 rview->is_stencil_sampler, 290 view->u.tex.first_level, view->u.tex.last_level, 291 0, util_max_layer(&tex->resource.b.b, view->u.tex.first_level)); 292 } else { 293 r600_blit_decompress_depth(&rctx->b.b, tex, NULL, 294 view->u.tex.first_level, view->u.tex.last_level, 295 0, util_max_layer(&tex->resource.b.b, view->u.tex.first_level), 296 0, u_max_sample(&tex->resource.b.b)); 297 } 298 } 299} 300 301void r600_decompress_depth_images(struct r600_context *rctx, 302 struct r600_image_state *images) 303{ 304 unsigned i; 305 unsigned depth_texture_mask = images->compressed_depthtex_mask; 306 307 while (depth_texture_mask) { 308 struct r600_image_view *view; 309 struct r600_texture *tex; 310 311 i = u_bit_scan(&depth_texture_mask); 312 313 view = &images->views[i]; 314 assert(view); 315 316 tex = (struct r600_texture *)view->base.resource; 317 assert(tex->db_compatible); 318 319 if (r600_can_sample_zs(tex, false)) { 320 r600_blit_decompress_depth_in_place(rctx, tex, 321 false, 322 view->base.u.tex.level, 323 view->base.u.tex.level, 324 0, util_max_layer(&tex->resource.b.b, view->base.u.tex.level)); 325 } else { 326 r600_blit_decompress_depth(&rctx->b.b, tex, NULL, 327 view->base.u.tex.level, 328 view->base.u.tex.level, 329 0, util_max_layer(&tex->resource.b.b, view->base.u.tex.level), 330 0, u_max_sample(&tex->resource.b.b)); 331 } 332 } 333} 334 335static void r600_blit_decompress_color(struct pipe_context *ctx, 336 struct r600_texture *rtex, 337 unsigned first_level, unsigned last_level, 338 unsigned first_layer, unsigned last_layer) 339{ 340 struct r600_context *rctx = (struct r600_context *)ctx; 341 unsigned layer, level, checked_last_layer, max_layer; 342 343 if (!rtex->dirty_level_mask) 344 return; 345 346 for (level = first_level; level <= last_level; level++) { 347 if (!(rtex->dirty_level_mask & (1 << level))) 348 continue; 349 350 /* The smaller the mipmap level, the less layers there are 351 * as far as 3D textures are concerned. */ 352 max_layer = util_max_layer(&rtex->resource.b.b, level); 353 checked_last_layer = last_layer < max_layer ? last_layer : max_layer; 354 355 for (layer = first_layer; layer <= checked_last_layer; layer++) { 356 struct pipe_surface *cbsurf, surf_tmpl; 357 358 surf_tmpl.format = rtex->resource.b.b.format; 359 surf_tmpl.u.tex.level = level; 360 surf_tmpl.u.tex.first_layer = layer; 361 surf_tmpl.u.tex.last_layer = layer; 362 cbsurf = ctx->create_surface(ctx, &rtex->resource.b.b, &surf_tmpl); 363 364 r600_blitter_begin(ctx, R600_DECOMPRESS); 365 util_blitter_custom_color(rctx->blitter, cbsurf, 366 rtex->fmask.size ? rctx->custom_blend_decompress : rctx->custom_blend_fastclear); 367 r600_blitter_end(ctx); 368 369 pipe_surface_reference(&cbsurf, NULL); 370 } 371 372 /* The texture will always be dirty if some layers aren't flushed. 373 * I don't think this case occurs often though. */ 374 if (first_layer == 0 && last_layer == max_layer) { 375 rtex->dirty_level_mask &= ~(1 << level); 376 } 377 } 378} 379 380void r600_decompress_color_textures(struct r600_context *rctx, 381 struct r600_samplerview_state *textures) 382{ 383 unsigned i; 384 unsigned mask = textures->compressed_colortex_mask; 385 386 while (mask) { 387 struct pipe_sampler_view *view; 388 struct r600_texture *tex; 389 390 i = u_bit_scan(&mask); 391 392 view = &textures->views[i]->base; 393 assert(view); 394 395 tex = (struct r600_texture *)view->texture; 396 assert(tex->cmask.size); 397 398 r600_blit_decompress_color(&rctx->b.b, tex, 399 view->u.tex.first_level, view->u.tex.last_level, 400 0, util_max_layer(&tex->resource.b.b, view->u.tex.first_level)); 401 } 402} 403 404void r600_decompress_color_images(struct r600_context *rctx, 405 struct r600_image_state *images) 406{ 407 unsigned i; 408 unsigned mask = images->compressed_colortex_mask; 409 410 while (mask) { 411 struct r600_image_view *view; 412 struct r600_texture *tex; 413 414 i = u_bit_scan(&mask); 415 416 view = &images->views[i]; 417 assert(view); 418 419 tex = (struct r600_texture *)view->base.resource; 420 assert(tex->cmask.size); 421 422 r600_blit_decompress_color(&rctx->b.b, tex, 423 view->base.u.tex.level, view->base.u.tex.level, 424 view->base.u.tex.first_layer, 425 view->base.u.tex.last_layer); 426 } 427} 428 429/* Helper for decompressing a portion of a color or depth resource before 430 * blitting if any decompression is needed. 431 * The driver doesn't decompress resources automatically while u_blitter is 432 * rendering. */ 433static bool r600_decompress_subresource(struct pipe_context *ctx, 434 struct pipe_resource *tex, 435 unsigned level, 436 unsigned first_layer, unsigned last_layer) 437{ 438 struct r600_context *rctx = (struct r600_context *)ctx; 439 struct r600_texture *rtex = (struct r600_texture*)tex; 440 441 if (rtex->db_compatible) { 442 if (r600_can_sample_zs(rtex, false)) { 443 r600_blit_decompress_depth_in_place(rctx, rtex, false, 444 level, level, 445 first_layer, last_layer); 446 if (rtex->surface.has_stencil) { 447 r600_blit_decompress_depth_in_place(rctx, rtex, true, 448 level, level, 449 first_layer, last_layer); 450 } 451 } else { 452 if (!r600_init_flushed_depth_texture(ctx, tex, NULL)) 453 return false; /* error */ 454 455 r600_blit_decompress_depth(ctx, rtex, NULL, 456 level, level, 457 first_layer, last_layer, 458 0, u_max_sample(tex)); 459 } 460 } else if (rtex->cmask.size) { 461 r600_blit_decompress_color(ctx, rtex, level, level, 462 first_layer, last_layer); 463 } 464 return true; 465} 466 467static void r600_clear(struct pipe_context *ctx, unsigned buffers, 468 const struct pipe_scissor_state *scissor_state, 469 const union pipe_color_union *color, 470 double depth, unsigned stencil) 471{ 472 struct r600_context *rctx = (struct r600_context *)ctx; 473 struct pipe_framebuffer_state *fb = &rctx->framebuffer.state; 474 475 if (buffers & PIPE_CLEAR_COLOR && rctx->b.gfx_level >= EVERGREEN) { 476 evergreen_do_fast_color_clear(&rctx->b, fb, &rctx->framebuffer.atom, 477 &buffers, NULL, color); 478 if (!buffers) 479 return; /* all buffers have been fast cleared */ 480 } 481 482 if (buffers & PIPE_CLEAR_COLOR) { 483 int i; 484 485 /* These buffers cannot use fast clear, make sure to disable expansion. */ 486 for (i = 0; i < fb->nr_cbufs; i++) { 487 struct r600_texture *tex; 488 489 /* If not clearing this buffer, skip. */ 490 if (!(buffers & (PIPE_CLEAR_COLOR0 << i))) 491 continue; 492 493 if (!fb->cbufs[i]) 494 continue; 495 496 tex = (struct r600_texture *)fb->cbufs[i]->texture; 497 if (tex->fmask.size == 0) 498 tex->dirty_level_mask &= ~(1 << fb->cbufs[i]->u.tex.level); 499 } 500 } 501 502 /* if hyperz enabled just clear hyperz */ 503 if (fb->zsbuf && (buffers & PIPE_CLEAR_DEPTH)) { 504 struct r600_texture *rtex; 505 unsigned level = fb->zsbuf->u.tex.level; 506 507 rtex = (struct r600_texture*)fb->zsbuf->texture; 508 509 /* We can't use hyperz fast clear if each slice of a texture 510 * array are clear to different value. To simplify code just 511 * disable fast clear for texture array. 512 */ 513 if (r600_htile_enabled(rtex, level) && 514 fb->zsbuf->u.tex.first_layer == 0 && 515 fb->zsbuf->u.tex.last_layer == util_max_layer(&rtex->resource.b.b, level)) { 516 if (rtex->depth_clear_value != depth) { 517 rtex->depth_clear_value = depth; 518 r600_mark_atom_dirty(rctx, &rctx->db_state.atom); 519 } 520 rctx->db_misc_state.htile_clear = true; 521 r600_mark_atom_dirty(rctx, &rctx->db_misc_state.atom); 522 } 523 } 524 525 r600_blitter_begin(ctx, R600_CLEAR); 526 util_blitter_clear(rctx->blitter, fb->width, fb->height, 527 util_framebuffer_get_num_layers(fb), 528 buffers, color, depth, stencil, 529 util_framebuffer_get_num_samples(fb) > 1); 530 r600_blitter_end(ctx); 531 532 /* disable fast clear */ 533 if (rctx->db_misc_state.htile_clear) { 534 rctx->db_misc_state.htile_clear = false; 535 r600_mark_atom_dirty(rctx, &rctx->db_misc_state.atom); 536 } 537} 538 539static void r600_clear_render_target(struct pipe_context *ctx, 540 struct pipe_surface *dst, 541 const union pipe_color_union *color, 542 unsigned dstx, unsigned dsty, 543 unsigned width, unsigned height, 544 bool render_condition_enabled) 545{ 546 struct r600_context *rctx = (struct r600_context *)ctx; 547 548 r600_blitter_begin(ctx, R600_CLEAR_SURFACE | 549 (render_condition_enabled ? 0 : R600_DISABLE_RENDER_COND)); 550 util_blitter_clear_render_target(rctx->blitter, dst, color, 551 dstx, dsty, width, height); 552 r600_blitter_end(ctx); 553} 554 555static void r600_clear_depth_stencil(struct pipe_context *ctx, 556 struct pipe_surface *dst, 557 unsigned clear_flags, 558 double depth, 559 unsigned stencil, 560 unsigned dstx, unsigned dsty, 561 unsigned width, unsigned height, 562 bool render_condition_enabled) 563{ 564 struct r600_context *rctx = (struct r600_context *)ctx; 565 566 r600_blitter_begin(ctx, R600_CLEAR_SURFACE | 567 (render_condition_enabled ? 0 : R600_DISABLE_RENDER_COND)); 568 util_blitter_clear_depth_stencil(rctx->blitter, dst, clear_flags, depth, stencil, 569 dstx, dsty, width, height); 570 r600_blitter_end(ctx); 571} 572 573static void r600_copy_buffer(struct pipe_context *ctx, struct pipe_resource *dst, unsigned dstx, 574 struct pipe_resource *src, const struct pipe_box *src_box) 575{ 576 struct r600_context *rctx = (struct r600_context*)ctx; 577 578 if (rctx->screen->b.has_cp_dma) 579 r600_cp_dma_copy_buffer(rctx, dst, dstx, src, src_box->x, src_box->width); 580 else 581 util_resource_copy_region(ctx, dst, 0, dstx, 0, 0, src, 0, src_box); 582} 583 584/** 585 * Global buffers are not really resources, they are are actually offsets 586 * into a single global resource (r600_screen::global_pool). The means 587 * they don't have their own buf handle, so they cannot be passed 588 * to r600_copy_buffer() and must be handled separately. 589 */ 590static void r600_copy_global_buffer(struct pipe_context *ctx, 591 struct pipe_resource *dst, unsigned 592 dstx, struct pipe_resource *src, 593 const struct pipe_box *src_box) 594{ 595 struct r600_context *rctx = (struct r600_context*)ctx; 596 struct compute_memory_pool *pool = rctx->screen->global_pool; 597 struct pipe_box new_src_box = *src_box; 598 599 if (src->bind & PIPE_BIND_GLOBAL) { 600 struct r600_resource_global *rsrc = 601 (struct r600_resource_global *)src; 602 struct compute_memory_item *item = rsrc->chunk; 603 604 if (is_item_in_pool(item)) { 605 new_src_box.x += 4 * item->start_in_dw; 606 src = (struct pipe_resource *)pool->bo; 607 } else { 608 if (item->real_buffer == NULL) { 609 item->real_buffer = 610 r600_compute_buffer_alloc_vram(pool->screen, 611 item->size_in_dw * 4); 612 } 613 src = (struct pipe_resource*)item->real_buffer; 614 } 615 } 616 if (dst->bind & PIPE_BIND_GLOBAL) { 617 struct r600_resource_global *rdst = 618 (struct r600_resource_global *)dst; 619 struct compute_memory_item *item = rdst->chunk; 620 621 if (is_item_in_pool(item)) { 622 dstx += 4 * item->start_in_dw; 623 dst = (struct pipe_resource *)pool->bo; 624 } else { 625 if (item->real_buffer == NULL) { 626 item->real_buffer = 627 r600_compute_buffer_alloc_vram(pool->screen, 628 item->size_in_dw * 4); 629 } 630 dst = (struct pipe_resource*)item->real_buffer; 631 } 632 } 633 634 r600_copy_buffer(ctx, dst, dstx, src, &new_src_box); 635} 636 637static void r600_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst, 638 uint64_t offset, uint64_t size, unsigned value, 639 enum r600_coherency coher) 640{ 641 struct r600_context *rctx = (struct r600_context*)ctx; 642 643 if (rctx->screen->b.has_cp_dma && 644 rctx->b.gfx_level >= EVERGREEN && 645 offset % 4 == 0 && size % 4 == 0) { 646 evergreen_cp_dma_clear_buffer(rctx, dst, offset, size, value, coher); 647 } else if (rctx->screen->b.has_streamout && offset % 4 == 0 && size % 4 == 0) { 648 union pipe_color_union clear_value; 649 clear_value.ui[0] = value; 650 651 r600_blitter_begin(ctx, R600_DISABLE_RENDER_COND); 652 util_blitter_clear_buffer(rctx->blitter, dst, offset, size, 653 1, &clear_value); 654 r600_blitter_end(ctx); 655 } else { 656 uint32_t *map = r600_buffer_map_sync_with_rings(&rctx->b, r600_resource(dst), 657 PIPE_MAP_WRITE); 658 map += offset / 4; 659 size /= 4; 660 for (unsigned i = 0; i < size; i++) 661 *map++ = value; 662 } 663} 664 665void r600_resource_copy_region(struct pipe_context *ctx, 666 struct pipe_resource *dst, 667 unsigned dst_level, 668 unsigned dstx, unsigned dsty, unsigned dstz, 669 struct pipe_resource *src, 670 unsigned src_level, 671 const struct pipe_box *src_box) 672{ 673 struct r600_context *rctx = (struct r600_context *)ctx; 674 struct pipe_surface *dst_view, dst_templ; 675 struct pipe_sampler_view src_templ, *src_view; 676 unsigned dst_width, dst_height, src_width0, src_height0, src_widthFL, src_heightFL; 677 unsigned src_force_level = 0; 678 struct pipe_box sbox, dstbox; 679 680 /* Handle buffers first. */ 681 if (dst->target == PIPE_BUFFER && src->target == PIPE_BUFFER) { 682 if ((src->bind & PIPE_BIND_GLOBAL) || 683 (dst->bind & PIPE_BIND_GLOBAL)) { 684 r600_copy_global_buffer(ctx, dst, dstx, src, src_box); 685 } else { 686 r600_copy_buffer(ctx, dst, dstx, src, src_box); 687 } 688 return; 689 } 690 691 assert(u_max_sample(dst) == u_max_sample(src)); 692 693 /* The driver doesn't decompress resources automatically while 694 * u_blitter is rendering. */ 695 if (!r600_decompress_subresource(ctx, src, src_level, 696 src_box->z, src_box->z + src_box->depth - 1)) { 697 return; /* error */ 698 } 699 700 dst_width = u_minify(dst->width0, dst_level); 701 dst_height = u_minify(dst->height0, dst_level); 702 src_width0 = src->width0; 703 src_height0 = src->height0; 704 src_widthFL = u_minify(src->width0, src_level); 705 src_heightFL = u_minify(src->height0, src_level); 706 707 util_blitter_default_dst_texture(&dst_templ, dst, dst_level, dstz); 708 util_blitter_default_src_texture(rctx->blitter, &src_templ, src, src_level); 709 710 if (util_format_is_compressed(src->format) || 711 util_format_is_compressed(dst->format)) { 712 unsigned blocksize = util_format_get_blocksize(src->format); 713 714 if (blocksize == 8) 715 src_templ.format = PIPE_FORMAT_R16G16B16A16_UINT; /* 64-bit block */ 716 else 717 src_templ.format = PIPE_FORMAT_R32G32B32A32_UINT; /* 128-bit block */ 718 dst_templ.format = src_templ.format; 719 720 dst_width = util_format_get_nblocksx(dst->format, dst_width); 721 dst_height = util_format_get_nblocksy(dst->format, dst_height); 722 src_width0 = util_format_get_nblocksx(src->format, src_width0); 723 src_height0 = util_format_get_nblocksy(src->format, src_height0); 724 src_widthFL = util_format_get_nblocksx(src->format, src_widthFL); 725 src_heightFL = util_format_get_nblocksy(src->format, src_heightFL); 726 727 dstx = util_format_get_nblocksx(dst->format, dstx); 728 dsty = util_format_get_nblocksy(dst->format, dsty); 729 730 sbox.x = util_format_get_nblocksx(src->format, src_box->x); 731 sbox.y = util_format_get_nblocksy(src->format, src_box->y); 732 sbox.z = src_box->z; 733 sbox.width = util_format_get_nblocksx(src->format, src_box->width); 734 sbox.height = util_format_get_nblocksy(src->format, src_box->height); 735 sbox.depth = src_box->depth; 736 src_box = &sbox; 737 738 src_force_level = src_level; 739 } else if (!util_blitter_is_copy_supported(rctx->blitter, dst, src)) { 740 if (util_format_is_subsampled_422(src->format)) { 741 742 src_templ.format = PIPE_FORMAT_R8G8B8A8_UINT; 743 dst_templ.format = PIPE_FORMAT_R8G8B8A8_UINT; 744 745 dst_width = util_format_get_nblocksx(dst->format, dst_width); 746 src_width0 = util_format_get_nblocksx(src->format, src_width0); 747 src_widthFL = util_format_get_nblocksx(src->format, src_widthFL); 748 749 dstx = util_format_get_nblocksx(dst->format, dstx); 750 751 sbox = *src_box; 752 sbox.x = util_format_get_nblocksx(src->format, src_box->x); 753 sbox.width = util_format_get_nblocksx(src->format, src_box->width); 754 src_box = &sbox; 755 } else { 756 unsigned blocksize = util_format_get_blocksize(src->format); 757 758 switch (blocksize) { 759 case 1: 760 dst_templ.format = PIPE_FORMAT_R8_UNORM; 761 src_templ.format = PIPE_FORMAT_R8_UNORM; 762 break; 763 case 2: 764 dst_templ.format = PIPE_FORMAT_R8G8_UNORM; 765 src_templ.format = PIPE_FORMAT_R8G8_UNORM; 766 break; 767 case 4: 768 dst_templ.format = PIPE_FORMAT_R8G8B8A8_UNORM; 769 src_templ.format = PIPE_FORMAT_R8G8B8A8_UNORM; 770 break; 771 case 8: 772 dst_templ.format = PIPE_FORMAT_R16G16B16A16_UINT; 773 src_templ.format = PIPE_FORMAT_R16G16B16A16_UINT; 774 break; 775 case 16: 776 dst_templ.format = PIPE_FORMAT_R32G32B32A32_UINT; 777 src_templ.format = PIPE_FORMAT_R32G32B32A32_UINT; 778 break; 779 default: 780 fprintf(stderr, "Unhandled format %s with blocksize %u\n", 781 util_format_short_name(src->format), blocksize); 782 assert(0); 783 } 784 } 785 } 786 787 dst_view = r600_create_surface_custom(ctx, dst, &dst_templ, 788 /* we don't care about these two for r600g */ 789 dst->width0, dst->height0, 790 dst_width, dst_height); 791 792 if (rctx->b.gfx_level >= EVERGREEN) { 793 src_view = evergreen_create_sampler_view_custom(ctx, src, &src_templ, 794 src_width0, src_height0, 795 src_force_level); 796 } else { 797 src_view = r600_create_sampler_view_custom(ctx, src, &src_templ, 798 src_widthFL, src_heightFL); 799 } 800 801 u_box_3d(dstx, dsty, dstz, abs(src_box->width), abs(src_box->height), 802 abs(src_box->depth), &dstbox); 803 804 /* Copy. */ 805 r600_blitter_begin(ctx, R600_COPY_TEXTURE); 806 util_blitter_blit_generic(rctx->blitter, dst_view, &dstbox, 807 src_view, src_box, src_width0, src_height0, 808 PIPE_MASK_RGBAZS, PIPE_TEX_FILTER_NEAREST, NULL, 809 FALSE, FALSE, 0); 810 r600_blitter_end(ctx); 811 812 pipe_surface_reference(&dst_view, NULL); 813 pipe_sampler_view_reference(&src_view, NULL); 814} 815 816static bool do_hardware_msaa_resolve(struct pipe_context *ctx, 817 const struct pipe_blit_info *info) 818{ 819 struct r600_context *rctx = (struct r600_context*)ctx; 820 struct r600_texture *dst = (struct r600_texture*)info->dst.resource; 821 unsigned dst_width = u_minify(info->dst.resource->width0, info->dst.level); 822 unsigned dst_height = u_minify(info->dst.resource->height0, info->dst.level); 823 enum pipe_format format = info->src.format; 824 unsigned sample_mask = 825 rctx->b.gfx_level == CAYMAN ? ~0 : 826 ((1ull << MAX2(1, info->src.resource->nr_samples)) - 1); 827 struct pipe_resource *tmp, templ; 828 struct pipe_blit_info blit; 829 830 /* Check basic requirements for hw resolve. */ 831 if (!(info->src.resource->nr_samples > 1 && 832 info->dst.resource->nr_samples <= 1 && 833 !util_format_is_pure_integer(format) && 834 !util_format_is_depth_or_stencil(format) && 835 util_max_layer(info->src.resource, 0) == 0)) 836 return false; 837 838 /* Check the remaining requirements for hw resolve. */ 839 if (util_max_layer(info->dst.resource, info->dst.level) == 0 && 840 util_is_format_compatible(util_format_description(info->src.format), 841 util_format_description(info->dst.format)) && 842 !info->scissor_enable && 843 (info->mask & PIPE_MASK_RGBA) == PIPE_MASK_RGBA && 844 dst_width == info->src.resource->width0 && 845 dst_height == info->src.resource->height0 && 846 info->dst.box.x == 0 && 847 info->dst.box.y == 0 && 848 info->dst.box.width == dst_width && 849 info->dst.box.height == dst_height && 850 info->dst.box.depth == 1 && 851 info->src.box.x == 0 && 852 info->src.box.y == 0 && 853 info->src.box.width == dst_width && 854 info->src.box.height == dst_height && 855 info->src.box.depth == 1 && 856 dst->surface.u.legacy.level[info->dst.level].mode >= RADEON_SURF_MODE_1D && 857 (!dst->cmask.size || !dst->dirty_level_mask) /* dst cannot be fast-cleared */) { 858 r600_blitter_begin(ctx, R600_COLOR_RESOLVE | 859 (info->render_condition_enable ? 0 : R600_DISABLE_RENDER_COND)); 860 util_blitter_custom_resolve_color(rctx->blitter, 861 info->dst.resource, info->dst.level, 862 info->dst.box.z, 863 info->src.resource, info->src.box.z, 864 sample_mask, rctx->custom_blend_resolve, 865 format); 866 r600_blitter_end(ctx); 867 return true; 868 } 869 870 /* Shader-based resolve is VERY SLOW. Instead, resolve into 871 * a temporary texture and blit. 872 */ 873 memset(&templ, 0, sizeof(templ)); 874 templ.target = PIPE_TEXTURE_2D; 875 templ.format = info->src.resource->format; 876 templ.width0 = info->src.resource->width0; 877 templ.height0 = info->src.resource->height0; 878 templ.depth0 = 1; 879 templ.array_size = 1; 880 templ.usage = PIPE_USAGE_DEFAULT; 881 templ.flags = R600_RESOURCE_FLAG_FORCE_TILING; 882 883 tmp = ctx->screen->resource_create(ctx->screen, &templ); 884 if (!tmp) 885 return false; 886 887 /* resolve */ 888 r600_blitter_begin(ctx, R600_COLOR_RESOLVE | 889 (info->render_condition_enable ? 0 : R600_DISABLE_RENDER_COND)); 890 util_blitter_custom_resolve_color(rctx->blitter, tmp, 0, 0, 891 info->src.resource, info->src.box.z, 892 sample_mask, rctx->custom_blend_resolve, 893 format); 894 r600_blitter_end(ctx); 895 896 /* blit */ 897 blit = *info; 898 blit.src.resource = tmp; 899 blit.src.box.z = 0; 900 901 r600_blitter_begin(ctx, R600_BLIT | 902 (info->render_condition_enable ? 0 : R600_DISABLE_RENDER_COND)); 903 util_blitter_blit(rctx->blitter, &blit); 904 r600_blitter_end(ctx); 905 906 pipe_resource_reference(&tmp, NULL); 907 return true; 908} 909 910static void r600_blit(struct pipe_context *ctx, 911 const struct pipe_blit_info *info) 912{ 913 struct r600_context *rctx = (struct r600_context*)ctx; 914 struct r600_texture *rdst = (struct r600_texture *)info->dst.resource; 915 916 if (do_hardware_msaa_resolve(ctx, info)) { 917 return; 918 } 919 920 /* Using SDMA for copying to a linear texture in GTT is much faster. 921 * This improves DRI PRIME performance. 922 * 923 * resource_copy_region can't do this yet, because dma_copy calls it 924 * on failure (recursion). 925 */ 926 if (rdst->surface.u.legacy.level[info->dst.level].mode == 927 RADEON_SURF_MODE_LINEAR_ALIGNED && 928 rctx->b.dma_copy && 929 util_can_blit_via_copy_region(info, false, rctx->b.render_cond != NULL)) { 930 rctx->b.dma_copy(ctx, info->dst.resource, info->dst.level, 931 info->dst.box.x, info->dst.box.y, 932 info->dst.box.z, 933 info->src.resource, info->src.level, 934 &info->src.box); 935 return; 936 } 937 938 assert(util_blitter_is_blit_supported(rctx->blitter, info)); 939 940 /* The driver doesn't decompress resources automatically while 941 * u_blitter is rendering. */ 942 if (!r600_decompress_subresource(ctx, info->src.resource, info->src.level, 943 info->src.box.z, 944 info->src.box.z + info->src.box.depth - 1)) { 945 return; /* error */ 946 } 947 948 if (rctx->screen->b.debug_flags & DBG_FORCE_DMA && 949 util_try_blit_via_copy_region(ctx, info, rctx->b.render_cond != NULL)) 950 return; 951 952 r600_blitter_begin(ctx, R600_BLIT | 953 (info->render_condition_enable ? 0 : R600_DISABLE_RENDER_COND)); 954 util_blitter_blit(rctx->blitter, info); 955 r600_blitter_end(ctx); 956} 957 958static void r600_flush_resource(struct pipe_context *ctx, 959 struct pipe_resource *res) 960{ 961 struct r600_texture *rtex = (struct r600_texture*)res; 962 963 assert(res->target != PIPE_BUFFER); 964 965 if (!rtex->is_depth && rtex->cmask.size) { 966 r600_blit_decompress_color(ctx, rtex, 0, res->last_level, 967 0, util_max_layer(res, 0)); 968 } 969} 970 971void r600_init_blit_functions(struct r600_context *rctx) 972{ 973 rctx->b.b.clear = r600_clear; 974 rctx->b.b.clear_render_target = r600_clear_render_target; 975 rctx->b.b.clear_depth_stencil = r600_clear_depth_stencil; 976 rctx->b.b.resource_copy_region = r600_resource_copy_region; 977 rctx->b.b.blit = r600_blit; 978 rctx->b.b.flush_resource = r600_flush_resource; 979 rctx->b.clear_buffer = r600_clear_buffer; 980 rctx->b.blit_decompress_depth = r600_blit_decompress_depth; 981} 982