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                 cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 2, 0);
 174                 cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_ZPASS_DONE) | EVENT_INDEX(1);
 175                 cs->buf[cs->cdw++] = va;
 176                 cs->buf[cs->cdw++] = (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                 cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 2, 0);
 183                 cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_SAMPLE_STREAMOUTSTATS) | EVENT_INDEX(3);
 184                 cs->buf[cs->cdw++] = va;
 185                 cs->buf[cs->cdw++] = (va >> 32UL) & 0xFF;
 186                 break;
 187         case PIPE_QUERY_TIME_ELAPSED:
 188                 cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE_EOP, 4, 0);
 189                 cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_TS_EVENT) | EVENT_INDEX(5);
 190                 cs->buf[cs->cdw++] = va;
 191                 cs->buf[cs->cdw++] = (3 << 29) | ((va >> 32UL) & 0xFF);
 192                 cs->buf[cs->cdw++] = 0;
 193                 cs->buf[cs->cdw++] = 0;
 194                 break;
 195         case PIPE_QUERY_PIPELINE_STATISTICS:
 196                 if (!ctx->num_pipelinestat_queries) {
 197                         cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 0, 0);
 198                         cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_PIPELINESTAT_START) | EVENT_INDEX(0);
 199                 }
 200                 ctx->num_pipelinestat_queries++;
 201                 cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 2, 0);
 202                 cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_SAMPLE_PIPELINESTAT) | EVENT_INDEX(2);
 203                 cs->buf[cs->cdw++] = va;
 204                 cs->buf[cs->cdw++] = (va >> 32UL) & 0xFF;
 205                 break;
 206         default:
 207                 assert(0);
 208         }
 209         cs->buf[cs->cdw++] = PKT3(PKT3_NOP, 0, 0);
 210         cs->buf[cs->cdw++] = r600_context_bo_reloc(ctx, &ctx->rings.gfx, query->buffer.buf, RADEON_USAGE_WRITE);
 211
 212         if (!r600_is_timer_query(query->type)) {
 213                 ctx->num_cs_dw_nontimer_queries_suspend += query->num_cs_dw;
 214         }
 215 }
 216
 217 static void r600_emit_query_end(struct r600_common_context *ctx, struct r600_query *query)
 218 {
 219         struct radeon_winsys_cs *cs = ctx->rings.gfx.cs;
 220         uint64_t va;
 221
 222         /* The queries which need begin already called this in begin_query. */
 223         if (!r600_query_needs_begin(query->type)) {
 224                 ctx->need_gfx_cs_space(&ctx->b, query->num_cs_dw, FALSE);
 225         }
 226
 227         va = r600_resource_va(ctx->b.screen, (void*)query->buffer.buf);
 228         /* emit end query */
 229         switch (query->type) {
 230         case PIPE_QUERY_OCCLUSION_COUNTER:
 231         case PIPE_QUERY_OCCLUSION_PREDICATE:
 232                 va += query->buffer.results_end + 8;
 233                 cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 2, 0);
 234                 cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_ZPASS_DONE) | EVENT_INDEX(1);
 235                 cs->buf[cs->cdw++] = va;
 236                 cs->buf[cs->cdw++] = (va >> 32UL) & 0xFF;
 237                 break;
 238         case PIPE_QUERY_PRIMITIVES_EMITTED:
 239         case PIPE_QUERY_PRIMITIVES_GENERATED:
 240         case PIPE_QUERY_SO_STATISTICS:
 241         case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 242                 va += query->buffer.results_end + query->result_size/2;
 243                 cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 2, 0);
 244                 cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_SAMPLE_STREAMOUTSTATS) | EVENT_INDEX(3);
 245                 cs->buf[cs->cdw++] = va;
 246                 cs->buf[cs->cdw++] = (va >> 32UL) & 0xFF;
 247                 break;
 248         case PIPE_QUERY_TIME_ELAPSED:
 249                 va += query->buffer.results_end + query->result_size/2;
 250                 /* fall through */
 251         case PIPE_QUERY_TIMESTAMP:
 252                 cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE_EOP, 4, 0);
 253                 cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_TS_EVENT) | EVENT_INDEX(5);
 254                 cs->buf[cs->cdw++] = va;
 255                 cs->buf[cs->cdw++] = (3 << 29) | ((va >> 32UL) & 0xFF);
 256                 cs->buf[cs->cdw++] = 0;
 257                 cs->buf[cs->cdw++] = 0;
 258                 break;
 259         case PIPE_QUERY_PIPELINE_STATISTICS:
 260                 assert(ctx->num_pipelinestat_queries > 0);
 261                 ctx->num_pipelinestat_queries--;
 262                 if (!ctx->num_pipelinestat_queries) {
 263                         cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 0, 0);
 264                         cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_PIPELINESTAT_STOP) | EVENT_INDEX(0);
 265                 }
 266                 va += query->buffer.results_end + query->result_size/2;
 267                 cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 2, 0);
 268                 cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_SAMPLE_PIPELINESTAT) | EVENT_INDEX(2);
 269                 cs->buf[cs->cdw++] = va;
 270                 cs->buf[cs->cdw++] = (va >> 32UL) & 0xFF;
 271                 break;
 272         default:
 273                 assert(0);
 274         }
 275         cs->buf[cs->cdw++] = PKT3(PKT3_NOP, 0, 0);
 276         cs->buf[cs->cdw++] = r600_context_bo_reloc(ctx, &ctx->rings.gfx, query->buffer.buf, RADEON_USAGE_WRITE);
 277
 278         query->buffer.results_end += query->result_size;
 279
 280         if (r600_query_needs_begin(query->type)) {
 281                 if (!r600_is_timer_query(query->type)) {
 282                         ctx->num_cs_dw_nontimer_queries_suspend -= query->num_cs_dw;
 283                 }
 284         }
 285
 286         r600_update_occlusion_query_state(ctx, query->type, -1);
 287 }
 288
 289 static void r600_emit_query_predication(struct r600_common_context *ctx, struct r600_query *query,
 290                                         int operation, bool flag_wait)
 291 {
 292         struct radeon_winsys_cs *cs = ctx->rings.gfx.cs;
 293
 294         if (operation == PREDICATION_OP_CLEAR) {
 295                 ctx->need_gfx_cs_space(&ctx->b, 3, FALSE);
 296
 297                 cs->buf[cs->cdw++] = PKT3(PKT3_SET_PREDICATION, 1, 0);
 298                 cs->buf[cs->cdw++] = 0;
 299                 cs->buf[cs->cdw++] = PRED_OP(PREDICATION_OP_CLEAR);
 300         } else {
 301                 struct r600_query_buffer *qbuf;
 302                 unsigned count;
 303                 uint32_t op;
 304
 305                 /* Find how many results there are. */
 306                 count = 0;
 307                 for (qbuf = &query->buffer; qbuf; qbuf = qbuf->previous) {
 308                         count += qbuf->results_end / query->result_size;
 309                 }
 310
 311                 ctx->need_gfx_cs_space(&ctx->b, 5 * count, TRUE);
 312
 313                 op = PRED_OP(operation) | PREDICATION_DRAW_VISIBLE |
 314                                 (flag_wait ? PREDICATION_HINT_WAIT : PREDICATION_HINT_NOWAIT_DRAW);
 315
 316                 /* emit predicate packets for all data blocks */
 317                 for (qbuf = &query->buffer; qbuf; qbuf = qbuf->previous) {
 318                         unsigned results_base = 0;
 319                         uint64_t va = r600_resource_va(ctx->b.screen, &qbuf->buf->b.b);
 320
 321                         while (results_base < qbuf->results_end) {
 322                                 cs->buf[cs->cdw++] = PKT3(PKT3_SET_PREDICATION, 1, 0);
 323                                 cs->buf[cs->cdw++] = (va + results_base) & 0xFFFFFFFFUL;
 324                                 cs->buf[cs->cdw++] = op | (((va + results_base) >> 32UL) & 0xFF);
 325                                 cs->buf[cs->cdw++] = PKT3(PKT3_NOP, 0, 0);
 326                                 cs->buf[cs->cdw++] = r600_context_bo_reloc(ctx, &ctx->rings.gfx, qbuf->buf, RADEON_USAGE_READ);
 327                                 results_base += query->result_size;
 328
 329                                 /* set CONTINUE bit for all packets except the first */
 330                                 op |= PREDICATION_CONTINUE;
 331                         }
 332                 } while (qbuf);
 333         }
 334 }
 335
 336 static struct pipe_query *r600_create_query(struct pipe_context *ctx, unsigned query_type)
 337 {
 338         struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 339         struct r600_query *query;
 340         bool skip_allocation = false;
 341
 342         query = CALLOC_STRUCT(r600_query);
 343         if (query == NULL)
 344                 return NULL;
 345
 346         query->type = query_type;
 347
 348         switch (query_type) {
 349         case PIPE_QUERY_OCCLUSION_COUNTER:
 350         case PIPE_QUERY_OCCLUSION_PREDICATE:
 351                 query->result_size = 16 * rctx->max_db;
 352                 query->num_cs_dw = 6;
 353                 break;
 354         case PIPE_QUERY_TIME_ELAPSED:
 355                 query->result_size = 16;
 356                 query->num_cs_dw = 8;
 357                 break;
 358         case PIPE_QUERY_TIMESTAMP:
 359                 query->result_size = 8;
 360                 query->num_cs_dw = 8;
 361                 break;
 362         case PIPE_QUERY_PRIMITIVES_EMITTED:
 363         case PIPE_QUERY_PRIMITIVES_GENERATED:
 364         case PIPE_QUERY_SO_STATISTICS:
 365         case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 366                 /* NumPrimitivesWritten, PrimitiveStorageNeeded. */
 367                 query->result_size = 32;
 368                 query->num_cs_dw = 6;
 369                 break;
 370         case PIPE_QUERY_PIPELINE_STATISTICS:
 371                 /* 11 values on EG, 8 on R600. */
 372                 query->result_size = (rctx->chip_class >= EVERGREEN ? 11 : 8) * 16;
 373                 query->num_cs_dw = 8;
 374                 break;
 375         /* Non-GPU queries. */
 376         case R600_QUERY_DRAW_CALLS:
 377         case R600_QUERY_REQUESTED_VRAM:
 378         case R600_QUERY_REQUESTED_GTT:
 379         case R600_QUERY_BUFFER_WAIT_TIME:
 380                 skip_allocation = true;
 381                 break;
 382         default:
 383                 assert(0);
 384                 FREE(query);
 385                 return NULL;
 386         }
 387
 388         if (!skip_allocation) {
 389                 query->buffer.buf = r600_new_query_buffer(rctx, query_type);
 390                 if (!query->buffer.buf) {
 391                         FREE(query);
 392                         return NULL;
 393                 }
 394         }
 395         return (struct pipe_query*)query;
 396 }
 397
 398 static void r600_destroy_query(struct pipe_context *ctx, struct pipe_query *query)
 399 {
 400         struct r600_query *rquery = (struct r600_query*)query;
 401         struct r600_query_buffer *prev = rquery->buffer.previous;
 402
 403         /* Release all query buffers. */
 404         while (prev) {
 405                 struct r600_query_buffer *qbuf = prev;
 406                 prev = prev->previous;
 407                 pipe_resource_reference((struct pipe_resource**)&qbuf->buf, NULL);
 408                 FREE(qbuf);
 409         }
 410
 411         pipe_resource_reference((struct pipe_resource**)&rquery->buffer.buf, NULL);
 412         FREE(query);
 413 }
 414
 415 static void r600_begin_query(struct pipe_context *ctx, struct pipe_query *query)
 416 {
 417         struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 418         struct r600_query *rquery = (struct r600_query *)query;
 419         struct r600_query_buffer *prev = rquery->buffer.previous;
 420
 421         if (!r600_query_needs_begin(rquery->type)) {
 422                 assert(0);
 423                 return;
 424         }
 425
 426         /* Non-GPU queries. */
 427         switch (rquery->type) {
 428         case R600_QUERY_DRAW_CALLS:
 429                 rquery->begin_result = rctx->num_draw_calls;
 430                 return;
 431         case R600_QUERY_REQUESTED_VRAM:
 432         case R600_QUERY_REQUESTED_GTT:
 433                 rquery->begin_result = 0;
 434                 return;
 435         case R600_QUERY_BUFFER_WAIT_TIME:
 436                 rquery->begin_result = rctx->ws->query_value(rctx->ws, RADEON_BUFFER_WAIT_TIME_NS);
 437                 return;
 438         }
 439
 440         /* Discard the old query buffers. */
 441         while (prev) {
 442                 struct r600_query_buffer *qbuf = prev;
 443                 prev = prev->previous;
 444                 pipe_resource_reference((struct pipe_resource**)&qbuf->buf, NULL);
 445                 FREE(qbuf);
 446         }
 447
 448         /* Obtain a new buffer if the current one can't be mapped without a stall. */
 449         if (r600_rings_is_buffer_referenced(rctx, rquery->buffer.buf->cs_buf, RADEON_USAGE_READWRITE) ||
 450             rctx->ws->buffer_is_busy(rquery->buffer.buf->buf, RADEON_USAGE_READWRITE)) {
 451                 pipe_resource_reference((struct pipe_resource**)&rquery->buffer.buf, NULL);
 452                 rquery->buffer.buf = r600_new_query_buffer(rctx, rquery->type);
 453         }
 454
 455         rquery->buffer.results_end = 0;
 456         rquery->buffer.previous = NULL;
 457
 458         r600_emit_query_begin(rctx, rquery);
 459
 460         if (!r600_is_timer_query(rquery->type)) {
 461                 LIST_ADDTAIL(&rquery->list, &rctx->active_nontimer_queries);
 462         }
 463 }
 464
 465 static void r600_end_query(struct pipe_context *ctx, struct pipe_query *query)
 466 {
 467         struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 468         struct r600_query *rquery = (struct r600_query *)query;
 469
 470         /* Non-GPU queries. */
 471         switch (rquery->type) {
 472         case R600_QUERY_DRAW_CALLS:
 473                 rquery->end_result = rctx->num_draw_calls;
 474                 return;
 475         case R600_QUERY_REQUESTED_VRAM:
 476                 rquery->end_result = rctx->ws->query_value(rctx->ws, RADEON_REQUESTED_VRAM_MEMORY);
 477                 return;
 478         case R600_QUERY_REQUESTED_GTT:
 479                 rquery->end_result = rctx->ws->query_value(rctx->ws, RADEON_REQUESTED_GTT_MEMORY);
 480                 return;
 481         case R600_QUERY_BUFFER_WAIT_TIME:
 482                 rquery->end_result = rctx->ws->query_value(rctx->ws, RADEON_BUFFER_WAIT_TIME_NS);
 483                 return;
 484         }
 485
 486         r600_emit_query_end(rctx, rquery);
 487
 488         if (r600_query_needs_begin(rquery->type) && !r600_is_timer_query(rquery->type)) {
 489                 LIST_DELINIT(&rquery->list);
 490         }
 491 }
 492
 493 static unsigned r600_query_read_result(char *map, unsigned start_index, unsigned end_index,
 494                                        bool test_status_bit)
 495 {
 496         uint32_t *current_result = (uint32_t*)map;
 497         uint64_t start, end;
 498
 499         start = (uint64_t)current_result[start_index] |
 500                 (uint64_t)current_result[start_index+1] << 32;
 501         end = (uint64_t)current_result[end_index] |
 502               (uint64_t)current_result[end_index+1] << 32;
 503
 504         if (!test_status_bit ||
 505             ((start & 0x8000000000000000UL) && (end & 0x8000000000000000UL))) {
 506                 return end - start;
 507         }
 508         return 0;
 509 }
 510
 511 static boolean r600_get_query_buffer_result(struct r600_common_context *ctx,
 512                                             struct r600_query *query,
 513                                             struct r600_query_buffer *qbuf,
 514                                             boolean wait,
 515                                             union pipe_query_result *result)
 516 {
 517         unsigned results_base = 0;
 518         char *map;
 519
 520         /* Non-GPU queries. */
 521         switch (query->type) {
 522         case R600_QUERY_DRAW_CALLS:
 523         case R600_QUERY_REQUESTED_VRAM:
 524         case R600_QUERY_REQUESTED_GTT:
 525         case R600_QUERY_BUFFER_WAIT_TIME:
 526                 result->u64 = query->end_result - query->begin_result;
 527                 return TRUE;
 528         }
 529
 530         map = r600_buffer_map_sync_with_rings(ctx, qbuf->buf,
 531                                                 PIPE_TRANSFER_READ |
 532                                                 (wait ? 0 : PIPE_TRANSFER_DONTBLOCK));
 533         if (!map)
 534                 return FALSE;
 535
 536         /* count all results across all data blocks */
 537         switch (query->type) {
 538         case PIPE_QUERY_OCCLUSION_COUNTER:
 539                 while (results_base != qbuf->results_end) {
 540                         result->u64 +=
 541                                 r600_query_read_result(map + results_base, 0, 2, true);
 542                         results_base += 16;
 543                 }
 544                 break;
 545         case PIPE_QUERY_OCCLUSION_PREDICATE:
 546                 while (results_base != qbuf->results_end) {
 547                         result->b = result->b ||
 548                                 r600_query_read_result(map + results_base, 0, 2, true) != 0;
 549                         results_base += 16;
 550                 }
 551                 break;
 552         case PIPE_QUERY_TIME_ELAPSED:
 553                 while (results_base != qbuf->results_end) {
 554                         result->u64 +=
 555                                 r600_query_read_result(map + results_base, 0, 2, false);
 556                         results_base += query->result_size;
 557                 }
 558                 break;
 559         case PIPE_QUERY_TIMESTAMP:
 560         {
 561                 uint32_t *current_result = (uint32_t*)map;
 562                 result->u64 = (uint64_t)current_result[0] |
 563                               (uint64_t)current_result[1] << 32;
 564                 break;
 565         }
 566         case PIPE_QUERY_PRIMITIVES_EMITTED:
 567                 /* SAMPLE_STREAMOUTSTATS stores this structure:
 568                  * {
 569                  *    u64 NumPrimitivesWritten;
 570                  *    u64 PrimitiveStorageNeeded;
 571                  * }
 572                  * We only need NumPrimitivesWritten here. */
 573                 while (results_base != qbuf->results_end) {
 574                         result->u64 +=
 575                                 r600_query_read_result(map + results_base, 2, 6, true);
 576                         results_base += query->result_size;
 577                 }
 578                 break;
 579         case PIPE_QUERY_PRIMITIVES_GENERATED:
 580                 /* Here we read PrimitiveStorageNeeded. */
 581                 while (results_base != qbuf->results_end) {
 582                         result->u64 +=
 583                                 r600_query_read_result(map + results_base, 0, 4, true);
 584                         results_base += query->result_size;
 585                 }
 586                 break;
 587         case PIPE_QUERY_SO_STATISTICS:
 588                 while (results_base != qbuf->results_end) {
 589                         result->so_statistics.num_primitives_written +=
 590                                 r600_query_read_result(map + results_base, 2, 6, true);
 591                         result->so_statistics.primitives_storage_needed +=
 592                                 r600_query_read_result(map + results_base, 0, 4, true);
 593                         results_base += query->result_size;
 594                 }
 595                 break;
 596         case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 597                 while (results_base != qbuf->results_end) {
 598                         result->b = result->b ||
 599                                 r600_query_read_result(map + results_base, 2, 6, true) !=
 600                                 r600_query_read_result(map + results_base, 0, 4, true);
 601                         results_base += query->result_size;
 602                 }
 603                 break;
 604         case PIPE_QUERY_PIPELINE_STATISTICS:
 605                 if (ctx->chip_class >= EVERGREEN) {
 606                         while (results_base != qbuf->results_end) {
 607                                 result->pipeline_statistics.ps_invocations +=
 608                                         r600_query_read_result(map + results_base, 0, 22, false);
 609                                 result->pipeline_statistics.c_primitives +=
 610                                         r600_query_read_result(map + results_base, 2, 24, false);
 611                                 result->pipeline_statistics.c_invocations +=
 612                                         r600_query_read_result(map + results_base, 4, 26, false);
 613                                 result->pipeline_statistics.vs_invocations +=
 614                                         r600_query_read_result(map + results_base, 6, 28, false);
 615                                 result->pipeline_statistics.gs_invocations +=
 616                                         r600_query_read_result(map + results_base, 8, 30, false);
 617                                 result->pipeline_statistics.gs_primitives +=
 618                                         r600_query_read_result(map + results_base, 10, 32, false);
 619                                 result->pipeline_statistics.ia_primitives +=
 620                                         r600_query_read_result(map + results_base, 12, 34, false);
 621                                 result->pipeline_statistics.ia_vertices +=
 622                                         r600_query_read_result(map + results_base, 14, 36, false);
 623                                 result->pipeline_statistics.hs_invocations +=
 624                                         r600_query_read_result(map + results_base, 16, 38, false);
 625                                 result->pipeline_statistics.ds_invocations +=
 626                                         r600_query_read_result(map + results_base, 18, 40, false);
 627                                 result->pipeline_statistics.cs_invocations +=
 628                                         r600_query_read_result(map + results_base, 20, 42, false);
 629                                 results_base += query->result_size;
 630                         }
 631                 } else {
 632                         while (results_base != qbuf->results_end) {
 633                                 result->pipeline_statistics.ps_invocations +=
 634                                         r600_query_read_result(map + results_base, 0, 16, false);
 635                                 result->pipeline_statistics.c_primitives +=
 636                                         r600_query_read_result(map + results_base, 2, 18, false);
 637                                 result->pipeline_statistics.c_invocations +=
 638                                         r600_query_read_result(map + results_base, 4, 20, false);
 639                                 result->pipeline_statistics.vs_invocations +=
 640                                         r600_query_read_result(map + results_base, 6, 22, false);
 641                                 result->pipeline_statistics.gs_invocations +=
 642                                         r600_query_read_result(map + results_base, 8, 24, false);
 643                                 result->pipeline_statistics.gs_primitives +=
 644                                         r600_query_read_result(map + results_base, 10, 26, false);
 645                                 result->pipeline_statistics.ia_primitives +=
 646                                         r600_query_read_result(map + results_base, 12, 28, false);
 647                                 result->pipeline_statistics.ia_vertices +=
 648                                         r600_query_read_result(map + results_base, 14, 30, false);
 649                                 results_base += query->result_size;
 650                         }
 651                 }
 652 #if 0 /* for testing */
 653                 printf("Pipeline stats: IA verts=%llu, IA prims=%llu, VS=%llu, HS=%llu, "
 654                        "DS=%llu, GS=%llu, GS prims=%llu, Clipper=%llu, "
 655                        "Clipper prims=%llu, PS=%llu, CS=%llu\n",
 656                        result->pipeline_statistics.ia_vertices,
 657                        result->pipeline_statistics.ia_primitives,
 658                        result->pipeline_statistics.vs_invocations,
 659                        result->pipeline_statistics.hs_invocations,
 660                        result->pipeline_statistics.ds_invocations,
 661                        result->pipeline_statistics.gs_invocations,
 662                        result->pipeline_statistics.gs_primitives,
 663                        result->pipeline_statistics.c_invocations,
 664                        result->pipeline_statistics.c_primitives,
 665                        result->pipeline_statistics.ps_invocations,
 666                        result->pipeline_statistics.cs_invocations);
 667 #endif
 668                 break;
 669         default:
 670                 assert(0);
 671         }
 672
 673         ctx->ws->buffer_unmap(qbuf->buf->cs_buf);
 674         return TRUE;
 675 }
 676
 677 static boolean r600_get_query_result(struct pipe_context *ctx,
 678                                         struct pipe_query *query,
 679                                         boolean wait, union pipe_query_result *result)
 680 {
 681         struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 682         struct r600_query *rquery = (struct r600_query *)query;
 683         struct r600_query_buffer *qbuf;
 684
 685         util_query_clear_result(result, rquery->type);
 686
 687         for (qbuf = &rquery->buffer; qbuf; qbuf = qbuf->previous) {
 688                 if (!r600_get_query_buffer_result(rctx, rquery, qbuf, wait, result)) {
 689                         return FALSE;
 690                 }
 691         }
 692
 693         /* Convert the time to expected units. */
 694         if (rquery->type == PIPE_QUERY_TIME_ELAPSED ||
 695             rquery->type == PIPE_QUERY_TIMESTAMP) {
 696                 result->u64 = (1000000 * result->u64) / rctx->screen->info.r600_clock_crystal_freq;
 697         }
 698         return TRUE;
 699 }
 700
 701 static void r600_render_condition(struct pipe_context *ctx,
 702                                   struct pipe_query *query,
 703                                   boolean condition,
 704                                   uint mode)
 705 {
 706         struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 707         struct r600_query *rquery = (struct r600_query *)query;
 708         bool wait_flag = false;
 709
 710         rctx->current_render_cond = query;
 711         rctx->current_render_cond_cond = condition;
 712         rctx->current_render_cond_mode = mode;
 713
 714         if (query == NULL) {
 715                 if (rctx->predicate_drawing) {
 716                         rctx->predicate_drawing = false;
 717                         r600_emit_query_predication(rctx, NULL, PREDICATION_OP_CLEAR, false);
 718                 }
 719                 return;
 720         }
 721
 722         if (mode == PIPE_RENDER_COND_WAIT ||
 723             mode == PIPE_RENDER_COND_BY_REGION_WAIT) {
 724                 wait_flag = true;
 725         }
 726
 727         rctx->predicate_drawing = true;
 728
 729         switch (rquery->type) {
 730         case PIPE_QUERY_OCCLUSION_COUNTER:
 731         case PIPE_QUERY_OCCLUSION_PREDICATE:
 732                 r600_emit_query_predication(rctx, rquery, PREDICATION_OP_ZPASS, wait_flag);
 733                 break;
 734         case PIPE_QUERY_PRIMITIVES_EMITTED:
 735         case PIPE_QUERY_PRIMITIVES_GENERATED:
 736         case PIPE_QUERY_SO_STATISTICS:
 737         case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 738                 r600_emit_query_predication(rctx, rquery, PREDICATION_OP_PRIMCOUNT, wait_flag);
 739                 break;
 740         default:
 741                 assert(0);
 742         }
 743 }
 744
 745 void r600_suspend_nontimer_queries(struct r600_common_context *ctx)
 746 {
 747         struct r600_query *query;
 748
 749         LIST_FOR_EACH_ENTRY(query, &ctx->active_nontimer_queries, list) {
 750                 r600_emit_query_end(ctx, query);
 751         }
 752         assert(ctx->num_cs_dw_nontimer_queries_suspend == 0);
 753 }
 754
 755 void r600_resume_nontimer_queries(struct r600_common_context *ctx)
 756 {
 757         struct r600_query *query;
 758
 759         assert(ctx->num_cs_dw_nontimer_queries_suspend == 0);
 760
 761         LIST_FOR_EACH_ENTRY(query, &ctx->active_nontimer_queries, list) {
 762                 r600_emit_query_begin(ctx, query);
 763         }
 764 }
 765
 766 /* Get backends mask */
 767 void r600_query_init_backend_mask(struct r600_common_context *ctx)
 768 {
 769         struct radeon_winsys_cs *cs = ctx->rings.gfx.cs;
 770         struct r600_resource *buffer;
 771         uint32_t *results;
 772         unsigned num_backends = ctx->screen->info.r600_num_backends;
 773         unsigned i, mask = 0;
 774         uint64_t va;
 775
 776         /* if backend_map query is supported by the kernel */
 777         if (ctx->screen->info.r600_backend_map_valid) {
 778                 unsigned num_tile_pipes = ctx->screen->info.r600_num_tile_pipes;
 779                 unsigned backend_map = ctx->screen->info.r600_backend_map;
 780                 unsigned item_width, item_mask;
 781
 782                 if (ctx->chip_class >= EVERGREEN) {
 783                         item_width = 4;
 784                         item_mask = 0x7;
 785                 } else {
 786                         item_width = 2;
 787                         item_mask = 0x3;
 788                 }
 789
 790                 while(num_tile_pipes--) {
 791                         i = backend_map & item_mask;
 792                         mask |= (1<<i);
 793                         backend_map >>= item_width;
 794                 }
 795                 if (mask != 0) {
 796                         ctx->backend_mask = mask;
 797                         return;
 798                 }
 799         }
 800
 801         /* otherwise backup path for older kernels */
 802
 803         /* create buffer for event data */
 804         buffer = (struct r600_resource*)
 805                 pipe_buffer_create(ctx->b.screen, PIPE_BIND_CUSTOM,
 806                                    PIPE_USAGE_STAGING, ctx->max_db*16);
 807         if (!buffer)
 808                 goto err;
 809         va = r600_resource_va(ctx->b.screen, (void*)buffer);
 810
 811         /* initialize buffer with zeroes */
 812         results = r600_buffer_map_sync_with_rings(ctx, buffer, PIPE_TRANSFER_WRITE);
 813         if (results) {
 814                 memset(results, 0, ctx->max_db * 4 * 4);
 815                 ctx->ws->buffer_unmap(buffer->cs_buf);
 816
 817                 /* emit EVENT_WRITE for ZPASS_DONE */
 818                 cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 2, 0);
 819                 cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_ZPASS_DONE) | EVENT_INDEX(1);
 820                 cs->buf[cs->cdw++] = va;
 821                 cs->buf[cs->cdw++] = (va >> 32UL) & 0xFF;
 822
 823                 cs->buf[cs->cdw++] = PKT3(PKT3_NOP, 0, 0);
 824                 cs->buf[cs->cdw++] = r600_context_bo_reloc(ctx, &ctx->rings.gfx, buffer, RADEON_USAGE_WRITE);
 825
 826                 /* analyze results */
 827                 results = r600_buffer_map_sync_with_rings(ctx, buffer, PIPE_TRANSFER_READ);
 828                 if (results) {
 829                         for(i = 0; i < ctx->max_db; i++) {
 830                                 /* at least highest bit will be set if backend is used */
 831                                 if (results[i*4 + 1])
 832                                         mask |= (1<<i);
 833                         }
 834                         ctx->ws->buffer_unmap(buffer->cs_buf);
 835                 }
 836         }
 837
 838         pipe_resource_reference((struct pipe_resource**)&buffer, NULL);
 839
 840         if (mask != 0) {
 841                 ctx->backend_mask = mask;
 842                 return;
 843         }
 844
 845 err:
 846         /* fallback to old method - set num_backends lower bits to 1 */
 847         ctx->backend_mask = (~((uint32_t)0))>>(32-num_backends);
 848         return;
 849 }
 850
 851 void r600_query_init(struct r600_common_context *rctx)
 852 {
 853         rctx->b.create_query = r600_create_query;
 854         rctx->b.destroy_query = r600_destroy_query;
 855         rctx->b.begin_query = r600_begin_query;
 856         rctx->b.end_query = r600_end_query;
 857         rctx->b.get_query_result = r600_get_query_result;
 858
 859         if (((struct r600_common_screen*)rctx->b.screen)->info.r600_num_backends > 0)
 860             rctx->b.render_condition = r600_render_condition;
 861
 862         LIST_INITHEAD(&rctx->active_nontimer_queries);
 863 }