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/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); 80 } 81 82 if (op & R600_SAVE_FRAMEBUFFER) 83 util_blitter_save_framebuffer(rctx->blitter, &rctx->framebuffer.state); 84 85 if (op & R600_SAVE_TEXTURES) { 86 util_blitter_save_fragment_sampler_states( 87 rctx->blitter, util_last_bit(rctx->samplers[PIPE_SHADER_FRAGMENT].states.enabled_mask), 88 (void**)rctx->samplers[PIPE_SHADER_FRAGMENT].states.states); 89 90 util_blitter_save_fragment_sampler_views( 91 rctx->blitter, util_last_bit(rctx->samplers[PIPE_SHADER_FRAGMENT].views.enabled_mask), 92 (struct pipe_sampler_view**)rctx->samplers[PIPE_SHADER_FRAGMENT].views.views); 93 } 94 95 if (op & R600_DISABLE_RENDER_COND) 96 rctx->b.render_cond_force_off = true; 97} 98 99static void r600_blitter_end(struct pipe_context *ctx) 100{ 101 struct r600_context *rctx = (struct r600_context *)ctx; 102 103 rctx->b.render_cond_force_off = false; 104} 105 106static unsigned u_max_sample(struct pipe_resource *r) 107{ 108 return r->nr_samples ? r->nr_samples - 1 : 0; 109} 110 111static void r600_blit_decompress_depth(struct pipe_context *ctx, 112 struct r600_texture *texture, 113 struct r600_texture *staging, 114 unsigned first_level, unsigned last_level, 115 unsigned first_layer, unsigned last_layer, 116 unsigned first_sample, unsigned last_sample) 117{ 118 struct r600_context *rctx = (struct r600_context *)ctx; 119 unsigned layer, level, sample, checked_last_layer, max_layer, max_sample; 120 struct r600_texture *flushed_depth_texture = staging ? 121 staging : texture->flushed_depth_texture; 122 const struct util_format_description *desc = 123 util_format_description(texture->resource.b.b.format); 124 float depth; 125 126 if (!staging && !texture->dirty_level_mask) 127 return; 128 129 max_sample = u_max_sample(&texture->resource.b.b); 130 131 /* XXX Decompressing MSAA depth textures is broken on R6xx. 132 * There is also a hardlock if CMASK and FMASK are not present. 133 * Just skip this until we find out how to fix it. */ 134 if (rctx->b.chip_class == R600 && max_sample > 0) { 135 texture->dirty_level_mask = 0; 136 return; 137 } 138 139 if (rctx->b.family == CHIP_RV610 || rctx->b.family == CHIP_RV630 || 140 rctx->b.family == CHIP_RV620 || rctx->b.family == CHIP_RV635) 141 depth = 0.0f; 142 else 143 depth = 1.0f; 144 145 /* Enable decompression in DB_RENDER_CONTROL */ 146 rctx->db_misc_state.flush_depthstencil_through_cb = true; 147 rctx->db_misc_state.copy_depth = util_format_has_depth(desc); 148 rctx->db_misc_state.copy_stencil = util_format_has_stencil(desc); 149 rctx->db_misc_state.copy_sample = first_sample; 150 r600_mark_atom_dirty(rctx, &rctx->db_misc_state.atom); 151 152 for (level = first_level; level <= last_level; level++) { 153 if (!staging && !(texture->dirty_level_mask & (1 << level))) 154 continue; 155 156 /* The smaller the mipmap level, the less layers there are 157 * as far as 3D textures are concerned. */ 158 max_layer = util_max_layer(&texture->resource.b.b, level); 159 checked_last_layer = last_layer < max_layer ? last_layer : max_layer; 160 161 for (layer = first_layer; layer <= checked_last_layer; layer++) { 162 for (sample = first_sample; sample <= last_sample; sample++) { 163 struct pipe_surface *zsurf, *cbsurf, surf_tmpl; 164 165 if (sample != rctx->db_misc_state.copy_sample) { 166 rctx->db_misc_state.copy_sample = sample; 167 r600_mark_atom_dirty(rctx, &rctx->db_misc_state.atom); 168 } 169 170 surf_tmpl.format = texture->resource.b.b.format; 171 surf_tmpl.u.tex.level = level; 172 surf_tmpl.u.tex.first_layer = layer; 173 surf_tmpl.u.tex.last_layer = layer; 174 175 zsurf = ctx->create_surface(ctx, &texture->resource.b.b, &surf_tmpl); 176 177 surf_tmpl.format = flushed_depth_texture->resource.b.b.format; 178 cbsurf = ctx->create_surface(ctx, 179 &flushed_depth_texture->resource.b.b, &surf_tmpl); 180 181 r600_blitter_begin(ctx, R600_DECOMPRESS); 182 util_blitter_custom_depth_stencil(rctx->blitter, zsurf, cbsurf, 1 << sample, 183 rctx->custom_dsa_flush, depth); 184 r600_blitter_end(ctx); 185 186 pipe_surface_reference(&zsurf, NULL); 187 pipe_surface_reference(&cbsurf, NULL); 188 } 189 } 190 191 /* The texture will always be dirty if some layers or samples aren't flushed. 192 * I don't think this case occurs often though. */ 193 if (!staging && 194 first_layer == 0 && last_layer == max_layer && 195 first_sample == 0 && last_sample == max_sample) { 196 texture->dirty_level_mask &= ~(1 << level); 197 } 198 } 199 200 /* reenable compression in DB_RENDER_CONTROL */ 201 rctx->db_misc_state.flush_depthstencil_through_cb = false; 202 r600_mark_atom_dirty(rctx, &rctx->db_misc_state.atom); 203} 204 205static void r600_blit_decompress_depth_in_place(struct r600_context *rctx, 206 struct r600_texture *texture, 207 bool is_stencil_sampler, 208 unsigned first_level, unsigned last_level, 209 unsigned first_layer, unsigned last_layer) 210{ 211 struct pipe_surface *zsurf, surf_tmpl = {{0}}; 212 unsigned layer, max_layer, checked_last_layer, level; 213 unsigned *dirty_level_mask; 214 215 /* Enable decompression in DB_RENDER_CONTROL */ 216 if (is_stencil_sampler) { 217 rctx->db_misc_state.flush_stencil_inplace = true; 218 dirty_level_mask = &texture->stencil_dirty_level_mask; 219 } else { 220 rctx->db_misc_state.flush_depth_inplace = true; 221 dirty_level_mask = &texture->dirty_level_mask; 222 } 223 r600_mark_atom_dirty(rctx, &rctx->db_misc_state.atom); 224 225 surf_tmpl.format = texture->resource.b.b.format; 226 227 for (level = first_level; level <= last_level; level++) { 228 if (!(*dirty_level_mask & (1 << level))) 229 continue; 230 231 surf_tmpl.u.tex.level = level; 232 233 /* The smaller the mipmap level, the less layers there are 234 * as far as 3D textures are concerned. */ 235 max_layer = util_max_layer(&texture->resource.b.b, level); 236 checked_last_layer = last_layer < max_layer ? last_layer : max_layer; 237 238 for (layer = first_layer; layer <= checked_last_layer; layer++) { 239 surf_tmpl.u.tex.first_layer = layer; 240 surf_tmpl.u.tex.last_layer = layer; 241 242 zsurf = rctx->b.b.create_surface(&rctx->b.b, &texture->resource.b.b, &surf_tmpl); 243 244 r600_blitter_begin(&rctx->b.b, R600_DECOMPRESS); 245 util_blitter_custom_depth_stencil(rctx->blitter, zsurf, NULL, ~0, 246 rctx->custom_dsa_flush, 1.0f); 247 r600_blitter_end(&rctx->b.b); 248 249 pipe_surface_reference(&zsurf, NULL); 250 } 251 252 /* The texture will always be dirty if some layers or samples aren't flushed. 253 * I don't think this case occurs often though. */ 254 if (first_layer == 0 && last_layer == max_layer) { 255 *dirty_level_mask &= ~(1 << level); 256 } 257 } 258 259 /* Disable decompression in DB_RENDER_CONTROL */ 260 rctx->db_misc_state.flush_depth_inplace = false; 261 rctx->db_misc_state.flush_stencil_inplace = false; 262 r600_mark_atom_dirty(rctx, &rctx->db_misc_state.atom); 263} 264 265void r600_decompress_depth_textures(struct r600_context *rctx, 266 struct r600_samplerview_state *textures) 267{ 268 unsigned i; 269 unsigned depth_texture_mask = textures->compressed_depthtex_mask; 270 271 while (depth_texture_mask) { 272 struct pipe_sampler_view *view; 273 struct r600_pipe_sampler_view *rview; 274 struct r600_texture *tex; 275 276 i = u_bit_scan(&depth_texture_mask); 277 278 view = &textures->views[i]->base; 279 assert(view); 280 rview = (struct r600_pipe_sampler_view*)view; 281 282 tex = (struct r600_texture *)view->texture; 283 assert(tex->db_compatible); 284 285 if (r600_can_sample_zs(tex, rview->is_stencil_sampler)) { 286 r600_blit_decompress_depth_in_place(rctx, tex, 287 rview->is_stencil_sampler, 288 view->u.tex.first_level, view->u.tex.last_level, 289 0, util_max_layer(&tex->resource.b.b, view->u.tex.first_level)); 290 } else { 291 r600_blit_decompress_depth(&rctx->b.b, tex, NULL, 292 view->u.tex.first_level, view->u.tex.last_level, 293 0, util_max_layer(&tex->resource.b.b, view->u.tex.first_level), 294 0, u_max_sample(&tex->resource.b.b)); 295 } 296 } 297} 298 299void r600_decompress_depth_images(struct r600_context *rctx, 300 struct r600_image_state *images) 301{ 302 unsigned i; 303 unsigned depth_texture_mask = images->compressed_depthtex_mask; 304 305 while (depth_texture_mask) { 306 struct r600_image_view *view; 307 struct r600_texture *tex; 308 309 i = u_bit_scan(&depth_texture_mask); 310 311 view = &images->views[i]; 312 assert(view); 313 314 tex = (struct r600_texture *)view->base.resource; 315 assert(tex->db_compatible); 316 317 if (r600_can_sample_zs(tex, false)) { 318 r600_blit_decompress_depth_in_place(rctx, tex, 319 false, 320 view->base.u.tex.level, 321 view->base.u.tex.level, 322 0, util_max_layer(&tex->resource.b.b, view->base.u.tex.level)); 323 } else { 324 r600_blit_decompress_depth(&rctx->b.b, tex, NULL, 325 view->base.u.tex.level, 326 view->base.u.tex.level, 327 0, util_max_layer(&tex->resource.b.b, view->base.u.tex.level), 328 0, u_max_sample(&tex->resource.b.b)); 329 } 330 } 331} 332 333static void r600_blit_decompress_color(struct pipe_context *ctx, 334 struct r600_texture *rtex, 335 unsigned first_level, unsigned last_level, 336 unsigned first_layer, unsigned last_layer) 337{ 338 struct r600_context *rctx = (struct r600_context *)ctx; 339 unsigned layer, level, checked_last_layer, max_layer; 340 341 if (!rtex->dirty_level_mask) 342 return; 343 344 for (level = first_level; level <= last_level; level++) { 345 if (!(rtex->dirty_level_mask & (1 << level))) 346 continue; 347 348 /* The smaller the mipmap level, the less layers there are 349 * as far as 3D textures are concerned. */ 350 max_layer = util_max_layer(&rtex->resource.b.b, level); 351 checked_last_layer = last_layer < max_layer ? last_layer : max_layer; 352 353 for (layer = first_layer; layer <= checked_last_layer; layer++) { 354 struct pipe_surface *cbsurf, surf_tmpl; 355 356 surf_tmpl.format = rtex->resource.b.b.format; 357 surf_tmpl.u.tex.level = level; 358 surf_tmpl.u.tex.first_layer = layer; 359 surf_tmpl.u.tex.last_layer = layer; 360 cbsurf = ctx->create_surface(ctx, &rtex->resource.b.b, &surf_tmpl); 361 362 r600_blitter_begin(ctx, R600_DECOMPRESS); 363 util_blitter_custom_color(rctx->blitter, cbsurf, 364 rtex->fmask.size ? rctx->custom_blend_decompress : rctx->custom_blend_fastclear); 365 r600_blitter_end(ctx); 366 367 pipe_surface_reference(&cbsurf, NULL); 368 } 369 370 /* The texture will always be dirty if some layers aren't flushed. 371 * I don't think this case occurs often though. */ 372 if (first_layer == 0 && last_layer == max_layer) { 373 rtex->dirty_level_mask &= ~(1 << level); 374 } 375 } 376} 377 378void r600_decompress_color_textures(struct r600_context *rctx, 379 struct r600_samplerview_state *textures) 380{ 381 unsigned i; 382 unsigned mask = textures->compressed_colortex_mask; 383 384 while (mask) { 385 struct pipe_sampler_view *view; 386 struct r600_texture *tex; 387 388 i = u_bit_scan(&mask); 389 390 view = &textures->views[i]->base; 391 assert(view); 392 393 tex = (struct r600_texture *)view->texture; 394 assert(tex->cmask.size); 395 396 r600_blit_decompress_color(&rctx->b.b, tex, 397 view->u.tex.first_level, view->u.tex.last_level, 398 0, util_max_layer(&tex->resource.b.b, view->u.tex.first_level)); 399 } 400} 401 402void r600_decompress_color_images(struct r600_context *rctx, 403 struct r600_image_state *images) 404{ 405 unsigned i; 406 unsigned mask = images->compressed_colortex_mask; 407 408 while (mask) { 409 struct r600_image_view *view; 410 struct r600_texture *tex; 411 412 i = u_bit_scan(&mask); 413 414 view = &images->views[i]; 415 assert(view); 416 417 tex = (struct r600_texture *)view->base.resource; 418 assert(tex->cmask.size); 419 420 r600_blit_decompress_color(&rctx->b.b, tex, 421 view->base.u.tex.level, view->base.u.tex.level, 422 view->base.u.tex.first_layer, 423 view->base.u.tex.last_layer); 424 } 425} 426 427/* Helper for decompressing a portion of a color or depth resource before 428 * blitting if any decompression is needed. 429 * The driver doesn't decompress resources automatically while u_blitter is 430 * rendering. */ 431static bool r600_decompress_subresource(struct pipe_context *ctx, 432 struct pipe_resource *tex, 433 unsigned level, 434 unsigned first_layer, unsigned last_layer) 435{ 436 struct r600_context *rctx = (struct r600_context *)ctx; 437 struct r600_texture *rtex = (struct r600_texture*)tex; 438 439 if (rtex->db_compatible) { 440 if (r600_can_sample_zs(rtex, false)) { 441 r600_blit_decompress_depth_in_place(rctx, rtex, false, 442 level, level, 443 first_layer, last_layer); 444 if (rtex->surface.has_stencil) { 445 r600_blit_decompress_depth_in_place(rctx, rtex, true, 446 level, level, 447 first_layer, last_layer); 448 } 449 } else { 450 if (!r600_init_flushed_depth_texture(ctx, tex, NULL)) 451 return false; /* error */ 452 453 r600_blit_decompress_depth(ctx, rtex, NULL, 454 level, level, 455 first_layer, last_layer, 456 0, u_max_sample(tex)); 457 } 458 } else if (rtex->cmask.size) { 459 r600_blit_decompress_color(ctx, rtex, level, level, 460 first_layer, last_layer); 461 } 462 return true; 463} 464 465static void r600_clear(struct pipe_context *ctx, unsigned buffers, 466 const union pipe_color_union *color, 467 double depth, unsigned stencil) 468{ 469 struct r600_context *rctx = (struct r600_context *)ctx; 470 struct pipe_framebuffer_state *fb = &rctx->framebuffer.state; 471 472 if (buffers & PIPE_CLEAR_COLOR && rctx->b.chip_class >= EVERGREEN) { 473 evergreen_do_fast_color_clear(&rctx->b, fb, &rctx->framebuffer.atom, 474 &buffers, NULL, color); 475 if (!buffers) 476 return; /* all buffers have been fast cleared */ 477 } 478 479 if (buffers & PIPE_CLEAR_COLOR) { 480 int i; 481 482 /* These buffers cannot use fast clear, make sure to disable expansion. */ 483 for (i = 0; i < fb->nr_cbufs; i++) { 484 struct r600_texture *tex; 485 486 /* If not clearing this buffer, skip. */ 487 if (!(buffers & (PIPE_CLEAR_COLOR0 << i))) 488 continue; 489 490 if (!fb->cbufs[i]) 491 continue; 492 493 tex = (struct r600_texture *)fb->cbufs[i]->texture; 494 if (tex->fmask.size == 0) 495 tex->dirty_level_mask &= ~(1 << fb->cbufs[i]->u.tex.level); 496 } 497 } 498 499 /* if hyperz enabled just clear hyperz */ 500 if (fb->zsbuf && (buffers & PIPE_CLEAR_DEPTH)) { 501 struct r600_texture *rtex; 502 unsigned level = fb->zsbuf->u.tex.level; 503 504 rtex = (struct r600_texture*)fb->zsbuf->texture; 505 506 /* We can't use hyperz fast clear if each slice of a texture 507 * array are clear to different value. To simplify code just 508 * disable fast clear for texture array. 509 */ 510 if (r600_htile_enabled(rtex, level) && 511 fb->zsbuf->u.tex.first_layer == 0 && 512 fb->zsbuf->u.tex.last_layer == util_max_layer(&rtex->resource.b.b, level)) { 513 if (rtex->depth_clear_value != depth) { 514 rtex->depth_clear_value = depth; 515 r600_mark_atom_dirty(rctx, &rctx->db_state.atom); 516 } 517 rctx->db_misc_state.htile_clear = true; 518 r600_mark_atom_dirty(rctx, &rctx->db_misc_state.atom); 519 } 520 } 521 522 r600_blitter_begin(ctx, R600_CLEAR); 523 util_blitter_clear(rctx->blitter, fb->width, fb->height, 524 util_framebuffer_get_num_layers(fb), 525 buffers, color, depth, stencil); 526 r600_blitter_end(ctx); 527 528 /* disable fast clear */ 529 if (rctx->db_misc_state.htile_clear) { 530 rctx->db_misc_state.htile_clear = false; 531 r600_mark_atom_dirty(rctx, &rctx->db_misc_state.atom); 532 } 533} 534 535static void r600_clear_render_target(struct pipe_context *ctx, 536 struct pipe_surface *dst, 537 const union pipe_color_union *color, 538 unsigned dstx, unsigned dsty, 539 unsigned width, unsigned height, 540 bool render_condition_enabled) 541{ 542 struct r600_context *rctx = (struct r600_context *)ctx; 543 544 r600_blitter_begin(ctx, R600_CLEAR_SURFACE | 545 (render_condition_enabled ? 0 : R600_DISABLE_RENDER_COND)); 546 util_blitter_clear_render_target(rctx->blitter, dst, color, 547 dstx, dsty, width, height); 548 r600_blitter_end(ctx); 549} 550 551static void r600_clear_depth_stencil(struct pipe_context *ctx, 552 struct pipe_surface *dst, 553 unsigned clear_flags, 554 double depth, 555 unsigned stencil, 556 unsigned dstx, unsigned dsty, 557 unsigned width, unsigned height, 558 bool render_condition_enabled) 559{ 560 struct r600_context *rctx = (struct r600_context *)ctx; 561 562 r600_blitter_begin(ctx, R600_CLEAR_SURFACE | 563 (render_condition_enabled ? 0 : R600_DISABLE_RENDER_COND)); 564 util_blitter_clear_depth_stencil(rctx->blitter, dst, clear_flags, depth, stencil, 565 dstx, dsty, width, height); 566 r600_blitter_end(ctx); 567} 568 569static void r600_copy_buffer(struct pipe_context *ctx, struct pipe_resource *dst, unsigned dstx, 570 struct pipe_resource *src, const struct pipe_box *src_box) 571{ 572 struct r600_context *rctx = (struct r600_context*)ctx; 573 574 if (rctx->screen->b.has_cp_dma) { 575 r600_cp_dma_copy_buffer(rctx, dst, dstx, src, src_box->x, src_box->width); 576 } 577 else if (rctx->screen->b.has_streamout && 578 /* Require 4-byte alignment. */ 579 dstx % 4 == 0 && src_box->x % 4 == 0 && src_box->width % 4 == 0) { 580 581 r600_blitter_begin(ctx, R600_COPY_BUFFER); 582 util_blitter_copy_buffer(rctx->blitter, dst, dstx, src, src_box->x, src_box->width); 583 r600_blitter_end(ctx); 584 } else { 585 util_resource_copy_region(ctx, dst, 0, dstx, 0, 0, src, 0, src_box); 586 } 587} 588 589/** 590 * Global buffers are not really resources, they are are actually offsets 591 * into a single global resource (r600_screen::global_pool). The means 592 * they don't have their own buf handle, so they cannot be passed 593 * to r600_copy_buffer() and must be handled separately. 594 */ 595static void r600_copy_global_buffer(struct pipe_context *ctx, 596 struct pipe_resource *dst, unsigned 597 dstx, struct pipe_resource *src, 598 const struct pipe_box *src_box) 599{ 600 struct r600_context *rctx = (struct r600_context*)ctx; 601 struct compute_memory_pool *pool = rctx->screen->global_pool; 602 struct pipe_box new_src_box = *src_box; 603 604 if (src->bind & PIPE_BIND_GLOBAL) { 605 struct r600_resource_global *rsrc = 606 (struct r600_resource_global *)src; 607 struct compute_memory_item *item = rsrc->chunk; 608 609 if (is_item_in_pool(item)) { 610 new_src_box.x += 4 * item->start_in_dw; 611 src = (struct pipe_resource *)pool->bo; 612 } else { 613 if (item->real_buffer == NULL) { 614 item->real_buffer = 615 r600_compute_buffer_alloc_vram(pool->screen, 616 item->size_in_dw * 4); 617 } 618 src = (struct pipe_resource*)item->real_buffer; 619 } 620 } 621 if (dst->bind & PIPE_BIND_GLOBAL) { 622 struct r600_resource_global *rdst = 623 (struct r600_resource_global *)dst; 624 struct compute_memory_item *item = rdst->chunk; 625 626 if (is_item_in_pool(item)) { 627 dstx += 4 * item->start_in_dw; 628 dst = (struct pipe_resource *)pool->bo; 629 } else { 630 if (item->real_buffer == NULL) { 631 item->real_buffer = 632 r600_compute_buffer_alloc_vram(pool->screen, 633 item->size_in_dw * 4); 634 } 635 dst = (struct pipe_resource*)item->real_buffer; 636 } 637 } 638 639 r600_copy_buffer(ctx, dst, dstx, src, &new_src_box); 640} 641 642static void r600_clear_buffer(struct pipe_context *ctx, struct pipe_resource *dst, 643 uint64_t offset, uint64_t size, unsigned value, 644 enum r600_coherency coher) 645{ 646 struct r600_context *rctx = (struct r600_context*)ctx; 647 648 if (rctx->screen->b.has_cp_dma && 649 rctx->b.chip_class >= EVERGREEN && 650 offset % 4 == 0 && size % 4 == 0) { 651 evergreen_cp_dma_clear_buffer(rctx, dst, offset, size, value, coher); 652 } else if (rctx->screen->b.has_streamout && offset % 4 == 0 && size % 4 == 0) { 653 union pipe_color_union clear_value; 654 clear_value.ui[0] = value; 655 656 r600_blitter_begin(ctx, R600_DISABLE_RENDER_COND); 657 util_blitter_clear_buffer(rctx->blitter, dst, offset, size, 658 1, &clear_value); 659 r600_blitter_end(ctx); 660 } else { 661 uint32_t *map = r600_buffer_map_sync_with_rings(&rctx->b, r600_resource(dst), 662 PIPE_TRANSFER_WRITE); 663 map += offset / 4; 664 size /= 4; 665 for (unsigned i = 0; i < size; i++) 666 *map++ = value; 667 } 668} 669 670void r600_resource_copy_region(struct pipe_context *ctx, 671 struct pipe_resource *dst, 672 unsigned dst_level, 673 unsigned dstx, unsigned dsty, unsigned dstz, 674 struct pipe_resource *src, 675 unsigned src_level, 676 const struct pipe_box *src_box) 677{ 678 struct r600_context *rctx = (struct r600_context *)ctx; 679 struct pipe_surface *dst_view, dst_templ; 680 struct pipe_sampler_view src_templ, *src_view; 681 unsigned dst_width, dst_height, src_width0, src_height0, src_widthFL, src_heightFL; 682 unsigned src_force_level = 0; 683 struct pipe_box sbox, dstbox; 684 685 /* Handle buffers first. */ 686 if (dst->target == PIPE_BUFFER && src->target == PIPE_BUFFER) { 687 if ((src->bind & PIPE_BIND_GLOBAL) || 688 (dst->bind & PIPE_BIND_GLOBAL)) { 689 r600_copy_global_buffer(ctx, dst, dstx, src, src_box); 690 } else { 691 r600_copy_buffer(ctx, dst, dstx, src, src_box); 692 } 693 return; 694 } 695 696 assert(u_max_sample(dst) == u_max_sample(src)); 697 698 /* The driver doesn't decompress resources automatically while 699 * u_blitter is rendering. */ 700 if (!r600_decompress_subresource(ctx, src, src_level, 701 src_box->z, src_box->z + src_box->depth - 1)) { 702 return; /* error */ 703 } 704 705 dst_width = u_minify(dst->width0, dst_level); 706 dst_height = u_minify(dst->height0, dst_level); 707 src_width0 = src->width0; 708 src_height0 = src->height0; 709 src_widthFL = u_minify(src->width0, src_level); 710 src_heightFL = u_minify(src->height0, src_level); 711 712 util_blitter_default_dst_texture(&dst_templ, dst, dst_level, dstz); 713 util_blitter_default_src_texture(rctx->blitter, &src_templ, src, src_level); 714 715 if (util_format_is_compressed(src->format) || 716 util_format_is_compressed(dst->format)) { 717 unsigned blocksize = util_format_get_blocksize(src->format); 718 719 if (blocksize == 8) 720 src_templ.format = PIPE_FORMAT_R16G16B16A16_UINT; /* 64-bit block */ 721 else 722 src_templ.format = PIPE_FORMAT_R32G32B32A32_UINT; /* 128-bit block */ 723 dst_templ.format = src_templ.format; 724 725 dst_width = util_format_get_nblocksx(dst->format, dst_width); 726 dst_height = util_format_get_nblocksy(dst->format, dst_height); 727 src_width0 = util_format_get_nblocksx(src->format, src_width0); 728 src_height0 = util_format_get_nblocksy(src->format, src_height0); 729 src_widthFL = util_format_get_nblocksx(src->format, src_widthFL); 730 src_heightFL = util_format_get_nblocksy(src->format, src_heightFL); 731 732 dstx = util_format_get_nblocksx(dst->format, dstx); 733 dsty = util_format_get_nblocksy(dst->format, dsty); 734 735 sbox.x = util_format_get_nblocksx(src->format, src_box->x); 736 sbox.y = util_format_get_nblocksy(src->format, src_box->y); 737 sbox.z = src_box->z; 738 sbox.width = util_format_get_nblocksx(src->format, src_box->width); 739 sbox.height = util_format_get_nblocksy(src->format, src_box->height); 740 sbox.depth = src_box->depth; 741 src_box = &sbox; 742 743 src_force_level = src_level; 744 } else if (!util_blitter_is_copy_supported(rctx->blitter, dst, src)) { 745 if (util_format_is_subsampled_422(src->format)) { 746 747 src_templ.format = PIPE_FORMAT_R8G8B8A8_UINT; 748 dst_templ.format = PIPE_FORMAT_R8G8B8A8_UINT; 749 750 dst_width = util_format_get_nblocksx(dst->format, dst_width); 751 src_width0 = util_format_get_nblocksx(src->format, src_width0); 752 src_widthFL = util_format_get_nblocksx(src->format, src_widthFL); 753 754 dstx = util_format_get_nblocksx(dst->format, dstx); 755 756 sbox = *src_box; 757 sbox.x = util_format_get_nblocksx(src->format, src_box->x); 758 sbox.width = util_format_get_nblocksx(src->format, src_box->width); 759 src_box = &sbox; 760 } else { 761 unsigned blocksize = util_format_get_blocksize(src->format); 762 763 switch (blocksize) { 764 case 1: 765 dst_templ.format = PIPE_FORMAT_R8_UNORM; 766 src_templ.format = PIPE_FORMAT_R8_UNORM; 767 break; 768 case 2: 769 dst_templ.format = PIPE_FORMAT_R8G8_UNORM; 770 src_templ.format = PIPE_FORMAT_R8G8_UNORM; 771 break; 772 case 4: 773 dst_templ.format = PIPE_FORMAT_R8G8B8A8_UNORM; 774 src_templ.format = PIPE_FORMAT_R8G8B8A8_UNORM; 775 break; 776 case 8: 777 dst_templ.format = PIPE_FORMAT_R16G16B16A16_UINT; 778 src_templ.format = PIPE_FORMAT_R16G16B16A16_UINT; 779 break; 780 case 16: 781 dst_templ.format = PIPE_FORMAT_R32G32B32A32_UINT; 782 src_templ.format = PIPE_FORMAT_R32G32B32A32_UINT; 783 break; 784 default: 785 fprintf(stderr, "Unhandled format %s with blocksize %u\n", 786 util_format_short_name(src->format), blocksize); 787 assert(0); 788 } 789 } 790 } 791 792 dst_view = r600_create_surface_custom(ctx, dst, &dst_templ, 793 /* we don't care about these two for r600g */ 794 dst->width0, dst->height0, 795 dst_width, dst_height); 796 797 if (rctx->b.chip_class >= EVERGREEN) { 798 src_view = evergreen_create_sampler_view_custom(ctx, src, &src_templ, 799 src_width0, src_height0, 800 src_force_level); 801 } else { 802 src_view = r600_create_sampler_view_custom(ctx, src, &src_templ, 803 src_widthFL, src_heightFL); 804 } 805 806 u_box_3d(dstx, dsty, dstz, abs(src_box->width), abs(src_box->height), 807 abs(src_box->depth), &dstbox); 808 809 /* Copy. */ 810 r600_blitter_begin(ctx, R600_COPY_TEXTURE); 811 util_blitter_blit_generic(rctx->blitter, dst_view, &dstbox, 812 src_view, src_box, src_width0, src_height0, 813 PIPE_MASK_RGBAZS, PIPE_TEX_FILTER_NEAREST, NULL, 814 FALSE); 815 r600_blitter_end(ctx); 816 817 pipe_surface_reference(&dst_view, NULL); 818 pipe_sampler_view_reference(&src_view, NULL); 819} 820 821static bool do_hardware_msaa_resolve(struct pipe_context *ctx, 822 const struct pipe_blit_info *info) 823{ 824 struct r600_context *rctx = (struct r600_context*)ctx; 825 struct r600_texture *dst = (struct r600_texture*)info->dst.resource; 826 unsigned dst_width = u_minify(info->dst.resource->width0, info->dst.level); 827 unsigned dst_height = u_minify(info->dst.resource->height0, info->dst.level); 828 enum pipe_format format = info->src.format; 829 unsigned sample_mask = 830 rctx->b.chip_class == CAYMAN ? ~0 : 831 ((1ull << MAX2(1, info->src.resource->nr_samples)) - 1); 832 struct pipe_resource *tmp, templ; 833 struct pipe_blit_info blit; 834 835 /* Check basic requirements for hw resolve. */ 836 if (!(info->src.resource->nr_samples > 1 && 837 info->dst.resource->nr_samples <= 1 && 838 !util_format_is_pure_integer(format) && 839 !util_format_is_depth_or_stencil(format) && 840 util_max_layer(info->src.resource, 0) == 0)) 841 return false; 842 843 /* Check the remaining requirements for hw resolve. */ 844 if (util_max_layer(info->dst.resource, info->dst.level) == 0 && 845 util_is_format_compatible(util_format_description(info->src.format), 846 util_format_description(info->dst.format)) && 847 !info->scissor_enable && 848 (info->mask & PIPE_MASK_RGBA) == PIPE_MASK_RGBA && 849 dst_width == info->src.resource->width0 && 850 dst_height == info->src.resource->height0 && 851 info->dst.box.x == 0 && 852 info->dst.box.y == 0 && 853 info->dst.box.width == dst_width && 854 info->dst.box.height == dst_height && 855 info->dst.box.depth == 1 && 856 info->src.box.x == 0 && 857 info->src.box.y == 0 && 858 info->src.box.width == dst_width && 859 info->src.box.height == dst_height && 860 info->src.box.depth == 1 && 861 dst->surface.u.legacy.level[info->dst.level].mode >= RADEON_SURF_MODE_1D && 862 (!dst->cmask.size || !dst->dirty_level_mask) /* dst cannot be fast-cleared */) { 863 r600_blitter_begin(ctx, R600_COLOR_RESOLVE | 864 (info->render_condition_enable ? 0 : R600_DISABLE_RENDER_COND)); 865 util_blitter_custom_resolve_color(rctx->blitter, 866 info->dst.resource, info->dst.level, 867 info->dst.box.z, 868 info->src.resource, info->src.box.z, 869 sample_mask, rctx->custom_blend_resolve, 870 format); 871 r600_blitter_end(ctx); 872 return true; 873 } 874 875 /* Shader-based resolve is VERY SLOW. Instead, resolve into 876 * a temporary texture and blit. 877 */ 878 memset(&templ, 0, sizeof(templ)); 879 templ.target = PIPE_TEXTURE_2D; 880 templ.format = info->src.resource->format; 881 templ.width0 = info->src.resource->width0; 882 templ.height0 = info->src.resource->height0; 883 templ.depth0 = 1; 884 templ.array_size = 1; 885 templ.usage = PIPE_USAGE_DEFAULT; 886 templ.flags = R600_RESOURCE_FLAG_FORCE_TILING; 887 888 tmp = ctx->screen->resource_create(ctx->screen, &templ); 889 if (!tmp) 890 return false; 891 892 /* resolve */ 893 r600_blitter_begin(ctx, R600_COLOR_RESOLVE | 894 (info->render_condition_enable ? 0 : R600_DISABLE_RENDER_COND)); 895 util_blitter_custom_resolve_color(rctx->blitter, tmp, 0, 0, 896 info->src.resource, info->src.box.z, 897 sample_mask, rctx->custom_blend_resolve, 898 format); 899 r600_blitter_end(ctx); 900 901 /* blit */ 902 blit = *info; 903 blit.src.resource = tmp; 904 blit.src.box.z = 0; 905 906 r600_blitter_begin(ctx, R600_BLIT | 907 (info->render_condition_enable ? 0 : R600_DISABLE_RENDER_COND)); 908 util_blitter_blit(rctx->blitter, &blit); 909 r600_blitter_end(ctx); 910 911 pipe_resource_reference(&tmp, NULL); 912 return true; 913} 914 915static void r600_blit(struct pipe_context *ctx, 916 const struct pipe_blit_info *info) 917{ 918 struct r600_context *rctx = (struct r600_context*)ctx; 919 struct r600_texture *rdst = (struct r600_texture *)info->dst.resource; 920 921 if (do_hardware_msaa_resolve(ctx, info)) { 922 return; 923 } 924 925 /* Using SDMA for copying to a linear texture in GTT is much faster. 926 * This improves DRI PRIME performance. 927 * 928 * resource_copy_region can't do this yet, because dma_copy calls it 929 * on failure (recursion). 930 */ 931 if (rdst->surface.u.legacy.level[info->dst.level].mode == 932 RADEON_SURF_MODE_LINEAR_ALIGNED && 933 rctx->b.dma_copy && 934 util_can_blit_via_copy_region(info, false)) { 935 rctx->b.dma_copy(ctx, info->dst.resource, info->dst.level, 936 info->dst.box.x, info->dst.box.y, 937 info->dst.box.z, 938 info->src.resource, info->src.level, 939 &info->src.box); 940 return; 941 } 942 943 assert(util_blitter_is_blit_supported(rctx->blitter, info)); 944 945 /* The driver doesn't decompress resources automatically while 946 * u_blitter is rendering. */ 947 if (!r600_decompress_subresource(ctx, info->src.resource, info->src.level, 948 info->src.box.z, 949 info->src.box.z + info->src.box.depth - 1)) { 950 return; /* error */ 951 } 952 953 if (rctx->screen->b.debug_flags & DBG_FORCE_DMA && 954 util_try_blit_via_copy_region(ctx, info)) 955 return; 956 957 r600_blitter_begin(ctx, R600_BLIT | 958 (info->render_condition_enable ? 0 : R600_DISABLE_RENDER_COND)); 959 util_blitter_blit(rctx->blitter, info); 960 r600_blitter_end(ctx); 961} 962 963static void r600_flush_resource(struct pipe_context *ctx, 964 struct pipe_resource *res) 965{ 966 struct r600_texture *rtex = (struct r600_texture*)res; 967 968 assert(res->target != PIPE_BUFFER); 969 970 if (!rtex->is_depth && rtex->cmask.size) { 971 r600_blit_decompress_color(ctx, rtex, 0, res->last_level, 972 0, util_max_layer(res, 0)); 973 } 974} 975 976void r600_init_blit_functions(struct r600_context *rctx) 977{ 978 rctx->b.b.clear = r600_clear; 979 rctx->b.b.clear_render_target = r600_clear_render_target; 980 rctx->b.b.clear_depth_stencil = r600_clear_depth_stencil; 981 rctx->b.b.resource_copy_region = r600_resource_copy_region; 982 rctx->b.b.blit = r600_blit; 983 rctx->b.b.flush_resource = r600_flush_resource; 984 rctx->b.clear_buffer = r600_clear_buffer; 985 rctx->b.blit_decompress_depth = r600_blit_decompress_depth; 986} 987