src/gallium/drivers/freedreno/freedreno_gmem.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_inlines.h"
  33 #include "util/u_format.h"
  34
  35 #include "freedreno_gmem.h"
  36 #include "freedreno_context.h"
  37 #include "freedreno_resource.h"
  38 #include "freedreno_query_hw.h"
  39 #include "freedreno_util.h"
  40
  41 /*
  42  * GMEM is the small (ie. 256KiB for a200, 512KiB for a220, etc) tile buffer
  43  * inside the GPU.  All rendering happens to GMEM.  Larger render targets
  44  * are split into tiles that are small enough for the color (and depth and/or
  45  * stencil, if enabled) buffers to fit within GMEM.  Before rendering a tile,
  46  * if there was not a clear invalidating the previous tile contents, we need
  47  * to restore the previous tiles contents (system mem -> GMEM), and after all
  48  * the draw calls, before moving to the next tile, we need to save the tile
  49  * contents (GMEM -> system mem).
  50  *
  51  * The code in this file handles dealing with GMEM and tiling.
  52  *
  53  * The structure of the ringbuffer ends up being:
  54  *
  55  *     +--<---<-- IB ---<---+---<---+---<---<---<--+
  56  *     |                    |       |              |
  57  *     v                    ^       ^              ^
  58  *   ------------------------------------------------------
  59  *     | clear/draw cmds | Tile0 | Tile1 | .... | TileN |
  60  *   ------------------------------------------------------
  61  *                       ^
  62  *                       |
  63  *                       address submitted in issueibcmds
  64  *
  65  * Where the per-tile section handles scissor setup, mem2gmem restore (if
  66  * needed), IB to draw cmds earlier in the ringbuffer, and then gmem2mem
  67  * resolve.
  68  */
  69
  70 static uint32_t bin_width(struct fd_context *ctx)
  71 {
  72         if (is_a4xx(ctx->screen))
  73                 return 1024;
  74         if (is_a3xx(ctx->screen))
  75                 return 992;
  76         return 512;
  77 }
  78
  79 static uint32_t
  80 total_size(uint8_t cbuf_cpp[], uint8_t zsbuf_cpp[2],
  81                    uint32_t bin_w, uint32_t bin_h, struct fd_gmem_stateobj *gmem)
  82 {
  83         uint32_t total = 0, i;
  84
  85         for (i = 0; i < 4; i++) {
  86                 if (cbuf_cpp[i]) {
  87                         gmem->cbuf_base[i] = align(total, 0x4000);
  88                         total = gmem->cbuf_base[i] + cbuf_cpp[i] * bin_w * bin_h;
  89                 }
  90         }
  91
  92         if (zsbuf_cpp[0]) {
  93                 gmem->zsbuf_base[0] = align(total, 0x4000);
  94                 total = gmem->zsbuf_base[0] + zsbuf_cpp[0] * bin_w * bin_h;
  95         }
  96
  97         if (zsbuf_cpp[1]) {
  98                 gmem->zsbuf_base[1] = align(total, 0x4000);
  99                 total = gmem->zsbuf_base[1] + zsbuf_cpp[1] * bin_w * bin_h;
 100         }
 101
 102         return total;
 103 }
 104
 105 static void
 106 calculate_tiles(struct fd_context *ctx)
 107 {
 108         struct fd_gmem_stateobj *gmem = &ctx->gmem;
 109         struct pipe_scissor_state *scissor = &ctx->max_scissor;
 110         struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
 111         uint32_t gmem_size = ctx->screen->gmemsize_bytes;
 112         uint32_t minx, miny, width, height;
 113         uint32_t nbins_x = 1, nbins_y = 1;
 114         uint32_t bin_w, bin_h;
 115         uint32_t max_width = bin_width(ctx);
 116         uint8_t cbuf_cpp[4] = {0}, zsbuf_cpp[2] = {0};
 117         uint32_t i, j, t, xoff, yoff;
 118         uint32_t tpp_x, tpp_y;
 119         bool has_zs = !!(ctx->resolve & (FD_BUFFER_DEPTH | FD_BUFFER_STENCIL));
 120
 121         if (has_zs) {
 122                 struct fd_resource *rsc = fd_resource(pfb->zsbuf->texture);
 123                 zsbuf_cpp[0] = rsc->cpp;
 124                 if (rsc->stencil)
 125                         zsbuf_cpp[1] = rsc->stencil->cpp;
 126         }
 127         for (i = 0; i < pfb->nr_cbufs; i++) {
 128                 if (pfb->cbufs[i])
 129                         cbuf_cpp[i] = util_format_get_blocksize(pfb->cbufs[i]->format);
 130                 else
 131                         cbuf_cpp[i] = 4;
 132         }
 133
 134         if (!memcmp(gmem->zsbuf_cpp, zsbuf_cpp, sizeof(zsbuf_cpp)) &&
 135                 !memcmp(gmem->cbuf_cpp, cbuf_cpp, sizeof(cbuf_cpp)) &&
 136                 !memcmp(&gmem->scissor, scissor, sizeof(gmem->scissor))) {
 137                 /* everything is up-to-date */
 138                 return;
 139         }
 140
 141         if (fd_mesa_debug & FD_DBG_NOSCIS) {
 142                 minx = 0;
 143                 miny = 0;
 144                 width = pfb->width;
 145                 height = pfb->height;
 146         } else {
 147                 minx = scissor->minx & ~31; /* round down to multiple of 32 */
 148                 miny = scissor->miny & ~31;
 149                 width = scissor->maxx - minx;
 150                 height = scissor->maxy - miny;
 151         }
 152
 153         bin_w = align(width, 32);
 154         bin_h = align(height, 32);
 155
 156         /* first, find a bin width that satisfies the maximum width
 157          * restrictions:
 158          */
 159         while (bin_w > max_width) {
 160                 nbins_x++;
 161                 bin_w = align(width / nbins_x, 32);
 162         }
 163
 164         /* then find a bin width/height that satisfies the memory
 165          * constraints:
 166          */
 167         DBG("binning input: cbuf cpp: %d %d %d %d, zsbuf cpp: %d; %dx%d",
 168                 cbuf_cpp[0], cbuf_cpp[1], cbuf_cpp[2], cbuf_cpp[3], zsbuf_cpp[0],
 169                 width, height);
 170         while (total_size(cbuf_cpp, zsbuf_cpp, bin_w, bin_h, gmem) > gmem_size) {
 171                 if (bin_w > bin_h) {
 172                         nbins_x++;
 173                         bin_w = align(width / nbins_x, 32);
 174                 } else {
 175                         nbins_y++;
 176                         bin_h = align(height / nbins_y, 32);
 177                 }
 178         }
 179
 180         DBG("using %d bins of size %dx%d", nbins_x*nbins_y, bin_w, bin_h);
 181
 182         gmem->scissor = *scissor;
 183         memcpy(gmem->cbuf_cpp, cbuf_cpp, sizeof(cbuf_cpp));
 184         memcpy(gmem->zsbuf_cpp, zsbuf_cpp, sizeof(zsbuf_cpp));
 185         gmem->bin_h = bin_h;
 186         gmem->bin_w = bin_w;
 187         gmem->nbins_x = nbins_x;
 188         gmem->nbins_y = nbins_y;
 189         gmem->minx = minx;
 190         gmem->miny = miny;
 191         gmem->width = width;
 192         gmem->height = height;
 193
 194         /*
 195          * Assign tiles and pipes:
 196          *
 197          * At some point it might be worth playing with different
 198          * strategies and seeing if that makes much impact on
 199          * performance.
 200          */
 201
 202 #define div_round_up(v, a)  (((v) + (a) - 1) / (a))
 203         /* figure out number of tiles per pipe: */
 204         tpp_x = tpp_y = 1;
 205         while (div_round_up(nbins_y, tpp_y) > 8)
 206                 tpp_y += 2;
 207         while ((div_round_up(nbins_y, tpp_y) *
 208                         div_round_up(nbins_x, tpp_x)) > 8)
 209                 tpp_x += 1;
 210
 211         /* configure pipes: */
 212         xoff = yoff = 0;
 213         for (i = 0; i < ARRAY_SIZE(ctx->pipe); i++) {
 214                 struct fd_vsc_pipe *pipe = &ctx->pipe[i];
 215
 216                 if (xoff >= nbins_x) {
 217                         xoff = 0;
 218                         yoff += tpp_y;
 219                 }
 220
 221                 if (yoff >= nbins_y) {
 222                         break;
 223                 }
 224
 225                 pipe->x = xoff;
 226                 pipe->y = yoff;
 227                 pipe->w = MIN2(tpp_x, nbins_x - xoff);
 228                 pipe->h = MIN2(tpp_y, nbins_y - yoff);
 229
 230                 xoff += tpp_x;
 231         }
 232
 233         for (; i < ARRAY_SIZE(ctx->pipe); i++) {
 234                 struct fd_vsc_pipe *pipe = &ctx->pipe[i];
 235                 pipe->x = pipe->y = pipe->w = pipe->h = 0;
 236         }
 237
 238 #if 0 /* debug */
 239         printf("%dx%d ... tpp=%dx%d\n", nbins_x, nbins_y, tpp_x, tpp_y);
 240         for (i = 0; i < 8; i++) {
 241                 struct fd_vsc_pipe *pipe = &ctx->pipe[i];
 242                 printf("pipe[%d]: %ux%u @ %u,%u\n", i,
 243                                 pipe->w, pipe->h, pipe->x, pipe->y);
 244         }
 245 #endif
 246
 247         /* configure tiles: */
 248         t = 0;
 249         yoff = miny;
 250         for (i = 0; i < nbins_y; i++) {
 251                 uint32_t bw, bh;
 252
 253                 xoff = minx;
 254
 255                 /* clip bin height: */
 256                 bh = MIN2(bin_h, miny + height - yoff);
 257
 258                 for (j = 0; j < nbins_x; j++) {
 259                         struct fd_tile *tile = &ctx->tile[t];
 260                         uint32_t n, p;
 261
 262                         assert(t < ARRAY_SIZE(ctx->tile));
 263
 264                         /* pipe number: */
 265                         p = ((i / tpp_y) * div_round_up(nbins_x, tpp_x)) + (j / tpp_x);
 266
 267                         /* slot number: */
 268                         n = ((i % tpp_y) * tpp_x) + (j % tpp_x);
 269
 270                         /* clip bin width: */
 271                         bw = MIN2(bin_w, minx + width - xoff);
 272
 273                         tile->n = n;
 274                         tile->p = p;
 275                         tile->bin_w = bw;
 276                         tile->bin_h = bh;
 277                         tile->xoff = xoff;
 278                         tile->yoff = yoff;
 279
 280                         t++;
 281
 282                         xoff += bw;
 283                 }
 284
 285                 yoff += bh;
 286         }
 287
 288 #if 0 /* debug */
 289         t = 0;
 290         for (i = 0; i < nbins_y; i++) {
 291                 for (j = 0; j < nbins_x; j++) {
 292                         struct fd_tile *tile = &ctx->tile[t++];
 293                         printf("|p:%u n:%u|", tile->p, tile->n);
 294                 }
 295                 printf("\n");
 296         }
 297 #endif
 298 }
 299
 300 static void
 301 render_tiles(struct fd_context *ctx)
 302 {
 303         struct fd_gmem_stateobj *gmem = &ctx->gmem;
 304         int i;
 305
 306         ctx->emit_tile_init(ctx);
 307
 308         if (ctx->restore)
 309                 ctx->stats.batch_restore++;
 310
 311         for (i = 0; i < (gmem->nbins_x * gmem->nbins_y); i++) {
 312                 struct fd_tile *tile = &ctx->tile[i];
 313
 314                 DBG("bin_h=%d, yoff=%d, bin_w=%d, xoff=%d",
 315                         tile->bin_h, tile->yoff, tile->bin_w, tile->xoff);
 316
 317                 ctx->emit_tile_prep(ctx, tile);
 318
 319                 if (ctx->restore) {
 320                         fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_MEM2GMEM);
 321                         ctx->emit_tile_mem2gmem(ctx, tile);
 322                         fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_NULL);
 323                 }
 324
 325                 ctx->emit_tile_renderprep(ctx, tile);
 326
 327                 fd_hw_query_prepare_tile(ctx, i, ctx->ring);
 328
 329                 /* emit IB to drawcmds: */
 330                 OUT_IB(ctx->ring, ctx->draw_start, ctx->draw_end);
 331                 fd_reset_wfi(ctx);
 332
 333                 /* emit gmem2mem to transfer tile back to system memory: */
 334                 fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_GMEM2MEM);
 335                 ctx->emit_tile_gmem2mem(ctx, tile);
 336                 fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_NULL);
 337         }
 338 }
 339
 340 static void
 341 render_sysmem(struct fd_context *ctx)
 342 {
 343         ctx->emit_sysmem_prep(ctx);
 344
 345         fd_hw_query_prepare_tile(ctx, 0, ctx->ring);
 346
 347         /* emit IB to drawcmds: */
 348         OUT_IB(ctx->ring, ctx->draw_start, ctx->draw_end);
 349         fd_reset_wfi(ctx);
 350 }
 351
 352 void
 353 fd_gmem_render_tiles(struct fd_context *ctx)
 354 {
 355         struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
 356         bool sysmem = false;
 357
 358         if (ctx->emit_sysmem_prep) {
 359                 if (ctx->cleared || ctx->gmem_reason || (ctx->num_draws > 5)) {
 360                         DBG("GMEM: cleared=%x, gmem_reason=%x, num_draws=%u",
 361                                 ctx->cleared, ctx->gmem_reason, ctx->num_draws);
 362                 } else if (!(fd_mesa_debug & FD_DBG_NOBYPASS)) {
 363                         sysmem = true;
 364                 }
 365         }
 366
 367         /* close out the draw cmds by making sure any active queries are
 368          * paused:
 369          */
 370         fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_NULL);
 371
 372         /* mark the end of the clear/draw cmds before emitting per-tile cmds: */
 373         fd_ringmarker_mark(ctx->draw_end);
 374         fd_ringmarker_mark(ctx->binning_end);
 375
 376         fd_reset_wfi(ctx);
 377
 378         ctx->stats.batch_total++;
 379
 380         if (sysmem) {
 381                 DBG("rendering sysmem (%s/%s)",
 382                         util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 383                         util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 384                 fd_hw_query_prepare(ctx, 1);
 385                 render_sysmem(ctx);
 386                 ctx->stats.batch_sysmem++;
 387         } else {
 388                 struct fd_gmem_stateobj *gmem = &ctx->gmem;
 389                 calculate_tiles(ctx);
 390                 DBG("rendering %dx%d tiles (%s/%s)", gmem->nbins_x, gmem->nbins_y,
 391                         util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 392                         util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 393                 fd_hw_query_prepare(ctx, gmem->nbins_x * gmem->nbins_y);
 394                 render_tiles(ctx);
 395                 ctx->stats.batch_gmem++;
 396         }
 397
 398         /* GPU executes starting from tile cmds, which IB back to draw cmds: */
 399         fd_ringmarker_flush(ctx->draw_end);
 400
 401         /* mark start for next draw/binning cmds: */
 402         fd_ringmarker_mark(ctx->draw_start);
 403         fd_ringmarker_mark(ctx->binning_start);
 404
 405         fd_reset_wfi(ctx);
 406
 407         /* reset maximal bounds: */
 408         ctx->max_scissor.minx = ctx->max_scissor.miny = ~0;
 409         ctx->max_scissor.maxx = ctx->max_scissor.maxy = 0;
 410
 411         ctx->dirty = ~0;
 412 }
 413
 414 /* tile needs restore if it isn't completely contained within the
 415  * cleared scissor:
 416  */
 417 static bool
 418 skip_restore(struct pipe_scissor_state *scissor, struct fd_tile *tile)
 419 {
 420         unsigned minx = tile->xoff;
 421         unsigned maxx = tile->xoff + tile->bin_w;
 422         unsigned miny = tile->yoff;
 423         unsigned maxy = tile->yoff + tile->bin_h;
 424         return (minx >= scissor->minx) && (maxx <= scissor->maxx) &&
 425                         (miny >= scissor->miny) && (maxy <= scissor->maxy);
 426 }
 427
 428 /* When deciding whether a tile needs mem2gmem, we need to take into
 429  * account the scissor rect(s) that were cleared.  To simplify we only
 430  * consider the last scissor rect for each buffer, since the common
 431  * case would be a single clear.
 432  */
 433 bool
 434 fd_gmem_needs_restore(struct fd_context *ctx, struct fd_tile *tile,
 435                 uint32_t buffers)
 436 {
 437         if (!(ctx->restore & buffers))
 438                 return false;
 439
 440         /* if buffers partially cleared, then slow-path to figure out
 441          * if this particular tile needs restoring:
 442          */
 443         if ((buffers & FD_BUFFER_COLOR) &&
 444                         (ctx->partial_cleared & FD_BUFFER_COLOR) &&
 445                         skip_restore(&ctx->cleared_scissor.color, tile))
 446                 return false;
 447         if ((buffers & FD_BUFFER_DEPTH) &&
 448                         (ctx->partial_cleared & FD_BUFFER_DEPTH) &&
 449                         skip_restore(&ctx->cleared_scissor.depth, tile))
 450                 return false;
 451         if ((buffers & FD_BUFFER_STENCIL) &&
 452                         (ctx->partial_cleared & FD_BUFFER_STENCIL) &&
 453                         skip_restore(&ctx->cleared_scissor.stencil, tile))
 454                 return false;
 455
 456         return true;
 457 }