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/u_format.h"
  33 #include "util/u_helpers.h"
  34
  35 #include "freedreno_draw.h"
  36 #include "freedreno_context.h"
  37 #include "freedreno_state.h"
  38 #include "freedreno_resource.h"
  39 #include "freedreno_query_acc.h"
  40 #include "freedreno_query_hw.h"
  41 #include "freedreno_util.h"
  42
  43 static void
  44 resource_read(struct fd_batch *batch, struct pipe_resource *prsc)
  45 {
  46         if (!prsc)
  47                 return;
  48         fd_batch_resource_used(batch, fd_resource(prsc), false);
  49 }
  50
  51 static void
  52 resource_written(struct fd_batch *batch, struct pipe_resource *prsc)
  53 {
  54         if (!prsc)
  55                 return;
  56         fd_batch_resource_used(batch, fd_resource(prsc), true);
  57 }
  58
  59 static void
  60 fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
  61 {
  62         struct fd_context *ctx = fd_context(pctx);
  63         struct fd_batch *batch = fd_context_batch(ctx);
  64         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
  65         struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
  66         unsigned i, prims, buffers = 0, restore_buffers = 0;
  67
  68         /* for debugging problems with indirect draw, it is convenient
  69          * to be able to emulate it, to determine if game is feeding us
  70          * bogus data:
  71          */
  72         if (info->indirect && (fd_mesa_debug & FD_DBG_NOINDR)) {
  73                 util_draw_indirect(pctx, info);
  74                 return;
  75         }
  76
  77         if (!info->count_from_stream_output && !info->indirect &&
  78             !info->primitive_restart &&
  79             !u_trim_pipe_prim(info->mode, (unsigned*)&info->count))
  80                 return;
  81
  82         /* if we supported transform feedback, we'd have to disable this: */
  83         if (((scissor->maxx - scissor->minx) *
  84                         (scissor->maxy - scissor->miny)) == 0) {
  85                 return;
  86         }
  87
  88         /* TODO: push down the region versions into the tiles */
  89         if (!fd_render_condition_check(pctx))
  90                 return;
  91
  92         /* emulate unsupported primitives: */
  93         if (!fd_supported_prim(ctx, info->mode)) {
  94                 if (ctx->streamout.num_targets > 0)
  95                         debug_error("stream-out with emulated prims");
  96                 util_primconvert_save_rasterizer_state(ctx->primconvert, ctx->rasterizer);
  97                 util_primconvert_draw_vbo(ctx->primconvert, info);
  98                 return;
  99         }
 100
 101         /* Upload a user index buffer. */
 102         struct pipe_resource *indexbuf = NULL;
 103         unsigned index_offset = 0;
 104         struct pipe_draw_info new_info;
 105         if (info->index_size) {
 106                 if (info->has_user_indices) {
 107                         if (!util_upload_index_buffer(pctx, info, &indexbuf, &index_offset))
 108                                 return;
 109                         new_info = *info;
 110                         new_info.index.resource = indexbuf;
 111                         new_info.has_user_indices = false;
 112                         info = &new_info;
 113                 } else {
 114                         indexbuf = info->index.resource;
 115                 }
 116         }
 117
 118         if (ctx->in_blit) {
 119                 fd_batch_reset(batch);
 120                 fd_context_all_dirty(ctx);
 121         }
 122
 123         batch->blit = ctx->in_blit;
 124         batch->back_blit = ctx->in_shadow;
 125
 126         /* NOTE: needs to be before resource_written(batch->query_buf), otherwise
 127          * query_buf may not be created yet.
 128          */
 129         fd_batch_set_stage(batch, FD_STAGE_DRAW);
 130
 131         /*
 132          * Figure out the buffers/features we need:
 133          */
 134
 135         mtx_lock(&ctx->screen->lock);
 136
 137         if (fd_depth_enabled(ctx)) {
 138                 if (fd_resource(pfb->zsbuf->texture)->valid) {
 139                         restore_buffers |= FD_BUFFER_DEPTH;
 140                 } else {
 141                         batch->invalidated |= FD_BUFFER_DEPTH;
 142                 }
 143                 buffers |= FD_BUFFER_DEPTH;
 144                 resource_written(batch, pfb->zsbuf->texture);
 145                 batch->gmem_reason |= FD_GMEM_DEPTH_ENABLED;
 146         }
 147
 148         if (fd_stencil_enabled(ctx)) {
 149                 if (fd_resource(pfb->zsbuf->texture)->valid) {
 150                         restore_buffers |= FD_BUFFER_STENCIL;
 151                 } else {
 152                         batch->invalidated |= FD_BUFFER_STENCIL;
 153                 }
 154                 buffers |= FD_BUFFER_STENCIL;
 155                 resource_written(batch, pfb->zsbuf->texture);
 156                 batch->gmem_reason |= FD_GMEM_STENCIL_ENABLED;
 157         }
 158
 159         if (fd_logicop_enabled(ctx))
 160                 batch->gmem_reason |= FD_GMEM_LOGICOP_ENABLED;
 161
 162         for (i = 0; i < pfb->nr_cbufs; i++) {
 163                 struct pipe_resource *surf;
 164
 165                 if (!pfb->cbufs[i])
 166                         continue;
 167
 168                 surf = pfb->cbufs[i]->texture;
 169
 170                 if (fd_resource(surf)->valid) {
 171                         restore_buffers |= PIPE_CLEAR_COLOR0 << i;
 172                 } else {
 173                         batch->invalidated |= PIPE_CLEAR_COLOR0 << i;
 174                 }
 175
 176                 resource_written(batch, surf);
 177
 178                 buffers |= PIPE_CLEAR_COLOR0 << i;
 179
 180                 if (fd_blend_enabled(ctx, i))
 181                         batch->gmem_reason |= FD_GMEM_BLEND_ENABLED;
 182         }
 183
 184         /* Mark SSBOs as being written.. we don't actually know which ones are
 185          * read vs written, so just assume the worst
 186          */
 187         foreach_bit(i, ctx->shaderbuf[PIPE_SHADER_FRAGMENT].enabled_mask)
 188                 resource_written(batch, ctx->shaderbuf[PIPE_SHADER_FRAGMENT].sb[i].buffer);
 189
 190         foreach_bit(i, ctx->shaderimg[PIPE_SHADER_FRAGMENT].enabled_mask) {
 191                 struct pipe_image_view *img =
 192                         &ctx->shaderimg[PIPE_SHADER_FRAGMENT].si[i];
 193                 if (img->access & PIPE_IMAGE_ACCESS_WRITE)
 194                         resource_written(batch, img->resource);
 195                 else
 196                         resource_read(batch, img->resource);
 197         }
 198
 199         foreach_bit(i, ctx->constbuf[PIPE_SHADER_VERTEX].enabled_mask)
 200                 resource_read(batch, ctx->constbuf[PIPE_SHADER_VERTEX].cb[i].buffer);
 201         foreach_bit(i, ctx->constbuf[PIPE_SHADER_FRAGMENT].enabled_mask)
 202                 resource_read(batch, ctx->constbuf[PIPE_SHADER_FRAGMENT].cb[i].buffer);
 203
 204         /* Mark VBOs as being read */
 205         foreach_bit(i, ctx->vtx.vertexbuf.enabled_mask) {
 206                 assert(!ctx->vtx.vertexbuf.vb[i].is_user_buffer);
 207                 resource_read(batch, ctx->vtx.vertexbuf.vb[i].buffer.resource);
 208         }
 209
 210         /* Mark index buffer as being read */
 211         resource_read(batch, indexbuf);
 212
 213         /* Mark indirect draw buffer as being read */
 214         if (info->indirect)
 215                 resource_read(batch, info->indirect->buffer);
 216
 217         /* Mark textures as being read */
 218         foreach_bit(i, ctx->tex[PIPE_SHADER_VERTEX].valid_textures)
 219                 resource_read(batch, ctx->tex[PIPE_SHADER_VERTEX].textures[i]->texture);
 220         foreach_bit(i, ctx->tex[PIPE_SHADER_FRAGMENT].valid_textures)
 221                 resource_read(batch, ctx->tex[PIPE_SHADER_FRAGMENT].textures[i]->texture);
 222
 223         /* Mark streamout buffers as being written.. */
 224         for (i = 0; i < ctx->streamout.num_targets; i++)
 225                 if (ctx->streamout.targets[i])
 226                         resource_written(batch, ctx->streamout.targets[i]->buffer);
 227
 228         resource_written(batch, batch->query_buf);
 229
 230         list_for_each_entry(struct fd_acc_query, aq, &ctx->acc_active_queries, node)
 231                 resource_written(batch, aq->prsc);
 232
 233         mtx_unlock(&ctx->screen->lock);
 234
 235         batch->num_draws++;
 236
 237         prims = u_reduced_prims_for_vertices(info->mode, info->count);
 238
 239         ctx->stats.draw_calls++;
 240
 241         /* TODO prims_emitted should be clipped when the stream-out buffer is
 242          * not large enough.  See max_tf_vtx().. probably need to move that
 243          * into common code.  Although a bit more annoying since a2xx doesn't
 244          * use ir3 so no common way to get at the pipe_stream_output_info
 245          * which is needed for this calculation.
 246          */
 247         if (ctx->streamout.num_targets > 0)
 248                 ctx->stats.prims_emitted += prims;
 249         ctx->stats.prims_generated += prims;
 250
 251         /* any buffers that haven't been cleared yet, we need to restore: */
 252         batch->restore |= restore_buffers & (FD_BUFFER_ALL & ~batch->invalidated);
 253         /* and any buffers used, need to be resolved: */
 254         batch->resolve |= buffers;
 255
 256         DBG("%p: %x %ux%u num_draws=%u (%s/%s)", batch, buffers,
 257                 pfb->width, pfb->height, batch->num_draws,
 258                 util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 259                 util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 260
 261         if (ctx->draw_vbo(ctx, info, index_offset))
 262                 batch->needs_flush = true;
 263
 264         for (i = 0; i < ctx->streamout.num_targets; i++)
 265                 ctx->streamout.offsets[i] += info->count;
 266
 267         if (fd_mesa_debug & FD_DBG_DDRAW)
 268                 fd_context_all_dirty(ctx);
 269
 270         fd_batch_check_size(batch);
 271
 272         if (info == &new_info)
 273                 pipe_resource_reference(&indexbuf, NULL);
 274 }
 275
 276 /* Generic clear implementation (partially) using u_blitter: */
 277 static void
 278 fd_blitter_clear(struct pipe_context *pctx, unsigned buffers,
 279                 const union pipe_color_union *color, double depth, unsigned stencil)
 280 {
 281         struct fd_context *ctx = fd_context(pctx);
 282         struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
 283         struct blitter_context *blitter = ctx->blitter;
 284
 285         fd_blitter_pipe_begin(ctx, false, true, FD_STAGE_CLEAR);
 286
 287         util_blitter_common_clear_setup(blitter, pfb->width, pfb->height,
 288                         buffers, NULL, NULL);
 289
 290         struct pipe_stencil_ref sr = {
 291                 .ref_value = { stencil & 0xff }
 292         };
 293         pctx->set_stencil_ref(pctx, &sr);
 294
 295         struct pipe_constant_buffer cb = {
 296                 .buffer_size = 16,
 297                 .user_buffer = &color->ui,
 298         };
 299         pctx->set_constant_buffer(pctx, PIPE_SHADER_FRAGMENT, 0, &cb);
 300
 301         if (!ctx->clear_rs_state) {
 302                 const struct pipe_rasterizer_state tmpl = {
 303                         .cull_face = PIPE_FACE_NONE,
 304                         .half_pixel_center = 1,
 305                         .bottom_edge_rule = 1,
 306                         .flatshade = 1,
 307                         .depth_clip_near = 1,
 308                         .depth_clip_far = 1,
 309                 };
 310                 ctx->clear_rs_state = pctx->create_rasterizer_state(pctx, &tmpl);
 311         }
 312         pctx->bind_rasterizer_state(pctx, ctx->clear_rs_state);
 313
 314         struct pipe_viewport_state vp = {
 315                 .scale     = { 0.5f * pfb->width, -0.5f * pfb->height, depth },
 316                 .translate = { 0.5f * pfb->width,  0.5f * pfb->height, 0.0f },
 317         };
 318         pctx->set_viewport_states(pctx, 0, 1, &vp);
 319
 320         pctx->bind_vertex_elements_state(pctx, ctx->solid_vbuf_state.vtx);
 321         pctx->set_vertex_buffers(pctx, blitter->vb_slot, 1,
 322                         &ctx->solid_vbuf_state.vertexbuf.vb[0]);
 323         pctx->set_stream_output_targets(pctx, 0, NULL, NULL);
 324         pctx->bind_vs_state(pctx, ctx->solid_prog.vp);
 325         pctx->bind_fs_state(pctx, ctx->solid_prog.fp);
 326
 327         struct pipe_draw_info info = {
 328                 .mode = PIPE_PRIM_MAX,    /* maps to DI_PT_RECTLIST */
 329                 .count = 2,
 330                 .max_index = 1,
 331                 .instance_count = 1,
 332         };
 333         ctx->draw_vbo(ctx, &info, 0);
 334
 335         util_blitter_restore_constant_buffer_state(blitter);
 336         util_blitter_restore_vertex_states(blitter);
 337         util_blitter_restore_fragment_states(blitter);
 338         util_blitter_restore_textures(blitter);
 339         util_blitter_restore_fb_state(blitter);
 340         util_blitter_restore_render_cond(blitter);
 341         util_blitter_unset_running_flag(blitter);
 342
 343         fd_blitter_pipe_end(ctx);
 344 }
 345
 346 static void
 347 fd_clear(struct pipe_context *pctx, unsigned buffers,
 348                 const union pipe_color_union *color, double depth, unsigned stencil)
 349 {
 350         struct fd_context *ctx = fd_context(pctx);
 351         struct fd_batch *batch = fd_context_batch(ctx);
 352         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
 353         unsigned cleared_buffers;
 354         int i;
 355
 356         /* TODO: push down the region versions into the tiles */
 357         if (!fd_render_condition_check(pctx))
 358                 return;
 359
 360         if (ctx->in_blit) {
 361                 fd_batch_reset(batch);
 362                 fd_context_all_dirty(ctx);
 363         }
 364
 365         /* pctx->clear() is only for full-surface clears, so scissor is
 366          * equivalent to having GL_SCISSOR_TEST disabled:
 367          */
 368         batch->max_scissor.minx = 0;
 369         batch->max_scissor.miny = 0;
 370         batch->max_scissor.maxx = pfb->width;
 371         batch->max_scissor.maxy = pfb->height;
 372
 373         /* for bookkeeping about which buffers have been cleared (and thus
 374          * can fully or partially skip mem2gmem) we need to ignore buffers
 375          * that have already had a draw, in case apps do silly things like
 376          * clear after draw (ie. if you only clear the color buffer, but
 377          * something like alpha-test causes side effects from the draw in
 378          * the depth buffer, etc)
 379          */
 380         cleared_buffers = buffers & (FD_BUFFER_ALL & ~batch->restore);
 381         batch->cleared |= cleared_buffers;
 382         batch->invalidated |= cleared_buffers;
 383
 384         batch->resolve |= buffers;
 385         batch->needs_flush = true;
 386
 387         mtx_lock(&ctx->screen->lock);
 388
 389         if (buffers & PIPE_CLEAR_COLOR)
 390                 for (i = 0; i < pfb->nr_cbufs; i++)
 391                         if (buffers & (PIPE_CLEAR_COLOR0 << i))
 392                                 resource_written(batch, pfb->cbufs[i]->texture);
 393
 394         if (buffers & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
 395                 resource_written(batch, pfb->zsbuf->texture);
 396                 batch->gmem_reason |= FD_GMEM_CLEARS_DEPTH_STENCIL;
 397         }
 398
 399         resource_written(batch, batch->query_buf);
 400
 401         list_for_each_entry(struct fd_acc_query, aq, &ctx->acc_active_queries, node)
 402                 resource_written(batch, aq->prsc);
 403
 404         mtx_unlock(&ctx->screen->lock);
 405
 406         DBG("%p: %x %ux%u depth=%f, stencil=%u (%s/%s)", batch, buffers,
 407                 pfb->width, pfb->height, depth, stencil,
 408                 util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 409                 util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 410
 411         /* if per-gen backend doesn't implement ctx->clear() generic
 412          * blitter clear:
 413          */
 414         bool fallback = true;
 415
 416         if (ctx->clear) {
 417                 fd_batch_set_stage(batch, FD_STAGE_CLEAR);
 418
 419                 if (ctx->clear(ctx, buffers, color, depth, stencil)) {
 420                         if (fd_mesa_debug & FD_DBG_DCLEAR)
 421                                 fd_context_all_dirty(ctx);
 422
 423                         fallback = false;
 424                 }
 425         }
 426
 427         if (fallback) {
 428                 fd_blitter_clear(pctx, buffers, color, depth, stencil);
 429         }
 430
 431         fd_batch_check_size(batch);
 432 }
 433
 434 static void
 435 fd_clear_render_target(struct pipe_context *pctx, struct pipe_surface *ps,
 436                 const union pipe_color_union *color,
 437                 unsigned x, unsigned y, unsigned w, unsigned h,
 438                 bool render_condition_enabled)
 439 {
 440         DBG("TODO: x=%u, y=%u, w=%u, h=%u", x, y, w, h);
 441 }
 442
 443 static void
 444 fd_clear_depth_stencil(struct pipe_context *pctx, struct pipe_surface *ps,
 445                 unsigned buffers, double depth, unsigned stencil,
 446                 unsigned x, unsigned y, unsigned w, unsigned h,
 447                 bool render_condition_enabled)
 448 {
 449         DBG("TODO: buffers=%u, depth=%f, stencil=%u, x=%u, y=%u, w=%u, h=%u",
 450                         buffers, depth, stencil, x, y, w, h);
 451 }
 452
 453 static void
 454 fd_launch_grid(struct pipe_context *pctx, const struct pipe_grid_info *info)
 455 {
 456         struct fd_context *ctx = fd_context(pctx);
 457         struct fd_batch *batch, *save_batch = NULL;
 458         unsigned i;
 459
 460         batch = fd_bc_alloc_batch(&ctx->screen->batch_cache, ctx, true);
 461         fd_batch_reference(&save_batch, ctx->batch);
 462         fd_batch_reference(&ctx->batch, batch);
 463
 464         mtx_lock(&ctx->screen->lock);
 465
 466         /* Mark SSBOs as being written.. we don't actually know which ones are
 467          * read vs written, so just assume the worst
 468          */
 469         foreach_bit(i, ctx->shaderbuf[PIPE_SHADER_COMPUTE].enabled_mask)
 470                 resource_written(batch, ctx->shaderbuf[PIPE_SHADER_COMPUTE].sb[i].buffer);
 471
 472         foreach_bit(i, ctx->shaderimg[PIPE_SHADER_COMPUTE].enabled_mask) {
 473                 struct pipe_image_view *img =
 474                         &ctx->shaderimg[PIPE_SHADER_COMPUTE].si[i];
 475                 if (img->access & PIPE_IMAGE_ACCESS_WRITE)
 476                         resource_written(batch, img->resource);
 477                 else
 478                         resource_read(batch, img->resource);
 479         }
 480
 481         /* UBO's are read */
 482         foreach_bit(i, ctx->constbuf[PIPE_SHADER_COMPUTE].enabled_mask)
 483                 resource_read(batch, ctx->constbuf[PIPE_SHADER_COMPUTE].cb[i].buffer);
 484
 485         /* Mark textures as being read */
 486         foreach_bit(i, ctx->tex[PIPE_SHADER_COMPUTE].valid_textures)
 487                 resource_read(batch, ctx->tex[PIPE_SHADER_COMPUTE].textures[i]->texture);
 488
 489         /* For global buffers, we don't really know if read or written, so assume
 490          * the worst:
 491          */
 492         foreach_bit(i, ctx->global_bindings.enabled_mask)
 493                 resource_written(batch, ctx->global_bindings.buf[i]);
 494
 495         if (info->indirect)
 496                 resource_read(batch, info->indirect);
 497
 498         mtx_unlock(&ctx->screen->lock);
 499
 500         batch->needs_flush = true;
 501         ctx->launch_grid(ctx, info);
 502
 503         fd_batch_flush(batch, false, false);
 504
 505         fd_batch_reference(&ctx->batch, save_batch);
 506         fd_batch_reference(&save_batch, NULL);
 507         fd_batch_reference(&batch, NULL);
 508 }
 509
 510 void
 511 fd_draw_init(struct pipe_context *pctx)
 512 {
 513         pctx->draw_vbo = fd_draw_vbo;
 514         pctx->clear = fd_clear;
 515         pctx->clear_render_target = fd_clear_render_target;
 516         pctx->clear_depth_stencil = fd_clear_depth_stencil;
 517
 518         if (has_compute(fd_screen(pctx->screen))) {
 519                 pctx->launch_grid = fd_launch_grid;
 520         }
 521 }