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_batch_resource_used(batch, fd_resource(prsc), true);
  59 }
  60
  61 static void
  62 fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
  63 {
  64         struct fd_context *ctx = fd_context(pctx);
  65         struct fd_batch *batch = fd_context_batch(ctx);
  66         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
  67         struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
  68         unsigned i, prims, buffers = 0, restore_buffers = 0;
  69
  70         /* for debugging problems with indirect draw, it is convenient
  71          * to be able to emulate it, to determine if game is feeding us
  72          * bogus data:
  73          */
  74         if (info->indirect && (fd_mesa_debug & FD_DBG_NOINDR)) {
  75                 util_draw_indirect(pctx, info);
  76                 return;
  77         }
  78
  79         if (!info->count_from_stream_output && !info->indirect &&
  80             !info->primitive_restart &&
  81             !u_trim_pipe_prim(info->mode, (unsigned*)&info->count))
  82                 return;
  83
  84         /* if we supported transform feedback, we'd have to disable this: */
  85         if (((scissor->maxx - scissor->minx) *
  86                         (scissor->maxy - scissor->miny)) == 0) {
  87                 return;
  88         }
  89
  90         /* TODO: push down the region versions into the tiles */
  91         if (!fd_render_condition_check(pctx))
  92                 return;
  93
  94         /* emulate unsupported primitives: */
  95         if (!fd_supported_prim(ctx, info->mode)) {
  96                 if (ctx->streamout.num_targets > 0)
  97                         debug_error("stream-out with emulated prims");
  98                 util_primconvert_save_rasterizer_state(ctx->primconvert, ctx->rasterizer);
  99                 util_primconvert_draw_vbo(ctx->primconvert, info);
 100                 return;
 101         }
 102
 103         /* Upload a user index buffer. */
 104         struct pipe_resource *indexbuf = NULL;
 105         unsigned index_offset = 0;
 106         struct pipe_draw_info new_info;
 107         if (info->index_size) {
 108                 if (info->has_user_indices) {
 109                         if (!util_upload_index_buffer(pctx, info, &indexbuf, &index_offset))
 110                                 return;
 111                         new_info = *info;
 112                         new_info.index.resource = indexbuf;
 113                         new_info.has_user_indices = false;
 114                         info = &new_info;
 115                 } else {
 116                         indexbuf = info->index.resource;
 117                 }
 118         }
 119
 120         if (ctx->in_blit) {
 121                 fd_batch_reset(batch);
 122                 fd_context_all_dirty(ctx);
 123         }
 124
 125         batch->blit = ctx->in_blit;
 126         batch->back_blit = ctx->in_shadow;
 127
 128         /* NOTE: needs to be before resource_written(batch->query_buf), otherwise
 129          * query_buf may not be created yet.
 130          */
 131         fd_batch_set_stage(batch, FD_STAGE_DRAW);
 132
 133         /*
 134          * Figure out the buffers/features we need:
 135          */
 136
 137         mtx_lock(&ctx->screen->lock);
 138
 139         if (fd_depth_enabled(ctx)) {
 140                 if (fd_resource(pfb->zsbuf->texture)->valid) {
 141                         restore_buffers |= FD_BUFFER_DEPTH;
 142                 } else {
 143                         batch->invalidated |= FD_BUFFER_DEPTH;
 144                 }
 145                 buffers |= FD_BUFFER_DEPTH;
 146                 resource_written(batch, pfb->zsbuf->texture);
 147                 batch->gmem_reason |= FD_GMEM_DEPTH_ENABLED;
 148         }
 149
 150         if (fd_stencil_enabled(ctx)) {
 151                 if (fd_resource(pfb->zsbuf->texture)->valid) {
 152                         restore_buffers |= FD_BUFFER_STENCIL;
 153                 } else {
 154                         batch->invalidated |= FD_BUFFER_STENCIL;
 155                 }
 156                 buffers |= FD_BUFFER_STENCIL;
 157                 resource_written(batch, pfb->zsbuf->texture);
 158                 batch->gmem_reason |= FD_GMEM_STENCIL_ENABLED;
 159         }
 160
 161         if (fd_logicop_enabled(ctx))
 162                 batch->gmem_reason |= FD_GMEM_LOGICOP_ENABLED;
 163
 164         for (i = 0; i < pfb->nr_cbufs; i++) {
 165                 struct pipe_resource *surf;
 166
 167                 if (!pfb->cbufs[i])
 168                         continue;
 169
 170                 surf = pfb->cbufs[i]->texture;
 171
 172                 if (fd_resource(surf)->valid) {
 173                         restore_buffers |= PIPE_CLEAR_COLOR0 << i;
 174                 } else {
 175                         batch->invalidated |= PIPE_CLEAR_COLOR0 << i;
 176                 }
 177
 178                 resource_written(batch, surf);
 179
 180                 buffers |= PIPE_CLEAR_COLOR0 << i;
 181
 182                 if (fd_blend_enabled(ctx, i))
 183                         batch->gmem_reason |= FD_GMEM_BLEND_ENABLED;
 184         }
 185
 186         /* Mark SSBOs as being written.. we don't actually know which ones are
 187          * read vs written, so just assume the worst
 188          */
 189         foreach_bit(i, ctx->shaderbuf[PIPE_SHADER_FRAGMENT].enabled_mask)
 190                 resource_written(batch, ctx->shaderbuf[PIPE_SHADER_FRAGMENT].sb[i].buffer);
 191
 192         foreach_bit(i, ctx->shaderimg[PIPE_SHADER_FRAGMENT].enabled_mask) {
 193                 struct pipe_image_view *img =
 194                         &ctx->shaderimg[PIPE_SHADER_FRAGMENT].si[i];
 195                 if (img->access & PIPE_IMAGE_ACCESS_WRITE)
 196                         resource_written(batch, img->resource);
 197                 else
 198                         resource_read(batch, img->resource);
 199         }
 200
 201         foreach_bit(i, ctx->constbuf[PIPE_SHADER_VERTEX].enabled_mask)
 202                 resource_read(batch, ctx->constbuf[PIPE_SHADER_VERTEX].cb[i].buffer);
 203         foreach_bit(i, ctx->constbuf[PIPE_SHADER_FRAGMENT].enabled_mask)
 204                 resource_read(batch, ctx->constbuf[PIPE_SHADER_FRAGMENT].cb[i].buffer);
 205
 206         /* Mark VBOs as being read */
 207         foreach_bit(i, ctx->vtx.vertexbuf.enabled_mask) {
 208                 assert(!ctx->vtx.vertexbuf.vb[i].is_user_buffer);
 209                 resource_read(batch, ctx->vtx.vertexbuf.vb[i].buffer.resource);
 210         }
 211
 212         /* Mark index buffer as being read */
 213         resource_read(batch, indexbuf);
 214
 215         /* Mark indirect draw buffer as being read */
 216         if (info->indirect)
 217                 resource_read(batch, info->indirect->buffer);
 218
 219         /* Mark textures as being read */
 220         foreach_bit(i, ctx->tex[PIPE_SHADER_VERTEX].valid_textures)
 221                 resource_read(batch, ctx->tex[PIPE_SHADER_VERTEX].textures[i]->texture);
 222         foreach_bit(i, ctx->tex[PIPE_SHADER_FRAGMENT].valid_textures)
 223                 resource_read(batch, ctx->tex[PIPE_SHADER_FRAGMENT].textures[i]->texture);
 224
 225         /* Mark streamout buffers as being written.. */
 226         for (i = 0; i < ctx->streamout.num_targets; i++)
 227                 if (ctx->streamout.targets[i])
 228                         resource_written(batch, ctx->streamout.targets[i]->buffer);
 229
 230         resource_written(batch, batch->query_buf);
 231
 232         list_for_each_entry(struct fd_acc_query, aq, &ctx->acc_active_queries, node)
 233                 resource_written(batch, aq->prsc);
 234
 235         mtx_unlock(&ctx->screen->lock);
 236
 237         batch->num_draws++;
 238
 239         prims = u_reduced_prims_for_vertices(info->mode, info->count);
 240
 241         ctx->stats.draw_calls++;
 242
 243         /* TODO prims_emitted should be clipped when the stream-out buffer is
 244          * not large enough.  See max_tf_vtx().. probably need to move that
 245          * into common code.  Although a bit more annoying since a2xx doesn't
 246          * use ir3 so no common way to get at the pipe_stream_output_info
 247          * which is needed for this calculation.
 248          */
 249         if (ctx->streamout.num_targets > 0)
 250                 ctx->stats.prims_emitted += prims;
 251         ctx->stats.prims_generated += prims;
 252
 253         /* any buffers that haven't been cleared yet, we need to restore: */
 254         batch->restore |= restore_buffers & (FD_BUFFER_ALL & ~batch->invalidated);
 255         /* and any buffers used, need to be resolved: */
 256         batch->resolve |= buffers;
 257
 258         DBG("%p: %x %ux%u num_draws=%u (%s/%s)", batch, buffers,
 259                 pfb->width, pfb->height, batch->num_draws,
 260                 util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 261                 util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 262
 263         if (ctx->draw_vbo(ctx, info, index_offset))
 264                 batch->needs_flush = true;
 265
 266         for (i = 0; i < ctx->streamout.num_targets; i++)
 267                 ctx->streamout.offsets[i] += info->count;
 268
 269         if (fd_mesa_debug & FD_DBG_DDRAW)
 270                 fd_context_all_dirty(ctx);
 271
 272         fd_batch_check_size(batch);
 273
 274         if (info == &new_info)
 275                 pipe_resource_reference(&indexbuf, NULL);
 276 }
 277
 278 /* Generic clear implementation (partially) using u_blitter: */
 279 static void
 280 fd_blitter_clear(struct pipe_context *pctx, unsigned buffers,
 281                 const union pipe_color_union *color, double depth, unsigned stencil)
 282 {
 283         struct fd_context *ctx = fd_context(pctx);
 284         struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
 285         struct blitter_context *blitter = ctx->blitter;
 286
 287         fd_blitter_pipe_begin(ctx, false, true, FD_STAGE_CLEAR);
 288
 289         util_blitter_common_clear_setup(blitter, pfb->width, pfb->height,
 290                         buffers, NULL, NULL);
 291
 292         struct pipe_stencil_ref sr = {
 293                 .ref_value = { stencil & 0xff }
 294         };
 295         pctx->set_stencil_ref(pctx, &sr);
 296
 297         struct pipe_constant_buffer cb = {
 298                 .buffer_size = 16,
 299                 .user_buffer = &color->ui,
 300         };
 301         pctx->set_constant_buffer(pctx, PIPE_SHADER_FRAGMENT, 0, &cb);
 302
 303         if (!ctx->clear_rs_state) {
 304                 const struct pipe_rasterizer_state tmpl = {
 305                         .cull_face = PIPE_FACE_NONE,
 306                         .half_pixel_center = 1,
 307                         .bottom_edge_rule = 1,
 308                         .flatshade = 1,
 309                         .depth_clip_near = 1,
 310                         .depth_clip_far = 1,
 311                 };
 312                 ctx->clear_rs_state = pctx->create_rasterizer_state(pctx, &tmpl);
 313         }
 314         pctx->bind_rasterizer_state(pctx, ctx->clear_rs_state);
 315
 316         struct pipe_viewport_state vp = {
 317                 .scale     = { 0.5f * pfb->width, -0.5f * pfb->height, depth },
 318                 .translate = { 0.5f * pfb->width,  0.5f * pfb->height, 0.0f },
 319         };
 320         pctx->set_viewport_states(pctx, 0, 1, &vp);
 321
 322         pctx->bind_vertex_elements_state(pctx, ctx->solid_vbuf_state.vtx);
 323         pctx->set_vertex_buffers(pctx, blitter->vb_slot, 1,
 324                         &ctx->solid_vbuf_state.vertexbuf.vb[0]);
 325         pctx->set_stream_output_targets(pctx, 0, NULL, NULL);
 326         pctx->bind_vs_state(pctx, ctx->solid_prog.vp);
 327         pctx->bind_fs_state(pctx, ctx->solid_prog.fp);
 328
 329         struct pipe_draw_info info = {
 330                 .mode = PIPE_PRIM_MAX,    /* maps to DI_PT_RECTLIST */
 331                 .count = 2,
 332                 .max_index = 1,
 333                 .instance_count = 1,
 334         };
 335         ctx->draw_vbo(ctx, &info, 0);
 336
 337         util_blitter_restore_constant_buffer_state(blitter);
 338         util_blitter_restore_vertex_states(blitter);
 339         util_blitter_restore_fragment_states(blitter);
 340         util_blitter_restore_textures(blitter);
 341         util_blitter_restore_fb_state(blitter);
 342         util_blitter_restore_render_cond(blitter);
 343         util_blitter_unset_running_flag(blitter);
 344
 345         fd_blitter_pipe_end(ctx);
 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 = fd_context_batch(ctx);
 354         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
 355         unsigned cleared_buffers;
 356         int i;
 357
 358         /* TODO: push down the region versions into the tiles */
 359         if (!fd_render_condition_check(pctx))
 360                 return;
 361
 362         if (ctx->in_blit) {
 363                 fd_batch_reset(batch);
 364                 fd_context_all_dirty(ctx);
 365         }
 366
 367         /* pctx->clear() is only for full-surface clears, so scissor is
 368          * equivalent to having GL_SCISSOR_TEST disabled:
 369          */
 370         batch->max_scissor.minx = 0;
 371         batch->max_scissor.miny = 0;
 372         batch->max_scissor.maxx = pfb->width;
 373         batch->max_scissor.maxy = pfb->height;
 374
 375         /* for bookkeeping about which buffers have been cleared (and thus
 376          * can fully or partially skip mem2gmem) we need to ignore buffers
 377          * that have already had a draw, in case apps do silly things like
 378          * clear after draw (ie. if you only clear the color buffer, but
 379          * something like alpha-test causes side effects from the draw in
 380          * the depth buffer, etc)
 381          */
 382         cleared_buffers = buffers & (FD_BUFFER_ALL & ~batch->restore);
 383         batch->cleared |= cleared_buffers;
 384         batch->invalidated |= cleared_buffers;
 385
 386         batch->resolve |= buffers;
 387         batch->needs_flush = true;
 388
 389         mtx_lock(&ctx->screen->lock);
 390
 391         if (buffers & PIPE_CLEAR_COLOR)
 392                 for (i = 0; i < pfb->nr_cbufs; i++)
 393                         if (buffers & (PIPE_CLEAR_COLOR0 << i))
 394                                 resource_written(batch, pfb->cbufs[i]->texture);
 395
 396         if (buffers & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
 397                 resource_written(batch, pfb->zsbuf->texture);
 398                 batch->gmem_reason |= FD_GMEM_CLEARS_DEPTH_STENCIL;
 399         }
 400
 401         resource_written(batch, batch->query_buf);
 402
 403         list_for_each_entry(struct fd_acc_query, aq, &ctx->acc_active_queries, node)
 404                 resource_written(batch, aq->prsc);
 405
 406         mtx_unlock(&ctx->screen->lock);
 407
 408         DBG("%p: %x %ux%u depth=%f, stencil=%u (%s/%s)", batch, buffers,
 409                 pfb->width, pfb->height, depth, stencil,
 410                 util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 411                 util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 412
 413         /* if per-gen backend doesn't implement ctx->clear() generic
 414          * blitter clear:
 415          */
 416         bool fallback = true;
 417
 418         if (ctx->clear) {
 419                 fd_batch_set_stage(batch, FD_STAGE_CLEAR);
 420
 421                 if (ctx->clear(ctx, buffers, color, depth, stencil)) {
 422                         if (fd_mesa_debug & FD_DBG_DCLEAR)
 423                                 fd_context_all_dirty(ctx);
 424
 425                         fallback = false;
 426                 }
 427         }
 428
 429         if (fallback) {
 430                 fd_blitter_clear(pctx, buffers, color, depth, stencil);
 431         }
 432
 433         fd_batch_check_size(batch);
 434 }
 435
 436 static void
 437 fd_clear_render_target(struct pipe_context *pctx, struct pipe_surface *ps,
 438                 const union pipe_color_union *color,
 439                 unsigned x, unsigned y, unsigned w, unsigned h,
 440                 bool render_condition_enabled)
 441 {
 442         DBG("TODO: x=%u, y=%u, w=%u, h=%u", x, y, w, h);
 443 }
 444
 445 static void
 446 fd_clear_depth_stencil(struct pipe_context *pctx, struct pipe_surface *ps,
 447                 unsigned buffers, double depth, unsigned stencil,
 448                 unsigned x, unsigned y, unsigned w, unsigned h,
 449                 bool render_condition_enabled)
 450 {
 451         DBG("TODO: buffers=%u, depth=%f, stencil=%u, x=%u, y=%u, w=%u, h=%u",
 452                         buffers, depth, stencil, x, y, w, h);
 453 }
 454
 455 static void
 456 fd_launch_grid(struct pipe_context *pctx, const struct pipe_grid_info *info)
 457 {
 458         struct fd_context *ctx = fd_context(pctx);
 459         struct fd_batch *batch, *save_batch = NULL;
 460         unsigned i;
 461
 462         batch = fd_bc_alloc_batch(&ctx->screen->batch_cache, ctx, true);
 463         fd_batch_reference(&save_batch, ctx->batch);
 464         fd_batch_reference(&ctx->batch, batch);
 465
 466         mtx_lock(&ctx->screen->lock);
 467
 468         /* Mark SSBOs as being written.. we don't actually know which ones are
 469          * read vs written, so just assume the worst
 470          */
 471         foreach_bit(i, ctx->shaderbuf[PIPE_SHADER_COMPUTE].enabled_mask)
 472                 resource_written(batch, ctx->shaderbuf[PIPE_SHADER_COMPUTE].sb[i].buffer);
 473
 474         foreach_bit(i, ctx->shaderimg[PIPE_SHADER_COMPUTE].enabled_mask) {
 475                 struct pipe_image_view *img =
 476                         &ctx->shaderimg[PIPE_SHADER_COMPUTE].si[i];
 477                 if (img->access & PIPE_IMAGE_ACCESS_WRITE)
 478                         resource_written(batch, img->resource);
 479                 else
 480                         resource_read(batch, img->resource);
 481         }
 482
 483         /* UBO's are read */
 484         foreach_bit(i, ctx->constbuf[PIPE_SHADER_COMPUTE].enabled_mask)
 485                 resource_read(batch, ctx->constbuf[PIPE_SHADER_COMPUTE].cb[i].buffer);
 486
 487         /* Mark textures as being read */
 488         foreach_bit(i, ctx->tex[PIPE_SHADER_COMPUTE].valid_textures)
 489                 resource_read(batch, ctx->tex[PIPE_SHADER_COMPUTE].textures[i]->texture);
 490
 491         /* For global buffers, we don't really know if read or written, so assume
 492          * the worst:
 493          */
 494         foreach_bit(i, ctx->global_bindings.enabled_mask)
 495                 resource_written(batch, ctx->global_bindings.buf[i]);
 496
 497         if (info->indirect)
 498                 resource_read(batch, info->indirect);
 499
 500         mtx_unlock(&ctx->screen->lock);
 501
 502         batch->needs_flush = true;
 503         ctx->launch_grid(ctx, info);
 504
 505         fd_batch_flush(batch, false, false);
 506
 507         fd_batch_reference(&ctx->batch, save_batch);
 508         fd_batch_reference(&save_batch, NULL);
 509         fd_batch_reference(&batch, NULL);
 510 }
 511
 512 void
 513 fd_draw_init(struct pipe_context *pctx)
 514 {
 515         pctx->draw_vbo = fd_draw_vbo;
 516         pctx->clear = fd_clear;
 517         pctx->clear_render_target = fd_clear_render_target;
 518         pctx->clear_depth_stencil = fd_clear_depth_stencil;
 519
 520         if (has_compute(fd_screen(pctx->screen))) {
 521                 pctx->launch_grid = fd_launch_grid;
 522         }
 523 }