src/gallium/drivers/radeon/r600_query.c

   1 /*
   2  * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
   3  * Copyright 2014 Marek Olšák <marek.olsak@amd.com>
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * on the rights to use, copy, modify, merge, publish, distribute, sub
   9  * license, and/or sell copies of the Software, and to permit persons to whom
  10  * the Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
  19  * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
  20  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
  21  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  22  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  23  */
  24
  25 #include "r600_cs.h"
  26 #include "util/u_memory.h"
  27
  28
  29 struct r600_query_buffer {
  30         /* The buffer where query results are stored. */
  31         struct r600_resource                    *buf;
  32         /* Offset of the next free result after current query data */
  33         unsigned                                results_end;
  34         /* If a query buffer is full, a new buffer is created and the old one
  35          * is put in here. When we calculate the result, we sum up the samples
  36          * from all buffers. */
  37         struct r600_query_buffer                *previous;
  38 };
  39
  40 struct r600_query {
  41         /* The query buffer and how many results are in it. */
  42         struct r600_query_buffer                buffer;
  43         /* The type of query */
  44         unsigned                                type;
  45         /* Size of the result in memory for both begin_query and end_query,
  46          * this can be one or two numbers, or it could even be a size of a structure. */
  47         unsigned                                result_size;
  48         /* The number of dwords for begin_query or end_query. */
  49         unsigned                                num_cs_dw;
  50         /* linked list of queries */
  51         struct list_head                        list;
  52         /* for custom non-GPU queries */
  53         uint64_t begin_result;
  54         uint64_t end_result;
  55 };
  56
  57
  58 static bool r600_is_timer_query(unsigned type)
  59 {
  60         return type == PIPE_QUERY_TIME_ELAPSED ||
  61                type == PIPE_QUERY_TIMESTAMP ||
  62                type == PIPE_QUERY_TIMESTAMP_DISJOINT;
  63 }
  64
  65 static bool r600_query_needs_begin(unsigned type)
  66 {
  67         return type != PIPE_QUERY_GPU_FINISHED &&
  68                type != PIPE_QUERY_TIMESTAMP;
  69 }
  70
  71 static struct r600_resource *r600_new_query_buffer(struct r600_common_context *ctx, unsigned type)
  72 {
  73         unsigned j, i, num_results, buf_size = 4096;
  74         uint32_t *results;
  75
  76         /* Non-GPU queries. */
  77         switch (type) {
  78         case R600_QUERY_DRAW_CALLS:
  79         case R600_QUERY_REQUESTED_VRAM:
  80         case R600_QUERY_REQUESTED_GTT:
  81         case R600_QUERY_BUFFER_WAIT_TIME:
  82                 return NULL;
  83         }
  84
  85         /* Queries are normally read by the CPU after
  86          * being written by the gpu, hence staging is probably a good
  87          * usage pattern.
  88          */
  89         struct r600_resource *buf = (struct r600_resource*)
  90                 pipe_buffer_create(ctx->b.screen, PIPE_BIND_CUSTOM,
  91                                    PIPE_USAGE_STAGING, buf_size);
  92
  93         switch (type) {
  94         case PIPE_QUERY_OCCLUSION_COUNTER:
  95         case PIPE_QUERY_OCCLUSION_PREDICATE:
  96                 results = r600_buffer_map_sync_with_rings(ctx, buf, PIPE_TRANSFER_WRITE);
  97                 memset(results, 0, buf_size);
  98
  99                 /* Set top bits for unused backends. */
 100                 num_results = buf_size / (16 * ctx->max_db);
 101                 for (j = 0; j < num_results; j++) {
 102                         for (i = 0; i < ctx->max_db; i++) {
 103                                 if (!(ctx->backend_mask & (1<<i))) {
 104                                         results[(i * 4)+1] = 0x80000000;
 105                                         results[(i * 4)+3] = 0x80000000;
 106                                 }
 107                         }
 108                         results += 4 * ctx->max_db;
 109                 }
 110                 ctx->ws->buffer_unmap(buf->cs_buf);
 111                 break;
 112         case PIPE_QUERY_TIME_ELAPSED:
 113         case PIPE_QUERY_TIMESTAMP:
 114                 break;
 115         case PIPE_QUERY_PRIMITIVES_EMITTED:
 116         case PIPE_QUERY_PRIMITIVES_GENERATED:
 117         case PIPE_QUERY_SO_STATISTICS:
 118         case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 119         case PIPE_QUERY_PIPELINE_STATISTICS:
 120                 results = r600_buffer_map_sync_with_rings(ctx, buf, PIPE_TRANSFER_WRITE);
 121                 memset(results, 0, buf_size);
 122                 ctx->ws->buffer_unmap(buf->cs_buf);
 123                 break;
 124         default:
 125                 assert(0);
 126         }
 127         return buf;
 128 }
 129
 130 static void r600_update_occlusion_query_state(struct r600_common_context *rctx,
 131                                               unsigned type, int diff)
 132 {
 133         if (type == PIPE_QUERY_OCCLUSION_COUNTER ||
 134             type == PIPE_QUERY_OCCLUSION_PREDICATE) {
 135                 bool old_enable = rctx->num_occlusion_queries != 0;
 136                 bool enable;
 137
 138                 rctx->num_occlusion_queries += diff;
 139                 assert(rctx->num_occlusion_queries >= 0);
 140
 141                 enable = rctx->num_occlusion_queries != 0;
 142
 143                 if (enable != old_enable) {
 144                         rctx->set_occlusion_query_state(&rctx->b, enable);
 145                 }
 146         }
 147 }
 148
 149 static void r600_emit_query_begin(struct r600_common_context *ctx, struct r600_query *query)
 150 {
 151         struct radeon_winsys_cs *cs = ctx->rings.gfx.cs;
 152         uint64_t va;
 153
 154         r600_update_occlusion_query_state(ctx, query->type, 1);
 155         ctx->need_gfx_cs_space(&ctx->b, query->num_cs_dw * 2, TRUE);
 156
 157         /* Get a new query buffer if needed. */
 158         if (query->buffer.results_end + query->result_size > query->buffer.buf->b.b.width0) {
 159                 struct r600_query_buffer *qbuf = MALLOC_STRUCT(r600_query_buffer);
 160                 *qbuf = query->buffer;
 161                 query->buffer.buf = r600_new_query_buffer(ctx, query->type);
 162                 query->buffer.results_end = 0;
 163                 query->buffer.previous = qbuf;
 164         }
 165
 166         /* emit begin query */
 167         va = r600_resource_va(ctx->b.screen, (void*)query->buffer.buf);
 168         va += query->buffer.results_end;
 169
 170         switch (query->type) {
 171         case PIPE_QUERY_OCCLUSION_COUNTER:
 172         case PIPE_QUERY_OCCLUSION_PREDICATE:
 173                 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
 174                 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE) | EVENT_INDEX(1));
 175                 radeon_emit(cs, va);
 176                 radeon_emit(cs, (va >> 32UL) & 0xFF);
 177                 break;
 178         case PIPE_QUERY_PRIMITIVES_EMITTED:
 179         case PIPE_QUERY_PRIMITIVES_GENERATED:
 180         case PIPE_QUERY_SO_STATISTICS:
 181         case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 182                 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
 183                 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_SAMPLE_STREAMOUTSTATS) | EVENT_INDEX(3));
 184                 radeon_emit(cs, va);
 185                 radeon_emit(cs, (va >> 32UL) & 0xFF);
 186                 break;
 187         case PIPE_QUERY_TIME_ELAPSED:
 188                 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
 189                 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_TS_EVENT) | EVENT_INDEX(5));
 190                 radeon_emit(cs, va);
 191                 radeon_emit(cs, (3 << 29) | ((va >> 32UL) & 0xFF));
 192                 radeon_emit(cs, 0);
 193                 radeon_emit(cs, 0);
 194                 break;
 195         case PIPE_QUERY_PIPELINE_STATISTICS:
 196                 if (!ctx->num_pipelinestat_queries) {
 197                         radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
 198                         radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_PIPELINESTAT_START) | EVENT_INDEX(0));
 199                 }
 200                 ctx->num_pipelinestat_queries++;
 201                 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
 202                 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_SAMPLE_PIPELINESTAT) | EVENT_INDEX(2));
 203                 radeon_emit(cs, va);
 204                 radeon_emit(cs, (va >> 32UL) & 0xFF);
 205                 break;
 206         default:
 207                 assert(0);
 208         }
 209         r600_emit_reloc(ctx, &ctx->rings.gfx, query->buffer.buf, RADEON_USAGE_WRITE);
 210
 211         if (!r600_is_timer_query(query->type)) {
 212                 ctx->num_cs_dw_nontimer_queries_suspend += query->num_cs_dw;
 213         }
 214 }
 215
 216 static void r600_emit_query_end(struct r600_common_context *ctx, struct r600_query *query)
 217 {
 218         struct radeon_winsys_cs *cs = ctx->rings.gfx.cs;
 219         uint64_t va;
 220
 221         /* The queries which need begin already called this in begin_query. */
 222         if (!r600_query_needs_begin(query->type)) {
 223                 ctx->need_gfx_cs_space(&ctx->b, query->num_cs_dw, FALSE);
 224         }
 225
 226         va = r600_resource_va(ctx->b.screen, (void*)query->buffer.buf);
 227         /* emit end query */
 228         switch (query->type) {
 229         case PIPE_QUERY_OCCLUSION_COUNTER:
 230         case PIPE_QUERY_OCCLUSION_PREDICATE:
 231                 va += query->buffer.results_end + 8;
 232                 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
 233                 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE) | EVENT_INDEX(1));
 234                 radeon_emit(cs, va);
 235                 radeon_emit(cs, (va >> 32UL) & 0xFF);
 236                 break;
 237         case PIPE_QUERY_PRIMITIVES_EMITTED:
 238         case PIPE_QUERY_PRIMITIVES_GENERATED:
 239         case PIPE_QUERY_SO_STATISTICS:
 240         case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 241                 va += query->buffer.results_end + query->result_size/2;
 242                 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
 243                 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_SAMPLE_STREAMOUTSTATS) | EVENT_INDEX(3));
 244                 radeon_emit(cs, va);
 245                 radeon_emit(cs, (va >> 32UL) & 0xFF);
 246                 break;
 247         case PIPE_QUERY_TIME_ELAPSED:
 248                 va += query->buffer.results_end + query->result_size/2;
 249                 /* fall through */
 250         case PIPE_QUERY_TIMESTAMP:
 251                 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE_EOP, 4, 0));
 252                 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_TS_EVENT) | EVENT_INDEX(5));
 253                 radeon_emit(cs, va);
 254                 radeon_emit(cs, (3 << 29) | ((va >> 32UL) & 0xFF));
 255                 radeon_emit(cs, 0);
 256                 radeon_emit(cs, 0);
 257                 break;
 258         case PIPE_QUERY_PIPELINE_STATISTICS:
 259                 assert(ctx->num_pipelinestat_queries > 0);
 260                 ctx->num_pipelinestat_queries--;
 261                 if (!ctx->num_pipelinestat_queries) {
 262                         radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
 263                         radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_PIPELINESTAT_STOP) | EVENT_INDEX(0));
 264                 }
 265                 va += query->buffer.results_end + query->result_size/2;
 266                 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
 267                 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_SAMPLE_PIPELINESTAT) | EVENT_INDEX(2));
 268                 radeon_emit(cs, va);
 269                 radeon_emit(cs, (va >> 32UL) & 0xFF);
 270                 break;
 271         default:
 272                 assert(0);
 273         }
 274         r600_emit_reloc(ctx, &ctx->rings.gfx, query->buffer.buf, RADEON_USAGE_WRITE);
 275
 276         query->buffer.results_end += query->result_size;
 277
 278         if (r600_query_needs_begin(query->type)) {
 279                 if (!r600_is_timer_query(query->type)) {
 280                         ctx->num_cs_dw_nontimer_queries_suspend -= query->num_cs_dw;
 281                 }
 282         }
 283
 284         r600_update_occlusion_query_state(ctx, query->type, -1);
 285 }
 286
 287 static void r600_emit_query_predication(struct r600_common_context *ctx, struct r600_query *query,
 288                                         int operation, bool flag_wait)
 289 {
 290         struct radeon_winsys_cs *cs = ctx->rings.gfx.cs;
 291
 292         if (operation == PREDICATION_OP_CLEAR) {
 293                 ctx->need_gfx_cs_space(&ctx->b, 3, FALSE);
 294
 295                 radeon_emit(cs, PKT3(PKT3_SET_PREDICATION, 1, 0));
 296                 radeon_emit(cs, 0);
 297                 radeon_emit(cs, PRED_OP(PREDICATION_OP_CLEAR));
 298         } else {
 299                 struct r600_query_buffer *qbuf;
 300                 unsigned count;
 301                 uint32_t op;
 302
 303                 /* Find how many results there are. */
 304                 count = 0;
 305                 for (qbuf = &query->buffer; qbuf; qbuf = qbuf->previous) {
 306                         count += qbuf->results_end / query->result_size;
 307                 }
 308
 309                 ctx->need_gfx_cs_space(&ctx->b, 5 * count, TRUE);
 310
 311                 op = PRED_OP(operation) | PREDICATION_DRAW_VISIBLE |
 312                                 (flag_wait ? PREDICATION_HINT_WAIT : PREDICATION_HINT_NOWAIT_DRAW);
 313
 314                 /* emit predicate packets for all data blocks */
 315                 for (qbuf = &query->buffer; qbuf; qbuf = qbuf->previous) {
 316                         unsigned results_base = 0;
 317                         uint64_t va = r600_resource_va(ctx->b.screen, &qbuf->buf->b.b);
 318
 319                         while (results_base < qbuf->results_end) {
 320                                 radeon_emit(cs, PKT3(PKT3_SET_PREDICATION, 1, 0));
 321                                 radeon_emit(cs, (va + results_base) & 0xFFFFFFFFUL);
 322                                 radeon_emit(cs, op | (((va + results_base) >> 32UL) & 0xFF));
 323                                 r600_emit_reloc(ctx, &ctx->rings.gfx, qbuf->buf, RADEON_USAGE_READ);
 324                                 results_base += query->result_size;
 325
 326                                 /* set CONTINUE bit for all packets except the first */
 327                                 op |= PREDICATION_CONTINUE;
 328                         }
 329                 } while (qbuf);
 330         }
 331 }
 332
 333 static struct pipe_query *r600_create_query(struct pipe_context *ctx, unsigned query_type)
 334 {
 335         struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 336         struct r600_query *query;
 337         bool skip_allocation = false;
 338
 339         query = CALLOC_STRUCT(r600_query);
 340         if (query == NULL)
 341                 return NULL;
 342
 343         query->type = query_type;
 344
 345         switch (query_type) {
 346         case PIPE_QUERY_OCCLUSION_COUNTER:
 347         case PIPE_QUERY_OCCLUSION_PREDICATE:
 348                 query->result_size = 16 * rctx->max_db;
 349                 query->num_cs_dw = 6;
 350                 break;
 351         case PIPE_QUERY_TIME_ELAPSED:
 352                 query->result_size = 16;
 353                 query->num_cs_dw = 8;
 354                 break;
 355         case PIPE_QUERY_TIMESTAMP:
 356                 query->result_size = 8;
 357                 query->num_cs_dw = 8;
 358                 break;
 359         case PIPE_QUERY_PRIMITIVES_EMITTED:
 360         case PIPE_QUERY_PRIMITIVES_GENERATED:
 361         case PIPE_QUERY_SO_STATISTICS:
 362         case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 363                 /* NumPrimitivesWritten, PrimitiveStorageNeeded. */
 364                 query->result_size = 32;
 365                 query->num_cs_dw = 6;
 366                 break;
 367         case PIPE_QUERY_PIPELINE_STATISTICS:
 368                 /* 11 values on EG, 8 on R600. */
 369                 query->result_size = (rctx->chip_class >= EVERGREEN ? 11 : 8) * 16;
 370                 query->num_cs_dw = 8;
 371                 break;
 372         /* Non-GPU queries. */
 373         case R600_QUERY_DRAW_CALLS:
 374         case R600_QUERY_REQUESTED_VRAM:
 375         case R600_QUERY_REQUESTED_GTT:
 376         case R600_QUERY_BUFFER_WAIT_TIME:
 377                 skip_allocation = true;
 378                 break;
 379         default:
 380                 assert(0);
 381                 FREE(query);
 382                 return NULL;
 383         }
 384
 385         if (!skip_allocation) {
 386                 query->buffer.buf = r600_new_query_buffer(rctx, query_type);
 387                 if (!query->buffer.buf) {
 388                         FREE(query);
 389                         return NULL;
 390                 }
 391         }
 392         return (struct pipe_query*)query;
 393 }
 394
 395 static void r600_destroy_query(struct pipe_context *ctx, struct pipe_query *query)
 396 {
 397         struct r600_query *rquery = (struct r600_query*)query;
 398         struct r600_query_buffer *prev = rquery->buffer.previous;
 399
 400         /* Release all query buffers. */
 401         while (prev) {
 402                 struct r600_query_buffer *qbuf = prev;
 403                 prev = prev->previous;
 404                 pipe_resource_reference((struct pipe_resource**)&qbuf->buf, NULL);
 405                 FREE(qbuf);
 406         }
 407
 408         pipe_resource_reference((struct pipe_resource**)&rquery->buffer.buf, NULL);
 409         FREE(query);
 410 }
 411
 412 static void r600_begin_query(struct pipe_context *ctx, struct pipe_query *query)
 413 {
 414         struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 415         struct r600_query *rquery = (struct r600_query *)query;
 416         struct r600_query_buffer *prev = rquery->buffer.previous;
 417
 418         if (!r600_query_needs_begin(rquery->type)) {
 419                 assert(0);
 420                 return;
 421         }
 422
 423         /* Non-GPU queries. */
 424         switch (rquery->type) {
 425         case R600_QUERY_DRAW_CALLS:
 426                 rquery->begin_result = rctx->num_draw_calls;
 427                 return;
 428         case R600_QUERY_REQUESTED_VRAM:
 429         case R600_QUERY_REQUESTED_GTT:
 430                 rquery->begin_result = 0;
 431                 return;
 432         case R600_QUERY_BUFFER_WAIT_TIME:
 433                 rquery->begin_result = rctx->ws->query_value(rctx->ws, RADEON_BUFFER_WAIT_TIME_NS);
 434                 return;
 435         }
 436
 437         /* Discard the old query buffers. */
 438         while (prev) {
 439                 struct r600_query_buffer *qbuf = prev;
 440                 prev = prev->previous;
 441                 pipe_resource_reference((struct pipe_resource**)&qbuf->buf, NULL);
 442                 FREE(qbuf);
 443         }
 444
 445         /* Obtain a new buffer if the current one can't be mapped without a stall. */
 446         if (r600_rings_is_buffer_referenced(rctx, rquery->buffer.buf->cs_buf, RADEON_USAGE_READWRITE) ||
 447             rctx->ws->buffer_is_busy(rquery->buffer.buf->buf, RADEON_USAGE_READWRITE)) {
 448                 pipe_resource_reference((struct pipe_resource**)&rquery->buffer.buf, NULL);
 449                 rquery->buffer.buf = r600_new_query_buffer(rctx, rquery->type);
 450         }
 451
 452         rquery->buffer.results_end = 0;
 453         rquery->buffer.previous = NULL;
 454
 455         r600_emit_query_begin(rctx, rquery);
 456
 457         if (!r600_is_timer_query(rquery->type)) {
 458                 LIST_ADDTAIL(&rquery->list, &rctx->active_nontimer_queries);
 459         }
 460 }
 461
 462 static void r600_end_query(struct pipe_context *ctx, struct pipe_query *query)
 463 {
 464         struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 465         struct r600_query *rquery = (struct r600_query *)query;
 466
 467         /* Non-GPU queries. */
 468         switch (rquery->type) {
 469         case R600_QUERY_DRAW_CALLS:
 470                 rquery->end_result = rctx->num_draw_calls;
 471                 return;
 472         case R600_QUERY_REQUESTED_VRAM:
 473                 rquery->end_result = rctx->ws->query_value(rctx->ws, RADEON_REQUESTED_VRAM_MEMORY);
 474                 return;
 475         case R600_QUERY_REQUESTED_GTT:
 476                 rquery->end_result = rctx->ws->query_value(rctx->ws, RADEON_REQUESTED_GTT_MEMORY);
 477                 return;
 478         case R600_QUERY_BUFFER_WAIT_TIME:
 479                 rquery->end_result = rctx->ws->query_value(rctx->ws, RADEON_BUFFER_WAIT_TIME_NS);
 480                 return;
 481         }
 482
 483         r600_emit_query_end(rctx, rquery);
 484
 485         if (r600_query_needs_begin(rquery->type) && !r600_is_timer_query(rquery->type)) {
 486                 LIST_DELINIT(&rquery->list);
 487         }
 488 }
 489
 490 static unsigned r600_query_read_result(char *map, unsigned start_index, unsigned end_index,
 491                                        bool test_status_bit)
 492 {
 493         uint32_t *current_result = (uint32_t*)map;
 494         uint64_t start, end;
 495
 496         start = (uint64_t)current_result[start_index] |
 497                 (uint64_t)current_result[start_index+1] << 32;
 498         end = (uint64_t)current_result[end_index] |
 499               (uint64_t)current_result[end_index+1] << 32;
 500
 501         if (!test_status_bit ||
 502             ((start & 0x8000000000000000UL) && (end & 0x8000000000000000UL))) {
 503                 return end - start;
 504         }
 505         return 0;
 506 }
 507
 508 static boolean r600_get_query_buffer_result(struct r600_common_context *ctx,
 509                                             struct r600_query *query,
 510                                             struct r600_query_buffer *qbuf,
 511                                             boolean wait,
 512                                             union pipe_query_result *result)
 513 {
 514         unsigned results_base = 0;
 515         char *map;
 516
 517         /* Non-GPU queries. */
 518         switch (query->type) {
 519         case R600_QUERY_DRAW_CALLS:
 520         case R600_QUERY_REQUESTED_VRAM:
 521         case R600_QUERY_REQUESTED_GTT:
 522         case R600_QUERY_BUFFER_WAIT_TIME:
 523                 result->u64 = query->end_result - query->begin_result;
 524                 return TRUE;
 525         }
 526
 527         map = r600_buffer_map_sync_with_rings(ctx, qbuf->buf,
 528                                                 PIPE_TRANSFER_READ |
 529                                                 (wait ? 0 : PIPE_TRANSFER_DONTBLOCK));
 530         if (!map)
 531                 return FALSE;
 532
 533         /* count all results across all data blocks */
 534         switch (query->type) {
 535         case PIPE_QUERY_OCCLUSION_COUNTER:
 536                 while (results_base != qbuf->results_end) {
 537                         result->u64 +=
 538                                 r600_query_read_result(map + results_base, 0, 2, true);
 539                         results_base += 16;
 540                 }
 541                 break;
 542         case PIPE_QUERY_OCCLUSION_PREDICATE:
 543                 while (results_base != qbuf->results_end) {
 544                         result->b = result->b ||
 545                                 r600_query_read_result(map + results_base, 0, 2, true) != 0;
 546                         results_base += 16;
 547                 }
 548                 break;
 549         case PIPE_QUERY_TIME_ELAPSED:
 550                 while (results_base != qbuf->results_end) {
 551                         result->u64 +=
 552                                 r600_query_read_result(map + results_base, 0, 2, false);
 553                         results_base += query->result_size;
 554                 }
 555                 break;
 556         case PIPE_QUERY_TIMESTAMP:
 557         {
 558                 uint32_t *current_result = (uint32_t*)map;
 559                 result->u64 = (uint64_t)current_result[0] |
 560                               (uint64_t)current_result[1] << 32;
 561                 break;
 562         }
 563         case PIPE_QUERY_PRIMITIVES_EMITTED:
 564                 /* SAMPLE_STREAMOUTSTATS stores this structure:
 565                  * {
 566                  *    u64 NumPrimitivesWritten;
 567                  *    u64 PrimitiveStorageNeeded;
 568                  * }
 569                  * We only need NumPrimitivesWritten here. */
 570                 while (results_base != qbuf->results_end) {
 571                         result->u64 +=
 572                                 r600_query_read_result(map + results_base, 2, 6, true);
 573                         results_base += query->result_size;
 574                 }
 575                 break;
 576         case PIPE_QUERY_PRIMITIVES_GENERATED:
 577                 /* Here we read PrimitiveStorageNeeded. */
 578                 while (results_base != qbuf->results_end) {
 579                         result->u64 +=
 580                                 r600_query_read_result(map + results_base, 0, 4, true);
 581                         results_base += query->result_size;
 582                 }
 583                 break;
 584         case PIPE_QUERY_SO_STATISTICS:
 585                 while (results_base != qbuf->results_end) {
 586                         result->so_statistics.num_primitives_written +=
 587                                 r600_query_read_result(map + results_base, 2, 6, true);
 588                         result->so_statistics.primitives_storage_needed +=
 589                                 r600_query_read_result(map + results_base, 0, 4, true);
 590                         results_base += query->result_size;
 591                 }
 592                 break;
 593         case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 594                 while (results_base != qbuf->results_end) {
 595                         result->b = result->b ||
 596                                 r600_query_read_result(map + results_base, 2, 6, true) !=
 597                                 r600_query_read_result(map + results_base, 0, 4, true);
 598                         results_base += query->result_size;
 599                 }
 600                 break;
 601         case PIPE_QUERY_PIPELINE_STATISTICS:
 602                 if (ctx->chip_class >= EVERGREEN) {
 603                         while (results_base != qbuf->results_end) {
 604                                 result->pipeline_statistics.ps_invocations +=
 605                                         r600_query_read_result(map + results_base, 0, 22, false);
 606                                 result->pipeline_statistics.c_primitives +=
 607                                         r600_query_read_result(map + results_base, 2, 24, false);
 608                                 result->pipeline_statistics.c_invocations +=
 609                                         r600_query_read_result(map + results_base, 4, 26, false);
 610                                 result->pipeline_statistics.vs_invocations +=
 611                                         r600_query_read_result(map + results_base, 6, 28, false);
 612                                 result->pipeline_statistics.gs_invocations +=
 613                                         r600_query_read_result(map + results_base, 8, 30, false);
 614                                 result->pipeline_statistics.gs_primitives +=
 615                                         r600_query_read_result(map + results_base, 10, 32, false);
 616                                 result->pipeline_statistics.ia_primitives +=
 617                                         r600_query_read_result(map + results_base, 12, 34, false);
 618                                 result->pipeline_statistics.ia_vertices +=
 619                                         r600_query_read_result(map + results_base, 14, 36, false);
 620                                 result->pipeline_statistics.hs_invocations +=
 621                                         r600_query_read_result(map + results_base, 16, 38, false);
 622                                 result->pipeline_statistics.ds_invocations +=
 623                                         r600_query_read_result(map + results_base, 18, 40, false);
 624                                 result->pipeline_statistics.cs_invocations +=
 625                                         r600_query_read_result(map + results_base, 20, 42, false);
 626                                 results_base += query->result_size;
 627                         }
 628                 } else {
 629                         while (results_base != qbuf->results_end) {
 630                                 result->pipeline_statistics.ps_invocations +=
 631                                         r600_query_read_result(map + results_base, 0, 16, false);
 632                                 result->pipeline_statistics.c_primitives +=
 633                                         r600_query_read_result(map + results_base, 2, 18, false);
 634                                 result->pipeline_statistics.c_invocations +=
 635                                         r600_query_read_result(map + results_base, 4, 20, false);
 636                                 result->pipeline_statistics.vs_invocations +=
 637                                         r600_query_read_result(map + results_base, 6, 22, false);
 638                                 result->pipeline_statistics.gs_invocations +=
 639                                         r600_query_read_result(map + results_base, 8, 24, false);
 640                                 result->pipeline_statistics.gs_primitives +=
 641                                         r600_query_read_result(map + results_base, 10, 26, false);
 642                                 result->pipeline_statistics.ia_primitives +=
 643                                         r600_query_read_result(map + results_base, 12, 28, false);
 644                                 result->pipeline_statistics.ia_vertices +=
 645                                         r600_query_read_result(map + results_base, 14, 30, false);
 646                                 results_base += query->result_size;
 647                         }
 648                 }
 649 #if 0 /* for testing */
 650                 printf("Pipeline stats: IA verts=%llu, IA prims=%llu, VS=%llu, HS=%llu, "
 651                        "DS=%llu, GS=%llu, GS prims=%llu, Clipper=%llu, "
 652                        "Clipper prims=%llu, PS=%llu, CS=%llu\n",
 653                        result->pipeline_statistics.ia_vertices,
 654                        result->pipeline_statistics.ia_primitives,
 655                        result->pipeline_statistics.vs_invocations,
 656                        result->pipeline_statistics.hs_invocations,
 657                        result->pipeline_statistics.ds_invocations,
 658                        result->pipeline_statistics.gs_invocations,
 659                        result->pipeline_statistics.gs_primitives,
 660                        result->pipeline_statistics.c_invocations,
 661                        result->pipeline_statistics.c_primitives,
 662                        result->pipeline_statistics.ps_invocations,
 663                        result->pipeline_statistics.cs_invocations);
 664 #endif
 665                 break;
 666         default:
 667                 assert(0);
 668         }
 669
 670         ctx->ws->buffer_unmap(qbuf->buf->cs_buf);
 671         return TRUE;
 672 }
 673
 674 static boolean r600_get_query_result(struct pipe_context *ctx,
 675                                         struct pipe_query *query,
 676                                         boolean wait, union pipe_query_result *result)
 677 {
 678         struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 679         struct r600_query *rquery = (struct r600_query *)query;
 680         struct r600_query_buffer *qbuf;
 681
 682         util_query_clear_result(result, rquery->type);
 683
 684         for (qbuf = &rquery->buffer; qbuf; qbuf = qbuf->previous) {
 685                 if (!r600_get_query_buffer_result(rctx, rquery, qbuf, wait, result)) {
 686                         return FALSE;
 687                 }
 688         }
 689
 690         /* Convert the time to expected units. */
 691         if (rquery->type == PIPE_QUERY_TIME_ELAPSED ||
 692             rquery->type == PIPE_QUERY_TIMESTAMP) {
 693                 result->u64 = (1000000 * result->u64) / rctx->screen->info.r600_clock_crystal_freq;
 694         }
 695         return TRUE;
 696 }
 697
 698 static void r600_render_condition(struct pipe_context *ctx,
 699                                   struct pipe_query *query,
 700                                   boolean condition,
 701                                   uint mode)
 702 {
 703         struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 704         struct r600_query *rquery = (struct r600_query *)query;
 705         bool wait_flag = false;
 706
 707         rctx->current_render_cond = query;
 708         rctx->current_render_cond_cond = condition;
 709         rctx->current_render_cond_mode = mode;
 710
 711         if (query == NULL) {
 712                 if (rctx->predicate_drawing) {
 713                         rctx->predicate_drawing = false;
 714                         r600_emit_query_predication(rctx, NULL, PREDICATION_OP_CLEAR, false);
 715                 }
 716                 return;
 717         }
 718
 719         if (mode == PIPE_RENDER_COND_WAIT ||
 720             mode == PIPE_RENDER_COND_BY_REGION_WAIT) {
 721                 wait_flag = true;
 722         }
 723
 724         rctx->predicate_drawing = true;
 725
 726         switch (rquery->type) {
 727         case PIPE_QUERY_OCCLUSION_COUNTER:
 728         case PIPE_QUERY_OCCLUSION_PREDICATE:
 729                 r600_emit_query_predication(rctx, rquery, PREDICATION_OP_ZPASS, wait_flag);
 730                 break;
 731         case PIPE_QUERY_PRIMITIVES_EMITTED:
 732         case PIPE_QUERY_PRIMITIVES_GENERATED:
 733         case PIPE_QUERY_SO_STATISTICS:
 734         case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 735                 r600_emit_query_predication(rctx, rquery, PREDICATION_OP_PRIMCOUNT, wait_flag);
 736                 break;
 737         default:
 738                 assert(0);
 739         }
 740 }
 741
 742 void r600_suspend_nontimer_queries(struct r600_common_context *ctx)
 743 {
 744         struct r600_query *query;
 745
 746         LIST_FOR_EACH_ENTRY(query, &ctx->active_nontimer_queries, list) {
 747                 r600_emit_query_end(ctx, query);
 748         }
 749         assert(ctx->num_cs_dw_nontimer_queries_suspend == 0);
 750 }
 751
 752 void r600_resume_nontimer_queries(struct r600_common_context *ctx)
 753 {
 754         struct r600_query *query;
 755
 756         assert(ctx->num_cs_dw_nontimer_queries_suspend == 0);
 757
 758         LIST_FOR_EACH_ENTRY(query, &ctx->active_nontimer_queries, list) {
 759                 r600_emit_query_begin(ctx, query);
 760         }
 761 }
 762
 763 /* Get backends mask */
 764 void r600_query_init_backend_mask(struct r600_common_context *ctx)
 765 {
 766         struct radeon_winsys_cs *cs = ctx->rings.gfx.cs;
 767         struct r600_resource *buffer;
 768         uint32_t *results;
 769         unsigned num_backends = ctx->screen->info.r600_num_backends;
 770         unsigned i, mask = 0;
 771         uint64_t va;
 772
 773         /* if backend_map query is supported by the kernel */
 774         if (ctx->screen->info.r600_backend_map_valid) {
 775                 unsigned num_tile_pipes = ctx->screen->info.r600_num_tile_pipes;
 776                 unsigned backend_map = ctx->screen->info.r600_backend_map;
 777                 unsigned item_width, item_mask;
 778
 779                 if (ctx->chip_class >= EVERGREEN) {
 780                         item_width = 4;
 781                         item_mask = 0x7;
 782                 } else {
 783                         item_width = 2;
 784                         item_mask = 0x3;
 785                 }
 786
 787                 while(num_tile_pipes--) {
 788                         i = backend_map & item_mask;
 789                         mask |= (1<<i);
 790                         backend_map >>= item_width;
 791                 }
 792                 if (mask != 0) {
 793                         ctx->backend_mask = mask;
 794                         return;
 795                 }
 796         }
 797
 798         /* otherwise backup path for older kernels */
 799
 800         /* create buffer for event data */
 801         buffer = (struct r600_resource*)
 802                 pipe_buffer_create(ctx->b.screen, PIPE_BIND_CUSTOM,
 803                                    PIPE_USAGE_STAGING, ctx->max_db*16);
 804         if (!buffer)
 805                 goto err;
 806         va = r600_resource_va(ctx->b.screen, (void*)buffer);
 807
 808         /* initialize buffer with zeroes */
 809         results = r600_buffer_map_sync_with_rings(ctx, buffer, PIPE_TRANSFER_WRITE);
 810         if (results) {
 811                 memset(results, 0, ctx->max_db * 4 * 4);
 812                 ctx->ws->buffer_unmap(buffer->cs_buf);
 813
 814                 /* emit EVENT_WRITE for ZPASS_DONE */
 815                 radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 2, 0));
 816                 radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_ZPASS_DONE) | EVENT_INDEX(1));
 817                 radeon_emit(cs, va);
 818                 radeon_emit(cs, va >> 32);
 819
 820                 r600_emit_reloc(ctx, &ctx->rings.gfx, buffer, RADEON_USAGE_WRITE);
 821
 822                 /* analyze results */
 823                 results = r600_buffer_map_sync_with_rings(ctx, buffer, PIPE_TRANSFER_READ);
 824                 if (results) {
 825                         for(i = 0; i < ctx->max_db; i++) {
 826                                 /* at least highest bit will be set if backend is used */
 827                                 if (results[i*4 + 1])
 828                                         mask |= (1<<i);
 829                         }
 830                         ctx->ws->buffer_unmap(buffer->cs_buf);
 831                 }
 832         }
 833
 834         pipe_resource_reference((struct pipe_resource**)&buffer, NULL);
 835
 836         if (mask != 0) {
 837                 ctx->backend_mask = mask;
 838                 return;
 839         }
 840
 841 err:
 842         /* fallback to old method - set num_backends lower bits to 1 */
 843         ctx->backend_mask = (~((uint32_t)0))>>(32-num_backends);
 844         return;
 845 }
 846
 847 void r600_query_init(struct r600_common_context *rctx)
 848 {
 849         rctx->b.create_query = r600_create_query;
 850         rctx->b.destroy_query = r600_destroy_query;
 851         rctx->b.begin_query = r600_begin_query;
 852         rctx->b.end_query = r600_end_query;
 853         rctx->b.get_query_result = r600_get_query_result;
 854
 855         if (((struct r600_common_screen*)rctx->b.screen)->info.r600_num_backends > 0)
 856             rctx->b.render_condition = r600_render_condition;
 857
 858         LIST_INITHEAD(&rctx->active_nontimer_queries);
 859 }