src/mesa/drivers/dri/i965/genX_blorp_exec.c

   1 /*
   2  * Copyright © 2011 Intel Corporation
   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
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include <assert.h>
  25
  26 #include "intel_batchbuffer.h"
  27 #include "intel_mipmap_tree.h"
  28 #include "intel_fbo.h"
  29
  30 #include "brw_context.h"
  31 #include "brw_state.h"
  32
  33 #include "blorp/blorp_genX_exec.h"
  34
  35 #if GEN_GEN <= 5
  36 #include "gen4_blorp_exec.h"
  37 #endif
  38
  39 #include "brw_blorp.h"
  40
  41 static void *
  42 blorp_emit_dwords(struct blorp_batch *batch, unsigned n)
  43 {
  44    assert(batch->blorp->driver_ctx == batch->driver_batch);
  45    struct brw_context *brw = batch->driver_batch;
  46
  47    intel_batchbuffer_begin(brw, n);
  48    uint32_t *map = brw->batch.map_next;
  49    brw->batch.map_next += n;
  50    intel_batchbuffer_advance(brw);
  51    return map;
  52 }
  53
  54 static uint64_t
  55 blorp_emit_reloc(struct blorp_batch *batch,
  56                  void *location, struct blorp_address address, uint32_t delta)
  57 {
  58    assert(batch->blorp->driver_ctx == batch->driver_batch);
  59    struct brw_context *brw = batch->driver_batch;
  60    uint32_t offset;
  61
  62    if (GEN_GEN < 6 && brw_ptr_in_state_buffer(&brw->batch, location)) {
  63       offset = (char *)location - (char *)brw->batch.state.map;
  64       return brw_state_reloc(&brw->batch, offset,
  65                              address.buffer, address.offset + delta,
  66                              address.reloc_flags);
  67    }
  68
  69    assert(!brw_ptr_in_state_buffer(&brw->batch, location));
  70
  71    offset = (char *)location - (char *)brw->batch.batch.map;
  72    return brw_batch_reloc(&brw->batch, offset,
  73                           address.buffer, address.offset + delta,
  74                           address.reloc_flags);
  75 }
  76
  77 static void
  78 blorp_surface_reloc(struct blorp_batch *batch, uint32_t ss_offset,
  79                     struct blorp_address address, uint32_t delta)
  80 {
  81    assert(batch->blorp->driver_ctx == batch->driver_batch);
  82    struct brw_context *brw = batch->driver_batch;
  83    struct brw_bo *bo = address.buffer;
  84
  85    uint64_t reloc_val =
  86       brw_state_reloc(&brw->batch, ss_offset, bo, address.offset + delta,
  87                       address.reloc_flags);
  88
  89    void *reloc_ptr = (void *)brw->batch.state.map + ss_offset;
  90 #if GEN_GEN >= 8
  91    *(uint64_t *)reloc_ptr = reloc_val;
  92 #else
  93    *(uint32_t *)reloc_ptr = reloc_val;
  94 #endif
  95 }
  96
  97 static uint64_t
  98 blorp_get_surface_address(struct blorp_batch *blorp_batch,
  99                           struct blorp_address address)
 100 {
 101    /* We'll let blorp_surface_reloc write the address. */
 102    return 0ull;
 103 }
 104
 105 #if GEN_GEN >= 7 && GEN_GEN < 10
 106 static struct blorp_address
 107 blorp_get_surface_base_address(struct blorp_batch *batch)
 108 {
 109    assert(batch->blorp->driver_ctx == batch->driver_batch);
 110    struct brw_context *brw = batch->driver_batch;
 111    return (struct blorp_address) {
 112       .buffer = brw->batch.state.bo,
 113       .offset = 0,
 114    };
 115 }
 116 #endif
 117
 118 static void *
 119 blorp_alloc_dynamic_state(struct blorp_batch *batch,
 120                           uint32_t size,
 121                           uint32_t alignment,
 122                           uint32_t *offset)
 123 {
 124    assert(batch->blorp->driver_ctx == batch->driver_batch);
 125    struct brw_context *brw = batch->driver_batch;
 126
 127    return brw_state_batch(brw, size, alignment, offset);
 128 }
 129
 130 static void
 131 blorp_alloc_binding_table(struct blorp_batch *batch, unsigned num_entries,
 132                           unsigned state_size, unsigned state_alignment,
 133                           uint32_t *bt_offset, uint32_t *surface_offsets,
 134                           void **surface_maps)
 135 {
 136    assert(batch->blorp->driver_ctx == batch->driver_batch);
 137    struct brw_context *brw = batch->driver_batch;
 138
 139    uint32_t *bt_map = brw_state_batch(brw,
 140                                       num_entries * sizeof(uint32_t), 32,
 141                                       bt_offset);
 142
 143    for (unsigned i = 0; i < num_entries; i++) {
 144       surface_maps[i] = brw_state_batch(brw,
 145                                         state_size, state_alignment,
 146                                         &(surface_offsets)[i]);
 147       bt_map[i] = surface_offsets[i];
 148    }
 149 }
 150
 151 static void *
 152 blorp_alloc_vertex_buffer(struct blorp_batch *batch, uint32_t size,
 153                           struct blorp_address *addr)
 154 {
 155    assert(batch->blorp->driver_ctx == batch->driver_batch);
 156    struct brw_context *brw = batch->driver_batch;
 157
 158    /* From the Skylake PRM, 3DSTATE_VERTEX_BUFFERS:
 159     *
 160     *    "The VF cache needs to be invalidated before binding and then using
 161     *    Vertex Buffers that overlap with any previously bound Vertex Buffer
 162     *    (at a 64B granularity) since the last invalidation.  A VF cache
 163     *    invalidate is performed by setting the "VF Cache Invalidation Enable"
 164     *    bit in PIPE_CONTROL."
 165     *
 166     * This restriction first appears in the Skylake PRM but the internal docs
 167     * also list it as being an issue on Broadwell.  In order to avoid this
 168     * problem, we align all vertex buffer allocations to 64 bytes.
 169     */
 170    uint32_t offset;
 171    void *data = brw_state_batch(brw, size, 64, &offset);
 172
 173    *addr = (struct blorp_address) {
 174       .buffer = brw->batch.state.bo,
 175       .offset = offset,
 176
 177       /* The VF cache designers apparently cut corners, and made the cache
 178        * only consider the bottom 32 bits of memory addresses.  If you happen
 179        * to have two vertex buffers which get placed exactly 4 GiB apart and
 180        * use them in back-to-back draw calls, you can get collisions.  To work
 181        * around this problem, we restrict vertex buffers to the low 32 bits of
 182        * the address space.
 183        */
 184       .reloc_flags = RELOC_32BIT,
 185
 186 #if GEN_GEN == 11
 187       .mocs = ICL_MOCS_WB,
 188 #elif GEN_GEN == 10
 189       .mocs = CNL_MOCS_WB,
 190 #elif GEN_GEN == 9
 191       .mocs = SKL_MOCS_WB,
 192 #elif GEN_GEN == 8
 193       .mocs = BDW_MOCS_WB,
 194 #elif GEN_GEN == 7
 195       .mocs = GEN7_MOCS_L3,
 196 #elif GEN_GEN > 6
 197 #error "Missing MOCS setting!"
 198 #endif
 199    };
 200
 201    return data;
 202 }
 203
 204 /**
 205  * See vf_invalidate_for_vb_48b_transitions in genX_state_upload.c.
 206  */
 207 static void
 208 blorp_vf_invalidate_for_vb_48b_transitions(struct blorp_batch *batch,
 209                                            const struct blorp_address *addrs,
 210                                            UNUSED uint32_t *sizes,
 211                                            unsigned num_vbs)
 212 {
 213 #if GEN_GEN >= 8 && GEN_GEN < 11
 214    struct brw_context *brw = batch->driver_batch;
 215    bool need_invalidate = false;
 216
 217    for (unsigned i = 0; i < num_vbs; i++) {
 218       struct brw_bo *bo = addrs[i].buffer;
 219       uint16_t high_bits =
 220          bo && (bo->kflags & EXEC_OBJECT_PINNED) ? bo->gtt_offset >> 32u : 0;
 221
 222       if (high_bits != brw->vb.last_bo_high_bits[i]) {
 223          need_invalidate = true;
 224          brw->vb.last_bo_high_bits[i] = high_bits;
 225       }
 226    }
 227
 228    if (need_invalidate) {
 229       brw_emit_pipe_control_flush(brw, PIPE_CONTROL_VF_CACHE_INVALIDATE | PIPE_CONTROL_CS_STALL);
 230    }
 231 #endif
 232 }
 233
 234 #if GEN_GEN >= 8
 235 static struct blorp_address
 236 blorp_get_workaround_page(struct blorp_batch *batch)
 237 {
 238    assert(batch->blorp->driver_ctx == batch->driver_batch);
 239    struct brw_context *brw = batch->driver_batch;
 240
 241    return (struct blorp_address) {
 242       .buffer = brw->workaround_bo,
 243    };
 244 }
 245 #endif
 246
 247 static void
 248 blorp_flush_range(UNUSED struct blorp_batch *batch, UNUSED void *start,
 249                   UNUSED size_t size)
 250 {
 251    /* All allocated states come from the batch which we will flush before we
 252     * submit it.  There's nothing for us to do here.
 253     */
 254 }
 255
 256 static void
 257 blorp_emit_urb_config(struct blorp_batch *batch,
 258                       unsigned vs_entry_size,
 259                       UNUSED unsigned sf_entry_size)
 260 {
 261    assert(batch->blorp->driver_ctx == batch->driver_batch);
 262    struct brw_context *brw = batch->driver_batch;
 263
 264 #if GEN_GEN >= 7
 265    if (brw->urb.vsize >= vs_entry_size)
 266       return;
 267
 268    gen7_upload_urb(brw, vs_entry_size, false, false);
 269 #elif GEN_GEN == 6
 270    gen6_upload_urb(brw, vs_entry_size, false, 0);
 271 #else
 272    /* We calculate it now and emit later. */
 273    brw_calculate_urb_fence(brw, 0, vs_entry_size, sf_entry_size);
 274 #endif
 275 }
 276
 277 void
 278 genX(blorp_exec)(struct blorp_batch *batch,
 279                  const struct blorp_params *params)
 280 {
 281    assert(batch->blorp->driver_ctx == batch->driver_batch);
 282    struct brw_context *brw = batch->driver_batch;
 283    struct gl_context *ctx = &brw->ctx;
 284    bool check_aperture_failed_once = false;
 285
 286 #if GEN_GEN >= 11
 287    /* The PIPE_CONTROL command description says:
 288     *
 289     * "Whenever a Binding Table Index (BTI) used by a Render Taget Message
 290     *  points to a different RENDER_SURFACE_STATE, SW must issue a Render
 291     *  Target Cache Flush by enabling this bit. When render target flush
 292     *  is set due to new association of BTI, PS Scoreboard Stall bit must
 293     *  be set in this packet."
 294    */
 295    brw_emit_pipe_control_flush(brw,
 296                                PIPE_CONTROL_RENDER_TARGET_FLUSH |
 297                                PIPE_CONTROL_STALL_AT_SCOREBOARD);
 298 #endif
 299
 300    /* Flush the sampler and render caches.  We definitely need to flush the
 301     * sampler cache so that we get updated contents from the render cache for
 302     * the glBlitFramebuffer() source.  Also, we are sometimes warned in the
 303     * docs to flush the cache between reinterpretations of the same surface
 304     * data with different formats, which blorp does for stencil and depth
 305     * data.
 306     */
 307    if (params->src.enabled)
 308       brw_cache_flush_for_read(brw, params->src.addr.buffer);
 309    if (params->dst.enabled) {
 310       brw_cache_flush_for_render(brw, params->dst.addr.buffer,
 311                                  params->dst.view.format,
 312                                  params->dst.aux_usage);
 313    }
 314    if (params->depth.enabled)
 315       brw_cache_flush_for_depth(brw, params->depth.addr.buffer);
 316    if (params->stencil.enabled)
 317       brw_cache_flush_for_depth(brw, params->stencil.addr.buffer);
 318
 319    brw_select_pipeline(brw, BRW_RENDER_PIPELINE);
 320    brw_emit_l3_state(brw);
 321
 322 retry:
 323    intel_batchbuffer_require_space(brw, 1400);
 324    brw_require_statebuffer_space(brw, 600);
 325    intel_batchbuffer_save_state(brw);
 326    check_aperture_failed_once |= intel_batchbuffer_saved_state_is_empty(brw);
 327    brw->batch.no_wrap = true;
 328
 329 #if GEN_GEN == 6
 330    /* Emit workaround flushes when we switch from drawing to blorping. */
 331    brw_emit_post_sync_nonzero_flush(brw);
 332 #endif
 333
 334    brw_upload_state_base_address(brw);
 335
 336 #if GEN_GEN >= 8
 337    gen7_l3_state.emit(brw);
 338 #endif
 339
 340 #if GEN_GEN >= 6
 341    brw_emit_depth_stall_flushes(brw);
 342 #endif
 343
 344 #if GEN_GEN == 8
 345    gen8_write_pma_stall_bits(brw, 0);
 346 #endif
 347
 348    const unsigned scale = params->fast_clear_op ? UINT_MAX : 1;
 349    if (brw->current_hash_scale != scale) {
 350       brw_emit_hashing_mode(brw, params->x1 - params->x0,
 351                             params->y1 - params->y0, scale);
 352    }
 353
 354    blorp_emit(batch, GENX(3DSTATE_DRAWING_RECTANGLE), rect) {
 355       rect.ClippedDrawingRectangleXMax = MAX2(params->x1, params->x0) - 1;
 356       rect.ClippedDrawingRectangleYMax = MAX2(params->y1, params->y0) - 1;
 357    }
 358
 359    blorp_exec(batch, params);
 360
 361    brw->batch.no_wrap = false;
 362
 363    /* Check if the blorp op we just did would make our batch likely to fail to
 364     * map all the BOs into the GPU at batch exec time later.  If so, flush the
 365     * batch and try again with nothing else in the batch.
 366     */
 367    if (!brw_batch_has_aperture_space(brw, 0)) {
 368       if (!check_aperture_failed_once) {
 369          check_aperture_failed_once = true;
 370          intel_batchbuffer_reset_to_saved(brw);
 371          intel_batchbuffer_flush(brw);
 372          goto retry;
 373       } else {
 374          int ret = intel_batchbuffer_flush(brw);
 375          WARN_ONCE(ret == -ENOSPC,
 376                    "i965: blorp emit exceeded available aperture space\n");
 377       }
 378    }
 379
 380    if (unlikely(brw->always_flush_batch))
 381       intel_batchbuffer_flush(brw);
 382
 383    /* We've smashed all state compared to what the normal 3D pipeline
 384     * rendering tracks for GL.
 385     */
 386    brw->ctx.NewDriverState |= BRW_NEW_BLORP;
 387    brw->no_depth_or_stencil = !params->depth.enabled &&
 388                               !params->stencil.enabled;
 389    brw->ib.index_size = -1;
 390
 391    if (params->dst.enabled) {
 392       brw_render_cache_add_bo(brw, params->dst.addr.buffer,
 393                               params->dst.view.format,
 394                               params->dst.aux_usage);
 395    }
 396    if (params->depth.enabled)
 397       brw_depth_cache_add_bo(brw, params->depth.addr.buffer);
 398    if (params->stencil.enabled)
 399       brw_depth_cache_add_bo(brw, params->stencil.addr.buffer);
 400 }