src/gallium/drivers/freedreno/freedreno_draw.c

   1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
   2
   3 /*
   4  * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice (including the next
  14  * paragraph) shall be included in all copies or substantial portions of the
  15  * Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  23  * SOFTWARE.
  24  *
  25  * Authors:
  26  *    Rob Clark <robclark@freedesktop.org>
  27  */
  28
  29 #include "pipe/p_state.h"
  30 #include "util/u_draw.h"
  31 #include "util/u_string.h"
  32 #include "util/u_memory.h"
  33 #include "util/u_prim.h"
  34 #include "util/u_format.h"
  35 #include "util/u_helpers.h"
  36
  37 #include "freedreno_draw.h"
  38 #include "freedreno_context.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_used(batch, fd_resource(prsc), false);
  51 }
  52
  53 static void
  54 resource_written(struct fd_batch *batch, struct pipe_resource *prsc)
  55 {
  56         if (!prsc)
  57                 return;
  58         fd_resource(prsc)->valid = true;
  59         fd_batch_resource_used(batch, fd_resource(prsc), true);
  60 }
  61
  62 static void
  63 fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
  64 {
  65         struct fd_context *ctx = fd_context(pctx);
  66         struct fd_batch *batch = ctx->batch;
  67         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
  68         struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
  69         unsigned i, prims, buffers = 0, restore_buffers = 0;
  70
  71         /* for debugging problems with indirect draw, it is convenient
  72          * to be able to emulate it, to determine if game is feeding us
  73          * bogus data:
  74          */
  75         if (info->indirect && (fd_mesa_debug & FD_DBG_NOINDR)) {
  76                 util_draw_indirect(pctx, info);
  77                 return;
  78         }
  79
  80         if (!info->count_from_stream_output && !info->indirect &&
  81             !info->primitive_restart &&
  82             !u_trim_pipe_prim(info->mode, (unsigned*)&info->count))
  83                 return;
  84
  85         /* if we supported transform feedback, we'd have to disable this: */
  86         if (((scissor->maxx - scissor->minx) *
  87                         (scissor->maxy - scissor->miny)) == 0) {
  88                 return;
  89         }
  90
  91         /* TODO: push down the region versions into the tiles */
  92         if (!fd_render_condition_check(pctx))
  93                 return;
  94
  95         /* emulate unsupported primitives: */
  96         if (!fd_supported_prim(ctx, info->mode)) {
  97                 if (ctx->streamout.num_targets > 0)
  98                         debug_error("stream-out with emulated prims");
  99                 util_primconvert_save_rasterizer_state(ctx->primconvert, ctx->rasterizer);
 100                 util_primconvert_draw_vbo(ctx->primconvert, info);
 101                 return;
 102         }
 103
 104         /* Upload a user index buffer. */
 105         struct pipe_resource *indexbuf = NULL;
 106         unsigned index_offset = 0;
 107         struct pipe_draw_info new_info;
 108         if (info->index_size) {
 109                 if (info->has_user_indices) {
 110                         if (!util_upload_index_buffer(pctx, info, &indexbuf, &index_offset))
 111                                 return;
 112                         new_info = *info;
 113                         new_info.index.resource = indexbuf;
 114                         new_info.has_user_indices = false;
 115                         info = &new_info;
 116                 } else {
 117                         indexbuf = info->index.resource;
 118                 }
 119         }
 120
 121         if (ctx->in_blit) {
 122                 fd_batch_reset(batch);
 123                 fd_context_all_dirty(ctx);
 124         }
 125
 126         batch->blit = ctx->in_blit;
 127         batch->back_blit = ctx->in_shadow;
 128
 129         /* NOTE: needs to be before resource_written(batch->query_buf), otherwise
 130          * query_buf may not be created yet.
 131          */
 132         fd_batch_set_stage(batch, FD_STAGE_DRAW);
 133
 134         /*
 135          * Figure out the buffers/features we need:
 136          */
 137
 138         mtx_lock(&ctx->screen->lock);
 139
 140         if (fd_depth_enabled(ctx)) {
 141                 if (fd_resource(pfb->zsbuf->texture)->valid)
 142                         restore_buffers |= FD_BUFFER_DEPTH;
 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_DEPTH;
 151                 buffers |= FD_BUFFER_STENCIL;
 152                 resource_written(batch, pfb->zsbuf->texture);
 153                 batch->gmem_reason |= FD_GMEM_STENCIL_ENABLED;
 154         }
 155
 156         if (fd_logicop_enabled(ctx))
 157                 batch->gmem_reason |= FD_GMEM_LOGICOP_ENABLED;
 158
 159         for (i = 0; i < pfb->nr_cbufs; i++) {
 160                 struct pipe_resource *surf;
 161
 162                 if (!pfb->cbufs[i])
 163                         continue;
 164
 165                 surf = pfb->cbufs[i]->texture;
 166
 167                 resource_written(batch, surf);
 168
 169                 if (fd_resource(surf)->valid)
 170                         restore_buffers |= PIPE_CLEAR_COLOR0 << i;
 171
 172                 buffers |= PIPE_CLEAR_COLOR0 << i;
 173
 174                 if (surf->nr_samples > 1)
 175                         batch->gmem_reason |= FD_GMEM_MSAA_ENABLED;
 176
 177                 if (fd_blend_enabled(ctx, i))
 178                         batch->gmem_reason |= FD_GMEM_BLEND_ENABLED;
 179         }
 180
 181         /* Mark SSBOs as being written.. we don't actually know which ones are
 182          * read vs written, so just assume the worst
 183          */
 184         foreach_bit(i, ctx->shaderbuf[PIPE_SHADER_FRAGMENT].enabled_mask)
 185                 resource_written(batch, ctx->shaderbuf[PIPE_SHADER_FRAGMENT].sb[i].buffer);
 186
 187         foreach_bit(i, ctx->shaderimg[PIPE_SHADER_FRAGMENT].enabled_mask) {
 188                 struct pipe_image_view *img =
 189                         &ctx->shaderimg[PIPE_SHADER_FRAGMENT].si[i];
 190                 if (img->access & PIPE_IMAGE_ACCESS_WRITE)
 191                         resource_written(batch, img->resource);
 192                 else
 193                         resource_read(batch, img->resource);
 194         }
 195
 196         foreach_bit(i, ctx->constbuf[PIPE_SHADER_VERTEX].enabled_mask)
 197                 resource_read(batch, ctx->constbuf[PIPE_SHADER_VERTEX].cb[i].buffer);
 198         foreach_bit(i, ctx->constbuf[PIPE_SHADER_FRAGMENT].enabled_mask)
 199                 resource_read(batch, ctx->constbuf[PIPE_SHADER_FRAGMENT].cb[i].buffer);
 200
 201         /* Mark VBOs as being read */
 202         foreach_bit(i, ctx->vtx.vertexbuf.enabled_mask) {
 203                 assert(!ctx->vtx.vertexbuf.vb[i].is_user_buffer);
 204                 resource_read(batch, ctx->vtx.vertexbuf.vb[i].buffer.resource);
 205         }
 206
 207         /* Mark index buffer as being read */
 208         resource_read(batch, indexbuf);
 209
 210         /* Mark indirect draw buffer as being read */
 211         if (info->indirect)
 212                 resource_read(batch, info->indirect->buffer);
 213
 214         /* Mark textures as being read */
 215         foreach_bit(i, ctx->tex[PIPE_SHADER_VERTEX].valid_textures)
 216                 resource_read(batch, ctx->tex[PIPE_SHADER_VERTEX].textures[i]->texture);
 217         foreach_bit(i, ctx->tex[PIPE_SHADER_FRAGMENT].valid_textures)
 218                 resource_read(batch, ctx->tex[PIPE_SHADER_FRAGMENT].textures[i]->texture);
 219
 220         /* Mark streamout buffers as being written.. */
 221         for (i = 0; i < ctx->streamout.num_targets; i++)
 222                 if (ctx->streamout.targets[i])
 223                         resource_written(batch, ctx->streamout.targets[i]->buffer);
 224
 225         resource_written(batch, batch->query_buf);
 226
 227         list_for_each_entry(struct fd_acc_query, aq, &ctx->acc_active_queries, node)
 228                 resource_written(batch, aq->prsc);
 229
 230         mtx_unlock(&ctx->screen->lock);
 231
 232         batch->num_draws++;
 233
 234         prims = u_reduced_prims_for_vertices(info->mode, info->count);
 235
 236         ctx->stats.draw_calls++;
 237
 238         /* TODO prims_emitted should be clipped when the stream-out buffer is
 239          * not large enough.  See max_tf_vtx().. probably need to move that
 240          * into common code.  Although a bit more annoying since a2xx doesn't
 241          * use ir3 so no common way to get at the pipe_stream_output_info
 242          * which is needed for this calculation.
 243          */
 244         if (ctx->streamout.num_targets > 0)
 245                 ctx->stats.prims_emitted += prims;
 246         ctx->stats.prims_generated += prims;
 247
 248         /* any buffers that haven't been cleared yet, we need to restore: */
 249         batch->restore |= restore_buffers & (FD_BUFFER_ALL & ~batch->cleared);
 250         /* and any buffers used, need to be resolved: */
 251         batch->resolve |= buffers;
 252
 253         DBG("%p: %x %ux%u num_draws=%u (%s/%s)", batch, buffers,
 254                 pfb->width, pfb->height, batch->num_draws,
 255                 util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 256                 util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 257
 258         if (ctx->draw_vbo(ctx, info, index_offset))
 259                 batch->needs_flush = true;
 260
 261         for (i = 0; i < ctx->streamout.num_targets; i++)
 262                 ctx->streamout.offsets[i] += info->count;
 263
 264         if (fd_mesa_debug & FD_DBG_DDRAW)
 265                 fd_context_all_dirty(ctx);
 266
 267         fd_batch_check_size(batch);
 268
 269         if (info == &new_info)
 270                 pipe_resource_reference(&indexbuf, NULL);
 271 }
 272
 273 /* Generic clear implementation (partially) using u_blitter: */
 274 static void
 275 fd_blitter_clear(struct pipe_context *pctx, unsigned buffers,
 276                 const union pipe_color_union *color, double depth, unsigned stencil)
 277 {
 278         struct fd_context *ctx = fd_context(pctx);
 279         struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
 280         struct blitter_context *blitter = ctx->blitter;
 281
 282         fd_blitter_pipe_begin(ctx, false, true, FD_STAGE_CLEAR);
 283
 284         util_blitter_common_clear_setup(blitter, pfb->width, pfb->height,
 285                         buffers, NULL, NULL);
 286
 287         struct pipe_stencil_ref sr = {
 288                 .ref_value = { stencil & 0xff }
 289         };
 290         pctx->set_stencil_ref(pctx, &sr);
 291
 292         struct pipe_constant_buffer cb = {
 293                 .buffer_size = 16,
 294                 .user_buffer = &color->ui,
 295         };
 296         pctx->set_constant_buffer(pctx, PIPE_SHADER_FRAGMENT, 0, &cb);
 297
 298         if (!ctx->clear_rs_state) {
 299                 const struct pipe_rasterizer_state tmpl = {
 300                         .cull_face = PIPE_FACE_NONE,
 301                         .half_pixel_center = 1,
 302                         .bottom_edge_rule = 1,
 303                         .flatshade = 1,
 304                         .depth_clip = 1,
 305                 };
 306                 ctx->clear_rs_state = pctx->create_rasterizer_state(pctx, &tmpl);
 307         }
 308         pctx->bind_rasterizer_state(pctx, ctx->clear_rs_state);
 309
 310         struct pipe_viewport_state vp = {
 311                 .scale     = { 0.5f * pfb->width, -0.5f * pfb->height, depth },
 312                 .translate = { 0.5f * pfb->width,  0.5f * pfb->height, 0.0f },
 313         };
 314         pctx->set_viewport_states(pctx, 0, 1, &vp);
 315
 316         pctx->bind_vertex_elements_state(pctx, ctx->solid_vbuf_state.vtx);
 317         pctx->set_vertex_buffers(pctx, blitter->vb_slot, 1,
 318                         &ctx->solid_vbuf_state.vertexbuf.vb[0]);
 319         pctx->set_stream_output_targets(pctx, 0, NULL, NULL);
 320         pctx->bind_vs_state(pctx, ctx->solid_prog.vp);
 321         pctx->bind_fs_state(pctx, ctx->solid_prog.fp);
 322
 323         struct pipe_draw_info info = {
 324                 .mode = PIPE_PRIM_MAX,    /* maps to DI_PT_RECTLIST */
 325                 .count = 2,
 326                 .max_index = 1,
 327                 .instance_count = 1,
 328         };
 329         ctx->draw_vbo(ctx, &info, 0);
 330
 331         util_blitter_restore_constant_buffer_state(blitter);
 332         util_blitter_restore_vertex_states(blitter);
 333         util_blitter_restore_fragment_states(blitter);
 334         util_blitter_restore_textures(blitter);
 335         util_blitter_restore_fb_state(blitter);
 336         util_blitter_restore_render_cond(blitter);
 337         util_blitter_unset_running_flag(blitter);
 338
 339         fd_blitter_pipe_end(ctx);
 340 }
 341
 342 /* TODO figure out how to make better use of existing state mechanism
 343  * for clear (and possibly gmem->mem / mem->gmem) so we can (a) keep
 344  * track of what state really actually changes, and (b) reduce the code
 345  * in the a2xx/a3xx parts.
 346  */
 347
 348 static void
 349 fd_clear(struct pipe_context *pctx, unsigned buffers,
 350                 const union pipe_color_union *color, double depth, unsigned stencil)
 351 {
 352         struct fd_context *ctx = fd_context(pctx);
 353         struct fd_batch *batch = ctx->batch;
 354         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
 355         struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
 356         unsigned cleared_buffers;
 357         int i;
 358
 359         /* TODO: push down the region versions into the tiles */
 360         if (!fd_render_condition_check(pctx))
 361                 return;
 362
 363         if (ctx->in_blit) {
 364                 fd_batch_reset(batch);
 365                 fd_context_all_dirty(ctx);
 366         }
 367
 368         /* for bookkeeping about which buffers have been cleared (and thus
 369          * can fully or partially skip mem2gmem) we need to ignore buffers
 370          * that have already had a draw, in case apps do silly things like
 371          * clear after draw (ie. if you only clear the color buffer, but
 372          * something like alpha-test causes side effects from the draw in
 373          * the depth buffer, etc)
 374          */
 375         cleared_buffers = buffers & (FD_BUFFER_ALL & ~batch->restore);
 376
 377         /* do we have full-screen scissor? */
 378         if (!memcmp(scissor, &ctx->disabled_scissor, sizeof(*scissor))) {
 379                 batch->cleared |= cleared_buffers;
 380         } else {
 381                 batch->partial_cleared |= cleared_buffers;
 382                 if (cleared_buffers & PIPE_CLEAR_COLOR)
 383                         batch->cleared_scissor.color = *scissor;
 384                 if (cleared_buffers & PIPE_CLEAR_DEPTH)
 385                         batch->cleared_scissor.depth = *scissor;
 386                 if (cleared_buffers & PIPE_CLEAR_STENCIL)
 387                         batch->cleared_scissor.stencil = *scissor;
 388         }
 389         batch->resolve |= buffers;
 390         batch->needs_flush = true;
 391
 392         mtx_lock(&ctx->screen->lock);
 393
 394         if (buffers & PIPE_CLEAR_COLOR)
 395                 for (i = 0; i < pfb->nr_cbufs; i++)
 396                         if (buffers & (PIPE_CLEAR_COLOR0 << i))
 397                                 resource_written(batch, pfb->cbufs[i]->texture);
 398
 399         if (buffers & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
 400                 resource_written(batch, pfb->zsbuf->texture);
 401                 batch->gmem_reason |= FD_GMEM_CLEARS_DEPTH_STENCIL;
 402         }
 403
 404         resource_written(batch, batch->query_buf);
 405
 406         list_for_each_entry(struct fd_acc_query, aq, &ctx->acc_active_queries, node)
 407                 resource_written(batch, aq->prsc);
 408
 409         mtx_unlock(&ctx->screen->lock);
 410
 411         DBG("%p: %x %ux%u depth=%f, stencil=%u (%s/%s)", batch, buffers,
 412                 pfb->width, pfb->height, depth, stencil,
 413                 util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 414                 util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 415
 416         /* if per-gen backend doesn't implement ctx->clear() generic
 417          * blitter clear:
 418          */
 419         bool fallback = true;
 420
 421         if (ctx->clear) {
 422                 fd_batch_set_stage(batch, FD_STAGE_CLEAR);
 423
 424                 if (ctx->clear(ctx, buffers, color, depth, stencil)) {
 425                         if (fd_mesa_debug & FD_DBG_DCLEAR)
 426                                 fd_context_all_dirty(ctx);
 427
 428                         fallback = false;
 429                 }
 430         }
 431
 432         if (fallback) {
 433                 fd_blitter_clear(pctx, buffers, color, depth, stencil);
 434         }
 435 }
 436
 437 static void
 438 fd_clear_render_target(struct pipe_context *pctx, struct pipe_surface *ps,
 439                 const union pipe_color_union *color,
 440                 unsigned x, unsigned y, unsigned w, unsigned h,
 441                 bool render_condition_enabled)
 442 {
 443         DBG("TODO: x=%u, y=%u, w=%u, h=%u", x, y, w, h);
 444 }
 445
 446 static void
 447 fd_clear_depth_stencil(struct pipe_context *pctx, struct pipe_surface *ps,
 448                 unsigned buffers, double depth, unsigned stencil,
 449                 unsigned x, unsigned y, unsigned w, unsigned h,
 450                 bool render_condition_enabled)
 451 {
 452         DBG("TODO: buffers=%u, depth=%f, stencil=%u, x=%u, y=%u, w=%u, h=%u",
 453                         buffers, depth, stencil, x, y, w, h);
 454 }
 455
 456 static void
 457 fd_launch_grid(struct pipe_context *pctx, const struct pipe_grid_info *info)
 458 {
 459         struct fd_context *ctx = fd_context(pctx);
 460         struct fd_batch *batch, *save_batch = NULL;
 461         unsigned i;
 462
 463         batch = fd_batch_create(ctx);
 464         fd_batch_reference(&save_batch, ctx->batch);
 465         fd_batch_reference(&ctx->batch, batch);
 466
 467         mtx_lock(&ctx->screen->lock);
 468
 469         /* Mark SSBOs as being written.. we don't actually know which ones are
 470          * read vs written, so just assume the worst
 471          */
 472         foreach_bit(i, ctx->shaderbuf[PIPE_SHADER_COMPUTE].enabled_mask)
 473                 resource_read(batch, ctx->shaderbuf[PIPE_SHADER_COMPUTE].sb[i].buffer);
 474
 475         foreach_bit(i, ctx->shaderimg[PIPE_SHADER_COMPUTE].enabled_mask) {
 476                 struct pipe_image_view *img =
 477                         &ctx->shaderimg[PIPE_SHADER_COMPUTE].si[i];
 478                 if (img->access & PIPE_IMAGE_ACCESS_WRITE)
 479                         resource_written(batch, img->resource);
 480                 else
 481                         resource_read(batch, img->resource);
 482         }
 483
 484         /* UBO's are read */
 485         foreach_bit(i, ctx->constbuf[PIPE_SHADER_COMPUTE].enabled_mask)
 486                 resource_read(batch, ctx->constbuf[PIPE_SHADER_COMPUTE].cb[i].buffer);
 487
 488         /* Mark textures as being read */
 489         foreach_bit(i, ctx->tex[PIPE_SHADER_COMPUTE].valid_textures)
 490                 resource_read(batch, ctx->tex[PIPE_SHADER_COMPUTE].textures[i]->texture);
 491
 492         if (info->indirect)
 493                 resource_read(batch, info->indirect);
 494
 495         mtx_unlock(&ctx->screen->lock);
 496
 497         ctx->launch_grid(ctx, info);
 498
 499         fd_gmem_flush_compute(batch);
 500
 501         fd_batch_reference(&ctx->batch, save_batch);
 502         fd_batch_reference(&save_batch, NULL);
 503 }
 504
 505 void
 506 fd_draw_init(struct pipe_context *pctx)
 507 {
 508         pctx->draw_vbo = fd_draw_vbo;
 509         pctx->clear = fd_clear;
 510         pctx->clear_render_target = fd_clear_render_target;
 511         pctx->clear_depth_stencil = fd_clear_depth_stencil;
 512
 513         if (has_compute(fd_screen(pctx->screen))) {
 514                 pctx->launch_grid = fd_launch_grid;
 515         }
 516 }