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_screen *screen)
  71 {
  72         if (is_a4xx(screen))
  73                 return 1024;
  74         if (is_a3xx(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 < MAX_RENDER_TARGETS; 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_batch *batch)
 107 {
 108         struct fd_context *ctx = batch->ctx;
 109         struct fd_gmem_stateobj *gmem = &ctx->gmem;
 110         struct pipe_scissor_state *scissor = &batch->max_scissor;
 111         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
 112         uint32_t gmem_size = ctx->screen->gmemsize_bytes;
 113         uint32_t minx, miny, width, height;
 114         uint32_t nbins_x = 1, nbins_y = 1;
 115         uint32_t bin_w, bin_h;
 116         uint32_t max_width = bin_width(ctx->screen);
 117         uint8_t cbuf_cpp[MAX_RENDER_TARGETS] = {0}, zsbuf_cpp[2] = {0};
 118         uint32_t i, j, t, xoff, yoff;
 119         uint32_t tpp_x, tpp_y;
 120         bool has_zs = !!(batch->resolve & (FD_BUFFER_DEPTH | FD_BUFFER_STENCIL));
 121         int tile_n[ARRAY_SIZE(ctx->pipe)];
 122
 123         if (has_zs) {
 124                 struct fd_resource *rsc = fd_resource(pfb->zsbuf->texture);
 125                 zsbuf_cpp[0] = rsc->cpp;
 126                 if (rsc->stencil)
 127                         zsbuf_cpp[1] = rsc->stencil->cpp;
 128         }
 129         for (i = 0; i < pfb->nr_cbufs; i++) {
 130                 if (pfb->cbufs[i])
 131                         cbuf_cpp[i] = util_format_get_blocksize(pfb->cbufs[i]->format);
 132                 else
 133                         cbuf_cpp[i] = 4;
 134         }
 135
 136         if (!memcmp(gmem->zsbuf_cpp, zsbuf_cpp, sizeof(zsbuf_cpp)) &&
 137                 !memcmp(gmem->cbuf_cpp, cbuf_cpp, sizeof(cbuf_cpp)) &&
 138                 !memcmp(&gmem->scissor, scissor, sizeof(gmem->scissor))) {
 139                 /* everything is up-to-date */
 140                 return;
 141         }
 142
 143         if (fd_mesa_debug & FD_DBG_NOSCIS) {
 144                 minx = 0;
 145                 miny = 0;
 146                 width = pfb->width;
 147                 height = pfb->height;
 148         } else {
 149                 minx = scissor->minx & ~31; /* round down to multiple of 32 */
 150                 miny = scissor->miny & ~31;
 151                 width = scissor->maxx - minx;
 152                 height = scissor->maxy - miny;
 153         }
 154
 155         bin_w = align(width, 32);
 156         bin_h = align(height, 32);
 157
 158         /* first, find a bin width that satisfies the maximum width
 159          * restrictions:
 160          */
 161         while (bin_w > max_width) {
 162                 nbins_x++;
 163                 bin_w = align(width / nbins_x, 32);
 164         }
 165
 166         if (fd_mesa_debug & FD_DBG_MSGS) {
 167                 debug_printf("binning input: cbuf cpp:");
 168                 for (i = 0; i < pfb->nr_cbufs; i++)
 169                         debug_printf(" %d", cbuf_cpp[i]);
 170                 debug_printf(", zsbuf cpp: %d; %dx%d\n",
 171                                 zsbuf_cpp[0], width, height);
 172         }
 173
 174         /* then find a bin width/height that satisfies the memory
 175          * constraints:
 176          */
 177         while (total_size(cbuf_cpp, zsbuf_cpp, bin_w, bin_h, gmem) > gmem_size) {
 178                 if (bin_w > bin_h) {
 179                         nbins_x++;
 180                         bin_w = align(width / nbins_x, 32);
 181                 } else {
 182                         nbins_y++;
 183                         bin_h = align(height / nbins_y, 32);
 184                 }
 185         }
 186
 187         DBG("using %d bins of size %dx%d", nbins_x*nbins_y, bin_w, bin_h);
 188
 189         gmem->scissor = *scissor;
 190         memcpy(gmem->cbuf_cpp, cbuf_cpp, sizeof(cbuf_cpp));
 191         memcpy(gmem->zsbuf_cpp, zsbuf_cpp, sizeof(zsbuf_cpp));
 192         gmem->bin_h = bin_h;
 193         gmem->bin_w = bin_w;
 194         gmem->nbins_x = nbins_x;
 195         gmem->nbins_y = nbins_y;
 196         gmem->minx = minx;
 197         gmem->miny = miny;
 198         gmem->width = width;
 199         gmem->height = height;
 200
 201         /*
 202          * Assign tiles and pipes:
 203          *
 204          * At some point it might be worth playing with different
 205          * strategies and seeing if that makes much impact on
 206          * performance.
 207          */
 208
 209 #define div_round_up(v, a)  (((v) + (a) - 1) / (a))
 210         /* figure out number of tiles per pipe: */
 211         tpp_x = tpp_y = 1;
 212         while (div_round_up(nbins_y, tpp_y) > 8)
 213                 tpp_y += 2;
 214         while ((div_round_up(nbins_y, tpp_y) *
 215                         div_round_up(nbins_x, tpp_x)) > 8)
 216                 tpp_x += 1;
 217
 218         /* configure pipes: */
 219         xoff = yoff = 0;
 220         for (i = 0; i < ARRAY_SIZE(ctx->pipe); i++) {
 221                 struct fd_vsc_pipe *pipe = &ctx->pipe[i];
 222
 223                 if (xoff >= nbins_x) {
 224                         xoff = 0;
 225                         yoff += tpp_y;
 226                 }
 227
 228                 if (yoff >= nbins_y) {
 229                         break;
 230                 }
 231
 232                 pipe->x = xoff;
 233                 pipe->y = yoff;
 234                 pipe->w = MIN2(tpp_x, nbins_x - xoff);
 235                 pipe->h = MIN2(tpp_y, nbins_y - yoff);
 236
 237                 xoff += tpp_x;
 238         }
 239
 240         for (; i < ARRAY_SIZE(ctx->pipe); i++) {
 241                 struct fd_vsc_pipe *pipe = &ctx->pipe[i];
 242                 pipe->x = pipe->y = pipe->w = pipe->h = 0;
 243         }
 244
 245 #if 0 /* debug */
 246         printf("%dx%d ... tpp=%dx%d\n", nbins_x, nbins_y, tpp_x, tpp_y);
 247         for (i = 0; i < 8; i++) {
 248                 struct fd_vsc_pipe *pipe = &ctx->pipe[i];
 249                 printf("pipe[%d]: %ux%u @ %u,%u\n", i,
 250                                 pipe->w, pipe->h, pipe->x, pipe->y);
 251         }
 252 #endif
 253
 254         /* configure tiles: */
 255         t = 0;
 256         yoff = miny;
 257         memset(tile_n, 0, sizeof(tile_n));
 258         for (i = 0; i < nbins_y; i++) {
 259                 uint32_t bw, bh;
 260
 261                 xoff = minx;
 262
 263                 /* clip bin height: */
 264                 bh = MIN2(bin_h, miny + height - yoff);
 265
 266                 for (j = 0; j < nbins_x; j++) {
 267                         struct fd_tile *tile = &ctx->tile[t];
 268                         uint32_t p;
 269
 270                         assert(t < ARRAY_SIZE(ctx->tile));
 271
 272                         /* pipe number: */
 273                         p = ((i / tpp_y) * div_round_up(nbins_x, tpp_x)) + (j / tpp_x);
 274
 275                         /* clip bin width: */
 276                         bw = MIN2(bin_w, minx + width - xoff);
 277
 278                         tile->n = tile_n[p]++;
 279                         tile->p = p;
 280                         tile->bin_w = bw;
 281                         tile->bin_h = bh;
 282                         tile->xoff = xoff;
 283                         tile->yoff = yoff;
 284
 285                         t++;
 286
 287                         xoff += bw;
 288                 }
 289
 290                 yoff += bh;
 291         }
 292
 293 #if 0 /* debug */
 294         t = 0;
 295         for (i = 0; i < nbins_y; i++) {
 296                 for (j = 0; j < nbins_x; j++) {
 297                         struct fd_tile *tile = &ctx->tile[t++];
 298                         printf("|p:%u n:%u|", tile->p, tile->n);
 299                 }
 300                 printf("\n");
 301         }
 302 #endif
 303 }
 304
 305 static void
 306 render_tiles(struct fd_batch *batch)
 307 {
 308         struct fd_context *ctx = batch->ctx;
 309         struct fd_gmem_stateobj *gmem = &ctx->gmem;
 310         int i;
 311
 312         ctx->emit_tile_init(batch);
 313
 314         if (batch->restore)
 315                 ctx->stats.batch_restore++;
 316
 317         for (i = 0; i < (gmem->nbins_x * gmem->nbins_y); i++) {
 318                 struct fd_tile *tile = &ctx->tile[i];
 319
 320                 DBG("bin_h=%d, yoff=%d, bin_w=%d, xoff=%d",
 321                         tile->bin_h, tile->yoff, tile->bin_w, tile->xoff);
 322
 323                 ctx->emit_tile_prep(batch, tile);
 324
 325                 if (batch->restore) {
 326                         fd_hw_query_set_stage(ctx, batch->gmem, FD_STAGE_MEM2GMEM);
 327                         ctx->emit_tile_mem2gmem(batch, tile);
 328                         fd_hw_query_set_stage(ctx, batch->gmem, FD_STAGE_NULL);
 329                 }
 330
 331                 ctx->emit_tile_renderprep(batch, tile);
 332
 333                 fd_hw_query_prepare_tile(ctx, i, batch->gmem);
 334
 335                 /* emit IB to drawcmds: */
 336                 ctx->emit_ib(batch->gmem, batch->draw);
 337                 fd_reset_wfi(ctx);
 338
 339                 /* emit gmem2mem to transfer tile back to system memory: */
 340                 fd_hw_query_set_stage(ctx, batch->gmem, FD_STAGE_GMEM2MEM);
 341                 ctx->emit_tile_gmem2mem(batch, tile);
 342                 fd_hw_query_set_stage(ctx, batch->gmem, FD_STAGE_NULL);
 343         }
 344 }
 345
 346 static void
 347 render_sysmem(struct fd_batch *batch)
 348 {
 349         struct fd_context *ctx = batch->ctx;
 350
 351         ctx->emit_sysmem_prep(batch);
 352
 353         fd_hw_query_prepare_tile(ctx, 0, batch->gmem);
 354
 355         /* emit IB to drawcmds: */
 356         ctx->emit_ib(batch->gmem, batch->draw);
 357         fd_reset_wfi(ctx);
 358 }
 359
 360 void
 361 fd_gmem_render_tiles(struct fd_batch *batch)
 362 {
 363         struct fd_context *ctx = batch->ctx;
 364         struct pipe_framebuffer_state *pfb = &batch->framebuffer;
 365         bool sysmem = false;
 366
 367         if (ctx->emit_sysmem_prep) {
 368                 if (batch->cleared || batch->gmem_reason || (batch->num_draws > 5)) {
 369                         DBG("GMEM: cleared=%x, gmem_reason=%x, num_draws=%u",
 370                                 batch->cleared, batch->gmem_reason, batch->num_draws);
 371                 } else if (!(fd_mesa_debug & FD_DBG_NOBYPASS)) {
 372                         sysmem = true;
 373                 }
 374         }
 375
 376         /* close out the draw cmds by making sure any active queries are
 377          * paused:
 378          */
 379         fd_hw_query_set_stage(ctx, batch->draw, FD_STAGE_NULL);
 380
 381         fd_reset_wfi(ctx);
 382
 383         ctx->stats.batch_total++;
 384
 385         if (sysmem) {
 386                 DBG("rendering sysmem (%s/%s)",
 387                         util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 388                         util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 389                 fd_hw_query_prepare(ctx, 1);
 390                 render_sysmem(batch);
 391                 ctx->stats.batch_sysmem++;
 392         } else {
 393                 struct fd_gmem_stateobj *gmem = &ctx->gmem;
 394                 calculate_tiles(batch);
 395                 DBG("rendering %dx%d tiles (%s/%s)", gmem->nbins_x, gmem->nbins_y,
 396                         util_format_short_name(pipe_surface_format(pfb->cbufs[0])),
 397                         util_format_short_name(pipe_surface_format(pfb->zsbuf)));
 398                 fd_hw_query_prepare(ctx, gmem->nbins_x * gmem->nbins_y);
 399                 render_tiles(batch);
 400                 ctx->stats.batch_gmem++;
 401         }
 402
 403         fd_ringbuffer_flush(batch->gmem);
 404
 405         fd_reset_wfi(ctx);
 406
 407         ctx->dirty = ~0;
 408 }
 409
 410 /* tile needs restore if it isn't completely contained within the
 411  * cleared scissor:
 412  */
 413 static bool
 414 skip_restore(struct pipe_scissor_state *scissor, struct fd_tile *tile)
 415 {
 416         unsigned minx = tile->xoff;
 417         unsigned maxx = tile->xoff + tile->bin_w;
 418         unsigned miny = tile->yoff;
 419         unsigned maxy = tile->yoff + tile->bin_h;
 420         return (minx >= scissor->minx) && (maxx <= scissor->maxx) &&
 421                         (miny >= scissor->miny) && (maxy <= scissor->maxy);
 422 }
 423
 424 /* When deciding whether a tile needs mem2gmem, we need to take into
 425  * account the scissor rect(s) that were cleared.  To simplify we only
 426  * consider the last scissor rect for each buffer, since the common
 427  * case would be a single clear.
 428  */
 429 bool
 430 fd_gmem_needs_restore(struct fd_batch *batch, struct fd_tile *tile,
 431                 uint32_t buffers)
 432 {
 433         if (!(batch->restore & buffers))
 434                 return false;
 435
 436         /* if buffers partially cleared, then slow-path to figure out
 437          * if this particular tile needs restoring:
 438          */
 439         if ((buffers & FD_BUFFER_COLOR) &&
 440                         (batch->partial_cleared & FD_BUFFER_COLOR) &&
 441                         skip_restore(&batch->cleared_scissor.color, tile))
 442                 return false;
 443         if ((buffers & FD_BUFFER_DEPTH) &&
 444                         (batch->partial_cleared & FD_BUFFER_DEPTH) &&
 445                         skip_restore(&batch->cleared_scissor.depth, tile))
 446                 return false;
 447         if ((buffers & FD_BUFFER_STENCIL) &&
 448                         (batch->partial_cleared & FD_BUFFER_STENCIL) &&
 449                         skip_restore(&batch->cleared_scissor.stencil, tile))
 450                 return false;
 451
 452         return true;
 453 }