src/gallium/drivers/radeon/r600_pipe_common.c

   1 /*
   2  * Copyright 2013 Advanced Micro Devices, Inc.
   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  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * 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 NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors: Marek Olšák <maraeo@gmail.com>
  24  *
  25  */
  26
  27 #include "r600_pipe_common.h"
  28 #include "r600_cs.h"
  29 #include "tgsi/tgsi_parse.h"
  30 #include "util/u_draw_quad.h"
  31 #include "util/u_memory.h"
  32 #include "util/u_format_s3tc.h"
  33 #include "util/u_upload_mgr.h"
  34 #include "vl/vl_decoder.h"
  35 #include "vl/vl_video_buffer.h"
  36 #include "radeon/radeon_video.h"
  37 #include <inttypes.h>
  38
  39 #ifndef HAVE_LLVM
  40 #define HAVE_LLVM 0
  41 #endif
  42
  43 /*
  44  * pipe_context
  45  */
  46
  47 void r600_draw_rectangle(struct blitter_context *blitter,
  48                          int x1, int y1, int x2, int y2, float depth,
  49                          enum blitter_attrib_type type,
  50                          const union pipe_color_union *attrib)
  51 {
  52         struct r600_common_context *rctx =
  53                 (struct r600_common_context*)util_blitter_get_pipe(blitter);
  54         struct pipe_viewport_state viewport;
  55         struct pipe_resource *buf = NULL;
  56         unsigned offset = 0;
  57         float *vb;
  58
  59         if (type == UTIL_BLITTER_ATTRIB_TEXCOORD) {
  60                 util_blitter_draw_rectangle(blitter, x1, y1, x2, y2, depth, type, attrib);
  61                 return;
  62         }
  63
  64         /* Some operations (like color resolve on r6xx) don't work
  65          * with the conventional primitive types.
  66          * One that works is PT_RECTLIST, which we use here. */
  67
  68         /* setup viewport */
  69         viewport.scale[0] = 1.0f;
  70         viewport.scale[1] = 1.0f;
  71         viewport.scale[2] = 1.0f;
  72         viewport.translate[0] = 0.0f;
  73         viewport.translate[1] = 0.0f;
  74         viewport.translate[2] = 0.0f;
  75         rctx->b.set_viewport_states(&rctx->b, 0, 1, &viewport);
  76
  77         /* Upload vertices. The hw rectangle has only 3 vertices,
  78          * I guess the 4th one is derived from the first 3.
  79          * The vertex specification should match u_blitter's vertex element state. */
  80         u_upload_alloc(rctx->uploader, 0, sizeof(float) * 24, &offset, &buf, (void**)&vb);
  81         vb[0] = x1;
  82         vb[1] = y1;
  83         vb[2] = depth;
  84         vb[3] = 1;
  85
  86         vb[8] = x1;
  87         vb[9] = y2;
  88         vb[10] = depth;
  89         vb[11] = 1;
  90
  91         vb[16] = x2;
  92         vb[17] = y1;
  93         vb[18] = depth;
  94         vb[19] = 1;
  95
  96         if (attrib) {
  97                 memcpy(vb+4, attrib->f, sizeof(float)*4);
  98                 memcpy(vb+12, attrib->f, sizeof(float)*4);
  99                 memcpy(vb+20, attrib->f, sizeof(float)*4);
 100         }
 101
 102         /* draw */
 103         util_draw_vertex_buffer(&rctx->b, NULL, buf, blitter->vb_slot, offset,
 104                                 R600_PRIM_RECTANGLE_LIST, 3, 2);
 105         pipe_resource_reference(&buf, NULL);
 106 }
 107
 108 void r600_need_dma_space(struct r600_common_context *ctx, unsigned num_dw)
 109 {
 110         /* Flush if there's not enough space. */
 111         if ((num_dw + ctx->rings.dma.cs->cdw) > ctx->rings.dma.cs->max_dw) {
 112                 ctx->rings.dma.flush(ctx, RADEON_FLUSH_ASYNC, NULL);
 113                 assert((num_dw + ctx->rings.dma.cs->cdw) <= ctx->rings.dma.cs->max_dw);
 114         }
 115 }
 116
 117 static void r600_memory_barrier(struct pipe_context *ctx, unsigned flags)
 118 {
 119 }
 120
 121 void r600_preflush_suspend_features(struct r600_common_context *ctx)
 122 {
 123         /* Disable render condition. */
 124         ctx->saved_render_cond = NULL;
 125         ctx->saved_render_cond_cond = FALSE;
 126         ctx->saved_render_cond_mode = 0;
 127         if (ctx->current_render_cond) {
 128                 ctx->saved_render_cond = ctx->current_render_cond;
 129                 ctx->saved_render_cond_cond = ctx->current_render_cond_cond;
 130                 ctx->saved_render_cond_mode = ctx->current_render_cond_mode;
 131                 ctx->b.render_condition(&ctx->b, NULL, FALSE, 0);
 132         }
 133
 134         /* suspend queries */
 135         ctx->queries_suspended_for_flush = false;
 136         if (ctx->num_cs_dw_nontimer_queries_suspend) {
 137                 r600_suspend_nontimer_queries(ctx);
 138                 r600_suspend_timer_queries(ctx);
 139                 ctx->queries_suspended_for_flush = true;
 140         }
 141
 142         ctx->streamout.suspended = false;
 143         if (ctx->streamout.begin_emitted) {
 144                 r600_emit_streamout_end(ctx);
 145                 ctx->streamout.suspended = true;
 146         }
 147 }
 148
 149 void r600_postflush_resume_features(struct r600_common_context *ctx)
 150 {
 151         if (ctx->streamout.suspended) {
 152                 ctx->streamout.append_bitmask = ctx->streamout.enabled_mask;
 153                 r600_streamout_buffers_dirty(ctx);
 154         }
 155
 156         /* resume queries */
 157         if (ctx->queries_suspended_for_flush) {
 158                 r600_resume_nontimer_queries(ctx);
 159                 r600_resume_timer_queries(ctx);
 160         }
 161
 162         /* Re-enable render condition. */
 163         if (ctx->saved_render_cond) {
 164                 ctx->b.render_condition(&ctx->b, ctx->saved_render_cond,
 165                                           ctx->saved_render_cond_cond,
 166                                           ctx->saved_render_cond_mode);
 167         }
 168 }
 169
 170 static void r600_flush_from_st(struct pipe_context *ctx,
 171                                struct pipe_fence_handle **fence,
 172                                unsigned flags)
 173 {
 174         struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 175         unsigned rflags = 0;
 176
 177         if (flags & PIPE_FLUSH_END_OF_FRAME)
 178                 rflags |= RADEON_FLUSH_END_OF_FRAME;
 179
 180         if (rctx->rings.dma.cs) {
 181                 rctx->rings.dma.flush(rctx, rflags, NULL);
 182         }
 183         rctx->rings.gfx.flush(rctx, rflags, fence);
 184 }
 185
 186 static void r600_flush_dma_ring(void *ctx, unsigned flags,
 187                                 struct pipe_fence_handle **fence)
 188 {
 189         struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 190         struct radeon_winsys_cs *cs = rctx->rings.dma.cs;
 191
 192         if (!cs->cdw) {
 193                 return;
 194         }
 195
 196         rctx->rings.dma.flushing = true;
 197         rctx->ws->cs_flush(cs, flags, fence, 0);
 198         rctx->rings.dma.flushing = false;
 199 }
 200
 201 static enum pipe_reset_status r600_get_reset_status(struct pipe_context *ctx)
 202 {
 203         struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 204         unsigned latest = rctx->ws->query_value(rctx->ws,
 205                                                 RADEON_GPU_RESET_COUNTER);
 206
 207         if (rctx->gpu_reset_counter == latest)
 208                 return PIPE_NO_RESET;
 209
 210         rctx->gpu_reset_counter = latest;
 211         return PIPE_UNKNOWN_CONTEXT_RESET;
 212 }
 213
 214 bool r600_common_context_init(struct r600_common_context *rctx,
 215                               struct r600_common_screen *rscreen)
 216 {
 217         util_slab_create(&rctx->pool_transfers,
 218                          sizeof(struct r600_transfer), 64,
 219                          UTIL_SLAB_SINGLETHREADED);
 220
 221         rctx->screen = rscreen;
 222         rctx->ws = rscreen->ws;
 223         rctx->family = rscreen->family;
 224         rctx->chip_class = rscreen->chip_class;
 225
 226         if (rscreen->family == CHIP_HAWAII)
 227                 rctx->max_db = 16;
 228         else if (rscreen->chip_class >= EVERGREEN)
 229                 rctx->max_db = 8;
 230         else
 231                 rctx->max_db = 4;
 232
 233         rctx->b.transfer_map = u_transfer_map_vtbl;
 234         rctx->b.transfer_flush_region = u_default_transfer_flush_region;
 235         rctx->b.transfer_unmap = u_transfer_unmap_vtbl;
 236         rctx->b.transfer_inline_write = u_default_transfer_inline_write;
 237         rctx->b.memory_barrier = r600_memory_barrier;
 238         rctx->b.flush = r600_flush_from_st;
 239
 240         if (rscreen->info.drm_major == 2 && rscreen->info.drm_minor >= 43) {
 241                 rctx->b.get_device_reset_status = r600_get_reset_status;
 242                 rctx->gpu_reset_counter =
 243                         rctx->ws->query_value(rctx->ws,
 244                                               RADEON_GPU_RESET_COUNTER);
 245         }
 246
 247         LIST_INITHEAD(&rctx->texture_buffers);
 248
 249         r600_init_context_texture_functions(rctx);
 250         r600_streamout_init(rctx);
 251         r600_query_init(rctx);
 252         cayman_init_msaa(&rctx->b);
 253
 254         rctx->allocator_so_filled_size = u_suballocator_create(&rctx->b, 4096, 4,
 255                                                                0, PIPE_USAGE_DEFAULT, TRUE);
 256         if (!rctx->allocator_so_filled_size)
 257                 return false;
 258
 259         rctx->uploader = u_upload_create(&rctx->b, 1024 * 1024, 256,
 260                                         PIPE_BIND_INDEX_BUFFER |
 261                                         PIPE_BIND_CONSTANT_BUFFER);
 262         if (!rctx->uploader)
 263                 return false;
 264
 265         if (rscreen->info.r600_has_dma && !(rscreen->debug_flags & DBG_NO_ASYNC_DMA)) {
 266                 rctx->rings.dma.cs = rctx->ws->cs_create(rctx->ws, RING_DMA,
 267                                                          r600_flush_dma_ring,
 268                                                          rctx, NULL);
 269                 rctx->rings.dma.flush = r600_flush_dma_ring;
 270         }
 271
 272         return true;
 273 }
 274
 275 void r600_common_context_cleanup(struct r600_common_context *rctx)
 276 {
 277         if (rctx->rings.gfx.cs) {
 278                 rctx->ws->cs_destroy(rctx->rings.gfx.cs);
 279         }
 280         if (rctx->rings.dma.cs) {
 281                 rctx->ws->cs_destroy(rctx->rings.dma.cs);
 282         }
 283
 284         if (rctx->uploader) {
 285                 u_upload_destroy(rctx->uploader);
 286         }
 287
 288         util_slab_destroy(&rctx->pool_transfers);
 289
 290         if (rctx->allocator_so_filled_size) {
 291                 u_suballocator_destroy(rctx->allocator_so_filled_size);
 292         }
 293 }
 294
 295 void r600_context_add_resource_size(struct pipe_context *ctx, struct pipe_resource *r)
 296 {
 297         struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 298         struct r600_resource *rr = (struct r600_resource *)r;
 299
 300         if (r == NULL) {
 301                 return;
 302         }
 303
 304         /*
 305          * The idea is to compute a gross estimate of memory requirement of
 306          * each draw call. After each draw call, memory will be precisely
 307          * accounted. So the uncertainty is only on the current draw call.
 308          * In practice this gave very good estimate (+/- 10% of the target
 309          * memory limit).
 310          */
 311         if (rr->domains & RADEON_DOMAIN_GTT) {
 312                 rctx->gtt += rr->buf->size;
 313         }
 314         if (rr->domains & RADEON_DOMAIN_VRAM) {
 315                 rctx->vram += rr->buf->size;
 316         }
 317 }
 318
 319 /*
 320  * pipe_screen
 321  */
 322
 323 static const struct debug_named_value common_debug_options[] = {
 324         /* logging */
 325         { "tex", DBG_TEX, "Print texture info" },
 326         { "texmip", DBG_TEXMIP, "Print texture info (mipmapped only)" },
 327         { "compute", DBG_COMPUTE, "Print compute info" },
 328         { "vm", DBG_VM, "Print virtual addresses when creating resources" },
 329         { "trace_cs", DBG_TRACE_CS, "Trace cs and write rlockup_<csid>.c file with faulty cs" },
 330         { "info", DBG_INFO, "Print driver information" },
 331
 332         /* shaders */
 333         { "fs", DBG_FS, "Print fetch shaders" },
 334         { "vs", DBG_VS, "Print vertex shaders" },
 335         { "gs", DBG_GS, "Print geometry shaders" },
 336         { "ps", DBG_PS, "Print pixel shaders" },
 337         { "cs", DBG_CS, "Print compute shaders" },
 338         { "tcs", DBG_TCS, "Print tessellation control shaders" },
 339         { "tes", DBG_TES, "Print tessellation evaluation shaders" },
 340         { "noir", DBG_NO_IR, "Don't print the LLVM IR"},
 341         { "notgsi", DBG_NO_TGSI, "Don't print the TGSI"},
 342         { "noasm", DBG_NO_ASM, "Don't print disassembled shaders"},
 343
 344         /* features */
 345         { "nodma", DBG_NO_ASYNC_DMA, "Disable asynchronous DMA" },
 346         { "nohyperz", DBG_NO_HYPERZ, "Disable Hyper-Z" },
 347         /* GL uses the word INVALIDATE, gallium uses the word DISCARD */
 348         { "noinvalrange", DBG_NO_DISCARD_RANGE, "Disable handling of INVALIDATE_RANGE map flags" },
 349         { "no2d", DBG_NO_2D_TILING, "Disable 2D tiling" },
 350         { "notiling", DBG_NO_TILING, "Disable tiling" },
 351         { "switch_on_eop", DBG_SWITCH_ON_EOP, "Program WD/IA to switch on end-of-packet." },
 352         { "forcedma", DBG_FORCE_DMA, "Use asynchronous DMA for all operations when possible." },
 353         { "precompile", DBG_PRECOMPILE, "Compile one shader variant at shader creation." },
 354
 355         DEBUG_NAMED_VALUE_END /* must be last */
 356 };
 357
 358 static const char* r600_get_vendor(struct pipe_screen* pscreen)
 359 {
 360         return "X.Org";
 361 }
 362
 363 static const char* r600_get_device_vendor(struct pipe_screen* pscreen)
 364 {
 365         return "AMD";
 366 }
 367
 368 static const char* r600_get_chip_name(struct r600_common_screen *rscreen)
 369 {
 370         switch (rscreen->info.family) {
 371         case CHIP_R600: return "AMD R600";
 372         case CHIP_RV610: return "AMD RV610";
 373         case CHIP_RV630: return "AMD RV630";
 374         case CHIP_RV670: return "AMD RV670";
 375         case CHIP_RV620: return "AMD RV620";
 376         case CHIP_RV635: return "AMD RV635";
 377         case CHIP_RS780: return "AMD RS780";
 378         case CHIP_RS880: return "AMD RS880";
 379         case CHIP_RV770: return "AMD RV770";
 380         case CHIP_RV730: return "AMD RV730";
 381         case CHIP_RV710: return "AMD RV710";
 382         case CHIP_RV740: return "AMD RV740";
 383         case CHIP_CEDAR: return "AMD CEDAR";
 384         case CHIP_REDWOOD: return "AMD REDWOOD";
 385         case CHIP_JUNIPER: return "AMD JUNIPER";
 386         case CHIP_CYPRESS: return "AMD CYPRESS";
 387         case CHIP_HEMLOCK: return "AMD HEMLOCK";
 388         case CHIP_PALM: return "AMD PALM";
 389         case CHIP_SUMO: return "AMD SUMO";
 390         case CHIP_SUMO2: return "AMD SUMO2";
 391         case CHIP_BARTS: return "AMD BARTS";
 392         case CHIP_TURKS: return "AMD TURKS";
 393         case CHIP_CAICOS: return "AMD CAICOS";
 394         case CHIP_CAYMAN: return "AMD CAYMAN";
 395         case CHIP_ARUBA: return "AMD ARUBA";
 396         case CHIP_TAHITI: return "AMD TAHITI";
 397         case CHIP_PITCAIRN: return "AMD PITCAIRN";
 398         case CHIP_VERDE: return "AMD CAPE VERDE";
 399         case CHIP_OLAND: return "AMD OLAND";
 400         case CHIP_HAINAN: return "AMD HAINAN";
 401         case CHIP_BONAIRE: return "AMD BONAIRE";
 402         case CHIP_KAVERI: return "AMD KAVERI";
 403         case CHIP_KABINI: return "AMD KABINI";
 404         case CHIP_HAWAII: return "AMD HAWAII";
 405         case CHIP_MULLINS: return "AMD MULLINS";
 406         default: return "AMD unknown";
 407         }
 408 }
 409
 410 static const char* r600_get_name(struct pipe_screen* pscreen)
 411 {
 412         struct r600_common_screen *rscreen = (struct r600_common_screen*)pscreen;
 413
 414         return rscreen->renderer_string;
 415 }
 416
 417 static float r600_get_paramf(struct pipe_screen* pscreen,
 418                              enum pipe_capf param)
 419 {
 420         struct r600_common_screen *rscreen = (struct r600_common_screen *)pscreen;
 421
 422         switch (param) {
 423         case PIPE_CAPF_MAX_LINE_WIDTH:
 424         case PIPE_CAPF_MAX_LINE_WIDTH_AA:
 425         case PIPE_CAPF_MAX_POINT_WIDTH:
 426         case PIPE_CAPF_MAX_POINT_WIDTH_AA:
 427                 if (rscreen->family >= CHIP_CEDAR)
 428                         return 16384.0f;
 429                 else
 430                         return 8192.0f;
 431         case PIPE_CAPF_MAX_TEXTURE_ANISOTROPY:
 432                 return 16.0f;
 433         case PIPE_CAPF_MAX_TEXTURE_LOD_BIAS:
 434                 return 16.0f;
 435         case PIPE_CAPF_GUARD_BAND_LEFT:
 436         case PIPE_CAPF_GUARD_BAND_TOP:
 437         case PIPE_CAPF_GUARD_BAND_RIGHT:
 438         case PIPE_CAPF_GUARD_BAND_BOTTOM:
 439                 return 0.0f;
 440         }
 441         return 0.0f;
 442 }
 443
 444 static int r600_get_video_param(struct pipe_screen *screen,
 445                                 enum pipe_video_profile profile,
 446                                 enum pipe_video_entrypoint entrypoint,
 447                                 enum pipe_video_cap param)
 448 {
 449         switch (param) {
 450         case PIPE_VIDEO_CAP_SUPPORTED:
 451                 return vl_profile_supported(screen, profile, entrypoint);
 452         case PIPE_VIDEO_CAP_NPOT_TEXTURES:
 453                 return 1;
 454         case PIPE_VIDEO_CAP_MAX_WIDTH:
 455         case PIPE_VIDEO_CAP_MAX_HEIGHT:
 456                 return vl_video_buffer_max_size(screen);
 457         case PIPE_VIDEO_CAP_PREFERED_FORMAT:
 458                 return PIPE_FORMAT_NV12;
 459         case PIPE_VIDEO_CAP_PREFERS_INTERLACED:
 460                 return false;
 461         case PIPE_VIDEO_CAP_SUPPORTS_INTERLACED:
 462                 return false;
 463         case PIPE_VIDEO_CAP_SUPPORTS_PROGRESSIVE:
 464                 return true;
 465         case PIPE_VIDEO_CAP_MAX_LEVEL:
 466                 return vl_level_supported(screen, profile);
 467         default:
 468                 return 0;
 469         }
 470 }
 471
 472 const char *r600_get_llvm_processor_name(enum radeon_family family)
 473 {
 474         switch (family) {
 475         case CHIP_R600:
 476         case CHIP_RV630:
 477         case CHIP_RV635:
 478         case CHIP_RV670:
 479                 return "r600";
 480         case CHIP_RV610:
 481         case CHIP_RV620:
 482         case CHIP_RS780:
 483         case CHIP_RS880:
 484                 return "rs880";
 485         case CHIP_RV710:
 486                 return "rv710";
 487         case CHIP_RV730:
 488                 return "rv730";
 489         case CHIP_RV740:
 490         case CHIP_RV770:
 491                 return "rv770";
 492         case CHIP_PALM:
 493         case CHIP_CEDAR:
 494                 return "cedar";
 495         case CHIP_SUMO:
 496         case CHIP_SUMO2:
 497                 return "sumo";
 498         case CHIP_REDWOOD:
 499                 return "redwood";
 500         case CHIP_JUNIPER:
 501                 return "juniper";
 502         case CHIP_HEMLOCK:
 503         case CHIP_CYPRESS:
 504                 return "cypress";
 505         case CHIP_BARTS:
 506                 return "barts";
 507         case CHIP_TURKS:
 508                 return "turks";
 509         case CHIP_CAICOS:
 510                 return "caicos";
 511         case CHIP_CAYMAN:
 512         case CHIP_ARUBA:
 513                 return "cayman";
 514
 515         case CHIP_TAHITI: return "tahiti";
 516         case CHIP_PITCAIRN: return "pitcairn";
 517         case CHIP_VERDE: return "verde";
 518         case CHIP_OLAND: return "oland";
 519         case CHIP_HAINAN: return "hainan";
 520         case CHIP_BONAIRE: return "bonaire";
 521         case CHIP_KABINI: return "kabini";
 522         case CHIP_KAVERI: return "kaveri";
 523         case CHIP_HAWAII: return "hawaii";
 524         case CHIP_MULLINS:
 525 #if HAVE_LLVM >= 0x0305
 526                 return "mullins";
 527 #else
 528                 return "kabini";
 529 #endif
 530         default: return "";
 531         }
 532 }
 533
 534 static int r600_get_compute_param(struct pipe_screen *screen,
 535         enum pipe_compute_cap param,
 536         void *ret)
 537 {
 538         struct r600_common_screen *rscreen = (struct r600_common_screen *)screen;
 539
 540         //TODO: select these params by asic
 541         switch (param) {
 542         case PIPE_COMPUTE_CAP_IR_TARGET: {
 543                 const char *gpu;
 544                 const char *triple;
 545                 if (rscreen->family <= CHIP_ARUBA || HAVE_LLVM < 0x0306) {
 546                         triple = "r600--";
 547                 } else {
 548                         triple = "amdgcn--";
 549                 }
 550                 switch(rscreen->family) {
 551                 /* Clang < 3.6 is missing Hainan in its list of
 552                  * GPUs, so we need to use the name of a similar GPU.
 553                  */
 554 #if HAVE_LLVM < 0x0306
 555                 case CHIP_HAINAN:
 556                         gpu = "oland";
 557                         break;
 558 #endif
 559                 default:
 560                         gpu = r600_get_llvm_processor_name(rscreen->family);
 561                         break;
 562                 }
 563                 if (ret) {
 564                         sprintf(ret, "%s-%s", gpu, triple);
 565                 }
 566                 /* +2 for dash and terminating NIL byte */
 567                 return (strlen(triple) + strlen(gpu) + 2) * sizeof(char);
 568         }
 569         case PIPE_COMPUTE_CAP_GRID_DIMENSION:
 570                 if (ret) {
 571                         uint64_t *grid_dimension = ret;
 572                         grid_dimension[0] = 3;
 573                 }
 574                 return 1 * sizeof(uint64_t);
 575
 576         case PIPE_COMPUTE_CAP_MAX_GRID_SIZE:
 577                 if (ret) {
 578                         uint64_t *grid_size = ret;
 579                         grid_size[0] = 65535;
 580                         grid_size[1] = 65535;
 581                         grid_size[2] = 1;
 582                 }
 583                 return 3 * sizeof(uint64_t) ;
 584
 585         case PIPE_COMPUTE_CAP_MAX_BLOCK_SIZE:
 586                 if (ret) {
 587                         uint64_t *block_size = ret;
 588                         block_size[0] = 256;
 589                         block_size[1] = 256;
 590                         block_size[2] = 256;
 591                 }
 592                 return 3 * sizeof(uint64_t);
 593
 594         case PIPE_COMPUTE_CAP_MAX_THREADS_PER_BLOCK:
 595                 if (ret) {
 596                         uint64_t *max_threads_per_block = ret;
 597                         *max_threads_per_block = 256;
 598                 }
 599                 return sizeof(uint64_t);
 600
 601         case PIPE_COMPUTE_CAP_MAX_GLOBAL_SIZE:
 602                 if (ret) {
 603                         uint64_t *max_global_size = ret;
 604                         uint64_t max_mem_alloc_size;
 605
 606                         r600_get_compute_param(screen,
 607                                 PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE,
 608                                 &max_mem_alloc_size);
 609
 610                         /* In OpenCL, the MAX_MEM_ALLOC_SIZE must be at least
 611                          * 1/4 of the MAX_GLOBAL_SIZE.  Since the
 612                          * MAX_MEM_ALLOC_SIZE is fixed for older kernels,
 613                          * make sure we never report more than
 614                          * 4 * MAX_MEM_ALLOC_SIZE.
 615                          */
 616                         *max_global_size = MIN2(4 * max_mem_alloc_size,
 617                                 rscreen->info.gart_size +
 618                                 rscreen->info.vram_size);
 619                 }
 620                 return sizeof(uint64_t);
 621
 622         case PIPE_COMPUTE_CAP_MAX_LOCAL_SIZE:
 623                 if (ret) {
 624                         uint64_t *max_local_size = ret;
 625                         /* Value reported by the closed source driver. */
 626                         *max_local_size = 32768;
 627                 }
 628                 return sizeof(uint64_t);
 629
 630         case PIPE_COMPUTE_CAP_MAX_INPUT_SIZE:
 631                 if (ret) {
 632                         uint64_t *max_input_size = ret;
 633                         /* Value reported by the closed source driver. */
 634                         *max_input_size = 1024;
 635                 }
 636                 return sizeof(uint64_t);
 637
 638         case PIPE_COMPUTE_CAP_MAX_MEM_ALLOC_SIZE:
 639                 if (ret) {
 640                         uint64_t *max_mem_alloc_size = ret;
 641
 642                         /* XXX: The limit in older kernels is 256 MB.  We
 643                          * should add a query here for newer kernels.
 644                          */
 645                         *max_mem_alloc_size = 256 * 1024 * 1024;
 646                 }
 647                 return sizeof(uint64_t);
 648
 649         case PIPE_COMPUTE_CAP_MAX_CLOCK_FREQUENCY:
 650                 if (ret) {
 651                         uint32_t *max_clock_frequency = ret;
 652                         *max_clock_frequency = rscreen->info.max_sclk;
 653                 }
 654                 return sizeof(uint32_t);
 655
 656         case PIPE_COMPUTE_CAP_MAX_COMPUTE_UNITS:
 657                 if (ret) {
 658                         uint32_t *max_compute_units = ret;
 659                         *max_compute_units = rscreen->info.max_compute_units;
 660                 }
 661                 return sizeof(uint32_t);
 662
 663         case PIPE_COMPUTE_CAP_IMAGES_SUPPORTED:
 664                 if (ret) {
 665                         uint32_t *images_supported = ret;
 666                         *images_supported = 0;
 667                 }
 668                 return sizeof(uint32_t);
 669         case PIPE_COMPUTE_CAP_MAX_PRIVATE_SIZE:
 670                 break; /* unused */
 671         case PIPE_COMPUTE_CAP_SUBGROUP_SIZE:
 672                 if (ret) {
 673                         uint32_t *subgroup_size = ret;
 674                         *subgroup_size = r600_wavefront_size(rscreen->family);
 675                 }
 676                 return sizeof(uint32_t);
 677         }
 678
 679         fprintf(stderr, "unknown PIPE_COMPUTE_CAP %d\n", param);
 680         return 0;
 681 }
 682
 683 static uint64_t r600_get_timestamp(struct pipe_screen *screen)
 684 {
 685         struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
 686
 687         return 1000000 * rscreen->ws->query_value(rscreen->ws, RADEON_TIMESTAMP) /
 688                         rscreen->info.r600_clock_crystal_freq;
 689 }
 690
 691 static int r600_get_driver_query_info(struct pipe_screen *screen,
 692                                       unsigned index,
 693                                       struct pipe_driver_query_info *info)
 694 {
 695         struct r600_common_screen *rscreen = (struct r600_common_screen*)screen;
 696         struct pipe_driver_query_info list[] = {
 697                 {"num-compilations", R600_QUERY_NUM_COMPILATIONS, {0}, PIPE_DRIVER_QUERY_TYPE_UINT64,
 698                  PIPE_DRIVER_QUERY_RESULT_TYPE_CUMULATIVE},
 699                 {"num-shaders-created", R600_QUERY_NUM_SHADERS_CREATED, {0}, PIPE_DRIVER_QUERY_TYPE_UINT64,
 700                  PIPE_DRIVER_QUERY_RESULT_TYPE_CUMULATIVE},
 701                 {"draw-calls", R600_QUERY_DRAW_CALLS, {0}},
 702                 {"requested-VRAM", R600_QUERY_REQUESTED_VRAM, {rscreen->info.vram_size}, PIPE_DRIVER_QUERY_TYPE_BYTES},
 703                 {"requested-GTT", R600_QUERY_REQUESTED_GTT, {rscreen->info.gart_size}, PIPE_DRIVER_QUERY_TYPE_BYTES},
 704                 {"buffer-wait-time", R600_QUERY_BUFFER_WAIT_TIME, {0}, PIPE_DRIVER_QUERY_TYPE_MICROSECONDS,
 705                  PIPE_DRIVER_QUERY_RESULT_TYPE_CUMULATIVE},
 706                 {"num-cs-flushes", R600_QUERY_NUM_CS_FLUSHES, {0}},
 707                 {"num-bytes-moved", R600_QUERY_NUM_BYTES_MOVED, {0}, PIPE_DRIVER_QUERY_TYPE_BYTES,
 708                  PIPE_DRIVER_QUERY_RESULT_TYPE_CUMULATIVE},
 709                 {"VRAM-usage", R600_QUERY_VRAM_USAGE, {rscreen->info.vram_size}, PIPE_DRIVER_QUERY_TYPE_BYTES},
 710                 {"GTT-usage", R600_QUERY_GTT_USAGE, {rscreen->info.gart_size}, PIPE_DRIVER_QUERY_TYPE_BYTES},
 711                 {"temperature", R600_QUERY_GPU_TEMPERATURE, {100}},
 712                 {"shader-clock", R600_QUERY_CURRENT_GPU_SCLK, {0}, PIPE_DRIVER_QUERY_TYPE_HZ},
 713                 {"memory-clock", R600_QUERY_CURRENT_GPU_MCLK, {0}, PIPE_DRIVER_QUERY_TYPE_HZ},
 714                 {"GPU-load", R600_QUERY_GPU_LOAD, {100}}
 715         };
 716         unsigned num_queries;
 717
 718         if (rscreen->info.drm_major == 2 && rscreen->info.drm_minor >= 42)
 719                 num_queries = Elements(list);
 720         else
 721                 num_queries = 9;
 722
 723         if (!info)
 724                 return num_queries;
 725
 726         if (index >= num_queries)
 727                 return 0;
 728
 729         *info = list[index];
 730         return 1;
 731 }
 732
 733 static void r600_fence_reference(struct pipe_screen *screen,
 734                                  struct pipe_fence_handle **ptr,
 735                                  struct pipe_fence_handle *fence)
 736 {
 737         struct radeon_winsys *rws = ((struct r600_common_screen*)screen)->ws;
 738
 739         rws->fence_reference(ptr, fence);
 740 }
 741
 742 static boolean r600_fence_finish(struct pipe_screen *screen,
 743                                  struct pipe_fence_handle *fence,
 744                                  uint64_t timeout)
 745 {
 746         struct radeon_winsys *rws = ((struct r600_common_screen*)screen)->ws;
 747
 748         return rws->fence_wait(rws, fence, timeout);
 749 }
 750
 751 static bool r600_interpret_tiling(struct r600_common_screen *rscreen,
 752                                   uint32_t tiling_config)
 753 {
 754         switch ((tiling_config & 0xe) >> 1) {
 755         case 0:
 756                 rscreen->tiling_info.num_channels = 1;
 757                 break;
 758         case 1:
 759                 rscreen->tiling_info.num_channels = 2;
 760                 break;
 761         case 2:
 762                 rscreen->tiling_info.num_channels = 4;
 763                 break;
 764         case 3:
 765                 rscreen->tiling_info.num_channels = 8;
 766                 break;
 767         default:
 768                 return false;
 769         }
 770
 771         switch ((tiling_config & 0x30) >> 4) {
 772         case 0:
 773                 rscreen->tiling_info.num_banks = 4;
 774                 break;
 775         case 1:
 776                 rscreen->tiling_info.num_banks = 8;
 777                 break;
 778         default:
 779                 return false;
 780
 781         }
 782         switch ((tiling_config & 0xc0) >> 6) {
 783         case 0:
 784                 rscreen->tiling_info.group_bytes = 256;
 785                 break;
 786         case 1:
 787                 rscreen->tiling_info.group_bytes = 512;
 788                 break;
 789         default:
 790                 return false;
 791         }
 792         return true;
 793 }
 794
 795 static bool evergreen_interpret_tiling(struct r600_common_screen *rscreen,
 796                                        uint32_t tiling_config)
 797 {
 798         switch (tiling_config & 0xf) {
 799         case 0:
 800                 rscreen->tiling_info.num_channels = 1;
 801                 break;
 802         case 1:
 803                 rscreen->tiling_info.num_channels = 2;
 804                 break;
 805         case 2:
 806                 rscreen->tiling_info.num_channels = 4;
 807                 break;
 808         case 3:
 809                 rscreen->tiling_info.num_channels = 8;
 810                 break;
 811         default:
 812                 return false;
 813         }
 814
 815         switch ((tiling_config & 0xf0) >> 4) {
 816         case 0:
 817                 rscreen->tiling_info.num_banks = 4;
 818                 break;
 819         case 1:
 820                 rscreen->tiling_info.num_banks = 8;
 821                 break;
 822         case 2:
 823                 rscreen->tiling_info.num_banks = 16;
 824                 break;
 825         default:
 826                 return false;
 827         }
 828
 829         switch ((tiling_config & 0xf00) >> 8) {
 830         case 0:
 831                 rscreen->tiling_info.group_bytes = 256;
 832                 break;
 833         case 1:
 834                 rscreen->tiling_info.group_bytes = 512;
 835                 break;
 836         default:
 837                 return false;
 838         }
 839         return true;
 840 }
 841
 842 static bool r600_init_tiling(struct r600_common_screen *rscreen)
 843 {
 844         uint32_t tiling_config = rscreen->info.r600_tiling_config;
 845
 846         /* set default group bytes, overridden by tiling info ioctl */
 847         if (rscreen->chip_class <= R700) {
 848                 rscreen->tiling_info.group_bytes = 256;
 849         } else {
 850                 rscreen->tiling_info.group_bytes = 512;
 851         }
 852
 853         if (!tiling_config)
 854                 return true;
 855
 856         if (rscreen->chip_class <= R700) {
 857                 return r600_interpret_tiling(rscreen, tiling_config);
 858         } else {
 859                 return evergreen_interpret_tiling(rscreen, tiling_config);
 860         }
 861 }
 862
 863 struct pipe_resource *r600_resource_create_common(struct pipe_screen *screen,
 864                                                   const struct pipe_resource *templ)
 865 {
 866         if (templ->target == PIPE_BUFFER) {
 867                 return r600_buffer_create(screen, templ, 4096);
 868         } else {
 869                 return r600_texture_create(screen, templ);
 870         }
 871 }
 872
 873 bool r600_common_screen_init(struct r600_common_screen *rscreen,
 874                              struct radeon_winsys *ws)
 875 {
 876         char llvm_string[32] = {};
 877
 878         ws->query_info(ws, &rscreen->info);
 879
 880         if (HAVE_LLVM)
 881                 snprintf(llvm_string, sizeof(llvm_string),
 882                          ", LLVM %i.%i.%i", (HAVE_LLVM >> 8) & 0xff,
 883                          HAVE_LLVM & 0xff, MESA_LLVM_VERSION_PATCH);
 884
 885         snprintf(rscreen->renderer_string, sizeof(rscreen->renderer_string),
 886                  "%s (DRM %i.%i.%i%s)",
 887                  r600_get_chip_name(rscreen), rscreen->info.drm_major,
 888                  rscreen->info.drm_minor, rscreen->info.drm_patchlevel,
 889                  llvm_string);
 890
 891         rscreen->b.get_name = r600_get_name;
 892         rscreen->b.get_vendor = r600_get_vendor;
 893         rscreen->b.get_device_vendor = r600_get_device_vendor;
 894         rscreen->b.get_compute_param = r600_get_compute_param;
 895         rscreen->b.get_paramf = r600_get_paramf;
 896         rscreen->b.get_driver_query_info = r600_get_driver_query_info;
 897         rscreen->b.get_timestamp = r600_get_timestamp;
 898         rscreen->b.fence_finish = r600_fence_finish;
 899         rscreen->b.fence_reference = r600_fence_reference;
 900         rscreen->b.resource_destroy = u_resource_destroy_vtbl;
 901         rscreen->b.resource_from_user_memory = r600_buffer_from_user_memory;
 902
 903         if (rscreen->info.has_uvd) {
 904                 rscreen->b.get_video_param = rvid_get_video_param;
 905                 rscreen->b.is_video_format_supported = rvid_is_format_supported;
 906         } else {
 907                 rscreen->b.get_video_param = r600_get_video_param;
 908                 rscreen->b.is_video_format_supported = vl_video_buffer_is_format_supported;
 909         }
 910
 911         r600_init_screen_texture_functions(rscreen);
 912
 913         rscreen->ws = ws;
 914         rscreen->family = rscreen->info.family;
 915         rscreen->chip_class = rscreen->info.chip_class;
 916         rscreen->debug_flags = debug_get_flags_option("R600_DEBUG", common_debug_options, 0);
 917
 918         if (!r600_init_tiling(rscreen)) {
 919                 return false;
 920         }
 921         util_format_s3tc_init();
 922         pipe_mutex_init(rscreen->aux_context_lock);
 923         pipe_mutex_init(rscreen->gpu_load_mutex);
 924
 925         if (rscreen->info.drm_minor >= 28 && (rscreen->debug_flags & DBG_TRACE_CS)) {
 926                 rscreen->trace_bo = (struct r600_resource*)pipe_buffer_create(&rscreen->b,
 927                                                                                 PIPE_BIND_CUSTOM,
 928                                                                                 PIPE_USAGE_STAGING,
 929                                                                                 4096);
 930                 if (rscreen->trace_bo) {
 931                         rscreen->trace_ptr = rscreen->ws->buffer_map(rscreen->trace_bo->cs_buf, NULL,
 932                                                                         PIPE_TRANSFER_UNSYNCHRONIZED);
 933                 }
 934         }
 935
 936         if (rscreen->debug_flags & DBG_INFO) {
 937                 printf("pci_id = 0x%x\n", rscreen->info.pci_id);
 938                 printf("family = %i\n", rscreen->info.family);
 939                 printf("chip_class = %i\n", rscreen->info.chip_class);
 940                 printf("gart_size = %i MB\n", (int)(rscreen->info.gart_size >> 20));
 941                 printf("vram_size = %i MB\n", (int)(rscreen->info.vram_size >> 20));
 942                 printf("max_sclk = %i\n", rscreen->info.max_sclk);
 943                 printf("max_compute_units = %i\n", rscreen->info.max_compute_units);
 944                 printf("max_se = %i\n", rscreen->info.max_se);
 945                 printf("max_sh_per_se = %i\n", rscreen->info.max_sh_per_se);
 946                 printf("drm = %i.%i.%i\n", rscreen->info.drm_major,
 947                        rscreen->info.drm_minor, rscreen->info.drm_patchlevel);
 948                 printf("has_uvd = %i\n", rscreen->info.has_uvd);
 949                 printf("vce_fw_version = %i\n", rscreen->info.vce_fw_version);
 950                 printf("r600_num_backends = %i\n", rscreen->info.r600_num_backends);
 951                 printf("r600_clock_crystal_freq = %i\n", rscreen->info.r600_clock_crystal_freq);
 952                 printf("r600_tiling_config = 0x%x\n", rscreen->info.r600_tiling_config);
 953                 printf("r600_num_tile_pipes = %i\n", rscreen->info.r600_num_tile_pipes);
 954                 printf("r600_max_pipes = %i\n", rscreen->info.r600_max_pipes);
 955                 printf("r600_virtual_address = %i\n", rscreen->info.r600_virtual_address);
 956                 printf("r600_has_dma = %i\n", rscreen->info.r600_has_dma);
 957                 printf("r600_backend_map = %i\n", rscreen->info.r600_backend_map);
 958                 printf("r600_backend_map_valid = %i\n", rscreen->info.r600_backend_map_valid);
 959                 printf("si_tile_mode_array_valid = %i\n", rscreen->info.si_tile_mode_array_valid);
 960                 printf("cik_macrotile_mode_array_valid = %i\n", rscreen->info.cik_macrotile_mode_array_valid);
 961         }
 962         return true;
 963 }
 964
 965 void r600_destroy_common_screen(struct r600_common_screen *rscreen)
 966 {
 967         r600_gpu_load_kill_thread(rscreen);
 968
 969         pipe_mutex_destroy(rscreen->gpu_load_mutex);
 970         pipe_mutex_destroy(rscreen->aux_context_lock);
 971         rscreen->aux_context->destroy(rscreen->aux_context);
 972
 973         if (rscreen->trace_bo)
 974                 pipe_resource_reference((struct pipe_resource**)&rscreen->trace_bo, NULL);
 975
 976         rscreen->ws->destroy(rscreen->ws);
 977         FREE(rscreen);
 978 }
 979
 980 bool r600_can_dump_shader(struct r600_common_screen *rscreen,
 981                           const struct tgsi_token *tokens)
 982 {
 983         /* Compute shader don't have tgsi_tokens */
 984         if (!tokens)
 985                 return (rscreen->debug_flags & DBG_CS) != 0;
 986
 987         switch (tgsi_get_processor_type(tokens)) {
 988         case TGSI_PROCESSOR_VERTEX:
 989                 return (rscreen->debug_flags & DBG_VS) != 0;
 990         case TGSI_PROCESSOR_TESS_CTRL:
 991                 return (rscreen->debug_flags & DBG_TCS) != 0;
 992         case TGSI_PROCESSOR_TESS_EVAL:
 993                 return (rscreen->debug_flags & DBG_TES) != 0;
 994         case TGSI_PROCESSOR_GEOMETRY:
 995                 return (rscreen->debug_flags & DBG_GS) != 0;
 996         case TGSI_PROCESSOR_FRAGMENT:
 997                 return (rscreen->debug_flags & DBG_PS) != 0;
 998         case TGSI_PROCESSOR_COMPUTE:
 999                 return (rscreen->debug_flags & DBG_CS) != 0;
1000         default:
1001                 return false;
1002         }
1003 }
1004
1005 void r600_screen_clear_buffer(struct r600_common_screen *rscreen, struct pipe_resource *dst,
1006                               unsigned offset, unsigned size, unsigned value,
1007                               bool is_framebuffer)
1008 {
1009         struct r600_common_context *rctx = (struct r600_common_context*)rscreen->aux_context;
1010
1011         pipe_mutex_lock(rscreen->aux_context_lock);
1012         rctx->clear_buffer(&rctx->b, dst, offset, size, value, is_framebuffer);
1013         rscreen->aux_context->flush(rscreen->aux_context, NULL, 0);
1014         pipe_mutex_unlock(rscreen->aux_context_lock);
1015 }