src/gallium/drivers/freedreno/freedreno_draw.c

   1 /*
   2  * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors:
  24  *    Rob Clark <robclark@freedesktop.org>
  25  */
  26
  27 #include "pipe/p_state.h"
  28 #include "util/u_draw.h"
  29 #include "util/u_string.h"
  30 #include "util/u_memory.h"
  31 #include "util/u_prim.h"
  32 #include "util/format/u_format.h"
  33 #include "util/u_helpers.h"
  34
  35 #include "freedreno_blitter.h"
  36 #include "freedreno_draw.h"
  37 #include "freedreno_context.h"
  38 #include "freedreno_fence.h"
  39 #include "freedreno_state.h"
  40 #include "freedreno_resource.h"
  41 #include "freedreno_query_acc.h"
  42 #include "freedreno_query_hw.h"
  43 #include "freedreno_util.h"
  44
  45 static void
  46 resource_read(struct fd_batch *batch, struct pipe_resource *prsc)
  47 {
  48         if (!prsc)
  49                 return;
  50         fd_batch_resource_read(batch, fd_resource(prsc));
  51 }
  52
  53 static void
  54 resource_written(struct fd_batch *batch, struct pipe_resource *prsc)
  55 {
  56         if (!prsc)
  57                 return;
  58         fd_batch_resource_write(batch, fd_resource(prsc));
  59 }
  60
  61 static void
  62 batch_draw_tracking(struct fd_batch *batch, const struct pipe_draw_info *info)
  63 {
  64         struct fd_context *ctx = batch->ctx;
  65         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
  66         unsigned buffers = 0, restore_buffers = 0;
  67
  68         /* NOTE: needs to be before resource_written(batch->query_buf), otherwise
  69          * query_buf may not be created yet.
  70          */
  71         fd_batch_set_stage(batch, FD_STAGE_DRAW);
  72
  73         /*
  74          * Figure out the buffers/features we need:
  75          */
  76
  77         fd_screen_lock(ctx->screen);
  78
  79         if (ctx->dirty & (FD_DIRTY_FRAMEBUFFER | FD_DIRTY_ZSA)) {
  80                 if (fd_depth_enabled(ctx)) {
  81                         if (fd_resource(pfb->zsbuf->texture)->valid) {
  82                                 restore_buffers |= FD_BUFFER_DEPTH;
  83                         } else {
  84                                 batch->invalidated |= FD_BUFFER_DEPTH;
  85                         }
  86                         batch->gmem_reason |= FD_GMEM_DEPTH_ENABLED;
  87                         if (fd_depth_write_enabled(ctx)) {
  88                                 buffers |= FD_BUFFER_DEPTH;
  89                                 resource_written(batch, pfb->zsbuf->texture);
  90                         } else {
  91                                 resource_read(batch, pfb->zsbuf->texture);
  92                         }
  93                 }
  94
  95                 if (fd_stencil_enabled(ctx)) {
  96                         if (fd_resource(pfb->zsbuf->texture)->valid) {
  97                                 restore_buffers |= FD_BUFFER_STENCIL;
  98                         } else {
  99                                 batch->invalidated |= FD_BUFFER_STENCIL;
 100                         }
 101                         batch->gmem_reason |= FD_GMEM_STENCIL_ENABLED;
 102                         buffers |= FD_BUFFER_STENCIL;
 103                         resource_written(batch, pfb->zsbuf->texture);
 104                 }
 105         }
 106
 107         if (fd_logicop_enabled(ctx))
 108                 batch->gmem_reason |= FD_GMEM_LOGICOP_ENABLED;
 109
 110         for (unsigned i = 0; i < pfb->nr_cbufs; i++) {
 111                 struct pipe_resource *surf;
 112
 113                 if (!pfb->cbufs[i])
 114                         continue;
 115
 116                 surf = pfb->cbufs[i]->texture;
 117
 118                 if (fd_resource(surf)->valid) {
 119                         restore_buffers |= PIPE_CLEAR_COLOR0 << i;
 120                 } else {
 121                         batch->invalidated |= PIPE_CLEAR_COLOR0 << i;
 122                 }
 123
 124                 buffers |= PIPE_CLEAR_COLOR0 << i;
 125
 126                 if (fd_blend_enabled(ctx, i))
 127                         batch->gmem_reason |= FD_GMEM_BLEND_ENABLED;
 128
 129                 if (ctx->dirty & FD_DIRTY_FRAMEBUFFER)
 130                         resource_written(batch, pfb->cbufs[i]->texture);
 131         }
 132
 133         /* Mark SSBOs */
 134         if (ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & FD_DIRTY_SHADER_SSBO) {
 135                 const struct fd_shaderbuf_stateobj *so = &ctx->shaderbuf[PIPE_SHADER_FRAGMENT];
 136
 137                 foreach_bit (i, so->enabled_mask & so->writable_mask)
 138                         resource_written(batch, so->sb[i].buffer);
 139
 140                 foreach_bit (i, so->enabled_mask & ~so->writable_mask)
 141                         resource_read(batch, so->sb[i].buffer);
 142         }
 143
 144         if (ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & FD_DIRTY_SHADER_IMAGE) {
 145                 foreach_bit (i, ctx->shaderimg[PIPE_SHADER_FRAGMENT].enabled_mask) {
 146                         struct pipe_image_view *img =
 147                                         &ctx->shaderimg[PIPE_SHADER_FRAGMENT].si[i];
 148                         if (img->access & PIPE_IMAGE_ACCESS_WRITE)
 149                                 resource_written(batch, img->resource);
 150                         else
 151                                 resource_read(batch, img->resource);
 152                 }
 153         }
 154
 155         if (ctx->dirty_shader[PIPE_SHADER_VERTEX] & FD_DIRTY_SHADER_CONST) {
 156                 foreach_bit (i, ctx->constbuf[PIPE_SHADER_VERTEX].enabled_mask)
 157                         resource_read(batch, ctx->constbuf[PIPE_SHADER_VERTEX].cb[i].buffer);
 158         }
 159
 160         if (ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & FD_DIRTY_SHADER_CONST) {
 161                 foreach_bit (i, ctx->constbuf[PIPE_SHADER_FRAGMENT].enabled_mask)
 162                         resource_read(batch, ctx->constbuf[PIPE_SHADER_FRAGMENT].cb[i].buffer);
 163         }
 164
 165         /* Mark VBOs as being read */
 166         if (ctx->dirty & FD_DIRTY_VTXBUF) {
 167                 foreach_bit (i, ctx->vtx.vertexbuf.enabled_mask) {
 168                         assert(!ctx->vtx.vertexbuf.vb[i].is_user_buffer);
 169                         resource_read(batch, ctx->vtx.vertexbuf.vb[i].buffer.resource);
 170                 }
 171         }
 172
 173         /* Mark index buffer as being read */
 174         if (info->index_size)
 175                 resource_read(batch, info->index.resource);
 176
 177         /* Mark indirect draw buffer as being read */
 178         if (info->indirect)
 179                 resource_read(batch, info->indirect->buffer);
 180
 181         /* Mark textures as being read */
 182         if (ctx->dirty_shader[PIPE_SHADER_VERTEX] & FD_DIRTY_SHADER_TEX) {
 183                 foreach_bit (i, ctx->tex[PIPE_SHADER_VERTEX].valid_textures)
 184                         resource_read(batch, ctx->tex[PIPE_SHADER_VERTEX].textures[i]->texture);
 185         }
 186
 187         if (ctx->dirty_shader[PIPE_SHADER_FRAGMENT] & FD_DIRTY_SHADER_TEX) {
 188                 foreach_bit (i, ctx->tex[PIPE_SHADER_FRAGMENT].valid_textures)
 189                         resource_read(batch, ctx->tex[PIPE_SHADER_FRAGMENT].textures[i]->texture);
 190         }
 191
 192         /* Mark streamout buffers as being written.. */
 193         if (ctx->dirty & FD_DIRTY_STREAMOUT) {
 194                 for (unsigned i = 0; i < ctx->streamout.num_targets; i++)
 195                         if (ctx->streamout.targets[i])
 196                                 resource_written(batch, ctx->streamout.targets[i]->buffer);
 197         }
 198
 199         resource_written(batch, batch->query_buf);
 200
 201         list_for_each_entry(struct fd_acc_query, aq, &ctx->acc_active_queries, node)
 202                 resource_written(batch, aq->prsc);
 203
 204         fd_screen_unlock(ctx->screen);
 205
 206         /* any buffers that haven't been cleared yet, we need to restore: */
 207         batch->restore |= restore_buffers & (FD_BUFFER_ALL & ~batch->invalidated);
 208         /* and any buffers used, need to be resolved: */
 209         batch->resolve |= buffers;
 210 }
 211
 212 static void
 213 fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
 214 {
 215         struct fd_context *ctx = fd_context(pctx);
 216
 217         /* for debugging problems with indirect draw, it is convenient
 218          * to be able to emulate it, to determine if game is feeding us
 219          * bogus data:
 220          */
 221         if (info->indirect && (fd_mesa_debug & FD_DBG_NOINDR)) {
 222                 util_draw_indirect(pctx, info);
 223                 return;
 224         }
 225
 226         if (!info->count_from_stream_output && !info->indirect &&
 227             !info->primitive_restart &&
 228             !u_trim_pipe_prim(info->mode, (unsigned*)&info->count))
 229                 return;
 230
 231         /* TODO: push down the region versions into the tiles */
 232         if (!fd_render_condition_check(pctx))
 233                 return;
 234
 235         /* emulate unsupported primitives: */
 236         if (!fd_supported_prim(ctx, info->mode)) {
 237                 if (ctx->streamout.num_targets > 0)
 238                         debug_error("stream-out with emulated prims");
 239                 util_primconvert_save_rasterizer_state(ctx->primconvert, ctx->rasterizer);
 240                 util_primconvert_draw_vbo(ctx->primconvert, info);
 241                 return;
 242         }
 243
 244         /* Upload a user index buffer. */
 245         struct pipe_resource *indexbuf = NULL;
 246         unsigned index_offset = 0;
 247         struct pipe_draw_info new_info;
 248         if (info->index_size) {
 249                 if (info->has_user_indices) {
 250                         if (!util_upload_index_buffer(pctx, info, &indexbuf, &index_offset, 4))
 251                                 return;
 252                         new_info = *info;
 253                         new_info.index.resource = indexbuf;
 254                         new_info.has_user_indices = false;
 255                         info = &new_info;
 256                 } else {
 257                         indexbuf = info->index.resource;
 258                 }
 259         }
 260
 261         struct fd_batch *batch = NULL;
 262         fd_batch_reference(&batch, fd_context_batch(ctx));
 263
 264         if (ctx->in_discard_blit) {
 265                 fd_batch_reset(batch);
 266                 fd_context_all_dirty(ctx);
 267         }
 268
 269         batch_draw_tracking(batch, info);
 270
 271         if (unlikely(ctx->batch != batch)) {
 272                 /* The current batch was flushed in batch_draw_tracking()
 273                  * so start anew.  We know this won't happen a second time
 274                  * since we are dealing with a fresh batch:
 275                  */
 276                 fd_batch_reference(&batch, fd_context_batch(ctx));
 277                 batch_draw_tracking(batch, info);
 278                 assert(ctx->batch == batch);
 279         }
 280
 281         batch->blit = ctx->in_discard_blit;
 282         batch->back_blit = ctx->in_shadow;
 283         batch->num_draws++;
 284
 285         /* Counting prims in sw doesn't work for GS and tesselation. For older
 286          * gens we don't have those stages and don't have the hw counters enabled,
 287          * so keep the count accurate for non-patch geometry.
 288          */
 289         unsigned prims;
 290         if (info->mode != PIPE_PRIM_PATCHES)
 291                 prims = u_reduced_prims_for_vertices(info->mode, info->count);
 292         else
 293                 prims = 0;
 294
 295         ctx->stats.draw_calls++;
 296
 297         /* TODO prims_emitted should be clipped when the stream-out buffer is
 298          * not large enough.  See max_tf_vtx().. probably need to move that
 299          * into common code.  Although a bit more annoying since a2xx doesn't
 300          * use ir3 so no common way to get at the pipe_stream_output_info
 301          * which is needed for this calculation.
 302          */
 303         if (ctx->streamout.num_targets > 0)
 304                 ctx->stats.prims_emitted += prims;
 305         ctx->stats.prims_generated += prims;
 306
 307         /* Clearing last_fence must come after the batch dependency tracking
 308          * (resource_read()/resource_written()), as that can trigger a flush,
 309          * re-populating last_fence
 310          */
 311         fd_fence_ref(&ctx->last_fence, NULL);
 312
 313         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
 314         DBG("%p: %ux%u num_draws=%u (%s/%s)", batch,
 315                 pfb->width, pfb->height, batch->num_draws,
 316                 util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 317                 util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 318
 319         if (ctx->draw_vbo(ctx, info, index_offset))
 320                 batch->needs_flush = true;
 321
 322         batch->num_vertices += info->count * info->instance_count;
 323
 324         for (unsigned i = 0; i < ctx->streamout.num_targets; i++)
 325                 ctx->streamout.offsets[i] += info->count;
 326
 327         if (fd_mesa_debug & FD_DBG_DDRAW)
 328                 fd_context_all_dirty(ctx);
 329
 330         fd_batch_check_size(batch);
 331         fd_batch_reference(&batch, NULL);
 332
 333         if (info == &new_info)
 334                 pipe_resource_reference(&indexbuf, NULL);
 335 }
 336
 337 static void
 338 batch_clear_tracking(struct fd_batch *batch, unsigned buffers)
 339 {
 340         struct fd_context *ctx = batch->ctx;
 341         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
 342         unsigned cleared_buffers;
 343
 344         /* pctx->clear() is only for full-surface clears, so scissor is
 345          * equivalent to having GL_SCISSOR_TEST disabled:
 346          */
 347         batch->max_scissor.minx = 0;
 348         batch->max_scissor.miny = 0;
 349         batch->max_scissor.maxx = pfb->width;
 350         batch->max_scissor.maxy = pfb->height;
 351
 352         /* for bookkeeping about which buffers have been cleared (and thus
 353          * can fully or partially skip mem2gmem) we need to ignore buffers
 354          * that have already had a draw, in case apps do silly things like
 355          * clear after draw (ie. if you only clear the color buffer, but
 356          * something like alpha-test causes side effects from the draw in
 357          * the depth buffer, etc)
 358          */
 359         cleared_buffers = buffers & (FD_BUFFER_ALL & ~batch->restore);
 360         batch->cleared |= buffers;
 361         batch->invalidated |= cleared_buffers;
 362
 363         batch->resolve |= buffers;
 364         batch->needs_flush = true;
 365
 366         fd_screen_lock(ctx->screen);
 367
 368         if (buffers & PIPE_CLEAR_COLOR)
 369                 for (unsigned i = 0; i < pfb->nr_cbufs; i++)
 370                         if (buffers & (PIPE_CLEAR_COLOR0 << i))
 371                                 resource_written(batch, pfb->cbufs[i]->texture);
 372
 373         if (buffers & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
 374                 resource_written(batch, pfb->zsbuf->texture);
 375                 batch->gmem_reason |= FD_GMEM_CLEARS_DEPTH_STENCIL;
 376         }
 377
 378         resource_written(batch, batch->query_buf);
 379
 380         list_for_each_entry(struct fd_acc_query, aq, &ctx->acc_active_queries, node)
 381                 resource_written(batch, aq->prsc);
 382
 383         fd_screen_unlock(ctx->screen);
 384 }
 385
 386 static void
 387 fd_clear(struct pipe_context *pctx, unsigned buffers,
 388                 const struct pipe_scissor_state *scissor_state,
 389                 const union pipe_color_union *color, double depth,
 390                 unsigned stencil)
 391 {
 392         struct fd_context *ctx = fd_context(pctx);
 393
 394         /* TODO: push down the region versions into the tiles */
 395         if (!fd_render_condition_check(pctx))
 396                 return;
 397
 398         struct fd_batch *batch = NULL;
 399         fd_batch_reference(&batch, fd_context_batch(ctx));
 400
 401         if (ctx->in_discard_blit) {
 402                 fd_batch_reset(batch);
 403                 fd_context_all_dirty(ctx);
 404         }
 405
 406         batch_clear_tracking(batch, buffers);
 407
 408         if (unlikely(ctx->batch != batch)) {
 409                 /* The current batch was flushed in batch_clear_tracking()
 410                  * so start anew.  We know this won't happen a second time
 411                  * since we are dealing with a fresh batch:
 412                  */
 413                 fd_batch_reference(&batch, fd_context_batch(ctx));
 414                 batch_clear_tracking(batch, buffers);
 415                 assert(ctx->batch == batch);
 416         }
 417
 418         /* Clearing last_fence must come after the batch dependency tracking
 419          * (resource_read()/resource_written()), as that can trigger a flush,
 420          * re-populating last_fence
 421          */
 422         fd_fence_ref(&ctx->last_fence, NULL);
 423
 424         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
 425         DBG("%p: %x %ux%u depth=%f, stencil=%u (%s/%s)", batch, buffers,
 426                 pfb->width, pfb->height, depth, stencil,
 427                 util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 428                 util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 429
 430         /* if per-gen backend doesn't implement ctx->clear() generic
 431          * blitter clear:
 432          */
 433         bool fallback = true;
 434
 435         if (ctx->clear) {
 436                 fd_batch_set_stage(batch, FD_STAGE_CLEAR);
 437
 438                 if (ctx->clear(ctx, buffers, color, depth, stencil)) {
 439                         if (fd_mesa_debug & FD_DBG_DCLEAR)
 440                                 fd_context_all_dirty(ctx);
 441
 442                         fallback = false;
 443                 }
 444         }
 445
 446         if (fallback) {
 447                 fd_blitter_clear(pctx, buffers, color, depth, stencil);
 448         }
 449
 450         fd_batch_check_size(batch);
 451         fd_batch_reference(&batch, NULL);
 452 }
 453
 454 static void
 455 fd_clear_render_target(struct pipe_context *pctx, struct pipe_surface *ps,
 456                 const union pipe_color_union *color,
 457                 unsigned x, unsigned y, unsigned w, unsigned h,
 458                 bool render_condition_enabled)
 459 {
 460         DBG("TODO: x=%u, y=%u, w=%u, h=%u", x, y, w, h);
 461 }
 462
 463 static void
 464 fd_clear_depth_stencil(struct pipe_context *pctx, struct pipe_surface *ps,
 465                 unsigned buffers, double depth, unsigned stencil,
 466                 unsigned x, unsigned y, unsigned w, unsigned h,
 467                 bool render_condition_enabled)
 468 {
 469         DBG("TODO: buffers=%u, depth=%f, stencil=%u, x=%u, y=%u, w=%u, h=%u",
 470                         buffers, depth, stencil, x, y, w, h);
 471 }
 472
 473 static void
 474 fd_launch_grid(struct pipe_context *pctx, const struct pipe_grid_info *info)
 475 {
 476         struct fd_context *ctx = fd_context(pctx);
 477         const struct fd_shaderbuf_stateobj *so = &ctx->shaderbuf[PIPE_SHADER_COMPUTE];
 478         struct fd_batch *batch, *save_batch = NULL;
 479
 480         batch = fd_bc_alloc_batch(&ctx->screen->batch_cache, ctx, true);
 481         fd_batch_reference(&save_batch, ctx->batch);
 482         fd_batch_reference(&ctx->batch, batch);
 483         fd_context_all_dirty(ctx);
 484
 485         fd_screen_lock(ctx->screen);
 486
 487         /* Mark SSBOs */
 488         foreach_bit (i, so->enabled_mask & so->writable_mask)
 489                 resource_written(batch, so->sb[i].buffer);
 490
 491         foreach_bit (i, so->enabled_mask & ~so->writable_mask)
 492                 resource_read(batch, so->sb[i].buffer);
 493
 494         foreach_bit(i, ctx->shaderimg[PIPE_SHADER_COMPUTE].enabled_mask) {
 495                 struct pipe_image_view *img =
 496                         &ctx->shaderimg[PIPE_SHADER_COMPUTE].si[i];
 497                 if (img->access & PIPE_IMAGE_ACCESS_WRITE)
 498                         resource_written(batch, img->resource);
 499                 else
 500                         resource_read(batch, img->resource);
 501         }
 502
 503         /* UBO's are read */
 504         foreach_bit(i, ctx->constbuf[PIPE_SHADER_COMPUTE].enabled_mask)
 505                 resource_read(batch, ctx->constbuf[PIPE_SHADER_COMPUTE].cb[i].buffer);
 506
 507         /* Mark textures as being read */
 508         foreach_bit(i, ctx->tex[PIPE_SHADER_COMPUTE].valid_textures)
 509                 resource_read(batch, ctx->tex[PIPE_SHADER_COMPUTE].textures[i]->texture);
 510
 511         /* For global buffers, we don't really know if read or written, so assume
 512          * the worst:
 513          */
 514         foreach_bit(i, ctx->global_bindings.enabled_mask)
 515                 resource_written(batch, ctx->global_bindings.buf[i]);
 516
 517         if (info->indirect)
 518                 resource_read(batch, info->indirect);
 519
 520         fd_screen_unlock(ctx->screen);
 521
 522         batch->needs_flush = true;
 523         ctx->launch_grid(ctx, info);
 524
 525         fd_batch_flush(batch);
 526
 527         fd_batch_reference(&ctx->batch, save_batch);
 528         fd_context_all_dirty(ctx);
 529         fd_batch_reference(&save_batch, NULL);
 530         fd_batch_reference(&batch, NULL);
 531 }
 532
 533 void
 534 fd_draw_init(struct pipe_context *pctx)
 535 {
 536         pctx->draw_vbo = fd_draw_vbo;
 537         pctx->clear = fd_clear;
 538         pctx->clear_render_target = fd_clear_render_target;
 539         pctx->clear_depth_stencil = fd_clear_depth_stencil;
 540
 541         if (has_compute(fd_screen(pctx->screen))) {
 542                 pctx->launch_grid = fd_launch_grid;
 543         }
 544 }