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_string.h"
  31 #include "util/u_memory.h"
  32 #include "util/u_prim.h"
  33 #include "util/u_format.h"
  34 #include "util/u_helpers.h"
  35
  36 #include "freedreno_draw.h"
  37 #include "freedreno_context.h"
  38 #include "freedreno_state.h"
  39 #include "freedreno_resource.h"
  40 #include "freedreno_query_acc.h"
  41 #include "freedreno_query_hw.h"
  42 #include "freedreno_util.h"
  43
  44 static void
  45 resource_read(struct fd_batch *batch, struct pipe_resource *prsc)
  46 {
  47         if (!prsc)
  48                 return;
  49         fd_batch_resource_used(batch, fd_resource(prsc), false);
  50 }
  51
  52 static void
  53 resource_written(struct fd_batch *batch, struct pipe_resource *prsc)
  54 {
  55         if (!prsc)
  56                 return;
  57         fd_batch_resource_used(batch, fd_resource(prsc), true);
  58 }
  59
  60 static void
  61 fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
  62 {
  63         struct fd_context *ctx = fd_context(pctx);
  64         struct fd_batch *batch = ctx->batch;
  65         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
  66         struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
  67         unsigned i, prims, buffers = 0;
  68
  69         if (!info->count_from_stream_output && !info->indirect &&
  70             !info->primitive_restart &&
  71             !u_trim_pipe_prim(info->mode, (unsigned*)&info->count))
  72                 return;
  73
  74         /* if we supported transform feedback, we'd have to disable this: */
  75         if (((scissor->maxx - scissor->minx) *
  76                         (scissor->maxy - scissor->miny)) == 0) {
  77                 return;
  78         }
  79
  80         /* TODO: push down the region versions into the tiles */
  81         if (!fd_render_condition_check(pctx))
  82                 return;
  83
  84         /* emulate unsupported primitives: */
  85         if (!fd_supported_prim(ctx, info->mode)) {
  86                 if (ctx->streamout.num_targets > 0)
  87                         debug_error("stream-out with emulated prims");
  88                 util_primconvert_save_rasterizer_state(ctx->primconvert, ctx->rasterizer);
  89                 util_primconvert_draw_vbo(ctx->primconvert, info);
  90                 return;
  91         }
  92
  93         /* Upload a user index buffer. */
  94         struct pipe_resource *indexbuf = NULL;
  95         unsigned index_offset = 0;
  96         struct pipe_draw_info new_info;
  97         if (info->index_size) {
  98                 if (info->has_user_indices) {
  99                         if (!util_upload_index_buffer(pctx, info, &indexbuf, &index_offset))
 100                                 return;
 101                         new_info = *info;
 102                         new_info.index.resource = indexbuf;
 103                         new_info.has_user_indices = false;
 104                         info = &new_info;
 105                 } else {
 106                         indexbuf = info->index.resource;
 107                 }
 108         }
 109
 110         if (ctx->in_blit) {
 111                 fd_batch_reset(batch);
 112                 fd_context_all_dirty(ctx);
 113         }
 114
 115         batch->blit = ctx->in_blit;
 116         batch->back_blit = ctx->in_shadow;
 117
 118         /* NOTE: needs to be before resource_written(batch->query_buf), otherwise
 119          * query_buf may not be created yet.
 120          */
 121         fd_batch_set_stage(batch, FD_STAGE_DRAW);
 122
 123         /*
 124          * Figure out the buffers/features we need:
 125          */
 126
 127         mtx_lock(&ctx->screen->lock);
 128
 129         if (fd_depth_enabled(ctx)) {
 130                 buffers |= FD_BUFFER_DEPTH;
 131                 resource_written(batch, pfb->zsbuf->texture);
 132                 batch->gmem_reason |= FD_GMEM_DEPTH_ENABLED;
 133         }
 134
 135         if (fd_stencil_enabled(ctx)) {
 136                 buffers |= FD_BUFFER_STENCIL;
 137                 resource_written(batch, pfb->zsbuf->texture);
 138                 batch->gmem_reason |= FD_GMEM_STENCIL_ENABLED;
 139         }
 140
 141         if (fd_logicop_enabled(ctx))
 142                 batch->gmem_reason |= FD_GMEM_LOGICOP_ENABLED;
 143
 144         for (i = 0; i < pfb->nr_cbufs; i++) {
 145                 struct pipe_resource *surf;
 146
 147                 if (!pfb->cbufs[i])
 148                         continue;
 149
 150                 surf = pfb->cbufs[i]->texture;
 151
 152                 resource_written(batch, surf);
 153                 buffers |= PIPE_CLEAR_COLOR0 << i;
 154
 155                 if (surf->nr_samples > 1)
 156                         batch->gmem_reason |= FD_GMEM_MSAA_ENABLED;
 157
 158                 if (fd_blend_enabled(ctx, i))
 159                         batch->gmem_reason |= FD_GMEM_BLEND_ENABLED;
 160         }
 161
 162         /* Mark SSBOs as being written.. we don't actually know which ones are
 163          * read vs written, so just assume the worst
 164          */
 165         foreach_bit(i, ctx->shaderbuf[PIPE_SHADER_FRAGMENT].enabled_mask)
 166                 resource_read(batch, ctx->shaderbuf[PIPE_SHADER_FRAGMENT].sb[i].buffer);
 167
 168         foreach_bit(i, ctx->constbuf[PIPE_SHADER_VERTEX].enabled_mask)
 169                 resource_read(batch, ctx->constbuf[PIPE_SHADER_VERTEX].cb[i].buffer);
 170         foreach_bit(i, ctx->constbuf[PIPE_SHADER_FRAGMENT].enabled_mask)
 171                 resource_read(batch, ctx->constbuf[PIPE_SHADER_FRAGMENT].cb[i].buffer);
 172
 173         /* Mark VBOs as being read */
 174         foreach_bit(i, ctx->vtx.vertexbuf.enabled_mask) {
 175                 assert(!ctx->vtx.vertexbuf.vb[i].is_user_buffer);
 176                 resource_read(batch, ctx->vtx.vertexbuf.vb[i].buffer.resource);
 177         }
 178
 179         /* Mark index buffer as being read */
 180         resource_read(batch, indexbuf);
 181
 182         /* Mark textures as being read */
 183         foreach_bit(i, ctx->tex[PIPE_SHADER_VERTEX].valid_textures)
 184                 resource_read(batch, ctx->tex[PIPE_SHADER_VERTEX].textures[i]->texture);
 185         foreach_bit(i, ctx->tex[PIPE_SHADER_FRAGMENT].valid_textures)
 186                 resource_read(batch, ctx->tex[PIPE_SHADER_FRAGMENT].textures[i]->texture);
 187
 188         /* Mark streamout buffers as being written.. */
 189         for (i = 0; i < ctx->streamout.num_targets; i++)
 190                 if (ctx->streamout.targets[i])
 191                         resource_written(batch, ctx->streamout.targets[i]->buffer);
 192
 193         resource_written(batch, batch->query_buf);
 194
 195         list_for_each_entry(struct fd_acc_query, aq, &ctx->acc_active_queries, node)
 196                 resource_written(batch, aq->prsc);
 197
 198         mtx_unlock(&ctx->screen->lock);
 199
 200         batch->num_draws++;
 201
 202         prims = u_reduced_prims_for_vertices(info->mode, info->count);
 203
 204         ctx->stats.draw_calls++;
 205
 206         /* TODO prims_emitted should be clipped when the stream-out buffer is
 207          * not large enough.  See max_tf_vtx().. probably need to move that
 208          * into common code.  Although a bit more annoying since a2xx doesn't
 209          * use ir3 so no common way to get at the pipe_stream_output_info
 210          * which is needed for this calculation.
 211          */
 212         if (ctx->streamout.num_targets > 0)
 213                 ctx->stats.prims_emitted += prims;
 214         ctx->stats.prims_generated += prims;
 215
 216         /* any buffers that haven't been cleared yet, we need to restore: */
 217         batch->restore |= buffers & (FD_BUFFER_ALL & ~batch->cleared);
 218         /* and any buffers used, need to be resolved: */
 219         batch->resolve |= buffers;
 220
 221         DBG("%p: %x %ux%u num_draws=%u (%s/%s)", batch, buffers,
 222                 pfb->width, pfb->height, batch->num_draws,
 223                 util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 224                 util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 225
 226         if (ctx->draw_vbo(ctx, info, index_offset))
 227                 batch->needs_flush = true;
 228
 229         for (i = 0; i < ctx->streamout.num_targets; i++)
 230                 ctx->streamout.offsets[i] += info->count;
 231
 232         if (fd_mesa_debug & FD_DBG_DDRAW)
 233                 fd_context_all_dirty(ctx);
 234
 235         fd_batch_check_size(batch);
 236
 237         if (info == &new_info)
 238                 pipe_resource_reference(&indexbuf, NULL);
 239 }
 240
 241 /* Generic clear implementation (partially) using u_blitter: */
 242 static void
 243 fd_blitter_clear(struct pipe_context *pctx, unsigned buffers,
 244                 const union pipe_color_union *color, double depth, unsigned stencil)
 245 {
 246         struct fd_context *ctx = fd_context(pctx);
 247         struct pipe_framebuffer_state *pfb = &ctx->batch->framebuffer;
 248         struct blitter_context *blitter = ctx->blitter;
 249
 250         fd_blitter_pipe_begin(ctx, false, true, FD_STAGE_CLEAR);
 251
 252         util_blitter_common_clear_setup(blitter, pfb->width, pfb->height,
 253                         buffers, NULL, NULL);
 254
 255         struct pipe_stencil_ref sr = {
 256                 .ref_value = { stencil & 0xff }
 257         };
 258         pctx->set_stencil_ref(pctx, &sr);
 259
 260         struct pipe_constant_buffer cb = {
 261                 .buffer_size = 16,
 262                 .user_buffer = &color->ui,
 263         };
 264         pctx->set_constant_buffer(pctx, PIPE_SHADER_FRAGMENT, 0, &cb);
 265
 266         if (!ctx->clear_rs_state) {
 267                 const struct pipe_rasterizer_state tmpl = {
 268                         .cull_face = PIPE_FACE_NONE,
 269                         .half_pixel_center = 1,
 270                         .bottom_edge_rule = 1,
 271                         .flatshade = 1,
 272                         .depth_clip = 1,
 273                 };
 274                 ctx->clear_rs_state = pctx->create_rasterizer_state(pctx, &tmpl);
 275         }
 276         pctx->bind_rasterizer_state(pctx, ctx->clear_rs_state);
 277
 278         struct pipe_viewport_state vp = {
 279                 .scale     = { 0.5f * pfb->width, -0.5f * pfb->height, depth },
 280                 .translate = { 0.5f * pfb->width,  0.5f * pfb->height, 0.0f },
 281         };
 282         pctx->set_viewport_states(pctx, 0, 1, &vp);
 283
 284         pctx->bind_vertex_elements_state(pctx, ctx->solid_vbuf_state.vtx);
 285         pctx->set_vertex_buffers(pctx, blitter->vb_slot, 1,
 286                         &ctx->solid_vbuf_state.vertexbuf.vb[0]);
 287         pctx->set_stream_output_targets(pctx, 0, NULL, NULL);
 288         pctx->bind_vs_state(pctx, ctx->solid_prog.vp);
 289         pctx->bind_fs_state(pctx, ctx->solid_prog.fp);
 290
 291         struct pipe_draw_info info = {
 292                 .mode = PIPE_PRIM_MAX,    /* maps to DI_PT_RECTLIST */
 293                 .count = 2,
 294                 .max_index = 1,
 295                 .instance_count = 1,
 296         };
 297         ctx->draw_vbo(ctx, &info, 0);
 298
 299         util_blitter_restore_constant_buffer_state(blitter);
 300         util_blitter_restore_vertex_states(blitter);
 301         util_blitter_restore_fragment_states(blitter);
 302         util_blitter_restore_textures(blitter);
 303         util_blitter_restore_fb_state(blitter);
 304         util_blitter_restore_render_cond(blitter);
 305         util_blitter_unset_running_flag(blitter);
 306
 307         fd_blitter_pipe_end(ctx);
 308 }
 309
 310 /* TODO figure out how to make better use of existing state mechanism
 311  * for clear (and possibly gmem->mem / mem->gmem) so we can (a) keep
 312  * track of what state really actually changes, and (b) reduce the code
 313  * in the a2xx/a3xx parts.
 314  */
 315
 316 static void
 317 fd_clear(struct pipe_context *pctx, unsigned buffers,
 318                 const union pipe_color_union *color, double depth, unsigned stencil)
 319 {
 320         struct fd_context *ctx = fd_context(pctx);
 321         struct fd_batch *batch = ctx->batch;
 322         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
 323         struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
 324         unsigned cleared_buffers;
 325         int i;
 326
 327         /* TODO: push down the region versions into the tiles */
 328         if (!fd_render_condition_check(pctx))
 329                 return;
 330
 331         if (ctx->in_blit) {
 332                 fd_batch_reset(batch);
 333                 fd_context_all_dirty(ctx);
 334         }
 335
 336         /* for bookkeeping about which buffers have been cleared (and thus
 337          * can fully or partially skip mem2gmem) we need to ignore buffers
 338          * that have already had a draw, in case apps do silly things like
 339          * clear after draw (ie. if you only clear the color buffer, but
 340          * something like alpha-test causes side effects from the draw in
 341          * the depth buffer, etc)
 342          */
 343         cleared_buffers = buffers & (FD_BUFFER_ALL & ~batch->restore);
 344
 345         /* do we have full-screen scissor? */
 346         if (!memcmp(scissor, &ctx->disabled_scissor, sizeof(*scissor))) {
 347                 batch->cleared |= cleared_buffers;
 348         } else {
 349                 batch->partial_cleared |= cleared_buffers;
 350                 if (cleared_buffers & PIPE_CLEAR_COLOR)
 351                         batch->cleared_scissor.color = *scissor;
 352                 if (cleared_buffers & PIPE_CLEAR_DEPTH)
 353                         batch->cleared_scissor.depth = *scissor;
 354                 if (cleared_buffers & PIPE_CLEAR_STENCIL)
 355                         batch->cleared_scissor.stencil = *scissor;
 356         }
 357         batch->resolve |= buffers;
 358         batch->needs_flush = true;
 359
 360         mtx_lock(&ctx->screen->lock);
 361
 362         if (buffers & PIPE_CLEAR_COLOR)
 363                 for (i = 0; i < pfb->nr_cbufs; i++)
 364                         if (buffers & (PIPE_CLEAR_COLOR0 << i))
 365                                 resource_written(batch, pfb->cbufs[i]->texture);
 366
 367         if (buffers & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
 368                 resource_written(batch, pfb->zsbuf->texture);
 369                 batch->gmem_reason |= FD_GMEM_CLEARS_DEPTH_STENCIL;
 370         }
 371
 372         resource_written(batch, batch->query_buf);
 373
 374         list_for_each_entry(struct fd_acc_query, aq, &ctx->acc_active_queries, node)
 375                 resource_written(batch, aq->prsc);
 376
 377         mtx_unlock(&ctx->screen->lock);
 378
 379         DBG("%p: %x %ux%u depth=%f, stencil=%u (%s/%s)", batch, buffers,
 380                 pfb->width, pfb->height, depth, stencil,
 381                 util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 382                 util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 383
 384         /* if per-gen backend doesn't implement ctx->clear() generic
 385          * blitter clear:
 386          */
 387         bool fallback = true;
 388
 389         if (ctx->clear) {
 390                 fd_batch_set_stage(batch, FD_STAGE_CLEAR);
 391
 392                 if (ctx->clear(ctx, buffers, color, depth, stencil)) {
 393                         if (fd_mesa_debug & FD_DBG_DCLEAR)
 394                                 fd_context_all_dirty(ctx);
 395
 396                         fallback = false;
 397                 }
 398         }
 399
 400         if (fallback) {
 401                 fd_blitter_clear(pctx, buffers, color, depth, stencil);
 402         }
 403 }
 404
 405 static void
 406 fd_clear_render_target(struct pipe_context *pctx, struct pipe_surface *ps,
 407                 const union pipe_color_union *color,
 408                 unsigned x, unsigned y, unsigned w, unsigned h,
 409                 bool render_condition_enabled)
 410 {
 411         DBG("TODO: x=%u, y=%u, w=%u, h=%u", x, y, w, h);
 412 }
 413
 414 static void
 415 fd_clear_depth_stencil(struct pipe_context *pctx, struct pipe_surface *ps,
 416                 unsigned buffers, double depth, unsigned stencil,
 417                 unsigned x, unsigned y, unsigned w, unsigned h,
 418                 bool render_condition_enabled)
 419 {
 420         DBG("TODO: buffers=%u, depth=%f, stencil=%u, x=%u, y=%u, w=%u, h=%u",
 421                         buffers, depth, stencil, x, y, w, h);
 422 }
 423
 424 static void
 425 fd_launch_grid(struct pipe_context *pctx, const struct pipe_grid_info *info)
 426 {
 427         struct fd_context *ctx = fd_context(pctx);
 428         struct fd_batch *batch, *save_batch = NULL;
 429         unsigned i;
 430
 431         batch = fd_batch_create(ctx);
 432         fd_batch_reference(&save_batch, ctx->batch);
 433         fd_batch_reference(&ctx->batch, batch);
 434
 435         mtx_lock(&ctx->screen->lock);
 436
 437         /* Mark SSBOs as being written.. we don't actually know which ones are
 438          * read vs written, so just assume the worst
 439          */
 440         foreach_bit(i, ctx->shaderbuf[PIPE_SHADER_COMPUTE].enabled_mask)
 441                 resource_read(batch, ctx->shaderbuf[PIPE_SHADER_COMPUTE].sb[i].buffer);
 442
 443         /* UBO's are read */
 444         foreach_bit(i, ctx->constbuf[PIPE_SHADER_COMPUTE].enabled_mask)
 445                 resource_read(batch, ctx->constbuf[PIPE_SHADER_COMPUTE].cb[i].buffer);
 446
 447         /* Mark textures as being read */
 448         foreach_bit(i, ctx->tex[PIPE_SHADER_COMPUTE].valid_textures)
 449                 resource_read(batch, ctx->tex[PIPE_SHADER_COMPUTE].textures[i]->texture);
 450
 451         mtx_unlock(&ctx->screen->lock);
 452
 453         ctx->launch_grid(ctx, info);
 454
 455         fd_gmem_flush_compute(batch);
 456
 457         fd_batch_reference(&ctx->batch, save_batch);
 458         fd_batch_reference(&save_batch, NULL);
 459 }
 460
 461 void
 462 fd_draw_init(struct pipe_context *pctx)
 463 {
 464         pctx->draw_vbo = fd_draw_vbo;
 465         pctx->clear = fd_clear;
 466         pctx->clear_render_target = fd_clear_render_target;
 467         pctx->clear_depth_stencil = fd_clear_depth_stencil;
 468
 469         if (has_compute(fd_screen(pctx->screen))) {
 470                 pctx->launch_grid = fd_launch_grid;
 471         }
 472 }